Quantum Annealing

Basics and
Hidetoshi Nishimori
Tokyo Institute of Technology

Motivation

Combinatorial optimization
Minimization of cost function :
Multivariable (discrete) & Single-valued

Ground state of Ising SG

H = J ij i j

Travelling salesman problem

N!

Configuration 1

N-2

N-1

Configuration 2

Minimize the cost function (=tour length)

4

Simulated Annealing (SA)

Generic, approximate algorithm
Phase-space search by thermal fluctuations
T (t )

cN
ln t

Quantum Annealing (QA)

Generic, approximate algorithm
Phase-space search by quantum fluctuations

Problems
Quantum vs thermal fluctuations:
Is QA useful for optimization purposes?
Yes, but be careful
Related with quantum computation?
Yes, equivalent.

Aharonov et al (2007)

Implementation
H (t ) = H classical + H quantum = J ij (t )
z
i

z
j

x
i

t
8

Experiments
H = J ij iz zj ix
i

Magnetic material

(Copyrighted material here)

Brooke, Bitko, Rosenbaum, Aeppli (1999)

10

Numerical evidence
H = J ij iz zj ix
i

11

T vs : Hopfield model
H = J ij i j (Finite T )

J ij = i j
=1

(copyrighted material here)

Amit, Gutfreunt, Sompolinsky (1985)

12

T vs : Hopfield model
H = J ij
z
i

z
j

x
i

(T = 0)

(copyrighted material here)

Nishimori & Nonomura (1996)

13

Frustrated system
Ferro interaction
Antiferro interaction

(copyrighted material here)

14

Master eqn vs. Schrdinger eqn

Spin glass (SK model) with 8 spins

3
(t ) =
t
Schrdinger

(copyrighted material here)

3
T (t ) =
t
Master /Equilibrium

Kadowaki & Nishimori (1998)

15

Monte Carlo for TSP

H (t ) =

H classical + 1

H quantum

H (0) = H quantum H ( ) = H classical

cf: classical SA T (0) = large T ( ) = 0

Residual energy : H ( ) Etrue

(copyrighted material here)

Martonak, Santoro &Tosatti (2004)

16

Monte Carlo for 3SAT

C1 = x1 x2 x3
C2 = x4 x1 x2

(copyrighted material here)

QA

( xi = true or false )

QA

F = C1 C2
Can F be true?

SA

Battaglia, Santoro, Tosatti (2005)

17

Theoretical foundation
H = J ij iz zj ix
i

18

Convergence theorem
H = H classical + H quantum = J ij iz zj (t ) ix
Convergence condition

(t ) = t

c / N

Morita & Nishimori

Control parameter

(Geman-Geman for SA)

19

Computational complexity
(t ) = t

c / N

= t = exp(aN ln )

cN
T (t ) =
=
ln t

t = exp(bN ln ln )

20

Classical-quantum mapping
Classical equilibrium state

Quantum ground state

A = (T ) A (T )

H ( )
A
e
(

e H ( z ) / 2
(T ) =
Z

H q (T ) = 0
TFIM

(Quasi) equilibrium
cN
T (t ) =
ln t

H q = 1 e H / 2 M (T )e H / 2

Adiabatic condition

(t ) = t
Somma, Batista & Ortiz (2007)

c / N

21

Remark: Type of the dynamics

Convergence condition (methods of proof)

(t ) = t

c / N

Real-time Schrdinger (adiabatic condition)

Imaginary-time Schrdinger (adiabatic condition)
Quantum Monte Carlo (reduction to SA by Suzuki-Trotter)

22

Adiabatic evolution

23

Quantum adiabatic evolution

E
1

H
0 (t )
t
=
(t ) 2

m (t )

Trivial initial state

H (t ) = 1

Non-trivial final state

x
i

z
z
J

ij i j
24

Computational complexity
Finite-size analysis
Adiabatic theorem
Gap scaling
Complexity

e aN

N b
t

e 2 aN (hard)

N 2b (easy)
25

View from quantum computation

Any problem hard classically
but easy quantum mechanically?
Exact cover
Maybe
No

(Fahri, Goldstone, Gutman, Lapan, Ludgrenm, Preda, 2001)

(Young, Knysh, Smelyanskiy, 2010)

26

XORSAT
No

(Jrg, Krzakala, Semerjian, Zamponi, 2010)

(copyrighted material here)

27

p-spin ferromagnet
1
H ( s ) = sN
N

i =1

p
z
i

(1 s )
i =1

QP

1st order transition at finite s

x
i

(s = t / )
F

s
1

Exponentially small energy gap.

e aN
Exponentially large time for adiabatic computation.

ebN

Jrg, Krzakala, Kurchan, Maggs, Pujos (2010)

The problem that quantum annealing cannot solve
Seki and Nishimori (2012) The problem that quantum annealing CAN solve

28

An additional quantum term

H ( , s ) = s ( H 0
1
H0 = N
N

i =1

) + (1 s) H TF
p
z
i

, H TF = ix
i =1

H ( , s ) = s (H 0 + (1 ) H AFF ) + (1 s ) H TF
H AFF = N

1
N

i =1

x
i

Conventional case: =1
Start: s=0, =any
Goal: s=, =1
0

29

Result
p=5

p=3

1st

p=11

1st

2nd

H ( , s ) = s (H 0 + (1 ) H AFF ) + (1 s ) H TF
30

Is

H AFF

1
=N
N

i
i =1

essential?
Seoane and Nishimori (2012)

H ( , s ) = s H 0 + (1 ) H
H

(k )
AFF

1
=N
N

(k )
AFF

) + (1 s) H

TF

i =1

It works fine for k>2 as well.

31

Summary

32

Summary
QA works fine as a generic, approximate algorithm.
Better than SA.
Negative evidence for a few difficult problems.
But there should be ways to avoid 1st order transitions.

33

Collaborators
Tadashi Kadowaki
Helmut G. Katzgraber
Yoshiki Matsuda
Satoshi Morita
Yoshihiko Nonomura
Masayuki Ohzeki
Masuo Suzuki
Sei Suzuki
Yuya Seki
Beatriz Seoane

34

History
Apolloni, Carvalho, de Falco (1989)
QA, algorithmic
Finnila, Gomez, Sebenik, Stenson, Doll (1994)
Schrdinger, continuous (Lennard-Jones)
Tanaka & Horiguchi (1997)
Image restoration, algorithmic
Kadowaki & Nishimori (1998)
Transverse-field Ising, Schrdinger
Fahri, Goldstone, Gutmann, Lapan, Lundgren, Preda (2001)
Adiabatic computation, complexity , not independent of KN35