Pith. sign in

REVIEW 4 major objections 4 minor 42 references

Moment selection in the NPA hierarchy is a synergy-dominated combinatorial problem: budget-aware optimizers, not greedy addition, find the tightest bounds.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 01:07 UTC pith:HYGA7D5J

load-bearing objection A useful moment-selection method paper anchored by a solid exhaustive I3322 benchmark, with broader claims that are promising but not yet fully secured. the 4 major comments →

arxiv 2607.14755 v1 pith:HYGA7D5J submitted 2026-07-16 quant-ph

Moment Optimization in the Navascu\'es-Pironio-Ac\'in Hierarchy

Francesco Flora , Losel Matos , Tim Heightman , Tam\'as Kriv\'achy , Adan Garriga , Antonio Ac\'in This is my paper
classification quant-ph
keywords NPA hierarchymoment selectionsemidefinite programmingBell inequalitiesquantum many-body certificationparallel temperingrestricted Boltzmann machineBayesian optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The Navascués–Pironio–Acín (NPA) hierarchy bounds quantum optimization problems by solving semidefinite programs over sets of operator moments, but the number of moments explodes with level. The paper reframes the practical question — for a fixed computational budget, which subset of candidate moments gives the tightest bound? — as a combinatorial subset-selection problem. Using the I3322 Bell inequality, where all 2^21 subsets can be enumerated, it shows the landscape is governed by strong higher-order synergy: the best k moments are not an extension of the best k−1, so greedy addition stalls. The paper introduces a marginal synergy diagnostic and three optimizers — Parallel Tempering, an RBM-based reinforcement-learning policy, and Bayesian Optimization — which all beat greedy addition at costs around two orders of magnitude below brute force, with the RBM closest to optimal. It then shows the framework transfers to 174 Bell inequalities and to certifying ground-state observables of the Heisenberg spin chain, where a warm-started search improves certified long-range correlation bounds by nearly two orders of magnitude.

Core claim

On its own terms, the paper claims moment selection in an NPA relaxation is a combinatorial problem whose landscape is dominated by strong higher-order synergy: the best k moments are not an extension of the best k−1. Exhaustive enumeration of the 21-moment I3322 adding set shows subsets of equal size spanning nearly the full NPA1–NPA2 gap, and the marginal synergy Δ(S_k) peaks exactly where the bound saturates. Three optimizers — Parallel Tempering, an RBM-based REINFORCE policy, and Bayesian Optimization — all beat greedy addition at roughly two orders of magnitude below brute force, with the RBM closest to optimal. In the Heisenberg chain, warm-started search over an enlarged NPA4 pool im

What carries the argument

The central machinery is the NPA moment matrix restricted to a chosen subset of operator words; the problem is to choose the k moments in the adding set that minimize the SDP bound. The paper's structural diagnostic is the marginal synergy Δ(S_k), the average degradation when one moment is removed from the best known k-subset; it identifies the synergy-dominated transition regime and gives a cheap saturation estimate. The three optimizers are Parallel Tempering (replica-exchange Monte Carlo), a Restricted Boltzmann Machine policy trained by REINFORCE with top-k Gumbel sampling, and Bayesian Optimization with a random-forest surrogate.

Load-bearing premise

The results stand on the premise that the heuristic optimizers, especially the RBM, find near-optimal moment subsets; the RBM's training step uses a proxy for the probability of its sample, not the exact value, so it carries no formal guarantee of finding good subsets.

