{"id":"a1ec0416-5ebb-44e7-ac48-22ede153d645","arxiv_id":"1908.10778","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Quantum circuit Born machines beat restricted Boltzmann machines with equal parameter counts on a finance-inspired generative modeling benchmark built from S&P 500 data.","lead":"This paper compares a quantum machine learning model called a quantum circuit Born machine with a classical restricted Boltzmann machine on probability distributions built from real stock market data. The quantum model wins on most test problems when both models get the same number of adjustable parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The QCBM-vs-RBM gap is confounded: QCBMs get CMA-ES on exact statevector probabilities while RBMs get PCD-1 sampled gradients; the Appendix A K_RBM sweep shows training protocol can change the ranking, so the expressivity claim is unproven.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the comparison does not control for optimizer strength or information access. The paper's own Appendix A provides direct evidence that the RBM training protocol affects the ranking at N=4, which shows the confound is not merely hypothetical. For N=6-10, the failure of PCD-100 to close the gap does not eliminate the confound, because PCD-100 remains a local approximate method and no strong classical baseline (e.g., exact likelihood training with a global optimizer) is attempted. The concern is concrete and testable, and it is central to the paper's advertised conclusion about quantum model expressivity. The reader's CONDITIONAL verdict is therefore appropriate; my stress-test does not move it, so I recommend UNCHANGED. The manuscript itself acknowledges limitations in Section V (e.g., not addressing generalization, benchmarks being application-inspired), which further supports a conditional rather than an unconditional reading.","tokens_in":11551,"tokens_out":6457,"duration_ms":71997,"concrete_test":"Fix the optimizer and loss: train the RBM (same N, N/2 hidden units) on the same 30 targets per N with CMA-ES minimizing the exact DKL of Eq. (4), computing P(v)=sum_h exp(-E(v,h))/Z by exhaustive enumeration of the 2^N x 2^(N/2) states (feasible for N<=10). Compare the bootstrapped median and per-instance scatter to Figs. 3-4. If the RBM gap to QCBM disappears or reverses, the claimed QCBM superiority is an optimizer artifact; if QCBM still wins on most instances, the expressivity claim survives this specific test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison is not model family versus model family; it is model family plus optimizer plus information access versus the same tuple for the classical side. QCBMs are trained with CMA-ES, a strong derivative-free evolutionary optimizer, on exact, noiseless Born probabilities obtained from full statevector simulation (Section III A). RBMs are trained with PCD-1, a stochastic local-gradient method that approximates the log-likelihood gradient with one Gibbs step from a persistent chain (Section III B). The equal-parameter-count construction in Fig. 1 does not remove this confound: parameter count is not expressivity, and training protocol is not controlled. The authors' own Appendix A shows sensitivity to the training protocol: increasing K_RBM from 1 to 10/100 changes the N=4 ranking, with RBM beating QCBM on most instances for K_RBM=10 and 100. That K_RBM=100 does not close the gap at N=6-10 is not decisive evidence against the confound, because PCD-100 is still a local approximate sampler and was not compared against a strong classical optimizer. Consequently the claim that QCBMs have superior performance 'under the same number of resources in terms of parameters' is not yet supported as a statement about quantum expressive power; the gap may largely reflect giving QCBMs a stronger optimizer and exact distributional information while giving RBMs local, sampled gradients. The paper's abstract scopes the claim to 'canonical training of the RBMs,' but the introduction and Fig. 1 frame the comparison as 'fair comparison in terms of model expressivities,' so the confound attacks the advertised conclusion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a benchmark for comparing quantum circuit Born machines (QCBMs) and restricted Boltzmann machines (RBMs) on a finance-inspired generative modeling task. Target distributions are constructed from S&P500 time-series data by solving cardinality-constrained portfolio optimization problems and assigning Boltzmann probabilities to efficient frontiers. Both models are designed with the same number of trainable parameters (N(N+3)/2 