Analysis of High-Fidelity Quantum Driving

The Hamiltonian used is a simple Landau-Zener:

HLZ = ( )z + ( )z
where is a re-scaled time: =

t
.
T

We can write H as:



HLZ

( ) ( )
=
( ) ( )

and we have the eigenvalue-eigenvector pairs:

2 +2
p
E1 ( ) = 2 + 2

|1 ( )i = N ORM

1
+2 +2
E2 ( ) =

p
2 + 2

|2 ( )i = N ORM

where N ORM is a (rather complicated) normalization factor.

Note that this means the minimum energy gap is:
p
E = 2 2 + 2
In the Analysis of the paper, we have ( ) = 4Erec ( 1/2), where Erec is the characteristic energy-scale of the system. Also, is treated to be a constant. The minimum
energy gap occurs when = 0, i.e. E = 2.
For now, assume and both vary. Then according to the paper, the general form of
Hs is as follows:
Hs = i
h

X |t n(t)ihn(t)| |n(t)iht n(t)|

n

This is still a mystery. I have plugged it into the adiabatic approximation, with/without
the assumption that the eigenstates are the same after the introduction of the auxiliary
Hamiltonian. Dont know why it works.

Analysis of High-Fidelity Quantum Driving

The paper claims that in fact,
Hs =

d
h
y
2 dt

where = arctan((t)/(t)). This is accurate, as checked in the Mathematica Workbook.

Note that:

h
Hs =
2

2(2 + 2 )

!
y

Now consider Quantum driving protocols for a two-level system: from generalized
Landau-Zener sweeps to superadiabatic control. (i.e. MBV (Malossi, Bason, Viteau))
Recall in Roland & Cerf that to get
|hE0 ; T |(T )i|2 1 2
, we require that at EACH TIME,



1  dH

i
h
1,0


g 2 dt
which, optimally, requires that at EACH TIME:


ds
N 1
= g 2 (t) = 1 4
s(1 s)
dt
N
which, upon integration, gives:
t=

h


i

N
1

tan1
N 1(2s 1) + tan1 N 1
2 N 1

where N = 2n .
Now then, to express the timescale T , as a function of , we require:
s=1
Z 1


N
dt
ds = t(s)
=
tan1 N 1
T () =
N 1
0 ds
s=0
which gives

for N = 2 (n = 1).

HOWEVER: The Hamiltonian used in Roland and Cerf is DIFFERENT to the Hamiltonian used in MBV (Malossi Bason Viteau).

NOTE : s in Roland-Cerf is different from in MBV. is simply a rescaled time:

= t/T . However, s is an arbitrary function of time. Therefore, (t) is always linear
and ( ) may not be. However, s(t) need not be linear, but (s) will be linear.
H = ( )z + ( )x
I want to try and prove that:
T () =

1
1

4 + 2

for this Hamiltonian.

We have the eigenvalue/eigenvector pairs as shown before, and
p
g( ) = E = 2 2 ( ) + 2 . Then:

 
  d 
1  dH
 
h
i
1,0
  dt 
g 2  d
 
 d 
1
g 2

=
=    dH
h i1,0 
T
dt
d

 dH
h i1,0 
d
= T
g 2
But recall that



d ( )
( )

= HLZ =
HLZ ( ) =

( )
0
( )
d


since we have assumed is constant in RC protocol.

CAREFUL: Here, the subscripts 1, 0 dont refer to the diabatic states |0i and |1i, but
rather, the adiabatic states derived earlier (which happen to be written one index higher
for no reason):
2 +2
p

E1 ( ) = 2 + 2
|1 ( )i = N ORM
1
+2 +2
E2 ( ) =

p
2 + 2

|2 ( )i = N ORM

Hence,


 dH  
)
(



= p 2
 d 1,0 
+ ( )2
(see Mathematica Notebook)

Analysis of High-Fidelity Quantum Driving

So:
1
T () 2
g

p
2 + ( )2

1
=
4(( )2 + 2 )
=

)
(

)
(

p
2 + ( )2

)
(
4(( )2 + 2 )3/2

As far as I understand it,

LZ Protocol - Just follow the evolution linearly
RC Protocol - Follow the ground state at optimal pace
Super-adiabatic - Introduce aux. Hamiltonian to counter transitions. Can be
absorbed into and , resulting in 0 and 0 , which just changes how
fast they change. (I dont know why this is sometimes called the super-adiabatic
linear protocol)
Super-adiabatic Tangent - Same auxiliary Hamiltonian, just a difference with how
it is absorbed. doesnt change, but 0 .
Composite Pulse Protocol - Extreme Limit of super-adiabatic protocol, closest
to quantum speed limit. Basically two delta pulses with half a Rabi oscillation
sandwiched in-between.
See Mathematica notebook for graphing super-adiabatic and . The graphs dont
quite match the ones in the paper.
Q: Yes, we know the equation for 0 in the super-adiabatic protocols, but how does ( )
itself vary? After all, it is allowed to vary in the super-adiabatic protocols.
According to MBV, For both of these [superadiabatic] protocols the discontinuities in
the derivative of at = 0 and = 1 lead to delta-functions in 0 at the beginning and
end of the time evolution.. I dont see this.
Do not understand Eq(4) in HFQD.

Getting 0 and 0
In HFQD they say they use an optical lattice potential of the form:


V0
2x
+ (t)
cos
2
dL
but then in the Methods section where they find 0 and 0 , this is replaced with a
cos2 potential for some reason.
Now back to HFQD, and the attempt to absorb the auxiliary Hamiltonian. Noting
that:
  i

 i
e4
0
0
e 4
(dL /4)x U
(dL /4)

x
=U
y =
0
e4i
0 e 4 i
the auxiliary Hamiltonian can be obtained simply by adding a shifted potential.
Combining two period potentials:
a cos2 x + b cos2 (x c) := d cos2 (x f ) + g
By collecting terms in cos 2x, sin 2x, and constants, we get:
a b cos 2c
+
=
2
2
b
sin 2c =
2
a b
+ =
2 2

d
cos 2f
2
d
sin 2f
2
d
+g
2

and the solution is:

p
a2 + b2 + 2ab cos 2c
a + b 1p 2
g=

a + b2 + 2ab cos 2c
2
2


1
b sin 2c
1
f = tan
2
a + b cos 2c
d=

In the paper, they claim that

!

dL
(x

)
x
4
Vtot (x) = V0 cos2
+ 4 cos2
dL
dL
s 
!
dL
(x

)
V0 2
2
= const. + 2
+ (2)2 cos2
2
dL


Ok, this makes sense. The 2ab cos 2c bit is equal to zero in this case, since c = 4 .

Analysis of High-Fidelity Quantum Driving

I think its true that the eigenstates/eigenvalues of H + Hs will be different to the

eigenstates/eigenvalues of H. The very fact that you can absorb Hs into 0 and 0 does this mean that in fact H and H + Hs share the same eigenspectrum? A: NO. 0
and 0 vary differently to each other, so eigenspectrum at a time will not necessarily
be the same.
Tried to find Hs for
H = ( )(z(1) + z(2) ) + ( )(x(1) + x(2) )
But the algebra is extremely complicated, see Mathematica Workbook.
Problem: Deriving 0 and 0 for one qubit relies on physical intuition related to translation operators on optical lattices. Would be nice to find some physical-situationindependent way to get 0 or 0 .
I think the superadiabatic linear protocol has a typo in both MBV and HFQD.
Dunno what all this talk about discontinuities is. Im not seeing them.