What would settle it

A concrete disproof would be a still-enumerable Bell scenario or larger adding set where the exhaustive optimal k-subsets are nested (each contains the previous one), the marginal synergy stays near zero, and greedy addition matches the optimum — or a larger-N demonstration that the RBM's biased gradient performs no better than random subset search.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Greedy moment addition is not a viable strategy in synergy-dominated regimes; optimal subsets are not nested, so individual moment assessments mislead.
  • Moment-selective relaxations can beat rigid level truncation: even a fraction of the NPA2 adding set improves on the full NPA1 bound for many inequalities.
  • Because convergence within the NPA1-to-NPA2 window is heterogeneous across Bell inequalities, the coarse 'NPA2 bound' label hides qualitatively different landscapes; moment budget k is the more meaningful resource.
  • For many-body certification, local physically motivated bases are compressible for local observables but not for long-range ones; broader pools plus warm-starting give certified bounds the local basis cannot reach.
  • The marginal synergy diagnostic can serve as a no-extra-cost convergence monitor: if it stays elevated while the cost curve flattens, further moment selection is still worthwhile.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if synergy-dominated landscapes are generic in noncommutative polynomial optimization, the same selection framework should transfer to steady-state certification, dimension witnesses, and the commutative Lasserre hierarchy — the paper notes the commutative extension but does not test it.
  • Editorial inference: the RBM's REINFORCE update uses a Bernoulli log-likelihood proxy rather than the exact top-k probability (Eq. 30), so its reported advantage could shrink once the biased gradient is corrected; a test would be to compare against an unbiased top-k gradient estimator on the same benchmark.
  • Editorial inference: the synergy diagnostic could be turned from a monitor into an active hierarchy-growing rule — adding moments in directions of high marginal synergy instead of full NPA levels — which the paper lists as future work and would make NPA relaxations themselves budget-aware.
  • Editorial inference: the nearly-two-orders-of-magnitude correlator improvement is demonstrated at small system sizes (N=9 or 10); whether the warm-started gain persists at larger N is an open testable question.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper studies moment selection in the NPA hierarchy as a combinatorial subset-selection problem. Given an initial moment set I and a candidate pool F, the task is to choose the k additional moments that yield the tightest semidefinite relaxation bound. On the I3322 Bell inequality, exhaustive enumeration of all 2^21 subsets provides ground truth, revealing a highly heterogeneous landscape with a sharp transition regime governed by higher-order synergistic interactions. The authors define a marginal-synergy diagnostic and develop three optimizers: parallel tempering, an RBM-based REINFORCE policy, and Bayesian optimization. All three outperform a greedy baseline on I3322, with the RBM closest to the optimal path. The methods are then applied to 174 Bell inequalities in the (4,4,2,2) scenario, where heterogeneous convergence budgets k* are reported, and to the 1D Heisenberg chain, where a warm-started NPA4 search is claimed to improve certified bounds on long-range correlations by nearly two orders of magnitude. The central claims are that moment selection is governed by higher-order synergy, that budget-aware selection substantially outperforms greedy methods, and that physically motivated local bases are internally compressible but not globally optimal.

Significance. If the results hold, the paper makes a useful and timely contribution: it reframes the NPA hierarchy as a budget-limited subset-selection problem, provides a concrete diagnostic for landscape structure, and benchmarks three practical optimizers. The exhaustive I3322 enumeration is a solid, reproducible benchmark that supports the landscape and synergy claims. The extension to the (4,4,2,2) scenario and to many-body certification demonstrates the framework's scope. However, the central generalization beyond I3322 rests on the near-optimality of heuristic optimizers, and this is not verified. The RBM's policy-gradient update is acknowledged to be biased, and the (4,4,2,2) convergence budgets are produced solely by that method. The Heisenberg headline improvement is presented without run-to-run statistics. These gaps do not invalidate the I3322 results, but they limit the strength of the transferability and optimality claims until addressed.