for N qubits/visible units). The QCBM is trained with CMA-ES on exact, noiseless Born probabilities obtained from full statevector simulation, while the RBM is trained with persistent contrastive divergence with K_RBM=1 Gibbs step (PCD-1). The main reported result is that QCBMs achieve lower KL divergence to the target distributions than RBMs on most of the 30 benchmark instances per system size, with the gap persisting for N=6, 8, 10. The paper also includes an appendix showing sensitivity of the RBM results to K_RBM=10 and 100.","tokens_in":11900,"tokens_out":4317,"duration_ms":44477,"significance":"If the central comparison were sound, the paper would provide a valuable real-data benchmark showing that QCBMs can outperform canonical RBM training on equal parameter counts. The construction of target distributions from real market data via quadratic programming is a useful contribution, and the authors are commendably explicit about their assumptions, such as the noiseless QVM, the infinite measurement precision, and the choice of a uniform baseline. However, the central claim is not yet supported because the comparison conflates model family with training protocol: QCBMs receive a global optimizer (CMA-ES) and exact distributional information, while RBMs receive a local, stochastic gradient method (PCD-1). The authors' own Appendix A shows that increasing K_RBM reverses the ranking at N=4, demonstrating that the outcome depends on the classical training protocol. The benchmark itself could be valuable, but the expressivity claim requires a matched-optimizer comparison or a carefully justified notion of 'canonical' training.","major_comments":[{"comment":"The comparison is not apples-to-apples: QCBMs are trained with CMA-ES, a global evolutionary optimizer, on exact, noiseless Born probabilities from full statevector simulation, while RBMs are trained with PCD-1, a stochastic local-gradient method that uses one Gibbs step from a persistent chain and sampled gradients. Because optimizer strength and information access differ jointly with the model family, the observed KL divergence gap cannot be attributed to model expressivity. A controlled experiment is needed, for example training the RBM by exact likelihood evaluation with a global optimizer (e.g., CMA-ES on the negative log-likelihood) or training the QCBM with a comparable local stochastic method, before the paper's central claim about quantum models being 'superior' is justified.","section":"Section III.A and III.B, Fig. 3"},{"comment":"The authors' own sensitivity analysis undermines the robustness of the main claim. Increasing K_RBM from 1 to 10/100 reverses the ranking at N=4: the scatter plots in Fig. 4 show most points above the identity line, meaning the RBM outperforms the QCBM on most instances. This directly contradicts the 'close to 100% win' language in Section IV and shows that the comparison's outcome depends materially on the chosen classical training protocol. The paper does not provide a principled justification for why K_RBM=1 is the canonical or fair choice, so the abstract's claim of 'superior performance on typical instances' is not robust.","section":"Appendix A, Fig. 4"},{"comment":"The uniform distribution baseline is used to argue that RBM performance approaches the no-learning limit for larger N. However, since the RBM is trained only with PCD-1 and no hyperparameter search or alternative optimizer is investigated at N=6, 8, and 10, the degradation could reflect insufficient optimization rather than limited expressivity. The paper itself acknowledges this ambiguity in the sentence 'this energy model with quadratic interactions and N/2 hidden units does not have enough expressive power or that there are difficulties in training such models,' but does not resolve it. The central claim that QCBMs are more expressive than RBMs cannot be separated from the confound without additional experiments that control for optimizer quality.","section":"Section IV, Fig. 3(a)"}],"minor_comments":[{"comment":"The list of expected return levels is written as {0.010, 0.015, 0.020, 0.025, 0.30, 0.35}; presumably the last two entries should be 0.030 and 0.035. Please correct this typo.","section":"Section IV"},{"comment":"The clip value epsilon introduced in the KL divergence definition is never specified. Please state the value used in the simulations, since it affects the loss landscape.","section":"Section III.A, Eq. (4)"},{"comment":"The symbol M is used both for the number of asset subsets in Section II (M = (N choose kappa)) and for the number of hidden units in the RBM in Fig. 1 (M = N/2). Please disambiguate the notation.","section":"Section II and Fig. 1"},{"comment":"In the caption, panels (b) and (c) are both labeled 'K_RBM=10'; the second should presumably be 'K_RBM=100'. Please correct.","section":"Fig. 4 caption"},{"comment":"The text 'In Appendix 4' should read 'In Appendix A'.