{"id":"f92d7b41-1303-4341-bf20-0b1f451cd933","arxiv_id":"2412.12398","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A variational Monte Carlo algorithm that trains a restricted Boltzmann machine quantum state by sampling a fitted Ising surrogate with a Trotterized quantum circuit, demonstrated on small spin and molecular systems.","lead":"A quantum-assisted Monte Carlo method trains neural-network quantum states by sampling from a fitted Ising 'surrogate' distribution using a Trotterized quantum circuit, avoiding mid-circuit measurements and storing only polynomially many parameters. The authors test it on small spin chains and active-space molecules, reporting chemical accuracy when combined with zero-variance extrapolation, but the efficiency claims outrun the provided error bounds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The practical algorithm drops the kappa residual from Eq. (5) and uses the quadratic surrogate phi as if it were the RBM diagonal; no error bound connects the O(n^2) fit to the energy estimate, so polynomial-efficiency claims lack support.","rationale":"The reader's weakest-assumption analysis correctly identifies the surrogate fit as the load-bearing point. My reading agrees and sharpens it: the issue is not merely the absence of an error bound for the fit, but the concrete ambiguity in how Eq. (8) is evaluated. If kappa is ignored, the estimator is biased by the fitting error epsilon(v); if kappa is used, the variance is controlled by the importance ratio rho_vv/phi, which is exactly the uncontrolled object. Either way, the central efficiency claim is not proven. The empirical demonstrations on n=8 (XXZ and LiH) and n=12 (H2O) are genuine evidence that the method works in those small, specific regimes, and the spectral-gap comparisons show a plausible mixing advantage for the quantum proposals. However, the abstract generalizes these findings to a polynomial-resource, constant-measurement algorithm of 'optimal computational efficiency,' and that generalization is not supported by the analysis in Sections II-IV and Appendices C-D. The Trotterized proposal's approximation is a lesser concern because Metropolis-Hastings acceptance corrects the stationary distribution; the spectral-gap numerics for proposals D-H are suggestive but not a proof of n-independent mixing. The 'constant measurement count' phrase is also internally inconsistent with N_s = O(Var/epsilon^2) if variance grows with n. For these reasons the reader's CONDITIONAL verdict is appropriate: the method merits further scrutiny and independent reproduction, but the strongest claims should be softened or accompanied by rigorous error bounds. No new information here changes that verdict, so I recommend UNCHANGED.","tokens_in":45052,"tokens_out":5689,"duration_ms":56426,"concrete_test":"Run the algorithm on n=20 and n=30 random RBM parameter instances using the paper's O(n^2) fitting protocol. For each sampled configuration v, compute the exact unnormalized rho_vv(v) and the fitted phi(v). Then: (1) evaluate max|log(rho_vv/phi)| over the sampled configurations; (2) compare the unweighted estimator mean(E_loc) with the kappa-weighted estimator of Eq. (8) (weight w(v)=rho_vv(v)/phi(v)) and with exact diagonalization; (3) compute the effective sample size N_eff = (sum w)^2 / sum w^2. If the unweighted and weighted estimates differ by more than the target accuracy (e.g., 1e-3 a.u. for molecular benchmarks), or if N_eff/N_s falls below 1% for any instance, then the O(n^2) surrogate does not control the energy estimate and the polynomial-efficiency claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claimed optimal efficiency rests on the surrogate distribution phi(v) (a quadratic Ising model) faithfully standing in for the RBM diagonal rho_vv(v) when estimating Eq. (8). Eq. (5) is only an identity if the non-constant prefactor kappa(v) is included, but the fitting protocol in Appendix D targets rho_vv approximately phi and never constructs or bounds kappa: it ends with rho_vv = phi*kappa + epsilon(v), where epsilon(v) is called the fitting error. Thus the estimator in Eq. (8) faces a dilemma. If kappa is omitted, as the applications appear to do when phi is described as 'a close approximant' to rho_vv, the estimate is biased by the uncontrolled fitting error. If kappa is instead included through the known RBM functional form, the estimator is unbiased only with perfect kappa, and its variance is governed by the importance ratio rho_vv/phi, exactly the uncontrolled quantity. Theorem II.1 guarantees only that some high-degree polynomial plus a kappa can represent rho_vv; it says nothing about the k=2 truncation actually used. Appendix D explicitly labels the q-largest configuration selection 'heuristic' and admits 'it might be possible to curate special instances where it might fail.' No error bound connects the O(n^2) fit to the bias or variance of mu_H. The abstract's claims of polynomial efficiency, constant measurement count, and optimal computational efficiency are therefore unsupported. The stated N_s = O(Var(mu)/epsilon^2) also conflicts with 'constant measurement count' unless the variance is shown to be n-independent.