Adiabatic Quantum Algorithms

advertisement
Adiabatic Evolution Algorithm
Chukman So (19195004)
Steven Lee (18951053)
December 3, 2009 for CS C191
In this presentation we will talk about the quantum mechanical
principle of a new quantum computation technique, based on
the adiabatic approximation. Two examples of its application
is introduced, and its equivalence to tradition unitary-based
quantum computer is demonstrated.
Adiabatic Evolution Algorithm
• Formulation
– For a Hamiltonian H
– Characterised by a some parameter λ (think box size in particle-in-box problem)
– Solve the eigensystem
H ( ) n ( )  En ( ) n ( )
• Adiabatic approximation
– If the parameter is t, n (t ) does not give the right evolution
– e.g. Start from certain 3 (t  0) , at a later time, may not be 3 (t )
– But if “slow enough”, this is approximately true
time evolution
n (t  0) 
 n (t )
• If n (t  0) is a ground state of the initial H (t  0)
time evolution will yield the ground state of H (t )
Quantum Computation by Adiabatic Evolution, E. Farhi, J Goldstone et al, Los Alamos arXiv 0001106
Adiabatic Evolution Algorithm
• A tool to obtain the fulfilling assignment to a clause
– E.g. Solve A OR B
– Start with some initial H (t  0)  H i
ground state  (t  0)  0 where 0 is the ground state of Hi
– We need a final H (t  T )  H f with the fulfilling assignment as lowest energy state
AB
H  1 00 00  0 01 01  0 10 10  0 11 11  00 00
Violating assignment
Energy = 1
Fulfilling assignment
Energy = 0
(this may seem useless, but clauses like this can be added = AND’ed)
– By slowly varying H(t) from t = 0 to T, the initial ground state can be evolved into a
fulfilling final state
– But how slow?
Adiabatic Approximation
• For a time-dependent Hamiltonian H(t)
– Time-dependent Schrodinger equation

 (t )
t
– For any given time, instantaneous eigenstates can be found
H (t )  (t )  i
H (t ) n (t )  En (t ) n (t )
– We can always expend  (t ) instantaneously using these kets, treating t just as a
parameter
i t
  En ( t ') dt '
 (t )  A (t )  (t ) e 0

n
n
n
Introduced without lose of generality
Introduction to Quantum Mechanics, Bransden & Ostlie 2006
Adiabatic Approximation
– The exact functional form of An (t ) is governed by TDSE; to make use of it we need
i
  En ( t ') dt ' 


 (t )   e 0
A
'
(
t
)

(
t
)

A
(
t
)
n (t )  An (t ) n (t )
n
n
 n
t

t
n

i t

En ( t ') dt '
H (t )  (t )  e 0
A (t ) E (t )  (t )
t

n
n
 i


E
(
t
)
n




n
n
– Putting into TDSE, two terms cancel, leaving
i
t
En ( t ') dt '
m (t )  e 0
  A 'n (t ) n (t )   m (t )

n
i
t
 A 'm (t )e 0

Em ( t ') dt '
0 En (t ') dt '  A (t )   (t ) 
e
n
n
 n

t



i
t
En ( t ') dt ' 



0
 m (t )  e
A
(
t
)

(
t
)
n
 n

t


n

i
t
( En ( t ')  Em ( t ')) dt '


0
A 'm (t )    An (t )e
m (t )
n (t )

t
nm

i
t
Need to find this: TISE
Adiabatic Approximation

n (t ) for m ≠ n, we differentiate TISE on both sides by t
t


H
(
t
)

(
t
)

En (t ) n (t ) 



