PHD Thesis Nonlinear Trajectory Navigation

NONLINEAR TRAJECTORY NAVIGATION
by
Sang H. Park
A dissertation submitted in partial fulfillment

of the requirements for the degree of
Doctor of Philosophy
(Aerospace Engineering)
in The University of Michigan
2007
Doctoral Committee:
Associate Professor Daniel J. Scheeres, Chair
Professor Alfred O. Hero III
Professor Pierre T. Kabamba
Professor N. Harris McClamroch
Research Scientist Paul W. Chodas, Jet Propulsion Laboratory
c Sang H. Park
All Rights Reserved

2007
To my parents.
ii
ACKNOWLEDGEMENTS
During the past five years at Michigan so many things have happened and there are
so many people to thank. First and foremost, its my parents who have encouraged me to
pursue PhD studies. I thank them for their encouragements and supports throughout my
academic career. To Prof. Daniel Scheeres, who has been my PhD advisor and a life-long
mentor: it is his guidance and help that made this dissertation exist. I want to thank him,
but no matter how much say here, I would not feel I have said enough. He has taught me
the concept of how much one can owe someone so much. Hence, instead of thanking him,
I promise that I will do the same as I have learned from him. Thank you for teaching me
this valuable lesson! To Sophia Lim, who has patiently encouraged my studies and gave
me the motivation for completing this dissertation: I thank you. Also, I am very grateful
to my doctoral committee members, Dr. Paul Chodas, Prof. Alfred Hero III, Prof. Pierre
Kabamba, and Prof. Harris McClamroch, for their helpful advice and critical comments.
A part of research described in this dissertation was sponsored by the Jupiter Icy Moon
Orbiter project through a grant from the Jet Propulsion Laboratory, California Institute of
Technology which is under contract with the National Aeronautics and Space Administration. I thank Dr. John Aiello, Dr. Lou DAmario, Mr. Try Lam, Dr. Chris Potts,
Dr. Ryan Russell, Dr. Jon Sims, and Dr. Mau Wong from the Jet Propulsion Laboratory
for their helpful comments and suggestions. During my graduate studies, I have spent almost a year at the Jet Propulsion Laboratory as an intern/on-call employee. I was exposed
to many interesting space mission projects: Pioneer anomaly, estimation of the parameiii
terized post-Newtonian parameters, Cassinis synthetic aperture radar and altimetry data
types, and Magellan orbit determination. I thank all the members of the Guidance, Navigation and Control Group and the Radio Science Systems Group for this great opportunity.
Many thanks to Dr. John Anderson, Dr. Sami Asmar, Dr. Shyam Bhaskaran, Dr. Al Cangahuala, Dr. Paul Chodas, Dr. Bob Gaskell, Dr. Moriba Jah, Ms. Eunice Lau, Dr. Michael
Lisano, Mr. Ian Roundhill, and Dr. Slava Turyshev. I have learned so much about deepspace spacecraft navigation and implementation of the radiometric measurements. I have
also spent a few months at the Applied Physics Laboratory as a NASA/APL summer intern, where I have worked on the pre-flight navigation of the NASAs Radiation Belt Storm
Probe. I thank Ms. Linda Butler, Ms. Julie Cutrufelli, Dr. Wayne Dellinger, Dr. David
Dunham, Dr. Robert Farquhar, Dr. Yanping Guo, Dr. Jose Guzman, Mr. Gene Heyler, Mr.
Daniel OShaughnessy, Mr. Gabe Rogers, Dr. Tom Strikwerda, and Dr. Robin Vaughn. It
was a great opportunity to learn about Earth-orbiting missions. Thank you all!
I am also grateful to the collaborators on the General Relativity study during my early
stage of the PhD program. I thank Prof. Ephraim Fischbach and Prof. James Longuski
from Purdue University and Dr. Giacomo Giampieri from Imperial College. Special
thanks to my former advisors, Prof. Robert Melton and Prof. David Spencer, from the
Pennsylvania State University who have given me many reasons to study the astrodynamics and have encouraged me to continue graduate studies at the University of Michigan.
Special thanks to Prof. Carlos Cesnik from the University of Michigan for being a great
mentor and a teacher. Last but not least, I thank all my friends at the University of Michigan who I had technical discussions and have given me the motivation and encouragement
for my PhD research: Steve Broschart, Nalin Chaturvedi, Vincent Guibout, Ji Won Mok,
Rafael Palacios-Nieto, Leo Rios-Reyes, Yoshifumi Suzuki, Benjamin Villac, and many
others. THANK YOU ALL!
iv
PREFACE
This dissertation was submitted in partial fulfilment of the requirements for the degree
of Doctor of Philosophy in Aerospace Engineering at the University of Michigan. The
doctoral committee members were
Dr. Paul W. Chodas, The Jet Propulsion Laboratory, California Institute of Technology,
Prof. Alfred O. Hero III, Electrical Engineering, The University of Michigan,
Prof. Pierre T. Kabamba, Aerospace Engineering, The University of Michigan,
Prof. N. Harris McClamroch, Aerospace Engineering, The University of Michigan,
Prof. Daniel J. Scheeres (Chair), Aerospace Engineering, The University of Michigan,
and the PhD thesis was defended on November 27, 2006. The following list of papers are
either published (or submitted) full journal articles or proceedings presented at technical
conferences. Note that some of these papers are based on the studies during my early stage
of the PhD program and the contents are not discussed in this dissertation. However, these
studies gave me the theoretical insights and technical background needed for this study
and have led to the baseline for the topics of this thesis.
Journal Papers
[76] R.S. Park and D.J. Scheeres, Nonlinear Semi-Analytic Methods for Trajectory
Estimation, submitted to the Journal of Guidance, Control, and Dynamics, November 2006.
[69] R.S. Park and D.J. Scheeres, Nonlinear Mapping of Gaussian State Uncertainties: Theory and Applications to Spacecraft Control and Navigation, Journal of
Guidance, Control, and Dynamics, Vol. 29, No. 6, 2006.
[90] D.J. Scheeres, F.-Y. Hsiao, R.S. Park, B.F. Villac, and J.M. Maruskin, Fundamental Limits on Spacecraft Orbit Uncertainty and Distribution Propagation, accepted for publication, Journal of the Astronautical Sciences, 2005.
[77] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach, Estimating Parameterized Post-Newtonian Parameters from Spacecraft Radiometric
Tracking Data, Journal of Spacecraft and Rockets, Vol. 42, No. 3, 2005.
[53] J. Longuski, E. Fischbach, D.J. Scheeres, G. Giampieri, and R.S. Park, Deflection of Spacecraft Trajectories as a New Test of General Relativity: Determining the
PPN Parameters and , Physical Review D, Vol. 69, No. 042001, 2004.
Conference Papers and Industrial Reports

[72] R.S. Park and D.J. Scheeres, Nonlinear Semi-Analytic Method for Spacecraft
Navigation, paper presented at AAS/AIAA Astrodynamics Specialist Conference, Keystone, Colorado, August 21-24, 2006, AIAA-2006-6399.
[71] R.S. Park, L.A. Cangahuala, P.W. Chodas, and I.M. Roundhill, Covariance
Analysis of Cassini Titan Flyby using SAR and Altimetry Data, paper presented at
vi
AAS/AIAA Astrodynamics Specialist Conference, Keystone, Colorado, August 2124, 2006, AIAA-2006-6398.
[73] R.S. Park, D.J. Scheeres, Nonlinear Semi-Analytic Methods for Spacecraft Trajectory Design, Control, and Navigation, paper presented at New Trends in Astrodynamics and Applications Conference, Princeton, New Jersey, August 16-18, 2006.
[67] R.S. Park, Expected Navigation Performance of the Radiation Belt Storm Probe
Mission, APL Internal Report, SEG-06-022, August 2006.
[70] R.S. Park and D.J. Scheeres, Nonlinear Mapping of Gaussian State Uncertainties: Theory and Applications to Spacecraft Control and Navigation, paper presented
at AAS/AIAA Astrodynamics Specialist Conference, Lake Tahoe, California, August
7-11, 2005, AAS-05-404.
[75] R.S. Park and D.J. Scheeres, Nonlinear Mapping of Gaussian State Uncertainties, paper presented at 15th Workshop on JAXA Astrodynamics and Flight Mechanics, Kanagawa, Japan, July 25-26, 2005.
[91] D.J. Scheeres, F.-Y. Hsiao, R.S. Park, B.F. Villac, and J.M. Maruskin, Fundamental Limits on Spacecraft Orbit Uncertainty and Distribution Propagation, invited
paper presented at the Shuster Symposium, Grand Island, New York, June 2005,
AAS-05-471.
[74] R.S. Park and D.J. Scheeres, Nonlinear Mapping of Gaussian State Covariance and Orbit Uncertainties, paper presented at AAS/AIAA Space Flight Mechanics
Meeting, Copper Mountain, Colorado, January 23-27. 2005, AAS-05-170.
[78] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach, Orbit Design for General Relativity Experiments: Heliocentric and Mercury-centric
vii
Cases, paper presented at AAS/AIAA Astrodynamics Specialist Conference, Providence, Rhode Island, August 16-19, 2004, AIAA-2004-5394.
[68] R.S. Park, E. Fischbach G. Giampieri, J.M. Longuski, and D.J. Scheeres, A
Test of General Relativity: Estimating PPN parameters and from Spacecraft Radiometric Tracking Data, Nuclear Physics B - Proceedings Supplement, Proceedings of the Second International Conference on Particle and Fundamental Physics
in Space, 134(2004), 2004.
[79] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach, Estimating General Relativity Parameters from Radiometric Tracking of Heliocentric
Trajectories, paper presented at AAS/AIAA Space Flight Mechanics Meeting, Ponce,
Puerto Rico, February 2003, AAS-03-205.
[17] C.E.S. Cesnik, R.S. Park, and R. Palacios, Effective Cross-Section Distribution
of Anisotropic Piezocomposite Actuators for Wing Twist, paper presented at the
SPIE 10th International Symposium on Smart Structures and Materials, San Diego,
CA, March 2003.
viii
TABLE OF CONTENTS
DEDICATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ii
ACKNOWLEDGEMENTS . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iii
PREFACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
xii
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi

LIST OF APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvii
NOTATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xviii
CHAPTERS
I. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1
1.2
Spacecraft Navigation and Uncertainty Propagation . . . . . . .

Specific Applications of Uncertainty Propagation to Spacecraft
Navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Scope of this Thesis . . . . . . . . . . . . . . . . . . . . . . . .
Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . .
4
8
11
II. RELATIVE MOTION OF GENERAL NONLINEAR DYNAMICAL

SYSTEMS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
13
1.3
1.4
2.1
2.2
2.3
2.4
2.5
General Trajectory Dynamics and Solution Flows . . . . . . . .

Higher Order Taylor Series Approximations and Solutions . . . .
2.2.1 Complexity of the Higher Order Solutions . . . . . . .
Dynamics and Properties of a Hamiltonian System . . . . . . . .
Symplecticity of the Higher Order Solutions of a Hamiltonian
System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Convergence of the Higher Order Solutions . . . . . . . . . . . .
III. EVOLUTION OF PROBABILITY DENSITY FUNCTIONS IN NONLINEAR DYNAMICAL SYSTEMS . . . . . . . . . . . . . . . . . . .

ix
13
16
25
27
32
43
46
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
Review of Probability Theory and Random Processes . . . . . .

The Gaussian Probability Distribution . . . . . . . . . . . . . .
Dynamics of the Mean and Covariance Matrix . . . . . . . . . .
The Fokker-Planck Equation . . . . . . . . . . . . . . . . . . .
Integral Invariance of Probability . . . . . . . . . . . . . . . . .
Solution of the Fokker-Planck Equation for a Hamiltonian System
On the Relation of Phase Volume and Probability . . . . . . . .
Time Invariance of the Probability Density Function of the Higher
Order Hamiltonian Systems . . . . . . . . . . . . . . . . . . . .
3.9 Nonlinear Mapping of the Gaussian Distribution . . . . . . . . .
3.10 Monte-Carlo Simulations . . . . . . . . . . . . . . . . . . . . .
3.11 Unscented Transformation . . . . . . . . . . . . . . . . . . . . .
46
49
52
53
56
60
63
IV. NONLINEAR TRAJECTORY NAVIGATION . . . . . . . . . . . . .
73
4.1
4.2
4.3
4.4
The Concept of Statistically Correct Trajectory . . . . . . . . . .

Nonlinear Statistical Targeting . . . . . . . . . . . . . . . . . .
4.2.1 On the Theoretic and Practical Aspects of Nonlinear
Statistical Targeting . . . . . . . . . . . . . . . . . . .
Higher Order Bayesian Filter with Gaussian Boundary Conditions
Implementation of a Nonlinear Filter . . . . . . . . . . . . . . .
4.4.1 Extended Kalman Filter . . . . . . . . . . . . . . . .
4.4.2 Higher-Order Numerical Extended Kalman Filter . . .
4.4.3 Higher-Order Analytic Extended Kalman Filter . . . .
4.4.4 Unscented Kalman Filter . . . . . . . . . . . . . . . .
65
67
69
71
75
79
84
88
98
100
101
106
109
V. NONLINEAR SPACE MISSION ANALYSIS . . . . . . . . . . . . . 111

5.1
5.2
5.3
5.4
5.5
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Nonlinear Propagation of Phase Volume . . . . . . . . . . . . .
Nonlinear Orbit Uncertainty Propagation . . . . . . . . . . . . .
5.3.1 Two-Body Problem: Earth-to-Moon Hohmann Transfer
5.3.2 Hill Three-Body Problem: about Europa . . . . . . . .
Nonlinear Statistical Targeting . . . . . . . . . . . . . . . . . .
5.4.2 Hill Three-Body Problem: about Europa . . . . . . . .
Nonlinear Trajectory Navigation . . . . . . . . . . . . . . . . .
5.5.1 Halo Orbit: Sun-Earth System . . . . . . . . . . . . .
5.5.2 Halo Orbit: Earth-Moon System . . . . . . . . . . . .
5.5.3 Potential Applications and Challenges . . . . . . . . .
111
112
120
120
123
125
125
128
131
131
141
152
VI. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS . . . . 156
6.1
6.2
Concluding Remarks and Key Contributions . . . . . . . . . . . 156

Future Research and Recommendations . . . . . . . . . . . . . . 159
APPENDICES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
xi
LIST OF FIGURES
Figure
3.1
Normalizedtwo-dimensional
Gaussian probability density function: p(x,y) =
(1/2) exp 21 (x2 + y2 ) . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2
Integral invariance of a phase volume. . . . . . . . . . . . . . . . . . .
57
4.1
Illustration of the statistically correction trajectory. . . . . . . . . . . .
75
4.2
Illustration of the nonlinear statistical targeting. . . . . . . . . . . . . .
79
4.3
Propagated mean and 1- error ellipsoid projected onto the spacecraft

position plane: comparison of the STT-approach and Monte-Carlo simulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
87
5.1
Hill three-body trajectory plot at Europa for 6.775 days: circled points
are computed at t {0, 0.881, 2.26, 4.42, 5.38, 5.74} days. . . . . . . . 113
5.2
Trajectory norms and higher order solution magnitudes: circled points

are computed at t {0, 0.881, 2.26, 4.42, 5.38, 5.74} days. . . . . . . . 115
5.3
Phase volume projections: solid line represents integrated, dotted line

represents the 1st order, dash-dot line represents the 2nd order, and
dashed line represents the 3rd order solutions. . . . . . . . . . . . . . 117
5.4
Phase volume projections: solid line represents integrated, dotted line

represents the 1st order, dash-dot line represents the 2nd order, and
dashed line represents the 3rd order solutions. . . . . . . . . . . . . . 118
5.5
Phase volume projections. . . . . . . . . . . . . . . . . . . . . . . . . 119
5.6
Two-body problem: Hohmann transfer trajectory. . . . . . . . . . . . . 121
5.7
Two-body problem: comparison of the computed mean and covariance

at apoapsis using STT-approach and Monte-Carlo simulations. . . . . . 121
xii
5.8
Hill three-body problem: a safe trajectory at Europa. . . . . . . . . . . 124
5.9
Hill three-body problem: comparison of the computed mean and covariance at periapsis using STT-approach and Monte-Carlo simulations. . . 124
5.10
Two-body problem: computed Vk using the linear and nonlinear methods.126
5.11
Two-body problem: deviated position and velocity means at the target. . 127
5.12
Two-body problem: Monte-Carlo simulation using the linear and nonlinear methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.13
Hill three-body problem: computed Vk using the linear and nonlinear

methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.14
Hill three-body problem: deviated position and velocity means at the

target. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.15
Hill three-body problem: Monte-Carlo simulation using the linear and

nonlinear methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
5.16
Nominal halo orbit about the Sun-Earth L1 point. . . . . . . . . . . . . 132
5.17
Nominal halo orbit about the Sun-Earth L1 point in x-y plane. . . . . . . 132
5.18
Sun-Earth halo orbit: covariance matrix computed after one orbital period.134
5.19
Sun-Earth halo orbit: comparison of the uncertainties computed using

the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements
are taken every 20 days. . . . . . . . . . . . . . . . . . . . . . . . . . . 135
5.20
Sun-Earth halo orbit: comparison of the absolute errors computed using

5.21

5.22

xiii
5.23

the HNEKFs for the cases m = {1, 2, 3}. Measurements are taken every
20 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.24

20 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
5.25

the HAEKFs for the cases m = {1, 2, 3}. Measurements are taken every
20 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.26

20 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
5.27

the EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements
are taken every 20 days based on the halo orbit Case 2. . . . . . . . . . 140
5.28

are taken every 20 days based on the halo orbit Case 2. . . . . . . . . . 140
5.29
Nominal halo orbit about the Earth-Moon L1 point. . . . . . . . . . . . 143
5.30
Nominal halo orbit about the Earth-Moon L1 point in x-y plane. . . . . 143
5.31
Earth-Moon halo orbit: comparison of the uncertainties computed using

5.32
Earth-Moon halo orbit: comparison of the absolute errors computed using the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 2 days. . . . . . . . . . . . . . . . . . . . . . . . 144
5.33

are taken every 2 days assuming zero initial mean. . . . . . . . . . . . . 147
5.34
Earth-Moon halo orbit: comparison of the absolute errors computed using the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 2 days assuming zero initial mean. . . . . . . . . 147
xiv
5.35

are taken every 2 days assuming zero initial mean and small initial covariance matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.36
Earth-Moon halo orbit: comparison of the absolute errors computed using the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 2 days assuming zero initial mean and small initial
covariance matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
5.37

are taken every 6 hours. . . . . . . . . . . . . . . . . . . . . . . . . . . 149
5.38
Earth-Moon halo orbit: comparison of the absolute errors computed using the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 6 hours. . . . . . . . . . . . . . . . . . . . . . . 149
5.39

2 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.40
Earth-Moon halo orbit: comparison of the absolute errors computed using the HNEKFs for the cases m = {1, 2, 3}. Measurements are taken
every 2 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.41

2 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.42
Earth-Moon halo orbit: comparison of the absolute errors computed using the HAEKFs for the cases m = {1, 2, 3}. Measurements are taken
every 2 days. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.43

are taken every 2 days based on the halo orbit Case 2. . . . . . . . . . . 153
5.44
Earth-Moon halo orbit: comparison of the absolute errors computed using the EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements are taken every 2 days based on the halo orbit Case 2. . . . . . . 153
A.1
Families of halo orbits about the Sun-Earth L1 point in non-dimensional

frame. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
xv
LIST OF TABLES
Table
5.1
Local nonlinearity index. . . . . . . . . . . . . . . . . . . . . . . . . . 122
5.2
Halo orbit maximum amplitudes with respect to the Sun-Earth L1 point. 131
5.3
Halo orbit maximum amplitudes with respect to the Earth-Moon L1 point. 142
A.1
Properties of planets and satellites. . . . . . . . . . . . . . . . . . . . . 164
A.2
Properties of three-body systems. . . . . . . . . . . . . . . . . . . . . . 166
xvi
LIST OF APPENDICES
Appendix
A.
EQUATIONS OF MOTION OF ASTRODYNAMICS PROBLEMS . . . . 163

A.1 The Two-Body Problem . . . . . . . . . . . . . . . . .
A.2 The Three-Body Problem . . . . . . . . . . . . . . . .
A.2.1 The Circular Restricted Three-Body Problem
A.2.2 The Hill Three-Body Problem . . . . . . . .
A.2.3 Halo Orbit . . . . . . . . . . . . . . . . . .
B.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
163
164
165
166
167
PROPERTIES OF PROBABILITY DENSITY FUNCTIONS . . . . . . . 169

B.1 Integral Invariance of the PDF of a Linear Hamiltonian Dynamical System with Gaussian Boundary Conditions . . . . . . . . . 169
C.
THE LINEAR KALMAN FILTER . . . . . . . . . . . . . . . . . . . . . 171

C.1 Kalman Filter Essentials . . . . . . . . . . . . . . . . . . . . . . 171
C.2 Kalman Filter Derivation . . . . . . . . . . . . . . . . . . . . . 173
D.
VECTORIZATION OF HIGHER ORDER TENSORS . . . . . . . . . . . 180

D.1 Specifications for MATLAB . . . . . . . . . . . . . . . . . . . . 180
D.2 Specifications for C or C++ . . . . . . . . . . . . . . . . . . . . 181
xvii
NOTATION
Scalars, vectors, and matrices:

Scalars are denoted by upper or lower case Roman or Greek letters in italic type,
e.g., p, N , , or .
Vectors are denoted by lower case Roman or Greek letters in boldface type, e.g.,
x or . The vector x is composed of elements xi or the vector is composed of
elements i . Components of a vector are denoted with a boldface superscript, e.g.,
given x = [rT , vT ]T , xr = r and xv = v, where xr and xv themselves are vectors.
When a vector contains conflicting superscripts, parentheses and brackets are used
to clarify a component of a vector, e.g., (x )i is an ith component of x .
Matrices are denoted by upper case Roman or Greek letters in boldface type, e.g.,
A or . The matrix A is composed of elements Aij , which indicates the ith-row and
jth-column entry of A. When a matrix contains conflicting superscripts, parentheses
and brackets are used to clarify a component of a matrix, e.g., (P )ij is an ith-row
and jth column entry of P . A component of a matrix operation is denoted using
superscripts with parentheses or brackets, i.e., Aij Bjk = (AB)ik , or Aij ()Bjk () =
[A()B()]ik .
Subscripts:
A plain subscript of a scalar, a vector, or a matrix denotes the time which the variable
is computed, e.g., xi0 = xi (t0 ), x0 = x(t0 ), or Hk = H(tk ).
A boldface subscript of a scalar denotes the row-wise partial derivative, e.g., Hx =
H(x)/x or Hxx = 2 H(x)/xx.
Superscripts:
: denotes the transpose of a vector or a matrix, e.g., xT or T .
: denotes the inverse of a matrix with full-rank, e.g., 1 .
: denotes predicted value of a scalar, a vector, or a matrix from a filter, e.g., x ,

m , or P .
: denotes updated value of a scalar, a vector, or a matrix from a filter, e.g., x+ , m+ ,

or P+ .
xviii
i,1 ,p
: denotes the pth order partial derivative of a scalar or a vector, e.g., yi,1 p (x)
= p yi (x)/x1 xp , where xj indicates the j th component of x.
is reserved to denote an th solution of an iteration.

is reserved to denote a th sample chosen from a phase volume or a probability
distribution.
Exceptions:
Vk R3 is a vector denoting an impulsive correction maneuver applied to a spacecraft trajectory.
Partial derivatives of a scalar are denoted with subscripts, e.g., Hi = H(x)/xi or
Hijk = 3 H(x)/xi xj xk .
Dimensions:
A Lagrangian system has a dimension n, e.g., a Lagrangian system with position
R3 and velocity R3 has a dimension n = 3.
A Hamiltonian system has a dimension 2n, e.g., a Hamiltonian system with generalized coordinate R3 and generalized momenta R3 has a dimension 2n = 6.
When a Lagrangian system is transformed into a full-state system, it has a dimension
N = 2n, e.g., when a Lagrangian system with position R3 and velocity R3 is
transformed into a full-state system, it has a dimension N = 6.
Mathematical symbols:
det : denotes a determinant.
E() : denotes an expectation.
E(|) : denotes a conditional expectation.
exp : denotes an exponential function.
lim : denotes a limit.
sup : denotes a supremum.
Trace(): denotes the trace of a matrix.
a b : denotes a dot product of a and b.
| | : denotes an absolute value.
k k : denotes a norm.
: denotes an element of a vector or a set.
: denotes a subset.
xix
CHAPTER I
INTRODUCTION
"As far as the laws of mathematics refer to reality, they are not certain; and
as far as they are certain, they do not refer to reality."
- Albert Einstein
1.1 Spacecraft Navigation and Uncertainty Propagation

Given a dynamical system, the evolution of a particular state can be completely characterized by the systems governing equations of motion. In reality, however, the state is
always associated with some errors that may be due to uncertain system models, inputs,
or measurements. Hence, studying a deterministic trajectory may not provide sufficient
information about the trajectory. For this reason, the problem of uncertainty propagation
has received much attention in many engineering and scientific disciplines. Given an initial state and its associated uncertainties (usually described with a mean and a covariance
matrix or a probability density function), the goal of uncertainty propagation is to predict the state and its statistical properties at some future time, or possibly along the entire
trajectory, considering the statistical properties of the initial state. Except under certain assumptions, however, uncertainty propagation is an extremely difficult process if we want
a complete statistical description. This is because it generally requires one to solve par-
2
tial differential equations such as the Fokker-Planck equation or to carry out particle-type
studies such as Monte-Carlo simulations. Therefore, in practice, an approximation method
is usually required.
For spacecraft trajectory design and operations, uncertainty propagation usually refers
to orbit uncertainty propagation, where the mean and covariance matrix of the spacecraft
state are determined. Conventionally, the usual assumptions in spacecraft trajectory problems are:
A linearized model sufficiently approximates the relative dynamics of neighboring
trajectories with respect to a reference trajectory.
The covariance matrix can be determined as the solution of a Riccati equation while
assuming the reference trajectory is the mean trajectory.
The orbit uncertainty can be completely characterized by a Gaussian probability
distribution.
These are usually good assumptions for spacecraft applications as we are usually given a
reference (nominal) trajectory with a high degree of precision. The objective of trajectory
navigation and spacecraft control is to follow a reference trajectory while minimizing some
pre-defined optimality constraints, such as the number of trajectory correction maneuvers,
flight time, fuel usage, etc. The basic underlying concept is to stay close enough to the
reference trajectory so that the linear dynamics assumption applies. This can be achieved
by taking a sufficient number of measurements along the trajectory so that the deviation
and statistics can be accurately mapped linearly using the state transition (fundamental)
matrix.
The trajectory navigation problem is, at heart, a local problem, meaning that we wish
to locate, control, and predict the spacecraft trajectory relative to a nominal path. When
3
uncertainties are small we often only need a linear characterization. For large uncertainties, however, the region of phase space where the spacecraft may be located is large and
may require that we incorporate nonlinear local dynamics. For example, consider an interplanetary trajectory where the nominal trajectory leads to Mars. In reality, it is often
the case that launch errors are large and lead to the actual trajectory deviating from the
nominal. Also, orbit determination (OD) has a limited ability to locate the spacecraft, so
even after tracking, the spacecrafts state is still only defined as a probability distribution.
To correct the deviated trajectory, correction maneuvers must be made to target back to
the Mars aim point. These target maneuvers have errors as well, due to the uncertainty
of where the spacecraft lies; thus later corrections must be planned for and executed to
provide additional corrections. Ideally, these sequences of correction maneuvers converge
and a final, small maneuver days before arrival is all that is needed to hit the aim point
with sufficient accuracy. However, when OD errors are large, the trajectory dynamics unstable, only a limited number of measurements available, or time spans are long, the use
of linear dynamics models can result in additional errors and lead to more and larger correction maneuvers being required, inaccurate uncertainty predictions, and poorer filtering
performance. It is these issues that this thesis deals with and solves.
The linear dynamics assumption, when applied to spacecraft navigation, simplifies
the problem a great deal, due mainly to the existence of a closed-form solution for the
local dynamics. However, astrodynamics problems are nonlinear in general and the linear
assumption can sometimes fail to characterize the true spacecraft dynamics and statistics
when a system is subject to a highly unstable environment or when mapped over a long
duration of time. Hence, in such cases, an alternate method which accounts for the system
nonlinearity must be implemented.
When such nonlinearities are important, the best known technique for propagating or-
4
bit uncertainty is a Monte-Carlo (MC) simulation, which approximates the probability
distribution by averaging over a large set of random samples. A Monte-Carlo simulation
can provide true statistics in the limit, but is computationally intensive and only solves
for the statistics of a specific epoch and its associated uncertainties. Hence, for mission
operations, these difficulties make Monte-Carlo simulations inefficient for practical spacecraft applications. One way to simplify the implementation would be to propagate each
random sample analytically based on the linearized model using the state transition matrix. It is, however, inapplicable for highly nonlinear trajectories or problems with large
initial errors, since the true trajectory (in a statistical sense) may not lie entirely within the
linear regime. Other approaches to orbit uncertainty propagation have also been considered. Junkins et al. [47, 48] analyzed the effect of the coordinate system on the propagated
statistics, and found that using osculating orbit elements improved future prediction. Their
propagation method was still based on a linear assumption and system nonlinearity was
not incorporated in the mapping. Another approach to orbit uncertainty propagation is
a reduced Monte-Carlo method, such as approximating a probability distribution by the
line-of-variation (LOV), which is a line chosen along the most uncertain direction of the
uncertainty distribution [60]. An additional example of a reduced Monte-Carlo method is
the unscented transformation, which approximates the probability distribution by nonlinearly integrating a set of deterministically chosen sample points [45, 44].
1.2 Specific Applications of Uncertainty Propagation to Spacecraft

Navigation
In the following we list a number of applications of orbit uncertainty propagation for
spacecraft trajectory problems. These applications apply to both traditional linear mappings and to nonlinear mapping methods:
5
Pre-mission covariance analysis:
Covariance analysis is a design tool which is often used in spacecraft missions to
characterize the navigation performance considering the statistical properties of the
system and measurements. Examples of covariance analysis are predictions of how
well estimation of some parameters can be made [53, 58, 59, 77] or how well a
spacecraft state can be estimated using radiometric measurements, such as range
and Doppler data [71, 89]. Orbit uncertainty propagation in a covariance analysis
refers to the computation of a covariance matrix which characterizes the level of accuracy to which a spacecraft orbit can be estimated from an OD process, and provide
predictions of delivery accuracy. Therefore, improved orbit uncertainty propagation
can provide a more accurate determination of a covariance matrix, and thus, a more
realistic estimation and mission scenario can be simulated.
Mission prediction of future uncertainties and confidence regions:
Trajectory prediction is an important problem in mission design as well as in celestial mechanics. The goal is to propagate the estimated spacecraft state to a future
time and construct a confidence region where a spacecraft, or a celestial body, should
be located with some probability level. For example, consider a small-body, such
as an asteroid or a comet. We want to map both the bodys state and its probability
distribution so that we can project its uncertainties onto a target plane to obtain a
confidence region [60]. When the confidence region is large, additional measurements may be necessary to reduce the error bounds. Hence, it is not sufficient to
only consider a deterministic trajectory, but the statistical properties of a trajectory
are also important.
6
Design and planning of statistical correction maneuvers:
In mission operations, a spacecraft often deviates from the nominal trajectory due
to uncertainties in the state and measurements and unmodeled accelerations acting
on the spacecraft. Therefore, a spacecraft performs series of correction maneuvers
to converge to a target or back to the nominal trajectory. Conventionally, trajectory
correction maneuvers are computed based on the solution of a deterministic trajectory. The uncertainty propagation then verifies that the applied correction maneuver
delivers the spacecraft to a desired target with tolerable error bounds. Using a statistical targeting method, however, it is also possible to incorporate the trajectorys
statistical information to improve targeting performance. For example Ref. [84]
discusses the computation of statistical correction maneuvers that optimizes the trajectory maintenance problem within an unstable dynamical environment.
Computation of probability density functions using first few moments:
In conventional navigation, a spacecraft state is usually modeled with Gaussian
statistics, and hence, its probability distribution is completely characterized by the
first two moments (mean and covariance matrix). Considering nonlinear trajectory
dynamics, however, the propagated probability distribution no longer preserves the
Gaussian structure. An approximation to the mapped probability distribution can be
made using the first few moments of the system. This can provide useful information in case mission operations do not require a complete statistical description of a
trajectory.
Filtering and orbit determination:
A filter is usually composed of two parts, prediction and update. Orbit uncertainty
propagation relates to the prediction problem while the uncertain distribution influ-
7
ences the update part. In conventional trajectory navigation, a spacecraft is initially
assumed to lie within a certain probability ellipsoid and the spacecraft trajectory
is sequentially estimated until the solution converges. Mission operations usually
implement an extended Kalman filter for trajectory navigation; however, when the
trajectory is significantly nonlinear, it may be necessary to consider a filter that incorporates system nonlinearity. Examples of such filters include:
Divided difference filter: approximates state using polynomial approximations
from a multi-dimensional interpolation formula [63],
Gaussian sum filter: approximates the conditional probability density using
the sum of Gaussian distributions [38, 95],
Hammerstein filter: incorporates the system nonlinearity using a linear Hammerstein system [11, 28, 41, 42],
Higher order filter: applies higher order Taylor series expansion to estimate
the mean and covariance matrix [6, 14, 55],
Particle filter: approximates the conditional probability density using ensemble of random sample points [5, 7, 15, 21, 34, 35], and
Volterra filter: applies Volterra series to estimate the higher order moments for
system identification and estimation [1, 3, 64, 65, 66].
These filters, however, have not been implemented to real spacecraft trajectory navigation problems, except for a few special cases. This is mainly because the extended
Kalman filter has, so far, provided sufficient accuracy for mission operations and the
development of software tools for a nonlinear filter can be quite costly, and thus,
may not be cost-effective. However, once a nonlinear filtering capability is devel-
8
oped and made feasible, it can provide more accurate science return and potentially
reduced mission cost.
A posteriori reconstruction:
Given a set of measurements of a spacecraft, the OD process reconstructs the spacecraft trajectory to some level of accuracy. This is usually carried out using the batch
least-squares approximation, which minimizes the sum of squared measurement
residuals. From the trajectory reconstruction, an a posteriori spacecraft trajectory
can be estimated and its statistics can be characterized. This is often useful for
interpretation of scientific measurements.
Among the many applications of orbit uncertainty propagation, this thesis mainly focuses
on the trajectory navigation problem where we consider the problems of nonlinear orbit
prediction, nonlinear statistical maneuver design, and nonlinear trajectory filtering and
orbit determination.
1.3 Scope of this Thesis

The goal of this thesis is to develop an analytical framework for nonlinear trajectory
navigation. Throughout this dissertation, we assume a trajectory can be modeled as a
Hamiltonian dynamical system with no diffusion (i.e., no process noise). Although this is
not a completely realistic assumption, it is a standard and quite useful first approach to this
problem.
For a given reference trajectory, the nonlinear relative motion can be approximated by
applying a Taylor series expansion of the solution function in terms of the initial conditions [62, 94]. The nonlinear local trajectory dynamics is then characterized by the higher
order Taylor series terms that are extensions of the state transition matrix (STM) to higher
orders. The theory and implementation of this approach are reasonably straightforward
9
and can be easily adapted to many spacecraft applications that are based on linear theory. However, there have been almost no studies on the use of higher order analysis for
trajectory navigation problems. The significance of such a higher order approach is that
it provides an analytic expression of the local nonlinear trajectory as a function of initial
conditions. Thus, trajectory propagation only requires a simple algebraic manipulation.
Using this nonlinear local solution we solve the Fokker-Planck equation for deterministic (i.e., diffusion-less) Hamiltonian systems and establish the time invariance property
of the probability density function. Also, assuming the local nonlinear solutions are computed and the initial probability distribution is precisely known, we derive an analytic
expression for propagation of the orbit uncertainties. Note that this analytic formulation
determines the probability distribution from the solution of a nonlinear function, rather
than from an empirical sampling technique such as a Monte-Carlo simulation. Assuming a Gaussian initial state with this approach, the statistics (mean and covariance matrix)
computed using the higher order approach provide good agreement with Monte-Carlo simulations over a reasonable time period in the presence of strong nonlinearities.
Using this result, we first consider the design of a statistically correct trajectory. Given
a trajectory with an initial probability distribution, the mean trajectory will, depending on
the systems nonlinearity, deviate from the reference (nominal) trajectory. To compensate
for this, we introduce the concept of the statistically correct trajectory by solving for the
initial state where the trajectory will satisfy a desired target state condition on average
(e.g., the final state of a boundary value problem), not from a deterministic solution. This
is a seemingly counter-intuitive approach since this implies that the initial state must be
different from the solution of the deterministic boundary value problem. However, this
should yield a better, more realistic (and practical), trajectory design according to probability theory.
10
As an extension of the statistically correct trajectory, we define a nonlinear statistical
targeting method where we solve for a correction maneuver based on the mean trajectory.
The usual linear targeting method solves for a maneuver for a deterministic trajectory. Due
to orbit uncertainties, however, the final target will be offset from the desired target and
additional correction maneuvers may be needed. If we solve for a correction maneuver
using our nonlinear approach the number of correction maneuvers may be reduced, as it
delivers the mean trajectory to the target.
Our approach can also be applied to Bayesian filtering. We present a general filtering
algorithm for optimal estimation of the conditional density function incorporating nonlinearity in the filtering process. We then derive practical Kalman-type filters based on our
formulation. The first type, the higher-order numerical Extended Kalman filter, is based
on an extension of the extended Kalman filter, where we integrate the nonlinear flow and
the higher order solutions between each measurement update. This is in a sense similar to
the second order filter by Athans et al. [6], but can be generalized to higher orders. The
second type is called the higher-order analytic extended Kalman filter and is an extension
of the linear Kalman filter, which assumes that the nonlinear flow and the higher order
solutions are available prior to filtering, and hence, requires no on-line integration. For
a nonlinear spacecraft trajectory we show that these filters have superior performance as
compared with the conventional EKF.
The following is a list of the key contributions of this thesis:
A general theory for nonlinear relative motion is developed.
A semi-analytic method for orbit uncertainty propagation is derived by applying the
solutions of higher order relative dynamics with a known initial probability distribution.
11
The concept of a statistically correct trajectory is introduced by incorporating navigation information in the trajectory design process.
A nonlinear statistical targeting method is developed by analytically computing a
correction maneuver that hits the target on average.
An optimal solution of the posterior conditional density function is presented by
solving Bayes rule for state and measurement probability density functions.
Practical Kalman-type filters are derived by incorporating nonlinear dynamical effects in the uncertainty propagation.
1.4 Thesis Organization

This thesis starts from the basics of a general dynamical system and probability theory,
and poses trajectory problems in an analytic framework.
In Chapter II we develop an analytic trajectory propagation method. We first discuss
the dynamical aspects of general astrodynamics problems and define the nonlinear relative
dynamics with respect to a reference trajectory. We then present how the relative motion of
a spacecraft can be completely characterized by computing the forward and inverse state
transition tensors. A review of Hamiltonian dynamical systems is also given and their
unique properties and facts are discussed.
In Chapter III, we present a review of probability theory and give a discussion of the
Fokker-Planck equation, which governs the evolution of the probability density function.
The solution of the Fokker-Planck equation is then analyzed for a deterministic system
and the time invariance of the probability density function for a Hamiltonian dynamical
system is derived. This is then combined with results from Chapter II and applied to orbit
uncertainty propagation as a function of the initial state and associated uncertainties.
12
Chapter IV presents several applications where orbit uncertainty propagation can be
utilized. We introduce the concept of the statistically correct trajectory, and extend the
idea to a nonlinear statistical targeting problem. Also, we discuss how the higher order
solutions can be implemented in a Bayesian filtering algorithm to compute the optimal
posterior conditional density function. We then extend this idea and derive Kalman-type
filters based on numerical and analytical propagation of the orbit statistics.
Chapter V gives several examples and simulations of our methods based on the twobody, the Hill three-body, and the circular restricted three-body problems. The examples
assume realistic initial conditions and errors, and we compare the nonlinear STT-approach
to conventional linear methods.
Finally, conclusions and future work are presented in Chapter VI. Our study shows
that a nonlinear relative trajectory of sufficient order recovers Monte-Carlo simulation results. Also, nonlinear statistical targeting provides a statistically more accurate correction
maneuver than the conventional linear method. For a nonlinear filtering problem, we show
that our higher order filters provide faster convergence and a superior solution as compared
to linear filters.
CHAPTER II
RELATIVE MOTION OF GENERAL NONLINEAR
DYNAMICAL SYSTEMS
In this chapter, relative motion about a nominal spacecraft trajectory is presented. We

first express the general dynamics of a spacecraft as first order differential equations and
define the solution flow which represents the nominal path of a spacecraft. We then derive
the relative motion by applying a Taylor series expansion to a nominal trajectory and discuss the computation of higher order Taylor series terms that describe the local nonlinear
motion. Also presented are the properties of a Hamiltonian system concerning astrodynamics problems that can be modeled as a Hamiltonian system.
2.1 General Trajectory Dynamics and Solution Flows

In this thesis, we consider astrodynamics problems that can be modeled as:
r(t) = f[t, r(t), r (t)],
(2.1)
where r(t) R3 represents the position vector and r (t) = v(t) R3 represents the
velocity vector.1 Considering Eqn. (2.1) and v = r , it is apparent that the system dynamics
can be transformed into first order differential equations. To show this, let x = [rT , vT ]T
1
Examples of this model consist of problems such as the two-body problem, Hill three-body problem,
and circular restricted three-body problem. The governing differential equations for these examples are
given in Appendix A.
13
14
represent the spacecraft state vector. The governing differential equations for x can be
written as:
v(t)
r (t)
x(t)
=
=
.
v(t)
f[t, r(t), v(t)]
(2.2)
By letting g(t) = [vT (t) , fT (t)]T , the general dynamics of a spacecraft can be stated as:
x(t)
= g[t, x(t)],
(2.3)
with dimension N = 6 and an initial state x0 = x(t0 ).

