Solving The Schrödinger Equation Has Everything Been Tried

Solving the
Schrdinger
Equation
Has Everything Been Tried?
P780.9781848167247-tp.indd 1 7/25/11 11:49 AM

This page intentionally left blank
Solving the
Schrdinger
Equation
Editor
Paul Popelier
Imperial College Press

ICP
P780.9781848167247-tp.indd 2 7/25/11 11:49 AM

Published by
Imperial College Press
57 Shelton Street
Covent Garden
London WC2H 9HE
Distributed by
World Scientific Publishing Co. Pte. Ltd.
5 Toh Tuck Link, Singapore 596224
USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601
UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.
SOLVING THE SCHRDINGER EQUATION

Copyright 2011 by Imperial College Press
All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means,
electronic or mechanical, including photocopying, recording or any information storage and retrieval
system now known or to be invented, without written permission from the Publisher.
For photocopying of material in this volume, please pay a copying fee through the Copyright
Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to
photocopy is not required from the publisher.
ISBN-13 978-1-84816-724-7
ISBN-10 1-84816-724-5
Typeset by Stallion Press

Email: enquiries@stallionpress.com
Printed in Singapore.
Catherine - Solving the Schrodinger Eqn.pmd 1 9/7/2011, 5:06 PM

July 20, 2011 9:6 9in x 6in b1189-fm Solving the Schrodinger Equation
To D.P.B.
v
The richest interactions occur between two almost identical but

opposing constituents.
vi
Contents
Preface xv
1. Intracule Functional Theory 1

Deborah L. Crittenden and Peter M.W. Gill
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Intracules . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Electron Correlation Models . . . . . . . . . . . . . . 13
1.4 Dynamic and Static Correlation . . . . . . . . . . . . . 16
1.5 Dispersion Energies . . . . . . . . . . . . . . . . . . . 18
1.6 Future Prospects . . . . . . . . . . . . . . . . . . . . . 21
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2. Explicitly Correlated Electronic Structure Theory 25

Frederick R. Manby
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 25
2.1.1 Basis-set expansions . . . . . . . . . . . . . . 25
2.2 F12 Theory . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.1 MP2-F12 . . . . . . . . . . . . . . . . . . . . 29
2.2.2 Explicitly correlated coupled-cluster
theory . . . . . . . . . . . . . . . . . . . . . 30
2.3 Five Thoughts for F12 Theory . . . . . . . . . . . . . . 31
2.3.1 Thought 1: Do we need (products of)
virtuals? . . . . . . . . . . . . . . . . . . . . 31
2.3.2 Thought 2: Are there better two-electron
basis sets? . . . . . . . . . . . . . . . . . . . 34
vii
viii Contents
2.3.3 Thought 3: Do we need the resolution

of the identity? . . . . . . . . . . . . . . . . . 35
2.3.4 Thought 4: Could we have explicit correlation
for higher excitations? . . . . . . . . . . . . . 38
2.3.5 Thought 5: Can we avoid three-electron errors
in two-electron systems? . . . . . . . . . . . . 39
2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . 40
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3. Solving Problems with Strong Correlation Using

the Density Matrix Renormalization Group (DMRG) 43
Garnet Kin-Lic Chan and Sandeep Sharma
3.1 The Problem of Strong Correlation . . . . . . . . . . . 43
3.2 The Density Matrix Renormalization Group
Wavefunction . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Locality and Entanglement in the DMRG . . . . . . . . 47
3.4 Other Properties of the DMRG . . . . . . . . . . . . . 50
3.5 Relation to the Renormalization Group . . . . . . . . . 51
3.6 Dynamic Correlation the Role of Canonical
Transformations . . . . . . . . . . . . . . . . . . . . . 53
3.7 What Can the DMRG Do? A Brief History . . . . . . . 54
3.8 The Future: Higher Dimensional Analogues . . . . . . 57
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4. Reduced-Density-Matrix Theory for Many-electron Correlation 61

David A. Mazziotti
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 62
4.2 Variational 2-RDM Method . . . . . . . . . . . . . . . 63
4.2.1 Energy as a 2-RDM functional . . . . . . . . 63
4.2.2 Positivity conditions . . . . . . . . . . . . . . 64
4.2.3 Semidefinite programming . . . . . . . . . . . 67
4.2.4 Applications . . . . . . . . . . . . . . . . . . 69
4.3 Contracted Schrodinger Theory . . . . . . . . . . . . . 73
4.3.1 ACSE and cumulant reconstruction . . . . . . 74
4.3.2 Solving the ACSE for ground
and excited states . . . . . . . . . . . . . . . . 75
4.3.3 Applications . . . . . . . . . . . . . . . . . . 77
4.4 Parametric 2-RDM Method . . . . . . . . . . . . . . . 80
4.4.1 Parametrization of the 2-RDM . . . . . . . . . 81
Contents ix
4.4.2 Applications . . . . . . . . . . . . . . . . . . 83
4.5 Looking Ahead . . . . . . . . . . . . . . . . . . . . . 85
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5. Finite Size Scaling for Criticality of the Schrodinger Equation 91

Sabre Kais
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 92
5.2 Criticality for Large-dimensional Models . . . . . . . . 93
5.3 Finite Size Scaling: A Brief History . . . . . . . . . . . 95
5.4 Finite Size Scaling for the Schrodinger Equation . . . . 97
5.5 The Hulthen Potential . . . . . . . . . . . . . . . . . . 100
5.5.1 Analytical solution . . . . . . . . . . . . . . . 100
5.5.2 Basis set expansion . . . . . . . . . . . . . . 101
5.5.3 Finite element method . . . . . . . . . . . . . 101
5.5.4 Finite size scaling results . . . . . . . . . . . 102
5.6 Finite Size Scaling and Criticality
of M-electron Atoms . . . . . . . . . . . . . . . . . . . 105
5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . 107
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6. The Generalized Sturmian Method 111

James Avery and John Avery
6.1 Description of the Method . . . . . . . . . . . . . . . . 111
6.1.1 The introduction of Sturmians
into quantum theory . . . . . . . . . . . . . . 111
6.1.2 Generalized Sturmians . . . . . . . . . . . . . 114
6.1.3 The generalized Sturmian method applied
to atoms . . . . . . . . . . . . . . . . . . . . 117
6.1.4 Goscinskian configurations . . . . . . . . . . 118
6.1.5 Goscinskian secular equations for atoms
and atomic ions . . . . . . . . . . . . . . . . 120
6.2 Advantages: Some Illustrative Examples . . . . . . . . 120
6.2.1 The large-Z approximation: restriction
of the basis set to an R-block . . . . . . . . . 121
6.2.2 Validity of the large-Z approximation . . . . . 126
6.2.3 Core ionization energies . . . . . . . . . . . . 129
6.3 Limitations of the Method; Prospects for the Future . . 130
6.3.1 Can the generalized Sturmian method
be applied to N-electron molecules? . . . . . . 133
x Contents
6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . 137

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7. Slater-Type Orbital Basis Sets: Reliable and Rapid

Solution of the Schrodinger Equation for Accurate
Molecular Properties 141
Philip E. Hoggan
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 142
7.1.1 Context of this chapter . . . . . . . . . . . . . 142
7.1.2 Atomic orbitals . . . . . . . . . . . . . . . . . 144
7.1.3 Problems to be solved when using
Slater-type orbitals . . . . . . . . . . . . . . . 145
7.1.4 Strategy for Slater basis sets . . . . . . . . . . 147
7.2 Some Dates: The Story So Far of Slater-Type
Orbitals . . . . . . . . . . . . . . . . . . . . . . . . . 148
7.3 Computer Programs Using Slater-Type Orbitals . . . . 149
7.3.1 Numerical grid methods . . . . . . . . . . . . 151
7.3.2 Configuration interaction . . . . . . . . . . . 151
7.4 Slater Orbitals and Gaussian Orbitals . . . . . . . . . . 151
7.5 Types of Exponentially Decaying Orbitals, Based
on Eigenfunctions for One-Electron Atoms . . . . . . . 154
7.5.1 Orbitals which are linear combinations
of Slater-type orbitals . . . . . . . . . . . . . 155
7.6 Types of Integral Over Slater Orbitals . . . . . . . . . . 156
7.6.1 One-electron integrals . . . . . . . . . . . . . 156
7.6.2 Two-electron integrals . . . . . . . . . . . . . 157
7.6.3 Three- and four-electron integrals . . . . . . . 158
7.7 Integration Methods in the Literature . . . . . . . . . . 158
7.7.1 Single-center expansion . . . . . . . . . . . . 159
7.7.2 Gaussian expansion . . . . . . . . . . . . . . 160
7.7.3 Gaussian transform method . . . . . . . . . . 160
7.7.4 Fourier-transform method . . . . . . . . . . . 160
7.7.5 Use of Sturmians . . . . . . . . . . . . . . . . 161
7.7.6 Elliptic coordinate method . . . . . . . . . . . 161
7.7.7 Monte Carlo integration . . . . . . . . . . . . 162
7.8 General Two-Electron Exponential Type Orbital
Integrals in Poly-Atomics Without Orbital Translations 163
7.8.1 Introduction . . . . . . . . . . . . . . . . . . 163
7.8.2 Basis sets . . . . . . . . . . . . . . . . . . . . 164
Contents xi
7.8.3 Programming strategy . . . . . . . . . . . . . 164

7.8.4 Avoiding ETO translations for two-electron
integrals over three- and four-centers. . . . . . 165
7.8.5 Numerical results of Coulomb resolutions:
efficiency . . . . . . . . . . . . . . . . . . . . 167
7.8.6 Perspectives and conclusions . . . . . . . . . 169
7.8.7 Angular momentum relations . . . . . . . . . 171
7.9 When are Slater-Type Orbitals Advantageous?
Some Applications . . . . . . . . . . . . . . . . . . . . 171
7.9.1 The NMR nuclear shielding tensor . . . . . . 171
7.9.2 Explicitly correlated methods for molecules . 176
7.9.3 Trial wave-functions for quantum Monte Carlo
simulations over STO . . . . . . . . . . . . . 178
7.10 Highly Accurate Calculations Using STOs . . . . . . . 181
7.11 Closing Remarks . . . . . . . . . . . . . . . . . . . . 181
7.12 Appendix A: How STOs were Translated: Products on
Two Atoms . . . . . . . . . . . . . . . . . . . . . . . . 183
7.12.1 Review of BCLFs . . . . . . . . . . . . . . . 184
7.13 Appendix B: Brief Time-Line of Events in Molecular
Work Over Slater-Type Orbitals to Date . . . . . . . . 186
7.14 Appendix C: Main Results of Podolanskis Paper
of 1931 with Additional Comments . . . . . . . . . . . 187
7.15 Appendix D: Potentials and Auxiliary Overlaps
for Coulomb Resolution . . . . . . . . . . . . . . . . . 189
7.16 Appendix E: Analysis of Nuclear Dipole Integrals
for NMR in a Slater Basis . . . . . . . . . . . . . . . . 191
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . 193
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 194
8. Modern Ab Initio Valence Bond Methods 201

Philippe C. Hiberty and Sason Shaik
8.1 Basic Principles and Survey of Modern Methods . . . . 202

8.1.1 VB vs. MO wave functions in the two-electron/
two-center case . . . . . . . . . . . . . . . . . 202
8.1.2 WritingVB functions beyond the two-electron/
two-center case . . . . . . . . . . . . . . . . . 205
8.1.3 Some landmark improvements of the early
VB method . . . . . . . . . . . . . . . . . . . 206
xii Contents
8.2 Strengths of the Valence Bond Approach . . . . . . . . 211

8.2.1 Interpretability combined with accuracy
of the wave functions . . . . . . . . . . . . . 211
8.2.2 A simple solution to the symmetry dilemma . 213
8.2.3 Calculations of diabatic energy curves along
a reaction coordinate . . . . . . . . . . . . . . 214
8.2.4 Quantitative evaluation of common chemical
paradigms . . . . . . . . . . . . . . . . . . . 218
8.3 Present Capabilities and Expected Improvements . . . . 224
8.3.1 Evaluation of Hamiltonian matrix elements . . 224
8.3.2 Direct VBSCF/BOVB algorithm . . . . . . . . 224
8.3.3 Current calculations of medium-sized
molecular systems . . . . . . . . . . . . . . . 225
8.3.4 Mixed Valence Bond Quantum Monte
Carlo methods . . . . . . . . . . . . . . . . . 225
8.3.5 Prospective . . . . . . . . . . . . . . . . . . . 226
8.4 Concluding Remarks . . . . . . . . . . . . . . . . . . 228
8.5 Appendix A: The Myth of VB failures . . . . . . . . 229
8.6 Appendix B: Some Available VB
Software Packages . . . . . . . . . . . . . . . . . . . . 230
8.6.1 The XMVB program . . . . . . . . . . . . . . 230
8.6.2 The TURTLE software . . . . . . . . . . . . . 230
8.6.3 The VB2000 software . . . . . . . . . . . . . 230
8.6.4 The CRUNCH software . . . . . . . . . . . . 231
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 231
9. Quantum Monte Carlo Approaches for Tackling

Electronic Correlation 237
Massimo Mella and Gabriele Morosi
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 238
9.2 Variational Monte Carlo (VMC): A Possible
Way Toward Explicitly Correlated Electronic
Wave Functions . . . . . . . . . . . . . . . . . . . . . 239
9.2.1 Numerical integrals in VMC . . . . . . . . . . 241
9.2.2 Optimization of trial wave functions . . . . . . 247
9.2.3 Analytical forms for trial wave
functions T . . . . . . . . . . . . . . . . . . 252
9.3 Diffusion Monte Carlo: How to Extract the Best
Information from Inaccurate Wave Functions . . . . . . 254
Contents xiii
9.3.1 Generalities . . . . . . . . . . . . . . . . . . 254

9.3.2 Improved projectors . . . . . . . . . . . . . . 258
9.3.3 DMC, state symmetry and excited states . . . 259
9.4 Computing Observables Different from State Energy . . 261
9.4.1 Exact calculation of position dependent
observables . . . . . . . . . . . . . . . . . . . 261
9.4.2 Calculation of atomic forces in VMC/DMC . . 262
9.4.3 Computing the expectation value of ultra-local
operators: electron and spin density
on nuclei . . . . . . . . . . . . . . . . . . . . 264
9.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . 266
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 268
10. Solving the Schrodinger Equation on Real-Space Grids

and with Random Walks 271
Thomas L. Beck and Joel H. Dedrick
10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 272
10.2 Solving the Schrodinger Equation Using Grids in Real
Space . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
10.2.1 Basics of grid methods . . . . . . . . . . . . . 275
10.2.2 Multiscale (multigrid) approaches . . . . . . . 279
10.3 New Ways of Thinking about Large-Scale
Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 282
10.3.1 An encounter with Silicon valley . . . . . . . 283
10.3.2 The Borg-ing of computation . . . . . . . . . 283
10.3.3 The worlds least efficient computer (yours) . . 283
10.3.4 The worlds most efficient computer
(also yours) . . . . . . . . . . . . . . . . . . . 284
10.3.5 The end of a 20-year free ride . . . . . . . . . 285
10.3.6 Can we change sides? . . . . . . . . . . . . . 286
10.3.7 Algorithm desiderata for the massively parallel
future . . . . . . . . . . . . . . . . . . . . . . 286
10.3.8 What are we looking for? . . . . . . . . . . . 288
10.4 Random Walks for Solving the Schrodinger
Equation . . . . . . . . . . . . . . . . . . . . . . . . . 292
10.4.1 Traditional diffusion quantum Monte Carlo . . 292
10.4.2 Another angle . . . . . . . . . . . . . . . . . 296
10.4.3 Stochastic differential equations and the
FeynmanKac approach . . . . . . . . . . . . 298
xiv Contents
10.4.4 Obtaining E(x,y)? . . . . . . . . . . . . . . . 302

10.4.5 A pipe dream . . . . . . . . . . . . . . . . . . 305
10.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . 306
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 307
11. Changes in Dense Linear Algebra Kernels: Decades-

Long Perspective 313
Piotr Luszczek, Jakub Kurzak, and Jack Dongarra
11.1 The Schrodinger Connection . . . . . . . . . . . . . . 313
11.2 A Stroll Down the Memory Lane . . . . . . . . . . . . 315
11.3 A Decompositional Approach . . . . . . . . . . . . . . 318
11.4 Vector Processors . . . . . . . . . . . . . . . . . . . . 319
11.5 RISC Processors . . . . . . . . . . . . . . . . . . . . . 322
11.6 Clusters . . . . . . . . . . . . . . . . . . . . . . . . . 325
11.7 Multicore Processors . . . . . . . . . . . . . . . . . . 332
11.8 Multicore Processors Redux . . . . . . . . . . . . . . . 334
11.9 Error Analysis and Operation Count . . . . . . . . . . 338
11.10 Future Directions for Research and Hardware Design . 339
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Index 343
Preface
In the first of his four lectures on wave mechanics, Schrodinger wrote:
Substituting from (12) and (8) in (10) and replacing p by (. . . )

we obtain
82 m
2 + (E V) = 0 (13)
h2
(. . . ) A simplification in the problem of the mechanical waves
consists in the absence of boundary conditions. I thought the
latter simplification fatal when I first attacked these equa-
tions. Being insufficiently versed in mathematics, I could not
imagine how proper vibration frequencies could appear without
boundary conditions. Later on I recognised that the more com-
plicated form of the coefficients (i.e. the appearance of V(x,y,z))
takes charge, so to speak, of what is ordinarily brought about
by boundary conditions, namely, the selection of definite values
of E.
(Dr. Erwin Schrodinger Four Lectures on Wave Mechanics.
Delivered at the Royal Institution, London, on 5, 7, and 14 March,
1928)
Here we are: equation number 13 arrived in Knowledge Space and

is there to stay. A simple looking equation, one of many in Schrodingers
paper, and one that can be written even simpler as H = E. Although
short in appearance, this equation is immensely powerful. It governs the
behaviour and properties of all matter at ambient conditions. It is, therefore,
the mother equation of chemistry. Who could have thought, though, that
some 85 years later, people are still thinking about solving it faster and more
accurately? Are the standard methods to solve the molecular electronic
Schrodinger equation, which are implemented in popular commercial ab
xv
xvi Preface
initio programs, the best we can achieve? Or is there a new and powerful idea
lurking at the surface of Knowledge Space, which leads to a better method,
more accurate and faster, and independent from (chemical) experiment? Is
this new idea based on the combination of two or more existing ideas? This
book asks these difficult and ambitious questions to its contributing authors
and to the reader.
This book invited its authors to elucidate the non-standard method that
they specialised in, explain its strength and weakness, and then speculate
about what is needed to widen the application radius of the method. Actually
achieving this may take years and involve several people. This book hopes to
inspire readers and researchers by putting non-standard approaches together
in one place. I believe that this has never been done. The format and style
in which the chapters are written should make it possible to read the whole
book through. It should be emphasised that this text was not designed as a
review. Instead, it is meant to be a collection of personal accounts capturing
the aspiration and perhaps frustration of experts of non-standard methods.
So, what can we learn from Schrodingers aforementioned quote, other
than that it is good to catch up with a sufficient amount of mathematics (or
spend Christmas holidays frolicking with an old flame in a mountain resort,
where he discovered his equation)? One lesson is to trust the potential of an
idea, often based on an analogy or a vivid picture. In fact, in Schrodingers
case this was the symbolic proportion:
Ordinary mechanics : Wave mechanics
= Geometrical optics : Undulatory optics.
His derivation of Eq. (13), developed in the pages leading up to the
excerpt above, is based on this analogy. Schrodinger could describe what
his new wave mechanics would look like based on this intuitive analogy.
As he worked out the maths behind this intuitive development he panicked
for a moment, due to his lack of mathematical knowledge1 . Fortunately, he
ended up with an equation that worked. Moreover, Schrodinger presented
quantum mechanics with a completely new formalism, dual to the older
matrix mechanics, which Heisenberg had proposed. The latter, and other
members of the Copenhagen clan, did not like wave mechanics much:
it was too intuitive and not as elegant and deep as matrix mechanics.
Matrices were of course wonderfully abstract mathematical entities to the
theoretical physicists of that generation. However, this abstraction did not
endow matrix mechanics with any authority over wave mechanics unless
1 Quantum Mechanics textbooks typically gloss over this concern. Is there something deeper in the
reassurance of V(x, y, z) acting as a boundary condition?
Preface xvii
by a vague sense of mysterious elegance. To the contrary, matrix mechanics

has remained rather useless to chemistry; solving the hydrogen atom by this
method took Pauli pages and pages of calculation.
Returning to trusting the potential of an idea, this trust was also present
in the development of Valence Bond and Molecular Orbital theories. Both
describe the same reality, but from diametrically opposed views. Of utmost
importance here is the confidence that a visual and intuitive approach
to a problem, or even thinking about it in words, can be an excellent
starting point for a derivation that leads to a new and powerful com-
putational scheme. It would be unnatural to derive Valence Bond theory
from Molecular Bond theory, or vice versa. Of course, one can formally
link the two, as one can do with matrix mechanics and wave mechanics.
The point is that equivalent but very different approaches originate from
different ideas, each based on an insight originally conceived as an image
or stated in words. Armed with this confidence in the potential of a fresh
idea one should ask: have we been imaginative enough in our intuition,
pictures and words, when trying to solve the Schrodinger equation?
Speaking of imagination, one cannot emphasise enough the importance
of thinking outside the box or what I call let go thinking. One could
object to taking a square root of a negative number or, instead, embrace
this idea, give the result a name, work with it and derive the full conse-
quences. Algebra has been helped tremendously by the let go concept of
an imaginary number. Of course, thinking outside the box is not a guar-
antee for success or a breakthrough. Quaternions, for example, are even
more imaginative than imaginary numbers but they have not lived up to
the rich expectations that their originator, Hamilton, fostered. Even less
enthused in terms of application or deep insight should one be when it
comes to octonions, or non-integer derivatives.
Still, the very attitude of let go is powerful when it comes to letting go
of a deeply ingrained assumption. An assumption that one cannot justify
or that one does not even realise one is making. A nice example can be
found in the planetary motion of our solar system. The Greeks revered
symmetry, to the point of adhering religious connotations to it. Symmetry
had a transcendental or even divine air. Invariance or constancy can also be
regarded as a form of symmetry. Surely the motion of celestial bodies must
be governed by symmetry, given the divine character of heavenly objects.
Ptolemaic astronomy (that had entered Greek astronomy) is a scheme in
which a planet moves in a circle and at constant motion, both, indeed,
signatures of symmetry. In Copernicus time this model became horribly
complicated in order to explain the large number of observations gathered
xviii Preface
by the Arabs since the ancient Greek times. The Ptolemaic theory needed
77 circles to describe the motion of the sun, moon, and the five planets
then known. Kepler broke with the tradition of 2,000 years, that circles
must be used to describe heavenly motions. He showed that a single ellipse
would do. An ellipse is not as symmetrical as a circle, and therefore not as
heavenly. A circle can be seen as an ellipse in which the two foci have
collapsed to one (and hence the two radii as well). Actually, in a deeper
way, an ellipse is more heavenly than a circle because it captures Nature in
a minimal model. In such a model there is no need for corrections within
corrections. All falls in place by letting go of the constraint that a planet
must move in a circular orbit. I now wonder where our circles are in
quantum chemistry. Which constraints are we holding on to?
This book invites its authors and the readers to abandon the usual lines
of thoughts and presumptions that we do not perhaps realise we are making.
The most powerful theories are minimal, not simple. Simple means that
we impose an unwarranted constraint onto what we are trying to explain.
Minimal means that we discovered the most essential, but unconstrained
concept that governs the observed data of interest. This economy of prin-
ciple or assumption always pays off, but obtaining a minimal theory requires
much imagination and audacity.
Returning to quantum chemistry, one may have the impression that
the only truly predictive computational schemes are built on brute force
foundations. The core idea behind configuration interaction is brute force
in nature. The explosion in computational work it leads to warrants clever
but inevitably approximate computational schemes. This is only vaguely
reminiscent of the combinatorial explosion encountered in the calculation
of a determinant of a large matrix by the definition of an alternating sum of
permuted terms. This idea leads to intractable calculations for a matrix as
small as a 30 30 matrix, for example. Yet calculating such a determinant
is perfectly feasible with LU decomposition in most reasonable times. This
is where the power of the idea makes apparently impossible tasks possible
after all. Then we can ask again: why can we not think of a method to solve
the Schrodinger equation of a protein beyond Coupled-Cluster Singles and
Doubles (Triples)/ Complete Basis Set (CCSD(T)/CBS) quality in a few
seconds?
Regarding the content of this book, one can see that there are eleven
chapters, covering ten ideas (or methods) not prevalent in current main-
stream quantum chemistry. Unfortunately, some methods are not included,
due to a lack of available authors. Otherwise, there would have been
extra chapters on Bohmian mechanics, the series solution method, the
Preface xix
pseudospectral method and the iterative configuration interaction method.

Another, perhaps more striking omission is that of Density Functional
Theory (DFT). The only chapter on DFT in the spirit of this book would
have to explain fresh ideas on finding the ultimate exact functional. Much
has been written on the theoretical conditions that this functional should
obey. Ironically, the most successful functionals in current use seem to
violate one of more of these conditions. If that is, to some readers, a reason
to have included a fresh chapter on DFT then I apologise. If this book ever
reaches a second edition a chapter on DFT should be considered, as well
as a recorded roundtable discussion on the various methods presented in it,
as in the book Quo vadis Quantum Mechanics?
I am very grateful, indeed, for the hard work and enthusiasm that the
authors have put in this book. Without them, this book would not be possible.
Thank you all and may you not buckle under the pressures imposed by
ignorant politicians, policy-makers and managers, serving an ever emptier
and short-term society. Being driven by good old fashioned curiosity alone
should be respectable again.
Finally, Id like to quote from a book I recently read. At the very end
of his book entitled The Trouble with Physics on the rise of string theory,
the fall of a science and what comes next, Lee Smolin concludes, . . . I am
going to turn off the phone and the BlackBerry, put on [. . . my favourite
music . . . ], erase the blackboard, get out some good chalk, open a new
notebook, take out my favourite pen, sit down, and start thinking. One
can query if modern electronic structure theory is in the same situation
as this part of physics, which pursues a grand unified theory. But even if
the situation is vaguely similar, many of us should buy some good chalk
also. I only hope that reading this book in full will help, as well as the idea
exchange that it hopes to catalyse, one day . . .
Paul Popelier
Manchester, 11 September 2010

July 19, 2011 11:28 9in x 6in b1189-ch01 Solving the Schrodinger Equation
Chapter 1
Intracule Functional Theory
Deborah L. Crittenden and Peter M.W. Gill
Research School of Chemistry,

Australian National University, Canberra ACT 0200, Australia
Density functional theory (DFT) has become the most popular by far
of the panoply of methods in quantum chemistry and the reason for
this is simple. Where other schemes had become bogged down in mind-
numbingly expensive and detailed treatments of the electron correlation
problem, DFT simply shrugged, pointed at the HohenbergKohn theorem,
and asserted that the correlation energy can be written as an integral of a
certain function of the one-electron density. The only thing that irritated
the wavefunction people more than the cavalier arrogance of that assertion
was the astonishing accuracy of the energies that it yields.
Well, most of the time. Occasionally, DFT fails miserably and,
although the reasons for its lapses are now understood rather well, it
remains a major challenge to correct these fundamental deficiencies, while
retaining the winsome one-electron foundation upon which DFT rests.
Does this mean that, for truly foolproof results, we have no option but
to return to the bog of many-body theory? One might think so, at least
from a cursory inspection of the current textbooks. But we feel differently,
and in this chapter we present an overview of an attractive alternative that
lies neither in the one-electron world of DFT, nor in the many-electron
world of coupled-cluster theory. Our approach nestles in the two-electron
Fertile Crescent that bridges these extremes, a largely unexplored land
that would undoubtedly have been Goldilocks choice.
We present results that demonstrate that the new approach Intracule
Functional Theory is capable of predicting the correlation energies of
small molecules with an accuracy that rivals that of much more expensive
post-HartreeFock schemes. We also show that it easily and naturally
models van der Waals dispersion energies. However, we also show that
1
2 D.L. Crittenden and P.M.W. Gill
its current versions struggle to capture static correlation energies and that
this is an important area for future development.
Finally, we peer into the probable future of the field, speculating on the
directions in which we and others are likely to take it. We conclude that,
although the approach is conceptually attractive and has shown consid-
erable promise, the investigations hitherto have scarcely scratched the sur-
face and there are ample opportunities for fresh ideas from creative
minds.
1.1. Introduction
In the late 1920s, Hartree [1] was among the first to realize that the newly
derived Schrodinger equation [2] describing quantum electronic motion
could be solved for multi-particle systems if the wavefunction, a com-
plicated multidimensional object that explicitly couples the motion of all
particles in the system, is approximated by a product
(r1 , r2 , . . . , rn ) = 1 (r1 )2 (r2 ) n (rn ) (1.1)
of single-particle functions (spin-orbitals). Physically, the Hartree wave-
function implies that each electron moves independently in the electro-
static field created by all of the others. Shortly thereafter, both Slater [3]
and Fock [4] pointed out that Hartrees wavefunction lacks the antisym-
metry required by the Pauli Principle [5], but that this can be rectified by
adopting the determinant form

1 (r1 ) 2 (r1 ) . . . n (r1 )

1 (r2 ) 2 (r2 ) . . . n (r2 )

(r1 , r2 , . . . , rn ) = .. .. . .. . (1.2)
. . . . .

1 (rn ) 2 (rn ) n (rn )
Unfortunately, the resulting HartreeFock (HF) model neglects the inter-

electron correlations that influence chemically important phenomena such
as bond making and breaking, electron gain and loss, and the response of
a molecule to an external electric and/or magnetic field. For example, in
the homolytic fission of a single bond, the two formerly paired electrons
migrate in opposite directions and this cannot be accurately described by a
single determinant.
The difference between a molecules HF energy and its exact energy is
Ec , the correlation energy, and the challenge of its determination is known
Intracule Functional Theory 3
as the electron correlation problem and has been the focus of ongoing
research efforts for almost a century. Currently, methods for recovering Ec
fall into two broad classes.
Wavefunction-based methods are based upon the mathematical obser-
vation that an improved wavefunction can be constructed from the occupied
and unoccupied orbitals that arise from solving the HF equations. These
methods are guaranteed eventually to converge to the exact result, but
their convergence is hampered because they are effectively approximating
cusps in the true wavefunction by sums of smooth functions. In practice,
wavefunction-based post-HF methods are typically limited in applicability
to systems containing a few dozen non-hydrogen atoms.
Density-based methods are a popular low-cost alternative. They are
based upon the HohenbergKohn theorem [6], which states that the energy
of the ground state of a system is a universal functional of its electron density
(r). Unfortunately, the theorem gives little insight into the construction
of the functional and, despite the efforts of many researchers over many
years, its form remains unknown. Many approximate functionals have been
devised, each with its own strengths and weaknesses, but none yet has
proven accurate for all types of chemical problems. The major systematic
weaknesses [7] of density functional theory (DFT) stem from its inability
to deal with intrinsically two-electron phenomena such as bond cleavage
and static correlation.
Comparing these two alternatives wavefunction-based and density-
based models reveals a vast and largely unexplored intermediate ground
between the complexity of wavefunction schemes (which depend explicitly
on the coordinates of every electron) and the simplicity of density schemes
(which depend only on the one-electron density). The most obvious entry
point and this is our present strategy is to develop approaches that
incorporate two-electron information but retain the computational advan-
tages enjoyed by DFT. We will use atomic units throughout.
1.2. Intracules
A reasonable starting point for the development of a two-electron analogue

of DFT is the two-electron density

2 (r1 , r2 ) = |(r1 , . . . , rn )|2 dr3 . . . drn , (1.3)
which gives the joint probability of finding one electron at r1 and another at
r2 . How might one extract the correlation energy from this six-dimensional
object? Intuitively, one may expect the statistical correlation between the
motions of two electrons to depend strongly on their separation and this
leads naturally to the position intracule [8]

P(u) = 2 (r1 , r2 )(r12 u)dr1 dr2 , (1.4)
(where is the Dirac delta distribution and r12 |r12 | |r1 r2 |)

which gives the probability density of finding two electrons separated by a
distance u.
Example 1.1: The Position Intracule for a He-like Ion

In most modern calculations, the molecular orbitals (MOs) are expanded in a
basis of Gaussian functions. If we model the 1s orbital in a He-like ion by the
single Gaussian exp(r 2 ), the HF wavefunction is
(r1 , r2 ) = (2/)3/2 exp[(r12 + r22 )]
and one finds from Eq. (1.4) that the position intracule is

P(u) = 2 (r1 , r2 )(r12 u)dr1 dr2

= (2/) 3
exp[2(r12 + r22 )](r12 u)dr1 dr2

= (2/)3 exp[2(r12 + |r1 u|2 )] dr1 du ,
by writing r2 = r1 u

= (/)3/2 exp(u2 )du ,
by integrating over r1
= (/)3/2 4u2 exp(u2 ), by integrating over the angular part of u
As the Gaussian exponent increases, the 1s orbital shrinks and P(u) contracts
toward the origin.
The Gaussian

exponent that minimizes the HF energy of the He atom is
= 338 9
2
and the resulting intracule reveals that the electrons are most
likely to be found with a separation u 1 but that they are unlikely to be
found close to one another (u 0) or at large separations (u > 3).
Continued
P(u)
1
0.8
0.6
0.4
0.2
u
0.5 1 1.5 2 2.5 3 3.5 4
Fig. 1.1. Position intracule for the helium atom.
Unfortunately, although it is easy enough to compute P(u), even in large

molecules [9], it is not an optimal source of two-electron information for
electron correlation. This can be demonstrated by considering two-electron
systems that are confined by a tunable parameter. For example, in the
helium-like ions (i.e. H , He, Li+ , . . . ), where the confinement parameter
is the nuclear charge Z, the exact [10] and HF [11] energies are
5
Eexact = Z2 + Z 0.15767 + O(Z1 ) (1.5)
8
5
EHF = Z2 + Z 0.11100 + O(Z1 ) (1.6)
8
and therefore, as the confinement parameter grows, the correlation energy
Ec = Eexact EHF approaches a limiting value (46.67 mEh ). Anal-
ogous behaviour is found in other such systems [12] and, indeed, it can be
proven [13] that the correlation energy of two electrons always approaches
a limiting value when they are confined to an infinitesimal volume. This
constancy contrasts sharply with the behaviour of the position intracule
P(u), which approaches a delta distribution at u = 0 as the two electrons
are sequeezed closer together. Such analysis indicates that P(u) does not
possess the qualitative behaviour required to capture Ec in these simple
systems.
Example 1.2: The Momentum Intracule for a He-like Ion

Modelling the 1s orbital in a He-like ion by the Gaussian exp(r2 ) yields
the HF momentum wavefunction
(p1 , p2 ) = (2)3/2 exp[(p21 + p22 )/4]
Continued
and one finds from Eq. (1.7) that the momentum intracule

M(v) = 2 (p1 , p2 )(p12 v)dp1 dp2

= (2)3 exp[(p21 + p22 )/2](p12 v)dp1 dp2

= (2)3 exp[(p21 + |p1 v|2 )/2]dp1 dv ,
by writing p2 = p1 v

= (4)3/2 exp(v2 /4)dv ,
by integrating over p1
3/2
= (4) 2
4v exp(v /4),2
by integrating over the angular part of v

is a Maxwell distribution. As increases, high relative momenta become more
likely and M(v) broadens.
Using the energy-minimizing exponent = 338 9
2
yields the momentum
intracule which reveals that the electrons are most likely to be moving
with a relative momentum v 2 but that they are unlikely to have
very similar momenta (v 0) or very different momenta (v > 6).
M(v)
0.5
0.4
0.3
0.2
0.1
v
1 2 3 4 5 6 7 8
Fig. 1.2. Momentum intracule for the helium atom.
A decade ago, Rassolov observed that the relative momentum p12

|p12 | |p1 p2 | also plays a role in electron correlation [14]. Intuitively,
this makes sense: high relative velocities reduce interaction times for the
electrons and therefore reduce the extent of their correlation. This infor-
mation is captured by the momentum intracule [15, 16]

M(v) = 2 (p1 , p2 )(p12 v) dp1 dp2 , (1.7)
where the two-electron momentum density is

2 (p1 , p2 ) = |(p1 , . . . , pn )|2 dp3 . . . dpn (1.8)
and the momentum wavefunction

3n/2
(p1 , . . . , pn ) = (2) (r1 , . . . , rn )ei(p1 r1 ++pn rn ) dr1 . . . drn
(1.9)
is the Fourier transform of the position-space wavefunction. Unfortunately,
the momentum intracules of the He-like ions become flatter as Z increases,
suggesting that like P(u) but for the opposite reason M(v) is not an
optimal source of two-electron information for correlation. However, the
opposing trends in the position and momentum intracules suggest that we
may be able to model correlation through a product variable involving both
r12 and p12 .
Yet again, however, it seems that we are thwarted because, although
one can form a momentum-space wavefunction from its position-space
counterpart using a Fourier transform, the Heisenberg Uncertainty Principle
forbids the construction of a joint phase-space wavefunction. Likewise,
although 2 (r1 , r2 ) and 2 (p1 , p2 ) can be easily obtained, there exists no
comparable joint probability density P(r1 , r2 , p1 , p2 ).
Nevertheless, although a genuine density in phase-space is prohibited,
it is possible to concoct quasi-densities with some of the properties that the
genuine article would possess. The most famous of these are the Wigner
quasi-densities [17]
Wn (r1 , . . . , rn , p1 , . . . , pn )

1
= 3n (r1 + q1 , . . . , rn + qn )

(r1 q1 , . . . , rn qn )e2i(p1 q1 ++pn qn ) dq1 . . . dqn (1.10)
and the Husimi quasi-densities [18]. Besley [19] has studied the latter but
we will confine our attention here to the former.
Being a function of 6n coordinates, the full Wigner quasi-density is
even more complicated than the wavefunction. However, because we are
primarily interested in two-electron information, it is natural to integrate

over all but two of the electrons to use instead the second-order reduced
Wigner quasi-density
W2 (r1 , r2 , p1 , p2 )

1
= 6 2 (r1 + q1 , r2 + q2 , r1 q1 , r2 q2 )e2i(p1 q1 +p2 q2 ) dq1 dq2 ,

(1.11)
where 2 is the reduced second-order density matrix [20]. W2 is a simpler
object than Wn but it is nonetheless a function of 12 variables and is con-
ceptually formidable. Ideally, we would like to extract from it only the
information that is directly relevant to a description of electron corre-
lation. It obviously contains information about the relative position r12
and momentum p12 variables but it also knows about the dynamical angle
12 between the vectors r12 and p12 , giving insight into the nature of the
electrons mutual orbit, as illustrated below.
By analogy with Eqs. (1.4) and (1.7), we can extract the quasi-density
for r12 , p12 and 12 to form the Omega intracule [21]
(u, v, )

= W2 (r1 , r2 , p1 , p2 ) (r12 u) (p12 v) (12 ) dr1 dr2 dp1 dp2

1
= 2 (r1 + q1 , r2 + q2 , r1 q1 , r2 q2 )e2i(p1 q1 +p2 q2 )
6
(r12 u)(p12 v)(12 )dq1 dq2 dr1 dr2 dp1 dp2

1
= 2 (r, r + q + u, r + q, r + u)eivq (uv ) drdqdu dv ,
83
(1.12)
where, as before, u and v are arbitrary vectors of length u and v, respectively,
and uv is the angle between them. At this point, things do not look very
12 = 0 12 = /2 12 =
Fig. 1.3. Physical interpretation of the dynamical angle 12 .

practical, for the Omega intracule is written as a ten-dimensional integral

over the second-order density matrix. However, as we will see in a moment,
things become much more tractable after we introduce a Gaussian basis set.
The Omega intracule is the grandfather of a family of lower-dimensional
intracules and each of its descendants is obtained by appropriate integration.
This is illustrated diagrammatically below where we also introduce s =
r12 p12 and x = r12 p12 . Both of these variables have dimensions of
angular momentum and units of Plancks constant.
If the MOs are expanded in a basis set {i }, the reduced second-order
density matrix is

2 (r1 , r2 , r1 , r2 ) = abcd a (r1 )b (r2 )c (r1 )d (r2 ), (1.13)
abcd
where the abcd are two-particle density matrix (2PDM) elements. Thus,
from Eq. (1.12), the Omega intracule is

(u, v, ) = abcd [abcd] , (1.14)
abcd
where the Omega integrals are

1
[abcd] = a (r)b (r + q + u)c (r + q)d (r + u)
8 3
eivq (uv ) drdqdu dv . (1.15)
Fig. 1.4. The intracule family tree.

If the i are s-type Gaussians centred at A, B, C and D, with exponents

, , and , it can be shown [22] that

1
[ssss] = K exp(R) i0 ( x + y cos t)dt, (1.16)
0
where
2 u2 v2 sin
K= exp(2 u2 2 v2 iuv cos ) (1.17)
( + )3/2 ( + )3/2

R= |A D|2 + |B C|2 (1.18)
+ +
x = (Pu)2 + (iQv)2 + 2(Pu)(iQv) cos cos
(1.19)
y = 2(Pu)(iQv) sin sin
1 1
2 = + 42 = + =
+ + + + + +
(1.20)
2 2
P= (A D) + (B C)
+ +
(1.21)
A + D B + C
Q= P Q = PQ cos
+ +
and i0 (z) = z1 sinh z. The integral in Eq. (1.16) can be evaluated by qua-
drature or series expansion [22]. However, if the Gaussian centres are colli-
near, it can be found in closed form and, if they are concentric, it reduces to
[ssss] = K. (1.22)
Integrals over the p, d, . . . basis functions may be obtained by systematic
Boys differentiation [23] of the [ssss] integral. However, it is more efficient
to use recursion and a 18-term recurrence relation has been developed for
this purpose [24].
Example 1.3: The Intracule Family for a He-like Ion

As in the previous examples, the HF wavefunction of a He-like ion in the basis
of a single Gaussian is
(r1 , r2 ) = (2/)3/2 exp[(r12 + r22 )]
and one finds from Eq. (1.22) that the Omega intracule is
(u, v, ) = (1/)u2 exp(u2 ) v2 exp(v2 /4) sin .
Continued
In this simple system, we find that (u, v, ) is proportional to the product of

P(u) and M(v), implying that u and v are statistically independent. However,
such systems are the exception, not the rule.
The lower intracules can be constructed easily from (u, v, ), as shown
below:
Intracule Construction Explicit form

W(u, v) 0 (u, v, )d (2/)u2 exp(u2 )v2 exp(v2 /4)

(s, ) 0 (u, s/u, )u1 du (1/)s2 K0 (s) sin

P(u) 0 W(u, v)dv (/)3/2 4u2 exp(u2 )

M(v) 0 W(u, v)du (4)3/2 4v2 exp(v2 /4)

A(s) 0 (s, )d (2/)s2 K0 (s)

D(x) x (s, )(s sin )1 ds (1/)xK1 (x)

() 0 (s, )ds (1/2) sin
Here K0 and K1 are modified Bessel functions of the second kind [25].
Each of the three one-dimensional intracules, A(s), D(x) and (), whose
graphs are shown below, is independent of the exponent , that is, they
are invariant with respect to dilation. As such, they apply not only to the
helium atom but, equally, to any helium-like ion. This will be important in
Section 1.3.
The attentive reader may wonder why, if u and v are statistically independent
in this system, the angle intracule () is not constant. After all, if the relative
positions and momenta of the two electrons are independent, one might have
expected the angle between r12 and p12 to be equally likely to take any value
between 0 and . The fact that this is not the case is a purely geometrical
(Jacobian) effect: as r12 and p12 range independently over their respective
domains, dynamical angles 12 close to /2 arise far more often than angles
close to 0 or . The fact that there are many more points on the Earths surface
with latitudes near 0 (equatorial regions) than with latitudes near 90 (polar
regions) arises from the same geometrical effect.
A(s) D(x) ()
0.5
0.3 0.3
0.25 0.25 0.4
0.2 0.2 0.3
0.15 0.15
0.2
0.1 0.1
0.05 0.05 0.1
s x
2 4 6 8 10 6 4 2 2 4 6 0.5 1 1.5 2 2.5 3
Fig. 1.5. Action, dot and angle intracules for a He-like ion.
Example 1.4: The Wigner Intracule for the Be Atom

The STO3G basis for the Be atom consists of two three-fold contracted
Gaussian-type functions. Total contraction coefficient arrays for the 1s and
2s molecular orbitals (C1 and C2 , respectively) are calculated by mul-
tiplying the normalized contraction coefficients by the appropriate MO
coefficients.
Gaussian Contraction 1s MO 2s MO 1s array 2s array

Index exponent coefficient coefficient coefficient C1 C2
1 30.167871 1.4158460 1.4057909 0.4160898

2 5.495115 1.3693446 0.9928982 0.2938807 1.3596198 0.4024239
3 1.487193 0.4267649 0.4237341 0.1254180
4 1.3148331 0.08747241 0.0022866 0.0905572
5 0.3055389 0.11701478 0.0261377 1.0351471 0.0030585 0.1211275
6 0.09937070 0.088312776 0.0023083 0.0914167
The HF two-particle density matrix elements can be constructed using
abcd = Ca1 Cb1 Cc1 Cd1 + 4Ca1 Cb1 Cc2 Cd2 + Ca2 Cb2 Cc2 Cd2 2Ca1 Cb2 Cc2 Cd1 ,
where Cak denotes the ath element of the Ck array. The Wigner intracule is
then assembled through

6
6
6
6
W(u, v) = abcd [abcd]W
a=1 b=1 c=1 d=1
and it is illustrated in the following contour plot:

This intracule possesses three maxima. The first, near (u, v) (0.7, 3),
describes electrons that are close together and moving fast; it arises from
observing the two 1s electrons. The second, near (u, v) (2.5, 0.8), describes
electrons that are well-separated and moving relatively slowly; it arises from
observing the two 2s electrons. The third, near (u, v) (2, 2), describes
electrons that are moderately far apart and moving at a moderate pace; it arises
from observing a 1s electron and a 2s electron. The third maximum is the
largest because there are 4 C2 = 6 ways to choose two of the electrons, and
four of these choices involve a 1s and a 2s electron.
Continued
3
v
2
0
0 1 2 3 4 5
u
Fig. 1.6. Wigner intracule for the beryllium atom.
1.3. Electron Correlation Models
The Omega intracule for an atomic or molecular system provides an impres-

sively detailed picture of the dynamical behaviour of its electrons. We know,
for any given values of u, v and , the (quasi-)probability of finding two of
its electrons at a distance u, moving with a relative speed v at a dynamical
angle . This now allows us to return to our original question: can we
exploit this information to predict electron correlation energies?
The foundation of DFT methods is the HohenbergKohn theorem
[6], which assures us that the correlation energy is a functional of the
one-electron density (r). We now make an analogous conjecture [21]:
that the correlation energy is a functional of the Omega intracule, i.e.
Ec = F [(u, v, )]. (1.23)
To prove this remains an interesting open challenge and there are surely
many possible lines of attack. Our earliest attempts sought to show that
the Hamiltonian can be reconstructed (apart from unimportant translations
and rotations) from the intracule. If this can be shown, it proves the con-
jecture, for the correlation energy is certainly a functional of the Hamil-
tonian. However, even in the absence of a proof, we feel that the correlation-
relevant information in the Omega intracule is much more accessible than
that in the one-electron density and, therefore, it should be easier to recover
Ec from (u, v, ) than from (r). We call this idea Intracule Functional
Theory (IFT).
Although one can imagine many ways to extract Ec from (u, v, ),
one of the simplest is to contract the intracule with an appropriate kernel,
writing

Ec = (u, v, )G(u, v, )ddvdu. (1.24)
0 0 0
In such a formulation, the correlation kernel G(u, v, ) acts as a weighting
function, assigning high priority to regions of intracule space where the
electrons are strongly correlated, and low priority to regions where corre-
lation is weak. The thought experiment summarized in the diagram below
helps to guide our thinking about this. In situations where both u and v
are small, the electrons are close together and moving relatively slowly
and so we anticipate a large correlation contribution. Conversely, corre-
lation effects should be small when the electrons are far apart and moving
quickly. In intermediate cases, where one of u and v is large and the other is
small, we expect moderate correlation effects. This picture fits nicely with
the conclusion in the preceding section that correlation in the He-like ions
depends in some way on the product r12 p12 .
If the wavefunction is expanded in a Gaussian basis, then combining
Eqs. (1.14) and (1.24) yields

Ec = abcd [abcd]G (1.25)
abcd
large u
medium correlation weak correlation
small u
strong correlation medium correlation
small v large v
Fig. 1.7. Four interaction scenarios for two electrons.

and it can be shown that the correlation integrals are given by

1
[abcd]G =
8( + ) ( + )3/2
3/2

exp(2 u2 2 v2 iu v P u iQ v R)
G(u, v, ) du dv. (1.26)
The four-parameter generalized Gaussian kernel [26]
G4 (u, v, ) = c exp(20 u2 20 v2 i0 uv cos ) (1.27)
has the attractive property that it leads to correlation integrals that can be
found in closed form. For example,
3
[ssss]G = c
( + )3/2 ( + )3/2 (4l2 m2 + h2 )3/2
2 2
m P + hPQ cos l2 Q2
exp R . (1.28)
4l2 m2 + h2
where l2 = 2 + 20 , m2 = 2 + 20 and h = + 0 . Except where
otherwise indicated, the numerical results below use this kernel with
two-particle density matrices abcd from (spin-unrestricted) UHF/6-311G
wavefunctions.
Optimization of the parameters (c, 0 , 0 and 0 ) in the G4 kernel
against the exact correlation energies [27] of the ground states of the first
18 atoms [28, 29] and the 56 small molecules in Poples G1 data set [30]
revealed that, whereas 0 plays a critical role in capturing the correlation
energies in these systems, 0 is unimportant and can be set to zero without
affecting the results significantly. Accordingly, we optimized and explored
the two simpler kernels
G2 (u, v, ) = c exp(i0 uv cos ) (c = 0.07695, 0 = 0.8474)
(1.29)
G3 (u, v, ) = c exp(20 u2 i0 uv cos ) (c = 0.2113, 0 = 1.0374,
0 = 0.5578) (1.30)
The correlation energies predicted by the G2 and G3 kernels (denoted
Ec2 and Ec3 , respectively) are plotted below against the exact correlation
energies of the 18 atoms and 56 molecules described above.
The first thing that one learns from these scatterplots is that these simple
kernels are surprisingly successful at capturing the principal correlation
effects in these 74 systems. It is very encouraging to find that the G2
Fig. 1.8. Comparison of correlation energies from the G2 kernel (left) and G3
kernel (right) with exact correlation energies.
kernel which is simply an exponential in x = r12 p12 performs so

well. On closer inspection, we see that the G2 kernel tends to overestimate
Ec in -bonded molecules such as CH4 and Si2 H6 and to underestimate
in -bonded molecules such as N2 and HCCH. The overestimation in the
-bonded systems is substantially reduced by the G3 kernel, whose extra
exp(20 u2 ) factor decreases the predicted correlation energies in spatially
extended systems. However, even the G3 kernel still underestimates Ec in
compact, -bonded molecules.
Why are the unsaturated molecules problematic? It appears that it is
because a significant fraction of Ec in these systems is static, rather than
dynamic, in nature. Though precise definitions are elusive, static cor-
relation is associated with the presence of low-lying excited states and
the resulting inadequacy of a single determinant wavefunction, whereas
dynamic correlation results from the intricate dance of the electrons as they
strive to avoid close encounters with one another. Evidently, our G2 and
G3 correlation models are effective at modelling dynamical correlation but
struggle to capture the static component.
1.4. Dynamic and Static Correlation
The total correlation energy, which is defined [31] as the difference

Ec = Eexact EUHF/CBS (1.31)
between the exact and UHF energies at the complete basis set (CBS) limit,
can be partitioned into a static part
Estat = ECASSCF(val)/CBS EUHF/CBS (1.32)
and a dynamic part

Edyn = Eexact ECASSCF(val)/CBS (1.33)
where CASSCF(val) refers to the Complete Active Space SCF method [32]
within a full-valence active space.Although this partition is just one of many
that have been suggested, it has the twin virtues of conceptual simplicity
and computational tractability, at least for smallish systems.
One of the simplest and most instructive systems in which both Estat and
Edyn are significant is partially dissociated H2 and the graph below shows
how Ec (solid black), Edyn (dashed black [33]), and Ec3 (solid grey, from

Eq. (1.30)) evolve as the bond length R varies from 0.2 to 3.8 A. The cusp in
the Edyn curve arises from the well-known RHF UHF instability around

1.2 Angstrom. It is clear that the Ec3 model reproduces the behaviour of Ec
poorly, but that it bears some similarity to the Edyn curve. This confirms our
earlier observation that our simple IFT models capture primarily dynamic,
rather than static, correlation energy.
If we re-fit the G3 kernel to the Edyn curve, we obtain the new param-
eters c = 0.090, 0 = 0.85 and 0 = 0.525 and the resulting Edyn 3 energies
(dashed grey) match Edyn with near-mEh accuracy. Continuing in this vein,
we can abandon the HF/6-311G two-particle density matrix in favour of
the CASSCF(val)/6-311G one and, by re-fitting the G3 kernel again, we
obtain the parameters c = 0.102, 0 = 1.02, 0 = 0.43. The resulting
energies match Edyn with sub-mEh accuracy at all bond lengths. This sug-
gests that combining an IFT-based treatment of dynamic correlation with a
full-valence multireference method will produce a method that is capable
of estimating Ec very accurately.
-10
e (mEh)
-20
-30
-40
-50
-60
0.2 0.8 1.4 2 2.6 3.2 3.8
R ()
Fig. 1.9. Correlation energy in the H2 molecule as a function of bond length.

Of course, multireference methods are exponentially expensive, but a

hierarchy of approximations can be constructed to reduce the computational
cost of this step [34]. Furthermore, the Density Matrix Renormalization
Group method (see Chapter 3) provides an alternate route to obtaining
static correlation energies and, in some cases, it dramatically outperforms
conventional wavefunction methods.
Although the absolute error in the IFT dynamic correlation energy is
small everywhere, its relative error grows exponentially with R. This is
because, whereas the correlation energy from any kernel of the form (1.30)
decays exponentially with R, the true Ec decays as R6 . Accordingly, we
now turn our attention to London dispersion.
1.5. Dispersion Energies
London Model of Dispersion Energy

The Hamiltonian for two Coulomb-coupled oscillators with force constant
k = 42 , separated by R, is
12 + 22 1 1 1 1
H = +22 (r12 +r22 )+ + .
2 |R| |R + r1 | |R r2 | |R + r1 r2 |
If R
1, the sum of the Coulomb interactions is dominated by the dipole-
dipole term and we can write
12 + 22 x1 x2 + y1 y2 2z1 z2
H = + 22 (r12 + r22 ) + .
2 R3
If we transform to extracule and intracule coordinates, i.e.
r1 + r2 r1 r 2
S= T= ,
2 2
then H becomes fully separable and its lowest eigenvalue is

1 1 1 1
E=2 + 2 +2 2 + 2 + 2 + .
4R3 4R3 2R3 2R3
Because R is large, we can expand E as a power series in 1/R to obtain
3
E = 6 +
323 R6
and subtracting the energy of the uncoupled oscillators yields the celebrated
London dispersion energy
3
Ec .
323 R6

At large bond lengths (R
5 A), the UHF energy of H2 rapidly
approaches the energy of two non-interacting H atoms and fails to capture
the long-range dynamic correlation that is responsible for the weak van der
Waals attraction. This long-range correlation energy can be rationalized by
considering a multipole expansion of the Coulomb operator, as pioneered
by London in the early 1930s [35, 36].
Can we use IFT to model dispersion? To answer this, we begin by consid-
ering the simple system two Coulomb-coupled harmonic oscillators
that London used to model dispersion effects. He showed that its dispersion
energy is asymptotically Ec 3/(323 R6 ) and his derivation is outlined
in the box above. Therefore, we must devise kernels that recover this asymp-
totic dispersion energy from this systems intracules. Because we favour
kernels that depend on x = r12 p12 , we confine our attention to the D(x)
intracule and seek kernels that satisfy

Ec D(x)G(x)dx. (1.34)

or equivalently, by Parsevals Theorem,

Ec D(k)G(k)dk. (1.35)

where the hats indicate Fourier transforms.
Example 1.6: The Intracule Family for the H H Complex

The UHF wavefunction for very stretched H H in a single-Gaussian basis is
(r1 , r2 ) = (2/)3/2 exp[(|r1 R/2|2 + |r2 + R/2|2 )]
and one finds from Eq. (1.16) that the Omega intracule is
(u, v, ) = (1/)u2 exp[(u2 + R2 )]i0 (2Ru)v2 exp(v2 /4) sin .
By integrating appropriately, we can find some of the lower intracules in closed
form, viz.
W(u, v) = (2/)u2 exp[(u2 + R2 )]i0 (2Ru)v2 exp(v2 /4)
P(u) = (/)3/2 4u2 exp[(u2 + R2 )]i0 (2Ru)
M(v) = (4)3/2 4v2 exp(v2 /4)
() = (1/2) sin
Continued
and the others in their Fourier representation, viz.

s2 sin k2 j0 (ks) R2 k2
(s, ) = exp dk
0 (1 + k 2 )3/2 1 + k2

2s2 k2 j0 (ks) R2 k2
A(s) = exp dk
0 (1 + k2 )3/2 1 + k2

1 cos kx R2 k 2
D(x) = exp dk,
2 (1 + k2 )3/2 1 + k2
where j0 (z) = z1 sin z. We note that the Omega intracule, and therefore all
the lower intracules, are non-negative everywhere. As R 0, they reduce to
the intracules of the He-like ions (see Example 1.3).
Fortunately, all of the intracules (or their Fourier transforms) of

Londons model can be found in closed form and they are shown in
Example 1.5. By combining Eq. (1.35) with the expression for D(x), we
find that we require

G(k) T 2 k2 3
exp dk , (1.36)
(1 + k ) 1+k
2 3/2 2 32T 6

where T = R is large. It is not difficult to show that Eq. (1.36) is satisfied
by any kernel of the form
3 5
G(k) |k| g(k), (1.37)
64
where g(k) is an even function with g(0) = 1 and g( ) = 0. An obvious
example of this is
3 45

G(k) = |k|5 exp(|k|) G(x) = (x + i)6 + (x i)6 .
64 8
(1.38)
By construction, a dispersion kernel G(x) such as the one in Eq. (1.38)
is guaranteed (for any ) to yield the correct asymptotic dispersion energy
when applied to Londons system of harmonic oscillators. But does it
provide us with a useful dispersion model when applied to chemical
systems? To begin to answer this question, one should apply it to pairs
of small atoms whose C6 dispersion coefficients (i.e. coefficients of R6 )
are known accurately from experiment or high-level theory. We have done
this [37], using the UHF/6-311G density matrices of the H, He, Li and Be
atoms, and the results are shown in Table 1.1 below. Given that the exact
Table 1.1. C6 dispersion coefficients (in atomic units) for

pairs of small atoms.
Exact values From Eq. (1.38)

H He Li Be H He Li Be
H 6.5 12.2
He 2.8 1.5 5.2 3.3
Li 66.5 22.5 1395 74.2 17.2 1534
Be 34.8 13.2 478 213 76.6 23.0 847 657
C6 coefficients range over three orders of magnitude, the discovery that the
IFT estimates are usually accurate to within a factor of two is a promising
start. Once again, this demonstrates the fundamental suitability of IFT for
capturing intrinsically two-electron correlation effects.
1.6. Future Prospects
In the teething stages of the development of DFT, much progress was made
through a primarily empirical approach. Indeed, between Slaters intro-
duction of X theory [38] in 1951 and the publication of the Hohenberg
Kohn theorem [6] 13 years later, it was not even realized that DFT was a
theoretically justifiable theory: rather, it was embraced simply because it
was a model that worked, surprisingly often.
In some ways, contemporary IFT has evolved similarly, and now stands
at a similar point. It is clearly capable of yielding chemically useful quan-
titative predictions but, for the moment, it lacks the solid foundation of a
HohenbergKohn analogue. This deficiency may deter the purist, but the
pragmatist finds it difficult to resist the allure of a model that seeks to
rationalize the correlation phenomenon through a simple, quasi-classical
two-electron picture.
So, what are the likely directions for the development of IFT in the near
future?
As functional manufacture has become an industry within DFT, we
foresee the construction of new and improved kernels as one of the most
obvious threads of future IFT research. To ensure that this progress is
rational, we expect that properties of the exact kernel will also be derived
and that these will be used as guides.
However, we also foresee the real possibility that the kernel ansatz (1.24)
may be obsolesced by the discovery of alternative methods for extracting
Ec from the Omega intracule. Perhaps such methods will be found as by-
products of the construction of a rigorous proof of the central IFT con-
jecture (1.23).
Of course, it is also possible that the Omega intracule family tree does not
contain the ultimate intracule and that, in the future, it will be replaced
by a different, and quantum mechanically rigorous, family. We are opti-
mistic about this because it has been shown recently that the Dot intracule
D(x) is actually a first-order (in h) approximation to the true density of the
x variable. Furthermore, the exact density X(x) has also been discovered
[3941] and it is no more difficult to extract from the wavefunction than
is D(x).
Finally, we conclude with a statement that is surpassingly obvious and
yet often overlooked. If we are to refine and enrich our understanding of the
electron correlation phenomenon, we must continue to unearth and analyze
simple systems where the phenomenon is most clearly exposed and most
readily comprehended. The helium atom, the hydrogen molecule and the
uniform electron gas have all proven to be rich veins in the past but our
quest for deeper understanding must be an ongoing one and there is no
doubt whatever that there is much to be learned from other prototypical
systems [42].
Bibliography
[1] D.R. Hartree, Proc. Cam. Phil. Soc. 24, 89 (1928).

[2] E. Schrodinger, Ann. Phys. 79, 361 (1926).
[3] J.C. Slater, Phys. Rev. 34, 1293 (1929).
[4] V. Fock, Z. Physik. 61, 126 (1930).
[5] W. Pauli, Z. Physik. 31, 765 (1925).
[6] P. Hohenberg and W. Kohn, Phys. Rev. B 136, 864 (1964).
[7] A.J. Cohen, P. Mori-Sanchez, and W. Yang, Science 321, 792 (2008).
[8] C.A. Coulson and A.H. Neilson, Proc. Phys. Soc. (London) 78, 831 (1961).
[9] A.M. Lee and P.M.W. Gill, Chem. Phys. Lett. 313, 271 (1999).
[10] J.D. Baker, D.E. Freund, R. Nyden Hill, and J.D. Morgan III, Phys. Rev. A 41, 1247
(1990).
[11] J. Linderberg, Phys. Rev. 121, 816 (1961).
[12] P.F. Loos and P.M.W. Gill, J. Chem. Phys. 131, 241101 (2009).
[13] P.F. Loos and P.M.W. Gill, Phys. Rev. Lett. 105, 113001 (2010).
[14] V.A. Rassolov, J. Chem. Phys. 110, 3672 (1999).
[15] K.E. Banyard and C.E. Reed, J. Phys. B 11, 2957 (1978).
[16] N.A. Besley, A.M. Lee, and P.M.W. Gill, Mol. Phys. 100, 1763 (2002).
[17] E. Wigner, Phys. Rev. 40, 749 (1932).
[18] K. Husimi, Proc. Phys. Math. Soc. Japan 22, 264 (1940).
[19] N.A. Besley, Chem. Phys. Lett. 409, 63 (2005).

[20] E.R. Davidson, Reduced Density Matrices in Quantum Chemistry (Academic, New
York, 1976).
[21] P.M.W. Gill, D.L. Crittenden, D.P. ONeill, and N.A. Besley, Phys. Chem. Chem.
Phys. 8, 15 (2006).
[22] D.L. Crittenden and P.M.W. Gill, J. Chem. Phys. 127, 014101 (2007).
[23] S.F. Boys, Proc. Roy. Soc. (London) A200, 542 (1950).
[24] J.W. Hollett and P.M.W. Gill, Phys. Chem. Chem. Phys. 13, 2972 (2011).
[25] NIST handbook of mathematical functions, edited by F.W.J. Olver, D.W. Lozier,
R.F. Boisvert, and C.W. Clark (Cambridge University Press, New York, 2010).
[26] D.L. Crittenden, E.E. Dumont, and P.M.W. Gill, J. Chem. Phys. 127, 141103 (2007).
[27] D.P. ONeill and P.M.W. Gill, Mol. Phys. 103, 763 (2005).
[28] E.R. Davidson, S.A. Hagstrom, S.J. Chakravorty, V. Meiser Umar, and C. Froese
Fischer, Phys. Rev. A 44, 7071 (1991).
[29] S.J. Chakravorty, S.R. Gwaltney, E.R. Davidson, F.A. Parpia, and C. Froese Fischer,
Phys. Rev. A 47, 3649 (1993).
[30] J.A. Pople, M. Head-Gordon, D.J. Fox, K. Raghavachari, and L.A. Curtiss, J. Chem.
Phys. 90, 5622 (1989).
[31] J.A. Pople and J.S. Binkley, Mol. Phys. 29, 599 (1975).
[32] B.O. Roos and P.R. Taylor, Chem. Phys. 48, 157 (1980).
[33] V.A. Rassolov, M.A. Ratner, and J.A. Pople, J. Chem. Phys. 112, 4014 (2000).
[34] D.L. Crittenden and P.M.W. Gill, to be published.
[35] F. London, Z. Physik. Chem. B 11, 222 (1930).
[36] F. London, Trans. Faraday. Soc. 33, 8 (1937).
[37] D.L. Crittenden, P. Xu, and P.M.W. Gill, to be published.
[39] Y.A. Bernard and P.M.W. Gill, New J. Phys. 11, 083015 (2009).
[40] Y.A. Bernard and P.M.W. Gill, J. Phys. Chem. Lett. 1, 1254 (2010).
[41] Y.A. Bernard, D.L. Crittenden and P.M.W. Gill, J. Phys. Chem. A 114, 11984 (2010).

Chapter 2
Explicitly Correlated Electronic

Structure Theory
Frederick R. Manby
Centre for Computational Chemistry,

School of Chemistry, University of Bristol, Bristol BS8 1TS, UK
A major problem of wavefunction-based electronic structure theory is the

slow convergence of correlation energies with respect to the size of the
one-particle basis set. The situation can be dramatically improved through
incorporation of terms in the wavefunction that depend explicitly on the
interelectronic distances; and after a decade of intense development such
explicitly correlated electronic structure theories are ready for widespread
use. In this chapter I briefly summarise the essential elements of explicitly
correlated methods, and then present five thoughts on how the field might
develop in the future.
2.1. Introduction
2.1.1. Basis-set expansions

Basis-set expansions are used throughout quantum chemistry. An expansion
of a function f of some variables x

n
f(x) = ci gi (x)
i=1
requires the definition of a basis set of functions gi and a corresponding

vector of coefficients ci . Important examples for quantum chemistry include
the expansion of molecular orbitals in a basis set of atomic orbitals and
expansion of the full, many-body wavefunction in a basis set of Slater
25
26 F.R. Manby
determinants. A key question is how accurate can such an expansion be?

In particular, if we establish some positive definite measure, n , of the
accuracy of the expansion of length n, is there a finite value n such that
n < for any positive ? In such cases the expansion is said to be complete
and we have the reassurance that arbitrary accuracy is accessible, at least in
principle. This has been established, for example, for expansions of orbitals
in Slater-type and Gaussian-type basis sets [1, 2], and for expansions of
N-electron antisymmetric functions in a basis set of Slater determinants [3].
The electronic structure theory of molecules has been dominated by the
idea of expanding wavefunctions in terms of products of one-electron func-
tions. This approach is well founded and eminently practical, since: (1) it is
reasonable to start by regarding the problem in a mean-field approximation,
where the wavefunction really is a product of one-electron functions; and
(2) the concept leads to theories for the calculation of wavefunctions and
energies that only require four different types of molecular integral. These
are the one-electron integrals for overlap, kinetic energy and interaction
with nuclei; and the two-electron integrals that express the electrostatic
interaction between electrons.
The two-electron integrals are quite complicated (six-dimensional and
inseparable) and quite numerous (the number of two-electron integrals
grows with the fourth power of the size of the molecule). In the case of
molecules the use of Gaussian-type orbitals enormously simplifies the eval-
uation of these integrals; in fact the simplification is so great that currently
known alternative choices for atomic-orbital expansions are rarely used for
polyatomic systems (but see Chapter 7).1
Expansions of Gaussians can be set up in such a way as to be com-
plete in the formal sense discussed above, and for nearly all electronic
structure applications there is no theoretical objection. In devising atomic-
orbital basis sets one can use the angular-momentum expansion, in which
successive sets of functions that saturate a given angular momentum l are
added. Alternatively, one can use the principal expansion in which one
function of each angular momentum is added at each stage, to give a series
of the kind 2s1p 3s2p1d 4s3p2d1f . For a first-row atom like
carbon, the minimal 2s1p basis is called single-zeta, the second double-zeta,
etc. It is possible to form sequences of basis sets in which energies con-
verge remarkably uniformly to the basis-set limit, best exemplified by the
1 It should be pointed out that for crystalline solids, or indeed for any periodic system, plane-waves can
be used, and then the treatment of two-electron terms is even more straightforward. Moreover, for pure
density functional theory, it is practical to use Slater-type orbitals, since in a density-fitted approach to
the Coulomb problem the need for multicentred Coulomb integrals can be avoided [4].
Explicitly Correlated Electronic Structure Theory 27
correlation-consistent basis sets of Dunning and co-workers [5], denoted cc-

pVnZ. These basis sets lead to correlation energies with errors that reliably
decay as = O(n3 ), and this fact can be exploited to extrapolate cor-
relation energies from two successive basis sets to give a more accurate
estimate of the basis-set limit [6].
A careful analysis of the errors and costs involved illustrates clearly
why basis-set convergence is such a central issue in molecular electronic
structure theory. The number of basis functions in cc-pVnZ scales as m =
O(n3 ). Hence the error in the correlation energy scales as = O(m1 ).
Since the number of two-electron integrals to evaluate scales as m4 , the
computer time must scale similarly: t = O(m4 ). And hence we find that
the error decays as = O(t 1/4 ); or to put it in more immediate terms, an
order-of-magnitude improvement in accuracy costs a ten-thousand times
more computer time.
One way to circumvent this slow convergence is to avoid calculating cor-
relation energies from many-electron wavefunctions all together. This is a
key feature of the density functional approach to electronic structure [7, 8],
and also of efforts to use intracules to compute correlation energies (see
Chapter 1). There is an enormous amount of literature on the strengths and
weakness of density functional theory, but here I will limit myself to one
of each. The relevant advantage is that the energies converge rapidly with
respect to the size of the one-electron basis set, because only one-electron
functions have to be described. Insofar as there is a many-electron wave-
function, it is of the mean-field type, and is represented as a single deter-
minant of orbitals. The disadvantage is that the error necessarily introduced
through the use of an approximate density functional cannot be controlled,
or in other words, accuracy cannot be systematically improved.
In fact, there have been some impressive efforts to introduce the sys-
tematically improvable accuracy of wavefunction-based theory to density
functional theory (see for example Refs. [912]), but it is perhaps worth
pointing out that with this comes the inextricable disadvantage of slow
convergence with respect to orbital basis set.
The approach to be discussed in the present chapter retains the benefits
of rigorous wavefunction-based electronic structure theory, but replaces the
expansion in terms of products of one-electron functions with something
better.
The history of such methods is rooted in the early days of quantum
mechanics [1315] and the current state of the art has been documented
very thoroughly in a number of reviews [1619]. Here, instead, the emphasis
will be on new possible directions for explicitly correlated theory, some
28 F.R. Manby
based on my own ideas and some on those of others. Nevertheless, some

background about the structure of modern F12-type explicitly correlated
methods is essential.
2.2. F12 Theory
We begin by considering second-order MllerPlesset perturbation theory

(MP2 see Jensen [20] for the basics and Szabo and Ostlund [21] or
Helgaker et al. [22] for a more detailed exposition). The Hamiltonian is
partitioned into the HartreeFock Hamiltonian

N
H0 = f (i) (2.1)
i=1
at zeroth order and a correction H1 = H H0 . Here H is the Fock operator;

the zeroth-order wavefunction is the HartreeFock Slater determinant, and
the HartreeFock energy is given by the sum of the zeroth and first-order
energy EHF = E0 +E1 . Correlation effects first appear in the wavefunction
at first order and in the energy at second order, and EMP2 = E0 + E1 + E2 .
MP2 is a pair theory, i.e., the correlation energy is a sum of contribu-
tions from each electron pair, which can be calculated independently. The
first-order wavefunction consists of a linear combination of all determi-
nants that can be obtained by a double excitation from the HartreeFock
ground state. A given electron pair occupying the state |ij is excited into
its corresponding first-order pair function, written
1 ij
|uij = Tab |ab, (2.2)
2
ab
ij
where Tab is the amplitude of the excitation from |ij to |ab. Here and
throughout, indices i, j, . . . refer to spin-orbitals that are occupied in the
HartreeFock reference state |0; and a, b, . . . refer to unoccupied or virtual
spin-orbitals.
It is also convenient (particularly on moving to coupled-cluster theory)
to view the first-order correction to the wavefunction in second-quantised
form:
1 ij
|1 = T2 |0 = Tab a b j i|0. (2.3)
4
ijab
Here, the operators i and a are the annihilation operator for a particle in
spin-orbital i and the creation operator for spin-orbital a, respectively. Thus
the action of the operator a b j i on |0 produces a doubly excited deter-

minant, in which orbitals i and j have been replaced by a and b, and which
might conveniently be denoted |ab ij . Only doubly excited determinants
contribute to the first-order correction |1.
2.2.1. MP2-F12
In MP2-F12 theory the basic idea is to supplement the product-of-virtuals
expansion in Eq. (2.3) with explicitly correlated terms formed as a product
of occupied orbitals and a correlation factor, f12 , which depends explicitly
on the distance between two electrons, r12 :
1 ij 1 ij
|uij = Tab |ab + Q12 Tkl f12 |kl (2.4)
2 2
ab kl
The operator Q12 ensures strong orthogonality to the occupied Hartree

Fock orbitals; this condition is required in order to maintain the simple
pair form of second-order perturbation theory. Without this operator, the
theory would be equivalent, but more complicated. Development of prac-
tical theories based on this wavefunction form was pioneered by Kutzelnigg
and Klopper [2328], who originally used the linear correlation factor
f12 = r12 .
In the notation of second quantization we can recognise that the con-
ventional T2 operator is supplemented by an explicitly correlated term:
|1 = T2 |0 + T2 |0, (2.5)
where T2 incorporates the conventional double excitations and T2 brings in

the explicitly correlated terms. This operator can be conveniently written
down using a representation of the Q12 using an infinite, virtual basis set
, :

Q12 = ||. (2.6)

Using this form, the pair function of Eq. (2.4) becomes

1 ij 1 ij
|uij = Tab |ab + Tkl ||f12 |kl
2 2
ab kl
1 ij 1 ij
= Tab |ab + kl
Tkl F |, (2.7)
2 2
ab kl
30 F.R. Manby
kl = |f |kl. The T operator can now be cast in the form

where F 12 2
1
T2 =
ij kl
Tkl F j i. (2.8)
4
ijkl
This form is convenient for deriving methods, but hardly physically trans-
parent, so it is worth noting that the operator produces excitations of two
electrons into states formed from the product of two occupied orbitals (k
and l) multiplied by the correlation factor f12 and projected to give a con-
figuration strongly orthogonal to the occupied space.
In the final working equations it is of course essential to remove any
explicit reference to the infinite
basis , ,
and this can be done by strategic
replacements of the kind || = 1 i |ii|. This process introduces
many-electron integrals which are very expensive to evaluate. Kutzelnigg
suggested removing all such many-electron integrals by a strategy equiv-
alent to replacing these exact resolutions of the identity by approximate
ones, initially in the MO basis set:

|| |aa|.
a
A very important refinement was made by Klopper and Sansom [29], who
introduced the idea of performing the approximate resolution of the identity
in a separate, auxiliary basis set. Among many other technical developments
(reviewed elsewhere [16, 17, 30]) which followed, one stands out as partic-
ularly significant: Ten-no discovered that a short-range correlation factor,
in particular the exponential f12 = exp(r12 ), led to a huge improvement
in accuracy compared to f12 = r12 [31], and now this form is used by all
leading groups. The general theory in its modern form is clearly presented
in [30].
2.2.2. Explicitly correlated coupled-cluster theory

Coupled-cluster theory with single and double excitations can be handled
using similar techniques. The conventional CCSD wavefunction ansatz
| = eT1 +T2 |0 (2.9)
is extended to include the explicitly correlated doubles to give

| = eT1 +T2 +T2 |0. (2.10)
The first such method (using f12 = r12 and resolutions of the identity (RIs)
in the molecular-orbital basis set) was given by Noga et al. [32, 33] but
more recently general implementations of CCSD-F12 using approximate

RIs in auxiliary basis sets have been described [3436].
One drawback of these methods is the need to use multiple approximate
resolutions of the identity. The full theory places high demands on the com-
pleteness of the auxiliary basis set, and the computer cost scales higher than
linearly with respect to the size of this basis. For these reasons a great deal
of effort has been invested in the development of approximate CCSD-F12
schemes, such as CCSD(R12) [37], CCSD(F12) [38,39], CCSD(2)R12 [40],
CCSD-F12a and -F12b [41, 42] and CCSD(F12 ) [43]. All of these can be
supplemented by a conventional perturbative treatment of triple excitations.
Using automatic code generation, it has been possible to implement
higher-order explicitly correlated coupled-cluster methods, such as
CCSDTQ-F12, using a cluster operator T1 + T2 + T2 + T3 + T4 augmented
with explicitly correlated terms only at the two-particle level [44, 45]. The
first, and thus far only, foray into the idea of full explicitly correlated treat-
ments of triples was made by Kohn [46].
Approximate CCSD-F12 schemes cost barely more than the conven-
tional CCSD calculation in the same basis, but achieve results much closer
to the basis-set limit (explicitly correlated calculations in a triple-zeta basis
are at least as good as conventional calculations using quintuple-zeta [38]).
The state of the art in explicitly correlated coupled-cluster theory has been
described in detail in two recent reviews [18, 19].
2.3. Five Thoughts for F12 Theory
The field of explicitly correlated electronic structure theory has undergone a

major period of innovation and development in the years following the key
paper of Klopper and Samson [29]. The result is that explicitly correlated
MP2 and coupled-cluster methods can now be used routinely in software
packages such as Molpro [47] and Turbomole [48]. Here I would like to
present a selective list of five speculations on how one might proceed with
explicitly correlated electronic structure theory in the future.
2.3.1. Thought 1: Do we need (products of) virtuals?

The thought occurred to me and to others [49] that since F12 theories
perform so well, perhaps one could abandon the use of products of virtual
orbitals altogether. This would offer potentially huge advantages, because
the number of basis functions required to saturate the occupied space is
obviously far smaller than are needed for an accurate treatment of electron
32 F.R. Manby
correlation. Moreover, the amplitudes in purely explicitly correlated wave-

functions have only occupied indices, presumably offering the opportunity
for considerable reduction in computational cost.
For example at the MP2 level one can set up the first-order pair functions
in the form
ij
|uij = Q12 Tkl, g12 |kl, (2.11)
kl

where g12 is a two-particle basis function that depends on the interelectronic
distance r12 . (The superscript is a label, not a power). The pair function can

be expected to be dominated by the diagonal terms g12 |ij, but the additional
summation over occupied orbitals kl is included to ensure invariance of the
theory under unitary rotations of the occupied orbitals [28].
The number of parameters to determine in such an expansion is about
o4 g, if there are o occupied orbitals and g geminal basis functions. The
number in conventional MP2 is o2 v2 , where v is the number of virtual
functions. The number of amplitudes in this explicitly correlated ansatz
is therefore smaller than in the standard method when g < (v/o)2 ; one
needs only a few (let us say ten) functions to give a flexible description of
a one-dimensional function of the kind typically encountered in quantum
chemistry, so this requirement is satisfied even at the double-zeta level.
Unfortunately, such methods are not sufficiently accurate to be useful,
either in an MP2-type formulation or in the simple coupled-cluster approxi-
mation, CC2 [49].Although between 80 and 90% of the correlation energy is
recovered, we found considerable problems with consistency. In particular,
the correlation energies obtained in such a scheme are of no practical value
in computing reaction energies.
A major factor in the failure of the ansatz of Eq. (2.11) is the neglect of
orbital relaxation, the importance of which can be seen in a simple calcu-
lation on the helium atom. We consider first a simple uncorrelated wave-
function of the form e(r1 +r2 ) . Optimization of the exponent gives the well
known screening result, 0 = 2 5/16, and the energy 2.848 Eh , com-
pared to the exact value of around 2.904 Eh . We will refer (slightly inac-
curately) to this difference of 56 mEh as the correlation energy. Using the
correlated wavefunction e0 (r1 +r2 ) f(r12 ) where f(r12 ) is a variationally
optimised function of the interelectronic distance, a correlation energy of
only 29 mEh is obtained. However, if is optimised together with f in
the wavefunction e(r1 +r2 ) f(r12 ) then a correlation energy of 43 mEh
is obtained. These calculations are closely analogous to the full calcula-
tions based on HartreeFock orbitals, and show that for the helium atom,
inclusion of orbital relaxation effects increases the fraction of correlation

energy recovered from around half to over three-quarters.
Unfortunately when one tries to make an orbitally relaxed formalism
in general, integrals inevitably arise of the form ij|w12 |ax, where w12 is
one of various benign two-electron operators, a is a virtual orbital, and x
is an auxiliary function used to perform the RI. Products of virtuals can be
introduced in the simplest approximation using only integrals of the form
1
ij|r12 |ab. Since the auxiliary basis (x) is typically larger than the virtual
basis (b), it can be seen that nothing is really gained by excluding the
products of virtuals. Furthermore, the use of a short-range correlation factor
g rules out long-range dispersion between mono-electronic fragments, as
will be illustrated below for the simple case of two hydrogen atoms.
Following Slater and Kirkwood [50] we consider two well-separated
hydrogen atoms, with the distance between them, R, large enough that we
can neglect exchange effects. The geometrical parameters are shown in
Fig. 2.1 and the internuclear axis is taken to be aligned with the z-axis. The
zeroth-order Hamiltonian is given by the sum of one-electron Hamiltonians
for the fragments, and the zeroth-order wavefunction is simply the product
of 1s orbitals on each atom: |0 = |1sA (1)1sB (2). The interaction can be
described as a small perturbation through the Hamiltonian
1 1 1 1
H1 = + +
rB1 rA2 R r12
and, since rA1 and rB2 are both much smaller than the internuclear separation
R, the operator can be accurately represented as a Taylor expansion in 1/R.
The leading term is
1
H1 (xA1 xB2 + yA1 yB2 2zA1 zB2 ).
R3
We now model the first-order wavefunction as
|1 = f(r12 )|0.
Fig. 2.1. Geometrical parameters used in the description of dispersion effects

between two well-separated hydrogen atoms.
34 F.R. Manby
The first-order energy vanishes, but at second order there are terms like
1|H1 |0 = 1sA (1)1sB (2)f(r12 )|H1 |1sA (1)1sB (2)
(|1sA (1)|2 |f(r12 )H1 |1sB (2)|2 ). (2.12)
we obtain
Given that R is so large, we can approximate r12 R
T(R)
1|H1 |0 = f(R) B ,
A
which is clearly zero since the dipole moments A = B = 0. Even in the
case of interaction of non-spherical fragments, for which these dipoles need
not vanish, there is clearly no transferable way to ensure that the correlation
factor f reproduces the correct long range behaviour; one would need to
know the answer in advance. Similarly, the other second-order term in the
2 , so again there appears no way
energy, 1|H0 |1, is proportional to f(R)
that a reasonable description of long-range dispersion could emerge.
In conclusion, it seems that orbital relaxation effects are essential for
an effective description of dynamic correlation, but including this effect is
more expensive than retaining the conventional virtual products. And long-
range dispersion seems impossible in any ansatz where correlation only
arises from products of occupied orbitals and a function of the interelec-
tronic distance. Perhaps progress can be made by considering more flexible
correlation factors of pair-function ansatze, such as those considered in the
next section.
2.3.2. Thought 2: Are there better two-electron basis sets?

In 1969 Boys and Handy [51] presented an analysis of the electron corre-
lation factor from which it is instructive to quote:
The simplest idea of a correlation function is f(r12 ). . . . However,
the shape of [the] apparent physical repulsion may change from
point to point and in general it is necessary to make fij dependent
on r1 , and r2 . In fact, the simpler r12 function is not general
enough even to explain an apparent repulsion which is different
in different directions.
The detailed form of the correlation factor then deduced was used in the
seminal work on the transcorrelated method [52] and also forms the basis
for correlation factors (or Jastrow factors) used in quantum Monte Carlo
calculations.
Building on the basic idea of using a more flexible correlation factor,
Knowles and I investigated an ansatz in which the correlation factor was
expanded in terms of the average and relative coordinates of the two elec-
= (r1 + r2 )/2 and r = r2 r1 [53]. At the time we were reluctant
trons, R
to use the RI approximation for the many-electron integrals but, in the
absence of a practical alternative, the method remained applicable only to
two-electron systems.
Now we realise that the main problem with MP2-R12 theory was the
choice f12 = r12 , not the decision to approximate the integrals using
RIs [54], so it seems a good time to revisit this type of theory. A flexible
ansatz can be made in the R, r coordinates, in the form
ij
|uij = Q12 p (r )|kl.
tklPp P (R) (2.13)
klPp
Here the P (R) are atom-centred basis functions, whose number would
therefore scale linearly with system size. The functions in r , on the other
hand, only have to span a length-scale characteristic of electron correlation,
and the number of p (r ) needed should not scale with system size. The
derivation of the theory (for example MP2 or coupled-cluster theory) would
be the same as for any model with more than one correlation factor [55]; the
challenges are first, scaling; and second, integrals. The number of param-
eters is on the order o4 Mm, if there are M functions in R and m functions
in r . Since M scales linearly with system size, it can be seen that there are
O(N 5 ) parameters as a function of system size.
The resolution of the identity ensures that only two-electron integrals
will appear in the working equations; but there will be several new types
r )|(1)(2). Some effort would be
of integrals, such as (1)(2)|P(R)p(
involved in making these available, but in the end we know that if all of the
functions are Gaussians, efficient computation is possible.
The scaling problem above could be circumvented if instead of
expanding the correlation factor in R, r one directly expanded the first-
order wavefunction:
ij
|uij = Q12 p (r ).
t | P (R) (2.14)
Pp
Pp
Then the number of parameters scales only as o2 Mm, which is clearly an

improvement over o2 v2 , or even over o4 , since m does not scale with system
size.
2.3.3. Thought 3: Do we need the resolution of the identity?