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a quantum-assisted variational Monte Carlo protocol for training Restricted Boltzmann Machine neural quantum states. A quadratic Ising surrogate distribution phi(v) is fitted from O(n^2) configurations of the RBM diagonal rho_vv; quantum circuits based on Trotterized time evolution are used as MCMC proposal distributions to sample phi(v); energy and gradients are estimated from these samples; ground states of the XXZ spin model and LiH/H2O electronic Hamiltonians are computed and benchmarked against exact diagonalization and CASSCI. The abstract claims that the algorithm scales linearly in circuit width and depth, needs a constant number of measurements, avoids mid-circuit measurements, requires polynomial storage, and is optimally efficient.","tokens_in":45311,"tokens_out":7661,"duration_ms":70600,"significance":"If the efficiency claims were rigorously supported, the paper would be a valuable contribution to quantum-assisted NQS training: the idea of using a surrogate distribution to make the RBM diagonal samplable, and the quantum-circuit proposal for faster mixing, are appealing. The numerical benchmarks for small systems show good agreement with exact references, and Appendix B gives a standard but correct monomial-expansion proof of representability. However, the central resource claims are not backed by a control on the surrogate fitting error, so the paper's main advertised contribution is currently not established.","major_comments":[{"comment":"The central estimator is not connected to Eq. (6) by any controllable approximation. Eq. (5) defines rho_vv(X) ∝ kappa(v,X) phi_v(l(X),J(X)), and Eq. (8) samples from phi with kernel kappa. However, the fitting protocol in Appendix D fits log rho_vv to a quadratic polynomial and terminates with rho_vv = phi*kappa + epsilon(v), where epsilon(v) is called the fitting error; the algorithm never constructs or bounds kappa(v). If the applications implicitly omit kappa and treat phi as 'a close approximant' to rho_vv, then Eq. (8) is biased by an uncontrolled fitting error. If kappa is instead included through the known RBM functional form, the estimator is unbiased only with perfect kappa, and its variance is governed by the importance ratio rho_vv/phi, which is not bounded. Theorem II.1 guarantees only the existence of some high-degree polynomial with a prefactor kappa bounded by the exponentially large expression M(2^n − sum_j C(n,j)); it says nothing about the k=2 truncation actually used. Appendix D itself labels the q-largest configuration selection 'heuristic' and admits 'it might be possible to curate special instances where it might fail.' Thus the polynomial-efficiency and constant-measurement claims do not follow from the presented analysis.","section":"Section III, Eq. (8); Appendix D"},{"comment":"The 'constant in measurement count' claim is not established by the paper's own sample-complexity formula. Step 2 of Section III gives N_s = O(Var(mu_H)/epsilon^2), and Section IV states that the number of circuit queries is O(N_s) and is 'independent of system size n.' But Var(mu_H) is never shown to be independent of n; because mu_H depends on the ratio rho_vv/phi through kappa, the variance may grow with n and with the misfit of the quadratic surrogate. Without a bound on Var(mu_H), the statement that the measurement count is constant in system size is unsupported.","section":"Section III, step 2; Section IV"},{"comment":"The abstract's claim that the algorithm 'scales linearly with circuit width and circuit-depth' is inconsistent with the Trotter analysis in Section IV. The Trotter number is N_trot approximately O((sum_{i=1}^{n^2+2n} ||chi_i||_1 tau)^{1+1/p}/eps^{1/p}), where the sum runs over O(n^2) terms. Even if a single Trotter layer has depth O(n), the total circuit depth is O(n) N_trot and therefore grows superlinearly with n unless the couplings l_i and J_ij are assumed to vanish or are uniformly bounded in a way not stated in the paper. If 'linear in circuit-depth' is intended per Trotter layer, this should be stated explicitly rather than as an overall scaling claim.","section":"Section IV; Abstract"},{"comment":"The