nishimori
DESCRIPTION
NishimoriTRANSCRIPT
Quantum Annealing ― Basics and …
Hidetoshi Nishimori
Tokyo Institute of Technology
2
Motivation
Combinatorial optimization
∑−= jiijJH σσ
Ground state of Ising SG
Minimization of cost function : Multivariable (discrete) & Single-valued
3
Travelling salesman problem
Minimize the cost function (=tour length)
Configuration 1 Configuration 2 N N-1
N-2 N!
4
Simulated Annealing (SA) • Generic, approximate algorithm
• Phase-space search by thermal fluctuations
tcNtTln
)( ≥t
cNtTln
)( ≥
5
Quantum Annealing (QA) • Generic, approximate algorithm
• Phase-space search by quantum fluctuations
6
Problems
¾Quantum vs thermal fluctuations:
Is QA useful for optimization purposes?
¾Related with quantum computation?
Yes, but be careful…
Yes, equivalent. Aharonov et al (2007)
7
Implementation
∑∑ Γ−−=+= xi
zj
ziij tJHHtH σσσ )()( quantumclassical
t
Γ
8
Experiments
∑ ∑Γ−−=i
xi
zj
ziijJH σσσ
9
Magnetic material
Brooke, Bitko, Rosenbaum, Aeppli (1999)
10
(Copyrighted material here)
Numerical evidence ∑ ∑Γ−−=
i
xi
zj
ziijJH σσσ
11
) (Finite ∑−= TJH jiij σσ
Amit, Gutfreunt, Sompolinsky (1985)
T vs Γ : Hopfield model
∑=
=p
jiijJ1µ
µµξξ
12
(copyrighted material here)
)0( ∑ ∑ =Γ−−= TJHi
xi
zj
ziij σσσ
Nishimori & Nonomura (1996)
T vs Γ: Hopfield model
13
(copyrighted material here)
Frustrated system
Ferro interaction Antiferro interaction
14
(copyrighted material here)
Master eqn vs. Schrödinger eqn
Kadowaki & Nishimori (1998)
Spin glass (SK model) with 8 spins
ttT 3)( =
tt 3)( =Γ
Schrödinger
Master /Equilibrium
15
(copyrighted material here)
16 16 16 16 16 16 16
Monte Carlo for TSP
true)( :energy Residual EH −τ
Martonak, Santoro &Tosatti (2004)
quantumclassical 1)( HtHttH −+=ττ
cf: classical SA 0)( large)0( =⇒= τTT
classicalquantum )()0( HHHH =⇒= τ
16
(copyrighted material here)
Monte Carlo for 3SAT
Battaglia, Santoro, Tosatti (2005)
QA
QA
SA
true?be Can
)falseor true(
21
2142
3211
FCCF
xxxxCxxxC
i
∧==
∨∨=∨∨=
17
(copyrighted material here)
Theoretical foundation ∑ ∑Γ−−=
i
xi
zj
ziijJH σσσ
18
Convergence theorem
∑∑ Γ−−=+= xi
zj
ziij tJHHH σσσ )(quantumclassical
Nctt /)( −=Γ Morita & Nishimori
(Geman-Geman for SA)
Convergence condition
Control parameter
19
20 20 20 20 20 20 20 20 20 20
)lnlnexp( ln
)(
)lnexp( )( /
δδ
δδ
bNtt
cNtT
aNttt Nc
=⇒==
=⇒==Γ −
Computational complexity
21 21 21 21 21 21 21 21 21 21 21 21 21
Classical-quantum mapping
Classical equilibrium state Quantum ground state
ZeA
AH
T∑ −
=)()( σβσ )()( TATA ψψ=
0)( =THq ψ2/2/ )(1 HH
q eTMeH ββ −−=
Adiabatic condition (Quasi) equilibrium
tcNtTln
)( = Nctt /)( −=Γ
Somma, Batista & Ortiz (2007)
∑−
=σ
σβ
σψZ
eTzH 2/)(
)(
TFIM
22 22 22
Remark: Type of the dynamics
Nctt /)( −=Γ
Convergence condition (methods of proof)
9Real-time Schrödinger (adiabatic condition)
9Imaginary-time Schrödinger (adiabatic condition)
9Quantum Monte Carlo (reduction to SA by Suzuki-Trotter)
22
Adiabatic evolution
23
24 24 24 24 24 24 24 24
Quantum adiabatic evolution
t
E
zj
ziij
xi JtttH σσ
τσ
τ ∑∑ −−−= 1)(
Trivial initial state Non-trivial final state
∆ εψψ
=∆
∂∂
2
0
)(
)()(
t
tt
Htm
τ1
24
25 25 25 25 25 25 25 25 25 25 25 25 25
Computational complexity
Adiabatic theorem 2−∆∝τ
∝∆−
−
b
aN
Ne
Gap scaling
Finite-size analysis
Complexity ∝(easy) (hard)
2
2
b
aN
Ne
τ
E
t
25
26 26 26 26 26 26 26 26 26 26 26 26 26 26 26
View from quantum computation
Any problem hard classically but easy quantum mechanically?