Definition 2.1.1 (Solution Flow). A solution flow, (t; x0 , t0 ), is a map of the initial state
x0 at time t0 to a state x at time t and is defined as:
x(t) = (t; x0 , t0 ),
(2.4)
where t0 and x0 are free variables and the solution flow is governed by:
d(t; x0 , t0 )
= g[t, (t; x0 , t0 )],
dt
(2.5)
(t0 ; x0 , t0 ) = x0 .
(2.6)
Definition 2.1.2 (Phase Volume). A phase volume is a subset of Euclidean space RN that
is compact (closed and bounded)2 ; e.g., the closed unit interval [0, 1] is a phase volume in
R.
Suppose we are given an initial phase volume B0 = B(t0 ). Using the solution flow
notation, the evolution of B0 can be defined as the mapping of every point in an initial set
as:
B(t) = {x | x = (t; x0 , t0 ) x0 B0 }.
2
(2.7)
A phase volume is usually defined for the phase space of a Hamiltonian system (in a generalized
coordinate-momenta coordinate frame), but in this thesis, a phase volume is also defined for a Lagrangian
system (in a position-velocity coordinate frame).
15
If B0 is defined locally with respect to a nominal initial condition 0 , Eqn. (2.7) represents
all possible trajectories in a compact and closed neighborhood of the nominal trajectory
(tk ; 0 , t0 ).
Definition 2.1.3 (Inverse Solution Flow). Suppose we are given the solution flow of a
nominal initial state x0 for some time interval [t0 , tk ], i.e., xk = (tk ; x0 , t0 ). The inverse
solution flow (t, x; t0 ) is defined as a map of x at a time t (t0 t tk ) to a fixed state
x0 at time t0 :
x0 = (t, x; t0 ).
(2.8)
The inverse flow can be defined using the solution flow as (t, x; t0 ) = (t0 ; x, t). Also,
(t, x; t0 ) satisfies a partial differential equation:
dx0

=
+
g(t, x) = 0.
dt
t
x
(2.9)
Remark 2.1.4. Combining the definitions of the forward and inverse flows, an obvious, but
important, identity exists:
x0 = [t, (t; x0 , t0 ); t0 ].
(2.10)
Definition 2.1.5 (Relative Motion and Dynamics). Given a nominal initial state x0 , the
relative motion x(t) with respect to the reference (nominal) trajectory is defined as:
x(t) = (t; x0 + x0 , t0 ) (t; x0 , t0 ),
(2.11)
where x0 represents a deviation in x0 . The relative motion satisfies the equations of

motion:
x(t)
= g[t, (t; x0 + x0 , t0 )] g[t, (t; x0 , t0 )].
(2.12)
16
Remark 2.1.6. Using tensor notation, Eqns. (2.11) and (2.12) can be stated as:
xi (t) = i (t; x0 + x0 , t0 ) i (t; x0 , t0 ),
(2.13)
x i (t) = gi [t, (t; x0 + x0 , t0 )] gi [t, (t; x0 , t0 )].
(2.14)
2.2 Higher Order Taylor Series Approximations and Solutions

Definition 2.2.1 (Taylor Series Expansion). Given an infinitely differentiable, real function s(x) with s, x RN , the Taylor series expansion about a point x = a is defined as
[8, 10, 62]:
" N
#j
X
X
1
i 1
N
i 1
N
k
k
s (x , , x ) =
(x a ) k s ( , , )
j! k=1
j=0
(2.15)
l =al
where represents a dummy variable. In vector form, the Taylor series expansion can be
stated as:
1
j i
s (x) =
[(x a) ] s () ,
j!
=a
j=0
i
(2.16)
where represents a gradient operator:
.
1
N
(2.17)
Definition 2.2.2 (Radius of Convergence). Consider the Taylor series expansion given in
Eqn. (2.16). A Taylor series is essentially a power series, and thus, the solution may converge or diverge depending on the size of kx ak. To ensure that the series is convergent,
we define the radius of convergence Rc as:
the series si (x) converges absolutely x such that 0 kx ak Rc and
the series si (x) diverges x such that kx ak > Rc .
17
Remark 2.2.3. We can re-order the j = 0 term in Eqn. (2.16) and rewrite the Taylor series
expansion as:
1
j i
s (x) s (a) =
[(x a) ] s () .
j!
=a
j=1
i
(2.18)
In this form, it is apparent that the relative motion and dynamics defined in Eqns. (2.11)
and (2.12), respectively, can be approximated along a nominal trajectory solution.
Remark 2.2.4. An M th order Taylor series is defined as:
M
X
1
j i
s (x) s (a) =
[(x a) ] s () .
j!
=a
j=1
i
(2.19)
Note that this is an approximation of Eqn. (2.18) by truncating the solution up to order M .
Definition 2.2.5 (Relative Trajectory Dynamics). Suppose we are given a system with the
governing differential equations g[t, x(t)] which is at least M times differentiable. By
applying the M th order Taylor series expansion about the reference trajectory (t; x0 , t0 ),
where x0 is the initial condition, the relative trajectory dynamics can be stated using the
Einstein summation convention as:
M
X
1 i,1 p 1
g
x xp ,
x (t) =
p!
p=1
i
(2.20)
where j {1, , N }, superscripts j denote the j th component of the state vector, and
g
i,1 p
(t, x) =
p gi [t, (t)]
.
1 p j =j (t;x0 ,t0 )
(2.21)
Then, for a given initial deviation x0 with respect to x0 , the relative trajectory motion x(t; x0 , t0 ) can be computed by integrating Eqn. (2.20) along the nominal flow
(t; x0 , t0 ). Note that this formulation is an extension of the conventional linear dynamics
theory.
18
Definition 2.2.6 (State Transition Tensors). Suppose we are given a reference solution flow
(t; x0 , t0 ) computed according to the governing differential equations g[t, x(t)], where x0
represents the nominal initial state. Moreover, assume g[t, x(t)] is at least m times differentiable. By applying the mth order Taylor series expansion about the initial condition,
the relative trajectory motion can be stated as:
m
X
1 i,1 p 1
x (t) =
(t,t0 ) x0 x0p ,
p!
p=1
i
(2.22)
where
i,
(t,t10 ) p (t; x0 , t0 )
p i (t; 0 , t0 )
.
01 0 p j =xj
0
(2.23)
The higher order partial derivatives of the solution flow, i.e., Eqn. (2.23), relate deviations
in the initial state at time t0 to deviations in the state at time t (i.e., x0 7 x(t)): denoted
by the subscript (t, t0 ). We call these partial derivatives the state transition tensors (STTs).
Note that assuming the STTs are available, the relative solution flow is analytic in x0 .
Throughout this thesis, we note that an M th order solution refers to the integrated
relative motion according to Eqn. (2.20) and an mth order solution refers to the analytically mapped solution according to Eqn. (2.22). Moreover, unless stated otherwise we
assume m = M . In conventional practice, the higher order series are usually truncated at
M = m = 1, which is a first order or linear analysis. Then, Eqns. (2.21) and (2.23) are
the usual linear dynamics matrix and the state transition matrix (STM), respectively.
It is important to note that the relative trajectory motion x0 7 x(t) computed according to Eqns. (2.20) and (2.22) are fundamentally different. This is not only because
the relative motion is numerically integrated in Eqn. (2.20) and is analytically mapped
(assuming the STTs are available) in Eqn. (2.22), but also due to the difference in their
accuracies. To clarify these differences, consider an M th order relative motion. This integrated solution is the M th order solution of a particular initial deviation x0 and is accurate
19
up to order M since the M th order partial derivatives of the dynamics gi,1 M are directly
incorporated in the integration.3 On the other hand, an mth order solution assumes that the
STTs are available and the relative motion is mapped analytically, and thus, the solution
accuracy depends on the accuracy of the STTs. If we consider the M = m = 1 case,
both approaches yield identical results. However, in general, numerically integrated and
analytically mapped solutions at the same order do not share the same level of accuracy.
Specifically, the analytic solution generally requires the STTs up to infinite order to capture
the full nonlinear effects captured by an M th order solution, which is discussed in Remark
2.2.11 in more detail. If we consider a sufficiently large m, however, the analytic solution should yield the same level of accuracy as the numerical solution, and ultimately give
an accurate approximation of the true nonlinear relative motion defined in Eqn. (2.11).
Considering this fact, in this thesis, we extend the analytic method to develop analytic
techniques for trajectory navigation problems.
Example 2.2.7. The summation convention applied to a second order expansion of the
solution is written as:
2
X
1 i,1 p 1
x (t) =
(t,t0 ) x0 x0p ,
p!
p=1
i
N
X
1 =1
1
1
i,
(t,t0 ) x0
N X
N
X
1 i,1 2 1 2
+
x x ,
2 (t,t0 ) 0 0
=1 =1
1
(2.24)
for 1 , 2 {1, , N }.
Remark 2.2.8. The partial derivative of the mth order solution, Eqn. (2.22), can be stated
as:
N
X
1
(xi )
i,1 p1 j x01 x0p1 .

=
j
(p 1)!
(x0 )
p=1
(2.25)
Note that this is not analytic in x0 as the integration must be carried for different initial conditions.
20
Remark 2.2.9. The time derivative of the local trajectory solution, Eqn. (2.20), can be
obtained by directly differentiating Eqn. (2.22):
x i (t) =
m
X
1 i,1 p 1
x0 x0p .
p!
p=1
(2.26)
To analyze the deviation x as an analytic function of the initial deviations x0 , we

must solve for the STTs along the reference trajectory. To obtain differential equations for
the STTs, first substitute Eqn. (2.22) into Eqn. (2.20), which gives the equation of x i as
a function of the STTs and the initial conditions. By equating this with Eqn. (2.26) and
balancing terms of the same order in x0 , the differential equations for the STTs ( i,1 p )
can be obtained.4 Using the Einstein summation convention, the ODEs up to fourth order
can be stated as:
i,a = gi, ,a ,
i,ab = gi, ,ab + gi, ,a ,b ,
(2.27)
(2.28)
i,abc = gi, ,abc + gi, ,a ,bc + ,ab ,c + ,ac ,b

+ gi, ,a ,b ,c ,
(2.29)
i,abcd = gi, ,abcd + gi, ,abc ,d + ,abd ,c + ,acd ,b + ,ab ,cd
+ ,ac ,bd + ,ad ,bc + ,a ,bcd + gi, ,ab ,c ,d

+ ,ac ,b ,d + ,ad ,b ,c + ,a ,bc ,d + ,a ,bd ,c
+ ,a ,b ,cd ) + gi, ,a ,b ,c ,d ,
(2.30)
where all indices are over {1, , N }. At t = t0 , the initial conditions of STTs are
i,a
(t0 ,t0 ) = 1 if i = a and all other initial STTs are zero.
Another way of obtaining the differential equations of the STTs is through direct differentiations of Eqn.
(2.23) according to the chain rule. An example of this approach is given in Example 2.2.10.
21
Example 2.2.10. The second order differential equations for the STTs can be obtained as
follows:
i
d i,ab
d 2 xik
2 gi
g
(tk ,t0 ) =
=
=
,
b
b
a
a
b
dt
dt x0 x0
x0 x0
x0 xa0
,a
) (gi, ) ,a
i, ,a
i, (
g
= g
=
+
,
xb0
xb0
xb0
= gi, ,ab + gi, ,a ,b .
For computation, the STTs are put into a vectorized form as first order differential
equations and are numerically integrated along the reference trajectory.5 Moreover, the
numerical integration cost can be significantly reduced by considering the symmetry of
the STTs, excluding the m = 1 case. For example, a system with a dimension N = 6
requires integration of 1554 number of equations for the 3rd order solution; however,
considering the symmetry of the STTs, the total number of integrated solutions reduce to
504. The numerical complexity of the higher order solution is further discussed in 2.2.1.
Remark 2.2.11. In general, an analytic mth order solution requires STTs up to infinite
order if we want to accurately approximate the numerically integrated M th order solution, where M = m. To illustrate this point, consider the second order analytic solution
discussed in Example 2.2.7:
1 i,ab a b
a
xi (t) = i,a
(t,t0 ) x0 + (t,t0 ) x0 x0 ,
2
i,ab
which depends on the first and second order STTs, i,a
(t,t0 ) and (t,t0 ) , respectively. These
STTs, however, do not capture all first and second order dynamics. For example, from
Eqns. (2.27-2.30), we observe that there are first order dynamics terms in the third and
fourth order STT differential equations, i.e., gi, ,abc and gi, ,abcd , that are not included
in the second order analytic solution; the same analogy applies to the second order dynamics, e.g., gi, ,a ,b from the third order STT differential equation. Hence, in general,
5
See Appendix D for more detail.
22
we need STTs up to infinite order to accurately approximate the numerical solution at a
given order. Note that this does not apply to the case m = 1 since the higher order STTs
(m 2) are initially zero, and hence, the contributions due to the first order dynamics are
zero, e.g., gi, ,abc = gi, ,abcd = 0.
Once the STTs are computed by integrating along a nominal trajectory, they serve
an identical role to the STM except that the higher order effects are now included, and
thus, the solution is nonlinear. The main significance of the STTs is that they allow the
local nonlinear motion of a spacecraft trajectory to be mapped analytically. Given a set
of STTs, the evolution of a state relative to the nominal trajectory is a simple algebraic
manipulation, and any neighboring trajectories within the radius of convergence can be
mapped analytically with respect to the reference solution. The solution to the original
dynamical system is then found by adding the deviations to the reference solution, or
i (t; x0 + x0 , t0 ) = i (t; x0 , t0 ) + xi (t).
Similar to the state transition tensors, we can also define a series mapping a deviation
backward in time.
Definition 2.2.12 (Inverse State Transition Tensor). Suppose we are given the analytic solution flow of a nominal initial state x0 for some time interval [t0 , tk ], i.e., xk = (tk ; x0 , t0 ),
so that the inverse solution flow is defined as x0 = (t, x; t0 ). The inverse series mapping
deviations in the state x at time t (t0 t tk ) to deviations in the initial state x0 along the
reference trajectory are:
xi0 =
m
X
1 i,1 p 1
(t0 ,t) x xp ,
p!
p=1
(2.31)
where j {1, , N } and

i,
(t01,t) p (t, x; t0 )
p i (t, ; t0 )
.
1 p j =xj
We call these higher order partials the inverse state transition tensors (ISTTs).
(2.32)
23
The ISTTs serve a role identical to the inverse of the STM and can be computed by
using a similar integration approach as in the STT computation. For example, consider
the state transition matrix = /x0 , or ia = i,a . As shown in Eqn. (2.27), the
differential equations for satisfy:
= g[t, x(t)] ,
(2.33)
and the differential equations for 1 can be stated as:

1 = 1 g[t, x(t)] .
(2.34)
Note that we assume the inverse 1 always exists and is well defined. The first order
STTs are i,a = (1 )ia , and hence, the differential equations for i,a yield:
i,a = i, g,a .
(2.35)
If the STTs have been found already, however, it is more convenient to compute the
ISTTs via series reversion [62]. The inverse series can be computed by substituting Eqn.
(2.22) into Eqn. (2.31) and collecting the terms of same order in x0 . Hence, the ISTTs,
i,
i,
(t01,t) p , are also analytic in the STTs, (t,t10 ) p . After carrying out the series reversion, the
ISTTs mapping from t to t0 up to fourth order are:
i,a =
1 (t, t0 )
ia
i,ab = i, ,j1 j2 j1 ,a j2 ,b ,
(2.36)
(2.37)
i,abc = i, ,j1 j2 j3 + i, ,j1 ,j2 j3 + ,j1 j2 ,j3 + ,j1 j3 ,j2

j1 ,a j2 ,b j3 ,c ,
(2.38)
24
i,abcd = i, ,j1 j2 j3 j4 + i, ,j1 j2 j3 ,j4 + ,j1 j2 j4 ,j3 + ,j1 j3 j4 ,j2

+ ,j1 j2 ,j3 j4 + ,j1 j3 ,j2 j4 + ,j1 j4 ,j2 j3 + ,j1 ,j2 j3 j4
+ i, ,j1 j2 ,j3 ,j4 + ,j1 j3 ,j2 ,j4 + ,j1 j4 ,j2 ,j3

+ ,j1 ,j2 j3 ,j4 + ,j1 ,j2 j4 ,j3 + ,j1 ,j2 ,j3 j4
j1 ,a j2 ,b j3 ,c j4 ,d ,
(2.39)
where all indices are {1, , N }, = (t0 ,t) and = (t,t0 ) are used for concise notations,
and (t, t0 ) in Eqn. (2.36) represents the STM. Note that Eqns. (2.36-2.39) are analytic
in the STTs and require no integration. Hence, given a reference trajectory with STT
solutions for some time interval [t0 , tf ], any arbitrary trajectories in the neighborhood of
the reference solution can be mapped nonlinearly forward from t0 and backward from tf
to any time t [t0 , tf ].
Now, consider the first order case. Given (t, t0 ) for the entire trajectory t [t0 , tf ],
the linear map from any given time tr to ts , where tr , ts [t0 , tf ], can be represented as:
(ts , tr ) = (ts , t0 )1 (tr , t0 ).
(2.40)
Analogously, applying the higher order forward and inverse state transition tensors, the
i,
STTs (ts1,tr ) p which nonlinearly map the deviations from arbitrary time tr to ts (tr , ts
[t0 , tf ] and tr ts ), can be represented as:
i,a
(ts ,tr ) =
ia
,a
(ts , t0 )1 (tr , t0 )
= i,
s r ,
i, ,ab
r,a r,b ,
+ i,
i,ab
s
(ts ,tr ) = s r
(2.41)
(2.42)
,a ,bc
i, ,abc
r r + r,ab r,c + r,ac r,b
+ i,
i,abc
s
(ts ,tr ) = s r
+ i,
r,a r,b r,c ,
s
(2.43)
25
,abc ,d
i, ,abcd
i,abcd
+ i,
r r + r,abd r,c + r,acd r,b + r,ab r,cd
s
(ts ,tr ) = s r
,ab ,c ,d
+ r,ac r,bd + r,ad r,bc + r,a r,bcd + i,
r r r
s
+ r,ac r,b r,d + r,ad r,b r,c + r,a r,bc r,d + r,a r,bd r,c
+ r,a r,b r,cd + i,

r,a r,b r,c r,d ,
s
(2.44)
where all indices are {1, , N } and r = (t0 ,tr ) and s = (ts ,t0 ) for concise notation.
i,
Note that the ISTTs mapping from ts to tr , (tr1,ts ) p , can computed by applying Eqns.
(2.36-2.39).
We have shown that once the STTs are computed for the entire reference trajectory,
any map from an arbitrary point in the relative space to some future time, or vice versa,
i,
becomes a simple algebraic manipulation. Note that (ts1,tr ) p can also be computed by
integrating the differential equations given in Eqns. (2.27-2.30) for the time interval [tr , ts ].
2.2.1 Complexity of the Higher Order Solutions
When computing the state transition tensors by integrating Eqns. (2.27-2.30), one concern is numerical precision if the integration is over a long time period. This problem, however, can be remedied by segmenting the reference trajectory arbitrarily to meet the desired
numerical accuracy, since in most space missions, the reference trajectory is known with
high precision. For example, given a nominal solution flow (tf ; x0 , t0 ) for some time interval [t0 , tf ]. Assuming this nominal trajectory is precisely known, we can choose points
{x0 , x(t1 ), x2 (t2 ), } from (tf ; x0 , t0 ), where t0 tj tf , and integrate the STTs for
time intervals [tj , tj+1 ]. In this way, numerical round-off errors in the higher order STTs
from an integration can be reduced significantly. When computing the ISTTs via series
reversion, it is apparent from Eqns. (2.36-2.39) that a small error in the forward STTs can
result in a significant error int he ISTTs. For example, a small error in can result in an
inaccurate determination of 1 , which yields accumulated errors in the computation of
26
i,
the higher order ISTTs (t01,t) p . In case of a Hamiltonian system, an integration method,
such as a symplectic or variational integrator, which preserves the Hamiltonian structure
can be implemented to compute more precisely.6 Thus, numerical errors in the STTs
and ISTTs can be reduced.
A main concern in computing the STTs is the number of computations as we consider
higher order solutions, especially for a large m. Specifically, assuming a system with a diP
q+1
mension N = 6, the mth order analysis requires integration of = m
equations.
q=0 6
For example, when m = 3, a total of 1554 equations must be integrated simultaneously.
However, the higher order solutions can be computed off-line, and when an orbit is periodic (e.g., an elliptical or halo orbit), these only need to be computed once. In general, if
we consider the symmetry of the higher order partials the number of integrated equations
for the mth order system reduce to:
m
X
N 1+j
= N
.
j
j=0
(2.45)
For example, when m = 3, we need to integrate a total of 504 equations.

An additional burden is the computation of the partials of the dynamics. We note that
there are symbolic manipulators available which provide automatic differentiations, and
also note that many of these partials vanish for spacecraft applications. For the problems
R symbolic toolbox is used for automatic differconsidered in this dissertation, MATLAB
entiation of the dynamics, and once computed, the partials are stored as an m-file. Also,
consider the governing differential equations given in Eqn. (2.2). All the partial derivatives
of v vanish for m 2.
For a Hamiltonian system, the symplectic or the variational integrator preserves that (det ) = 1 for all
times.
27
2.3 Dynamics and Properties of a Hamiltonian System

For spacecraft applications, the general dynamics model Eqn. (2.3) is usually stated
in a Lagrangian frame, i.e., in a position and velocity coordinate frame. However, many
astrodynamics problems can be restated as a Hamiltonian system, which provides several
useful properties that simplify the problem [16, 18]. This section presents these important
properties of a Hamiltonian system and extends the linear Hamiltonian dynamical theory
to the higher order Hamiltonian dynamics.
Definition 2.3.1 (Lagrangian System). The holonomic form of Lagranges equations are
defined as [29, 30, 31]:
d
dt
L
q i
L
= 0,
qi
(2.46)
t) is called the Lagrangian function

for i {1, n}. The scalar function L = L(q, q,
and the qi are called the generalized coordinates. The Lagrangian function L is defined as
L = T + V , where T represents the kinetic energy and V represents the potential energy.
Example 2.3.2. Consider the normalized two-body problem (GM = 1), where the kinetic
energy is defined as T (r) = 21 r r and the potential energy is defined as V (r) = 1/krk.
The Lagrangian function for the two-body problem yields:
L(r, r ) =
1
1
r r +
,
2
krk
and the Lagrangian equations of motion becomes:

r(t) =
1
r,
krk3
which satisfies the general dynamical system Eqn. (2.1).

Definition 2.3.3 (Hamiltonian System). A system is called Hamiltonian if there exists
a smooth scalar function H(q, p, t) such that the governing equations of motion can be
28
stated as [4, 13, 54]:
H
,
pi
H
= i.
q
q i =
(2.47)
p i
(2.48)
The scalar function H, which is time-varying in general, is called a Hamiltonian function,

and qi and pi = L/ q i are called the generalized coordinates and generalized momenta,
respectively.
Given a Lagrangian system Eqn. (2.46) a Hamiltonian function H(q, p, t) can be defined by the Legendre transformation as:
H(q, p, t) =
n
X
t).
pi q i L(q, q,
(2.49)
i=1
After transforming a Lagrangian system into Hamiltonian form, there exist many useful
properties that simplify the problem. The rest of this section presents these properties.7
Definition 2.3.4 (Symplecticity). A 2n 2n matrix M is called symplectic if:
MT JM = J,
where
(2.50)
0nn Inn
J = J2n2n =
.
Inn 0nn
(2.51)
Here, J is called the symplectic unit matrix.

Property 2.3.5. Properties of the symplectic unit matrix:
JT J = JJ = I,
(2.52)
Ji Jj = Ji Jj = ij ,
(2.53)
det(J) = 1.
(2.54)
The details of the definitions, properties, and proofs presented in this section can be found in Refs.
[30, 81].
29
where ij represents the Kronecker delta function.
Property 2.3.6. If a 2n 2n matrix M is symplectic, then:
M1 = JMT J.
(2.55)
Hence the inverse of a symplectic matrix can be computed without a matrix inversion.
Property 2.3.7. The governing equations of motion of a Hamiltonian system can be stated
as:
q
x = = JHxT ,
p
(2.56)
where Hx represents the row-wise partial derivatives of H. In tensor notation:

x i = Ji H ,
(2.57)
where {1, . . . , 2n}. Note that H is an th component of a vector Hx , which is

composed of partial derivatives H/xi .8
Property 2.3.8. Liouvilles theorem states that if a 2n 2n matrix M is symplectic, then
det(M) = 1.
Property 2.3.9. If a 2n 2n matrix M is symplectic, then M1 and MT are symplectic as
well.
Ann Bnn
Property 2.3.10. If M =
then M is symplectic if and only if:
Cnn Dnn
1. ATnn Cnn and BTnn Dnn are symmetric and
2. DTnn Ann BTnn Cnn = Inn .
See the Notation section for details.
30
Definition 2.3.11 (Canonical Transformation). A transformation of phase space (q, p) 7
(Q, P) that preserves the Hamiltonian structure of the dynamical system is called a canonical transformation.
Property 2.3.12. Consider a Hamiltonian system H(q, p, t) and let Q = Q(q, p, t) and
(Q, P)
P = P(q, p, t). Also let MQP =
be symplectic, which yields det(MQP ) = 1, and
(q, p)
thus, there exists a unique inverse. This leads to q = q(Q, P, t) and p = p(Q, P, t), where
(q, p)
Mqp =
is symplectic. Then:
(Q, P)
q
q
Q
+
P
Q
Q
q
p
= Mqp
(2.58)
=
.
p
+ p P
p
P
Q
Q
P
T
Applying M1
qp = JMqp J gives:
q
T
= JMqp
H
P
p
=
,
P
(2.59)
for a Hamiltonian function K(Q, P, t). Therefore, a transformation (q, p) 7 (Q, P) is

(Q, P)
canonical if and only if M =
is a symplectic matrix.
(q, p)
Property 2.3.13. The Hamiltonian system with q(q0 , p0 , t) and p(q0 , p0 , t) is a canonical
transformation between (q0 , p0 ) and (q, p) with t as an independent variable, as the state
(q, p)
transition matrix (t, t0 ) =
is a symplectic matrix.
(q0 , p0 )
Property 2.3.14. Given a Hamiltonian system (q, p, H(q, p, t)), the linearized model becomes:
q0
q
.
= (t, t0 )
p0
p
(2.60)
The state transition matrix is symplectic, and thus, the transformation is canonical, which
yields a Hamiltonian system. Therefore, the linearized dynamics of a Hamiltonian system
are Hamiltonian.
31
Property 2.3.15. Properties of the state transition matrix of a Hamiltonian system:
1 (t, t0 ) = (t0 , t) = JT J,
(2.61)
det[(t, t0 )] = 1,
(2.62)
d
= JHxx .
dt
(2.63)
Hence the inverse of the state transition matrix (the first order ISTTs discussed in 2.2)
can be computed without a matrix inversion. Also, the determinant of the STM (the first
order STTs) can be used to check the numerical accuracy of an integration of the higher
order STTs.
By applying properties of a Hamiltonian system, we can also compute the inverse of
a STM (first order ISTTs) of a Lagrangian system without a matrix inversion. Consider
L(q, q,
t)), and suppose there exists a unique state transition
a Lagrangian system, (q, q,
matrix such that:
q
q0
=
.
q
q 0
(2.64)
Also consider the linearized Hamiltonian system:
q
q0
=
.
p
p0
(2.65)
t) and p = L/ q,
and let T be
Recall the Legendre transformation H = pT q L(q, q,
the Jacobian of the Legendre transformation, i.e.,
(q, p)
I
= 2L
(q, q)
q q
q(t)
q(t)
.
= T(t)
q(t)
p(t)
T =
0
2L ,
q
q
(2.66)
(2.67)
32
By substituting Eqn. (2.67) into Eqn. (2.65), we get:
(t, t0 ) = T1 (t)(t, t0 )T(t0 ),
(2.68)
(t, t0 ) = T(t)(t, t0 )T1 (t0 ),
(2.69)
1 (t, t0 ) = T1 (t0 )JTT (t0 )T (t, t0 )TT (t)JT(t).
(2.70)
which gives:
2L
2L
is symmetric and
= I, according Property 2.3.10, T is symplectic. Apply q
q q
q
ing the symplectic condition for T, Eqn. (2.70) simplifies to:
If
1 (t, t0 ) = JT (t, t0 )J,
(2.71)
which indicates that is symplectic as well.

Therefore, we can take advantage of a Hamiltonian system by computing the inverse
of a STM (1 ) without a matrix inversion. This property is especially useful for the
computation of the ISTTs of a Lagrangian system, Eqn. (2.36-2.39), since the errors from
a matrix inversion can be avoided.
2.4 Symplecticity of the Higher Order Solutions of a Hamiltonian

System
In 2.3, we have shown that the solution flows of a Hamiltonian and a linearized Hamiltonian system are canonical transformations (Properties 2.3.13 and 2.3.14); thus the transformations are symplectic. In this section, we extend the symplecticity of a linearized
Hamiltonian system and show that the higher order relative solutions are also canonical
transformations.
Consider a Hamiltonian system x i = gi = Ji H with an initial state x0 , where x =
[qT , pT ]T and H = H(x, t). Given the initial deviation x0 , the M th order differential
33
equations for the relative dynamics of a Hamiltonian system can be stated as:
M
X
1 i
x (t) =
J H1 p x1 xp ,
p!
p=1
i
(2.72)
where we apply the M th order Taylor series expansion given in Eqn. (2.20) and the higher
order partial derivatives of Hamiltonian dynamics:
i,1 p
(t, x) = J H1 p = J
p H[t, (t)]
.
1 p j =j (t;x0 ,t0 )
(2.73)
To show that an integrated relative solution x(t, x) computed according to Eqn.

(2.72) is symplectic, we can show that there exists a Hamiltonian function K(x, t) such
that the relative equations of motion can be written as:
x i = Ji K ,
M
X
1
K =
H1 p x1 xp .
p!
p=1
(2.74)
(2.75)
For example, when M = 1 the new Hamiltonian can be defined as K(x, t) = 21 Hab xa xb ,
so that the first order relative dynamics can be written as x i (t) = Ji H1 x1 . For higher
order equations, the new Hamiltonian function can be stated explicitly as [33]:
K(q, p, t) =
p
M
X
X
p=2 i1 , ,i2n =0
i1 ++i2n =p
1
p H(q0 , p0 , t)
(x1 )i1 (x2n )i2n .
i1 ! i2n ! (q1 )i1 (qn )in (p1 )in+1 (pn )i2n
(2.76)
Therefore, the relative solution x(t, x0 ) can be written in a Hamiltonian form, and thus,
the solution flow is symplectic. Note that x(t, x0 ) is defined with respect to a nominal
solution of the original Hamiltonian system (x(x0 ), H(x, t)). Hence the relative dynamics
x(t)
must be integrated together with x(t)
so that the higher order partials of H in Eqn.
(2.76) are computed along the nominal solution.
34
We have shown that the solution flow computed according to the differential equations
Eqn. (2.72) is symplectic. This, however, does not necessarily mean that an analytic
relative solution flow with respect to the original Hamiltonian flow is symplectic. To
clarify this point, consider the analytic relative flow Eqn. (2.22):
xi (t) =
i, p
where the STTs (t,t10 )
m
X
1 i,1 p 1
(t,t0 ) x0 x0p ,
p!
p=1
(2.77)
are computed along the nominal solution flow (t; x0 , t0 ) of the
original Hamiltonian system (x, H(x, t)) and x0 is an initial deviation in x0 . As discussed
in 2.2, the computation of the relative solution flow Eqn. (2.77) is fundamentally different than the relative solution integrated according to Eqn. (2.72). In limit (m ),
the analytic map x0 7 x(t), Eqn. (2.77), is a canonical transformation since x(t) is
essentially the solution flow of a Hamiltonian system (x(x0 ), K(x, t)) with m in
1
Eqn. (2.76). Also, when m = 1, the first order STTs i,
(t,t0 ) are symplectic, and thus, the
relative solution flow is canonical. However, for an arbitrary m, the map x0 7 x(t) is
not necessarily guaranteed to be a canonical transformation.
To show this, consider the second order relative solution (m = 2):
1 i,1 2 1 2
1
1
xi (t) = i,
(t,t0 ) x0 + (t,t0 ) x0 x0 ,
2
(2.78)
and suppose that the map x0 7 x(t) is symplectic. Then, the Jacobian of Eqn. (2.78)
must satisfy the symplectic condition Eqn. (2.50):
ij
(x ) (x )
J
= J ,i ,j ,
j
(xi0 )
(x0 )
(2.79)
where i,j represents the Jacobian. For m = 2 case, i,j is defined as:
i,j = i,j + i,1 j x01 .
(2.80)
35
Assuming the symplectic condition is valid, the total time derivative of Eqn. (2.79) must
vanish, so that:
0
ij
d
=
dt
(x )
(xi0 )
(x )
(xj0 )
(x )
(xi0 )
d
dt
(x )
(xj0 )
= ,i J ,j + ,i J ,j .
(2.81)
Considering a Hamiltonian system, i.e., x i = gi = Ji H , the time derivatives of the first

and second order STTs are (Eqns. (2.27) and (2.28)):
i,a = Ji H ,a ,
i,ab = Ji H ,ab + Ji H ,a ,b ,
(2.82)
(2.83)
where the partials of Hamiltonian dynamics gi,1 p = Ji H1 p are substituted:
gi, = Ji H j (t;x0 ,t0 ) ,
gi, = Ji H j (t;x0 ,t0 ) .