n
t
t


 E

 H

m 
n  H n   m  n n  En n 
t
t
 t

 t

H


m
n  Em m
n  En m
n
for m  n
t
t
t

1
H
1
H
m
n 
m
n  
m
n
t
En  Em
t
mn
t
– To find m (t )
– Putting this back, we have
A 'm (t )   An (t )
nm
1
mn
i mn dt '
H
m
n e 0
t
– So far everything is exact – no approximations
t
Adiabatic Approximation
A 'm (t )   An (t )
nm
i mn dt '
H
m
n e 0
t
t
1
mn
– Adiabatic Approximation
• Assuming the initial wavefunction is a pure eigenstate i
only one Ai (t  0)  1 , all other zero
• Assuming (a priori) at later time, other amplitudes stay small
i.e. Ai (t )  1 for all time, all other  0
(justified later by looking at the evolution)
– Then we can simplify:
A ' f i (t ) 
1
 fi
i  fi dt '
H
f
i e 0
t
t
– Integrating with time:
i  fi ( t '') dt ''
H (t ')
Af i (t )   dt '
f
i e 0
0
 fi (t ')
t '
t
1
t'
Adiabatic Approximation
i  fi ( t '') dt ''
H (t ')
Af i (t )   dt '
f
i e 0
0
 fi (t ')
t '
t'
1
t
– Now we can try to justify our a priori assumption
– a crude way to approximate the order of this integral:
ignore time dependence
Af i (t )
1
f
 fi (t )
H (t )
i
t

t
0
dt ' e
i fi ( t ) t '
 ei fi (t )t  1 
H (t )

i 
 i (t ) 
 fi (t ) f t
fi


1
Pf (t )  Af i (t )
2
H (t )

i
f
2
4
 fi (t )
t
1
H (t )

i
f
2
4
 fi (t )
t
2
H (t )
 2 4
f
i
 fi (t )
t
2
2
4
1  cos(
2
e
fi
i fi ( t ) t
(t )t ) 
1
2
Adiabatic Approximation
– i.e. For our a priori assumption to work, we require
H (t )
Pf (t )  2 4
f
i
 fi (t )
t
4
H (t )

i
2
 fi 4 (t ) f t
4
H ( )
i

( E f  Ei ) 2
f
1
f , t
2
d
dt
H ( )
i

( E f  Ei ) 2
f
2
1
Putting back  fi  E f  Ei
1
where  (t ) 
t
T
T is the total ramp time
from i to f state
T
– For adiabatic approximation to work, T must be large enough / ramp slow enough
Adiabatic Approximation
– This measure is important
• Determines how fast the computation can be performed
• Since
1
T
( E f  Ei )2
the smaller the gap is, the more likely a transition is
• The 1st excited state dominates
• T chosen wrt. smallest gap during evolution
• If states cross & matrix element  f
H ( )
i non-zero → computation fail

which makes choosing the initial H i important
What is SAT?
• Boolean satisfiability problem (SAT)
C1  C2 
 CM
– Clause: A disjunction of literals
Ck  x1   x2
– Literal: a variable or negation of variables
• Basically a huge Boolean expression, which we try to find a valid set of values
for the variables to make the given problem TRUE overall
• Adiabatic approximation setup:
– N-bit problem maps to n variables; use time evolution to solve for problem
• SAT is NP-complete
NP-complete
• Nondeterministic polynomial time (NP)
– Verifiable in polynomial time by deterministic Turing machine
– Solvable in polynomial time by nondeterministic Turing machine
• NP-complete is a class of problems having two properties:
– Being NP
– Problem (in class) can be solved quickly (polynomial time) → all NP problems can
be solved quickly as well
• Showing that a NPC problem reducing to a given NP problem is sufficient to
show the problem is NPC
• P != NP? So far most believe that is not the case, thus NP-complete problems
are at best deterministically solvable in exponential time
SAT quantum algorithm
• Create a time-dependent Hamiltonian which is a linear ramp between the
initial/starting Hamiltonian and final/problem Hamiltonian
– Idea is to, given enough time T, to slowly evolve the initial ground state (easy to
find) to final ground state (hard to find)
t

t
H  t   1   H i    H
 T
T 
H  s   1  s  H i  sH f
f
• Note n-bit SAT problems mean that the Hamiltonian we are working with exist
in a Hilbert space spanned by N = 2n basis vectors
• Thus finding ground state of problem Hamiltonian in general requires
exponential time
• Adiabatic approximation efficiency all depends on T, which is related to gmin
Quantum Computation by Adiabatic Evolution, E. Farhi, J Goldstone et al, Los Alamos arXiv 0001106
Initial Hamiltonian Hi
• Set up an initial Hamiltonian whose ground state is easy to find
k 
Hi

