Anders W Sandvik, Boston University - Ulm...624. WE-Heraeus-Seminar, Bad Honnef, September 19-22,...
Transcript of Anders W Sandvik, Boston University - Ulm...624. WE-Heraeus-Seminar, Bad Honnef, September 19-22,...
624. WE-Heraeus-Seminar, Bad Honnef, September 19-22, 2016 Simulating Quantum Processes and Devices
Quantum annealing in imaginary time
Anders W Sandvik, Boston University
C. De Grandi, A. Polkovnikov, A. W. Sandvik, PRB 84, 224303 (2011) C.-W. Liu, A. Polkovnikov, A. W. Sandvik, PRB 87, 174302 (2013)C.-W. Liu, A. Polkovnikov, A. W. Sandvik, PRB 89, 054307 (2014)C.-W. Liu, A. Polkovnikov, A. W. Sandvik, PRL 114, 147203 (2015)
Cheng-Wei Liu (BU → HP)Anatoli Polkovnikov (BU)
Na Xu (BU)
OutlineIntroduction- Quantum annealing for quantum computing/optimization- Transverse-field Ising model, quantum phase transition
- Kibble-Zurek velocity scaling- Demonstration for classical Ising model
Dynamic scaling at phase transitions
- Exact numerical solutions for small systems- Dynamical quantum Monte Carlo in imaginary time
Quantum dynamics in real and imaginary time
- Scaling of success probability in real and imaginary time- Problem hardness in real and imaginary time
[Going all the way to the classical spin-glass ground state]
- Imaginary-time QMC and KZ scaling- Implications for quantum computing
Quantum spin glass transition
h
Tglass phase
Thermal and Quantum Annealing
Adiabatic Theorem:For small v, the system stays in the ground state of H[s(t)]
Useful paradigm for quantum computing?
Can quantum annealing be more efficient than thermal annealing?Kadowaki, Nishimory (PRE 1998), Farhi et a (Science 2002),….
Simulated (Thermal) AnnealingReduce T as a function of time in a Monte Carlo simulation- efficient way to equilibrate a simulation- powerful as optimization algorithm
Reduce quantum fluctuations as a function of time- start with simple quantum system H0 (s=0):- end with a complicated classical potential H1 (s=1)
Quantum Annealing
s = s(t) = vt, v = 1/tmax
H(s) = (1� s)H0 + sH1 [H0, H1] 6= 0
Quantum Annealing & Quantum ComputingThe D-wave “quantum annealer”; ~1000 flux qubits- Claimed to solve some hard optimization problems- Is it really doing quantum annealing?
- Is quantum annealing really better than simulated annealing (on a classical computer)?
Hamiltonian implemented in D-wave quantum annealer....
Hamiltonian of the D-Wave Device
→ Studies of dynamics of transverse-field Ising models
||Matthias Troyer
Performance of simulated annealing, simulated quantum annealing and D-Wave on hard spin glass instances
1
9
0 1 2 3
4 5 6 7
8 9 10
12 13 14
16 18 19
20 21 22 23
24 25 26 27
28 29 30 31
32 33 34 35
36 37 38 39
40 41 42 43
44 45 46 47
48 49 50 51
52 53 54 55
56 57 58 59
60 61 62 63
64 65 67
68 69 70 71
72 73 74 75
76 77 78 79
80 81 82 83
85 86 87
88 89 90 91
92 93 94 95
96 97 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
116
117
118
119
120
121
122
123
124
125
126
127
128
130
131
132
133
135
136
137
138
139
140
141
144
145
146
147
148
149
150
151
152
153
154
155
156
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
304
305
306
307
308
309
310
312
313
314
315
316
317
318
319
320
321
323
324
325
326
327
328
329
330
331
332
333
334
336
337
338
339
340
341
342
343
344
345
347
348
349
350
351
352
353
354
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
388
389
390
391
392
393
394
395
396
397
398
400
401
402
403
404
405
406
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
428
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
460
461
462
463
464
465
466
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
493
494
495
496
497
498
499
500
501
502
504
505
506
507
508
509
510
511
1115
17
66
84
98
115
129
134
142
143
157
227
249
274
303
311
322
335
346
355
387
399
407
424
425
426
427
429
430
431
459
467
491
492
503
FIG. 6. The 476-qubit DW2 device architecture andqubit connectivity.