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is likely to attract attention, but the confound between model family and training protocol is load-bearing. I recommend that the authors either provide a matched-optimizer comparison (e.g., train the RBM with CMA-ES on exact likelihood, or train the QCBM with a local stochastic method) or substantially soften the claim to describe a benchmark of specific pipelines (QCBM+CMA-ES versus RBM+PCD-1) rather than a statement about quantum expressive power. The self-citation pattern is not inappropriate, but the abstract currently overstates what Appendix A shows. The benchmark construction itself is a useful contribution that could support a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper builds a new finance-inspired benchmark for generative modeling and reports that QCBMs beat RBMs at matched parameter count. The benchmark is genuinely new and useful. The conclusion, as scoped in the abstract to \"canonical training\" of RBMs, mostly holds. But the paper also frames the comparison as one of model expressivity, and that framing is not supported. The quantum side gets CMA-ES on exact, noiseless Born probabilities; the classical side gets PCD-1 with sampled gradients. Your skeptic is right about the confound.\n\nWhat the paper does well: the construction of target distributions from real S&P 500 data via a probabilistic Markowitz model with cardinality constraints is a legitimate way to generate hard, discrete, application-derived benchmarks. Matching the parameter count between QCBM and RBM is a sensible first step. The paper is also transparent about its idealizations: noiseless simulation, exact probabilities for the QCBM, and a large training set that sidesteps generalization questions.\n\nThe soft spots are real but not fatal to the narrower claim. The central comparison is model family plus optimizer plus information access, not model family alone. Appendix A shows that increasing K_RBM from 1 to 10 or 100 flips the ranking at N=4, so training protocol matters. The authors do not show that RBM performance plateaus with a stronger classical optimizer, and PCD-100 is still a local, approximate sampler. Also, the QCBM receives exact distributional information via statevector simulation, which is a huge advantage over sampled gradients. For N=6, 8, 10 the QCBM still wins even at K_RBM=100, but that does not close the confound. The paper should report results with a stronger classical baseline, for example CD-k with many steps, a global optimizer for the RBM, or at least a discussion of how optimizer strength affects the gap. Missing code and data is another soft spot; the benchmark construction is described well enough to reproduce, but the exact instances and hyperparameters are not released. The market temperature T and return levels are somewhat ad hoc, though the paper acknowledges the benchmark is application-inspired rather than a faithful financial model.\n\nWho this is for: people working on quantum generative modeling or benchmarking quantum vs classical ML. They will find the benchmark construction useful and the comparison instructive, even if the expressivity claim needs reining in. I would send it to peer review. The benchmark is a contribution, and the confound is addressable with additional experiments. A serious referee should ask for stronger classical baselines, code/data release, and an abstract that does not overclaim. With those revisions, the paper would be solid as a benchmark proposal and a cautionary tale about fair comparison.","headline":"A useful new finance-inspired benchmark, but the QCBM-vs-RBM expressivity claim is undercut by an optimizer/information confound; worth reviewing with major revisions.","tokens_in":12437,"tokens_out":2000,"would_cite":true,"duration_ms":23818,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Under equal parameter counts, quantum circuit Born machines beat canonical restricted Boltzmann machines on real financial portfolio benchmarks.","keywords":["quantum circuit Born machines","restricted Boltzmann machines","generative modeling","portfolio optimization","Kullback-Leibler divergence","near-term quantum computing","unsupervised learning","application-inspired benchmarks"],"falsifier":"Train the RBM with the same CMA-ES