Now substitute Eqn. (2.80) into Eqn. (2.81) and collect the terms of the same order in x0 .
The symplectic condition yields:
0ij = ( ,i + ,i x0 )J (,j + ,j x0 ) + (,i + ,i x0 )J ( ,j + ,j x0 )
zeroth order terms
z
}|
{
,i ,j
,i ,j
= J + J
first order terms
z
}|
{
+ ,i J ,j x0 + ,i J ,j x0 + ,j ,i J x0 + ,j ,i J x0
second order terms
}|
{
z
,i ,j
,j ,i
+ J x0 x0 + J x0 x0 ,
(2.84)
where superscripts {, , , } {1, , 2n} are dummy variables. We shall analyze

each order of terms in Eqn. (2.84) to verify the symplectic condition.
36
Zeroth order terms:
The zeroth order terms in Eqn. (2.84) are from the first order relative dynamics and vanish
to zero:
,i J ,j + ,i J ,j = J Hk k,i J ,j + ,i J J Hk k,j ,
= Hk k,i ,j Hk ,i k,j ,
= Hk k,i ,j Hk ,i k,j ,
0ij ,
where the first step substitutes Eqn. (2.82) for time derivatives of the STTs, second and
third steps use the symplectic property Ji Jj = Ji, Jj = ij , and the last step uses the
symmetry of H and the fact that the dummy variables {, , k} can be swapped.
First order terms:
Substituting Eqns. (2.82) and (2.83) for the time derivatives of the STTs and re-ordering
give:
,i J ,j x0 + ,i J ,j x0 + ,j ,i J x0 + ,j ,i J x0
= J H1 1 ,j ,i J x0 + J H1 1 ,i + J H1 2 1 ,i 2 , J ,j x0
+ J H1 1 ,i J ,j x0 + J H1 1 ,j + J H1 2 1 ,j 2 , ,i J x0 ,
where the superscripts {, , 1 , 2 , , , } are dummy variables. Applying the symplectic properties Ji Jj = Ji Jj = ij simplifies to:
H1 1 ,j ,i x0 + H1 1 ,i ,j x0 + H1 2 1 ,i 2 , ,j x0
+ H1 1 ,i ,j x0 H1 1 ,j ,i x0 H1 2 1 ,j 2 , ,i x0 ,
=
H1 ,j 1 ,i + H1 2 1 ,i 2 , ,j + H1 1 ,i ,j
H1 1 ,j ,i H1 ,i 1 ,j H1 2 1 ,j 2 , ,i x0 ,
37
where the dummy variables and are swapped with and , respectively. Apply ii = 1
to get:
H1 ,j 1 ,i + H1 1 ,i ,j + H1 2 1 ,i 2 , ,j
H1 1 ,j ,i H1 ,i 1 ,j H1 2 1 ,j 2 , ,i x0 0ij ,
since the second-rank tensor H1 and the third-rank tensor H1 2 are symmetric. Therefore, the second order terms in Eqn. (2.83) vanish to zero.
Second order terms:
Substitute Eqn. (2.83) for the time derivatives of the STTs in the second order terms of
Eqn. (2.84) and multiply out the terms to get:
,i J ,j x0 x0 + ,j ,i J x0 x0
=
J H1 1 ,i + J H1 2 1 ,i 2 , J ,j x0 x0
+ J H1 1 ,j + J H1 2 1 ,j 2 , ,i J x0 x0 ,
= J J H1 1 ,i ,j x0 x0 + J J H1 2 1 ,i 2 , ,j x0 x0
+ J J H1 1 ,j ,i x0 x0 + J J H1 2 1 ,j 2 , ,i x0 x0 ,
where the superscripts {, , 1 , 2 , , , } are dummy variables. Applying the symplectic property Ji Jj = Ji Jj = ij gives:
H1 1 ,i ,j x0 x0 + H1 2 1 ,i 2 , ,j x0 x0
H1 1 ,j ,i x0 x0 H1 2 1 ,j 2 , ,i x0 x0
=
H1 1 ,i ,j + H1 2 1 ,i 2 , ,j
H1 ,j 1 ,i H1 2 1 ,j 2 , ,i x0 x0 ,
(2.85)
where we apply ii = 1 and swap the dummy variable with . Considering the symmetry
of H1 , Eqn. (2.85) simplify to:
H1 2 1 ,i 2 , ,j H1 2 1 ,j 2 , ,i x0 x0 ,
38
which does not vanish since when i 6= j, despite the symmetry of H1 2 :
H1 2 1 ,i 2 , ,j H1 2 1 ,j 2 , ,i
6= 0ij .
We have shown that the second order analytic relative flow is not a canonical transformation. This result is generally true for any higher order solutions m 2 that are
computed analytically with respect to a Hamiltonian solution flow. However, we have also
shown that the second order solution is symplectic up to first order, O(x). Based on this
observation, we introduce the following conjecture.
Conjecture 2.4.1. An mth order analytic solution of the relative trajectory motion of a
Hamiltonian system is symplectic up to order (m 1).
To show this is true for a general mth order case, we can apply the similar approach
used for the m = 2 case. Recall the sufficiency condition:
,i J ,j + ,i J ,j = 0ij ,
where
i,j
m
X
p=1
i,j1 p1 x01 x0p1 .

(p 1)!
(2.86)
Substituting Eqn. (2.86) into the symplectic sufficiency condition yields:

#
" m
" m
#
X
X
1
1
,i1 p1 x01 x0p1 J

,j1 q1 x01 x0q1
(p
1)!
(q
1)!
p=1
q=1
" m
#
" m
#
X
X
1
1
,i1 p1 x01 x0p1 J

,j1 q1 x01 x0q1 .
+
(p 1)!
(q 1)!
p=1
q=1
Now multiply out the terms and re-order to find:
m X
m
X
p=1 q=1
m X
m
X
p=1 q=1
(p 1)! (q 1)!
J
(p 1)! (q 1)!
,i1 p1 ,j1 q1 x01 x0p1 x01 x0q1
,i1 p1 ,j1 q1 x01 x0p1 x01 x0q1 . (2.87)
39
To show that an mth order solution is symplectic up to order (m1), it suffices to show that
O(xm1
) terms in Eqn. (2.87) vanish. This, however, is intuitively true since including
0
m1
O(xm
) terms. In other words,
0 ) terms in Eqn. (2.87) contribute no additional O(x0
if we consider an (m + 1)th, or higher, order solution, the contributions due to O(xm1

)
0
are the same as from an mth order solution. Moreover, in limit (m ), we know that
the symplectic condition must be satisfied, which indicates that the contributions due to
O(xm1
) terms must vanish at some order of solution. Therefore, the contributions due
0
to O(xm1
) terms must vanish at least for an mth order solution, and thus the Conjecture
0
2.4.1 is true. From a different perspective, for an mth order solution to be symplectic at
order m, it requires (m + 1)th order effects. This is as expected, as the symplecticity
condition is nonlinear. Hence satisfying the condition at one order requires contributions
from the next highest order, at least.
We can also show that Conjecture 2.4.1 is true more explicitly. From Eqn. (2.87) we
find that there are 2m terms that are O(xm1 ):
m terms
z
"
}|
#{
,i
,i
,i
,j
,j
,j
1
m1
2
m1
1
m1
1
J
+
+ +
x01 x0m1
(m 1)!0!
(m 2)!1!
0!(m 1)!
"
#
,i1 m1 ,j
,i2 m1 ,j1
,i ,j1 m1
+ J
+
+ +
x01 x0m1
(m 1)!0!
(m 2)!1!
0!(m 1)!
|
{z
}
m terms
m
X
= J
p=1
+J
if p=m, then ,j
z
}|
{
1
,i1 p1 ,jp m1 x01 x0m1

(m p)!(p 1)!
m
X
p=1
,ip m1 ,j1 p1 x01 x0m1 .

(m p)!(p 1)! | {z }
if p=m, then ,i
(2.88)
40
Recall the time derivative of the STTs from 2.2:
,i1 p1 = g,1 $1 + g,1 2 $2 + + g,1 p $p ,
= J H1 $1 + J H1 2 $2 + + J H1 p $p ,
p
X
J H1 q $q ,
=
(2.89)
q=1
where $j are some functions of the STTs up to order p, i.e., i,1 p . For example:
$1 = $1 (i, 1 , , p1 ) = 1 ,i1 p1 ,
$p = $p (i, 1 , , p1 ) = 1 ,i 2 ,1 p ,p1 .
Note that $j is symmetric in the superscripts of ,i1 p1 excluding . Hence, we define
$j = $j (i, 1 , , p1 ) to denote that its symmetric in (i, 1 , , p1 ). Using Eqn.
(2.91) we have:
,i1 p1 =
p
X
J H1 q $q (i, 1 , , p1 ),
(2.90)
J H1 q $q (j, 1 , , p1 ).
(2.91)
q=1
p
,j1 p1 =
X
q=1
Now substitute Eqns. (2.90) and (2.91) into Eqn. (2.88) to find:
" p
#
m
,jp m1
X
X
J
J H1 q $q (i, 1 , , p1 ) x01 x0m1
(m p)!(p 1)! q=1
p=1
" p
#
m
,ip m1
X
X
+ J
J H1 q $q (j, 1 , , p1 ) x01 x0m1 .
(m
p)!(p
1)!
p=1
q=1
Applying the symplectic identity property, Ji Jj = Ji Jj = ij , gives:
" p
#
m
X
X
1
,jp m1
H1 q $q (i, 1 , , p1 ) x01 x0m1
(m
p)!(p
1)!
q=1
p=1
" p
#
m
X
X
1
,ip m1
H1 q $q (j, 1 , , p1 ) x01 x0m1 .
(m p)!(p 1)!
q=1
p=1
(2.92)
41
Without further simplification, we hypothesize that Eqn. (2.92) must vanish and consider
the m = 3 case as an example to show this is satisfied.
If we let m = 3, Eqn. (2.92) can be stated as:
+
+
1 ,j1 2
[H1 $1 (i)] x01 x02
2!0!
1 ,j2
[H1 $1 (i, 1 ) + H1 2 $2 (i, 1 )] x01 x02

1!1!
1 ,j
[H1 $1 (i, 1 , 2 ) + H1 2 $2 (i, 1 , 2 ) + H1 2 3 $3 (i, 1 , 2 )] x01 x02
0!2!
1 ,i1 2
[H1 $1 (j)] x01 x02

2!0!
1 ,i2
[H1 $1 (j, 1 ) + H1 2 $2 (j, 1 )] x01 x02

1!1!
1 ,i
[H1 $1 (j, 1 , 2 ) + H1 2 $2 (j, 1 , 2 ) + H1 2 3 $3 (j, 1 , 2 )] x01 x02 .
2!0!
(2.93)
Substituting expressions for $j gives:
: a
1 ,j1 2
H1 1 ,i x01 x02
2
+ ,j2 H1 1 ,i1 + H1 2 1 ,i 2 ,1 x01 x02

"
#
b
:
1 ,j
1 2
1 ,i
H
+ H1 2 1 ,i 2 ,1 2 + 1 ,1 2 ,i2 + 1 ,2 2 ,i1 x01 x02
+
2
: c
1 ,j
,i 2 ,1 3 ,2
H 1
x01 x02
+
2 12 3
: b
1 ,i1 2
1 ,j
H1 x01 x02
2
,i2 H1 1 ,j1 + H1 2 1 ,j 2 ,1 x01 x02
: a
1 ,j 2 ,1 2
1 ,i
,j
,j
,j
1
1
2
2
1
2
2
1
1
1
2
x01 x02
H
+ H1 2

2
: c
1 2
1
1
,j
2 ,1 3 ,2
x0 x0 .
,i H
1
2
3
2
42
Now simplify to find:
{}
}|
z
X
1 {2
1 ,i1
1 ,i 2 ,1
X
X
,j2 HX
+
H
x0 x0
1
1 2
X
X
1 ,j
1 ,2 2 ,i1
1 ,i
1 ,1 2 ,i2
2 ,1 2
x01 x02
+ H1 2
1 ,j
2 ,
1 2
,i H1 2
+ 1 ,1 2 ,j2 + 1 ,2 2 ,j1 x01 x02
2
X
X 1 ,j2 + H 1 ,j 2 ,2 x1 x2 .
,i1 HX
1 XX
1 2
0
0
X
{z
}
|
{}
Note that {} terms were the non-vanishing terms from the m = 2 case. Simplifying once
again results in:
1
,j2 H1 2 1 ,i 2 ,1 x01 x02 + ,j H1 2 1 ,1 2 ,i2 + 1 ,2 2 ,i1 x01 x02
2
1
,i1 H1 2 1 ,j 2 ,2 x01 x02 ,i H1 2 1 ,1 2 ,j2 + 1 ,2 2 ,j1 x01 x02
2
1 ,i 2 ,1 ,j2
= H1 2
,i1 1 ,j 2 ,2 x01 x02
+ H1 2 1 ,1 2 ,i2 ,j x01 x02 H1 2 ,i 1 ,1 2 ,j2 x01 x02 ,

= 0ij .
Hence, the m = 3 case is symplectic up to order 2, which satisfies our conjecture.
Although an mth order solution is only symplectic up to O(xm1 ), we can still achieve
an approximate symplectic condition. That is, given a sufficiently high order analytic map
x(t) 7 x0 computed relative to a Hamiltonian system (x(x0 ), H(x, t)), an accurate
approximation of the true nonlinear relative motion can be made. Therefore, we can say
that the analytic flow of a local relative solution is almost symplectic. This indicates
that, when an initial deviation is within the radius of convergence and m is large enough,
symplectic properties presented in 2.3 can be applied.
43
2.5 Convergence of the Higher Order Solutions

Given the Taylor series expansion of a solution Eqn. (2.22), it is non-trivial to determine what order of solution suffices to represent the local nonlinear motion for a given
set of initial deviations. If the initial conditions lie outside the region of convergence,
the higher order solutions diverge. In this dissertation, we assume the solution is within
the radius of convergence. Moreover, the convergence rate of the higher order solutions
vary along the trajectory, which directly depends on the fact that the systems nonlinearity
varies along the trajectory. This is an interesting fact since it indicates that different orders of solution can provide different levels of approximation of the relative motion of the
neighboring trajectory. In this section, we discuss a systematic way to find the necessary
order of Taylor series that captures the local nonlinear dynamics.
Consider an initial phase volume B0 defined locally with respect to a nominal initial
state 0 . We can propagate B0 by nonlinearly mapping all points x0 B0 , but since
B0 is defined locally, we can also propagate each point relative to the nominal trajectory
(t; 0 , t0 ). Suppose we want to propagate B0 linearly to some future time. One important
issue in linear propagation is the significance of a coordinate system. For example, in the
two-body problem, the orbit elements are constant (except for the mean anomaly) whereas
the Cartesian coordinates are not. Hence, depending on the choice of a coordinate system,
different linear propagations can give different levels of approximation of the nonlinear
motion. This problem has been studied by Junkins et al. [47, 48], where the level of
nonlinearity (or linearity) of a trajectory is quantified using the nonlinearity index.
Definition 2.5.1 (Nonlinearity Index). Given an initial phase volume B0 with respect to
a nominal initial state x0 , the nonlinearity index measures the level of nonlinearity (or
44
linearity) of the propagated phase volume. The nonlinearity index is defined as:
(t, t0 ) ,
sup
=1, ,
k (t, t0 ) (t, t0 )kf

,
k(t, t0 )kf
(2.94)
where is the number of initial sample points, k kf represents the Frobenius norm,
(t, t0 ) is the state transition matrix computed along the reference trajectory (t; x0 , t0 ),
and (t, t0 ) represents the state transition matrix computed along the th sample trajectory (t; x0 , t0 ), where x0 represents the th sample point chosen from the boundary of
B0 . 9
The nonlinearity index computes the level of maximum linear deviation from the reference trajectory. However, our focus is more on deciding the sufficient order of the higher
order solution. For this reason, we apply a slightly different approach where we propagate
a relative trajectory instead of a state transition matrix.
Definition 2.5.2 (Local Nonlinearity Index). Given an initial phase volume B0 with respect
to a nominal initial state x0 , define the local nonlinearity index as:
m (t, t0 ) ,
sup
i=1, ,N
=1, ,
|(xi )m (t; x0 , t0 ) xi (t; x0 , t0 )|

,
|xi (t; x0 , t0 )|
(2.95)
where | | represents an absolute value, (xi )m represents the ith component of the mth order analytic solution Eqn. (2.22) computed with respect to the nominal trajectory (t; x0 , t0 ),
xi represents the ith component of the nonlinearly integrated solution vector according
to the governing equations of motion Eqn. (2.12), and x0 represents the th sample state
vector chosen from the boundary of the initial phase volume. Note that the superscripts
i {1, , N } and {1, , }, where N is the system dimension and is the
number of initial sample points that characterize B0 .10
9
For space mission analysis, the initial phase volume B0 can be considered as an error ellipsoid, so that
the sample points can be chosen from the worst-case initial conditions (e.g., boundary points of the 3-
ellipsoid).
10
If we have a single deviated state x0 with respect to x0 , the local nonlinearity index can be computed
by considering x0 as the only sample point. In other words, x0 can be considered as B0 .
45
In other words, we find the state that deviates the most from its reference value over all
initial samples and each component of the state vector. Hence, this is in a sense a MonteCarlo simulation which compares the higher order solution and the integrated solution,
which is a one time operation for the given reference trajectory and initial set B0 . The
computed value of m then tells how well the mth order solution can approximate the true
nonlinear motion. Using Eqn. (2.22), m can be restated as:
P
p
m
1 i,1 p 1
i
x
(t;
x
,
t
)
0
0
0 0
p=1 p!
m
(t, t0 ) =
sup
. (2.96)
|xi (t; x0 , t0 )|
i=1, ,N
k=1, ,
As we consider higher order Taylor series, and if the series converge, m will converge
to zero, and by increasing the order of solution we can compute the percent difference
between the true and STT solutions. Assuming the initial sample points are within the
radius of convergence, m 0 as m . As a result, a higher order solution needs to
be considered if m > , where stipulates the precision of approximation desired.
CHAPTER III
EVOLUTION OF PROBABILITY DENSITY
FUNCTIONS IN NONLINEAR DYNAMICAL
SYSTEMS
The Gaussian distribution is widely used for astrodynamics applications due to its simplicity and its invariance under linear operations. When we consider mapping a Gaussian
random vector under nonlinear orbital dynamics, however, the Gaussian structure is no
longer preserved, which is an important issue, but usually not considered in conventional
trajectory navigation. In this chapter, we start from the basics of probability theory and
discuss the integral invariance of the probability function of a Hamiltonian dynamical system via solutions of the Fokker-Planck equation. We then apply the higher order solutions
from Chapter II and present an exact analytic representation of a non-Gaussian probability density, and an approximation of this distribution using the first few moments of
the original Gaussian distribution. Implicit in our discussion is that a well-defined initial
probability density function for our dynamical state exists. This is a standard and usual
assumption for spacecraft navigation and is uncontroversial.
3.1 Review of Probability Theory and Random Processes

In this section, we give a few definitions from probability theory that are used throughout this thesis. The formal definitions and discussion can be found in Refs. [32, 51, 82, 88].
46
47
Definition 3.1.1 (Probability Density Function). Given a continuous random vector x
RN , the probability of x in some volume B can be computed by:
Z
Z
Pr(x B) =
p()d =
p()d 1 d 2 d N ,
B
(3.1)
where a function p(x) is called a probability density function (PDF).1 Note that the integral
is over d = d 1 d 2 d N .
Remark 3.1.2. Since Pr(x B) 0 for all B, the PDF p(x) must be nonnegative for all x.
Remark 3.1.3. Given a PDF of a random vector x RN , the PDF must satisfy:
Z
p()d = 1.
(3.2)
Remark 3.1.4. Given a constant RN , Pr(x = ) = 0 for all x.

Definition 3.1.5 (Marginal Density Function). Given two random variables x R and
y R, the marginal density function of x is defined as:
Z
p(x) =
p(x, y)dy.
(3.3)
Definition 3.1.6 (Law of Total Expectation). For a continuous random vector x RN

with density p(x), the expected value of an arbitrary function g(x) can be computed by:
Z
E[g(x)] =
g()p()d,
(3.4)
where E[] represents the expectation operator.

Definition 3.1.7. Given a random vector x RN with a PDF p(x), the mean and covariance matrix are defined as:
Z
m = E[x] =
p()d,
(3.5)
Z
( m)( m)T p()d,
P = E[(x m)(x m) ] =
(3.6)
Formally, the PDF should be written as pX (x, t), indicating that pX is a probability density function of a
random process X, where x X. In this thesis, the subscript is dropped for concise notation and p(x, t) is
used to represent pX (x, t).
48
or in tensor notation:
Z
m
i p()d,
= E[x ] =
(3.7)
Pij = E[(xi mi )(xj mj )] = E[xi xj ] mi mj ,

Z
=
i j p()d mi mj .
(3.8)
Remark 3.1.8. In general, an mth order moment is defined as:

E[x1 xm ],
(3.9)
and an mth order central moment is defined as:

E[(x1 m1 ) (xm mm )].
(3.10)
Definition 3.1.9 (Characteristic Function). The joint characteristic function (JCF) of a

continuous random vector x RN is defined as:
T
(u) = E[eju x ],
where j =
(3.11)
1. The higher moments can be computed by:

1
E[x ] = j
E[x1 x2 ] = j 2
E[x1 x2 x3 ] = j 3
..
.
1 2
E[x x x ] = j
(u)
,
u1 u=0
2 (u)
,
u1 u2 u=0
3 (u)
,
u1 u2 u3 u=0
m (u)
.
u1 u2 um u=0
(3.12)
(3.13)
(3.14)
(3.15)
Remark 3.1.10. A joint characteristic function is related to a probability density function

by:
Z
T
eju x p(x)dx,
(u) =
(3.16)
49
and the probability density function can be recovered using the Fourier transform:
Z
1
T
p(x) =
eju x (u)du,
(3.17)
N
(2)
where N is the dimension of x.
Definition 3.1.11 (Conditional Density Function). Given two random variables x R and
y R, the conditional density function of y given x is defined as:
p(y|x) =
p(x, y)
,
p(x)
(3.18)
for x with p(x) > 0.

Definition 3.1.12 (Transition Density Function). Given a random process x(t) RN , the
transition density function is defined as:
p(xk , tk |x0 , t0 ) = p(xk |x0 ) = p[x(tk )|x(t0 ) = x0 ],
(3.19)
which represents a PDF of x at time tk given x(t0 ) = x0 .

Definition 3.1.13 (Markov Process). A random process x(t) RN is a Markov process if
for all 0 tk1 tk :
p(xk |xk1 , , x1 , x0 ) = p(xk |xk1 ).
(3.20)
3.2 The Gaussian Probability Distribution

Definition 3.2.1 (Gaussian Probability Density Function). Let x be a Gaussian random
vector, x N (m, P), where m is the mean vector and P is the covariance matrix. The
Gaussian probability density function for x is defined as:
1
1
T 1
exp (x m) P (x m) ,
p(x) = p
2
(2)N det P
(3.21)
where N is the dimension of the state. The Gaussian probability density function for the
case N = 2 is shown in Figure 3.1 in non-dimensional units.
50
0.15
p(x,y)
0.1
0.05
0
3
2
1
0
1
2
3
Figure 3.1: Normalized two-dimensional

Gaussian probability density function: p(x,y) =
1
2
2
(1/2) exp 2 (x + y )
An important property of the Gaussian distribution is that the statistics of the Gaussian
random vector x can be completely described by the first two moments, i.e., m and P
[20, 57, 93]. In other words, the higher moments of x, such as E[xi xj xk ] and E[xi xj xk xl ],
can all be computed as functions of m and P. In this thesis, we implement the JCF defined
in Definition 3.1.9 for higher moment computation.
Definition 3.2.2 (Gaussian Joint Characteristic Function). For a Gaussian random vector
the joint characteristic function is defined as:
1 T
T
(u) = exp ju m u Pu .
2
(3.22)
For a nonzero mean Gaussian random vector, x N (m, P), using Eqn. (3.15), the
51
first four moments can be computed by:
E[xi ] = mi ,
(3.23)
E[xi xj ] = mi mj + Pij ,
(3.24)
E[xi xj xk ] = mi mj mk + mi Pjk + mj Pik + mk Pij ,
(3.25)
E[xi xj xk xl ] = mi mj mk ml
+ mi mj Pkl + mi mk Pjl + mj mk Pil + mi ml Pjk + mj ml Pik + mk ml Pij

+ Pij Pkl + Pik Pjl + Pil Pjk .
(3.26)
Remark 3.2.3. The mth order central moments of x N (m, P) become [96]:
1. If m is odd, E[(x1 m1 ) (xm mm )] = 0,
2. If m is even, where m = 2k (k 1),
E[(x1 m1 ) (xm mm )] =
P1 2 P2k1 2k ,
(3.27)
perm
where the sum is computed over all permutations of {1, , 2k} and yields
(2k 1)!
,
1)!
2k1 (k
(3.28)
terms in the sum.

In conventional trajectory navigation, a spacecraft state uncertainty is usually modeled
using a Gaussian distribution. Assuming a spacecraft state x N (m, P) R6 , we can
construct a confidence region (or a phase volume) within which the spacecraft is located
with a certain probability. The usual model is:
EM = {x | (x m)T P1 (x m) M 2 },
(3.29)
where M R and EM represents a 6-dimensional ellipsoid, and depending on the value

of M , the confidence region EM is called an M - error ellipsoid. If we consider a scalar
52
state x R N (m, 2 ):2
Pr(x E1 ) = 0.6827,
Pr(x E2 ) = 0.9545,
Pr(x E3 ) = 0.9973.
It is important to note that the probability of a spacecraft in an M - error ellipsoid depends

on the dimension of a state. For example, if x R3 N (m, P):3
Pr(x E1 ) = 0.3182,
Pr(x E2 ) = 0.8696,
Pr(x E3 ) = 0.9919.
In general, as the dimension of x increases, Pr(x EM ) decreases.
3.3 Dynamics of the Mean and Covariance Matrix

To study how an initial probability ellipsoid evolves dynamically we can solve for the
mean and covariance matrix as functions of time. Consider the local trajectory dynamics,
i.e., x = (t; x0 + x0 , t0 ) (t; x0 , t0 ), and its mean, covariance, and higher moments.
Differentiate the relative mean and covariance matrix to find:
m i = E[ x i ],
(3.30)
ij
P
= E[ x i xj + xi x j ] m i mj + mi m j .
(3.31)
Substituting the M th order relative dynamics from Eqn. (2.20), we find:

"M
#
M
X 1
X
1 i,1 p
i
i,1 p 1
p
m = E
g
x x
=
g
E [x1 xp ] , (3.32)
p!
p!
p=1
p=1
ij
P
=
M
M
X
X
1 j,1 p 1
1 i,1 p 1
E g
x xp xj +
g
E x xp xi
p!
p!
p=1
p=1
i j
m m + mi m j .
(3.33)
Probabilities are computed by integrating the normalized scalar Gaussian density function, i.e.,
2
1 ex /2 dx.
M
2
3
Probabilities are computed by integrating the normalized 3-dimensional Gaussian density function, i.e.,
RM RM RM
2
2
2
1
e(x +y +z )/2 dxdydz.
M M M (2)3/2
RM
53
We observe from Eqns. (3.32) and (3.33) that the mean and covariances of the relative
motion cannot be computed unless we apply certain assumptions, such as linear dynamics
and Gaussian distribution. This is because, at each step of the integration, higher order expectations in Eqns. (3.32) and (3.33) are required, which are not directly available. These
higher order expectations can be computed, if we integrate the higher order moments, or
the probability density function, together with Eqns. (3.32) and (3.33); however, this is
not considered in this thesis. Also, we can incorporate the second order terms to derive the
prediction equations as given in the truncated second order filter or the Gaussian second
order filter;[6, 55] however, we want to generalize the propagation method by including
the higher order effects in order to compute more accurate orbit uncertainties. In 3.9,
we present a nonlinear analytic method to propagate the mean and covariance matrix by
mapping the initial statistics using the state transition tensors discussed in 2.2.
3.4 The Fokker-Planck Equation

In this section, we give a formal definition of the Fokker-Planck equation (also known
as the forward Kolmogorov equation). The Fokker-Planck equation (FPE) is a partial differential equation that satisfies the propagation of a transition probability density function,
or a probability density function. Hence, the solution of the FPE gives a complete statistical description of a trajectory.
Definition 3.4.1 (It Stochastic Equation). Most orbital dynamics problems can be written
using the It stochastic differential equation:
dx(t) = g[x(t), t]dt + G[x(t), t]d(t),
(3.34)
which is an extension of the general dynamical system in Eqn. (2.3), where G is an N by-s matrix characterizing the diffusion, and is an s-dimensional Brownian motion or
54
Wiener process with zero mean and diffusion matrix Q, i.e.,
E[d(t)dT (t)] = Q(t)dt,
Z t2
T
E [(t2 ) (t1 )][(t2 ) (t1 )]
=
Q(t)dt.
(3.35)
(3.36)
t1
For example, the diffusion vector can be modeled as stochastic acceleration or process
noise in case of a spacecraft orbit determination process. Note that systems with deterministic inputs (e.g., controls) can be simply rewritten by including them in the state vector.
Remark 3.4.2. The solution to the It stochastic differential equation is:
Z t
Z t
x(t) = x0 +
g[x( ), ]d +
G[x( ), ]d( ).
t0
(3.37)
t0
Definition 3.4.3 (The Fokker-Planck Equation for the Transition Probability Density Function). Let p(x, t|x0 , t0 ) be the transition probability density function of the stochastic process x(t). Given a system satisfying the It stochastic differential equation, the time evolution of the transition probability density function must satisfy the Fokker-Planck equation
[26, 55, 87]:
N
X

p(x, t|x0 , t0 )
=
p(x, t|x0 , t0 )gi (x, t)
i
t
x
i=1
N
N
ij o
1 X X 2 n
T
p(x,
t|x
,
t
)
G(x,
t)Q(t)G
(x,
t)
,
0 0
2 i=1 j=1 xi xj
(3.38)
where the initial condition is simply p(x, t0 |x0 , t0 ) = (x x0 ), (Dirac delta function).
From the FPE for the transition density function, we can also derive the FPE for the
probability density function. To show this, consider x in Eqn. (3.38) as a dummy variable.
Then we can view p(x, t|x0 , t0 ) as a random variable as a function of x0 . Computing the
expectation of the transition probability density gives:
Z
p(x, t) =
p(x, t|x0 , t0 )p(x0 , t0 )dx0 ,
(3.39)
55
where p(x0 , t0 ) is now an initial condition. Multiplying Eqn. (3.38) by p(x0 , t0 ) and integrating over x0 gives:
The left-hand side of Eqn. (3.38):
Z
Z
p(x, t|x0 , t0 )
p(x0 , t0 )dx0 =
p(x, t|x0 , t0 )p(x0 , t0 )dx0 ,
t
t
p(x, t)
=
.
t
The first term in the right-hand side of Eqn. (3.38):

Z
N
X
p(x,
t|x
,
t
)g
(x,
t)
p(x0 , t0 )dx0
0
0
xi
i=1
Z
N
X
i
[p(x, t|x0 , t0 )p(x0 , t0 )] dx0 g (x, t) ,
=
xi
i=1
N
X

=
p(x, t)gi (x, t) .
i
x
i=1
The second term in the right-hand side of Eqn. (3.38):

Z
N
N
ij o
1 X X 2 n
T
p(x,
t|x
,
t
)
G(x,
t)Q(t)G
(x,
t)
p(x0 , t0 )dx0
0 0
2 i=1 j=1 xi xj
Z
N
N
ij
1 X X 2
T
=
p(x, t|x0 , t0 )p(x0 , t0 )dx0 G(x, t)Q(t)G (x, t)
,
2 i=1 j=1 xi xj
N
N
ij o
1 X X 2 n
T
p(x,
t)
G(x,
t)Q(t)G
(x,
t)
.
=
2 i=1 j=1 xi xj
Thus, the probability density function satisfies the Fokker-Planck equation.

Definition 3.4.4 (The Fokker-Planck Equation for the Probability Density Function). The
probability density function must satisfy the Fokker-Planck equation:
N
X
p(x, t)

i
=
p(x,
t)g
(x,
t)
t
xi
i=1
N
N
ij o
1 X X 2 n
T
+
p(x, t) G(x, t)Q(t)G (x, t)
.
2 i=1 j=1 xi xj
(3.40)
56
Remark 3.4.5. The FPE for the systems with no stochastic terms can be simplified by
setting (t) = 0:
N
X
p(x, t)

i
=
p(x,
t)g
(x,
t)
.
i
t
x
i=1
(3.41)
Note that solutions of the FPE give the true evolution of the probability density function. However, including these partial differential equations in the trajectory navigation
problem introduces additional difficulties and is usually avoided for practical reasons. In
this thesis, we consider systems with no process noise and the probability density function
satisfies the Fokker-Planck equation defined in Eqn. (3.41). This is a reasonable model for
astrodynamics problems with no thrusters and no dissipative forces acting on the spacecraft.4
3.5 Integral Invariance of Probability

Definition 3.5.1 (Integral Invariance). Consider a dynamical system with the governing
equations of motion x = g(x, t) and let I(x, t) be an integral of a vector field M (x, t) over
some volume B:
Z
I(x, t) =
M (x, t)dx.
(3.42)
The integral I(x, t) is called an integral invariant if it is constant for all time, i.e., dI/dt =
0. In general, the sufficiency condition for integral invariance can be explicitly stated as
[30]:
N
X
M (x, t)

=
M (x, t)gi (x, t) ,
i
t
x
i=1
(3.43)
which is known as Liouvilles equation.

4
For interplanetary spacecraft motion such as Sun-Earth halo orbits or cruise from Earth to Mars, effects
attributable to stochastic acceleration can be reduced to 1012 km/s2 or less, negligibly small compared to
conservative forces acting on a spacecraft.
57
Hamiltonian
Flow
Final
Phase
Volume
Initial
Phase
Volume
Figure 3.2: Integral invariance of a phase volume.
For example, the phase volume in a Hamiltonian system,

Z
B =
dx1 dx2 dx2n ,
dx =
B
(3.44)
is an integral invariant according to Liouvilles theorem and is illustrated in Figure 3.2.