Maybe (Fahri, Goldstone, Gutman, Lapan, Ludgrenm, Preda, 2001)
No (Young, Knysh, Smelyanskiy, 2010)
“Exact cover”
26
“XORSAT”
No (Jörg, Krzakala, Semerjian, Zamponi, 2010)
27
(copyrighted material here)
28 28 28 28 28 28 28 28 28 28 28
p-spin ferromagnet
( ) )/( 11
)(11
τσσ tssN
sNsHN
i
xi
pN
i
zi =−−−= ∑∑
==
Jörg, Krzakala, Kurchan, Maggs, Pujos (2010) “The problem that quantum annealing cannot solve”
• 1st order transition at finite s
• Exponentially small energy gap.
• Exponentially large time for adiabatic computation.
aNe−∝∆
bNe∝τ
0 1 s
QP F
Seki and Nishimori (2012) “The problem that quantum annealing CAN solve” 28
29 29 29 29 29 29 29 29 29 29 29 29 29
An additional quantum term
( ) TFAFF0 )1()1(),( HsHHssH −+−+= λλλ2
1AFF
1= ∑
=
N
i
xiN
NH σ
Conventional case: λ=1
Start: s=0, λ=any Goal: s=1, λ=1
( ) TF0 )1( ),( HsHssH −+=λ
∑∑==
−=−=N
i
xi
pN
i
zi H
NNH
1TF
10 ,
1 σσ
s
λ 0
1
1
0 29
30 30 30 30 30 30 30 30 30 30
Result
1st 1st 2nd
p=3 p=5 p=11
( ) TFAFF0 )1()1(),( HsHHssH −+−+= λλλ30
Is essential?
kN
i
xi
k
NNH = ∑
=1
)(AFF
1 σ
( ) TF)(
AFF0 )1()1(),( HsHHssH k −+−+= λλλ
It works fine for k>2 as well.
Seoane and Nishimori (2012)
2
1AFF
1= ∑
=
N
i
xiN
NH σ
31
Summary
32
33 33 33 33 33 33 33 33 33 33 33 33
Summary
9QA works fine as a generic, approximate algorithm.
9“Better” than SA.
9Negative evidence for a few difficult problems.
9But there should be ways to avoid 1st order transitions.
33
34 34 34 34 34 34 34 34 34 34 34
Collaborators
zTadashi Kadowaki
zHelmut G. Katzgraber
zYoshiki Matsuda
zSatoshi Morita
zYoshihiko Nonomura
zMasayuki Ohzeki
zMasuo Suzuki
zSei Suzuki
zYuya Seki
zBeatriz Seoane
34
35 35 35 35 35 35 35 35 35 35 35 35 35 35
History
9Apolloni, Carvalho, de Falco (1989)
“QA”, algorithmic
9Finnila, Gomez, Sebenik, Stenson, Doll (1994)
Schrödinger, continuous (Lennard-Jones)
9Tanaka & Horiguchi (1997)
Image restoration, algorithmic
9Kadowaki & Nishimori (1998)
Transverse-field Ising, Schrödinger
9Fahri, Goldstone, Gutmann, Lapan, Lundgren, Preda (2001)
“Adiabatic computation”, complexity , not independent of KN