Appendix C: The D-Wave Two Chip
1. The Chimera
The Chimera graph of the D-Wave Two (DW2) chipused in this study is shown in Fig. 6. The chip is an8⇥8 array of unit cells where each unit cell is a balancedK4,4 bipartite graph of superconducting flux qubits. Inthe ideal Chimera graph the degree of each (internal)vertex is 6. On our chip, only a subset of 476 qubits, isfunctional. The temperature of the device ⇠ 15mK.
The chip is designed to solve a very specific typeof problems, namely, Ising-type optimization problemswhere the cost function is that of the Ising Hamiltonian[see Eq. (1) of the main text]. The Ising spins, s
i
= ±1are the variables to be optimized over and the sets {J
ij
}and {h
i
} are the programmable parameters of the costfunction. In addition, hiji denotes a sum over all activeedges of the Chimera graph.
2. The D-Wave Two annealing schedule
The DW2 performs the annealing by implementing thetime-dependent Hamiltonian
H(t) = �A(t)X
i
�x
i
+B(t)HIsing , (C1)
with t 2 [0, ta] where the allowed range of annealing timesta, due to engineering restrictions, is between 20µs and20ms. The annealing schedules A(t) and B(t) used inthe device are shown in Fig. 7. There are four anneal-ing lines, and their synchronization becomes harder forfaster annealers. The filtering of the input control linesintroduces some additional distortion in the annealingcontrol.
FIG. 7. Annealing schedule of the D-Wave chip. Thefunctions A(t) and B(t) are the amplitudes of the (transverse-field) driver and classical Ising Hamiltonians, respectively.Also shown is the temperature in units of energy (kB = 1).
3. Gauge averaging on the D-Wave device
Calibration inaccuracies stemming mainly from thedigital to analog conversions of the problem parameters,cause the couplings J
ij
and hi
realized on the DW2 chipto be slightly di↵erent from the intended programmedvalues (with a typical ⇠ 5% variation). Therefore, in-stances encoded on the device will be generally di↵erentfrom the intended instances. Additionally, other, moresystematic biases exist which cause spins to prefer oneorientation over another regardless of the encoded pa-rameters. To neutralize these e↵ects, it is advantageousto perform multiple annealing rounds (or ‘programmingcycles’) on the device, where each such cycle correspondsto a di↵erent encoding or ‘gauge’ of the same probleminstance onto the couplers of the device [9]. To real-ize these di↵erent encodings, we use a gauge freedomin realizing the Ising spin glass: for each qubit we canfreely define which of the two qubits states correspondsto s
i
= +1 and si
= �1. More formally this correspondsto a gauge transformation that changes spins s
i
! ⌘i
si
,with ⌘
i
= ±1 and the couplings as Jij
! ⌘i
⌘j
Jij
andhi
! ⌘i
hi
. The simulations are invariant under such agauge transformation, but due to calibration errors whichbreak the gauge symmetry, the results returned by theDW2 are not.
4. Performance of the DW2 chip as a function ofannealing time
Since the DW2 chip is a putative quantum annealer, itis only natural to ask how its performance, namely, thetypical time-to-solution ts it yields, depends on annealingtime ta. Ideally, the longer ta is, the better the perfor-mance we expect [52]. However, in practice, decoheringinteractions with the environment are present which be-
D-wave Chimera-lattice setup, picture fromMartin-Mayor & Hen, arXiv:1502.02494
Interactions Jij are programmable- restricted to “Chimera lattice”
Solves optimization problems mappedonto frustrated Ising model
H1 =NX
i=1
NX
j=1
Jij�zi �
zj , �z
i 2 {�1,+1}
Tune the strength of the field
adiabatically from s=0 to s=1 s = s(t) = vt, v = 1/t
max
H(s) = (1� s)H0 + sH1
H0 = �NX
i=1
�x
i
= �NX
i=1
(�+i
+ ��i
)
[H0, H1] 6= 0
Quantum Phase Transition
Ground state changes qualitatively as s changes- trivial (easy to prepare) for s=0- complex (solution of hard optimization problem) at s=1
→ expect a quantum phase transition at some s=sc
- trivial x-oriented ferromagnet at s=0 (→→→)- z-oriented (↑↑↑or ↓↓↓, symmetry broken) at s=1- sc=1/2 in 1D, appr. 0.25 in 2D
as in the clean transverse-field Ising ferromagnet
(N ! 1)
How long does it take (versus problem size N)?
Have to pass through sc and beyond adiabatically
H(s) = �sX
hiji
�z
i
�z
i+1 � (1� s)NX
i=1
�x
i
One can expect a quantum phase transition in the system
H(s) = (1� s)H0 + sH1 [H0, H1] 6= 0
Quantum Dynamics
What can imaginary time tell us about real-time dynamics?