optimizer on exact likelihood values, or with persistent contrastive divergence using 10 to 100 Gibbs steps, and check whether its KL divergence drops to the QCBM's level for $N=6,8,10$. The paper's own appendix shows that for $N=4$ the longer chains already close much of the gap, so the decisive observation is whether the same happens at larger system sizes.","tokens_in":11355,"feed_emoji":"📈","tokens_out":11796,"duration_ms":112644,"temperature":0.7,"pith_summary":"This paper sets out to compare a quantum generative model, the quantum circuit Born machine (QCBM), against a classical workhorse, the restricted Boltzmann machine (RBM), on benchmarks built from real stock-market data rather than synthetic distributions. The authors construct target distributions from a probabilistic version of mean-variance portfolio optimization with cardinality constraints, then train both models with the same number of parameters. Under canonical RBM training, they report that the QCBM reaches substantially lower distribution-fitting error (KL divergence) on typical instances for system sizes $N=4,6,8,10$, and wins nearly every case-by-case comparison. If correct, this is evidence that quantum generative models can be at least as expressive as classical ones on real-world data with equal resources, using circuit layouts that fit near-term ion-trap hardware. The smallest case, $N=4$, narrows when the classical model is given longer Gibbs sampling chains.","feed_headline":"At equal parameters, quantum models beat classical on market data","feed_subtitle":"Quantum circuit Born machines outperform restricted Boltzmann machines on 30 financial benchmarks at equal model size.","key_machinery":"The load-bearing device is an equal-parameter benchmark built from real financial data. A two-layer hardware-efficient circuit -- $X$ and $Z$ single-qubit rotations followed by a fully connected layer of parametrized two-qubit $XX_{ij}(\\chi)$ entangling gates -- defines the QCBM's Born-rule distribution, and an RBM with $M=N/2$ hidden units is sized so both models share exactly $N(N+3)/2$ parameters. The target distributions come from a Boltzmann law over efficient-frontier risks: for each of $\\binom{N}{\\kappa}$ cardinality-constrained portfolios with $\\kappa=N/2$, the risk $\\sigma^\\kappa_{\\rho,i}$ at return level $\\rho$ is weighted by $p(\\sigma^\\kappa_{\\rho,i})\\propto e^{-\\sigma^\\kappa_{\\rho,i}/T}$, with market temperature $T$. Drawing samples from this distribution yields the benchmark that both models must learn, scored by clipped KL divergence. The pairing of a global optimizer (CMA-ES) acting on exact wavefunction probabilities for the quantum model with a local stochastic-gradient routine (PCD-1) for the classical model is the mechanism that produces the reported gap.","core_discovery":"The paper's central claim is that, at equal parameter counts, QCBMs represent the portfolio-derived target distributions more accurately than RBMs trained in the standard way. Both models are allotted $N(N+3)/2$ trainable parameters: the quantum ansatz uses $2N$ single-qubit rotations and $N(N-1)/2$ fully connected two-qubit $XX$ entangling gates, while the classical model uses $N$ visible units, $N/2$ hidden units, and the associated biases and weights. Training the QCBM with CMA-ES on noiseless statevector Born probabilities and the RBM with persistent contrastive divergence with one Gibbs step, the paper reports bootstrapped median KL divergences consistently lower for the QCBM at $N=6,8,10$, with most scatter points falling below the equality line. The RBM's performance drifts toward the uniform-distribution baseline as $N$ grows, whereas the QCBM stays far from that baseline. The authors read this as an expressivity or trainability advantage of the quantum model by the standard of canonical RBM training, noting that $N=4$ is the case where longer classical Gibbs chains close much of the gap.","pith_inferences":["Editorial inference: the comparison does not isolate model family from training procedure; the QCBM uses a global optimizer on exact probabilities while the RBM uses a short local sampling routine, and the paper's appendix shows this matters at $N=4$, where longer classical Gibbs chains close most of the gap.","Editorial inference: the target distributions are sparse and peaked, concentrated on $\\kappa$-of-$N$ bitstrings; a natural test is to run both models on dense or high-entropy targets to see whether the quantum advantage is specific to sparse, peaked distributions.","Editorial inference: equal parameter count is only one figure of merit; comparing sample