{"id":"ceae2104-b619-4698-8623-48c676e5c32d","arxiv_id":"2508.09870","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":2,"one_line_summary":"Boolean Fourier polynomials significantly outperform neural networks in generalizing to the sign structures of frustrated spin systems, suggesting these sign structures are learnable.","lead":"A physics and machine learning team shows that sign structures of frustrated quantum magnets, which typically fool neural networks, can be represented as Boolean polynomials that generalize far better, suggesting the sign problem for neural quantum states may be learnable. This matters because it points to a new route for variational simulation of frustrated spin systems.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fourier-polynomial generalization advantage may be an artifact of supervised benchmark design; practical transfer to NQS optimization is explicitly unshown.","rationale":"The reader's weakest assumption is that the sign structures have a compact Fourier representation; the paper's evidence is limited to numerical examples. I agree that this is a key empirical hypothesis. However, the more load-bearing concern is that the reported 'dramatic generalization' may be an artifact of the supervised benchmark setting itself: the polynomial ansatz is not variational, and the abstract states it cannot yet be used in variational optimization. Even if the Fourier representation is compact for the example systems, the learning complexity claim only matters for the NQS pipeline, where the sign structure is not given as labeled data but must be discovered through energy minimization. The paper's own limitation statement flags this gap, and the central claim's practical relevance depends on closing it. Since the full text is unavailable, I cannot verify whether the benchmark controls for equal sample access, model capacity, or the use of ground-truth labels. Therefore the reader's UNVERDICTED verdict is appropriate; my concern does not change the verdict, but it sharpens why the claim should not be accepted without a matched, variational-realistic comparison.","tokens_in":629,"tokens_out":3180,"duration_ms":34830,"concrete_test":"Re-run the paper's supervised learning benchmark under strictly matched conditions: train the Fourier polynomial (with the same sparse-regression or compressed-sensing coefficient selection described in the paper) and the neural network on identical training sets generated from the exact sign structure, with the same number of labeled samples, same input representation, and same noise model. Then (a) compare test-set accuracy after equal training samples; (b) repeat the comparison in the actual NQS setting by replacing exact labels with sign expectations sampled from a variational Monte Carlo state, and restrict the polynomial ansatz to a form that can be optimized variationally (e.g., evaluate only via local energy gradients). If the Fourier advantage vanishes or shrinks under equal data/query access and noisy labels, the central inference fails; if it persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that Boolean Fourier polynomials dramatically outperform neural networks in generalizing sign structures, so the learning complexity of sign structures is not an insurmountable obstacle—rests on the unverified assumption that the reported advantage reflects a property of the physical target (a low-degree or sparse Fourier spectrum) rather than an artifact of the supervised benchmark. The abstract explicitly concedes that the polynomial ansatz 'cannot yet be directly used in the context of variational optimization.' If the Fourier coefficients were fitted using ground-truth sign labels obtained from exact diagonalization and then evaluated on held-out configurations, that measures interpolation of a known function, not the ability to discover the sign structure from the unlabeled or indirectly sampled data available in a variational NQS calculation. The leap from 'a function class containing the target can generalize when trained with labels' to 'the learning complexity is not an obstacle' requires that the learner in the actual setting has comparable sample/query access. The only evidence for compactness is the paper's numerical examples; no analytic bound or phase-dependent argument for Fourier sparsity of frustrated ground-state sign structures is presented in the abstract, and full text is unavailable for verification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the sign structure of spin-1/2 ground states using Boolean Fourier analysis, representing sign configurations as polynomial functions on the Boolean hypercube. The authors argue that this representation offers a new language for understanding the learning complexity of sign structures. They report that Boolean Fourier polynomial ansätze dramatically outperform neural networks in supervised generalization for sign prediction on frustrated systems, although they explicitly state that such ansätze cannot yet be used directly in variational optimization. The paper also suggests that augmenting data with Boolean functions can improve neural-network sign prediction. The claims are presented as initial demonstrations rather than as a complete solution to the NQS sign problem.","tokens_in":969,"tokens_out":2328,"duration_ms":34001,"significance":"If the central claims hold up under full-text scrutiny, the paper would make a valuable contribution. It introduces a concrete, mathematically grounded alternative to neural-network parametrizations of sign structures, and it potentially reframes a known obstacle in neural quantum states as a function-class generalization problem. The explicit admission that the polynomial ansatz is not yet variational is honest and appropriately limits the scope. The promise of data augmentation with Boolean functions is also a useful practical direction. However, the significance depends on whether the reported generalization advantage is robust to protocol choices (Fourier degree, coefficient selection, baselines) and whether it reflects a property of the physical systems rather than an artifact of the supervised benchmark.","major_comments":[{"comment":"The central numerical claim—that Boolean Fourier polynomials dramatically outperform neural