Time evolution | (t)i = U(t, t0)| (t0)i
Difficult to study numerically for a many-body system- exact diagonalization of small systems- DMRG/MPS/TEBD for 1D systems (moderate sizes and times)
| (⌧)i = U(⌧, ⌧0)| (⌧0)i U(⌧, ⌧0) = T⌧exp
�Z ⌧
⌧0
d⌧ 0H[s(⌧ 0)]
�
Can be implemented in Quantum Monte CarloDe Grandi, Polkovnikov, Sandvik, PRB2011
Schrödinger dynamics in imaginary time t=iτAlternative approach:
Time evolution operator with time-dependent H
U(t, t0) = Ttexp
i
Z t
t0
dt0H[s(t0)]
�
Real and imaginary time quantum dynamics
How different? Which one is more adiabatic?
Example: linear ramp of transverse-field Ising ferromagnet
Start from eigenstate of H(s=0) at t=0
- Instantaneous ground state: | 0(t)i = | 0(s[t])i
- Actual state during evolution: | (t)i
Distance between these states given by log-fidelity
� ln[F (t)] = �1
2ln�|h 0(t)| (t)i|2
�
Integrate Schrödinger equation numerically for small L- compare real and imaginary time
2D square-lattice system; N=L2
H(s) = �sX
hiji
�z
i
�z
j
� (1� s)NX
i=1
�x
i
s 2 [0, 1], s = vt
Example: 4×4 lattice
Main peakreflects quantumphase transitionat Sc≈0.25
0 0.2 0.4 0.6 0.8 1t/tmax
0.00
0.02
0.04
0.06
0.08
0.10
-ln(F)
real time imaginary time
v = 1/40
Example: 4×4 lattice
Main peakreflects quantumphase transitionat Sc≈0.25
Imaginary timemore efficient inreaching ground state for s→1
0 0.2 0.4 0.6 0.8 1t/tmax
0.000
0.005
0.010
0.015
0.020
0.025
-ln(F)
real time imaginary time
v = 1/80
Example: 4×4 lattice
Main peakreflects quantumphase transitionat Sc≈0.25
Imaginary timemore efficient inreaching ground state for s→1
0 0.2 0.4 0.6 0.8 1t/tmax
0.000
0.001
0.002
0.003
0.004
0.005
-ln(F)
real time imaginary time
v = 1/160
Example: 4×4 lattice
Dynamic exponent z is same in real and imaginary time De Grandi, Polkovnikov, Sandvik, PRB 2011
Differences between real and imaginary time are of order v2 or v3
(depends on observable)
Same dynamic susceptibilitiesaccessed in realand imaginarytime
0 0.2 0.4 0.6 0.8 1t/tmax
0.0000
0.0002
0.0004
0.0006
0.0008
0.0010
0.0012
-ln(F)
real time imaginary time
v = 1/320
Use imaginary time for large systems
Quantum Monte Carlo Algorithm
Simpler scheme: evolve with just a H-product(Liu, Polkovnikov, Sandvik, PRB 2013)
Implemented in quantum Monte Carlo as:
| (⌧)i =1X
n=0
Z ⌧
⌧0
d⌧n
Z ⌧n
⌧0
d⌧n�1 · · ·Z ⌧2
⌧0
d⌧1[�H(⌧n)] · · · [�H(⌧1)]| (0)i
| (⌧)i = U(⌧, ⌧0)| (⌧0)i
Schrödinger dynamic in imaginary time t=iτ
U(⌧, ⌧0) = T⌧exp
�Z ⌧
⌧0
d⌧ 0H[s(⌧ 0)]
�
How is this method implemented?
| (sM )i = H(sM ) · · ·H(s2)H(s1)| (0)i, si = i�s, �s =sMM
Time unit is ∝1/N, velocity is v / N�s
Difference in v-dependence between product evolution and imaginary-time Schrödinger dynamics is O(v2)- same critical scaling behavior, dynamic susceptibilities
h (0)|H(s1) · · ·H(s7)|H(s7) · · ·H(s1)| (0)i
QMC Algorithm Illustration
H1(i) = �(1� s)(�+i + ��
i )
H2(i, j) = �s(�zi �
zj + 1)
Transverse-field Ising model: 2 types of operators:
12345677654321 12345677654321 1234567765432112345677654321
Represented as “vertices”
Simple extension of ground-state projector QMC (fixed H)
Analyze results versus velocity v and system size
12345677654321
MC sampling of networks of vertices
N = 4
M = 7
Dynamic QMC Illustration
Computational advantage:All s=values in one simulation!