major comments (4)
  1. [§4.2, Eq. (30)] Equation (30) defines log πθ as the conditional Bernoulli log-likelihood of the visible units, not the log-probability of the top-k Gumbel sample. The REINFORCE estimator (25) is therefore biased for J(θ) in Eq. (24), and the paper provides no convergence guarantee for the resulting policy. This matters because the RBM is the best method on I3322 and is the only optimizer used in §5.2 to produce the (4,4,2,2) convergence budgets k*. The claim that the RBM 'achieves the closest approach to optimal' and the heterogeneity analysis both rest on an estimator whose bias is unquantified. Please either provide a corrected estimator with exact top-k log-probability, or validate the biased proxy on a second exhaustive benchmark (e.g., a smaller scenario) showing that the bias does not affect landscape conclusions.
  2. [§5.2, Fig. 6] All 171 convergence curves are produced by a single RBM run per inequality, with no seed statistics, on a landscape with no ground truth. On I3322 the RBM tracks the exhaustive optimum, but this does not transfer automatically to N=40. The observed heterogeneity in k* (between k≈9 and k≈33) could reflect run-to-run variation or the biased gradient in Eq. (30) rather than landscape structure. Please report variance over independent seeds for at least a subsample of inequalities, and/or compare with PT/BO on a subset, so the reader can separate optimizer artifacts from properties of the moment-selection landscape.
  3. [§5.3.2, Fig. 9] The claimed two-orders-of-magnitude improvement (Δ_C from 7.13×10^-5 to ≈10^-6) is presented without error bars or seed statistics for the warm-started NPA4 PT run. Figure 8 reports error bars for the energy case, but the analogous information is missing for the observable, which is the headline result. Given the large variance of the independent NPA4 search (orange curve) and the warm-start dependence, one favorable run is insufficient to establish that the improvement reflects the moment-selection landscape. Please provide multiple independent runs (or a seed table) and state whether the crossing at k≈1000 and the final Δ_C ≈ 10^-6 are stable across runs.
  4. [§3.3, Fig. 2(b)] The statement that the peak of the random-subset synergy Δ(k) provides a 'cheap upper bound' on the saturation budget is an empirical observation on I3322 and is not established for other landscapes. Since the diagnostic is presented as a general budget guide, please test this claim on at least one additional scenario (e.g., a subset of the (4,4,2,2) inequalities with known NPA2 reference) or explicitly restrict its validity to the I3322 benchmark.
minor comments (4)
  1. [§4.1 / Algorithm 1] The temperature ordering is inconsistent: the text states T1 < T2 < ··· < TR, while Algorithm 1's requirement line writes T1 > ··· > TR. Please harmonize.
  2. [§4.2] The hidden-node count is given as int(|SN|/2), but |SN| = 2^N, making this 2^{N-1}, not N/2 as stated. The intended expression appears to be int(N/2), given the following sentence about O(N) total nodes.
  3. [Figure 2 caption] The caption uses 'Left' and 'Right' to refer to the two panels, whereas other figures use (a)/(b) labels. Please make the panel references consistent.
  4. [§3.1] The definition of I3322 follows the correlator representation of [8], but the signs in Eq. (18) could be checked once more against the reference to avoid reader confusion.

Circularity Check

0 steps flagged

No significant circularity: central benchmarks are external and the main claims are not forced by construction.

full rationale

The paper's derivation chain is self-contained against external benchmarks rather than self-referential. The I3322 analysis uses exhaustive enumeration over all 2^21 subsets to define ground-truth optimal paths; the reported gains of PT, RBM, and BO are measured against this external ground truth, not against values derived from the methods' own fitted parameters. The synergy diagnostic Delta(S_k) is defined directly from the SDP objective f_gamma and from independently enumerated or sampled subsets; it is not fitted to the paper's later claims, and it is used descriptively. In the (4,4,2,2) scenario, the convergence budgets k* are measured relative to the full NPA2 bound, which is an independently computed reference, and the RBM's biased policy-gradient proxy (Eq. 30) is a heuristic limitation rather than a circular step: the paper does not redefine the target as the RBM's output. In the Heisenberg application, the local basis of Ref. [40] is used as the object of study, and the claimed improvement over it is obtained by comparing against the Bethe-ansatz exact energies and the full local-basis relaxation; the warm-started NPA4 result is an empirical observation, not a quantity equivalent to the input local basis. The author overlap with Ref. [40] is worth noting, but the paper explicitly shows that this basis is suboptimal, so the citation is not load-bearing in the direction of the central claims. Concerns about heuristic near-optimality and missing seed statistics are correctness or robustness risks, not circularity, because they concern unverified assumptions about optimizer performance rather than a definitional or fitted equivalence between inputs and outputs.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The free parameters are optimizer hyperparameters, evaluation budgets, compression fractions, thresholds, and one unspecified energy-window constraint. The central physics claims do not rest on fitted physical constants, but the method-comparison claims do depend on the chosen budgets and hyperparameters.