By comparing Eqn. (3.41) with Eqn. (3.43), we see that p(x, t) satisfies the sufficiency
condition for the probability to be an integral invariant. Hence, this implies that probability
of any dynamical system with no diffusion term is an integral invariant.
Definition 3.5.2 (Time Invariance). Consider a dynamical system with the governing
equations of motion x = g(x, t). A scalar function f (x, t) is called a time invariant if
its total time derivative is zero for all t:
df (x, t)
= 0.
dt
(3.45)
Remark 3.5.3. If an integral I(x, t) is an integral invariant, it satisfies the time invariance
condition as well.
58
Suppose we are given a nominal initial state x0 with a PDF p(x0 , t0 ). From the fundamental theorem of calculus and integral invariance of the probability we have:
Z
Pr(x B) =
p(x, t)dx,
Z
=
Z
(3.46)
(t; x0 , t0 )
p[(t; x0 , t0 ), t] det
dx0 ,
x0
B0
p(x0 , t0 )dx0 ,
(3.47)
(3.48)
B0
where x(t) = (t; x0 , t0 ) is the solution flow defined in 2.1. By equating Eqns. (3.47)
and (3.48) we find:
Z
B0
x
p(x0 , t0 ) dx0 = 0.
p[(t; x0 , t0 ), t] det
x0
(3.49)
Hence, the probability density functions at t0 and at t are related by [19, 93]:
(t;
x
,
t
)
0
0
,
p[(t; x0 , t0 ), t] = p(x0 , t0 ) det
x0
(3.50)
where the value of | det(x/x0 )| depends on the system dynamics.
Theorem 3.5.4. Given a system x(t)

= g[x(t), t] with x(t0 ) = x0 and an associated
probability density function p(x0 , t0 ), the probability density function p(x, t) satisfies the
time invariant condition:
p(x, t) = p(x0 , t0 ),
(3.51)
Z t
exp
x g(x, )d
= 1,
(3.52)
if and only if:
t0
where x denotes the gradient vector:
x =
assuming a system with dimension N .
,
x1
xN
(3.53)
59
Proof. Under standard theories of ordinary differential equations, assume the solution flow
x = (t; x0 , t0 ) is continuous and has continuous partial derivatives with respect to x0 ,
and let x0 = (t, x; t0 ) be the inverse solution flow. The current state PDF p(x, t) can be
related to the initial state PDF by applying Eqn. (3.50):
p(x, t) = p[(t, x; t0 ), t0 ] |det T| ,
(3.54)
where
T =
(t, x; t0 )
x
(t; x0 , t0 )
x0
1
.
(3.55)
To show the sufficiency condition, suppose the PDF satisfies the time invariance so that
| det T| = 1 and consider the Fokker-Planck equation for the PDF of a diffusion-less
system:
N
X
p(x, t)

i
=
p(x,
t)g
(x,
t)
.
i
t
x
i=1
(3.56)
Re-ordering Eqn. (3.56) gives:

N
X gi (x, t) X
p(x, t)
p(x, t)
i
+ p(x, t)
+
g
(x,
t)
= 0,
i
i
t
x
x
i=1
i=1
(3.57)
X
p(x, t)
p(x, t)
+ p(x, t)x g(x, t) +
gi (x, t)
= 0.
t
xi
i=1
(3.58)
The Lagrange system of Eqn. (3.58) can be stated as [30, 93]:

dp(x, t)
dx1
dx2
dxN
dt
=
= 1 = 2 = = N ,
1
p(x, t)x g(x, t)
g
g
g
(3.59)
where we assume a dimension N . The first equality provides an obvious relation:

dp(x, t)
= p(x, t)x g(x, t).
dt
(3.60)
Note that this can also be derived directly by substituting the Fokker-Planck equation into
60
the total time derivative of a PDF:
dp(x, t)
p
=
dt
x
p
=
x
x p
+ ,
t
t
g(x, t)
p(x, t)x g(x, t) +
N
X
i=1
p(x, t)
gi (x, t)
xi
!
,
p
p
=
g(x, t) p(x, t)x g(x, t)
g(x, t),
x
x
= p(x, t)x g(x, t).
The solution to Eqn. (3.60) is:
Z t
p(x, t) = p(x0 , t0 ) exp

x g(x, )d
t0
(3.61)
xi0 = i (x,t;t0 )
By comparing Eqns. (3.54) and (3.61), we have:

Z t
|det T| = exp
x g(x, )d ,
(3.62)
t0
which gives:
Z t
exp
x g(x, )d
= +1.
(3.63)
t0
Also, the necessity condition is trivially satisfied since, if Eqn. (3.63) is true, we get
|det T| = 1, and thus, the PDF is a time invariant. Note that the PDF of any dynamical
system is not necessarily a time invariant (constant), although the probability satisfies the
integral invariance condition.
3.6 Solution of the Fokker-Planck Equation for a Hamiltonian System

Consider a PDF of a Hamiltonian system, H(q, p, t) with dimension N = 2n, and re
call the dynamics equations x(t)
= JHxT . Assuming no diffusion in the dynamics (e.g., no
stochastic accelerations), the Fokker-Planck equation for the PDF, Eqn.(3.41), simplifies
61
to:
T
p(x, t)
p(x, t)
x + p(x, t)Trace JHxx .
=
t
x
(3.64)
The second term in the right-hand side vanishes since:
Hpp
T
Hpq
Trace JHxx
= Trace
,
Hqq Hqp
n
! n
!
X
X
=
Hpi qi
Hqi pi ,
i=1
i=1
= 0.
(3.65)
and reduces the FPE to:

dp(x, t)
p p
=
+ x = 0.
dt
t x
(3.66)
Hence the solution of the FPE is:

p(x, t) = p[(t; x0 , t0 ), t] = p(x0 , t0 ),
(3.67)
where p(x0 , t0 ) is assumed to be specified.

We can also show this directly from Eqn. (3.50). For a Hamiltonian system the mapping from x0 to x, (t; x0 , t0 ), is a canonical transformation, and thus, det(x/x0 ) = +1
for all t according to Liouvilles theorem. Therefore, for a Hamiltonian system, the PDF
must satisfy:
p[(t; x0 , t0 ), t] = p(x0 , t0 ),
(3.68)
or we can also state this in terms of the inverse flow x0 = (t, x; t0 ) as:
p(x, t) = p[(t, x; t0 ), t0 ].
(3.69)
In other words, the current state PDF can be completely characterized by the initial PDF,
or vice versa. This means that if the solution is known as a function of initial conditions
62
(i.e., is integrated) and the PDF is known at any one time, it can be found for all time. This
derivation is of particular interest since, as we will see later, it provides a direct way to
evolve the current state statistics as functions of the initial state and its statistics.
This is an important result since it shows that not only is the probability a time invariant, but the probability density function also satisfies the time invariance condition.
This result is not true for just any dynamical system, but it is generally true for astrodynamics problems represented in Lagrangian form, such as two-body problem, Hill threebody problem, and circular restricted three-body problem, as they can be transformed into
Hamiltonian systems. This time invariance of a PDF, in conjunction with Liouvilles theorem, provides another proof that the probability over some volume B in a Hamiltonian
system is indeed an integral invariant.
Now consider a special case of practical interest. Suppose our system is Hamiltonian
and the initial state is Gaussian with mean m0 = m(t0 ) and covariance matrix P0 = P(t0 ),
which are constants, so that:
1
T
p(x0 , t0 ) = p
exp (x0 m0 ) 0 (x0 m0 ) ,
2
(2)2n |P0 |
1
(3.70)
where 0 = P(t0 )1 . Using Eqn. (3.69), the current state PDF can be stated as:
1
T
p(x, t) = p
exp [(t, x; t0 ) m0 ] 0 [(t, x; t0 ) m0 ] .
2
(2)2n |P0 |
1
(3.71)
The PDF Eqn. (3.71) is a valid probability density function since it satisfies:
Z
Z
p(x, t)dx =
p[(t, x; t0 ), t]dx,
dx0 ,
=
p(x0 , t0 ) det
x0
= 1,
63
where we apply the symplectic property of a Hamiltonian system (| det(x/x0 )| = 1).
Also, the PDF is time invariant as it satisfies:
dp(x, t)
=
dt
dp[(t, x; t0 ), t0 ]
p(, t0 ) d(t, x; t0 )
=
= 0 ,
dt
dt
(3.72)
since (t, x; t0 ) = x0 are integrals of motion of our system, and thus, their total time
derivative is zero. A formal proof of the time invariance of a PDF for a linear Hamiltonian
system is presented in Appendix B.
An interesting observation that can be made from Eqn. (3.71) is that the maximum
value of the PDF (i.e., the mode) is always located at the propagated initial mean according
to the deterministic map (i.e., x = (t; m0 , t0 )). However, the solution flow is nonlinear
in general, and thus, the current state mean vector may no longer be located at the mode.
It is apparent from Eqn. (3.71) that once the solution flow (t; x0 , t0 ) is represented by
higher order terms and an analytic expression can be obtained as a function of the initial
state, statistical moments of the current state can be obtained that are, by definition, more
accurate than the propagated statistics from linear theory. Obtaining an analytic framework
for statistics propagation is discussed in 3.9.
3.7 On the Relation of Phase Volume and Probability

Suppose we are given an initial confidence region B0 for the initial spacecraft state
x0 N (m0 , P0 ), and without loss of generality, consider B0 to be a 1- error ellipsoid E1 :
B0 =
x0 | (x0 m0 )T P1
0 (x0 m0 ) 1 .
(3.73)
As the probability of a deterministic system is an integral invariant:

Pr(x0 B0 ) = Pr[x(t) B(t)],
(3.74)
where B(t) = {x(t) | x(t) = (t; x0 , t0 ) x0 B0 }. This indicates that the propagated
phase volume is no longer ellipsoidal since the system dynamics are nonlinear in general.
64
Thus, this implies that the confidence region at some future time can be defined by nonlinearly mapping the boundary of the initial error ellipsoid, B(t0 ). This mapped confidence
region retains its probability value but loses its relation to the moments of the statistical
distribution.
If we wish to use the confidence region to characterize the future distribution, this approach is computationally intensive since we must integrate many samples chosen from
the boundary of B(t0 ) according to the governing equations of motion. For this reason,
in practice, one usually works with the simple linear model with its penalty for ignoring
the higher order statistics. The linear method works well as long as the linear model sufficiently approximates the true dynamics as the propagated phase volume and the statistics
are both available and are equivalent. That is, the propagated N - ellipsoid of p(x0 , t0 ) is
the same as the N - ellipsoid of p(x, t). However, this is no longer true when the system is
subject to a highly unstable environment or when mapped over a long duration of time, as
the linear solution will fail then to characterize the true dynamics. For this reason we implement the STT approach to allow us to analytically characterize both phase volume and
statistics. In other words, once we have the time solution of the STTs, computing the phase
volume incorporating the higher order effects becomes a simple algebraic manipulation,
which provides a more accurate solution than the linear case.
Considering the uniqueness of solutions, the boundary of an initial phase volume must
map to the outer boundary points of the phase volume computed at a later time. Hence we
only need to analyze the behavior of the surface of this N -dimensional object, the surface
being an (N 1)-dimensional object. After the boundary of B0 is integrated to some
time t, it can be projected onto the position and velocity planes to compute the confidence
region B(t), where Pr[x(t) B(t)] is the same as Pr(x0 B0 ).
65
3.8 Time Invariance of the Probability Density Function of the Higher

Order Hamiltonian Systems
In 3.6, we showed that a PDF of a Hamiltonian system is constant for all times.
Hence, a PDF of the higher order Hamiltonian system must satisfy the time invariance
condition as well. In this section, we give a detailed derivation of this property by considering the symplectic structure of the higher order Hamiltonian system.
Consider the relative dynamics of a Hamiltonian system:
M
X
1 i,1 p 1
x (t) =
g
x xp ,
p!
p=1
i
(3.75)
M
X
1 i
J H1 p x1 xp ,
=
p!
p=1
(3.76)
= ci (x, t).
(3.77)
To show the integral invariance of a PDF p(x, t) of the relative motion x(t), it suffices
to check if Theorem 3.5.4 is satisfied:
Z t
exp
x c(x, )d
= 1.
(3.78)
t0
Lets consider the integrand of the Eqn.(3.78) and substitute Eqn. (3.76) to get:
PM
1 1
J H1 p x1 xp
p!
..
.
p=1
x c(x, t) = x
P
1 2n
M
J H1 p x1 xp
p=1
p!
(3.79)
where the indices are {1, 2, , 2n}. We need to show that for every order M Eqn. (3.79)
vanishes. Without loss of generality, consider the case where p = q and carry out the dot
66
product to get:
#
" 1
J H1 q x1 xq
J2n H1 q x1 xq
1
+ +
q!
(x1 )
(x2n )

1 1
=
q J H11 q1 x1 xq1 +
q!
+ q J2n H2n1 q1 x1 xq1 ,

!
X
2n
q
=
Ji Hi1 q1 x1 xq1 .
(3.80)
q!
i=1
Note that Hi1 q1 is symmetric and Jab = Jba and Jaa = 0. Hence, Eqn. (3.80)
trivially vanishes, i.e.,
!
X
2n
q
Ji Hi1 q1 x1 xq1
= 0.
q!
i=1
(3.81)
This can be shown explicitly as:

2n
X
Ji Hi1 q1 x1 xq1
i=1
= J11 H111 q1 + J12 H121 q1 + + J12n H12n1 q1

+ J21 H211 q1 + J22 H221 q1 + + J22n H22n1 q1
..
.
+ J2n1 H2n11 q1 + J2n2 H2n21 q1 + + J2n2n H2n2n1 q1 ,
= 0,
since Jab Hab1 q1 + Jba Hba1 q1 = 0 and Jaa Haa1 q1 = 0. Note that this fact is
equivalent to:
Ji Hi
T
= Trace JHxx
= 0.
Hence, the condition Eqn. (3.78) is satisfied, and thus, the PDF p(x, t) of the higher order
Hamiltonian system is a time invariant so that dp(x, t)/dt = 0.
67
Remark 3.8.1. The PDF of a higher order Hamiltonian system is a time invariant, and
hence, from Eqn. (3.50), we find:
[(t;
x
,
t
)]
((t,
x;
t
))
0
0
0
det
= det
= 1,
(x0 )
(x)
(3.82)
which is computed along the nominal relative trajectory (t; x0 , t0 ).
3.9 Nonlinear Mapping of the Gaussian Distribution

In 3.3 we discussed the difficulties of mapping the orbit uncertainties from a direct
integration. In this section, we assume the STTs are computed for a given reference trajectory and present an analytic mapping of orbit uncertainties as functions of the initial
statistics.
Lets first consider computation of the mean. Assuming the system is symplectic and
applying the results from Eqns. (3.68) and (3.69), we have the following four equivalent
expressions for the mean of the state x(t):
Z
E[x(t)] =
x(t) p(x, t)dx,
(3.83)
(t; x0 , t0 ) p(x0 , t0 )dx0 ,
(3.84)
(t; x0 , t0 ) p[(t; x0 , t0 ), t]dx0 ,
(3.85)
x(t) p[(t, x; t0 ), t0 ]dx.
(3.86)
Z
=
Z
=
Z
=
We observe that Eqn. (3.84) is suitable for computation of the state uncertainties using the
STT formulation because the solution flow can be expanded using the Taylor series and
higher moments can be computed using the JCF of the initial Gaussian distribution.
Consider the Gaussian boundary condition for the PDF, Eqn. (3.21). Assuming a
nonzero mean for the initial state, the PDF for the state x0 can be obtained via a linear
transformation, x0 = x0 + m0 m0 , where m0 is the initial mean and m0 is the initial
68
mean of the deviation. We note that these variables are constants. Applying the change of
variable to the PDF yields:
1
T
p(x0 , t0 ) = p
exp (x0 m0 ) 0 (x0 m0 ) . (3.87)
2
(2)N det P0
1
Since the expectation of the nominal trajectory does not change, by definition, it is easier
to instead analyze the statistics of the relative motion.
Using the state transition tensor notation and applying Eqn. (3.84), the current state
mean and covariance matrix can be stated as:
m
X
1 i,1 p 1

m (t) =
(t,t0 ) E x0 x0p ,
(3.88)
p!
p=1
m m
!
X X 1 i, j,
(t,t10 ) p (t,t10 ) q E[x01 x0p x01 x0q ] mi (t)mj (t),

Pij (t) =
p!q!
p=1 q=1
i
(3.89)
where {j , j } {1, , N } and the higher order expectations are defined as:
Z
E[x01
x0p ]
01 0 p p(0 , t0 )d(0 ).
(3.90)
This result is in general true for any dynamical system with any initial distribution; however, note that we arrived at this result from the very basics of probability and time invariance of the PDF. When m = 1, this propagation gives the ordinary first-order covariance
propagation, that is:
m(t) = (t, t0 )m0 ,
P(t) = (t, t0 )P0 (t, t0 )T m(t)m(t)T ,
(3.91)
(3.92)
where (t, t0 ) is the usual state transition matrix which linearly maps the deviation from
t0 to t. For the cases where m > 1 it is apparent from Eqn. (3.89) that we need to compute
2mth-order Gaussian moments, which is, however, a one time operation for the entire
69
trajectory since we only need to compute the moments of the initial distribution. If we
consider a different initial distribution these higher order moments have to be computed
again.
Considering that the initial distribution is Gaussian, i.e., Eqn. (3.87), the mean and
covariance matrix can thus be obtained as functions of time once we have the time solution
of STTs and the initial first two moments. Thus, the computation of the mean and the
covariance matrix is an algebraic operation. If we consider a zero initial mean, all the odd
moments of the initial conditions vanish, which is a property of the Gaussian distribution,
and the above equations simplify a great deal. Also, it is clear from Eqn. (3.88) that the
mean of the future trajectory will not be zero, indicating that the mean trajectory deviates
from the reference trajectory, whereas in the linear analysis the mean and the reference
trajectory coincide.
Example 3.9.1. Consider the case where m = 2 with zero initial mean. The mean and
covariance matrix become functions of the initial covariance matrix P0 , that is:
mi (t) =
1 i,ab ab
P ,
2 (t,t0 ) 0
i
1 i,ab j, h ab
a b
j,
a b
a
mi (t)mj (t).
P
+
P
+
P
P
P
P
Pij (t) = i,a
P
+
0
0
0
0
(t,t0 ) (t,t0 ) 0
4 (t,t0 ) (t,t0 ) 0 0
Considering this fact, the mean trajectory deviates from the deterministic trajectory as
m(t) 6= 0.
3.10 Monte-Carlo Simulations

In conventional trajectory navigation, Monte-Carlo simulation is often implemented
to analyze the nonlinear behavior of trajectory statistics. Monte-Carlo simulation is an
empirical method and approximates the orbit statistics by averaging over a large set of
random samples.
70
Definition 3.10.1 (Monte-Carlo Simulation). Given an initial distribution p(x0 , t0 ), the
Monte-Carlo simulation for the mean and covariance matrix are defined as [47, 51]:
1 X i
m (t) =
(t; x0 , t0 ),
=1
i
1 X i
P (t) =
(t; x0 , t0 ) mi (t) j (t; x0 , t0 ) mj (t) ,
1 =1
ij
(3.93)
(3.94)
where represents the number of random samples and x0 represents the th random
sample that is chosen according to the initial PDF p(x0 , t0 ). For example, if x0 is Gaussian,
each sample point can be drawn using the Gaussian random number generator.
Remark 3.10.2. The Monte-Carlo simulation for the relative motion can be stated as:
mi (t) =
Pij (t) =
1 X i
x (t; x0 , t0 ),
=1
(3.95)
1 X i
x (t; x0 , t0 ) mi (t) xj (t; x0 , t0 ) mj (t) , (3.96)
1 =1
where the random samples x0 are chosen according to the initial PDF p(x0 , t0 ).
Based on the law of large numbers and convergence of the statistics, Eqns. (3.93)
and (3.94) become the true mean and true covariance matrix as we consider more sample
trajectories, i.e., . In general, a sufficiently large number of samples give the true
probability distribution by taking higher order moments.
Precision orbit prediction often relies on Monte-Carlo simulations to predict the future
state for nonlinear dynamical situations. However, there are three critical disadvantages
when using this approach: (a) the number of sample trajectories may grow quite large to
obtain convergence of the statistics, (b) the simulation needs to be repeated for different
initial distributions, and (c) it does not provide an analytic framework. These problems
make the Monte-Carlo simulation computationally intensive and statistics-specific. For
71
practical problems, the initial samples are mapped based on a linearized model, assuming the initial error is small and is within the linear regime. One way to improve this
linear assumption is to apply the STT approach to capture the system nonlinearity in the
propagation model.
As discussed in 2.2.1, the computation of the mth order STTs for a system with a
P N 1+j
dimension N generally requires N m
differential equations to be integrated.
j=0
j
Monte-Carlo simulation, on the other hand, requires one to integrate N equations until the solution converges and it is often difficult to approximate a sufficient number of
samples for solution convergence. The number of integrated equations for the STTs may
exceed that of the Monte-Carlo simulation; however, the importance of the STT approach
comes from the fact that the STTs need be integrated only once and then can be used for
varying-epoch statistics, whereas the Monte-Carlo analysis needs to be recomputed for
each set of initial statistics.
3.11 Unscented Transformation

Another method for nonlinear orbit uncertainty propagation is the unscented transformation (UT), which was proposed by Julier et al. [45]. The UT is based on the idea that,
for a given system, it may be is easier to approximate the probability distribution than
the nonlinear transformation [43, 44, 46]. That is, instead of performing a higher order
analysis, the probability distribution at a future time can be approximated by nonlinearly
integrating a few samples that are deterministically chosen from the initial distribution.
Moreover, the weight of each sample does not have to lie in the range [0, 1], which is
different from the Monte-Carlo simulation.
Given an initial distribution with a mean m0 and a covariance matrix P0 , the UT is
72
initialized with (2N + 1) sample points:
X00 = m0 ,
ii
(N + )P0 ,
hp
ii
= m0
(N + )P0 ,
X0i = m0 +
X0i+N
(3.97)
hp
(3.98)
(3.99)
with sample weights:

W00 = /(N + )
(3.100)
W0i = Wki+N = 1/[2(N + )],
(3.101)
where <, X0j are the sample vectors with associated weights W0j , and
hp
(N + )P0
ii
are the ith row of the matrix square root of [(N + )P0 ]. With this initialization, the UT is
defined as (assuming no process noise):
X i (t) = (t; X0i , t0 ),
2N
X
m(t) =
W0i X i (t),
P(t) =
i=0
2N
X
T
W0i X i (t) m(t) X i (t) m(t) .
(3.102)
(3.103)
(3.104)
i=0
The UT requires N (2N + 1) differential equations to be integrated, which is on the order

of solving the Riccati equation, making it fast compared to the higher order STT approach
or the Monte-Carlo simulation. However, it is important to note that the UT is not analytic
with respect to initial statistics so that it must be carried out for different initial conditions,
and is limited to only a second order approximation of the dynamics.
CHAPTER IV
Generally, spacecraft navigation is composed of trajectory design, control, and estimation. To illustrate this point, consider an interplanetary trajectory to Mars from Earth.
Initially, a nominal trajectory is provided by the mission design which satisfies the required
mission objective. However, upon the launch of a spacecraft from Earth, which usually
introduces large errors, the spacecraft deviates from the previously defined nominal trajectory. Also, the spacecraft state is only known up to some uncertainties since the spacecraft
state is determined through an estimation/filtering process based on measurements such as
radiometric range, Doppler, or optical, that also have associated errors.
For this reason, a series of correction maneuvers must be applied to either target back
to the original nominal trajectory or to an alternate trajectory that also satisfies the mission requirement, which must be redesigned by considering the current spacecraft state.
Usually, a single correction maneuver cannot achieve this goal since 1) a spacecraft state is
uncertain, 2) each correction maneuver has associated uncertainties, 3) measurements have
associated uncertainties, and 4) the system model governing the spacecraft motion is not
perfect. Hence, when a correction maneuver is applied to refine its trajectory, the spacecraft is not precisely targeted back to a desired orbit due to attributed errors. Therefore,
the trajectory navigation is the continuation of trajectory design, targeting, and estimation,
73
74
and our goal is to more effectively perform these processes.
In conventional trajectory navigation, the nominal trajectory and correction maneuvers
are usually computed assuming the spacecraft state, control input, and the system model
are perfectly known [9, 83, 99]; hence, these computations are deterministically carried
out and related uncertainties are not taken into account. However, if the trajectory uncertainties are available and can be characterized, it may be is more robust to incorporate them
into trajectory and maneuver design processes. Also, conventional trajectory estimation
is built around the linear theory [23, 61, 100], which assumes the trajectory dynamics are
locally linear. In general, there are two types of estimation methods. One type is the batch
least-squares approximation, or linear regression, which reconstructs the entire spacecraft
trajectory by minimizing the measurement residual errors given a set of measurements. In
practice, the square-root information filter is often implemented as a batch estimator for
orbit reconstruction/determination problem which is numerically more stable and robust
than the least-squares filter. Another type is the sequential estimation, such as the extended
Kalman filter, and is usually divided into two parts: prediction and update.1 In this setting,
the spacecraft uncertainties are linearly propagated and measurements are updated using
linearized measurement models. Hence, system nonlinearities are not incorporated in the
filtering process.
In this chapter we discuss the trajectory navigation where our nonlinear uncertainty
propagation technique can be utilized and implemented. We first introduce the concept
of a statistically correct trajectory where we incorporate statistical information of an orbit
into the trajectory design. We then extend this idea and present a method of nonlinear
statistical targeting by computing the correction maneuver that gives a statistically more
1
The batch estimation is also divided into prediction and update parts, but the prediction part is performed
over the entire batch of measurements and the update incorporates the information contents of the entire
batch. This is in contrast to sequential filters which alternate prediction and update at each measurement
time.
75
accurate target solution at a desired time. Our uncertainty propagation method is also implemented using nonlinear filtering algorithms. We first derive an analytic expression of
the posterior density function for an optimal nonlinear filtering problem by using Bayes
rule and by incorporating higher order solutions of the relative spacecraft motion. We then
present two Kalman-type filters by directly applying the higher order state transition tensors to the Kalman filter algorithm.
E[x(t)+x(t)]
statistically
correct
trajectory
neighboring
deterministic
trajectory
x(tf )+x(tf )
x(t 0)+x(t 0)
x(tf )
x(t 0)
x(tf )
x(t 0)
p(t 0)
reference
trajectory
mean of
reference
trajectory
desired
target
m(tf )=E[x(tf )]
Figure 4.1: Illustration of the statistically correction trajectory.
4.1 The Concept of Statistically Correct Trajectory

Conventional mission design usually relies on the deterministic solution of a boundary
value problem; no statistical information is taken into account in the design process. The
idea of a statistically correct trajectory is to improve on this deterministic trajectory by
incorporating trajectory navigation information. Consider Figure 4.1, which illustrates the
76
concept of a statistically correct trajectory. Suppose the deterministic reference trajectory
satisfies the desired target state at tf so that:
xf = x(tf ) = (tf ; x0 , t0 ).
(4.1)
In practice, however, x0 = x(t0 ) is always associated with a non-zero uncertainty (e.g.,

x0 N (m0 , P0 )), and thus, it is inevitable that the mean, mf = m(tf ) = E[(tf ; x0 , t0 )],
deviates from the desired target, i.e., mf xf 6= 0. More specifically,
E[(tf ; x0 , t0 )] 6= (tf ; E[x0 ], t0 ).
(4.2)
Definition 4.1.1. Suppose we are given an initial state x0 with the probability density function p(x0 , t0 ) and it deterministically reaches the desired target xf = x(tf ) = (tf ; x0 , t0 ).
The statistically correct trajectory is the expectation of a neighboring trajectory (tk ; x0 +
x0 , t0 ) which satisfies:
E[(tk ; x0 + x0 , t0 )] xf = 0,
(4.3)
where x0 + x0 is assumed to have the same probability distribution as x0 , but the mean of
x0 is shifted by x0 to satisfy Eqn. (4.3). In other words, we are varying the mean of the
initial distribution p(x0 , t0 ) so that the propagated p(x0 , t0 ) has the mean located at xf . In
terms of relative motion, this can be viewed as x0 being a constant vector and x0 is a new
random vector with the same probability distribution as x0 with a mean shifted by x0 . The
goal of the statistically correct trajectory is then to find the mean of x0 , m0 = E[x0 ]
such that Eqn. (4.3) is satisfied. Note that time tk does not have to equal tf in general.
Remark 4.1.2. Considering the relative motion, the condition for the statistically correct
trajectory can be restated as:
E[(tk ; x0 + x0 , t0 ) (tk ; x0 , t0 )] [xf (tk ; x0 , t0 )]
= E[x(tk ; x0 , t0 )] c = 0,
(4.4)
77
where c = xf (tk ; x0 , t0 ) and does not depend on x0 . Applying the STT solutions,
E[x(tk ; x0 , t0 )] can be computed as:
m
X
1 i,1 p 1

E[x (tk ; x0 , t0 )] =
(tk ,t0 ) E x0 x0p ,
p!
p=1
i
(4.5)
where x0 is the random variable and has the same probability distribution as x0 . The
solution of the statistically correct trajectory is then the mean of x0 that satisfies Eqn.
(4.4).
Definition 4.1.3 (Differential Correction or Newtons Method). For y <N , suppose we
want to find a solution of y = h(x), i.e., find x <N such that y = h(x) or h(x)y = 0. To
achieve this goal, construct a convergent sequence (x) such that lim h[(x) ] y = 0.
Applying a Taylor series expansion gives:
h[(x) + (x) ] y = 0,
h[(x) ] +
(x) y 0,
x x=(x)
(x)
1
h
Newtons method is defined as:
(x)+1
h
= (x)
x
{h[(x) ] y} .
x=(x)
{h[(x) ] y} ,
(4.6)
x=(x)
and its convergence varies depending on the initial guess of the solution.
Remark 4.1.4. Assuming the Jacobian of Eqn. (4.3) exists and is invertible, the statistically
correct trajectory can be solved by applying Newtons method:
1
[e(t
;
x
,
t
)]
k
0 0
e[tk ; (x0 ) , t0 ],
(x0 )+1 = (x0 )
(x0 )
(4.7)
x0 =(x0 )
where
e[tk ; (x0 ) , t0 ] = E{[tk ; x0 + (x0 ) , t0 ]} xf ,
= E{x[tk ; (x0 ) , t0 ]} c.
(4.8)
78
Note that the superscript indicates the solution from the th iteration and the iteration
is carried out until m0 = E[(x0 ) ] is sufficiently small. The iteration can be initialized
with:
E[(x0 )1 ] = 1 (tk , t0 )m(tk ),
(4.9)
where m(tk ) = E[x(tk ; x0 , t0 )] is the deviated mean at tk due to the uncertainties in the
initial state.
Remark 4.1.5. If we consider the case where tk = tf , Eqn. (4.3) simplifies to:
E[(tf ; x0 + x0 , t0 ) xf ] = E[x(tf ; x0 , t0 )] = 0,
and Newtons method becomes:
(x0 )+1
(E[x(tf ; x0 , t0 )])
= (x0 )
(x0 )
m[tf ; (x0 ) , t0 ],
(4.10)
x0 =(x0 )
with the initial guess of

E[(x0 )1 ] = 1 (tf , t0 )m(tf ),
(4.11)
where m(tf ) = E[x(tf ; x0 , t0 )] is the deviated mean of the initial state.

It is important to note that if we fix the target time to be tk = tf the solution exists
for cases with sufficiently small initial uncertainties. However, by varying tk , there is
more freedom in solving for a x0 that converges. Once m0 = E[x0 ] is computed, the
statistically correct initial state x0 + E[x0 ] can be determined for the given initial probability distribution. If we consider the initial state to be Gaussian, the deviated mean, i.e.,
mk = E[x(tk ; x0 , t0 )], can be represented as a power series in the initial mean m0
and covariance P0 using Eqn. (3.88). Hence, once the STTs are integrated for the entire
trajectory up to tk , the Jacobian of mk can be computed analytically and the iteration
process becomes a trivial problem assuming convergence. In general, for a fixed xf , there
79
are varying mk = E[x(tk ; m0 , P0 )] that satisfies the statistically correct trajectory condition. Note that as P0 of the initial Gaussian distribution approaches zero, we recover the
conventional deterministic solution.
reference
trajectory
mr(tf )=0
p(x0,t0)
r(t0)
v(t0)
mean of
reference
trajectory
target
r(tf )
r(tk )
V k
r(tf )
r(tk ) = mr(tk )
v(tk ) = mv(tk )
deterministially
corrected
trajectory
mean of
deterministically
corrected trajectory
Figure 4.2: Illustration of the nonlinear statistical targeting.
4.2 Nonlinear Statistical Targeting

As an extension of the statistically correct trajectory, we describe how to practically
design a nonlinear statistical correction maneuver that uses this concept. For this example,
we focus on position targeting (i.e., interception) based on a single impulsive maneuver;
however, the result can be generalized to target the full state with two or more maneuvers.
Figure 4.2 illustrates the concept of nonlinear statistical targeting. In particular, suppose
we are given a reference trajectory (tf ; x0 , t0 ) and let rf = r(tf ) = r (tf ; x0 , t0 ) be the
fixed target position and vf = v(tf ) = v (tf ; x0 , t0 ) be the corresponding velocity vector, which can vary. Note that r represents the solution flow of the position components
(i.e., r {1, , N/2}). Also, suppose the initial state has a probability density function p(x0 , t0 ). Under this assumption, the mean spacecraft trajectory deviates from the
80
reference trajectory. At time tk , let the estimated mean position and velocity be r(tk ) =
mr (tk ) and v(tk ) = mv (tk ), respectively. The goal of nonlinear statistical targeting is
to design a maneuver Vk which satisfies E[r (tf ; rk + mrk , vk + mvk + Vk , tk )] = rf .
In other words, we solve for Vk such that mr (tf ) = E[r (tf ; rk + mrk , vk + mvk +
Vk , tk )] rf = 0, which can be stated analytically similar to Eqn. (3.88). Note that
Vk R3 , which is an exception of the notation used in this thesis.
In order to make a distinction between the conventional deterministic method and
our statistical method, we first discuss the correction maneuver design assuming perfect
knowledge of the initial state.
Definition 4.2.1 (Nonlinear Deterministic Targeting). Consider a reference trajectory
which hits the desired target rf = r(tf ) = r (tf ; x0 , t0 ) at tf , where r denotes position components of x. Suppose the spacecraft happens to be offset from the reference
trajectory at some time tk (t0 tk tf ) with deviations rk and vk due to errors in the
initial conditions. In order to re-target the spacecraft to hit the desired target, the goal of
conventional nonlinear targeting is to find a correction maneuver Vk such that:
r (ts ; rk + rk , vk + vk + Vk , tk ) rf = 0,
(4.12)
where ts tk . Note that Vk must be solved iteratively since r is nonlinear in general,

and note that the time of interception ts does not necessarily have to be tf . In practice, ts is
usually chosen based on some optimality constraints, such as one that minimizes kVk k.
Remark 4.2.2. At the linear level, the mean trajectory can be propagated as:
rr
rv
(tf , tk ) (tf , tk ) r(tk )
r(tf )
,
vr
vv
v(tk )
(tf , tk ) (tf , tk )
v(tf )
(4.13)
where is the usual state transition matrix mapping the deviations from tk to deviations
at tf .
81
Definition 4.2.3 (Linear Deterministic Targeting). Assuming the interception time is ts ,
the linear deterministic targeting problem solves:
r(ts ) =
rr
rv
(ts , tk ) (ts , tk )
r(tk )
= 0,
(4.14)
v(tk ) + Vk
which gives the the linear correction maneuver:

Vk = (rv )1 rr r(tk ) v(tk ).
(4.15)
With a deterministic correction maneuver, the deviated position is zero; however, we

know that the true trajectory (in a statistical sense) will likely miss the desired target,
depending on the associated uncertainties at tk , the transit period tf tk , system nonlinearity, etc., which may be significant.2 For this reason, additional maneuvers may be
needed. Our proposed statistical approach is to instead satisfy E[r(tf )] = mr (tf ) = 0.
In other words, instead of deterministically mapping the deviation, we apply the concept
of the statistically correct trajectory.
Definition 4.2.4 (Nonlinear Statistical Targeting). Suppose we are given a reference solution which hits the desired target rf = r(tf ) = r (tf ; x0 , t0 ) at time tf . However, due
to uncertainties associated with x0 , suppose at time tk the spacecraft is deviated from the
reference trajectory by position deviation of r(tk ) and velocity deviation of v(tk ), and
let p(xk , tk ) be the probability distribution of the state xk = [rTk , vTk ]T at time tk . The goal
of nonlinear statistical targeting is to find the correction maneuver Vk that satisfies:
E[r (ts ; rk + rk , vk + vk + Vk , tk )] rf = 0,
(4.16)
where ts tk in general and ts does not necessarily equal tf .