“Asymmetric” expectation values
hAik = h (0)|1Y
i=M
H(si)MY
i=k
H(si)AkY
i=1
H(si)| (0)ihAik = h (0)|1Y
i=M
H(si)MY
i=k
H(si)AkY
i=1
H(si)| (0)i
Collect data, do scaling analysis...
Same leading-order (in v)behavior as conventionalexpectation values
Test on clean 2D Ising modelin transverse field
Using H-product dynamics
Animation of single configuration
Dynamic Critical Exponent and GapDynamic exponent z at a phase transition- relates time and length scales
Continuous quantum phase transition- excitation gap at the transition depends on the system size and z as
� ⇠ 1
Lz=
1
Nz/d, (N = Ld)
At a continuous transition (classical or quantum):- large (divergent) correlation length
⇠r ⇠ |�|�⌫ , ⇠t ⇠ ⇠zr ⇠ |�|�⌫z δ = distance from critical point (in T or other param)
Classical (thermal) phase transition- Fluctuations regulated by temperature T>0Quantum (ground state, T=0) phase transition- Fluctuations regulated by parameter g in Hamiltonian
Classical and quantum phase transitions
There are many similarities between classical and quantum transitions- and also important differences
ord
er p
aram
eter
T [g]Tc [gc]
(a)
ord
er p
aram
eter
T [g]Tc [gc]
(b)
FIGURE 3. Temperature (T ) or coupling (g) dependence of the order parameter (e.g., the magnetizationof a ferromagnet) at a continuous (a) and a first-order (b) phase transition. A classical, thermal transitionoccurs at some temperature T = Tc, whereas a quantum phase transition occurs at some g= gc at T = 0.
where the spin correlations decay exponentially with distance [60].While quasi-1D antiferromagnets were actively studied experimentally already in the
1960s and 70s, these efforts were further stimulated by theoretical developments in the1980s. Haldane conjectured [55], based on a field-theory approach, that the Heisenbergchain has completely different physical properties for integer spin (S = 1,2, . . .) and“half-odd integer” spin (S = 1/2,3/2, . . .). It was known from Bethe’s solution that theS = 1/2 chain has a gapless excitation spectrum (related to the power-law decayingspin correlations). Haldane suggested the possibility of the S= 1 chain instead having aground state with exponentially decaying correlations and a gap to all excitations; a kindof spin liquid state [26]. This was counter to the expectation (based on, e.g., spin wavetheory) that increasing S should increase the tendency to ordering. Haldane’s conjecturestimulated intense research activities, theoretical as well as experimental, on the S = 1Heisenberg chain and 1D systems more broadly. There is now completely conclusiveevidence from numerical studies that Haldane was right [61, 62, 63]. Experimentally,there are also a number of quasi-one-dimensional S = 1/2 [64] and S = 1 [65] (andalso larger S [66]) compounds which show the predicted differences in the excitationspectrum. A rather complete and compelling theory of spin-S Heisenberg chains hasemerged (and includes also the VBS transitions for half-odd integer S), but even to thisdate various aspects of their unusual properties are still being worked out [67]. There arealso many other variants of spin chains, which are also attracting a lot of theoretical andexperimental attention (e.g., systems including various anisotropies, external fields [68],higher-order interactions [69], couplings to phonons [70, 71], long-range interactions[72, 73], etc.). In Sec. 4 we will use exact diagonalization methods to study the S= 1/2Heisenberg chain, as well as the extended variant with frustrated interactions (and alsoincluding long-range interactions). In Sec. 5 we will investigate longer chains using theSSE QMC method. We will also study ladder-systems consisting of several coupledchains [9], which, for an even number of chains, have properties similar to the Haldanestate (i.e., exponentially decaying spin correlations and gapped excitations).