networks in generalization—cannot be assessed from the abstract because the protocol for selecting the Fourier degree d and the sparsity threshold is not stated. If these hyperparameters were chosen using the target sign data (e.g., by optimizing on the training set or selecting the best model on a validation set), the comparison would be a form of post hoc capacity tuning rather than an unbiased comparison of ansatz classes. The full text must specify whether d and the coefficient selection rule are fixed a priori, derived from system-independent arguments, or tuned on the target labels. This is load-bearing for the 'dramatically outperform' conclusion.","section":"Abstract, 'dramatically outperform' claim"},{"comment":"The abstract infers from numerical examples that sign structures are learnable because polynomial ansätze generalize well. This inference rests on the empirical hypothesis that the sign structures of the studied frustrated Hamiltonians have low-degree or sparse Boolean Fourier spectra. No analytic bound, scaling argument, or phase-dependent characterization is presented in the abstract. The full text should provide evidence for this compactness (e.g., spectral decay plots, degree-cutoff sensitivity, or comparison across system sizes). Without such evidence, the claim is limited to the specific instances studied and cannot support the broader statement about the complexity of sign structures in general.","section":"Abstract, 'complexity of sign structures is not an insurmountable curse'"},{"comment":"The supervised-learning result measures interpolation of a known sign function from labeled configurations. The actual NQS setting requires discovering the sign structure from unlabeled or indirectly sampled data, or from energy minimization. The abstract concedes that the polynomial ansatz cannot yet be used in that setting, so the jump from 'a function class can generalize when trained with labels' to 'the learning complexity is not an obstacle for NQS' is not yet made. The full text should either explicitly restrict the conclusion to supervised learning or provide a concrete route (e.g., a variational parametrization with learnable Fourier coefficients) and state what sample/query access the learner would need. The current hedging is commendable, but the significance claim outruns the demonstrated scope.","section":"Abstract, 'cannot yet be directly used in variational optimization'"}],"minor_comments":[{"comment":"The phrase 'spin-1/2 magnetic systems' is very broad; the abstract should name the specific frustrated Hamiltonians (e.g., square-lattice J1-J2, Kagome, etc.) and the system sizes used. This would improve reproducibility and let readers judge the scope.","section":"Abstract, general clarity"},{"comment":"This sentence is vague. It is unclear whether the augmentation adds transformed sign labels, auxiliary symmetries, or Fourier-feature inputs. One sentence of clarification would help.","section":"Abstract, 'augmenting data with Boolean functions'"},{"comment":"No information is given about error bars, number of independent runs, or baselines for the neural networks. The full text should report these to support the 'dramatically outperform' comparison.","section":"General"}],"recommendation":"uncertain","confidential_remarks":"This review is based solely on the abstract because the full text was not made available. The core claims are plausible and the authors are appropriately cautious, but the technical validity of the Fourier truncation choices and the comparison protocol cannot be verified from the abstract. I recommend that the editor ensure the full manuscript is available for a complete review before any decision. My 'uncertain' recommendation reflects the lack of full-text evidence rather than a specific flaw in the work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: the abstract reports that Boolean Fourier polynomials generalize markedly better than neural networks at predicting sign structures of frustrated spin-1/2 ground states, and that data augmentation with Boolean functions helps NNs too. That is a concrete, new result relative to the prior work showing NN generalization failure, and it opens a representational family worth exploring.\n\nWhat the paper does well: it frames sign structures as polynomial functions on the Boolean hypercube, which is a clean and natural language for the problem. The claims are appropriately hedged — the abstract explicitly says these polynomials \"potentially serve\" and \"cannot yet be directly used\" in variational optimization. That is honest. The data-augmentation finding is a useful practical aside, not just decoration.\n\nSoft spots, in proportion: the central interpretive step — from \"a function class can generalize when trained on labeled sign data\" to \"learning complexity is not an insurmountable curse\" — rests on a supervised benchmark. In variational NQS one does not have ground-truth sign labels; you have an unlabeled or indirectly sampled setting. The paper itself concedes it cannot yet be used there, so the stress-test concern is real, but it is a limit on the strength of the conclusion, not a flaw in the reported experiments. I cannot check the harder details from the abstract: how the Fourier degree and coefficient sparsity were chosen, whether the selection used the target sign data, and what the actual error bars and baselines look like. The abstract gives no visibility into those choices. That is a genuine unknown, and it is the difference between a solid empirical demonstration and a possibly overfit comparison.