free parameters (8)
  • PT evaluation budget = R=5, E=20, L=50 -> 5000 SDP evaluations
    Chosen by hand; the comparison of methods and the 'orders of magnitude below brute force' claim depend on it.
  • RBM evaluation budget = 500 gradient steps, M=10 -> 501 SDP evaluations
    Chosen to fit a runtime/training budget; the 'RBM closest to optimal' claim depends on this budget.
  • BO evaluation budget = ninit=50, niter=200, nc=100 -> 250 SDP evaluations
    Chosen by hand; BO's reported trade-off depends on these values.
  • RBM learning rate and cooling = η=0.01, decay=0.98, Tstart=1.0, α=0.99
    Hyperparameters set by hand in Sec. 5.1; they affect RBM performance.
  • BO exploration coefficient = β=0.7
    Hand-chosen trade-off parameter in Eq. (31).
  • Heisenberg compression fractions = p ∈ {0.3, 0.5, 0.7}
    Chosen to probe fractional compression; the compressibility claim is defined relative to these fractions.
  • Convergence thresholds = ε = 5% and 1%
    Chosen thresholds in Eq. (32) defining k*; the reported heterogeneity depends on them.
  • Energy-window constraint (Fig. 9)
    The observable SDP uses an energy-window constraint, but its exact bandwidth is not reported; results may depend on it.
axioms (6)
  • standard math The NPA moment-matrix constraints (positivity, normalization, linearized algebraic relations) are valid relaxations whose bounds converge to the exact quantum optimum.
    Invoked throughout Sec. 2.2 and used as the foundation of all experiments.
  • domain assumption The SDP solver returns globally optimal values for each fixed relaxation to numerical precision.
    All fγ evaluations in Sec. 5 assume the SDP solve is reliable; solver and tolerances are not specified.
  • domain assumption The exact ground-state energy of the 1D Heisenberg chain is given by the Bethe ansatz.
    Used as the external benchmark in Sec. 5.3 for energy certificates.
  • ad hoc to paper The RBM conditional Bernoulli log-likelihood (Eq. 30) is a useful REINFORCE score-function direction even though it is not the exact probability of a top-k sample.
    Acknowledged in Sec. 4.2 as a proxy; the method's theoretical grounding rests on this assumption.
  • ad hoc to paper The peak of the random-subset synergy curve provides a cheap upper bound on the saturation budget k.
    Empirical claim in Sec. 3.3; no proof is given that the random-subset peak bounds the true saturation point.
  • domain assumption The candidate pools F=NPA2 and F=NPA4 define the relevant universe of moments for improving relaxations.
    The Bell and Heisenberg studies fix these pools; conclusions about optimality are relative to these pools.

pith-pipeline@v1.3.0-alltime-deepseek · 22799 in / 14772 out tokens · 123619 ms · 2026-08-02T01:07:39.452907+00:00 · methodology