The resolution of the identity (see Section 2.2) is one of the key ele-
ments of modern explicitly correlated methods. Avoidance of the six-index,
36 F.R. Manby
three-electron and eight-index, four-electron integrals is clearly crucial

in order to progress to systems of more than a few atoms. As described
above, the key innovation came in the form of Kutzelniggs approximate
resolutions of the identity. The unit operator can be approximated in an
orthonormal basis set of one-electron functions |x as

1 x = |xx|. (2.15)
x
A typical three-electron integral appearing in MP2-F12 theory is
1
ijm|f12 r23 |mlk, where the repeated m-indices arise from a projection
onto the occupied orbital space. This integral can then be approximated by
inserting the operator x2 between the two-electron operators:
1 1
ijm|f12 r23 |mlk ijm|f12 x2 r23 |mlk

1
= ijmx|f12 24 r23 |mlkx, (2.16)
x
where in the second line the operator 24 permutes the electron labels two
and four. Applying this operator and factorizing the now decoupled four-
electron integral one obtains

1 1
ijm|f12 r23 |mlk ijmx|f12 r34 |mxkl
x

1
= ij|f12 |mxmx|r12 |kl, (2.17)
x
so it can be seen that the RI has approximately resolved the original three-
electron integral as a sum of products of two-electron integrals.
The RI can equally well be seen as a fitting procedure [56] where
products of three orbitals are expanded in the auxiliary basis |x. To see
this, we can write the integral in the previous equation as a one-electron
overlap integral:
1
ijm|f12 r23 |mlk = |, (2.18)
where

(r2 ) = dr1 i (1)m (1)f12 j (2)
(2.19)
1
(r2 ) = dr3 m (3)k (3)r23 l (2).
Because of this structure, the auxiliary basis for a calculation on an atom

can become saturated only when it contains functions with three times
the maximum angular momentum of the occupied orbitals, 3locc . For light
elements in the s- and p-block there is no problem, because this implies only
up to f-functions in the fitting set; but for transition metals, the occupied
d-orbitals lead to the need for i-functions in the RI basis; and, worse still,
for lanthanides and actinides, the auxiliary basis requires functions with
l = 9.
An alternative strategy is to avoid the RI altogether and directly apply
density fitting to the three-electron integrals. In conventional density fitting
orbital products i (r )j (r ) |ij) are approximately expanded in a basis
set of auxiliary functions |A) (typically taken as atom-centered Gaussians).
The approximation to the orbital product then has the form
ij
|ij) |ij) = DA |A), (2.20)
A
ij
where the fitting coefficients DA are found by minimizing the Coulomb
energy of the fitting residual (ij
ij|ij
ij)/2.
Writing our typical three-electron integral in a Mulliken-like notation
1 1
ijm|f12 r23 |mlk (im|f12 | jl |r23 |mk) (2.21)
we can easily suggest an approximation of the form

1
(im|f12 | jl |r23 |mk) (im|f |r1 |mk).
12 | jl (2.22)
23
This approximation can, however, be expected to be poor, because errors

in the approximated integrals scale linearly with errors in the fitted orbital
products. Exactly the same situation arises when one tries to use density
fitting for two-electron integrals with operators different from that used to
ij
determine the coefficients DA . But it was clearly pointed out by Dunlap (see
for example Ref. [57] and references therein) how to avoid this problem
through the use of robust density fitting formulae, and this has proved
invaluable in a number of areas [56, 58, 59].
In brief: one obtains a robust density fitting expression by starting from
the naive version above, and explicitly subtracting the linear contributions
to the difference between the approximate and exact integrals.
The resulting fitting formula has the form
1
(im|f12 | jl |r23 |r1 |mk)
|mk) (im|f12 | jl + (im|f
12 | jl |r1 |mk)
23 23
|r1 |mk) 2(im|f
12 | jl
+ (im|f |r1 |mk),
12 | jl (2.23)
23 23
38 F.R. Manby
or, in terms of basic integrals and fitting coefficients

1 1 jl
(im|f12 | jl |r23 |mk) (im|f12 |B|r23 |C)DB DC
mk
1
+ (A|f12 | jl |r23 |C)DA
im mk
DC
1 jl
+ (A|f12 |B|r23 |mk)DA
im
DB
1 jl
2(A|f12 |B|r23 |C)DA
im mk
DB DC . (2.24)
This expression involves only three- and four-index integrals and interme-
diates, and the various contractions between integrals and coefficients scale
only as O(N 5 ). The conventional RI approach has an O(N 6 ) cost. Using
RI, as noted above, functions with up to 3locc are needed in the auxiliary
basis set, but using DF, only functions up to 2locc are required. Therefore,
direct density fitting of three-electron integrals would appear to be a viable
alternative to the resolution of the identity.
2.3.4. Thought 4: Could we have explicit correlation

for higher excitations?
The preliminary answer has to be yes, since Kohn has described such
a method for the perturbative treatment of triple excitations [46]. Here,
an additional explicitly correlated term is introduced based on the combi-
nation of a conventional double excitation and a particle-hole correlation
factor; this has the effect of explicitly correlating two of the particles in the
connected three-particle scattering event.
But it is perhaps interesting to consider the problem more generally: is
there any value in a fully three-particle explicitly correlated method? This
would involve excitations from a state |ijk to some general three-particle
virtual state, |3 . The theory can be formally derived, as in doubles the-
ories, by introducing a resolution of the identity in a complete, orthonormal
virtual basis
3
|3 = |S , (2.25)

where we have defined overlap integrals S 3
= |3 . Thus, for
purposes of formal derivation, a triples cluster operator can be written
in the form

T3 = Tijk3 S
3
k j i
ijk3
ijk
= T k j i . (2.26)
ijk
Because of this formal use of an infinite virtual space, the basic equa-
tions are exactly as in conventional coupled cluster theory with a complete
treatment of triples [60]; but here the challenge is to take these equations
and convert them into a computable form by eliminating all references to
the infinite virtual space, resorting to resolution-of-the-identity approxima-
tions as necessary. This is definitely a technical challenge, which cannot be
addressed without a very considerable investment of effort.
One can speculate about useful possible forms for |3 . The scattering
picture of triple excitations (see for example [61]) suggests a formalism
of the structure f12 f23 |lmn or f12 |lma for the triples function. The fully
connected version f12 f13 f23 |lmn would presumably lead to very difficult
integrals; and the most general type of expression f123 |lmn would lead to
many-electron integrals of a kind that could not be resolved by conventional
RIs. But given that the Hamiltonian contains only two-particle interactions,
and based on the success of Kohns work, there are grounds to be optimistic
that a form such as f12 f23 |lmn might be accurate enough. If this could be
used in place of rather than in addition to the conventional expansion,
very significant savings could be made, and explicitly correlated triples
models with the order of o6 amplitudes, instead of o3 v3 , at least offer an
interesting prospect.
2.3.5. Thought 5: Can we avoid three-electron errors in

two-electron systems?
In a HartreeFock calculation on the hydrogen atom, it makes no difference
whether you use exact or approximate two-electron integrals, or ignore them
altogether, because the direct and exchange two-electron contributions to
the final energy precisely cancel out. Obviously, this is not the case if
the treatment is different for the Coulomb and exchange terms, and this
kind of self-interaction effect is a serious problem in approximate density
functional theory.
In an explicitly correlated configuration-interaction calculation on the
helium atom, all three-electron integrals precisely cancel out. A CCSD
40 F.R. Manby
calculation on helium is equivalent to configuration interaction and so it too

should be independent of three- and four-electron integrals. Nevertheless
if one performs calculations using standard approximations [CCSD-F12a,
F12b or (F12 )] one finds that the resulting energies do depend on the aux-
iliary basis set used, suggesting a dependence on three-electron integrals.
This anomaly can be readily understood: there are some three-electron
integrals that should exactly cancel out for a two-electron system, but which
are treated differently in currently available approximations. It would surely
improve stability of these methods if this issue was resolved.
2.4. Conclusions
Treatment of dynamic correlation using conventional, orbital-based elec-

tronic structure theory is beset by basis-set convergence difficulties. These
can be greatly alleviated through the use of explicitly correlated terms in the
wavefunction, which bypass the slow convergence of orbital expansions by
describing the important two-electron features of the wavefunction directly.
Immense strides have been made in the past decade towards making these
methods useful and usable. Here, five lines of thought have been set out on
how to push the concept of explicit correlation further, to develop methods
with even higher accuracy and efficiency in the field of molecular electronic
structure theory.
Acknowledgments
I would like to thank my friends and collaborators for illuminating discus-

sions on explicitly correlated wavefunctions, in particular Hans-Joachim
Werner, Mihaly Kallay, Wim Klopper, Seiichiro Ten-no, Ed Valeev and
Peter Knowles. I am especially grateful to David Tew for very helpful dis-
cussions and suggestions during the preparation of this chapter.
Bibliography
[1] B. Klahn and W.A. Bingel, Theo. Chim. Acta 44, 9 (1977).
[2] B. Klahn and W.A. Bingel, Theo. Chim. Acta 44, 27 (1977).
[3] P.-O. Lowdin, Phys. Rev. 97, 1474 (1955).
[4] E.J. Baerends, D.E. Ellis, and P. Ros, Chem. Phys. 2, 41 (1973).
[5] T.H. Dunning, Jr., J. Chem. Phys. 90, 1007 (1989).
[6] T. Helgaker, W. Klopper, H. Koch, and J. Noga, J. Chem. Phys. 106, 9639 (1997).
[7] R.G. Parr and W. Yang, Density-functional theory of atoms and molecules (Oxford
University Press, New York, 1994).
[8] W. Koch and M.C. Holthausen, A Chemists Guide to Density Functional Theory
(Wiley-VCH, New York, 2000).
[9] F. Furche, J. Chem. Phys. 129, 114105 (2008).
[10] J. Harl and G. Kresse, Phys. Rev. Lett. 103, 1 (2009).
[11] B.G. Janesko, T.M. Henderson, and G.E. Scuseria, J. Chem. Phys. 130, 081105
(2009).
[12] R.J. Bartlett, Chem. Phys. Lett. 484, 1 (2009).
[13] E.A. Hylleraas, Z. Phys. 54, 347 (1929).
[14] H. Hettema, Quantum Chemistry: Classic Scientific Papers, volume 8 of 20th Century
Chemistry (World Scientific, Singapore, 2000), for an English translation of [13].
[15] T. Helgaker and W. Klopper, Theor. Chim. Acta 103, 180 (1999), for a modern
perspective on [13].
[16] W. Klopper, F.R. Manby, S. Ten-no, and E.F. Valeev, Int. Rev. Phys. Chem. 25, 427
(2006).
[17] T. Helgaker, W. Klopper, and D.P. Tew, Mol. Phys. 106, 2107 (2008).
[18] D. Tew, C. Hattig, R. Bachorz, and W. Kopper, in Recent Progress in Coupled Cluster
Methods edited by P. Carsky, J. Paldus, and J. Pittner (Springer, Dordrecht, 2010),
535572.
[19] H.-J. Werner, T.B. Adler, G. Knizia, and F.R. Manby, in Recent Progress in Coupled
Cluster Methods edited by P. Carsky, J. Paldus, and J. Pittner (Springer, Dordrecht,
2010), 573620.
[20] F. Jensen, Introduction to Computational Chemistry (John Wiley & Sons, Chichester,
second edition, 2007).
[21] A. Szabo and N.S. Ostlund, Modern Quantum Chemistry (McGraw-Hill, New York,
1982).
[22] T. Helgaker, P. Jrgensen, and J. Olsen, Molecular Electronic Structure Theory (John
Wiley & Sons, Chichester, 2000).
[23] W. Klopper and W. Kutzelnigg, Chem. Phys. Lett. 134, 17 (1987).
[24] W. Klopper and W. Kutzelnigg, J. Phys. Chem. 94, 5625 (1990).
[25] W. Kutzelnigg and W. Klopper, J. Chem. Phys. 94, 1985 (1991).
[26] V. Termath, W. Klopper, and W. Kutzelnigg, J. Chem. Phys. 94, 2002 (1991).
[27] W. Klopper and W. Kutzelnigg, J. Chem. Phys. 94, 2020 (1991).
[28] W. Klopper, Chem. Phys. Lett. 186, 583 (1991).
[29] W. Klopper and C.C.M. Samson, J. Chem. Phys. 116, 6397 (2002).
[30] H.-J. Werner, T.B. Adler, and F.R. Manby, J. Chem. Phys. 126, 164102 (2007).
[31] S. Ten-no, J. Chem. Phys. 126, 014108 (2007).
[32] J. Noga, W. Kutzelnigg, and W. Klopper, Chem. Phys. Lett. 199, 497 (1992).
[33] J. Noga and W. Kutzelnigg, J. Chem. Phys. 101, 7738 (1994).
[34] T. Shiozaki, M. Kamiya, S. Hirata, and E.F. Valeev, Phys. Chem. Chem. Phys. 10,
3358 (2008).
[35] T. Shiozaki, M. Kamiya, S. Hirata, and E.F. Valeev, J. Chem. Phys. 129, 071101
(2008).
[36] A. Kohn, G.W. Richings, and D.P. Tew, J. Chem. Phys. 129, 201103 (2008).
[37] H. Fliegl, W. Klopper, and C. Hattig, J. Chem. Phys. 122, 084107 (2005).
42 F.R. Manby
[38] D.P. Tew, W. Klopper, C. Neiss, and C. Hattig, Phys. Chem. Chem. Phys. 9, 1921
(2007).
[39] D.P. Tew, W. Klopper, and C. Hattig, Chem. Phys. Lett. 452, 326 (2008).
[40] E.F. Valeev, Phys. Chem. Chem. Phys. 10, 106 (2008).
[41] T.B. Adler, G. Knizia, and H.-J. Werner, J. Chem. Phys. 127, 221106 (2007).
[42] G. Knizia, T.B. Adler, and H.-J. Werner, J. Chem. Phys. 130, 054104 (2009).
[43] C. Hattig, D.P. Tew, and A. Kohn, J. Chem. Phys. 132, 231102 (2010).
[44] T. Shiozaki, M. Kamiya, S. Hirata, and E.F. Valeev, J. Chem. Phys. 130, 054101
(2009).
[45] T. Shiozaki, E.F. Valeev, and S. Hirata, J. Chem. Phys. 131, 044118 (2009).
[46] A. Kohn, J. Chem. Phys. 130, 131101 (2009).
[47] H.-J. Werner, P.J. Knowles, R. Lindh, F.R. Manby, M. Schutz, et al., Molpro, version
2009.1, a package of ab initio programs, 2009, see http://www.molpro.net.
[48] TURBOMOLE V6.2 2010, a development of University of Karlsruhe and
Forschungszentrum Karlsruhe GmbH, 19892007, TURBOMOLE GmbH, since
2007; available from http://www.turbomole.com.
[49] S. Hofener, D.P. Tew, W. Klopper, and T. Helgaker, Chem. Phys. 356, 25 (2009).
[50] J.C. Slater and J.G. Kirkwood, Phys. Rev. 37, 682 (1931).
[51] S.F. Boys and N.C. Handy, Proc. Roy. Soc. A309, 209 (1969).
[52] S.F. Boys and N.C. Handy, Proc. Roy. Soc. A310, 43 (1969).
[53] F.R. Manby and P.J. Knowles, Chem. Phys. Lett. 310, 561 (1999).
[54] A.J. May, E. Valeev, R. Polly, and F.R. Manby, Phys. Chem. Chem. Phys. 7, 2710
(2005).
[55] E.F. Valeev, J. Chem. Phys. 125, 244106 (2006).
[56] F.R. Manby, J. Chem. Phys. 119, 4607 (2003).
[57] B.I. Dunlap, Phys. Chem. Chem. Phys. 2, 2113 (2000).
[58] A.J. May and F.R. Manby, J. Chem. Phys. 121, 4479 (2004).
[59] M. Schutz and F.R. Manby, Phys. Chem. Chem. Phys. 5, 3349 (2003).
[60] J. Noga and R.J. Bartlett, J. Chem. Phys. 86, 7041 (1987).
[61] P.E. Maslen, A.D. Dutoi, M.S. Lee, Y. Shao, and M. Head-Gordon, Mol. Phys. 103,
425 (2005).
Chapter 3
Solving Problems with Strong

Correlation Using the Density Matrix
Renormalization Group (DMRG)
Garnet Kin-Lic Chan and Sandeep Sharma
Department of Chemistry and Chemical Biology,

Baker Laboratory, Ithaca, New York 14850, USA.
This chapter is concerned with the problem of strongly correlated

electrons in quantum chemistry. We describe how a technique known
as the density matrix renormalization group (DMRG) can tackle com-
plicated chemical problems of strong correlation by capturing the local
nature of the correlations. We analyse the matrix product state structure
of the DMRG wavefunction that encodes one-dimensional aspects of
locality. We also discuss the connection to the traditional ideas of the
renormalization group. We finish with a survey of applications of the
DMRG, its strengths and weaknesses in chemical applications, and its
recent promising generalization to tensor network states.
3.1. The Problem of Strong Correlation
In quantum chemistry, strongly correlated states are often defined through

a partitioning of the wavefunction. In intermediate normalization, this is
| = |HF + |corr (3.1)
HF | = 1, (3.2)

where |HF is the reference HartreeFock determinant and |corr is the
correlation correction. We say | is strongly correlated if coefficients in
the determinant expansion of |corr are large, i.e. on the order of unity
43
44 G.K.-L. Chan and S. Sharma
(or greater). For the purposes of this chapter, strong correlation may be
taken to be synonymous with multireference.
Strongly correlated states most commonly arise from near-degeneracy
in the underlying orbitals. As a familiar example, the hydrogen molecule
at equilibrium (in a minimal basis) possesses energetically well-separated
bonding g and antibonding u orbitals, and | is well approximated by
the single configuration g2 . However, as the bond is stretched, g and u
become near-degenerate and | evolves into a strongly correlated superpo-
sition of configurations with different occupancies across the two orbitals.
While the correlation in stretched H2 is strong, it can be exactly described
by many methods, such as doubles configuration interaction and coupled
cluster doubles theory. This is because |corr contains only one deter-
minant with large weight, u2 . In larger problems, however, the number of
significant determinants in |corr rises very rapidly. For example, consider
a set of hydrogen atoms arranged in a square lattice (Fig. 3.1). As we expand
the lattice constant (i.e. stretch the bonds) we recover a large degeneracy in
the underlying orbitals and |corr consists of a superposition of many con-
figurations distributed across all the orbitals. The strongly correlated super-
position of such a large set of configurations can now no longer be described
by simple theories. Of course, such a hydrogen lattice problem is artificial
from a chemistry standpoint, but one can readily find realistic examples
of large-scale strongly correlated electronic structure. For example,
molecules with multiple transition metals contain many near-degenerate 3d
orbitals which experience only limited overlap with neighbouring orbitals,
and are thus much like the atomic orbitals in the expanded hydrogen
lattice.
Fig. 3.1. A stretched hydrogen molecule involves strong correlation because of

the near degeneracy of the molecular orbitals. In a stretched square lattice of
hydrogen atoms the number of degenerate orbitals becomes very large leading
to an exponential number of important configurations.
Solving Problems with Strong Correlation Using the DMRG 45
The above examples illustrate that from a quantum chemistry view,

the primary challenge of strong correlation in a set of of active orbitals is
the very large number of determinants which contribute significantly to the
wavefunction. As the number of strongly correlated orbitals increases, the
number of important determinants increases exponentially, and the strong
correlation problem thus appears intractable.
A little thought, however, shows that the apparent exponential com-
plexity of strong correlation cannot really hold. While quantum mechanics
admits many unusual states such as Schrodinger cat states most of
quantum phase space is not explored by physical ground-states. Indeed,
Nature imposes many constraints on the structure of quantum states. One
of the most important is the principle of locality. Regardless of how com-
plicated a physical system may be, be it a complicated transition metal
molecule at a stretched region of the potential energy surface, or a complex
phase of a high temperature superconducting material, the response of the
system to an external perturbation always remains local. Phrased more
chemically, reactions on one side of the material do not cause changes a
macroscopic distance away.
As a consequence of locality, although the wavefunction in a strongly
correlated state may consist of an exponentially large number of deter-
minants, the coefficients of the expansion must be highly structured to
reflect locality. In fact, they should be parametrizable by an amount of
information that is proportional to the size of the system. The challenge
of strongly correlated electronic structure may therefore be translated
into a simple mathematical problem how do we encode locality in a
wavefunction that involves an expansion, in principle, across all possible
determinants?
In recent years, work in the area of the density matrix renormalization
group (DMRG), and its related modern successor of tensor networks, has
provided a conceptual and in some cases practical solution to this question.
In this chapter, we first describe the density matrix renormalization group,
which perfectly encodes locality in strongly correlated topologically one-
dimensional problems. In such systems, the DMRG is now the method of
choice for difficult problems of electronic structure.At the end of the chapter
we describe the modern development of tensor networks, which provide a
general solution to the problem of encoding locality in strong correlated
states, but where progress so far has mainly been at the conceptual rather
than practical level. In the course of our discussion, we will also take a
detour to briefly survey some applications and different formulations of the
DMRG.
3.2. The Density Matrix Renormalization Group Wavefunction
The density matrix renormalization group was first conceived by White as

a way to tackle strongly correlated quantum lattices [13]. In the current
context, we view the DMRG as a special wavefunction ansatz; we return
to the renormalization group structure in Section 3.5.
To understand the wavefunction structure of the DMRG, we establish
some basic notation. Consider an N-electron wavefunction expanded in a
complete basis of determinants, through the full configuration interaction
(FCI) expansion,

| = n1 n2 n3 ...nk |n1 n2 n3 . . . nk , (3.3)
{n}
{n} = {|vac, | , | , | }, (3.4)

{n} = N. (3.5)
i
Here |n1 . . . nk is the occupation number representation of the determinant,
where ni is the occupation of orbital i. The total number of orbitals is k and
N is the total number of electrons.
The dimension of the coefficient tensor in the above expansion is
4k , which is intractable for values of k much larger than ten. In a strongly
correlated problem, we expect a large fraction of these coefficient to be non-
zero, and thus we cannot impose a sparsity pattern on as is normally done
in configuration interaction. Instead, we need to find an ansatz where is
expressed more compactly and which does not assume sparsity. In general,
we would want such an ansatz to require only a polynomial amount of
information as a function of the number of orbitals in the system, k.
A very simple ansatz would be to approximate the high-dimensional
coefficient tensor by a tensor product of vectors. In elemental form, this
would be
n1 n2 n3 ...nk n1 n2 n3 . . . nk . (3.6)
n
Note that the vector is not an orbital vector, but rather an occupancy
vector of length four that reflects the different occupancies of a given
orbitals. Also, n1 , n2 are taken to represent elements of different vectors.
The above product ansatz contains only 4k parameters and is certainly very
tractable. However, it is also not, in general, very accurate!
To improve the ansatz we need to increase the flexibility of the vectors
n . We can introduce additional auxiliary indices, making each vector into
a tensor, i.e.
n iin . (3.7)
The new indices i, i are auxiliary in the sense that they do not appear in
the final coefficient tensor and must be contracted over in some fashion.
The simplest arrangement is to contract the indices sequentially from one
n tensor to the next. We then have

n1 n2 n3 ...nk in11 in12i2 in23i3 . . . ink1
k
. (3.8)
i1 i2 i3 ...ik1
More compactly, we can use matrix notation,
n1 n2 n3 ...nk n1 n2 n3 . . . nk , (3.9)
where we understand e.g. n2 n3 to denote the matrix product between the
two involving the auxiliary indices. For simplicity, we will assume that the
dimensions of all auxiliary indices are the same, and we call this dimension
M. Then the tensors n are of dimension 4 M M (except for the first
and the last) and the total number of parameters in the wavefunction ansatz
is O(4M 2 k).
This approximation (3.9) is, in fact, the DMRG wavefunction. It is com-
monly referred to as the DMRG wavefunction with M states. In calculations
it is typically used in a variational fashion, where the components iin are
the coefficients to be varied. Note that by increasing the dimension M,
we make the ansatz arbitrarily flexible, and eventually exact. Because the
wavefunction coefficients are obtained as a series of matrix products, the
ansatz is also referred to in the literature as the matrix product state [47].
Combining the above ansatz for the coefficient tensor explicitly with the
Slater determinants yields the full DMRG wavefunction,

|DMRG = n1 n2 n3 . . . nk |n1 n2 n3 . . . nk . (3.10)
{n}
3.3. Locality and Entanglement in the DMRG
The DMRG encodes locality in strong correlated quasi-one-dimensional

problems. How is this achieved?
Consider, for simplicity, a set of orbitals arranged in a line, as shown
in Fig. 3.2. (Concretely, we may imagine a one-dimensional hydrogen
chain in a minimal, orthonormalized basis). The DMRG encodes locality
through the special structure of contractions of auxiliary indices. From
Eq. (3.9), we see the first auxiliary index i1 is associated with occupancies
n1 and n2 , and the first two components in11 and in12i2 . If i1 were not
present, the wavefunction would factorize between the n1 occupancy and
the other occupancies in the system. Consequently, i1 and its subsequent
Fig. 3.2. This figure illustrates how the matrix product state of the DMRG wave-
function encodes locality of the problem by sequentially contracting auxiliary
indices that connect adjacent orbitals.
contraction gives rise to correlations between occupancies n1 and n2 . Sim-

ilarly, auxiliary index i2 is associated with occupancies n2 and n3 , and
subsequent contraction of i2 introduces correlations between occupancies
n2 , n3 (Fig. 3.2). This pattern is repeated for the other auxiliary indices
in the wavefunction. Note that while there is no direct coupling between
non-neighbouring occupancies (such as n1 and n3 ) via auxiliary indices,
there is correlation between them by virtue of the correlation between n1 ,
n2 , and n2 , n3 .
We see that the DMRG encodes a sequential structure to the correlation.
If the orbitals considered are local and the underlying topology is quasi-one-
dimensional, then this is exactly the structure of correlation commensurate
with locality.
In fact, in one dimension orbital topologies, the DMRG can be con-
sidered a perfect encoding of locality. This optimality can be proved in
a rigorous manner, where it is related to the area law [8]. This provides
a more rigorous expression of the local nature of correlation in terms of
entanglement. Consider a system partitioned into two parts A and B, where
A contains orbitals 1 . . . a, and B, orbitals a + 1 . . . k. Any wavefunction
can be expressed in the product Fock space FA FB as

| = p |ap |bp , (3.11)
p
where {|ap } and {|bp } are orthonormal bases spanning FA and FB ,

respectively.
The entanglement entropy between A and B is then given
as S = p p ln p . The area law in one-dimensional systems states
that the entanglement entropy of the ground-state of a gapped, local (finite
interaction range) Hamiltonian is independent of the system size [8].
To see how this relates to the DMRG, consider the DMRG wavefunction
for a system partitioned into two parts as above, where we denote the
components for the subsystems by A and B, respectively,

|DMRG = An1 An2 . . . Ana Bna+1 . . . Bnk |n1 n2 . . . nk . (3.12)
{n}
We now re-express (3.12) in bi-partite form (3.11). First, we rewrite the
DMRG wavefunction in a canonical form [9] where the components satisfy
orthogonality conditions that will ensure the orthonormality of the sub-
system bases {|ap }, {|bp } in Eq. (3.11). This can be achieved by inserting
identities XX1 into the matrix product,

|DMRG = An1 X1 X11 An2 X2 X21 . . . Ana Xna Xn1
a
{n}
1 1 nk
Bna+1 Xa+1 Xa+1 . . . Xk1 Xk1 B )|n1 n2 . . . nk

= An1 An2 . . . Ana Bna+1 . . . Bnk |n1 n2 . . . nk . (3.13)
{n}
Through appropriate choices of X, the new components An , Bn can be

made to satisfy

(An ) An = I (3.14)
n

Bn (Bn ) = I. (3.15)
n
Next, we carry out a singular value decomposition of Ana = U na V , thus
giving

|DMRG = An1 An2 . . . U na V Bna+1 . . . Bnk |n1 n2 . . . nk , (3.16)
{n}
where is the diagonal matrix of singular values. By comparison with the
bi-partite form Eq. (3.11), we identify the orthonormal subsystem bases as

|ap = (An1 . . . Ana1 U na )1p |n1 . . . na (3.17)
{n}

|bp = (V Bna+1 . . . Bnk )p1 |na+1 . . . nk . (3.18)
{n}
The crucial point is that the number of singular values in Eq. (3.16) is
M, the dimension of the auxiliary indices in the DMRG wavefunction, and
this is independent of system size. Consequently, the DMRG can encode
a fixed amount of entanglement entropy independent of system size. Com-
bined with the one-dimensional area law, we see that we only need to
arrange for M to be sufficiently large to capture the appropriate amount of

entanglement for the class of Hamiltonians considered, but this does not
depend on the size of the problem, and this is the natural consequence of
correctly encoding locality.
The sequential local encoding of correlation in the DMRG is both
a strength and a weakness. On the one hand, it is the correct structure
for one-dimensional topologies, and the structure leads to highly efficient
algorithms to evaluate expectation values such as the energy. However,
for systems which are not topologically one-dimensional, the DMRG no
longer provides an optimal encoding of locality. While the complexity of
strongly correlated states formally appears to scale exponentially with all
three dimensions of the problem, the DMRG wavefunction removes the
exponential scaling associated with only one of the dimensions. A complete
reduction to polynomial complexity for general orbital topologies requires
more flexible wavefunction structures. We return to this later in the chapter
when we discuss tensor networks. Nonetheless, even in its current form, the
DMRG wavefunction is still a powerful wavefunction for strongly corre-
lated states in arbitrary dimensions, and a perfect wavefunction for strong
correlation in quasi-one-dimensional problems, and this is what is observed
in its practical application to quantum chemistry.
3.4. Other Properties of the DMRG
The DMRG has many formal properties which are beneficial for quantum
chemical applications. Here we briefly discuss a few:
Variational: Since we have an explicit wavefunction, the expectation

value of the energy provides a variational upper bound to the true energy
and in practice DMRG energies are evaluated in this way. With effi-
cient algorithms, the DMRG variational energy can be evaluated in
O(M 3 k3 ) + O(M 2 k4 ) time, where k is the number of orbitals. As M is
increased, the DMRG energy converges from above to the exact energy.
Multireference: In the DMRG wavefunction there is no division into
occupied and virtual orbitals, all orbitals appear on an equal footing in
the ansatz (Eq. 3.10). In particular, the HartreeFock reference has no
special significance here. For this reason, we expect (and observe) the
ansatz to be very well-balanced for describing nondynamic correlation
in multireference problems (see e.g. [1012]). Conversely, the ansatz
is inefficient for describing dynamic correlation, since to treat dynamic
correlation one would benefit from the knowledge of which orbitals are
in the occupied and virtual spaces. We return to this in Section 3.6.
Size-consistency: The DMRG ansatz is size-consistent when using a
localised basis. To see this in an informal way, let us assume that we
have two DMRG wavefunctions |A and |B for subsystems A and B
separately. Both A and B have a matrix product structure, i.e.

|A = na1 . . . nak |na1 . . . nak (3.19)
{na }

|B = nb1 . . . nbk |nb1 . . . nbk . (3.20)
{nb }
Their product is also a DMRG wavefunction with a matrix product
structure. This then describes the combined system AB in a size-
consistent way, i.e.
|AB = |A |B = na1 . . . nak nb1 . . .
{na }{nb }
nbk
|na1 . . . nak nb1 . . . nbk (3.21)
3.5. Relation to the Renormalization Group
In Section 3.2 we motivated the construction of the DMRG from the decom-
position of a high-dimensional tensor. In the original formulation, however,
the DMRG was derived from the numerical renormalization group ideas of
Wilson [1,2,13]. This alternative viewpoint is quite helpful and we describe
it briefly.
Consider again the model problem of a linear chain of k hydrogen atoms
in a minimal orthonormal basis. In a renormalization group approach, we
build up the electronic structure of the hydrogen chain one atom at a time.
For the first atom, any state |i1 in the Fock space F1 of the first basis
function, {n1 }, can be written as
|i1 = in11 |n1 . (3.22)
n1
For states |i2 , |i3 in the Fock spaces F2 , F3 of two and three hydrogen
atoms, respectively, nn
|i2 = i21 2 |n1 n2 (3.23)
n1 n2

|i3 = in31 n2 n3 |n1 n2 n3 . (3.24)
n1 n2 n3
Finally, for a state in the Fock space of the full chain Fk , we recover the
full configuration interaction representation of Eq. (3.3).
In the above, we expanded states in the product basis of occupancies
of the individual atoms. However, imagine that we solve the Schrodinger
equation of successively longer hydrogen chains, first with one atom, then
two, and so on. We would like to reuse information from the eigenstates
of the k 1 atom subchain to construct eigenstates of the k atom chain.
Since the k 1 chain eigenstates |ik1 form a complete basis for Fk1 ,
rather than expanding in the occupancy basis of Fk , we can instead use the
basis {|ik1 } {|nk }. For example, for |i3 F3 , instead of (3.24), we
can write
n
|i3 = i23i3 |i2 n3 , (3.25)
i2 n3
where the coefficients in this intermediate expansion in23 and in the occu-
pancy basis n1 n2 n3 are related via
nn n
in31 n2 n3 = i21 2 i23i3 . (3.26)
i2
Extrapolating, a state of k hydrogen atoms |ik can be written in terms of
the states of the intermediate k 1 atom chain |ik1 , which themselves can
be written in turns of the intermediate states |ik2 , and this is repeated all
the way to |i1 . This leads to a nested relationship between the coefficients
in the intermediate expansion and occupancy basis of the k atom chain.

ink1 n2 n3 ...nk = in11 in22i3 in33i4 . . . ink1
k
ik . (3.27)
i1 i2 i3 ...ik1
For ground state ik = 1 and Eq. (3.27) takes the same form as the
DMRG wavefunction. However, there are two details which we have yet
to discuss. Firstly, Eq. (3.27) is exact, since we used the complete basis
for each (intermediate) Fock space of the subchains. As derived above, the
dimension of the auxiliary ip index, associated with the p atom subchain,
is not M, but grows exponentially as p = 1 . . . k 1. Secondly, the compo-
nents are here constructed from the Hamiltonian eigenstates of interme-
diate subchains, rather than being variational coefficients as in the DMRG
wavefunction.
Regarding the first point, Wilson noted that the eigenstates |ip of an
intermediate p subchain span an increasing set of energies as the subchain
length p increases. If we are interested in only a few low energy eigen-
states of the full k chain problem, it would be unnecessary to use a com-
plete basis for the intermediate Fock spaces. Instead, for each intermediate
chain, we can truncate the expansion to a fixed number M of low-energy

eigenstates. (In renormalization group terminology, the adaptive truncation
is a renormalization. More specifically, the Hamiltonian matrix elements
change in each step as we move to a new truncated basis, thus we view
the interactions of the problem as being effectively renormalized. The
usage of the term group is largely historical, dating back to the Gell-
ManLow differential equation in quantum electrodynamics which has a
Lie-group structure. Here, we can formally view the set of all the trun-
cations as forming a semi-group). Wilsons choice of low-energy eigen-
states for the renormalization leads to the numerical renormalization group
technique [13].
Regarding the second point, White recognized that a sub-chain of
hydrogen atoms embedded in a longer chain should be viewed as an open
system, and is characterized not by eigenstates of its Hamiltonian, but rather
by its statistical density matrix [1,2]. The natural basis for truncation for the
subsystem is then not the low-energy Hamiltonian eigenstates, but rather
eigenstates of the statistical density matrix with large weights! The use of
the density matrix based renormalization is what gives the density matrix
renormalization group method its name. At the variational minimum, the
density matrix based truncation can be shown to be equivalent to varia-
tionally optimizing the components (see, e.g. [9]), and this establishes
the equivalence between the matrix product state wavefunction and the
density matrix renormalization group viewpoints.
3.6. Dynamic Correlation the Role of Canonical

Transformations
As we mentioned in Section 3.4, the DMRG wavefunction treats all orbital

occupancies on an even footing. This is very different from a reference
based quantum chemistry method, where a set of the orbitals is considered
largely unoccupied (virtuals), and these are to be excited into. The exci-
tation picture of correlation is appropriate for dynamic correlation, but is
very different from the structure of correlation contained in the DMRG
wavefunction. Experience with coupled cluster theory demonstrates that
the appropriate structure of dynamic correlation is achieved through an
exponential operator.
An exponential description of dynamic correlation may be combined
with a DMRG description of the strong correlation if we use the DMRG to
describe correlation in an active space, and use the exponential operator to
excite out of the active-space. In this case, the wavefunction becomes

| = eA |DMRG , (3.28)
where A connects the active and virtual spaces, a kind of state-specific
multireference coupled cluster ansatz. However, the basic problems of this
ansatz are well-known in coupled cluster theory. Unlike in standard single-
reference coupled cluster theory, because the reference function is not a
single Slater determinant, the exponential operator eA has to contain semi-
internal excitations that relax the active space. (An example of a semi-
internal excitation would be an operator of the form aa1 av1 aa2 aa3 , where
a1 , a2 , a3 are active orbital labels, and v1 is a virtual orbital label). These
semi-internal excitation operators do not commute and give rise to a non-
terminating BakerCampbellHausdorff (BCH) expansion of the effective
Hamiltonian
H = eA HeA
1
= H + [H, A] + [[H, A], A] + , (3.29)
2
where each successive commutator contains operators of increasing com-
plexity. This has so far limited the use of wavefunctions of the general form
(3.28).
We have recently proposed a method that we term canonical trans-
formation theory (see, e.g. the review article [14]). This uses cumulant
decompositions and operator decompositions in a recursive way to effi-
ciently evaluate the non-terminating BCH expansion. However, this is cer-
tainly not the only viable approach, and questions still remain over the best
strategy to handle dynamic correlation in a strong correlation scenario
regardless of whether the DMRG is used as the strong correlation method.
This area of intersection between dynamic and strong correlation techniques
remains a topic for much future work.
3.7. What Can the DMRG Do? A Brief History
After its introduction in 1992 by White [1], the density matrix renor-
malization group was soon applied to many problems involving model
Hamiltonians in condensed matter. Early applications in conjunction with
semi-empirical Hamiltonians focused on the Hubbard and PariserParr
Pople (PPP) models for conjugated systems, see e.g. Refs. [1521]. As
a representative example, we consider the work by Fano et al. [18] who
performed DMRG studies on cyclic polyenes (Cm Hm , m = 4n + 2,
n = 1, 2, . . . ) (also known as annulenes) in the PPP model [2224]. These

calculations obtained an excellent, near-exact treatment of the electronic
correlation. Given the well-known difficulties of single-reference quantum
chemistry methods such as coupled cluster theory to describe correlation in
these systems (see e.g. studies by Paldus and coworkers [25,26]) such early
work demonstrated the promise of the DMRG method for multireference
chemistry problems.
In 1999, White and Martin carried out the first application of the DMRG
method to a molecule using the full ab initio quantum chemical Hamil-
tonian [27]. In particular, they demonstrated that a DMRG calculation with
a moderate number of states could recover near full CI accuracy for the
water molecule in a double-zeta with polarization basis. The White and
Martin paper was significant also in formulating the intermediates that are
necessary for an efficient implementation of the DMRG in quantum chem-
istry. After these initial papers, other groups in quantum chemistry started
to investigate the DMRG method. Currently, several implementations exist,
including that of Mitrushenkov et al. [28], ours [29], that of Reiher [30]
derived from the earlier implementation of Legeza and Hess [31], and more
recently, Zgid and Nooijen [32], and Kurashige and Yanai [33].
Over the last few years the DMRG has been applied to a wide variety of
molecular problems in quantum chemistry. In the early stages, these were
typically small molecular problems for which benchmark molecular results
were available. Some examples include the single-triplet gap in HHeH [34],
benchmark water and nitrogen curves in small basis sets [29, 35], singlet-
triplet gaps in methylene [32], and the ionic-covalent curve crossing in
LiF [36]. More recently, thanks to efficient implementations, it has become
possible to use the DMRG to obtain near-exact solutions to molecular
problems beyond the capabilities of full configuration interaction. Some
examples of this include our exact solution of the Schrodinger equation
for water in a triple-zeta, double-polarization basis [37], as well as for
the nitrogen binding curve at the all-electron double-zeta with polarization
basis level [11]. However, while the use of the DMRG to obtain near-exact
solutions to the Schrodinger equation for problems beyond conventional
FCI is interesting, the early molecular studies highlighted a more signif-
icant point, namely the types of correlation that the DMRG wavefunction
is efficient at capturing. For example, in our studies of the nitrogen binding
curve, we compared DMRG calculations to high level coupled cluster with
up to hextuple excitations (CCSDTQPH). While the CCSDTQPH energies
and the DMRG energies were comparable at the equilibrium geometry,
the DMRG energies for a fixed number of states M retained essentially
the same accuracy across the entire potential energy curve. This and other
studies demonstrated the ability of the DMRG wavefunction to capture
multireference correlation in a balanced way, as we described in Section 3.4.
Conversely, when moving from a small basis to a larger basis DMRG calcu-
lation for the same molecule (e.g. from a double-zeta to a triple-zeta basis
for the water molecule, as in [29]), the number of states that needed to be
kept in the DMRG ansatz to achieve a given accuracy had to be increased
significantly, demonstrating that dynamical correlation is not efficiently
captured by the DMRG wavefunction. Thus, as emphasized several times
in this chapter, the most promising domain of application of the DMRG
method must to be to solve active-space strong correlation, multireference,
problems. With current DMRG technology, a nearly-exact treatment of the
complete active space correlation for arbitrary molecules with up to roughly
30 active orbitals and electrons can be achieved.
Given the strength of the DMRG method for large-scale multireference
electronic structure, a clear domain of application must be to complicated
transition metal problems. Although such applications are still at a rel-
atively early stage, the Reiher group has performed some preliminary
studies [30, 3841]. For example, they used the DMRG method [30] to
calculate the spin-gap of the Cu2 O2 core of tyrosinase, a problem which
had evaded conventional complete active space methods due to the need for
a large active space. More recently, Kurashige and Yanai, not only obtained
correctly converged DMRG energies for the same Cu2 O2 system that had
been studied earlier but not fully converged by the Reiher group, but also
carried out a near exact solution of the complete active space problem for the
Cr2 molecule correlating an active space of 24 electrons in 30 orbitals [33].
Finally, we recently reported a description of the Cu2 O2 electronic structure
problem that included dynamic correlation via canonical transformation
theory on top of the density matrix renormalization group, along the lines
of Section 3.6.
One of the directions of our own group in recent years has been to use
the DMRG as an efficient local multireference method for long molecules.
In these ideal settings, the DMRG method obtains near-exact active space
solutions of the Schrodinger equation for problem sizes inconceivable using
other techniques, e.g. for 100 orbital, 100 electron active spaces. In our
first demonstration, we showed how the DMRG could exactly describe the
simultaneous bond-breaking of 49 bonds in a hydrogen chain, a problem
nominally requiring a 50 electron, 50 orbital active space. In more recent
works, we have used our local DMRG method to study excited states in con-
jugated systems which have significant multireference character, ranging
from unusual polyradical character in the acene series [42], to magnetism

in poly-phenylcarbenes [43] to the nature of dark states in light-harvesting
pigments [44].
3.8. The Future: Higher Dimensional Analogues
As we have emphasized in this chapter, the DMRG wavefunction encodes

locality perfectly in topologically one-dimensional problems, by virtue of
the structure of contractions of the auxiliary indices in Eq. (3.9). While
this allows us to push back the barriers of strong correlation from the
current limits of complete active space methods, there are still many dif-
ficult problems of correlation which will require a more general encoding
of locality that applies to non-one-dimensional topologies. Consider, for
example, the molecule shown in Fig. 3.3, a giant Keplerate magnet with
over 30 iron centres. An active space approach, using only the 3d and 4s
orbitals on each iron atom, would require 180 active orbitals. While this
size of active space seems extremely large, if the topology of the connec-
tions were one-dimensional, this problem could, in fact, be treated in the
Fig. 3.3. A keplerate magnet contains more than 30 iron spin-centres. Because
of the non-linearity of the correlation DMRG does not give a good description of
electronic structure of this molecule, but this might be a candidate for solution by
a more general tensor network.
density matrix renormalization group. Our inability to treat such problems

is thus related to our inability to efficiently express the principle of locality
in general non-one-dimensional topologies.
The natural approach to higher dimensions would be to extend each
component tensor in Eq. (3.9) with additional indices, that could be con-
nected in ways reflecting the underlying general topology of the orbitals
and the structure of the entanglement. This extension gives rise to the tensor
network class of wavefunctions [45]. Here, two of the prominent examples
are the Projected Entangled Pair States (PEPS), which is a direct general-
ization of the DMRG to higher dimensions, and the multiscale entanglement
renormalization ansatz or MERA, which is framed more as a generalization
of the original concept of the renormalization group [46]. Here we will focus
only on the description of PEPS. A graphical illustration of the PEPS wave-
function is shown in Fig. 3.4. It is more difficult to write down algebraically,
but schematically the wavefunction tensor for a square lattice topology of
orbitals is decomposed as
nxy
n00 n01 ...nkk = ixy ix +1 jx1y jxy
y
{nxy } {ixy ,jxy } {xy}

=
n00 ,n01 ,... i00 ,i01 ,... j00 ,j01 ,...
in0000i10 j10 j00 in0101i11 j11 j01 . . . , (3.30)
Fig. 3.4. Schematic representation of Projected Entangled Pair States (PEPS)

which are a direct generalization of DMRG for higher dimensions. Unlike in the
DMRG, each orbital can connect to more than two other orbitals with bonds which
represent the auxiliary indices or renormalized bases.
where x, y range from 0 to k and denote the coordinates of the orbitals on the
lattice. Promisingly, the general arguments that demonstrate the optimality
of the DMRG in one-dimensional topologies appear to apply to the PEPS
wavefunction in two- and three-dimensional topologies.
There remain, however, many challenges before efficient calculations
using tensor network wavefunctions are practical. Most of the difficulties
arise from the proliferation of auxiliary indices. For example, unlike in
the case of the DMRG wavefunction, the exact variational evaluation of
the energy in PEPS formally requires exponential time! However, as is
well known from coupled cluster theory, it is not necessary for the energy
(and other observables) to be evaluated as variational expectation values
and indeed approximate polynomial time algorithms to evaluate the PEPS
energy have been introduced [45]. Still, the most pressing questions before
these higher dimensional analogues of the DMRG become widely used,
however, are (i) what is the best approximate algorithm for evaluating
expectation values, and (ii) are there modifications to the tensor network
form which facilitate more efficient manipulation. Should these questions
be satisfactorily solved, this would open the way to the application of
tensor network states such as PEPS and MERA as a general way to solve
strongly correlated electronic structure problems, of arbitrary complexity,
in quantum chemistry.
Bibliography
[1] S.R. White. Phys. Rev. Lett. 69, 2863 (1992).

[2] S.R. White. Phys. Rev. B 48, 10345 (1993).
[3] U. Sychollwock, Rev. Mod. Phys. 77, 259 (2005).
[4] M. Fannes, B. Nachtergaele, and R.F. Werner, Comm. Math. Phys. 144, 443 (1992).
[5] M. Fannes, B. Nachtergaele, and R.F. Werner, J. Funct. Anal. 120, 511 (1994).
[6] S. Ostlund and S. Rommer, Phys. Rev. Lett. 75, 3537 (1995).
[7] S. Rommer and S. Ostlund, Phys. Rev. B 55, 2164 (1997).
[8] M.B. Hastings, Phys. Rev. B 76, 035114 (2007).
[9] G.K.-L. Chan, Phys. Chem. Chem. Phys. 10, 3454 (2008).
[10] G.K.-L. Chan, J. Chem. Phys. 120, 3172 (2004).
[11] G.K.-L. Chan, M. Kallay, and J. Gauss, J. Chem. Phys. 121, 6110 (2004).
[12] J. Hachmann, W. Cardoen, and G.K.-L. Chan, J. Chem. Phys. 125, 144101 (2006).
[13] K.G. Wilson, Rev. Mod. Phys. 47, 773 (1975).
[14] E. Neuscamman, T. Yanai, and G.K.-L. Chan, Int. Rev. Phys. Chem. 29, 231 (2010).
[15] S. Ramasesha, S.K. Pati, H.R. Krishnamurthy, Z. Shuai, and J.L. Bredas, Synth. Met.
85, 1019 (1997).
[16] D. Yaron, E.E. Moore, Z. Shuai, and J.L. Bredas, J. Chem. Phys. 108, 7451 (1998).
[17] Z. Shuai, J.L. Bredas, A. Saxena, and A.R. Bishop, J. Chem. Phys. 109, 2549 (1998).
[18] G. Fano, F. Ortolani, and L. Ziosi, J. Chem. Phys. 108, 9246 (1998).
[19] G.L. Bendazzoli, S. Evangelisti, G. Fano, F. Ortolani, and L. Ziosi, J. Chem. Phys.
110, 1277 (1999).
[20] C. Raghu, Y. Anusooya Pati, and S. Ramasesha, Phys. Rev. B 65, 155204 (2002).
[21] C. Raghu, Y. Anusooya Pati, and S. Ramasesha, Phys. Rev. B 66, 035116 (2002).
[22] R. Pariser and R. Parr, J. Chem. Phys. 21, 466 (1953).
[23] R. Pariser and R. Parr, J. Chem. Phys. 21, 767 (1953).
[24] J.A. Pople, Trans. Faraday Soc. 49, 1375 (1953).
[25] J. Paldus, M. Takahashi, and R.W.H. Cho, Phys. Rev. B 30, 4267 (1984).
[26] J. Paldus, J. Czek, and M. Takahashi, Phys. 30, 2193 (1984).
[27] S.R. White and R.L. Martin, J. Chem. Phys. 110, 4127 (1999).
[28] A.O. Mitrushenkov, G. Fano, F. Ortolani, R. Linguerri, and P. Palmieri, J. Chem.
Phys. 115, 6815 (2001).
[29] G.K.-L. Chan and M. Head-Gordon, J. Chem. Phys. 116, 4462 (2002).
[30] K.H. Marti, I.M. Ondk, G. Moritz, and M. Reiher, J. Chem. Phys. 128, 014104
(2008).
[31] O. Legeza, J. Roder, and B.A. Hess, Phys. Rev. B 67, 125114 (2003).
[32] D. Zgid and M. Nooijen, J. Chem. Phys. 128, 014107 (2008).
[33] Y. Kurashige and T. Yanai, J. Chem. Phys. 130, 234114 (2009).
[34] S. Daul, I. Ciofini, C. Daul, and S.R. White, Int. J. Quantum Chem. 79, 331 (2000).
[35] A.O. Mitrushenkov, R. Linguerri, P. Palmieri, and G. Fano, J. Chem. Phys. 119, 4148
(2003).
[36] O. Legeza, J. Roder, and B.A. Hess, Mol. Phys. 101, 2019 (2003).
[37] G.K.-L. Chan and M. Head-Gordon, J. Chem. Phys. 118, 8551 (2003).
[38] G. Moritz, B.A. Hess, and M. Reiher, J. Chem. Phys. 122, 024107 (2005).
[39] G. Moritz and M. Reiher, J. Chem. Phys. 124, 034103 (2006).
[40] G. Moritz, A. Wolf, and M. Reiher, J. Chem. Phys. 123, 184105 (2005).
[41] G. Moritz and M. Reiher, J. Chem. Phys. 126, 244109 (2007).
[42] J. Hachmann, J.J. Dorando, M. Aviles, and G.K.-L. Chan, J. Chem. Phys. 127, 134309
(2007).
[43] T. Yanai, Y. Kurashige, D. Ghosh, and G.K.-L. Chan, Int. J. Quantum Chem. (2009).
In press.
[44] D. Ghosh, J. Hachmann, T. Yanai, and G.K.-L. Chan, J. Chem. Phys. 128, 144117
(2008).
[45] F. Verstraete, V. Murg, and J.I. Cirac, Adv. Phy. 57, 143 (2008).
[46] G. Vidal. Understanding Quantum Phase Transitions, Series in Condensed Matter
Physics, edited by L.D. Carr (Taylor & Francis, Boca Raton, 2010) pp. 115138.
Chapter 4
Reduced-Density-Matrix Theory
for Many-electron Correlation
David A. Mazziotti
Department of Chemistry and The James Franck Institute,

The University of Chicago, Chicago, IL 60637 USA
For 50 years progress towards the direct calculation of the ground-

state two-electron reduced density matrix (2-RDM) was stymied from
an inability to constrain the 2-RDM to represent an N-electron wave-
function. Recent advances in theory and optimization have realized three
methods for the direct calculation of the 2-RDM: (i) the variational
2-RDM method in which the 2-RDM is constrained explicitly through
N-representability constraints known as positivity conditions, (ii) the
parametric 2-RDM method in which the 2-RDM is constrained implicitly
though the parametrization of the 2-RDM as a functional of itself, and
(iii) the solution of the contracted Schrodinger equation (CSE) or its
anti-Hermitian part (ACSE), in which p-RDMs for p > 2 are built from
the 2-RDM by a cumulant-based reconstruction. Advantages of the 2-
RDM methods include: (i) the treatment of strong electron correlation by
the variational 2-RDM method, where traditional wavefunction methods
would require as many as a billion times more determinants than feasible
with the largest supercomputers, (ii) the balanced description of single-
and multi-reference correlation of the ACSE method which matches or
exceeds the accuracy of traditional multi-reference wavefunction-based
methods at a lower computational scaling, and (iii) the combination of
accuracy and efficiency through the parametric 2-RDM method, which
approaches the accuracy of coupled cluster methods with single, double,
and triple excitations at the computational cost of configuration interaction
with single and double excitations. Collectively, the 2-RDM methods have
been applied to studying strong electron correlation in acene chains and
hydrogen lattices, resolving the energy barriers in bicyclobutanes ring
61
62 D.A. Mazziotti
opening, computing the conical intersections in methylenes triplet excited

states, and examining hydroxyurea derivatives for treating sickle-cell
anemia. In this chapter we will discuss the theoretical foundations, prac-
tical advantages, and some recent applications of each 2-RDM method.
4.1. Introduction
Because electrons are indistinguishable with only pairwise interactions, the

energy of any atom or molecule can be expressed as a linear functional of
the two-electron reduced density matrix (2-RDM) [14]. This formulation
suggests the tantalizing possibility of employing the 2-RDM rather than
the many-electron wavefunction to compute the ground-state energy of an
atom or molecule. In 1955 Mayer [4] performed an encouraging pencil-
and-paper calculation, but Tredgold [5] soon discovered that for a simple
system the energy as a functional of the 2-RDM could be optimized sub-
stantially below the exact ground-state energy. Tredgold [5], Coleman [7],
Coulson [6] and others realized that for an N-electron problem the trial
2-RDM was assuming a form that did not correspond to an N-electron
wavefunction, that is the trial 2-RDM at the minimum energy could not be
obtained from the integration of an N-electron density matrix. The 2-RDM
must be constrained by additional rules (or conditions) to derive from an
N-electron wavefunction, and Coleman described these necessary and suf-
ficient rules as N-representability conditions [79].
The 2-RDM calculations of Mayer and Tredgold employed four basic
requirements for a density matrix of indistinguishable fermions [7]: the
matrix should be (i) normalized to conserve particle number, (ii) Hermitian,
(iii) antisymmetric under particle exchange, and (iv) positive semidefinite to
keep probabilities nonnegative. A matrix is positive semidefinite if and only
if all of its eigenvalues are nonnegative. These conditions are not sufficient
to guarantee that the 2-RDM may be computed from the integration (or
contraction) of an N-electron density matrix. A 2-RDM that can be derived
from an N-electron density matrix is said to be N-representable. What addi-
tional conditions are needed to restrict the 2-RDM to be N-representable,
and how can these conditions be efficiently imposed within a variational
calculation of the energy?
Interest in this question reemerged in the 1990s in the calculation
of the 2-RDM by an iterative solution of the contracted Schrodinger
equation [1016]. From recent progress in both theory and optimization,
Reduced-Density-Matrix Theory for Many-electron Correlation 63
two complementary approaches to the direct calculation of the 2-RDM

have emerged: (i) variational minimization of the ground-state energy as
a functional of the 2-RDM [1744] and (ii) solution of the contracted
Schrodinger equation [1016, 4552], or its anti-Hermitian part [5366].
Furthermore, in class (i) of variational calculations two approaches to
constraining the 2-RDM have been developed: (a) the variational 2-
RDM methods in which the 2-RDM is constrained by necessary N-
representability conditions known as positivity conditions [8, 17], and (b)
the parametric variational 2-RDM methods in which the 2-RDM is con-
strained to be nearly N-representable by its parametrization [6775].
4.2. Variational 2-RDM Method
After showing that the energy of a many-electron atom or molecule can be

written as a linear functional of the 2-RDM in Section 4.2.1, we develop
a systematic hierarchy of N-representability constraints for the 2-RDM
known as p-positivity conditions in Section 4.2.2. Minimizing the ground-
state energy with respect to the 2-RDM constrained by positivity conditions
requires a special type of optimization known as semidefinite programming,
which we develop in Section 4.2.3. Finally, in Section 4.2.4 we apply the
variational 2-RDM method to studying strong electron correlation in acene
chains and hydrogen lattices.
4.2.1. Energy as a 2-RDM functional

Because electrons interact pairwise, the many-electron Hamiltonian for any
atom or molecule can be written as

2 i,j
H = Kk,l ai aj al ak , (4.1)
i,j,k,l
where the a and the a are the second-quantized creation and annihilation
operators, the indices refer to members of a spin-orbital basis set, and the
two-electron reduced Hamiltonian matrix 2 K is the matrix representation
of the operator

1 1 Zj 1
2
K = 12 + 1 . (4.2)
N 1 2 r1j 2 r12
j
64 D.A. Mazziotti
The expectation value of the Hamiltonian operator yields the many-electron

energy
i,j i,j
E= 2
Kk,l 2 Dk,l (4.3)
E = Tr(2 K 2 D) (4.4)
as a functional of the reduced Hamiltonian matrix and the two-electron

reduced density matrix (2-RDM), where
i,j
2
Dk,l = |ai aj al ak |. (4.5)
Both the energy as well as the one- and two-electron properties of an atom
or molecule can be computed from a knowledge of the 2-RDM. To perform
a variational optimization of the ground-state energy, we must constraint
the 2-RDM to derive from integrating an N-electron density matrix. These
necessary yet sufficient constraints are known as N-representability condi-
tions [1, 2].
4.2.2. Positivity conditions

General p-particle N-representability conditions on the 2-RDM are
derivable from metric (or overlap) matrices. From the ground-state wave-
function | and a set of p-particle operators {Ci1 ,i2 ,...,ip }, a set of basis
functions can be defined
i1 ,i2 ,...,ip | = |Ci1 ,i2 ,...,ip (4.6)
for which the metric (or overlap) matrix M with elements

i ,i ,...,i
Mj11 ,j22 ,...,jpp = i1 ,i2 ,...,ip |j1 ,j2 ,...,jp (4.7)
= |Ci1 ,i2 ,...,ip Cj1 ,j2 ,...,jp | (4.8)
must be positive semidefinite. We indicate that a matrix has this property

by the notation M 0. For a p-RDM, which is parameterized by a wave-
function these vector-space restrictions are always satisfied. More gen-
erally, however, these conditions, known as p-positivity conditions [17, 19,
25, 26, 34, 35], offer a systematic approach for imposing N-representability
conditions on an RDM without using the wavefunction.
4.2.2.1. Two-positivity
When p = 2, we may choose the Ci,j in three distinct ways: (i) to create
one particle in the jth orbital and one particle in the ith orbital, that is
Ci,j = ai aj , (ii) to annihilate one particle in the jth orbital and one particle
in the ith orbital (or create holes in each of these orbitals), Ci,j = ai aj , and
(iii) to annihilate one particle in the jth orbital and create one particle in
the ith orbital, that is Ci,j = ai aj . These three choices for the Ci,j produce
the following three different metric matrices for the 2-RDM:
i,j
2
Dk,l = |ai aj al ak |, (4.9)
i,j
2
Qk,l = |ai aj al ak |, (4.10)
i,j
2
Gk,l = |ai aj al ak |, (4.11)
which must be positive semidefinite if the 2-RDM is N-representable

[8,17,34].All three matrices contain equivalent information in the sense that
rearranging the creation and annihilation operators produces linear map-
pings between the elements of the three matrices; particularly, the two-hole
RDM 2 Q and the particle-hole RDM 2 G may be written in terms of the
two-particle RDM 2 D as follows
i,j i,j j i,j
2
Qk,l = 2 2 Ik,l 4 1 Dki 1 Il + 2 Dk,l (4.12)
and
2 i,j j i,l
Gk,l = 1 Il 1 Dki 2 Dk,j , (4.13)
where 1 I and 2 I are the one- and two-particle identity matrices and
denotes the Grassmann wedge product [14]. While all three matrices are
interconvertible, the nonnegativity of the eigenvalues of one matrix does
not imply the nonnegativity of the eigenvalues of the other matrices, and
hence, the restrictions 2 Q 0 and 2 G 0 provide two important N-
representability conditions in addition to 2 D 0. These conditions physi-
cally restrict the probability distributions for two particles, two holes, and
one particle and one hole to be nonnegative with respect to all unitary trans-
formations of the one-particle basis set. Collectively, the three restrictions
are known as the two-positivity conditions [17, 19, 25, 26, 34].
66 D.A. Mazziotti
Because 2 D 0 and 2 Q 0 imply 1 D 0 and 1 Q 0 by contraction

1 2 i,j
Dk =
1 i
Dk,j (4.14)
N 1
j
1 i,j
Qk =
1 i 2
Qk,j (4.15)
rN 1
j
the two-positivity conditions imply the one-positivity conditions. The r in
the contraction of the two-hole RDM denotes the rank of the one-particle
basis set. In general, the p-positivity imply the q-positivity conditions for
q p. The one-positivity conditions from the metric matrices for the one-
particle and one-hole RDMs, 1 D and 1 Q, restrict the occupation numbers
ni (or eigenvalues) of the 1-RDM to lie in the interval ni [0, 1]. Coleman
showed this condition on the eigenvalues to be both necessary and sufficient
for the N-representability of the 1-RDM [7].
4.2.2.2. Three-positivity
The conditions that a 3-RDM be three-positive follow from writing the
operators in Eq. (4.6) as products of three second-quantized operators [17,
19,34,76]. The resulting basis functions lie in four vector spaces according
to the number of creation operators in the product. Basis functions between
these vector spaces are orthogonal because they are contained in Hilbert
spaces with different numbers of particles. The four metric matrices that
must be constrained to be positive semidefinite for three-positivity [17] are
given by
3 i,j,k
Dp,q,r = |ai aj ak ar aq ap | (4.16)
3 i,j,k
Ep,q,r = |ai aj ak ar aq ap | (4.17)
3 i,j,k
Fp,q,r = |ai aj ak ar aq ap | (4.18)
p,q,r = |ai aj ak ar aq ap |.
3
Qi,j,k
(4.19)
Physically, because 3 D is the metric (or overlap) matrix of basis functions

in which three particles have been killed, the condition that 3 D by pos-
itive semidefinite 3 D 0 restricts the probability distribution for three
particles to be nonnegative. Because 3 E is the metric matrix of basis func-
tions in which two particles and one hole have been killed, the condition
3 E 0 restricts the probability distribution for two particles and one hole
to be nonnegative. Similarly, the conditions 3 F 0 and 3 Q 0 restrict the

probability distributions for one particle and two holes and three holes
to be nonnegative.
As in Eqs. (4.12) and (4.13) for the two-positive metric matrices, the
three-positive metric matrices are connected by linear mappings which can
be derived by rearranging the second-quantized operators. For example,
the mapping from 3 D to 3 Q may be written with the Grassmann wedge
product [14, 77] as
p,q,r = 6 Ip,q,r 18 Dp Iq,r + 9 Dp,q Ir Dp,q,r .

3 3 i,j,k 1 i 2 j,k 2 i,j 1 k 3 i,j,k
Qi,j,k (4.20)
where 1 I, 2 I, and 3 I are the one-, two-, and three-particle identity matrices.
Similar mappings can be derived to express 3 E and 3 F as functionals of 3 D.
Contraction of the three-positivity matrices in Eq. (4.16) generates the one-
and two-positivity metric matrices, and hence, the three-positivity condi-
tions imply the one- and two-positivity conditions. A 2-RDM is defined to
be three-positive if it arises from the contraction of a three-positive 3-RDM:
2 i,j 1 3 i,j,k
Dp,q = Dp,q,k . (4.21)
N 2
k
The three-positivity conditions have been examined in variational 2-RDM
calculations on spin [17, 76, 78] and molecular [34, 35] systems where they
give highly accurate energies and 2-RDMs.
4.2.3. Semidefinite programming

Variational calculation of the energy with respect to the 2-RDM constrained
by two-positivity conditions requires minimizing the energy in Eq. (4.3)
while restricting the 2 D, 2 Q, and 2 G to be not only positive semidefinite but
also interrelated by the linear mappings in Eqs. (4.124.15). This is a special
optimization problem known as a semidefinite program. The solution of a
semidefinite program is known as semidefinite programming [37, 7981].
In the mid-1990s a powerful family of algorithms, known as primal-
dual interior-point algorithms, was developed for solving semidefinite pro-
grams [79]. The phrase interior-point means that the method keeps the trial
primal and dual solutions on the interior of the feasible set throughout
the solution process. In these algorithms a good initial guess for the 2-
RDM is a scalar multiple of the two-particle identity matrix. Advantages
of the interior-point methods are: (i) rapid quadratic convergence from
the identity matrix to the optimal 2-RDM for a set of positivity condi-
tions, and (ii) a rigorous criterion in the duality gap for convergence to
the global minimum. These benefits, however, are accompanied by large
68 D.A. Mazziotti
memory requirements and a significant number of floating-point operations

per iteration, specifically O(nm3 + n2 m2 ) where n is the number of vari-
ables and m is the number of constraints. With m and n proportional to the
number of elements in the 2-RDM ( r4 ), the method scales approximately
as r16 where r is the rank of the one-particle basis set [19, 21]. The vari-
ational 2-RDM method has been explored for minimal basis sets with the
primal-dual interior-point algorithm, but the computational scaling signif-
icantly limits both the number of active electrons and the size of the basis
set [1821, 23, 24].
The author has developed a large-scale semidefinite programming algo-
rithm for solving the semidefinite program in the variational 2-RDM
method [25, 26, 34, 37]. The optimization challenge in the 2-RDM method
is to constrain the metric matrices to be positive semidefinite while the
ground-state energy is minimized. The algorithm constrains the solution
matrix M to be positive semidefinite by a matrix factorization
M = RR , (4.22)
where for the two-positivity conditions M contains the 2 D, 2 Q, and
2 G matrices. Such a matrix factorization was previously considered in
the context of 2-RDM theory by Rosina [82], Harriman [83], and the
author [14], and it was recently employed for solving large-scale semidef-
inite programs in combinatorial optimization [84]. The applications in
Mazziotti [25, 26] and Burer and Choi [85] are the first to apply the matrix
factorization to semidefinite programs with multiple diagonal blocks in
the solution matrix M. The linear constraints, including the trace, the
contraction, and the interrelations between the metric matrices, become
quadratic in the new independent variables R. Therefore, the factorization
in Eq. (4.22) converts the semidefinite program into a nonlinear program
where the energy must be minimized with respect to R, while nonlinear
constraint equalities are enforced.
We solve the nonlinear formulation of the semidefinite program by
the augmented Lagrange multiplier method for constrained nonlinear opti-
mization [25, 26, 37, 84, 86]. Consider the augmented Lagrangian function
1
L(R) = E(R) i ci (R) + ci (R)2 , (4.23)

i i
where R is the matrix factor for the solution matrix M, E(R) is the ground-
state energy as a function of R, {ci (R)} is the set of equality constraints, {i }
is the set of Lagrange multipliers, and is the penalty parameter. For an
appropriate set of multipliers {i }, the minimum of the Lagrangian function

with respect to R corresponds to the minimum of the energy E(R) subject
to the nonlinear constraints ci (R). The positive third term in the augmented
Lagrangian function, known as the quadratic penalty function, tends to zero
as the constraints are satisfied.
The cost of the algorithm is dominated by r 6 floating-point opera-
tions [25], mainly from the matrix multiplication of the block-diagonal
R matrix with itself, where r is the rank of the one-particle basis set.
Storage of the factorized 2-RDM, several copies of its gradient, and the
Lagrange multipliers scales as r4 . In comparison with the primal-dual
interior-point approach, which scales as r 16 and r8 in floating-point opera-
tions and memory storage, the first-order nonlinear algorithm for the vari-
ational 2-RDM method [25, 26, 37] provides a significant improvement in
computational efficiency.
4.2.4. Applications
Because the N-representability conditions are independent of a reference
wavefunction, the variational 2-RDM method can capture strong electron
correlation effects in molecules. To illustrate this ability, we discuss pre-
vious applications of the variational 2-RDM method to (i) the dissociation
of the N2 molecule [34], (ii) the prediction of the metal-to-insulator tran-
sition in the H64 lattice [41], and (iii) the emergence of polyradical character
in acene chains [38].
4.2.4.1. Dissociation of the N2 molecule

The ground-state energy of the nitrogen molecule as a function of bond
length is examined with 2-RDM and wavefunction methods, coupled cluster
with single, double, and triple excitations (CCSDT) and full configuration
interaction (FCI) in Fig. 4.1 [34]. Minimal Slater-type orbital basis sets
(STO-6G) are employed, and all valence electrons are correlated. The
variational lower-bound 3POS curve is essentially indistinguishable from
the FCI curve. The 2-RDM method with three-positivity (3POS) has a

maximum error of 1.4 mH at R = 1.7 A. The two-positivity (2POS) and
two-positivity plus T1 T2 conditions yield maximum errors of 23.6 mH and

4.6 mH at R = 1.7 A. Around equilibrium the three-positivity improves
the energies from two-positivity plus T1 T2 and two-positivity by one and
two orders of magnitude respectively, and it is an order of magnitude more
accurate than CCSDT.
70 D.A. Mazziotti
108.5 3POS
CCSDT
Energy (a.u.)
108.55
FCI
108.6
108.65
108.7
1 1.2 1.4 1.6 1.8 2

N-N distance
Fig. 4.1. This figure compares the 3POS and CCSDT potential energy curves
denoted by dashed lines with the FCI curve denoted by a solid line. The variational
lower-bound 3POS curve is essentially indistinguishable from the FCI curve. Bond

distance is reported in Angstroms (A).
Fig. 4.2. Sketch of the 444 hydrogen cube.
4.2.4.2. Metal to insulator transition in the H64 lattice

For the 444 cube, RHF and MP2 provide incorrect dissociation limits,
and the coupled-cluster calculations diverge for non-equilibrium inter-

atomic distances R 1.5 A. In contrast, as shown in Fig. 4.2, the 2-RDM
method ensures that upon dissociation the cluster energies per atom con-
verge to the energy of an isolated H atom in the given basis set [41]. Impor-
tantly, the dissociation of the 444 cube would require approximately
1018 equally weighted determinants in a conventional wavefunction cal-
culation, which is a billion times more determinants than treatable with a
standard multi-reference self-consistent-field calculation.
The metal-to-insulator transition for the 444 cube is shown in
Fig. 4.4 [41]. In the variational 2-RDM method the magnitude of the
off-diagonal 1-RDM elements in the atomic-orbital basis set, measured by
0.3
RHF
MP2
Energy per atom (a.u.) 0.35 DQG
0.4
0.45
0.5
0.55
0.5 1 1.5 2 2.5 3 3.5
R ()
Fig. 4.3. Potential energy curve for the symmetric dissociation of the 444
hydrogen cube, reported per atom, as a function of the distance between the closest
atoms.
0.25
RHF
DQG
0.2
0.15

0.1
0.05
0
0.5 1 1.5 2 2.5 3 3.5
R ()
Fig. 4.4. Metal-to-insulator transition in the 444 hydrogen cube under the
change of the distance R between closest atoms.
harmonic average , decays as the interatomic distance R increases while

in the RHF method the harmonic average rapidly converges to a positive
limit. Note that the aggregate measure decreases more gradually than the
measure 1 D1n for two hydrogen atoms at diagonal vertices.
4.2.4.3. Polyradical character of the acene chains

The active spaces necessary to describe the electrons in the n-acene
series where n = 2 8 (see Fig. 4.5) become prohibitively large for n > 4
72 D.A. Mazziotti
Fig. 4.5. Sketch of general acene chains of length n.
2
2acene
3acene
4acene
5acene
Occupation number
1.5 6acene
7acene
8acene
0.5
0
5 10 15 20 25 30
Orbital index
Fig. 4.6. Natural orbital occupation numbers for the n-acene series (n = 2 8).
The basis set is double- and calculations are performed with an active space that
includes the 4n + 2 lowest lying molecular orbitals.
because the dimension of the CI Hamiltonian scales exponentially with the

number of electrons. For example, even a calculation of tetracene (4-acene)
would require computing the lowest eigenvector of a Hamiltonian with
dimension 1.12 108 , which already presents a formidable challenge
in terms of both storage and time for traditional CASSCF approaches.
While a CASSCF study of octacene would demand the diagonalization of
a Hamiltonian of dimension 1.47 1017 , the approximate ground-state
energy and 2-RDM can be readily computed within the framework of the
variational 2-RDM method with two-positivity conditions.
In [40] we studied the change in the natural-orbital occupation numbers
with both chain length and basis-set size. Half of the available orbitals of
the double- basis set were included in the active space. Figure 4.6 shows the
natural-orbital occupation numbers for acenes with chain lengths n ranging
from two to eight. The onset of biradical character is evident because the
difference between the occupation numbers of the highest-occupied and
lowest-unoccupied natural orbitals decreases with increasing chain length.

The difference is as large as 1.67 for napthalene and as small as 0.36 for
octacene. Furthermore, around n = 8 we begin to observe the emergence
of polyradical character in the chains. While the increase in basis set from
STO-3G to double- substantially decreases the energy, it does not sig-
nificantly change either the occupation numbers or the trends in radical
character.
4.3. Contracted Schrodinger Theory
Integration (or contraction in a matrix formulation) of the density-matrix

version of the Schrodinger equation over all electrons save two produces the
contracted Schrodinger equation (CSE) [1016,4552]. The CSE was first
obtained in a coordinate representation in 1976 by Cohen and Frishberg [87]
and Nakatsuji [88]. The anti-Hermitian part of the CSE (ACSE) was first
derived by Harriman [89] in 1979; later in 1979 the ACSE was obtained by
Kutzelnigg [49, 90] who called it the generalized Brillouin condition. The
ACSE enforces the set of hypervirial relations for all one- and two-body
operators, which were developed by Hirshfelder [91].
The CSE is an important ingredient for 2-RDM methods because it is
a stationary-state condition for both ground and excited states [40]. By
Nakatsujis theorem, if the RDMs in the CSE are N-representable, then for
energetically non-degenerate states both ground and excited states
there is a one-to-one mapping between the solutions of the Schrodinger
equation and solutions of the CSE [14,88]. Although an exact Nakatsuji-like
theorem has not been proven for the ACSE, the ACSE implies a significant
part of the CSE [40]. Despite their importance as stationary-state conditions,
however, neither the CSE nor the ACSE can be solved for the 2-RDM
without additional information because they both depend on higher-particle
RDMs.
A significant advance occurred in the 1990s when it was recognized
that the indeterminacy of these equations could be removed by recon-
structing the higher RDMs as functionals of the 2-RDM [10, 12, 14]. In
1993 Colemenero, Perez de Valle, and Valdemoro [10] introduced a set
of reconstruction functionals based on particlehole duality, and in 1996
Nakatsuji and Yasuda [12] confirmed these functionals by Greens function
techniques. In 1998 Mazziotti [14, 15] systematized and generalized these
reconstructions by developing a cumulant theory for RDMs [15,16,45,47].
Cumulant reconstruction of RDMs has been applied to solving both the
CSE [10, 12, 14] and ACSE [5366].
74 D.A. Mazziotti
In Section 4.3.1 the ACSE and the cumulant reconstruction of the 3-

RDM required for its solution are presented, and in Section 4.3.2 we develop
a system of differential equations for solving the ACSE for both ground and
excited states. Finally, in the last section we discuss recent applications that
demonstrate some of the advantages of the ACSE method including: (i)
its treatment of excited states dominated by single, double, or even higher
excitations of the ground state [40, 66], (ii) its accurate dissociation of
chemical bonds with single-, double-, or triple-bond character [58, 61],
(iii) its applicability to arbitrary spin states [64], (iv) its size extensivity
due to the cumulant RDM reconstruction [15], (v) its balanced treatment
of single- and multi-reference correlation for accurate energy differences
between different molecular species and states, (vi) its quadratic scaling
ra2 in the number of active orbitals that enables the treatment of larger
active spaces than multi-reference configuration interaction with similar
or improved accuracy, and (vii) its automatic generation of all one- and
two-particle properties from the 1- and 2-RDMs without additional com-
putational effort.
4.3.1. ACSE and cumulant reconstruction

Projection of the N-electron Schrodinger equation onto all two-electron
transitions from the wavefunction yields the contracted Schrodinger
equation (CSE) [1, 10, 13, 14]
ij
|ai aj al akH| = E 2Dkl . (4.24)
If the CSE is separated into Hermitian and anti-Hermitian components [34],

both parts must vanish individually for Eq. (4.24) to be satisfied. The anti-
Hermitian portion of the CSE (ACSE) can be expressed as
|[ai aj al ak ,H]| = 0, (4.25)
where the square brackets indicate a commutator.

Because the Hamiltonian contains pairwise electronic interactions

2 p,q
H = Ks ap as +
1 p
Vs,t ap aq at as , (4.26)
p,s p,q,s,t
where 1K represents the kinetic and nuclear attraction energies of a single

p,q
electron and 2 Vs,t = pq|st is an electron-electron repulsion integral in
physicists notation [92], the ACSE depends only upon the 2- and 3-RDMs.
To eliminate the 3-RDM from the ACSE approximately, we can reconstruct
the 3-RDM from the 2-RDM according to its cumulant expansion [15, 16,
45, 47, 48, 93]

3 i,j,k j i,j j
Ds,t,u 6 1Dsi 1Dt 1Duk + 9 2Ds,t 2 1Dsi 1Dt 1Duk , (4.27)
where denotes the anti-symmetric Grassmann (or wedge) product [14,
94]. The missing term in the reconstruction, known as the connected (or
cumulant) part 3 of the 3-RDM, contains information not expressible as
wedge products of the 1- and 2-RDMs [15, 16, 45, 47, 48, 93]. Although
the connected 3-RDM can be approximated in terms of the 2-RDM, it is
neglected in the multi-reference formulation of the ACSE in [58].
The cumulant reconstruction [15, 16, 45, 47] is also an essential part of
the canonical transformation (CT) method [95, 96], which has been shown
to be a solution of the ACSE in the Heisenberg representation [54]. Despite
their theoretical connections, the ACSE and CT methods are practically
very different with distinct fundamental variables (the 2-RDM (ACSE)
versus an effective Hamiltonian (CT)), convergence behaviors, results, and
capabilities [54, 61]. In general, reconstruction is an important component
of any method within contracted Schrodinger theory, that is a theory using
the CSE, or a part of the CSE such as the ACSE, as a stationary-state
condition [40].
4.3.2. Solving the ACSE for ground and excited states

Consider a sequence of infinitesimal two-body unitary transformations of
an initial wavefunction ()
|( + ) = eS() |(), (4.28)
where the transformations are ordered by a continuous time-like variable .
For the transformation to be unitary, the two-body operator S, defined by

Ss,t () 2 s,t
2 p,q p,q
S() = (4.29)
p,q,s,t
with
2 i,j
k,l = ai aj al ak (4.30)
must be anti-Hermitian, S = S.

In the limit, as shown in [5355, 58], the changes in the energy

and its 2-RDM with are governed by the following differential equations:
dE
= ()|[H, S()]|(). (4.31)
d
76 D.A. Mazziotti
and
i,j
d 2 Dk,l
= ()|[2 k,l , S()]|().
i,j
(4.32)
d
To minimize the energy along , we select the following elements of the
p,q
two-particle matrix 2 Ss,t (), which minimize dE/d along its gradient
with respect to these elements [34]:
Sr,s () = ()|[2 s,t ,H]|().
2 p,q p,q
(4.33)
Importantly, the left side of Eq. (4.33) is simply the residual of the ACSE.
If the residual in the ACSE vanishes, the unitary transformations become
the identity operator, and the energy and 2-RDM cease to change with .
Using the cumulant reconstruction of the 3-RDM in Eq. (4.27) permits
us to express these equations approximately in terms of the 2-RDM.
Hence, Eqs. (4.31)(4.33) collectively provide a system of differential
equations [5355, 58] for evolving an initial 2-RDM to a final 2-RDM
that solves the ACSE for stationary states. In practice, the equations are
evolved in until either (i) the energy or (ii) the least-squares norm of
the ACSE increases. The ACSE can be seeded with an initial 2-RDM from
either (i) a HartreeFock calculation or (ii) any correlated calculation (i.e. a
multi-configuration self-consistent-field (MCSCF) calculation [58]). Con-
vergence to the ACSEs solution is efficient in both cases [53, 54, 58].
As demonstrated in the recent extension of the ACSE to excited
states [63], even though the unitary rotations are selected in Eq. (4.33)
to minimize the energy, the system of differential equations in Eqs. (4.31)
(4.33) is capable of producing energy and 2-RDM solutions of the ACSE
for both ground and excited states. Because excited states correspond to
local energy minima of the ACSE and the gradient in Eq. (4.33) leads to a
local rather than global energy minimum, an excited-state solution can be
readily obtained from a good guess for the initial 2-RDM. A guess will be
good when it is closer to the minimum of the desired solution of the ACSE
than to any other minimum. Such 2-RDM guesses can be generated from
multi-configuration self-consistent-field (MCSCF) calculations. The initial
MCSCF 2-RDM directs the optimization of the ACSE to a desired excited
state because it contains important multi-reference correlation effects that
identify the state.
Seeding the ACSE with an MCSCF 2-RDM yields a balanced treatment
of both single- and multi-reference correlation [40, 58, 61, 64, 65]. Because
the ACSE with reconstruction incorporates many high orders of a renor-
malized perturbation theory, its energies are significantly more accurate
than those from second or third orders of a multi-reference many-body
perturbation theory [40, 58, 61, 64]. Furthermore, in the absence of strong
correlation the ACSE can be compared to coupled cluster methods where it
yields energies that are between the accuracies of coupled cluster with
single and double excitations (CCSD) and coupled cluster with single,
double, and triple excitations (CCSDT) [55]. In addition to its balance of
moderate and strong correlation effects, the ACSE has advantages in com-
putational scaling. It scales like r6 , where r is the rank of the one-electron
basis set, but its accuracy is between that of CCSD and CCSDT where the
latter scales as r7 . Moreover, while multi-reference wavefunction methods
scale exponentially with the number ra of active orbitals of the active space,
the ACSE only scales quadratically ra2 [58, 61]. This significant reduction
in computational cost allows the ACSE to treat larger active spaces than
traditional methods.
4.3.3. Applications
Applications of the ACSE to the ground state have been made to a number
of systems and reactions including: (i) the electrocyclic ring-opening of
bicyclobutane to gauche-1,3-butadiene [60], (ii) the relative energies of the
cis-trans isomers of HO 3 [57], (iii) the sigmatropic shift of hydrogen in
propene and acetone enolate [61], and (iv) the study of vinylidene carbene
reactions [65]. These calculations demonstrate that the ACSE yields a
balanced description of single- and multi-reference (strong) correlation
effects in both the presence or absence of strong electron correlation. In
contrast, traditional wavefunction methods tend to be optimal in either
the presence (multi-reference perturbation methods) or absence (coupled-
cluster methods) of strong correlation. An equally accurate description of
correlation in both limits is extremely important in practical applications
where energy differences must be computed between molecular species or
states with significantly different degrees of electron correlation.
4.3.3.1. Energy barriers of bicyclobutanes transition states

In the ring opening of bicyclobutane to gauche-1,3-butadiene the energy
barrier for the disrotatory pathway includes the energy of a strongly cor-
related biradial transition state [60]. The ACSE method enables the direct
calculation of multi-reference correlation energies and 2-RDMs without the
fully correlated many-electron wavefunction. Qualitatively, the Woodward
Hoffmann rules indicate that the electrocyclic reaction of bicyclobutane to
form gauche-1,3-butadiene prefers the conrotatory pathway to the disro-
tatory pathway. The solution of the ACSE in the 6-311G** basis set predicts
78 D.A. Mazziotti
Fig. 4.7. The ring opening of bicyclobutane to gauche-1,3-butadiene can occur

by conrotatory or disrotatory pathways where the energy barrier for the disrotatory
pathway includes the energy of a strongly correlated biradial transition state. The
solution of the ACSE in the 6-311G** basis set predicts 41.2 and 55.7 kcal/mol
reaction barriers for the conrotatory and disrotatory pathways respectively.
41.2 and 55.7 kcal/mol reaction barriers for the conrotatory and disrotatory
pathways, respectively (Fig. 4.7). The ACSE energy barrier of 55 kcal/mol
appears to resolve a 10 kcal/mol energy discrepancy between coupled
cluster and multi-reference perturbation methods in the literature [60].
4.3.3.2. Conical intersections in CH2 s triplet excited states

Because excited states generally contain more multi-reference correlation
than ground states, the ACSE method is especially applicable to treating
both the energies and properties of excited states. In 2009 the ACSE
methodology was extended to treat both excited states [63] and arbi-
trary spin states [64]. Initial benchmark calculations included the excited
states of hydrogen fluoride where the ACSE matched the accuracy of the
computationally more expensive multi-reference configuration-interaction
method with single and double excitations plus Davidsons Q correction
(MRCI+Q) [63]. Importantly, unlike MRCI+Q the ACSE is size extensive,
that is it has the correct scaling with system size. Calculations have shown
that this property becomes important in only slightly larger molecular
systems. Recently, the generalization of the ACSE for excited states and
arbitrary spin has been applied to computing (i) the conical intersection
in the triplet excited states of methylene [66] and (ii) the photochemical
conversion of the allyl radical to the cyclopropyl radical.
The conical intersection between methylenes excited triplet states 1
3A and 2 3 B is shown in Fig. 4.8. In each plot the points represent the
2 1
38.4 3
MCSCF 1 B1
3
MCSCF 13A2
38.5 MCSCF 2 B1
3
Full CI 1 B1
Full CI 13A2
3
38.6 Full CI 2 B1
Energy (H)
38.7
38.8
38.9
39
39.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
R ()
(a)
38.4 3
MRMP2 1 B1
3
MRMP2 1 A2
38.5 MRMP2 23B1
3
Full CI 1 B1
Full CI 13A2
3
38.6 Full CI 2 B1
Energy (H)
38.7
38.8
38.9
39
39.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
R ()
(b)
Fig. 4.8. (Continued)

80 D.A. Mazziotti
38.4 3
ACSE 1 B1
3
ACSE 1 A2
3
38.5 ACSE 2 B1
3
Full CI 1 B1
3
Full CI 1 A2
3
38.6 Full CI 2 B1
Energy (H)
38.7
38.8
38.9
39
39.1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
R ()
(c)
Fig. 4.8. The (a) MCSCF (top), (b) MRMP2 (middle), and (c) ACSE (bottom)
potential energy curves for the 1 3 B1 , 1 3A2 , and 2 3 B1 states of methylene, as
functions of R, plotted against those from FCI, given by data points. Reproduced
from [66].
values computed by full configuration interaction (FCI) in the correlation-

consistent polarized valence double-zeta (cc-pVDZ) basis set. The FCI
results are compared with those from (a) multi-reference self-consistent
field (MCSCF), (b) second-order multi-reference perturbation theory
(MRPT2), and (c) the ACSE. The ACSE improves the accuracy of MCSCF
by two orders of magnitude and MRMP2 by one order of magnitude. The
average differences in the MCSCF, MRMP2, and ACSE energies from
FCI are 93.89, 12.27, and 1.51 mH respectively [66]. Similar accuracy was
obtained at the conical intersection. The location of the intersection com-
puted by the ACSE corroborates Yarkonys results [97] with a high degree
of accuracy. Furthermore, the ACSE produces nearly N-representable 2-
RDMs from which both one- and two-particle properties can be computed
without any additional calculations.
4.4. Parametric 2-RDM Method
Both the variational 2-RDM method in Section 4.2 and the contracted
Schrodinger methods in Section 4.3 have the ability to capture strong
electron correlation that is especially difficult for traditional wavefunction

methods. However, it would also be useful to have a 2-RDM-based
method that is designed for treating moderate electron correlation with
a combination of speed and accuracy which exceeds that conventional
wavefunction-based methods. In this section we develop a parametric varia-
tional 2-RDM method [6775] in which the calculations can be performed
at the speed of configuration interaction with single and double excita-
tions (CISD) with an accuracy approaching that of the more expensive
coupled cluster methods (CCSDT) where single, double, and triple exci-
tations are included. Although for simplicity in this section we derive
a family of parametric 2-RDM methods starting from a model wave-
function [67], these methods have recently been derived from con-
tractions of the cumulant reconstructions in Section 4.3 [75], which
provides a rigorously size extensive theory based on the properties of
cumulant RDMs.
4.4.1. Parametrization of the 2-RDM

Consider the correlated wavefunction | with double excitations from a
HartreeFock wavefunction |0 :
1 ab
| = c0 |0 + cij aa ab aj ai |0 , (4.34)
4
ijab
known as configuration interaction with double excitations (CID), where

i, j and a, b denote occupied and unoccupied spin orbitals respectively and
the second-quantized operators ai (ai ) create (destroy) an electron in the
spin orbital i. With the CID wavefunction the 2-RDM can be parameterized
in terms of the expansion coefficients {c0 , cijab }. Energies from a CID wave-
function, or any truncated configuration interaction wavefunction, are not
size extensive. The lack of size extensivity arises from the following class
of 2-RDM elements [67, 68].

