Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection...
Transcript of Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection...
![Page 1: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/1.jpg)
Classical boson sampling algorithms and the outlook for experimental boson sampling
A. Neville, C. Sparrow, R. Clifford, E. Johnston, P. Birchall,
A. Montanaro, A. Laing
&
P. Clifford1, R. Clifford2
University of Bristol
1University of Oxford2University of Bristol
![Page 2: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/2.jpg)
2
Nature Physics 13, 1153–1157 (2017)arXiv:1705.00686
Proc. SODA (2018)arXiv:1706.01260
![Page 3: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/3.jpg)
• Boson sampling• Mathematical problem
• Linear optical experiment
• Previous expectations
• Our results• (Approximate) Markov chain
Monte Carlo sampler
• Exact sampler
• Experimental time “complexity”• Photon loss
• Revised expectations
• Algorithm outlines
• Bonus – estimated mixing time
• Rejection sampling?
3
Outline
![Page 4: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/4.jpg)
4
Boson samplingInput:
Collision Free Subspace (CFS)
Output: Sample with prob.Sample with prob.
matrix made from taking rows of according torows of according to
How many copies of each row
Aaronson & Arkhipov, arXiv:1011.3245 (2010)
AA: Efficient classical algorithm for (even approximate) boson sampling would have surprising complexity theoretic consequences
![Page 5: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/5.jpg)
5
Boson sampling – example (n=3)
![Page 6: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/6.jpg)
6
Boson sampling – Matrix permanents
Time complexity: (Ryser or Glynn)
Time complexity:
-hard
Time complexity: (Fast matrix multiplication)
Time complexity:
![Page 7: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/7.jpg)
7
Boson sampling – Linear optical experiment
multiphoton transformation
Aaronson & Arkhipov, arXiv:1011.3245 (2010)
The experiment “naturally” performs boson sampling
No feed-forward required
![Page 8: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/8.jpg)
8
Previous expectations – quantum computational supremacy
Aaronson & Arkhipov, arXiv:1309.7460 (2013)
Aaronson & Arkhipov, arXiv:1011.3245 (2010)
Preskill, arXiv:1203.5813 (2012)
Barkhofen et al., Phys. Rev. Lett 118 (2017)
Goldstein et al., Phys. Rev. B 95 (2017)
Latmiral et al., New J. Phys 18 (2016)
![Page 9: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/9.jpg)
9
Brute force sampling
matrix permanents in the CFS
>300 years
![Page 10: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/10.jpg)
10
Our results – classical sampling algorithms
• Metropolised independence sampling (MIS) Neville et al. Nature Physics (2017)
• Markov chain Monte Carlo• Collision free subspace• Experimentally verified approximate sampling• 100 matrix permanents per sample for (and probably beyond)100 matrix permanents per sample for (and probably beyond)100 matrix permanents per sample for (and probably beyond)
• Exact boson samplingClifford & Clifford, Proc. SODA (2018)
• Full space with collisions• Rigorously proven exact sampling• Equivalent of 2 matrix permanents per sampleEquivalent of 2 matrix permanents per sampleEquivalent of 2 matrix permanents per sample
![Page 11: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/11.jpg)
11
Our results – sampler timing (on a laptop)
![Page 12: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/12.jpg)
12
Experimental time “complexity” – photon loss
![Page 13: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/13.jpg)
13
Source rep. rate:
Revised expectations
1 day of supercomputer time:58 photon boson sampling
A,B: Wang et al. Nature Photonics (2017)
![Page 14: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/14.jpg)
14
MCMC - Metropolised independence sampling (MIS)
• Markov chain:
• Proposal distribution: Distinguishable particle distribution
Neville et al. Nature Physics (2017)
Elementwise
![Page 15: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/15.jpg)
15
Metropolised independence sampling – Thinning
n=25
Neville et al. Nature Physics (2017)
![Page 16: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/16.jpg)
16
MIS – “Verification” overview
Neville et al. Nature Physics (2017)
• Likelihood ratio test• Alternative hypothesis: distinguishable particle sampling• Used by experimentalists – photon distinguishability is a source of error• Relevant to MIS – proposals from distinguishable particle distribution•
• 2-sample Bootstrap Kolmogorov-Smirnov (KS) test• Null hypothesis: 2 samples are drawn from same distribution• Comparison samplers: Brute force, rejection, alternative MIS (longer
thinning interval), distinguishable particles•
![Page 17: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/17.jpg)
17
MIS – “Verification” part 1
Neville et al. Nature Physics (2017)
n=20
![Page 18: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/18.jpg)
18
MIS – “Verification” part 2
n=7
Neville et al. Nature Physics (2017)
![Page 19: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/19.jpg)
19
Exact classical algorithms for boson sampling
Rough idea:
1. Expand the sample space
2. Compute marginal probabilities for row and column choices
3. Incrementally sample rows and columns using probabilistic chain rule
4. Exploit Laplace expansion of the permanent for further speedup
Clifford & Clifford, Proc SODA (2018)
![Page 20: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/20.jpg)
20
Clifford & Clifford Algorithm A
Recall: How many of each row to include. Recall:
Rewrite: Which rows to include. Rewrite: Which rows to include.