0 comments
read the original abstract

The Navascu\'es-Pironio-Ac\'in (NPA) hierarchy provides a convergent sequence of semidefinite programming (SDP) relaxations for noncommutative polynomial optimisation, ubiquitous in quantum physics. However, its practical applicability is limited by the combinatorial growth in operator moments required at each level. Since not all moments contribute equally to bound tightness, selecting moments within a fixed computational budget is a relevant problem. We reframe moment selection as combinatorial subset selection and show it is governed by strong higher-order synergistic interactions among moments, quantified through a marginal synergy diagnostic adapted from complex systems theory. We develop and compare three optimisation methods: Parallel Tempering (PT), an RBM-based reinforcement learning policy, and Bayesian Optimisation (BO). On the $I_{3322}$ Bell inequality benchmark, all three substantially outperform greedy approaches at costs around two orders of magnitude below brute force, with the RBM achieving the closest approach to optimal throughout the hard transition regime. We apply the framework to the 174 Bell inequalities in the $(4,4,2,2)$ scenario, finding heterogeneous convergence behaviour across inequalities, and to the one-dimensional Heisenberg spin chain, demonstrating that physically motivated monomial bases are internally compressible and are not globally optimal in general. A budget-aware search over a broader pool improves certified bounds on long-range correlations by nearly two orders of magnitude. These results establish a scalable framework for moment selection in noncommutative polynomial optimisation, with broad applications across quantum physics and quantum information.

Figures

Figures reproduced from arXiv: 2607.14755 by Adan Garriga, Antonio Ac\'in, Francesco Flora, Losel Matos, Tam\'as Kriv\'achy, Tim Heightman.

Figure 1
Figure 1. Figure 1: Complete distribution of I3322 relaxation values across all 2 21 moment subsets, stratified by Hamming weight k. The structural transition in the range 5 < k ≤ 19 — where distributions are widest and the optimal value improves most rapidly — identifies the combinatorially active regime of the landscape. The wide spread of bound values at each k, with subsets of the same size spanning nearly the full range … view at source ↗
Figure 2
Figure 2. Figure 2: Synergy as a diagnostic for subset saturation. Left: Best violation f(Sk) (blue, left axis) and marginal synergy ∆(Sk) (red, right axis) as a function of k, with Sk taken as the globally optimal k-subset of the I3322 NPA1→NPA2 adding set. The cost curve flattens abruptly at k = 8 as the NPA2 bound is reached (green dotted line), and the synergy attains its maximum ∆(S ⋆ 8 ) ≈ 0.084 at exactly the same step… view at source ↗
Figure 3
Figure 3. Figure 3: shows the performance of each method with uncertainty bands. PT tracks the optimal curve closely from [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Solution quality and computational cost of the three optimization methods on the I3322 moment￾selection problem. (a) Mean I3322 violation achieved by PT (orange), RBM (green), and BO (purple) as a function of k, averaged over 30 independent runs. The solid black curve is the ground-truth optimal path from exhaustive enumeration; the dashed black curve is the greedy baseline. PT tracks the optimum closely f… view at source ↗
Figure 5
Figure 5. Figure 5: Distance to the ground-truth optimal path. For each method and budget k ∈ {1, . . . , 20}, we plot f(x k r ) − f ∗ (k) across Nrep = 30 independent runs, where f ∗ (k) ≡ minxk∈Sk N f(x k ) is the exhaustive-enumeration optimum at budget k. Solid lines show the mean across runs; shaded bands show ±1 standard deviation. (a) Gap in log10 scale. Mean and standard deviation are computed on the log-transformed g… view at source ↗
Figure 6
Figure 6. Figure 6: Convergence analysis of 171 Bell inequalities in the (4, 4, 2, 2) scenario within the NPA1–NPA2 window (N = 40 moments). Inequalities with |f NPA1 γ − f NPA2 γ | < 10−4 (already converged at NPA1) are excluded from the analysis. (a) Normalised relaxation value as a function of Hamming weight k for each of the 171 inequalities, obtained by running the RBM optimizer across the full range k ∈ {0, . . . , 40}.… view at source ↗
Figure 7
Figure 7. Figure 7: Fractional compression of the local-basis relaxation for the 1D Heisenberg chain. For each system size N and fraction p ∈ {0.3, 0.5, 0.7}, PT selects a subset of size k = ⌊p |A(N)|⌋ from the adding set A = Blocal \ NPA1 via a warm-start sweep. The dashed line shows the gap of the full NPA2 truncation for comparison. (a) Energy discrepancy |Efull(N) − Epartial(N, p)| relative to the full local-basis relaxat… view at source ↗
Figure 8
Figure 8. Figure 8: Energy optimality test beyond the local basis (N = 9, 1D Heisenberg chain). At matched monomial budget k, three PT strategies are compared: optimisation restricted to the local pool Alocal = Blocal \ NPA1 (blue); optimisation over the enlarged pool ANPA4 = NPA4 \ NPA1 with standard warm-start (orange); and the same enlarged-pool search additionally seeded at each k from the best local-pool solution (green,… view at source ↗
Figure 9
Figure 9. Figure 9: Observable optimality test beyond the local basis (N = 10, half-chain correlator CN/2). The same three PT strategies as in [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 3 linked inside Pith