2
In practice, the validity of this deterministic correction maneuver is checked by running Monte-Carlo
simulations and making sure that the deviated mean is within some specified error bound with respect to the
target rf .
82
Remark 4.2.5. The condition for nonlinear statistical targeting can be restated as, find uk
such that:
E[r (ts ; rk + rk , vk + vk + Vk , tk ) r (ts ; x0 , t0 )] cr
= E[r (ts ; rk + rk , vk + uk , tk ) r (ts ; x0 , t0 )] cr ,
= E[r(ts ; rk , uk , tk )] cr = 0,
where cr = rf r (ts ; x0 , t0 ) and uk = vk + Vk so that xTk = [rTk , uTk ]. Note that
cr is constant for a fixed ts . Since the position is not changed instantaneously (i.e., rk is
constant), we only need to solve for uk such that:
m
!
X 1 r,
(ts 1,tk ) p E xk1 xkp

cr = 0.
p!
p=1
(4.17)
This nonlinear statistical targeting problem can be solved numerically using MonteCarlo simulations in an iterative way; however, a Monte-Carlo simulation does not provide
an analytic framework and is difficult to implement. By applying the higher order STT
solutions, depending on the initial probability distribution, the deviated position mean can
be determined analytically. If we assume the distribution at time ts can be approximated
by a Gaussian distribution, the position mean in Eqn. (4.17) is simply a power series in
the mean m(tk ) and covariance matrix P(tk ). Note that the order of the solution can
be chosen depending on the systems nonlinearity, which can be specified by the local
nonlinearity index. Therefore, we can compute Vk analytically while incorporating the
system nonlinearity, finding analytical results similar to a Monte-Carlo simulation.
In general, as long as the Jacobian of the higher order expectations in Eqn. (4.17)
exists and is invertible, the correction maneuver can be solved by iteration according to:
r
1
[e (ts ; rk , uk , tk )]
er [ts ; rk , (uk ) , tk ],
(uk )+1 = (uk )
(uk )
uk =(uk )
(4.18)
83
where
er [ts ; rk , (uk ) , tk ] = E[r (ts ; rk + rk , vk + uk , tk )] rf ,
and the superscript indicates an th iteration. For an initialization of the iteration, we can
use the solution from the linear correction maneuver as given in Eqn. (4.15), i.e.,
(uk )1 = [rv (ts , tk )]1 rr (ts , tk )rk .
(4.19)
Note that if we consider the case m = 1 in Eqn. (4.18), the solution is identical to the
linear targeting maneuver.
Remark 4.2.6. Consider the case where ts = tf . The nonlinear statistical targeting condition is:
mrf (tf ; rk , uk , tk )
m
X
1 r,1 p 1

(tf ,tk ) E xk xkp ,
=
p!
p=1
= 0,
and the iteration formula becomes:
1
[m
(t
;
r
,
u
,
t
)
f f
k
k k
(uk )+1 = (uk )
(uk )
mrf [tf ; rk , (uk ) , tk ],

uk =(uk )
(4.20)
with the initialization:
(uk )1 = [rv (tf , tk )]1 rr (tf , tk )rk .
(4.21)
Example 4.2.7. Assume xk N (mk , Pk ). The 3rd order STT propagated mean is:
mrf =
3
X
1 r,1 p 1

(tf ,tk ) E xk xkp ,
p!
p=1
= r,1 mk1 +
+
1 r,1 2
(mk1 mk2 + Pk1 2 )

2!
1 r,1 2 3
(mk1 mk2 mk3 + mk1 Pk2 3 + mk2 Pk1 3 + mk3 Pk1 2 ) ,

3!
(4.22)
84
where j {1, , N }. In this example, the solution of the nonlinear statistical targeting
is mk , which satisfies mrf = 0, where Pk is a constant covariance matrix at the time of
the maneuver (i.e., tk ) and can include maneuver execution uncertainties. Note that the
partial derivatives of mrf with respect to mk can be computed as:
(mrf )
(mjk )
1 r,1 2 j
1
2
1 2
r,1 j
1
= r,j
).
(tf ,tk ) + (tf ,tk ) mk + (tf ,tk ) (mk mk + Pk
2
(4.23)
Assuming that the deviated mean computed using an mth order STT solution is in
good agreement with a Monte-Carlo simulation result, only one correction maneuver is
required to hit the desired target on average according to probability theory. In terms of
Vk cost, the linear or nonlinear deterministic targeting method predicts that it is better
to perform correction maneuvers at an early stage of the trajectory, since the deviated
mean is smaller at the time of the correction maneuver. This yields a trade-off between
the final deviation and the Vk cost (i.e., if a maneuver is performed earlier the trajectory
deviates more). Using the nonlinear statistical method, the final deviated position mean is
zero. Moreover, we will see later that the nonlinear statistical targeting gives an optimal
time (minimum Vk ) to perform the correction maneuver, which is not found using the
deterministic method.
This assumes that the error due to Vk can be cast into the velocity covariance, which
is a simple transformation in case we assume vk and Vk are locally Gaussian.
4.2.1 On the Theoretic and Practical Aspects of Nonlinear Statistical Targeting
Although, in theory, the nonlinear statistical targeting provides a maneuver that is statistically more accurate than the linear statistical theory, there are practical, as well as
fundamental, problems that must be discussed. First we note that the nonlinear statistical
targeting depends on statistical knowledge about the initial state whereas the linear correction maneuver is independent of the statistics, and hence, there is only one maneuver
85
(assumed impulsive) in the deterministic targeting problem. In practical implementation
of these maneuvers, the statistical part of the linear maneuver is then checked by MonteCarlo runs to ensure that the spacecraft resides within the necessary error bound at the
target, which is often carried out based on larger-than-estimated initial uncertainties so
that a more conservative distribution can be treated at the target. However, we must keep
in mind that at the time of the maneuver, past navigation data is the only information
we have about the trajectory, and that the linear correction ignores a component of this
information in the design process, the level of uncertainties or covariance.
For a nonlinear maneuver, however, dependency on the navigation data makes nonlinear targeting more difficult for maneuver designers to implement since the calibrated
initial uncertainties will provide a different maneuver.3 This is precisely the place where
we can make an important, yet distinct, comparison between the two methods. Our study
shows that the nonlinearly propagated mean provides a more conservative solution than
the linear solution. We mean conservative in that the resulting projection of the covariance
matrix using the nonlinear correction gives more dispersed uncertainties at a future time
since the covariance matrix is positively affected (i.e., the covariance matrix is increased)
by the deviated mean. Hence, it may not be necessary to calibrate the initial uncertainties
to encompass possible additional errors.
Another concern one may have is, what if the navigation data is perceived to be too
accurate, so the covariance matrix is increased to compensate for potential solution errors,
or there are other uncertainty data from different navigation sources? If there is only one
initial distribution (i.e., only one filter solution from a navigation team), we are actually
increasing the region of event space to obtain a higher probability by increasing the initial
uncertainties. Hence, the covariance matrix should remain unchanged in the maneuver
3
Calibrated initial uncertainties mean increased initial covariances, which can occur if the initial state
errors are believed to be smaller than expected.
86
design process. When the uncertainties are increased to check the linear maneuver using a
Monte-Carlo simulation, it actually means that a larger initial error ellipsoid is considered
to increase the probability, not the covariance matrix. This indicates that both methods are,
in a way, utilizing the navigation data. However, it is important to note that the nonlinear
statistical targeting includes the navigation data in the maneuver design process, which is
usually not considered in conventional mission operations. If there are initial distributions
from different navigation sources (i.e., multiple filter solutions from different navigation
teams), it will be the maneuver designers decision to choose which navigation data to be
used. If the navigation data to be used is extremely accurate and if the system behaves
linearly, both correction maneuvers are essentially the same; however, there are usually
significant levels of uncertainties associated at an initial epoch, and correction maneuvers
computed from linear and nonlinear targeting methods will be different. The importance
of the nonlinear method is that the number of correction maneuvers can be reduced when
the statistical-based nonlinear correction is used because the spacecraft will more likely
lie within the necessary confidence region for a longer period.
In nonlinear statistical targeting, we instead aim for the target on average with a known
probability distribution at the time of the maneuver whereas the linear correction aims for
the target on the reference trajectory. From the integral invariance of the probability, it
is easy to show that the reference trajectory always depicts the highest value (i.e., mode)
of the probability density function for an initial Gaussian state, and hence, the linear correction maneuver is in a sense the most probable targeting correction for the initial state.
However, the target point is a single event in the probability space, and hence, has a zero
probability. On the other hand, the mean is, by definition, the expected or averaged value
of the deviated state. It is true that an infinitesimal volume around the reference will have
the highest probability, but this probability is negligible when compared to the initial phase
87
0.07
0.06
0.05
Nominal
Trajectory
ycoordinate
0.04
0.03
0.02
0.01
0
0.01
m=1
m=2
m=3
m=4
MonteCarlo
0.02
0.03
0.14
0.12
0.1
0.08
0.06
0.04
0.02
xcoordinate
Figure 4.3: Propagated mean and 1- error ellipsoid projected onto the spacecraft position
plane: comparison of the STT-approach and Monte-Carlo simulations.
volume we consider. To make a clear comparison, we must consider the probability of the
entire confidence region. In other words, both the target and mean have zero probabilities
in the probability space, but statistically speaking, the mean characterizes final location
more accurately given the initial phase volume.
As an example, consider Figure 4.3, which shows the propagated mean and 1- error
ellipsoid projected onto the spacecraft position plane considering an initial Gaussian distribution with zero mean.4 Here, m = i case represents the ith order STT solution, and
hence, m = 1 represents the linearly propagated uncertainties. One direct observation
from Figure 4.3 is that the linear solution gives a poor characterization of the propagated
uncertainties whereas higher order STT solutions (m 2) give more accurate description
of the propagated uncertainties as compared to the Monte-Carlo solution. Also, since the
initial distribution is Gaussian, the mean computed using the linear solution is located at
4
A detailed discussion of Figure 4.3 is given in 5.3.
88
the mode of the propagated PDF, or at the deterministic solution flow, whereas the higher
order solutions approximate the propagated mean. The difference between the linear and
nonlinear targeting methods is that the nonlinear statistical targeting applies a correction
maneuver so that the deviated mean intercepts the target. Hence, in this example, a correction maneuver designed using the nonlinear targeting method can be quite different from
the solution of the linear targeting method. Statistically, however, targeting the nonlinearly
propagated mean is more accurate than targeting the mode.
From a different perspective, we can propagate uncertainties to the final state, and
approximate the PDF as a Gaussian, using the first two moments. Then, in essence, the
maneuver design approach targets the mode of the statistical distribution mapped forward
in time. This takes advantage of the fact that the mean of a distribution does not follow the
dynamical equations of motion.
4.3 Higher Order Bayesian Filter with Gaussian Boundary Conditions

In this section, we derive a Bayesian filter which incorporates the system nonlinearity by applying the integral invariance property of the probability density function and
the higher order relative trajectory solutions. Here, we assume the spacecraft dynamics
model is known with perfect knowledge and only the initial state and measurements have
random errors that are modeled as Gaussian. This can be considered as the case where a
spacecraft is relatively quiet with no thrusters turned on (i.e., no process noise) and precise ephemerides are given. Also, we consider a measurement function which is a linear
function in state since our goal is to show the importance of the systems dynamical nonlinearity acting on the filtering process. Examples of nonlinear measurement models can
be found ins Refs. [53, 77, 89]. In this section, a higher order Bayesian filter is derived
89
based on a single measurement taken at some future time, tk , and expressions for posterior conditional mean and covariance matrix of the state after the measurement update are
presented. The results can be generalized to multiple measurement updates.
Suppose we want to estimate a spacecraft state x(tk ) RN given an initial state x0 with
a probability density function p(x0 , t0 ) and a set of measurements z1:k = {z1 , z2 , , zk },
which denotes all the measurements taken over the interval [t0 , tk ] and zj represents the
measurement vector taken at time tj . The goal of an optimal filtering problem is to compute the posterior conditional density at tk conditioned on a set of measurements, i.e.,
p(xk |z1:k ). The general Bayesian filtering problem is a two-step process, prediction and
update, defined as follows [21, 37]:
Definition 4.3.1 (Bayesian Prediction). General Bayesian prediction is defined as:
Z
p(xk |z1:k1 ) =
p(xk |xk1 , z1:k1 )p(xk1 |z1:k1 )dxk1 .
(4.24)
Definition 4.3.2 (Bayesian Measurement Update). General Bayesian measurement update

is defined as:
p(xk |z1:k ) =
p(zk |xk , z1:k1 )p(xk |z1:k1 )

.
p(zk |z1:k1 )
(4.25)
It is evident from Eqns. (4.24) and (4.25) that the Bayesian formulation is not a recursive process since we need to store and re-process all the measurements in order to
compute the terms p(xk |xk1 , z1:k1 ) and p(zk |xk , z1:k1 ), which are conditioned on z1:k1 .
This difficulty can be remedied by applying the following two assumptions.
Assumption 4.3.3 (Markov Process). The spacecraft state is a Markovian process:
p(xk |x0 , x1 , , xk1 ) = p(xk |xk1 ).
(4.26)
Assumption 4.3.4. Each measurement depends only on the state at the time the measurement is taken and is independent of previous measurements. Thus, the density conditioned
90
on xk and zk is independent of the previous states and measurements:
p(z1 , , zk |x1 , , xk ) =
k
Y
p(zi |xi ).
(4.27)
i=1
Example 4.3.5. Consider two measurements conditioned on two states:

p(z1 , z2 , x1 , x2 )
,
p(x1 , x2 )
p(z1 |z2 , x1 , x2 )p(z2 , x1 , x2 )
=
,
p(x1 , x2 )
p(z1 |z2 , x1 , x2 )p(z2 |x1 , x2 )p(x1 , x2 )
=
,
p(x1 , x2 )
p(z1 , z2 |x1 , x2 ) =
= p(z1 |x1 )p(z2 |x2 ).

Based on these two assumptions, we drop the conditioning on z1:k1 and we write
p(xk |xk1 , z1:k1 ) = p(xk |xk1 ). This leads to the following recursive prediction and
update equations:
Definition 4.3.6 (Recursive Bayesian Prediction).
Z
p(xk |z1:k1 ) =
p(xk |xk1 )p(xk1 |z1:k1 )dxk1 .
(4.28)
Definition 4.3.7 (Recursive Bayesian Measurement Update).

p(xk |z1:k ) =
p(zk |xk )p(xk |z1:k1 )

.
p(zk |z1:k1 )
(4.29)
Remark 4.3.8. Note that applying the law of total probability, the denominator in Eqn.
(4.29) can be stated as:
Z
p(zk |z1:k1 ) =
p(zk |xk )p(xk |z1:k1 )dxk ,
(4.30)
which is simply the integral of the numerator in Eqn. (4.29) over xk .

Now, consider the following discrete system realization of the relative dynamics:
xik+1
m
X
1 i,1 p 1
=
(tk+1 ,tk ) xk xkp + wik ,
p!
p=1
zk = Hk xk + vk ,
(4.31)
(4.32)
91
where zk is a single measurement taken at time tk , Hk is a row vector characterizing a
measurement function at time tk , vk is a measurement noise with zero mean5 and E[v2k ] =
k2 , and wk is the process noise which we assume to be zero. To ease the notation, let
k = x(tk ; x0 , t0 ) and zk = zk throughout this section unless noted otherwise. Our
system can be restated as:
i
k+1
m
X
1 i,1 p 1
=
(tk+1 ,tk ) k k p ,
p!
p=1
zk = Hk k + vk .
(4.33)
(4.34)
Suppose the initial state 0 = x0 can be characterized as Gaussian with zero mean and
covariance matrix P0 , i.e.,
1 T
p(0 , t0 ) = C0 exp 0 0 0 ,
2
(4.35)
p
where C0 = 1/ (2)N det P0 and 0 = P1
0 .
According to the recursive Bayesian prediction (4.28) we find:
Z
p(k , tk ) =
p(k , tk |0 , t0 )p(0 , t0 )d0 .
(4.36)
This is where the solution of the diffusion-less Fokker-Planck equation becomes useful.
Based on results from 3.5, the state PDF at time tk is related to the initial state PDF by:
p[(tk ; 0 , t0 ), tk ] = p(0 , t0 ) det

,

(4.37)
k
,
p(k , tk ) = p[0 (tk , ; t0 ), t0 ] det
0
(4.38)
or
and gives an analytic solution of the recursive Bayesian prediction equation. Considering
the initial Gaussian boundary conditions, the state PDF at time tk is:
5
The higher order Bayesian filter can be generalized for a case where the mean is non-zero.
92
1
T
p(k , tk ) = Ck exp 0 (tk , k ; t0 ) 0 0 (tk , k ; t0 ) ,
2
(
)
N
1 X
= Ck exp
,
2 ,=1 0 0 0
(4.39)
(4.40)
where 0 = 0 (tk , k ; t0 ) denotes a function of k and

Ck
.
= C0 det

(4.41)
Now suppose a single (scalar) measurement zk is taken at tk . In this case, the recursive
Bayesian equation, i.e., Eqn. (4.29), becomes:
p(zk |k )p(k )
,
p(z
|
)p(
)d
k
k
k
k
p(k |zk ) = R
(4.42)
where we drop the conditioning on z1:k1 . The PDF of the measurement conditioned on
the state in Eqn. (4.42) can be stated as:
(zk Hk )2
p(zk |k ) = Cz exp
2k2
(4.43)
where Czk = 1/ 2k . If p(zk |k ) is assumed to be Gaussian, the numerator of the

posterior conditional PDF then becomes:
pN (zk , k ) = p(zk |k )p(k ),
(4.44)
(zk Hk )2
1
= Ck Czk exp T0 (tk , k ; t0 )0 0 (tk , k ; t0 ) +
.
2
k2
(4.45)
Now consider the denominator of the Bayesian measurement update equation, pD =
p(zk |z1:k1 ) = p(zk ), which is the PDF of the measurement. Note that this is certainly
not Gaussian since:
" m
#
X 1 ,
(tk ,t1 0 ) p 0 1 0 p ,
E[zk ] = E[Hk + vk ] = H1 E
p!
p=1
(4.46)
93
and obviously shows non-Gaussian structure unless m = 1. Following Remark 4.3.8, we
have:
Z
pD (zk ) =
p(zk |k )p(k )dk ,
Z
1 T
(zk Hk )2
=
dk ,
Ck Czk exp 0 (tk , k ; t0 )0 0 (tk , k ; t0 ) +
2
k2
(4.47)
which is constant for the given measurement. However, we can map k back to the initial
epoch state and utilize the Gaussian form of the initial distribution to find its value. The
map (0 k ) is bijective, and thus, by mapping it to the initial epoch we can write:
pD (zk )
=
Z
Ce
Z
[zk Hk (tk ; 0 , t0 )]2

+
d0 ,
k2
2
zk 2zk Hk + Tk HT Hk
exp
d0 ,
2z2
!)
(
N
N
X
X
1
H1 H1 k k
exp 2 z2k 2zk
H1 k +
d0 ,
2k
=1
,=1
|
}
{z
1
C exp
2
T
0 0 0
2
Ce
T
0 0 0
2
T0 0 0
(4.48)
where C = C0 Czk and k = k (tk ; 0 , t0 ) is a function of 0 . Now, substitute Eqn. (4.33)
for k and decompose the second exponential function into linear and higher order terms
as:
"
m
!
N
X
X 1 ,
= 2 z2k 2zk
H1
(tk ,t01) p 0 1 0 p
2k
p!
=1
p=1
m
! m
!#
N
X
X 1 ,
X 1 ,
p
q
p
q
(tk ,t01) 0 1 0
(tk ,t1 0 ) 01 0
+
H1 H1
,
p!
q!
p=1
q=1
,=1
=
1
(zk H0 )2 + u(0 ),
2k2
(4.49)
94
where = (tk , t0 ) is the usual state transition matrix and
!
N
m
X
X
1
1 ,1 p 1
(tk ,t0 ) 0 0 p
u(0 ) = 2 2zk
H1
2k
p!
=1
p=2
N
m
X
X
1
1 ,1 p ,1 q 1

1 1
2
(tk ,t0 ) (tk ,t0 ) 0 0 p 01 0q
H H
.
2k ,=1
p!q!
p,q=1
p=q6=1
(4.50)
By substituting this result into Eqn. (4.48), the measurement PDF becomes:
)
(
Z
T0 0 0 (zk H0 )2
exp {u(0 )} d0 . (4.51)
pD (zk ) =
C exp
2
2k2
Remark 4.3.9 (Matrix Algebra). See Appendix C for details.
(y Mx)T R1 (y Mx) + (x m)T S1 (x m)

= yT R1 y + mT S1 m aT 1 a + (x 1 a)T (x 1 a),
(4.52)
where
= MT R1 M + S1 ,
a = MT R1 y + S1 m,
(4.53)
(4.54)
Moreover, the determinants are related by:

(det 1 )1/2 =
(det S)1/2 (det R)1/2

1/2 .
det(MSMT + R)
(4.55)
Applying the results from Remark 4.3.9 gives:
1
z2k
1
1
pD (zk ) =
C exp
+ 0 a 0 a
2 HP0 T HT + 2
exp {u(0 )} d0 ,
(4.56)
where
1 T T
H H + P1
0 ,
2
k
1 T T
a =
H zk .
k2
(4.57)
(4.58)
95
Finally, applying the determinant identity gives the following expression for the measurement PDF:
pD (zk ) =
Z
(zk )
(det )1/2 exp {u(0 )}

1
1 T
1
exp 0 a 0 a d0 ,
(2)N/2
2
(4.59)
where (zk ) is a constant for the given measurement and is defined as:
(
)
1
1
z2k
.
(zk ) =
1/2 exp 2
HP0 T HT + 2
(2)1/2 det(HP0 T HT + 2 )
(4.60)
Note that if we consider a linear system, Eqn. (4.59) simply reduces to pD (zk ) = (zk )
since u(0 ) = 0, and hence, the integral in Eqn. (4.59) becomes unity. In order to compute
its actual value, we can transform Eqn. (4.59) into an expectation form:
pD (zk ) = (zk ) E[exp {u(0 )}],
(4.61)
where u(0 ) is replaced with u(0 ) to denote that 0 now has an updated Gaussian probability density function of 0 = 0 N (1 a, 1 ). The mean and covariance matrix
of 0 can be re-written as:
1
1 a = P0 T HT HP0 T HT + k2
zk ,
1
1 = P0 P0 T HT HP0 T HT + k2
HP0 .
(4.62)
(4.63)
Note that Eqns. (4.62) and (4.63) are essentially the update equations for a linear model,
where 1 a is the updated mean and 1 is the updated covariance matrix. For computational purposes consider a Taylor series expansion of an exponential function:
u
X
ui
i=0
i!
= 1 + u1 +
u2 u3 u4 u5 u6 u7
+
+
+
+
+
+ .
2!
3!
4!
5!
6!
7!
(4.64)
96
We can apply this to calculate Eqn. (4.61) by considering a sufficient order of the solution.
That is, we can substitute u(), defined in Eqn. (4.50), into Eqn. (4.64) and truncate the
series by fixing the degree of the higher order moments. Using this result, the posteriori
conditional PDF of the state can be approximated by computing:
p(k |zk ) =
p(zk |k )p(k )
,
E[exp {u(0 )}]
(4.65)
and its complete form is:
k
det
p
p(k |zk ) =
0
k (2)N +1 E[exp {u(0 )}]
1 T
(zk Hk )2
exp 0 (tk , k ; t0 )0 0 (tk , k ; t0 ) +
. (4.66)
2
k2
(det P0 )1/2
If we consider a Hamiltonian system,6 the determinant of the Jacobian becomes unity, i.e.,
| det(k /0 )| = 1, and hence the posteriori conditional PDF simplifies to:
p(k |zk ) =
1 T
(zk Hk )2
p
exp 0 (k )0 0 (k ) +
,
2
k2
k (2)N +1 E[exp {u(0 )}]
(det P0 )1/2
(4.67)
where 0 (k ) = 0 (tk , k ; t0 ) is to denote a function of k and can be computed analytically with an STT theory.
We are interested in computing the first and second central moments of the posterior
conditional density, i.e., the updated mean and updated covariance matrix, which are defined as:
+
= E[k |zk ],
k
(4.68)
+ T
= E[(k +
P+
k )(k k ) |zk ],
k
(4.69)
where plus signs indicate updated values.

6
Assuming sufficiently high order solution is considered, the relative flow is almost symplectic.
97
First consider the updated mean equation. By directly applying expectation and mapping it back to the initial state, we find:7
Z
(k+ )i
ki p(k |zk )dk ,
m
!
Z
(zk )
(det )1/2 X 1 i,1 p 1
0 0 p exp {u(0 )}
pD (zk ) (2)N/2
p!
p=1
1
T
1
1
exp 0 a 0 a d0 ,
2
i
E[k (0 ) exp{u(0 )}]
=
.
(4.70)
E[exp{u(0 )}]
Applying a similar method, the updated covariance matrix equation can be stated as:
Z
ij
(P+
k)
ki kj p(k |zk )dk
(k+ )i (k+ )j ,
! m
!
Z X
m
X
1 i,1 p 1
1
0 0 p
j,1 p 01 0q exp {u(0 )}
pD p=1 p!
q!
p=1
1/2
(det )
1
1 T
1
exp 0 a 0 a d0 (k+ )i (k+ )j ,

(2)N/2
2
E[ki (0 )kj (0 ) exp{u(0 )}]

(k+ )i (k+ )j ,
E[exp{u(0 )}]
(4.71)
0 N (1 a, 1 ). An ith order moment can be stated as:

E[k1 ki |zk ] =
E[k1 (0 ) ki (0 ) exp{u(0 )}]

,
E[exp{u(0 )}]
(4.72)
which is analytic in 0 and is completely described by a Gaussian distribution. Since

0 is Gaussian with the mean 1 a the covariance matrix 1 , E[exp{u(0 )}] can be
computed analytically after truncating the exponential series Eqn. (4.64) depending on the
desired accuracy of the solution.
We note that if we consider linear system dynamics, Eqns. (4.70) and (4.71) simplify
to the conventional Kalman filter algorithm for the initial state estimation:
7
Note that this process is similar to the computation of the measurement PDF, pD = p(zk ).
98
1
+
= 1 a = P0 T HT HP0 T HT + k2
zk ,
k
(4.73)
1
P+
= 1 = P0 P0 T HT HP0 T HT + k2
HP0 .
k
(4.74)
4.4 Implementation of a Nonlinear Filter

In 4.3 we have shown that, in theory, the higher order solutions of the relative dynamics can be used to approximate a posterior conditional density function by using Bayes
rule and we have derived an optimal nonlinear filter. This formulation, however, can be difficult to implement due to the series expansion of the exponential function in Eqn. (4.72),
which can be quite complicated. Hence, we present two sub-optimal nonlinear filters by
directly incorporating the higher order solutions into the Kalman filtering algorithm, which
are simpler and easier to implement than the Bayesian formulation.
Although the Kalman filter algorithm can be derived from Bayes rule of conditional
densities, as pointed out by Julier et al. [44], Kalmans original derivation did not come
from the Bayesian approach [49], but rather from estimations of a few expectations involving a state and a measurement. To show this, consider the following general system
model in discrete form:
xk+1 = (tk+1 ; xk , tk ) + wk ,
(4.75)
zk+1 = h(xk+1 , tk+1 ) + vk+1 ,
(4.76)
where xk is the true spacecraft state, is the solution flow, wk is the white process noise
perturbing the spacecraft dynamics, zk is the actual measurement, h is the measurement
function, and vk is white measurement noise characterizing the observation error. The
process noise and measurement noise are assumed to be non-correlated, i.e., E[vi wTj = 0],
99
with the autocorrelations:
E[wi wTj ] = Qi ij ,
(4.77)
E[vi vTj ] = Ri ij ,
(4.78)
for all discrete time indexes i and j, where ij represents the Dirac delta function. Here,
Qi and Ri are also known as the diffusion and measurement noise matrices, respectively.
Definition 4.4.1 (Kalman Filter Algorithm). Given the system model Eqns. (4.75) and
(4.76), suppose we are given a state xk with mean m+
k = E[xk ] and covariance matrix
+
+ T
P+
k = E[(xk mk )(xk mk ) ] at time tk . The Kalman algorithm is defined as follows:
Kalman Filter Prediction Equations:

m
k+1 = E[(tk+1 ; xk , tk ) + wk ],
(4.79)
T
P
k+1 = E [(tk+1 ; xk , tk ) + wk ][(tk+1 ; xk , tk ) + wk ]
T
(m
k+1 )(mk+1 ) ,
n
k+1 = E[h(xk+1 , tk+1 ) + vk+1 ],
(4.80)
(4.81)
where n
k = E[hk ] is the expectation of the measurement computed at tk .
Kalman Filter Update Equations:
zz
1
Kk+1 = Pxz
k+1 (Pk+1 ) ,
(4.82)
m+
k+1 = mk+1 + Kk+1 zk+1 nk+1 ,
(4.83)
T
zz
P+
k+1 = Pk+1 Kk+1 Pk+1 Kk+1 ,
(4.84)
where Kk is known as the Kalman gain matrix, Pxz

k is the cross-covariance matrix of the
state and the measurement, Pzz
k is the covariance matrix of the measurement, zk is the
observation, and the difference between the actual and predicted measurement (i.e., zk
n
k ) is called the residual or innovation.
100
4.4.1 Extended Kalman Filter
For estimation problems, the linear Kalman filter (LKF) is probably the most well
known filtering technique. The LKF allows one to compute the minimum mean-squareerror (MMSE) solution; however, it can only be used for linear systems, and in general,
cannot be used for trajectory navigation. In conventional spacecraft trajectory navigation, the extended Kalman filter (EKF) is usually implemented.8 The EKF is based on
the Kalman filter algorithm given in Eqns. (4.79-4.84), but assumes the true trajectory is
within the boundary where the linear approximation about a reference trajectory can sufficiently model the trajectory dynamics and its statistics. Under this assumption, the mean
trajectory is propagated according to the deterministic solution flow and the covariance
matrix is linearly mapped assuming Gaussian statistics [2, 12, 100].
Definition 4.4.2 (Extended Kalman Filter Algorithm).
EKF Prediction Equations:
+
m
k+1 = (tk+1 ; mk , tk ),
(4.85)
+ T
P
k+1 = (tk+1 , tk )Pk (tk+1 , tk ) + Qk ,
(4.86)
n
k+1 = h(mk+1 , tk+1 ).
(4.87)
EKF Update Equations:

zz
1
Kk+1 = Pxz
k+1 (Pk+1 ) ,
T
T
1
= P
k+1 Hk+1 (Hk+1 Pk+1 Hk+1 + Rk+1 ) ,
m+
k+1 = mk+1 + Kk+1 (zk+1 nk+1 ),
(4.88)
(4.89)
T
zz
P+
k+1 = Pk+1 Kk+1 Pk+1 Kk+1 ,
= P
k+1 Kk+1 Hk+1 Pk+1 ,
8
(4.90)
In practice, the extended Kalman filter is implemented for trajectory navigation often in square-root
information filter (SRIF) or in U-D filter formulation for numerical precision.
101
where h(m
k+1 , tk+1 ) is the measurement function evaluated at tk+1 a function of mk+1 and
Hk = hk /xk is the measurement partial computed at tk .

Among the many important properties of the extended Kalman filter, we point out two
which will be discussed in Chapter V in more detail. Considering the gain Eqn. (4.88) and
the mean update Eqn. (4.89), we observe that as the a priori covariance matrix becomes
more accurate (i.e., P
k+1 0) the filter values the residual less (i.e., the actual measurement is trusted less). On the other hand, as the measurement becomes more accurate
(i.e., Rk+1 0) the filter values the residual more (i.e., the actual measurement is trusted
more). Therefore, optimally weighting the residual is a critical component of maximizing
the filter performance.
4.4.2 Higher-Order Numerical Extended Kalman Filter
In deriving the higher-order numerical extended Kalman filter (HNEKF), we assume
that the reference trajectory and its higher order state transition tensors are integrated for
each time interval between the measurements according to Eqns. (2.27-2.30). Under this
assumption the local trajectory motion can be mapped analytically over this time interval
while incorporating nonlinear effects, and the same analogy applies when mapping the
trajectory statistics. We note that this process is numerically quite intensive considering
higher order solutions; however, this can yield a more accurate filter solution.
Once the higher order state transition tensors are available for some time interval
[tk , tk+1 ], the mean and covariance matrix of the relative dynamics at tk can be mapped
analytically to tk+1 as functions of the probability distribution at tk as discussed in 3.9.
From tk to tk+1 , the propagated mean and covariance can be stated as:
mik+1 (xk ) = E xik+1 ,

m
X
1 i,1 p 1

=
(tk+1 ,tk ) E xk xkp ,
p!
p=1
(4.91)
102
i
j
j
i
Pij
k+1 (xk ) = E (xk+1 mk+1 )(xk+1 mk+1 ) ,
m m
!
X X 1 i, j,
p 1
q
1
p
q
1
1
=
E[xk xk xk xk ]
p!q! (tk+1 ,tk ) (tk+1 ,tk )
p=1 q=1
mik+1 mjk+1 ,
(4.92)
where {j , j } {1, , N }. Now, the only unknowns in Eqns. (4.91-4.92) are the
expectations (i.e., moments) of the deviations. Even if the state at time tk is Gaussian,
except for the case m = 1, it is obvious that the mapped trajectory distribution is no longer
Gaussian due to system nonlinearity, and hence exact computation requires computation
of the higher order moments.
In particle-based filters, this problem is remedied by using an ensemble of sample
points to approximate the probability distribution, whereas a more formal approach is
to use the Edgeworth/Gram-Chalier [55] or Laplace approximations to approximate the
posterior density function. In trajectory navigation, however, the Gaussian assumption
has proven to provide a sufficiently accurate statistical approximation. Hence, we assume
that the updated estimates are Gaussian and we implement the joint characteristic function9
to compute the higher order moments up to 2mth-order as apparent from Eqn. (4.92). As
the order of the solution increases, i.e., m , the higher order solution yields the true
Monte-Carlo mean and covariance matrix as discussed in 3.10.
+
Now, suppose at time tk , the state estimate has mean m+
k and covariance matrix Pk .
Also, let x(tk ) = m+

k + xk be the true trajectory we want to estimate. Following the
Kalman filter algorithm, the HNEKF algorithm is given as follows:
By assuming the updated state can be approximated with Gaussian statistics, the higher order moments
are functions of the first two moments. If we consider a zero initial mean, all the odd moments of the
initial conditions vanish, which is the unique property of the Gaussian distribution, and the equations for the
propagated mean and covariance matrix simplify a great deal.
103
HNEKF Prediction Equations:
i
i
(m
= E[i (tk+1 ; m+
k+1 )
k + xk , tk ) + wk ],
ij
(P
k+1 )
i
= i (tk+1 ; m+
k , tk ) + mk+1 (xk ),
m
X
1 i,1 p 1

= i (tk+1 ; m+
,
t
)
+
(4.93)
k
k
p!
p=1
i
j
+
= E [ (tk+1 ; mk + xk , tk ) + wik ][j (tk+1 ; m+
k + xk , tk ) + wk ]
i
j
(m
k+1 ) (mk+1 ) ,
!
m m
X X 1 i, j,
q
p 1
1
p
q
1
1
E[xk xk xk xk ]
=
p!q! (tk+1 ,tk ) (tk+1 ,tk )
p=1 q=1
mik+1 (xk )mjk+1 (xk ) + Qij

k,
(4.94)
(4.95)
i
(n
= E[hi (tk+1 ; m+
k+1 )
k + xk , tk ) + vk+1 ],
i
= hi (tk+1 ; m+
k , tk ) + nk+1 (xk ),
m
X
1 i,1 p 1

= hi (tk+1 ; m+
,
t
)
+
h(tk+1 ,tk ) E xk xkp ,
k
k
p!
p=1
(4.96)
+
where the STTs (i.e., i (tk+1 ; m+
k , tk )) are computed along the solution flow (tk+1 ; mk , tk )
and
i,1 p
h(tk+1
,tk )
p hik+1
(xk1 ) (xkp ) xk+1 =(tk+1 ;m+ ,tk )
(4.97)
Note that h
(tk+1 ; m+
k , tk )
denotes that the measurement function is evaluated at tk+1 as

i,
p
1
10
a function of the solution flow (tk+1 ; m+
The partial derivatives h(tk+1
k , tk ).
,tk ) up to
fourth order are defined as:

i,
,a
hi,a
(tk+1 ,tk ) = hk+1 k+1 ,
(4.98)
i,
,ab
i, ,a ,b
hi,ab
(tk+1 ,tk ) = hk+1 k+1 + hk+1 k+1 k+1 ,
i,
,abc
i,
,a ,bc
,ab ,c
,ac ,b
hi,abc
=
h
+
h
k+1 k+1
k+1
k+1 k+1
k+1 k+1
k+1 k+1
(tk+1 ,tk )
(4.99)
,a ,b ,c
+ hi,
k+1 k+1 k+1 k+1 ,
(4.100)
+
It is important to note that h(tk+1 ; m+
k , tk ) 6= h(mk+1 , tk+1 ) in general since mk+1 6= (tk+1 ; mk , tk )
for general nonlinear systems.
10
104
i,
,abcd
i,
,abc ,d
,abd ,c
,acd ,b
,ab ,cd
hi,abcd
=
h
+
h
k+1 k+1
k+1 k+1 k+1 + k+1 k+1 + k+1 k+1 + k+1 k+1
(tk+1 ,tk )
,bd
,ad ,bc
,a ,bcd
i,
,c ,d
+ ,ac
+
h
,ab
k+1 k+1
k+1 k+1
k+1 k+1
k+1
k+1 k+1 k+1
,b ,d
,ad ,b ,c
,a ,bc ,d
,a ,bd ,c
+ ,ac
k+1 k+1 k+1 + k+1 k+1 k+1 + k+1 k+1 k+1 + k+1 k+1 k+1
,b ,cd
i, ,a ,b ,c
,d
+ ,a
k+1 k+1 k+1 ) + hk+1 k+1 k+1 k+1 k+1 ,
(4.101)
where k+1 = (tk+1 ,tk ) is used for a concise notation and that these are similar to the
differential equations of the STTs given in Eqns. (2.27-2.30). Note that this prediction
step is a simple algebraic operation once the STTs are computed for the time interval
[tk , tk+1 ].
HNEKF Update Equations:
ij
T ij
(Pzz
= E (z
,
k+1 )
k+1 nk+1 )(zk+1 nk+1 )
i
j
i
j
= E (z
)
(z
)
(n
k+1
k+1
k+1 ) (nk+1 ) ,
j
j
+
i
= E [hi (tk+1 ; m+
k + xk , tk ) + vk+1 ][h (tk+1 ; mk + xk , tk ) + vk+1 ]
i
j
(n
k+1 ) (nk+1 ) ,
!
m X
m
X
1 i,1 p j,1 q
p 1
q
ij
1
h
=
Rk+1 +
h
E[xk xk xk xk ]
p!q! (tk+1 ,tk ) (tk+1 ,tk )
p=1 q=1
i
j
(n
k+1 ) (nk+1 ) ,
(4.102)
ij
T ij
(Pxz
= E (x
,
k+1 )
k+1 mk+1 )(zk+1 nk+1 )
i
j
i
j
= E (x
)
(z
)
(m
k+1
k+1
k+1 ) (nk+1 ) ,
j
j
+
i
= E [i (tk+1 ; m+
k + xk , tk ) + wk ][h (tk+1 ; mk + xk , tk ) + vk+1 ]
j
)i (n
(m
k+1 ) ,
m k+1
!
m
XX
1 i,1 q j,1 p
=
(tk+1 ,tk ) h(tk+1 ,tk ) E[xk1 xkp xk1 xkq ]
p!q!
p=1 q=1
j
i
(m
k+1 ) (nk+1 ) ,
(4.103)
105
zz
1
Kk+1 = Pxz
k+1 (Pk+1 ) ,
(4.104)
m+
k+1 = mk+1 + Kk+1 (zk+1 nk+1 ),
(4.105)
zz
T
P+
k+1 = Pk+1 Kk+1 Pk+1 Kk+1 .
(4.106)
Note that if we consider the measurement function Eqn. (4.76) to be linear in xk , Eqn.
(4.96) simplifies to:
i
i
(n
= hi (tk+1 ; m+
k+1 )
k , tk ) + (nk+1 ) ,
m
X
1 ,1 p 1

i,
i
+
= h (tk+1 ; mk , tk ) + hk+1
p!
p=1
i,
= hi (tk+1 ; m+
k , tk ) + hk+1 (mk+1 ) ,
= hi (m
k+1 , tk+1 ),
(4.107)
i,
and gives (n
k+1 ) = hk+1 (mk+1 ) . Applying this result, Eqns. (4.102) and (4.103)
simplify to:
ij
(Pzz
=
k+1 )
!
m X
m
X
1
,1 p ,1 q
p 1
q
j,
1
hi,
(tk+1
k+1 hk+1
,tk ) (tk+1 ,tk ) E[xk xk xk xk ]
p!q!
p=1 q=1
i
j
+ Rij
k+1 (nk+1 ) (nk+1 ) ,
i, j,
i
j
= Rij
+
h
h
E[x
x
]
(n
k+1 k+1
k+1
k+1 k+1
k+1 ) (nk+1 ) ,
ij
(Pxz
k+1 )
T
ij
= (Hk+1 P
(4.108)
k+1 Hk+1 + Rk+1 ) ,
!
m
m
X X 1 i, ,
p 1
q
1
p
q
1
1
(tk+1
=
hj,
k+1
p!q!
p=1 q=1
i
j
(m
k+1 ) (nk+1 ) ,
j
i
= E[xik+1 xk+1 ]hj,
k+1 (mk+1 ) (nk+1 ) ,
T
ij
= (P
k+1 Hk+1 ) ,
(4.109)
which indicates that the measurement prediction and update equations are identical to the
EKF algorithm.11 Moreover, for a linear measurement function, it is apparent that we
11
Note that when m = 1, the HNEKF becomes the EKF algorithm as shown in Eqns. (4.85-4.90).
106
can implement the Potters algorithm to develop a square-root filter for numerical stability
[12, 56]. This, however, is not obvious if we consider a nonlinear measurement function as
the Potters algorithm depends on there being a Cholesky decomposition, or alike, of the
covariance matrix, which may not be true for the higher order tensors. An extension to a
square-root filter, if possible, is not considered in this thesis since the focus our study is to
establish a general filter setup which incorporates the higher order dynamics and statistics.
4.4.3 Higher-Order Analytic Extended Kalman Filter
From the derivation of the HNEKF, it is obvious that we can also derive a higher-order
analytic extended Kalman filter (HAEKF) by assuming that the reference trajectory and
the higher order solutions (i.e., STTs) are computed over some time span prior to filtering.
The filter algorithm is similar to the HNEKF except that the point of series expansion is
now with respect to the initial reference trajectory, not the updated mean as in the HNEKF
algorithm.
Suppose the STTs are computed for the time interval of [t0 , tf ] and let xk = (tk ; x0 , t0 )
represent the reference trajectory for tk [t0 , tf ], where x0 has mean m+
0 and covariance
k + xk be the true trajectory we want to estimate.
matrix P+
0 . Moreover, let x(tk ) = x
Following the Kalman filter algorithm, the HAEKF algorithm is given as follows:
HAEKF Prediction Equations:
i
(m
= E[i (tk+1 ; xk + xk , tk ) + wik ],
k+1 )
= i (tk+1 ; xk , tk ) + mik+1 (xk ),

m
X
1 i,1 p 1

= i (tk+1 ; xk , tk ) +
p!
p=1
(4.110)
107
ij
(P
= E [i (tk+1 ; xk + xk , tk ) + wik ][j (tk+1 ; xk + xk , tk ) + wjk ]
k+1 )
j
)i (m
(4.111)
(m
k+1 ) ,
m k+1
!
m
X X 1 i, j,
p 1
q
1
p
q
1
1
=
(tk+1
p!q!
p=1 q=1
mik+1 (xk )mjk+1 (xk ) + Qij

k,
(4.112)
i
= E[hi (tk+1 ; xk + xk , tk ) + vik+1 ],
(n
k+1 )
= hi (tk+1 ; xk , tk ) + nik+1 (xk ),

m
X
1 i,1 p 1

i
= h (xk+1 , tk+1 ) +
h(tk+1 ,tk ) E xk xkp ,
p!
p=1
where the STTs are computed along xk+1 = (tk+1 ; xk , tk ) and
p hik+1
i,1 p
h(tk+1 ,tk ) =
.
p
1
(xk ) (xk ) xk+1 =xk+1
(4.113)
(4.114)
HAEKF Update Equations:
ij
T ij
(Pzz
= E (z
,
k+1 )
k+1 nk+1 )(zk+1 nk+1 )
i
j
i
j
= E (z
(n
k+1 ) (zk+1 )
k+1 ) (nk+1 ) ,
= E [hi (tk+1 ; xk + xk , tk ) + vik+1 ][hj (tk+1 ; xk + xk , tk ) + vjk+1 ]