2.4. Models with quantum phase transitions in two dimensions
The existence of different types of ground states implies that phase transitions canoccur in a system at T = 0 as some parameter in the hamiltonian is varied (which
146
Downloaded 27 Feb 2012 to 128.197.40.223. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
In both cases phase transitions can be- first-order (discontinuous): finite correlation length ξ as g→gc or g→gc- continuous: correlation length diverges, ξ~|g-gc|-ν or ξ~|T-Tc|-ν
3
Classical (thermal) phase transition- Fluctuations regulated by temperature T>0Quantum (ground state, T=0) phase transition- Fluctuations regulated by parameter g in Hamiltonian
Classical and quantum phase transitions
There are many similarities between classical and quantum transitions- and also important differences
ord
er p
aram
eter
T [g]Tc [gc]
(a)
ord
er p
aram
eter
T [g]Tc [gc]
(b)
FIGURE 3. Temperature (T ) or coupling (g) dependence of the order parameter (e.g., the magnetizationof a ferromagnet) at a continuous (a) and a first-order (b) phase transition. A classical, thermal transitionoccurs at some temperature T = Tc, whereas a quantum phase transition occurs at some g= gc at T = 0.
where the spin correlations decay exponentially with distance [60].While quasi-1D antiferromagnets were actively studied experimentally already in the
1960s and 70s, these efforts were further stimulated by theoretical developments in the1980s. Haldane conjectured [55], based on a field-theory approach, that the Heisenbergchain has completely different physical properties for integer spin (S = 1,2, . . .) and“half-odd integer” spin (S = 1/2,3/2, . . .). It was known from Bethe’s solution that theS = 1/2 chain has a gapless excitation spectrum (related to the power-law decayingspin correlations). Haldane suggested the possibility of the S= 1 chain instead having aground state with exponentially decaying correlations and a gap to all excitations; a kindof spin liquid state [26]. This was counter to the expectation (based on, e.g., spin wavetheory) that increasing S should increase the tendency to ordering. Haldane’s conjecturestimulated intense research activities, theoretical as well as experimental, on the S = 1Heisenberg chain and 1D systems more broadly. There is now completely conclusiveevidence from numerical studies that Haldane was right [61, 62, 63]. Experimentally,there are also a number of quasi-one-dimensional S = 1/2 [64] and S = 1 [65] (andalso larger S [66]) compounds which show the predicted differences in the excitationspectrum. A rather complete and compelling theory of spin-S Heisenberg chains hasemerged (and includes also the VBS transitions for half-odd integer S), but even to thisdate various aspects of their unusual properties are still being worked out [67]. There arealso many other variants of spin chains, which are also attracting a lot of theoretical andexperimental attention (e.g., systems including various anisotropies, external fields [68],higher-order interactions [69], couplings to phonons [70, 71], long-range interactions[72, 73], etc.). In Sec. 4 we will use exact diagonalization methods to study the S= 1/2Heisenberg chain, as well as the extended variant with frustrated interactions (and alsoincluding long-range interactions). In Sec. 5 we will investigate longer chains using theSSE QMC method. We will also study ladder-systems consisting of several coupledchains [9], which, for an even number of chains, have properties similar to the Haldanestate (i.e., exponentially decaying spin correlations and gapped excitations).
2.4. Models with quantum phase transitions in two dimensions
The existence of different types of ground states implies that phase transitions canoccur in a system at T = 0 as some parameter in the hamiltonian is varied (which
146
Downloaded 27 Feb 2012 to 128.197.40.223. Redistribution subject to AIP license or copyright; see http://proceedings.aip.org/about/rights_permissions
In both cases phase transitions can be- first-order (discontinuous): finite correlation length ξ as g→gc or g→gc- continuous: correlation length diverges, ξ~|g-gc|-ν or ξ~|T-Tc|-ν
3
Exponentially small gap at a first-order(discontinuous) transition
� ⇠ e�aL
Exactly how does z enter in the adiabatic criterion?
Important issue for quantum annealing!P. Young et al. (PRL 2008)
Kibble-Zurek Velocity and ScalingThe adiabatic criterion for passing through a continuous phase transition involves exponents z and ν:
Kibble 1978- defects in early universeZurek 1981- classical phase transitionsPolkovnikov 2005 + others- quantum phase transitions
Same criterion for classicaland quantum phase transitions- adiabatic (quantum)- quasi-static (classical)
A(�, v,N) = N�/⌫0g(�N1/⌫0
, vNz0+1/⌫0), ⌫0 = d⌫, z = z/d
Will use for spin glasses of interest in quantum computing
Apply to well-understood classical system first...
Must have v < vKZ, withvKZ ⇠ L�(z+1/⌫)
Generalized finite-size scaling hypothesis
A(�, v, L) = L�/⌫g(�L1/⌫ , vLz+1/⌫)δ = distance from critical point (in T or other param)
Classical simulated annealing
Repeat many times, collect averages, analyze,....