  1. [1]

    Device- independent security of quantum cryptography against collective attacks.Phys

    Antonio Acín, Nicolas Brunner, Nicolas Gisin, Serge Massar, Stefano Pironio, and Valerio Scarani. Device- independent security of quantum cryptography against collective attacks.Phys. Rev. Lett., 98:230501, Jun 2007

    2007

  2. [2]

    On the computational complexity of Ising spin glass models.Journal of Physics A: Mathematical and General, 15(10):3241–3253, 1982

    Francisco Barahona. On the computational complexity of Ising spin glass models.Journal of Physics A: Mathematical and General, 15(10):3241–3253, 1982

    1982

  3. [3]

    Global fluctuations and Gumbel statistics.Physical Review Letters, 95:170601, 2005

    Eric Bertin. Global fluctuations and Gumbel statistics.Physical Review Letters, 95:170601, 2005

    2005

  4. [4]

    H. Bethe. Zur Theorie der Metalle. I. Eigenwerte und Eigenfunktionen der linearen Atomkette.Zeitschrift für Physik, 71:205–226, 1931

    1931

  5. [5]

    Robinson.Operator Algebras and Quantum Statistical Mechanics 1

    Ola Bratteli and Derek W. Robinson.Operator Algebras and Quantum Statistical Mechanics 1. Springer, Berlin, Heidelberg, 2nd edition, 1987

    1987

  6. [6]

    Random forests.Machine Learning, 45(1):5–32, 2001

    Leo Breiman. Random forests.Machine Learning, 45(1):5–32, 2001

    2001

  7. [7]

    Bell nonlocality

    Nicolas Brunner, Daniel Cavalcanti, Stefano Pironio, Valerio Scarani, and Stephanie Wehner. Bell nonlocality. Reviews of Modern Physics, 86(2):419–478, 2014

    2014

  8. [8]

    A relevant two-qubit Bell inequality inequivalent to the CHSH inequality

    Daniel Collins and Nicolas Gisin. A relevant two-qubit Bell inequality inequivalent to the CHSH inequality. Journal of Physics A: Mathematical and General, 37(5):1775–1787, 2004

    2004

  9. [9]

    Zambrini Cruzeiro and N

    E. Zambrini Cruzeiro and N. Gisin. Complete list of tight bell inequalities for two parties with four binary settings. Phys. Rev. A, 99:022104, Feb 2019

    2019

  10. [10]

    Earl and Michael W

    David J. Earl and Michael W. Deem. Parallel tempering: theory, applications, and new perspectives.Physical Chemistry Chemical Physics, 7(23):3910–3916, 2005. 21 Moment Optimization in the Navascués-Pironio-Acín hierarchy

    2005

  11. [11]

    MIT Press, Cambridge, MA, 2016

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, Cambridge, MA, 2016

    2016

  12. [12]

    Deep learning in classical and quantum physics.arXiv preprint arXiv:2508.10666, 2025

    Timothy Heightman and Marcin Płodzién. Deep learning in classical and quantum physics.arXiv preprint arXiv:2508.10666, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Geoffrey E. Hinton. Training products of experts by minimizing contrastive divergence.Neural Computation, 14(8):1771–1800, 2002