(n
)i (n )j ,
k+1 mk+1m
!
X
X
1
i,1 p j,1 q
p 1
q
1
h(tk+1
=
Rij
k+1 +
,tk ) h(tk+1 ,tk ) E[xk xk xk xk ]
p!q!
p=1 q=1
i
j
(n
k+1 ) (nk+1 ) ,
(4.115)
ij
T ij
(Pxz
= E (x
,
k+1 )
k+1 mk+1 )(zk+1 nk+1 )
j
i
j
i
(m
= E (x
k+1 ) (nk+1 ) ,
k+1 ) (zk+1 )
= E [i (tk+1 ; xk + xk , tk ) + wik ][hj (tk+1 ; xk + xk , tk ) + vjk+1 ]

j
)i (n
(m
k+1 ) ,
m k+1
!
m
XX
1 i,1 q j,1 p
=
(tk+1 ,tk ) h(tk+1 ,tk ) E[xk1 xkp xk1 xkq ]
p!q!
p=1 q=1
j
i
(m
k+1 ) (nk+1 ) ,
(4.116)
108
zz
1
Kk+1 = Pxz
k+1 (Pk+1 ) ,
(4.117)
m+
k+1 = mk+1 + Kk+1 (zk+1 nk+1 ),
(4.118)
zz
T
P+
k+1 = Pk+1 Kk+1 Pk+1 Kk+1 .
(4.119)
As in the HNEKF case, the update equations for the HAEKF becomes the same as the
EKF when we consider a measurement function that is linear in xk :
i
i
(n
= hi (xk+1 , tk+1 ) + (n
k+1 )
k+1 ) ,
m
X
1 ,1 p 1

i,
i
= h (xk+1 , tk+1 ) + hk+1
p!
p=1
= hi (xk+1 , tk+1 ) + hi,

k+1 (mk+1 ) ,
ij
(Pzz
k+1 )
= hi (m
, tk+1 ),
(4.120)
k+1
!
m
m
X X 1 , ,
i, j,
p
q
p
q
1
1
1
1
=
hk+1 hk+1
(tk+1
p!q!
p=1 q=1
i
j
+ Rij
k+1 (nk+1 ) (nk+1 ) ,
i, j,
i
j
= Rij
+
h
h
E[x
x
]
(n
k+1
k+1
k+1 k+1
k+1
k+1 ) (nk+1 ) ,
ij
(Pxz
k+1 )
T
ij
= (Hk+1 P
(4.121)
k+1 Hk+1 + Rk+1 ) ,
!
m X
m
X
1 i,1 p ,1 q
(tk+1 ,tk ) (tk+1 ,tk ) E[xk1 xkp xk1 xkq ]

=
hj,
k+1
p!q!
p=1 q=1
i
j
(m
k+1 ) (nk+1 ) ,
i
j
= E[xik+1 xk+1 ]hj,
k+1 (mk+1 ) (nk+1 ) ,
T
ij
= (P
k+1 Hk+1 ) .
(4.122)
Also, note that when m = 1 (i.e., first order), the HAEKF becomes the linear Kalman
filter (LKF), not the EKF.12 The superiority of the EKF over the LKF is clearly demonstrated in Maybeck [55]. However, when the true trajectory is within the convergence
radius of the reference trajectory, we shall see later that the HAEKF can provide a more
12
We call it the higher-order analytic extended Kalman filter, not higher-order linear Kalman filter, since
the prediction equations are nonlinear in general.
109
accurate solution and faster convergence than the EKF.
4.4.4 Unscented Kalman Filter
The unscented Kalman filter (UKF), first introduced by Julier and Uhlmann [44, 46,
98], has been implemented in diverse fields of engineering, science, and economics due to
its simplicity while providing faster convergence and better accuracy than the EKF. The
first implementation of the UKF to a trajectory navigation problem was discussed in Refs.
[39, 40], where Mars aerobraking spacecraft state is estimated using inertial measurement
unit data. The UKF is based on the unscented transformation discussed in 3.11, which
deterministically chooses the sample points to approximate the probability distribution
while keeping the computational cost at the order of the linear methods. The UKF provides
a good approximation of the true probability distribution and lower expected errors than
the EKF, and it does not require calculation of the Jacobian matrix. However, it depends
on how the initial sample points are chosen and parameterized, and generally captures only
the first three moments of an arbitrary distribution.13
Here, we do not go through the detailed derivation, as thorough discussions can be
found in Refs. [43, 44, 45, 46, 97]. We only present the UKF algorithm for additive
(linear) process and measurement noises with zero mean.
The UKF is initialized with the following deterministically chosen sample points:
13
Xk0 = m+
k,
(4.123)
Wk0 = /(N + ),
hp
ii
+
Xki = m+
(N
+
)P
,
k
k
(4.124)
Wki = 1/[2(N + )],

hp
ii
Xki+N = m+
(N
+
)P
k ,
k
(4.126)
In case of a Gaussian distribution, UKF captures the first four moments.
(4.125)
(4.127)
110
Wki+N = 1/[2(N + )],
where <, Xkj are the sample points with associated weights Wkj , and
(4.128)
hp
ii
(N + )P(t0 )
is the ith row of the matrix square root of [(N + )P(t0 )]. With this initialization, the UKF
algorithm is given as follows:
UKF Prediction Equations:
i
= (tk+1 ; Xki , tk ),
Xk+1
2N
X
i
mk+1 =
Wki Xk+1
,
P
k+1 =
i=0
2N
X
i
i
T
Wki Xk+1
m
Xk+1 m
+ Qk ,
k+1
k+1
(4.129)
(4.130)
(4.131)
i=0
i
Zk+1
n
k+1
i
= h(Xk+1
),
2N
X
i
=
Wki Zk+1
.
(4.132)
(4.133)
i=0
UKF Update Equations:

Pzz
k+1
Pxz
k+1 =
Kk+1 =
2N
X
i=0
2N
X
i
i
T
Wki Zk+1
n
Z
n
+ Rk ,
k+1
k+1
k+1
(4.134)
i
i
T
Wki Xk+1
m
Zk+1 n
,
k+1
k+1
(4.135)
i=0
zz
1
Pxz
k+1 (Pk+1 ) ,
m+
k+1 = mk+1 + Kk+1 zk+1 nk+1 ,
zz
T
P+
k+1 = Pk+1 Kk+1 Pk+1 Kk+1 .
(4.136)
(4.137)
(4.138)
CHAPTER V
NONLINEAR SPACE MISSION ANALYSIS
5.1 Motivation
In this chapter, we apply the theoretical results derived previously and present examples and simulations of space mission problems. In these examples we implement our
nonlinear navigation results, the higher order solutions of relative dynamics to trajectory
and uncertainty propagations, statistical targeting, and higher order filtering.
The first example considered is an orbit about Europa in a Hill three-body formulation
(in the Jupiter-Europa system). Europa is one of the Jovian satellites, and currently, there is
a high interest in Europa by the science community as it is believed to have a vast reservoir
of water beneath its surface. This problem was originally motivated by the Jupiter Icy
Moon Orbiter (JIMO) mission (now the Europa Orbiter mission) which planned to study
the Jovian satellites Europa, Ganymede, and Callisto. We have specifically analyzed orbits
about Europa since its dynamical environment is more nonlinear than the other Jovian
satellites. The second problem is a Hohmann transfer orbit from the Earth to the Moon,
which is the simplest realization of an interplanetary trajectory. For example, considering
the Earth as the only gravitating body with uniform gravitational field, a cruise to the
Moon can be viewed as a Hohmann transfer orbit. The last problem is a halo orbit in the
Sun-Earth and Earth-Moon systems. A halo orbit is a periodic orbit found about Lagrange
111
112
points, and in the past, there have been a number of missions placed on halo orbits, e.g.,
the ISEE-3/ICE, SOHO, and Genesis missions. A halo orbit is a fundamentally fascinating
trajectory as it balances the perturbing forces of two massive bodies while preserving its
periodicity. Also, the halo orbit requires relatively little maintenance and has many useful
applications, such as solar activity monitoring station [27], a communication relay [24],
and a trajectory transit point [50, 52].
5.2 Nonlinear Propagation of Phase Volume

To show the effect of nonlinearity on a solution flow, this section presents a comparison
of linearly and nonlinearly propagated phase volumes.1 Consider the planar Hill threebody problem applied to the Jupiter-Europa system in Lagrangian form (see Appendix A
for the governing equations of motion and required constants). Suppose a spacecraft is
initially located slightly below the L2 point with the initial conditions:
x(t0 ) = [ x(t0 ), y(t0 ), u(t0 ), v(t0 ) ]T ,
r(t0 ) = [ x(t0 ), y(t0 ) ]T = [ 13581.17, 1321.85 ]T km,
v(t0 ) = [ u(t0 ), v(t0 ) ]T = [ 44.56297, 12.84600 ]T m/s.
Figure 5.1 shows the position and velocity flows of this initial state plotted over tf = 12
units of time ( 6.775 days). This is a safe trajectory since it does not collide with
Europa or escape from the system [80]. Such a trajectory is similar to a capture trajectory
at Europa. The boundary line of the shaded area in Figure 5.1(a) represents the zerovelocity curve2 with Jacobi constant, J = 2.15. Figure 5.2(a) shows the distance (||r||)
and speed (||v||) of the spacecraft as functions of time. The circled points are computed at
t {0, 0.881, 2.26, 4.42, 5.38, 5.74} days and will be considered later when we compare
1
2
Note that phase volumes are propagated deterministically assuming the initial state is perfectly known.
The zero-velocity curve is a boundary where a spacecraft with a given energy cannot cross.
113
x 10
1.5
ycoordinate (km)
0.5
L2
0.5
1.5
2
1.5
0.5
0.5
xcoordinate (km)
1.5
2
4
x 10
(a) Position plane.
1500
vcoordinate (m/s)
1000
500
500
1000
1500
2000
1500
1000
500
500
1000
1500
ucoordinate (m/s)
(b) Velocity plane.
Figure 5.1: Hill three-body trajectory plot at Europa for 6.775 days: circled points are
computed at t {0, 0.881, 2.26, 4.42, 5.38, 5.74} days.
114
the propagated phase volumes. We consider this trajectory to be the reference trajectory
and analyze its neighboring phase volume with the STTs.
In order to show the effect of the higher order STTs, first consider the following initial
deviation in x0 :
x0 = [ 0 km, 10 km, 0 m/s, 0.1 m/s ]T .
The STTs, up to fourth order, are integrated with respect to the nominal flow (t; x0 , t0 )
over the time interval [t0 , tf ]. Figure 5.2(b) shows the 1st-4th order STT effects as functions of time. That is, the norm of each order of STT solutions:
1st order effect = i,1 x01 ,
1 i,1 2 1 2
x0 x0 ,
2
1 i,1 2 3 1 2 3
3rd order effect =
x0 x0 x0 ,
6
1 i,1 2 3 4 1 2 3 4
4th order effect =
x0 x0 x0 x0 ,
24
2nd order effect =
so that the sums represent the higher order contributions to the relative solution x. It
shows that the higher order terms become larger than the lower order terms shortly after
6 days, which indicates that the higher order solutions diverge. This divergence of the
higher order solutions can be remedied by segmenting the orbit into piecewise trajectories;
however, that is not considered in this example.
Now consider an initial uncertainty distribution x0 N (0, P0 ), where
P0 = diag[ (10 km)2 , (10 km)2 , (0.1 m/s)2 , (0.1 m/s)2 ].
With this initial covariance matrix, we define the initial phase volume with respect to the
nominal initial state x0 as a 1- error ellipsoid:
B0 = {x0 | xT0 P1
0 x0 1},
115
||r || (km)
12000
10000
ref
8000
6000
4000
2000
0
3
4
Time (days)
3
4
Time (days)
1000
ref
||v || (m/s)
1500
500
(a) Distance and speed of the spacecraft plotted as functions of time.
10
10
Higher order effect
10
10
10
10
10
10
1st order effect

2nd order effect
3rd order effect
4th order effect
12
10
14
10
3
4
Time (days)
(b) Effect of higher order solutions
Figure 5.2: Trajectory norms and higher order solution magnitudes: circled points are
computed at t {0, 0.881, 2.26, 4.42, 5.38, 5.74} days.
116
or
x20
y20
u20
v20
+
+
+
1,
(10 km)2 (10 km)2 (0.1 m/s)2 (0.1 m/s)2
(5.1)
which is a four dimensional hyper-ellipsoid. From several simulations based on integrated

nonlinear trajectories, we find that the position errors in P0 dominate the outer boundary
solutions of the future phase volume in this example. For this reason, we analyze the flow
of a string from the 4-dimensional object such that:
x20
y20
+
= 1,
(10 km)2 (10 km)2
which represents a boundary solution of B0 by simply changing the inequality sign in Eqn.
(5.1) with an equality sign and by assuming v(t0 ) = 0, i.e., u0 = v0 = 0. In other
words, we study the evolution of a circular cross-section of the initial phase volume. The
projections onto the position and velocity planes are shown in Figures 5.3 and 5.4 for the
integrated3 and 1st-3rd order analytic solutions.4 We note that this is essentially the flow
of a string on the phase volume boundary. Several important facts can be observed from
these plots. First is that the convergence radius of the STTs varies with time. Specifically,
the convergence radius varies with the location of reference solution in phase space. As
an example, consider Figure 5.4(a). The linear phase curve solution computed at t = 4.42
days is closer to the true solution than the phase curve computed at t = 0.881 days (at
periapsis). One explanation for this is due to the difference in the higher order effects as
shown in Figure 5.2(b). Another observation is that the third order solutions are almost
identical to the true solution.
Another interesting observation is that the phase curve gets twisted along the trajectory.
Figures 5.3(b) and 5.4(b) show the twisting motion of the phase curve, which is shown by
3
Numerically integrated according to the nonlinear governing equations of motion.

Note that the 2nd and 3rd order solutions are sometimes overlapped with the integrated solution, which
indicates that they are good approximations of the true dynamics.
4
10
10
5
0
5
xcoordinate (km)
t = 2.26 days
10
ycoordinate (km)
400
200
0
200
400
100
0
100
xcoordinate (km)
t = 5.38 days
500
0
500
1000
1000 500
0
500
xcoordinate (km)
t = 0.881 days
1000
0
1000
200
ycoordinate (km)
t = 0 days
10
ycoordinate (km)
ycoordinate (km)
ycoordinate (km)
ycoordinate (km)
117
0
200
400
xcoordinate (km)
t = 4.42 days
20
0
20
200
0
200
xcoordinate (km)
t = 5.74 days
200
100
0
100
500
0
500
xcoordinate (km)
1000
0
50
200
0
200
x coordinate (km)
t = 4.415 days
40
20
0
20
200
0
200
x coordinate (km)
t = 4.426 days
20
0
20
200
0
200
x coordinate (km)
y coordinate (km)
50
t = 4.409 days
40
20
0
20
40
y coordinate (km)
t = 4.404 days
20
0
20
y coordinate (km)
y coordinate (km)
y coordinate (km)
y coordinate (km)
(a) Phase volume projected onto the position plane.
40
20
0
20
40
200
0
200
x coordinate (km)
t = 4.421days
200
0
200
x coordinate (km)
t = 4.432 days
200
0
200
x coordinate (km)
(b) Phase volume projected onto the position plane with smaller time increments.
Figure 5.3: Phase volume projections: solid line represents integrated, dotted line represents the 1st order, dash-dot line represents the 2nd order, and dashed line represents
the 3rd order solutions.
0
1
1
1
0
1
20
0
20
ucoordinate (m/s)
t = 5.38 days
40
20
0
20
40
60
80
100
0
100
ucoordinate (m/s)
t = 0.881 days
300
200
100
0
100
vcoordinate (m/s)
0.5
0
0.5
ucoordinate (m/s)
t = 2.26 days
vcoordinate (m/s)
t = 0 days
1
200
vcoordinate (m/s)
vcoordinate (m/s)
vcoordinate (m/s)
vcoordinate (m/s)
118
500
0
ucoordinate (m/s)
t = 4.42 days
500
20
0
20
5
0
5
ucoordinate (m/s)
t = 5.74 days
10
5
0
5
ucoordinate (m/s)
50
0
50
20
0
20
2
0
2
ucoordinate (m/s)
t = 4.449 days
20
0
20
0.5
0
0.5
1
ucoordinate (m/s)
t = 4.46 days
1.5
20
0
20
2
0
2
ucoordinate (m/s)
vcoordinate (m/s)
vcoordinate (m/s)
t = 4.438 days
vcoordinate (m/s)
vcoordinate (m/s)
vcoordinate (m/s)
vcoordinate (m/s)
(a) Phase volume projected onto the velocity plane
t = 4.443 days
20
0
20
1
0
1
ucoordinate (m/s)
t = 4.455 days
20
0
20
2
1
0
1
2
ucoordinate (m/s)
t = 4.466 days
20
0
20
4
2
0
2
4
ucoordinate (m/s)
(b) Phase volume projected onto the velocity plane with smaller time increments.
Figure 5.4: Phase volume projections: solid line represents integrated, dotted line represents the 1st order, dash-dot line represents the 2nd order, and dashed line represents
the 3rd order solutions.
119
t = 0 days
t = 2.26 days
y coordinate (km)
y coordinate (km)
10
5
0
5
5
0
5
x coordinate (km)
t = 4.42 days
30
20
10
0
10
20
200
0
200
400
10
100
0
100
x coordinate (km)
t = 5.74 days
y coordinate (km)
y coordinate (km)
10
10
400
200
100
0
100
200
0
200
x coordinate (km)
500
0
500
x coordinate (km)
1000
(a) Phase volume projected onto the position plane for different initial outer
boundaries using the 3rd order solution.
t = 2.258 days
vcoordinate (m/s)
vcoordinate (m/s)
t = 0 days
0.05
0
0.05
1
0
1
20
20
0
20
ucoordinate (m/s)
t = 5.742 days
vcoordinate (m/s)
vcoordinate (m/s)
0.05
0
0.05
ucoordinate (m/s)
t = 4.421 days
10
0
10
20
50
0
50
30
5
0
5
ucoordinate (m/s)
10
5
0
5
ucoordinate (m/s)
(b) Phase volume projected onto the velocity plane for different initial outer
boundaries using the 3rd order solution.
Figure 5.5: Phase volume projections.
120
the crossing motion of the phase curve. This is interesting since the linear solution only
predicts whether the phase curve gets stretched or contracted, but cannot lead to such a
trajectory. The points cross each other when projected to the position and velocity planes;
however, they are not intersecting in the actual 4-dimensional space. This is similar to a
twisted ball projected onto a plane.
Figures 5.5(a) and 5.5(b) show the projection of the phase volume onto the position and
velocity planes using the third order STT solutions while varying the initial conditions.
The initial conditions for the contour plots are computed by solving Eqn. (5.1) while
p
varying kr0 k = x20 + y20 from 0 km to 10 km with 0.2 km increment. The outer
boundary solutions are the same as in Figures 5.3(a) and 5.4(a) (except for the first plot
in Figure 5.5(b)), where kr0 k = 10 km resulting u0 = v0 = 0 m/s and the inner
p
most solutions correspond to kr0 k = 0 km resulting kv0 k = u20 + v20 = 0.1 m/s. In
addition to the twisting motion of the phase volume, the non-concentric distribution shown
in the contour plot is an interesting result.
5.3 Nonlinear Orbit Uncertainty Propagation

In this section, we present the simulations of the nonlinear orbit uncertainty propagation discussed in 3.9, where we compare linear and nonlinear approaches. The first
example is based on the planar two-body problem where a spacecraft is on the Earthto-Moon Hohmann transfer. The second example is based on the planar Hill three-body
problem in Jupiter-Europa system.
Figure 5.6 shows a Hohmann transfer from near Earth (20, 000 km) to the Moon
(384, 400 km). The reference trajectory is propagated for the transfer period ( 5.24
days) and the initial statistics are assumed to be Gaussian with a zero mean and position
121
x 10
2
1.5
ycoordinate (km)
0.5
Earth
0.5
0.5
0.5
1.5
2.5
3.5
xcoordinate (km)
4
5
x 10
Figure 5.6: Two-body problem: Hohmann transfer trajectory.
x 10
3
ycoordinate (km)
Linearly
Propagated
Mean
Nonlinearly
Propagated
Means
1
~2400 km
m=1
m=2
m=3
m=4
MonteCarlo
3
3.5
3.6
3.7
3.8
3.9
xcoordinate (km)
4.1
4.2
5
x 10
Figure 5.7: Two-body problem: comparison of the computed mean and covariance at
apoapsis using STT-approach and Monte-Carlo simulations.
122
Table 5.1: Local nonlinearity index.

m
1
2
3
4
Two-Body Hill Three-Body

1.06
3.57
0.04
0.29
0.007
0.28
0.001
0.06
uncertainty of 100 km and velocity uncertainty of 0.1 m/s:

P0 = diag[ (100 km)2 , (100 km)2 , (0.1 m/s)2 , (0.1 m/s)2 ].
With this initial distribution, the initial confidence region is defined with respect to the
initial state as:
B0 = {x0 | xT0 P1
0 x0 1}.
To show the effect of nonlinearity on the relative motion, Table 5.1 shows the local nonlinearity index computed at the apoapsis. When computing m , eight sample points corresponding to the eigenvectors of the initial error ellipsoid are considered. As predicted
by the local nonlinearity index, for this Hohmann transfer example, the higher order series
are convergent and the second order solution (i.e., 2 ) provides results superior to the first
order solution (i.e., 1 ).
The initial uncertainties are then propagated to the apoapsis using the higher order
STT approach and the Monte-Carlo simulation. Figure 5.7 shows the propagated mean
and 1- error ellipsoid plotted with respect to the target state (i.e, apoapsis) where the
Monte-Carlo result is based on an ensemble of 106 sample points. The second and higher
order solutions are overlapped with the Monte-Carlo solution, which indicates that they
provide far more accurate estimates of the mean and the dispersion of the samples (i.e.,
projection of the covariance matrix) than the linear (m = 1) solution. Also, note that the
123
linear method predicts that the mean is located at the mode of the propagated distribution
(or the deterministic solution flow) whereas the true mean is deviated from the mode by
2400 km. This shows that incorporating the system nonlinearity provides a more accurate
description of the propagated probability distribution.
5.3.2 Hill Three-Body Problem: about Europa
Consider the same planar Hill three-body problem discussed in 5.2. The reference
trajectory in non-dimensional coordinates is shown in Figure 5.8, where the final position
is located at a periapsis. The initial conditions used are:
r(t0 ) = [ 0.69010031015662 0.06716709529872 ],
v(t0 ) = [ 0.11045639526249 0.03184084790390 ],
which are the same as in 5.2, but given in non-dimensional units. The initial state is
assumed to be Gaussian with zero mean and a diagonal covariance matrix with position
error of 10 km (5.1 104 in normalized units) and velocity error of 0.1 m/s (2.5 104
in normalized units):
P0 = diag[ (10 km)2 , (10 km)2 , (0.1 m/s)2 , (0.1 m/s)2 ],
which defines the initial confidence region as B0 = {x0 | xT0 P1
0 x0 1}. The local nonlinearity indices for this problem are computed at periapsis and are given in Table 5.1. The
result shows that the fourth order solution provides accuracy better than 10 percent error.
Note that the level of accuracy can be improved when a different final time is considered
since the strongest nonlinearity is at periapsis.
As in the two-body problem case, the initial uncertainties are propagated using the
higher order STT approach and a Monte-Carlo simulation based on an ensemble of 106
initial samples. Figure 5.9 shows the propagated mean and 1- covariance matrix plotted
124
0.3
ycoordinate
0.2
0.1
Europa
t0
0.1
0.2
0.1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
xcoordinate
Figure 5.8: Hill three-body problem: a safe trajectory at Europa.
0.07
0.06
Linearly
Propagated
Mean
0.05
ycoordinate
0.04
0.03
Nonlinearly
Propagated
Means
0.02
0.01
0
~0.0081
0.01
m=1
m=2
m=3
m=4
MonteCarlo
0.02
0.03
0.14
0.12
0.1
0.08
0.06
0.04
0.02
xcoordinate
Figure 5.9: Hill three-body problem: comparison of the computed mean and covariance at
periapsis using STT-approach and Monte-Carlo simulations.
125
with respect to the periapsis. The result shows that the linear (first order) solution captures
the semi-major axis of the true covariance projection (i.e., Monte-Carlo solution) and the
semi-minor axis depends on the deviated mean. Also, the linearly propagated mean (i.e.,
mode) is offset from the mean of the true probability distribution by 0.0081 160 km
and lies inside the true covariance projection. As a result, it is clear that the higher order
solution provides a superior result as compared to the linear case. We can see this from
the location of the deviated mean and the covariance projection of the Monte-Carlo result.
This is important since a more conservative and statistically more accurate error bound
can be computed. Note that, assuming the STT series are convergent, we can analytically
propagate different values of the initial mean and covariance matrix without any additional
numerical integrations.
5.4 Nonlinear Statistical Targeting

In this section, we present simulations of the nonlinear statistical targeting discussed in
4.2 where we consider the two-body Hohmann transfer and the Hill three-body problem
discussed in the previous section, 5.3.
In the two-body case, we assume that the 3rd order solution is considered to be the
truth model (i.e., solution based on the 3rd order STT series) as the local nonlinearity
index predicts that it captures most nonlinear effects. Thus, the Vk computed using the
STT-based nonlinear statistical targeting method satisfies the statistically correct trajectory
concept. To simulate nonlinear statistical targeting, we first solve for the 3rd order solution
of the entire Hohmann transfer (t0 to tf ) and compute the deviated mean for the entire
trajectory. Note that the deviated mean is the difference between the mean computed
using the 3rd order solution and the reference trajectory. We then re-solve for the 3rd order
126
10
Linear
Nonlinear
2
10
10
||V || (m/s)
10
10
10
10
20
40
60
80
100
120
140
160
180
(degrees)
Figure 5.10: Two-body problem: computed Vk using the linear and nonlinear methods.
solutions from the deviated mean computed at tk to tf (tk t0 ) and find the correction Vk
using Newtons method as discussed in 4.2. We can avoid integrating the STT solutions
for each tk by applying the series reversion discussed in 2.2; however, this is not used in
this particular example. At every instance of a Vk execution, the state is assumed to be
Gaussian with a zero mean and position and velocity uncertainties of 100 km and 0.1 m/s,
respectively, which is the same as the initial errors.
Figure 5.10 shows the magnitude of Vk corrections computed using the linear and
nonlinear methods as functions of the true anomaly value which corresponds to the time tk
the impulse is applied. Both solutions become essentially the same after 90 degrees;
indicating small nonlinearity for the later part of the transfer. As expected, the linear
solution Vk grows as the correction maneuver is made at a later time due to a larger
deviation in the mean trajectory. We note that the nonlinear statistical targeting method
gives an optimal time (i.e., minimum kVk k) to perform a correction maneuver that targets
127
||mr(t )|| (km)
10
Linear Solutions
10
10
10
10
10
10
20
40
60
80
100
120
140
160
180
120
140
160
180
(degrees)
2
||mv(t )|| (m/s)
10
10
10
10
Linear
Nonlinear
10
20
40
60
80
100
(degrees)
Figure 5.11: Two-body problem: deviated position and velocity means at the target.
Target
MC mean based on linear correction
MC mean based on nonlinear correction
600
400
ycoordinate (km)
~166 km
200
0
200
MC 1 covariance
projection based on
nonlinear correction
MC 1 covariance
projection based on
linear correction
400
600
800
2000
1500
1000
500
500
1000
1500
xcoordinate (km)
Figure 5.12: Two-body problem: Monte-Carlo simulation using the linear and nonlinear
methods.
128
the position mean. Figure 5.11 shows the deviated position and velocity means at the
target. The nonlinear deviated position mean is not shown since it is zero by definition.
When the correction maneuver is made at an early stage of the trajectory using the linear
theory, the position mean may deviate quite a bit at the final target. For the deviation in
the velocity mean, the difference between the linear and nonlinear solutions is very small.
Lastly, for a verification purpose, Figure 5.12 shows Monte-Carlo simulations of the
linearly and nonlinearly corrected trajectories, where the correction maneuvers are performed at = 90 degrees. Both cases are based on ensembles of 106 sample points. We
observe that, on average, the nonlinear correction maneuver computed using the 3rd order
STT approach intercepts the target whereas the linear correction results in a deviation of
166 km. Hence, this confirms that our nonlinear analytic targeting approach gives a
correction maneuver that satisfies the statistically correct trajectory.
5.4.2 Hill Three-Body Problem: about Europa
Similar to the two-body example, in the Hill three-body problem example, we assume
the uncertainties computed using the 4th order STT solution is the truth model. Once the
mean trajectory is computed for the time interval [t0 , tf ], at every tk [t0 , tf ], we compute
Vk using the nonlinear and linear methods. At every time of maneuver, we assume the
spacecraft state has the same Gaussian statistics as the initial state, i.e., zero mean and
a covariance matrix with 10 km uncertainties and 0.1 m/s uncertainties for position and
velocity components, respectively.
Figure 5.13 shows the magnitude of Vk applied at tk [t0 , tf ] that are solved using
both the linear and nonlinear methods. The correction maneuvers are plotted as functions
of time and we observe little nonlinearity from tk 15 hours onwards. In this case,
however, Vk fluctuates around tk 4 hours in both cases, unlike the two-body case. This
129
10
10
10
||V || (m/s)
10
10
10
10
Linear
Nonlinear
5
10
10
15
20
Time (hours)
Figure 5.13: Hill three-body problem: computed Vk using the linear and nonlinear methods.
is due to a sudden change in the velocity direction, and hence, the system nonlinearity is
varied. There is also an optimal Vk (i.e., minimum kVk k), which occurs around tk 15
hours. Figure 5.14 shows the deviated position and velocity means. As in the two-body
case, when the correction maneuver is made at an early stage of the trajectory using the
linear theory, the position mean may deviate noticeably. The overall difference between
the linear and nonlinearly computed velocity mean is very small. A high fluctuation around
tk 4 in the nonlinearly solved mv (tf ) is due to the higher correction maneuvers in that
time frame.
As in the two-body case, Figure 5.15 shows the Monte-Carlo simulation results for
the correction maneuvers computed, at time tk = 3 hours, using the linear and nonlinear
methods. The result shows that linearly corrected mean results in a deviation of 60 km
whereas the nonlinear correction maneuver intercepts the target on average.
130
||mr(t )|| (km)
10
Linear Solutions
2
10
10
10
10
10
15
20
Time (hours)
4
||mv(t )|| (m/s)
10
Linear
Nonlinear
10
10
10
10
10
10
15
20
Time (hours)
Figure 5.14: Hill three-body problem: deviated position and velocity means at the target.
800
Target
MC using linear correction
MC using nonlinear correction
600
ycoordinate (km)
400
200
MC 1 covariance
projection based on
nonlinear correction
~60 km
0
200
400
MC 1 covariance
projection based on
linear correction
600
800
200
150
100
50
50
100
150
200
250
xcoordinate (km)
Figure 5.15: Hill three-body problem: Monte-Carlo simulation using the linear and nonlinear methods.
131
Table 5.2: Halo orbit maximum amplitudes with respect to the Sun-Earth L1 point.
Cases
1
2
Ax (km) Ay (km)
245924 668228
246069 668416
Az (km)
137908
139015
5.5 Nonlinear Trajectory Navigation