1 cd 2
2
Dijab = cijab 1 |ckl | . (4.35)
4
klcd
When {i, j} {k, l} = and {a, b} {c, d} = , the 2-RDM contains

unconnected terms that cause the energy to scale quadratically with the
number N of electrons.
82 D.A. Mazziotti
Equation (4.35) can be corrected by inserting a tensor f that removes

the terms with an incorrect scaling:

1 ijkl cd 2
2
Dijab = cijab 1 fabcd |ckl | . (4.36)
4
klcd
Because of its role, this tensor has been called the topological factor [67
69, 98]. If all elements of f are set to one, we obtain CID as in Eq. (4.35),
but if all elements of f are set to zero, we obtain a coupled electron-
pair approximation (CEPA). While the CID energy is not size extensive,
its 2-RDM is N-representable; in contrast, the energy from CEPA is size
extensive, but its 2-RDM is not N-representable. Selection of an optimal f
requires us to consider the N-representability of the 2-RDM.
The most important N-representability conditions are the two-positivity
conditions. The two-positivity conditions imply N-representability condi-
tions known as the CauchySchwarz inequalities. From nonnegativity of
2 D and 2 Q we have
ij ij
(2 Dab )2 2 Dij 2 Dab
ab
(4.37)
ij ij
(2 Qab )2 2 Qij 2 Qab
ab . (4.38)
These inequalities become equalities in two cases: (i) when correlation

is absent, both sides of the inequalities trivially vanish and (ii) when the
number of particles (holes) equals two, Eq. (4.37) (Eq. (4.38)) becomes
an equality. Equation (4.37) (Eq. (4.38)) is an inequality mainly due to
unconnected terms (terms scaling as N p where p 2) that appear on the
right side when the number of particles (holes) is greater than two. Hence,
if the 2-RDM parametrization from Eqs. (4.34) and (4.36) is substituted
into Eqs. (4.37) and (4.38), we can equate the connected parts of the left
and right sides to solve for two topological factors (D and Q) [67, 75].
Repeating this process for an average of the four inequalities from the 2 G
matrix provides a more direct derivation of Kollmars topological factor
(K) [67, 75].
Table 4.1 defines the elements of the topological factors for CID and
ijkl
CEPA as well as D, Q, and K. Because a topological factor fabcd only
depends on the number of indices shared by cijab and ckl cd in Eq. (4.36), its
possible values can be divided into nine classes, labeled by no /nv , where
no is the number of occupied orbitals shared by {ij} and {kl} and nv is
Table 4.1. Elements of the topological factors for CID, CEPA,

and the D, Q, K, and M parametric 2-RDM methods defined.
ijkl
Topological factors fabcd (or fno /nv )
2-RDM
methods 0/0 1/0 2/0 0/1 0/2 1/1 2/1 1/2 2/2
CID 1 1 1 1 1 1 1 1 1
CEPA 0 0 0 0 0 0 0 0 0
D 0 1 1 0 0 1 1 1 1
Q 0 0 0 1 1 1 1 1 1
K 0 1/2 1 1/2 1 3/4 1 1 1
M 0 0 1 0 1 1 1 1 1
the number of virtual orbitals shared by {ab} and {cd}. For the energy
functional to be size extensive, the topological factor must vanish for the
class no /nv = 0/0. Unlike the other factors, the D and Q factors do not
maintain particle-hole symmetry, that is in general fno /nv = fnv /no . To
restore particle-hole symmetry with exact results for two particles or two
holes (when single excitations are included), we must set the other classes
to one, except for 0/1, 1/0, and 1/1. Because either D or Q has a factor of
zero for 0/1 or 1/0 while both D and Q have factors of one for 1/1, we set
the factor to one for 1/1 and zero for 0/1 and 1/0. These choices generate a
new topological factor (M) in Table 4.1 [67, 75].
4.4.2. Applications
4.4.2.1. Correlation energies at equilibrium geometries
Correlation energies from parametric 2-RDM methods as well as traditional
wavefunction methods are reported in Table 4.2 for several molecules in
the polarized quadruple-zeta (cc-pVQZ) basis set [75, 99]. Molecules NH3
and HCN are given in the correlation-consistent polarized triple-zeta (cc-
pVTZ) basis set [99]. The K and M methods recover much more correlation
energy than CISD, which is not size extensive. Furthermore, the M method
improves significantly upon CCSD with energies that are closer to those
from CCSD(T). The K method improves slightly upon CCSD. The 2-RDMs
from the parametric methods are nearly N-representable; for example, with
the M method for N2 the lowest eigenvalues of 2 D, 2 Q, and 2 G, 5.0
104 , 3.0 104 , and 4.1 104 , are 3-to-4 orders of magnitude
smaller than the largest eigenvalues.
84 D.A. Mazziotti
Table 4.2. Correlation energies from parametric 2-RDM methods as well as tra-
ditional wavefunction methods are reported for molecules in the cc-pVQZ basis
set except for NH3 and HCN in the cc-pVTZ basis set. The M 2-RDM methods
improves significantly upon CCSD. Energies are given in Hartrees (H).
Correlation energy (H)

Wavefunction methods 2-RDM Methods
HF
Molecules Energy (H) CCSD CCSD(T) CISD K M
H2 O 76.0648 0.2860 0.2950 0.2744 0.2868 0.2904

CH2 38.8947 0.1712 0.1765 0.1647 0.1729 0.1761
N2 108.9911 0.3931 0.4133 0.3657 0.3957 0.4032
CO 112.7888 0.3805 0.3990 0.3556 0.3837 0.3906
NH3 56.2179 0.2476 0.2553 0.2368 0.2487 0.2522
HCN 92.9081 0.3492 0.3671 0.3237 0.3518 0.3586
100.15
Energy (a.u.)
100.2
CCSD
2-RDM (K)
100.25 2-RDM (M)
CCSD(T)
CR-CC(T)
100.3
1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8

H-F distance
Fig. 4.9. The potential energy curves for hydrogen fluoride in the cc-pVQZ basis
set from the K, M, CCSD, CCSD(T), and CR-CC(2,3) methods. The energy results
of the M functional are nearly indistinguishable from those of the computationally
more expensive CR-CC(2,3). The length of the H-F bond is given in Angstroms

(A).
4.4.2.2. Dissociation of the HF molecule

In Fig. 4.9 the dissociation curve for hydrogen fluoride is presented in a
cc-pVQZ basis set [75, 99]. Because FCI calculations are not available for
such a large basis set, we compare the energy errors from the K and M
parametric 2-RDM methods and the CCSD and CCSD(T) coupled-cluster
methods relative to the energies from the completely renormalized coupled
cluster method with perturbative triple excitations [CR-CC(2,3)] [100].
Both CCSD(T) and CR-CC(2,3), scaling approximately as r 7 , contain a

perturbative treatment of triple excitations. For hydrogen fluoride at 2.8 A,
the CCSD, CCSD(T), K, and M methods yield energy errors of 27.5,
34.6, 15.3, and 2.6 mHs respectively. Furthermore, the NPEs from CCSD,
CCSD(T), K, and M are 19.1, 33.4, 7.8, and 2.1 mHs. The M functional
performs better in this basis set than in the smaller 6-311G** basis set.
Figure 4.9 displays the potential energy curves for hydrogen fluoride from
the K, M, CCSD, CCSD(T), and CR-CC(2,3) methods. The energy results
of the M functional are nearly indistinguishable from those of the compu-
tationally more expensive CR-CC(2,3).
4.4.2.3. Isomerization of nitrosomethane

The parametric 2-RDM method with the K parametrization was recently
applied to the isomerization of nitrosomethane to trans-formaldoxime,
which can occur by (i) a single 1,3-hydrogen shift or (ii) two successive
1,2-hydrogen shifts [74]. The potential energy surface from the 2-RDM
method with the K functional in the aug-cc-pVTZ basis set is shown in
Fig. 4.10, where we present energies in kcal/mol of the stationary points
relative to nitrosomethane. The solid line follows the channel describing
successive 1,2-hydrogen shifts while the dashed line follows a single 1,3-
hydrogen shift to cis-formaldoxime followed by a rotation with a barrier
of about 3 kcal/mol to trans-formaldoxime, which is globally the lowest
energy species on the surface. We note that the barrier of the 1,3-shift is
slightly higher than the barrier of the first 1,2-shift, with heights relative to
nitrosomethane of 65.0 and 55.2 kcal/mol, respectively as predicted by the
2-RDM method.
4.5. Looking Ahead
The collection of 2-RDM methods offers a new paradigm for the com-
putation of electron correlation in quantum systems [1]. While the wave-
function scales exponentially in the number N of electrons, the 2-RDM
scales polynomially in N. Consequently, for many-electron quantum
systems the 2-RDM theory offers a significant reduction in computational
cost even in the presence of strong electron correlation. The 2-RDM has
long been employed as a tool for analysis of quantum information, but, as
discussed in the Introduction, efforts to compute the 2-RDM directly were
hindered by the N-representability problem, that is the 2-RDM must be
constrained to correspond to an N-electron system [7].
86 D.A. Mazziotti
Fig. 4.10. Critical points on the potential energy surface for the isomerization of
nitrosomethane to trans-formaldoxime as computed by the 2-RDM method in the
aug-cc-pVTZ basis set. The dashed line represents a 1,3-hydrogen shift; the solid
line represents successive 1,2-shifts. The figure shows that 1,2-shift is energetically
more favorable than the 1,3-shift by about 10 kcal/mol. All relative energies are
reported in kcal/mol.
Recent advances [1] have enabled the direct computation of the 2-RDM
without the many-electron wavefunction by the methods discussed in this
chapter: (i) the constrained and parametric variational 2-RDM methods
and (ii) the solution of the contracted Schrodinger equation or its anti-
Hermitian part. Importantly, as seen with the acene chains [38] and the
hydrogen lattices [41], these 2-RDM approaches permit the treatment
of strong electron correlation in systems that are too large to treat with
traditional electronic structure methods. Although recent wavefunction
methods for strong correlation such as density-matrix renormalization
group are often limited to systems with a well-defined, one-dimensional
ordering of the electronic orbitals (i.e. linear systems) [101], the variational
2-RDM method is applicable to a broader range of molecules including
systems with arbitrary orbital orderings and geometries. The 2-RDM-based
methods have been applied to study: (i) chemical reactions and materials
[38, 40, 41, 6066, 73, 74], (ii) quantum phase transitions [104, 105], (iii)
motions of electrons and nuclei [39, 106, 107], (iv) molecular conduc-
tivity [102, 103], and (v) high-temperature superconductivity [108].
While significant progress has been made, there remain many important
opportunities for further advancements in theory and applications. A sam-
pling of future extensions of recent work might include: (i) improvements
in the computational efficiency of the first-order semidefinite-programming
algorithms [25, 26, 37], (ii) enhancements of existing linear-scaling para-
metric 2-RDM methods [73] for the better treatment of medium-to-large
molecular systems, and (iii) generalizations of existing non-equilibrium
steady-state ACSE methods [102, 103] to treat electron correlation in
molecular conductors explicitly. It is hoped that the present chapter may
serve as a starting point for these and other new developments in 2-RDM
mechanics that will further enhance our ability to study and understand
quantum molecular systems and processes.
Acknowledgments
The author expresses his appreciation to Dudley Herschbach, Herschel

Rabitz, John Coleman, and Alexander Mazziotti for their support and
encouragement. He thanks the members of his research group, past and
present, who have contributed to the development of 2-RDM methods.
Finally, the author also gratefully acknowledges the NSF, the Henry-
Camille Dreyfus Foundation, the Alfred P. Sloan Foundation, Microsoft
Corporation, and the David-Lucile Packard Foundation for their support.
Bibliography
[1] Reduced-Density-Matrix Mechanics: With Application to Many-electron Atoms and

Molecules, edited by D. A. Mazzioti (John Wiley & Sons, New York, 2007).
[2] A.J. Coleman and V.I. Yukalov, Reduced Density Matrices: Coulsons Challenge
(Springer-Verlag, New York, 2000).
[3] P.O. Lowdin, Phys. Rev. 97, 1474 (1955).
[4] J.E. Mayer, Phys. Rev. 100, 1579 (1955).
[5] R.H. Tredgold, Phys. Rev. 105, 1421 (1957).
[6] C.A. Coulson, Rev. Mod. Phys. 32, 170 (1960).
[7] A.J. Coleman, Rev. Mod. Phys. 35 668 (1963).
[8] C. Garrod and J. Percus, J. Math. Phys. 5, 1756 (1964).
[9] R.M. Erdahl, Int. J. Quantum Chem. 13, 697 (1978).
[10] F. Colmenero, C. Perez del Valle, and C. Valdemoro, Phys. Rev. A 47, 971 (1993);
F. Colmenero and C. Valdemoro, Phys. Rev. A 47, 979 (1993).
88 D.A. Mazziotti
[11] F. Colmenero and C. Valdemoro, Int. J. Quantum Chem. 51, 369 (1994).
[12] H. Nakatsuji and K. Yasuda, Phys. Rev. Lett. 76, 1039 (1996).
[13] K. Yasuda and H. Nakatsuji, Phys. Rev. A 56, 2648 (1997).
[14] D.A. Mazziotti, Phys. Rev. A 57, 4219 (1998).
[15] D.A. Mazziotti, Chem. Phys. Lett. 289, 419 (1998); Int. J. Quantum Chem. 70, 557
(1998).
[16] D.A. Mazziotti, Phys. Rev. A 60, 3618 (1999); 4396 (1999).
[17] D.A. Mazziotti and R. M. Erdahl, Phys. Rev. A 63, 042113 (2001).
[18] M. Nakata, H. Nakatsuji, M. Ehara, M. Fukuda, K. Nakata, and K. Fujisawa, J.
Chem. Phys. 114, 8282 (2001).
[20] M. Nakata, M. Ehara, and H. Nakatsuji, J. Chem. Phys. 116, 5432 (2002).
[22] G. Gidofalvi and D.A. Mazziotti, Phys. Rev. A 69, 042511 (2004).
[23] T. Juhasz and D.A. Mazziotti, J. Chem. Phys. 121, 1201 (2004).
[24] Z. Zhao, B.J. Braams, H. Fukuda, M.L. Overton, and J.K. Percus, J. Chem. Phys.
120, 2095 (2004).
[25] D.A. Mazziotti, Phys. Rev. Lett. 93, 213001 (2004).
[26] D.A. Mazziotti, J. Chem. Phys. 121, 10957 (2004).
[27] G. Gidofalvi and D.A. Mazziotti, J. Chem. Phys. 122, 094107 (2005).
[31] J.R. Hammond and D.A. Mazziotti, Phys. Rev. A 73, 012509 (2006).
[32] D.A. Mazziotti, Acc. Chem. Res. 39, 207 (2006).
[33] G. Gidofalvi and D.A. Mazziotti, J. Phys. Chem. A 110, 5481 (2006); J. Chem.
Phys. 125, 144102 (2006).
[36] M. Nakata, B.J. Braams, K. Fujisawa, M. Fukuda, J.K. Percus, M. Yamashita, and
Z. Zhao, J. Chem. Phys. 128, 164113 (2008).
[37] D.A. Mazziotti, Math. Model. Num. Anal. 41, 249 (2007).
[39] E. Kamarchik and D.A. Mazziotti, Phys. Rev. A 79, 012502 (2009).
[40] L. Greenman and D.A. Mazziotti, J. Chem. Phys. 130, 184101 (2009).
[41] A.V. Sinitskiy, L. Greenman, and D.A. Mazziotti, J. Chem. Phys. 133, 014104
(2010).
[42] M.V. Mihailovic and M. Rosina, Nucl. Phys. A237, 221 (1975).
[43] C. Garrod, V. Mihailovic, and M. Rosina, J. Math. Phys. 10, 1855 (1975).
[44] R.M. Erdahl, Reports Math. Phys. 15, 147 (1979).
[45] W. Kutzelnigg and D. Mukherjee, J. Chem. Phys. 110, 2800 (1999).
[46] K. Yasuda, Phys. Rev. A 59, 4133 (1999).
[47] D.A. Mazziotti, Chem. Phys. Lett. 326, 212 (2000).
[48] D.A. Mazziotti in Many-electron Densities and Density Matrices, edited by
J. Cioslowski (Kluwer, Boston, 2000) pp. 139163.
[49] W. Kutzelnigg and D. Mukherjee, J. Chem. Phys. 114, 2047 (2001).
[50] D.A. Mazziotti, J. Chem. Phys. 116, 1239 (2002); Phys. Rev. E 65, 026704 (2002).
[51] M.D. Benayoun, A.Y. Lu, and D.A. Mazziotti, Chem. Phys. Lett. 387, 485 (2004).
[52] D.R. Alcoba, F.J. Casquero, L.M. Tel, E. Perez-Romero, and C. Valdemoro, Int. J.
Quantum Chem. 102, 620 (2005).
[55] D.A. Mazziotti, J. Chem. Phys. 126, 184101 (2007).
[56] C. Valdemoro, L.M. Tel, D.R. Alcoba, and E. Perez-Romero, Theor. Chem. Acc.
118, 503509 (2007).
[57] D.A. Mazziotti, J. Phys. Chem. A 111, 12635 (2007).
[59] C. Valdemoro, L.M. Tel, E. Perez-Romero, and D.R. Alcoba, Int. J. Quantum Chem.
108, 1090 (2008).
[60] D.A. Mazziotti, J. Phys. Chem. A 112, 13684 (2008).
[61] J.J. Foley IV, A.E. Rothman, and D.A. Mazziotti, J. Chem. Phys. 130, 184112
(2009).
[62] C. Valdemoro, D.R. Alcoba, L.M. Tel, and E. Perez-Romero, Int. J. Quantum Chem.
109, 2622 (2009).
[64] A.E. Rothman, J.J. Foley IV, and D.A. Mazziotti, Phys. Rev. A 80, 052508 (2009).
[65] L. Greenman and D.A. Mazziotti, J. Phys. Chem. A 114, 583 (2010).
[66] J.W. Snyder Jr., A.E. Rothman, J.J. Foley IV, and D.A. Mazziotti, J. Chem. Phys.
132, 154109 (2010).
[68] C. Kollmar, J. Chem. Phys. 125, 084108 (2006).
[69] A.E. DePrince III and D.A. Mazziotti, Phys. Rev. A. 76, 042501 (2007).
[70] A.E. DePrince III, E. Kamarchik, and D.A. Mazziotti, J. Chem. Phys. 128, 234103
(2008).
[71] A.E. DePrince III and D.A. Mazziotti, J. Phys. Chem. B 112, 16158 (2008).
[72] A.E. DePrince III and D.A. Mazziotti, J. Chem. Phys. 130, 164109 (2009).
[76] R.M. Erdahl and B. Jin in Many-electron Densities and Density Matrices, edited
by J. Cioslowski (Kluwer, Boston, 2000) pp. 5784.
[77] D.A. Mazziotti, Phys. Rev. E 65, 026704 (2002).
[78] J.R. Hammond and D.A. Mazziotti, Phys. Rev. A 71, 062503 (2005).
[79] L. Vandenberghe and S. Boyd, SIAM Review 38, 49 (1996).
[80] S. Wright, Primal-Dual Interior-Point Methods (SIAM, Philadelphia, 1997).
[81] Y. Nesterov and A.S. Nemirovskii, Interior Point Polynomial Method in Convex
Programming: Theory and Applications (SIAM, Philadelphia, 1993).
[82] M.V. Mihailovic and M. Rosina, Nucl. Phys. A 130, 386 (1969).
[83] J.E. Harriman, Phys. Rev. A 17, 1257 (1978).
[84] S. Burer and R.D.C. Monteiro, Math. Programm. Ser. B 95, 329 (2003).
[85] S. Burer and C. Choi, Optim. Methods Soft. 21, 493 (2006).
[86] R. Fletcher, Practical Methods of Optimization (John Wiley & Sons, New York,
1987).
[87] L. Cohen and C. Frishberg, Phys. Rev. A 13, 927 (1976).
90 D.A. Mazziotti
[88] H. Nakatsuji, Phys. Rev. A 14, 41 (1976).

[89] J.E. Harriman, Phys. Rev. A 19, 1893 (1979).
[90] W. Kutzelnigg, Chem. Phys. Lett. 64, 383 (1979).
[91] J.O. Hirschfelder, J. Chem. Phys. 33, 1462 (1960).
[92] A. Szabo and N. Ostlund, Modern Quantum Chemistry: Introduction to Advanced
Electronic Structure Theory (Dover, New York, 1996).
[93] A.E. DePrince and D.A. Mazziotti, J. Chem. Phys. 127, 104104 (2007).
[94] W. Slebodzinski, Exterior Forms and their Applications (Polish Scientific Pub-
lishers, Warsaw, 1970).
[95] T. Yanai and G.K. Chan, J. Chem. Phys. 124, 194106 (2006).
[96] G.K. Chan and T. Yanai, in Reduced-Density-Matrix Mechanics: With Application
to Many-electron Atoms and Molecules, edited by D. A. Mazziotti (John Wiley &
Sons, New York, 2007) pp. 343384.
[97] D.R. Yarkony, Rev. Mod. Phys. 68, 4 (1996).
[98] R. Ahlrichs, P. Scharf, and C. Ehrhardt, J. Chem. Phys. 82, 890 (1985).
[99] T.H. Dunning Jr., J. Chem. Phys. 90, 1007 (1989).
[100] P. Piecuch and M. Wloch, J. Chem. Phys. 123, 224105 (2005).
[101] J. Hachmann, W. Cardoen, and G. K. Chan, J. Chem. Phys. 125, 144101 (2006).
[102] J.E. Subotnik, T. Hansen, M.A. Ratner, and A. Nitzan, J. Chem. Phys. 130, 144105
(2009).
[103] A.E. Rothman and D.A. Mazziotti, J. Chem. Phys. 132, 104112 (2010).
[105] C.A. Schwerdtfeger and D.A. Mazziotti, J. Chem. Phys. 130, 224102 (2009).
[106] E. Kamarchik and D.A. Mazziotti, Phys. Rev. A 75, 013203 (2007).
[107] E. Kamarchik and D.A. Mazziotti, Phys. Rev. Lett. 99, 243002 (2007).
[108] J.C. Phillips, Proc. Nat. Acad. Sci. 107, 1307 (2010).
Chapter 5
Finite Size Scaling for Criticality of the

Schrdinger Equation
Sabre Kais
Department of Chemistry and Physics,

Purdue University, West Lafayette, IN 47907, USA
By solving the Schrodinger equation one obtains the whole energy

spectrum, both the bound and the continuum states. If the Hamiltonian
depends on a set of parameters, these could be tuned to a transition from
bound to continuum states. The behavior of systems near the threshold,
which separates bound-states from continuum states, is important in
the study of such phenomenon as: ionization of atoms and molecules,
molecule dissociation, scattering collisions, and stability of matter. In
general, the energy is non-analytic as a function of the Hamiltonian param-
eters or a bound-state does not exist at the threshold energy. The overall
goal of this chapter is to show how one can predict, generate and identify,
new class of stable quantum systems using large-dimensional models and
the finite size scaling approach. Within this approach, the finite size corre-
sponds not to the spatial dimension but to the number of elements in a com-
plete basis set used to expand the exact eigenfunction of a given Hamil-
tonian. This method is efficient and very accurate for estimating the critical
parameters, {i }, for stability of a given Hamiltonian, H(i ). We present
two methods of obtaining critical parameters using finite size scaling for
a given quantum Hamiltonian: the finite element method and the basis set
expansion method. The long term goal of developing finite size scaling
is treating criticality from first principles at quantum phase transitions. In
the last decade considerable attention has concentrated on a new class of
phase transitions, transitions which occur at the absolute zero of temper-
ature. These are quantum phase transitions which are driven by quantum
91
92 Sabre Kais
fluctuations as a consequence of Heisenbergs uncertainty principle. These

new transitions are tuned by parameters in the Hamiltonian. Finite size
scaling might be useful in predicting the quantum critical parameters for
systems going through quantum phase transitions.
5.1. Introduction
Weakly bound states represent an interesting field of research in atomic

and molecular physics. The behavior of systems near the binding threshold
is important in the study of ionization of atoms and molecules, molecule
dissociation, and scattering collisions. Moreover, the stability of atomic
and molecular systems in external electric, magnetic, and laser fields
is of fundamental importance in atomic and molecular physics and has
attracted considerable experimental and theoretical attention over the past
decades [14]. A superintense laser field can change the nature of atomic
and molecular systems and their anions; the stabilization in superstrong
fields is accompanied by splitting of the electron distribution into distinct
lobes, with locations governed by the quiver amplitude and polarization of
the laser field. This localization markedly alters electron-nucleus interac-
tions as well as reduces electron-electron repulsions and hence suppresses
autoionization. In molecules, it can also enhance chemical bonding. This
localization markedly reduces the ionization probability and can enhance
chemical bonding when the laser strength becomes sufficiently strong and
can give rise to new stable multiply charged negative ions such as H ,
He and H 2 [3, 57].
In general, the energy is non-analytical, an analytic function is a function
that is locally given by a convergent power series, as a function of the
Hamiltonian parameters or a bound-state does not exist at the threshold
energy. It has been suggested for some time, based on large-dimensional
models, that there are possible analogies between critical phenomena and
singularities of the energy [810].
Phase transitions are associated with singularities of the free energy.
These singularities occur only in the thermodynamic limit [11, 12] where
the dimension of the system approaches infinity. However calculations are
done only on finite systems. A finite size scaling (FSS) approach is needed
in order to extrapolate results from finite systems to the thermodynamic
limit [13]. FSS is not only a formal way to understand the asymptotic
behavior of a system when the size tends to infinity, but a theory that
also gives us numerical methods [1420] capable of obtaining accurate
results for infinite systems by studying the corresponding small systems
Finite Size Scaling for Criticality of the Schrodinger Equation 93
[2132]. Applications include expansion in Slater-type basis functions [30],

Gaussian-type basis functions [33] and recently finite elements [34].
5.2. Criticality for Large-dimensional Models
Large dimension models were originally developed for specific theories

in the fields of nuclear physics, critical phenomena and particle physics
[35, 36]. Subsequently, with the pioneering work of Herschbach [10, 37],
they found wide use in the field of atomic and molecular physics [38].
In this method one takes the dimension of space, D, as a variable, solves
the problem at some dimension D = 3, where the physics becomes much
simpler, and then uses perturbation theory or other techniques to obtain an
approximate result for D = 3 [10].
It is possible to describe stability and symmetry breaking of elec-
tronic structure configurations of atoms and molecules as phase transi-
tions and critical phenomena. This analogy was revealed by using dimen-
sional scaling method and the large-dimensional limit model of electronic
structure configurations [3942].
To study the behavior of a given system near the critical point, one has
to rely on model calculations which are simple, capture the main physics of
the problem and which belong to the same universality class [16, 17]. Here
we will illustrate the phase transitions and symmetry breaking using the
large dimension model. In the application of dimensional scaling to elec-
tronic structure, the large-D limit reduces to a semi-classical electrostatic
problem in which the electrons are assumed to have fixed positions relative
both to the nuclei and to each other in the D-scaled space [10]. This configu-
ration corresponds to the minimum of an effective potential which includes
Coulomb interactions as well as centrifugal terms arising from the gener-
alized D-dependence kinetic energy. Typically, in the large-D regime the
electronic structure configuration undergoes symmetry breaking for certain
ranges of nuclear charges or molecular geometries [46].
In order to illustrate the analogy between symmetry breaking and phase
transitions we present as an example the results for the two-electron atoms
in the HartreeFock (HF) approximation [39]. In the HF approximation
at the D limit, the dimensional-scaled effective Hamiltonian for
the two-electron atom in an external weak electric field E can be written
as [47, 48],

1 1 1 1 1 1
H = + 2 Z + + 2 E (r1 r2 ). (5.1)
2 r1 2 r2 r1 r2 (r1 + r22 )1/2
94 Sabre Kais
where r1 and r2 are the electron-nucleus radii, and Z is the nuclear charge.
The ground state energy at the large-D limit is then given by E (Z, E ) =
min{r1 ,r2 } H .
In the absence of an external electric field, E = 0; Herschbach and
coworkers [49] have found that these equations have a symmetric solution
with the two electrons equidistant from the nucleus, with r1 = r2 = r.
This symmetric solution represents a minimum in the region where all the
eigenvalues of the Hessian matrix are positive, Z Zc = 2. For values of
Z smaller than Zc , the solutions become unsymmetrical with one electron
much closer to the nucleus than the other (r1 = r2 ). In order to describe
this symmetry breaking, it is convenient to introduce new variables (r, )
of the form: r1 = r; r2 = (1 )r, where = (r1 r2 )/r1 = 0 measures
the deviation from the symmetric solution.
By studying the eigenvalues of the Hessian matrix, one finds that the
solution is a minimum of the effective potential for the range, 1 Z Zc .
We now turn to the question of how to describe the system near the critical
point. To answer this question, a complete mapping between this problem
and critical phenomena in statistical mechanics is readily feasible with the
following analogies:
nuclear charge (Z) temperature (T)
external electric field (E ) ordering field (h)
ground state energy (E (Z, E )) free energy (f(T, h))
asymmetry parameter () order parameter (m)
stability limit point (Zc , E = 0) critical point (Tc , h = 0).
Using the above scheme, we can define the critical exponents (, ,
and ) for the electronic structure of the two-electron atom in the following
way:
(Z, E = 0) (Z) ; Z 0
E (Z, E = 0) | Z | ; Z 0
(5.2)
E (Zc , ) sgn() ; 0

E |E =0 | Z | ; Z 0
where Z Z Zc . These critical exponents describe the nature of the
singularities in the above quantities at the critical charge Zc . The values
obtained for these critical exponents are known as classical or mean-field
critical exponents: = 21 ; = 2; = 3; = 1.
This analogy between symmetry breaking and phase transitions was
also generalized to include the large-dimensional model of the N-electron
atoms [40], simple diatomic molecules [41,43], both linear and planar one-
electron systems [42] as well as three-body Coulomb systems of the general
form ABA [44].
The above simple large-D picture helps to establish a connection to
phase transitions. However, the next question to be addressed is: how to
carry out such an analogy to D = 3? This question will be examined in the
subsequent sections using the finite size scaling approach.
5.3. Finite Size Scaling: A Brief History
Ice tea, boiling water and other aspects of two-phase coexistence are
familiar features of daily life.Yet phase transitions do not exist at all in finite
systems! They appear in the thermodynamic limit: the volume V and
particle number N in such a way that their ratio, which is the density
= N/V , approaches a finite quantity. In statistical mechanics, the exis-
tence of phase transitions is associated with singularities of the free energy
per particle in some region of the thermodynamic space. These singular-
ities occur only in the thermodynamic limit [11, 12]. This fact could be
understood by examining the partition function Z.

Z= eE(
)/kB T, (5.3)
microstate
where E(
) is the energies of the states, kB is the Boltzmann constant and
T is the temperature. For a finite system, the partition function is a finite
sum of analytical terms, and therefore it is itself an analytical function.
The Boltzmann factor is an analytical function of T except at T = 0. For
T > 0, it is necessary to take an infinite number of terms in order to obtain
a singularity in the thermodynamic limit [11, 12].
In practice, real systems have a large but finite volume and particle
numbers (N 1023 ), and phase transitions are observed. More dramatic
even is the case of numerical simulations, where sometimes systems with
only a few number (hundreds, or even tens) of particles are studied, and
critical phenomena are still present. Finite size scaling theory, which
was pioneered by Fisher [13], addresses the question of why finite systems
apparently describe phase transitions and what is the relation of this phe-
nomena with the true phase transitions in corresponding infinite systems.
Moreover, finite size scaling is not only a formal way to understand the
asymptotic behavior of a system when the size tends to infinity. In fact, the
theory gives us numerical methods capable of obtaining accurate results for
96 Sabre Kais
infinite systems by studying the corresponding small systems (see [1517]

and references therein).
In order to understand the main idea of finite size scaling, let us consider
a system defined in a D-dimensional volume V of a linear dimension L
(V = LD ). In a finite size system, if quantum effects are not taken into
consideration, there are in principle three length scales: the finite geometry
characteristic size L, the correlation length , which may be defined as
the length scale covering the exponential decay er/ with distance r of
the correlation function, and the microscopic length a which governs the
range of the interaction. Thermodynamic quantities thus may depend on
the dimensionless ratios /a and L/a. The finite size scaling hypothesis
assumes that, close to the critical point, the microscopic length drops out.
If in the thermodynamic limit, L , we consider that there is only
one parameter (say temperature T ) in the problem and the infinite system has
a second order phase transition at a critical temperature Tc , a thermodynamic
quantity G develops a singularity as a function of the temperature T in the
form:
G(T) = lim GL (T) |T Tc | , (5.4)
L
whereas it is regular in the finite system, GL (T) has no singularity.
When the size L increases, the singularity of G(T) starts to develop. For
example, if the correlation length diverges at Tc as:
(T) = lim L (T) |T Tc | , (5.5)
L
then L (T) has a maximum which becomes sharper and sharper, then FSS
ansatz assumes the existence of scaling function FK such that:

L
GL (T) G(T)FK , (5.6)
(T)
where FK (y) y/ for y 0+. Since the FSS ansatz, Eq. (5.6), should be
valid for any quantity which exhibits an algebraic singularity in the bulk,
we can apply it to the correlation length itself. Thus, the correlation length
in a finite system should have the form:
L (T) L (L1/ |T Tc |). (5.7)
The special significance of this result was first realized by Nightingale [51],
who showed how it could be reinterpreted as a renormalization group trans-
formation of the infinite system. The phenomenological renormalization
(PR) equation for finite systems of sizes L and L
is given by:
L (T) L
(T
)
= (5.8)
L L
and has a fixed point at T (L,L ). It is expected that the succession of points

{T (L,L ) } will converge to the true Tc in the infinite size limit.

The finite size scaling theory combined with transfer matrix calculations
had been, since the development of the phenomenological renormalization
in 1976 by Nightingale [51], one of the most powerful tools to study critical
phenomena in two-dimensional lattice models. For these models the par-
tition function and all the physical quantities of the system (free energy,
correlation length, response functions, etc) can be written as a function of
the eigenvalues of the transfer matrix [52]. In particular, the free energy
takes the form:
f(T) = T ln 1 (5.9)
and the correlation length is:
1
(T) = , (5.10)
ln (2 /1 )
where 1 and 2 are the largest and the second largest eigenvalues of the
transfer matrix. In this context, critical points are related to the degeneracy
of these eigenvalues. For finite transfer matrix, the largest eigenvalue is
isolated (non degenerated) and phase transitions can occur only in the limit
L , where the size of the transfer matrix goes to infinity and the
largest eigenvalues can be degenerated. In the next section, we will see that
these ideas of finite size scaling can be generalize to quantum mechanics,
in particular addressing the criticality of the Schrodinger equation.
5.4. Finite Size Scaling for the Schrdinger Equation
The finite size scaling method is a systematic way to extract the critical
behavior of an infinite system from analysis on finite systems [30]. It
is efficient and accurate for the calculation of critical parameters of the
Schrodinger equation. Lets assume we have the following Hamiltonian:
H = H0 + V , (5.11)
where H0 is -independent and V is the -dependent term. We are inter-
ested in the study of how the different properties of the system change when
the value of varies. A critical point, c , will be defined as a point for which
a bound state becomes absorbed or degenerate with a continuum.
Without loss of generality, we will assume that the Hamiltonian,
Eq. (5.11), has a bound state, E , for > c which becomes equal to
zero at = c . As in statistical mechanics, we can define some critical
98 Sabre Kais
exponents related to the asymptotic behavior of different quantities near

the critical point. In particular, for the energy we can define the critical
exponent as:
E ( c ) . (5.12)
+
c
The existence or absence of a bound state at the critical point is related to

the type of the singularity in the energy. Using statistical mechanics termi-
nology, we can associate first order phase transitions with the existence
of a normalizable eigenfunction at the critical point. The absence of such a
function could be related to continuous phase transitions [30].
In quantum calculations, the variational method is widely used to
approximate the solution of the Schrodinger equation. To obtain exact
results one should expand the exact wave function in a complete basis
set and take the number of basis functions to infinity. In practice, one trun-
cates this expansion at some order N. In the present approach, the finite
size corresponds not to the spatial dimension, as in statistical mechanics,
but to the number of elements in a complete basis set used to expand the
exact eigenfunction of a given Hamiltonian. We will compare two methods
to obtain the matrix elements needed to apply the FSS ansatz. The size of
our system for the basis set expansion will correspond to the dimension
of the Hilbert space. For a given complete basis set n , the ground-state
eigenfunction has the following expansion:

= an ()n , (5.13)
n
where n is the set of quantum numbers. We have to truncate the series at
order N and the expectation value of any general operator O at order N is
given by:

N
ON = an(N) am
(N)
On,m , (5.14)
n,m
where On,m are the matrix elements of O in the basis set {n }.
For the finite element method (FEM), the wavefunction n (r) in the n-th
element is expressed in terms of local shape functions. For our calculations,
we use Hermite interpolation polynomials with two nodes and three degrees
of freedom. This choice ensures the continuity of the wavefunction and its
first two derivatives. Then in n-th element the wavefunction is [34]:
2 =

n (r) = i (r)ni + i (r)ni + i (r)ni , (5.15)

i=1
with indicating the nodal index of the element; i = 1 for the left and i = 2
=
for the right border of the element. The functions i (r), i (r), and i (r) are

the (fifth degree) Hermite interpolation polynomials. Then ni , ni , and ni

are the undetermined values values of the wavefunction and its first and
second derivative on the nodal points. The size for the case of solving the
equation with the FEM will be the number of elements used.
Since O is not analytical at = c , then we define a critical exponent,
O , if the general operator has the following relation:
O ( c )O for +
c, (5.16)
where + c represents taking the limit of approaching the critical
point from larger values of . As in the FSS ansatz in statistical mechanics
[16,53], we will assume that there exists a scaling function for the truncated
magnitudes such that:
(N)
O O FO (N| c | ), (5.17)
with the scaling function FO being particular for different operators but all
having the same unique scaling exponent .
To obtain the critical parameters, we define the following function:

ln(ON N
/O )
O (; N, N
) = . (5.18)
ln(N
/N)
At the critical point, the expectation value is related to N as a power law,
O N O / , and Eq. (5.18) becomes independent of N. For the energy
operator O = H and using the critical exponent for the corresponding
exponent O we have:

H (c ; N, N
) = . (5.19)

In order to obtain the critical exponent from numerical calculations,
it is convenient to define a new function [30]:
H (; N, N
)
(, N, N
) = , (5.20)
H (; N, N
) V (; N, N
)

which at the critical point is independent of N and N

and takes the value
of . Namely, for = c and any values of N and N
we have
(c , N, N
) = , (5.21)
and the critical exponent is readily given by Eq. (5.19).
100 Sabre Kais
5.5. The Hulthen Potential
To illustrate the application of the FSS method in quantum mechanics, let

us give an example of the criticality of the Hulthen potential. The Hulthen
potential behaves like a Coulomb potential for small distances whereas for
large distances it decreases exponentially so that the capacity for bound
states is smaller than that of Coulomb potential. Thus, they have the same
singularity but shifted energy levels. They always lie lower in the Coulomb
case than in the Hulthen case, where there remains only space for a finite
number of bound states [55]. Here, we present the FSS calculations using
two methods: finite elements and basis set expansion; each used to obtaining
quantum critical parameters for the Hulthen Hamiltonian. First, we give the
analytical solution, then FSS with basis set expansion and finite element
solution.
5.5.1. Analytical solution

The Hulthen potential has the following form [54, 55]:
er/a
V(r) = , (5.22)
a2 1 er/a
where is the coupling constant, and a is the scaling parameter. For small
values of r/a the potential V(r) 1a /r, whereas for large values of r/a
the potential approaches zero exponentially fast, therefore the scale a in
the potential regulates the infinite number of levels that would otherwise
appear with a large-distance Coulomb behavior.
Schrodinger radial differential equation in the dimensionless variable
r = r/a becomes:
1 d2 er
+ ( 2
+ ) = 0. (5.23)
2 dr 2 1 er
We only consider the case for l = 0 for the Hulthen potential. Here we used
the abbreviations 2 = Ea2 0 (in atomic units m = h = 1). The com-
plete solutions for the wavefunctions are written in term of hypergeometric
functions as follows [55]:
= N0 er (1 er )2 F1 (2 + 1 + n, 1 n, 2 + 1; er ), (5.24)
1
where the normalization factor is given by N0 = [( + n)(2 + n)] 2
[(2 + n)/ (2 + 1)(n)]. It follows that the energy levels are given by:
1 (2 n2 )2
En = ; n = 1, 2, 3, . . . , nmax . (5.25)
a2 8n2
We can make the following comments concerning the energy levels

obtained for the Hulthen potential. There exists a critical value for the
coupling c to have the given energy levels, c = n2 /2. It follows directly
from the first observation that the number of levels nmax allowed is finite
and it depends on the size of the coupling constant n2max 2. As
the potential is well behaved, which canbe seen as follows: in this limit we
get the obvious inequality 2 2 n. It follow that we can set
2
0 in Eq. (5.24) to obtain:

0 = (1 er )2 F1 (1 + n, 1 n, 1; er ), (5.26)
which is the wave function at threshold. This wave function is not normal-
izable as expected when the energy exponent = 2, E ( c ) . For
the ground state, the asymptotic limit of the probability density for r >> 1
and c becomes:
P(r) er/ , | c | , (5.27)
with a characteristic length and exponent = 1. The Hulthen potential has
a finite capacity determined by the critical coupling, c . The potential admits
bound states between the range of values for the coupling: = [1/2, ).
5.5.2. Basis set expansion

For the Hulthen potential, the wavefunction can be expanded in the fol-
lowing Slater basis (see Chapter 7 for details on SRT):

1/4
n (r) = er/2 L(2)
n (r). (5.28)
(n + 1)(n + 2)
(2)
Ln (r) is the Laguerre polynomial of degree n and order two. The kinetic
term can be obtained analytically. However, the potential term needs to be
calculated numerically [57].
Figure 5.1 shows the results for the plot (, N, N
) as a function of
with different N and N
; all the curves will cross exactly at the critical
point.
5.5.3. Finite element method

The FEM is a numerical technique which gives approximate solutions
to differential equations. In the case of quantum mechanics, the differ-
ential equation is formulated as a boundary value problem [61,62]. For our
purposes, we are interested in solving the time-independent Schrodinger
102 Sabre Kais
equation with finite elements. We will require our boundary conditions to

be restricted to the Dirichlet type. For this problem, we will use two interpo-
lation methods: linear interpolation and Hermite interpolation polynomials
to solve for this potential.
We start by integration by parts and impose the boundary conditions for
the kinetic energy and reduce it to the weak form [34]:

1 2
r (r)
(r)dr. (5.29)
2 0
For the potential energy:

er
r 2 (r)(r) dr. (5.30)
0 1 er
We calculated the local matrix elements of the potential energy by using
a four point Gaussian quadrature to evaluate the integral. We set the cutoff
for the integration to rc . To include the integration to infinity, we added an
infinite element approximation. To do so, we approximate the solution of
the wave function in the region of [rc , ) to be an exponentially decaying
function with the form (r) = (rc )er .
The local matrices are then assembled to form the complete solution and
by invoking the variational principle on the nodal values i we obtain a gen-
eralized eigenvalue problem representing the initial Schrodinger equation:
Hij |j = Uij |j . (5.31)
The solution of Eq. (5.31) is achieved using standard numerical methods
(see Chapter 10 for details [63]).
5.5.4. Finite size scaling results

The finite size scaling equations are valid only as asymptotic expressions,
but unique values of c , , and can be obtained as a succession of values
as a function of N. The lengths of the elements are set h = 0.5. The plots
of , Fig. 5.1, the basis set expansion is giving values very close to the
analytical solution of the Hulthen potential. For the plot of for the FEM
estimation of c is producing results very close to the exact values using
Hermite interpolation. The intersection of these curves indicate the c on the
abscissa. The ordinate gives the critical exponents (in plots). In Fig. 5.2,
we observed the behavior of the pseudocritical parameters, (N) (N) (N)
c , c , c ,
as a function of 1/N. The three curves monotonically converge to limiting
values for the Hermite interpolation and the basis set expansion.
To check the validity of our finite size scaling assumptions, we per-
formed a data collapse [64] calculation of the Hulthen potential. In the
Fig. 5.1. Plot of , obtained by FSS method, as a function of . Using the number
of basis N from 8 to 48 in steps of two. For FEM the number elements used were
from 100 to 380 in steps of 20.
Fig. 5.2. Extrapolated values for the critical exponents and the critical parameter .
The solid red dots at 1/N = 0 are the extrapolated critical values. The left side
is the basis set method while the right is the FEM with Hermite interpolation
polynomials.
104 Sabre Kais
data collapse analysis, we examine the main assumption we have made

in Eq. (5.17) for the existence of a scaling function for each truncated
(N)
magnitude O with a unique scaling exponent .
Since the O(N)
is analytical in , then from Eq. (5.17) the asymptotic
behavior of the scaling function must have the form:
FO (x) xO / . (5.32)
For large values of N, at the c , we have:
O(N) (c ) N O / . (5.33)
Because the same argument of regularity holds for the derivatives of the
truncated expectation values, we have:

m O(N)
N (O m)/ , (5.34)
m
=c
O(N) is analytical in , then using Eq. (5.34), the Taylor expansion

could be written as:
O(N) () N O / GO (N 1/ ( c )), (5.35)
where GO is an analytical function of its argument. This equivalent

expression for the scaling of a given expectation value has a correct form
to study the data collapse in order to test FSS hypothesis. If the scaling
Eq. (5.17) or Eq. (5.35) holds, then near the critical point the physical quan-
tities will collapse to a single universal curve when plotted in the appropriate
form O(N) N O / against N 1/ ( c ). If the operator O is the Hamil-
tonian then we will have a data collapse when plotting E0 N / against
N 1/ ( c ). In Fig. 5.3 we plot the results corresponding to the basis
set method (right panel) and Hermite interpolation (left panel), which have
been calculated with c = 0.49999, = 1.9960 and = 0.99910 for the
basis set method and for the Hermite interpolation we have c = 0.50000,
= 2.00011 and = 1.000322. The data collapse study do in fact support
our FSS assumptions. We have conveniently summarized our results for the
critical parameters for the analytical, linear interpolation, Hermite interpo-
lation and the basis set expansion in Table 5.1.
We have successfully obtained the critical exponents and the critical
parameter for the Hulthen potential using FSS with the basis set method
and the FEM. The results are in excellent agreement with the analytical
Fig. 5.3. Data collapse study of the basis set method and FEM. The left is the
basis set method and the right being the FEM.
Table 5.1. Critical parameters for the hulthen

potential.
Analytical Linear Hermite Basis Set
0.5 (exact) 0.50184 0.50000 0.49999

2 (exact) 1.99993 2.00011 1.9960
1 (exact) 1.00079 1.00032 0.99910
solution even for the very simplistic linear interpolation used for the FEM
calculations. However, the ability of the FEM to describe the wavefunction
locally in terms of elements affords a very natural way to extend its use for
FSS purposes.
5.6. Finite Size Scaling and Criticality of M-electron Atoms
Let us examine the criticality of the M-electrons atomic Hamiltonian as a

function of the nuclear charge Z. The scaled Hamiltonian takes the form:
M M
1 1 1
H() = i2 + , (5.36)
2 ri rij
i=1 i<j=1
where rij are the interelectron distances, and = 1/Z is the inverse of
the nuclear charge. For this Hamiltonian, a critical point means the value
of the parameter, c , for which a bound state energy becomes absorbed or
degenerate with the continuum.
106 Sabre Kais
To carry out the FSS procedure, one has to choose a convenient basis
set to obtain the two lowest eigenvalues and eigenvectors of the finite
Hamiltonian matrix. For M = 2, one can choose the following basis set
functions:
1 j
ijk, (x1 , x2 ) = r1i r2 e(r1 +r2 ) + r1 r2i e(r1 +r2 ) r12
j k
F (12 ,
),
2
(5.37)
where and are fixed parameters; we have found numerically that = 2
and = 0.15 is a good choice for the ground state [21], r12 is the interelec-
tronic distance and F (12 ,
) is a suitable function of the angle between
the positions of the two electrons 12 and the Euler angles
= (, , ).
This function F is different for each orbital-block of the Hamiltonian. For
the ground state F0 (12 ,
) = 1 and F1 (12 ,
) = sin(12 ) cos() for the
2p23 P state. These basis sets are complete for each -subspace. The com-
plete wave function is then a linear combination of these terms multiplied
by variational coefficients determined by matrix diagonalization [21]. In the
truncated basis set at order N, all terms are included such that N i+j +k.
Using FSS calculations with N = 6, 7, 8, . . . , 13 gives the extrapolated
values of c = 1.0976 0.0004 which is in excellent agreement with the
best estimate of c = 1.09766079 using large-order perturbation calcula-
tions [65]. Since the critical charge Zc = 1/c 0.91 indicates that the
hydrogen anion H is stable, Z = 1 > Zc .
For three-electron atoms, M = 3, one can repeat the FSS procedure
with the following Hylleraas-type basis set [22]:

j l m n (r1 +r2 ) r3
ijklmn (x1 , x2 , x3 ) = CA r1i r2 r3k r12 r23 r31 e e 1 , (5.38)
where the variational parameters, = 0.9 and = 0.1, were chosen to
obtain accurate results near the critical charge Z 2, 1 is the spin function
with spin angular moment 1/2:
1 = (1)(2)(3) (1)(2)(3), (5.39)
C is a normalization constant and A is the usual three-particle antisym-
metrizer operator [22]. The FSS calculations gives c = 0.48 0.03. Since
Zc 2.08 the anions He and H are unstable.
One can extend this analysis and calculate the critical charges for
M-electron atoms in order to perform a systematic check of the stability
of atomic dianions. In order to have a stable doubly negatively charged
atomic ion one should require the surcharge, Se (M) M Zc (M) 2. We
have found that the surcharge never exceeds two. The maximal surcharge,
Se (86) = 1.48, is found for the closed-shell configuration of element Rn
and can be related to the peak of electron affinity of the element N = 85.
The FSS numerical results for M-electron atoms show that at most, only
one electron can be added to a free atom in the gas phase. The second extra
electron is not bound by singly charged negative ion because the combined
action of the repulsive potential surrounding the isolated negative ion and
the Pauli exclusion principle. However, doubly charged atomic negative
ions might exist in a strong magnetic field of the order few atomic units,
where 1 a.u. = 2.3505109 G and superintese laser fields.
5.7. Conclusions
In this chapter, we have shown how the finite size scaling ansatz can be
combined with the variational method to extract information about critical
behavior of quantum Hamiltonians. This approach is based on taking the
number of elements in a complete basis set or the finite element method as
the size of the system. As in statistical mechanics, finite size scaling can then
be used directly to the Schrodinger equation. This approach is general and
gives very accurate results for the critical parameters, for which the bound
state energy becomes absorbed or degenerate with a continuum. To illus-
trate the applications in quantum calculations, we have presented detailed
calculations for the simple case of Hulthen potential and few electron atoms.
For atomic systems we have shown that finite size scaling can be used to
explain and predict the stability of atomic anions: at most, only one electron
can be added to a free atom in the gas phase.
Recently, there has been an ongoing experimental and theoretical search
for doubly charged negative molecular dianions [1]. In contrast to atoms,
large molecular systems can hold many extra electrons because the extra
electrons can stay well separated. However, such systems are challenging
from both theoretical and experimental points of view. The present finite
size scaling approach might be useful in predicting the general stability of
molecular dianions.
The approach can be generalized to complex systems by calculating
the matrix elements needed for FSS analysis by ab initio, density func-
tional methods, orbital free density functional (OF-DFT) [66,67] approach,
density matrices [68, 69] and other electronic structure methods [70]. The
implementation should be straightforward. We need to obtain the matrix
elements to calculate a as a function of the number of elements used in
solving for the system. In the finite element using mean field equations
(like HartreeFock or KohnSham methods) the solution region will be
discretized into elements composed of tetrahedrons.
108 Sabre Kais
The field of quantum critical phenomena in atomic and molecular

physics is still in its infancy and there are many open questions about the
interpretations of the results including whether or not these quantum phase
transitions really do exist. The possibility of exploring these phenomena
experimentally in the field of quantum dots [71] and systems in superin-
tense laser fields [72] offers an exciting challenge for future research. This
finite size scaling approach is general and might provide a powerful way
in determining critical parameters for the stability of atomic and molecular
systems in external fields, and for design and control electronic properties
of materials using artificial atoms.
The critical exponents calculated with finite size scaling indicate the
nature of the transitions from bound to continuum states. Study of the
analytical behavior of the energy near the critical point show that the open
shell system, such as the lithium-like atoms, is completely different from
that of a closed shell system, such as the helium-like atoms. The transition in
the closed shell systems from a bound state to a continuum resemble a first-
order phase transition, E ( c )1 , while for the open shell system, the
transition of the valence electron to the continuum is a continuous phase
transition, E ( c )2 . For closed shell systems, one can show that
H(c ) has a square-integrable eigenfunction corresponding to a threshold
energy, the existence of a bound state at the critical coupling constant c
implies that for < c , E() approaches E(c ) linearly in ( c ) as
c . However, for open shell systems, the wave function is not square-
integrable at at c . This difference in critical exponents might be helpful in
developing a new atomic classification schemes based on the type of phase
transitions and criticality of the system.
Acknowledgments
I would like to thank Pablo Serra, Juan Pablo Neirotti, Marcelo Carignano,
Winton Moy and Qi Wei for their valuable contributions to this ongoing
research of developing and applying finite size scaling to quantum problems
and Ross Hoehn for critical reading of the chapter. I would also like to thank
the Army Research Office (ARO) for financial support of this project.
Bibliography
[1] M.K. Scheller, R.N. Compton, and L.S. Cederbaum, Science 270, 1160 (1995).
[2] V.G. Bezchastov, P. Schmelcher, and L.S. Cederbaum, Phys. Chem. Chem. Phys. 5,
4981 (2003).
[3] M. Gavrila, in Atoms in Super Intense Laser Fields, edited by M. Gavrila (Academic,
New York, 1992), p. 435.
[4] Q. Wei, S. Kais and N. Moiseyev J. Chem. Phys. 124, 201108 (2006).
[5] E. van Duijn and H.G. Muller, Phys. Rev. A 56, 2182 (1997).
[6] E. van Duijn and H.G. Muller, Phys. Rev. A 56, 2192 (1997).
[7] Q. Wei, S. Kais, and D. Herschbach, J. Chem. Phys. 127, 094301 (2007).
[8] F.H. Stillinger and D.K. Stillinger, Phys. Rev. A 10, 1109 (1974).
[9] J. Katriel and E. Domany, Int. J. Quantum Chem. 8, 559 (1974).
[10] D.R. Hershbach, J. Avery, and O. Goscinsky, Dimensional Scaling in Chemical
Physics (Kluwer, Dordercht, 1993).
[11] C.N. Yang and T.D. Lee, Phys. Rev. 87, 404 (1952).
[12] T.D. Lee and C.N. Yang, Phys. Rev. 87, 410 (1952).
[13] M.E. Fisher, in Critical Phenomena, Proceedings of the 51st Enrico Fermi Summer
School, Varenna, Italy, edited by M.S. Green (Academic, New York, 1971); M.E.
Fisher and M.N. Barber, Phys. Rev. Lett. 28, 1516 (1972).
[14] B. Widom, in Critical Phenomena in Fundamental Problems in Statistical
Mechanics, edited by E.G.D. Cohen (Elsevier, New York, 1975).
[15] M.N. Barber, in Phase Transitions and Critical Phenomena Vol. 8, edited by C. Domb
and J.L. Lebowits (Academic, London, 1983).
[16] V. Privman, Finite Size Scaling and Numerical Simulations of Statistical Systems
(World Scientific, Singapore, 1990).
[17] J.L. Cardy, Finite-Size Scaling (Elsevier Science Publishers, New York, 1988).
[18] M.P. Nightingale, Physica 83A, 561 (1976).
[19] P.J. Reynolds, H.E. Stanley, and W. Klein, J. Phys. A 11, L199 (1978).
[20] P.J. Reynolds, H.E. Stanley, and W. Klein, Phys. Rev. B 21, 1223 (1980).
[21] J.P. Neirotto, P. Serra, and S. Kais, Phys. Rev. Lett., 79, 3142 (1997).
[22] P. Serra, J.P. Neirotti, and S. Kais, Phys. Rev. Lett. 80, 5293 (1998).
[23] S. Kais, J.P. Neirotti, and P. Serra, Int. J. Mass Spectrometry 182/183, 23 (1999).
[24] P. Serra, J.P. Neirotti, and S. Kais, Phys. Rev. A 57, R1481 (1998).
[25] P. Serra, J.P. Neirotti, and S. Kais, J. Chem. Phys. 102, 9518 (1998).
[26] J.P. Neirotti, P. Serra, and S. Kais, J. Chem. Phys. 108, 2765 (1998).
[27] Q. Shi and S. Kais, Mol. Phys. 98, 1485 (2000).
[28] S. Kais and Q Shi, Phys. Rev. A62, 60502 (2000).
[29] S. Kais, and P. Serra, Rev. Phys. Chem. 19, 97 (2000).
[30] S. Kais, and P. Serra, Adv. Chem. Phys. 125, 1 (2003).
[31] P. Serra, and S. Kais, Chem. Phys. Lett. 372, 205209 (2003).
[32] A. Ferron, P. Serra, and S. Kais, J. Chem. Phys. 120, 84128419 (2004).
[33] W. Moy, P. Serra, and S. Kais, Mol. Phys. 106, 203 (2008).
[34] W. Moy, M. Carignano, and S. Kais, J. Phys. Chem. 112, 54485452 (2008).
[35] For reviews see A. Chartterjee, Phys. Reports 186, 249 (1990).
[36] E. Witten, Phys. Today 33 (7), 38 (1980).
[37] D.R. Herschbach, J. Chem. Phys. 84, 838 (1986).
[38] C.A. Tsipis, V.S. Popov, D.R. Hershbach, and J.S. Avery, New Methods in Quantum
Theory (Kluwer, Dordrecht 1996).
[39] P. Serra and S. Kais, Phys. Rev. Lett. 77, 466 (1996).
[40] P. Serra and S. Kais, Phys. Rev. A 55, 238 (1997).
[41] P. Serra and S. Kais, Chem. Phys. Lett. 260, 302 (1996).
110 Sabre Kais
[42] P. Serra and S. Kais, J. Phys. A. 30, 1483 (1997).

[43] Q. Shi, S. Kais, F. Remacle, and R.D. Levine, ChemPhysChem 2, 434 (2001).
[44] Q. Shi and S. Kais, Int. J. Quantum Chem. 85, 307 (2001).
[45] S. Kais, S.M. Sung, and D.R. Hershbach, Int. J. Quan. Chem. 49, 657 (1994).
[46] D.D. Frantz and D. R. Herschbach, Chem. Phys. 126, 59 (1988).
[47] J.G. Loeser, J. Chem. Phys. 86, 5635 (1987).
[48] M. Cabrera, A.L. Tan, and J.G. Loeser, J. Phys. Chem. 97, 2467 (1993).
[49] D.Z. Goodson, and D.R. Hershbach, J. Chem. Phys. 86, 4997 (1987)
[50] H.E. Stanley, Introduction to Phase Transitions and Critical Phenomena (Oxford
University Press, New York, 1971).
[51] M.P. Nightingale, Physica 83A, 561 (1976).
[52] C.J. Thompson, Classical Statistical Mechanics (Clarendon Press, London, 1988).
[53] B. Derrida, B. Derrida, and L. De Seze, J. Physique 43, 475 (1982).
[54] L. Hulthen, Arkiv f or Matematik Astronomi och Fysik 28A, 5 (1942).
[55] S. Flugge, Practical Quantum Mechanics (Spring-Verlag, New York, 1974).
[56] P.E. Hoggan, Chapter 7 of this book.
[57] E. Antillon, W. Moy, Q. Wei, and S. Kais, J. Chem. Phys. 131, 104105 (2009).
[58] S. Kais and D.R. Herschbach, J. Chem. Phys. 100, 4367 (1994).
[59] M. Le Bellac, Quantum and Statistical Field Theory, (Oxford University Press,
Oxford, 1991).
[60] K.S. Gupta and S.G. Rajeev, Phys. Rev. D 48, 5940, (1993).
[61] Darrel W. Pepper and Juan C. Heinrich, The Finite Element Method (Taylor & Francis,
New York, 2006).
[62] J.N. Reddy, An Introduction to the Finite Element Method (McGraw-Hill, St. Louis,
1993).
[63] T.L. Beck and J.H. Dedrick, Chapter 10 of this book.
[64] P. Serra and S. Kais, Chem. Phys. Lett. 319, 273 (2000).
[65] J.D. Baker, D.E. Freund, R.N. Hill, and J.D. Morgan III, Phys. Rev. A 41, 1247 (1990).
[66] V. Lignres and E.A. Carter, in Handbook of Materials Modeling, edited by S.Yip
(Springer, Dordrecht, 2005) pp. 137148.
[67] V. Gavini, J. Knap, K. Bhattacharya, and M. Ortiz, J. Mech. Phys. Solids 55(4), 669
(2007).
[68] D.A. Mazziotti, Acc. Chem. Res. 39, 207 (2006).
[69] D. A. Mazziotti, J. Chem. Phys. 121, 10957 (2004).
[70] Ortiz, J.V. Int. J. Quantum Chem. 26, 1 (1992).
[71] J.X. Wang and S. Kais, Phys. Rev. B66, 081101 (2002).
[72] Q. Wei, S. Kais, and D. Herschbach, J. Chem. Phys. 129, 214110 (2008).
Chapter 6
The Generalized Sturmian Method
James Avery and John Avery
Department of Computer Science and Department of Chemistry,

University of Copenhagen, 1017, Copenhagen, Denmark
The generalized Sturmian method makes use of basis sets that are solutions
to an approximate wave equation with a weighted potential. The weighting
factors are chosen in such a way as to make all the members of the basis
set isoenergetic. In this chapter we will show that when the approximate
potential is taken to be that due to the attraction of the bare nucleus, the
generalized Sturmian method is especially well suited for the calculation
of large numbers of excited states of few-electron atoms and ions. Using
the method we shall derive simple closed-form expressions that approx-
imate the excited state energies of ions. The approximation improves with
increasing nuclear charge. The method also allows automatic generation
of near-optimal symmetry adapted basis sets, and it avoids the Hartree
Fock SCF approximation. Programs implementing the method may be
freely downloaded from our website, sturmian.kvante.org [1].
6.1. Description of the Method
6.1.1. The introduction of Sturmians into quantum theory

One of the very early triumphs of quantum theory was the exact solution
of the Schrodinger equation for hydrogen-like atoms:

1 2 Z
En n,l,m (x) = 0 (6.1)
2 r
111
112 J. Avery and J. Avery
In Eq. (6.1) and throughout the chapter, atomic units are used. The energies
and wavefunctions are given respectively by
Z2
En = , n = 1, 2, 3, . . . (6.2)
2n2
and
n,l,m (x) = Rn,l (r)Yl,m (, ) (6.3)
Here Yl,m (, ) is a spherical harmonic, and
R1,0 (r) = 2(Z/1)3/2 eZr/1
R2,0 (r) = 2(Z/2)3/2 (1 Zr/2)eZr/2
2
R2,1 (r) = (Z/2)3/2 (Zr/2) eZr/2
3
.. .. ..
. . . (6.4)
It was natural to try to use hydrogen-like orbitals as building blocks to
represent the wave functions of more complicated atoms. However, to the
great disappointment of the early workers in atomic theory, it was soon
realized that unless the continuum was included, the hydrogen-like orbitals
did not form a complete set; and the continuum proved to be prohibitively
difficult to use in practical calculations. This dilemma led Hloien, Shull
and Lowdin [2] to introduce basis functions that have exactly the same
form as hydrogen-like orbitals except that Z/n is replaced by a constant,
k, which is the same for all the members of the basis set. This type of basis
set came to be called Coulomb Sturmians, the name being given to them by
A. Rotenberg [3] to emphasize their connection with the SturmLiouville
theory of orthonormal sets of functions. Coulomb Sturmian basis sets are
complete without the inclusion of the continuum: any square-integrable
solution to a one-electron Schrodinger equation can be represented as a
linear superposition of them. If the potential in the one-electron Schrodinger
equation has some similarity to a Coulomb potential for example if it is
a screened Coulomb potential the convergence of such a series is rapid.
The members of a Coulomb Sturmian basis set are solutions to a one-
electron equation of the form

1 2 nk k2
+ n,l,m (x) = 0. (6.5)
2 r 2
If we compare Eq. (6.5) with (6.1) we can see that with the substitu-
tions Z/n k and En k2 /2, Eq. (6.1) is converted into Eq. (6.5).
The Generalized Sturmian Method 113
Table 6.1. One-electron Coulomb

Sturmian radial functions. If k is replaced
by Z/n they are identical to the familiar
hydrogen-like radial wave functions.
n l Rn,l (r)
1 0 2k 3/2 ekr
2 0 2k3/2 (1 kr)ekr
3/2
2 1 2k
kr ekr
3
2

3 0 2k3/2 1 2kr + 2(kr)3 ekr

3 1 2k3/2 2 3 2 kr 1 kr2 ekr

3 2 2k 3/2 2 (kr)2 ekr
3 5
Therefore, if we interpret k2 /2 as the energy, the solutions to the Coulomb

Sturmian wave equation (6.5) are just the familiar hydrogen-like orbitals
with Z/n replaced by k. In other words, they have the form
n,l,m (x) = Rn,l (r)Yl,m (, ) (6.6)
where Z/n in Eq. (6.4) is replaced by a constant k, which is the same for all
the members of the basis set. Since k 2 /2 is interpreted as the energy, all the
members of a Coulomb Sturmian basis set correspond to the same energy:
they are isoenergetic. The first few Coulomb Sturmian wave functions are
shown in Table 6.1.
You can easily see that if you make the substitution k Z/n
for the radial functions in this table, you will just get the familiar
hydrogen-like atomic orbitals; but Coulomb Sturmian basis sets have
very different properties! They obey a potential-weighted orthonormality
relation:

1 k
d 3 x (x) (x) = (n, l, m), (6.7)
r n
from which it follows that

3 2 + k 2
d x (x) (x) = . (6.8)
2k2
The Coulomb Sturmian basis sets behave quite differently from the usual
sets of eigenfunctions to the zeroth-order Hamiltonian often used in
quantum theory. Equation (6.1) is the usual type of eigenvalue problem with
which everyone in the physical sciences is familiar. By contrast, Eq. (6.5)
is an entirely different problem, sometimes called a conjugate eigenvalue
problem: Each member of a set of solutions corresponds to the same energy
k 2 /2, k being a constant that is the same for all the members of the set. The
quantity that plays the role of the usual eigenvalue is now a weighting factor
attached to the potential, which is chosen in such a way as to make all the
members of the basis set isoenergetic. Because of their useful properties,
Coulomb Sturmian basis sets are widely used in atomic theory, and there
exists a large literature discussing their properties and applications [213].
6.1.2. Generalized Sturmians

In 1968, Osvaldo Goscinski generalized the Sturmian concept by intro-
ducing basis sets that are solutions to an approximate many-particle
Schrodinger equation with a weighted potential:

N
1
2 + V0 (x1 , . . . , xN ) E (x1 , . . . , xN ) = 0, (6.9)
2mj j
j=1
the weighting factor once again being chosen in such a way as to

make all of the solutions correspond to the same energy. When N = 1,
V0 (x) = Z/r, and = nk/Z, this equation reduces to Eq. (6.5), obeyed
by Coulomb Sturmians. Basis sets of this kind have many advantages,
especially the advantage of spanning an appropriate Hilbert space, and
they could potentially be used in a wide variety of problems; but until now
the applications of generalized Sturmians have been very limited because
most physicists and theoretical chemists are unfamiliar with them. In fact
Osvaldo Goscinski did not publish his pioneering 1968 paper until very
recently [14]. It was only printed as an internal report of the Uppsala Uni-
versity Quantum Chemistry Group, and was known to just a small circle of
people. The idea remained dormant, and unfortunately little use was made
of generalized Sturmian basis sets in practical calculations. Today, however,
the generalized Sturmian method is an idea whose time has come!
In the generalized Sturmian method, the basis functions are chosen to
be isoenergetic solutions to an approximate Schrodinger equation with a
weighted potential (references [1518]).
A set of generalized Sturmian basis functions can be shown (see [14],
or [16], Chapter 1) to obey the following potential-weighted orthonormality
relations:

2E p2
d (x)V0 (x) (x) = , (6.10)

where we let denote a particular state and where we have introduced
the abbreviated notation x (x1 , x2 , . . . , xN ). To obtain the generalized
Sturmian secular equations, we begin by substituting the superposition

(x) = (x)B (6.11)

into the Schrodinger equation (6.22). This yields:

N
1 2
j + V(x) E (x)B = 0. (6.12)

2m j
j=1
We now split the potential V(x) into two parts, V(x) = V0 (x) + V (x), and
introduce the definitions

1
T
0
d (x)V0 (x) (x)
p
(6.13)
1
T d (x)V (x) (x).
p
From the potential-weighted orthonormality relations (6.10) it follows
that T 0 is diagonal:
p
T0 = T
0
= . (6.14)

Next, we notice that since all of the isoenergetic configurations in the basis
set obey (6.9), Eq. (6.12) can be rewritten as

[V(x) V0 (x)] (x)B = 0. (6.15)

We then multiply by a conjugate function from our basis set and integrate
over all space and spin coordinates:

d (x)[V(x) V0 (x)] (x)B = 0. (6.16)

Making use of Eqs. (6.13)(6.14), we obtain

p T0 p T + p T0 B = 0. (6.17)

Using (6.14) to derive
p T0 = p2 , (6.18)
and finally, dividing by p , we obtain the generalized Sturmian eigen-

problem:

0
T + T p B = 0. (6.19)

Generalized Sturmian basis sets can come in many species and vari-
eties: Every choice of the approximate potential V0 (x) (which should be
chosen to resemble V(x) as closely as possible) leads to a particular set of
shapes for the N-particle basis functions (x). Solving Eq. (6.9), which
is done once and for all for a particular V0 , specifies the functions up to
an undetermined scaling parameter p . Solving the generalized Sturmian
eigenproblem (6.19) then yields as eigenvalues the scaling parameters p
and as eigenfunctions

(x) = B (p , x), (6.20)

where each p scales the entire basis to give all the N-particle basis
functions the same energy E . If the generalized Sturmian basis { } is
complete, then Eq. (6.19) has exactly the same eigenfunctions as the
Schrodinger equation, and the energies are
p2
E = . (6.21)
2
In practice, one of course always uses a finite basis, so solutions are approx-
imate. However, we shall see that the automatic scaling allows us to obtain
good accuracy with few basis functions, as well as to obtain many excited
states at once.
It is remarkable to see how completely Eq. (6.19) differs from the con-
ventional secular equations used in quantum theory:
(1) The kinetic energy term has vanished.
(2) The matrix representing the approximate potential V0 (x) is diagonal.
(3) The roots are not energies but values of the scaling parameter, p ,
which is proportional to the square roots of the binding energies
(Eq. (6.10)).
(4) Before the secular equation is solved, only the shapes of the basis
functions are known, but not the values of the scaling parameters p .
(5) Solution of the secular equations yields a near-optimum basis set
appropriate for each state, as well as the states themselves and their
corresponding energies.
(6) The Hamiltonian formalism is nowhere to be seen!
In the present chapter, we review the generalized Sturmian method applied

to atoms and atomic ions, as well as the large-Z approximation that was
introduced by us in [17]. These methods have been described in much more
detail in our recent book [16]. Other work on Sturmians and generalized
Sturmians can be found in references [10][15].
The large-Z approximation yields extremely simple closed form expres-
sions for the approximate energies of both the ground states and excited
states of atoms and atomic ions. The accuracy of the large-Z approxi-
mation for few-electron systems is such that even for moderate values of
Z, inaccuracies are much smaller than relativistic corrections. An approx-
imate method for making relativistic corrections is introduced below in
Section 6.2.2. It is shown that the corrected energies rapidly approach the
experimental ones as Z increases.
6.1.3. The generalized Sturmian method applied to atoms

In atomic units, the non-relativistic Schrodinger equation for an N-electron
atom or atomic ion with nuclear charge Z is given by

1
+ V(x) E (x) = 0, (6.22)
2
where is the th electronic state with E the corresponding energy, and
where x stands for all the coordinates including spin.

N
Z
1
N N
V(x) = + . (6.23)
rj rij
j=1 j>i i=1
and
1
2
N
1
j , (6.24)
2 2
j=1
where j is the index of an individual electron in the system.

6.1.4. Goscinskian configurations

When the generalized Sturmian method is applied to atoms or atomic ions,
a wonderful thing happens: We wish to use a basis set consisting of solu-
tions to

1 p2
+ V0 (x) E (x) = 0 E = . (6.25)
2 2
It turns out that if we choose V0 (x) to be the Coulomb attraction potential of

the nucleus, then exact solutions to Eq. (6.25) can be found with the greatest
ease! Furthermore, the weighting factors are obtained automatically.
And as if this were not enough, there is a final bonus: the basis functions
(x) are automatically normalized! How can all this be possible? Read on
and see.
As just mentioned, we let V0 (x) be the electrostatic attraction potential
of the nucleus:

N
Z
N
1
V0 (x) = and V (x) = . (6.26)
rj rij
j=1 j>i i=1
Now we claim that with this choice of V0 (x), the weighting factors are
determined automatically, and Eq. (6.25) is satisfied by Slater determinants
of the form:

1 (1) 2 (1) N (1)

(2) (2) (2)
1
(x) = .
1 2 N
N! ..
.
.. .
..

(N) (N) (N)
1 2 N
= |1 2 N |, (6.27)
where the s are just the familiar hydrogen-like spin-orbitals,
nlm,+1/2 (xj ) = Rnl (rj )Ylm (j , j )(j)

(6.28)
nlm,1/2 (xj ) = Rnl (rj )Ylm (j , j )(j),
but with the weighted charges Q (Ref. [16], Chapter 3) chosen according
to the rules in the following box, where n1 , n2 , . . . , nN are the prin-
cipal quantum numbers of the hydrogen-like spin-orbitals in the configu-
ration . The Goscinskian configurations will be exact solutions to (6.25)
provided that:
p
Q = Z =
R

p 2E (6.29)

1 1 1
R 2
+ 2 + + 2 .
n1 n2 nN
At this point the reader may be muttering I dont believe it. Well, if
you dont believe it, think about this: The energy E will then be related to
the weighted nuclear charges Q by

p2 1 2 2 Q2 Q2 Q2
E = = Q R = + + + 2 . (6.30)
2 2 2n21 2n22 2nN
Each of the hydrogen-like spin-orbitals obeys a one-electron Schrodinger

equation of the form:

1 2 Q2 Q
j + 2 (xj ) = 0. (6.31)
2 2n rj
From Eq. (6.31) it follows that

N
1 j2 (x)
2
j=1

Q2 Q2 Q Q
= + + 2 + + + (x)
2n21 2nN r1 rN

= E V0 (x) (x). (6.32)
Now compare Eq. (6.32) with (6.25): they are the same! Thus Eq. (6.25)
will indeed be satisfied by the configurations shown in Eq. (6.27), pro-
vided that the effective nuclear charges Q are chosen according to the rule
given in Eq. (6.29). We shall call such a set of isoenergetic solutions to
(6.25) with V0 (x) chosen to be the nuclear attraction potential a set
of Goscinskian configurations to honor Professor Osvaldo Goscinskis
important early contributions to the generalized Sturmian method [14].
6.1.5. Goscinskian secular equations for atoms and atomic ions

Recall the solution of from Eq. (6.29)
p
= and p = 2E . (6.33)
Z R
This, and the potential-weighted orthonormality relations (6.10), give
T0 = ZR . (6.34)
Thus, the matrix T0 is diagonal and independent of p . It can be shown
(see [16], Appendix A) that T is also independent of p , although it is not
diagonal. Inserting Eq. (6.34) into the generalized Sturmian eigenproblem
we obtain the generalized Sturmian secular equations:

ZR + T p B = 0. (6.35)

We note that the only thing that requires any effort to calculate in
Eq. (6.35) is the interelectron repulsion matrix T the rest is trivial.
We shall call T the energy-independent interelectron repulsion

matrix. Its elements are pure numbers that depend only on the number
of electrons N. Having generated T , we can use it to calculate the
properties of a large number of states for an entire isoelectronic series.
6.2. Advantages: Some Illustrative Examples
We have just seen the remarkable ways in which the generalized Sturmian
secular equations differ from the usual secular equations that result from
diagonalizing the matrix representation of the Hamiltonian of a system: We
should especially notice that the eigenvalues are not energies, but values
of a parameter p , which is related to the energies by E = p2 /2. In the
case of Goscinskians, the configurations become pure functions (p x)
of p x, i.e. p acts as a scaling parameter of the space. Thus, in the solution
of the secular equations, an automatic scaling of the basis functions occurs:
For tightly-bound states, the atomic orbitals correspond to large values of
the effective charge, Q = p /R , and are contracted in space, whereas for
loosely-bound states the orbitals are spatially diffuse. It turns out, in fact,
that the Slater exponents that are automatically obtained by solution of
the generalized Sturmian secular equations are very nearly optimal. Thus,
when the generalized Sturmian method is applied to atoms and atomic
ions, not only is there no initial HartreeFock calculation, there is also no

preliminary worry about what Slater exponents would be appropriate. The
generalized Sturmian method using Goscinskians thus offers us a rapid
and convenient method for calculating the spectra and other properties of
few-electron atoms and atomic ions.
The accuracy of the method can be judged from Tables 6.2 and 6.3, com-
paring computed energies to experiment [19] and to exact nonrelativistic
values calculated by Nakatsuji and coworkers [20]. For a fixed number of
electrons, the calculated values approach the exact solution to the nonrela-
tivistic Schrodinger equation as the nuclear charge Z increases. However,
in the tables, we do not see the calculation approach the experiment for
the heavier elements. This is due to the fact that relativistic effects, which
deepen the binding energy, rapidly become important with increasing Z.
Had Nakatsuji and coworkers calculated exact nonrelativistic energies for
the heavier ions, our values would have approached his, both sets of values
being much less tightly bound than the experimental values.
Figures 6.1 and 6.2 illustrate this trend. However, we can notice in
Tables 6.2 and 6.3, that the exact nonrelativistic value for He is more tightly
bound than the experimental value. This is due to the small role of relativistic
effects and the large correction for the moving nucleus in the light helium
atom. Nakatsuji and coworkers also made a calculation taking into account
the motion of the nucleus, with correspondingly less tightly bound results.
For 1s2s 1 S, adding motion of the nucleus shifts the nonrelativistic energy
from 2.1460 to 2.1457, i.e. the relativistic effects deepen the binding
energy by 104 Hartrees to 2.1458. In Tables 6.2 and 6.3, both Nakatsuji
and coworkers values and our values neglect nuclear motion.
The Goscinskian basis is unsuitable for calculating ground states of the
helium-like isoelectronic series to high precision (see Section 6.3). These
states can be calculated much more accurately using a different generalized
Sturmian basis. Values calculated using 102 isoenergetic configurations
based on Coulomb Sturmians are given by us in Reference [16], Table F.1.
For the helium ground state we obtain 2.90250 Hartrees.
6.2.1. The large-Z approximation: restriction of the basis

set to an R-block
One of the great advantages of the generalized Sturmian method when it
is applied to atoms and atomic ions is that it leads naturally to an approx-
imation that allows us to write down the energies of atomic states with so
Table 6.2. 1 S excited state energies (in Hartrees) for the two-electron isoelec-
tronic series. The basis set used consisted of 40 generalized Sturmians of the
Goscinski type, and the whole table was computed in a few milliseconds. Exper-
imental values are taken from the NIST tables [19] (http://physics.nist.gov/asd),
and the exact nonrelativistic results of Nakatsuji and coworkers [20] are also given
for comparison.
He Li+ Be2+ B3+ C4+ N5+
1s2s 1 S 2.1429 5.0329 9.1730 14.564 21.206 29.098

Nakatsuji et al. 2.1460
expt. 2.1458 5.0410 9.1860 14.582 21.230 29.131
1s3s 1 S 2.0603 4.7297 8.5099 13.402 19.406 26.521
expt. 2.0611 4.7339 8.5183 13.415 19.425 26.548
1s4s 1 S 2.0332 4.6276 8.2837 13.003 18.785 25.629
expt. 2.0334 4.6299 8.2891 18.801 25.654
1s5s 1 S 2.0210 4.5811 8.1806 12.820 18.500 25.220
expt. 2.0210 4.5825 18.513 25.241
1s6s 1 S 2.0144 4.5562 8.1250 12.721 18.346 24.998
expt. 2.0144 4.5571
1s7s 1 S 2.0105 4.5412 8.0917 12.662 18.253 24.865
expt. 2.0104 4.5418
1s8s 1 S 2.0080 4.5315 8.0701 12.624 18.194 24.779
expt. 2.0079
1s9s 1 S 2.0063 4.5248 8.0554 12.598 18.153 24.720
expt. 2.0062
1s10s 1 S 2.0051 4.5201 8.0449 12.579 18.124 24.678
expt. 2.0050
1s11s 1 S 2.0042 4.5166 8.0371 12.566 18.102 24.647
expt. 2.0041
1s12s 1 S 2.0034 4.5140 8.0312 12.555 18.086 24.624
expt. 2.0034
Table 6.3. 3 S excited state energies calculated with 36 Goscinskians. The cal-
culation of similar tables for 1 P, 3 P, 1 D, 3 D, doubly excited autoionizing states,
etc., is equally easy, rapid, and of comparable accuracy. Tables are given in
Chapters 3 and 4 in [16], but may easily be reproduced using our programs, as
shown in Tutorial 1 on [1].
He Li+ Be2+ B3+ C4+ N5+
1s2s 3 S 2.1736 5.1079 9.2937 14.730 21.417 29.353

expt. 2.1750 5.1109 9.2983 14.738 21.429 29.375
1s3s 3 S 2.0682 4.7504 8.5442 13.450 19.466 26.594
expt. 2.0685 4.7522 8.5480 13.457 19.478 26.614
1s4s 3 S 2.0363 4.6360 8.2983 13.023 18.811 25.661
expt. 2.0363 4.6373 8.3015 13.030 18.822 25.680
1s5s 3 S 2.0225 4.5854 8.1880 12.831 18.514 25.237
expt. 2.0224 4.5862 8.1905 18.524 25.254
1s6s 3 S 2.0153 4.5586 8.1293 12.728 18.354 25.008
expt. 2.0152 4.5592 18.364
1s7s 3 S 2.0111 4.5427 8.0944 12.667 18.259 24.872
expt. 2.0109 4.5431 18.268
1s8s 3 S 2.0084 4.5325 8.0719 12.627 18.197 24.784
expt. 2.0082 4.5328 18.206
1s9s 3 S 2.0066 4.5255 8.0567 12.600 18.156 24.724
expt. 2.0064
1s10s 3 S 2.0053 4.5206 8.0458 12.581 18.126 24.681
expt. 2.0051
1s11s 3 S 2.0044 4.5170 8.0378 12.567 18.103 24.649
expt.
1s12s 3 S 2.0035 4.5142 8.0317 12.556 18.087 24.625
expt.

Z2
0.58
0.60
0.62
0.64
Z
10 20 30 40 50
Fig. 6.1. This figure shows energies for the lowest 3 S state of the helium-like
isoelectronic series, divided by Z2 to make the details easier to see for large Z. The
values are calculated in the large-Z approximation, which here limits the basis to
a single configuration. The lower (solid) line is corrected for relativistic effects as
discussed in the text; the dots indicate experimental values from the NIST tables. It
is easy to visually verify that, for Z > 10, the relativistic correction is much larger
than calculational errors due to the large-Z approximation.
little effort that the calculation can literally be carried out on the back of an
envelope! We call this approximation the Large-Z Approximation.
If interelectron repulsion is entirely neglected, i.e. when disregarding
the second term in Eq. (6.19), the calculated energies E of become
those of a set of N completely independent electrons moving in the field of
the bare nucleus:
p2 1 Z2 Z2 Z2
E = Z2 R 2 = 2 2 2 . (6.36)
2 2 2n1 2n2 2nN
Equation (6.36) is not the large-Z approximation:
In the large-Z approximation, we do not neglect interelectron repulsion,

but we restrict the basis set to those Goscinskian configurations that would
be degenerate if interelectron repulsion were entirely neglected, i.e. the
basis is a set of configurations all of which correspond to the same value
of R .

2
Z
1.0
1.1
1.2
1.3
1.4
Z
10 20 30 40 50
Fig. 6.2. The ground state of the carbon-like isoelectronic series. As Z grows,
the approximation approaches the exact solution to the nonrelativistic Schrodinger
equation. Due to the increased role of interelectron repulsion in the carbon-like
series, this takes longer than for the helium-like series. However, at around Z = 18,
the inaccuracy of the large-Z approximation becomes smaller than the relativistic
correction.
In that case, the term ZR in (6.19) is a multiple of the identity

matrix, and the eigenvectors B are the same as those that would be
obtained by diagonalizing the energy-independent interelectron repulsion
matrix T , since eigenfunctions are unchanged by adding a multiple of the
unit matrix.

T B = 0. (6.37)

The roots are shifted by an amount equal to the constant by which the
identity matrix is multiplied:
p = ZR + = ZR | | (6.38)
and the energies become
1 1
E = p2 = (ZR | |)2 . (6.39)
2 2
Table 6.4. Roots of the ground state R-block of the interelectron

repulsion matrix for the He-like, Li-like, Be-like, B-like and C-like
isoelectronic series.
Li-like Be-like B-like C-like

| | term | | term | | term | | term
0.681870 2S 0.986172 1S 1.40355 2P 1.88151 3P
0.729017 2P 1.02720 3P 1.44095 4P 1.89369 1D
1.06426 1P 1.47134 2D 1.90681 1S
1.09169 3P 1.49042 2S 1.91623 5S
1.10503 1D 1.49395 2P 1.995141 3D
1.13246 1S 1.52129 4S 1.96359 3P
1.54037 2D 1.98389 3 yS
1.55726 2P 1.98524 1D
1.99742 1P
2.04342 3P
He-like 2.05560 1D
0.441942 1S 2.07900 1S
Since the roots are always negative, we may use the form | | in
place of to make explicit the fact that interelectron repulsion reduces
the binding energies, as of course it must. The roots are pure numbers
that can be calculated once and for all and stored. Values of these roots for
N = 2, 3, . . . , 10 are shown in Tables 6.4 and 6.5, together with their cor-
responding spectroscopic terms. From the roots, a great deal of information
about atomic states can be found with almost no effort: Given the values
of the principal quantum numbers n1 , n2 , . . . , nN , and given the value of
| |, which can be looked up in a table, the calculation of the energies for
the entire isoelectronic series is completely effortless!
The eigenfunctions corresponding to the spectroscopic terms in
Tables 6.4 and 6.5 are symmetry adapted RusselSaunders states and can
be used as basis functions for more exact calculations. The classification is
done automatically by the method discussed in [16], Sections 3.4 and 3.5.
Tutorial 2 on our website sturmian.kvante.org [1] shows in detail
how to do this.
6.2.2. Validity of the large-Z approximation

In Fig. 6.1, the large-Z approximation energy E = 12 (ZR | |)2
for the lowest triplet states of the helium-like isoelectronic series is plotted
Table 6.5. Roots of the ground state R-block of the interelectron

repulsion matrix T for the N-like, O-like, F-like and Ne-like isoelec-
tronic series.
N-like O-like F-like Ne-like

| | term | | term | | term | | term
2.41491 4S 3.02641 3P 3.68415 2P 4.38541 1S
2.43246 2D 3.03769 1D 3.78926 2S
2.44111 2P 3.05065 1S
2.49314 4P 3.11850 3P
2.52109 2D 3.14982 1P
2.53864 2S 3.24065 1S
2.54189 2P
2.61775 2P
against spectroscopically determined energies. In order to better see the

details, we plot E /Z2 . Figure 6.2 shows E /Z2 for the ground state of the
six-electron isoelectronic series.
As the nuclear charge Z increases, the energies and wave functions
calculated with the large-Z approximation approach the exact solutions
to the non-relativistic Schrodinger equation. However, relativistic effects
begin to be pronounced at around Z = 10, and become progressively
more so as Z increases. Therefore the calculated values first approach the
experimental ones, but begin to differ as relativity becomes more and more
important.
It is possible to make a rough correction for the relativistic effect on the
energies by multiplying them by an easily-calculated factor f (Z), so that
E becomes
1
f (Z)(ZR + )2 . (6.40)
2
The correction factor f (Z) is the ratio between the relativistic and non-
relativistic energies of a configuration, assuming interelectron repulsion to
be completely neglected such that the energy is equal to that of N inde-
pendent electrons moving in the field of the nucleus.
In the nonrelativistic case, the energy of a hydrogen-like spin-orbital is
Z2
given by 2n 2 , and thus the total energy of an N-electron configuration is
2 Z R .
1 2 2
In the relativistic case, the exact solution to the Dirac equation for
hydrogen-like atoms can be found in [21], or in [16], Eqs. (7.35) through
(7.40). The ratio of the relativistic energy Erel and the nonrelativistic energy
Enonrel for a multiconfigurational state

= B (6.41)

is
2
Erel B |H0 | rel
f (Z) = = 2
Enonrel B |H0 | nonrel
2
B |H0 | rel
= 1 . (6.42)
2 Z2 B
2 R 2

Here, H0 is a sum of one-electron Hamiltonian operators corresponding to

single electrons moving in the field of the bare nucleus, i.e. interelectron
repulsion is completely neglected.
In the figures, the lines are calculated in the large-Z approximation. The
upper (dashed) line is not corrected for relativistic effects, while the lower
(solid) line is corrected. The dots are experimental values of the energies
taken from the NIST Atomic Spectra Database [19]. It can be seen from
Figs. 6.1 and 6.2 that agreement between the energies calculated from the
large-Z approximation and experimental energies become progressively
better as Z increases, provided that the rough relativistic correction is made.
We note that the large-Z approximation, despite its great simplicity, well
approximates non-relativistic energies: Even for modest values of nuclear
charge, the error of the large-Z approximation is much smaller than the error
due to neglecting relativity. Further, relativistic effects may be accounted for
by means of an easily calculated factor, yielding energies that correspond
well with experiment.
The second example presented here (the ground state of the carbon-like
isoelectronic series) is a case not easily approximated using a small number
of Goscinskian basis functions, because interelectron repulsion plays a large
role. Nevertheless, it can be seen that even in this somewhat difficult case,
the large-Z approximation gives surprisingly reasonable results. The large-
Z approximation is not only extremely simple, but it is characterized by
a small number of parameters the roots of the interelectron repulsion
matrix. These roots are dimensionless and independent of energy and of
nuclear charge. They can be calculated once and for all, and they contain
information concerning many states of the entire isoelectronic series. Once
the roots are obtained, calculating approximate atomic energies, and a
number of other properties, become tasks that can be carried out by pen
and paper.
Relative error
Gosc.rel.
LargeZrel.
Gosc. nonrel.
0.015
LargeZnonrel.
0.010
0.005
Z
5 10 15 20 25 30
E E
Fig. 6.3. Ground state relative errors calcEexp exp compared to experiment
for the helium-like isoelectronic series. The large-Z approximation energies
12 (Z 2 .441942)2 are compared to results using a fuller Goscinskian basis.
The two dotted lines are the nonrelativistic values, while the solid lines are cor-
rected for relativistic effects using Eq. (6.42). For very large values of Z, errors due
to quantum electrodynamic effects cause a systematic overestimation of binding
energy.
6.2.3. Core ionization energies

The large-Z approximation can be used to estimate a number of additional
properties. For example, using the approximation, we can calculate by hand
the core-ionization energies, i.e. the energies required to remove an electron
from the inner shell of an atom. From (6.39) we can see that this energy
will be given by
1
E = (ZR | |)2 (ZR | |)2 , (6.43)
2
where the unprimed quantities refer to the original ground state, while the
primed quantities refer to the core-ionized states. Since
R 2 R 2 = 1, (6.44)
Eq. (6.43) can be written in the form
Z2 | |2 | |2
E = Z R | | R | | + . (6.45)
2 2
Thus we can see that within the framework of the large-Z approximation,
the quantity E Z2 /2 is linear in Z for an isoelectronic series. This
Z2
E
2 10 20 30 40
0 Z
20
N=2
40
60
80
N = 10
100
120
N = 18
Fig. 6.4. For isoelectronic series, Eq. (6.45) indicates that within the large-Z
approximation, the quantity E Z2 /2 is exactly linear in Z, as is illustrated
above.
quantity represents the contribution of interelectron repulsion to the core

ionization energy, since if interelectron repulsion is completely neglected,
the core ionization energy is given by E = Z2 /2. Core ionization energies
calculated from Eqs. (6.43)(6.45) are shown in Fig. 6.4. Between N = 10
and N = 18 the lines in the figure become more closely spaced. This is
because a new shell starts to fill at N = 11. A table showing qualitative
agreement between experiment and the core-ionization energies calculated
in the large-Z approximation is given in Chapter 5 of [16]. Detailed calcu-
lations can be found on our website [1] in Tutorials 3 and 5.
6.3. Limitations of the Method; Prospects for the Future
We mentioned above that the generalized Sturmian method using

Goscinskian configurations offers a very rapid and convenient method for
calculating the spectra, wave functions and other properties of few-electron
atoms and atomic ions. But why is the method limited to systems with a
small number of electrons? The reason for this is that Goscinskian configu-
rations are solutions to Eq. (6.25) with V0 chosen as the Coulomb attraction
of the bare nucleus, as shown in Eq. (6.26). As the number of electrons
N becomes large, this zeroth-order potential becomes progressively more
unrealistic, because the effects of interelectron repulsion become progres-
sively more important.
How can we correct this defect? One way to extend the range of the
method is to use a V0 in Eq. (6.25) that in some form includes interelectron
repulsion effects. This will make it less straightforward to obtain the gen-
eralized Sturmian configurations , depending on the complexity of the
chosen V0 , in general requiring a self-consistent field iteration. However,
the useful properties of the generalized Sturmian basis are retained, and the
extra initial work would lead to improved convergence.
Another possibility is to extend the method by using a basis set consisting
of isoenergetic configurations
(x) = |1 2 N |, (6.46)
constructed from orbitals satisfying

1 2 k2
j + + v(rj ) (xj ) = 0, (6.47)
2 2
where v(rj ) is the nuclear attraction potential, corrected by a repulsive
potential due to the core electrons:
Z
v(rj ) = + vc (rj ). (6.48)
rj
This introduces interelectron repulsion effects even earlier in the calcula-
tions. The potential vc (rj ) can be found by performing a fast preliminary
calculation using Goscinskian configurations. From this preliminary cal-
culation, a spherically-averaged core density, (rj ) can be obtained, and
from this vc (rj ) may be calculated by means of the relationship

1
vc (rj ) = drj rj2 (rj ) , r> Max[rj , rj ]. (6.49)
0 r>
The orbitals (xj ) can be built up from Coulomb Sturmians, so that (6.47)
becomes:

1 k2

j +2
+ v(rj ) (xj )C, = 0. (6.50)

2 2
Multiplying from the left by a conjugate Coulomb Sturmian, we obtain:

1 2 k2
0= d xj (xj ) j +
3
+ v(rj ) (xj )C,

2 2

= k + d xj (xj )v(rj ) (xj ) C,
2 3

= k 2 k t C, (6.51)

or

k

t C, = 0, (6.52)

where

1
t d 3 xj (xj )v(rj ) (xj ). (6.53)
k
After solving Eq. (6.52) to obtain the coefficients C, , we can next use
the isoenergetic configurations (x) = |1 2 N | as basis functions
for solving the Schrodinger equation for an atom or atomic ion. This can
be written in the form

N
1 2 k2
j + + V(x) (x) = 0, (6.54)
2 2
j=1
with

N
Z
1
N N
V(x) = + , (6.55)
rj rij
j=1 i>j j=1
and with

N
k2 Nk2
E = = . (6.56)
2 2
j=1
Thus we write

(x) = (x)B, . (6.57)

Substituting this into the N-electron Schrodinger equation, and taking
the scalar product with a conjugate configuration, we obtain the secular
equations:

N 2
1 k
dx (x) j2 + + V(x) (x)B, = 0.