    2002

  14. [14]

    Geoffrey E. Hinton. A practical guide to training restricted Boltzmann machines. In Grégoire Montavon, Genevieve B. Orr, and Klaus-Robert Müller, editors,Neural Networks: Tricks of the Trade, pages 599–619. Springer, 2012

    2012

  15. [15]

    Hinton, Simon Osindero, and Yee-Whye Teh

    Geoffrey E. Hinton, Simon Osindero, and Yee-Whye Teh. A fast learning algorithm for deep belief nets.Neural Computation, 18(7):1527–1554, 2006

    2006

  16. [16]

    Iris A. M. Huijben, Wouter Kool, Max B. Paulus, and Ruud J. G. Van Sloun. A review of the Gumbel-max trick and its extensions for discrete stochasticity in machine learning.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1353–1371, 2022

    2022

  17. [17]

    Exchange Monte Carlo method and application to spin glass simulations

    Koji Hukushima and Koji Nemoto. Exchange Monte Carlo method and application to spin glass simulations. Journal of the Physical Society of Japan, 65(6):1604–1608, 1996

    1996

  18. [18]

    L. Hulthén. Über das Austauschproblem eines Kristalles.Arkiv för Matematik, Astronomi och Fysik, 26A(11):1– 106, 1938

    1938

  19. [19]

    Hoos, and Kevin Leyton-Brown

    Frank Hutter, Holger H. Hoos, and Kevin Leyton-Brown. Sequential model-based optimization for general algorithm configuration. InLearning and Intelligent Optimization, volume 6683 ofLecture Notes in Computer Science, pages 507–523. Springer, 2011

    2011

  20. [20]

    Daniel Gelatt, and Mario P

    Scott Kirkpatrick, C. Daniel Gelatt, and Mario P. Vecchi. Optimization by simulated annealing.Science, 220(4598):671–680, 1983

    1983

  21. [21]

    Lasserre

    Jean B. Lasserre. Global optimization with polynomials and the problem of moments.SIAM Journal on Optimization, 11(3):796–817, 2001

    2001

  22. [22]

    Levin and Yuval Peres.Markov Chains and Mixing Times

    David A. Levin and Yuval Peres.Markov Chains and Mixing Times. American Mathematical Society, Providence, RI, 2nd edition, 2017

    2017

  23. [23]

    Ising formulations of many NP problems.Frontiers in Physics, 2:5, 2014

    Andrew Lucas. Ising formulations of many NP problems.Frontiers in Physics, 2:5, 2014

    2014

  24. [24]

    Maddison, Andriy Mnih, and Yee Whye Teh

    Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. The concrete distribution: A continuous relaxation of discrete random variables. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  25. [25]

    Rosenbluth, Marshall N

    Nicholas Metropolis, Arianna W. Rosenbluth, Marshall N. Rosenbluth, Augusta H. Teller, and Edward Teller. Equation of state calculations by fast computing machines.The Journal of Chemical Physics, 21(6):1087–1092, 1953

    1953

  26. [26]

    World Scientific, Singapore, 1987

    Marc Mézard, Giorgio Parisi, and Miguel Angel Virasoro.Spin Glass Theory and Beyond. World Scientific, Singapore, 1987

    1987

  27. [27]

    Certifying steady-state properties of open quantum systems.Phys

    Luke Mortimer, Donato Farina, Grazia Di Bello, David Jansen, Andreas Leitherer, Pere Mujal, and Antonio Acín. Certifying steady-state properties of open quantum systems.Phys. Rev. Res., 7:033237, Sep 2025

    2025

  28. [28]

    Bounding the set of quantum correlations.Physical Review Letters, 98:010401, 2007

    Miguel Navascués, Stefano Pironio, and Antonio Acín. Bounding the set of quantum correlations.Physical Review Letters, 98:010401, 2007

    2007

  29. [29]

    A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations.New Journal of Physics, 10:073013, 2008