1
2
1   
k 
k 
with  x
x
0 1


1
0


H i k  xk  x  x xk  x
1  1
xk  0 
  and
2  1
1 1
xk  0 
 
2  1
• Noticing that 3-SAT is equivalent to SAT:
H i ,C  H B c   H B c   H B c 
i
H i   H i ,C
C
j
k
Initial Hamiltonian Hi
• Ground state for HB is xk = 0 for all kth bit
x1  0 x2  0
1
xn  0  n /2
2
  z
1
z1
z2
z2
zn
n
H i   d k H i k 
k 1
• Reason why we use a ground state in the x-axis instead of the z-axis is to
prevent gmin from becoming zero, else adiabatic approximation fails
zn
Problem Hamiltonian Hf

hC ziC , z jC , zkC

• Energy function of clause C: 0 if the bits satifsy the clause, else 1
• Total energy can be defined as sum of individual HC’s
• Hf can be defined as follows:
H f ,C  z1 z2
zn
  h z
C
iC

, z jC , zkC z1 z2
zn
H f   H f ,C
C
• Ground state is solution to SAT problem
• If no solution exists, will minimize number of violated clauses (lowest energy)
1-bit problem
• Consider a problem with one 1-bit clause satisfied with 1 bit
H i  H i1
 12
Hi   1
 2
Hf  0 0
 12 
1 
2 
1 0
Hf 

0
0


• Setup time-dependent ramped Hamiltonian
H  s   1  s  H i  sH f
1 1  s s  1
H s  

2  s 1 1 s 
• Eigenvalues:
1
E 2  E  s (1  s )  0
2
1  1  2 s (1  s)
E
2
1-bit problem
E  1  2 s (1  s )
E1 
E  2s 2  2s  1
1  1  2 s(1  s)
2
E0 
g min 
1
2
 when s  12 
1  1  2 s (1  s )
2
Quantum Computation by Adiabatic Evolution, E. Farhi, J Goldstone et al, Los Alamos arXiv 0001106
Grover Problem
• A quantum search problem
– Locate a specific entry in unstructured database 0
– Using the following notation for states
 


 ... 
 mz1  mz 2  mz 3  ...  mzn
A total of n bits, each a spin measure in z
– Given a quantum oracle Hamiltonian
1 
Hf  
0  
H f  1  0 0
if   0
if   0
Lowest energy state
– To find 0 , we start with an initial Hamiltonian for which ground state is known
n

 0 1 
1  if   
  h
Hi   
2 


i.e. Lowest state = Hadamard state
if   h
0  
Hi  1  h h
How fast is Adiabetic Quantum Computation?, W. van Dam, M. Mosca et al, Los Alamos arXiv 0206003
Grover Problem
– Linear ramp between the two Hamiltonian:
t

t
H (t )  1   H i    H f  1  s  H i   s  H f
 T
T 
 (1  s)(1  h h )  s(1  0 0 )
 1  (1  s) h h  s 0 0
– Using Adiabatic Approx.
• Solve instantaneous eigenvalues
• Find out two lowest states separation
• Get the bound on ramping rate
H ( )
i

( E f  Ei ) 2
f
T
1
E 2
where s 
t
T
Grover Problem
• Solve instantaneous eigenvalues
E  H 
E     (1  s) h h   s 0 0 
– dot both sides with h and 0
 ( E  1) h   (1  s ) h   s h 0 0 

( E  1) 0   (1  s ) 0 h h   s 0 

( E  s ) h    s h 0 0 

( E  1  s ) 0   (1  s ) 0 h h 
 ( E  s ) h    s h 0
1  s
0 h h 
E 1 s
 ( E  s )( E  1  s )  s (1  s ) 0 h
2
or h   0
corresponding to E=1 roots
need to solve this
Grover Problem
– Where the dot product 0 h
2
is well defined:
 0 1
0 h   mz1  mz 2  ... mzn  


1
2n /2
 m 0
z1
1
2
  m  0
z2

1
0 1
2
 ... 
  ...  m  0
zn
0 1 

2 
1

mz1 must be either 0 or 1
1
2n /2
– Therefore the eigenvalues are given by ( E  s)( E  1  s)  s(1  s) 0 h