2 2
j=1
(6.58)
We next introduce the k-independent matrix

N
1 1 2 k2
S 2 dx (x) j + (x), (6.59)
k 2 2
j=1
which can be interpreted as a generalized ShibuyaWulfman matrix, and

another k-independent matrix:

1
T dx (x)V(x) (x). (6.60)
k
In terms of these matrices, the secular equations become:

[T k S ] B, = 0. (6.61)

This gives us a spectrum of k-values from which the energies of the various
states, E = Nk 2 /2, can be obtained. It seems quite likely that this
procedure would allow the generalized Sturmian method for atoms and
atomic ions to be extended to larger values of N. Some steps in this direction
have already been taken by us and by Professor Gustavo Gasaneo and his
students at Universidad Nacional del Sur in Argentina.
What developments are necessary in order to apply the generalized
Sturmian method to complex chemical problems? Once we have found
a generalized Sturmian basis that converges well, most of the standard
techniques in quantum chemistry can be employed in the same way that
they are currently used with bases obtained from initial HartreeFock cal-
culations. Two obvious steps are to use the frozen core approximation to
factor out correlation of core electrons, and use standard perturbation theory
based techniques to reduce the computational efforts necessary for con-
figuration interaction. Using the generalized Sturmian method with, for
example, coupled cluster methods requires some work, but may be well
worth the effort due to improved convergence properties compared to using
HartreeFock based configurations.
6.3.1. Can the generalized Sturmian method be applied

to N-electron molecules?
If we wish to use a many-center Coulomb Sturmian basis set to treat
molecules, we can introduce the notation
(x) (x Xa ), (6.62)
where
(a, l, m, n). (6.63)
In a molecule, the one-electron nuclear attraction potential has the form:

Za
v(xj ) = . (6.64)
a
|xj Xa |
We can build up solutions to the one-electron equation

1 2 k2
j + + v(xj ) (xj ) = 0 (6.65)
2 2
from superpositions of many-center Coulomb Sturmians:

(xj ) = (xj )C, . (6.66)

Thus, we have

1 k2

j +
2
+ v(xj ) (xj )C, = 0. (6.67)

2 2
Taking the scalar product with a conjugate Coulomb Sturmian yields:

1 k2
d 3 xj (xj ) j2 + + v(xj ) (xj )C, = 0. (6.68)

2 2
If we let

1 1 k2
S d 3 xj (xj ) j2 + (xj ), (6.69)
k2 2 2
and

1
W d 3 xj (xj )v(xj ) (xj ), (6.70)
k
then the one-electron secular equations take the form:

k

W S C = 0. (6.71)

The integrals S are the well-studied ShibuyaWulfman integrals, which

can be generated using a variety of algorithms.
Historical note: The ShibuyaWulfman integrals were first introduced

by T. Shibuya and C. Wulfman in 1965 in connection with their famous
momentum-space treatment of many-center one-electron problems [22].
These integrals can conveniently be evaluated in momentum space using
the Fock projection, which establishes a relationship between hyper-
spherical harmonics and the Fourier transforms of Coulomb Sturmians.
The problem of evaluating these integrals, as well as many other integrals

needed in molecular problems, can then be converted into a problem of
hyperangular integration [8, 16, 23, 24].
It can be shown that W and S are related through the sum rule
[9, 10]:

Za
W = S S . (6.72)

n
In Eq. (6.72), = (a, ) runs over all the atoms a in the molecule and the full
Coulomb Sturmian basis set centered at Xa . If the basis set is truncated, the
relationship is only approximate. Solutions to the two-center one-electron
Schrodinger equation using the method just described have been studied
by Aquilanti and his group in Perugia [5, 8, 15] and by Koga and his group
in Japan [9, 10].
We now introduce the N-electron configurations of the form
(x) = |1 2 N |, (6.73)
where the molecular orbitals j are solutions to (6.71). We would like to
use these configurations to build up solutions to the N-electron Schrodinger
equation

N 2
1 k Nk 2
j2 + + V(x) (x) = 0 with E = ,
2 2 2
j=1
(6.74)
where

Za
1N N
V(x) = + . (6.75)
|xj Xa | rij
j=1 a i>j j=1
Then, in a manner exactly similar to equations (6.54)(6.58), we are led to

secular equations of the form

[T k S ] B, = 0, (6.76)

the only difference being that in the calculation of T , the molecular

potential is used in place of the atomic potential.
In the case of diatomic molecules, we begin by picking a value of the
parameter s = kR, where R is the interatomic distance and k is the exponent
E/Hartree
0 R/Bohr
1 2 3 4 5 6
1
Etot
2
He+H+
3 Ee
7
Li +
Fig. 6.5. This figure shows the electronic energy Ee and the total energy Etot of
the HeH+ ion as a function of the internuclear separation R = s/k. The calculation
was performed with a single configuration using a one-electron basis set consisting
of three Coulomb Sturmians on each center. For R 0, the electronic energy
approaches the energy calculated for the Li+ ion using the generalized Sturmian
method with a single configuration [16]. In the separated atom region, the total
energy approaches that of He when calculated in the same way. Our calculation
exhibits a shallow minimum at R = 1.35 Bohrs, which can be compared to the
equilibrium bond length of 1.3782 Bohrs resulting from a HF/STO-3G calculation
quoted by Szabo and Ostlund [25], and with the value of 1.46 Bohrs obtained
in a benchmark calculation by Wolniewicz [26]. Since our pilot calculation uses
only one configuration, it makes sense that we obtain a result comparable to the
HartreeFock calculation.
of the Coulomb Sturmian basis set. Neither R nor k is known at this point,
but only their product s. For the diatomic case, all of the integrals involved
in Eqs. (6.71) and (6.61) are pure functions of s. Having chosen s, we can
thus solve the one-electron secular equations and obtain the coefficients
C and the spectrum of ratios k/ . We are then able to solve Eq. (6.76),
which gives us the eigenvectors B as well as a spectrum of k-values, and
thus energies Nk 2 /2. From a k-value, we also get the unscaled distance
R = s/k. We repeat the procedure for a range of s-values and interpolate to
find the solutions as functions of R. Figure 6.5 shows our pilot calculation
on the HeH+ two-electron molecular ion using the method described above.
This is an extremely simple calculation, using only one configuration, but
we are actively working to explore the method further. We chose HeH+ for
the pilot calculation rather than H2 because, as is well known, the correct
dissociation curve for H2 needs at least two configurations.
In the case of polyatomic molecules, one can choose a set of angles

between the nuclei; these are left fixed under scaling of the coordinate
system. The procedure is then similar to that described for the diatomic case.
6.4. Discussion
In this chapter, we have concentrated on the use of isoenergetic configura-

tions of the Goscinskian type. However, the generalized Sturmian method
can be applied to wide variety of problems, including cases where the
masses are unequal and the force field very different from Coulomb inter-
actions. All that is needed is a set of isoenergetic solutions to Eq. (6.9),
where V0 determines the shapes of the resulting generalized Sturmian basis
functions.
We have performed calculations based on isoenergetic configurations
formed from Coulomb Sturmians, as is discussed in our recent book
(see [16], Appendix F). Basis sets of this type have the advantage that
we have rigorous proof of their completeness. Unlike most other methods,
this thus gives assurance that we can approximate the exact solution to
the Schrodinger equation with arbitrary precision, given arbitrary compu-
tation time. In Section 6.3, we outline how to build a generalized Sturmian
basis for molecules. Professor Gustavo Gasaneo and his coworkers are
exploring Sturmian bases that include interelectron repulsion effects as
well as Sturmians that include the continuum.
We have seen that the generalized Sturmian method using Goscinskian
configurations as a basis offers an extremely rapid and convenient way of
calculating atomic spectra and other properties of few-electron atoms and
ions. We obtain good solutions to many states at once, using only a very
small basis and without any need for SCF calculations. The wave function
is in a form that is easy to interpret by inspection or to analyze automatically
by computer (see Tutorial 2 of [1]). Furthermore, T , the matrix which
represents interelectron repulsion in a basis of Goscinskian configurations,
consists of pure numbers of universal applicability. The matrix depends
only on the number of electrons in an atom or ion, and it is independent of
energy and nuclear charge. The energy-independent interelectron repulsion
matrix, T , can be calculated once and for all, and used in a wide variety
of cases.
The results shown in Section 6.2 illustrate the degree of speed and
accuracy that can be expected for few-electron atoms and ions. Neutral
atoms are the worst case when using Goscinskians. However, in the neutral
helium atom, where Nakatsujis results [20] are available, our results agree
well with his as can be seen in Tables 6.2 and 6.3. Had Nakatsuji and
coworkers made calculations on the whole isoelectronic series, agreement
would be progressively better for the heavier ions in the series.
We find that in order to obtain good agreement with experiment, it is
necessary to include relativistic effects. For the few-electron systems treated
here, the crude relativistic correction of Eq. (6.42) gives very good results.
For the two-electron isoelectronic series, the ground state was obtained
with relative error compared to experiment of 3.5 103 for Z = 2 (the
worst case) to roughly 106 for Z 8, and excited states were obtained
with relative errors between 104 and 106 . The complete calculation of
all the states (found in [18], Chapter 4) required only 77ms of computation.
It should be noted, that for very large values of Z, quantum electrodynamic
effects become important, and neglecting them will cause an overestimation
of the binding energies. If more precision is required, we can treat the
system by means of the DiracCoulomb equation. Calculations using a
fully relativistic analogue to the Goscinskian configurations can be found
in Chapter 7 of our book [16].
As the number of electrons grow, there is a decrease in accuracy when
using similar computation time. Already for five-electron systems, the cal-
culations in Chapter 4 of [18] yield ground states for the Z = N case with
less accuracy than the HartreeFock limit. This suggests to us that in order
to solve systems with many electrons accurately, while retaining efficiency,
we need Sturmian basis functions that incorporate interelectron repulsion
in V0 (x), as is discussed in Section 6.3.
The generalized Sturmian method using Goscinskians leads to an
extremely simple and convenient approximation, the large-Z approxi-
mation, which is described in Section 6.2. The approximation leads to a
remarkably simple closed-form expression, E = 21 (ZR | |)2 , for the
energies of states in terms of the appropriate roots of the energy-independent
interelectron repulsion matrix. As the name suggests, the large-Z approxi-
mation is not very accurate when Z = N, especially in the case of ground
states. It underestimates the binding energy of the ground state of neutral
helium by 2% and of neutral argon by 5% (Fig. 4.5 in [18]), but it improves
rapidly with increasing Z N (Fig. 4.3 in [18]). For excited states of few-
electron atoms, the large-Z approximation gives surprisingly good results
even for modest values of Z N. Given the interelectron repulsion roots
, which are dimensionless quantities that depend only on the number of
electrons and can be precalculated, we can calculate electronic states for
entire isoelectronic series with a pencil and a scrap of paper.
It is our hope that in the future the method may be extended to give
accurate calculations for atoms where interelectron repulsion effects are
comparable in importance to nuclear attraction. We are also in the process
of extending the method to molecules.
Bibliography
[1] J. Avery and J. Avery, The Generalized Sturmian Library. Available at

http://sturmian.kvante.org, (2006).
[2] H. Shull and P.-O. Lowdin, J. Chem. Phys. 30, 617 (1959).
[3] M. Rotenberg, Adv. At. Mol. Phys. 6, 233 (1970).
[4] A. Sherstyuk, Teor. Mat. Fiz. 56(2), 272 (1983).
[5] V. Aquilanti, S. Cavalli, C. Coletti, and G. Grossi, Chem. Phys. 209, 405 (1996).
[6] J. Avery and D.R. Herschbach, Int. J. Quantum Chem. 41, 673 (1992).
[7] J. Avery, Int. J. Quant. Chem. 100(2), 121 (2004). doi: 10.1002/qua.10820.
[8] A. Caligiana, PhD thesis, University of Perugia, Italy (2003).
[9] T. Koga and T. Matsuhashi, J. Chem. Phys. 87(8), 4696 (1987).
[10] T. Koga and T. Matsuhashi, J. Chem. Phys. 89, 983 (1988).
[11] R. Shakeshaft, J. Phys. B: At. Mol. Phys. 18(17), (1985).
[12] R. Shakeshaft, Phys. Rev. A: Gen. Phys. 34(6), 5119 (1986).
[13] R. Shakeshaft, Phys. Rev. A. 14(5), 1626 (1976).
[14] O. Goscinski, Advan. Quant. Chem. 41, 51 (2002).
[15] V. Aquilanti, S. Cavalli, D. De Fazio, and G. Grossi, in New Methods in Quantum
Theory edited by C. Tsipis, V. Popov, D. Herschbach and J. Avery (Kluwer,
Dordrecht, 1996) pp. 233250.
[16] J. Avery and J. Avery, Generalized Sturmians and Atomic Spectra (World Scientific,
New York, 2007).
[17] J. Avery and J. Avery, Adv. Quant. Chem. 49, 103 (2005).
[18] J. Avery, Masters thesis, Department of Computer Science, University of Copen-
hagen (DIKU), (2008). Available at http://sturmian.kvante.org/thesis/speciale.pdf
and http://sturmian.kvante.org/thesis/sturmian-1.0.pdf.
[19] Y. Ralchenko,A. Kramida, J. Reader, and N.A. Team. NISTAtomic Spectra Database.
http://physics.nist.gov/asd, (2008).
[20] H. Nakashima, Y. Hijikata, and H. Nakatsuji, J. Chem. Phys. 128, 154108 (2008).
[21] A. Akhiezer and V. Berestetskii, Quantum Electrodynamics (Interscience, New York,
1965).
[22] T. Shibuya and C. Wulfman, Proc. Roy. Soc. A. 286, 376 (1965).
[23] J. Avery, Hyperspherical Harmonics; Applications in Quantum Theory (Kluwer,
Dordrecht, 1989).
[24] J. Avery, J. Math. Chem. 24, 169 (1998).
[25] A. Szabo and N. Ostlund, Modern Quantum Chemistry (Dover, New York, 1996).
[26] L. Wolniewicz, J. Chem. Phys. 43, 1087 (1965).

Chapter 7
Slater-Type Orbital Basis Sets: Reliable

and Rapid Solution of the Schrdinger
Equation for Accurate Molecular
Properties
Philip E. Hoggan
LASMEA, UMR 6602 CNRS,

Universite Blaise Pascal, 24 avenue des Landais,
BP 80026, 63171 Aubiere CEDEX, France
It is easy to prove that atomic and molecular orbitals must decay expo-
nentially at long-range. They should also possess cusps when an electron
approaches another particle (a peak where the ratio of orbital gradient to
function gives the particle charge).
Therefore, hydrogen-like or Slater-type orbitals are the natural basis
functions in quantum molecular calculations. Over the past four decades,
the difficult integrals led computational chemists to seek alternatives.
Consequently, Slater-type orbitals were replaced by Gaussian expansions
in molecular calculations (although they decay more rapidly and have
no cusps). From the 1990s on, considerable effort on the Slater integral
problem by several groups has led to efficient algorithms which have
served as the tools of new computer programs for polyatomic molecules.
The key ideas for integration: one-center expansion, Gauss transform,
Fourier transform, use of Sturmians and elliptical coordinate methods are
presented here, together with their advantages and disadvantages, and the
latest developments within the field.
Recent advances using symbolic algebra, pre-calculated and stored
factors and the state-of-the art with regard to parallel calculations are
reported.
141
142 P.E. Hoggan
At times, high accuracy is not required and at others speed is

unimportant.
A recent approximation separating the variables of the Coulomb
operator will be described, as well as its usefulness in molecular
calculations.
There is a renewed interest in the use of Slater orbitals as basis
functions for configuration interaction (CI) and Hylleraas-CI atomic and
molecular calculations, and in density functional and density matrix
theories. In a few special cases, e.g. three particles, symmetry conditions
lead to simple explicit pair-correlated wave-functions.
Similarly, advantages of this basis are considerable for both absolute
energy and fixed-node error in quantum Monte Carlo (QMC). The model
correlated functions may be useful to build up Jastrow factors.
These considerations will be dealt with in the context of modern com-
puter hardware and its rapid development.
In memory of Daniel Rinaldi (11 May 194227 May 2010)

He introduced me to work on molecular integrals over Slater orbitals,
as well as doing so much more for me and for his other students.
7.1. Introduction
7.1.1. Context of this chapter

The aim of this book is to explain the essence and the strengths and weak-
nesses of some less routine methods that are used to solve the Schrodinger
equation for molecules. It expects the authors to make very clear to readers
what the essential ideas are behind current alternative methods and thus
catalyse creativity in the pursuit of a new (possible hybrid) method that
combines the best of all worlds (methods). The present chapter will now
begin explaining use of Slater basis sets. It is limited to the stationary case
(in the non-relativistic limit, although exponentially decaying functions are
also used for the DiracCoulomb equation).
Here, direct use of analytical Slater-type orbital (STO) basis sets is
presented. They are a natural choice, since the relationship to physical
atomic orbitals is clear. Nevertheless, they are frequently expanded in terms
of Gaussians.
It is important to explain what problems arise in molecular calculations,
although quite a few groups worldwide go to the trouble of doing this math-
ematically taxing work when Gaussian functions are so easy to manipulate,
because of their simple products. For well-chosen applications, this hard
work is amply rewarded by accurate results.
Slater-Type Orbital Basis Sets: Reliable Molecular Properties 143
Assuming the molecular backbone is readily determined and that the

nuclei have known, fixed positions, the interesting question is how to
determine the electron distribution. The constituent atoms also have known
atomic number and the user decides which atomic orbitals to consider,
subject to including all those wholly or partially occupied by electrons in
the separate atoms (so-called ab initio work).
This chapter fixes molecular geometry and uses the variation principle
on total energy in the Schrodinger equation, minimizing it with respect
to a linear combination coefficients over Slater functions (LCAO) itera-
tively. That is, HartreeFock (HF) calculations are performed on a basis
of exponentially decaying atomic orbitals, such as Slater-type orbitals. As
in any HF calculation, the atoms and their orbitals are a good starting
point for molecular electron distributions and are used in linear combi-
nations to describe molecular orbitals. As Lucretius1 would have put it,
the atoms came together willy-nilly, (except that today we would add
that they assemble and stay together according to criteria of energetic sta-
bility) it is true, nevertheless, that apart from a few of the most peripheral
electronic orbitals, most retain the analytic atomic form, namely hydrogen-
like (or Slater-type orbitals). Therefore, all that is required from atoms
in molecules is the idea that core density is little affected going from
free atoms to molecules, so this is treated as transferable in a suitable
AO basis and (possibly) relaxed in the last couple of SCF cycles. We will
return to atoms in molecules and describe diatomics in molecules in more
detail.
This LCAO-molecular orbital approach treats the electrons indepen-
dently: they are uncorrelated apart from an exchange phenomenon recalled
below. The total energy is built up of matrix elements (integrals) in the AO
basis. These integrals are (real) numerical values, that can readily be handled
by computer programs. The bottleneck stems from two-electron integrals.
Gaussian functions have a simple product theorem that reduces these terms
to single-center expressions, whereas the product of Slater orbitals on dis-
tinct atoms generally leads to an infinite sum about one center (translation
or addition theorem).
It must be emphasized that the accuracy required varies with SCF con-
vergence. In fact, a good approximation without orbital translation is satis-
factory for all iterations except the last couple. Furthermore, only the sum
is required so many terms can be batched.
1 De Rerum Natura: work based on that of Democritus and the Greek atomists of the fourth century
BC, who were referring to molecules by the term atom and interpreting odour etc.
144 P.E. Hoggan
Basically, the problem is reduced to N 2 complexity until the last itera-

tions which are N 4 subject to some neglect vide infra (i.e. the calculation
time depends on the square of the number N of AOs for most of its duration).
The same may be applied to KohnSham density functional theory
(DFT), with final complexity limited to N 3 .
Exchange correlation is illustrated by Hunds rule in footnote.2
First recall what most students in the field learn about atomic orbitals
and Slater exponents, placed in a more general context.
7.1.2. Atomic orbitals

The stationary Schrodinger equation for one-electron atoms has well-
known analytical solutions. These eigenfunctions of the corresponding
Hamiltonian are called hydrogen-like functions. The orbitals are chosen
to be real. Their angular factor is a spherical harmonic. These are eigen-
functions of angular momentum operators (L2 and its z-projection Lz ) and
stem from the spherical symmetry of atoms. The Lz eigenfunctions define
the quantum number m and are complex, unless m = 0 but may be com-
bined with the conjugate (also an eigenfunction with the same |m| value)
to give orbitals as real wave-functions.
The radial factor decays exponentially, i.e. contains a factor
exp( r/n). This makes the hydrogen-like functions square integrable
and gives an eigenvalue (energy) En for the one-electron Hamiltonian,
completely characterized by the principal quantum number (n) which is a
positive integer.
This expression En = 1/2 (/n)2 was correctly given by the Bohr
model in 1913, however, individual electrons are now known to be correctly
described only by the hydrogen-like orbitals. The energy is quantized (fixed
by integer n 1) and the electron delocalized (Born postulate: the square
of an orbital is interpreted as the density distribution).
The remaining radial factor is a polynomial in r, of order n1 (associated
Laguerre polynomial). This structure gives orthonormal functions. It also
proves the energy eigenvalue is an increasing function of the number of
nodes (n1 zeros of the polynomial). This is a general topological property
of the eigenfunctions, which relates to standing waves in classical physics
and to which we shall return, because it also applies to molecular wave
functions. The parameter is the average nuclear charge felt by the electron.
2A triplet states symmetric spin factors and orbital, anti-symmetric in electron pair exchange (Pauli
principle), imply vanishing density for small r12 . This is stabilising and indicates that electrons tend
to avoid each other. Conversely, a singlet orbital is symmetric with non-zero density even if electrons
collide.
In a one-electron atom, it is just the atomic number. These functions are

often generalized for use in the case of many electron atoms and molecules.
In that case, varies according to the screening effect of other electrons. The
hydrogen-like functions are then called atomic orbitals or basis functions.
A further generalisation is obtained by dropping the n-dependence of
the exponent. The resulting functions are called Coulomb Sturmians and
we shall return to them later for some special properties they possess.
Slater-type orbitals (STO) [1] are linear combinations of the hydrogen-
like functions. Instead of the polynomial in their radial factor they just have
an r n1 factor. This was aimed at facilitating their manipulation (which
it does) but, obviously Slater functions are not orthogonal over r, being
node-less in r. These functions are detailed in Section 7.5.
The applications where Sturmians are at an advantage concern reliable
description of the electron density near the atomic nuclei. This is particularly
useful for evaluating nuclear magnetic resonance (NMR) shielding tensors
and can provide chemical shifts to accuracies comparable with measured
values, particularly for spectra involving such dipolar nuclei as 15 N and
19 F vide infra (Section 7.9.1).
It has also been shown that electron correlation is better accounted for
by exponential type orbitals. Two cases are considered below, the fact
that configuration interaction (CI) requires many fewer exponential than
Gaussian functions and explicitly correlated (or geminal) exponential func-
tions (Section 7.9.2).
A further domain of successful application, which is now developing
fast, is the use of Slater basis trial wave-functions for the correlated ground
state obtained in quantum Monte Carlo simulations (Section 7.9.3).
Nowadays, there is great interest in weakly bound systems, often treated
with difficulty using DFT. It is possible to tailor functionals on the bench-
marks available from accurate Slater-type orbital basis calculations, to
ensure the analytical asymptote which is exponential decay. Nevertheless,
in view of the low energies involved and the fact that they are expressed
as a difference between two large total energies, the strategy so far favored
has been the use of quantum Monte Carlo methods in order to account for
the majority of essential correlation energy.
QMC, however is hellish long, especially when three-body (e-e-N) terms
are required. These highly accurate benchmarks often show the three-body
terms are absolutely essential.
7.1.3. Problems to be solved when using Slater-type orbitals

The versatility and properties of Slater-type orbitals make them natural
basis functions in quantum molecular calculations. Nevertheless, their use
146 P.E. Hoggan
has been rather restricted, mostly due to mathematical integration diffi-

culties. Even today there are no simple general algorithms to solve all the
integrals appearing in a HartreeFock (HF)3 or configuration interaction
(CI) molecular calculation, where integrals involving up to four atomic
centers may appear. More on post HF treatment and electron correlation is
given in Section 7.9.2.
Despite these difficulties, research on Slater orbitals has continued. This
is explained by the requirement for large basis sets of Gaussian orbitals
(GTO) and large wave function expansions to perform more accurate cal-
culations of energy and properties of ever larger interesting systems.
In 1981, at a congress in Tallahassee on exponential type orbitals; Milan
Randic described the situation:
Gaussian functions are not the first choice in theoretical chem-

istry. They are used (. . .) primarily because molecular integrals
can be evaluated, not because they possess desirable properties.
Today this may be a valid reason for their use, but tomorrow
they may be thought of as bastard surrogates, which served their
purpose in the transition period, have no longer viable merits and
will fall into oblivion [2].
This is not the case yet, and may never be, since although Slater-type
orbitals are now viable as an alternative, it sometimes remains faster and
more convenient to use Gaussians. Perhaps a new method will eventually
take over: orbital free/density based or another?
Using GTO expansions of STOs instead of analytic STOs was a prag-
matic solution and originally intended to facilitate numerical integration
in the calculation of the first molecules on early mainframe computers.
The GTO expansion together with the popular distribution of computery
programs like Gaussian (g09) have contributed to the use of GTOs for
accurate calculations of large systems. The size limits in systems studied
have receded, e.g. HF calculations of clusters of hundreds of atoms, CI cal-
culations including hundreds of thousands of Slater determinants. In spite
of the rapid development of the computer technology and the availability
of supercomputers, computational times are unreasonably long, so that the
computational chemist is often restricted to test and model calculations.
This motivates the search for basis functions, where fewer would give a
3 It is assumed the reader is familiar with this SCF method (see Atkins molecular quantum mechanics).
Note, in passing, it is the work of Douglas Hartree and his father, William, and Bertha Swirles (drawing
on the Fock operator).
good CI, in particular. The possibility of using Slater orbitals, where a

minimal basis would consist in one function per atomic orbital has added
some forward impulse to theoretical and computational chemistry.
Since the integration difficulties are purely mathematical in nature, i.e.
definite integrations, it was worthwhile pursuing investigations of Slater
orbitals.
The purpose of this chapter is to explain the key ideas about Slater
orbitals for readers outside the field. The history of Slater orbitals and the
first computer programs using them is presented and computer programs
currently used are listed. The STO and GTO are defined and compared.
The methods used in the literature are explained recalling the key ideas on
which these methods are based. The latest developments within the field
are reported.
7.1.4. Strategy for Slater basis sets

When treating a molecule with at least four atoms, the stumbling block
with Slater-type orbital basis sets is expressing their products on different
atoms, which lead to very slowly converging infinite sums.
It is therefore opportune to reduce the number of sums to evaluate. First,
consider purely ab initio methods, in which no integral is neglected. In this
context, strategies based on atoms in molecules (AIM), or better, diatomics
in molecules (DIM) are in common use.
Atoms in molecules cater well for core electron density, which is almost
unaffected by participation of the atoms in molecules, metals etc. Care is
required in deciding how best to define the core and, naturally, there are no
products on different atoms involved. Therefore, what can be achieved by
AIM is a reduction in the number of orbitals that must be considered to be
delocalized in the molecule.
Diatomics in molecules can be used as building bricks for bigger
systems. The diatomic unit presents the advantage of possessing analytic
closed formulae for all the Coulomb integrals and some exchange integrals,
the remaining exchange integrals being evaluated by sums of simple inte-
grals. Here, the only products of Slater-type orbitals required (required for
larger molecules) on different atoms are those giving three- and four-center
integrals based on the diatomic brick. Obviously, the whole set of diatomic
units in the complete molecule needs to be systematically covered.
Here, the number of infinite sums is reduced by one compared with
single center strategies. None is required within the diatomic unit (instead
of one, in the single center strategy). Only one in three-center cases and,
finally, two in four-center cases.
148 P.E. Hoggan
Chemical intuition can directly be related to these units, in the sense that
the difference between bound and unbound atom pairs is obviously related
to the value if integrals involved, as is the gap between the bonding and anti-
bonding orbitals, when they are well-localized to atom pairs. The key role
is played by the value of exchange integrals, which decrease exponentially
at long range.
The diatomic unit is also useful to implement two approximations, where
accuracy is chosen by the user. The first is the use of the Schwarz inequality,
which defines a product of two center integrals as an upper limit for the
value of the corresponding three- and four-center terms. The second is a
resolution of the Coulomb operator, which may be applied whatever the
basis function used and which is due to P. Gill [70].
In the current state of the art, both of these leave some integrals for
more accurate treatment. The inequality is simply used for screening negli-
gible integrals and the Coulomb resolution may be limited to micro Hartree
accuracy for atomic orbitals with high angular momentum (l 2). This lim-
itation, may, however, be overcome by techniques that reduce the angular
momentum of orbital products (see Section 7.8.7).
At present, it is necessary to evaluate a few of the ugly infinite sums
but there is reason to believe this will shortly be a thing of the past. This
text will relegate the analysis involved (which is a specialist subject) to
Appendix A (see Section 7.12).
7.2. Some Dates: The Story So Far of Slater-Type Orbitals
The history of STOs is an essential part of developing theoretical chemistry.

Slater [1] simplified the hydrogen-like orbitals, obtaining those that bear
his name. Curiously, Slater called these orbitals Hartree orbitals in 1928.
A time-line will be found in Appendix B (Section 7.13).
Very soon, with Slater at MIT, researchers broached the problem of
evaluating the two-electron integrals in the STO basis. During the 1950s
the Chicago group led by Mulliken took on the task of evaluating all the
molecular integrals. Roothaan treated the Coulomb and Hybrid two-center
integrals [6, 7], Rudenberg the exchange integrals [8].4 This exchange
integral algorithm was so efficient that it became the sole technique, also
4 The two-center two-electron integrals are classified according to the centers a, b. Writing them
according to the charge distributions [(1)|(2)] the Coulomb integrals are [aa|bb], the hybrids
[aa|ab] and the exchange integrals [ab|ab]. The most difficult are the exchange integrals because the
charge distribution of every electron is located over two centers.
used for Coulomb integrals in the group (K Rudenberg, during the 50th
Sanibel Symposium, 2010).
Among the many authors who were working around the world on the
solution of the necessary integrals was M Kotani in Japan [9], who wrote
the famous integral tables which bear his name and were widely used.
Coulson in Oxford proposed a method to evaluate the three- and four-center
integrals [10]; Lowdin in Uppsala [11] and a young American scientist
called Harris [12] were involved. Work in the early 1950s mostly focused
on integrals over STO.
The interest was to make the first theoretical calculations of some
molecules starting with the diatomic systems H2 , N2 . For three-center
molecules the problem of integration was encountered (orbital translation).
Mulliken and Roothaan called this The bottleneck of Quantum Chem-
istry [14], Mulliken mentioning it in his Nobel Lecture in 1966 on the
molecular orbital method.
Boys in Cambridge published his landmark paper [15] containing the
evaluation of three- and four-center integrals using Gaussian function
expansions of the STO. This bold step led to great simplifications, based
on the so-called product theorem: the product of two Gaussian functions
located on different centers is a new Gaussian function located on a new
center. Thus, four-center electron distributions could be reduced to single-
center distributions and analytical evaluation was greatly facilitated. Boys
regarded this as an existence theorem for a closed GTO product rule. It
was to change the course of molecular computations. Note that the product
theorem for Slater orbitals leads to complicated infinite sums, making eval-
uation awkward compared with the simple closed forms for Gaussians.
In 1954 Boys, Shavitt et al. [16] expanded Slater orbitals into Gaussians
to perform quantum mechanical calculations. In 1963 Clementi presented
the so-called basis set using Slater exponents/orbitals [17]. Later, Pople
would base his programs on Boys pragmatism (see [23]).
7.3. Computer Programs Using Slater-Type Orbitals
The first manual calculation of a molecule, the N2 molecule, was done by

Scherr in 1956. It was the work of two men for two years. Afterwards, this
calculation was reproduced by the first digital computer calculation [18,19],
taking 35 minutes. Many of those involved are still keen participants in the
field who describe how exciting this groundbreaking period was, when
everything needed to be created to use the new computers.
150 P.E. Hoggan
In 1962 Shull initiated the Quantum Chemistry Program Exchange

(QCPE) at Indiana University, distributing software free to academics.
The first automatic computer program for molecular electronic structure
calculation was POLYATOM [20] using GTOs in the SCF-LCAO scheme.
The program was developed at MIT in 1963 when Slater was there. In 1963
the program IBMOL [21] was developed by Clementi and others when he
visited the Chicago group.
In 1968 the STO code was developed by Scrocco and Tomasi in Pisa [4].
Preliminary work by Scrocco is reported in Italian as early as 1962. It was
also used by Berthier in Paris.
The program ALCHEMY was originally developed in 1968 using Slater
orbitals by McLean and Yoshimine at IBM in San Jose [22], afterwards, the
new ALCHEMY II by Bagus and others used GTOs.
In the 1970s, the advent of Gaussian [23] saw development in the STO
field hibernate somewhat for a couple of decades. By the 1990s several
groups around the world developed new STO computer programs which
are distributed. The program STOP by Bouferguene and Hoggan [24] was
published first in 1996. It is based on the single center strategy and was first
presented in 1994 (Eighth International Congress of Quantum Chemistry
(ICQC) in Prague). It has been updated regularly, including last year (2010).
Then, in 1998 a program was written using Fourier Transforms by
Steinborn, Weniger, Homeier et al., in Regensburg [25]. The program
SMILES by Fernandez Rico, Lopez, Ema,Aguado, Ramrez et al. in Madrid
was first available in 1998 and new versions have appeared, the latest in
2004 for the HF and CI calculations of molecules [26].
The program CADPAC [27] in Cambridge uses techniques like density
fitting, involving auxiliary Slater-type orbital basis sets to perform Hartree
Fock and Density Functional Theory (DFT) calculations with a reduced
number of indices in requisite integrals. The authors aimed to obtain better
nuclear magnetic resonance (NMR) chemical shifts on the basis possessing
nuclear cusps.5
Despite this, Slater work was often considered as tilting at windmills,
i.e. fighting unwinnable or futile battles, before 2000.6 The presentation of
STOP in 1994 met with considerable scepticism, exacerbated by the rustic
basis sets available for Slaters compared to Gaussians then. However, by
5 i.e. the ratio of radial derivative to the value of the function, when r tends to zero gives the exponent:
nuclear cusp condition.
6 This expression has been borrowed in English from Don Quixote by Cervantes.
the turn of the century, better bases and CI work had made presentations of
such work far more convincing.
7.3.1. Numerical grid methods

In the density functional theory field, the program ADF (Amsterdam
Density Functional) [28, 29] begun in 1973 by Baerends et al. uses Slater
orbitals for their calculations. This much-used package offers a very
extensive series of atomic basis sets for input, including most elements. It
adopts a numerical grid strategy, the potential being resolved by numerical
Green function methods before grid integration.
7.3.2. Configuration interaction

The program ATMOL by Bunge et al. performs large highly accurate CI
calculations on atoms using Slater orbitals [30]. Early in the third mil-
lennium much interest is concentrated in generating more efficient algo-
rithms for correlation in molecules, use of non-integer Slater orbitals,
numerical solution of integrals when working in momentum space and
in the electron correlation with Hylleraas wave functions.
7.4. Slater Orbitals and Gaussian Orbitals
This chapter will focus on analytical manipulation of atomic orbitals. Nev-

ertheless, a brief mention in passing deservedly goes to numerical orbital
techniques, that the interested reader will find detailed by Talman [13].
It is well known that hydrogen-like orbitals are the solution of the
Schrodinger equation for a one-electron atom. For helium and atoms with
more electrons, the Schrodinger equation has no analytical solution due
to the potential term 1/rij which correlates the (otherwise) independent
electrons. It is assumed that for systems with two or more electrons, the
exponential er will be the asymptote of the formal solution (i.e. the same
asymptotic behaviour of orbitals is expected for many electron atoms as
that obtained for the analytic one-electron eigenfunctions). The same holds
for molecular orbitals.
Asymptotic behaviour of energy terms is also characteristic.7
7 Electron repulsion energy (e.g. two-center): the Coulomb terms decay in 1/R for large interatomic
distance R and the exchange terms in ekR . This implied exchange is purely a quantum (spin related)
phenomenon. For large R exchange is zero, the classical Coulomb law remains.
152 P.E. Hoggan
0.9 STO
0.8 GTO
0.7
0.6
Wave function 0.5
0.4
0.3
0.2
0.1
-3.0 -2.5 -2.0 -1.5 -1.0 -0.5 0 0.5 1.0 1.5 2.0 2.5
r
Fig. 7.1. Comparison of the shape of a STO with a GTO 1s function. Radius (r)
in atomic units.
The hydrogen-like orbitals have nodes, i.e. the 2s orbital is of the form
(1 br)er , and higher quantum number orbitals are similar but STOs are
node-less. A related problem appears for Gaussians.
In 1928 Slater [1] noted that the radial polynomial factors make cal-
culations messy and proposed the use of single powers of r, i.e. linear
combinations of hydrogen-like terms.
A picture which helps to visualize the differences between Slater and
Gaussian orbitals is the representation of the 1s orbital function of both
types (with suitable exponents), see Fig. 7.1.
STOs represent the electron density well near the nucleus (cusp)
and at long range (correct asymptotic decay). STOs thus resemble the
physical atomic orbitals provided a suitable exponent has been obtained by
optimization.
Conversely, the GTOs have no cusp (zero radial derivative at the nucleus)
and decay too fast.
To reproduce a 1s STO using three GTOs (the so-called minimal GTO
basis) an orbital is obtained with the (bell-shape) of a Gauss curve, no cusp,
see Fig. 7.2.
To reproduce a single STO many GTOs are necessary, but the electron
cusp at the nucleus is missing. This is one of the reasons for the slow con-
vergence of the wave function solutions to the exact (HF or CI) result.
In general, if the basis function is not built up of eigenfunctions of the
Schrodinger equation, its convergence is slower. In GTOs more Slater deter-
minants are needed for a given accuracy.
Another advantage of Slater orbitals is the size of the basis, one orbital
per electron is of reasonable quality and multiple-zeta basis sets converge
0.9 STO
0.8 GTOs
0.7
Wave function
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.5 1.0 1.5 2.0 2.5
r
Fig. 7.2. Construction of a 1s STO with three GTOs. Radius (r) in atomic units.
fast to the HartreeFock limit. This is the lowest energy solution for inde-
pendent electrons. A basis approaching this limit is said to saturate i.e.
adding higher angular momentum functions leads to little improvement.
A Slater-type orbital basis saturates much faster than GTOs. Therefore, the
number of integrals to be evaluated in an STO basis is dramatically smaller.
CI is also spectacularly more efficient (see Section 7.9.2). Finally, con-
ceptually the Slater orbitals give a more intuitive description of the atomic
orbitals and of the molecular orbitals (MO) using them.
The disadvantages of Slater orbitals have already been mentioned: the
three- and four-center two-electron integrals are the bottleneck. There is
no general analytical solution for them, which would be the most effective
and fastest way of calculation. Instead, there are a number of approximate
methods of calculation, involving infinite series, or truncated approxima-
tions to the Coulomb operator itself. They will be treated in Section 7.7.8.
Radial Slater functions do not represent the bonding region adequately
and higher angular momentum functions should be added.
It is nevertheless possible to use linear combinations restoring radial
nodes. This approach is advocated particularly for ADF, where the
hydrogen-like basis is obtained by fixing the coefficients for combining
Slater functions.
Another disadvantage is that some of the two-center integrals since
the times of Roothaan and Rudenberg have been solved for a co-axial
154 P.E. Hoggan
conformation of the atomic coordinate systems (the z-axes are directed

towards each other) that is not the molecular frame. Therefore, orbital rota-
tions are necessary. These problems have been solved, but require additional
calculations [31].
Slater orbitals are used in atomic calculations, especially in highly
accurate calculations of atoms using Hylleraas wave functions (with explicit
r12 dependence, and also in diatomics. They are used in DFT and in Density
matrix theories. Traditionally they have been used in semi-empirical calcu-
lations, where of course the three- and four-center integrals were neglected.
Gaussian orbitals are often used in ab initio work but they are poor
for the calculation of properties where the density near nucleus has to be
well described. The radial dependence is unsatisfactory and the number of
integrals increases with the dimension of the basis dramatically.
The major advantage of GTOs is the existence of a product theorem. For
many years, workers have improved the calculation of the necessary inte-
grals, having achieved a considerable speed-up. For example, the Coulomb
operator with a Laplace transform enables three- and four-center integrals
to be calculated using two-center integral techniques [46].
Concluding, the main defect of GTO expansions is the absence of cusp
and rapid decay which slows the convergence and the large number of
integrals to be computed.
7.5. Types of Exponentially Decaying Orbitals, Based

on Eigenfunctions for One-Electron Atoms
The atom-centered Slater orbitals are defined as:
nlm (r) = rn1 er Ylm (, ), (7.1)
where n, l, m are the quantum numbers (recall l integer n 1 and l

m l). The radial factor is g(r) = rn1 er , for integer n 1. The
angular factors are the spherical harmonics defined using the Condon
Shortley phase:
1/2
2l + 1 (l m)!
Ylm (, ) = (1) m
Plm (cos )eim , (7.2)
4 (l + m)!
where Plm (cos ) are the associated Legendre functions. The spherical har-
monics are eigenfunctions of the angular momentum operator L2 and its
z-projection Lz .
The complex spherical harmonics are used mainly in atoms and in devel-
oping theories because it is easier to work out general formulae and deriva-
tions with them. The real spherical harmonics are linear combinations of
the complex ones. Orbitals are chosen to be real in atoms and molecules.
Note that they are written using polar coordinates (suitable for atoms).
Cartesian Slater-type orbitals are very seldom used compared with
Cartesian Gaussians, which are an almost systematic choice.
When the principal quantum number n in Eq. (7.1) is a non-integer we
have the NISTOs (non integer slater orbitals). The main difficulty when
working with these orbitals is during the derivations a binomial has to be
used with an non-integer power what leads to an infinite expansion. These
orbitals are widely investigated in the present [32]. The additional flexibility
of using non-integer quantum numbers results in a lowering in the energy
and better density. It is also the possible to transform from polar coordinates
to elliptical coordinates.
Elliptical Slater orbitals have been used extensively as basis functions
for two-center molecules [3335]. These orbitals are known to lead to lower
energy results [36]. Using = 1 = ra + rb and = 1 = ra rb ,
nlm (r) = n l ( 2 1)m/2 (1 2 )m/2 e eim , (7.3)
where , , are the elliptical coordinates.
7.5.1. Orbitals which are linear combinations

of Slater-type orbitals
The most commonly used are B-functions [37], hydrogen-like orbitals,
Sturmians [38]. More general functions have been tried by Guseinov, but
have not proven more useful [51]. The B-functions are related to Bessel
functions. The orbitals have some helpful properties, like a compact Fourier
transform. Written in the form
n
(2n j 1)!
Bnlm (r) = (r)l+j1 er Ylm (, )
2 2n+l1 (n + l)!(n j)!(j 1)!
j=1
(7.4)
one can see that they are a linear combination of Slater orbitals. The angular
parts are the spherical harmonics.
The Fourier transform B(p) of B(r) is expressed simply as:

2 2 n+l1 (i|p|)l
Bnlm (p) = Y m (p , p ). (7.5)
( 2 + |p|2 )n+l+1 l
156 P.E. Hoggan
The hydrogen-like orbitals which are solutions of the Schrodinger equation

for the hydrogen atom have a radial part which is a Laguerre polynomial.
The polynomial and the exponent coefficient depend on the atomic number
Z and the principal quantum number n:

2l+2 2Zr l Zr m
nlm (r) = Nnl Lnl1 r e n Yl (, ), (7.6)
n
therefore, hydrogen-like orbitals do not form a complete set (for finite n),
they need orbitals of the continuum to be complete. This is important for the
convergence of the solutions. Shull and Lowdin [39] realized that this was
due to the dependence of Z with n that dilates the orbitals and they pro-
posed the following orbitals where these were substituted by adjustable
parameters, i.e. usual orbital exponents:
l r m
nlm (r) = Nnl L2l+2
nl1 (2r)r e Yl (, ). (7.7)
These orbitals were subsequently called Coulomb Sturmians because they

fulfill the so-called SturmLiouville theorem for eigenfunctions of such
differential equations, with central Coulomb attraction. These orbitals are
therefore a complete set.
Section 7.7 on literature integration methods will show how these kinds
of orbital have been used.
7.6. Types of Integral Over Slater Orbitals
Due to the form of the Hamiltonian and of its expectation value we find
the following kinds of integral. First the integrals which appear when using
HartreeFock and CI wave functions, in general ab initio methods. The
integrals are classified according the number of electrons and atomic centers
involved. In order of increasing difficulty, there are:
7.6.1. One-electron integrals

These are the one- and two-center overlap integrals a|b, kinetic energy
integrals a| 2 /2|b and two-center nuclear attraction terms a|1/rb |b.
Another case of one-electron integral is the three-center nuclear
attraction, originating from the nuclear attraction operators in the Hamil-
tonian: a|1/rc |b.
This term is often abbreviated NAI and is also used in DFT work.
7.6.2. Two-electron integrals

They can have up to four atomic centers because of the determinant
giving the wave-function and thus the four orbitals which form the integral
involving the Coulomb operator.8
Over Slater-type orbitals, the following are available in closed form:
Single center: any (may be pre-calculated and stored). Numerical value
depends on a factor involving , which is known on input.
Two-center: all Coulomb integrals and those exchange integrals
involving atomic orbitals with equal exponent values. Other exchange inte-
grals may be evaluated to arbitrary accuracy but require an extrapolated
infinite sum (Podolanski [5] and Summary in Appendix C, Section 7.14).
The predominant methodology involves separating the variables in prolate
spheroidal coordinates (see [6]).
The Podolanski text on two-center exchange integrals was published in
German in 1931 and seems to have been little quoted, although it sets out an
efficient scheme for evaluating two-center repulsion terms in HartreeFock
calculations (see Appendix C, Section 7.14).
Three-center: almost none. One orbital translation (and the con-
comitant infinite sum) is required to operate in a two-centered system.
Four-center: none. Two orbital translations (and the concomitant
infinite sum) is required to operate in a two-centered system. It will be
shown later that the three- and four-center cases may be approximated well
using Gills Coulomb resolution.
Since the two-center integrals can be calculated quite cheaply, the
Schwarz inequality may be used to establish an upper bound on three- and
four-center terms useful for screening the method used to evaluate them
and even to establish a cutoff.
2 2 (b) 2 (d) 2
g(a) (1) h(c) (2) g (1) h (2)
d1 d2 d1 d2
r12 r12
(a) 2
g (1)g(b) (2)h(c) (1)h(d) (2)
d1 d2 . ,
r12
where a, b, c, d are distinct nuclei on which the functions are centered.

The two-center integrals have been the most thoroughly investigated,
they have the following nomenclature: The Coulomb integrals where the
8 1/r two independent electrons, possibly two distinct orbitals in each of the bra and ket which may
12
all be on different atoms i.e. up to four atomic centers.
158 P.E. Hoggan
charge distribution of each electron is located at a center: [aa|bb]. Hybrid

integrals, one charge distribution is located at one center and the other over
two centers [aa|ab] and their equivalents [bb|ab]. The exchange integral is
more difficult, since it leads, in cases of different exponents, to an infinite
sum. Both electrons are located on two centers: [ab|ab]. To solve these
integrals a change to elliptical coordinates is useful. The Coulomb potential
in elliptic coordinates must be written after a change of variable to avoid
singularities (see Section 7.6).
In actual calculations, the Coulomb and Hybrid integrals are evaluated
analytically, numerous methods exist. The exchange integrals are calculated
with great accuracy.
The three-center integrals are of several types [aa|bc], [ab|ac]. For dif-
ferent exponents there is no analytical solution in closed algebraic form,
i.e. all require numerical quadrature.
The four-electron integrals are of the type [ab|cd]. All require quadra-
tures.
7.6.3. Three- and four-electron integrals

They appear in the Hylleraas-CI method [40] when using one inter-
electronic distance rij per configuration. For the two-center case they have
been solved generally by Budzinski [41]. Three- and more atomic center
integrals have not been solved yet.
These can be multi-center integrals, as every electron from right and
left in the expectation value expression may be on a different atomic
center. These integrals are of the type, i.e. the easier [aa|r12 r13 |ab|bb], to
the most difficult [ab|r12 r13 |ab|ab]. Four-electron integrals [aa|r12 r13 /r14 |
bb|ab|bb], and so on.
For three- and more atomic centers, three- and four-electron integrals
occur, with as many atomic centers as the molecule has (up to eight). These
integrals are still not solved. Interest nowadays focuses on the solution of
two- and three-center molecules using explicitly correlated methods.
7.7. Integration Methods in the Literature
It is useful to obtain a gold-standard expression of the integrals from

symbolic programming, to guide numerical evaluation. This has been
carried out by choosing the appropriate number of required digits and
expanding numerous terms of infinite series in such symbolic programming
systems as Maple by Harris and Pinchon [144] [54], in particular.
A detailed repertoire is located on their websites. In this section the main

evaluation methods of the three- and four-center integrals over Slater
orbitals from the literature will be explained. The methods are approximate
because they consist in transformations, expansions or include numerical
integrations (quadratures). Therefore they are not as effective as ana-
lytical integration. Nevertheless, using these methods to evaluate such
integrals is possible, and the programs are competitive with those using
Gaussians.
7.7.1. Single-center expansion

In the single-center expansion method, the Slater-type orbitals located
at different centers are expanded about only one of them and then inte-
grated as for atoms. The translation method consists in selecting an
atom as origin then the translating the distant orbitals from their atom to
this origin. Therefore, both methods are essentially the same. To expand
one function centered in A at another point B the following expression
is used:

Ai = Ai Bj d Bj . (7.8)
j=1
This formula is due to Smeyers [42]. In the brackets are the requisite coef-
ficients. The various methods of single-center expansion differ in the tech-
nique to calculate these coefficients.
This approach was first proposed by Barnett and Coulson [10] in 1956
using radial orbitals (s-orbitals) and was called the zeta function method
because of expansions in terms of successive derivatives with respect to
exponents. The terms have alternate sign.
The method is similar to Lowdins alpha function method [11]. Harris
and Michels [43] extended the method to angular general orbitals in 1965.
This method has been used by many and Appendix A (Section 7.12) details
it. Signs of terms give oscillating sums and poor convergence.
The idea is the translation of an orbital from one point to another. Trans-
lation of a spherical harmonic is a finite expansion; on the other hand, trans-
lation of the radial part leads to an infinite series. This situation can be best
explained following Guseinov [44]:
n
1
l
n,l,m (, rA ) = Vnlm,n l m (, RAB ) n ,l ,m (, rB ), (7.9)
n =1 l =0 m =l
160 P.E. Hoggan
where V are the coefficients of the expansion. The method is very stable but
it requires computation of many terms to obtain sufficient correct decimal
digits, therefore this method gives lengthy computation times.
7.7.2. Gaussian expansion

This is the BoysShavitt method [45], which consists in solving some
integrals over Slater orbitals by expanding them into a finite series of
Gaussians:

NG
er = ci ei r
2
(7.10)
i=1
ci and i are obtained by minimizing the least squares difference between
exponential and sum. This method and some improvements of this method
are used at present in the program SMILES [26]. As NG is usually larger
than the number of the primitives when using only Gaussian basis sets,
the number of integrals to calculate is large. The method is very stable and
robust. It requires lengthy computation times to get accurate integral values.
It can be made numerically accurate [54], yet the basis will not have the
analytic properties of STOs and this can be a source of problems when first
and higher derivatives are required. It should not be used when alternatives
are available and is often reserved for speeding up four-center terms.
7.7.3. Gaussian transform method

The Gaussian transform method by Shavitt and Karplus 1965 [46] has
probably been the most widely used method. It consists in the Laplace
transform of the exponential function, here exemplified by the simplest one
i.e. a 1s orbital:

er = (s3/2 e /(4s) ds)esr .
2 2
(7.11)
2 0
Every Slater exponential within the integral is transformed into a Gaussian.
To integrate this, one has to solve the integrals over the variable s which
have a special form. This integral is solved numerically. Therein lies the
methods disadvantage. Furthermore, higher angular momentum functions
require the derivatives with respect to which can amplify any inaccuracy.
7.7.4. Fourier-transform method

The B-functions Eq. (7.4) proposed by Filter and Steinborn in 1978
[37] have a highly compact Fourier transform (Eq. 7.5). The group of
Steinborn has developed this method [25]. The evaluation of integrals using
B-functions leads to some integrals including a Bessel function of the first
kind which is oscillatory (i.e. like the sum it replaces in Section 7.7.1):

r n er Jl+1/2 (rx)dr (7.12)
0
To evaluate these accurately, extrapolation methods are used [47, 48] due
to Sidi [49] and [50], i.e. substituting this integral by a sine integral which
has the same behavior. This evaluation requires numerical integration (e.g.
GaussLegendre quadrature, summing the integrand at polynomial roots).
7.7.5. Use of Sturmians

The Sturmians were proposed by Shull and Lowdin in 1956 [39]. Smeyers
used Sturmians to evaluate three-center nuclear attraction integrals using
the single-center expansion [42]. Guseinov also uses them extensively [51].
The Sturmians Eq. (7.7) satisfy the SturmLiouville equation and theorem
implying they comprise a complete orthonormal set:

2n m
Sn,l (, r) =
2 m 2
Sn,l (, r). (7.13)
r
The so-called Coulomb Sturmians orthogonalize the Coulomb potential in
their argument. This generally applies to the attraction term, at least for
one-electron functions. Geminals useful for explicit correlation have also
been used. A seminal text by Avery gives more details to the interested
reader on this subject [38], see also Chapter 6 of the present book.
7.7.6. Elliptic coordinate method

From the Slater-type orbital expression, define the radial factor g(r):
g(r) = r n1 er .
Then, (from the spectral forms in [63]), the potential due to this distri-
bution is immediately written (with g short for g(r)):
l (g) = r2 F(r),
1
F(r) = du g(ru) u +
l+2
du g(ru) u1l . (7.14)
0 1
F(r) is given in Eq. (7.14), with a suitable integration variable; u.
This expression is used to write all radially dependent one- and two-
center integrals in analytical closed form.
162 P.E. Hoggan
y1a y1b
1 (1, 1, 1)
r1a r1b 1a= 1b= 1
1a R 1b
a b
x1a z1a z1b
x1b
Fig. 7.3. Transformation from polar to elliptical coordinates.
Subsequently, the elliptic coordinate method is the transformation of the

polar orbital coordinates into elliptical ones , according to Fig. 7.3. The
two coordinate systems pointed to each other so that the elliptical angle
coincides with polar angle . This transformation is:
R R
r1a = (1 + 1 ), r1b = (1 1 ) (7.15)
2 2
1 + 1 1 1 1 1
cos 1a = , cos 1b = (7.16)
1 + 1 1 1
[(21 1)(1 21 )]1/2 [(21 1)(1 21 )]1/2
sin 1a = , sin 1a = .
1 + 1 1 1
(7.17)
The volume element and the domain change are:

2
r 2 dr sin d d
0 0 0
+1 2
R3
d1 d1 (21 21 ) d1 . (7.18)
8 1 1 0
The method is used by many authors. It was recently used for three-electron
integrals [52].
7.7.7. Monte Carlo integration

In this stochastic procedure the expected value of the integral is set at the
exact value and adjusted according to trials governed by variational criteria.
7.8. General Two-Electron Exponential Type Orbital Integrals

in Polyatomic Molecules Without Orbital Translations
7.8.1. Introduction
The Coulomb resolution now will be presented. This is a readily controlled
approximation to separate the variables in the 1/r12 which, in recent work
by Gill and by Hoggan [70, 74] is shown to spell the end of exponential
orbital translations and ensuing integral bottlenecks.
This section advocates the use of atomic orbitals which have direct
physical interpretation, i.e. hydrogen-like orbitals. They are exponential
type orbitals (ETOs).
Until 2008, such orbital products on different atoms were difficult to
manipulate for the evaluation of two-electron integrals. The difficulty was
mostly due to cumbersome orbital translations involving slowly convergent
infinite sums. These are completely eliminated using Coulomb resolutions.
They provide an excellent approximation that reduces these integrals to a
sum of one-electron overlap-like integral products that each involve orbitals
on at most two centers. Such two-center integrals are separable in prolate
spheroidal coordinates. They are thus readily evaluated. Only these integrals
need to be re-evaluated to change basis functions.
The above is still valid for three-center integrals. In four-center integrals,
the resolutions require translating one potential term per product. This is
outlined here.
Numerical results are reported for the H2 dimer and CH3 F molecule.
The choice between Gaussian and exponential basis sets for molecules
is usually made for reasons of convenience at present. In fact, it appears
to be constructive to regard them as being complementary, depending on
the specific physical property required from molecular electronic structure
calculations.
As regards exponential type orbitals (ETOs) such as Slater functions,
much analysis suggests it is difficult to evaluate two-electron integrals
because the general three- and four-center integrals evaluated by the usual
methods require orbital translations. Some workers avoid the problem using
large GTO expansions, e.g. SMILES [53, 54].
It would be helpful to devise a separation of variables for integration.
This would eliminate orbital translations and therefore present major advan-
tages, although some other translations remain involving a simple analytic
potential.
The present section describes a breakthrough in two-electron integral
calculations, as a result of Coulomb operator resolutions. This separates
164 P.E. Hoggan
the independent variables of the operator and gives rise to simple analytic
potentials. The two-center integrals are replaced by sums of overlap-like
one-electron integral products. One potential term in these products requires
translation in four-center terms, which is significantly simpler to carry
out than that of the orbitals. This implies a speed-up for all basis sets,
including Gaussians. The improvement is most spectacular for exponential
type orbitals. A change of basis set is also facilitated as only these one-
electron integrals need to be changed. The Gaussian and exponential type
orbital basis sets are, therefore interchangeable in a given program. The
timings of exponential type orbital calculations are no longer significantly
longer than for a Gaussian basis, when a given accuracy is sought for
molecular electronic properties.
For STOs, nano-Hartree accuracy of Coulomb resolutions is acces-
sible for AO angular momentum to l = 2, beyond this falls to at worst
milli-Hartree accuracy, they should thus be used systematically until the
last stages of SCF, when high accuracy required implies that precautions
described in Section 7.8.7 should be taken for high angular momentum, or
treating the coupling relations.
7.8.2. Basis sets

Although the majority of electronic quantum chemistry uses Gaussian
expansions of atomic orbitals [15, 45], the present work uses exponential
type orbital (ETO) basis sets which satisfy Katos conditions for atomic
orbitals: they possess a cusp at the nucleus and decay exponentially at long
distances from it [5557]. It updates a real chemistry interest beginning
around 1970 and detailed elsewhere [3, 4, 17, 29, 46, 58, 59]. Slater-type
orbitals (STOs) [60, 61] are considered here.
STOs allow us to use routines from the STOP package [24, 62] directly.
The integrals may be evaluated after Gaussian expansion or expressed as
overlaps to obtain speed up [63]. Exponents may be optimized solving a
secular determinant as in [38, 64].
7.8.3. Programming strategy

Firstly, the ideal ab initio code would rapidly switch from one type of basis
function to another.
Secondly, the chemistry of molecular electronic structure must be
used to the very fullest extent. This implies using atoms in molecules
(AIM): atom cores are conserved in an implementation due to Rico
et al. [53]. We prefer to use diatomics in molecules (DIM) using the

concept by Tully [65] from the outset, implemented in our previous
work [62, 66]. The natural choice of atomic orbitals, i.e. the Stur-
mians or hydrogen-like orbitals lend themselves to the AIM approach.
To a good approximation, core eigenfunctions for the atomic hamil-
tonian remain unchanged in the molecule. Otherwise, atom pairs are
the natural choice, particularly if the Coulomb resolution recently advo-
cated by Gill is used. This leads us to products of auxiliary overlaps
which are either literally one- or two-centered, or have one factor of
the product where a simple potential function is translated to one-atomic
center.
The Slater basis set nightmare of the Gegenbauer addition theorem is
completely avoided (see, Appendix A, Section 7.12). Naturally, the series
of products required for, say a four-center two-electron integral may require
10 or even 20 terms to converge to chemical accuracy, when at least one
atom pair is bound but the auxiliaries are easy to evaluate recursively and
re-use. Unbound pairs may be treated using approximate methods.
Now the proposed switch in basis set may also be accomplished just by
re-evaluating the auxiliary overlaps. Furthermore, the exchange integrals
are greatly simplified in that the products of overlaps just involve a two-
orbital product instead of an exchange operator. The resulting cpu-time
growth of the calculation is N 2 for SCF, rather than N 4 . Further gains may
be obtained by extending the procedure to post-HartreeFock techniques
involving explicit correlation, since the r12 1 integrals involving more than
two electrons, that previously soon led to bottlenecks, are also just products
of overlaps. This Coulomb resolution is diagonal in Fourier space in some
cases.
7.8.4. Avoiding ETO translations for two-electron integrals over

three- and four-centers.
Previous work on separation of integration variables is difficult to apply, in
contrast to the case for Gaussians [67, 68]. Recent work by Gill et al. [69]
proposes a resolution of the Coulomb operator. This is based on separating
the variables of r112 , by determining suitable functions of r1 and r2 that treat
these variables equivalently and constitute a complete set which orthog-
onalizes the Coulomb operator. The associated potentials i provide an
expansion, or resolution of r112 similar to that of the identity (using the
summation convention):
|gi gi | = I. (7.19)
166 P.E. Hoggan
This is the completeness property of a set of orthonormal functions [gi ],

within a particular Hilbert space. Similarly, for the Coulomb operator,
suitable potentials give:
1
= |i i |. (7.20)
r12
This Coulomb resolution is based on a complete set of functions which may
be determined such that they impose the identity as matrix representation
of the Coulomb operator, r112 in this basis:

1
fi fj = ij . (7.21)
r12
The completeness relation for the associated potentials can also be written
in the form of Eq. (7.20) The functional expression of the above gives:
1
= i (r1 )i (r2 ). (7.22)
r12
The potential functions i , are solutions of Poissons equation. The
functions chosen may also be based on Coulomb Sturmians (see the work
by Avery, e.g. Chapter 6 of the present book and references therein).
Completeness of the functions fi allows us to expand a density in terms
of them (using Eqs. (7.19) and (7.21)):

1
(r)| = (r) fi (r)fi (r)|. (7.23)
r12
J is re-written, summing over i and j:

1
J12 = (r1 ) (r2 )
r12

1 1
= (r1 ) fi (r1 )fi (r1 )
fj (r2 )fj (r2 ) 1 (r2 ).
r r r
12 12 12
(7.24)
Introducing the orthogonalized operator from Eq. (7.21) to resolve the

two-electron integral into a sum of products of one-electron overlap-like
integrals:

1 1
J12 = (r1 ) fi (r1 )fi (r2 ) (r2 ) implied sum over i. (7.25)
r12 r12
And recalling the defining relation for potentials (i.e. one-electron functions
of a single radial variable):

1
fi (r) = |i (r) (7.26)
r
12
J12 = (r1 )i (r1 )i (r2 )(r2 ) with implied sum over i (7.27)
In addition, the potentials must ensure rapid convergence of the above
implied sum in the resulting expression for Coulomb integrals J12 as
products of auxiliaries i.e. overlap integrals, as detailed in [69].
This technique can be readily generalized to exchange and multi-center
two-electron integrals [74]. For two-center terms it is helpful to define
structure harmonics by Fourier transforms, limiting evaluation to non-zero
terms [70].
The requisite potentials and auxiliaries are given in Appendix D
(Section 7.15).
This assumes tacitly that the potential obtained from the Coulomb
operator resolution be centered on one of the atoms. Whilst this choice
can be made for one pair in a four-center product, it cannot for the second.
There remains a single translation for this potential in one auxiliary of the
two in a product representing a four-center integral and none otherwise.
This method obviates the need to evaluate infinite series that arise
from the orbital translations efficiently. They have been eliminated in the
Coulomb operator resolution approach, since only orbitals on two centers
remain in the one-electron overlap-like auxiliaries. These can be evaluated
with no orbital translation, in prolate spheroidal coordinates, or by Fourier
transformation [70, 74].
7.8.5. Numerical results of Coulomb resolutions: efficiency

First a test system is studied, built up of four hydrogen atoms. The second
example is the full RHF calculation of CH3 F using the Coulomb resolutions.
Consider the H2 molecule and its dimer/agregates. In an s-orbital
basis, all two-center integrals are known analytically, because they can
be integrated by separating the variables in prolate spheroidal coordinates.
A modest s-orbital basis is therefore chosen, simply for accuracy demon-
stration on a rapid calculation, for which some experimental data could be
corroborated.
The purpose of this section is to compare evaluations using the Coulomb
resolution to the exact values, obtained analytically. The IBM Fortran com-
piler used is assumed to be reliable to 14 decimals in double precision. The
168 P.E. Hoggan
worst values in the Coulomb resolution approximation have ten correct

decimals for two-center integrals with a 25-term sum.
Timings are then compared for translation of a Slater-type orbital basis
to a single center (STOP) [62] with the Poisson equation solution using
a DIM (diatomics in molecules or atom pair) strategy and finally to show
that the overlap auxiliary method is by far the fastest approach, for a given
accuracy (the choice adopted is a sufficient six decimals, for convenient,
reliable output).
7.8.5.1. H2 molecule with interatomic distance of 1.402 atomic

units (a.u.)
One- and two-center Coulomb integrals may be obtained analytically and
Coulomb resolution values compare well with them [69].
The total energy obtained for the isolated H2 molecule is
1.1284436 a.u. as compared to a HartreeFock limit estimate of
1.1336296 a.u. Nevertheless, the Van der Waals well, observed at 6.4 a.u.
with a depth of 0.057 kcal/mol (from Raman studies) is quite reasonably
reproduced [77].
7.8.5.2. Dimer geometry: rectangular and planar

The distance between two hydrogen atoms of neighboring molecules is
6 a.u. This alone justifies the expression dimer the geometry corre-
sponds to two almost completely separate molecules, however, the method
is applicable in any geometry (for 3 a.u. all three- and four-center integrals
evaluated by Coulomb resolution (OVERLAP) agree with those of STOP
to at least six decimals).
Timings on an IBM RS6000 Power 6 workstation, for the dimer (all
four-center integrals in milliseconds):
STOP: 12 POISSON: 10 OVERLAP: 2. Total dimer energy:
2.256998 a.u. This corresponds to a well-depth of 0.069 kcal/mol, which
may be considered reasonable in view of the basis set.
7.8.5.3. Selected exchange integrals for the CH3 F molecule

(evaluated using the Coulomb resolution)
Geometry and exponents are from [80]: Bond-lengths CH = 2.067 and
CF = 2.618 a.u. tetrahedral angles.
No symmetry is assumed but geometric relationships are observed,
as well as those due to m values, at least to the nano-Hartree accuracy
chosen. For illustrative purposes, three-center exchange integrals are tabu-

lated below in a real basis. Timings on IBM RS6000 Power 6 workstation
for all two-electron integrals: STOP: 1.21s, OVERLAP: 0.17s.
All the two-electron integrals are identical to better than six signif-
icant figures with those obtained using the STOP software package [62].
The factor limiting precision in this study is the accuracy of input. The
values of Slater exponents and geometric parameters are required to at
least the accuracy demanded of the integrals and the fundamental constants
are needed to greater precision than that available in some cases.
7.8.6. Perspectives and conclusions

A remarkable gain in simplicity is provided by Coulomb operator resolu-
tions [69], that now enables the exponential type orbital translations to be
completely avoided in ab initio molecular electronic structure calculations.
This breakthrough that Coulomb resolutions represent (in particular
with the convenient choice of Laguerre polynomials: see Appendix D,
Section 7.15) in the ETO algorithm strategy stems from a well-controlled
approximation, analogous to the resolution of the identity. The convergence
has been shown to be rapid in all cases (an extensive benchmark, with focus
on accuracy will soon be ready for publication and extends the prelim-
inary study [70]). The dependence of accuracy on angular momentum is
the current issue, for which there is always the possibility of momentum
reduction by coupling, leaving a few integrals to be evaluated by one-
dependent BCLF methods (see appropriate sections of this chapter and
appendices).
The applications to H2 dimer Van der Waals complexes and CH3 F uses
a general code within the STOP package [62]. They show the Coulomb
Table 7.3a. Orbital exponents.
AO No. n l m zeta
01 1 0 0 5.6727
02 2 0 0 1.6083
35 2 1 m 1.5679
06 1 0 0 8.5600
07 2 0 0 2.5600
810 2 1 m 2.5200
H 1 0 0 1.2400
July 20, 2011
170
9:7
Table 7.3b. Selected examples of three-center exchange integrals.
9in x 6in
Integral Value Integral Value
2sC 2sF |2sC 1sHa 0.4970 48510 101 2sF 1sHa |1sF 2sC 0.1014 05594 102
2sC 2sF |2sC 1sHa 0.8420 56635 102 2sF 1sHa |2sF 2sC 0.9341 35949 102
2sC 1sF |1sC 1sHa 0.5737 90540 103 2sF 1sHa |2pzF 2sC 0.8442 95091 102
2sC 1sF |2sC 1sHa 0.3789 18525 102 2sF 1sHa |1sF 2pzC 0.1813 23479 102
b1189-ch07
1sC 2pzF |2pzC 1sHa 0.1587 58344 102 2sF 1sHa |2sF 2pzC 0.1379 64387 101
P.E. Hoggan
2sC 2pzF |2pzC 1sHa 0.5258 34208 102 2sF 1sHa |2pzF 2pzC 0.1135 01125 101
2pzC 1sF |1sC 1sHa 0.1025 32536 102 1sHa 2sF |1sHa 2sC 0.1252 319411 101
2pzC 1sF |2sC 1sHa 0.6772 76818 102 1sHa 2sF |1sHa 2pzC 0.1591 49899 102
1sC 1sF |1sC 1sHa 0.1099 00118 106 1sHa 2pzF |1sHa 2pzC 0.1772 90873 102
1sC 1sF |2sC 1sHa 0.6794 54131 106 1sF 1sHb |2sF 1sC 0.2287 77210 104
1sC 2sF |1sC 1sHa 0.1446 31297 102 1sHb 2sF |1sHb 1sC 0.1939 63837 102
Solving the Schrodinger Equation

1sC 2sF |2sC 1sHa 0.4235 59085 102 2sC 1sHa |1sC 1sHb 0.2034 841982 101
2pzC 2sF |1sC 1sHa 0.1112 10955 101 1sC 1sHa |1sC 1sHb 0.7154 932331 102
2pzC 2sF |1sC 1sHa 0.6738 14908 101 2sC 1sHa |2sC 1sHb 0.1137 390852
resolution can be used to give fast and accurate results for basis sets of s
and p Slater-type orbitals. Generalization is in progress.
Numerical vales for the H2 dimer geometry and interaction energy agree
well with complete ab initio potential energy surfaces obtained using very
large Gaussian basis sets and data from vibrational spectroscopy [77].
7.8.7. Angular momentum relations

Analytical work by Dunlap [78] and numerical treatment by Talman [79]
give simplifications of angular momentum coupling that are indicating
methods of limiting the current stumbling block of high-angular momentum
functions for which the Coulomb resolution loses accuracy and where
orbital translation converges very poorly. In fact, there is a relation put
forward by Dunlap that alleviates the tedious and inaccurate evaluation
of derivatives required by the BCLF orbital translation coefficients (see
Appendix A; Section 7.12).
7.9. When are Slater-Type Orbitals Advantageous?

Some Applications
7.9.1. The NMR nuclear shielding tensor

More complete work is referred to here and the present description is a
brief summary [121, 122, 127]. In NMR, the nuclear shielding tensor is a
second order perturbation energy correction, for derivatives with respect to
the nuclear dipole moment and the external field.
The perturbed Fock matrix element when including the effect of the
external field contains both one and two electron terms. In this example,
we focus on the one electron terms.
The purpose of the present section is to give a case study of one of the
contributing energy integrals involving the dipole 1/rN 3 operator.
In the applied magnetic field, the question of gauge invariance must be

resolved. A method of circumventing the problem was devised by Ditch-
field using the London GIAO [128]. These gauge including atomic orbitals
reduce to STO for zero field and contain the required phase factor oth-
erwise [121, 122, 137].
The integrals were evaluated for GTO at zero field and nuclear shielding
tensor or chemical shifts have been available since Gaussian 72 based on
this pioneering work [23] and distributed to academics by QCPE. It is
nevertheless important that users input the appropriate structure in order
172 P.E. Hoggan
to obtain accurate chemical shifts corresponding to the species studied and

note that for work in solution (or in solids) some structural changes may
occur.
Define the nuclear shielding tensor as a second order energy pertur-
bation:

2 0 H(
N, B
0 ) 0
=
N
, (7.28)
N, B0,

N =0, B0 =0

with
N the nuclear dipole moment of nucleus N and B
0 the external field.
|0 is a closed shell ground state Slater determinant. and stand for
cartesian coordinates.
A coupled HartreeFock treatment of the above equation leads to [121,
122, 136]:

(0,1) (0) (1,1)
N
= Tr P h(1,0)
+ P h , (7.29)
(0,1)
where P (0) and P are the density matrix of zero order and first order
(1,0)
with respect to the external magnetic field, h is the core Hamiltonian
(1,1)
of the first order with respect to nuclear dipole moment and h is the
second order one-electron Hamiltonian with respect to the nuclear moment
and the external field B .
The non-zero orders in (7.29) involve integrals which are absent from
ab initio HartreeFock calculations. In this work, we focus our attention
on integrals involving 1/rN3 in their operator. These integrals appearing in
the second order expression for the approximate perturbed Hamiltonian:

2
r
r r r
(1,1) 0 e N , N,
h, =
4 2me 3
rN

r ) (
rN )
(R

+ . (7.30)
rN3
The integral which we have chosen to investigate in detail within the Fourier
transform approach, is the three-center one-electron integral:

r
r r r
N , N,
I = . (7.31)
3
rN
Here
rN is the instantaneous position of the electron with respect to the
nuclei N.
Analytical treatment
This is to be found in Appendix E in Section 7.16.
Application
The 15 N chemical shifts measured for a set of benzothiazoles are evaluated

with the above expressions. These molecules possess a ring nitrogen and
have been studied previously in our group [137]. The measurements were
made in natural abundance. The intensity of signals due to the nitrogen
must be amplified by a two-dimensional NMR technique involving cross-
polarization to benefit from the intensity of proton resonances coupled to
that of the 15 N in the molecule.
The in vivo NMR benefited from measurements by B. Combourieu
in [137]: these molecules are metabolized by bacteria and researchers in
the group try to follow the pathway by NMR. Since such studies are very
difficult to do, we tried calculating some chemical shifts accurately from
structures to assign them (see acknowledgements).
TheY substituent, generally a hydroxide, was found to be in the position
indicated (for mechanistic reasons, it is the only accessible and stabilized
position for ring hydroxylation which has been found to take place in vivo
after experiments in our group).
In solution, these molecules undergo a tautomeric equilibrium reaction
transferring a proton towards this nitrogen as shown in the figure below
(also used for nomenclature).
Summary of NDDO-PM3 fitted STO molecular-site calculations
on unprotonated tautomers (b). It is possible to conclude when the
Gaussian03/PBE 6-311++G(2d,p) calculation (c) differs substantially
from the measured value (a) (ppm/CH3 NO2 ) that the resonating nitrogen is
mostly protonated. This serves as a guideline for ab initio structures studied
for these equilibria.
H
N N
XH X
Y S S
Y
XBT PXBT
Fig. 7.4. Tautomerisation of benzothiazoles.

174 P.E. Hoggan
Table 7.4a. Chemical shift calculations for BT tautomers.
Molecule Substituent a b c
BT:benzothiazole: No X 72.5 71.8 61.4

OBT: X=O 238.8 238.9 133.3
OHOBT: X=O; Y=OH 240.4 239.9 135.3
ABT: X=NH 153.1 152.1 131.6
OHABT: X=NH; Y=OH 153.1 153.6 132.3
MBT: X=S 199.6 199.9 79.9
OHMBT: X=S; Y=OH 205.5 205.5 83.2
MTB: X=N(CH3 )CONHCH3 124.0 125.4 141.0
Table 7.4b. DFT Calculations. Differences between calculated and observed

15 N chemical shifts for commercial benzothiazoles and some metabolites (in
ppm). a-Measured values with respect to nitromethane standard in deuterated

methanol solvent (B. Combourieu in [137]) error bars of 2 ppm.
b-Coupled perturbed STO.
c-Gaussian [123] augmented with hydrogen-like AOs.
d-Gaussian [123].
Note. b through d involve solvation models, detailed below.
Molecule Substituent ac ab ad
BT:benzothiazole: No X 1.3 8.3 11.1

POBT: X=O 4.6 11.7 20.0
POHOBT: X=O; Y=OH 4.5 7.4 14.9
ABT: X=NH 1.1 3.8 21.5
OHABT: X=NH; Y=OH 4.5 10.1 20.8
PMBT: X=S 3.0 11.2 21.2
POHMBT: X=S; Y=OH 2.5 10.1 18.8
The above results prompted use of the protonated structure to obtain the
zero-order wave-function, in all cases apart from benzothiazole (BT) and
ABT. Below, the same cases are treated in the DFT work.
Note that the basis sets including hydrogen-like orbitals perform better
than the STO basis sets that in turn improve upon dense-core Gaussian basis
sets [6-311++G(2d,p)].
Basis sets augmented with hydrogen-like orbitals are within 5 ppm
of the experimental values (measured within 2 ppm) for the discrete
solvated model. This model explicitly includes several deuterated methanol

molecules to cater for the specific hydrogen bonding interactions.
a. Measured chemical shift for ring nitrogen.

b. STO: DFT PBE 6-311++G(2d,p) calculations with two discrete CD3 OD
molecules on OH, NH, and SH (one on N, O, S) for minimal total energy.
c. Gaussian 2003 as (b) with hydrogen-like orbital DFT PBE aug-
6.311G**(2d, p) calculations.
d. Gaussian 2003 DFT PBE 6-311++G(2d,p) calculations.
The content of this table is original and based on the previous work
of the author [125] i.e. geometries are re-optimized from the coordinates
of [125].
Conclusion
Another step on the way to ab initio ETO basis nuclear screening tensor
calculations has been accomplished.
It is essential to use a basis set which comprises orbitals with the correct
nuclear cusp behavior. This implies a non-zero value of the function at the
origin for spherically symmetric cases and satisfying Katos conditions.
Hydrogen-like atomic orbital basis sets therefore perform better than Slater-
type orbitals which are an improvement upon even large Gaussian basis
sets.
The NDDO-PM3 molecular site approach has the advantage of rapidity.
Calculations take about a minute instead of 5075 hours on the IBM-44P-
270. They cannot be systematically improved, however, once the site Slater
exponents have been fitted. Note that the 2s Slater exponent fluctuates wildly
in fits, providing further evidence that shielding must be of the form (2-r)
for the 2s ETO.
Fundamental work on orbital translation is also in progress to speed up
these calculations within the test-bed of the STOP programs [50, 80, 131].
The interplay of these discrete molecule solvent models and accurate
in vivo NMR measurements is satisfactory, in that the structures postu-
lated give calculated chemical shifts to similar accuracy as obtained for
experimental values (on the order 2 ppm). It should be stressed that energy
minimization in this case does evidence directional hydrogen bonds but can
lead to several possible solvent geometries. Further study, using molecular
dynamics techniques would be useful in the modeling of solvent shells and
176 P.E. Hoggan
is planned in the future. In view of the complex systems studied, this is

highly satisfactory.
7.9.2. Explicitly correlated methods for molecules

The application and development of such methods to determine accurate
ground and excited states, and properties of diatomic and triatomic
molecules is very promising and more interesting for the computational
chemist than the atomic case. There is nowadays a growing interest in this
field. Subroutines and programs which perform these calculations are often
requested in the community.
The investigation of these integrals is done within the molecular orbital
method (MO) [81], since the MO wave function is the simplest wave
function for a molecular system.As Coulson [82] discussed, the MO method
permits the visualization of electrons and nuclei and interpretation of indi-
vidual electrons and their orbital exponents better than the wave functions
written in elliptical coordinates. The wave functions constructed with ellip-
tical orbitals are of two types, the so-called JamesCoodlige [83] wave
functions (one-alpha exponent), recently extended to the two-exponent
case [35], and KolosWolniewicz [84,85] wave functions (with both orbital
exponents alpha, and beta). Both have been applied to H2 .
The elliptical wave functions are the natural representation of a two-
center problem but for three-center and larger molecules the use of the
MO method becomes necessary. Frost [86] used the MO method and the
correlated molecular method (CMO) in H2 calculations. Concerning the
extension of the method he wrote:
The extension of CMO-type wave functions to more complex

molecules does not seem feasible at the present time. The new
integrals which will be introduced would involve more than two
centers if more nuclei were involved and higher atomic orbitals
than 1s if more electrons were considered, and their evaluation
would be extremely difficult.
Recently, impressive calculations using Hylleraas wave functions have

been done for H2 , see Table 7.5. Hylleraas [35], the iterative complement
iteration method (ICI) [87], and explicitly correlated Gaussian (ECG) [88]
calculations of the hydrogen molecule, Hylleraas calculations on HeH+
and some other species [89] leading to 2.9710784698 a.u. using 9576
configurations and calculations of He2 using 4800 optimized ECG con-
figurations with energy 5.80748359014 a.u. [88] achieved the highest
Table 7.5. Highly accurate calculations on the H2 molecule with different

types of wave function at R = 1.4011 a.u.
Authors type w. f. Confs. Energy (a.u.)
1933 James and Coolidge JC 5 1.1735

1960 Kolos and Roothaan KR 1.17214
1968 Kolos and Wolniewicz KW 1.174475
1995 Wolniewicz KW 833 1.17447467
2006 Sims and Hagstrom JC 7034 1.17447593139984
2007 Nakatsuji ICI 6776 1.17447571400027
2008 Cencek and Szalewicz ECG,opt 4800 1.17447571400135
known accuracy in molecules (picoHartree accuracy is better than that of

chemical measurements, e.g. a micro cm1 , a nano eV or micro cal/mol.
Although one must recall that in the calculation of properties according to
Drake [90], only half of the digits of the energy are kept).
Note also that input exponents, distances and some fundamental con-
stants may limit accuracy of calculations compared with measurements
and that molecules may not be rigid. Eventually, dynamics and the Born
Oppenheimer approximation must be accounted for.
Hylleraas-CI (Hy-CI) was applied in 1976 to the LiH molecule by Clary
[34] using elliptical STOs. For two-center molecules the three-electron
and four-electron integrals occurring in the Hy-CI have been developed by
Budzinski [41].
Another type of explicitly correlated wave function is that using
Gaussian orbitals. Clementi et al. extended the Hy-CI to molecules using
Gaussian orbitals [91], and applied it to the calculation of H3 . The ECG
wave function is appropriate also for molecules [88, 92], as the inter-
electronic distance r12 is a Gaussian exponent. This leads to results which
are comparable with Hylleraas calculations [88].
The R12 -wave function proposed by Kutzelnigg and Klopper [93, 94]
has the merit of satisfying the (electron-electron) cusp condition, to
use Gaussian functions avoiding the three- and four-center integration
problems, and to include precisely r12 , involving electrons one and two,
close to the nucleus, with higher probability that r12 = 0, also these elec-
trons are present in any system starting from the helium atom. The r12 vari-
ation influences energy. The R12 wave function, developed for molecular
calculations is nowadays widely used and combined with all kinds of
methods. The three- and four-electron integrals occurring are calculated
178 P.E. Hoggan
in terms of two-electron integrals. Due to the use of a single r12 value, the
accuracy achieved for atomic calculations is less than that of Hy and Hy-CI
calculations. Recent improvements of the method [9597] can achieve
micro-Hartree accurate energy results for chemically interesting systems.
Short wave function expansions lead to very good results. When a large
number of configurations are used (up to 10,000) the energy results are
beyond pico-hartree accuracy, while the CI wave function would need in
the order of millions of configurations.
7.9.3. Trial wave-functions for quantum Monte Carlo simulations

over STO
Here, exponential type orbital basis sets, in particular Slater-type orbitals
(STO) or CoulombSturmians are considered. These provide trial wave-
functions with exact atomic nodes.
Results are given here for a singlet-triplet transition in the acrolein
molecule. This work has been extended to a full paper, currenlty in press
[146].
Experimentally, a UV transition is studied from the non bonding n orbital
to the anti-bonding orbital of the CO moeity.
This work uses VB Slater-type orbital basis sets optimized by Ema
et al. [26] and is compared with a recent study by Caffarel et al. [143]
which used GTOs with a (cubic) polynomial to mimic the nuclear cusp.
The QMC calculations are carried out using CHAMP adapted by
J. Toulouse to run on at least 8,192 processors on a Bluegene, with linear
scaling of calculation time (see, for example [126]).
7.9.3.1. Basis sets

The VB1 STO basis set due to Ema et al. comprises analytical Slater-type
orbitals denoted (5s,3p,1d) for C and O and also (3s,1p) orbitals for H.
In terms of function angular momentum, it is between two of the basis
sets used by Caffarel et al. [143].
They are large (cusp corrected) GTO basis i.e. those denoted basis two
(with (3s,2p,1d) on O and C) and basis three (BS3) (with (4s,3p,2d,1f) on
O and C), respectively.
7.9.3.2. Computational procedure

The procedure starts by generating an initial wave function using
the quantum chemistry program GAMESS or MOLPRO, typically a
restricted HartreeFock (HF) wave function or a multi-configurational self-

consistent-field wave function. We use Slater-type basis sets, each Slater
function being approximated by a fit to Gaussian functions in GAMESS. In
the preliminary restricted HartreeFock step, a STO-10G expansion was
used to expand the Ema basis set in order to obtain variational wave-
functions using the GAMESS software.
Subsequently, the analytical STO functions are actually used in the
quantum Monte Carlo code CHAMP (by Umrigar, Filippi et al., both for
variational Monte Carlo (VMC) Jastrow factor optimisation and diffusion
(DMC) work.
This trial wave-function, using a Slater-type orbital basis, is multiplied
by a Jastrow factor, imposing the electron-electron cusp condition, but
with essentially all other free parameters chosen to be initially zero to
form our starting trial JastrowSlater wave function, and QMC calcula-
tions are performed with the program CHAMP on the BlueGene using
the true analytical Slater basis set rather than its Gaussian expansion.
The Jastrow, configuration state function (CSF) and orbital parameters are
simultaneously optimized with the linear energy minimization method in
variational Monte Carlo (VMC). Once the trial wave function has been
optimized, we perform a diffusion Monte Carlo (DMC) calculation within
the short-time and fixed-node approximations. In this algorithm, the many-
electron Schrodinger equation is propagated in imaginary time, and the
propagation continues until a desired statistical uncertainty on the energy,
decreasing with the square root of the number of Monte Carlo points,
is reached. As the propagation is done for a very large number of tra-
jectories (walkers), the resulting algorithm is ideal for massively-parallel
architectures.
All QMC calculations were carried out on the IBM BlueGene machines
in Munich (RZG) and Julich (FZJ), employing at least 512 processors (2048
cores) per run.
Results from GAMESS (i.e. variational wave-functions) at HF and CAS
levels, clearly show that at this (STO-10G) level of approximation, the
Caffarel basis three gives slightly better results than the Ema valence Bond
like VB1 basis (abbreviated vb):
E(HF VB1) = 190.8264 . . . E(HF BS3)

= 190.8355 . . .
E(CAS (6, 5) VB1) = 190.8901 . . . E(CAS(6, 5) BS3)
= 190.9008 . . .
180 P.E. Hoggan
Caffarels basis three gives about 0.01 a.u. lower total energy, at this
preliminary stage.
Examining the total DMC energies with the corresponding nodes (within
the fixed-node approximation for the ground state), the results below were
subsequently obtained:
E-DMC(HF VB1) = 191.8481(5) E-DMC(HF BS3) = 191.729(4)
E(CAS(6,5) vb) = 190.8479(5) E-DMC(CAS(6,5) 3) = 190.744(5)
The improvement in results is remarkable! (They are about 0.12 u.a.
lower in energy for the VB1 basis).
These results do not depend on whether the Jastrow factors are well
optimized or not. They are simply dictated by the initial nodal structure of
the wave-function that remains unchanged.
It is therefore possible to compare the basis sets directly.
These results show that the Slater-type orbital basis set is much more
appropriate for a trial wave-function than that on a Gaussian basis. It is pos-
sible to infer categorically that the nodal structure is substantially improved
towards the exact result when an STO basis set is used, compared to a cusp-
corrected Gaussian basis set.
The following results show the singlet-triplet gap calculation is far less
basis set sensitive in the case of Slater-type orbital basis, regardless of
Jastrow factor optimisation. The basis appears to saturate much sooner
and this may allow for limited optimization thus avoiding (at least in part)
the very time-consuming step in quantum Monte Carlo. Fixed-node and
variational accuracy are both improved over the cusp-corrected Gaussians.
The Slater-type orbital basis State Specific results are:
E-DMC (CAS(2,2)vb) = 191.8494(5) 191.7107(5) gap = 3.77(2)
E-DMC (CAS(6,5)vb) = 191.8479(5) 191.7154(5) gap = 3.61(2)
Whereas the previous (cusp-corrected Gaussian) State Specific results
are:
E-DMC (CAS(2,2) BS3) = 191.729(4) 191.608(2) gap = 3.29(12)
E-DMC (CAS(6,5) BS3) = 191.744(5) 191.596(2) gap = 4.03(14)
The gap energies fluctuate much less with the configurations taken into
account over STO compared to cusp-corrected GTO because the nodal
structure is closer to that of the exact solution.
When the basis is increased progressively from VB1 to VB3 through
VB2 there is a systematic improvement, both variationally and in the
accuracy of the nodal structure. This mirrors the improvement already

observed by increasing the Gaussian basis size similarly.
State averaged calculations using the Slater-type orbital basis sets further
improve the results.
7.9.3.3. Perspectives
Time-consuming optimisation of the Jastrow factor which introduced cor-
relation through explicit r12 dependence may be reduced by using a trial
wave-function over Slater-type orbitals that enable the HartreeFock limit
to be approached for a smaller basis (i.e. the basis saturates sooner) than
for Gaussians.
Recently, Gill proved a simple model two-particle closed-form
expression for correlation in hyper-spherical coordinates [142]. Co-
ordinates have also been given by Fano, for three-particles. A simple
expression would lead to considerable time-gains for Jastrow factor opti-
misation, since these terms dominate the calculation time to a large extent
when they are necessary.
7.10. Highly Accurate Calculations Using STOs
Another problem appearing in these calculations is digital erosion. For

many operations (in particular subtraction) involving numbers of similar
value some digits can be lost leading to erroneous results. Quadruple pre-
cision avoids this, about 30 decimal digits are correct on most computers.
Another possibility is high precision arithmetic software. Some programs
are available like Baileys MPFUN [98], the Brent and Miller program
packages [99, 100].
One example of the use of Slater orbitals, at the present five, are the
highly accurate calculations of small molecules using explicitly correlated
wave functions i.e. wave functions where the inter-electronic coordinate
rij is included explicitly in the wave function. These are the Hylleraas
and Hylleraas-CI wave functions, ICI method, compared with the explicit
correlated Gaussians ECG and the R12 method.
7.11. Closing Remarks
Early molecular electronic structure work, from the first diatomic molecules
tackled before 1930 to the advent of mainframe computers in the mid-1950s,
182 P.E. Hoggan
all used Slater-type orbitals (as a linear combination of the hydrogen-like

orbitals with direct physical interpretation as atomic orbitals). Boys then
suggested expansion of STOs in n Gaussians, reducing three- and four-
center integrals to single center overlap-like integrals. This was ideal for
numerical evaluation on the computers available. For convenient use of
hardware and software, this approach continued to be almost the only one
used for larger and larger molecules for many years. Some Slater work
continued through the mid-1960s/1970s but codes were rather slow, with
the exception of the numerical grid approach to DFT over Slaters developed
by Baerends et al. as ADF.
During the 1990s, some code taking advantage of new hardware and
the advent of symbolic algebra systems such as Maple to control accuracy
and check formulae gave rise to new Slater codes (STOP, SMILES).
It rapidly became obvious that CI requires fewer functions in Slater
basis sets than for GTOs. Furthermore, it seemed more constructive to
identify domains of application where users should choose Slater basis
sets (leaving others where users can exploit the well known software on
Gaussian basis sets). This work illustrates the advantages of Slater-type
orbitals for such work as electron correlation (CI or, even better, explicitly
correlated wave-functions), NMR chemical shifts and quantum Monte
Carlo work.
Regarding Slater-type orbital basis calculations, whilst the QMC work
is well-suited to parallel architectures, this is not at present the case of RHF
calculations. Tests on a set of organic molecules have been carried out on
64 processors by Ishida [145] and the present author independently and
show scaling is sub-linear (speed up of 2040). The problem stems from
difficulty in anticipating the calculation times for batches of integrals which
leads to idle time for processors.
What is required to continue towards well-written software and supply
an efficient and easily used program product is an injection of pace from
enthusiastic young workers who know modern programming languages
well (C etc.) and have at least some background in the molecular quantum
theory that required to use Slater-type orbital basis sets. There is also room
for more theoretical/mathematical development. In the mid- to long-term,
it will be constructive to provide both types of basis option in a given
software.
We conclude with the words of the late Gaston Berthier (19232009):
Gaussians are like medicine, you must use them as long as they cure you
but once they dont work any more, you much change them! (Gaston
Berthier, Interview, Paris, 2 June 1997).
7.12. Appendix A: How STOs were Translated:

Products on Two Atoms
The BarnettCoulsonLowdin functions (BCLFs see [103]) arise as coeffi-

cients in the series expansion of a Slater-type orbital centered at a distance
a from the origin, placed on an atomic nucleus where a set of Slater-type
orbitals are centered [111, 114]. This allows the one- and two-electron
multi-center integrals to be evaluated at a given origin in the molecule.
The series expansion obtained is infinite, since the molecular geometry
variable a (usually 120 a.u.) is fixed for an electronic structure calcu-
lation, whereas the instantaneous electron position variable r is independent
of it and 0 < r < . They are both radial vectors and generally cannot be
aligned.
Much work is already available on BCLFs [10, 24, 105, 107110, 112,
115] and references therein. Nevertheless, two bottlenecks are yet to be
efficiently resolved when Slater-type orbital translations are required:
(i) rapid and accurate generation of the BCLFs themselves, and

(ii) acceleration of the convergence of the infinite series generated, which
typically do not converge quickly.
In the present work, the first item, (i) is thoroughly addressed.

The value of the screening parameter generally exceeds one and should
not exceed the atomic number. In practice, the lower limit for is related to
the
first ionization potential I in atomic units, i.e., must not be less than
2 I. These limitations are helpful in establishing the numerical behavior
of the BCLFs.
In this work, we treat the problem of efficient computation of BCLFs.
Our aim is to develop a computational procedure by which a whole sequence
of BCLFs can be computed fast and accurately. In Section 7.12.1, we present
an up-to-date review of properties of BCLFs. In more complete work on
the subject [116], we discuss and evaluate possible strategies for computing
them, and conclude that recursion relations can be used efficiently for this
purpose, provided that the modified spherical Bessel functions In+1/2 (x)
and Kn+1/2 (x) can be computed fast and accurately.
A method by which a whole array of BCLFs can be computed simultane-
ously, quickly and accurately is also detailed in previous work [116]. In this
recent work, we also discuss the details of the programming of our method.
It is important to note that, in our method, we do not compute In+1/2 (x) and
Kn+1/2 (x) directly. Taking into account the asymptotic behaviour of I (x)
and K (x) as , we compute some appropriately scaled versions of
184 P.E. Hoggan
these functions instead. The scaling we use enables us to avoid the under-
flows and overflows that may occur in direct computation of In+1/2 (x) and
Kn+1/2 (x) for large values of n; it is thus an important ingredient of our
method. This also allows us to scale the BCLFs appropriately. In order to
end up with BCLFs that have double-precision accuracy, in our method,
we compute both the functions In+1/2 (x) and Kn+1/2 (x) and the BCLFs in
extended precision arithmetic, the idea being that the quadruple-precision
arithmetic is shown to suffice and it is offered with some high-level pro-
gramming language compilers used for scientific applications, e.g. Fortran
77 and C. As the number of arithmetic operations required is very small
(of the order of wN, where N is the number of BCLFs computed and w is
a small integer), the use of quadruple-precision arithmetic cannot increase
the cost of the computation time-wise. We provide an error analysis for the
procedure we use to compute the scaled modified spherical Bessel func-
tions, which shows that the procedure is indeed very accurate in previous
work [116].
Finally, in [116], we also provide three appendices that contain several
results that seem to be new. In the first, we analyze the asymptotic behavior
of the modified Bessel functions I (x) and K (x) as . We derive
two sets of full asymptotic expansions that have some quite interesting
properties. The scalings we use in [116] are based on the results of this
appendix. In the second, we obtain explicit power series expansions for
products of modified spherical Bessel functions. In the third appendix
of [116], we derive asymptotic expansions of BCLFs as their order tends to
infinity.
7.12.1. Review of BCLFs

7.12.1.1. Definition and properties of BCLFs
Let n be a non-negative integer, a and r two real positive numbers, a real
positive number. Of these, a and are finite, while r assumes values from
0 to infinity. With R defined as in

R = a2 + r 2 2ar cos (7.32)
consider the function Rn1 eR . Letting x = cos so that x [1, +1],
its expansion in Legendre polynomials P (x) may be expressed as

1
Rn1 eR = (2 + 1)An+1/2 (, a, r) P (x), 1 x 1,
ar
=0
(7.33)
An+1/2 being the BCLFs. From this relation, it is seen that Rn1 eR serves
as a generating function for the the BCLFs. Since
+1
2
P2 (x) dx = , = 0, 1, . . . , (7.34)
1 2 +1
we immediately deduce from (7.33) that
+1
ar
A+1/2 (, a, r) =
n
Rn1 eR P (x) dx, = 0, 1, . . . .
2 1
(7.35)
Clearly, the An+1/2 (, a, r) are symmetric functions of a and r, that is,
An+1/2 (, a, r) = An+1/2 (, r, a), (7.36)
because the function Rn1 eR is.
A simple expression for BCLFs with n = 0 and = 0, 1, . . . , is known
(see Abramowitz and Stegun [102], p. 445, formula 10.2.35):
A0+1/2 (, a, r) = I+1/2 ()K+1/2 ( ); = min{a, r},
= max{a, r}. (7.37)
Here, I+1/2 (x) and K+1/2 (x) are the modified spherical Bessel functions.
The functions I+1/2 (x) and K+1/2 (x) satisfy three-term recursion rela-
tions in that are given in this work, and are defined for all integer values
of . Those I+1/2 (x) with 0 are called modified spherical Bessel
functions of the first kind, while those with < 0 are called modified
spherical Bessel functions of the second kind. The K+1/2 (x) are called
modified spherical Bessel functions of the third kind. Each of the two pairs
[I+1/2 (x) and I1/2 (x)] and [I+1/2 (x) and K+1/2 (x)] is a linearly
independent set of solutions of the modified spherical Bessel equation of
order (see Abramowitz and Stegun [102], Chapter 10) of the first and
third kind, respectively. Because I+1/2 (x) and K+1/2 (x) are defined for
all integer values of , we let (7.37) define A0+1/2 (, a, r) for < 0 as
well. This is an important step that enables us to define An+1/2 (, a, r) for
< 0 as well, which is what we consider next (see [106]).
From the integral representation in (7.35), it follows that, for n 0,
n
+1/2 (, a, r) =
An+1 A (, a, r), (7.38)
+1/2
and hence
n
An+1/2 (, a, r) = (1)n n A0+1/2 (, a, r). (7.39)

186 P.E. Hoggan
From (7.37), it is obvious that A0+1/2 (, a, r) = A0+1/2 (1, a, r). By a

simple manipulation of the integral representation in (7.35), it can be shown
analogously that An+1/2 (, a, r) satisfy the homogeneity relation
An+1/2 (, a, r) = n An+1/2 (1, a, r), n 0. (7.40)
This relation shows that An+1/2 (, a, r) are actually functions of two vari-
ables, namely, of a and r, and can be computed directly from the functions
An (a, r) that are defined as in
An (a, r) = An+1/2 (1, a, r). (7.41)
From (7.40) and (7.41), it follows that An+1/2 (, a, r) can be computed
from An (a, r) via
An+1/2 (, a, r) = n An (a, r). (7.42)
Invoking (7.42), it is easy to show that (7.38) can be rewritten as

A (a, r) = nA (a, r) a + r
n+1 n
An (a, r), n 0. (7.43)
a r
7.13. Appendix B: Brief Time-Line of Events in Molecular Work

Over Slater-Type Orbitals to Date
1928 Slater and London.