In this section, we present simulations of the nonlinear Kalman filters discussed in
4.4. We give several examples based on halo orbits of the Sun-Earth and Earth-Moon
systems [22, 25, 36, 52]. A halo orbit is a Lissajous-type periodic orbit where the inplane and out-of-plane frequencies are the same, which we compute based on the circular
restricted three-body problem (see Appendix A for details).
5.5.1 Halo Orbit: Sun-Earth System
Consider a halo orbit about the Sun-Earth L1 point in a non-dimensionalized frame,
which can be dimensionalized by applying the length scale of ` = 1 AU = 1.49597870691
108 km, where AU stands for astronomical unit, and time scale of = 1/E , where E
p
is the mean motion of the Earth about the Sun (i.e., S /AU 3 = 1.991 107 s1 ). Figures 5.16 and 5.17 show the reference (nominal) trajectory for one orbital period (177.86
days) in 3-dimensions and in the x-y plane, respectively, which corresponds to the Case 1
given in Table 5.2. The initial conditions for these orbits are (in non-dimensional units):
rcase1 (t0 ) = [ 0.988884102845168, 0.0, 0.000921858528329094 ]T ,
vcase1 (t0 ) = [ 0.0, 0.00893471471659142, 0.0 ]T ,
rcase2 (t0 ) = [ 0.98888423093423, 0.0, 0.000929261736280955 ]T ,
vcase2 (t0 ) = [ 0.0, 0.00893688204973967, 0.0 ]T .
132
zcoordinate (km)
x 10
1
L1
0
1
6
4
5
x 10
2
0
ycoordinate (km) 2
4
6
1.48
1.483
1.482
8
1.481
x 10
xcoordinate (km)
Figure 5.16: Nominal halo orbit about the Sun-Earth L1 point.
x 10
L1
6
ycoordinate (km)
4
2
0
2
4
6
8
1.472 1.474 1.476 1.478 1.48 1.482 1.484 1.486 1.488 1.49
xcoordinate (km)
x 10
Figure 5.17: Nominal halo orbit about the Sun-Earth L1 point in x-y plane.
133
For the measurement model, we assume a simple linear model where only the ycoordinate is observed, i.e.,
zk = yk + vk ,
(5.2)
where yk represents the vertical position component of the state vector and vk represents
the measurement error. This measurement model can be viewed as a range measurement
obtained by optical imaging of the Earth relative to distant stars or a Very Long Baseline
Interferometry (VLBI) measurement. The measurement noise is assumed to be 0.1 m for
each range measurement. This linear assumption simplifies the problem a great deal since
the measurement sensitivity does not require the computation of the higher order partials.
This way, it is easier to understand the effect of the nonlinear orbit uncertainty propagation
on filter performance.
Initially, the spacecraft state is assumed to be a zero mean Gaussian with position
uncertainties of 100 km and velocity uncertainties of 0.1 m/s.5 The initial mean and covariance matrix are mapped using the STT approach for m = {1, 3}, unscented transformation, and Monte-Carlo simulations based on 106 sample points. Figure 5.18 shows
the mean and the projection of the 1- covariance matrix onto the x-y plane after one orbital period. Assuming the MC simulation is the true solution, the result shows that the
3rd order solution is the most accurate approximation, whereas the linear solution fails to
characterize the orbit uncertainty distribution.
We now consider the same initial uncertainties, but assume the initial guess (mean) is
off by 100 km for the position components and 0.1 m/s for the velocity components so
that they lie on the boundary of the initial 1- ellipsoid. A set of pseudo-measurements
are computed based on the reference trajectory with a 20-day increment. Using the same
5
The initial covariance matrix is a diagonal matrix with (100 km)2 and (0.1 m/s)2 for position components
and velocity components, respectively.
134
x 10
m=1
m=3
UT
MC
ycoordinate (km)
4
2
0
2
4
6
8
1.477 1.4775 1.478 1.4785 1.479 1.4795
1.48
1.4805 1.481 1.4815
xcoordinate (km)
x 10
(a) Projected onto the x-y plane.
x 10
2.5
m=1
m=3
UT
MC
ycoordinate (km)
1.5
1
0.5
0
0.5
1
1.5
1.479
1.4791
1.4792
1.4793
1.4794
1.4795
xcoordinate (km)
1.4796
1.4797
8
x 10
(b) Larger plot of (a).
Figure 5.18: Sun-Earth halo orbit: covariance matrix computed after one orbital period.
135
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
(m/s)
10
10
(km)
10
10
10
10
2
10
10
10
100
200
300
Time (days)
10
400
200
Time (days)
400
Figure 5.19: Sun-Earth halo orbit: comparison of the uncertainties computed using the
EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 20
days.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
||Velocity Error|| (m/s)
||Position Error|| (km)
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
100
200
300
Time (days)
400
10
100
200
300
Time (days)
400
Figure 5.20: Sun-Earth halo orbit: comparison of the absolute errors computed using the
days.
136
measurements, the initial mean and covariance matrix are mapped and solved using the
EKF, UKF, 3rd order HNEKF, and 3rd order HAEKF. For the HAEKF, since the trajectory is periodic, the STTs are computed and stored for only one orbital period, which is
divided into two segments for numerical consistency, and reversion of series is applied to
map states analytically. Figure 5.19 shows the a priori and a posteriori position and ve
locity root-sum-square errors, where R = xx + yy + zz and V = uu + vv + ww ,

and ii represents (i, i) component of the covariance matrix. A sudden drop in the uncertainties right after 100 days is due to the fact that the initial covariance matrix is quite large
and requires at least six independent measurements to obtain a well-defined (i.e., reduced
to the measurement noise level in all directions) a posteriori covariance matrix. The result
shows that the EKF overestimates the solution accuracy (i.e., the uncertainties are smaller
than they are in actuality) while the UKF, HNEKF, and HAEKF provide conservative uncertainty estimates. Figure 5.20 shows the magnitude of the absolute position and velocity
errors, i.e., the magnitude of the difference between the updated mean and the true state.
The result shows that the EKF does not perform well as compared to the higher order filters. This clearly explains the importance of nonlinear orbit uncertainty propagation. The
covariance matrix computed by using the first order method (i.e., EKF) overestimates the
solution accuracy, and hence, the residual is trusted less. On the other hand, the UKF and
the higher order filters predict more conservative uncertainties and more effectively balance the a priori uncertainties and the actual measurements (i.e., measurements are valued
more than the a priori information in this case). Figures 5.21 and 5.22 are based on the
same filter setup except that the measurements are updated every 5 days. It shows that
there is not much difference in the propagated uncertainties, but the absolute errors are
computed more accurately in UKF and higher order filter runs.
Figures 5.23 and 5.24 show the HNEKF results for cases m {1, 2, 3}. As mentioned
137
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
(m/s)
10
10
(km)
10
10
10
10
10
10
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
100
200
300
Time (days)
10
400
200
Time (days)
400
days.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
100
200
300
Time (days)
400
10
100
200
300
Time (days)
400
days.
138
10
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
10
(m/s)
10
10
(km)
10
10
10
10
2
10
10
10
100
200
300
Time (days)
10
400
200
Time (days)
400
HNEKFs for the cases m = {1, 2, 3}. Measurements are taken every 20 days.
10
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
1
2
10
10
10
10
10
10
10
10
10
10
10
10
10
10
100
200
300
Time (days)
400
10
100
200
300
Time (days)
400
139
10
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
10
(m/s)
10
10
(km)
10
10
10
10
2
10
10
10
100
200
300
Time (days)
10
400
200
Time (days)
400
HAEKFs for the cases m = {1, 2, 3}. Measurements are taken every 20 days.
10
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
1
2
10
10
10
10
10
10
10
10
10
10
10
10
10
10
100
200
300
Time (days)
400
10
100
200
300
Time (days)
400
140
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
(m/s)
10
10
(km)
10
10
10
10
2
10
10
10
100
200
300
Time (days)
10
400
200
Time (days)
400
EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements are taken every 20
days based on the halo orbit Case 2.
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
100
200
300
Time (days)
400
10
100
200
300
Time (days)
400
EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements are taken every 20 days
based on the halo orbit Case 2.
141
earlier, note that the case m = 1 is identical to the EKF formulation. The result shows that
the higher order filters, m {2, 3}, provide superior filter performance over the first order
case and it is observed that the second order effect contains most of the system nonlinearity,
indicating that the second order filter is sufficient for an accurate nonlinear filter in our
example. Figure 5.25 and 5.26 show the HAEKF uncertainties and absolute error plots,
respectively, for m {1, 2, 3}. The uncertainties for m = 1 are similar to the EKF solution
and for m = 2 are similar to the case m = 3 as shown in Figure 5.19. The absolute error
plot shows that all three filters provide good estimation performance even for the case
m = 1. This is expected since the pseudo-measurements are computed based on the
reference trajectory which the STTs are computed based on. In other words, the reference
trajectory can be thought of as a regression solution for the simulated measurements.
In order to analyze the higher order effect, the pseudo-measurements are now generated from the Case 2 halo orbit given in Table 5.2. Figures 5.27 and 5.28 show the
simulated filter solutions. The results show that the higher order solutions are superior
over the linear filters, i.e., EKF and HAEKF for m = 1. As expected, this indicates that
the linear Kalman filter is only feasible when the reference trajectory is sufficiently close
to the true trajectory. The HAEKFs for m > 1, however, have more flexibility in the reference trajectory. The overall filter convergence is slightly slower than the previous cases
since the initial mean is assumed to be the same as in the previous cases, and thus, it is
farther away from the true trajectory (i.e., the trajectory which the pseudo-measurements
are generated).
5.5.2 Halo Orbit: Earth-Moon System
As another example of nonlinear filtering, we present a similar simulation as in 5.5.1
based on a halo orbit about the Earth-Moon L1 point. This is a much more nonlinear
142
Table 5.3: Halo orbit maximum amplitudes with respect to the Earth-Moon L1 point.
Cases
1
2
Ax (km) Ay (km)
6934
21612
6968
21667
Az (km)
2132
2665
system since the mass ratio constant is much larger, i.e., EM = 0.01215 > SE =
3.003 106 . For this system, the length scale is ` = 384400 km and the time scale
is = 4.3691 days. Figures 5.29 and 5.30 show the reference trajectory for one orbital
period (11.99 days), which is the Case 1 given in Table 5.3. Note that this orbit has
a much smaller orbit size as well as orbital period than the Sun-Earth halo orbit. The
corresponding initial conditions are (in non-dimensional units):
rcase1 (t0 ) = [ 0.823386040115578, 0.0, 0.00554618294369076 ]T ,
vcase1 (t0 ) = [ 0.0, 0.126839229387039, 0.0 ]T ,
rcase2 (t0 ) = [ 0.823384852005846, 0.0, 0.00693385322129738 ]T ,
vcase2 (t0 ) = [ 0.0, 0.127125596330242, 0 ]T .
Also, we assume the same measurement model considered in 5.5.1, i.e., zk = yk + vk ,
with 0.1 m accuracy.
At epoch, the initial state is assumed to be Gaussian with:
m+
= [ 100 km, 100 km, 100 km, 0.1 m/s, 0.1 m/s, 0.1 m/s ]T ,
0
P+
= diag[ (100 km)2 , (100 km)2 , (100 km)2
0
(0.1 m/s)2 , (0.1 m/s)2 , (0.1 m/s)2 ],
and we assume the measurements are taken every 2 days, which we consider to be a
baseline case. Note that the mean vector m+
0 lies on the boundary of the 1- ellipsoid,
which simply means that the initial guess is offset from the initial state of the true trajectory. This navigation scenario is simulated using the EKF, UKF, 3rd order HNEKF,
zcoordinate (km)
143
2000
0
2000
2.5
L1
2
1.5
1
4
x 10
0.5
0
0.5
ycoordinate (km)
1
3.3
1.5
3.25
3.2
2.5
x 10
xcoordinate (km)
Figure 5.29: Nominal halo orbit about the Earth-Moon L1 point.
2.5
x 10
L1
2
1.5
ycoordinate (km)
1
0.5
0
0.5
1
1.5
2
2.5
3.1
3.2
3.3
xcoordinate (km)
3.4
3.5
5
x 10
Figure 5.30: Nominal halo orbit about the Earth-Moon L1 point in x-y plane.
144
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
Figure 5.31: Earth-Moon halo orbit: comparison of the uncertainties computed using the
days.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
20
Time (days)
30
10
10
20
Time (days)
30
Figure 5.32: Earth-Moon halo orbit: comparison of the absolute errors computed using
the EKF, UKF, HNEKF (m = 3), and HAEKF (m = 3). Measurements are taken every 2
days.
145
and 3rd order HAEKF,6 and Figure 5.31 shows the a priori and a posteriori position and
velocity root-sum-squares errors (i.e., R and V ) and we observe that the UKF, HNEKF,
and HAEKF provide conservative uncertainty estimates whereas EKF overestimates the
solution accuracy. Figure 5.32 shows the absolute errors of the position and velocity of
k k, where xk represents our reference halo orbit which the
the trajectory, i.e., km+
k x
pseudo-measurements are generated about. The result shows that the EKF provides a slow
convergence as compared to the higher order filters. Again, this is because the EKF overestimates the solution accuracy whereas the higher order filters compute the Kalman gain
more effectively by balancing the a priori state error and the measurement residual.
Figures 5.33 and 5.34 are based on the exact same filter setup except that the initial
mean is assumed to be on the true trajectory, i.e., m+
0 = 0. In this case, the result shows
that the EKF provides a faster solution convergence than the higher order filters. This is
expected since the propagated initial mean of the EKF is the true trajectory and can be
considered as an already converged solution, and thus, the EKF only improves the covariance matrix.7 The higher order filters, on the other hand, propagate the mean nonlinearly,
which deviates from the reference trajectory because of the large initial uncertainties. The
higher order filters then try to simultaneously reduce the errors in the state and the covariance matrix, which requires additional measurements to obtain a converged solution.
Therefore, the EKF solution is, in a sense, a false result which is obtained by disregarding
the system nonlinearity. To make this case more realistic, the initial uncertainties (i.e.,
6
For HAEKF runs, as in the Sun-Earth case, the STTs are computed and stored for only one orbital
period since the orbit is periodic; the orbit is divided into two segments and integrated independently for
numerical consistency.
7
If started out with small initial uncertainties and large initial deviations from the true state, a linear filter
solution usually does not converge since the filter trusts the a priori covariance strictly and does not include
the measurement (residual) contribution in the update process. If started out with large initial uncertainties
and also large initial deviations from the true state, but within the initial error bounds, a linear filter solution
usually converges since the filter has enough flexibility to adjust so that the true solution can be found from
sequential updates.
146
P+
0 ) should be reduced to a level where the linear assumption gives a good approximation
of the dynamics. This is because we want to show that when a linear approximation is
sufficient, both nonlinear filters and EKF should result in similar filter solutions. This case
is shown in Figures 5.35 and 5.36, where the initial state is assumed to have zero mean
with a covariance matrix:
= diag[ (1 km)2 (1 km)2 , (1 km)2 (0.01 m/s)2 (0.01 m/s)2 (0.01 m/s)2 ].
P+
0
We observe that both linear and nonlinear filters provide the same level of filter performance.
Now lets consider the baseline case with different measurement scheduling. Figures
5.37 and 5.38 show the uncertainties and absolute errors where the measurements are now
updated every 6 hours (n.b., 8 times more frequently than the baseline case). Overall, the
result shows that both linear and nonlinear filters provide essentially the same filter output
since measurements are taken frequently enough to maintain the phase volume within a
linear boundary.
Figures 5.39 and 5.40 show uncertainties and absolute errors based on the HNEKF
result for cases m {1, 2, 3}. As in the Sun-Earth halo orbit case, the higher order
filters, m {2, 3}, provide superior filter performance over the EKF, and we observe
that the second order solution (i.e., m = 2) captures most of the trajectory dynamics in
this case. Figures 5.41 and 5.42 show the HAEKF uncertainties and absolute error plots,
respectively, for m {1, 2, 3}. The uncertainties for m = 1 are similar to the EKF
solution and for m = 2 are similar to the case m = 3 as shown in Figure 5.31.
The absolute error plot shows that all three filters provide good filter solutions and
convergence even for the first order case, despite that fact the initial state, m+
0 , is not
within the linear regime. This is expected since the m = 1 case is, in a sense, similar to
147
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
days assuming zero initial mean.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
4
10
10
5
10
10
20
Time (days)
30
10
10
20
Time (days)
30
days assuming zero initial mean.
148
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
2
10
(m/s)
10
(km)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
days assuming zero initial mean and small initial covariance matrix.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
0
10
20
Time (days)
30
10
20
Time (days)
30
days assuming zero initial mean and small initial covariance matrix.
149
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
hours.
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
EKF
UKF
HNEKF (m=3)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
20
Time (days)
30
10
10
20
Time (days)
30
hours.
150
10
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
10
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
10
HNEKF (m=1)
HNEKF (m=2)
HNEKF (m=3)
10
10
20
Time (days)
30
10
10
20
Time (days)
30
the HNEKFs for the cases m = {1, 2, 3}. Measurements are taken every 2 days.
151
10
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
10
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
10
HAEKF (m=1)
HAEKF (m=2)
HAEKF (m=3)
10
10
20
Time (days)
30
10
10
20
Time (days)
30
the HAEKFs for the cases m = {1, 2, 3}. Measurements are taken every 2 days.
152
the EKF run of the previous case where we assumed the initial mean is zero, and thus, the
solution is false since the filter completely ignores the system nonlinearity. To be more
specific, because the reference trajectory is the true trajectory,8 the solution converges no
matter how offset the initial state is from the reference trajectory. We shall see that this
is no longer true if the pseudo-measurements are generated based on a trajectory that is
different than the reference trajectory.
In order to show this, the pseudo-measurements are generated from the Case 2 halo
orbit given in Table 5.2 while keeping the Case 1 halo orbit as the reference trajectory.
Figures 5.43 and 5.44 show that the higher order solutions are superior over the linear
filters, i.e., EKF and HAEKF for m = 1. As in the Sun-Earth simulation, we observe
that the linear Kalman filter (i.e., first order HAEKF) solution does not converge, which
implies that the first order HAEKF is only feasible when the reference trajectory is sufficiently close to the true trajectory. We note that the HAEKFs for m > 1 incorporates
the system nonlinearity and provide solution essentially equivalent to the numerical higher
order filters.
5.5.3 Potential Applications and Challenges
In this study, the EKF required integration of N + N 2 = 42 equations and the UKF
required integration of (2N + 1)N = 78 equations between each measurement update,
and in the actual filter runs, the EKF was slightly faster than the UKF. The HNEKFs
for m > 1 provided superior results over the linear filters; however, we note that the
computational load increases significantly as m increases. For example, the third order
HNEKF requires integration of 1554 equations for STT computation. On the other hand,
the HAEKF does not require any integration in the the actual filtering process. The most
8
In other words, the reference trajectory is the reconstructed (i.e., a regression solution) solution of the
simulated data.
153
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
10
(m/s)
10
10
10
(km)
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
10
10
10
10
20
Time (days)
10
30
10
20
Time (days)
30
EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements are taken every 2 days
based on the halo orbit Case 2.
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
10
10
10
10
10
10
10
10
10
10
10
10
EKF
UKF
HAEKF (m=1)
HAEKF (m=3)
10
10
20
Time (days)
30
10
10
20
Time (days)
30
the EKF, UKF, and HAEKFs for the cases m = {1, 3}. Measurements are taken every 2
days based on the halo orbit Case 2.
154
expensive numerical operation in the HAEKF is the higher order moment computation;
however, we note that there exist various techniques for efficient computation of moments.
The examples presented are filter initialization problems, where an initial state is assumed to have a large covariance matrix and has a poor initial guess that may be far from
the true trajectory. The goal is then to sequentially find a converged solution with accurate covariances. Our examples show that a nonlinear filter can have more flexibility in
the initial guess and obtain a converged solution whereas a linear filter diverges when the
initial guess is far from the true trajectory. Also, it is apparent that a nonlinear filter can
provide a more accurate filter solution than a linear filter even when the initial state is
precisely known. For example, when a trajectory is propagated over a long time period
or when there exists strong system nonlinearity, the number of measurements to maintain
a certain error limit can be reduced since a nonlinear filter provides a faster convergence
than a linear filter. Also, the integration between the dynamics and filtering can be more
effectively modeled.
For a given reference trajectory, we have shown that the higher-order analytic EKF is
essentially equivalent to higher order numerical filters (i.e., HNEKF and UKF) if we consider a sufficient order of solutions and assume the series is within the radius of convergence. Hence, for missions with pre-determined reference trajectories, the HAEKF may
be suitable for the trajectory navigation while obtaining faster convergence and a computationally faster filter9 than the EKF. Such applications of the HAEKF consist of interplanetary cruises, trajectories about complex dynamical environments, spacecraft launches,
orbit insertions, atmospheric re-entries, etc.
We have shown examples where the the trajectory dynamics are assumed to be known
with perfect knowledge by setting the process noise to be zero. This is a valid assumption if
9
The HAEKF process is faster than the EKF or other numerical filters since the higher order solutions
can be stored onboard before the launch or up-linked to a spacecraft during the mission.
155
a spacecraft is quiet and thrusters are turned off. However, this not true for some problems,
such as spacecraft launch or atmospheric re-entry since the un-modeled accelerations (e.g.,
rocket thrusters or atmospheric drag) can be significant. In practice, a covariance matrix
is usually integrated including a process noise matrix according to the Riccati equation.10
This problem is not considered in this dissertation since the focus of our study is to show
the importance of the system nonlinearity on a filtering problem. However, we note that, in
discrete form, the process noise matrix can be estimated by integrating it over some time
interval [55, 56], which is simple if we assume an additive white noise. This, however, can
be a difficult process if we consider a nonlinear stochastic acceleration and apply no linear
assumption.
In general the differential equation for the covariance matrix can be written as: P(t)
= A(t)P(t) +
T
P(t)A (t) + Q(t), where A(t) represents the linear dynamics matrix.
10
CHAPTER VI
CONCLUSIONS AND FUTURE RESEARCH
DIRECTIONS
This last chapter summarizes the theory developed in this thesis and gives an overview
of the key contributions made for general space applications. Also, we present several
future research directions/projects where our study can be used as a baseline and can be
extended to practical problems.
6.1 Concluding Remarks and Key Contributions

The main contribution of this thesis is the development of a nonlinear analytic theory of
spacecraft navigation. To achieve this goal, we have developed higher order state transition
tensors, nonlinear orbit uncertainty propagation, the concept of the statistically correct
trajectory, nonlinear statistical targeting, and nonlinear Bayesian and Kalman filters.
Higher-Order Relative Dynamics:
In this dissertation, we have developed an analytic expression for the solution of
relative dynamics by solving for state transition tensors that describe the localized
nonlinear motion. These tensors are computed by applying a Taylor series expansion and by numerically integrating them along a nominal solution flow. Once the
state transition tensors are available for some time interval, the inverse series can be
156
157
computed via a series reversion, which requires no numerical integration. Then, assuming an initial condition is within the radius of convergence and a sufficient order
of solution is considered, the nonlinear relative motion can be completely characterized and the deviations can be mapped analytically from an arbitrary point in the
relative space to some future time, or vice versa. In this way, we can also propagate
a phase volume analytically by mapping the boundary of the phase volume using the
higher state transition tensors. A convergence criteria for these higher order series
is discussed by introducing the local nonlinearity index. Also presented is the symplecticity of the higher solutions of a Hamiltonian system. As many astrodynamics
problems can be transformed into a Hamiltonian form, we can take the advantage of
useful properties that are available from a Hamiltonian dynamical system.
Time Invariance of Probability Density Function:
From the Fokker-Planck equation for the probability density function, we have shown
that the probability function of any dynamical system (without diffusion) satisfies
the integral invariance condition. Then, by applying the integral invariance of the
probability and conservation of the phase volume, we presented the relation between
the probability function and the phase volume, since they possess the same statistical
information, but are represented in different ways. We have also solved the FokkerPlanck equation for a deterministic system and derived a sufficiency condition for
the time invariance of a probability density function. This result was then applied to
show that the probability density function of a higher order Hamiltonian system is a
time invariant.
Nonlinear Orbit Uncertainty Propagation:
Applying the higher order state transition tensors and the time invariance of the prob-
158
ability density function, we have derived an analytic representation of the nonlinear
uncertainty propagation as a function of an initial distribution. When a sufficient
order of the higher order solutions are considered, we have shown that our analytic
approach can replace Monte-Carlo simulations while providing the same level of accuracy. This serves as the baseline of this thesis and was applied to several trajectory
navigation problems.
The Statistically Correct Trajectory and Nonlinear Statistical Targeting:
The analytical uncertainty propagation method enabled us to introduce the concept
of the statistically correct trajectory, where we solve for an initial state that satisfies
the target condition on average. As a practical application of the statistically correct
trajectory, we have developed the nonlinear statistical targeting method by utilizing
the statistical property of the trajectory in the maneuver design process, and thus
providing a statistically more accurate solution. When the initial uncertainties are
small, the nonlinear method essentially becomes the linear solution; however, when
there are sufficiently large initial uncertainties the solution gives superior statistical
performance. The results from the two-body and Hill three-body examples show
that there appear to exist an optimal time to perform a correction maneuver, a result
that is not possible using the linear method.
Nonlinear Filtering:
As an extension of the nonlinear orbit uncertainty propagation, we have derived
an analytic expression for the posterior conditional density function by solving the
Bayes rule for conditional probability distributions. This showed that the optimal
nonlinear filtering problem can be solved by applying the higher order relative dynamics solutions. For practical purposes, we then derived two Kalman-type filters,
159
called higher-order numerical extended Kalman filter and higher-order analytic extended Kalman filter, by directly applying the higher order solutions to the Kalman
filter algorithm. These higher order filters were compared with the conventional
extended Kalman filter and the unscented Kalman filter based on halo orbits computed in restricted three-body problem frame about the Sun-Earth and Earth-Moon
L1 points. The filter simulations were carried out assuming the dynamics of the
system are perfectly known, but there are errors in the initial state and in the measurements. The results showed that a higher order filter provides faster convergence,
a superior filter solution, and more flexility in the initial guess over linear filters.
Also, the Gaussian assumption of the a posteriori state yielded a sufficient approximation even for nonlinear filters. For the cases where the reference trajectory was
relatively close to the true trajectory, the higher order analytic filter provided solutions essentially equivalent to both the UKF and HNEKF, and yielded a much faster
filter process.
6.2 Future Research and Recommendations

There are a number of open questions and future research directions for the research
presented in this thesis. In the following we discuss several potential applications where
our analytic trajectory navigation techniques can be implemented and utilized.
Low-Thrust Trajectory:
The problem of low-thrust trajectory is one area where both the statistically correct
trajectory and the higher-order analytic extended Kalman filter can be applied. The
challenge in the low-thrust trajectory problem is that the nature of the small, continuous thrust is stochastic in general, and may be modeled as non-additive white
noise. Thus, a critical research for this problem would be to incorporate the nonlin-
160
ear noise effect into our analytic trajectory navigation framework, or to effectively
estimate the process noise (i.e., stochastic acceleration). Additional applications of
this type are spacecraft launch, atmospheric re-entry/aero-braking, and near solar
trajectories.
Autonomous Trajectory Navigation:
In recent proposals for mission to the Moon and Mars, precision entry, descent, and
landing problems are often discussed. This can be considered as a problem where
we have a fairly accurate initial solution, which is subject to a highly nonlinear trajectory environment. However, we are still given a desired landing site or a reference
trajectory. The nature of these problems must consider autonomous navigation techniques since it is not possible to perform open-loop control from a ground station
during the critical periods of the mission. This is another main research area where
our method can be directly implemented. Once a reference trajectory is computed,
an autonomous navigation processor that incorporates the trajectory nonlinearity in
a filtering process can be developed and a faster convergence may be feasible in
practice. Note that the processor would not require any numerical integration if
it uses the higher-order analytic extended Kalman filter (HAEKF) technique, and
thus, it is also possible to implement semi-analytic Monte-Carlo/particle type filtering techniques by mapping each random sample analytically using the STT approach
instead of numerically integrating the samples. The analytic method would require
less computational power; however, one must study the trade-off between large data
storage and numerical integration onboard a spacecraft.
Small-Body Collision/Encounter Analysis and Space Surveillance:
In small-body collision/encounter studies, the uncertainties associated with an initial
161
state are usually the dominating error source and process noise is almost negligible
(depending on the size of a small-body). However, we are still given an estimated
orbit from trajectory reconstruction, which we can consider as the reference trajectory. After integrating this reference orbit to some arbitrary time in the future,
which can be determined according to the local nonlinearity index, we can establish
a nonlinear filtering algorithm which includes the higher order statistics whenever
a measurement (e.g., a radar/optical measurement) becomes available. One challenge with this approach is that there exist a large number of small-bodies in our
solar system, and thus, computation of the higher order solutions can be computationally quite intensive. However, it is also important to note that this only needs
to be carried out once over a long time period. For the space surveillance problem
(e.g., low Earth space debris), the same analogy applies as in the small-body study.
One critical difference is that the process noise due to the atmospheric drag must be
estimated as a part of a filtering process.
Robust Computation of the State Transition Tensors:
Assuming the trajectory dynamics can be modeled as a Hamiltonian system, the
numerical errors (e.g., truncation error) associated with computing the higher order
state transition tensors can be significant reduced by implementing the variational
or symplectic integrator. This is possible since these integrators preserve the Hamiltonian structure in each step of integration, and thus, the nominal trajectory solution
and the state transition matrix can be computed with high accuracy, which yields a
more robust computation of the higher order solutions and can be integrated for a
long period of time.
APPENDICES
162
163
APPENDIX A
EQUATIONS OF MOTION OF ASTRODYNAMICS
PROBLEMS
A.1 The Two-Body Problem

The two-body problem describes the relative motion of two bodies, represented as
point particles, under their mutual gravitational attractions. The simplest generalization
of the two-body problem is when one of the bodies has negligibly small mass and the
central body has a uniform gravitational force, such as a spacecraft orbiting about the Earth
(assuming Earth is spherical with uniform gravitational force). This approximation is what
we usually referred to as the two-body problem. The governing equations of motion for
the two-body problem are defined as:
In Hamiltonian Form:
H =
1 2
GM
(px + p2y + p2z )
,
2
r
(A.1)
where the Hamiltonian H is simply the specific energy of an orbit, r = krk is the radius,
G is the universal gravitational constant, and M is the mass of the central body. The state
vector x = [ qT , pT ]T , where q = [ x , y , z ]T is the position (generalized coordinate)
vector and p = [ px , py , pz ]T is the velocity (generalized momentum) vector.
164
Table A.1: Properties of planets and satellites.
Planets
Sun
Earth
Moon
Europa
Jupiter
Titan
GM (km3 /s2 ) Radius (km)

1.3271011
695990
398600
6378
4903
1738
3201
1565
1.267108
71492
9028
2575
In Lagrangian Form:
r =
GM
r,
r3
(A.2)
where r = [ x , y , z ]T is the position vector and v = [ x , y , z ]T is the velocity vector. Note

that for two-body problem, both the Lagrangian and Hamiltonian equations of motion
are identical to each other, i.e., the generalized momentum are the same as the velocity
components. The properties of planets and satellites considered in this thesis are given in
Table A.1.
A.2 The Three-Body Problem

The three-body problem is an extension of the full two-body problem, where a third
body is added to the mutual gravitational potential, and hence, it describes the dynamics
of three point mass particles in space. The three-body problem is particularly useful for
trajectory design and analysis since it can approximate the majority of the dynamical environments that are present in our solar system. The two most widely used models for the
three-body problem are the circular restricted three-body problem (CR3BP) and the Hill
three-body problem.
165
A.2.1 The Circular Restricted Three-Body Problem
The CR3BP assumes the third body has negligible mass compared to the other two
bodies and both bodies are in a mutually circular orbit, such as the Sun-Earth-spacecraft
system. Assuming the coordinate system is centered at the center of mass in a rotating
frame the governing equations of motion in non-dimensional units are defined as follows:
1
1 2
(px + p2y + p2z ) (xpy ypx ) + (x2 + y2 ) V (x, y, z),
2
2
1 2
1
+ ,
=
(x + y2 ) +
2
r1
r2
p
=
(x + )2 + y2 + z2 ,
H =
(A.3)
(A.4)
r1
r2 =
(x 1 + )2 + y 2 + z 2 ,
(A.5)
(A.6)
where = M2 /(M1 + M2 ) are shown in Table A.2 and M2 M1 are the body masses.
Note that q = [ x , y , z ]T is the generalized coordinate vector and p = [ px , py , pz ]T is
the generalized momentum vector.
In Lagrangian Form:
V
,
x
V
y + 2x =
,
y
V
z =
,
z
x 2y =
(A.7)
(A.8)
(A.9)
where v = [ x , y , z ]T is the velocity vector. The generalized momentum is related to the

velocity components as:
x y
p = y + x
(A.10)
166
Table A.2: Properties of three-body systems.
Systems
Sun-Earth
3.003106
Earth-Moon
0.01215
Jupiter-Europa 2.526105
Distance (km)
(s1 )
149597871
1.991107
384400
2.649106
670900
2.048105
Note that, according to Eqn. 2.70 in 2.3, the inverse of the state transition matrix in a
Lagrangian system can be computed as:
T T T
1 = T1
0 JT0 T JT,
where
I
(q, p)
T =
=
2L
(q, q)
q q
(A.11)
0
I 0
=
2
,
L
I
q 2
0 1 0
1 0 0 .
0 0 0
(A.12)
(A.13)
The units can be dimensionalized by applying the length scale of ` = distance between
the two massive bodies, and the time scale of = 1/, where is the mean motion of the
secondary body about the primary body. Moreover, there exists a Jacobi constant which is
preserved for all time:
J =
1 2
(x + y 2 + z 2 ) V (x, y, z).
2
(A.14)
A.2.2 The Hill Three-Body Problem

The Hill three-body problem is an approximation of the CR3BP where the mass of the
primary is assumed to be much larger than the secondary, i.e., M2 M1 or 1/3 1.
167
After shifting the coordinate center to the secondary point mass, the governing equations
of motion can be defined as follows:
1 2
1
(px + p2y + p2z ) (xpy ypx ) + (x2 + y2 ) VH (x, y, z),
2
2
1
1 2
= p
+ (3x z2 ),
2
2
2
2
x +y +z
H =
VH
(A.15)
(A.16)
In Lagrangian Form:
VH
,
x
VH
y + 2x =
,
y
VH
z =
,
z
x 2y =
(A.17)
(A.18)
(A.19)
and a Jacobi constant exists:

J =
1 2
(x + y 2 + z 2 ) VH (x, y, z).
2
(A.20)
The units can be dimensionalized by applying the length scale of ` = (GM2 / 2 )1/3 and
the time scale of = 1/, where M2 is the mass of the secondary body and is the mean
motion of the secondary body about the primary body.
A.2.3 Halo Orbit
In the circular restricted three-body problem, there exists a special type of orbit called
a halo orbit, which is of particular interest for space missions. A halo orbit is a Lissajoustype1 periodic orbit where the in-plane and out-of-plane frequencies are the same [22, 25,
36, 52]. Figure A.1 shows families of halo orbits about the Sun-Earth L1 point in a nondimensional frame.2 As shown, there exist many families of halo orbits; however, the
1
A Lissajous orbit is an orbit where the in-plane and out-of-plane frequencies are not necessarily the
same. There is also an orbit called a Lyapunov orbit, which is periodic in the rotating frame of the primary
bodies, i.e., periodic in the x-y plane only.
2
In the CR3BP in a rotating frame, there exist five equilibrium points called the Lagrangian points. The
L1 point is a collinear Lagrange point located between the primary and secondary bodies.
168
x 10
8
6
zcoordinate
4
2
0
2
4
6
8
0.99
0.995
xcoordinate
ycoordinate
3
x 10
Figure A.1: Families of halo orbits about the Sun-Earth L1 point in non-dimensional
frame.
computation of a halo orbit requires a numerical technique since the problem is not integrable in closed form. The examples considered in this dissertation implemented a third
order analytic solution as the initial guess and applied differential corrections to obtain
convergence to the true halo orbit solution [85, 86].
169
APPENDIX B
PROPERTIES OF PROBABILITY DENSITY
FUNCTIONS
B.1 Integral Invariance of the PDF of a Linear Hamiltonian Dynamical System with Gaussian Boundary Conditions
Here, we show the integral invariance of the Gaussian probability density function
subject to a linear Hamiltonian dynamical system. Consider a linearized Hamiltonian
system:
q
x =
= JHxx x = Ax,
p
(B.1)
which gives the solution flow:

x(t; x0 , t0 ) = (t, t0 )x0 ,
= JHxx = A.
(B.2)
(B.3)
Now, without loss of generality, suppose x = (q, p) N (0, P) so that:
(det )1/2
1 T
p(x, t) =
exp x x ,
(2)N/2
2
(B.4)
with initial conditions:
(det 0 )1/2
1 T
p(x0 , t0 ) =
exp x0 0 x0 ,
(2)N/2
2
(B.5)
170
where = P1 , which is usually called the information matrix. In order to show the
integral invariance, it suffices to show:
1/2 12 xT x
(det ) e
dp(x, t)
=
dt
2(2)N/2
z(
}|
){
(det )
+ xT x
x T x + xT x
,
(det )
= 0,
(B.6)
and thus we only have to show {} = 0.