2D classical IsingMetropolis MC
simulations
100 102 104 106 108
v Lz+1/ν
10-4
10-3
10-2
10-1
100
<m
2 > L
2β/ν
L = 12L = 24L = 48L = 64L = 96L = 128L = 192L = 256L = 500L = 1024polynomial fitpower-law fit
104 105 10610-3
10-2L = 128
Velocity Scaling, 2D Ising ModelRepeat process many times, average data for T=Tc
Used known 2D Ising exponentsβ=1/8, ν=1
Result: z ≈2.1767(5)consistent withvalues obtainedin other ways
Adjusted z foroptimal scalingcollapse
Liu, Polkovnikov,Sandvik, PRB 2014
Can also be done for quantum systems in imaginary time
hm2(� = 0, v, L)i = L�2�/⌫f(vLz+1/⌫)
0.15 0.20 0.25 0.30S
0.0
0.2
0.4
0.6
0.8
1.0
U L = 12L = 16L = 32L = 48L = 56L = 60
0.24 0.25
0.8
0.9
2D Transverse-Ising, Scaling Example
A(�, v, L) = L�/⌫g(�L1/⌫ , vLz+1/⌫)
Example: Binder cumulant
Step function should form,jump from U=0 to 1 at sc
- crossing points for finite system size
Do similar studies for quantum spin glasses
If z, ν known, sc not: use
vLz+1/⌫= constant
for 1-parameter scaling
z = 1, ⌫ ⇡ 0.70
N=8
3-regular graphs with anti-ferro couplingsN spins, randomly connected, coordination-number 3
- sc ≈ 0.37 from quantum cavity approximation- QMC consistent with this sc, power-law gaps at sc
The quantum model was studied by Farhi, Gosset, Hen, Sandvik, Shor, Young, Zamponi, PRA 2012
More detailed studies with quantum annealing
Analyze <q2> using QMC and velocity scaling
Edwards-Anderson spin-glass order parameter
q =1
N
NX
i=1
�zi (1)�
zi (2)
(1) and (2) are independent simulations (replicas)
Classical model has mean-field glass transition- Tc known exactly (Krazakala et al.)
Extracting Quantum-glass transition
Using Binder cumulantU(s, v,N) = U [(s� sc)N
1/⌫0, vNz0+1/⌫0
]
But now we don’t know the exponents. Use
v / N�↵, ↵ > z0 + 1/⌫0
- do several α- check for consistency
Consistent with previouswork, but smaller errors
Next, critical exponents...
sc = 0.3565 +/- 0.0012
Best result for α=17/12
4
0.000 0.001 0.002 0.003 0.004
1/N
0.335
0.340
0.345
0.350
0.355
s c(N)
0.33 0.34 0.35 0.36
s
0.0
0.1
0.2
0.3
0.4
U
192
256
320
384
448
512
576
N
FIG. 2: (Color online) Crossing points between Binder cumu-lants for 3-regular graphs with N and N+64 spins, extractedusing the curves shown in the inset. The results were obtainedin quenches with v ⇠ N�↵ for ↵ = 17/12. The curve in themain panel is a power-law fit for extrapolating sc.
were in good agreement with this estimate. The errorbars on these calculations is several percent.
We find sc
using r = 1 QAQMC with v / N�↵, where↵ exceeds the KZ exponent z0 + 1/⌫0 (which is unknownbut later computable for a posteriori verification). Thenhq2i ⇠ N�2�/⌫0
at sc
because f(x) in Eq. (8) approachesa constant when x ! 0. As illustrated in Fig. 2, quench-ing past the estimated s
c
, we use a curve crossing analysisof the Binder cumulant, U = (3� hq4i/hq2i2)/2, and ob-tain s
c
= 0.3565(12). This value agrees well with theprevious results but has smaller uncertainty.
Performing additional quenches to sc
using protocolswith both r = 1 and r = 2/3 in Eq. (5) we extract allthe critical exponents. An example of scaling collapsefor r = 1 is shown in Fig. 3. Here all exponents aretreated as adjustable parameters for obtaining optimaldata collapse. The vertical and horizontal scalings givethe ratio �/⌫0 and the KZ exponent z0+1/⌫0, respectively,and the slope in the linear regime is the exponent (9).Combining results for r = 1 and r = 2/3 we obtain theexponents � = 0.54(1), ⌫0 = 1.26(1), and z0 = 0.52(2).