Expand the space by relaxing the ordering requirement:
Sample from
Clifford & Clifford, Proc SODA (2018)
(The permanent is invariant under permutations of rows)
![Page 21: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/21.jpg)
21
Clifford & Clifford Algorithm A
- Set of k-combinations from
- Submatrix of with rows and columns Submatrix of with rows and columns Submatrix of with rows and columns
Marginals:
Chain rule:
Clifford & Clifford, Proc SODA (2018)
![Page 22: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/22.jpg)
22
Clifford & Clifford Algorithm A
1. Sample from
2. Sample from
…
n. Sample from
Sample from
n. Sample from
Sample from
Running time:
Clifford & Clifford, Proc SODA (2018)
![Page 23: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/23.jpg)
23
Further expand the space:
Introduce - auxiliary column labels (permutation)Introduce
Clifford & Clifford Algorithm B
Define a new marginal pmf:
Clifford & Clifford, Proc SODA (2018)
Sample from
(The permanent is invariant under permutations of columns)
![Page 24: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/24.jpg)
24
Clifford & Clifford Algorithm B
0. Sample from
1. Sample from
2. Sample from
…
n. Sample from
0. Sample from
n. Sample from
Sample from
Sample from
Running time: ?
Clifford & Clifford, Proc SODA (2018)
![Page 25: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/25.jpg)
25
Clifford & Clifford Algorithm B
Trick:
(Glynn formula for the permanent (Gray code ordered))
(Laplace expansion of the permanent)
+
Evaluate in time Evaluate in time Evaluate in time
Clifford & Clifford, Proc SODA (2018)
![Page 26: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/26.jpg)
26
Clifford & Clifford Algorithm B
Running time:
Clifford & Clifford, Proc SODA (2018)
0. Sample from
1. Sample from
2. Sample from
…
n. Sample from
0. Sample from
n. Sample from
Sample from
Sample from
![Page 27: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/27.jpg)
27
Bonus: MIS mixing time
Jun S. Liu, Monte Carlo strategies in scientific computing (2001)
• For suggests that 50 burn in states required to sample from distribution at most ~1% close to targetFor suggests that 50 burn in states required to sample from distribution at most ~1% close
![Page 28: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/28.jpg)
Could approximate rejection sampling work?
28
With thanks to the QIP PC
Aaronson & Arkhipov, arXiv:1309.7460 (2013)
Translation: Rejection sampling with uniform proposal could be good* to approximately sample from a distribution at least close to the target distribution with probability over boson sampling matrices with and small approximately sample from a distribution at least close to the target distribution with probability over boson sampling matrices with and small probability over boson sampling matrices with and small
*with accepts per reject
![Page 29: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/29.jpg)
29
Rejection sampling - How many permanents per sample?
• For need permanent computations per sample for reliable TVD • Seems slower than MIS for same , but a comparison requires subtle considerations
For need permanent computations per sample for reliable TVD For need permanent computations per sample for reliable TVD Seems slower than MIS for same , but a comparison requires subtle considerations
![Page 30: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/30.jpg)
30
Rejection sampling - summary
• We tested to see whether rejection sampling could perform approximate sampling for small and with small TVD to the boson sampling distribution
• It seems to work remarkably well, but still probably slower than MIS
• Clifford and Clifford sampler faster than both and samples exactly for all input unitaries
sampling for small and with small TVD to the boson sampling distribution
![Page 31: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/31.jpg)
31
Future work
• Verification schemes• Using the exact sampler• Without computing additional permanents?
(Agresti et al. arXiv:1712.06863)
• Loss tolerance scheme?
• Improve classical algorithms as much as possible• E.g. Include realistic experimental error
(García-Patrón et al. arXiv:1712.10037, Renema et al. arXiv:1707.02793)
• Reduce linear optical circuit depth?
![Page 32: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/32.jpg)
32
Conclusion
• Classical boson sampling algorithms vastly outperform current state of the art quantum experiments
• Increased lower bound for the number of photons required to demonstrate quantum computational supremacy by boson sampling
• Decreased upper bound on quantum advantage by boson sampling as a function of and
• It’s important to optimise classical algorithms as much as possible –2018 may be the year of the first quantum computational supremacy claim, it would be sad if it had to be retracted
function of and
![Page 33: Classical boson sampling algorithms and the outlook for ...€¦ · Could approximate rejection sampling work? 28 With thanks to the QIP PC Aaronson & Arkhipov, arXiv:1309.7460 (2013)](https://reader033.fdocuments.net/reader033/viewer/2022051911/6000d8292abe4e705c4169a5/html5/thumbnails/33.jpg)
33
Thank you