Recall from the basic probability theory that P = E[xxT ]. Taking the time derivative
gives:
T + x x T ],
P(t)
= E[ xx
= E[AxxT + xxT AT ],
= A(t)P(t) + P(t)AT (t),
(B.7)
which is the Riccati equation for the covariance matrix. Now consider the identity, P =
I. Taking the total time derivative of P and substituting Eqn. (B.7) give:
= AT A.
(B.8)
The time derivative of the information matrix can be stated as [92]:

d(det )
= 2(det )Trace(A),
dt
N
X
2H
2H
Trace(A) =
i i
= 0.
i pi
q
p
q
i=1
(B.9)
(B.10)
Then Eqn. (B.7) simplifies to:
{} = 0 xT AT x + xT AT A x + xT Ax ,
= xT AT x xT AT x xT Ax + xT Ax,
= 0,
which satisfies the necessary condition for the integral invariance.
(B.11)
171
APPENDIX C
THE LINEAR KALMAN FILTER
In this appendix, following a similar derivation as in Maybeck [56], we give a formal

derivation of the current state linear Kalman filter based on the Bayes rule for the posterior
conditional density function.
C.1 Kalman Filter Essentials

Here, we first present two useful matrix identities that will be used to derive the Kalman
Filter in the next section.
Lemma C.1.1 (Matrix Inversion Lemma). Given matrices PN N , HmN , and Rmm , the
following matrix inversion identities exist:
P1 + HT R1 H
P1 + HT R1 H
1
= P PHT HPHT + R
HP,
1
HT R1 = PHT HPHT + R
,
(C.1)
(C.2)
1
R1 H P1 + HT R1 H
= HPHT + R
HP,
(C.3)
1 T
1
H P1 + HT R1 H
H = R R HPHT + R
R,
(C.4)
where we assume PN N and Rmm are semi-positive definite matrices.
172
Proof. Define:
A =
P1
N N
HTmN
HmN Rmm
DN N FN m
=
.
GTN m Emm
A1
(C.5)
(C.6)
Setting AA1 = I(N +m)(N +m) , and solving for A1 , we get:

D =
1
P1 + HT R1 H
,
1
F = PHT HPHT + R
,
1
GT = R1 H P1 + HT R1 H
,
1
E = HPHT + R
.
Substituting these matrices into A1 and evaluating A1 A = I(N +m)(N +m) , the matrix
inversion identities, Eqns. (C.1-C.4), can be attained.
Remark C.1.2 (Determinant Identity). Suppose we are given:
= HT R1 H + P1 .
(C.7)
(det P)1/2 (det R)1/2

1 1/2
det
=
1/2
det(HPHT + R)
(C.8)
The following identity holds:
Proof. Recall from basic linear algebra, given AN N and BN N , the following determinant identities hold:
det(AB) = det(A) det(B),
det(A) = det(AT ).
(C.9)
(C.10)
173
Also, given CN m and Dmm :
A C
det
= det(A) det(D).
0 D
Now consider:
(C.11)
PH
P
=
,
T
HP HPH + R
1
T 1
T
1
PH
I
0
(P + H R H)
,
T
T
1
0
HPH + R
(HPH + R) HP I
which gives det = det[(HT R1 H + P1 )1 ] det(HPHT + R). The following matrix
partition holds for a positive-definite symmetric matrix:
X11 X12
=
T
X12 X22
1
1
0
I
0
I X11 X12 X11
, (C.12)
T
1
T
1
1
0
I
0 (X22 X12 X11 X12 )
X12 X11 I
which gives det = (det P)(det R). Combining det from two approaches, we get:
(det P)(det R)
,
det(HPHT + R)
(C.13)
(det P)1/2 (det R)1/2

=
1/2 .
det(HPHT + R)
(C.14)
det[(HT R1 H + P1 )1 ] =
det
1 1/2
C.2 Kalman Filter Derivation

Consider the following system dynamics:
dx(t) = F(t)x(t)dt + G(t)d(t),
(C.15)
174
where x is the state vector with a dimension N , F is an N N matrix characterizing the
system dynamics, G is an N s noise input matrix characterizing system diffusion, and
is an s-dimensional Brownian motion with:
E[(t)] = 0,
E[d(t)dT (t)] = Q(t)dt,
Z t2
T
E [(t2 ) (t1 )][(t2 ) (t1 )]
=
Q( )d.
(C.16)
(C.17)
(C.18)
t1
Since measurements are usually obtained in discrete form, we will consider the following
linear dynamics and linear measurements in discrete form:
xk = (tk , tk1 )xk1 + wk1 ,
(C.19)
z k = Hk x k + v k ,
(C.20)
where wk1 and vk are white noises with E[wk ] = E[vk ] = 0, E[wk wTk ] = Qk , and
E[vk vTk ] = Rk . Note that wk1 can be understood as a linear stochastic integral:
Z
tk
wk1 =
(tk , )G( )d( ),
(C.21)
tk1
and E[wk wTk ] can be computed as:

Z
E[wk wTk ]
= Qk =
tk
(tk , )G( )Q( )GT ( )T (tk , )d.
(C.22)
tk1
Using Bayes rule, the posterior conditional density (i.e., measurement update probability density function) can be stated as:
p(xk |z1:k1 )p(zk |xk , z1:k1 )
,
p(zk |z1:k1 )
p(xk |z1:k1 )p(zk |xk , z1:k1 )
= R
,
p(x
|z
)p(z
|x
,
z
)dx
k
1:k1
k
k
1:k1
k
p(xk |z1:k ) =
(C.23)
(C.24)
where z1:k = {z1 , z2 , , zk }. Now suppose a state at tk1 can be characterized as Gaus+
sian, i.e., xk1 N (m+
k1 , Pk1 ). The system equations, Eqn. (C.19) and (C.20) are both
175
linear and process (wk ) and measurement (vk ) noises are Gaussian. Thus, it is apparent that
the probability density functions in Eqn. (C.23) must be Gaussian as well. First consider
p(xk |z1:k1 ) in the numerator, which describes the evolution of the probability density of
xk . The mean and covariance matrix can be mapped to tk by computing:
m
= E [xk |z1:k1 ] = (tk , tk1 )m+
(C.25)
k
k1 ,
h
i
T
Pk = E xk mk xk mk |z1:k1 ,
Z tk
+
T
= (tk , tk1 )Pk1 (tk , tk1 ) +
(tk , )G( )Q( )GT ( )T (tk , )d,
tk1
(C.26)
which gives:
1/2
(det
1
T
k)
p(xk |z1:k1 ) =
exp xk mk k xk mk
, (C.27)
(2)N/2
2
1
where
k = (Pk ) . Now, consider the second numerator in the Bayes rule equation.
The mean and covariance matrix of the measurement at tk yields:

E [zk |xk , z1:k1 ] = Hk xk ,
E (zk Hk xk )(zk Hk xk )T |xk , z1:k1 = Rk ,
(C.28)
(C.29)
and the probability density function can be stated as:
(det Rk )1/2
1
T
1
p(zk |xk , z1:k1 ) =
exp (zk Hk xk ) Rk (zk Hk xk ) . (C.30)
(2)m/2
2
Remark C.2.1. The measurement probability density function, p(zk |z1:k1 ) can be directly
computed as:
E [zk |z1:k1 ] = E [Hk xk + vk |z1:k1 ] = Hk m
k,
(C.31)
T
E [Pzk |z1:k1 ] = E (zk Hk m
k )(zk Hk mk ) |z1:k1 ,
T
= E (Hk xk Hk m
k + vk )(Hk xk Hk mk + vk ) |z1:k1 ,
T
= Hk P
k H k + Rk ,
(C.32)
176
and the resulting probability density function is:
1
1
e 2 (zk Hk mk ) (Hk Pk Hk +Rk ) (zk Hk mk )
p(zk |z1:k1 ) =
.
T
1/2
(2)m/2 [det(Hk P
k Hk + Rk )]
Note that the measurement probability density function, p(zk |z1:k1 ), can be computed
directly by applying the linearity of the measurement function and the Gaussian assumption. However, we carry out the integration given in Eqn. (C.33) to show that these two
approaches yield the same solution.
Now, consider the denominator of the posterior conditional density function:
Z
p(zk |z1:k1 ) =
p(zk |xk , z1:k1 )p(xk |z1:k1 )dxk
}|
{
z
T
T 1
1
exp{ 2 (zk Hk xk ) Rk (zk Hk xk ) + (xk mk ) k (xk mk ) }
Z
=
(2)
N +m
2
1/2
(det Rk )1/2 (det
k)
dxk .
(C.33)
{} =
T 1
T 1
T T 1
zk Rk zk xTk HTk R1
k zk zk Rk Hk xk + xk Hk Rk Hk xk
T
T
T
+ xT
x
m
+
m
m
,
k
k
k
k
k
k
k
k
k
k
k
T T 1
T 1
T
T 1
T
xk (Hk Rk Hk +
k )xk 2xk (Hk Rk zk + k mk ) + zk Rk zk + mk k mk ,
(C.34)
where we let Pk = P
k and mk = mk to ease the notation. Now define:
1
k = HTk R1
k Hk + Pk ,
1
ak = HTk R1
k zk + Pk mk .
(C.35)
(C.36)
177
Then Eqn. (C.34) can be re-written as:
{} =
T 1
xk k xk 2xTk ak + zTk R1
k zk + mk Pk mk ,
T
1
1
T
xk k xk 2xTk k 1
k a k + a k k k k a k
T 1
T 1
aTk 1
k ak + zk Rk zk + mk Pk mk ,
T 1
1
1
1
T
T
zk Rk zk + mTk P1
m
a
+
(x
a
)
(x
a
)
.
k
k
k
k
k
k
k
k
k
k
k
k
The measurement density function then becomes:

p(zk |z1:k1 )
R 1
1
1
T 1
T 1
T
1
1
T
e 2 {zk Rk zk +mk Pk mk ak k ak } e 2 (xk k ak ) k (xk k ak )
dxk ,
=
N +m
(2) 2 (det Rk )1/2 (det Pk )1/2
1
1
T 1
T 1
T
1/2
e 2 {zk Rk zk +mk Pk mk ak k ak } (2)N/2 |1 |
=
,
N +m
(2) 2 (det Rk )1/2 (det Pk )1/2
1
1
e 2
1/2
(2)m/2 Hk Pk HTk + Rk
}|
{
1
T 1
T 1
T
zk Rk zk + mk Pk mk ak k ak , (C.37)
where the third equality applies the result from Remark C.1.2. Using the matrix inversion
lemma, we can factor out the exponential function in Eqn. (C.37) as:
1
T 1
T
{} = zTk R1
k zk + mk Pk mk ak k ak ,
h
1 i 1
T
T 1
T 1
T 1
= zk Rk zk + mk Pk mk zk Rk Rk Rk Hk Pk Hk + Rk
Rk Rk z k
1
T
T
2mTk P1
P
H
H
P
H
+
R
zk
k
k
k
k
k
k
k
i
h
1
1
T
T
H
P
+
R
H
P
H
P
P
H
mTk P1
k k Pk mk ,
k
k k k
k
k k
k
h
1 i
T
1
T
T 1
zk
+
R
H
P
H
m
z
R
z
+
m
P
= zTk R1
k
k k k
k
k
k
k k
k
k
h
1 i
T
T
1
T
T
T T
Hk mk ,
zk mk Pk Hk Hk Pk Hk + Rk
2mk Hk Hk Pk Hk + Rk
1
zk
zk 2mTk HTk Hk Pk HTk + Rk
= zTk Hk Pk HTk + Rk
1
Hk mk .
+ mTk HTk Hk Pk HTk + Rk
178
Applying this result, the measurement probability density function becomes:
p(zk |z1:k1 ) =
1
1
(zk Hk mk )
21 (zk Hk mk )T (Hk Pk HT
k +Rk )
,
e
T
m/2
1/2
(2)
det(Hk Pk Hk + Rk )]
(C.38)
which is identical to Eqn. (C.33) as expected.
At this point, simplifying the numerator of the posterior conditional density function
becomes straightforward as the process is similar computing p(zk |z1:k1 ). The complete
representation of the updated probability density function can be stated as:
p(xk |z1:k1 )p(zk |xk , z1:k1 )
,
p(xk |z1:k1 )p(zk |xk , z1:k1 )dxk
1
1
1
1
T
=
exp (xk k ak ) k (xk k ak ) ,
1/2
2
(2)N/2 (det 1
k )
p(xk |z1:k ) = R
(C.39)
which gives:
m+
= E [xk |z1:k ] = 1
k
k ak ,
P
k
1
T
T
= m
+
P
H
H
P
H
+
R
(zk Hk m
k
k
k
k
k
k
k
k ),
h
i
+ T
= E xk m+
x
m
|z
= 1
k
1:k
k
k
k ,
1
T
T
= P
Hk P
k Pk Hk Hk Pk Hk + Rk
k,
(C.40)
(C.41)
where the matrix inversion lemma is applied again.

The conventional Kalman Filter algorithm is then stated as follows:
Definition C.2.2. Prediction:
= (tk , tk1 )m+
m
k1 ,
k
T
= (tk , tk1 )P+
P
k1 (tk , tk1 ) +
k
(C.42)
Z
tk
(tk , )G( )Q( )GT ( )T (tk , )d.
tk1
(C.43)
179
Definition C.2.3. Update:
m+
= m
k
k + Kk zk Hk mk ,
(C.44)
P+
= P
k
k Kk Hk P k ,
(C.45)
1
T
T
Kk = P
,
k Hk Hk Pk Hk + Rk
(C.46)
zz 1
is called the Kalman gain.
where Kk = Pxz
k (Pk )
180
APPENDIX D
VECTORIZATION OF HIGHER ORDER TENSORS
When integrating the higher order state transition tensors, it is more convenient to put
them into a vectorized form.
D.1 Specifications for MATLAB

Consider a system with dimension N . Let S(i) be the ith element of a tensor 1 ,2 p ,
where the element number increases from the last index p to the first index 1 . For
example, a system with dimension 2, the second order tensor 1 ,2 can be stated as:
S(1) = 1,1 , S(2) = 1,2 , S(3) = 2,1 , S(4) = 2,2 . Now, let U (i) be the ith element
number of the tensor in a vector form. The previous example can be stated as: U (1) = 1 ,
U (2) = 2 , U (3) = 1,1 , U (4) = 1,2 , U (5) = 2,1 , U (6) = 2,2 . Note that U (1) = 1
and U (2) = 2 include zeroth-order tensors, i.e., the state components.
Example D.1.1. An element of a second order tensor can be stated as:
i,j = S[N (i 1) + j],
= U [N + N (i 1) + j],
= U (iN + j).
181
Example D.1.2. An element of a third order tensor can be stated as:
i,jk = S[N 2 (i 1) + N (j 1) + k],
= U [N + N 2 + N 2 (i 1) + N (j 1) + k],
= U (iN 2 + jN + k).
Example D.1.3. An element of a fourth order tensor can be stated as:
i,jkl = S[N 3 (i 1) + N 2 (j 1) + N (k 1) + l],
= U (iN 3 + jN 2 + kN + l).
In general, an element of a tensor can be stated as:
"
#
p
X
1 ,2 p = S 1 +
N pi (i 1) ,
= U
p
X
i=1
(D.1)
N pi i
(D.2)
i=1
where j {1, , N }. Note that MATLAB initializes an array with an integer 1, and
thus, this generalization cannot be applied to C++ or alike.
D.2 Specifications for C or C++

In C or C++ programming language, an array is initialized with an integer 0. In general,
an element of a tensor can be stated as:
!
p
X
1 ,2 p = S
N pi i ,
i=1
"
p
= U
(" i=1
p
= U
X
i=1
where j {0, , N 1}.
(D.3)
!
N pi i
p
X
i=1
!
N pi
#
1 ,
(D.4)
N pi (i + 1) 1 ,
(D.5)
BIBLIOGRAPHY
182
183
BIBLIOGRAPHY
[1] G. Adomian. Stochastic Systems, volume 169. Academic Press, 1983.

[2] B.D.O. Anderson and J.B. Moore. Optimal Filtering. Prentice Hall, 1979.
[3] T.F. Andr, R.D. Nowak, and B.D. Van Veen. Low-rank estimation of higher order
statistics. IEEE Transactions on Signal Processing, 45(3):673685, 1997.
[4] V.I. Arnold. Mathematical Methods of Classical Mechanics. Springer-Verlag, second edition edition, 2000.
[5] M.S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle
filters for online nonlinear/non-Gaussian Bayesian tracking. IEEE Transactions on
Signal Processing, 50(2):174188, Feb. 2002.
[6] M. Athans, R.P. Wishner, and A. Bertolini. Suboptimal state estimation for
continuous-time nonlinear systems from discrete noisy measurements. IEEE Transactions on Automatic Control, AC-13(5):504514, October 1968.
[7] B. Azimi-Sadjadi and P.S. Krishnaprasad. Approximate nonlinear filtering and its
application in navigation. Automatica, 41:945956, 2005.
[8] R.R. Bate, D.D. Mueller, and J.E. White. Fundamentals of Astrodynamics. Dover
Publications, third edition, reprinted edition, 1971.
[9] R.H. Battin. An Introduction to the Mathematics and Methods of Astrodynamics.
AIAA Education Series. AIAA, revised edition, 1999.
[10] J.S. Bay. Fundamentals of Linear State Space Systems. McGraw-Hill, 1999.
[11] A.T. Bharucha-Reid. Random Integral Equations, volume 96. Academic Press,
1972.
[12] G.J. Bierman. Factorization Methods for Discrete Sequential Estimation, volume
128. Academic Press, 1977.
[13] A.M. Bloch, J. Baillieul, P. Crouch, and J. Marsden. Nonholonomic Mechanics and
Control. Springer-Verlag, 2003.
184
[14] D.G. Boden and B.A. Conway. A comparison of nonlinear filters for orbit determination. Astrodynamics Specialist Conference, August 1986, Williamsburg, VA.
AAS 86-2055.
[15] E. Blviken, P.J. Acklam, N. Christophersen, and J.-M. Strdal. Monte carlo filters
for non-linear state estimation. Automatica, 37:177183, 2001.
[16] H. Cabral and F. Diacu. Classical and Celestial Mechanics: The Recife Lectures.
Princeton University, 2002.
[17] C.E.S. Cesnik, R.S. Park, and R. Palacios. Effective cross-section distribution of
anisotropic piezocomposite actuators for wing twist. SPIE 10th International Symposium on Smart Structures and Materials, March 2003, San Diego, California.
[18] A.J. Chorin, R. Kupferman, and D. Levy. Optimal prediction for Hamiltonian partial differential equations. Journal of Computational Physics, 162:267297, 2000.
[19] J.L. Crassidis and J.L. Junkins. Optimal Estimation of Dynamics Systems. CRC
Press LLC, 2004.
[20] R. Deutsch. Nonlinear Transformations of Random Processes. Prentice Hall, 1962.
[21] A. Doucet, N. de Freitas, and N.J. Gordon. Sequential Monte Carlo Methods in
Practice. Springer-Verlag, 2001.
[22] D.W. Dunham and C.E. Roberts. Stationkeeping techniques for libration-point
satellites. Journal of the Astronautical Sciences, Special Issue on Libration-Point
Missions, 49(1), 2001.
[23] P.R. Escobal. Methods of Orbit Determination. Krieger Publishing Company,
reprinted edition, 1976.
[24] R.W. Farquhar. The utilization of halo orbits in advanced lunar operations. NASATM-X-65409, 1970.
[25] R.W. Farquhar. The flight of ISEE-3/ICE: Origins, mission history, and a legacy.
Journal of the Astronautical Sciences, Special Issue on Libration-Point Missions,
49(1), 2001.
[26] T.D. Frank. Nonlinear Fokker-Planck Equations. Springer-Verlag, 2005.
[27] H. Franz, P. Sharer, K. Ogilvie, and M. Desch. WIND nominal mission performance and extended mission design. Journal of the Astronautical Sciences, Special
Issue on Libration-Point Missions, 49(1), 2001.
[28] W. Greblicki. Nonlinearity estimation in Hammerstein systems based on ordered
observations. IEEE Transactions on Signal Processing, 44(5):12241233, 1996.
[29] D.T. Greenwood. Principles of Dynamics. Prentice Hall, second edition edition,
1988.
185
[30] D.T. Greenwood. Classical Dynamics. Dover Publications, 1997.
[31] D.T. Greenwood. Advanced Dynamics. Cambridge University Press, 2003.
[32] G.R. Grimmett and D.R. Stirzaker. Probability and Random Processes. Oxford
University Press, 2001.
[33] V.M. Guibout. The Hamilton-Jacobi Theory for Solving Two-Point Boundary Value
Problems: Theory and Numerics with Application to Spacecraft Formation Flight,
Optimal Control and the Study of Phase Space Structure. PhD thesis, The University of Michigan, 2004.
[34] F. Gustafsson, F. Gunnarsson, N. Bergman, U. Forssell J. Jansson, R. Karlsson,
and P.-J. Nordlund. Particle filters for positioning, navigation, and tracking. IEEE
Transactions on Signal Processing, 50(2):425437, Feb. 2002.
[35] Y.C. Ho and R.C.K. Lee. A Bayesian approach to problems in stochastic estimation
and control. IEEE Transactions on Automatic Control, 9, October 1964.
[36] K.C. Howell. Families of orbits in the vicinity of the collinear libration points.
Journal of the Astronautical Sciences, Special Issue on Libration-Point Missions,
49(1), 2001.
[37] C. Hue, J.-P. Le Cadre, and P. Prez. Sequential Monte Carlo methods for multiple target tracking and data fusion. IEEE Transactions on Signal Processing,
50(2):309325, Feb. 2002.
[38] K. Ito and K. Xiong. Gaussian filters for nonlinear filtering problems. IEEE Transactions on Automatic Control, 45(5):910927, May 2000.
[39] M.K. Jah, M.E. Lisano, G.H. Born, and P. Axelrad. Mars aerobraking spacecraft
state estimation by processing inertial measurement unit data. NASA/TP-2005213901, August 2005.
[40] M.K. Jah, M.E. Lisano, G.H. Born, and P. Axelrad. Mars aerobraking spacecraft
state estimation by processing inertial measurement unit data. SpaceOps 2006 Conference, June 2006, Rome, Italy. AIAA-2006-5920.
[41] J. Jeraj and V.J. Mathews. A stable adaptive Hammerstein filter employing partial
orthogonalization of the input signals. IEEE Transactions on Signal Processing,
54(4):14121420, 2006.
[42] J. Jeraj and V.J. Mathews. Stochastic mean-square performance analysis of an
adaptive Hammerstein filter. IEEE Transactions on Signal Processing, 54(6):2168
2177, 2006.
[43] S.J. Julier and J.K. Uhlmann. The scaled unscented transformation. American
Control Conference, pages 45554559, 2002, Anchorage, AK.
186
[44] S.J. Julier and J.K. Uhlmann. Unscented filtering and nonlinear estimation. Proceedings of the IEEE, 92(3):401422, March 2004.
[45] S.J. Julier, J.K. Uhlmann, and H.F. Durrant-Whyte. A new approach for filtering
nonlinear systems. American Control Conference, pages 16281632, June 1995,
Seattle, WA.
[46] S.J. Julier, J.K. Uhlmann, and H.F. Durrant-Whyte. A new method for the nonlinear
transformation of means and covariances in filters and estimators. IEEE Transactions on Automatic Control, 45(3):477482, March 2000.
[47] J.L. Junkins, M.F. Akella, and K.T. Alfriend. Non-Gaussian error propagation in
orbit mechanics. Journal of the Astronautical Sciences, 44(4):541563, 1996.
[48] J.L. Junkins and P. Singla. How nonlinear is it? a tutorial on nonlinearity of orbit
and attitude dynamics. Journal of the Astronautical Sciences, 52(1 and 2):760,
2004.
[49] R.E. Kalman. A new approach to linear filtering and prediction problems. Transactions of the ASME, Journal of Basic Engineering, Series D, 82, March 1960.
[50] W.S. Koon, M.W. Lo, J.E. Marsden, and S.D. Ross. Shoot the Moon. Space Flight
Mechanics Meeting, January 2000, Clearwater, Florida. AAS 00-166.
[51] H. Lass and P. Gottlieb. Probability and Statistics. Addison-Wesley, 1971.
[52] M.W. Lo, B.G. Williams, W.B. Bollman, D. Han, Y. Hahn, J.L. Bell, E.A. Hirst,
R.A. Corwin, P.E. Hong, K.C. Howell, B. Barden, and R. Wilson. Genesis mission
design. Journal of the Astronautical Sciences, Special Issue on Libration-Point
Missions, 49(1), 2001.
[53] J.M. Longuski, E. Fischbach, D.J. Scheeres, G. Giampieri, and R.S. Park. Deflection of spacecraft trajectories as a new test of general relativity: Determining the
PPN parameters and . Phys. Rev. D, 69(042001):042001104200115, 2004.
[54] J.E. Marsden and T.S. Ratiu. Introduction to Mechanics and Symmetry. SpringerVerlag, second edition edition, 1999.
[55] P.S. Maybeck. Stochastic Models, Estimation, and Control. Vol 2. Academic Press,
1982.
[56] P.S. Maybeck. Stochastic Models, Estimation, and Control. Vol 1. Republished by
Navtech Book & Software Store, 1994.
[57] P. McCullagh. Tensor Methods in Statistics. Chapman and Hall, 1987.
[58] K.D. Mease, J.D. Anderson, L.J. Wood, and L.K. White. Tests of general relativity
using Starprobe radio metric tracking data. Journal of Guidance, Control, and
Dynamics, 7(1):3644, 1983.
187
[59] K.D. Mease, L.J. Wood, M.J. Bergam, and L.K. White. Estimation of solar gravitational harmonics with Starprobe radiometric tracking data. Journal of the Astronautical Sciences, 31(1):322, January-March 1983.
[60] A. Milani, S.R. Chesley, P.W. Chodas, and G.B. Valsecchi. Asteroid Close Approaches: Analysis and Potential Impact Detection in Asteroids III. University of
Arizona Press, 2002.
[61] O. Montenbruck and E. Gill. Satellite Orbits. Springer, 2nd edition, 2001.
[62] F.R. Moulton. Differential Equations. The MacMillan Company, 1930.
[63] M. Nrgaard, N.K. Poulsen, and O. Ravn. New developments in state estimation
for nonlinear systems. Automatica, 36(11):16271638, November 2000.
[64] R.D. Nowak and B.D. Van Veen. Random and pseudorandom inputs for Volterra
filter identification. IEEE Transactions on Signal Processing, 42(8):21242135,
1994.
[65] R.D. Nowak and B.D. Van Veen. Invertibility of higher order moment matrices.
IEEE Transactions on Signal Processing, 43(3):705708, 1995.
[66] R.D. Nowak and B.D. Van Veen. Tensor product basis approximations for Volterra
filters. IEEE Transactions on Signal Processing, 44(1):3650, 1996.
[67] R.S. Park. Expected navigation performance of the Radiation Belt Storm Probe
mission. APL Internal Report, SEG-06-022, August 2006.
[68] R.S. Park, E. Fischbach, G. Giampieri, J.M. Longuski, and D.J. Scheeres. A test of
general relativity: Estimating PPN parameters and from spacecraft radiometric
tracking data. Proceedings Supplements of Nuclear Physics B, 134(2004):181183,
December 2003, Washington, D.C.
[69] R.S. Park and D.J. Scheeres. Nonlinear mapping of Gaussian statistics: Theory and
applications to spacecraft trajectory design. 29(6):13671375, 2006.
[70] R.S. Park and D.J. Scheeres. Nonlinear mapping of Gaussian state uncertainties:
Theory and application to spacecraft control and navigation. Astrodynamics Specialist Conference, August 2005, Lake Tahoe, California. AAS 05-404.
[71] R.S. Park and D.J. Scheeres. Covariance analysis of Cassini Titan flyby using SAR
and altimetry data. Astrodynamics Specialist Conference, August 2006, Keystone,
Colorado. AIAA-2006-6398.
[72] R.S. Park and D.J. Scheeres. Nonlinear semi-analytic method for spacecraft navigation. Astrodynamics Specialist Conference, August 2006, Keystone, Colorado.
AIAA-2006-6399.
188
[73] R.S. Park and D.J. Scheeres. Nonlinear semi-analytic methods for spacecraft trajectory design, control, and navigation. New Trends in Astrodynamics and Applications Conference, August 2006, Princeton, New Jersey.
[74] R.S. Park and D.J. Scheeres. Nonlinear mapping of Gaussian state covariance and
orbit uncertainties. Space Flight Mechanics Meeting, January 2005, Copper Mountain, Colorado. AAS 05-170.
[75] R.S. Park and D.J. Scheeres. Nonlinear mapping of Gaussian state uncertainties.
JAXA Astrodynamics and Flight Mechanics Meeting, July 2005, Kanagawa, Japan.
[76] R.S. Park and D.J. Scheeres. Nonlinear semi-analytic methods for trajectory estimation. submitted to the Journal of Guidance, Control, and Dynamics, November
2006.
[77] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach. A
test of general relativity: Estimating the parameterized post-Newtonian parameters from spacecraft radiometric measurements. Journal of Spacecraft and Rockets,
42(3):559568, 2005.
[78] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach. Orbit design for general relativity experiments: Heliocentric and Mercury-centric cases. Astrodynamics Specialist Conference, August 2004, Providence, Rhode Island. AIAA
2004-5394.
[79] R.S. Park, D.J. Scheeres, G. Giampieri, J.M. Longuski, and E. Fischbach. Estimating general relativity parameters from radiometric tracking of heliocentric trajectories. Space Flight Mechanics Meeting, February 2003, Ponce, Puerto Rico. AAS
03-205.
[80] M.E. Paskowitz and D.J. Scheeres. Robust capture and transfer trajectories for planetary satellite orbiters. Journal of Guidance, Control, and Dynamics, 29(2):342
353, 2006.
[81] H. Pollard. Celestial Mechanics. The Mathematical Association of America, 1976.
[82] Y.V. Prohorov and Y.A. Rozanov. Probability Theory. Springer-Verlag, 1969.
[83] J. Prussing and B. Conway. Orbital Mechanics. Oxford University Press, Inc., New
York, 1993.
[84] C.A. Renault and D.J. Scheeres. Optimal placement of statistical maneuvers in
an unstable orbital environment. Journal of Guidance, Control, and Dynamics,
26(5):758769, 2003.
[85] D.L. Richardson. Analytic construction of periodic orbits about the collinear points.
Celestial Mechanics, 22:241253, 1980.
189
[86] D.L. Richardson and N.D. Cary. A uniformly valid solution for motion about the
interior libration point of the perturbed elliptic-restricted problem. Astrodynamics
Specialist Conference, July 1975, Nassau, Bahamas. AAS 75-021.
[87] H. Risken. The Fokker-Planck Equation. Springer-Verlag, 1984.
[88] S.M. Ross. A First Course in Probability. Prentice Hall, 1998.
[89] D.J. Scheeres, D. Han, and Y. Hou. Influence of unstable manifolds on orbit uncertainty. Journal of Guidance, Control, and Dynamics, 24(3):573585, 2001.
[90] D.J. Scheeres, F.-Y. Hsiao, R.S. Park, B.F. Villac, and J.M. Maruskin. Fundamental
limits on spacecraft orbit uncertainty and distribution propagation. accepted for
publication in Journal of the Astronautical Sciences, 2006.
[91] D.J. Scheeres, F.-Y Hsiao, R.S. Park, B.F. Villac, and J.M. Maruskin. Fundamental
limits on spacecraft orbit uncertainty and distribution propagation. Shuster Symposium, June 2005, Grand Island, New York. AAS 05-471.
[92] D.J. Scheeres and M.W. Lo. Integrated trajectory and navigation design in unstable
orbital environments. InterPlanetary Network Progress Report, 42-150, 2003.
[93] T.T. Soong. Random Differential Equations in Science and Engineering, volume
103. Academic Press, 1973.
[94] T.T. Soong and N.A. Paul. A second- and higher order perturbation analysis of
two-body trajectories. AIAA Journal, 9(4), 1971.
[95] L. Trailovic. Variance estimation and ranking of target tracking position error modeled using Gaussian mixture distributions. Automatica, 41, 2005.
[96] Kostas Triantafyllopoulos. On the central moments of the multidimensional
Gaussian distribution. Mathematical Scientist, 28, 2003.
[97] R. van der Merwe, A. Doucet, N. de Freitas, and E. Wan. The unscented particle
filter. Technical Report CUED/F-INFENG/TR, August 2000.
[98] E. Wan and R. van der Merwe. Kalman Filtering and Neural Networks: The Unscented Kalman Filter, chapter 7. Wiley, 2001.
[99] W. Wiesel. Modern Astrodynamics. Aphelion Press, 2003.
[100] W. Wiesel. Modern Orbit Determination. Aphelion Press, 2003.
ABSTRACT
by
Sang H. Park
Chair: Daniel J. Scheeres
Trajectory navigation entails the solution of many different problems that arise due
to uncertain knowledge of the spacecraft state, including orbit prediction, correction maneuver design, and trajectory estimation. In practice, these problems are usually solved
based on an assumption that linear dynamical models sufficiently approximate the local
trajectory dynamics and their associated statistics. However, astrodynamics problems are
nonlinear in general and linear spacecraft dynamics models can fail to characterize the true
trajectory dynamics when the system is subject to a highly unstable environment or when
mapped over a long time period. This limits the performance of traditional navigation
techniques and can make it difficult to perform precision analysis or robust navigation.
This dissertation presents an alternate method for spacecraft trajectory navigation based
on a nonlinear local trajectory model and their statistics in an analytic framework. For a
given reference trajectory, we first solve for the higher order Taylor series terms that describe the localized nonlinear motion and develop an analytic expression for the relative
solution flow. We then discuss the nonlinear dynamical mapping of a spacecrafts probability density function by solving the Fokker-Planck equation for a deterministic system.
From this result we derive an analytic method for orbit uncertainty propagation which
can replicate Monte-Carlo simulations with the benefit of added flexibility in initial orbit
statistics.
Using this approach, we introduce the concept of the statistically correct trajectory
where we directly incorporate statistical information about an orbit state into the trajectory
design process. As an extension of this concept, we define a nonlinear statistical targeting method where we solve for a correction maneuver which intercepts the desired target
on average. Then we apply our results to a Bayesian filtering problem to obtain a general filtering algorithm for optimal estimation of the posterior conditional density function
incorporating nonlinearity into the filtering process. Finally, we derive practical Kalmantype filters by applying our nonlinear relative solutions into the standard filters and show
that these filters provide superior performance over linear filtering methods based on realistic trajectory and uncertainty models. The examples we consider are a conventional
Hohmann transfer from the Earth to Moon using a simple two-body model, a strongly unstable transfer trajectory in the Hill three-body problem from the vicinity of L2 through
several orbits, and to the navigation of a spacecraft in a halo orbit in the restricted threebody problem. For each of these examples we show the benefits of using our nonlinear
trajectory navigation techniques as compared to traditional linear navigation techniques.

PHD Thesis Nonlinear Trajectory Navigation

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

PHD Thesis Nonlinear Trajectory Navigation

Uploaded by

Copyright:

Available Formats

NONLINEAR TRAJECTORY NAVIGATION

A dissertation submitted in partial fulfillment

All Rights Reserved

Conference Papers and Industrial Reports

LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi

Spacecraft Navigation and Uncertainty Propagation . . . . . . .

II. RELATIVE MOTION OF GENERAL NONLINEAR DYNAMICAL

General Trajectory Dynamics and Solution Flows . . . . . . . .

III. EVOLUTION OF PROBABILITY DENSITY FUNCTIONS IN NONLINEAR DYNAMICAL SYSTEMS . . . . . . . . . . . . . . . . . . .

Review of Probability Theory and Random Processes . . . . . .

IV. NONLINEAR TRAJECTORY NAVIGATION . . . . . . . . . . . . .

The Concept of Statistically Correct Trajectory . . . . . . . . . .

V. NONLINEAR SPACE MISSION ANALYSIS . . . . . . . . . . . . . 111

VI. CONCLUSIONS AND FUTURE RESEARCH DIRECTIONS . . . . 156

Concluding Remarks and Key Contributions . . . . . . . . . . . 156

Integral invariance of a phase volume. . . . . . . . . . . . . . . . . . .

Illustration of the statistically correction trajectory. . . . . . . . . . . .

Illustration of the nonlinear statistical targeting. . . . . . . . . . . . . .

Propagated mean and 1- error ellipsoid projected onto the spacecraft

Trajectory norms and higher order solution magnitudes: circled points

Phase volume projections: solid line represents integrated, dotted line

Phase volume projections: solid line represents integrated, dotted line

Phase volume projections. . . . . . . . . . . . . . . . . . . . . . . . . 119

Two-body problem: Hohmann transfer trajectory. . . . . . . . . . . . . 121

Two-body problem: comparison of the computed mean and covariance

Hill three-body problem: a safe trajectory at Europa. . . . . . . . . . . 124

Two-body problem: computed Vk using the linear and nonlinear methods.126

Hill three-body problem: computed Vk using the linear and nonlinear

Hill three-body problem: deviated position and velocity means at the

Hill three-body problem: Monte-Carlo simulation using the linear and

Nominal halo orbit about the Sun-Earth L1 point. . . . . . . . . . . . . 132

Sun-Earth halo orbit: comparison of the uncertainties computed using

Sun-Earth halo orbit: comparison of the absolute errors computed using

Sun-Earth halo orbit: comparison of the uncertainties computed using

Sun-Earth halo orbit: comparison of the absolute errors computed using

Sun-Earth halo orbit: comparison of the uncertainties computed using

Sun-Earth halo orbit: comparison of the absolute errors computed using

Sun-Earth halo orbit: comparison of the uncertainties computed using

Sun-Earth halo orbit: comparison of the absolute errors computed using

Sun-Earth halo orbit: comparison of the uncertainties computed using

Sun-Earth halo orbit: comparison of the absolute errors computed using

Nominal halo orbit about the Earth-Moon L1 point. . . . . . . . . . . . 143

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Earth-Moon halo orbit: comparison of the uncertainties computed using

Families of halo orbits about the Sun-Earth L1 point in non-dimensional

Local nonlinearity index. . . . . . . . . . . . . . . . . . . . . . . . . . 122

Properties of planets and satellites. . . . . . . . . . . . . . . . . . . . . 164

Properties of three-body systems. . . . . . . . . . . . . . . . . . . . . . 166

EQUATIONS OF MOTION OF ASTRODYNAMICS PROBLEMS . . . . 163

PROPERTIES OF PROBABILITY DENSITY FUNCTIONS . . . . . . . 169

THE LINEAR KALMAN FILTER . . . . . . . . . . . . . . . . . . . . . 171

VECTORIZATION OF HIGHER ORDER TENSORS . . . . . . . . . . . 180

Scalars, vectors, and matrices:

: denotes the transpose of a vector or a matrix, e.g., xT or T .

: denotes the inverse of a matrix with full-rank, e.g., 1 .

: denotes predicted value of a scalar, a vector, or a matrix from a filter, e.g., x ,

: denotes updated value of a scalar, a vector, or a matrix from a filter, e.g., x+ , m+ ,

is reserved to denote an th solution of an iteration.