1929 Hylleraas: He atom.
1931 Podolanski: one- and two-center repulsion integrals.
1933 James and Coodlidge: Hylleraas calculations on H2 .
1949 Roothaan LCAO paper.
1950 Boys: first Gaussian expansion of STO published.
1951 Two-center Coulomb Integrals. Roothaan formulae.
1954 Boys and Shavitt Automated calculations.
1958 Tauber: Work on analytic two-center exchange integrals: Poisson
equation.
1962 Scrocco: first publishes STO work (in Italian), but with a program.
This follows early molecular work in 195153 [3, 4].
1963 Clementi produces tables of optimized single zeta basis sets for atoms.
Shavitt B-Functions described.
1970 The Journal of Chemical Physics published work on STO codes by
E. Scrocco and R. Stevens. GAUSSIAN 70 prepared for QCPE by J. Pople
and R. Ditchfield.
1973 E. J. Baerends: numerical integration over STO used for ADF DFT
code.
1978 Filter and Steinborn: Fourier transform work. B-functions and plane-
wave expansion of Coulomb operator.
1981 ETO Conference in Tallahassee. Weatherford and Jones.
1994 First (published) STOP (Slater-type orbital package, QCPE 667 1996)
code. Bouferguene and Hoggan.
2001 First SMILES (Slater molecular integrals for large electronic systems)
code. Fernandez Rico, Lopez et al.
2002 Complete analytic study of two-center repulsion integrals using
Maple symbolic algebra. F. Harris
2009 Gill: Coulomb resolution. Model exact pair correlated wave-
functions.
7.14. Appendix C: Main Results of Podolanskis Paper of 1931

with Additional Comments
See Annalen der Physik, 402, 7 (1931), pp. 868872 (in German):
In molecular quantum chemistry, major contributions to the electronic
interaction energy for diatomic systems are exchange integrals of the form:
(1) (2) (1) (2)
a (1)b (1)a (2)b (2)
Iex = d1 d2 ,
r12
where (1) and (2) are (hydrogen-like) atomic orbitals centered at a and
b. This integral was solved in 1927 for ground-state H2 by Sugiura.9
In general, these two-center exchange integrals may be expressed as
polynomials multiplied in turn by exponentials, logarithms, and the expo-
nential integral functions. The sum is finite for equal orbital exponents, ,
or for orbitals with the same principal quantum number.
Some standard integrals are required.10
9Y. Sugiura, Zeits f. Phys. 45 (1927), p. 484.

10 The integral for the most general functions involved can be evaluated analytically by adapting
Sugiuras work. The final formulae use quadratures of the form:
x
ex r(t)et dt,

where r is a rational function. Integration may be easily carried out, using the formula
x r (i) (x)
ex r(t)et dt = (r (i) = i-th derivative of r).
i+1
i=0
Or the Euler integral, when r = t n [102].

188 P.E. Hoggan
Introducing elliptic coordinates , , : the integral takes the general

form (1)
D6 w (1 , 1 )w(2) (2 , 2 ) (1) 1 (2) 2 im(1 2 )
Iex = e e e
64 r12
d1 d2 d1 d2 d1 d2 ,
where w is the orbital product, simplified by the exponential and volume
element change.
Expanding w(1) and w(2) in spherical functions Pim () of order m, gives
an infinite sum, except in the cases cited above:
()
w() = vi ()Pim () ( = 1, 2).
i
The products
gi() (x) = (1)m/2 v() m
i (x)Pi (x),
() ()
hi (x) = (1)m/2 vi (x)Rm i (x)
m
are all rational functions. The Ri appear in the splitting of spherical func-
tions of second kind in their expression as (cf. below)
x+1
Qm i (x) = Pi (x) log + Rm
m
i (x).
x1
The requisite polynomials are, using gi and hi
x
() x
gi (t)e t dt,
() () ()
Gi (x) = e

x
() x
hi (t)e
() () () t
Hi (x) = e dt

and finally from them and gi the functions
2 (1) (2)
a0 (x) = G (x)Gi (x),
2i + 1 i
i
2 (1) (2)
a1 (x) = G (x)Gi (1),
2i + 1 i
i
2 (1) (2)
a2 (x) = Gi (1)Gi (x),
2i + 1
i
2 (1)
(2) (2) (1)
b(x) = gi (x)Hi (x) + gi (x)Hi (x) ,
2i + 1
i
ak (x) ak (1) ak (x) ak (1)

Ak (x) = , k = 0, 1, 2.
x1 x+1
The exchange integral is therefore given by
2 D5 ((1) +(2) )
Iex = e J.
8

2(1) (2) (1)
J = a0 (1) + log (1) a1 (1)e2 Ei (2(1) )
+ (2)
a2 (1)e2 Ei (2(2) ) + a0 (1)e2( + ) Ei (2((1) + (2) ))
(2) (1) (2)
1 1
(1) x
(1) (2)
A2 (x)e x dx
(2)
+e A1 (x)e dx + +e

1
+ e + [b(x) A0 (x)]e( + )x dx,
(1) (2) (1) (2)
where = 0.57722 . . . is the Euler constant and

1 eu
Ei (x) = du
u
is the exponential integral function.
7.15. Appendix D: Potentials and Auxiliary Overlaps

for Coulomb Resolution
Define potential functions [70]:
i = 23/2 n l (r)Ylm (, ).
Omitting the spherical harmonic term gives radial factors:

+
n l (r) = hn (x)jl (rx)dx with jl (x) denoting the
0
spherical Bessel function (7.44)
Here, hn (x) is the nth member of any set of functions that are complete
and orthonormal on the interval [0, +), such as the nth order polynomial
function (i.e. polynomial factor of an exponential). The choice made in [69]
is to use parabolic cylinder functions (see also another application [54]),
i.e. functions with the even order Hermite polynomials as a factor. This is
not the only possibility and a more natural and convenient choice is based
190 P.E. Hoggan
on the Laguerre polynomials Ln (x). Define:

hn (x) = 2Ln (2x)ex . (7.45)
These polynomial functions are easy to use and lead to the following ana-
lytical expressions for the first two terms in the potential defined in (22):
tan1 (r)
V00 (r) = 2 (7.46)
r

tan1 (r) 2
V10 (r) = 2 . (7.47)
r (1 + r 2 )
Furthermore, higher n expressions of Vn0 (r) all resemble (23) (see [70]
Eq. (23)):

1 n 1
1 k sin(2ktan (r))
Vn0 (r) = 2 tan (r) + (1) (7.48)
r k
1
and analytical expressions of Vnl (r) with non-zero l are also readily obtained
by recurrence.
These radial potentials can generally be expressed in terms of hyperge-
ometric functions, whether the choice of polynomial is the present one, i.e.
Laguerre, or Hermite polynomials, as in [69]. This structure has been used to
confirm the results of [70] using a rapid code in C [71]. Spherical harmonics
are translated using Talmans approach [72]. The displaced potential in one
factor of the product of auxiliaries, from four-center integrals is readily
expanded in two-center overlaps, after applying Eulers hypergeometric
transformation [73, 74].
The auxiliary overlap integrals (r1 )i (r1 ) and i (r2 )(r2 ) will
involve densities obtained from atomic orbitals centered on two different
atoms in exchange multi-center two-electron integrals. The overlap inte-
grals required in an ETO basis are thus of the type:
a (r1 )b (r1 )i (r1 )

max
= N (n1 , n2 , ni , li , |mi |)s(n1 , l1 , m, n2 , l2 , ) (7.49)
=0
with = 1 R and = 2 R, Slater exponents. In three-center overlaps, N

is a normalized Racah coefficient [74]. In two-center cases the sum reduces
to a single normalisation term, N0 . A Fourier transform approach is also
being investigated, extending [70].
The real space core overlaps then take the form:

1 1
s(n1 , l1 , m, n2 , l2 , , ) = Dl1 ,l2 ,m Yij Ai ( + ) Bj ( ) .

2 2
ij
(7.50)
Yij is a matrix with integer elements uniquely determined from n, l and m.
Dl1 ,l2 ,m is a coefficient that is independent of the principal quantum
number. It is obtained upon expanding the product of two Legendre func-
tions in this coordinate system. Symmetry conditions imply that only
m1 = m2 = m lead to non-zero coefficients.

1 1
Ai ( + ) = exp ( + ) i d (7.51)
2 1 2
1
1 1
Bj ( ) = exp ( ) j d. (7.52)
2 1 2
Here, recurrence relations on the auxiliary integrals A and B lead to
those for the requisite core integrals [75, 76]. These integrals may be pre-
calculated and stored.
Such integrals appear for two-center exchange integrals and three- and
four-center integrals (although just in one factor for three-center Coulomb
terms). Note that exchange integrals require distinct orbitals a and b .
In the atomic case, they must have different values for at least one of
n, l, m or . In the two-center case, the functions centered at a and b may
be the same. The product does not correspond to a single-center density,
it is two-centered. Equation (7.49) then illustrates the relationship to the
one-electron two-center overlap integral, although it clearly includes the
extra potential term from the Coulomb operator resolution.
7.16. Appendix E: Analysis of Nuclear Dipole Integrals for NMR

in a Slater Basis
The operator (
r
rN r, rN, )/(rN3 ) can be expressed as a combination
of terms involving cartesian coordinates. These terms take the following

general form:
X,i XN,j
3
, (7.53)
rN
where XN,j stands for cartesian coordinates of the electron with respect to
the nuclei N.
192 P.E. Hoggan
Now, it is more convenient to express the cartesian coordinates as sums

of spherical polar coordinates with their complex conjugates. These coor-
dinates are of the general form:

2 1
x=r [Y (
r ,
r ) Y11 (
r ,
r )]
3 1

2 1
y = ir [Y (
r ,
r ) + Y11 (
r ,
r )] (7.54)
3 1
4
z = r Y10 (
r ,
r ).
3
The product of a STO by a cartesian coordinate can be expressed as a
combination of STO, since:
rYLM (
r ,
r )n,l
m
(,
r )
= rYLM (
r ,
r ) r n1 er Ylm (
r ,
r )

max
= (1) M
lm |L M| m + M r n er Ym+M (
r ,
r )
=min ,2

max
= (1) M
lm |L M| m + M n+1,
m+M
(,
r ). (7.55)
=min ,2
Here we have used the Gaunt coefficients [124, 139] and the Condon
and Shortley phase convention for spherical harmonics Ylm (
r ,
r ) [120].
Consequently, the integral (7.31) is reduced to a sum of terms of the form:
j
Y ( , )
1 r
N r
N
(7.56)
2
rN
with just a 1/r 2 dependence and where j = 1, 0, 1. Using the Fourier

transform formalism requires the integral representation of the operator
involved in (7.56). We obtain:
j j
Y1 (r
N , r
N ) 2i Y1 (p
, p
)
2
(
p ) = (7.57)
rN 2 p
This immediately allows us to write the inverse Fourier transform:
j i
pr
N j
Y1 (r
N , r
N ) i e Y1 (p
, p
)
= 2
.
dp (7.58)
rN2 2 |
p|
Now, this places us in a position to write the Fourier integral for the present
term in the NMR nuclear shielding tensor calculation. After expanding the
Slater-type orbitals in terms of B-functions and substituting (7.58) in (7.56),
the present integral becomes:
i
pR
N j
i e Y1 (p
, p
)
I= 2
2 |
p| (7.59)
B m1
(1 ,
r )|ei
p
r |B 2 (2 ,
r R
2 )(
r) d p
n1 ,l1
m
n2 ,l2
,
whereas the three-center nuclear attraction integral is:
i
pR
N
1 e
I= Bnm11,l1 (1 ,
r )|ei
p
r |Bnm22,l2 (2 ,
r R
2 )(
r) d p

.
2 2 |
p|2
(7.60)
The three-center dipolar integral (7.13) appears in a form closely related to
that of the three-center nuclear attraction integrals required at the HF-SCF
level of electronic structure calculation (and also used in electronic DFT
work). In both above integrals note the presence of the common factor in
B-function Fourier transform work first studied by the Steinborn group:
Eq. (7.4 and 7.5). See [25], i.e.:
Bnm11,l1 (1 ,
r )|ei
p
r |Bnm22,l2 (2 ,
r R
2 )(
r) . (7.61)
The analytical treatment developed here has not required any hypothesis
on the relative position of nucleus (aligned or not) and any restriction on
quantum numbers. Consequently, the equation (7.59) is completely general
and may be directly evaluated from routines available in a quantum calcu-
lation software.
Note that such an integral satisfies all applicability conditions of non-
linear transformations for extrapolation described byA. Sidi [133]. Previous
work on three-center nuclear integral evaluation [132] has been used to
develop an efficient program to compute this dipolar 1/rN 3 three-center
integral.
Acknowledgements
Maria Belen Ruiz of Erlangen University invited me to write a book chapter

for [101]. Her input to the present text, through various discussions and
exchanges has been very helpful.
Measured NMR data was supplied by Bruno Combourieu of Lyon
University.
194 P.E. Hoggan
I was introduced to work on Slater-type orbitals in the early 1980s by my

PhD supervisor, Daniel Rinaldi who was then the CNRS specialist for pro-
gramming such applications. We co-directed two PhDs by A. Bouferguene
and M. Fares that ended in rather depressing conclusions for the speed of
four-center integrals. This was eventually resolved by integral representa-
tions for the orbital translation coefficients (see Appendix A, Section 7.12)
and the program STOP was to be written, almost ex nihilo during A.
Bouferguenes post-doc with me in Caen, Normandy from 19931997.
He wrote most of the code himself on an IBM RS 6000 funded by Rhone
Poulenc.
I must therefore thank D. Rinaldi for getting me involved with STOs for
molecular structure in the first place, Professor P. Nortier (Paris V, then at
Rhone Poulenc) for his vision in funding us for equipment and, of course
Ahmed Bouferguene for all his hard work.
Use of CHAMP for QMC runs was made possible by Julien Toulouse of
Pierre and Marie Curie University, Paris. A great deal of useful discussion
with his colleague Peter Reinhardt was also helpful.
Writing of this work was greatly facilitated by the constant love and
support of my wife, Pascale, due to be joined by our twin daughters.
Bibliography

[2] M. Randic, in International Conference on ETO Multicenter Molecular Integrals,
Tallahassee (1981), edited by C.A. Weatherford and H. Jones (Reidel, Dortdrech,
1982) pp. 141155.
[3] E. Scrocco and O. Salvatti, Ric. Sci. 21, 1629 (1951); ibid 22, 1766 (1952); ibid 23,
98 (1953).
[4] C. Petrongolo, E. Scrocco, and J. Tomasi, J. Chem. Phys. 48, 407 (1968) and
references therein.
[5] J. Podolanski, Ann. d. Physik. 402, 868 (1931) (in German).
[6] C.C.J. Roothaan, J. Chem. Phys. 19, 1445 (1951).
[7] A.C. Wahl, P.E. Cade, and C.C.J. Roothaan, J. Chem. Phys. 41, 2578 (1964).
[8] K. Rudenberg, J. Chem. Phys. 19, 1459 (1951).
[9] M. Kotani, A. Amemiya, and T. Simose, Proc. Phys. Math. Soc. Japan. 20, extra
No. 1, 1 (1938). M. Kotani and A. Amemiya, Proc. Phys. Math. Soc. Japan. 22,
extra No. 1, 1 (1940).
[10] M.P. Barnett and C.A. Coulson, Phil. Trans. R. Soc. Lond. A. 243, 221 (1951).
[11] P.O. Lowdin, Adv. Phys. 5, 1 (1956).
[12] F.E. Harris, Rev. Mod. Phys. 35, 558 (1963).
[13] J. Talman, Phys. Rev. Lett. 84, 855 (2000).
[14] R.S. Mulliken and C.C.J. Roothaan, Proc. Natl. Acad. Sci. U.S. 45, 394 (1959).
[15] S.F. Boys, Proc. Roy. Soc. (London). A200, 542 (1950).
[16] I. Shavitt and M. Karplus M, J. Chem. Phys. 36, 550 (1962).
[17] E. Clementi and D.L. Raimondi, J. Chem. Phys. 38, 2686 (1963).
[18] C.W. Scherr, J. Chem. Phys. 23, 569 (1955).
[19] S.J. Smith and B.T. Sutcliffe, in Reviews in Computational Chemistry, edited by
D.B. Boyd and K.B. Lipkowitz 10, 271 (1997).
[20] POLYATOM: D.B. Newmann, H. Basch, R.L. Korregay, L.C. Snyder, J. Moskowitz,
C. Hornback, and P. Liebman, Quantum Chemistry Program Exchange, Indiana
University, No. 199; I.G. Csizmadia, M.C. Harrison, J.W. Moscowitz, B.T. Sutcliffe,
Theoret. Chim. Acta. 6, 191 (1966).
[21] IBMOL: D.J. David, CDC 6600 Version. Technical Report of C.C. ENSJF and Lab.
de Chimie ENS, Paris (1969). E. Clementi, D.R. Davis, J. Comput. Phys. 1, 223244
(1967); A. Veillard, IBMOL: Computation of wave-functions for molecules of
general geometry, Version 4; IBM Research Laboratory, San Jose.
[22] A.D. McLean, M. Yoshimine, B.H. Lengsfield, P.S. Bagus, and B. Liu, in Modern
Techniques in Computational Chemistry, MOTECC 91, edited by Clementi E.
(Elsevier, Leiden, 1991) pp. 233353 .
[23] W.J. Hehre, W.A. Lathan, R. Ditchfield, M.D. Newton, and J.A. Pople. GAUSSIAN
70: Ab Initio SCF-MO Calculations on Organic Molecules QCPE 11, Programme
number 236 (1973).
[24] A. Bouferguene, M. Fares, and P.E. Hoggan, Int. J. Quantum Chem. 57, 801
(1996).
[25] H.H.H. Homeier, E. Joachim Weniger, and E.O. Steinborn, Comp. Phys. Comm.
72, 269 (1992). H.H.H. Homeier and E.O. Steinborn, Comp. Phys. Comm. 77, 135
(1993).
[26] J. Fernandez Rico, R. Lopez, I. Ema, and G. Ramrez, J. Comp. Chem. 19, 1284
(1998). ibid, J. Comp. Chem. 25, 1347 (2004).
[27] CADPAC: The Cambridge Analytic Derivatives Package, R.D. Amos, I.L. Alberts,
J.S. Andrews, S.M. Colwell, N.C. Handy, D. Jayatilaka, P.J. Kowles, R. Kobayashi,
K.E. Laidig, G. Laming, A.M. Lee, P.E. Maslen, C.W. Murray, J.E. Rice, E.D.
Simandiras, A.J. Stone, M.D. Su, and D.J. Tozer, Cambridge, (1995).
[28] E.J. Baerends, D.E. Ellis, and P. Ros, Chem. Phys. 2, 17 (1973).
[29] G. Te Velde, F.M. Bickelhaupt, E.J. Baerends, C. Fonseca Guerra, S.J. A. Van Gis-
bergen, J.G. Snijders, and T. Ziegler, J. Comp. Chem. 22, 931 (2001). ADF: Ams-
terdam Density Functional, available at http://www.scm.com/.
[30] C.F. Bunge, R. Jauregui, and E. Ley-Koo, Can. J. Phys. 76, 421 (1998).
[31] D. Pinchon and P.E. Hoggan, J. Phys. A: Math. Theor. 40, 1597 (2007).
[32] T. Ozdogan, Int. J. Quantum Chem. 92, 419 (2003).
[33] S.A. Hagstrom and H. Shull, Rev. Mod. Phys. 35, 624 (1963).
[34] D.C. Clary, Mol. Phys. 34, 793 (1977).
[35] J.S. Sims and S.A. Hagstrom, J. Chem. Phys. 124, 094101 (2006).
[36] M.B. Ruiz and K. Peuker, in Recent Advances in Computational Chemistry:
Molecular Integrals over Slater Orbitals, edited by T. Ozdogan and M.B. Ruiz,
(Transworld Research Network, Kerala, 2008) pp. 100144.
[37] E. Filter and E.O. Steinborn, Phys. Rev. A. 18, 1 (1978).
196 P.E. Hoggan
[38] J. Avery, Hyperspherical Harmonics and Generalized Sturmians (Kluwer, Boston,

2000); J. Avery, Int. J. Quantum Chem. (2004) 100, 121130; E. Red and C.A.
Weatherford Int. J. Quantum Chem. (2004) 100, 208213.
[39] H. Shull and P.O. Lowdin, J. Chem. Phys. 30, 617 (1959).
[40] J.S. Sims and S.A. Hagstrom, Phys. Rev. A. 4, 908 (1971). J.S. Sims and S.A.
Hagstrom, J. Chem. Phys. 55, 4699 (1971).
[41] J. Budzinski, Int. J. Quantum Chem. 97, 832 (2004).
[42] Y.G. Smeyers, Theoret. Chim. Acta. 4, 452 (1966).
[43] F.E. Harris and H.H. Michels, J. Chem. Phys. 43, S165 (1965).
[44] I.I. Guseinov, J. Chem. Phys. 119, 4614 (2003).
[45] S.F. Boys, G.B. Cook, C.M. Reeves, and I. Shavitt, Nature. 178, 1207 (1956).
[46] I. Shavitt and M. Karplus, J. Chem. Phys. 43, 398 (1965).
[47] L. Berlu, H. Safouhi and P.E. Hoggan, Int. J. Quantum Chem. 99, 221 (2004).
[48] H. Safouhi and L. Berlu, J. Comp. Phys. 216, 19 (2006).
[49] D. Levin and A. Sidi, Appl. Math. Comp. 9, 175 (1981).
[50] A. Sidi, Math. Comp. 38, 517 (1982).
[51] I.I. Guseinov, Int. J. Quantum Chem. 90, 114 (2002).
[52] M.B. Ruiz, Habilitation Thesis, University of Erlangen-Nuremberg (2010).
[53] J. Fernandez Rico, R. Lopez, G. Ram rez, I. Ema, and E.V. Ludena, J. Comp. Chem.
24, 1355 (2004).
[54] D. Pinchon and P.E. Hoggan, Int. J. Quantum Chem. 109, 135 (2009).
[55] T. Kato, Commun. Pure Applied Math. 10, 151 (1957).
[56] S. Agmon, in Schrodinger Operators, edited by S. Graffi, Lecture Notes in Mathe-
matics 1159 (Springer-Verlag, Berlin, 1985) pp. 138.
[57] S. Agmon, Lectures on Exponential Decay of Solutions of Second Order Elliptic
Equations: Bounds on Eigenfunctions of N-Body Schrodinger Operators (Princeton
University, Princeton, 1982).
[58] R.M. Stevens, J. Chem. Phys. 52, 1397 (1970).
[59] H.W. Jones, International Conference on ETO Multicenter Integrals (Tallahassee,
USA 1981), edited by C.A. Weatherford and H.W. Jones (Reidel, Dordrecht, 1982);
Special Issue in memory of H.W. Jones, edited by C.A. Weatherford and P.E. Hoggan,
Int. J. Quantum Chem. 100, Volume 2 (2004).
[62] A. Bouferguene and P.E. Hoggan, QCPE, Programme No. 667, (1996).
[63] C.A. Weatherford, E. Red, D. Joseph, and P.E. Hoggan, Mol. Phys. 104, 1385 (2006).
[64] J. Avery and J. Avery, Hyperspherical Harmonics and Generalized Sturmians
(Kluwer, Boston, 2007).
[65] J.C. Tully, J. Chem. Phys. 58, 1396 (1973).
[66] C.A. Weatherford, E. Red, and P.E. Hoggan, Mol. Phys. 103, 2169 (2005).
[67] J.C. Cesco, J.E. Perez, C.C. Denner, G.O. Giubergia, and A.E. Rosso, Applied Num.
Math. 55, 173 (2005) and references therein.
[68] Y. Shao, C.A. White, and M. Head-Gordon, J. Chem. Phys. 114, 6572 (2001).
[69] S.A. Vagranov, A.T.B. Gilbert, E. Duplazes, and P.M.W. Gill, J. Chem. Phys. 128,
201104 (2008).
[70] P.M.W. Gill and A.T.B. Gilbert, Chem. Phys. 356, 86 (2009).
[71] D. Pinchon and P.E. Hoggan, Unpublished code.
[72] J.D. Talman, Phys. Rev. A. 48, 243 (1993).

[73] E.T. Whittaker and G.N. Watson, A Course in Modern Analysis, 4th Ed. (Cambridge
University Press, Cambridge, 1990).
[74] P.E. Hoggan, Int. J. Quantum Chem. 109, 2926 (2009).
[75] I.I. Guseinov, A. Ozmen, U. Atav, H. Uksel, Int. J. Quantum Chem. 67, 199 (1998).
[76] P.E. Hoggan, DSc Thesis, University of Nancy I. (1991) (in French available at
www.INIST.fr), Appendix 2.
[77] R.J. Hinde, J. Chem. Phys. 128, 154308 (2008).
[78] B.I. Dunlap, Phys. Rev. A. 128, 032502 (2002).
[79] J.D. Talman, Int. J. Quantum Chem. (2010) DOI: 10.1002/qua.22511
[80] N. Absi and P.E. Hoggan, Int. J. Quantum Chem. 106, 2881 (2006).
[81] R.S. Mulliken, J. Chem. Phys. 3, 375 (1935).
[82] C.A. Coulson, Trans. Faraday Soc. 33, 1479 (1937).
[83] H.M. James and A.S. Coolidge, J. Chem. Phys. 1, 825 (1933).
[84] W. Kolos and C.C.J. Roothaan, Rev. Mod. Phys. 32, 205 (1968).
[85] W. Kolos and L. Wolniewicz, J. Chem. Phys. 43, 2429 (1965). ibid 49, 404 (1968).
[86] A.A. Frost and J. Braunstein, J. Chem. Phys. 19, 1133 (1951).
[87] Y. Kurokawa, H. Nakashima, and H. Nakatsuji, Phys. Rev. A. 72, 062502 (2005).
[88] W. Cencek and K. Szalewicz, Int. J. Quantum Chem. 108, 2191 (2008).
[89] B.L. Zhou, J.M. Zhu, and Z.C. Yan, Phys. Rev. A. 73, 064503 (2006).
[90] G.W.F. Drake, Phys. Scr. T83, 83 (1999).
[91] D. Frey, A. Preiskorn, G.C. Lie, and E. Clementi, in Modern Techniques in Com-
putational Chemistry: MOTECC-90, edited by E. Clementi (ESCOM Science Pub-
lishers, Leiden, 1990), pp. 5797.
[92] J. Rychlewski and J. Komasa, in Explicitly Correlated Wave Functions in Chemistry
and Physics, edited by J. Rychlewski (Kluwer, Dordrecht, 2004), pp. 91147.
[93] W. Kutzelnigg, Theor. Chim. Acta. 68, 445 (1985).
[94] W. Klopper and W. Kutzelnigg, Chem. Phys. Lett. 134, 17 (1987).
[95] W. Klopper and J. Noga, in Explicitly Correlated Wave Functions in Chemistry and
Physics, edited by J. Rychlewski (Kluwer, Dordrecht, 2004), pp. 149183.
[96] W. Cardoen, R.J. Gdanitz and J. Simons, J. Phys. Chem. A. 110, 564 (2006).
[97] Explicitly Correlated Wave Functions in Chemistry and Physics, edited by J. Rych-
lewski (Kluwer, Dordrecht, 2004).
[98] D.H. Bailey and J.M. Borwein, J. Symbolic Comp. (2010 in press) High-precision
software directory. Available from: http://crd.lbl.gov/ dhbailey/mpdist/mpdist.html
[99] R.P. Brent. A Fortran multiple-precision arithmetic package (TOMS), 4, pp. 5770,
(1978).
[100] A.J. Miller, TOMS (2010). A. Millers Fortran software. Available from:
http://users.bigpond.net.au/amiller/
[101] Recent Advances in Computational Chemistry: Molecular Integrals over Slater
Orbitals, edited by Ozdogan T. and Ruiz M.B. (Transworld Research Network,
Kerala, 2008).
[102] M. Abramowitz and I.A. Stegun, Handbook of Mathematical Functions with For-
mulas, Graphs, and Mathematical Tables (US Government Printing Office, Wash-
ington, D.C., 1964).
[103] M.P. Barnett, J. Chem. Phys. 113, 6419 (2000).
[104] A. Bouferguene, J. Phys. A: Math. Gen. 38, 2899 (2005).
198 P.E. Hoggan
[105] A. Bouferguene and D. Rinaldi, Int. J. Quantum Chem. 50, 21 (1994).

[106] I.S. Gradshteyn and I.M. Ryzhik, Table of Integrals, Series, and Products (Academic
Press, New York, 1994).
[107] H.W. Jones, in Int. Conf. on ETO Multicenter integrals, Tallahasse, Florida, edited
by C.A. Weatherford and H.W. Jones, 1981, p. 53.
[108] H.W. Jones, Phys. Rev. A. 38, 1065 (1988).
[109] H.W. Jones, J. Comput. Chem. 38, 1217 (1991).
[110] H.W. Jones and J. Jain, Int. J. Quantum Chem. 23, 953 (1983).
[111] H.W. Jones and C.A. Weatherford, Int. J. Quantum Chem. 12, 483 (1978). (Inter-
national Symposium on Atomic, Molecular, and Solid-State Theory, Collision Phe-
nomena and Computational Methods, Flagler Beach, Florida, 1978).
[112] H.W. Jones and C.A. Weatherford, J. Mol. Struct. (Theochem) 199, 233 (1989).
[113] F.W.J. Olver, Asymptotics and Special Functions (Academic Press, New York,
1974).
[114] R.R. Sharma, Phys. Rev. A, 13, 517 (1976).
[115] A. Bouferguene, PhD thesis, University of Nancy I, France (1992).
[116] A. Sidi and P.E. Hoggan, J. Math. Anal. Appl., submitted.
[117] D. Antolovic and H.J. Silverstone, Int. J. Quantum Chem. 100, 146 (2004).
[118] L. Berlu, H. Safouhi, and P. E. Hoggan, Int. J. Quantum Chem. 99, 221 (2004).
[119] A.J. Cohen and N.C. Handy, J. Chem Phys. 117, 1470 (2002). M.A. Watson, A.J.
Cohen, and N.C. Handy, J. Chem. Phys. 119, 6475 (2003). M.A. Watson,A.J. Cohen,
N.C. Handy, and T. Helgaker, J. Chem. Phys. 120, 7252 (2004).
[120] E.U. Condon and G.H. Shortley, The Theory of Atomic Spectra (Cambridge Uni-
versity Press, Cambridge 1978).
[121] R. Ditchfield, J. Chem. Phys. 56, 5688 (1972).
[122] R. Ditchfield, Mol. Phys. 27, 789 (1974).
[123] M. J. Frisch et al. Gaussian 03, Revision C.02 (Gaussian, Inc., Wallingford CT,
2004).
[124] J.A. Gaunt, Phil. Trans. Roy. Soc. A(228), 151 (1929).
[125] P.E. Hoggan, Int. J. Quantum Chem. 100, 214 (2004).
[126] J. Toulouse, M. Caffarel, P. Reinhardt, P. E. Hoggan, and C. J. Umrigar. PTCP,
Proceedings of QSCP XV (in press 2011).
[127] N. Joudieh, PhD thesis, Faculte des Sciences de lUniversite de Rouen, Rouen,
France (1998).
[128] F. London, J. Phys. Radium. 8, 397 (1937).
[129] J. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
[130] D. Pinchon and P. E. Hoggan, Int. J. Quantum Chem. 107, 2186 (2007).
[131] E. Red and C. A. Weatherford, Int. J. Quantum Chem. 100, 204 (2004).
[132] H. Safouhi and P. E. Hoggan, Mol. Phys. 101, 19 (2003) and references therein.
[133] A. Sidi. Practical extrapolation methods: Theory and Applications, (Cambridge
University Press, Cambridge, 2003) Chapter 11 and references therein.
[134] E.O. Steinborn and E. Filter, Theor. Chim. Acta. 38, 273 (1975).
[135] E. O. Steinborn, H.H.H. Homeier, I. Ema, R. Lopez, and G. Ramrez, Int. J. Quantum
Chem. 76, 244 (2000).
[136] R. M. Stevens and W. N. Lipscomb, J. Chem. Phys. 40, 2238 (1964).
[137] L. Vieille, L. Berlu, B. Combourieu, and P.E. Hoggan, J. Theo. Comp. Chem. 1, 295
(2002).
[138] E.J. Weniger, J. Mathematical Physics. 26, 276 (1985).

[139] E.J. Weniger and E.O. Steinborn, Comput. Phys. Commun. 25, 149 (1982).
[140] E.J. Weniger and E.O. Steinborn, J. Math. Phys. 30, 774 (1989).
[141] H.-J. Werner and P.J. Knowles, MOLPRO, version 2006.1 a package of ab initio
programs, www.molpro.net
[143] T. Bouabca, N. Ben Amor, D. Maynau, and M. Caffarel, J. Chem. Phys. 130, 114107
(2009).
[144] http://www.physics.utah.edu/ harris/home.html
[145] K. Ishida, ICQC 13 (2009) and private communication.
[146] J. Toulouse, M. Caffarel, P.E. Hoggan, P. Reinhardt and C.J. Umrigar, PTCP (in
press, 2011).

Chapter 8
Modern Ab Initio Valence Bond Methods
Philippe C. Hibertya and Sason Shaikb

a Laboratoire de Chimie Physique, UMR CNRS 8000,
Universite de Paris-Sud, 91405 Orsay, CEDEX, France
b Institute of Chemistry and The Lise Meitner-Minerva
Center for Computational Quantum Chemistry,

The Hebrew University of Jerusalem, Jerusalem 91904, Israel
Quantum mechanics has provided chemistry with two general theories of

bonding: valence bond (VB) theory and molecular orbital (MO) theory.
VB theory is essentially a quantum mechanical formulation of the classical
concept of the chemical bond wherein the molecule is regarded as a set
of atoms held together by local bonds. This is a very appealing model
as it represents the quantum mechanical translation of the classical basic
concepts that are deeply rooted in chemistry, such as Lewis structural
formulas, chemical valency, hybrid orbitals, and resonance. MO theory,
on the other hand, uses a more physics-related language and has sprung
as a means to interpret the electronic spectra of molecules and deal with
excited states. However, with its canonical MOs delocalized over the entire
molecule, this theory bears little relationship to the familiar language of
chemists in terms of localized bonds and this is probably the reason why
it was initially eclipsed by VB theory, up to the mid-1950s. Then the
situation reversed and MO theory took over, among other reasons, because
of the efficient implementations, which provided the chemical community
with computational software of ever increasing speed and capabilities.
Nowadays, with the advent of modern computational ab initio VB
methods and the progress in computer and coding technologies, VB theory
is coming of age. Indeed, starting with the 1980s onward, several method-
ological advances in VB theory have been made, and allowed new and
more accurate applications. Thus, dynamic correlation has been incorpo-
rated into VB calculations, so that at present, sophisticated VB methods
201
202 P.C. Hiberty and S. Shaik
combine the accuracy of post-HF methods with the specific advantages

of VB theory such as extreme compactness of the wave functions that are
readily interpreted in terms of Lewis structures, ability to calculate dia-
batic states, resonance energies and so on. Moreover, VB theory has been
recently extended to handle species and reactions in solution, and is also
capable of treating transition metal complexes. These newly developed
tools have been used to verify and quantify fundamental concepts such as
aromaticity, resonance energies, hybridization and so on, and to develop
new ideas and models in chemical reactivity, that were not foreseen from
the empirical VB model. The combination of the lucid insight inherent
to VB theory and its new computational capabilities is discussed in this
chapter. We hope that this chapter makes a case also for using these modern
ab initio VB methods as routine tools in the service of chemistry.
8.1. Basic Principles and Survey of Modern Methods
8.1.1. VB vs. MO wave functions in the two-electron/two-center

case
The description of the chemical bond in VB theory can be illustrated with
the example of the H2 molecule. Let a and b be two atomic orbitals (AOs)
localized on the left and right hydrogen atom, respectively (Scheme 8.1).
To a first approximation, one may consider the fully covalent description of
the bond, as in the pioneering article of Heitler and London [1]. Dropping
normalization constants hereafter, the corresponding wave function, HL
in Eq. (8.1), displays two Slater determinants (see the definition of a Slater
determinant in next box), each representing a situation where both atoms
are neutral and bear electrons of opposite spins:

HL (H H) = a b a b (8.1)
The bar over an orbital indicates spin-down (), while the absence of a bar
indicates spin-up ().
HL-Wavefunction Covalent-Ionic Superposition MO-Wavefunction

in an AB Bond
u
H H H H A B A B A B g
1 2 3
Scheme 8.1. VB and MO wave functions for the two-electron/two-center case.

Modern Ab Initio Valence Bond Methods 203
Definition of Slater determinants

The Pauli principle requires the electronic wave function to be antisym-
metric with respect to the exchange of two electrons. A simple orbital-product
wave function, e.g. 1 (1)2 (2) for a two-electron case, does not satisfy this
requirement. The product can however, be antisymmetrized, as in D(2-e)
below:
1
D(2e) = [1 (1)2 (2) 2 (1)1 (2)] (B1)
2
Equation (B1) is then expressible as a determinant, called Slater determinant:

1 1 (1)1 (2)
D(2e) = (B2)
2 2 (1)2 (2)
Generalizing, a Slater determinant for a state involving N electrons is an
antisymmetrized product of N orbitals and therefore contains N! terms, cor-
responding to the N! possible orbital permutations applied to the diagonal
term, e.g. 1 (1)2 (2) in Eq. (B2). This can be written in the form of an
NN-dimensional determinant, multiplied by a factor (N!)1/2 . This deter-
minant is usually represented with a simplified notation, using the diagonal
term:
D(Ne) = |1 2 . . . N | (B3)
This notation contains all the information needed to generate the NN-
dimensional Slater determinant. The factor (N!)1/2 is implicitly assumed
in Eq. (B3).
By itself, each Slater determinant in Eq. (8.1) is not much lower in energy
than the two separate atoms, and is therefore practically non-bonding [2].
It is the superposition of the two determinants, or resonance between the
two spin arrangements, that creates bonding, as represented in 1 in Scheme
8.1. This early description was remarkably successful since it accounted
for 75% of the total bond energy of H2 . For a complete description, it is
necessary to include two ionic terms, as is done in Eq. (8.2) that represents
the full VB wave function for a general single bond between two atoms A
and B:

VB (A B) = C1 a b a b + C2 a a + C3 b b (8.2)
where the two last Slater determinants represent A B+ and A+ B ionic
situations, as in 2 (Scheme 8.1). Thus, the full VB wave function is a linear
combination of three VB functions (generally referred to as VB struc-
tures), each representing a particular bonding scheme. The AOs that are
used to construct the VB structures in Eq. (8.2), are defined as linear com-
binations of the basis functions, , centered on a single atom, Eq. (8.3):

i = Ti (8.3)

and taken from standard basis sets. The AOs are 1s types for hydrogen
atoms, and hybrids of ns and np basis functions for heavier atoms, giving
rise to the concept of hybridization. For example, the so resulting hybrid
atomic orbitals (HAOs) of carbon in a C-H bond in different molecules
will resemble the sp3 , sp2 or sp types, well known from the important
hybridization concept.
The VB description of the AB bond can be compared with the simple
MO description, in which a unique delocalized MO is doubly occupied, 3
in Scheme 8.1, Eq. (8.4):

MO (A B) = g g ; g (a + b ) (8.4)
where the coefficients of a and b in g are assumed to be equal for

simplicity.
This simple MO wave function is insufficient however, and a better
description of the bond requires configuration interaction (CI) between
MO and some excited configurations, as in Eq. (8.5), for an accurate
description of the AB bond:

MOCI (A B) = g g + g u g u + |u u | ;
u (a b ) (8.5)
Clearly, VB and MOCI may seem to speak different languages.

However simple algebra, consisting of expanding the MO determinants
of Eq. (8.5) into AO determinants like those of Eq. (8.2), shows that these
two wave functions are exactly equivalent and therefore provide the same
bonding energy, and any other property of the AB molecule [3, 4]. This
equivalence between MO-CI and VB descriptions is general, so that any
MO-CI wave function can be transformed into a VB wave function and vice
versa [5]. It is therefore clear that both MO-CI and VB methods, if pushed
to a sufficient level of sophistication, are accurate theories. For example,
despite some still existing erroneous views, VB methods at any level of
sophistication, perfectly account for the triplet state of dioxygen, the valence
ionization spectrum of methane and water, the aromaticity/antiaromaticity
dichotomy and so on [3, 4] (see Appendix A; Section 8.5).
2 4
1 3
Scheme 8.2. Covalent structure in the -bonding system of butadiene.
8.1.2. Writing VB functions beyond the two-electron/two-center

case
Writing VB wave functions for polyelectronic cases follows simple exten-
sions of the two-electron/two-center example discussed above. Let us con-
sider butadiene and restrict the description to the system. Denoting the
AOs of the C1 C4 carbons by a , b , c , and d , respectively, the fully
covalent VB structure cov for the system of butadiene displays two
covalent bonds: one between a and b , and one between c and d , as
represented in 4 in Scheme 8.2. It follows that the covalent VB structure
can be expressed in the form of Eq. (8.6) as a product of the bond wave
functions.

cov (4) = (a b a b )(c d c d ) (8.6)
Upon expansion of the product, one gets a sum of four determinants as
in Eq. (8.7):

cov (4) = a b c d a b c d a b c d + a b c d (8.7)
The product of bond functions in Eq. (8.6) involves so-called perfect
pairing, whereby the covalent structures of the molecule involves a product
of all the bond-pair wave functions of the individual bonds in the corre-
sponding Lewis structure. As a rule, such a perfect-pairing polyelectronic
VB structure having n bond pairs will be described by 2n determinants,
displaying all the possible spin permutations between the orbitals that are
singlet-coupled. Inclusion of the ionic components of the bonds results in
the final VB wave function expressed in Eq. (8.8) as a linear combination
of VB structures K :

VB = CK K (8.8)
K
One drawback of this description of a molecular system, as in Eq. (8.8),
in the early VB theory was the exponential growth in the number of VB
structures with the number of bonds in the molecule. This had in turn two
undesired consequences: (i) the wave function lost lucidity; and (ii) the cal-
culation became time-consuming. Thus, since the Hamiltonian VB matrix
elements are not as easily calculated as in MO-CI theory, owing to the

so-called N! problem (see box), the growth in the number of VB con-
figurations posed a formidable computational challenge. Additionally, the
early VB theory used the same AOs as in the free atoms, and consequently
the calculations suffered from the lack of efficient methods for orbital opti-
mization and for incorporation of dynamic correlation. As will be seen
below, these difficulties have been overcome by the new generation of VB
methods.
The non-orthogonality problem or the N! problem

In MO theory, two different Slater determinants have a zero-overlap, owing to
the orthogonality of the orbitals. In VB theory, the orbitals are generally non
orthogonal, and the overlap between two N-dimensional determinants is calcu-
lated as a sum of N! products of orbital overlaps, a complication that has been
termed the N! problem. Similarly, the Hamiltonian matrix elements between
Slater determinants are calculated by means of the very simple SlaterCondon
rules when the orbitals are orthogonal, whereas for non-orthogonal orbitals
these simple rules are not applicable, and there are many more terms multi-
plied by overlaps. It follows that the evaluation of these matrix elements is
much more time-consuming in the VB framework than in configuration inter-
action in the MO framework. However, the term the N! problem is actually
a misnomer, since the difficulty due to non-orthogonality does not imply that
the computational effort required to perform a non-orthogonal configuration
interaction scales as N! Modern ab initio VB methods generally scale as N 4 .
8.1.3. Some landmark improvements of the early VB method

To keep the advantage of VB theory while being at the same time efficient,
the new VB methods had to meet several challenges : (i) to provide compact
VB wave functions that would be clearly interpretable in terms of Lewis
structure, (ii) to be as little time-consuming as possible, and (iii) to provide
energetics as accurate as sophisticated MO-CI methods. All three chal-
lenges have been achieved by major modernization of the early VB method.
Some of these landmark improvements are briefly described below.
8.1.3.1. VB wave functions with semi-localized atomic orbitals

A great step for obtaining compact wave functions was made by Coulson
and Fischer [6], who proposed to describe the two-electron bond as a for-
mally covalent singlet-coupling between two orbitals a and b , the latter
being optimized with freedom to delocalize over the two centers. This is
exemplified in Eq. (8.9) for the AB bond:
CF (A B) = |a b | |a b | (8.9a)
a = a + b (8.9b)
b = b + a (8.9c)
Here a and b are purely localized AOs (or HAOs), while a and b
are delocalized orbitals. In fact, experience shows that the CoulsonFischer
orbitals a and b , which result from the energy minimization, are generally
not extensively delocalized (, < 1), and as such they can be viewed
as distorted atomic orbitals. However, minor as this may look, this slight
delocalization renders the CoulsonFischer wave function equivalent to the
VB (AB) wave function (Eq. (8.2)) with the three classical VB structures.
A straightforward expansion of the CoulsonFischer wave function leads
to the linear combination of the classical structures, in Eq. (8.10).

CF (A B) = (1 + ) a b a b

+2 a a + 2 b b (8.10)
Thus, the CoulsonFischer representation keeps the simplicity of the
covalent picture while treating the covalent/ionic balance by embedding
the effect of the ionic terms in a variational way, through the delocalization
tails of the VB orbitals. The CoulsonFischer idea has later been generalized
to polyatomic molecules and gave rise to the generalized valence bond
(GVB) [7] and spin-coupled (SC) [8] methods. The advantage of using wave
functions of CoulsonFischer type becomes obvious when one wishes to
treat all the bonds of a molecular system in a VB way. For example, the
GVB wave function representing methane with its four C-H bonds needs a
single formally covalent structure (Eq. (8.11)),

GVB = (1 h1 1 h1 )(2 h2 2 h2 )(3 h3 3 h3 )(4 h4 4 h4 )
(8.11)
where the i s are the four HAOs of the carbon atom, which are singlet-
coupled to the orbitals hi of the hydrogen atoms; both i and hi are
localized on their respective centers while bearing small delocalization tails
to the other centers. This is a great simplification compared with the mixed
covalent/ionic wave function that possesses 81 mixed structures based on
localized orbitals.
Letting all the orbitals in Eq. (8.11) be determined variationally leads
to four sp3 type HAOs i pointing in tetrahedral directions toward the
corresponding hydrogen atoms, as shown in Scheme 8.3. These HAOs,
h1
1
4 C 2
h4
3 h2
h3
Scheme 8.3. The schematic GVB wave function for methane.
which come out from a variational calculation without the input of any
qualitative preconception, clearly demonstrate the validity of the universally
used hybridization model. Incidentally, this GVB wave function is much
lower in energy than the simple MO wave function with its delocalized
canonical MOs [9].
8.1.3.2. Efficient orbital optimization by Self-Consistent-Field VB

Despite their remarkable usefulness, the VB methods which utilize
CoulsonFischer orbitals are not able to distinguish ionic from covalent
structures, both types of structures being implicitly embedded in a for-
mally covalent wave function. For example, in the transition state of the
SN 2 reaction Eq. (8.12),
X + H3 CY [X . . . CH3 . . . Y] XCH3 + Y (8.12)
the ionic structure 7 (Scheme 8.4) plays a fundamental role, and it is
important to be able to calculate its relative energy and relative weight
with respect to the covalent structures 5 and 6.
To account for the ionic structures, one has to go back to the classical
VB representation in terms of VB structures built with pure HAOs. The
compactness of the VB wave function is maintained by treating only the
active orbitals and electrons (those that are implied in bond-breaking/bond-
forming in the reaction, red in Scheme 8.4) in a VB fashion, while the other
orbitals (called inactive or spectator orbitals, black in Scheme 8.4) are
treated as doubly occupied bonding orbitals throughout the reaction, i.e.
treated in a MO way. This way, the transition state of an SN 2 reaction is
represented by a total of six VB structures, three of which being represented
in Scheme 8.4 while the three remaining ones are practically negligible.
The calculation of such a multi-structure VB wave function, in the spirit
of Eq. (8.8), can then be made by means of the Valence Bond Self Consistent
H H
X C Y X C Y
HH HH

6 XCH3 Y
5 X H3CY
H
X C Y
HH
7 X H3C+ Y
Scheme 8.4. The main VB structures necessary to describe an SN 2 transition

state.
field method (VBSCF) [10], which optimizes the coefficients CK and the
orbitals of the VB structures K simultaneously. The VBSCF method has
been implemented in ab initio codes by van Lenthe et al. [11], with effi-
cient algorithms of orbital optimization that get rid of the N! problem. The
algorithm has been further improved just recently by Wu et al. [12], and
even faster versions are currently in progress.
8.1.3.3. Improving the accuracy by including dynamic electron

correlation
The accuracy of the GVB, SCVB and VBSCF methods is comparable to that
of valence-CASSCF, which is sufficient for many applications. However,
none of these VB or MO methods are capable of yielding accurate reaction
barriers or dissociation energies. The reason for that becomes clear if one
examines the pictorial representation of the AB bond as described at the
VBSCF level in Scheme 8.5a (recall that the GVB, SC or MO-CASSCF
wave functions are practically equivalent to the VBSCF ones).
In VBSCF , the necessary VB structures (810) are present and their
coefficients and orbitals are optimized simultaneously, thus covering the
main part of electron correlation, so-called static correlation. Now the set
of HAOs is common to the three VB structures 810, with the consequence
that the active orbitals (red in Scheme 8.5a) have the same size and shape
whether they are singly occupied as in 8 or doubly occupied as in 9 or 10.
Similarly, the inactive orbitals (black in Scheme 8.5a) are optimized for
an average mean-field situation while they experience the fields of neutral
C1 A B + C2 A B + C3 A B
8 9 10
(a)
C'1 A B + C'2 A B + C'3 A B
8 9 10
(b)
Scheme 8.5. Representation of the AB bond by the VBSCF (a) and BOVB (b)
methods. The spectator orbitals in black may be lone pairs or bond orbitals of bonds
between A and/or B to substituents on A and B.
atoms in 8, versus ions in 9 and 10. Clearly, a better wave function would
be allowed to have different orbitals for different VB structures. Such a
wave function is represented in Scheme 8.5b, where it is seen that the
orbitals surrounding, e.g. A in 9 or B in 10, are drawn bigger than those
surrounding A and B in 8. This is the essence of the breathing-orbital
valence bond method (BOVB) [13], and this improvement, that keeps the
wave function as compact as in VBSCF, brings some dynamic correlation
that is necessary for getting accurate dissociation energies.
Another recent VB method that takes care of dynamic correlation is the
VBCI method [14]. This is a post-VBSCF approach, where the VBSCF
wave function serves as a reference wave function for the CI procedure.
Thus, excited VB structures are generated from the reference wave function
by replacing occupied (optimized VBSCF) orbitals with virtual orbitals, and
CI is performed between the reference VB structures and the excited ones.
To generate physically meaningful excited structures, the virtual orbitals are
constructed to be strictly localized, like the occupied VB orbitals. After the
CI has been done, the reference and all the excited VB structures that rep-
resent the same bonding scheme are condensed into a single structure. In this
manner, the extensive VBCI wave function is condensed to a minimal set of
fundamental structures, which ensures that VBCI keeps the VB advantage
of compactness.
A much faster variant than VBCI is the VBPT2 method [15], in which
the excited VB structures are treated by perturbation theory to second order
rather than by CI. According to our experience and to benchmark calcula-

tions [15], the BOVB, VBCI and VBPT2 methods all provide reasonably
accurate reaction barriers and dissociation energies. This latter feature is
crucially important for VB applications to chemical reactivity, as will be
exemplified below.
8.1.3.4. Inclusion of solvent effects

Coupling the VB method with the polarizable continuum model (PCM)
generates the VBPCM method, which was developed for exploring the
solutesolvent interactions at the ab initio VB level [16]. To incorporate
solvent effects into a VB scheme, the state wave function is expressed in
the usual form as a linear combination of VB structures, but now these VB
structures interact with one another in the presence of the polarizing field
of a solvent. VBPCM enables one to study the energy curve of the full VB
state as well as that of individual VB structures throughout the path of a
chemical reaction, and then reveal the solvent effect on the different VB
structures as well as on the total VB wave function.
Another solvation model, the SMx model developed by Truhlar and co-
workers [17], has also been incorporated into VB theory, leading to the
VBSM method [18]. For large systems such as biological systems, VB
theory has been combined with molecular mechanics (MM). The resulting
VB/MM method [19] utilizes the ab initio VB approach for the reactive
fragments and MM for the environment and has been applied to an SN 2
reaction in solution [19].
8.2. Strengths of the Valence Bond Approach
8.2.1. Interpretability combined with accuracy of the

wave functions
As mentioned above, modern ab initio VB methods like BOVB, VBCI or
VBPT2 can combine extreme compactness and clear physical meaning of
the wave functions with accuracy of the calculated energies. This combi-
nation of compactness and accuracy can be appreciated with the example
of the dissociation energy of difluorine, a well-known difficult test case. At
the MO (HartreeFock) level, even with very large basis sets, this molecule
is unbound [20], in contrast to an experimental dissociation energy of 38.3
kcal/mol [21]. At the GVB, SC, VBSCF and valence-CASSCF levels,
the calculated dissociation energy is about half the experimental value
[13]. Finally, a simple BOVB/cc-pVTZ calculation provides a dissociation
energy of 37.9 kcal/mol, very close to the experimental value, with only
three VB structures (810 in Scheme 8.5) [22]. By comparison, in the MO
framework, an MCSCF treatment must go far beyond the valence CASSCF
level, and the resulting dissociation energy oscillates between too small to
too large values until as many as 968 configurations are included and the
dissociation energies converges [23].
VB theory provides a clear picture of the very important concept of
electron correlation. For a single bond, the static electron correlation is
accounted for if the three VB structures are given optimized coefficients, as
in Scheme 8.5a, whereas the weights of the ionic structures are systemat-
ically overestimated at the HartreeFock level. This equilibration of ionic
vs. covalent coefficients is also called left-right correlation. On the other
hand, the subtle breathing-orbital effect, by which the orbitals rearrange
in size and shape to follow the charge fluctuation in the bond (as shown
in Scheme 8.5b) is associated with dynamic correlation, more precisely,
with the change in dynamic correlation that attends bond-breaking/bond-
formation of the AB bond (also called differential dynamic correlation).
This latter term is the dominant correlation term in three-electron/two-
center (3e/2c) bonds, in which there is no left-right correlation. To illus-
trate this point, consider theAB anion that possesses a three-electron bond,
noted [A B] . At the HartreeFock level, the bonding and antibonding
orbitals g and u (see Scheme 8.1) are doubly and singly occupied, respec-
tively. It turns out that the HartreeFock wave function is equivalent to the
VBSCF description, i.e. a resonance between two Lewis structures, as an
expansion of the HartreeFock determinant would show (Eq. (8.13)):

MO [A B] = g g u = b b a a a b ;
g (a + b ), u (a b ) (8.13)
Since the HartreeFock description involves the two Lewis structures
that can be drawn for this system with the right coefficients (5050 ifA and B
have equivalent electronegativities), no left-right correlation is necessary to
re-equilibrate these coefficients by CI, so that the only electron correlation
that has to be accounted for is dynamic [24]. In accord, a simple BOVB
description of the [A B] bond, in terms of two VB structures 11 and
12 each having their own set of optimized orbitals, contains all the physics
of 3e/2c interactions and provides accurate bonding energies [25]. Besides
the very simple picture provided by VB theory for dynamic correlation,
it should be noted that the contribution of this term to the 3e/2c bonding
energy is very large in all cases [26,27]. For example, the bonding energy
of F2 , as calculated at the HartreeFock level, is close to zero, compared
to 28.0 kcal/mol with a two-structure BOVB calculation and 30.2 kcal/mol

experimentally [21].
The BOVB method not only provides dissociation energies, but also
accurate dissociation curves at any interatomic distance [13b], and accurate
relative energies of transition states relative to reactants and products
[2831]. The root cause of this combination of accuracy and compactness
(e.g. six structures for an SN 2 reaction [29], and eight for a radical
abstraction reaction [31]) probably lies in the intuitive principle that if
an electronic state is described by all relevant VB structures and if each of
these VB structures has its optimal specific set of orbitals, then the relative
energy of this electronic state should be balanced throughout a potential
energy surface. Another point that favors compactness is that one does not
take care of all the electron correlation in a molecular system (as full CI
would do), but only the part of electron correlation that varies throughout
a reaction coordinate or potential surface is taken into account. It is in this
spirit that spectator orbitals can be held doubly occupied as in a simple MO
treatment, and that only the differential part of dynamic correlation is taken
into account.
8.2.2. A simple solution to the symmetry dilemma

The so-called symmetry dilemma has been first reported by Lowdin [32]
and it refers to an artefactual symmetry-breaking of the wave function,
which occurs in some MO-based methods, even sophisticated ones. The
problem is encountered each time a molecular system qualitatively corre-
sponds to a resonance between two (or more) Lewis structures of equal or
quasi-equal weights, as for example in a three-electron bond [A A] .
In such cases, an unphysical symmetry-broken solution may happen to be
lower in energy compared with the symmetry-adapted one, resulting in poor
relative energies, inaccurate molecular properties and so on. The problem
is general and exists in a variety of open-shell electronic states involving
radicals of allylic types, core-ionized diatoms, charged clusters, etc.
The symmetry dilemma is clearly illustrated with the example of the
[A A] species (Eq. (8.13) and Scheme 8.6 with A = B). As has been
said, the symmetry-adapted HartreeFock wave function for [A A]
corresponds to two identical VB structures, like in Scheme 8.6 but with
the same set of orbitals for the two structures. In such a case, the wave
function is stabilized by resonance, but the orbitals are not optimal for
each individual VB structure. Another solution, upon which a Hartree
Fock calculation may possibly converge, optimizes the orbitals for one VB
A B A B
11 12
Scheme 8.6. The 2-structure BOVB description of the [A B] bond.
structure (e.g. 11), to the detriment of the other (12). Then the VB structures
have different energies and 11 ends up having a larger coefficient than 12. In
this symmetry-broken solution, the resonance energy is diminished relative
to the symmetry-adapted solution. Here is the dilemma: at the HartreeFock
level, one cannot have simultaneously good orbitals and full resonance
energy [32]. A classical remedy consists of imposing the symmetry and
doing CI. However, in many cases there is only quasi-symmetry (e.g. in
[A B] with A = B), and in such a case there is no way to avoid the
artificial favoring of one structure over the other. As a consequence, it is
then very difficult to correct the initial deficiency by subsequent CI.
While the problem is currently solved in the MO framework with
elaborate methods such as coupled-cluster calculations using Brueckner
orbitals [33], the symmetry-breaking artefact vanishes at the BOVB level.
Indeed, as this method provides a superposition of two VB structures each
having its optimal set of orbitals, the BOVB wave function involves at the
same time both optimal orbitals and full resonance effect at any molecular
geometry, and the root cause for the symmetry-breaking disappears. It
follows that the BOVB method is, by nature, free from the symmetry-
breaking artefact. Historically, the first calculation of that kind was done by
Jackels and Davidson in 1976 for the NO2 radical [34]. A standard BOVB
calculation was later performed for the potential surface of the HOOH
anion [35].
8.2.3. Calculations of diabatic energy curves along a reaction

coordinate
8.2.3.1. General model
There are two fundamental questions that any model of chemical reactivity
would have to answer: What are the origins of the barriers? And what are
the factors that determine reaction mechanisms? Since chemical reactivity
involves bond breaking and making, VB theory with its focus on the bond
as the key constituent of the wave function, is able to provide a lucid model
that answers these two questions in a unified manner. The centerpiece of
R* P*
Gr
Gp
r B
p
E
R Erp
P
Reaction Coordinate
Fig. 8.1. VBSCD for a general reaction R P. R and P are ground states of
reactants and products, R* and P* are promoted excited states.
the VB model is the VB state correlation diagram (VBSCD), displayed in

Fig. 8.1 [36], which traces the energy of the VB configurations along the
reaction coordinate, and provides a mechanism for the barrier formation
and generation of a transition state in an elementary reaction.
This diagram applies to elementary reactions wherein the barrier can
be described as the interplay of two major VB states, that of the reactants
and that of the products. It displays the ground state energy profile for the
reacting system (bold curve), as well as the energy profiles for individual
VB states (thinner curves); these latter curves are also called sometimes
diabatic curves, while the full state energy curve (in bold) is called adia-
batic. Thus, starting from the reactant geometry on the left, theVB structure
r that represents the reactants electronic state, R, has the lowest energy
and it merges with the ground state. Then, as one deforms the reacting
molecules towards the product geometry, r gradually rises and finally
reaches an excited state P* that represents the VB structure of the reactants
in the product geometry. A similar diabatic curve can be traced from P,
the VB structure of the products in its optimal geometry, to R*, the same
VB structure in the reactant geometry. Consequently, the two curves cross
somewhere in the middle of the diagram. At each point of the diagram, the
adiabatic ground state of the system (bold curve) is generated by mixing
of the individual VB states. This mixing is stabilizing, so that the ground
state is stabilized by a resonance energy term, labeled as B, in the region
of the crossing point of the diabatic curves, which corresponds to the tran-
sition state. The barrier is thus interpreted as arising from avoided crossing
between two diabatic curves, which represent the energy profiles of the VB
state curves of the reactants and products.
The nature of the R* and P* promoted states depends on the reaction
type and will be specified below using a few examples. In all cases, the
promoted state R* is the electronic image of P in the geometry of R,
while P* is the image of R at the geometry of P. The G terms are the
corresponding promotion energy gaps, B is the resonance energy of the
transition state (TS), E= is the energy barrier, and Erp is the reaction
energy. The simplest expression for the barrier is given by Eq. (8.14):
E= = f Gr B (8.14)
Here, the term fGr is the height of the crossing point, expressed as some
fraction (f ) of the promotion gap at the reactant side (Gr ).
A more explicit expression is Eq. (8.15):
E= f0 G0 + 0.5 Erp B;

G0 = 0.5(Gr + Gp ), f0 = (fr + fp ) (8.15)
which shows the effects of the two promotion gaps and f factors through
their average quantities, G0 and f0 .
Equation (8.15) expresses the barrier as a balance of the contributions of
an intrinsic term, f0 G0 B and a driving force term, 0.5 Erp . The model
is general and has been described in detail before [3,4,37], and applied to a
large number of reactions of different types. Here we will briefly summarize
some VB computational applications on hydrogen abstraction reactions and
various SN 2 reactions.
8.2.3.2. Application to hydrogen-abstraction reactions

Consider a general hydrogen abstraction reaction that involves cleavage
of a bond HY by a radical X (X, Y = a univalent atom or a molecular
fragment):
X + HY XH + Y (8.16)
Practically, r is a linear combination of covalent and ionic forms that

contribute to the Lewis structure X + H Y, as shown below:
r = C1 (X + H Y) + C2 (X + H+ : Y ) + C3 (X + H : Y+ )
(8.17)
This combination is maintained in r from R to P* throughout the

reaction coordinate, while the coefficients of the contributing structures
change and adapt themselves to the geometric change (e.g. at infinite HY
distance C1 = 1). The curve p , which stretches between P and R* is
defined in an analogous manner.
Since the promoted state R* is the VB structure of P in the geometry of
R, its electronic state is illustrated by Eq. (8.18):
R* = (X H) . . . Y (8.18)
where the HY bond is infinitely long, while the X radical (spin-up)
experiences some Pauli repulsion with the electron of H, which is 5050%
spin-up and spin-down. Thus, the XH fragment has a wave function which
is 75% triplet and 25% singlet, and hence the promotion gap, from the
ground state to R*, is 3/4 of the corresponding singlet-to-triplet excitation
of the XH bond.
In the case of identity reactions (X = Y), it has been demonstrated that
the promotion energy Gr required to go from R to R* is indeed proportional
to the singlet-triplet gap of the X-H bond [31,37], which in turn is propor-
tional to the XH bond energy D(XH). Actually, systematic VB ab initio
calculations by the VBCI method have shown that, to a good approximation,
Gr can be expressed as follows [31]:
Gr 2D (XH) (8.19)
In the general case of non-identity reaction, the XH andYH bond strengths
are different, one being the weakest (DW ) and the other the strongest (DS ).
VB calculations for a panel of 14 reactions showed that B is approximately
half of the weakest bonding energy, DW , or in other words, of the bond
energy of the bond that is broken in the reactants of the exothermic direction
of the reaction, while G0 in Eq. (8.15) is given by the sum of both bonding
energies [31c]:
B = 0.5DW ; G0 = DW + DS (8.20)
Moreover, the f factor appears to be relatively constant in both identity
and non-identity reactions. Thus, by taking f0 1/3 as in identity reactions
(accurate VB calculations yield f0 = 0.32-0.36) [31c], one gets the very
simple Eq. (8.21):
E= = K (DS 0.5DW ) + 0.5 Erp ; K 1/3 (8.21)
Equation (8.21) is valid for non-identity as well as identity reactions,
in which case DS = DW . The so-calculated barriers were shown to match
fairly well the corresponding CCSD(T) barriers for a series of identity
45 E (eq. 21)
40
35
30
R2 = 0.974
25
20
15
10
5 10 15 20 25 30 35 40 45
E (VB ab initio)
Fig. 8.2. VBSCD-derived barriers plotted against ab initio VB calculated barriers.

Energies in kcal/mol.
abstraction reactions (X = Y = H, CH3 , SiH3 , GeH3 , SnH3 , PbH3 ) [31a],

with an average deviation of 2.1 kcal/mol and a maximum deviation of 4.8
kcal/mol. While the limitations of this expression have been discussed in
detail, e.g. in case where the TS is not co-linear [3], still Eq. (8.21) yields
good orders of magnitudes and correctly reproduces the trends in the series.
For non-identity reactions (X = Y), Fig. 8.2 displays a good correlation
of the barriers calculated through Eq. (8.21) plotted against the VBCISD
calculations. Thus, it appears that the VBSCD model is able to assess semi-
quantitatively barriers for H-abstraction reactions in terms of the dissoci-
ation energies of the bonds of reactants and products.
Recent applications of Eq. (8.21) to the reactivity of the enzyme
cytochrome P450 in alkane hydroxylation shows that a good correlation
with DFT-computed barriers is achieved with Eq. (8.21) using a constant
B value, which is very close to 0.5DW [38].
8.2.4. Quantitative evaluation of common chemical paradigms

8.2.4.1. Direct calculation of resonance energies and hypercon-
jugation energies
Estimating resonance energies (REs) is a simple matter in the framework
of VB theory. The RE is a measure of the magnitude of contributions from
resonance structures other than the principal Lewis structure to the ground
state of a conjugated molecule.
Starting from the description of an electronic system in terms of inter-

acting VB structures (Eq. (8.8)), if we assume that a given Lewis structure
L is the most stable (i.e. has the greatest weight) among all resonance
contributors, the RE would be
RE = E(L ) E(VB ) (8.22)
in which VB is the fully delocalized ground state, and L is the reference
Lewis structure (which, depending on the reference state, may be repre-
sented by a single VB structure or by a group of VB structures). Equation
(8.22) has been used with ab initio VB calculations to calculate the res-
onance energies of benzene [39], butadiene [40], allyl radical and ions
[41, 42], transition states of organic reactions [3, 29, 31, 4346], and so on,
and to quantify the -aromaticity of cyclopropane [47]. Equation (8.22) has
also been used to calculate the resonance energy arising from the mixing
of covalent and ionic VB structures in a bond, leading to the discovery of
a new type of chemical bonding (see next subsection) [22, 48, 49].
Another technique for calculating resonance energies or delocalization
energies consists of defining the reference Lewis structure, by a so-called
block-localized wave function (BLW), in which the orbitals are doubly
occupied but optimized with some localization constraints. Thus, the
orbitals that represent e.g. a -bond in a conjugated system can be optimized
while being constrained to be strictly localized on the two bonded atoms,
and the orbitals that represent a lone pair are localized on a single atom. The
orbital optimization can be carried out at the HartreeFock level [50], but a
recent version using orbital optimization at the DFT level also exists [51].
The resonance energy is then calculated as the difference between the BLW
wave function representing the reference Lewis structure and the fully delo-
calized wave function of the ground state. More generally, the BLW method
can be used to calculate delocalization energies by defining a diabatic state
in which delocalization is turned off. In this latter state, the molecule or
interacting system is partitioned into subgroups, and each localized MO
is expanded in terms of basis functions belonging to only one subgroup.
As the BLW method involves optimization of non-orthogonal orbitals, and
since the BLW wave function represents a Lewis structure, the BLW tech-
nique can be considered as belonging to the VB family, actually the simplest
VB-variant.
The above BLW method has been used to calculate the resonance
energies of many organic molecules. For example, it has been used to
quantify the role of resonance in the rotational barriers of amides [52], and
in the acidities of carboxylic acids and enols as compared to alcohols [53].
It was also used to provide accurate estimations of the vertical and adiabatic
resonance energies of benzene [54], allyl radical and anions [55], and so on.
Calculations of delocalization energies by VBSCF or BLW methods
have also been used to get accurate estimates of the magnitudes of hyper-
conjugation. This has been applied to trace the origin of Saytzeffs rule
[56], the role of hyperconjugation in the rotational barrier of ethane [57]
or in the exceptional short bond length of tetrahedranyl-tetrahedrane [58],
and so on.
8.2.4.2. Characterization of a novel type of bonding:

Charge-shift bonds
The resonance energy arising from the mixing of covalent and ionic VB
structures in a bond AB can be calculated by means of Eq. (8.22), in
which L is the pure covalent structure (first term in Eq. (8.2)) while VB
is the full VB wave function involving both covalent and ionic terms (all
three terms in Eq. (8.2)). Such calculations were done at the BOVB and
VBCI levels [22, 48], and the resonance energy, referred to as covalent-
ionic resonance energy (RECI ), was determined. Thus, RECI was found to
be a minor component of the bonding energy in the two classical families
of covalent and ionic bonds, which are mainly stabilized by the purely
covalent interactions in the first case, and by purely electrostatic interac-
tions in the second. However, alongside these two classical types, a third
category of bonds appeared, in which RECI is the major component of the
bonding energy, even in some homonuclear bonds. This type of bonding was
called charge-shift bonding, because the electron-pair fluctuation plays
the dominant role in the bonding mechanism [22, 48]. An extreme case of
charge-shift bonding is the F2 molecule, in which the covalent repulsion is
repulsive at all distances. More generally, the charge-shift bonding territory
involves homopolar bonds of compact electronegative and/or lone-pair-rich
elements, heteropolar bonds of these elements among themselves and with
other atoms (for example, the metalloids, such as silicon and germanium),
hypercoordinated molecules, and bonds whose covalent components are
weakened by exchange-repulsion strain (as in [1.1.1] propellane) [49], no-
density bonds [22, 48d], and so on.
Charge-shift bonding has experimental consequences, such as the bar-
riers for halogen-transfer reactions having much larger barriers than the
corresponding hydrogen-transfer processes [48b], the rarity of silicenium
ions in condensed phases [59], the surprising strength of inverted bonds in
the series of propellanes molecules [49].
A related bonding mechanism was discovered in maximum spin clusters

devoid of any electron pairs between the constituent atoms, and nevertheless
quite strongly bound. Thus, for example, the 3 + 3
u state of Cu2 is bonded
despite of the fact that its and * orbitals are singly occupied and hence
the bond order of the two atoms is zero [60]. As the cluster grows, this no-
pair bonding increases and reaches 1819 kcal/mol per atom. The bonding
arises from covalent-ionic mixing of triplet configurations and constitutes
a mechanism of bound triplet pairs.
8.2.4.3. vs driving force for the D6h geometry of benzene

The regular hexagonal structure of benzene can be considered as a stable
intermediate in a reaction that interchanges two distorted Kekule type
isomers, each displaying alternating CC bond lengths as shown in Fig. 8.3.
It is well known that the D6h geometry of benzene is stable against a
Kekulean distortion (of b2u symmetry), but one may still wonder which
one of the two sets of bonds, or , is responsible for this resistance to a
b2u distortion. The frame, which is just a set of identical single bonds,
prefers by nature a regular geometry with equal CC bond lengths. It is
not obvious whether the electronic component has, by itself, the same
K1* K2*
K1* K2*
1B
2u()
1B
2u
K1 K2
1A ()
1g
K1 K2 1A
1g
RC
RC
(a) (b)
Fig. 8.3. VBSCDs showing the crossing and avoided crossing of the Kekule
structures of benzene along the bond alternating mode, b2u for: (a) -only curves,
(b) full + curves.
tendency and hence contributes to the D6h geometry of benzene. Could

the electronic component be actually distortive, but overwhelmed by the
propensity of the frame to maintain a D6h geometry? To answer this
question, consider in Fig. 8.3 the VBSCD that represents the interchange
of Kekule structures along the b2u reaction coordinate; the middle of the
b2u coordinate corresponds to the D6h structure, while its two extremes
correspond to the bond-alternated mirror image Kekule geometries.
Part (a) of the figure considers energies only. Starting from the left-
hand side, Kekule structure K1 correlates to the excited state K2 * in which
the bonds of the initial K1 structure are elongated, while the repulsive
non-bonding interactions between the bonds are reinforced. The same
argument applies if we start from the right-hand side, with structure K2
and follow it along the b2u coordinate; K2 will then rise and correlate
to K1 *. To get an estimate for the gap, we can extrapolate the Kekule
geometries to a complete distortion, in which the bonds of K1 andK2
would be completely separated (which in practice is prevented by the
frame that limits the distortion). At this asymptote the promotion energy,
Ki Ki (i = 1,2), is due to the unpairing of three bonds in the ground
state, Ki , and replacing these bonds by three non-bonding interactions, in
Ki *. According to qualitative VB theory [3], the latter are repulsive by a
quantity that amounts to half the size of a triplet repulsion. The fact that
such a distortion can never be reached is of no concern. What matters is that
this constitutes an asymptotic estimate of the energy gap G that correlates
the two Kekule structures, and that eventually determines if their mixing
results in a barrier or in a stable situation, in the style of the VBSCD above.
According to the VB rules [3], G is given by Eq. (8.23):
G(K K * ) = 3[0.75 EST (C = C)] = 9/4 EST (C = C) (8.23)
Since the EST value for an isolated bond is well over 100 kcal/mol,
Eq. (8.23) places the electronic system in the region of large gaps. Con-
sequently, the -component of benzene is predicted by the VBSCD model
to be an unstable transition state, 1A1g (), as illustrated in Fig. 8.3(a). This
-transition state prefers a distorted Kekulean geometry with bond alter-
nation, but is forced by the frame, with its strong symmetrizing driving
force, to adopt the regular D6h geometry. This prediction, which was derived
at the time based on qualitative considerations of G in the VBSCD of iso-
electronic series [61], was later confirmed by a variety of rigorous ab initio
separation methods [62]. The prediction was further linked [63] to
experimental data associated with the vibrational frequencies of the excited
states of benzene.
The spectroscopic experiments [63] show a peculiar phenomenon. This

phenomenon is both state specific, to the 1 B2u excited state, as well as
vibrational mode specific, to the bond-alternating mode, i.e. the Kekule
mode b2u . Thus, upon excitation from the 1A1g ground state to the 1 B2u
excited state, with exception of b2u all other vibrational modes behave
normally and undergo frequency lowering in the excited state, as expected
from the decrease in -bonding and disruption of aromaticity following a
* excitation. By contrast, the Kekule b2u mode undergoes an upward
shift of 257261 cm1 . As explained below, this phenomenon is predictable
from the VBSCD model and constitutes a critical test of distortivity in
the ground state of benzene. A simple vibronic coupling mechanism cannot
account for this mode specificity and state specificity.
Indeed, the VBSCD model is able to make predictions not only for the
ground state of an electronic system, but also for a selected excited state.
Thus, the mixing of the two Kekule structures K1 and K2 in Fig. 8.3(a) leads
to a pair of resonant and antiresonant states K1 K2 ; the1A1g ground state
K1 + K2 is the resonance-stabilized combination, and the 1 B2u excited
state K1 K2 is the antiresonant mixture (this is the first excited state
of benzene [64]). In fact, the VBSCD in Fig. 8.3(a) predicts that the cur-
vature of the 1A1g () ground state (restricted to the electronic system)
is negative, whereas by contrast, that of the 1 B2u () state is positive. Of
course, when the energy of the frame is added as shown in Fig. 8.3(b),
the net total driving force for the ground state becomes symmetrizing, with
a small positive curvature. By comparison, the 1 B2u excited state displays
now a steeper curve and is much more symmetrizing than the ground state,
having more positive curvature. As such, the VBSCD model predicts that
the 1A1g 1 B2u excitation of benzene should result in the reinforcement
of the symmetrizing driving force, which will be manifested as a frequency
increase of the Kekulean b2u mode.
In order to show how delicate the balance is between the and
opposing tendencies, we recently [65] derived an empirical equation,
Eq. (8.24), for 4n+2 annulenes:
E+ = 5.0(2n + 1) 5.4(2n), kcal/mol (8.24)
Here E+ stands for the total ( and ) distortion energies, the terms
5.0(2n+1) represent the resisting effect, which is 5.0 kcal/mol for an
adjacent pair of -bond, whereas the negative term, 5.4(2n), accounts
for the -distortivity. This expression predicts that for n = 7, namely the
C30 H30 annulene, the E+ becomes negative and the annulene undergoes
bond localization. If we increase the -distortivity coefficient by just a tiny
bit, namely to Eq. (8.25),

E+ = 5.0(2n + 1) 6.0(2n), kcal/mol (8.25)
the equation would predict now that already the annulene with n = 3,
namely C14 H14 , will undergo bond localization. This extreme sensitivity,
which is predicted to manifest in computations and experimental data
of annulenes, is a simple outcome of the VBSCD prediction that the
-component of these species behaves as a transition state with a propensity
towards bond-localization.
8.3. Present Capabilities and Expected Improvements
8.3.1. Evaluation of Hamiltonian matrix elements

Since the early days of VB theory, much progress has been made in
improving the algorithms and speeding up the evaluation of the Hamil-
tonian matrix elements between VB structures. Efficient algorithms that
deal with the non-orthogonality problem (see Box 8.2) have been imple-
mented [11], such that the calculation of Hamiltonian matrix elements
between non-orthogonal determinants scales as N 4 . Among the method-
ological contributions that brought VB to the modern era, one may cite
the Prosser and Hagstrom method of evaluating matrix elements [66], the
generalized SlaterCondon rules [67], the spin free VB method [68], the
left coset decomposition algorithm [69], the algebrant algorithm of Li et
al. [70], and the paired-permanent-determinant approach [71].
8.3.2. Direct VBSCF/BOVB algorithm

In the conventional VBSCF or BOVB calculations, the Hamiltonian matrix
elements are expressed explicitly in form of orbital integrals, and thus,
an integral transformation from basis functions to VB orbitals is required
for each iteration, resulting thereby in costly demands for both CPU
time and storage space. Furthermore, as discussed above, the conventional
VBSCF/BOVB method applies the super-CI method to optimize orbitals or
uses approximate NewtonRaphson type methods [11b], where the energy
gradients are approximately obtained. Both of the two algorithms require
costly computational efforts. To reduce the cost in the VBSCF/BOVB cal-
culations, Wu et al. developed an algorithm for evaluating analytical energy
gradients [72]. Most importantly, the formulas in this algorithm are orbital-
free, and thus may be performed in integral direct mode easily, without
integral transformation procedure. A further advantage of this algorithm is

that the computational cost mainly depends on the number of basis func-
tions, rather than the number of orbitals. This latter feature is fundamental
for speeding up calculations of BOVB type, which usually deal with many
more orbitals than VBSCF calculations, since BOVB wave functions allow
each VB structure to have its own orbital set. Thus, with the direct algo-
rithm, the cost of one BOVB iteration is almost the same as that of a VBSCF
iteration [72].
8.3.3. Current calculations of medium-sized molecular systems

With the advances in VB methods, made in the last two decades, ab
initio VB theory is now capable not only of accurate calculations of small
molecules, by using BOVB and VBCI, but also of VB calculations of
medium-size molecules, or even transition metal complexes. Just recently,
Wu et al. [73] have applied the VBSCF method to a study of the Diels
Alder reaction, and to the rotational barrier for the organometallic complex
(CO)4 Fe(CH2 =CH2 ). In these studies, a 6-31G* basis set was used for
atoms of the first and second rows of the periodic table, and a standard
Lanl2dz basis set was used for the Fe atom. All electrons except for the
ECP electrons of Fe were included in the VB calculation, summing up to
88 electrons in 46 VB orbitals in the case of the organometallic complexes.
In both DielsAlder and organometallic complex studies, the VBSCF
accuracy was found to be equivalent to the corresponding CASSCF calcu-
lations. Better accuracy can be achieved with higher levels of VB methods,
especially VBPT2, which is a very cheap post-VBSCF method. A new
version of VBPT2 that applies contraction technique of active space is in
progress, and will hopefully provide the long sought for powerful tool for
VB studies of many problems involving medium-sized molecules.
8.3.4. Mixed Valence Bond Quantum Monte Carlo methods

Quantum Monte Carlo (QMC) methods [74] offer interesting alternatives
to basis set ab initio methods. In QMC, the trial wave function is typically
constructed as the product of a determinantal expansion and a correlation
function, the latter containing explicit interparticle distance terms. The Dif-
fusion Monte Carlo (DMC) method belongs to the QMC family and is able
to provide extremely accurate BDEs, provided the trial function already
includes a good deal of electron correlation. Thus, truncated MCSCF wave
functions were used as trial wave functions but proved disappointing [75].
To get better DMC trial functions, several workers had the idea to use
trial wave functions related to VB theory [76, 77], expecting two advan-
tages: (i) for the same amount of electron correlation being taken into
account, VB expansions can be more compact than MCSCF ones; (ii)
because VB orbitals are generally localized on one or two centers, a VB-
based trial wave-function could be cheaper than a trial wave function
of the same expansion length based on MOs delocalized over the entire
molecule.
Lester et al. [77] calculated the BDE for the acetylenic CH bond by
performing a DMC calculation using a trial BOVB wave function with a
polarized triple- basis set of Slater orbitals. The accuracy is excellent,
with a CH BDE of 132.4 0.9 kcal/mol, practically equivalent to the
recommended experimental value of 132.8 0.7 kcal/mol. These values
are to be compared with DMC results obtained with single determinant
trial wave functions, using HartreeFock orbitals (137.5 0.5 kcal/mol)
and local spin density (LDA) KohnSham orbitals (135.6 0.5 kcal/mol).
Very recently, Goddard et al. [78] used simple GVB wave functions
as guess functions for DMC calculations, and applied this approach to the
adiabatic singlet-triplet splitting in methylene, the vertical and adiabatic
singlet-triplet splitting in ethylene, 2 + 2 cycloaddition, and Be2 bond
breaking. In all these cases, this approach was accurate within a few tenths
of a kcal/mol. Less accurate results, however, were found for the very
difficult test case of the NV transition energy of ethylene, for which
dynamic correlation is crucially important.
With the very recent (and actually on-going) progress of QMC algo-
rithms, trial wave functions of VBSCF quality yield results as accurate as
former BOVB trial functions. This improvement should allow quite large
systems to be treated by VB-QMC methods, with possibly up to 100 VB
structures, with an accuracy close to experimental error bars. Moreover,
work is in progress to use QMC methods to perform calculation of VB
type, allowing calculations of the weights of VB structures, and calcula-
tions of individual VB structures, i.e. diabatic states [79].
Last but not least, a further advantage of QMC methods, and soon of
VB-QMC methods, is that QMC algorithms scale as N 3 (N being the
number of electron), and possibly as N 2 in the near future [79].
8.3.5. Prospective
The question that may be posed at this point is whether VB theory will
ever return as a mainstream method that will be used by chemists as a
computational tool and/or as a conceptual framework? Let us try to discuss

these prospects by separating the two aspects; the future computational
capabilities, and the conceptual impact.
As shown above, VB methodology is constantly developing, and surely
in the not too far future we may expect to see computations of real reac-
tions like the DielsAlder reaction, or of organometallic species becoming
more and more common. In this respect the VBPT2 method shows a great
promise to be a future standard method, with the cost of VBSCF, but with
accuracy competing with CASPT2.
One of the present shortcomings of VBPT2 is the treatment of spin-state
of transition metal complexes in e.g. bioinorganic chemistry. At present,
the calculations of the nonheme oxo-iron complex, (NH3 )5 FeO2+ , indi-
cates that the method overestimates the stability of high-spin states [80].
However, spin-state ordering seems to be one of the most difficult hurdles
for computational chemistry, and even a large CASSCF(20,13) followed
by spectroscopy-oriented configuration interaction (SORCI) calculations
overestimate the stability of the high-spin quintet state of this reagent [81].
CASPT2 performs relatively well [82] by using zero order Hamiltonians
with energy-shifted orbital energies, a very large basis set, and a double-
shell for the occupied 3d orbitals of iron. This latter effect, which correlates
the non-bonding doubly occupied 3d orbitals, is analogous to the orbital
splitting used in BOVB, and can be utilized in the same manner in VBPT2,
thus making this method applicable to bioinorganic complexes. When this
is achieved, VB as a computational tool will be as versatile as the MO-based
methods, albeit never as fast as MP2.
Another future development is VBDFT, which combines the dynamic
correlation facility of DFT, with the static correlation inherent in VB. An
early version which has been tried already more than a decade ago [83],
showed some promise. However, it was already then clear that an efficient
way of implementing VBDFT method could be constructed by using den-
sities to get the energies of the VB structures, and then mixing them either
with these frozen densities or by iterating on the density during the mixing
procedure. Such a VBDFT method may combine speed and accuracy, and
may enable the calculations of reasonably large systems.
The constantly developing BLW method is now capable of incorporating
charge resonance effects within the MO-VB framework [84]. Once this
development is generalized it will add to the growing arsenal of VB-type
methods. This method will certainly be competitive with MP2.
Conceptually, VB theory is extremely versatile and applicable to various
domains in chemistry. There is no real hurdle to use VB theory as a
conceptual framework for chemistry, other than the saddening fact that for
historical reasons VB is not taught anymore in quantum chemistry courses
as a mainstream method. However, the recent monograph written on VB
theory [3], may ease the way for those who are willing to try and teach
or study elements of VB theory. Once this wall is broken, chemists will
find a beautiful theory, which can easily be incorporated into a thought
process.
8.4. Concluding Remarks
This short chapter has intended to show that ab initio VB theory has
enjoyed impressive progress during the past two or three decades. As a
result, ab initio VB algorithms are today much faster than they used to
be, and by several orders of magnitude. Moreover, modern VB has also
reasonably accurate computational methods, which can provide bonding
energies and reaction barriers with accuracies comparable to sophisticated
MO-CI methods. This has been achieved by incorporating dynamic cor-
relation effects in the VB calculations, and this without complicating the
wave functions that remain compact and easily interpretable. In addition,
the modern VB methods can be combined with a solvent model, and provide
thereby a method that can handle molecules and reactions in solution and in
proteins. Thus, from a quantitative point of view, VB theory enables today
the calculations of real chemical problems for organic molecules, as well
as molecules that contain transition metals, and all these can be done in the
gas phase or in solution. Further improvements in speed and capabilities of
VB methods are currently in progress. Especially promising is the VBPT2
method that emerges as a fast and accurate method, and the combination
of VB and QMC methods that may enable, in the future, to handle much
larger systems than those presented here.
Another aspect of VB theory that is emphasized in this chapter is insight.
Thus, despite the sophistication and accuracy of the above VB methods, all
of them rely on a compact wave function that includes a minimal number of
structures in the VB-structure set. The insight of this compact wave function
is projected by a set of applications including bonding in main group ele-
ments, quantitative evaluation of common paradigms such as resonance
energies, hyperconjugation, aromaticity and antiaromaticity in conjugated
systems, distortivity of the system of benzene and related molecules
[65], and general models of chemical reactivity. Many other applications
e.g. to photochemistry, excited states, polyradicals, etc. have appeared in a
recent monograph [3]. VB theory provides also a great deal of insights into
bonding in odd-electron systems, specifically on one-electron and three-

electron bonds [4, 24, 25, 85, 86].
In fact, VB theory gives rise to new bonding paradigms, like charge-
shift bonding, which concerns two-electron bonds which are neither
covalent nor ionic but with bonding energy that is dominated by the
covalent-ionic resonance interaction [22, 48, 49]. Another paradigm is the
ferromagnetic bonding that occurs in high-spin clusters, e.g. n+1 Lin ,
n+1 Cu , etc. that are devoid of electron pairs but still have significant bond
n
energies [60].
It is to be expected that this combination of chemical insight with
ever increasing speed and accuracy will place modern VB methods in an
important position among the ab initio methods that will be used in the
future.
8.5. Appendix A: The Myth of VB failures
One of the major alleged VB failures is associated with the dioxygen

molecule, O2 [3]. As we repeatedly stated, it is true that a nave appli-
cation of hybridization and perfect pairing approach (simple Lewis pairing)
without consideration of the important effect of four-electron repulsion
would put two electrons in one plane and four electrons in the other,
leading to the erroneous prediction of a 1 g ground state. However, proper
use of elementary VB principles, even at a simple qualitative level, shows
that a better bonding energy is obtained by forming two three-electron -
bonds, leading to the 3 g paramagnetic ground state in agreement with
experiment. In agreement with the qualitative analysis, early ab initio VB
calculations [7d] correctly provide the ordering of the ground state and
low-lying excited states of O2 .
Another alleged VB failure deals with the photoelectron spectroscopy
(PES) of CH4 . Starting from a nave application of the VB picture of CH4 ,
it follows that since methane has four equivalent localized bond orbitals
(LBOs), ergo the molecule should exhibit only one ionization peak in PES.
However, since the PES of methane shows two peaks, ergo VB theory
fails! This nave argument is obviously false, since a physically correct
representation of the CH+ 4 cation must be a linear combination of the four
VB configurations that correspond to one bond ionization, and not only
one configuration. Elementary application of symmetry point group theory
then shows that correct linear combinations are 2A1 and 2 T2 ; the latter
being a triply degenerate VB state [3]. Accordingly, VB theory predicts two
ionization peaks, in agreement with experiment.
8.6. Appendix B: Some Available VB Software Packages
Other than the GVB method that is implemented in many packages by now,
here are brief descriptions of the main VB software packages we are aware
of and with which we have had some experience to varying degrees.
8.6.1. The XMVB program

The XMVB software [87, 88] is a general program that is designed
to perform multistructure VB calculations. It can execute either non-
orthogonal CI, or non-orthogonal MCSCF calculations with simulta-
neous optimization of orbitals and coefficients of VB structures. Complete
freedom is given to the user to deal with HAOs, or orbitals of Coulson
Fischer type, so that calculations of VBSCF, SCVB, BLW, BOVB, VBCI
and VBPT2 types can be performed. The parallel version of XMVB, based
on the Message Passing Interface, is also available [87]. XMVB can be
used as a stand-alone program that is freely available from the author
(website: http://ctc.xmu.edu.cn/xmvb/). It has also been incorporated also
into GAMESS-US [89], and can be interfaced to GAUSSIAN [90].
8.6.2. The TURTLE software

TURTLE [91, 92] is also designed to perform multistructure VB calcula-
tions and can execute calculations of the VBSCF, SCVB, BLW or BOVB
types. Currently, TURTLE involves analytical gradients to optimize the
energies of individual VB structures or multistructure electronic states
with respect to the nuclear coordinates [93]. A parallel version has been
developed and implemented using the message-passing interface (MPI),
for the sake of making the software portable [94]. TURTLE is now imple-
mented in the GAMESS-UK program [95].
8.6.3. The VB2000 software

VB2000 [96, 97] is an ab initio VB package that can be used for performing
non-orthogonal CI, multi-structure VB with optimized orbitals, as well
as SCVB, GVB, VBSCF and BOVB. VB2000 can be used as a plug-in
module for GAMESS(US) [89] and Gaussian98/03 [90] so that some of
the functionalities of GAMESS and Gaussian can be used for calculating
VB wave functions. GAMESS also provides interface (option) for the access
of VB2000 module.
8.6.4. The CRUNCH software

CRUNCH (computational resource for understanding chemistry) was
originally written in Fortran by Gallup, and recently translated into C [98].
This program can perform multiconfiguration VB calculations with fixed
orbitals, plus a number of MO-based calculations like RHF, ROHF, UHF
(followed by MP2), orthogonal CI and MCSCF.
Bibliography
[1] W. Heitler, F. London, Z. Phys. 44, 55 (1927).