\n\nBottom line: this is a serious, well-scoped piece of work that deserves a careful referee. The mathematics is standard but the application is new and the experimental claims are specific. I would bring it to a reading group for NQS people and would likely cite it if I worked on sign representations. The main verdict should be: referee it, and ask for full details on model selection and for a clear statement that the comparison protocol did not use target labels to pick the Fourier truncation. If that holds, the paper is a useful contribution even if the variational promise remains open.","headline":"A well-posed new ansatz family for sign structures with a genuine generalization result; the main caveat is the leap from supervised benchmarks to variational optimization, which the paper itself concedes.","tokens_in":1326,"tokens_out":1107,"would_cite":true,"duration_ms":16179,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that Boolean Fourier polynomials, representing sign structures of frustrated spin-1/2 ground states, learn to predict signs from far fewer samples than neural networks, suggesting that the learning complexity of sign struct","keywords":["Boolean Fourier analysis","sign structure","frustrated magnetism","neural quantum states","generalization","spin-1/2 systems","machine learning for quantum many-body","supervised learning"],"falsifier":"Prepare a frustrated spin-1/2 ground state (e.g., on a kagome lattice) via exact diagonalization for a moderate system size, compute the full Boolean Fourier spectrum of its sign structure, and check whether coefficients beyond a low degree (say degree 3) are negligible. If low-degree truncation fails to predict signs on held-out configurations with accuracy comparable to the paper's reported results, the central claim is falsified.","tokens_in":642,"feed_emoji":"🧲","tokens_out":3222,"duration_ms":33447,"temperature":0.7,"pith_summary":"The paper asks whether the sign structure of frustrated quantum ground states is inherently hard to learn. It proposes representing the sign as a Boolean polynomial over spin configurations, and shows that low-degree, sparse such polynomials, trained on few samples, generalize far better than standard neural-network quantum-state architectures. If true, this shifts the bottleneck from “the sign structure is intractable” to “we need architectures that capture Boolean structure.” The paper also shows that augmenting neural networks with Boolean-function features improves their sign prediction.","feed_headline":"Boolean polynomials learn quantum signs neural nets miss","feed_subtitle":"For frustrated spin-1/2 ground states, few samples suffice for Fourier-polynomial sign prediction.","key_machinery":"Boolean Fourier analysis: a sign function $\\sigma:\\{0,1\\}^N \\to \\{\\pm1\\}$ is expanded as $\\sum_{S\\subseteq[N]} \\hat{\\sigma}(S)(-1)^{\\sum_{i\\in S}s_i}$. The degree and sparsity of this expansion control how many samples are needed to learn the function; the paper uses this to analyze sign structures and to construct polynomial ansätze.","core_discovery":"On numerical evidence, sign structures of frustrated spin-1/2 ground states admit compact representations in the Boolean Fourier basis, and polynomials built from this basis dramatically outperform neural networks in supervised sign prediction. This is presented as evidence that the learning complexity of sign structures is not an insurmountable obstacle for neural quantum states, and that augmenting neural networks with Boolean features improves their generalization.","pith_inferences":["If sign structures are generically low-degree in the Boolean basis, a provable sample-complexity bound connecting Fourier degree to generalization would be a natural next step, turning the empirical advantage into a theoretical one.","The same Fourier lens could be applied to other sign problems in quantum Monte Carlo, where the complexity of the sign structure directly controls the severity of the sign problem.","A testable extension: verify on a larger family of frustrated lattices whether the Fourier degree scales slowly with system size, which would indicate the approach works beyond the specific examples studied.","The polynomial representation may also aid classical simulation methods by providing compact classical descriptions of ground states that lie outside tensor-network territory."],"forward_implications":["Low-degree Boolean polynomial ansätze can predict signs of frustrated ground states from far fewer training samples than neural-network ansätze, suggesting a cheaper supervised route to sign-structure reconstruction.","The failure of standard neural-quantum-state architectures on frustrated systems is not due to the intrinsic complexity of the sign structure but to the architecture's inductive bias.","Data augmentation with Boolean-function features improves sign prediction by neural networks, a directly usable recipe.","This opens the possibility of designing new neural-quantum-state architectures that explicitly build in a Boolean-Fourier bias, potentially enabling variational optimization for frustrated systems."],"supporting_citations":[],"fun_headline_variants":["Fourier polynomials outperform neural nets on spin signs","Boolean Fourier analysis boosts quantum sign prediction","Polynomial ansätze beat neural networks on frustrated signs","Data augmentation with Boolean functions aids sign learning","Compact Fourier forms reveal learnable quantum sign structures"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The frustrated Hamiltonians examined have sign structures that are sparse or low-degree in the Boolean Fourier basis; if the true spectrum is dense or high-degree, the polynomial ansatz's generalization advantage will not hold.","fun_headline_variants_meta":{"raw":{"variants":["Fourier polynomials outperform neural nets on spin signs","Boolean Fourier analysis boosts quantum sign prediction","Polynomial ansätze beat neural networks on frustrated signs","Data augmentation with Boolean functions aids sign learning","Compact Fourier forms reveal learnable quantum sign structures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000169,"raw_usage":{"total_tokens":1061,"prompt_tokens":663,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":328}},"tokens_in":407,"tokens_out":398,"duration_ms":4701,"temperature":1.0,"reasoning_tokens":328,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:43:51.952588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Prepare a frustrated spin-1/2 ground state (e.g., on a kagome lattice) via exact diagonalization for a moderate system size, compute the full Boolean Fourier spectrum of its sign structure, and check whether coefficients beyond a low degree (say degree 3) are negligible. If low-degree truncation fails to predict signs on held-out configurations with accuracy comparable to the paper's reported results, the central claim is falsified.","supporting_citations":[],"review_version":1}