E 2  E  s (1  s )  s (1  s )2 n
E 2  E  s (1  s )(1  2 n )  0
1  1  4 s(1  s)(1  2 n )
E
or 1
2
2
as
Grover Problem
• Eigenvalue spectrum, from n=2 to 20
• Against s (or time)
Energy
E 1
n-2 degenerate states
E 
1  1  4s(1  s)(1  2 n )
E
2
1st excited state
1  4s(1  s)(1  2 n )
1  1  4s(1  s)(1  2 n )
E
2
ground state
s
Grover Problem
– i.e. t
H ( )
i

( E f  Ei )2
f
1
E 2
n
where E  1  4s (1  s )(1  2 )
E
E  1  4s (1  s )(1  2 n )
s
Grover Problem
– Allowing the ramping rate to adjust to the gap
1
s 0 E 2 ds
1
1
s 0 1  4s(1  s)(1  2 n ) ds
1
1/2
1
1

ds  
du
s 0
u

1/2
1
1
1




1    s2  s   
1   u2  
4 4
4


2 1/2
1
2 1/2
1
 
du  
du
 u 0  1 1  2
 u 0  2  u 2
  u
 4 
 1   n /2
2

tan 1 
 2

2

8


T   t
1
2

1

u

s




4


  4 1  2 n  4



1
1 1 1
2 n
 
 

4 4 2n  1
4
4 1  2 n
2


– Compared with conventional search, Tconventional  total bit combinations
Tquantum
2n/2
i.e. quantum quadratic speed up
2n
Grover Problem
• Choice of initial Hamiltonian Hi is important
– Bad choice changes gap dependence → longer ramp time
– e.g if we choose
n
1
H i   1   x(i ) 
i 1 2
– Eigenvalues calculation in quant-ph/0001106
Energy
1st excited state
Tquantum 2n
No quantum speed up
ground state
s
Quantum Computation by Adiabatic Evolution, E. Farhi, J Goldstone et al, Los Alamos arXiv 0001106
Approximating adiabatic with unitaries
• Discretize the interval 0 to T into M intervals
– Unitary written as product of M factors
d
i U  t , t0   H  t  U  t , t 0 
dt
 T   U T , 0    0 
where   T / M
U  T , 0   U  T , T    U  T  , T  2  U   , 0 
• Note that we want to make the intervals small enough so that the Hamiltonian is
near-constant in each discrete interval
if
H  t1   H  t2 
1
M
t1 , t2  l ,  l  1  
then U   l  1 , l   e iH l  
Quantum Computation by Adiabatic Evolution, E. Farhi, J Goldstone et al, Los Alamos arXiv 0001106
Approximating adiabatic with unitaries
• Then we substitute the Hamiltonian with the ramped Hamiltonian between the
initial and final Hamiltonians
 H f  Hi
1
• M is T times polynomial in n
• Trotter formula for self-adjoint matrices:
 ( A B )
e
 lim  e

 A/ n  B / n n
n
e
• Thus n in the above equation needs to be large enough to be used as sufficient
approximation
Approximating adiabatic with unitaries
H  l    uHi  vH f where u  1  l  T  , v  l  T
• Then by using a large K, we can approximate using Trotter formula:
if
K
then e

M 1   Hi   H f
 iH  l  
e
 iuH i / K  ivH f / K
e


2
K
Approximating adiabatic with unitaries
• Thus the whole equation can be written in 2K terms, half being each of these
terms:
e
 iu H i K
or e
 ivH f K
• Hi is sum of n commuting 1-bit operators, so related unitary can be written as
product of n 1-qubit unitary operators
• Hf is sum of commuting operators (each for each clause), so related unitary can
be written as product of unitary operators, each acting only on qubits related to
clause
• Thus total number of factors is T2 times polynomial in n
– If T is polynomial as well, then number of factors is also polynomial
Conclusion
• We have talked about:
– Physical principle of quantum adiabatic evolution algorithm
– Its equivalence to traditional unitary quantum computation
– Its application in two examples: a one-bit SAT problem, and Glover problem
• Much like tradition QC
– Adiabatic evolution leads to quantum speed up in specialised problems
– “Smartness” is needed
• picking unitary vs picking initial Hamiltonian
– No general rule
Download