[2] W. Kutzelnigg, in Theoretical Models of Chemical Bonding, Part 2 edited by Z.B.
Maksic, (Springer-Verlag, Heidelberg, 1990) pp. 143.
[3] S. Shaik and P.C. Hiberty, A Chemists Guide to Valence Bond Theory (Wiley-
Interscience, New York, 2007).
[4] S. Shaik and P.C. Hiberty, in Reviews in Computational Chemistry, Vol. 20, edited
by K.B. Lipkowitz and D.B. Boyd, (VCH, New York, 1995) pp. 1100.
[5] P.C. Hiberty and C. Leforestier, J. Amer. Chem. Soc. 100, 2012 (1978).
[6] C.A. Coulson and I. Fischer, Phil. Mag. 40, 386 (1949).
[7] (a) W.A. Goddard III, Phys. Rev. 157, 81 (1967).
(b) R.C. Ladner and W.A. Goddard III, J. Chem. Phys. 51, 1073 (1969).
(c) F.W. Bobrowicz and W.A. Goddard III, in Methods of Electronic Structure Theory,
Vol. 4, edited by H.F. Schaefer III, (Plenum, New York, 1977) pp. 79126.
(d) W.A. Goddard III, T.H. Dunning Jr., W.J. Hunt, and P.J. Hay, Acc. Chem. Res. 6,
368 (1973).
[8] (a) P.B. Karadakov, J. Gerratt, D.L. Cooper, and M. Raimondi, J. Chem. Phys. 97,
7637 (1992).
(b) D.L. Cooper, J. Gerratt, M. Raimondi, M. Sironi, and T. Thorsteinsson, Theor.
Chim. Acta. 85, 261 (1993).
(c) M. Sironi, M. Raimondi, R. Martinazzo, and F.A. Gianturco, in Valence Bond
Theory, edited by D.L. Cooper, (Elsevier, Amsterdam, 2002) pp. 261277.
(d) D.L. Cooper, J. Gerratt, and M. Raimondi, Int. Rev. Phys. Chem. 7, 59 (1988).
(e) D.L. Cooper, J. Gerratt, and M. Raimondi, in Valence Bond Theory and Chemical
Structure, edited by D.J. Klein and N. Trinajstic (Elsevier, Amsterdam, 1990) p. 287.
(f) D.L. Cooper, J. Gerratt, and M. Raimondi, in Advances in the Theory of Benzenoid
Hydrocarbons, edited by I. Gutman and S.J. Cyvin, Top. Current. Chem. 153, 41
(1990).
[9] F. Penotti, D.L. Cooper, J. Gerratt, and M. Raimondi, J. Mol. Struct. (Theochem)
169, 421 (1988).
[10] (a) J.H. van Lenthe and G.G. Balint-Kurti, J. Chem. Phys. 78, 5699 (1983).
(b) J.H. van Lenthe and G.G. Balint-Kurti, Chem. Phys. Lett. 76, 138 (1980).
[11] (a) J. Verbeek and J.H. van Lenthe, J. Mol. Struct. (Theochem) 229, 115, (1991).
(b) J.H. van Lenthe, J. Verbeek, and P. Pulay, Mol. Phys. 73, 1159 (1991).
[12] L. Song, J. Song, Y. Mo, and W. Wu, J. Comput. Chem. 30, 399 (2009).
[13] (a) P.C. Hiberty, J.P. Flament, and E. Noizet, Chem. Phys. Lett. 189, 259 (1992).
(b) P.C. Hiberty, S. Humbel, J.H. van Lenthe, and C.P. Byrman, J. Chem. Phys. 101,
5969 (1994).
(c) P.C. Hiberty and S. Shaik, Theor. Chem. Acc. 108, 255 (2002).
[14] (a) W. Wu, L. Song, Z. Cao, Q. Zhang, and S. Shaik, J. Phys. Chem. A 106, 2721
(2002).
(b) L. Song, W. Wu, Q. Zhang, and S. Shaik, J. Comput. Chem. 25, 472 (2004).
[15] Z. Chen, J. Song, S. Shaik, P.C. Hiberty, and W. Wu, J. Phys. Chem. A 113, 11560
(2009).
[16] L. Song, W. Wu, Q. Zhang, and S. Shaik, J. Phys. Chem. A 108, 6017 (2004).
[17] (a) C.J. Cramer and D.G. Truhlar, Chem. Rev. 99, 2161 (1999).
(b) C.P. Kelly, C.J. Cramer, and D.G. Truhlar, J. Chem. Theory Comput. 1, 1133
(2005).
(c) A.V. Marenich, R.M. Olson, C.P. Kelly, C.J. Cramer, and D.G. Truhlar, J. Chem.
Theory Comput. 3, 2011 (2007).
[18] P. Su, W. Wu, C.P. Kelly, C.J. Cramer, and D.G. Truhlar, J. Phys. Chem. A 112, 12761
(2008).
[19] (a) A. Shurki and H.A. Crown, J. Phys. Chem. B 109, 23638 (2005).
(b) A. Sharir-Ivry, H.A. Crown, W. Wu and A. Shurki, J. Phys. Chem. A 112, 2489
(2008).
[20] The bonding energy is found negative at the HartreeFock/6-31G level,

33.4 kcal/mol at an interatomic distance of 1.43 A (see Ref. 13c).
[21] K.P. Huber and G. Herzberg, Molecular Spectra and Molecular structures. IV. Con-
stants of Diatomic Molecules (van Nostrand, Reinhold, New York, 1979).
[22] S. Shaik, D. Danovich, B. Silvi, D. Lauvergnat, and P.C. Hiberty, Chem. Eur. J. 11,
6358 (2005) (see the Supporting Information Document).
[23] M.V. Rama Krishna and K.D. Jordan, Chem. Phys. 115, 405 (1987).
[24] P.C. Hiberty, S. Humbel, D. Danovich, and S. Shaik, J. Amer. Chem. Soc. 117, 9003
(1995).
[25] P.C. Hiberty, S. Humbel, and P. Archirel, J. Phys. Chem. 98, 11697 (1994).
[26] T. Clark, J. Amer. Chem. Soc. 110, 1672 (1988).
[27] P.M.W. Gill and L. Radom, J. Amer. Chem. Soc. 110, 4931 (1988).
[28] L. Song, W. Wu, P.C. Hiberty, D. Danovich, and S. Shaik, Chem. Eur. J. 9, 4540
(2003).
[29] L. Song, W. Wu, P.C. Hiberty, and S. Shaik, Chem. Eur. J. 12, 7458 (2006).
[30] P. Su, F. Ying, W. Wu, P.C. Hiberty, and S. Shaik, S. Chem. Phys. Chem. 8, 2603
(2007).
[31] (a) S. Shaik, W. Wu, K. Dong, L. Song, and P.C. Hiberty, J. Phys. Chem. A 105, 8226
(2001).
(b) L. Song, W. Wu, K. Dong, P.C. Hiberty, and S. Shaik, J. Phys. Chem. A 106,
11361 (2002).
(c) P. Su, L. Song, W. Wu, P.C. Hiberty, and S. Shaik, J. Amer. Chem. Soc. 126, 13539
(2004).
[32] P.O. Lowdin, Rev. Mod. Phys. 35, 496 (1963).
[33] K.A. Brueckner, Phys. Rev. 96, 508 (1954).
[34] C.F. Jackels and E.R. Davidson, J. Chem. Phys. 64, 2908 (1976).
[35] S. Humbel, I. Demachy, and P.C. Hiberty, Chem. Phys. Lett. 247, 126 (1995).
[36] S.S. Shaik, J. Amer. Chem. Soc. 103, 3692 (1981).
[37] (a) S. Shaik, and P.C. Hiberty, in Adv. Quant. Chem., Vol. 26, edited by P.-O. Lowdin,
(Academic Press, 1995).
(b) A. Pross, Theoretical and Physical Principles of Organic Reactivity (Wiley-
(c) S. Shaik and A. Shurki, Angew. Chem. Int. Ed. 38, 586 (1999).
(d) S. Shaik, Phys. Chem. Chem. Phys. 12, 8706 (2010).
[38] (a) S. Shaik, D. Kumar, and S.P. de Visser, J. Amer. Chem. Soc. 130, 10128, erratum
p. 14016 (2008).
(b) S. Shaik, W. Lai, H. Chen, and Y. Wang, Acc. Chem. Res. 43, 1154 (2010). In
these papers, anf factor of 0.3 was used together with vertical bond energies, i.e. not
including the geometrical and electronic relaxations of the dissociated radicals.
[39] J.M. Norbeck and G.A. Gallup, J. Amer. Chem. Soc. 96, 3386 (1974).
[40] A.F. Voter and W.A. Goddard III, J. Amer. Chem. Soc. 108, 2830 (1986).
[41] Y.R. Mo, Z.Y. Lin, W. Wu, and Q.N. Zhang, J. Phys. Chem. 100, 6469 (1996).
[42] M. Linares, S. Humbel, and B. Braida, J. Phys. Chem. A 112, 13249 (2008).
[43] G. Sini, G. Ohanessian, P.C. Hiberty, and S.S. Shaik, J. Amer. Chem. Soc. 112, 1407
(1990).
[44] P. Matre, F. Volatron, P.C. Hiberty, and S.S. Shaik, Inorg. Chem. 29, 3047 (1990).
[45] S.S. Shaik, E. Duzy, and A. Bartuv, J. Phys. Chem. 94, 6574 (1990).
[46] S. Shaik and A.C. Reddy, J. Chem. Soc., Faraday Trans. 90, 1631 (1994).
[47] W. Wu, B. Ma, J.I.-C. Wu, P.v.R. Schleyer, and Y.R. Mo, Chem. Eur. J. 15, 9730
(2009).
[48] (a) S. Shaik, P. Matre, G. Sini, and P.C. Hiberty, J. Amer. Chem. Soc. 114, 7861
(1992).
(b) P.C. Hiberty, C. Megret, L. Song, W. Wu, and S. Shaik, J. Amer. Chem. Soc. 128,
2836 (2006).
(c) P.C. Hiberty, R. Ramozzi, L. Song, W. Wu, and S. Shaik, Faraday Discuss. 135,
261 (2007).
(d) L. Zhang, F. Ying, W. Wu, P.C. Hiberty, and S. Shaik, Chem. Eur. J. 15, 2979
(2009).
[49] (a) W. Wu, J. Gu, J. Song, S. Shaik, and P.C. Hiberty, Angew. Chem. Int. Ed. 48, 1407
(2009).
(b) S. Shaik, D. Danovich, W. Wu, and P.C. Hiberty, Nature Chem. 1, 443 (2009).
(c) S. Shaik, Z. Chen, W. Wu, A. Stanger, D. Danovich, and P.C. Hiberty, Chem.
Phys. Chem. 10, 2658 (2009).
[50] Y. Mo and S.D. Peyerimhoff, J. Chem. Phys. 109, 1687 (1998).
[51] Y. Mo, L. Song, and Y. Lin, J. Phys. Chem. A 111, 8291 (2007).
[52] D. Lauvergnat and P.C. Hiberty, J. Amer. Chem. Soc. 119, 9478 (1997).
[53] P.C. Hiberty and C.P. Byrman, J. Amer. Chem. Soc. 117, 9875 (1995).
[54] (a) Y. Mo, J. Phys. Chem. A 113, 5163 (2009).
(b) Y. Mo and P.v.R. Schleyer, Chem. Eur. J. 12, 2009 (2006).
[55] Y.R. Mo, L.C. Song, and Y.C. Lin, J. Phys. Chem. A 111, 8291 (2007).
[56] B. Braida, V. Prana, and P.C. Hiberty, Angew. Chem. Int. Ed. 48, 5724 (2009).
[57] (a) Y. Mo, W. Wu, L. Song, M. Lin, Q. Zhang, and J. Gao, Angew. Chem. Int. Ed. 43,
1986 (2004).
(b) Y. Mo and J. Gao, Acc. Chem. Res. 40, 113 (2007).
[58] Y. Mo, Org. Lett. 8, 535 (2006).
[59] P. Su, L. Song, W. Wu, S. Shaik, and P.C. Hiberty, J. Phys. Chem. A 112, 2988
(2008).
[60] (a) S.P. de Visser, D. Danovich, W. Wu, and S. Shaik, J. Phys. Chem. A 106, 4961
(2002).
(b) D. Danovich, W. Wu, and S. Shaik, J. Amer. Chem. Soc. 121, 3165 (1999).
(c) D. Danovich and S. Shaik, J. Chem. Theory Comput. 6, 1479 (2010).
[61] S.S. Shaik and R. Bar, Nouv. J. Chim. 8, 411 (1984).
[62] (a) S.S. Shaik, P.C. Hiberty, J.-M. Lefour, and G. Ohanessian, J. Amer. Chem. Soc.
109, 363 (1987).
(b) S.S. Shaik, P.C. Hiberty, G. Ohanessian, and J.-M. Lefour, J. Phys. Chem. 92,
5086 (1988).
(c) P.C. Hiberty, D. Danovich, A. Shurki, and S. Shaik, J. Amer. Chem. Soc. 117,
7760 (1995).
[63] Y. Haas and S. Zilberg, J. Amer. Chem. Soc. 117, 5387 (1995).
[64] E.C. da Silva, J. Gerratt, D.L. Cooper, and M. Raimondi, J. Chem. Phys. 101, 3866
(1994).
[65] S. Shaik, A. Shurki, D. Danovich, and P.C. Hiberty, Chem. Rev. 101, 1501 (2001).
[66] F. Prosser and S. Hagstrom, Int. J. Quant. Chem. 2, 89 (1968).
[67] J. Verbeek and J.H. van Lenthe, Int. J. Quant. Chem. 40, 201 (1991).
[68] (a) X. Li and Q. Zhang, Int. J. Quant. Chem. 36, 599 (1989).
(b) R. McWeeny, Int. J. Quant. Chem. 34, 25 (1988).
[69] (a) J. Li and W. Wu, Theor. Chim. Acta 89, 105 (1994).
(b) J. Li, Theor. Chim. Acta 93, 35 (1996).
[70] (a) J. Li and R. Pauncz, Int. J. Quantum Chem. 62, 245 (1997).
(b) J. Li, Theor. Chim. Acta 93, 35 (1996).
(c) J. Li, J. Math. Chem. 17, 295 (1995).
(d) J. Li and W. Wu, Theor. Chim. Acta 89, 105 (1994).
[71] (a) W. Wu, A. Wu, Y. Mo, and Q. Zhang, Science in China (English Ed.) B39, 35
(1996).
(b) W. Wu, A. Wu, Y. Mo, M. Lin, and Q. Zhang, Int. J. Quant. Chem. 67, 287 (1998).
[72] L. Song, J. Song, Y. Mo, and W. Wu, J. Comput. Chem. 30, 399 (2009).
[73] W. Wu, personal communication.
[74] (a) B.L. Hammond, W.A. Lester Jr., and P.J. Reynolds, World Scientific Lecture and
Course Notes in Chemistry, Vol. 1 (World Scientific, Singapore, 1994).
(b) J.B. Anderson, in Reviews in Computational Chemistry, Vol. 13, edited by K.B.
Lipkowitz and D.B. Boyd, (John Wiley & Sons, New York, 1999) pp. 133182.
(c) D.M. Ceperley and L. Mitas, in New Methods in Computational Quantum
Mechanics, Vol. 93, edited by I. Prigogine and S.A. Rice, (John Wiley & Sons,
New York, 1996) pp. 138.
(d) W.M.C. Foulkes, L. Mitas, R.J. Needs, and G. Rajagopal, Rev. Mod. Phys. 73, 33
(2001).
[75] R.N. Barnett, Z. Sun, and W.A. Lester Jr., J. Chem. Phys. 114, 2013 (2001).
[76] (a) M. Casula, C. Attaccalite, and S. Sorella, J. Chem. Phys. 121, 7110 (2004).
(b) S. Sorella, M. Casula, and D. Rocca, J. Chem. Phys. 127, 014105 (2007).
[77] D. Domin, B. Braida, and W.A. Lester Jr., J. Phys. Chem. A 112, 8964 (2008).
[78] A.G. Anderson and W.A. Goddard III, J. Chem. Phys. 132, 164110 (2010).
[79] B. Braida, personal communication.
[80] W. Wu and S. Shaik, unpublished results.

[81] F. Neese, J. Inorg. Biochem. 100, 716 (2006).
[82] H. Chen, J. Song, W.Z. Lai, W. Wu, and S. Shaik, J. Chem. Theory Comput. 6, 940
(2010).
[83] W. Wu and S. Shaik, Chem. Phys. Lett. 301, 37 (1999).
[84] J. Gao, A. Cembran, and Y. Mo, J. Chem. Theory Comput. 6, 2402 (2010).
[85] B. Braida and P.C. Hiberty, J. Phys. Chem. A 104, 4628 (2000).
[86] B. Braida, D. Lauvergnat, and P.C. Hiberty, J. Chem. Phys. 115, 90 (2001).
[87] L. Song, Y. Mo, Q. Zhang, and W. Wu, J. Comput. Chem. 26, 514 (2005).
[88] L. Song, Y. Mo, Q. Zhang, and W. Wu, XMVB: An ab initio Non-orthogonal Valence
Bond Program (Xiamen University, Xiamen, 2003).
[89] M.W. Schmidt, K.K. Baldridge, J.A. Boatz, S.T. Elbert, M.S. Gordon, J.H. Jensen,
S. Koseki, N. Matsunaga, K.A. Nguyen, S. Su, T.L. Windus, M. Dupuis, and
J.A. Montgomery Jr., J. Comput. Chem. 14, 1347 (1993). http://www.msg.ameslab.
gov/GAMESS/GAMESS.html.
[90] M.J. Frisch, et al., Gaussian 03, Revision D.01, (Gaussian, Inc. Wallingford CT,
2004). http://www.gaussian.com/
[91] J. Verbeek, J.H. Langenberg, C.P. Byrman, F. Dijkstra, and J.H. van Lenthe,
TURTLE-A gradient VB/VBSCF program (19982004), (Theoretical Chemistry
Group, Utrecht University, Utrecht).
[92] J.H. van Lenthe, F. Dijkstra, and R.W.A. Havenith, in Theoretical and Computational
Chemistry, Vol. 10., edited by D.L. Cooper, (Elsevier, Amsterdam, 2002) pp. 79116.
[93] F. Dijkstra and J.H. van Lenthe, J. Chem. Phys. 113, 2100 (2000).
[94] D. Fokke and J.H. van Lenthe, J. Comput. Chem. 22, 665 (2001).
[95] GAMESS-UK is a package of ab initio programs written by M.F. Guest, J.H.
van Lenthe, J. Kendrick, K. Schoffel, P. Sherwood, R.J. Harrison, with contri-
butions from R.D. Amos, R.J. Buenker, M. Dupuis, N.C. Handy, I.H. Hillier,
P.J. Knowles, V. Bonacic-Koutecky, W. von Niessen, V.R. Saunders, A. Stone.
The package is derived from the original GAMESS code due to M. Dupuis,
D. Spangler, J. Wendoloski, NRCC Software Catalog, Vol. 1, Program No. QG01
(GAMESS), 1980. See: M.F. Guest, I.J. Bush, H.J.J. van Dam, P. Sherwood, J.M.H.
Thomas, J.H. van Lenthe, R.W.A. Havenith, J. Kendrick, Mol. Phys. 2005, 103, 719.
http://www.cfs.dl.ac.uk/gamess-uk/index.shtml.
[96] J. Li and R. McWeeny, Int. J. Quant. Chem. 89, 216 (2002).
[97] J. Li, B. Duke, and R. McWeeny, VB2000, Version 2.0, (SciNet Technologies, San
Diego, 2007). http://www.scinetec.com/
[98] G.A. Gallup, Valence Bond Methods (Cambridge University Press, Cambridge,
2002). http://www.unl.edu/ggallup/CRUNCH/.

Chapter 9
Quantum Monte Carlo Approaches

for Tackling Electronic Correlation
Massimo Mellaa,b and Gabriele Morosib

aSchool of Chemistry, Cardiff University, Main Building,
Park Place, Cardiff CF10 3AT (UK)
bDipartimento di Scienze Chimiche ed Ambientali,
Universita dellInsubria, via Lucini 3, 22100 Como (I)
Quantum Monte Carlo (QMC) methods sample the wave function, in

principle the exact one, instead of optimizing analytical functions as
standard ab initio approaches. They have emerged as suitable alternative to
ab initio methods to deal with the dynamical correlation in the description
of the electronic structure of atoms and molecules.
Differently from standard quantum chemistry approaches, QMC
enjoys several features that make it the method of choice for complicated
systems. Among these features, it is important to notice its intrinsically
parallelizable nature, the slow (N 3 ) scaling of the computational cost
with number of particles N, the limited amount of memory required and
its ability to deal with substantially different systems within the same
theoretical/algorithmic structure (e.g. it can be easily applied to both
bosons and fermions, to the description of systems containing electrons
and positrons as well as of vibrational properties of molecules).
One of the strongest points characterizing QMC is the fact that it
may use any kind of basis sets, albeit uncommon ones, depending on the
species. This fact allows QMC either to quickly converge to the exact
answer in the case of bosonic systems, or to easily recover 9095% of the
correlation energy in electronic species without the inverse cubic conver-
gence with respect to the size of the basis set that plagues more common
ab initio methods. Thanks to these characteristics, different flavours of
QMC have been applied to a wide set of species/problems spanning a
237
238 M. Mella and G. Morosi
range that stretches from molecules as small as water up to pieces of

bulk matter as large as silicon and germanium crystals and that includes
molecules such as porphyrins and C20 .
Despite the strong points highlighted, QMC suffers from a relatively
high computational cost mainly due to the necessity of evaluating many
times a reference wave function. Thus, a substantial reduction of this
cost may come from central processing units (CPU) particularly tai-
lored to compute exponentials, polynomials and rational functions, from
better algorithms that require less function evaluations and from variance
reduction techniques. Besides, more robust approaches to reduce the so-
called nodal error would help improve the already appreciable accuracy
afforded by QMC.
With the above issues improved, there are several avenues that would
become possible to pursue on a routine basis. Among these, we foresee
the calculations of intermolecular interactions for very large systems
(e.g. parts of DNA with or without interacting species), the calculation
of nuclear magnetic resonance (NMR) parameters for difficult systems,
the automatic optimization of molecular structures and, even better, the
chance of running molecular dynamics simulation a la CarParrinello
using QMC computed atomic forces. In these circumstances, the study
of phase transitions, bulk matter, interfaces and large biological systems
may reach an unexpected level of accuracy that is currently unavailable
due to methodological limitations.
9.1. Introduction
In their quest for highly accurate quantitative results, quantum chemists

have devised a menagerie of approaches to improve on the shortcomings
of widespread independent particle techniques such as the HartreeFock
method. In this respect, the big urge to avoid uncontrolled approxima-
tions has led to the creation of the so-called first principle approaches.
Here, the only available approximation is represented by a specific ana-
lytical choice for the wave function representing the electronic system,
while all the ancillary quantities (e.g. the matrix elements of electronic
operators) are computed exactly or, at worst, with controllable numerical
errors.
Owing to the linear nature of the Schrodinger equation, many of the
methods devised so far are built starting from a common idea, namely that it
is possible to write the exact wave function for a quantum system as a linear
combination of functions. In this respect, differences can be found in the
way linear coefficients are chosen, e.g. either solving secular equations or
Quantum Monte Carlo Approaches for Tackling Electronic Correlation 239
employing perturbation theory formulae. Alternatively, differences depend

on the manner of selecting the set of basis functions used to approximate
the exact one, as in the case of single and double configuration interaction,
CISD, where the linear expansion is truncated limiting the number of excita-
tions from occupied orbitals to virtual ones. Due to such differences, each
method has a reasonably well-defined realm of applications, with tech-
niques having been explicitly tailored to deal either with dynamic (strong)
or static (weak) correlation [1].
Apart from the intrinsic linearity of the Schrodinger equation, another
advantage of proceeding in the ways just mentioned is provided by the
ability of computers in handling matrix problems. For instance, solving
linear algebra eigenvalue-eigenvector problems has a cost that scales
according to some power (usually cubic) of the number of basis functions
employed in approximating the exact electronic state. However, this number
may steeply increase upon increasing the accuracy of the wave function or
the system size, so that the total cost of a calculation may become quickly
unaffordable due to memory limitation.
Having said that, one may quickly realize that computers are also
excellent in carrying out similar operations over and over, a feature that
may be exploited in the development of alternative approaches for tackling
the issue of describing electronic correlation. In particular, we wish to
provide a concise, but hopefully clear, description of an alternative set of
approaches for such a task. These approaches cumulatively go under the
name of quantum Monte Carlo (QMC) methods.
The general idea behind these methods is to exploit formal similarities
between statistical mechanics and quantum mechanics, thus allowing one
to adapt some of the numerical machinery developed for the description
of condensed phase systems to the realm of application of the latter. In
fact, Monte Carlo is a well-developed technique to estimate average values
(e.g. internal energy, specific heat, etc.) concerning sizable systems [2].
This approach has been bent to comply with the necessities of quantum
chemistry as described in the following sections.
9.2. Variational Monte Carlo (VMC): A Possible Way Toward

Explicitly Correlated Electronic Wave Functions
Very often, the first step carried out in quantum chemistry is the opti-
mization of a wave function describing the particular system under study.
With this available, one could then estimate properties such as binding
energy, electron density moments, and so on.
Bearing this idea in mind, it should therefore come as no surprise that

alternatives have been devised to circumvent some of the shortcomings of
standard quantum chemistry approaches.
Under the umbrella name variational Monte Carlo (VMC) several
tools are collected that allow one to optimize the vector of parameters
p defining an analytical wave function T (R, p) (usually dubbed trial
wave function or TWF). Here, the parameters p may be, for instance, the
coefficients of a linear expansion or the exponents of atomic basis func-
tions; R = (r1 , r2 ,..., rN ) is a 3N dimensional vector formed with the
coordinates of all electrons. The optimization of p must be carried out in
order for T (R, p) to best approximate an eigensolution of the time inde-
pendent Schrodinger equation. The same tools can also be used to estimate
expectation values that may be useful to provide indications on the physical
quantum system under investigation.
Mainstream quantum chemistry provides end users with almost black
box codes with user-friendly interfaces by means of which these two pro-
cesses are usually carried out starting with an orbital approximation for
T (R, p). This means that one makes the approximation that a good
TWF could be written as an antisymmetrized product of spin orbitals
i (ri , i ) = i (ri )(i). These are single particle functions describing the
spatial distribution assigned to an electron and its spin function. Alterna-
tively, a linear combination of such antisymmetrized products could be
employed (see for instance Chapter 2).
These analytical forms for the TWF allow one to write analytical for-
mulae from which it is possible to obtain the expectation value for the
Hamiltonian operator, H, to set up the machinery needed to optimize
the spatial part of the spin orbitals as well as the coefficients of a linear
combination of antisymmetrized products. Moreover, the chosen form for
T (R, p) also allows one to implement algorithms to compute energy
derivatives with respect, for instance, to nuclear positions or to external
fields in order to have access to minimum energy structures and molecular
response properties.
Despite all these achievements, the very choice of using TWFs based
on the idea of antisymmetrized products has an intrinsic limitation when it
comes to improve the absolute accuracy of the average energy for a system.
This is due to the slow convergence of the correlation energy with respect
to the size of the atomic basis set and of the length of the linear expansion.
This difficulty is largely due to the non-analytic behaviour of the exact wave
function where two electrons collide, a point around which the exact wave
function increases almost in linear fashion with respect to the inter-electron
distance rij . In other words, it possesses a cusp at the electron coalescence

point.
Since the original work on the ground and first excited state of
helium published in 1929 by Hylleraas [3], it has been known that the
explicit introduction of the interelectronic distances greatly speeds up
the convergence toward the exact energy and wave function for atomic
and molecular systems. It should not then be a surprise that this idea
has been developed further during the following 80 years by several
researchers, an effort described in other chapters in this book, and that
it has generated a new family of methods commonly dubbed R12 or
F12 depending on the particular analytical form chosen for the TWF.
Even in these new approaches, however, a few limitations are still
present in the freedom afforded to theoreticians in choosing the most
appropriate expressions to be used, due to the necessity of computing
analytically the required expectation values or matrix elements for the
Hamiltonian.
Dispensing with the requirement of analytical integration, instead,
would open up the complete freedom in choosing T (R, p). This, in turn,
is likely to foster an even faster convergence toward the exact energies and
a generation of TWFs with sufficient accuracy and flexibility to guarantee
a uniform quality of the computed molecular properties as a function of the
nuclear geometry.
In the following section, it is our intention to introduce the reader to the
alternative approaches provided by the family of variational Monte Carlo
methods. We will do so, initially at least, avoiding an explicit definition of
the analytical form for the TWF; in this way, we hope to make clear the
general applicability of the methods described.
9.2.1. Numerical integrals in VMC

A common task in quantum mechanics is the evaluation of multi-
dimensional integrals. These are needed to estimate average properties such
as energy and particle densities. However, grid-based integration methods
cannot be extended beyond six dimensions, therefore hindering the pos-
sibility of dealing with many-body systems or with non-separable TWFs.
Luckily, however, similar problems have been tackled and solved in sta-
tistical physics, where simulations of systems containing up to tens of
thousands of particles are commonly carried out. Here, we would like to
show how the same numerical methods can be used in the field of quantum
chemistry.
9.2.1.1. General introduction

In this section we concentrate on the general task of computing an expec-
tation value of the form

dRT (R, p)OT (R, p)
O(p) = , (9.1)
dRT (R, p)T (R, p)
which is a common occurrence in quantum mechanics. Here, we have
implicitly assumed our TWF T to be real, and indicated the parametric
dependence of the expectation value for the operator O (e.g. 2 ) on the
2
parameter set p. The first step of the general Monte Carlo approach is to
rewrite the above equation in a way that can be recognized as identical
to the calculation of a general expectation value of a position-dependent
observable over the coordinate space in statistical physics, i.e.

O = dRO(R)P(R). (9.2)
Here, P(R) is the probability distribution function for the appropriate

ensemble (e.g. P(R) (E)eE for the canonical ensemble). This task
is accomplished introducing the local form for the operator [4]
Oloc (R, p) = T (R, p)1 [OT (R, p)], (9.3)
which allows one to write the expectation value as

dRT2 (R, p)Oloc (R, p)
O(p) = = dRPT (R, p)Oloc (R, p),
dRT2 (R, p)
(9.4)

where we have identified PT (R, p) T2 (R, p)/ dRT2 (R, p) as a
properly defined, i.e. positive definite and normalized, probability distri-
bution function.
In this frame, the task has been transformed into a form that makes
apparent the similarity and that opens the way to the application of the
menagerie of numerical methods made available by years of development
in the field of statistical simulations [2]. To do so, one exploits the fact that
the analytical form of Oloc (R, p) is in general available or easily computable
since in many cases it requires at most to compute the second derivatives
of the TWF. Thus, one is left with the issue of generating a distribution of
points {Rj }PT (p) in configurational space that are distributed accordingly
to PT (R, p).
If we assume for the time being that such a device is indeed available
(vide infra for possible approaches), an estimate for the expectation value
Oloc (p) is provided by
1
M
Oloc (p) Oloc (Rj , p), (9.5)
M
j=1
where M is the total number of points sampled from the distribution
PT (R, p). Such an estimate will converge toward the exact value of
the expectation value for M . Normally, numerical integration
approaches based on a discrete grid of points, such as the generalized trape-
zoidal rule, have a well defined error term depending in a systematic fashion
on the discretization interval of the grid and a convergence to exact results
that is monotonic. This is not so in the case of statistical methods where
it becomes rather fluctuating in nature due to the underlying stochastic
machinery.
If it is reasonable to assume that all the sampled points {Rj }PT (p) are
statistically independent, it is possible to associate a statistical uncertainty
(or standard error) to the average value just computed. This reads

O2loc Oloc 2
(O) = , (9.6)
M1
where it is understood that the average values are computed with respect to
the distribution PT (R, p) and O2loc Oloc 2 represents the variance of
the local operator Oloc (R, p) over the same distribution.
Before turning to the task of distributing points according to PT (R, p),
it is interesting to notice a few formal properties associated with the
calculation of the expectation value of the Hamiltonian. In the limit
T (R, p) = 0 (R), 0 (R) being the exact wave function, and for
Oloc (R, p) = Hloc (R, p), the local operator does no longer depend on the
position in configuration space and it assumes a constant value, the energy
eigenvalue E0 . In this case, the variance of the local operator vanishes and
one reaches an infinite precision on the estimate provided above. It should
also be evident that the same property, usually called zero variance prin-
ciple, applies to any operator that commutes with the Hamiltonian, albeit
there are not many of those objects. This makes the use of the approach
described above to generate an estimate more and more efficient the closer
T (R, p) is to 0 (R).
For non-commuting operators, the property just introduced does not
apply; it is nevertheless possible to improve the statistical accuracy asso-
ciated with the estimate of an expectation value implementing an idea
recently developed into a practical approach by Assaraf and Caffarel [5].

In short, one starts by noticing that Oloc Aloc = Oloc if Aloc = 0
with Aloc being an auxiliary Hermitian operator of unknown form so far.
Albeit Aloc has no effect on the average value, it changes the variance of
the estimate according to
(Oloc Aloc )2 Oloc Aloc 2

= O2loc + A2loc 2Oloc Aloc Oloc 2 (9.7)
so that the standard error associated to the estimate may be reduced if
A2loc 2Oloc Aloc < 0. (9.8)
This clearly requires the two local operators to be highly correlated. As

exemplified by Assaraf and Caffarel, this approach can also be used to
eliminate the divergence of the variance in the estimator for particularly
hard observables such as 1/r 2 . In this cases, the variance of the local
operator is not bound, i.e. O2loc Oloc 2 = , despite the fact that the
average Oloc (p) has a well-defined value. This usually happens when a
local estimator contains terms that diverge as 1/r with > 1; in these
cases the variance contains terms of the form 1/r 2 which diverge faster
than 1/r 2 making the integrals unbound. With the appropriate choice for
the operator A, these diverging terms may be controlled or eliminated, thus
allowing one to obtain a meaningful estimate of the expectation value. As
an example of possible application for this approach, we mention that it
could be useful to estimate H n when n > 1.
9.2.1.2. Sampling of PT
In this section, we will dwell for a while on a few approaches that can
be used to generate a sample of points in configuration space distributed
accordingly to PT (R, p). To make sure that the concept of being distributed
as a specific distribution is as clear as possible, we assume for the time
being that our probability density is mono-dimensional and that it differs
from zero only over a finite interval on the real axis.
If we discretize such interval in subparts of width h and collect the
number of points generated during our sampling that fall inside a spe-
cific subinterval, we could build a histogram of the frequency, or prob-
ability, of hitting a region on the definition interval. In the limit of an
infinite number of points sampled and of infinitesimally small subintervals,
the frequency of falling in any subinterval would become proportional to
Fig. 9.1. Demonstrating the evolution of the Monte Carlo sampling of a 1D

distribution as a function of the number of random variates collected in the bin
grid.
its width so that one could define the value of the associated probability
distribution as
M(x, x + x)
p(x) = lim . (9.9)
Mx x
Here, M(x, x + x) is the number of points fallen in the interval [x, x + x]
(see Fig. 9.1) and the limit is taken in such a way that the number of samples
largely exceeds 1/x.
In discussing the approaches commonly used to sample points in 3N
dimensions, we start noticing that the analytical form of a distribution can
be quite complex (e.g. it contains several maxima). Thus, standard trans-
formation approaches [6] that start with random numbers {} distributed
according to a given distribution and transform the latter into a new set
{ f()} by means of an auxiliary function f(x) are not suitable for this
task. Unfortunately, this happens despite the advantageous fact that the new
set would contain independent variates (another name for random numbers)
if independent objects compose the original one.
The general framework used to produce a sample of points is based on
the theory of Markov chains (see Box 9.1), whose main idea is to iteratively
displace points in configuration space with an appropriate displacement
rule so that in the limit of many iterations the arbitrarily chosen initial
configurations will end up being distributed according to PT (R, p). The
displacement rule T(R R ) gives the probability that the systems being
in position R will jump to position R. As discussed in Box 9.1, the
stochastic transition matrix can be written as product of a displacement
probability D(R R ) and an acceptance probability A(R R ).
Box 9.1: Markov chains.

Given a system with l possible states with occupation probability represented

by a column vector (Pl(i) )T (P1(i) , P2(i) , P3(i) , . . . , Pl(i) ) = lm=1 p(i) m mm with
l (i)
m=1 pm = 1 and mm being the mth versor, a Markov chain is the sequence of
(i+1) (i) (i+1) P(0) . Here, M is a square matrix (transition
vectors Pl = MPl = (M) l
matrix) whose elements fulfil m=1 Mmn = 1 whatever n. The initial proba-
(0)
bility vector (Pl ) is chosen arbitrarily. Under very mild conditions, the state
(i+1) () () ()
vector Pl converges to a stationary point Pl given by Pl = MPl .
An approach to generate a Monte Carlo representation of this process is to
sample a new state mm with probability Mmn from the occupied mn . For a
continuous set of states (e.g. positions in a Cartesian space), the discrete sum
transforms into integrals giving P(R, i + 1) = dR T(R R )P(R , i)

with dRT(R R ) = 1. Thus, the task of sampling an arbitrary distri-
bution has now been transformed into the task of choosing an appropriate
transition matrix such that lim i P(R, i + 1) P(R, i) = dR T(R

R )P(R , i) dR T(R R)P(R, i) = 0.
Thus, the process of obtaining an analytical form for T(R R ) is
usually undertaken writing the transition matrix as a product T(R R ) =
A(R R )D(R R ) between the displacement D(R R ) and accep-
tance A(R R ) matrices. Imposing that at convergence the flow of prob-
ability between two specific states is zero (detailed balance condition), one
can write A(R R )D(R R )P(R ) = A(R R)D(R R)P(R)
) D(R R)P(R)
or A(RR
A(R R) = D(RR )P(R ) , giving the relative value of the accepting prob-
ability conditional to the particular choice of displacement matrix and limit
distribution.
In this way, the overall jump R R is broken down into a part that
can be interpreted as a random displacement with arbitrary distribution and
a correction (acceptance) step that eliminate the bias introduced by a non
optimal choice for D(R R ).
Is the choice of the analytical form for D(R R ) and A(R R )
inconsequential? If D(R R ) guarantees that all the states available to the
systems can be visited with non zero probability within a finite, albeit very
large, number of steps, the form of the displacement matrix controls only the
efficiency of the exploration. Consequently, it defines also the quality of
the statistical properties associated to the averages to be computed.
To comprehend why this happens, one must remember that the
sequential strategy employed to generate samples distributed according to
PT (R, p) does not produce independent configurations, but rather points in
configuration space that remember (i.e. are correlated with) previously

visited positions. This forces one to wait an initially unknown number of
steps before including a configuration related value into an average. Even
more important is the fact that the transition between regions in config-
uration space may be a rare event, perhaps due to an unwise choice for
D(R R ). This fact may preclude an ergodic sampling and introduce a
bias in the results.
As to A(R R ), the freedom available in picking any analytical
form that satisfies the requirements of being a properly defined probability
should be used to assure that it assumes the largest possible value for a given
D(R R ). A commonly used form, proposed by Nicholas Metropolis
[7, 8], is

D(R R)P(R)
A(R R ) = min 1, , (9.10)
D(R R) P(R )
which appears to guarantee the highest possible acceptance probability
for any chosen displacement matrix. As to the analytical form for the
displacement matrix, this is very often chosen to be the short time approx-
imation of a Langevin equation [9].
9.2.2. Optimization of trial wave functions

Independently of the approach used to compute expectation values, the
quality of the data obtained depends, in the end, on the quality of the model
wave function employed in describing the quantum states of a system.
It is, therefore, of utmost importance to choose the best analytical form,
or rather the numerical values of the set of parameters that define it, in the
best possible way.
In mainstream quantum chemistry, this task is usually carried out
making use of the variational principle, i.e. of the fact that E0 H for
any trial wave function that approximates the ground state of the quantum
system. Thus, one usually chooses p as the vector of parameters that produce
the lowest expectation value for the Hamiltonian. An alternative approach,
often used in quantum Monte Carlo, makes use of the fact that the variance
of the local Hamiltonian (also known as local energy) satisfies
Hloc
2
Hloc 2 0, (9.11)
with the identity being fulfilled only by the exact eigenstate [10].
Both approaches, which will be described in the following, guarantee
a uniform convergence (i.e. in a least squares sense) toward the exact
wave function provided that the trial wave function chosen to model the
quantum state has sufficient flexibility. However, both the cost of selecting
an appropriate parameter set and the quality of the results depend on the
strategy employed as described in the next sections.
9.2.2.1. Minimum variance

Historically, the minimum variance approach was the first proposed by
Conroy [10] to tackle the optimization of the parameter set defining the
analytical form of a trial wave function T (R, p). This happened at a
time when the needed integrals could only be numerically estimated on
a limited set of points appropriately chosen. The main reason for this initial
supremacy is related to the fact that the bound
Hloc
2
Hloc 2 = [Eloc (R, p) Hloc ]2 0 (9.12)
holds even if the expectation value is estimated on a limited set of configu-
rations {Ri ; i = 1, Nc } (usually, Nc is of the order of 103 104 ) which do
not even have to be distributed according to T2 (R, p). However, an alter-
native choice of distribution may introduce a bias in the optimal parameters
if the trial wave function does not contain the exact representation of the
eigenstate.
In general terms, a Monte Carlo estimate for the variance of T (R, p)
is provided by
Nc
wi (p)[Eloc (Ri , p) Eloc ]2
Hloc Hloc i=1
2 2
Nc (9.13)
i=1 wi (P)
with

Nc
Nc
Eloc (p) = wi (p)Eloc (R, p) wi (p) (9.14)
i=1 i=1
and where
wi (p) = T2 (Ri , p)/P0 (Ri ) (9.15)
are weights that measure the overlap between P0 , the probability density
used to distribute the configuration set, and the square of the trial wave
function.
A few comments are necessary with respect to the performance and
implementation of the above estimator. We begin mentioning that, although
in the limit of a large number of configurations the Monte Carlo estimate
of the variance converges to the correct value, the accuracy provided by a
finite sample is strongly dependent on the dispersion of the weights wi (p).
One could think about this issue as the equivalent of trying to hit a bulls-eye
throwing darts randomly at the target. The more diffused are the darts, the
higher their number to hit the maximum score.
In practical calculations, the distribution of configurations is usually
chosen as the square of a trial wave function with guessed parameters;
after a few minimization steps (e.g. using a conjugated gradients routine),
such distribution is updated running a Monte Carlo sampling of the newly
obtained T2 (R, p) in order to reduce the weight dispersion for the following
few minimization steps [11].
It is also important to notice that the variance of the Monte Carlo esti-
mator for the dispersion of the local energy may not be bound despite the
fact that the average variance of the local energy is itself always well defined
for an electronic system. In other words, whereas H 2 exists, H 4 may
not. This happens when electron-nucleus cusp conditions are not exactly
satisfied due to the presence of 1/r4 that diverge too quickly to be com-
pensated by the local volume element. To eliminate this difficulty, one has
to use a trial wave function that exactly satisfies all the interparticle cusp
conditions in order to compensate the divergence of the potential with an
opposite one in the local kinetic energy. The net effect of this requirement is
to forbid the use of Gaussian atomic basis sets unless local correction mea-
sures are taken [12]. Alternatively, one can favour the usage of Slater-type
basis sets [13].
As a final comment, we highlight the fact that minimizing the variance
of the local energy can also be interpreted as a fitting problem [11, 14].
Here, the parameters of the trial wave function should be chosen in order
to reduce the least square error between local and average energies. Under
this light, a relevant issue is the implicit assumption of normally distributed
errors made in the Monte Carlo estimator. A recent analysis of this problem
[14] has clearly identified the latter assumption as a problematic one due
to the inappropriate weighting assigned to outlier configurations. The latter
are necessarily present when optimizing molecular wave functions due to
a non-Gaussian shape of the probability distribution function for the local
energy. To correct for this difficulty, it has been proposed to minimize the
average absolute deviation of the local energy from its average

M
1
M |Eloc (Ri , p) Eloc (p)|, (9.16)
i=1
a choice that implicitly assumes a distribution for the values of the local
energy exp[|Eloc (R, p) Eloc (p)|] (i.e. with a slowly decaying tail at
low energy).
To conclude this section, it is important to stress that both the

minimization of the variance and of the absolute errors are effective and
efficient methodologies capable of producing highly accurate trial wave
functions when carried out with the appropriate care. Despite these positive
features, they carry an intrinsic limitation due to the fact that Ep
loc (p)
i
= 0,
an issue that may negatively affect, for instance, the calculation of potential
energy curves/surfaces generating a step-like discontinuous behaviour of
the latter.
9.2.2.2. Minimum energy

As mentioned at the beginning of the main section, energy minimization
is almost ubiquitous in mainstream quantum chemistry (e.g. in Hartree
Fock, HF, multi configuration self consistent field, MCSCF, configuration
interaction, CI, and complete active space-SCF, CASSCF) and it has been
so for many years. Instead, the application of the energy variational prin-
ciple in VMC has picked up momentum only in recent years due to dif-
ficulties encountered when trying to minimize the average value of the
local energy on a fixed configuration. These problems were a consequence
of the fact that the hypothesis of the variational theorem is fulfilled only
when the required integrals are computed exactly or with high numerical
accuracy. On a limited number of samples, the optimization may instead
be biased toward those configurations that have a lower than average local
energy, a fact that is likely to lead toward a strong distortion of the wave
function [4].
The key step that paved the way toward energy-optimized wave func-
tions was the efficient implementation of the expectation values required
to compute the gradient and Hessian of the average energy for a given
parameter set [15, 16]. It makes use of the two identities
ln
= (9.17)
p p
and

2 2 ln ln 2
= + , (9.18)
p2 p2 p
with p representing a generic parameter of the wave function. This set
of equations clearly paved the way for the usage of the Newton method
for the search of the minimum in parameter space. In this respect, the
great advantage provided by a quadratic method as Newtons one is its
intrinsic ability of defining the vector of the displacements in parameter
space by means of a single VMC run. Although this substantially reduces

the computational effort for the optimization with respect to gradient-only
approaches, the overall cost still remains higher than for the variance min-
imization due to a much larger sampling needed to estimate the Hessian
matrix.
Despite the fact that it is in principle possible to optimize all different
kinds of wave function parameters [15], the optimization of the linear
coefficients defining a molecular orbital was found to be unstable. Nev-
ertheless, the methodology proposed by Lin et al. [16] attracted interest
from several groups and there is now a general consensus that energy-
optimized wave functions can provide one with more accurate expectation
values than the variance optimized ones. This is particularly true for oper-
ators non-commuting with the Hamiltonian.
Apart from the increased accuracy for average properties, is there any
other reason why one should pay the additional cost connected with energy
minimization? The answer to this question is a resounding yes, and this is
due to the fact that the optimization of all parameters in T minimizing the
average energy brings the additional advantage that optimized wave func-
tions satisfy the hypothesis of the HellmannFeynman theorem. This, in
turn, allows one to carry out several theoretical simplifications, for instance,
in the calculation of atomic forces (vide infra Section 9.4.2), necessary for
the geometrical optimization of molecular structures.
All these advantages have consequently renewed interest in the energy
optimization of a trial wave function and sparked additional work in recent
years. This effort has led to more stable and less expensive approaches
for this task [17,18]. The clever idea behind the first of these alter-
native approaches [17] was to reduce the fluctuations of the estimators
for the required expectation values in a way similar to what presented in
Section 9.2.1.1. In short, the estimator for the elements of the Hessian
matrix were modified introducing additional terms with zero averages that
allowed the Hessian terms to be estimated as co-variances. The Hessian
matrix was also forcefully symmetrized. The advantage of this approach is
a reduced dispersion of values and a decrease of the sample size needed to
carry out the estimation. This point was demonstrated by optimizing wave
functions for the ground state of NO2 and for the excited 1 Bu state of the
decapentaene C10 H12 using non local pseudopotentials to eliminate core
electrons.
In the second approach [18], the dependence of the trial wave function
on the parameter set is linearized using a Taylor series. The change in
the parameters that minimizes the average energy is obtained solving a
generalized eigensystem whose matrix elements are computed between

the current wave function and its derivatives with respect to the analytical
parameters. Although additional tricks must be used to stabilize the pos-
sibly wild changes in the parameter set, the method has demonstrated to be
capable of optimizing both the linear and non-linear parameters of multi-
determinantal wave functions for C2 and Si2 . The latter systems are known
as difficult beasts to tame due to their peculiar bonding pattern.
9.2.3. Analytical forms for trial wave functions T

By now, the astute reader would have noticed that no mention has been
made so far on the specific form that a TWF should have; this was done in
order to stress that the applicability of the VMC methods is quite general
and it does not require any assumption on the analytical form of T apart
from the usual physical requirements. These are to be square integrable,
anti-symmetric, and to have discontinuous derivatives only on a limited
set of points in space. Thus, we conclude this section on VMC discussing
some of the analytical models commonly employed to approximate the
exact eigenstates of electronic quantum systems.
As for the constraints, we notice that whereas the square integrability
can be easily assured by properly choosing the long distance behaviour so
that T decays exponentially, the approach used to fulfil the constraint of
being antisymmetric may depend, in principle, on the analytic form elected
as an approximation. In this context, a useful simplification comes from
the observation that it is possible to avoid a complete antisymmetrization if
one is interested in molecular properties that are not spin-dependent. In this
case, it can be proven analytically that, once a reference spin function
2N N = N N
i=1 (i)i=N +1 (i) (9.19)
with
Sz 2N N = (2N N)2N N (9.20)
is chosen, the contribution to spin-independent observables coming from
terms generated exchanging electrons of different spin is identically zero.
In particular, neither the average energy nor the local energy variance are
affected if exchanges between different spin electrons are not included.
This means that there is no contamination from spin functions built with
different values of N , and it allows one to concentrate only on the anti-
symmetrization of electrons that have been assigned identical spin and a
limited set of symmetrization operations on min(N , N N ) electrons of
different spin.
In the most general case, the correct spatial symmetry can be enforced
using the Young tableaux [19], which for a chosen N and N value allow
one to build the minimum number of permutation operators that satisfy the
antisymmetry requirement for identical spin electrons in the absence of both
internal or external magnetic fields. Unfortunately, the cost of evaluating
all the permuted terms scales as
N (N 1) N (N 1)
2min(N ,NN ) 2 2 2 2 , (9.21)
thus making this approach fairly time consuming even though it remains
very general. In this respect, it should come as no surprise that it has been
found feasible only for small or exotic systems containing up to five
leptons [2024], for which, however, results have challenged and occa-
sionally superseded more standard molecular orbital approaches.
For larger systems, it has now become customary to exploit as much as
possible the information provided by ab initio electronic structure codes
and to write T as
Nconf

T = eJ
cl det[i,l ]det[j,l ] (9.22)
l=1
where eJ is a total symmetric correlation factor (i.e. a positive term

depending on interparticle distances), and det[i,l ] is a determinant of the
N spin orbitals assigned to the lth configuration function.

This form for the trial wave function affords several clear computa-
tional advantages with respect to a more general form that requires the use
of the Young tableaux. First, the correct antisymmetry between identical
spin electrons is properly enforced by the determinants with an order Ns3
operations despite the fact that to avoid contamination by higher spin states
one is required to use a ROHF-like occupation of spin orbitals. As discussed
in the following, this turns out not to be an issue if the trial wave function is
used in subsequent diffusion Monte Carlo calculations, which are capable
of projecting out the high spin components.
Second, the evaluation of its gradient and Laplacian with respect to the
electron coordinates is simplified by the sum of products form. Finally, it
is possible to exactly satisfy the interparticle cusp condition as described
in Chapter 2 of this book to eliminate the possible divergence of the
local energy, choosing correctly the correlation factor J. In particular,
different logarithmic slopes can be introduced for same spin (1/2) and
opposite spin (1/4) electrons. Notice also that the correlation factor may
be made dependent on the local chemical environment by introducing a
dependence on the electron-nucleus distance, and that the form shown

above has the intrinsic capability of describing both ground and excited
states provided that the correct linear combination of configurations is
chosen. In this respect, however, the optimization of an excited state may
be expected to require more effort if the latter is of the same symmetry of
the ground state due to a possible contamination by low lying states.
9.3. Diffusion Monte Carlo: How to Extract the Best

Information from Inaccurate Wave Functions
To set the stage for introducing the diffusion Monte Carlo method, we feel
it is important to reiterate a few key ideas about what has been described
in the preceding sections. First, VMC allows one the complete freedom in
choosing a trial wave function and to estimate the corresponding average
values of the properties. Second, the results still depend on the chosen TWF
despite the increased freedom of choice. Thus, one may strive to invent
methods that somewhat correct TWF deficiencies in a systematic way as
it is done for HartreeFock TWFs using Configuration Interaction or CC
theories.
In the realm of MC techniques, one can get much better results, in
principle exact ones, by means of the diffusion Monte Carlo method. The
latter is capable of projecting out the excited state components from any
TWF and of sampling the exact (or quasi-exact) ground state employing
a mathematical similarity between the Schrodinger and the generalized
diffusion equation first noticed by Fermi in 1945.
9.3.1. Generalities
The original derivation of the diffusion Monte Carlo method (DMC) [25]
was based on the observation that the time-dependent Schrodinger equation
in imaginary time = it
1
= 2 V(R) (9.23)
2
is analogous to a generalized diffusion equation in the presence of sinks
and sources,
C
= D 2 C k(R)C (9.24)
t
assuming a diffusion constant D = 1/2 and a rate constant k dependent
on the position R. In diffusion process molecules move around randomly,
owing to collisions, and disappear at sinks, while appearing at sources.

So, the similarity between the two equations requires interpreting as a
density of particles in configuration space usually called walkers. Without
the potential term, the above equation is a diffusion equation
C
= D 2 C (9.25)
t
that can be simulated by a random walk. In fact, it can be shown analyti-
(RR0 )2
cally that the function C(R, t) = C0 e 2t , which is the mathematical
representation of the random walk, is the solution of the diffusion equation
with the initial condition C(R, 0) = C0 (R R0 ), the latter representing
C0 walkers located in R0 .
Without the kinetic term, instead, the generalized diffusion equation is
a first-order kinetic equation
C
= k(R)C (9.26)
t
with a rate constant depending on the point in configuration space, that
changes the weight (or number) of the walkers. That this is a correct inter-
pretation can be shown by integrating the above equation with respect to t,
obtaining C(R, t) = C(R, 0)ek(R)t , or C(R, t) = C(R, 0)(1 k(R)t) at
first order in t.
Employing a Euler-like approach, the random walk and kinetic terms
can be alternated in their action as a way of correctly implementing the
time evolution provided the time increment is sufficiently short.
A more formal introduction to DMC is shown in Box 9.2.
Box 9.2: formal derivation of diffusion Monte Carlo method.

The time-dependent Schrodinger equation in imaginary time = it
(R, )
= (H Eref )(R, )

serves as a starting point for the development of the diffusion Monte Carlo
(DMC) method. Eref is the reference energy; the reason for introducing this
term to modify the potential values will be discussed in the following. The
formal solution
(R, + ) = e(HEref ) (R, )
Continued
represents the evolution of the wave function at imaginary time to a later

time + . The evolution can be represented as

(R , + ) = G(R , R; )(R , ),

where the Greens function G(R , R; ) is the spatial representation of the

projector e(HEref ) .
Expanding the wave function in the eigenfunctions of the Hamiltonian leads
to

(R, 0) = ci i
i=0

() = ci e(i Eref ) i .
i=0
If the initial state is not orthogonal to the ground state, in the long time limit
one obtains
lim () = c0 e(0 Eref ) 0

as the contributions from the excited states decay exponentially fast. So, the
reason to introduce Eref was that setting Eref = 0 allows one to keep the
normalization fixed to
() = c0 0 .
In conclusion, the DMC method samples the ground state wave function.
A random initial distribution of walkers drawn with VMC is represen-

tative of a generic wave function, but its evolution in the long time limit
results in a sample of the ground state wave function. To simulate the evo-
lution process of the ensemble of walkers, the total time is divided in short
time steps (short time approximation) in order to exploit the fact that
Trotter [26] approximated the propagator as
eH = e(T +V) eT eV (9.27)
owing to the non-commuting property of kinetic and potential operators.
A better approximation of the propagator in a symmetrized form,
formally second order in , is
e(T +V) e 2 eT e 2 .
V V
(9.28)
Thus, the evolution is split in a pure diffusion followed by a reweighting
process. However, a pure diffusion Monte Carlo method carrying the
weights along the simulation diverges: in the long time limit, a walker whose
weight has exponentially grown dominates the rest of the sample. To avoid
that, the reweighting process is substituted with a stochastic kinetic process
generating n copies with n = int[ + w], where is a uniform random
number from the interval [0,1]. Walkers experiencing an unfavourable
potential are deleted, while n copies of those feeling a favourable potential
are replicated. The fluctuation of the walker number introduces a bias that
can be avoided introducing a minimal stochastic reconfiguration of the
population [27].
For potentials unbound from below, the branching process can result in
wild fluctuations, but so-called importance sampling can cure this problem.
In fact, multiplying 0 by a known trial function T , a new mixed distri-
bution f(R, ) = 0 (R, )T (R) is sampled simulating the equation
f(R, ) 1
= 2 f(R, ) + [F(R)f(R, )]
2
(Eloc Eref )f(R, ) = Lf(R, ) (Eloc Eref )f(R, ), (9.29)
where F(R) = T1 T is the quantum force, Eloc = T1 HT is the local
energy, and L is the FokkerPlanck (Schmoluchowski) operator. First, a
drift term is introduced that drives the walkers towards regions where T is
large. Second, the reweighing term now depends on the local energy instead
of on the potential. Also, the fluctuations of the local energy of a decent trial
wave function are fairly small thanks to a cancellation of diverging terms in
both the potential and local kinetic energy. As approximate representations
for eL spoil the detailed balance condition, the walker move is usually
accepted with probability

|T (R )|2 G(R R, )

A(R R) = min 1, . (9.30)
|T (R)|2 G(R R , )
The energy mixed estimator

0 HT dR
E0 =
0 T dR
is the exact ground state energy since 0 is an eigenstate of H.
Notice that, starting for example from T2 , the simulation eventually
ends up with T 0 , not the pure distribution needed to estimate properties
that do not commute with the Hamiltonian. Strategies useful for the cal-
culation of average values for non-commuting operators are described in a
following section (see Section 9.4).
9.3.2. Improved projectors

The short time approximation introduces a bias that is O( n ); the correct
result can be obtained by several simulations using rather short time steps
and extrapolation to zero time-step. However, short time steps introduce
correlation in sampled data due to limited changes in the walker position,
thus decreasing the efficiency of the simulation.
To improve the evolution process more points can be sampled during the
single time-step. A long time-step is split in smaller ones that are evaluated
by a simpler trial wave function to reduce the computational cost [28].
Alternatively, a fourth-order approximation of the propagator [29] can also
be derived
2
e(T +V) e 6 eT 2 e 3 V eT 2 e 6
V V
(9.31)
with
2 [V(R)]2
V (R) = V(R) + . (9.32)
48 m
In the latter, an intermediate point R, where the potential and its gra-
dient must be evaluated, is introduced between the initial R and the final
R points. Unfortunately, this propagator depends on the potential and its
gradient, so it can be used only with potentials bound from below.
Another possibility to improve the efficiency of the simulation is to
select the final point by a better numerical scheme simulating the effect of
the evolution operator [30].
The solution of the importance-sampled Schrodinger equation is
obtained by iterating the action of the operator e(L+Eloc ) over an initial
distribution f (R,0). Using high order approximations requires potential and
local energy gradients, terms that may degrade the efficiency of the simu-
lation process. Thus, limiting ourselves to a Trotter-like approximation
e(L+Eloc ) = eEloc 2 eL eEloc 2

(9.33)
it is possible to introduce a multistep method and define a better final point
by a predictor-corrector technique. Again this technique is efficient only
when applied to smooth potentials.
In the end, the possibility of using long time-steps for Coulomb
potentials is still an open problem. It is particularly severe for atoms with
large Z, as core electrons require smaller time steps than valence ones.
Pseudo-potentials allow dealing with valence electrons only and using
longer time-steps. However, beyond the approximation inherent to the use
of pseudo-potentials, one has to further approximate the latter in a local
form: in this way DMC enables the calculation of very accurate results, but
on a model problem.
9.3.3. DMC, state symmetry and excited states

For boson systems the ground state wave function is symmetric, but for
fermions it is antisymmetric, that is, it cannot be interpreted as a density
of walkers, unless their diffusion process is constrained within a nodal
volume, introducing what is known as fermion sign problem. This means,
literally, rejecting any walker displacement that crosses a nodal surface,
i.e. a displacement during which the TWF changes its sign. The exact
nodal surface is known only for very few model systems, so one usually
takes the nodal surface of a trial wave function as a boundary to the dif-
fusion process (fixed-node approximation). It can be demonstrated that the
FN-DMC method is variational, i.e. the energy is an upper bound to the
exact one.
As discussed in earlier sections, the trial wave function is usually written
as a product of an antisymmetric part (a determinant or a determinant
expansion) times a symmetric correlation factor, that reduces the statistical
fluctuations improving the wave function accuracy, but has no influence on
the nodal surface. The nodal surface of this standard trial wave function
tends to be fairly accurate. For instance, DMC recovers 100% of the cor-
relation energy for LiH, roughly 95% for Li2 and approximately 90% for
C2 [31]. The correlation energy can be improved using large atomic basis
sets, different single particle basis sets and multi determinant expansions.
At present, among the single particle basis sets (HF, MC-SCF, NO) DFT
orbitals seems to be the best choice, at least for single determinant trial wave
functions. The use of exchange-correlation functionals with variable exact-
exchange component makes possible the minimization of the nodal error
[32]. The optimization of the trial wave function usually results in a better
DMC energy. It is performed within the variational Monte Carlo method,
so it samples more frequently the regions of the wave function that have a
larger probability density and are thus far from the nodes. In summary, the
nodes are improved only indirectly. To get a direct reduction of the fermion
sign error one has to optimize the nodal surfaces in DMC. Because the
gradient of the wave function is discontinuous at a nodal surface not coin-
cident with the exact one, smoothing out the kinks at the nodes improves the
nodes of the trial wave function. A self healing diffusion Monte Carlo algo-
rithm [33] can improve both the single particle orbitals and the coefficients
of a multi determinant expansion on the basis of the walker configuration
distribution generated in a DMC calculation.
QMC methods allow dealing with a wide variety of wave functions.

A natural generalization of a one-particle orbital is a two-particle or pair
orbital, sometimes called a geminal. It is a direct way to introduce cor-
relation, but the optimization of the antisymmetrized pair function (APF)
met numerical difficulties and so its use was limited to few simple systems,
in spite of its ability to represent a wave function in a very compact way
when compared to CI expansions. Casula and Sorella [34] showed that
these difficulties can be overcome in Monte Carlo methods, optimizing
a product of an APF times an explicitly correlated Jastrow factor similar
to F12 (see Chapter 2) by a stochastic reconfiguration technique. In APF,
the pair electron spins are singlet coupled, so systems partially polarized
are poorly described. To introduce the triplet pairing, Bajdich et al. [35]
wrote the wave function in Pfaffian form, which gives the determinant of a
skew symmetric matrix as a polynomial. The Pfaffian allows one to include
into a single, compact wave function pair orbitals, both singlet and triplet,
together with unpaired one-particle orbitals. This wave function and a multi-
Pfaffian expansion recover a large fraction of the correlation energy, while
being much more compact than expansions in determinants.
Several proposals to go beyond the fixed node approximation have been
suggested.
In the release node method [36] the antisymmetric wave function, dif-
ference of two positive components, is simulated by two distributions of
signed walkers that are propagated using a guide function everywhere pos-
itive, so decaying to two bosonic distributions. As the simulation goes on,
one gets a transient estimate while the fermionic signal decays and the noise
increases. Another possibility is to position for example a Gaussian on top
of each signed walker [37] building an antisymmetric function whose nodal
surface evolves during the simulation process: to define this adaptive wave
function one has to fill all the space, a problem more and more severe
the larger its dimensionality. Also, Kalos and Pederiva [38] have heavily
modified the DMC method introducing correlated pairs of random walkers
that carry opposite signs, using different functions to guide walkers of dif-
ferent signs, while the Gaussians used for members of a pair are correlated;
walkers can cancel so as to conserve their expected future contributions.
The overall computational complexity of this method, in principle exact,
has hampered its application beyond very simple model systems.
Small improvements to DMC energies may also be obtained by the
backflow technique, i.e. by replacing the positions of the electrons in the
wave function by quasiparticle coordinates in terms of the positions of
all the other particles [39]. It can only slightly modify the nodal surface, so
one has to start already from a good one.
Worth stressing is the fact that the lowest state of a given space and/or
spin symmetry can be easily computed provided a trial wave function of
appropriate symmetry. For excited states of same symmetry as lower states,
the assumption that the DMC energy is greater than or equal to the energy of
the lowest exact eigenfunction with the same symmetry as the trial function
is valid only if the trial function transforms according to a one-dimensional
irreducible representation of the symmetry group of the Hamiltonian [40].
Furthermore, each state requires an independent simulation and so it is
difficult to extract energy differences. To cope with these issues, a method
has been devised that combines DMC and the variational principle [41],
generating simultaneously matrix elements and so many states orthogonal
to lower states. However, this method suffers from statistical noise, which
is an occurrence directly connected with the sign error.
9.4. Computing Observables Different from State Energy
What can one do once DMC is sampling a distribution that is proportional to

0 ? Besides computing the energy, many other quantities are of interest in
quantum chemistry (e.g. multipole moments, particle distributions, electro-
static potential, etc.). For these, however, the associate operator O does not
commute with the Hamiltonian (i.e. [O, H] = 0) so that O0 = OFN and
one faces the task of somewhat correcting for the bias introduced employing
the mixed distribution f instead of the square of the exact wave function.
How this task could be approached is the content of the next subsection.
Additionally, we shall also discuss approaches recently implemented
to compute problematic quantities such as atomic forces and 0 with
the latter operator being the Dirac delta operator. Importantly, obtaining
accurate estimating for 0 still represents a difficult problem in quantum
chemistry.
9.4.1. Exact calculation of position dependent observables

For these properties, the error due to the difference between 0 and T can
be reduced using the variational Monte Carlo estimate [4]

0 A0 dR 0 AT dR T AT dR
=2
0 0 dR 0 T dR T T dR
2

0 T
+O . (9.34)

0 0 dR T T dR
Pure expectation values can be obtained by the forward walking technique

[4244], that is sampling the ratio 0 /T evaluating the asymptotical
number of descendents from a given walker. However, in systems with high-
dimensional configuration spaces a small population introduces systematic
biases [45].
More efficient evaluation of pure expectation values can be obtained
by other quantum Monte Carlo methods, like the reptation method [46]:
a new path is suggested by removing some configurations at one end of
the path and adding new configurations at other end. One extracts esti-
mates of the exact energy from reptiles heads and tails (their first and
last configurations of electrons, respectively), and estimates expectation
values for operators that do not commute with the Hamiltonian from their
middle configurations, which are distributed according to the true ground
state.
9.4.2. Calculation of atomic forces in VMC/DMC

Much of the chemical interpretation of a phenomenon comes from knowing
the structure of stable species and of the transition states separating them.
In other words, much of the work in applied quantum chemistry has to do
with the localization of stationary points on the potential energy surface of
an atomic ensemble. It should therefore be clear why much effort was made
in the 1980s to obtain efficient ways of computing atomic forces (i.e. the
potential energy surface, PES, gradient) and second derivatives with respect
to atomic coordinates for all but the highest level of theory. Nowadays, it is
in fact possible to localize stationary points at the SCF, density functional
theory (DFT) and MllerPlesset second order (MP2) levels for medium
and large molecular systems as a routine task without resorting to state of
the art computing facilities.
The above possibility clearly paves the way for a two-step hierar-
chical approach to the study of chemical transformation that in good
substance boils down to localizing stationary points at a low level of theory
(e.g. DFT) and to computing more accurate energies and properties using
more accurate, but also more expensive approaches (e.g. CCSD(T)). The
above mentioned approach has become so widespread to deserve its own
nomenclature or symbology, with the lowest level being separated from the
highest by a double slash // symbol (e.g. DFT/6-31D(d,p)//CCSDT/aug-
cc-pVTZ is a commonly used approach to study gas phase reactions). In
this respect, it should be clear that QMC could also be used in a similar
strategy as top notch electronic structure methods.
Despite the success afforded by this two-step approach, it however

remains possible that optimizing stable structures with higher theory level
may lead to qualitatively different results. The possibility of such an occur-
rence is readily exemplified by the case of the Zundel cation, H5 O+2 , which is
found to present a sizable barrier for proton transfer at the SCF level, which
disappears when using MP2 [47]. Even worse outcomes may happen when
a system is intrinsically multiconfigurational in nature, a conclusion that
has lead many groups to attempt the coding of, at least, energy gradients for
theory levels such as Configuration Interaction, MCSCF, CCSD and CC2.
In this context, it should therefore come as no surprise that a similar
effort has been carried out by the quantum Monte Carlo community, with
the clear goal of liberating the method from its subordination to different
quantum chemical techniques.
While initial efforts afforded only limited success due to a finite dif-
ference approach and the intrinsic stochastic nature of QMC [48], the
modern attempts of computing atomic forces Rnuc [Eloc (Rnuc )] have
all concentrated on eliminating the 1/r 2 divergence in the Helmann
Feynmann estimator,

nele
(xi R)
F = ZA . (9.35)
|ri R|3
i=1
This estimator is not cancelled by the volume element r2 dr in spherical coor-
dinates. Once again, this task is accomplished by formulating alternative
estimators that are free of such singularity as demonstrated by Assaraf and
Caffarel [49] who employed the estimator
(H EL )min
F [T , min ] = F + . (9.36)
T
Here, F is the above force operator and

nele
(xi R)
min = T ZA . (9.37)
|ri r|3
i=1
Unfortunately, even this approach has some limitations. First, it has a cubic
scaling of the computational cost needed to obtain a chosen standard error
with respect to the atomic number. Thus, the cost increases by 93 on
going from Z = 1 to Z = 9 somewhat limiting the application of this
approach to only light atoms. This is despite a successful attempt to ame-
liorate this scaling law made by Chiesa et al. [50], who employed antithetic
variates to filter out the diverging components, albeit only in the context
of VMC simulations. Noteworthy, the application of a similar trick in
DMC simulations has, however, not yet been attempted due to its non-local
nature.
Second, Assaraf and Caffarels approach does not take into account
that there are additional contributions to the force values that come from
the variation of the wave function with the nuclear coordinates, unless the
wave function used to represent the electronic state is exact or fully opti-
mized with respect to the average energy. Such contributions are known as
Pulays correction [51] and are familiar within the quantum chemistry com-
munity. They can also be introduced in the QMC context, as demonstrated
by Casalegno et al. [15], who reported substantial improvements in the
average values of the forces, equilibrium distances, harmonic frequencies
and their anharmonic corrections [52]. However, this comes at the cost of
an extensive optimization of some of the parameters of the wave function
by minimizing the energy, which no longer represents a major problem as
discussed in previous sections.
As an alternative, one could try to satisfy the HellmannFeynmann
hypothesis sampling the exact electron density, i.e. obtaining 02 instead
than the mixed distribution. Even this, however, is only a partial solution
of the problem given the fact that for fermionic systems the exact nodal
surfaces are not known in advance. In turn, this means that the 02 distri-
bution is exact only with respect to the boundary condition imposed by the
particular choice of trial wave function and that the related electron density
still contains an error difficult to estimate a priori [53].
In principle, the problem of structural determination can also be
confronted without force calculations, but only with energies with mod-
erate statistical uncertainties by a Bayesian inference method [54]. In this
case, however, foreseeing the realm of possible applications is however
made difficult by the lack of direct experience from a wide community of
researchers.
Before concluding this section, it is important to recall that the capa-
bility of computing reliable atomic forces would also pave the way for
the calculation of other energy derivatives, thus opening the chance for
a very accurate estimate of infrared intensities as well as other response
properties.
9.4.3. Computing the expectation value of ultra-local operators:

electron and spin density on nuclei
As mentioned earlier, there are a few observables that are troublesome for
classical quantum chemistry due to their own mathematical form. Among
these, we can certainly list the Dirac delta operator that is required to model
spectroscopic methods such electron spin resonance and nuclear magnetic
resonance by means of electronic structure calculations [55]. Whereas such
a task is quite simple to carry out whenever a model wave function is written
as a linear combination of molecular orbital determinants so as to require
only inexpensive post-processing of the wave function, the main issue con-
fronted by mainstream quantum chemistry is the selection of appropriate
models since it requires a substantial amount of correlation to be injected.
The requirements also include the development and use of specialized basis
sets capable of describing electron-nucleus coalescence regions with high
accuracy.
Clearly, QMC may play an important role in this arena thanks to its
intrinsic capability of building fully correlated, and yet compact, models.
However, the intrinsically stochastic nature of QMC and its discrete
representation for the electron density raise a few issues when it comes
to produce an estimate for operators such as the Diracs delta, which has a
very limited set of points over which it returns non-zero values.
In this respect, many initial attempts of estimating (r r ), with r
being the position of a specific nucleus, addressed the issue by substituting
a sequence of simple functions that weakly converge to the correct operator
for a limit value of some parameter [24, 56]. Despite some success, such
an approach is intrinsically biased due to the diverging variance of the
estimators employed in the calculation, thus limiting its application to small
systems for which simulations can always be run long enough to control
the statistical errors.
Alternatively, one could exploit the differential identity 4(r) = 2
(1/|r|) and integrate by parts to obtain an estimator that contains only
position-dependent quantities. Among the latter, however, a term propor-
tional to 1/r2 is also present, which clearly has an unbound variance
and which requires the regularization approach discussed above to be
applied [57].
More recently, Chiesa et al. [58] proposed an alternative approach
developed by Ceperley and Alder [59] to compute muon-sticking prob-
abilities into a practical form for estimating the coalescence probability
between electron-positron pairs by separating the contributions coming
from the different parts of the estimator. Given the excellent performance
demonstrated by the latter idea, Hakansson and Mella [60] morphed again
the same approach in order to improve its efficiency when it comes to
compute the electron and spin densities on top of nuclei. In short, the general
idea behind this scheme is to analytically eliminate the Diracs delta for a
specific electron by integrating over its coordinates. Since this operation

reduces the number of dimensions of the integral at the numerator by three,
the dimensionality of the latter must be readjusted multiplying it by the
integral of a fully positive and normalized function obtaining

(r r )T2 (r, r2 , . . . , rn )drdr2 . . . drn
(r r ) = 2
T (r, r2 , . . . , rn )drdr2 . . . drn
2
(r , r2 , . . . , rn )dr2 . . . drn
= 2T (9.38)
T (r, r2 , . . . , rn )drdr2 . . . drn

f(r r )T2 (r , r2 , . . . , rn )drdr2 . . . drn
= 2
T (r, r2 , . . . , rn )drdr2 . . . drn
In this way, any usual VMC scheme devoted to sample T2 can be
used to estimate the electron density, at least in principle, by averaging
f(rr )T2 (r ,r2 ,...,rn )
T2 (r,r2 ,...,rn )
. In practice, however, a robust estimate requires the
usage of an alternative sampling distribution that is strictly positive; this
is due to the fact that numerator and denominator in the new estimator do
not have completely overlapping zero sets. This fact introduces diverging
behaviour in some regions of configuration space. To make the overall
scheme efficient, the alternative distribution is built, shifting the deter-
minantal part of the trial wave function, i.e. re-using quantities that are
already computed; with the proper implementation, the new approach [60]
was found to be an order of magnitude more efficient than the approach
previously implemented [57].
9.5. Conclusions
We have made an attempt to introduce a few key ideas on the alternative

route provided by quantum Monte Carlo to tackle the problem of an accurate
solution of the electronic Schrodinger equation. Needless to say, the con-
cepts introduced are biased by our own interests and occasionally by our
own work as well.
Nevertheless, it is felt that they represent a reasonable amount of infor-
mation for the non-adepts to gather a critical view on the methods.
Deliberately, and certainly in quite an arbitrary fashion, we have men-
tioned only a few examples of calculations carried out with either VMC
and/or DMC in an explicit way. This choice is only related to our intention
of remaining as general (or abstract) as possible in our presentation, and it
is certainly not due to some hidden limitation of the methodologies. In fact,

QMC has been applied even to both ground and excited states of system as
large as porphyrins [61].
Calculations of atoms, molecules, clusters and solids have demonstrated
QMC applicability to real systems with hundreds of electrons despite the
limitations imposed by the fixed node approximation. Usually 9095%
of the correlation energy is provided and energy differences are found
typically within a few percent of experiment. The diffusion Monte Carlo
method may thus be considered the best method to deal with correlation
problems of large ensembles of particles, both bosons and fermions, inter-
acting with arbitrary potentials. It is computationally very expensive, but it
is easily parallelizable and is very well suited to take advantage of massive
parallel architectures. Its scalability with the number of particle allows
tackling problems that present ab initio methods cannot deal with. In spite
of the limitations of DMC (mainly the sign problem), the continuous devel-
opment of new methods, algorithms and software packages, makes it likely
to be one of the main tools of investigation in materials science in the next
future.
We have also made a substantial effort in reducing to the minimum the
list of references by providing only those key papers that represent focal
nodes of past developments and in somewhat limiting any bias toward future
developments in previous sections that may due to our preconception on
what is important. We hoped, implicitly, that the reader may see possible
links between the material presented and her/his current research interests.
Nevertheless, the spirit of this book requires us to terminate the exposition
providing our very view on future tasks. A sensible list would contain:
1. Better understanding of the nodes and their efficient description.

2. Benchmarking the accuracy of various basis sets and trial wave function
models in computing energetic quantities after optimization by mini-
mizing their average energy.
3. Defining more efficient functional forms for variational wave functions.
4. Automatic optimization of molecular structures and of transition states.
5. Calculation of vibrational frequencies.
6. Calculation of excited states.
7. Development of correlated sampling for computing interaction energy
between molecules as a possible approach to reduce the standard error
associated to this quantity.
8. Implementation of approaches suitable for the calculation of NMR
chemical shielding and coupling between active nuclei.
9. Implementation and optimization of model wave functions describing

the leptonic structure of positron containing molecules.
10. Investigation of systems containing transition metal atoms.
11. Providing very accurate results on selected systems to benchmark other
less computationally expensive methods.
Bibliography
[1] T. Helgaker, J. Olsen and P.L. Jorgensen, Molecular Electronic-Structure Theory

(Wiley, New York, 2000).
[2] B. Smit and D. Frenkel, Understanding Molecular Simulation: From Algorithms to
Applications (Academic Press, New York, 2002).
[3] E.A. Hylleraas, Z. Physik. 54, 347 (1929).
[4] B.L. Hammond, W.A. Lester, Jr., and P.J. Reynolds, Monte Carlo Methods in Ab
Initio Quantum Chemistry (World Scientific, Singapore, 1994).
[5] R. Assaraf and M. Caffarel, Phys. Rev. Lett. 83(23), 4682 (1999).
[6] W.H. Press, B.P. Flannery, S.A. Teukolsky and W.T. Vetterling, Numerical Recipes
(Cambridge University Press, Cambridge, 1992).
[7] N. Metropolis and S. Ulam, J. Am. Stat. Ass. 44, 335 (1949).
[8] N. Metropolis, A.W. Rosenbluth, M.N. Rosenbluth, A.M. Teller and E. Teller, J.
Chem. Phys. 21, 1087 (1953).
[9] L. Mitas, Computer Physics Communications 96(23), 107 (1996).
[10] H. Conroy, J. Chem. Phys. 41, 1331 (1964).
[11] C.J. Umrigar, K.G. Wilson, and J.W. Wilkins, Phys. Rev. Lett. 60(17), 1719 (1988).
[12] S. Manten and A. Luchow, J. Chem. Phys. 115(12), 5362 (2001).
[13] C. Diedrich, A. Luchow, and S. Grimme, J. Chem. Phys. 123(18), 184106 (2005).
[14] D. Bressanini, G. Morosi, and M. Mella, J. Chem. Phys. 116(13), 5345 (2002).
[15] M. Casalegno, M. Mella, and A.M. Rappe, J. Chem. Phys. 118(16), 7193 (2003).
[16] X. Lin, H.K. Zhang, and A.M. Rappe, J. Chem. Phys. 112(6), 2650 (2000).
[17] C.J. Umrigar and C. Filippi, Phys. Rev. Lett. 94(15), 150201 (2005).
[18] C.J. Umrigar, J. Toulouse, C. Filippi, S. Sorella and R.G. Hennig, Phys. Rev. Lett.
98(11), 110201 (2007).
[19] G.A. Gallup, Valence Bond Methods (Cambridge University Press, Cambridge,
2002).
[20] D. Bressanini, M. Mella, and G. Morosi, Chem. Phys. Lett. 240, 566 (1995).
[21] D. Bressanini, M. Mella, and G. Morosi, Chem. Phys. Lett. 272, 370 (1997).
[22] L. Bertini, D. Bressanini, M. Mella and G. Morosi, Int. J. Quantum Chem. 74(1), 23
(1999).
[23] L. Bertini, M. Mella, D. Bressanini and G. Morosi, J. Phys. B: Atomic Molecular &
Optical Physics 34(3), 257 (2001).
[24] M. Mella, G. Morosi, and D. Bressanini, J. Chem. Phys. 111(1), 108114, 1999.
[25] J.B. Anderson, J. Chem. Phys. 63, 1499 (1975).
[26] H.F. Trotter, Proc. Am. Math. Soc. 10, 545 (1959).
[27] R. Assaraf, M. Caffarel, and A. Khelif, Phys. Rev. E 61(4 Part B), 4566 (2000).
[28] J.B. Anderson, J. Chem. Phys. 82(6), 2662 1985.
[29] A.N. Drozdov, J. Chem. Phys. 108(16), 6580 (1998).

[30] P. Hakansson, M. Mella, D. Bressanini, G. Morosi and M. Patrone, J. Chem. Phys.
125(18), 184106 (2006).
[31] D. Bressanini, G. Morosi, and S. Tarasco, J. Chem. Phys. 123(20), 204109 (2005).
[32] J. Kolorenc, S. Hu, and L. Mitas, Phys. Rev. B 82(11), 115108 (2010).
[33] F.A. Reboredo, R.Q. Hood, and P.R.C. Kent, Phys. Rev. B: Condensed Matter and
Materials Physics 79(19), 195117 (2009).
[34] M. Casula and S. Sorella, J. Chem. Phys. 119(13), 65006511 (2003).
[35] M. Bajdich, L. Mitas, G. Drobny, L.K. Wagner and K.E. Schmidt, Phys. Rev. Lett.
96(13), 130201 (2006).
[36] D.M. Ceperley and B.J. Alder, J. Chem. Phys. 81(12), 58335844 (1984).
[37] R . Bianchi, D. Bressanini, P. Cremaschi and G. Morosi, J. Chem. Phys. 98(9), 7204
7209 (1993).
[38] M.H. Kalos and F. Pederiva, Phys. Rev. Lett. 85(17), 35473551 (2000).
[39] P.L. Rios, A. Ma, N.D. Drummond, M.D. Towler and R.J. Needs, Phys. Rev. E 74(6),
066701 (2006).
[40] W.M.C. Foulkes, R.Q. Hood, and R.J. Needs, Phys. Rev. B: Condensed Matter 60(7),
45584570 (1999).
[41] D.M. Ceperley and B. Bernu, J. Chem. Phys. 89(10), 63166328 (1988).
[42] M.H. Kalos, J. Comput. Phys. 1(2), 257276 (1966).
[43] M.H. Kalos, Phys. Rev. A 2(1), 250255, (1970).
[44] K.S. Liu, M.H. Kalos, and G.V. Chester, Phys. Rev. A 10, 303308 (1974).
[45] G.L. Warren and R.J. Hinde, Phys. Rev. E 73(5), 056706 (2006).
[46] S. Baroni and S. Moroni, Phys. Rev. Lett. 82(24), 47454748 (1999).
[47] L. Ojamae, I. Shavitt, and S.J. Singer, J. Chem. Phys. 109(13), 55475564 (1998).
[48] C. Filippi and C.J. Umrigar, Phys. Rev. B 61(24), R16291R16294 (2000).
[49] R. Assaraf and M. Caffarel, J. Chem. Phys. 113(10), 40284034 (2000).
[50] S. Chiesa, D.M. Ceperley, and S. Zhang, Phys. Rev. Lett. 94(3), 036404 (2005).
[51] P. Pulay, Molecular Physics 17(2), 197204 (1969).
[52] L. Myung Won, S.V. Levchenko, and A.M. Rappe, Molecular Physics 105(1922),
24932497 (2007).
[53] A. Badinski, P.D. Haynes, J.R. Trail and R.J. Needs, J. Phys. Cond. Melt. 22(7),
074202 (2010).
[54] L.K. Wagner and J.C. Grossman, Phys. Rev. Lett. 104(21), 210201 (2010) .
[55] Calculation of NMR and EPR Parameters: Theory and Applications, edited by
M. Kaupp, M. Buhl, and V.G. Malkin (Wiley-VCH, Weinheim, 2004).
[56] N. Jiang and D.M. Schrader, J. Chem. Phys. 109(21), 94309433 (1998).
[57] R. Assaraf, M. Caffarel, and A. Scemama, Phys. Rev. E 75(3), 035701 (2007).
[58] S. Chiesa, M. Mella, and G. Morosi, Phys. Rev. A 69(2), 022701 (2004).
[59] D.M. Ceperley and B.J. Alder, Phys. Rev. A 31(12), 19992004 (1985).
[60] P. Hakansson and M. Mella, J. Chem. Phys. 129(12), 124101 (2008).
[61] A. Aspuru-Guzik, O. El Akramine, J.C.Grossman and W.A. Lester Jr., J. Chem. Phys.
120(7), 30493050 (2004).