    Miguel Navascués, Stefano Pironio, and Antonio Acín. A convergent hierarchy of semidefinite programs characterizing the set of quantum correlations.New Journal of Physics, 10:073013, 2008

    2008

  30. [30]

    Pál and Tamás Vértesi

    Károly F. Pál and Tamás Vértesi. Maximal violation of a bipartite three-setting, two-outcome Bell inequality using infinite-dimensional quantum systems.Physical Review A, 82(2):022116, 2010

    2010

  31. [31]

    Convergent relaxations of polynomial optimization problems with noncommuting variables.SIAM Journal on Optimization, 20(5):2157–2180, 2010

    Stefano Pironio, Miguel Navascués, and Antonio Acín. Convergent relaxations of polynomial optimization problems with noncommuting variables.SIAM Journal on Optimization, 20(5):2157–2180, 2010

    2010

  32. [32]

    Certificates of quantum many-body properties assisted by machine learning.Physical Review Research, 5(1):013097, 2023

    Borja Requena, Gorka Muñoz Gil, Maciej Lewenstein, Vedran Dunjko, and Jordi Tura. Certificates of quantum many-body properties assisted by machine learning.Physical Review Research, 5(1):013097, 2023

    2023

  33. [33]

    Bootstrapping the stationary state of bosonic open quantum systems.arXiv preprint arXiv:2410.07384, 2024

    Gustave Robichon and Tilloy Antoine. Bootstrapping the stationary state of bosonic open quantum systems.arXiv preprint arXiv:2410.07384, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    Adams, and Nando De Freitas

    Bobak Shahriari, Kevin Swersky, Ziyu Wang, Ryan P. Adams, and Nando De Freitas. Taking the human out of the loop: a review of Bayesian optimization.Proceedings of the IEEE, 104(1):148–175, 2016. 22 Moment Optimization in the Navascués-Pironio-Acín hierarchy

    2016

  35. [35]

    Sutton and Andrew G

    Richard S. Sutton and Andrew G. Barto.Reinforcement Learning: An Introduction. MIT Press, Cambridge, MA, 2nd edition, 2018

    2018

  36. [36]

    Titsias and Christopher Yau

    Michalis K. Titsias and Christopher Yau. The Hamming ball sampler.Journal of the American Statistical Association, 112(520):1598–1611, 2017. arXiv preprint arXiv:1504.08133 (2015)

    Pith/arXiv arXiv 2017

  37. [37]

    Tsirelson

    Boris S. Tsirelson. Quantum analogues of the Bell inequalities. The case of two spatially separated domains. Journal of Soviet Mathematics, 36(4):557–570, 1987

    1987

  38. [38]

    Maddison, John Lawson, and Jascha Sohl-Dickstein

    George Tucker, Andriy Mnih, Chris J. Maddison, John Lawson, and Jascha Sohl-Dickstein. REBAR: low-variance, unbiased gradient estimates for discrete latent variable models.Advances in Neural Information Processing Systems, 30, 2017

    2017

  39. [39]

    Thomas F. Varley. A scalable synergy-first backbone decomposition of higher-order structures in complex systems. npj Complexity, 1(9), 2024

    2024

  40. [40]

    Certifying ground-state properties of many-body systems.Physical Review X, 14(3):031006, 2024

    Jie Wang, Jacopo Surace, Irénée Frérot, Benoît Legat, Marc-Olivier Renou, Victor Magron, and Antonio Acín. Certifying ground-state properties of many-body systems.Physical Review X, 14(3):031006, 2024

    2024

  41. [41]

    Simple statistical gradient-following algorithms for connectionist reinforcement learning

    Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3):229–256, 1992

    1992

  42. [42]

    Peter Wittek. Algorithm 950: Ncpol2sdpa—sparse semidefinite programming relaxations for polynomial opti- mization problems of noncommuting variables.ACM Transactions on Mathematical Software, 41(3):21, 2015. 23

    2015