Interestingly, the exponents are far from those ob-tained using Landau theory [41] and other methods [42]for large-d and fully connected (d = 1 [43]) Ising modelsin a transverse field; � = 1, ⌫0 = 2 and z0 = 1/4 (d
u
= 8)[41]. One might have expected the same mean-field ex-ponents for these systems, as in the classical case. AQMC calculation for the fully-connected model [44] wasnot in complete agreement with the Landau values. Itwas argued that z = 4, which, with d
u
= 8, agrees withour z0 ⇡ 1/2 for the 3-regular graphs. However, � wasclose to 1 and ⌫ = 1/4 (⌫0 = 2) was argued. It wouldbe interesting to study n-regular graphs and follow theexponents from n = 3 to large n.
Implications for quantum computing.—The critical ex-
10-1
100
101
102
103
104
v N z’+1/ν’
0.3
0.5
0.8
1.0
<q
2>
N 2
β/ν
’
128
256
384
512
768
1024
1536
N
FIG. 3: (Color online) Optimized scaling collapse of the or-der parameter in critical quenches of 3-regular graphs, givingthe exponents listed in the text. The line has slope given inEq. (9) and the points above it deviate due to high-velocitycross-overs [31] not captured by Eq. (8). For N ! 1 thelinear behavior should extend to infinity.
ponents contain information relevant to QA quantumcomputing. In the classical case the KZ exponent isz0 + 1/⌫0 = 1, while in the quantum system z0 + 1/⌫0 ⇡1.31. Thus, by Eq. (7) the adiabatic annealing time growsfaster with N in QA. Furthermore, since the critical or-der parameter scales asN�2�/⌫0
, the critical cluster is lessdense with QA, i.e., it is further from the state soughtwhen s ! 1 (the solution of the optimization problem).Thus, in both these respects QA performs worse than SAin passing through the critical point (while QA on thefully-connected model, with the exponents of Ref. [41],would reach s
c
faster than SA, though the critical clus-ter is still less dense). QA can be made faster than SAby following a protocol (5) with su�ciently large r, butthis may not be practical when s
c
is not known and thegoal is anyway to proceed beyond this point. While ourresults do not contain any quantitative information onthe process continuing from s
c
to s = 1, it is certainlydiscouraging that the initial stage of QA is ine�cient.
It would be interesting to run the D-Wave machine[11, 12] as well on a problem with a critical point andstudy velocity scaling. This would give valuable insightsinto the nature of the annealing process.
Acknowledgments.—We thank David Huse and A Pe-ter Young for stimulating discussions. This work wassupported by the NSF under grant No. PHY-1211284.
[1] S. Kirkpatrick, C. D. Gelatt Jr, M. P. Vecchi, Science220 671 (1983).
[2] V. Cerny, J. Optim. Theor. and Appl. 45: 41 (1985).[3] V. Granville, M. Krivanek, J.-P. Rasson, J.-P., IEEE
Trans. on Pattern Analysis and Machine Intelligence 16
↵ = 17/12
Study evolution to sc
- several system sizes N- several velocities
4
0.000 0.001 0.002 0.003 0.004
1/N
0.335
0.340
0.345
0.350
0.355
s c(N)
0.33 0.34 0.35 0.36
s
0.0
0.1
0.2
0.3
0.4
U
192
256
320
384
448
512
576
N
FIG. 2: (Color online) Crossing points between Binder cumu-lants for 3-regular graphs with N and N+64 spins, extractedusing the curves shown in the inset. The results were obtainedin quenches with v ⇠ N�↵ for ↵ = 17/12. The curve in themain panel is a power-law fit for extrapolating sc.
were in good agreement with this estimate. The errorbars on these calculations is several percent.
We find sc
using r = 1 QAQMC with v / N�↵, where↵ exceeds the KZ exponent z0 + 1/⌫0 (which is unknownbut later computable for a posteriori verification). Thenhq2i ⇠ N�2�/⌫0
at sc
because f(x) in Eq. (8) approachesa constant when x ! 0. As illustrated in Fig. 2, quench-ing past the estimated s
c
, we use a curve crossing analysisof the Binder cumulant, U = (3� hq4i/hq2i2)/2, and ob-tain s
c
= 0.3565(12). This value agrees well with theprevious results but has smaller uncertainty.
Performing additional quenches to sc
using protocolswith both r = 1 and r = 2/3 in Eq. (5) we extract allthe critical exponents. An example of scaling collapsefor r = 1 is shown in Fig. 3. Here all exponents aretreated as adjustable parameters for obtaining optimaldata collapse. The vertical and horizontal scalings givethe ratio �/⌫0 and the KZ exponent z0+1/⌫0, respectively,and the slope in the linear regime is the exponent (9).Combining results for r = 1 and r = 2/3 we obtain theexponents � = 0.54(1), ⌫0 = 1.26(1), and z0 = 0.52(2).