Chapter 10
Solving the Schrdinger Equation on

Real-Space Grids and with Random Walks
Thomas L. Becka and Joel H. Dedrickb

a Departmentsof Chemistry and Physics,
University of Cincinnati, Cincinnati, OH 45221-0172
b 2144 Garden Terrace, Mountain View, CA 94040
This chapter first discusses real-space grid methods for solving the Kohn
Sham equations of density functional theory. These approaches possess
advantages due to the relatively localized nature of the Hamiltonian
operator on a spatial grid. This computational locality and the physical
locality due to the decay of the one-particle density matrix allow for
the development of low-scaling algorithms. The localized nature of the
real-space representation leads to a drawback, however; iterative pro-
cesses designed to update the wave functions tend to stall due to the
long-wavelength components of the error. Multigrid methods aimed at
overcoming the stalling are discussed. The chapter then moves in a
different direction motivated both by 1) the relatively large computa-
tional and storage overheads of wave-function-based methods and 2)
possible new opportunities for computing based on special-purpose mas-
sively parallel architectures. Potential alternative approaches for large-
scale electronic structure are discussed that employ ideas from quantum
Monte Carlo and reduced density-matrix descriptions. Preliminary work
on a FeynmanKac method that solves directly for the one-particle
density matrix using random walks in localized regions of space is
outlined.
271
272 T.L. Beck and J.H. Dedrick
10.1. Introduction
Numerical solutions of the many-electron Schrodinger equation have been

approached from many different directions, as the chapters in this book
amply illustrate. The numerical approaches have included basis-set methods
for representing the wave function, with those basis sets constructed from
Gaussian functions [1], Slater-type orbitals [2], numerical atomic orbitals
[3], and plane waves [4]. More recently, real-space methods, in which
the wave functions are represented directly on a spatial grid, have been
developed [513]. Another approach that has gained in importance in recent
years is the quantum Monte Carlo (QMC) method. QMC employs random
walks to compute expectation values of various operators while sampling
the many-electron wave function [1419].
Besides choosing a numerical approach for solving the Schrodinger
equation, a basic physical question concerns choosing an appropriate
theoretical level to tackle a particular problem. Currently, for large-
scale problems, the most common choice is Density Functional Theory
(DFTKohnSham equations) [20]. This is because DFT has been advanced
to the point of yielding accurate results for a wide range of chemical
problems [21], and the computational cost scales rather gently with system
size (as N 3 or less) compared with other more accurate theoretical methods
[7,22]. The quantum chemical methods such as perturbation theories (MP2,
etc.), configuration interaction (CI), and the coupled-cluster (CC) method
can yield quite accurate results [1], but at a heavy computational cost
due to the harsh scaling with system size (N 5 or higher without spe-
cialized techniques). Progress is being made, however, on lower-scaling
MP2 calculations [23].
An example problem of current active interest is the quantum mechanical
modeling of monatomic ions in water and in or near proteins [2434]. While
significant progress has been made using approximate force fields and clas-
sical molecular dynamics simulations, recent work has shown that quantum
mechanical effects due to polarization and charge transfer involving the ion
and nearby waters may be more important than previously realized for accu-
rately describing the hydration environment [2932]. These findings may
have important ramifications not only for a basic understanding of solutions
but also for biological processes involving ion channels and transporters
(proteins that control ion fluxes across membranes) [3539].
In order to provide an accurate picture of ions in water, water itself
needs to be reliably modeled. A great deal of research has been directed at
modeling liquid water at the quantum mechanical level, and some notable
Solving the Schrodinger Equation on Real-Space Grids 273
surprises have emerged. Most of the quantum simulations have been con-
ducted at the DFT level with gradient-corrected exchange-correlation func-
tionals [40]. This level of theory allows for efficient numerical solution
of the KohnSham equations, but possesses a drawback too: namely, dis-
persion interactions are not properly represented [4144]. Dispersion inter-
actions are universal and are due to electron correlation effects [45, 46].
The interactions can be non-local, and thus the near-local gradient cor-
rected functional cannot mimic the exact behavior. Dispersion accounts for
about 30% of the binding energy of the water dimer [44]. Computations of
the phase diagram of water with the DFT simulations yield a density nearly
20% too low for liquid water at atmospheric pressure [41]. This is a rather
sobering result, considering the degree of computational effort required for
the simulations. If dispersion interactions are included at an approximate
level, the density is closer to the experimental value [42]. Nevertheless,
these results point out that we still have a long way to go in developing
accurate models of liquid water at the quantum mechanical level. Finally,
it has been shown that nuclear quantum effects are significant in water,
adding another layer of complexity [4749]. Progress is sorely needed, and
that progress wont come purely from larger parallel computers (although
such machines may stimulate part of the progress).
The QMC method offers one possible alternative for the long term
[1419]. In recent years, QMC has become a method of choice for accurate
quantum calculations on relatively large systems. QMC can provide predic-
tions of electronic energies that include nearly all of the electron correlation
[1419]. Extensive algorithmic progress has been made in extending the
method to larger systems, and several software packages exist for use by
a wider range of condensed matter scientists [5052]. The QMC method
scales more gently with system size than the other correlated electronic
structure methods (roughly as N 3 or less [15]), and Monte Carlo methods
are perhaps the easiest to implement on parallel machines. Still, the compu-
tational overhead is quite large. Also, although there has been some recent
progress [5355], there is no widely available method for computing forces
and modeling the dynamics of large systems.
This chapter presents a discussion reflecting a progression in one
authors (TLB) group from grid-based methods toward alternatives related
to the QMC approach. The chapter will first present an overview of the
development of real-space grid algorithms for DFT electronic structure
calculations. Over the last ten years, several reviews [68, 1113], and
two recent texts [56, 57], have appeared that discuss the methodology and
applications of real-space approaches. These numerical methods have now
reached a fairly high level of maturity, and are being applied to large-scale
calculations in chemistry and physics.
Challenging applications have included large biological molecules
[13, 58, 59] and novel nano-structured materials [6, 13], with system sizes
of well over 1,000 atoms. These large-scale applications illustrate important
aspects of frontier problems for quantum modeling. First, the systems are
very large, and thus require efficient algorithms that do not possess severe
scaling bottlenecks. Second, they are inhomogeneous. A biological system
might consist of a peptide [13, 59] or a short DNA strand [58] solvated in
water; each of these is a very large molecule embedded in a sea of many
smaller molecules, with strong intra-molecular and molecule-water inter-
actions. A nano-structure of current interest consists of a large organic
molecule sandwiched between two conducting electrode materials (for
example, gold); computing the current through the molecule as a function
of applied voltage can lead to insights into the possible switching behavior
of the molecular device [13, 60]. This last example displays another inho-
mogeneity most of the region between the electrodes, except for near
the organic molecule, is vacuum. Can we devise a computational method
that allows us to neglect this vacuum region in solving the Schrodinger
equation?
As we discuss in detail below, the motivation for the development of
real-space methods came from some drawbacks of traditional plane-wave
calculations. First, the real-space approach leads to a relatively spatially
localized (or banded) representation of the Hamiltonian operator, in contrast
to the plane-wave method. The localized representation in turn makes the
resulting algorithms more suitable for parallel computing, a major thrust of
modern computational science. Second, at the physical level, the effect of
moving one atom a small amount propagates only a short distance in space
for most systems [61, 62]. Thus a more localized representation coupled
with the physical localization can lead to efficient low-scaling algorithms
[6365].
Moving to the real-space representation comes with a cost, however.
Namely, the iterative methods typically used to update the wave func-
tions tend to stall (critical slowing down, or CSD) [66, 67]. This effect
occurs due to the long-wavelength components of the errors in the initial
approximation to the wave functions. The localized iterations cannot effi-
ciently remove that error. Multigrid methods [66, 67], developed in applied
mathematics in the 1970s by Achi Brandt and others, attempt to overcome
the stalling by passing the problem to a hierarchy of coarse grids, iterating

there, and correcting the fine-grid solutions. Such an approach can, at least
in principle, remove the long-wavelength errors. This is one of the few
cases in computational science in which you seemingly get something
for nothing (the FFT algorithm [68] is another); the coarse-grid processes
require only 1/8 the computational overhead of the next-finer grid for each
successive level, yet those coarse-grid iterations solve the hardest part of
the problem. This chapter will discuss some of the basics of the multigrid
method for solving the KohnSham equations.
The above-discussed limitations of the DFT method, and algorithmic
limitations related to the computational and storage overhead of the existing
codes, led us to search for new approaches, however. Thus, following the
overview of real-space grid methods, the discussion will veer in a different
direction related to QMC methods, motivated by the possibility of new
massively parallel computing architectures. The view presented is highly
preliminary, and is intended to stimulate possible progress along an alter-
native path. Two general goals guide the discussion: an emphasis on the
near-locality of the calculations (using reduced density matrix representa-
tions) and low storage overhead.
10.2. Solving the Schrdinger Equation Using Grids in Real

Space
10.2.1. Basics of grid methods

Plane-wave basis sets have been utilized in computational physics for a long
time [4, 56]. These basis sets possess many advantages: (1) they automati-
cally satisfy periodic boundary conditions (2) they yield a fully-numerical
approach in which a single parameter (the plane-wave cutoff energy) con-
trols the level of convergence (3) no spurious Pulay forces appear due to
a localized basis set and (4) they allow for the use of efficient FFT methods.
The CarParrinello (CP) method, first proposed in 1985 [69], ushered in
the ab initio condensed matter simulation field. In that method, a plane-
wave basis set is employed, and the electron degrees of freedom (orbital
values) are propagated on an equal footing with the nuclei based on an
extended Lagrangian formalism. (The Lagrangian is the difference between
the kinetic and potential energies.) If a small enough effective mass is used
for the electronic degrees of freedom in the CP simulations, little coupling
occurs between the electrons and nuclei, and the energy is approximately
conserved for the nuclear motions. After updating the orbitals, they need
to be re-orthogonalized, an N 3 scaling step. While the CP simulations are
costly, they have led to a whole generation of quantum models of complex
condensed phase systems at the DFT level. The CP method is described in
detail in [56]. Other algorithms converge the electronic degrees of freedom
to the ground state for each time step in a dynamics simulation [70].
Besides all of the advantages of plane-wave methods, there are some
drawbacks. The major one is that plane waves are completely delocalized
in space. Imagine moving one nucleus at a specific site in a large system.
The electrons then redistribute so as to screen out the effect of moving
the nucleus. For systems with a band gap, the screening is exponential
[62, 71]; even for metals at zero temperature, the screening is algebraic
(and likely exponential at finite temperature) [72]. Thus the disturbance
created by moving the nucleus a small amount is localized. In a plane-wave
calculation, all of the plane-wave coefficients need to be updated to recreate
the local disturbance. As another example, if plane waves are used to study
a molecule or finite cluster, they must add up to yield a near-zero result for
the electron density away from the region of interest. This basic physical
feature led to an interest in representing the orbitals directly on grids in real
space [7]. These real-space methods are also fully numerical in the sense
that a single parameter, the grid spacing, can be reduced to yield a desired
level of convergence. It is easy to see that if the physical effect of moving an
atom is relatively localized in space, the orbitals only need to be updated in
that local region. We mention that one of the linear-scaling approaches, the
Order-N Electronic Total Energy Package (ONETEP), employs a localized
psinc basis built from plane waves, and exploits advantages of both plane-
wave and localized real-space representations to optimize computational
efficiency [59].
To get an intuitive idea of grid methods, we will consider the sim-
plest possible case, solution of the Schrodinger equation in one dimension
(atomic units are assumed in this chapter):
1 d 2 (x)
+ V(x)(x) = E(x). (10.1)
2 dx2
For the present discussion, we will consider the finite-difference (FD)
method [7, 8, 73]. The resulting equations for the finite-element (FE) method
turn out to be quite similar [7476]; there are important differences between
the FD and FE methods, however, including the fact that the FE method is
a variational (localized basis set) method, while the FD method is not.
Perform a Taylor expansion of the wave function about the point x in

the forward and backward directions:
1
(x + h) (x) + (x)h + (x)h2 +
2
1
(x h) (x) (x)h + (x)h2 , (10.2)
2
where h is the grid spacing. Then solve for (x):
1
(x) 2 [(x h) 2(x) + (x + h)] + O h2 . (10.3)
h
This is the simplest, second-order FD approximation to the second-
derivative operator. Then we can express the Schrodinger equation on a
grid as
1
[(xi h) 2(xi ) + (xi + h)] + V(xi )(xi ) = E(xi ),
2h2
(10.4)
where here the label i has been added to emphasize that we only store
wave function values on the grid points. Equation (10.4) leads to a matrix
eigenvalue equation in which the matrix size is Ng , the total number of grid
points. It is a good exercise to write a short program to solve this matrix
equation for a simple case such as the harmonic oscillator in one dimension.
The grid needs to span the region in which the wave function is appreciably
nonzero, and that domain size can be estimated by looking at the analytical
result. You will likely find that the eigenvalues can be below the analytic
results (but they improve with increased resolution of the grid). In this sense
the FD method is not variational like a basis-set method. As we will see
below, however, we can view iterative processes to obtain the solution as
resulting from minimizing a high-dimensional function on the grid.
As the matrix size gets large for any realistic problem we might want
to solve, the matrix approach is not feasible. Typical grids for solving the
KohnSham equations for a large system might have 100 points on a side,
or 106 grid points in three dimensions! In response to this, a key point to
note is that the FD Hamiltonian in Eq. (10.4) is highly banded. That is, only
information along the diagonal and nearby points is required. Applying the
Hamiltonian to the wave function only requires an operation count that
is a small multiple of the number of grid points. The bandedness of the
Hamiltonian makes this approach also more suitable for the development
of parallel codes.
Many if not most problems in theoretical physics and chemistry can
be viewed from a variational standpoint: the differential equation that we
want to solve emerges by extremizing some action functional [77]. (A func-

tional is a mapping from a function to a number [78]). For the Schrodinger
equation, that action functional is:

1 d2
S [] = * 2 dx + * Vdx. (10.5)
2 dx
In this case, the action S is a functional of the wave function. It is clear
that this functional is just the expectation value of the energy for a given
wave function. If we simply minimized this functional, however, the wave
function would no longer be normalized. Thus, we need to add a con-
straint (using the Lagrange multiplier method, see Ref. [79]) to maintain
the normalization:
2
1 *d
S [] = dx + Vdx E
*
dx 1
*
(10.6)
2 dx2
Next, we take the functional derivative of S with respect to changes in
* , and set that derivative to zero (an excellent introduction to functional
derivatives is given as an appendix in the book by Parr and Yang, [20]):
S [] 1 d2
= + V E = 0. (10.7)
* 2 dx2
We see that we get back the Schrodinger equation with an energy eigenvalue
E. That eigenvalue comes from the normalization constraint.
With what we just did, we can obtain the ground state wave function and
eigenvalue, but what about the higher lying states? To extend this approach
to multiple states as we need to do, for example, in KohnSham DFT, then
we take the functional derivative with respect to each of the states, and we
maintain the orthonormality of all the states. That is, we need to make sure
each state is normalized, and each pair of states is orthogonal. A Gram
Schmidt procedure is a numerical approach to enforce these constraints
[79]. The problem with this (for large systems) is that enforcing the con-
straints is a global operation that needs to be performed over the whole
system volume, and that leads to N 3 scaling for the solver. These obser-
vations led to the development of methods that scale better with system
size [7, 10, 12, 22, 59, 80]. Those methods are based on the fact that the
one-particle density matrix (discussed below) decays in magnitude as we
move away from the diagonal element in real space.
How do we obtain an iterative procedure for solving for the eigenfunc-
tions and eigenvalues? Once we have the functional derivative of Eq. (10.7),
we can develop an effective dynamical equation that moves the action

downhill, namely a steepest descent equation [68]:
(xi ) S []
= * . (10.8)
(xi )
Here the functional derivative on the right side is a force that drives
the action downhill; once the minimum is obtained, we have solved the
Schrodinger equation to the accuracy specified by the FD approximation.
How this steepest descent equation translates into the various iterative
schemes (weighted Jacobi, GaussSeidel, SOR, etc.) is discussed in Refs.
[7, 8], and [67]. There are stability limits on the size of the time step for
a given grid spacing.
In practice, it has been found that the second-order approximation of
Eq. (10.4) is not adequate to obtain realistic properties in DFT calculations.
The simple Taylor expansion outlined in Eq. (10.2) can be extended out
to higher orders orders of six or higher have been found necessary to
yield reasonably accurate results on grids that can handle hundreds of atoms
[7, 73, 81]. The same basic principles discussed above apply, but the band
of the Hamiltonian matrix is wider, although still localized in space. Other
forms of high-order approximations are under active development [82, 83].
In addition, most algorithms have employed non-local pseudopotentials
to remove the core electrons [73] (one exception is the GPaw code of
Ref. [84]).
If we attempt to solve the Schrodinger equation with the above iter-
ative methods, we observe that the residual of the grid equations (reflecting
the degree to which the current approximation satisfies Eq. (10.4)) at first
decreases quickly in magnitude, and then the algorithm stalls due to the
CSD phenomenon mentioned above. This is true in any real-space repre-
sentation. If we use some iterative scheme to update the grid equations with
a grid spacing of h, it turns out that we can only remove errors with wave-
lengths that are a small multiple of h. Thus the long-wavelength errors do
not get removed by the iterative process, and many (hundreds or thousands
of) iterations are necessary to get to an acceptable convergence level. These
points are clearly discussed in Ref. [67].
10.2.2. Multiscale (multigrid) approaches

A solution to the CSD problem was developed in breakthrough work in the
early 1970s [66]. The basic idea is intuitive and relatively simple: pass the
problem to a coarser level, perform iterations there (thus removing longer-
wavelength modes of the error), correct the fine-grid problem using the
coarse-grid solution, and iterate a few more times on the fine grid. If this
process is continued recursively to successively coarser grids, in principle
errors of all wavelengths can be efficiently removed, and the solver can
obtain the solution with only several (maybe ten) iterations on the finest
grid. This is impressive!
The origin of the CSD problem is related to the update matrix in the
chosen iteration scheme (here well assume we are using the weighted
Jacobi iteration [67]). The eigenvalues of the update matrix determine
the rate of convergence for the modes of a given wavelength. The closer the
eigenvalue is to one, the slower the rate of convergence basically, the
update matrix decimates errors over a range of wavelengths, and the longer-
wavelength modes possess eigenvalues that approach one as the grid spacing
shrinks. The eigenvalues for those longer-wavelength modes (small k) are
given approximately by
2 2
k h
k 1 , (10.9)
2
where is the weighting parameter in the weighted Jacobi iteration scheme
(which must be less than one), and h is the grid spacing (see Ref. [8] for
the derivation). Thus we can see that, as the grid spacing gets smaller, the
eigenvalue for the longest-wavelength mode approaches one. By passing
to a coarser grid, the update matrix possesses long-wavelength eigenvalues
that decrease relative to the finer level, thus improving the convergence for
those modes of the error. This is a basic principle of the multigrid method.
An important point is that the coarse-grid problem needs to be con-
structed in a highly specific way in order for the algorithm to fully converge
to the exact numerical result on the grid; that is, if we iterate the multigrid
process many times, we should be able to drive the solution to machine
precision errors on the finest grid. Alternatively, we can say that, if we had
the exact fine grid solution and passed it to the next coarser grid, nothing
should happen.
Consider a simple problem such as numerical solution of the Poisson
equation of electrostatics:
2 (r) = 4(r), (10.10)
where is the electrostatic potential and is the charge density. On the
finest grid labeled by h, we can express this equation in FD form as
Lh U h = f h , (10.11)
where Lh is the FD Laplacian operator, U h is the exact grid solution for the
potential, and f h is the right side of Eq. (10.10). The current approximation
to the potential is written in lower-case, uh . To pass the problem to the

next coarser grid, the Laplacian takes the same FD form as shown above,
but with a larger grid spacing H = 2h. To pass the potential uh to the
coarse grid, we take a local weighted average of the fine-grid values; this
coarsening operation is termed restriction, and the corresponding operator
is labeled IhH (for fine-to-coarse, bottom-to-top). The inverse of restriction
is interpolation (IH h ), in which a function is passed from the coarse grid
to the fine grid. These operations are discussed in more detail in Refs. [8]
and [67].
It is relatively easy to see that we need an equation different from
Eq. (10.11) on the coarse grid. If we were to set the coarse-grid equation to
LH U H = f H = IhH f h (10.12)
and iterate the problem there, followed by a correction on the fine grid, we
would see that, even with the exact solution from the fine grid, there would
be a net correction. Achi Brandts idea [66] was to modify the coarse-grid
equations to remove this problem:
LH U H = f H + H , (10.13)
where
H = LH IhH uh IhH Lh uh . (10.14)
The grid function H is called the defect correction (which is only a property
of the current fine-grid approximation). Here we used the current approx-
imation to the potential uh to see that H changes as the solution evolves
towards the exact grid result (after which it does not change). Now notice
that, if we insert the exact solution from the fine grid, we obtain an identity:
LH U H = f H + LH IhH U h IhH Lh U h
= f H + LH IhH U h IhH f h
= LH IhH U h (10.15)
Thus, with the inclusion of the defect correction, we have an equation that
satisfies the important condition of zero-correction-at-convergence. So the
problem is passed to the coarser grid, iterated there using Eq. (10.13), and
then the fine-grid solution is corrected as follows:
uh uh + IH (u IhH uh )
h H
(10.16)
The above discussion pertains to two grids only; the process can be extended
to a range of coarse grids, with some minor modifications [8, 66]. Also,
similar strategies, with additional features, are available for solving eigen-
value problems [80, 8588]. The inclusion of the defect correction in
Eq. (10.13) is termed the Full Approximation Scheme (FAS). The FAS
method can be used to solve nonlinear problems also, such as the Poisson
Boltzmann equation of electrostatics [7].
For relatively simple linear problems like solving the Poisson equation,
the ideal multigrid solver behavior is observed, meaning a total of about ten
or fewer iterations on the finest grid, and linear-scaling computational cost.
For solving the nonlinear self-consistent eigenvalue problems in quantum
chemistry, however, other issues arise [88]. These issues have limited (to
some extent) the general utility of multiscale approaches for these tough
problems.
The main limitation in developing multiscale eigenvalue solvers has
been the fact that the higher eigenfunctions are oscillatory. Crudely
speaking, the coarsest grid must have enough resolution to approximately
represent the wiggles in the highest energy eigenfunction. If we go to even
coarser grids, the solver will stall or even diverge. Thus, we can obtain sig-
nificant enhancement in the solver efficiency by using a couple of coarser
grid levels to accelerate the solution process, but the advertised multigrid
efficiency is lost. In practice, that might mean the requirement of a few tens
of iterations on the finest scale to obtain adequate convergence, compared
with ten or fewer iterations for a Poisson problem.
A great deal of development has occurred related to real-space solvers in
DFT, and the progress has by no means been related solely to FD represen-
tations and multigrid solvers. Many of these developments and extensive
applications are covered in a recent Physica Status Solidi B issue [89]
and the review of Ref. [8]. Time-dependent extensions are discussed in
Refs. [90] and [91]. Having seen the progression from a few real-space
papers in the early 1990s to many large-scale real-space solver methods at
the present time, it is clear that real-space methods will continue to com-
prise one major avenue for further developments in large-scale electronic
structure.
10.3. New Ways of Thinking about Large-Scale Solutions
We first give a background discussion of the current status of com-

puting technology relevant to electronic structure calculations. That dis-
cussion suggests that there may be good reasons to consider developing
special-purpose computing architectures for electronic structure problems.
In addition, the discussion motivates the development of alternative algo-
rithms that are highly localized and parallel.
10.3.1. An encounter with Silicon valley

Several years ago one author (TLB) was contacted by the other author
(JHD), an entrepreneur and venture investor who has been involved in
the founding of a number of technology companies. Dedrick wanted to
evaluate whether a special-purpose computational system based on custom
semiconductor devices would allow, for the first time, accurate simulations
of bio-molecules at the quantum mechanical level. He had examined the
literature on electronic structure methods, and had come to the conclusion
that real-space multigrid methods held promise for large-scale applica-
tions. The initial discussions led to a meeting at Los Alamos National
Laboratory with several computational researchers involved in electronic
structure methods. A conclusion of the meeting was that, while real-space
methods hold several advantages (discussed above), they still require a fairly
large storage overhead for the multiple wave function values on the grid and
a large computational overhead due to the processing of the KohnSham
eigenvalue and/or density matrix formulations.
10.3.2. The Borg-ing of computation

It is an unfortunate truth that the computational needs of scientists have had
a minor and steadily decreasing impact on investment priorities in compu-
tational technology. The initial proliferation of architectural approaches has
been subsumed by the inexorable march of machines based on the same-
core technology as the desktop PC. General-purpose microprocessors have
steadily assimilated other approaches under the weight of the immense
R&D expenditure both demanded and afforded by the PC marketplace. As
an example, it is quite certain that the level of floating point performance of
current microprocessors has much less to do with solving the Schrodinger
equation than it has to do with winning benchmarks and creating compelling
3D video games (predicted to account for $34 billion in annual hardware
sales by 2012 [92]). The reader more interested in the former application
may find it daunting to accept the premise that computing technology will
continue to evolve in whichever direction is most appealing to adolescents.
Even having done so, it remains a difficult problem to discern how best to
take advantage of that insight.
10.3.3. The worlds least efficient computer (yours)

While the PC market has resulted in modern microprocessors that are
stunningly powerful by any historical standard, they are also stunningly
inefficient. Of the billion or so transistors on the largest current-generation

processor integrated circuits, the fraction devoted to actual computation
(the execution units that implement the arithmetic specified by the current
machine instruction) is well under 1%! The other 99% is consumed in an
all-out effort to guess what the user (program) will want to do next, and
take actions that increase (even a little) the likelihood that the machine
wont have to stop and wait for the data and resources needed for that next
instruction. This works well much of the time, but it is a probabilistic game,
and one of rapidly diminishing returns. A modern microprocessor can easily
wait idle for 50 or more machine cycles for every useful one, if the assumed
locality of memory accesses doesnt hold and each data element must be
fetched from main memory. So, spending 99% of cost on guessing right
is justifiable, especially given the vast size and winner-take-all nature of
the PC market.
Despite the prevalence of the microprocessor, there is a competing
implementation model for high performance, compute-intensive integrated
circuits that has also conquered a market worth tens of billions of dollars
annually. This second technology is on average 1,000% to 10,000% more
efficient, whether measured in terms of arithmetic operations delivered per
unit of power, cost, or space [93]. What is this paragon of compute effi-
ciency? It is the fully custom, single-purpose device the so called Appli-
cation Specific Integrated Circuit or ASIC specifically, a so-called
direct mapped implementation.
10.3.4. The worlds most efficient computer (also yours)

One ubiquitous and computationally intensive application almost univer-
sally addressed using the direct-mapped ASIC approach is the baseband
portion of a radio. The typical smartphone may contain six to ten dis-
tinct radios (e.g. four-band voice, 12-channel GPS, Bluetooth, wireless
LAN, etc.). Each includes a very compute-intensive arithmetic pipeline
for functions such as up/down conversion, adaptive channel estimation,
(de)modulation, and error correction.
Such a proliferation of radio standards would seem to argue for a
software implementation known in the literature as a software-defined
radio (SDR). While conceptually known for several decades, SDR archi-
tectures have never achieved meaningful commercial acceptance, precisely
due to unacceptable inefficiency. Such an implementation would require
quite a powerful processor a modern desktop CPU might (barely) be up
to the job. Unfortunately the user experience of a battery-powered handset
containing a 150-watt microprocessor would be both uncomfortably warm

and unacceptably brief. The fact that, despite their huge computational cost,
these radios routinely operate for many hours on a small battery demon-
strates that the promised efficiency gain available with a direct-mapped
hardware implementation is real.
By far the most critical factor in the huge efficiency difference
between general-purpose microprocessors and direct-mapped ASIC
implementations is that in the latter case, the system is designed for only
one application, and that application is known at system design time. The
challenge faced by the microprocessor architect is a formidable one: his/her
product must be equally adept at running computational physics codes and
running the word processing programs used to write about them. Allowing
the system to be optimized for a single, known function is the basis for
an efficiency difference that, if realized, would be the rough equivalent
of putting the worlds current largest supercomputer inside the case of a
desktop PC. The current fastest supercomputer is the Jaguar Cray XT5
machine at Oak Ridge National Laboratory. It is comprised of 224,162
general-purpose microprocessors, each with one to two GB of memory [94].
A special-purpose design in current semiconductor technology [95] could
integrate perhaps 25,000 single-precision FLP units per chip, providing
a nominal peak performance increase of 10,000x. A subtle but important
corollary however, is that the economic value of the (narrower) solution
must justify the considerable initial development expense.
10.3.5. The end of a 20-year free ride

Since the mid 1980s, the roughly 50% annual growth in the performance
of individual microprocessors [96] has provided similarly large increases
in the performance of all algorithms. Shortly after the beginning of the
21st century an inflection point was reached: in a speech in 2001, an Intel
executive pointed out that an extrapolation of increasing power density (and
thus temperature) of microprocessor chips would exceed that of a rocket
nozzle by 2006, and of the suns surface by 2012. Clearly something had
to give, and the result is that going forward we can expect an exponential
increase in the number of processors (cores) available to us, but not in their
individual performance. This means we need to make a similar change
in the way we think about structuring our algorithms. Rather than working
toward furthering the efficiency of a few hundred geniuses (very powerful
computers), we need to instead find ways to attack the problem with a
billion ants.
10.3.6. Can we change sides?

It is reasonable at this point to inquire as to whether a machine aimed
at solving the Schrodinger equation could be implemented in this dras-
tically more efficient way. Clearly we meet the first requirement i.e.
a willingness to specify a single problem to be solved. And it would
be hard to imagine a more socially and economically valuable contri-
bution than building a machine that could finally realize a much-hyped
dream, namely that of changing the core activity in drug development
from a search/discovery process to a constructive design process. However,
there are several strict preconditions on the form of an algorithm, which
together determine whether direct-mapping will deliver the desired effi-
ciency improvements.
10.3.7. Algorithm desiderata for the massively parallel future

Our goal here is to impart to the reader a qualitative understanding of the
very different way algorithms must be structured in order to benefit from
the trend toward massive parallelization. It is beyond the scope of this
chapter to enter into a detailed discussion of parallel algorithms or com-
puter architecture. Rather than attempt a quantitative prescription for a new
computational cost function, we will offer some qualitative prescriptions
that can be viewed as rules of thumb. While true in the asymptote, we
readily stipulate that many exceptions can be found. However we believe
that the cost function associated with todays direct-mapped hardware
implementation is a reasonable ansatz for tomorrows cluster consisting
of tens of millions of very simple processors.
(1) A Natural, Fine-grained Decomposition. Good parallel algorithms
do not evolve from serial algorithms they are parallel from the
beginning. The first and most fundamental decision to be made is
to choose the dimensions along which the problem can be divided
into a large number of smaller problems (informally, chunks). In
making this choice, focus on the regularity of communications patterns
between the chunks, and the uniformity of processing across them, not
on whether decomposition nominally results in more calculations. Con-
ceptually we assign each chunk to a separate processing element. In
the parallel algorithm literature, this is called the data layout problem
[97]. We emphasize the importance of thinking fine-grain here. The
best way to align an algorithm to the future trajectory of computer archi-
tecture is to take parallelism to the extreme. Therefore we advocate,
when reading the literature on the subject, to focus on distributed rather
than shared-memory machines, as they are consistent with such an arbi-

trarily fine decomposition. Ideally, each processor should have exactly
one unit of work to do; in reality, each should have either zero or one
units of work at any given time. Having chosen a decomposition, we
should design the rest of the algorithm with a bottom up point of
view. Create a set of instructions that can be locally and independently
executed in each region. The ideal algorithm will exhibit the character-
istics that follow.
(2) Common code running in lockstep across all chunks. This implies
(ideally) no data-dependent flow control. For SIMD machines, it is
often noted that the cost of a CASE statement [98] is the sum of the cost
of all the branches, since in lockstep execution each processor element
must implement all possible branches. While it may seem draconian,
this understates the real cost, since it ignores the cost of implementing
the control structure itself. We do not intend to imply here that the
separate processors assigned to each chunk must actually be subject to
lockstep (SIMD) control, only that if the algorithm allows it, maximum
implementation efficiency will result.
(3) Implicitly addressed communications. The destination for data transfers
between chunks should be implicit, or stated another way, relative to the
source chunk rather than data dependent or absolute. If, for example,
chunks correspond to a square tiling of a two-dimensional domain, we
might transfer data north one chunk or west two chunks, but not to
a location dependent on a data value.
(4) Strongly localized communications. Moving data costs far more than
computing it, and moving it further costs proportionally more. So, we
strive to constrain communications to the smallest possible number of
adjacent chunks. When global communications cant be avoided, try
to structure them as either broadcasts (a single value transmitted to all
chunks), or a reduction that can be structured hierarchically.
(5) Insensitivity to Communications Latency. Work hard to avoid inner
loops containing inter-chunk communications. Instead, try to structure
algorithms so as to be insensitive to the latency of inter-chunk infor-
mation flow.
(6) No probabilistic optimizations. Sparse matrices (referring here to those
whose pattern of sparsity is not fixed) save a great deal of computational
effort on a sequential machine. However, the potential savings cannot
be realized if each matrix entry is assigned its own processor. Design
for the worst case, not the average case, since each chunk must wait
for the slowest of them.
(7) Moderate per-chunk Storage Requirements. An arbitrage is always

available between communications capacity and storage capacity.
While we will decline to attempt a definition of moderate, it should be
noted that semiconductor memories have high overhead costs. Imple-
menting many small memories costs much more than one large one of
the same capacity. So, large-scale replication of data across chunks is
unlikely to be effective as a way to reduce communications costs.
The above discussion lays out a possible future for massively parallel com-
puting that differs from the current trend towards utilization of Graphics
Processing Units (GPUs). While the initial cost (in time and money) of
pursuing this course is high, the payoff may be significant, allowing for
desktop computing that challenges the fastest supercomputers today. The
proposed hardware structure emphasizes algorithm design that is parallel
from the start, is as local as possible, and requires little storage overhead.
While a massively parallel machine as described above is attractive,
caution is in order: we note that the history of computing contains examples
of special-purpose computing architectures that did not achieve widespread
use in the long term. Perhaps the best-known example is the Connection
Machine (an interesting account, written by the founder of Thinking
Machines, is given in Ref. [99], including Feynmans involvement in the
project). On the other hand, current examples of more successful special-
purpose computing projects are the GRAPE machine in Japan, dedicated
to gravitational and molecular dynamics studies, and D.E. Shaws Anton
molecular dynamics machine (links to webpages describing these com-
puters are given in Ref. [94]). What seems to have happened is that the
performance/cost of commodity processors improved so rapidly that any
new machine dedicated to a specific application was out of date by the time it
was constructed. Also, the development of the highly parallel and uniquely
structured software required by such machines was a significant barrier to
broad adoption. The arguments presented above suggest, however, that the
landscape has permanently changed due to fundamental limitations on the
progression of computing hardware.
10.3.8. What are we looking for?

Assuming that a massively parallel machine with rudimentary compute ele-
ments could be developed as described above, what kind of algorithm would
be well-suited for such a hardware scenario? Ever since the beginnings of
parallel computing, it has been recognized that Monte Carlo calculations

are perhaps the most directly parallel algorithms. Each random walker can
undergo its motion independent of the others, and statistical averages can be
assembled by collecting data from each processor after extensive sampling.
As discussed above, the scaling of QMC calculations tends to be much more
gentle than, for example, the alternative correlated electron methods. Some
recent developments even indicate that linear-scaling QMC codes may be
possible [100102]. The prefactor to the scaling is quite large, however,
and there is a large computational overhead for repeatedly constructing
the trial functions that guide the sampling. The significant payoff by fol-
lowing the QMC strategy is that highly accurate physical results can be
obtained.
For a long time, chemists and physicists have sought methods that
would solve electronic structure problems directly using the one- and/or
two-particle density matrices (DMs), rather than involving the entire 3N-
dimensional wave function, and there has been recent progress in this area
[103]. It is well known that the exact ground state energy can be expressed
solely in terms of the two-particle DM or 2-DM [20, 103]. Total energies at
the approximate HartreeFock or DFT levels can be expressed entirely in
terms of the 1-DM [20]. We also know that the exact ground state energy
can be expressed in terms of the one-particle Greens function (1-GF, see
below) [104, 105]. The 1-GF describes the propagation of a disturbance
in which a particle is added or removed from a many-particle equilibrium
system [104].
We can obtain the 1-DM (1 (r1 , r1 )) and the 2-DM (2 (r1 , r2 , r1 , r2 ))
from the many-electron wave function:

1 (r1 , r1 ) =
* (r1 , r2 , r3 , . . . , rN )

(r1 , r2 , r3 , . . . , rN )dr2 . . . drN (10.17)

2 (r1 , r2 , r1 , r2 ) =
* (r1 , r2 , r3 , . . . , rN )

(r1 , r2 , r3 , . . . , rN )dr3 . . . drN . (10.18)
We add the subscripts 1 and 2 to the DMs here in order to keep track of
the one and two particle forms. We can obtain the 1-DM from the 2-DM
by integration over the r2 coordinates, so the 1-DM information is con-
tained within the 2-DM. Note that the prime is omitted on the second r2 in
Eq. (10.18). We can restrict the variables in this way since the exact total
energy can be expressed as

1 2
E= r1 1 (r1 , r1 ) dr + v(r)(r)dr
2 r1 =r1

1
+ 2 (r1 , r2 ; r1 , r2 )dr1 dr2 . (10.19)
r12
The 2-DM in the last (electron-electron interaction) term contains only
diagonal elements, while the 1-DM contains off-diagonal elements. The
three terms of the energy correspond to the kinetic energy, the electron-
nucleus interaction energy (where v(r) is the external potential from the
nuclei), and the electron-electron interaction energy, respectively. Thus, in
principle, we only require a function of nine coordinates to compute the
exact ground state energy from the 2-DM.
This led to early optimism that we could do away with wave func-
tions and minimize the total energy with respect to variations of the 2-
DM. Quickly it was realized, however, that there are important conditions
on the 1-DM and 2-DM that must be satisfied. This problem was termed
the N-representability problem by Coleman [106] the 2-DM must be
derivable from an anti-symmetric N-electron wave function. A great deal
of work has gone into solving this thorny problem. For the relatively simple
case of the 1-DM derived from a single Slater determinant wave function,
the N-representability condition is satisfied by a 1-DM that maintains the
correct number of electrons and is idempotent (see below). Below we will
focus on the 1-DM and label it (r, r ).
As we discussed above, the 1-DM decays in space as we move away
from the point r = r . If there is a band gap (or HOMO/LUMO gap)
for the system, that decay is exponential, and typically the decay to near-
zero values occurs over a region that includes the point of interest and a
few nearby atoms. Typically (r, r ) oscillates, displaying negative values
at larger distances away from r = r . The diagonal element is of course
positive since this is the electron density (r) = [(r, r )]r=r .
For the time being, lets search for a Monte Carlo method to solve
directly for the 1-DM. This would lead to solutions for theories that can be
expressed in terms of the 1-DM such as HartreeFock or KohnSham DFT.
The authors believe that, over time, better and better DFT functionals will
be designed that incorporate challenging features like dispersion, and some
progress has already been made along these lines [42, 107109]. So solving
for the 1-DM itself holds some promise in the long term. In addition, future
Monte Carlo methods along the lines discussed below might be adapted to
iteratively compute the 2-DM as well. Computing the one-particle Greens
function with Monte Carlo methods would be a longer-term goal, and that
will be briefly discussed at the end of the chapter.
A clear discussion showing how the HartreeFock total energy can be
expressed in terms of the 1-DM is given in the book by Parr and Yang [20].
Similarly, the total energy in KohnSham DFT can be expressed in terms of
the 1-DM. Ref. [20] also discusses the constrained minimization that leads
to the HartreeFock equations in 1-DM form. There are two constraints,
one for conservation of the number of electrons, and one for idempotency
of the 1-DM:

(r, r)dr = N (10.20)

(r, r )(r , r )dr = (r, r ). (10.21)
The second constraint, Eq. (10.21), follows from the fact that, for a 1-DM
constructed from a wave function that is a single Slater determinant, the
density matrix is given by

N/2

(r, r ) = 2 i* (r)i (r ), (10.22)
i=1
where the i (r) are the eigenfunctions that solve the HartreeFock or Kohn
Sham equations. (Here we assume doubly occupied states with no net spin.)
The first constraint, Eq. (10.20), is global but easy to enforce: simply update
the density matrix by some process and then rescale to maintain the correct
particle number. The second, Eq. (10.21), is again global, and very difficult
to maintain during an iterative process. Various numerical approaches, such
as the McWeeny purification (discussed in Ref. [56], pgs. 4634), have been
implemented in large-scale codes to iteratively enforce idempotency as we
move toward convergence [59]. Is there another way to effectively enforce
this difficult constraint?
Another point to note is that, in order to construct the total energy and
the electron density (two of the primary goals of the calculation), say in a
KohnSham-type DFT calculation, all we need is the diagonal element of
(r, r ) (the electron density), and nearby points that determine the kinetic
energy:

1 2
Ts (DFT ) = r (r, r ) r=r dr (10.23)

2
The s subscript on the kinetic energy reminds us that, in KohnSham theory,
the kinetic energy is that for a non-interacting system moving in an effective
potential. For example, using only the diagonal element and some nearby
points, we could use a FD approximation for the action of the Laplacian on

the 1-DM, then set r = r and integrate over space to estimate the kinetic
energy. The remainder of the total energy is a functional of the electron
density itself.
In the next section, we will first briefly review the QMC method that
employs random walks to solve the Schrodinger equation, and then follow
a direction that might provide a start at answering the above question con-
cerning the idempotency. Our goal is a method that can use a very large
number of independent Monte Carlo trajectories to construct the 1-DM in a
way that maintains the idempotency constraint. Such an approach meshes
well with the above discussion of new massively parallel architectures.
10.4. Random Walks for Solving the Schrdinger Equation
The reader is referred to Chapter 9 for further discussion of QMC methods.

The present discussion is included for background to the discussion of the
FeynmanKac approach that follows.
10.4.1. Traditional diffusion quantum Monte Carlo

The basic idea of diffusion QMC (DMC) is simple. Consider the time-
dependent Schrodinger equation:
1 d2
i = H = + V. (10.24)
t 2 dx2
Next we replace the real time t with an imaginary time = it. Then we
obtain a partial differential equation that looks just like a diffusion equation
with a branching term (see below) due to the potential energy operator:
1 d2
= H = V (10.25)
2 dx2
If we are to solve this equation as a diffusion equation, we need to add a
constraint term to enforce the wave function norm (as we did above in the
grid context):
1 d2
= H = (V E) (10.26)
2 dx2
We can see that, as the pseudo-time approaches infinity, and the wave
function no longer depends on , we get back the original Schrodinger
equation. If we make an analogy to statistical mechanics, we can see that is
like the inverse temperature, so is just like T 0. Thus the above
diffusion process should sample the ground state wave function at long
pseudo-times. If we vary the parameter E in order to stabilize the diffusion
process (which is equivalent to approximately maintaining the norm of the
wave function), we should also obtain one estimate of the energy eigenvalue.
While the above picture is correct, it turns out that, for strongly inter-
acting systems like electrons and nuclei, the statistical noise gets completely
out of hand if Eq. (10.26) is solved directly with Monte Carlo methods. To
deal with this issue, Grimm and Storer [110] introduced an importance
sampling alternative to Eq. (10.26) that uses a trial function to enhance
the sampling in important regions. Instead of sampling directly the wave
function , we attempt to sample the product of and a trial function T
that is chosen to be as close as possible to the true wave function; we call
this composite function f = T . It is then a relatively easy exercise to
show that the following differential equation for f reduces to Eq. (10.26)
(here a one-dimensional notation is used for simplicity, but all the formulas
are easily generalized to the high-dimensional case.):

f 1 f
= Ff (ET (x) E)f
2 x x

1 f ln T2
= f (ET (x) E)f,
2 x x x

1 2 f ln T
= f (ET (x) E)f (10.27)
2 x2 x x
where F is a drift force that leads to enhanced sampling in regions where
the trial function has a large magnitude. The trial energy ET (x) is given by
ET (x) = HT (x)/T (x). (10.28)
The trial energy is a function of position since the trial wave function is
not the exact eigenfunction; it becomes increasingly smooth as the trial
function becomes more accurate, however. The drift force
ln T2
F= (10.29)
x
can be interpreted as minus the derivative of a (drift or guiding) potential
VD (x) = ln T2 , (10.30)
where the minimum in the potential occurs at the maximum values of
the trial function. Thus addition of this drift term drives the sampling
into the regions of large trial function values; if the trial function closely
approximates the true wave function, then the sampling noise is reduced
significantly.
What has been gained by transforming to Eq. (10.27)? Two improve-
ments have been made: (1) Now we have a diffusion process with drift
(coming from the trial function) that results in importance sampling in
the more important regions determined by the trial function, and (2) the
potential operator V(x) has been replaced by the trial energy ET (x). So
long as the trial function accurately reflects the important properties of the
true wave function, the trial energy is much smoother than the bare (singular)
potential V(x) that is typically the Coulomb potential in electronic structure
calculations. This importance sampling transformation has allowed for real-
istic calculations that would not have been possible without it.
It is relatively easy to generate the Monte Carlo trajectories that lead to
sampling of the distribution f . An equation like Eq. (10.27) is a diffusion
equation that is termed a forward Kolmogorov equation in the mathematics
literature [111], or a FokkerPlanck equation in physics and chemistry
[112]. The purpose of the random walks for this case is, at long times, to
produce sampling of the equilibrium distribution (and not to yield the actual
solution to the differential equation, f , that in turn would give the exact
wave function ). Below we will discuss an alternative view, the backward
equation, which does yield the solution to the differential equation. The
trajectories that yield the desired sampling are Langevin trajectories:

x+1 = x + b d + d, (10.31)
where
F ln T
b= = (10.32)
2 x
and is a Gaussian random number with unit variance. The last two terms
on the right side of Eq. (10.31) are the drift and diffusion terms, respectively.
We select the initial location with a probability determined by T2 and then
initiate the trajectories determined by Eq. (10.31). The sampling becomes
more accurate as the time step size d approaches zero. The Langevin-type
Eq. (10.31) is called a stochastic differential equation (SDE) in mathematics
[113] (written in discrete form here with a finite time step size d for imple-
mentation on a computer); see below for more discussion of SDEs. The code
for numerically solving the SDE is very simple, only requiring a Gaussian
random number (GRN) generator to produce the values. A walker is a
realization of a trajectory given by Eq. (10.31) on a computer.
A final aspect of the sampling relates to the (ET E)f term in

Eq. (10.27). Imagine that the diffusion and drift terms are small. Then
we would have the differential equation:
f
(ET E)f (10.33)

The solution to this equation is
f exp [(ET E)] . (10.34)
Thus we see that, depending on our choice for E, the sampling will lead
to exponential growth or decay of the sought distribution f . This term is
called the branching term in analogy to problems in nuclear reactions or
population growth/decay. What is typically done in DMC simulations is to
add a step in which walkers are created or destroyed based on a condition
so as to maintain a stable population [15]. Also, the parameter E is varied
to drive the system towards stability. This is analogous to maintaining the
norm of the wave function.
How do we compute expectation values during the sampling process?
Ref. [15] provides a clear description. The most important quantity we seek
is the ground state energy E; Ref. [15] shows that an accurate estimate is
obtained as the simple average of the trial energy ET (x) along the Monte
Carlo trajectories.
We mention an interesting alternative to Eq. (10.27) in which, instead
of solving for f , we solve for the difference between f and the square of
the trial function T2 = hT :g = f hT . The resulting differential equation
for g is

g 1 2 g ln T
= g
2 x2 x x
(ET E) g (ET E) hT . (10.35)
This approach may provide an iterative way to successively improve the trial
functions used in the DMC process. As an aside, we note that Eq. (10.35)
is quite different mathematically from Eq. (10.27). The last term on the
right in Eq. (10.35) is a source term, similar to the charge density in a
Poisson equation. In the mathematics literature Eq. (10.35) is referred to as
a Dirichlet problem, while Eq. (10.27) is a Cauchy problem [113]. Different
methods are required for solution of these two equations. Anderson has
proposed a QMC approach for Eq. (10.35) [114].
We will not go into details as to how the trial functions are chosen in
electronic structure calculations. Several excellent reviews are available that
discuss this issue in depth (for example, Ref. [15]). Suffice it to say that the
most common trial functions are built from HartreeFock or DFT orbitals,
and terms are added (called Jastrow functions) to improve the treatment of
electron correlation. The DMC approach outlined here can recover a very
high fraction of the correlation energy.
The above discussion has ignored one important point. We have
implicitly assumed that we are locating the ground state for a one-particle
system that ground state possesses no nodes. In reality, the wave function
for a large system depends on 3N coordinates, and the wave function must
be anti-symmetric with respect to the interchange of electron coordinates.
This leads to spatial nodes in the ground state wave function for a many-
electron system, and the exact nodal locations are unknown. The most
common approximation is then to set the nodes at locations determined
by the trial functions, and restrict the sampling to regions with a single
sign for the wave function [19]. There has been extensive theoretical and
computational work showing that such sampling adequately covers the con-
figuration space [15]. The fixed-node approximation is just that, however,
an approximation.
10.4.2. Another angle

Walter Kohn, in his Nobel Prize lecture [115], makes the provocative
statement: In general the many-electron wave function
(r1 , . . . , rN ) for a
system of N electrons is not a legitimate scientific concept, when N N0 ,
where N0 103 . He bases this conclusion on the exponential wall in the
number of parameters required to represent a many-electron wave function
as the size increases. Since QMC for many electrons is sampling this high-
dimensional wave function, and the largest current calculations are near
this limit of size, perhaps we should consider alternative approaches. This
chapter now moves towards a QMC-type approach for solving for the 1-DM,
from which we can in principle calculate the total energy and the electron
density in a DFT-based approach. As we will see, we dont get something
for nothing, and while the present approach has attractive features, there
are also major obstacles to overcome.
The basic idea is to write a differential equation for the 1-DM that mirrors
the DMC approach for the wave function. Before doing this, we need some
background. Lets assume we are considering a one-particle approach such
as KohnSham DFT, where the wave function is represented by a single
Slater determinant; then, the 1-DM is given by Eq. (10.22). We allow the
Hamiltonian to act on the 1-DM, and multiply and divide by the 1-DM on
the right side to obtain

N/2
Hx (x, y) = 2 i i (x)i (y)
i=1
N/2
2 i=1 i i (x)i (y)
= N/2 (x, y) = E(x, y)(x, y) (10.36)
2 i=1 i (x)i (y)
or
1 d 2 (x, y)
+ V(x)(x, y) = E(x, y)(x, y), (10.37)
2 dx2
where the Hamiltonian acts on the x coordinate, and we have assumed here
that the eigenfunctions are real (which is true for practical DFT calcula-
tions). The function E(x, y) tends to be relatively smooth, but certainly not
constant.
So instead of having an eigenvalue problem with a single parameter, E,
as in the traditional QMC approach, we now have a differential equation for
the 1-DM with a function E(x, y) that is spatially dependent. We maintain
the symbol E here in analogy to the Schrodinger equation. Dawson and
March [116] pointed out that the function E(x, y) is essentially the Lagrange
multiplier for the idempotency constraint Eq. (10.21). The principal dif-
ficulty introduced then by the reduced 1-DM representation is how to
determine the function E(x, y) as opposed to the simple E when the full
wave function is sampled.
Given the differential equation in Eq. (10.37), we can next invent a
diffusion-type equation analogous to Eq. (10.27), that, at equilibrium,
solves Eq. (10.37):
(x, y) 1 2 (x, y)
= 2
V(x) E(x, y) (x, y) (10.38)
2 x
We can go through the same exercise as in DMC to develop an importance-
sampling version of Eq. (10.38):

f(x, y) 1 2 f(x, y) ln T
= f
2 x2 x x

ET (x, y) E(x, y) f(x, y), (10.39)

where f = T , T is the trial 1-DM, and ET (x, y) = HT (x, y)/T (x, y).
The defect correction form analogous to Eq. (10.35) is

g(x, y) 1 2 g(x, y) ln T
= g
2 x2 x x

ET (x, y) E(x, y) g(x, y)

ET (x, y) E(x, y) hT (x, y) (10.40)

where g = f hT and hT = T2 . Thus, the same types of diffusion processes
can be set up to solve for the 1-DM, at least in principle.
The 1-DM description is a highly reduced one, however, and the 1-DM
possesses a relatively simple structure. For many electrons, the 1-DM
is positive near the diagonal element (which gives the electron density),
decays in space (often exponentially with distance), and oscillates (with a
nodal surface appearing in three-dimensional space some distance from the
diagonal).
The same type of fixed-node approximation can be made, where now
the nodes are determined by the trial 1-DM T . Here, we wont con-
sider numerical approaches for determining T ; it might be constructed
using HartreeFock or KohnSham DFT solutions, which would be rather
expensive numerically, or alternatively it might be built up from an approx-
imate DFT theory such as the Optimized ThomasFermi theory of Pratt,
Hoffman, and Harris [117]. That last approach would hold several advan-
tages numerically.
10.4.3. Stochastic differential equations and the

FeynmanKac approach
Now that we have simple forms (Eqs. 10.39 and 10.40) that might allow
us to solve for the 1-DM using a diffusion process, how should we go
about solving those equations? One avenue would be to utilize the forward
Kolmogorov (or FokkerPlanck-type) sampling as is done in DMC and
is discussed above. Then, for a chosen y location, we would attempt to
iteratively update the function E(x, y) so as to stabilize the number of
walkers. But there doesnt appear to be a simple recipe for computing the
expectation value of the energy in terms of averages of the local energy, as
for the DMC case above.
An alternative is to attempt to use Monte Carlo methods to solve directly
for the 1-DM itself. After studying the traditional QMC literature, and the
FeynmanKac formula for the ground state energy that appears in path
integral textbooks [118], one author (TLB) began to read more broadly
the literature on numerical solutions of differential equations using SDEs
(like Eq. (10.31) above). As mentioned above, the purpose of the random
walks in a FokkerPlanck approach is to properly sample the equilibrium
(here ground state) distribution; expectation values of operators of interest
can be computed as averages over these random walks. By a slight rear-
rangement of the diffusion-type equation (into the backward Kolmogorov
form), however, we can obtain the actual solution to the differential equation
(f in Eq. (10.39)). This came as a surprise to someone not well-versed in
SDEs. A helpful mathematics book by Freidlin [113] lays out the back-
ground theory to this approach this book is challenging for chemists and
physicists but is also clearly written and gives the solutions required for the
problems addressed here. The text by Gardiner [111] gives a clear physical
explanation of the backward equation, but does not develop the functional
integral formulas for its solution.
Functional integration was introduced into quantum mechanics by
Feynman with his path integral formula for the evolution of a quantum
system in time [119]. In the path (functional) integral method, quantities
are expressed in terms of averages over many paths linking the initial and
final points. The method was subsequently extended to equilibrium statis-
tical mechanics by going to imaginary time as we did above in looking at
both the grid methods and QMC. Mark Kac, inspired by Feynmans new
way of looking at quantum mechanics, examined the mathematical structure
of the formulation in imaginary time [120]. The resulting theory is called
the FeynmanKac approach. It is worthwhile to work through the chapter
on functional integration in Ref. [120] to begin to understand how solutions
to differential equations can emerge from the path integral approach.
Consider Eq. (10.39) above, but write out the derivative of the second
term to obtain

f(x, y) 1 2 f(x, y) ln T f
=
2 x2 x x
2
ln T
+ ET (x, y) E(x, y) f(x, y). (10.41)
x2
This equation is in the backward form (indicated by having the drift term
in front of the first derivative of f ) and can be written in shorthand as
f(, x)
= Lf(, x) c(x)f(, x), (10.42)

where we suppress the y dependence (we choose a value of y and keep this
fixed to solve for f as a function of x), and
1 2
L= + b(x). (10.43)
2 x2
The drift term in the backward form is
ln T
b(x) = , (10.44)
x
which we can see has the opposite sign from the forward case in Eq. (10.32).
We could view this as the walkers moving on an inverted potential, or
alternatively as moving backwards in time in relation to Eq. (10.31). The
potential operator c(x) includes the three terms of the last contribution (in
brackets) to Eq. (10.41).
What does the seemingly trivial rearrangement of Eq. (10.39) gain for
us? This is where the beautiful mathematics of the SDEs comes in as
claimed above, it allows us to then solve directly for the function sought,
here f , using random walks. The relation of the averages over random
walks to the solution of the differential equation is proved in Ref. [113].
Even after looking at these results for some time, it still seems somewhat
mysterious that adding up quantities computed along the random walks
yields a numerical approximation to the exact solution (to within statistical
errors based on the finite sampling, and finite time-step errors).
We take as our initial condition for f(, x) in Eq. (10.42) the values of
the trial function squared hT = T2 . Also, we attempt to solve the equation
only over a finite domain located within the volume defined by the first
nodal surface of T , and we assume that we know the values of f on that
surface (for example, we might set those values also to the square of the
trial function, which would be an approximation). We will label with k the
values of f on the boundaries D of the domain D
f(, x)|D = k(x). (10.45)
Then it is remarkable that the exact solution can be written as [113]

f(, x) = hT (Xx )<D exp c(Xsx )ds
0

D
+ k(XxD )>D exp c(Xsx )ds . (10.46)
0
Equation (10.46) requires some explanation, but the interpretation of the

process is clear. First, the Langevin-type equation that models the solution
process is just like Eq. (10.31):

x
X+1 = Xx + b d + d, (10.47)
except the sign of the function yielding b has changed. In addition we use
the upper-case X here to represent the current location of the walker that
started at the point x. To obtain the numerical solution f at the point x, we
initiate a large number of Langevin trajectories all starting from the point
x; the averaging in Eq. (10.46) is over the collection of trajectories guided
by the drift function b. The indicator function tests whether the trajectory
has hit the chosen boundary of the domain. In the first term on the right
side, its value is 1 if the trajectory does not hit the boundary during the
time , otherwise it is 0. The in the second term on the right is 1 if the
trajectory does hit the boundary, and 0 if it does not; D is the time it takes
the trajectory to hit the boundary.
We can summarize the solution process for determining f(, x) as
follows:
(1) Initiate a large number of Monte Carlo trajectories (based on
Eq. (10.47)), all starting from the point x.
(2) Integrate the potential operator along the trajectory up to the time
or the first time the trajectory hits the boundary D , whichever comes
first.
(3) If the trajectory does not hit the boundary, we include the first term on
the right side of Eq. (10.46) in the average by taking the product of the
exponential of minus the integral of the potential and the value of the
function hT computed at the final location of the trajectory.
(4) If the trajectory does hit the boundary before the time is reached, we
include the second term on the right side of Eq. (10.46) in the average
by taking the product of the exponential of minus the integral of the
potential and the value of the function k computed at the final location
of the trajectory, which for this case is on the boundary.
(5) Averaging over many trajectories then yields a statistical estimate of
the function f(, x)
We have tried this out for the analytical case of six non-interacting fermions
in a harmonic well. The trial function was taken as the analytical 1-DM with
a variable width different from the exact result. For the known solution, we
have available the function E(x, y), and that was inserted. The algorithm
converged nicely towards the exact result with increasing time . The sam-
pling domain was taken inside the first node of T . The exact (x, y), with
1.5
1
(x,0)
0.5
-4 -2 0 2 4
x
Fig. 10.1. FeynmanKac solution for the 1-DM for six non-interacting electrons
in a harmonic well. The y value was taken as 0. The exact 1-DM is shown as a solid
line, the trial 1-DM is a dashed line, and the diamonds are the numerical solution
generated by Monte Carlo sampling. All units are au.
y chosen at the center of the oscillator well, is shown in Fig. 10.1, along
with the trial function and the numerical solution.
10.4.4. Obtaining E(x,y)?

The above exercise only shows that, given our analytical knowledge of
the exact E(x, y), we can reproduce the known solution to the differential
equation. But a key issue is to somehow iteratively obtain a good approxi-
mation to E(x, y) using Monte Carlo sampling. Here we outline a possible
strategy for doing this; at this stage, the present discussion is speculative
and is work in progress. We can go back to Eq. (10.46) to see that what we
really have is an integral equation with the solution on both the left and
right sides. This is because E(x, y) = H/ and = f/T , and E(x, y)
appears in the potential operator c(x). The resulting problem is similar to
that encountered in the integral formulation of DFT (discussed in Ref. [20]
and first derived by Harris and Pratt [121]), and the optimized Thomas
Fermi theory (OTF) [117] provides a hint of how we might proceed.
In the OTF derivation, an effective potential was derived based on setting
the first-order variation of the density with a change in potential to zero:
= [u + u] [u] = 0. (10.48)
The above equation again uses the concept of functional, as we did in using a
variational approach for minimizing the action in Eq. (10.6). In Eq. (10.48)
the density is a functional of the potential u.
Here we carry out a similar strategy, namely by setting the variation
of f with respect to changes in the potential c(x) to zero. For the present
discussion we will omit the second term in Eq. (10.46) to keep the notation
simpler, but the results are easily generalized to include this term. To first
order, the functional variation in f is then

f(, x; c + c) = f(, x; c) h(Xx )<D cds
0

exp c(Xsx )ds . (10.49)
0
Here the current approximation to the potential is c, and the change is c.

We then set the derivative to 0 and can view this as an iterative process for
updating our approximation to E(x, y):

h(Xx )<D En (Xsx )ds exp cn (Xsx )ds
0 0
En+1 (x) = .

h(Xx )<D exp cn (Xsx )ds
0
(10.50)
The initial approximation to E = E0 would be taken as ET . Corresponding
boundary terms need to be added to the numerator and denominator to obtain
the full solution. (For the boundary term, the trajectory time D needs to be
included inside the averaging brackets.) Presumably we would perform the
sampling for short times to allow for local sampling around the point x.
This iterative scheme, or something like it, may provide a stochastic way to
effectively enforce the idempotency constraint that needs to be satisfied by
the 1-DM. It is certainly a different view of these constraints than that pro-
vided by directly enforcing Eq. (10.21) by some other numerical approach
(such as the McWeeny purification mentioned above).
The present discussion so far is conceptual, but work is underway to
try out the ideas outlined here. What would we have if all this, or related
ideas, worked out? We can imagine a huge number of independent random
walkers, one on each compute node, doing their work so as to compute
the electron density and nearby matrix elements of the 1-DM to construct
the total energy. The walkers would move in a localized region around the
chosen point y, up to near the first nodal surface in the trial 1-DM. We
would be attempting to compute the 1-DM, and the full 3N-dimensional
wave function would not be involved. Thus the methods cost would scale
strictly linearly with the number of electrons.
The ideas proposed here are in a sense intermediate between the standard
KohnSham eigenvalue approach and the integral formulation of DFT men-
tioned above [20]. That integral formulation reflects the true goal of DFT
namely, to represent the problem entirely in terms of the electron density.
There a path integral formalism is invoked, and the density is computed
from the inverse Laplace transform of the thermal Greens function. The
result is a beautiful formal expression in which the kinetic energy piece
contains an oscillatory integrand. Those oscillations can be problematic
computationally, however. The above approach leads to estimates of the
1-DM that do not contain the same problem with oscillations. Rather, the
sign problem is pushed into the trial 1-DM via the fixed-node approximation
using the trial 1-DM.
What are concerns with the above discussion? First, does Eq. (10.50)
provide an accurate estimate of E(x, y)? Is the iterative scheme of
Eq. (10.50) stable? What is the physical meaning of setting the first-order
variation to 0? Is the approximation of making the boundary values in
the second term of Eq. (10.46) equal to the trial function squared (that is
k = hT ) a reasonably accurate one for three-dimensional problems? Do we
need physical information from outside the first nodal surface to obtain an
accurate E(x, y) estimate?
Another question could be raised: for the low-dimensional 1-DM
problem, why not just use grid methods to solve for the 1-DM? Gen-
erally, grid methods might be considered more efficient for this task. A
first point to make is that, in the discussion here, we are not performing
a high-dimensional integration; rather, we are proposing to solve a differ-
ential equation using Monte Carlo methods (the FeynmanKac approach).
For example, we could attempt to solve Eq. (10.41) by iteration on a grid.
Several points can be made that argue against this approach, however: (1)
There are limitations in both accuracy and stability when integrating a
diffusion-type equation on a grid. (2) While the Monte Carlo approach
yields in principle the actual solution to the differential equation with suf-
ficient sampling, the constraints of normalization and idempotency would
have to be imposed during the iterations on a grid. Alternatively, we could
imagine minimizing the total energy by varying the 1-DM values, but we
are still stuck with the nonlocal constraint issue (Eq. (10.21)). The strategy
outlined above provides an alternative Monte Carlo approach for enforcing

those difficult constraints. (3) Finally, grid-based updates require higher
storage overhead and lead to increased latency (see point 5 in section 10.3.7
above) all 1-DM values must be stored and updated in sequence based
on a chosen solver method (Jacobi, etc.). Due to the above discussion,
the Monte Carlo approach presented here would thus appear to be more
appropriate for massively parallel implementation.
At the physical level, should these ideas pan out numerically, we would
have an approach for obtaining the 1-DM in DFT. In the introduction, we
discussed current limitations of DFT for accurate modeling of dispersion
interactions. Those limitations may get worked out, and likely will, over
time. A more ambitious goal would be to attempt to use related Monte Carlo
methods to get at the exact ground state energy. The next section poses a
challenge to possibly pursue that goal.
10.4.5. A pipe dream

We know that both the 2-DM [103] and the one-particle Greens function
(1-GF) [104] can in principle yield the exact ground state energy for a many-
electron system. Extensive work in quantum chemistry has been directed at
studies of the 2-DM [103]. The 1-GF can be viewed as the propagator that
examines the effect of inserting a particle at the location r and following
its motion until time twhen it is removed. Refs. [104] and [105] show that
this function can produce the exact ground state energy. Also, the real-time
one-particle Greens function G obeys the following differential equation
(Ref. [104], pg. 18):

1 2
i + G(1, 1 ) = (1 1 )
t1 2 x12

i dr2 v(r1 r2 )G2 (12; 1 2+ )|t2 =t1 (10.51)
The symbols 1, 1 here refer to the initial and final space-time points for the
particle trajectories, and the plus sign means t2 approaches t1 from above.
The last term contains the interaction potential v between the particles.
This equation can be transformed to the imaginary-time domain as we did
above. We see then that there is a differential equation for the 1-GF, but
it involves a higher-order Greens function, namely the 2-GF (leading to a
hierarchy of equations). Since we only need the 1-GF to express the exact
ground state energy, is there a way to estimate the 2-GF, perhaps by some
successive approximations scheme, so as to yield a good approximation to
the true 1-GF? Ref. [105] points out that the 1-GF can be thought of as the
time-dependent extension of the 1-DM the 1-DM is the zero-time limit
of the 1-GF. Also, in computing the ground state energy from the 1-GF,
we take the limit as 1 1 , but the time-dependent behavior of the 1-GF
is clearly important in allowing it to be used to compute the exact energy,
which is not available from the 1-DM.
The pipe dream would be to develop a stochastic differential equation
approach for solving Eq. (10.51) for the 1-GF, based on successive approx-
imations to the 2-GF. The 1-GF should possess localization properties
similar to the 1-DM, and thus relatively local sampling should again be
possible.
10.5. Summary
We summarize the conclusions of the chapter with a list of points made:
(1) We are reaching a wall in computer technology that will limit the
speed of commodity processors. Special purpose machines may need
to be developed in order to move forward in large-scale applications of
quantum mechanics to bio-molecules and nano-materials.
(2) We are reaching a wall in wave function approaches, as argued by
Kohn [115]. We should thus focus on methods based on the electron
density, the reduced density matrices, and/or Greens function-based
theories.
(3) Real-space grid methods were a first step towards exploiting the near-
locality of the 1-DM.
(4) But wave-function-based grid methods still require large computational
and storage overheads, and the special-purpose machine likely wont
allow for that.
(5) Stochastic differential equation methods for calculating the 1-DM, 2-
DM, and/or the 1-GF may provide a way out, along with clever physical
approximations.
(6) Monte Carlo is the most parallel algorithm (and typically requires
limited storage overhead), so lets exploit that, and sample only in local
regions of space to build up the 1-DM etc. We can imagine 107 walkers
on 107 compute nodes, all working so as to construct the reduced density
matrices in localized regions of space.
(7) Can we develop new ways of incorporating SDE methods into solving
for the one-particle Greens function of condensed matter theory?
(8) Has everything been tried? Certainly not. In fact, we might just be
entering a new computing era that may require some radical new ideas
in electronic structure calculations.
(9) A major goal is a new set of computational tools to accurately model
problems involving tens of thousands of atoms moving over nano-
or micro-second time scales. Applications might include drug design,
enzymes, ion channels, and the development of novel materials for
nano-scale engineering.
It is our hope that the ideas outlined here might serve to stimulate some alter-
native research directions for solving the Schrodinger equation on novel
massively parallel architectures yet to be developed.
Acknowledgments
We gratefully acknowledge the support of NSF grants CHE-0709560 and

CHE-1011746 for partial support of this research, and a grant of computing
time at the Ohio Supercomputer Center. We also thank Lawrence Pratt for
many helpful discussions.
Bibliography
[1] A. Szabo and N.S. Ostlund, Modern Quantum Chemistry (McGraw-Hill, NewYork,
1989).
[2] E.V. Lenthe and E. Baerends, J. Comput. Chem. 24, 1142 (2003).
[3] J. Soler, E. Artacho, J. Gale, A. Garcia, J. Junquera, P. Ordejon, and D. Sanchez-
Portal, J. Phys.-Cond. Matt. 14, 2745 (2002).
[4] M. Payne, M. Teter, D. Allan, T. Arias, and J. Joannopoulos, Rev. Mod. Phys. 64,
1045 (1992).
[5] M. Alemany, M. Jain, M. Tiago, Y. Zhou, Y. Saad, and J. Chelikowsky, Comput.
Phys. Commun. 177, 339 (2007).
[6] L. Kronik, A. Makmal, M. Tiago, M. Alemany, M. Jain, X. Huang, Y. Saad, and
J. Chelikowsky, Phys. Stat. Sol. B 243, 1063 (2006).
[7] T.L. Beck, Rev. Mod. Phys. 72, 1041 (2000).
[8] T.L. Beck, Rev. Comput. Chem. 26, 223 (2009).
[9] N. Modine, G. Zumbach, and E. Kaxiras, Phys. Rev. B 55, 10289 (1997).
[10] J. Fattebert, J. Phys.-Cond. Matt. 20, 294210 (2008).
[11] T. Torsti, T. Eirola, J. Enkovaara, T. Hakala, P. Havu, V. Havu, T. Hoynalanmaa,
J. Ignatius, M. Lyly, I. Makkonen, T. Rantala, J. Ruokolainen, K. Ruotsalainen, E.
Rasanen, H. Saarikoski, and M. Puska, Phys. Stat. Sol. B 243, 1016 (2006).
[12] D. Bowler, R. Choudhury, M. Gillan, and T. Miyazaki, Phys. Stat. Sol. B 243, 989
(2006).
[13] J. Bernholc, M. Hodak, and W. Lu, J. Phys.-Cond. Matt. 20, 294205 (2008).
[14] D. M. Ceperley and L. Mitas, Adv. Chem. Phys. 93, 1 (1996).
[15] W. Foulkes, L. Mitas, R. Needs, and G. Rajagopal, Rev. Mod. Phys. 73, 33 (2001).
[16] W. Lester, L. Mitas, and B. Hammond, Chem. Phys. Lett. 478, 1 (2009).
[17] M. Towler, Phys. Stat. Sol. B 243, 2573 (2006).
[18] R. Needs, M. Towler, N. Drummond, and P. Rios, J. Phys.-Cond. Matt. 22, 023201
(2010).
[19] J.B. Anderson, Rev. Comput. Chem. 13, 133 (1999).
[20] R.G. Parr and W.Yang, Density-Functional Theory of Atoms and Molecules (Oxford
University Press, Oxford, 1989).
[21] Y. Zhao, N. Schultz, and D. Truhlar, J. Chem. Theor. Comput. 2, 364 (2006).
[22] S. Goedecker, Rev. Mod. Phys. 71, 1085 (1999).
[23] B. Doser, D. Lambrecht, J. Kussmann, and C. Ochsenfeld, J. Chem. Phys. 130,
064107 (2009).
[24] T.L. Beck, M.E. Paulaitis, and L.R. Pratt, The Potential Distribution Theorem and
Models of Molecular Solutions (Cambridge University Press, New York, 2006).
[25] D. Asthagiri, L.R. Pratt, M.E. Paulaitis, and S.B. Rempe, J. Am. Chem. Soc. 126,
1285 (2004).
[26] D. Asthagiri, L.R. Pratt, and H.S. Ashbaugh, J. Chem. Phys. 119, 2702 (2003).
[27] D.M. Rogers and T.L. Beck, J. Chem. Phys. 129, 134505 (2008).
[28] D. Rogers and T. Beck, (2010) J. Chem. Phys. 132, 014505 (2010).
[29] Z. Zhao, D.M. Rogers, and T.L. Beck, J. Chem. Phys. 132, 014502 (2010).
[30] E. Guardia, I. Skarmoutsos, and M. Masia, J. Chem. Theor. Comput. 5, 1449 (2009).
[31] M. Dal Peraro, S. Raugei, P. Carloni, and M.L. Klein, Chem. Phys. Chem. 6, 1715
(2005).
[32] A.V. Marenich, R.M. Olson, A.C. Chamberlin, C.J. Cramer, and D.G. Truhlar, J.
Chem. Theor. Comput. 3, 2055 (2007).
[33] S.Y. Noskov and B. Roux, Biophys. Chem. 124, 279 (2006).
[34] D. Asthagiri, P.D. Dixit, S. Merchant, M.E. Paulaitis, L.R. Pratt, S.B. Rempe, and
S. Varma, Chem. Phys. Lett. 485, 1 (2010).
[35] A. Accardi and C. Miller, Nature 427, 803 (2004).
[36] J.Yin, Z. Kuang, U. Mahankali, and T. Beck, Proteins: Struct., Func., and Bioinform.
57, 414 (2004).
[37] Z. Kuang, U. Mahankali, and T.L. Beck, Proteins: Struct., Func., and Bioinform.
68, 26 (2007).
[38] D. Bucher, S. Raugei, L. Guidoni, M. Dal Peraro, U. Rothlisberger, P. Carloni, and
M. L. Klein, Biophys. Chem. 124, 292 (2006).
[39] S. Varma, D. Sabo, and S. Rempe, J. Molec. Biol. 376, 13 (2008).
[40] D. Asthagiri, L.R. Pratt, and J.D. Kress, Phys. Rev. E 68, 041505 (2003).
[41] M.J. McGrath, J.I. Siepmann, I.F. Kuo, C.J. Mundy, J. VandeVondele, J. Hutter,
F. Mohamed, and M. Krack, J. Phys. Chem. A 110, 640 (2006).
[42] J. Schmidt, J. VandeVondele, I. Kuo, D. Sebastiani, J. Siepmann, J. Hutter, and
C. Mundy, J. Phys. Chem. B 113, 11959 (2009).
[43] B. Santra, A. Michaelides, M. Fuchs, A. Tkatchenko, C. Filippi, and M. Scheffler,
J. Chem. Phys. 129, 194111 (2008).
[44] F. Sterpone, L. Spanu, L. Ferraro, S. Sorella, and L. Guidoni, J. Chem. Theor.
Comput. 4, 1428 (2008).
[45] J. Mahanty and B.W. Ninham, Dispersion Forces (Academic Press, New York,
1976).
[46] W. Kunz, P. Lo Nostro, and B.W. Ninham, Curr. Opin. Colloid Interface Sci. 9, 1
(2004).
[47] T.L. Beck, in Free energy calculations: Theory and applications in chemistry and
biology, edited by A. Pohorille and C. Chipot (Springer, New York, 2007) p. 389.
[48] E. Schwegler, J.C. Grossman, F. Gygi, and G. Galli, J. Chem. Phys. 121, 5400
(2004).
[49] P. Sit and N. Marzari, J. Chem. Phys. 122, 204510 (2005).
[50] CASINO available at http://www.tcm.phy.cam.ac.uk/mdt26/casino2.html.
[51] CHAMP available at http://pages.physics.cornell.edu/cyrus/champ.html.
[52] QWALK available at http://www.qwalk.org/wiki/index.php?title=Main Page.
[53] A. Badinski and R. Needs, Phys. Rev. E 76, 036707 (2007).
[54] A. Badinski, P. Haynes, J. Trail, and R. Needs, J. Phys.-Cond. Matt. 22, 074202
(2010).
[55] S. Chiesa, D.M. Ceperley, and S. Zhang, Phys. Rev. Lett. 94, 036404 (2005).
[56] R.M. Martin, Electronic Structure: Basic Theory and Practical Methods
(Cambridge University Press, New York, 2004).
[57] K. Hirose, T. Ono, Y. Fujimoto, and S. Tsukamoto, First-Principles Calculations in
Real-Space Formalism (Imperial College Press, London, 2005).
[58] T. Otsuka, T. Miyazaki, T. Ohno, D. Bowler, and M. Gillan, J. Phys.-Cond. Matt.
20, 294201 (2008).
[59] P. Haynes, C. Skylaris, A. Mostofi, and M. Payne, Phys. Stat. Sol. B 243, 2489
(2006).
[60] G. Feng, N. Wijesekera, and T. Beck, IEEE Trans. Nanotech. 6, 238 (2007).
[61] W. Kohn, Phys. Rev. Lett. 76, 3168 (1996).
[62] E. Prodan and W. Kohn, Proc. Natl. Acad. Sci USA 102, 11635 (2005).
[63] F. Shimojo, R. Kalia, A. Nakano, and P. Vashishta, Comput. Phys. Commun. 167,
151 (2005).
[64] Z. Zhao, J. Meza, and L. Wang, J. Phys.-Cond. Matt. 20, 294203 (2008).
[65] S. Burger and W. Yang, J. Phys.-Cond. Matt. 20, 294209 (2008).
[66] A. Brandt, Math. Comput. 31, 333 (1977).
[67] W.L. Briggs, V.E. Henson, and S.F. McCormick, A Multigrid Tutorial (SIAM,
Philadelphia, 2000).
[68] W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery, Numerical Recipes
in C (Cambridge University Press, Cambridge, 1992).
[69] R. Car and M. Parrinello, Phys. Rev. Lett. 55, 2471 (1985).
[70] G. Kresse and J. Furthmuller, Phys. Rev. B 54, 11169 (1996).
[71] S. Ismail-Beigi and T.A. Arias, Phys. Rev. Lett. 82, 2127 (1999).
[72] S. Goedecker, Phys. Rev. B 58, 3501 (1998).
[73] M. Alemany, M. Jain, L. Kronik, and J. Chelikowsky, Phys. Rev. B 69, 075101
(2004).
[74] S.C. Brenner and L.R. Scott, The Mathematical Theory of Finite Element Methods
(Springer, New York, 1994).
[75] J. Pask, B. Klein, P. Sterne, and C. Fong, Comput. Phys. Commun. 135, 1 (2001).
[76] J. Pask and P. Sterne, Model. Simul. Mater. Sci. Engr. 13, R71 (2005).
[77] C. Lanczos, The Variational Principles of Mechanics (University of Toronto Press,

Toronto, 1962).
[78] B.D. Reddy, Introductory Functional Analysis (Springer, New York, 1998).
[79] G.B. Arfken and H.J. Weber, Mathematical Methods for Physicists (Academic
Press, New York, 2001).
[80] G. Feng and T. Beck, Phys. Stat. Sol. B 243, 1054 (2006).
[81] E. Briggs, D. Sullivan, and J. Bernholc, Phys. Rev. B 54, 14362 (1996).
[82] E. Krotscheck, E. Hernandez, S. Janecek, M. Kaczmarski, and R. Wahl, Eur. Phys.
J. D 43, 173 (2007).
[83] H. Lee and M. Tuckerman, J. Chem. Phys. 126, 164501 (2007).
[84] A. Larsen, M. Vanin, J. Mortensen, K. Thygesen, and K. Jacobsen, Phys. Rev. B 80,
195112 (2009).
[85] A. Brandt, S. McCormick, and J. Ruge, SIAM J. Sci. Stat. Comput. 4, 244 (1983).
[86] S. Costiner and S. Taasan, Phys. Rev. E 51, 3704 (1995).
[87] S. Costiner and S. Taasan, Phys. Rev. E 52, 1181 (1995).
[88] N. Wijesekera, G. Feng, and T. Beck, Phys. Rev. B 75, 115101 (2007).
[89] E. Artacho, T.L. Beck, and E. Hernandez, Phys. Stat. Sol. B 243, 971 (2006).
[90] A. Castro, H. Appel, M. Oliveira, C. Rozzi, X. Andrade, R. Lorenzen, M. Marques,
E. Gross, and A. Rubio, Phys. Stat. Sol. B 243, 2465 (2006).
[91] K. Yabana, T. Nakatsukasa, J. Iwata, and G. Bertsch, Phys. Stat. Sol. B 243, 1121
(2006).
[92] J. Peddie, Gaming PC Market Bigger than Thought. Available at http://jonpeddie.
com/press-releases/details/gaming pc market bigger than thought/(2008).
[93] R.W. Brodersen, Technology, Architecture and Applications. Special Topic Session
SE2: Low-Voltage Design for Portable Systems. Institute of Electrical and Electronic
Engineers IEEE, Solid State Circuits Society, (2002).
[94] Jaguar. Available at http://en.wikipedia.org/wiki/Supercomputer.
[95] International Technology Roadmap for Semiconductors 2009 ITRS Winter Con-
ference Proceedings. Available at http://www.itrs.net/.
[96] J.L. Hennessy and D.A. Patterson, Computer Architecture, A Quantitative Approach
4th Ed. (Morgan Kaufmann Publishers, Burlington, 2007).
[97] J.W. Demmel, M.T. Heath, and H.A. van der Vorst, Acta Numerica 2, 111 (1993).
[98] CASE statement, C. Available at http://en.wikipedia.org/wiki/Switch statement.
[99] W.D. Hillis, Richard Feyman and the Connection Machine. Available at
http://www.longnow.org/essays/richard-feynman-connection-machine/.
[100] D. Alfe and M. Gillan, J. Phys.-Cond. Matt. 16, L305 (2004).
[101] A.J. Williamson, R.Q. Hood, and J.C. Grossman, Phys. Rev. Lett. 87, 246406 (2001).
[102] J. Kussmann and C. Ochsenfeld, J. Chem. Phys. 128, 134104 (2008).
[103] D.A. Mazziotti, Advances in Chemical Physics 134, 21 (2007).
[104] L.P. Kadanoff and G. Baym, Quantum Statistical Mechanics (Addison Wesley, New
York, 1962).
[105] E.K.U. Gross, E. Runge, and O. Heinonen, Many-Particle Theory (Adam Hilger,
New York, 1991).
[106] A.J. Coleman, Adv. Chem. Phys. 134, 3 (2007).
[107] E. Johnson and A. Becke, J. Chem. Phys. 128, 124108 (2008).
[108] J. Grafenstein and D. Cremer, J. Chem. Phys. 130, 124105 (2009).
[109] Q. Hill and C. Skylaris, Proc. Roy. Soc. A 465, 669 (2009).
[110] R.C. Grimm and R.G. Storer, J. Comput. Phys. 7, 134 (1971).
[111] C.W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry, and the
Natural Sciences (Springer-Verlag, Berlin, 1985).
[112] N.G. van Kampen, Stochastic Processes in Physics and Chemistry (Elsevier, New
York, 1992).
[113] M. Freidlin, Functional Integration and Partial Differential Equations (Princeton
University Press, Princeton, 1985).
[114] J. Anderson, J. Chem. Phys. 112, 9699 (2000).
[115] W. Kohn, Electronic Structure of Matter Wave Functions and Density Func-
tionals. Available at http://nobelprize.org/nobel prizes/chemistry/laureates/1998/
kohn-lecture.html.
[116] K. Dawson and N. March, J. Chem. Phys. 81, 5850 (1984).
[117] L. Pratt, G. Hoffman, and R. Harris, J. Chem. Phys. 92, 6687 (1990).
[118] L.S. Schulman, Techniques and Applications of Path Integration (Wiley-
[119] R.P. Feynman and A.R. Hibbs, Quantum Mechanics and Path Integrals
(McGraw-Hill, New York, 1965).
[120] M. Kac, Probability and Related Topics in Physical Sciences (American
Mathematical Society, Providence, 1959).
[121] R.A. Harris and L.R. Pratt, J. Chem. Phys. 82, 856 (1985).

Chapter 11
Changes in Dense Linear Algebra

Kernels: Decades-Long Perspective
Piotr Luszczeka , Jakub Kurzaka , and Jack Dongarraa,b

aUniversity
of Tennessee,
Knoxville, Tennessee, 37996
bOak Ridge National Laboratory
1, Bethel Valley Road, Oak Ridge TN 37831

and University of Manchester
Oxford Road, Manchester M13 9PL
Over the years, computational physics and chemistry served as an ongoing

source of problems that demanded the ever increasing performance from
hardware as well as the software that ran on top of it. Most of these
problems could be translated into solutions for systems of linear equa-
tions: the very topic of numerical linear algebra. Seemingly then, a set of
efficient linear solvers could be solving important scientific problems for
years to come. We argue that dramatic changes in hardware designs pre-
cipitated by the shifting nature of the marketplace of computer hardware
had a continuous effect on the software for numerical linear algebra. The
extraction of high percentages of peak performance continues to require
adaptation of software. If the past history of this adaptive nature of linear
algebra software is any guide then the future theme will feature changes
as well changes aimed at harnessing the incredible advances of the
evolving hardware infrastructure.
11.1. The Schrdinger Connection
Computational chemistry and physics have been in a continuous need of

high performance hardware and software to model molecular and particle
313
314 P. Luszczek, J. Kurzak, and J. Dongarra
interactions. Especially in high demand are numerical libraries for linear

algebra on dense matrices (matrices with most entries being non-zero).
Computational experiments of self-sustaining fusion reactions could
give us an informed perspective on how to build a device capable of
producing and controlling the high performance plasma [1]. Modeling the
heating response of plasma due to radio frequency (RF) waves in the fast
wave time scale leads to solving the generalized Helmholtz equation. The
time harmonic terms of effective approximations of the electric field, mag-
netic field, and distribution function as a time-averaged equilibrium satisfy
the equation. The Scientific Discovery through Advanced Computing
project (SciDAC) Numerical Computation of Wave Plasma-Interactions in
Multi-dimensional Systems developed and implemented a simulation code
that gives insight into how electromagnetic waves can be used for driving
current flow, heating and controlling instabilities in the plasma. The code is
called AORSA [2, 3, 4] and stands for All ORders Spectral Algorithm. The
resulting computation requires a solution of a system of linear equations
exceeding half a million unknowns [5].
In quantum chemistry, most of the scientific simulation codes result in
a numerical linear algebra problem that may readily be solved with the
ScaLAPACK library. For example, early versions of ParaGauss relied on
diagonalization of the KohnSham matrix and the parallelization method
of choice relied on the irreducible representations of the point group. The
submatrices diagonalize in parallel and the number of them depended on the
symmetry group. When using one of ScaLAPACKs parallel eigensolvers it
is possible to achieve speedup even for a KohnSham matrix with only one
block. A different use of the BLAS library occurs in UTChem an appli-
cation code that collects a number of methods that allow for accurate and
efficient calculations for computational chemistry of electronic structure
problems. Both the ground and excited states of molecular systems are
covered. In supporting a number of single-reference many-electron the-
ories such as configuration-interaction theory, coupled-cluster theory, and
MllerPlesset perturbation theory, UTChem derives working equations
using a symbolic manipulation program called Tensor Contraction Engine
(TCE). It automates the process of deriving final formulas and generation of
the execution program. The contraction of creation and annihilation oper-
ators according to Wicks theorem, consolidation of identical terms, and
reduction of the expressions into the form of tensor contractions controlled
by permutation operators are all done automatically by TCE. If tensor con-
tractions are treated as a collection of multi-dimensional summations of
the product of a few input arrays then the commutative, associative, and
distributive properties of the summation allow for a number of execution
Changes in Dense Linear Algebra Kernels: Decades-Long Perspective 315
orders, each of which having different execution rates when mapped to

a particular hardware architecture. Also, some of the execution orders
would result in calls to BLAS, which provides a substantial increase in
floating-point execution rate. The current TCE implementation generates
many-electron theories that are limited to non-relativistic HartreeFock
formulation with reference wave functions but it is possible to extend it to
relativistic two- and four-component reference wave functions.
11.2. A Stroll Down the Memory Lane
The key motivation in the design of efficient linear algebra algorithms for
advanced-architecture computers involves the storage and retrieval of data.
Designers wish to minimize the frequency with which data moves between
different levels of the memory hierarchy. Once data is in registers or the
fastest cache, all processing required for this data should be performed
before it gets evicted back to the main memory. Thus, the main algorithmic
approach for exploiting both vectorization and parallelism in our imple-
mentations uses block-partitioned algorithms, particularly in conjunction
with highly tuned kernels for performing matrix-vector and matrix-matrix
operations (the Level-2 and Level-3 BLAS). Block partitioning means that
the data is divided into blocks, each of which should fit within a cache
memory or a vector register file.
The computer architectures considered in this chapter are:
Vector machines
RISC computers with cache hierarchies
Parallel systems with distributed memory (the communication between
compute nodes happens by explicitly exchanging messages: the memory
is physically and programmatically distributed)
Multi-core computers
We briefly discuss these architectures in the order of these bullet points.