complexity, wall-clock training cost, and hardware execution time could change the practical ranking, as the paper itself lists these as open considerations."],"forward_implications":["If the reported gap is genuine, near-term quantum devices with a few tens of qubits could run generative models that match or beat canonical RBMs on real-world benchmark distributions of comparable size.","The benchmark construction itself -- Boltzmann distributions over mean-variance efficient frontiers -- is a reusable, application-inspired testbed for comparing other generative models, including variational autoencoders, GANs, and tensor-network Born machines, once parameter counts can be matched.","The results indicate that the RBM's weakness on these instances is not simply a shortage of parameters, since it tends toward the uniform baseline even with the same parameter count as the QCBM; the difficulty lies in training or in the expressive limits of the quadratic energy model.","For quantitative finance, the pipeline connects classical portfolio optimization to quantum generative modeling, though the paper is explicit that this is an application-inspired benchmark rather than a practical portfolio solver."],"supporting_citations":[{"why":"Introduces the quantum circuit Born machine and its data-driven quantum circuit learning loop, the quantum model under test.","marker":"[2]"},{"why":"Defines the mean-variance portfolio-selection problem whose efficient frontiers generate the benchmark distributions.","marker":"[24]"},{"why":"Provides the fast learning algorithm for RBMs that the paper cites as the basis of its classical training.","marker":"[31]"},{"why":"Introduces persistent contrastive divergence, the specific Gibbs-chain gradient estimator used for RBM training.","marker":"[32]"},{"why":"Supplies the CMA-ES optimizer used to train the QCBM's circuit parameters.","marker":"[34]"},{"why":"Describes the CMA-ES evolution strategy that serves as the classical solver in the quantum pipeline.","marker":"[35]"}],"fun_headline_variants":["Quantum models beat classical at equal size on market data","Equal parameters, better fit: QCBM vs RBM on S&P 500","Quantum wins on financial benchmarks with same parameters","QCBM outperforms RBM on market data at identical parameter count"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that equal parameter counts make the comparison fair even though the quantum model is trained with a global optimizer on exact probabilities and the classical model with a short local sampling routine; if the optimizer difference, rather than the model family, drives the lower error, the central claim collapses.","fun_headline_variants_meta":{"raw":{"variants":["Quantum models beat classical at equal size on market data","Equal parameters, better fit: QCBM vs RBM on S&P 500","Quantum wins on financial benchmarks with same parameters","QCBM outperforms RBM on market data at identical parameter count"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1513,"prompt_tokens":1043,"completion_tokens":470,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":659,"completion_tokens_details":{"reasoning_tokens":400}},"tokens_in":659,"tokens_out":470,"duration_ms":5563,"temperature":1.0,"reasoning_tokens":400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:34:19.170050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the RBM with the same CMA-ES optimizer on exact likelihood values, or with persistent contrastive divergence using 10 to 100 Gibbs steps, and check whether its KL divergence drops to the QCBM's level for $N=6,8,10$. The paper's own appendix shows that for $N=4$ the longer chains already close much of the gap, so the decisive observation is whether the same happens at larger system sizes.","supporting_citations":[{"cited_title":"Benedetti, D","cited_arxiv_id":null,"evidence_quote":"Introduces the quantum circuit Born machine and its data-driven quantum circuit learning loop, the quantum model under test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the mean-variance portfolio-selection problem whose efficient frontiers generate the benchmark distributions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fast learning algorithm for RBMs that the paper cites as the basis of its classical training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces persistent contrastive divergence, the specific Gibbs-chain gradient estimator used for RBM training."},{"cited_title":"net/tutorial/rbm.html","cited_arxiv_id":null,"evidence_quote":"Supplies the CMA-ES optimizer used to train the QCBM's circuit parameters."}],"review_version":1}