Interestingly, the exponents are far from those ob-tained using Landau theory [41] and other methods [42]for large-d and fully connected (d = 1 [43]) Ising modelsin a transverse field; � = 1, ⌫0 = 2 and z0 = 1/4 (d
u
= 8)[41]. One might have expected the same mean-field ex-ponents for these systems, as in the classical case. AQMC calculation for the fully-connected model [44] wasnot in complete agreement with the Landau values. Itwas argued that z = 4, which, with d
u
= 8, agrees withour z0 ⇡ 1/2 for the 3-regular graphs. However, � wasclose to 1 and ⌫ = 1/4 (⌫0 = 2) was argued. It wouldbe interesting to study n-regular graphs and follow theexponents from n = 3 to large n.
Implications for quantum computing.—The critical ex-
10-1
100
101
102
103
104
v N z’+1/ν’
0.3
0.5
0.8
1.0
<q
2>
N 2
β/ν
’
128
256
384
512
768
1024
1536
N
FIG. 3: (Color online) Optimized scaling collapse of the or-der parameter in critical quenches of 3-regular graphs, givingthe exponents listed in the text. The line has slope given inEq. (9) and the points above it deviate due to high-velocitycross-overs [31] not captured by Eq. (8). For N ! 1 thelinear behavior should extend to infinity.
ponents contain information relevant to QA quantumcomputing. In the classical case the KZ exponent isz0 + 1/⌫0 = 1, while in the quantum system z0 + 1/⌫0 ⇡1.31. Thus, by Eq. (7) the adiabatic annealing time growsfaster with N in QA. Furthermore, since the critical or-der parameter scales asN�2�/⌫0
, the critical cluster is lessdense with QA, i.e., it is further from the state soughtwhen s ! 1 (the solution of the optimization problem).Thus, in both these respects QA performs worse than SAin passing through the critical point (while QA on thefully-connected model, with the exponents of Ref. [41],would reach s
c
faster than SA, though the critical clus-ter is still less dense). QA can be made faster than SAby following a protocol (5) with su�ciently large r, butthis may not be practical when s
c
is not known and thegoal is anyway to proceed beyond this point. While ourresults do not contain any quantitative information onthe process continuing from s
c
to s = 1, it is certainlydiscouraging that the initial stage of QA is ine�cient.
It would be interesting to run the D-Wave machine[11, 12] as well on a problem with a critical point andstudy velocity scaling. This would give valuable insightsinto the nature of the annealing process.
Acknowledgments.—We thank David Huse and A Pe-ter Young for stimulating discussions. This work wassupported by the NSF under grant No. PHY-1211284.
[1] S. Kirkpatrick, C. D. Gelatt Jr, M. P. Vecchi, Science220 671 (1983).
[2] V. Cerny, J. Optim. Theor. and Appl. 45: 41 (1985).[3] V. Granville, M. Krivanek, J.-P. Rasson, J.-P., IEEE
Trans. on Pattern Analysis and Machine Intelligence 16
hq2(sc)i / N�2�/⌫0f(vNz0+1/⌫0
)
Velocity Scaling at the Glass Transition
β/ν‘ ≈ 0.43(2)z’+1/ν’ ≈ 1.3(2)
Significance of the exponents?
β/ν‘ ≈ 0.47(3)z’+1/ν’ ≈ 0.8(1)
Fully connected model (SK)- velocity scaling gives
Differ from valuesexpected for d=∞:(Read, Sachdev, Ye, 1995)
β/ν‘ = 1/2z’+1/ν’ = 3/4
Why is 3-regular model different?
Relevance to Quantum Computing
Reaching sc, the degree of ordering scales asp< hq2i > ⇠ N��/⌫0
�/⌫0 ⇡ 0.43
Classicalβ/ν‘ = 1/3z’+1/ν’ = 1
Let’s compare with the know classical exponents(finite-temperature transition of 3-regular random graphs)
Quantumβ/ν‘ ≈ 0.43z’+1/ν’ ≈ 1.3
h
Tglass phase
• It takes longer for quantum annealing to reach its critical point
• And the state is further from ordered (further from the optimal solution)
Proposal: Do velocity scaling with the D-wave machine!
The time needed to stay adiabatic up to sc scales as
t ⇠ Nz0+1/⌫ z0 + 1/⌫0 ⇡ 1.3