First, vector machines were introduced in the late 1970s and early 1980s.
They were able in one step to perform a single operation on a relatively large
number of operands stored in vector registers. Expressing matrix algorithms
as vector-vector operations was a natural fit for this type of machines.
However, some of the vector designs had a limited ability to load and store
the vector registers in main memory. A technique called chaining allowed
this limitation to be circumvented by moving data between the registers
before accessing main memory. Chaining required recasting linear algebra
in terms of matrix-vector operations.
Secondly, RISC computers were introduced in the late 1980s and early
1990s. While their clock rates might have been comparable to those of the
vector machines, the computing speed lagged behind due to their lack of
vector registers. Another deficiency was their creation of a deep memory
hierarchy with multiple levels of cache memory to alleviate the scarcity of
bandwidth that was, in turn, caused mostly by a limited number of memory
banks. The eventual success of this architecture is commonly attributed
to the right price point and astonishing improvements in performance over
time as predicted by Moores Law. With RISC computers, the linear algebra
algorithms had to be redone yet again. This time, the formulations had to
expose as many matrix-matrix operations as possible, which guaranteed
good cache reuse.
Thirdly, a natural way of achieving even greater performance levels
with both vector and RISC processors is by connecting them together with
a network and letting them cooperate to solve a problem bigger than would
be feasible on just one processor. This advance results in parallel systems
with distributed memory. Many hardware configurations followed this path,
so the matrix algorithms had to follow this as well. It was quickly discovered
that good local performance has to be combined with good global parti-
tioning of the matrices and vectors.
Any trivial divisions of matrix data quickly uncovered scalability
problems dictated by so-called Amdahls Law: the observation that the
time taken by the sequential portion of a computation provides the minimum
bound for the entire execution time, and therefore limits the gains achievable
from parallel processing. In other words, unless most of computations
can be done independently, the point of diminishing returns is reached,
and adding more processors to the hardware mix will not result in faster
processing.
Finally, the class of multi-core architectures includes both symmetric
multiprocessing (SMP) and single-chip multi-core machines, for the sake
of simplicity. Single-chip multi-core processors constitute a new paradigm
in commodity hardware: instead of increasing frequency and complexity
of a chip, new cores are added on a single die. This is probably an unfair
simplification, as the SMP machines usually have better memory systems.
But when applied to matrix algorithms, both yield good performance results
with very similar algorithmic approaches: these combine local cache reuse
and independent computation with explicit control of data dependences.
The initial success of vector computers in the 1970s was driven by raw
performance. The introduction of this type of computer systems started the
area of supercomputing (see Fig. 11.1). In the 1980s the availability of
Fig. 11.1. Performance of the fastest computer systems for the last six decades
compared to Moores Law. (DEC VAX-11/780 with its speed of 1 MIPS is not
features as it couldnt compete with its contemporaries from CDC and Cray.)
standard development environments and of application software packages

became more important. Next to performance these criteria determined
the success of massively parallel vector systems especially at industrial
customers. These and similar MPPs (massively parallel processors) became
successful in the early 1990s due to their better price/performance ratios,
which was enabled by the attack of the killer-micros (microprocessor
designs that eventually commoditized the processing market and canni-
balized the profits of custom processor companies and thus effectively
killed them in the marketplace). In the lower and medium market seg-
ments the MPPs were replaced by microprocessor based SMP systems in
the middle of the 1990s. Towards the end of the 1990s only the companies
that had entered the emerging markets for massive parallel database servers
and financial applications attracted enough business volume to be able to
support the hardware development for the numerical high end computing
market as well. Success in the traditional floating-point intensive engi-
neering applications was no longer sufficient for survival in the market.
The success of microprocessor based SMP concepts even for the very high-
end systems were the basis for the emerging cluster concepts in the early
2000s. Within the first half of this decade clusters of PCs and workstations
have become the prevalent architecture for many application areas in the
TOP500 (see www.top500.org) on all ranges of performance. However, the

Earth Simulator vector system demonstrated that many scientific applica-
tions could benefit greatly from other computer architectures. At the same
time there is renewed broad interest in the scientific HPC community for
new hardware architectures and new programming paradigms. The IBM
BlueGene/L system is one early example of a shifting design focus for
large-scale system.
11.3. A Decompositional Approach
At the basis of numerical solutions to linear systems of equations lies

a decompositional approach. The general idea is the following: given a
problem involving a matrix A, one factors or decomposes A into a product of
simpler matrices from which the problem can easily be solved. This divides
the computational problem into two parts: first determine an appropriate
decomposition, and then use it in solving the problem at hand. Consider
the problem of solving the linear system:
Ax = b, (11.1)
where A is a nonsingular matrix of order n. The decompositional approach

begins with the observation that it is possible to factor A in the form:
A = LU, (11.2)
where L is a lower triangular matrix (a matrix that has only zeros above the
diagonal) with ones on the diagonal, and U is upper triangular (with only
zeros below the diagonal). During the decomposition process, diagonal
elements of A (called pivots) are used to divide the elements below the
diagonal. If matrix A has a zero pivot, the process will break with division-
by-zero error. Also, small values of the pivots excessively amplify the
numerical errors of the process. So for numerical stability, the method
needs to interchange rows of the matrix or make sure pivots are as large
(in absolute value) as possible. This observation leads to a row permutation
matrix P and modifies the factored form to:
PA = LU. (11.3)
The solution can then be written in the form:
x = A1 b (11.4)
and the use of L and U factors suggests the following algorithm for solving
the system of equations:
1. Factor PA into LU (P is applied as we factor A).
2. Solve the system Ly = Pb (this comes by replacing Ux with y in
LUx = Pb).
3. Solve the system Ux = y.
This approach to matrix computations through decomposition has
proven very useful for several reasons. First, the approach separates the
computation into two stages: the computation of a decomposition, followed
by the use of the decomposition to solve the problem at hand. Such sepa-
ration can be important, for example, if different right hand sides are present
and need to be solved at different points in the process. The matrix needs
to be factored only once and reused for the different right hand sides. This
is particularly important because the factorization of A, step 1, requires
O(n3 ) operations, whereas the solutions, steps 2 and 3, require only O(n2 )
operations. Another aspect of the algorithms strength is in storage: the
L and U factors do not require extra storage, but can take over the space
occupied initially by the original matrix A.
For the discussion of coding this algorithm, we present only the com-
putationally intensive part of the process, which is step 1, the factorization
of the matrix.
11.4. Vector Processors
In the second half of the 1970s the introduction of vector computer systems
marked the beginning of modern Supercomputing. These systems offered
a performance advantage of at least one order of magnitude over con-
ventional systems of that time. Raw performance was the main if not
the only selling argument. In the first half of the 1980s the integration
of vector systems in conventional computing environments became more
important. Only the manufacturers, which provided standard programming
environments, operating systems and key applications, were successful
in getting industrial customers and survived. Performance was mainly
increased by improved chip technologies and by producing shared memory
multi-processor systems. They were able in one step to perform a single
operation on a relatively large number of operands stored in vector registers.
Expressing matrix algorithms as vector-vector operations was a natural fit
for this type of machines. However, some of the vector designs had a limited
ability to load and store the vector registers in main memory. A technique
called chaining allowed this limitation to be circumvented by moving data

between the registers before accessing main memory. Chaining required
recasting linear algebra in terms of matrix-vector operations.
Vector architectures exploit pipeline processing by running mathe-
matical operations on arrays of data in a simultaneous or pipelined fashion.
Most algorithms in linear algebra can be easily vectorized. Therefore, in
the late 1970s there was an effort to standardize vector operations for use
in scientific computations. The idea was to define some simple, frequently
used operations and implement them on various systems to achieve porta-
bility and efficiency. This package came to be known as the Level-1 Basic
Linear Algebra Subprograms (BLAS) or Level-1 BLAS.
The term Level-1 denotes vector-vector operations. As we will see,
Level-2 (matrix-vector operations), and Level-3 (matrix-matrix operations)
play important roles as well. In the 1970s, the algorithms of dense linear
algebra were implemented in a systematic way by the LINPACK project.
LINPACK is a collection of Fortran subroutines that analyze and solve linear
equations and linear least-squares problems. The package solves linear
systems whose matrices are general, banded, symmetric indefinite, sym-
metric positive definite, triangular, and square tridiagonal (only diagonal,
super-diagonal and sub-diagonal are present). In addition, the package com-
putes the QR (matrix Q is unitary or hermitian and R is upper trapezoidal)
and singular value decompositions of rectangular matrices and applies them
to least-squares problems. LINPACK uses column-oriented algorithms,
which increase efficiency by preserving locality of reference. By column
orientation, we mean that the LINPACK code always references arrays
down columns, not across rows. This is important since Fortran stores
arrays in column-major order. This means that as one proceeds down a
column of an array, the memory references proceed sequentially through
memory. Thus, if a program references an item in a particular block, the
next reference is likely to be in the same block.
The software in LINPACK was kept machine-independent partly
through the introduction of the Level-1 BLAS routines. Calling Level-1
BLAS did almost all of the computation. For each machine, the set of
Level-1 BLAS would be implemented in a machine-specific manner to
obtain high performance.
The Level-1 BLAS subroutines DAXPY, DSCAL, and IDAMAX are
used in the routine DGEFA (see Fig. 11.2).
It was presumed that the BLAS operations would be implemented in
an efficient, machine-specific way suitable for the computer on which the
subroutines were executed. On a vector computer, this could translate
subroutine dgefa(a,lda,n,ipvt,info)
integer lda,n,ipvt(1),info
double precision a(lda,1)
double precision t
integer idamax,j,k,kp1,l,nm1
c
c gaussian elimination with partial pivoting
c
info = 0
nm1 = n - 1
if (nm1 .lt. 1) go to 70
do 60 k = 1, nm1
kp1 = k + 1
c
c find l = pivot index
c
l = idamax(n-k+1,a(k,k),1) + k - 1
ipvt(k) = l
c
c zero pivot implies this column is already triangularized
c
if (a(l,k) .eq. 0.0d0) go to 40
c
c interchange if necessary
c
if (l .eq. k) go to 10
t = a(l,k)
a(l,k) = a(k,k)
a(k,k) = t
10 continue
c
c compute multipliers
c
t = -1.0d0/a(k,k)
call dscal(n-k,t,a(k+1,k),1)
c
c row elimination with column indexing
c
do 30 j = kp1, n
t = a(l,j)
if (l .eq. k) go to 20
a(l,j) = a(k,j)
a(k,j) = t
20 continue
call daxpy(n-k,t,a(k+1,k),1,a(k+1,j),1)
30 continue
go to 50
40 continue
info = k
50 continue
60 continue
70 continue
ipvt(n) = n
if (a(n,n) .eq. 0.0d0) info = n
return
end
Fig. 11.2. LINPACK variant of LU factorization (this is the original FORTRAN

66 code if LINPACK was written today it would have used Fortran 95).
into a simple, single vector operation. This avoided leaving the opti-
mization up to the compiler and explicitly exposing a performance-critical
operation.
In a sense, then, the beauty of the original code was regained with the
use of a new vocabulary to describe the algorithms: the BLAS. Over time,
the BLAS became a widely adopted standard and were most likely the first
to enforce two key aspects of software: modularity and portability. Again,
these are taken for granted today, but at the time they were not. One could
have the cake of compact algorithm representation and eat it too, because
the resulting Fortran code was portable.
Most algorithms in linear algebra can be easily vectorized. However,
to gain the most out of such architectures, simple vectorization is usually
not enough. Some vector computers are limited by having only one path
between memory and the vector registers. This creates a bottleneck if a
program loads a vector from memory, performs some arithmetic operations,
and then stores the results. In order to achieve top performance, the scope of
the vectorization must be expanded to facilitate chaining operations together
and to minimize data movement, in addition to using vector operations.
Recasting the algorithms in terms of matrix-vector operations makes it
easy for a vectorizing compiler to achieve these goals.
Thus, as computer architectures became more complex in the design of
their memory hierarchies, it became necessary to increase the scope of the
BLAS routines from Level-1 to Level-2 and Level-3.
11.5. RISC Processors
RISC computers were introduced in the late 1980s and early 1990s. While
their clock rates might have been comparable to those of the vector
machines, the computing speed lagged behind due to their lack of vector
registers. Another deficiency was their creation of a deep memory hierarchy
with multiple levels of cache memory to alleviate the scarcity of bandwidth
that was, in turn, caused mostly by a limited number of memory banks. The
eventual success of this architecture is commonly attributed to the right
price point and astonishing improvements in performance over time as pre-
dicted by Moores Law. With RISC computers, the linear algebra algorithms
had to be redone yet again. This time, the formulations had to expose as
many matrix-matrix operations as possible, which guaranteed good cache
reuse.
As mentioned before, the introduction in the late 1970s and early 1980s
of vector machines brought about the development of another variant of
algorithms for dense linear algebra. This variant was centered on the
multiplication of a matrix by a vector. These subroutines were meant to
give improved performance over the dense linear algebra subroutines in
LINPACK, which were based on Level-1 BLAS. In the late 1980s and
early 1990s, with the introduction of RISC-type microprocessors (the killer
micros) and other machines with cache-type memories, we saw the devel-
opment of LAPACK Level-3 algorithms for dense linear algebra. A Level-3
code is typified by the main Level-3 BLAS, which, in this case, is matrix
multiplication.
The original goal of the LAPACK project was to make the widely
used LINPACK library run efficiently on vector and shared-memory par-
allel processors. On these machines, LINPACK is inefficient because its
memory access patterns disregard the multilayered memory hierarchies
of the machines, thereby spending too much time moving data instead of
doing useful floating-point operations. LAPACK addresses this problem by
reorganizing the algorithms to use block matrix operations, such as matrix
multiplication, in the innermost loops (see the paper by E. Anderson and
J. Dongarra under Further Reading). These block operations can be opti-
mized for each architecture to account for its memory hierarchy, and so
provide a transportable way to achieve high efficiency on diverse modern
machines.
Here we use the term transportable instead of portable because,
for fastest possible performance, LAPACK requires that highly optimized
block matrix operations be implemented already on each machine. In other
words, the correctness of the code is portable, but high performance is
not if we limit ourselves to a single Fortran source code.
LAPACK can be regarded as a successor to LINPACK in terms of
functionality, although it doesnt always use the same function-calling
sequences. As such a successor, LAPACK was a win for the scientific
community because it could keep LINPACKs functionality while getting
improved use out of new hardware.
Most of the computational work in the algorithm from Fig. 11.3 is
contained in three routines:
DGEMM Matrix-matrix multiplication

DTRSM Triangular solve with multiple right hand sides
DGETF2 Unblocked LU factorization for operations within a block
column
One of the key parameters in the algorithm is the block size, called
NB here. If NB is too small or too large, poor performance can result
SUBROUTINE DGETRF( M, N, A, LDA, IPIV, INFO )

INTEGER INFO, LDA, M, N
INTEGER IPIV( * )
DOUBLE PRECISION A( LDA, * )
DOUBLE PRECISION ONE
PARAMETER ( ONE = 1.0D+0 )
INTEGER I, IINFO, J, JB, NB
EXTERNAL DGEMM, DGETF2, DLASWP, DTRSM, XERBLA
INTEGER ILAENV
EXTERNAL ILAENV
INTRINSIC MAX, MIN
INFO = 0
IF( M.LT.0 ) THEN
INFO = -1
ELSE IF( N.LT.0 ) THEN
INFO = -2
ELSE IF( LDA.LT.MAX( 1, M ) ) THEN
INFO = -4
END IF
IF( INFO.NE.0 ) THEN
CALL XERBLA( 'DGETRF', -INFO )
RETURN
END IF
IF( M.EQ.0 .OR. N.EQ.0 ) RETURN
NB = ILAENV( 1, 'DGETRF', ' ', M, N, -1, -1 )
IF( NB.LE.1 .OR. NB.GE.MIN( M, N ) ) THEN
CALL DGETF2( M, N, A, LDA, IPIV, INFO )
ELSE
DO 20 J = 1, MIN( M, N ), NB
JB = MIN( MIN( M, N )-J+1, NB )
* Factor diagonal and subdiagonal blocks and test for exact
* singularity.
CALL DGETF2( M-J+1, JB, A( J, J ), LDA, IPIV( J ), IINFO )
* Adjust INFO and the pivot indices.
IF( INFO.EQ.0 .AND. IINFO.GT.0 ) INFO = IINFO + J - 1
DO 10 I = J, MIN( M, J+JB-1 )
IPIV( I ) = J - 1 + IPIV( I )
10 CONTINUE
* Apply interchanges to columns 1:J-1.
CALL DLASWP( J-1, A, LDA, J, J+JB-1, IPIV, 1 )
IF( J+JB.LE.N ) THEN
* Apply interchanges to columns J+JB:N.
CALL DLASWP( N-J-JB+1, A( 1, J+JB ), LDA, J, J+JB-1, IPIV, 1 )
* Compute block row of U.
CALL DTRSM( 'Left', 'Lower', 'No transpose', 'Unit', JB,
$ N-J-JB+1, ONE, A( J, J ), LDA, A( J, J+JB ), LDA )
IF( J+JB.LE.M ) THEN
* Update trailing submatrix.
CALL DGEMM( 'No transpose', 'No transpose', M-J-JB+1,
$ N-J-JB+1, JB, -ONE, A( J+JB, J ), LDA,
$ A( J, J+JB ), LDA, ONE, A( J+JB, J+JB ), LDA )
END IF
END IF
20 CONTINUE
END IF
RETURN
END
Fig. 11.3. LAPACKs LU factorization routine DGETRF (FORTRAN 77

coding.)
hence the importance of the ILAENV function, whose standard implemen-

tation was meant to be replaced by a vendor implementation encapsulating
machine-specific parameters upon installation of the LAPACK library. At
any given point of the algorithm, NB columns or rows are exposed to a
well-optimized Level-3 BLAS. If NB is one, the algorithm is equivalent in
performance and memory access patterns to the LINPACKs version.
Matrix-matrix operations offer the proper level of modularity for perfor-
mance and transportability across a wide range of computer architectures,
including parallel systems with memory hierarchy. This enhanced perfor-
mance is primarily due to a greater opportunity for reusing data. There are
numerous ways to accomplish this reuse of data to reduce memory traffic
and to increase the ratio of floating-point operations to data movement
through the memory hierarchy. This improvement can bring a three- to
ten-fold improvement in performance on modern computer architectures.
The jury is still out concerning the productivity of writing and reading the
LAPACK code: how hard is it to generate the code from its mathematical
description? The use of vector notation in LINPACK is arguably more
natural than LAPACKs matrix formulation. The mathematical formulas
that describe algorithms are usually more complex if only matrices are
used, as opposed to mixed vector-matrix notation.
11.6. Clusters
Traditional design focus for MPP systems was the very high end of perfor-
mance. In the early 1990s the SMP systems of various workstation man-
ufacturers as well as the IBM SP series, which targeted the lower and
medium market segments, gained great popularity (see Fig. 11.4). Their
price/performance ratios were better due to the missing overhead in the
design for support of the very large configurations and due to cost advan-
tages of the larger production numbers. Due to the vertical integration of
performance it was no longer economically feasible to produce and focus
on the highest end of computing power alone. The design focus for new
systems shifted to the market of medium performance systems.
The acceptance of MPP systems not only for engineering applications
but also for new commercial applications especially for database applica-
tions emphasized different criteria for market success such as the stability of
the system, continuity of the manufacturer and price/performance. Success
in commercial environments became a new important requirement for a
successful Supercomputer business towards the end of the 1990s. Due to
these factors and the consolidation in the number of vendors in the market,
500
400
300 Vector
Systems
Single Proc.
SMP
200 Const.
Cluster
MPP
100
0
93
98
03
08
19
19
20
20
Fig. 11.4. Main architectural categories seen in the TOP500.
hierarchical systems built with components designed for the broader com-
mercial market did replace homogeneous systems at the very high end
of performance. The marketplace adopted clusters of SMPs readily, while
academic research focused on clusters of workstations and PCs.
In the early 2000s, clusters built with off-the-shelf components gained
more and more attention not only as academic research object but also com-
puting platforms with end-users of HPC computing systems. By 2004 these
groups of clusters represent the majority of new systems on the TOP500
in a broad range of application areas. One major consequence of this trend
was the rapid rise in the utilization of Intel processors in HPC systems.
While virtually absent in the high end at the beginning of the decade, Intel
processors are now used in the majority of HPC systems. Clusters in the
1990s were mostly self-made system designed and built by small groups
of dedicated scientist or application experts. This changed rapidly as soon
as the market for clusters based on PC technology matured. Nowadays the
large majority of TOP500-class clusters are manufactured and integrated by
either a few traditional large HPC manufacturers such as IBM or Hewlett-
Packard or numerous small, specialized integrators of such systems.
At the end of the 1990s clusters were common in academia but mostly as
research objects and not primarily as general purpose computing platforms
for applications. Most of these clusters were of comparable small scale and
as a result the November 1999 edition of the TOP500 listed only seven
cluster systems. This changed dramatically as industrial and commercial
customers started deploying clusters as soon as applications with less

stringent communication requirements permitted them to take advantage of
the better price/performance ratio-roughly an order of magnitude- of com-
modity based clusters.At the same time all major vendors in the HPC market
started selling this type of cluster to their customer base. In November 2004,
clusters were the dominant architectures in the TOP500 with 294 systems
at all levels of performance (see Fig. 11.4). Companies such as IBM and
Hewlett-Packard sell the majority of these clusters and a large number of
them are installed at commercial and industrial customers.
In addition, there still is generally a large difference in the usage of
clusters and their more integrated counterparts: clusters are mostly used
for capacity computing while the integrated machines primarily are used
for capability computing. The largest supercomputers are used for capa-
bility or turnaround computing where the maximum processing power is
applied to a single problem. The goal is to solve a larger problem, or to
solve a single problem in a shorter period of time. Capability computing
enables the solution of problems that cannot otherwise be solved in a rea-
sonable period of time (for example, by moving from a two-dimensional
to a three-dimensional simulation, using finer grids, or using more realistic
models). Capability computing also enables the solution of problems with
real-time constraints (e.g. predicting weather). The main figure of merit is
time to solution. Smaller or cheaper systems are used for capacity com-
puting, where smaller problems are solved. Capacity computing can be
used to enable parametric studies or to explore design alternatives; it is
often needed to prepare for more expensive runs on capability systems.
Capacity systems will often run several jobs simultaneously. The main
figure of merit is sustained performance per unit cost. Traditionally, vendors
of large supercomputer systems have learned to provide for the capacity
mode of operation as the precious resources of their systems were required
to be used as effectively as possible. By contrast, Beowulf clusters are
mostly operated through the Linux operating system (a small minority using
Microsoft Windows) where these operating systems either lack the tools or
these tools are relatively immature to use a cluster effectively for capability
computing. However, as clusters become on average both larger and more
stable in terms of continuous operation, there is a trend to use them also as
computational capability servers.
There are a number of choices of communication networks available
in clusters. Of course 100 Mb/s Ethernet or Gigabit Ethernet is always
possible, which is attractive for economic reasons, but it has the drawback
of a high latency (100 s) the time it takes to send the shortest message.
Alternatively, there are, for instance, networks that operate from user space,
like Myrinet, Infiniband. The network speeds as shown by these networks

are more or less on par with some integrated parallel systems. So, possibly
apart from the speed of the processors and of the software that is provided
by the vendors of traditional integrated supercomputers, the distinction
between clusters and the class of custom capability machines becomes
rather small and will, without a doubt, decrease further in the coming years.
And the advances of the Ethernet standard into the 100 Gb/s territory with
latencies well below 10 s make it even more so.
LAPACK was designed to be highly efficient on vector processors,
high-performance superscalar workstations, and shared-memory mul-
tiprocessors. LAPACK can also be used satisfactorily on all types of
scalar machines (PCs, workstations, and mainframes). However, LAPACK
in its present form is less likely to give good performance on other
types of parallel architectures for example, massively parallel Single
Instruction Multiple Data (SIMD) machines, or Multiple Instruction Mul-
tiple Data (MIMD) distributed-memory machines. The ScaLAPACK effort
was intended to adapt LAPACK to these new architectures.
By creating the ScaLAPACK software library, we extended the
LAPACK library to scalable MIMD, distributed-memory, highly parallel
computers. For such machines, the memory hierarchy includes the off-
processor memory of other processors, in addition to the hierarchy of reg-
isters, cache, and local memory on each processor.
Like LAPACK, the ScaLAPACK routines are based on block-
partitioned algorithms in order to minimize the frequency of data movement
between different levels of the memory hierarchy. The fundamental building
blocks of the ScaLAPACK library are distributed-memory versions of the
Level-2 and Level-3 BLAS, and a set of Basic Linear Algebra Commu-
nication Subprograms (BLACS) for communication tasks that arise fre-
quently in parallel linear algebra computations. In the ScaLAPACK rou-
tines, all interprocessor communication occurs within the distributed BLAS
and the BLACS, so the source code of the top software layer of ScaLAPACK
looks very similar to that of LAPACK (see Fig. 11.5).
In order to simplify the design of ScaLAPACK, and because the BLAS
have proven to be very useful tools outside LAPACK, we chose to build
a Parallel BLAS, or PBLAS (described in the paper by Choi et al.; see
Further Reading), whose interface is as similar to the BLAS as possible.
This decision has permitted the ScaLAPACK code to be quite similar, and
sometimes nearly identical, to the analogous LAPACK code.
It was our aim that the PBLAS would provide a distributed memory
standard, just as the BLAS provided a shared memory standard. This would
SUBROUTINE PDGETRF( M, N, A, IA, JA, DESCA, IPIV, INFO )

INTEGER BLOCK_CYCLIC_2D, CSRC_, CTXT_, DLEN_, DTYPE_, LLD_, MB_, M_, NB_, N_, RSRC_
PARAMETER ( BLOCK_CYCLIC_2D = 1, DLEN_ = 9, DTYPE_ = 1, CTXT_ = 2, M_ = 3, N_ = 4, MB_ = 5, NB_ = 6,
$ RSRC_ = 7, CSRC_ = 8, LLD_ = 9 )
DOUBLE PRECISION ONE
PARAMETER ( ONE = 1.0D+0 )
CHARACTER COLBTOP, COLCTOP, ROWBTOP
INTEGER I, ICOFF, ICTXT, IINFO, IN, IROFF, J, JB, JN, MN, MYCOL, MYROW, NPCOL, NPROW
INTEGER IDUM1( 1 ), IDUM2( 1 )
EXTERNAL BLACS_GRIDINFO, CHK1MAT, IGAMN2D, PCHK1MAT, PB_TOPGET, PB_TOPSET, PDGEMM, PDGETF2, PDLASWP, PDTRSM,
$ PXERBLA
INTEGER ICEIL
EXTERNAL ICEIL
INTRINSIC MIN, MOD
* Get grid parameters
ICTXT = DESCA( CTXT_ )
CALL BLACS_GRIDINFO( ICTXT, NPROW, NPCOL, MYROW, MYCOL )
* Test the input parameters
INFO = 0
IF( NPROW.EQ.-1 ) THEN
INFO = -(600+CTXT_)
ELSE
CALL CHK1MAT( M, 1, N, 2, IA, JA, DESCA, 6, INFO )
IF( INFO.EQ.0 ) THEN
IROFF = MOD( IA-1, DESCA( MB_ ) )
ICOFF = MOD( JA-1, DESCA( NB_ ) )
IF( IROFF.NE.0 ) THEN
INFO = -4
ELSE IF( ICOFF.NE.0 ) THEN
INFO = -5
ELSE IF( DESCA( MB_ ).NE.DESCA( NB_ ) ) THEN
INFO = -(600+NB_)
END IF
END IF
CALL PCHK1MAT( M, 1, N, 2, IA, JA, DESCA, 6, 0, IDUM1, IDUM2, INFO )
END IF
IF( INFO.NE.0 ) THEN
CALL PXERBLA( ICTXT, 'PDGETRF', -INFO )
RETURN
END IF
IF( DESCA( M_ ).EQ.1 ) THEN
IPIV( 1 ) = 1
RETURN
ELSE IF( M.EQ.0 .OR. N.EQ.0 ) THEN
RETURN
END IF
* Split-ring topology for the communication along process rows
CALL PB_TOPGET( ICTXT, 'Broadcast', 'Rowwise', ROWBTOP )
CALL PB_TOPGET( ICTXT, 'Broadcast', 'Columnwise', COLBTOP )
CALL PB_TOPGET( ICTXT, 'Combine', 'Columnwise', COLCTOP )
CALL PB_TOPSET( ICTXT, 'Broadcast', 'Rowwise', 'S-ring' )
CALL PB_TOPSET( ICTXT, 'Broadcast', 'Columnwise', ' ' )
CALL PB_TOPSET( ICTXT, 'Combine', 'Columnwise', ' ' )
* Handle the first block of columns separately
MN = MIN( M, N )
IN = MIN( ICEIL( IA, DESCA( MB_ ) )*DESCA( MB_ ), IA+M-1 )
JN = MIN( ICEIL( JA, DESCA( NB_ ) )*DESCA( NB_ ), JA+MN-1 )
JB = JN - JA + 1
* Factor diagonal and subdiagonal blocks and test for exact singularity.
CALL PDGETF2( M, JB, A, IA, JA, DESCA, IPIV, INFO )
IF( JB+1.LE.N ) THEN
* Apply interchanges to columns JN+1:JA+N-1.
CALL PDLASWP('Forward', 'Rows', N-JB, A, IA, JN+1, DESCA, IA, IN, IPIV )
CALL PDTRSM( 'Left', 'Lower', 'No transpose', 'Unit', JB, N-JB, ONE, A, IA, JA, DESCA, A, IA, JN+1, DESCA )
IF( JB+1.LE.M ) THEN
CALL PDGEMM( 'No transpose', 'No transpose', M-JB, N-JB, JB,-ONE, A, IN+1, JA, DESCA, A, IA, JN+1,DESCA,
$ ONE, A, IN+1, JN+1, DESCA )
END IF
END IF
* Loop over the remaining blocks of columns.
Fig. 11.5. ScaLAPACK variant of LU factorization (FORTRAN 77 coding makes

the code overly verbose due to lack of object oriented capabilities that could have
hidden much of the complexity).
DO 10 J = JN+1, JA+MN-1, DESCA( NB_ )

JB = MIN( MN-J+JA, DESCA( NB_ ) )
I = IA + J - JA
* Factor diagonal and subdiagonal blocks and test for exact singularity.
CALL PDGETF2( M-J+JA, JB, A, I, J, DESCA, IPIV, IINFO )
IF( INFO.EQ.0 .AND. IINFO.GT.0 ) INFO = IINFO + J - JA
* Apply interchanges to columns JA:J-JA.
CALL PDLASWP('Forward', 'Rowwise', J-JA, A, IA, JA, DESCA, I,I+JB-1, IPIV)
IF( J-JA+JB+1.LE.N ) THEN
* Apply interchanges to columns J+JB:JA+N-1.
CALL PDLASWP( 'Forward', 'Rowwise', N-J-JB+JA, A, IA, J+JB,DESCA, I, I+JB-1, IPIV )
CALL PDTRSM( 'Left', 'Lower', 'No transpose', 'Unit', JB,N-J-JB+JA, ONE, A, I, J, DESCA, A, I, J+JB,
$ DESCA )
IF( J-JA+JB+1.LE.M ) THEN
CALL PDGEMM( 'No transpose', 'No transpose', M-J-JB+JA,N-J-JB+JA, JB, -ONE, A, I+JB, J, DESCA, A,
$ I, J+JB, DESCA, ONE, A, I+JB, J+JB, DESCA )
END IF
END IF
10 CONTINUE
IF( INFO.EQ.0 ) INFO = MN + 1
CALL IGAMN2D(ICTXT, 'Rowwise', ' ', 1, 1, INFO, 1, IDUM1,IDUM2, -1,-1, MYCOL)
IF( INFO.EQ.MN+1 ) INFO = 0
CALL PB_TOPSET( ICTXT, 'Broadcast', 'Rowwise', ROWBTOP )
CALL PB_TOPSET( ICTXT, 'Broadcast', 'Columnwise', COLBTOP )
CALL PB_TOPSET( ICTXT, 'Combine', 'Columnwise', COLCTOP )
RETURN
END
Fig. 11.5. (Continued)
simplify and encourage the development of high-performance and portable

parallel numerical software, as well as providing manufacturers with just
a small set of routines to be optimized. The acceptance of the PBLAS
requires reasonable compromises between competing goals of functionality
and simplicity.
The PBLAS operate on matrices distributed in a two-dimensional block
cyclic layout. Because such a data layout requires many parameters to fully
describe the distributed matrix, we have chosen a more object-oriented
approach and encapsulated these parameters in an integer array called an
array descriptor. An array descriptor includes:
The descriptor type

The BLACS context (a virtual space for messages that is created to avoid
collisions between logically distinct messages)
The number of rows in the distributed matrix
The number of columns in the distributed matrix
The row block size
The column block size
The process row over which the first row of the matrix is distributed
The process column over which the first column of the matrix is dis-
tributed
The leading dimension of the local array storing the local blocks
By using this descriptor, a call to a PBLAS routine is very similar to a

call to the corresponding BLAS routine:
CALL DGEMM ( TRANSA, TRANSB, M, N, K, ALPHA,

A( IA, JA ), LDA, B( IB, JB ), LDB, BETA,
C( IC, JC ), LDC )
CALL PDGEMM( TRANSA, TRANSB, M, N, K, ALPHA, A,
IA, JA, DESC A, B, JB, DESC B, BETA, C, IC, JC,
DESC C )
DGEMM computes C = BETA * C + ALPHA * op( A ) * op( B ),

where op(A) is either A or its transpose depending on TRANSA, op(B)
is similar, op(A) is M-by-K, and op(B) is K-by-N. PDGEMM is the
same, with the exception of the way submatrices are specified. To pass
the submatrix starting at A(IA,JA) to DGEMM, for example, the actual
argument corresponding to the formal argument A is simply A(IA,JA).
PDGEMM, on the other hand, needs to understand the global storage
scheme of A to extract the correct submatrix, so IA and JA must be passed in
separately.
DESC A is the array descriptor for A. The parameters describing
the matrix operands B and C are analogous to those describing A. In a
truly object-oriented environment, matrices and DESC A would be syn-
onymous. However, this would require language support and detract from
portability.
Using message passing and scalable algorithms from the ScaLAPACK
library makes it possible to factor matrices of arbitrarily increasing size,
given machines with more processors. By design, the library computes more
than it communicates, so for the most part, data stay locally for processing
and travels only occasionally across the interconnect network.
But the number and types of messages exchanged between processors
can sometimes be hard to manage. The context associated with every dis-
tributed matrix lets implementations use separate universes for message
passing. The use of separate communication contexts by distinct libraries
(or distinct library invocations) such as the PBLAS insulates communi-
cation internal to the library from external communication. When more
than one descriptor array is present in the argument list of a routine in
the PBLAS, the individual BLACS context entries must be equal. In other
words, the PBLAS do not perform inter-context operations.
In the performance sense, ScaLAPACK did to LAPACK what LAPACK
did to LINPACK: it broadened the range of hardware where LU factor-
ization (and other codes) could run efficiently. In terms of code elegance,
the ScaLAPACKs changes were much more drastic: the same mathematical
operation now required large amounts of tedious work. Both the users
and the library writers were now forced into explicitly controlling data
storage intricacies, because data locality became paramount for perfor-
mance. The victim was the readability of the code, despite efforts to mod-
ularize the code according to the best software engineering practices of
the day.
11.7. Multicore Processors
The advent of multi-core chips brought about a fundamental shift in the

way software is produced. Dense linear algebra is no exception. The good
news is that LAPACKs LU factorization runs on a multi-core system and
can even deliver a modest increase of performance if multi-threaded BLAS
are used. In technical terms, this is the fork-join model of computation:
each call to BLAS (from a single main thread) forks a suitable number
of threads (parallel units of executions that share memory and are often
scheduled by the operating system), which perform the work on each core
and then join the main thread of computation. The fork-join model implies
a synchronization point at each join operation.
The bad news is that the LAPACKs fork-join algorithm gravely impairs
scalability even on small multi-core computers that do not have the memory
systems available in SMP systems. The inherent scalability flaw is the heavy
synchronization in the fork-join model: only a single thread is allowed to
perform the significant computation that occupies the critical section of
the code, leaving other threads idle. That results in lock-step execution: all
threads have to wait for the slowest one among them. It also prevents hiding
of inherently sequential portions of the code behind parallel ones. In other
words, the threads are forced to perform the same operation on different
data. If there is not enough data for some threads, they will have to stay idle
and wait for the rest of the threads that perform useful work on their data.
Clearly, another version of the LU algorithm is needed such that would
allow threads to stay busy all the time by possibly making them perform
different operations during some portion of the execution.
The multithreaded version of the algorithm recognizes the existence
of a so-called critical path in the algorithm: a portion of the code whose
execution depends on previous calculations and can block the progress of
the algorithm. The LAPACKs LU does not treat this critical portion of the
code in any special way: the DGETF2 subroutine is called by a single thread
and doesnt allow much parallelization even at the BLAS level. While one
thread calls this routine, the other ones wait idly. And since the performance
of DGETF2 is bound by memory bandwidth (rather than processor speed),
this bottleneck will exacerbate scalability problems as systems with more
cores are introduced.
The multithreaded version of the algorithm attacks this problem head-on
by introducing the notion of look-ahead: calculating things ahead of time
to avoid potential stagnation in the progress of the computations. This of
course requires additional synchronization and bookkeeping not present in
the previous versions a trade-off between code complexity and perfor-
mance. Another aspect of the multi-threaded code is the use of recursion
in the panel factorization. It turns out that the use of recursion can give
even greater performance benefits for tall panel matrices than it does for
the square ones.
The algorithm is the same for each thread (the SIMD paradigm), and the
matrix data is partitioned among threads in a cyclic manner using panels
with pw columns in each panel (except maybe the last). The pw parameter
corresponds to the blocking parameter NB of LAPACK. The difference is the
logical assignment of panels (blocks of columns) to threads. (Physically,
all panels are equally accessible, because the code operates in a shared
memory regime.) The benefits of blocking in a thread are the same as they
were in LAPACK: better cache reuse and less stress on the memory bus.
Assigning a portion of the matrix to a thread seems an artificial requirement
at first, but it simplifies the code and the bookkeeping data structures; most
importantly, it provides better memory affinity. It turns out that multi-core
chips are not symmetric in terms of memory access bandwidth, so min-
imizing the number of reassignments of memory pages to cores directly
benefits performance.
The standard components of LU factorization are represented by the
pfactor() and pupdate() functions (see Fig. 11.6). As one might
expect, the former factors a panel, whereas the latter updates a panel using
one of the previously factored panels.
The main loop makes each thread iterate over each panel in turn. If
necessary, the panel is factored by the owner thread while other threads
wait (if they happen to need this panel for their updates).
The look-ahead logic is inside the nested loop (prefaced by the comment
for each panel to be updated) that replaces DGEMM or PDGEMM from
previous algorithms. Before each thread updates one of its panels, it checks
whether its already feasible to factor its first unfactored panel. This min-
imizes the number of times the threads have to wait because each thread
constantly attempts to eliminate the potential bottleneck.
void SMP_dgetrf(int n, double *a, int lda, int *ipiv, int pw,
int tid, int tsize, int *pready,ptm *mtx, ptc *cnd) {
int pcnt, pfctr, ufrom, uto, ifrom, p;
double *pa = a, *pl, *pf, *lp;
pcnt = n / pw; /* number of panels */
pfctr = tid + (tid ? 0 : tsize); /* first panel that should be factored by this thread after the very first panel
(number 0) gets factored */
/* this is a pointer to the last panel */
lp = a + (size_t)(n - pw) * (size_t)lda;
/* for each panel (that is used as source of updates) */
for (ufrom = 0; ufrom < pcnt; ufrom++, pa += (size_t)pw * (size_t)(lda + 1)){
p = ufrom * pw; /* column number */
/* if the panel to be used for updates has not been factored yet; 'ipiv' does not be consulted, but it is to
possibly avoid accesses to 'pready'*/
if (! ipiv[p + pw - 1] || ! pready[ufrom]) {
if (ufrom % tsize == tid) { /* if this is this thread's panel */
pfactor( n - p, pw, pa, lda, ipiv + p, pready, ufrom, mtx, cnd );
} else if (ufrom < pcnt - 1) { /* if this is not the last panel */
LOCK( mtx );
while (! pready[ufrom]) { WAIT( cnd, mtx ); }
UNLOCK( mtx );
}
}
/* for each panel to be updated */
for (uto = first_panel_to_update( ufrom, tid, tsize ); uto < pcnt; uto += tsize) {
/* if there are still panels to factor by this thread and preceding panel has been factored; test to 'ipiv' could
be skipped but is in there to decrease number of accesses to 'pready' */
if (pfctr < pcnt && ipiv[pfctr * pw - 1] && pready[pfctr - 1]) {
/* for each panel that has to (still) update panel 'pfctr' */
for (ifrom = ufrom + (uto > pfctr ? 1 : 0); ifrom < pfctr; ifrom++) {
p = ifrom * pw;
pl = a + (size_t)p * (size_t)(lda + 1);
pf = pl + (size_t)(pfctr - ifrom) * (size_t)pw * (size_t)lda;
pupdate( n - p, pw, pl, pf, lda, p, ipiv, lp );
}
p = pfctr * pw;
pl = a + (size_t)p * (size_t)(lda + 1);
pfactor( n - p, pw, pl, lda, ipiv + p, pready, pfctr, mtx, cnd );
pfctr += tsize; /* move to this thread's next panel */
}
/* if panel 'uto' hasn't been factored (if it was, it certainly has been updated, so no update is necessary) */
if (uto > pfctr || ! ipiv[uto * pw]) {
p = ufrom * pw;
pf = pa + (size_t)(uto - ufrom) * (size_t)pw * (size_t)lda;
pupdate( n - p, pw, pa, pf, lda, p, ipiv, lp );
}
}
}
Fig. 11.6. Factorization for multi-threaded execution (C code.)
As was the case for ScaLAPACK, the multithreaded version detracts

from the inherent elegance of the LAPACKs version. Also in the same
spirit, performance is the main culprit: LAPACKs code will not run effi-
ciently on machines with ever-increasing numbers of cores. Explicit control
of execution threads at the LAPACK level rather than the BLAS level is
critical: parallelism cannot be encapsulated in a library call. The only good
news is that the code is not as complicated as ScaLAPACKs, and efficient
BLAS can still be put to a good use.
11.8. Multicore Processors Redux
The multicore processors do not resemble the SMP systems of the past,
nor do they resemble distributed memory systems. In comparison to SMPs,
multicores are much more starved for memory due to the fast increase in
the number of cores, which is not followed by a proportional increase in
bandwidth. Owing to that, data access locality is of much higher importance
in case of multicores. At the same time, they do follow to a large extent the
memory model where the main memory serves as a central (not distributed)
repository for data. For those reasons, the best performing algorithms or
multicores happen to be parallel versions of what used to be known as
out-of-core algorithms (algorithms developed in the past for situations
where data does not fit in the main memory and has to be explicitly moved
between the memory and the disc).
In dense linear algebra the Tile Algorithms are direct descendants of
out-of-core algorithms. The Tile Algorithms are based on the idea of
processing the matrix by square submatrices, referred to as tiles, of rel-
atively small size. This makes the operation efficient in terms of cache
and TLB use. The Cholesky factorization lends itself readily to tile for-
mulation, however the same is not true for the LU and QR factorizations.
The tile algorithms for them are constructed by factorizing the diagonal
tile first and then incrementally updating the factorization using the entries
below the diagonal tile. This is a very well known concept that dates back
to the work of Gauss. The idea was initially used to build out-of-core
algorithms and recently rediscovered as a very efficient method for imple-
menting linear algebra operations on multicore processors. (It is crucial to
note that the technique of processing the matrix by square tiles yields satis-
factory performance only when accompanied by data organization based on
square tiles. The layout is referred to as Square Block layout or, simply, Tile
Layout.)
For parallel execution those algorithms can be scheduled either statically
or dynamically. For static execution (Fig. 11.7) the work for each core is
predetermined and each core follows the cycle: check task dependencies
(and wait if necessary), perform a task, update dependencies, transition
to the next task (using a static transition function). For regular algorithms,
such as dense matrix factorizations, static scheduling is straightforward and
very robust.
An alternative approach, which emphasizes the ease of development,
is based on writing a serial algorithm and the use of a dynamic scheduler,
which traverses the code and queues tasks for parallel execution, while
automatically keeping track of data dependencies (Fig. 11.8). This approach
relies on the availability of such a scheduler, which is not trivial to develop,
but offers multiple advantages, such as pipelining/streaming of different
stages of the computation (e.g. factorization and solve).
#define A(m,n) &((PLASMA_Complex64_t*)A.mat)[A.bsiz*(m)+A.bsiz*A.lmt*(n)]

#define L(m,n) &((PLASMA_Complex64_t*)L.mat)[L.bsiz*(m)+L.bsiz*L.lmt*(n)]
#define IPIV(m,n) &IPIV[A.nb*(m)+A.nb*A.lmt*(n)]
void plasma_pzgetrf(plasma_context_t *plasma)
{
PLASMA_desc A, L;
int *IPIV,k, m, n, next_k, next_m, next_n, iinfo;
PLASMA_Complex64_t *work;
plasma_unpack_args_3(A, L, IPIV);
work = (PLASMA_Complex64_t*)plasma_private_alloc(plasma, L.mb*L.nb, L.dtyp);
ss_init(A.mt, A.nt, -1);
k = 0; n = PLASMA_RANK;
while (n >= A.nt) {k++; n = n-A.nt+k;}
m = k;
while (k < min(A.mt, A.nt) && n < A.nt) {

next_n = n; next_m = m; next_k = k;
next_m++;
if (next_m == A.mt) {
next_n += PLASMA_SIZE;
while (next_n >= A.nt && next_k < min(A.mt, A.nt)) {next_k++; next_n = next_n-A.nt+next_k;}
next_m = next_k;
}
if (n == k) {
if (m == k) {
ss_cond_wait(k, k, k-1);
CORE_zgetrf(k == A.mt-1 ? A.m-k*A.nb : A.nb, k == A.nt-1 ? A.n-k*A.nb : A.nb, L.mb, A(k, k), A.nb,
IPIV(k, k), &iinfo);
if (PLASMA_INFO == 0 && iinfo > 0 && m == A.mt-1)
PLASMA_INFO = iinfo + A.nb*k;
ss_cond_set(k, k, k);
}
else {
ss_cond_wait(m, k, k-1);
CORE_ztstrf(m == A.mt-1 ? A.m-m*A.nb : A.nb, k == A.nt-1 ? A.n-k*A.nb : A.nb, L.mb, A.nb, A(k, k),
A.nb, A(m, k), A.nb, L(m, k), L.mb, IPIV(m, k), work, L.nb, &iinfo);
if (PLASMA_INFO == 0 && iinfo > 0 && m == A.mt-1)
PLASMA_INFO = iinfo + A.nb*k;
ss_cond_set(m, k, k);
}
}
else {
if (m == k) {
ss_cond_wait(k, k, k);
ss_cond_wait(k, n, k-1);
CORE_zgessm(k == A.mt-1 ? A.m-k*A.nb : A.nb, n == A.nt-1 ? A.n-n*A.nb : A.nb, A.nb, L.mb, IPIV(k, k),
A(k, k), A.nb, A(k, n), A.nb);
}
else {
ss_cond_wait(m, k, k);
ss_cond_wait(m, n, k-1);
CORE_zssssm(A.nb, m == A.mt-1 ? A.m-m*A.nb : A.nb, n == A.nt-1 ? A.n-n*A.nb : A.nb, L.mb, A.nb,
A(k, n), A.nb, A(m, n), A.nb, L(m, k), L.mb, A(m, k), A.nb, IPIV(m, k));
ss_cond_set(m, n, k);
}
}
n = next_n; m = next_m; k = next_k;
}
plasma_private_free(plasma, work);
ss_finalize();
}
Fig. 11.7. Factorization for multicore execution using the SMPD programming
model with static scheduling of work (C code.)
#define A(m,n) &((PLASMA_Complex64_t*)A.mat)[A.bsiz*(m)+A.bsiz*A.lmt*(n)]

#define L(m,n) &((PLASMA_Complex64_t*)L.mat)[L.bsiz*(m)+L.bsiz*L.lmt*(n)]
#define IPIV(m,n) &IPIV[A.nb*(m)+A.nb*A.lmt*(n)]
void plasma_pzgetrf_dsched(PLASMA_desc A, PLASMA_desc L, int *IPIV)
{
int k, m, n;
PLASMA_Complex64_t *work;
plasma_context_t *plasma;
int temp1, temp2;
plasma = plasma_context_self();
for (k = 0; k < min(A.mt, A.nt); k++) {

temp1 = A.m-k*A.nb;
temp2 = A.n-k*A.nb;
DSCHED_Insert_Task(plasma->dsched, CORE_zgetrf_, 0x00,
sizeof(int), k == A.mt-1 ? &temp1 : &A.nb, VALUE,
sizeof(int), k == A.nt-1 ? &temp2 : &A.nb, VALUE,
sizeof(int), &L.mb, VALUE,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(k, k), INOUT | LOCALITY,
sizeof(int), &A.nb, VALUE,
sizeof(int)*A.mb, IPIV(k, k), OUTPUT,
sizeof(int*), &plasma->iinfo, OUTPUT,
0);
for (n = k+1; n < A.nt; n++) {
temp1 = A.m-k*A.nb;
temp2 = A.n-n*A.nb;
DSCHED_Insert_Task(plasma->dsched, CORE_zgessm_, 0x00,
sizeof(int), k == A.mt-1 ? &temp1 : &A.nb, VALUE,
sizeof(int), n == A.nt-1 ? &temp2 : &A.nb, VALUE,
sizeof(int)*A.mb, IPIV(k, k), INPUT,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(k, k), NODEP,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(k, n), INOUT | LOCALITY,
0);
}
for (m = k+1; m < A.mt; m++) {
temp1 = A.m-m*A.nb;
temp2 = A.n-k*A.nb;
DSCHED_Insert_Task(plasma->dsched, CORE_ztstrf_, 0x00,
sizeof(int), m == A.mt-1 ? &temp1 : &A.nb, VALUE,
sizeof(int), k == A.nt-1 ? &temp2 : &A.nb, VALUE,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(k, k), INOUT | LOCALITY,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(m, k), INOUT,
sizeof(PLASMA_Complex64_t)*L.mb*L.nb, L(m, k), OUTPUT,
sizeof(int)*A.mb, IPIV(m, k), OUTPUT,
sizeof(PLASMA_Complex64_t)*L.mb*L.nb, NULL, SCRATCH,
sizeof(int), &L.nb, VALUE,
sizeof(int*), &plasma->iinfo, OUTPUT,
0);
for (n = k+1; n < A.nt; n++) {
temp1 = A.m-m*A.nb;
temp2 = A.n-n*A.nb;
DSCHED_Insert_Task(plasma->dsched, CORE_zssssm_, 0x00,
sizeof(int), m == A.mt-1 ? &temp1 : &A.nb, VALUE,
sizeof(int), n == A.nt-1 ? &temp2 : &A.nb, VALUE,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(k, n), INOUT | LOCALITY,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(m, n), INOUT,
sizeof(PLASMA_Complex64_t)*L.mb*L.nb, L(m, k), INPUT,
sizeof(PLASMA_Complex64_t)*A.mb*A.nb, A(m, k), INPUT,
sizeof(int)*A.mb, IPIV(m, k), INPUT,
0);
}
}
}
}
Fig. 11.8. Factorization for multicore execution using dynamic task scheduling
(C code.)
11.9. Error Analysis and Operation Count
The key aspect of all of the implementations presented in this section is

their numerical properties.
It is acceptable to forgo elegance in order to gain performance. But
numerical stability is of vital importance and cannot be sacrificed, because
it is an inherent part of the algorithms correctness. While these are serious
considerations, there is some consolation to follow. It may be surprising to
some readers that all of the algorithms presented are the same, even though
its virtually impossible to make each excerpt of code produce exactly the
same output for exactly the same inputs. The fundamental reason for this
are the vagaries of the floating-point arithmetic in finite precision as it is
implemented in virtually all hardware. In essence, only a slight change in the
order in which the floating-point operations are performed causes a change
in the result: the change is on the order of the, so called, machine precision.
Machine precision comes from the number of decimal digits represented
in the floating-point format: for double precision there are 15 digits and so
the machine precision is about 1015 . LINPACK and LAPACK perform
the operations in different order because the latter merges the updates into
a single call to BLAS. And even though ScaLAPACK merges the updates
in a similar fashion as LAPACK does, the former performs its operations
only on the local portion of the matrix whereas the latter treats the matrix
as a single piece quantity. In other words, when LAPACK makes a single
update operation, ScaLAPACK could make as many as there are processors
involved in the computation.
When it comes to repeatability of results, the vagaries of floating-point
representation may be captured in a rigorous way by error bounds. One
way of expressing the numerical robustness of the previous algorithms is
with the following formula:
r/A e A1 r,
where error e = x y is the difference between the computed solution y
and the correct solution x, and r = Ay b is a so-called residual. The
previous formula basically says that the size of the error (the parallel bars
surrounding a value indicate a norm a measure of absolute size) is as
small as warranted by the quality of the matrix A. Therefore, if the matrix
is close to being singular in numerical sense (some entries are small with
respect to machine precision and the condition number of the matrix and
so they might be considered to be zero) the algorithms will not give an
accurate answer. But otherwise, a relatively good quality of the result may
be expected.
Another feature that is common to all the versions presented is the oper-
ation count: they all perform (2/3)n3 floating-point multiplications and/or
additions. The order of these operations is what differentiates them. There
exist algorithms that increase the amount of floating-point work to save
on memory traffic or network transfers (especially for distributed-memory
parallel algorithms.) But because the algorithms shown in this chapter have
the same operation count, it is valid to compare them for performance. The
computational rate (number of floating-point operations per second) may
be used instead of the time taken to solve the problem, provided that the
matrix size is the same. But comparing computational rates is sometimes
better because it allows a comparison of algorithms when the matrix sizes
differ. For example, a sequential algorithm on a single processor can be
directly compared with a parallel one working on a large cluster on a much
bigger matrix.
11.10. Future Directions for Research and Hardware Design
In this chapter we have looked at the evolution of the design of a simple

but important algorithm in computational science. The changes over the
past 30 years have been necessary to follow the lead of the advances in
computer architectures. In some cases these changes have been simple,
such as interchanging loops. In other cases, they have been as complex as
the introduction of recursion and look-ahead computations. In each case,
however, the codes ability to efficiently utilize the memory hierarchy is
the key to high performance on a single processor as well as on shared and
distributed memory systems.
The essence of the problem is the dramatic increase in complexity that
software developers have had to confront, and still do. Dual-core machines
are already common, and the number of cores is expected to roughly double
with each processor generation. But contrary to the assumptions of the old
model, programmers will not be able to consider these cores independently
(i.e. multi-core is not the new SMP) because they share on-chip resources
in ways that separate processors do not. This situation is made even more
complicated by the other nonstandard components that future architectures
are expected to deploy, including the mixing of different types of cores,
hardware accelerators, and memory systems.
When processor clock speeds flatlined in 2004, after more than 15 years
of exponential increases, the era of routine and near automatic performance
improvements that the HPC application community had previously enjoyed
came to an abrupt end. The air of crisis that followed in the wake of this new
regime continues to hang over computational science. To develop software

that will perform well on petascale systems with thousands of nodes and
millions of cores, the list of major challenges that must now be confronted
is formidable:
1. Dramatic escalation in the costs of intrasystem communication between

processors and/or levels of memory hierarchy.
2. Increased hybridization of processor architectures (mixing CPUs, GPUs,
etc.), in varying and unexpected design combinations.
3. High levels of parallelism and more complex constraints means that
cooperating processes must be dynamically and unpredictably scheduled
for asynchronous execution.
4. Software will not run at scale without much better resilience to faults
and far more robustness.
5. New levels of self-adaptivity will be required to enable software to mod-
ulate process speed in order to satisfy limited energy budgets.
After the industry-wide move from single to multi-core systems, dom-

inant mainstream computer architecture is now undergoing a second major
evolution: from homogeneous to heterogeneous platforms. With increased
frequency, the new systems are called Hybrid Multicores (HMCs). Todays
breed of HMCs simply feature a multi-core processor and a high end GPU.
In the future, the multi-core vendors are planning integration of GPU-like
technology directly into the multi-core chip. From the programmer per-
spective, this might alleviate the problem of dealing with two separate
memory address spaces: one attached to the multi-core and one attached to
the GPU. If such integration is realized and the performance levels are satis-
factory, then such hybrid computing device could be the prevalent hardware
design. Faced with a choice of either having an external GPU or an inte-
grated GPU-like device, the programmer would have to choose the more
productive solution give the problem at hand.
In a nutshell, the high performance computing (HPC) community will
soon be faced with machines supporting heterogeneities in all hardware
aspects processing elements of multiple types with different ISAs, mul-
tiple memory components with variable data transport interfaces, general
and specific accelerators for various purpose, power control system infras-
tructure integrated throughout and all in concurrent, simultaneous action
and interaction.
Finally, the proliferation of widely divergent design ideas shows that the
question of how to best combine all these new resources and components
is largely unsettled. When combined, these changes produce a picture of
a future in which programmers will have to overcome software design

problems vastly more complex and challenging than those in the past in
order to take advantage of the much higher degrees of concurrency and
greater computing power that new architectures will offer. The current
trends in software do not address such complexities. The message passing
paradigm epitomized by the MPI (Message Passing Interface) standard
quickly leads to management issues if every processing core corresponds to
a single MPI process. Only a hierarchical approach could possibly address
todays machine that features man hundreds of thousands of computational
cores. One such approach is a mix of MPI and OpenMP. The former con-
nects the multi-core nodes while the latter commands the computation
inside each node. Such a mix could potentially reduce the programming
complexity by one or two orders of magnitude in the number cores but the
issue of the attached accelerator (either a GPU or GPU-like device) is still
not addressed within a single programming framework. Existing PGAS
(Partitioned Global Address Space) languages such as Co-Array Fortran,
Titanium, and UPC have never been designed to address the hardware
hybridization phenomenon. Even the new languages of the breed such as
Chapel, Fortress, and X10 could potentially face the challenge of redesign
to fit in the changing hardware landscape. It is still too early to tell what
would come out of the current initiatives to retrofit the mainstay languages
of HPC, C and Fortran, with new facilities for handling hybrid computers.
At this point in time it is hard to predict which approach will prove to have
a lasting power.
So the bad news is that none of the presented code will work efficiently
someday. The good news is that we have learned various ways to mold
the original simple rendition of the algorithm to meet the ever-increasing
challenges of hardware designs.
Bibliography
[1] R. Aymar, V. Chuyanov, M. Huguet, andY. Shimomura, Nuclear Fusion 41(10) (2001).
[2] E.F. Jaeger, L.A. Berry, E. DAzevedo, D.B. Batchelor, M.D. Carter MD, K.F. White,
and H. Weitzner, Physics of Plasmas 9(5), (2002) 1873.
[3] E.F. Jaeger, L.A. Berry, J.R. Myra, D.B. Batchelor, E. DAzevedo, P.T. Bonoli, C.K.
Philips, D.N. Smithe, D.A. DIppolito, M.D. Carter, R.J. Dumont, J.C. Wright, and
R.W. Harvey, Phys. Rev. Lett. 90(19) (2003) 5001.
[4] E.F. Jaeger, R.W. Harvey, L.A. Berry, J.R. Myra, R.J. Dumont, C.K. Philips, D.N.
Smithe, R.F. Barrett, D.B. Batchelor, P.T. Bonoli, M.D. Carter, E.F. Dazevedo, D.A.
Dippolito, R.D. Moore, and J.C. Wright, Nuclear Fusion 46(7), (2006) S397.
[5] R.F. Barrett, T.H.F. Chan, E.F. DAzevedo, E.F. Jaeger, K. Wong, and R.Y. Wong,
Concurrency and Computation: Practice and Experience 22(5), (2010) 573.
Further Reading
A set of Level 3 Basic Linear Algebra Subprograms, J. J. Dongarra, J. Du Croz, I. S. Duff,

and S. Hammarling, ACM Trans. Math. Soft. 16 (1990), pp. 117.
A Proposal for a Set of Parallel Basic Linear Algebra Subprograms, J. Choi, J. Dongarra,
S. Ostrouchov, A. Petitet, D. Walker, and R. C. Whaley, University of Tennessee Computer
Science Technical Report, UT-CS-95-292, (May 1995).
An extended set of FORTRAN Basic Linear Algebra Subprograms, J. J. Dongarra, J. Du
Croz, S. Hammarling, and R. J. Hanson, ACM Trans. Math. Soft. 14 (1988), pp. 117.
Basic Linear Algebra Subprograms for FORTRAN usage, C. L. Lawson, R. J. Hanson,
D. Kincaid, and F. T. Krogh, ACM Trans. Math. Soft. 5 (1979), pp. 308323.
Implementation Guide for LAPACK, E. Anderson and J. Dongarra, University of Ten-
nessee Computer Science Technical Report, UT-CS-90-101 (April 1990).
LINPACK Users Guide, J. J. Dongarra, J. R. Bunch, C. B. Moler, and G. W. Stewart (SIAM,
Philadelphia, 1979). ISBN 0-89871-172-X.
LAPACK Users Guide (3rd Edition), E. Anderson, Z. Bai, C. Bischof, S. Blackford,
J. Demmel, J. Dongarra, J. Du Croz, A. Greenbaum, S. Hammaring, A. McKenney,
and D. Sorensen (SIAM, Philadelphia, 1999). ISBN 0-89871-447-8.
LAPACK Working Note 37: Two Dimensional Basic Linear Algebra Communication Sub-
programs, J. Dongarra and R. A. van de Geijn, University of Tennessee Computer Science
Technical Report, UT-CS-91-138 (October 1991).
LAPACK Working Note 19: Evaluating Block Algorithm Variants in LAPACK, E.
Anderson and J. Dongarra, University of Tennessee Computer Science Technical Report,
UT-CS-90-103 (April 1990).
Matrix computations with Fortran and paging, C. B. Moler, Communications of the ACM
15(4) (1972), pp. 268270.
Recursion leads to automatic variable blocking for dense linear-algebra algorithms, F. G.
Gustavson, IBM J. Res. Dev. 41 (1997), pp. 737756.
ScaLAPACK Users Guide, L. S. Blackford, J. Choi, A. Cleary, E. DAzevedo, J. Demmel,
I. Dhillon, J. Dongarra, S. Hammarling, G. Henry, A. Petitet, K. Stanley, D. Walker, and
R. C. Whaley (SIAM, Philadelphia, 1997). ISBN 0-89871-397-8.
July 19, 2011 11:29 9in x 6in b1189-Index Solving the Schrodinger Equation
Index
C6 dispersion coefficients, 20, 21 antiaromaticity, 204, 228

system, 205, 228 antibonding orbital, 44, 212
-distortivity, 223, 228 antisymmetrized, 203, 240, 260
separation, 222 antisymmetrized pair function (APF), 260
1 B state, 251 antisymmetrizer, 106
u
3 Cu , 221 antithetic variates, 263
2
3 + , 221
u Anton, 288
1,2-hydrogen shift, 85 AO, 143, 144, 164, 169, 174, 202207
2-RDM, 6165, 6770, 7277, 8087 AORSA, 314
4n+2 annulenes, 223 Application Specific Integrated Circuit
6-311G, 15, 17, 20, 77, 78, 84 (ASIC), 284, 285
6-31G*, 225 architecture, 179, 182, 267, 271, 275, 282,
284, 286, 288, 292, 307, 315318, 320,
acceptance probability, 245, 247 322, 323, 325, 327, 328, 339341
acene chains, 61, 63, 69, 71, 72, 85 arithmetic operation, 184, 284, 322
acetone enolate, 77 aromaticity, 202, 204, 219, 223, 228
ACSE, 61, 7378, 80, 87 asymptotic limit, 101
action functional, 278 ATMOL, 151
active electrons, 68 atomic forces, 238, 251, 261264
active space, 17, 53, 54, 56, 57, 71, 72, 74, atomic number, 143, 145, 156, 184, 263
77, 225, 250 atomic orbital, 25, 26, 44, 70, 113, 120,
ADF, 151, 153, 182, 187 141145, 147, 148, 151153, 157,
ALCHEMY, 150 163165, 171, 175, 176, 182, 188, 191,
alcohols, 219 202, 204, 206, 207, 272
algebrant algorithm, 224 atomic units, 3, 21, 100, 107, 112, 117,
allyl radical, 78, 219, 220 152, 153, 168, 184, 276
Amdahls Law, 316 atoms in molecules (AIM), 143, 147, 164,
amides, 219 165
amplitudes, 28, 32, 39, 92 autoionization, 92
analytical integration, 159, 241 auxiliary basis set, 30, 31, 3638, 40
angular-momentum expansion, 26 auxiliary Hermitian operator, 244
annihilation operator, 28, 63, 65, 314 auxiliary indices, 4649, 5759
anti-Hermitian, 61, 63, 7375, 85 average mean-field, 209
343
344 Index
b2u , 221223 breathing-orbital valence bond method

B-functions, 155, 160, 161, 187, 194 (BOVB), 210214, 220, 224227, 230
backflow technique, 260 Brueckner orbitals, 214
BakerCampbellHausdorff (BCH) butadiene, 77, 78, 205, 219
expansion, 54
band gap, 276, 290 C14 H14 , 224
banded, 274, 277, 320 C2 , 252, 259
banded representation, 274 cache memory, 315, 316, 322, 328
bandwidth, 316, 322, 333, 335 CADPAC, 150
basis set, 9, 16, 2527, 2931, 34, 3638, canonical ensemble, 242
40, 55, 63, 65, 66, 6870, 72, 73, canonical form, 49
7779, 8386, 91, 98, 100107, canonical MOs, 201, 208
111118, 121, 122, 124, 131, 133, canonical transformation, 53, 54, 56, 75
135137, 141, 142, 146, 147, 149152, canonical transformation theory, 54, 56
160, 163165, 168, 171, 174, 175, capability computing, 264, 327
178182, 187, 204, 211, 225227, 237, capacity computing, 327
240, 249, 259, 265, 267, 272, 275277 CarParrinello (CP), 238, 275, 276
basis set expansion, 25, 91, 98, 100102, carboxylic acids, 219
104 Cartesian space, 246
basis set limit, 26, 27, 31 CASPT2, 227
Bayesian inference method, 264 CASSCF, 16, 17, 72, 209, 211, 212, 225,
BCLF, 169, 171, 183186 227, 250
BDE, 225, 226 Cauchy problem, 295
benzene, 219223, 228 CauchySchwarz inequalities, 82
benzothiazoles, 173, 174 CC2, 32, 263
Beowulf clusters, 327 CCSD, 30, 31, 39, 40, 77, 8385, 262, 263
Bessel functions, 11, 155, 161, 184186, CCSDT, 69, 70, 77, 81, 8385, 217, 262
190 CH3 F, 163, 167169
bi-partite, 49 CH4 , 16, 229
bicyclobutane, 62, 77, 78 CH+ 4 , 229
binding energies, 116, 121, 126, 129, 138, chaining, 315, 320, 322
239, 273 CHAMP, 178, 179, 183
biradical, 72 charge-shift bonding, 220, 229
BLACS, 328, 330, 332 Cholesky factorization, 335
BLAS library, 314 cis-trans isomers, 77
block-localized wave function, 219, 220, closed shell system, 108
227, 230 (CO)4 Fe(CH2 =CH2 ), 225
Boltzmann factor, 95 column orientation, 320
bond ionization, 229 combinatorial optimization, 68
bond localization, 223, 224 commodity processor, 288, 306
bonding orbital, 44, 148, 178, 208, 212 communication, 286288, 315, 327, 328,
boson, 237, 259, 267 331, 340
bosonic distribution, 260 commute, 54, 243, 257, 261, 262
boundary condition, 102, 264, 275 compiler, 167, 184, 322
Boys differentiation, 10 complete active space, 56, 57, 250
BoysShavitt method, 160 complete basis, 16, 46, 52, 91, 98, 106,
branching term, 292, 295 107, 112
Index 345
complete basis set (CBS) limit, 16 coupled-cluster, 1, 28, 3032, 35, 39, 44,
completeness, 31, 137, 166 5355, 59, 61, 69, 70, 77, 78, 81, 83,
configuration, 30, 39, 40, 44, 46, 52, 55, 84, 133, 214, 272, 314
61, 69, 74, 76, 78, 79, 81, 93, 106, 115, covalent, 55, 202, 205208, 212, 216,
118121, 124, 127, 130133, 135138, 219221, 229
142, 145, 146, 151, 158, 176, 178180, covalent-ionic mixing, 221
204, 206, 212, 215, 221, 227, 229, 239, CPU, 224, 238, 284, 340
243, 245250, 253255, 259, 262, 263, creation operators, 28, 65, 66
266, 272, 296, 314, 316, 325 critical exponent, 94, 98, 99, 102104, 108
configuration interaction (CI), 39, 40, 46, critical phenomena, 9295, 97, 108
52, 55, 61, 69, 74, 78, 79, 81, 133, 142, critical slowing down, or CSD, 274
145, 146, 151, 204, 206, 239, 250, 254, critical temperature Tc , 96
263, 272, 314 CRUNCH, 231
configuration set, 248 cubic scaling, 263
configuration space, 243245, 247, 255, cumulant expansion, 75
262, 266 cumulant theory for RDMs, 73
conical intersection, 62, 78, 80 cusp-corrected Gaussian, 180
conjecture, 13, 22 cyclic polyenes, 54
conjugate eigenvalue problem, 114 cyclopropane, 219
conjugated systems, 54, 56, 219, 228 cyclopropyl radical, 78
Connection Machine, 288 cytochrome P450, 218
conrotatory pathway, 77, 78
continuous phase transitions, 98, 108 D6h , 221, 222
continuum, 91, 97, 105, 107, 108, 112, decapentaene C10 H12 , 251
137, 156, 211 decomposition, 49, 51, 54, 224, 286, 287,
continuum state, 91, 108 318320
contracted Schrodinger equation (CSE), defect correction, 281, 298
6163, 7375, 85 degenerate, 44, 73, 97, 105, 107, 124, 229
contraction, 12, 38, 47, 48, 57, 62, 6668, degenerate orbitals, 44
73, 81, 225, 314 delocalized, 144, 147, 201, 204, 207, 208,
core ionization energies, 129, 130 219, 226, 276
correlation energy, 1, 2, 4, 5, 13, 1519, dense linear algebra, 313, 320, 323, 332,
25, 27, 32, 33, 77, 83, 84, 145, 237, 335
240, 259, 260, 267, 296 density functional theory (DFT), 1, 3, 13,
correlation length , 96 21, 26, 27, 39, 144, 145, 150, 151, 154,
correlation-consistent basis sets, 27 156, 174, 175, 182, 187, 194, 218, 219,
Coulomb operator, 19, 142, 148, 153, 154, 227, 259, 262, 271273, 275, 276, 278,
157, 163, 165167, 169, 187, 192 279, 282, 289291, 296298, 302, 304,
Coulomb potential, 100, 112, 118, 158, 305
161, 258, 294 density matrix, 8, 9, 12, 15, 17, 18, 20, 43,
Coulomb Sturmians, 112114, 121, 131, 45, 46, 53, 54, 56, 58, 61, 62, 64, 73, 86,
133137, 145, 156, 161, 166, 178 107, 142, 154, 172, 271, 275, 278, 283,
CoulsonFischer orbitals, 207, 208 289, 291, 303, 306
coupled electron-pair approximation density matrix renormalization group, 18,
(CEPA), 82, 83 43, 45, 46, 53, 54, 56, 58, 86
coupled HartreeFock, 172 DGEMM, 323, 331, 334
346 Index
diatomics in molecules (DIM), 135, 143, effective potential, 93, 94, 291, 302
147, 165, 168 eigenstates, 52, 53, 247, 248, 252, 257
DielsAlder reaction, 225, 227 eigenvalues, 62, 65, 66, 83, 94, 97, 102,
differential dynamic correlation, 212, 213 106, 114, 116, 120, 144, 239, 243, 277,
diffusion equation, 254, 255, 292, 294, 278, 280283, 293, 297, 304
297, 299, 304 electric field, 2, 9294, 314
diffusion Monte Carlo (DMC), 179, 180, electrode, 274
225, 226, 253256, 259262, 264, 266, electron correlation problem, 1, 3
267, 292, 295298 electron distribution, 92, 143, 149
difluorine, 211 electron spin resonance, 265
dimensionless, 96, 100, 128, 138 electron-electron repulsion integral, 74
dioxygen, 204, 229 electron-nucleus cusp, 249
Dirac delta, 4, 261, 265 electronegativities, 212, 220
Dirac delta operator, 261, 265 electrostatic attraction potential, 118
Dirac equation, 127 elliptic coordinates, 158, 161, 162, 188
DiracCoulomb equation, 138, 142
elliptical Slater orbitals, 155
direct mapped implementation, 284286
energy barrier, 61, 77, 78, 216
Dirichlet boundary conditions, 102
entanglement, 4750, 58
Dirichlet problem, 295
entanglement entropy, 48, 49
discretization, 243
ergodic sampling, 247
discretize, 107, 244
error analysis, 185, 338
dispersion, 1, 1821, 33, 34, 248, 249,
251, 273, 290, 305 error bound, 338
displacement probability, 245 Euler angles, 106
displacement rule, 245 Euler constant, 190
disrotatory pathway, 77, 78 exact ground state energy, 62, 257, 289,
dissociation, 6971, 74, 83, 91, 92, 136, 290, 305
209213, 218 exchange integrals, 147, 148, 157, 158,
dissociation curves, 83, 136, 213 165, 167170, 187, 188, 190192
dissociation energy, 209213, 218 excited configurations, 204
distributed memory, 315, 316, 328, 330, excited states, 16, 56, 62, 7376, 78, 111,
335, 339 116, 117, 122, 123, 138, 176, 201, 215,
diverge, 70, 96, 244, 249, 257, 282 222, 223, 228, 229, 254, 256, 259, 261,
DNA strand, 274 267, 314
double excitation, 2830, 38, 61, 69, 74, expectation value, 50, 59, 64, 98, 99, 104,
77, 78, 81 156, 158, 240244, 247, 248, 250, 251,
double-bond, 74 262, 264, 272, 278, 295, 298, 299
double-polarization basis, 55 exponential operator, 53, 54
double-zeta, 26, 32, 55, 56, 72, 73, 79 exponential scaling, 50
doubly excited, 29, 123 exponential type orbitals (ETOs), 145,
drift force, 293 146, 163, 164, 169, 178
dual-core, 339 exponential wall, 296
dynamic correlation, 1619, 34, 40, 50, extended Lagrangian formalism, 275
51, 53, 54, 56, 201, 206, 210, 212, 213,
226228, 237, 239 F2 , 220
F
2 , 212
ECP, 225 F12, 28, 29, 31, 36, 40, 241, 260
Index 347
factorization, 68, 319, 321, 323, 324, 329, functional integral, 299
332337
fermion sign problem, 259 GAMESS, 178, 179, 230
fermionic systems, 264 GAMESS-US, 230
ferromagnetic bonding, 229 gauche-1,3-butadiene, 77, 78
FeynmanKac formula, 298 Gaunt coefficients, 193
FFT, 275 GaussLegendre quadrature, 161
finite element (FE), 91, 93, 98, 100102, GaussSeidel, 279
107, 276 Gaussian, 4, 5, 9, 10, 12, 14, 15, 19, 26,
finite size scaling (FSS), 91, 92, 95100, 35, 37, 93, 102, 141143, 145, 146,
102108 149152, 154, 155, 159, 160, 163165,
finite temperature, 276 171, 173177, 179182, 187, 230, 249,
finite-difference (FD), 263, 276, 277, 260, 272, 294
279282, 292 Gaussian functions, 4, 12, 35, 142, 143,
first order phase transitions, 98, 108 145, 146, 149, 177, 179, 272
first-order kinetic equation, 255 Gaussian quadrature, 102
first-order pair function, 28, 32 Gaussians random number (GRN), 294
fixed-node, 142, 179, 180, 259, 260, 267, Gegenbauer addition theorem, 165
296, 298, 304
geminal, 32, 145, 161, 260
fixed-node approximation, 179, 180, 259,
geminal basis functions, 32
260, 267, 296, 298, 304
generalized Sturmians, 111, 114122,
floating point performance, 283
130, 131, 133, 136138
floating-point operations, 68, 69, 323, 325,
generalized Sturmians secular equations,
338, 339
115, 120
FLP, 285
Fock operator, 28, 146 generalized trapezoidal rule, 243
Fock space, 48, 51, 52 generalized valence bond, 207
FokkerPlanck (Schmoluchowski) GIAO, 171
operator, 257 Gigabit Ethernet, 328
FokkerPlanck equation, 294 global minimum, 67, 76
force operator, 263 Goscinskian configurations, 118, 119,
fork-join model, 332 124, 130, 131, 137, 138
formaldoxime, 85, 86 GramSchmidt procedure, 278
Fortran, 167, 184, 231, 320324, 329, 341 GRAPE, 288
forward Kolmogorov equation, 294 Graphics Processing Units (GPUs), 288,
Fourier space, 165 340, 341
Fourier transform, 7, 19, 20, 134, 141, 150, Grassmann wedge product, 65, 67, 75
155, 160, 167, 172, 187, 191, 193, 194 Greens function, 73, 151, 256, 289, 291,
frozen core approximation, 133 304306
Full Approximation Scheme (FAS), 282 grid spacing, 276, 277, 279281
full configuration interaction (FCI), 46, ground state, 3, 15, 28, 6164, 68, 69,
52, 55, 69, 70, 79, 80, 83 7278, 94, 98, 101, 106, 117, 121,
functional, 1, 3, 13, 14, 21, 26, 27, 39, 125129, 138, 145, 172, 176, 180, 188,
6164, 67, 73, 82, 84, 85, 107, 142, 215, 217219, 222, 223, 229, 241, 247,
144, 145, 150, 151, 166, 259, 262, 267, 251, 254, 256, 257, 259, 262, 267, 276,
271273, 278, 279, 290, 292, 299, 303 278, 289, 290, 293, 295, 296, 298, 299,
functional derivative, 278, 279 305, 306, 314
348 Index
H2 , 17, 19, 136, 149, 163, 167169, 171, hybrid atomic orbitals (HAOs), 204,
176, 177, 187, 188, 202 207209, 230
Hamiltonian, 13, 18, 28, 33, 39, 48, 50, hybrid integrals, 158
5255, 63, 64, 72, 74, 75, 9193, 97, Hybrid Multicores (HMC), 340
98, 100, 104107, 113, 117, 120, 128, hybridization, 202, 204, 208, 229, 340,
144, 156, 165, 172, 205, 206, 224, 227, 341
240, 241, 243, 247, 251, 256, 257, 261, hydration, 272
262, 271, 274, 277, 279, 297 hydrogen, 3, 22, 33, 39, 44, 47, 5153, 56,
harmonic oscillator, 19, 20, 277 61, 63, 70, 71, 77, 78, 8386, 106,
harmonic well, 301, 302 111113, 118, 119, 127, 141, 143145,
Hartree orbitals, 148 148, 151153, 155, 156, 163, 165, 167,
HartreeFock (HF), 1, 2, 28, 29, 32, 39, 168, 174176, 182, 188, 202, 204, 207,
43, 50, 76, 81, 93, 107, 111, 121, 133, 216, 220
136, 138, 143, 146, 150, 153, 156, 157, hydrogen abstraction reactions, 216
165, 168, 172, 179, 181, 211214, 219, hydrogen fluoride, 78, 8385
226, 238, 250, 254, 289291, 296, 298, hydrogen lattices, 44, 61, 63
315 hydrogen molecule, 22, 44, 176
HartreeFock limit, 138, 153, 168, 181 hydrogen-like orbitals, 112, 113, 141, 143,
HartreeFock orbitals, 226 144, 148, 151, 152, 155, 156, 163, 165,
HCCH, 16 174, 175, 182, 188
He2 , 176 hydrogen-transfer, 220
HeH+ , 136, 176 hydroxylation, 173, 218
Heisenberg representation, 75 Hylleraas wave functions, 151, 154, 176,
Heisenberg Uncertainty Principle, 7, 92 181
helium, 5, 6, 11, 22, 32, 39, 40, 108, 121, hyperconjugation, 218, 220, 228
124126, 129, 138, 151, 177, 241 hypergeometric functions, 100, 191
helium atom, 5, 6, 11, 22, 32, 39, 108,
121, 138, 151, 177 IBM, 150, 167169, 175, 179, 183, 318,
HellmannFeynman theorem, 251 325327, 342
Hermite interpolation polynomials, 98, 99, IBM BlueGene, 179, 318
102, 103 IBMOL, 150
Hermitian, 6163, 7375, 85, 244, 320 idempotency constraint, 292, 297, 303
Hessian, 94, 250, 251 idempotent, 290
Hessian matrix, 94, 251 ILAENV function, 325
heteropolar bond, 220 imaginary time, 179, 254256, 292, 299,
hextuple excitations, 55 305
high performance computing (HPC), 318, importance sampling, 257, 293, 294, 297
326, 327, 339341 independent particle techniques, 238
Hilbert spaces, 66, 98, 114, 166 indicator function, 301
HO 3 , 77 indistinguishable, 62, 69, 70, 84, 85
HohenbergKohn theorem, 1, 3, 13, 21 infrared intensities, 264
holes, 38, 6567, 73, 82 integral equation, 302
HOMO/LUMO gap, 290 integral formulation of DFT, 302, 304
homopolar bond, 220 interatomic distance, 70, 71, 135, 151,
HPC computing, 326 168, 213
Hulthen potential, 100102, 104, 105, 107 interelectron repulsion, 120, 124128,
Husimi quasi-densities, 7 130, 131, 137, 138
Index 349
interior-point, 67 large-dimension models, 9194

intermediate expansion, 52 laser fields, 92, 107, 108
interpolation, 281 latency, 287, 305, 328
intracule, 1, 314, 1820, 22, 27 LCAO, 143, 150, 187
intracule functional theory (IFT), 1, 14, left coset decomposition algorithm, 224
1719, 21, 22 Lewis structure, 202, 205, 206, 212, 213,
intrasystem communication, 340 216, 218, 219
inverse Laplace transform, 304 Lie-group, 53
ionic forms, 216 LiH, 177, 259
ionization, 91, 92, 129, 130, 184, 204, 229 linear algebra, 239, 313316, 320, 322,
ionization potential, 184 323, 328, 332, 335
ionization spectrum, 204 linear-scaling, 87, 178, 276, 282, 289
ions in water, 272 LINPACK, 320, 321, 323, 325, 332, 338
ISA, 340 local estimator, 244
isoelectronic series, 120122, 124130, local spin density (LDA), 226, 331
138 localized bond orbitals (LBOs), 229
isoenergetic, 111, 113115, 119, 121, 131, lone-pair, 210, 219, 220
132, 137 long-wavelength, 271, 274, 275, 279, 280
look-ahead, 85, 333, 334, 339
Jacobi, 279, 280, 305 Lowdins alpha function method, 159
Jastrow factors, 34, 142, 179181, 260 lower triangular matrix/upper triangular,
318
Kekule mode, 223
Kekule type, 221 machine precision, 338
Keplerate magnet, 57 magnetic field, 2, 92, 107, 171, 172, 253,
kernel, 1421, 313, 315 314
kinetic energy, 26, 74, 93, 102, 116, 156, mainstream quantum chemistry, 240, 247,
249, 257, 275, 290292, 304 250, 265
kinetic energy integrals, 156 many-body theory, 1
KohnSham, 107, 144, 226, 271273, Markov chains, 245, 246
275, 277, 278, 283, 290, 291, 296, 298, massive parallelization, 286
304, 314 maximum spin cluster, 221
KohnSham method, 107 Maxwell distribution, 6
KohnSham orbitals, 226 McWeeny purification, 291, 303
Kollmars topological factor, 82 mean-field, 26, 27, 94, 107, 209
KolosWolniewicz wave functions, 176 memory hierarchy, 315, 316, 322, 323,
325, 328, 339, 340
Lagrange multipliers, 68, 69, 278, 297 memory limitation, 239
Lagrangian function, 68, 69 metal-to-insulator transition, 6971
Laguerre polynomial, 101, 144, 156, 169, methylene, 55, 62, 78, 80, 226
191 metric matrix, 6568
Langevin trajectories, 294, 301 microprocessors, 283, 285, 323
Lanl2dz, 225 MO-CI, 204, 206, 228
Laplace transform, 154, 160, 304 molecular dianions, 107
Laplacian, 253, 280, 281, 292 molecular orbitals, 4, 12, 25, 30, 44, 72,
large-Z approximation, 117, 121, 135, 141, 143, 149, 151, 153, 176, 201,
124130, 138 251, 253, 265
350 Index
MllerPlesset perturbation theory, 28, N! problem, 206

314 N-representability, 290
MllerPlesset second order (MP2), 28, N-representability conditions, 6265, 69,
29, 31, 32, 35, 36, 70, 227, 231, 262, 82, 290
263, 272 nuclear charge, 5, 93, 94, 105, 111, 117,
MOLPRO, 178 119, 121, 127, 128, 137, 144
momentum-space, 7, 134, 151 nuclear magnetic resonance (NMR), 145,
monotonic, 243 150, 171, 173, 175, 182, 183, 192, 194,
Monte Carlo estimator, 249 238, 265, 267
Moores Law, 316, 317, 322 nuclear motion, 121, 276
MP2-F12, 29, 36 nuclear physics, 93
multi-configuration self-consistent-field numerical libraries, 313
(MCSCF), 76, 79, 80, 179, 212, 225, numerical quadrature, 158
226, 230, 231, 250, 263
multi-core computers, 315, 332 O2 , 56, 229
multiconfigurational state, 128 object-oriented, 329331
multigrid methods, 271, 274, 275, 280, occupation number, 46
283 occupied orbitals, 3, 29, 30, 32, 34, 36, 37,
multiple-zeta basis sets, 152 50, 72, 81, 82, 239
multiply charged negative ions, 92 octacene, 72, 73
multireference, 17, 18, 44, 50, 5456, 61, one-center overlap integrals, 156
70, 7479 one-positivity conditions, 66, 67
muon, 265 open shell system, 108
operation count, 277, 338, 339
N2 , 16, 69, 83, 84, 149 orbital free density functional (OF-DFT),
NAI, 156 107
nano-structure, 274 orbital relaxation, 3234
NDDO-PM3, 173, 175 Order-N Electronic Total Energy Package
near-degeneracy, 44 (ONETEP), 276
N-electron wavefunction, 46, 61, 62, 290 orthonormal, 36, 38, 48, 49, 51, 112, 144,
Newton method, 250 161, 166, 190
NewtonRaphson, 224 orthonormality, 49, 113115, 120, 278
NISTOs (non integer slater orbitals), 155 orthonormalized basis, 47
nitrosomethane, 85, 86 oscillatory, 161, 282, 304
NO2 , 214 overall jump, 246
NO2 , 173, 251 overlap, 26, 36, 38, 44, 64, 66, 156,
nodal surface, 259, 260, 264, 298, 300, 163169, 182, 190192, 206, 248
304
nodal values, 102 paired-permanent-determinant approach,
non-analytical, 92 224
non-commuting operator, 243, 251, 257 parallel computers, 273, 328
non-Gaussian shape, 249 parallel computing, 274, 275, 288, 289
non-integer Slater orbitals, 151 paramagnetic ground state, 229
nondynamic correlation, 50 PariserParrPople (PPP) models, 54
nonlinear, 68, 69, 282 Pauli exclusion principle, 107
nonnegativity, 65, 82 Pauli principle, 2, 144, 203
normalization constant, 106, 202 permutation matrix, 318
Index 351
perturbation theory, 28, 29, 76, 77, 79, 93, promotion gap, 216, 217
133, 210, 239, 272, 314 propagator, 256, 258, 305
perturbative triple excitations, 84 propellanes, 220
Pfaffian form, 260 propene, 77
phase diagram, 273 protein, 228, 272
photochemical conversion, 78 pseudo-potentials, 251, 258
photochemistry, 228 pseudo-time , 292
photoelectron spectroscopy (PES), 229, pseudocritical parameters, 102
262 pseudopotentials, 279
pico-hartree accuracy, 177, 178 psinc, 276
pipelining/streaming, 336 Pulay force, 275
pivots, 318 Pulays correction, 264
Plancks constant, 9
plane-wave, 26, 187, 272, 274276 quadratic convergence, 67
Poissons equation, 166, 168, 187, 280, quadruple precision, 181, 184, 185
282, 295 quadruple-zeta, 83
PoissonBoltzmann equation, 282 Quantum Chemistry Program Exchange
polarizable continuum model (PCM), 211 (QCPE), 150
poly-phenylcarbenes, 57 quantum dot, 108
POLYATOM, 150 quantum electrodynamics, 53, 129, 138
polynomial, 46, 50, 59, 98, 99, 101103, Quantum Monte Carlo (QMC), 34, 142,
144, 145, 152, 156, 161, 169, 178, 185, 145, 178180, 182, 225, 226, 228, 237,
188191, 238, 260 239, 247, 262, 263, 266, 271273, 275,
polyradical, 57, 69, 71, 73, 228 289, 292, 295299
Poples G1 data set, 15 quantum Monte Carlo simulations, 145
porphyrins, 238, 267 quantum numbers, 98, 118, 126, 144, 152,
positive semidefinite, 62, 6468 154156, 188, 192, 194
positivity conditions, 61, 6369, 72, 82 quantum phase transitions, 86, 91, 92, 108
positron, 237, 265, 268 quasi-densities, 7, 8
potential energy surface (PES), 45, 85, 86, quasiparticle coordinates, 260
171, 213, 250, 262 quintuple-zeta, 31
p-positivity conditions, 63, 64
predictor-corrector technique, 258 Racah coefficient, 191
price/performance ratios, 317, 325, 327 radial differential equation, 100
primal-dual interior-point algorithms, radial functions, 113
6769 random numbers, 245, 257, 294
principal quantum number, 118, 126, 144, random walks, 255, 271, 272, 292, 294,
155, 156, 188, 192 299, 300
principle of locality, 45, 58 reaction mechanisms, 214
probability distribution, 6567, 242, 245, real-space grid method, 271, 275, 283, 306
249 reduced second-order density matrix, 8, 9
product ansatz, 46 relativistic corrections, 117, 124, 125, 128,
product Fock space, 48 138
product theorem, 143, 149, 154 release node method, 260
Projected Entangled Pair States (PEPS), renormalization group, 18, 43, 45, 46, 51,
58, 59 53, 54, 56, 58, 86, 96
promoted state, 216, 217 reptation method, 262
352 Index
residual, 37, 76, 279, 338 semi-internal excitations, 54

resolution of the grid, 277 semiconductor, 283, 285, 288
resolution of the identity, 30, 31, 35, 36, semidefinite programming, 63, 67, 68, 87
38, 39, 169 separation of integration variables, 165
resonance, 145, 150, 173, 201203, serial algorithm, 286, 335
212216, 218220, 223, 227229, 238, shielding tensor, 145, 171, 172, 194
265 short-range correlation factor, 30, 33
resonance energies (REs), 202, 214216, Si2 H6 , 16
218220, 228 Si2 , 252
response properties, 240, 264 sigmatropic shift, 77
restriction, 64, 65, 121, 194, 281 sign problem, 259, 267, 304
ring-opening, 62, 77, 78 silicenium ion, 220
RISC computers, 315, 316, 322 SIMD machine, 287, 328
ROHF, 231, 253 single-bond, 2, 74, 203, 212, 221
rotational barriers, 219, 220, 225 single thread, 332, 333
RusselSaunders states, 126 single-reference, 54, 55, 61, 74, 76, 77,
314
SN 2 reaction, 208, 211, 213, 216 singlet-to-triplet excitation, 217
sampling, 86, 244247, 249, 251, 254, singular value decomposition, 49, 320
257, 261, 262, 264, 266, 267, 272, 289, singularities, 92, 9496, 98, 100, 158, 263
293298, 300304, 306 size extensive, 74, 78, 8183
Saytzeffs rule, 220 size-consistency, 51
scalability, 267, 316, 332, 333 Slater determinant, 26, 28, 47, 54, 118,
ScaLAPACK library, 314, 328, 331 172, 202, 203, 206, 290, 291, 296
scales, 27, 31, 35, 37, 38, 68, 69, 72, 77, SlaterCondon rules, 206, 224
81, 85, 96, 100, 116, 184, 206, 224, 226, Slater-type, 26, 69, 93, 141, 143, 145,
239, 253, 271274, 278, 282, 283, 288, 147150, 153, 155, 157, 159, 161, 164,
291, 304, 306, 307, 314, 318, 327, 340 168, 171, 175, 178184, 187, 194, 249,
scaling, 35, 50, 61, 68, 74, 77, 78, 81, 82, 272
84, 87, 9193, 9597, 99, 100, 102, Slater-type orbital (STO), 12, 26, 69, 73,
104, 105, 107, 108, 116, 117, 120, 137, 136, 141143, 145150, 152, 153, 155,
178, 182, 184, 185, 237, 263, 271, 272, 157, 159, 160, 161, 164, 168, 171,
274, 276, 278, 282, 289 173175, 177184, 187, 193, 194, 272
scaling parameters, 87, 100, 116, 117, 120 smartphone, 284
scattering collisions, 91, 92 SMILES, 150, 160, 163, 182, 187
SCF, 17, 111, 137, 143, 146, 150, 164, SMx model, 211
165, 194, 250, 259, 262, 263 software-defined radio (SDR), 284
Schrodinger cat states, 45 SOR, 279
Schwarz inequality, 82, 148, 157 source term, 295
screened Coulomb potential, 112 sparsity, 46, 287
SCVB, 209, 230 spectator orbitals, 208, 210, 213
second-quantized, 28, 63, 66, 67, 81 spectroscopy-oriented configuration
second-quantized operators, 63, 66, 67, 81 interaction (SORCI), 227
secular equation, 115117, 120, 133136, spectrum, 91, 133, 136, 204
238 spherical harmonic, 112, 144, 154, 155,
self-consistent field, 70, 76, 79, 131, 179, 159, 190, 191, 193
208, 209, 250 spherically-averaged core density, 131
Index 353
spin coordinates, 115, 117 three-positivity, 66, 67, 69

spin free VB method, 224 tile algorithm, 335
spin orbitals, 2, 28, 63, 81, 118, 119, 127, TOP500, 318, 326, 327
240, 253 topological property, 144
spin-coupled (SC), 207, 209 topology, 48, 57, 58
square-integrable, 108, 112, 144, 252 trace, 68, 215, 220
statistical mechanics, 94, 95, 9799, 107, transition matrix, 245, 246
239, 292, 299 transition metals, 37, 44, 45, 56, 202, 225,
statistical physics, 241, 242 227, 228, 268
steepest descent equation, 279 transition state (TS), 77, 78, 208, 209, 213,
stochastic, 162, 243, 245, 257, 260, 263, 215, 216, 219, 222, 224, 262, 267
265, 294, 298, 303, 306 transportable, 323
stochastic differential equation (SDE), trial energy, 293295
294, 298, 306 trial function, 225, 226, 261, 289,
stochastic transition matrix, 245 293296, 300302, 304
STOP, 150, 164, 168, 169, 175, 182, 183, trial wave function, 145, 178181, 225,
187 226, 240, 247254, 257259, 261, 264,
storage, 69, 72, 224, 271, 275, 283, 288, 266, 267, 293
305, 306, 315, 319, 331, 332 triple-bond, 74
strong correlation, 4345, 50, 53, 54, 56,
triple-zeta, 31, 55, 56, 83, 226
57, 77, 86, 239
triplet state, 78, 126, 144, 204
strong orthogonality, 29
Trotter-like approximation, 258
SturmLiouville theorem, 156, 161
Turbomole, 31
SturmLiouville theory, 112
TURTLE, 230
Sturmians, 111122, 130, 131, 133138,
141, 145, 155, 156, 161, 165, 166, 178 TWF, 240242, 252, 254, 259
s-type Gaussians, 10 two-center nuclear attraction terms, 156
superconductivity, 86 two-center overlap integrals, 156, 191, 192
superposition, 44, 112, 115, 134, 203, 214 two-dimensional lattice models, 97
superscalar, 328 two-electron density, 3, 291
symmetric multiprocessing (SMP), 316, two-electron integrals, 26, 27, 3537, 39,
317, 325, 326, 332, 335, 339 143, 148, 153, 157, 163, 165167, 169,
symmetrized, 251, 256 178, 183, 191
symmetry breaking, 93, 94, 213, 214 two-electron reduced density matrix, 61,
symmetry dilemma, 213 62, 64
two-positivity conditions, 6568, 72, 82
Taylor expansion, 33, 104, 277, 279
temperature (T), 9496 UHF, 1517, 19, 20, 231
Tensor Contraction Engine (TCE), 314, uniform convergence, 247
315 uniform electron gas, 22
tensor networks, 43, 45, 50, 5759 unitary, 32, 65, 75, 76, 320
the bottleneck of Quantum Chemistry, 149 unitary rotations, 32, 76
thermodynamic limit, 92, 95, 96 unitary transformations, 65, 75, 76
Thinking Machines, 288 universality class, 93
ThomasFermi theory, 298, 302
thought experiment, 14 valence Bond, 179, 201, 207, 208, 210,
three-body Coulomb systems, 95 211, 225
354 Index
Valence Bond Self Consistent field method water, 55, 56, 95, 204, 238, 272274
(VBSCF), 209212, 220, 224227, 230 wavefunction, 15, 7, 10, 1416, 18, 19,
variance, 238, 243, 244, 247252, 265, 22, 2530, 32, 33, 35, 40, 43, 4559,
294 61, 62, 64, 69, 70, 74, 75, 77, 80, 81,
variational, 32, 47, 50, 52, 53, 59, 6164, 8386, 98101, 105, 112
6770, 72, 80, 81, 85, 86, 98, 102, 106, wavefunction ansatz, 30, 46, 47
107, 162, 179, 180, 207, 208, 239241, weakly bound states, 92
247, 250, 259, 261, 267, 276, 277, 303 weight dispersion, 249
variational coefficients, 52, 106 weighted potential, 111, 113115
variational Monte Carlo (VMC), 179, Wicks theorem, 314
239241, 250252, 254, 256, 259, Wigner intracule, 12, 13
261263, 266 Wigner quasi-density, 7, 8
VB state correlation diagram (VBSCD),
215, 218, 221224 X theory, 21
VB/MM, 211 XMVB, 230
VB2000, 230
VBCI method, 210, 211, 217 Young tableaux, 253
VBPCM, 211
VBPT2, 210, 211, 225, 227, 228, 230 zero temperature, 91, 276
vector machines, 315, 316, 322, 323 zero variance principle, 243
vector-vector operations, 315, 319, 320 zero-correction-at-convergence, 281
vibrational spectroscopy, 171 zeroth-order Hamiltonian, 33, 113, 227
vinylidene carbene, 77 Zundel cation, H5 O+2 , 263
walkers, 179, 255260, 262, 289, 294,

295, 298, 300, 301, 303, 304, 306

Solving The Schrödinger Equation Has Everything Been Tried

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Solving The Schrödinger Equation Has Everything Been Tried

Uploaded by

Copyright:

Available Formats

Solving the

P780.9781848167247-tp.indd 1 7/25/11 11:49 AM

Imperial College Press

P780.9781848167247-tp.indd 2 7/25/11 11:49 AM

British Library Cataloguing-in-Publication Data

SOLVING THE SCHRDINGER EQUATION

Typeset by Stallion Press

Catherine - Solving the Schrodinger Eqn.pmd 1 9/7/2011, 5:06 PM

The richest interactions occur between two almost identical but

1. Intracule Functional Theory 1

2. Explicitly Correlated Electronic Structure Theory 25

2.3.3 Thought 3: Do we need the resolution

3. Solving Problems with Strong Correlation Using

4. Reduced-Density-Matrix Theory for Many-electron Correlation 61

5. Finite Size Scaling for Criticality of the Schrodinger Equation 91

6. The Generalized Sturmian Method 111

6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . 137

7. Slater-Type Orbital Basis Sets: Reliable and Rapid

7.8.3 Programming strategy . . . . . . . . . . . . . 164

8. Modern Ab Initio Valence Bond Methods 201

8.1 Basic Principles and Survey of Modern Methods . . . . 202

8.2 Strengths of the Valence Bond Approach . . . . . . . . 211

9. Quantum Monte Carlo Approaches for Tackling

9.3.1 Generalities . . . . . . . . . . . . . . . . . . 254

10. Solving the Schrodinger Equation on Real-Space Grids

10.4.4 Obtaining E(x,y)? . . . . . . . . . . . . . . . 302

11. Changes in Dense Linear Algebra Kernels: Decades-

In the first of his four lectures on wave mechanics, Schrodinger wrote:

Substituting from (12) and (8) in (10) and replacing p by (. . . )

Here we are: equation number 13 arrived in Knowledge Space and

by a vague sense of mysterious elegance. To the contrary, matrix mechanics

pseudospectral method and the iterative configuration interaction method.

This page intentionally left blank

Intracule Functional Theory

Deborah L. Crittenden and Peter M.W. Gill

Research School of Chemistry,

2 D.L. Crittenden and P.M.W. Gill

Unfortunately, the resulting HartreeFock (HF) model neglects the inter-

Intracule Functional Theory 3

A reasonable starting point for the development of a two-electron analogue

4 D.L. Crittenden and P.M.W. Gill

(where is the Dirac delta distribution and r12 |r12 | |r1 r2 |)

Example 1.1: The Position Intracule for a He-like Ion

Intracule Functional Theory 5

Fig. 1.1. Position intracule for the helium atom.

Unfortunately, although it is easy enough to compute P(u), even in large

Example 1.2: The Momentum Intracule for a He-like Ion

6 D.L. Crittenden and P.M.W. Gill

by integrating over the angular part of v

Fig. 1.2. Momentum intracule for the helium atom.

A decade ago, Rassolov observed that the relative momentum p12

Intracule Functional Theory 7

where the two-electron momentum density is

and the momentum wavefunction

8 D.L. Crittenden and P.M.W. Gill

primarily interested in two-electron information, it is natural to integrate

Fig. 1.3. Physical interpretation of the dynamical angle 12 .

Intracule Functional Theory 9

practical, for the Omega intracule is written as a ten-dimensional integral

where the Omega integrals are

Fig. 1.4. The intracule family tree.

10 D.L. Crittenden and P.M.W. Gill

If the i are s-type Gaussians centred at A, B, C and D, with exponents

In this simple system, we find that (u, v, ) is proportional to the product of