numerical demonstrations are limited to n=8 spins and to active spaces of 8 and 12 spin orbitals, and they do not test the resource scaling advertised in the abstract. In particular, they provide no evidence that the O(n^2)-configuration fitting step of Appendix D yields a reliable surrogate as n grows, especially in view of the admitted heuristic nature of the q-largest configuration search. The agreement with exact diagonalization and CASSCI is encouraging for small n, but it is not a substitute for an error bound or for scaling tests.","section":"Section V"}],"minor_comments":[{"comment":"The notation shifts between v_i in {1,-1} and integer-indexed binary configurations; Appendix A should state the explicit mapping between the two conventions.","section":"Section II; Appendix A"},{"comment":"There are several typos and incomplete formulas: 'doesnt', 'precludes the need to include the factor ... in Eq.9' has mismatched parentheses, and 'See SectionFinAppendix' is missing a space.","section":"Section III; Appendix D"},{"comment":"The algorithm is described as returning the q-largest configurations, but the text later admits the procedure is heuristic and may fail; the output should be labeled as an approximate list to avoid ambiguity.","section":"Appendix D.2"},{"comment":"The text states that a k-qubit R_zz...z gate requires 'only 2k two-qubit entangling gates,' but the CNOT decomposition in Fig. 7 uses 2(k-1) CNOT gates; this should be corrected for consistency with the main-text statement about R_zz.","section":"Appendix F"},{"comment":"The notation O(2(nm+n+m)+1) for the total number of observables is confusing; since the leading term is O(nm), the display should be simplified or clarified.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core algorithmic idea is genuinely new: fit a quadratic Ising surrogate to the RBM diagonal using O(n^2) configurations, then use a Trotterized Hamiltonian-evolution proposal inside Metropolis-Hastings to sample that surrogate, and train a complex-valued RBM with the resulting configurations. That construction is not in the prior literature, and the empirical mixing improvement (spectral gap decay, l2 error on the sampled distribution) is shown across several proposals with a reported factor of ~3 in gap slope. The benchmarks are real: XXZ spin chains, LiH and H2O CASSCI ground states, including distorted geometries, with ZVE bringing errors below chemical accuracy. The monomial-expansion theorem (Appendix B) is standard and the proof is fine; the analytical gradients in Appendix E are a plus.\n\nThe soft spot is exactly where the abstract makes its loudest claims. The polynomial-efficiency and \"constant measurement count\" statement are not supported by the analysis. The identity rho = kappa*phi is only true if kappa is kept, but the fitting protocol in Appendix D ends with rho = phi*kappa + epsilon and never bounds epsilon; the algorithm then uses phi as a \"close approximant\" to rho, effectively dropping kappa, so the estimator in Eq. (8) is biased by an uncontrolled fitting error. Theorem II.1 is an existence statement about some high-degree polynomial plus a kappa; it says nothing about the k=2 truncation actually used. The q-largest configuration selection is explicitly admitted heuristic and can fail on curated instances. The stated N_s = O(Var/eps^2) contradicts \"constant measurements\" unless the variance is shown to be n-independent, and no such bound is given. Trotter error is analyzed for the circuit but not tied to the quality of the sampled phi. So the resource claims should be read as conjectures, not theorems.\n\nThat said, the paper is honest enough to include the heuristic caveat in Appendix D, the numerical work is reproducible in principle, and the central idea is worth taking seriously. It is a methods paper with a plausible algorithmic contribution, not a proof of optimal scaling.\n\nWho benefits: people working on NQS training, quantum-assisted MCMC, and hybrid quantum-classical algorithms for chemistry. A serious referee should engage with it, but the next revision needs to either prove a bound on the surrogate error under stated assumptions or substantially soften the abstract. I would send it to review with a request for major revision.","headline":"A genuinely new surrogate-network plus Trotterized-proposal scheme for training complex RBMs, with credible small-system results, but the optimal-efficiency claims outrun the error analysis.","tokens_in":45918,"tokens_out":2701,"would_cite":false,"duration_ms":26219,"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":"Restricted Boltzmann machine representations of quantum states can be trained by quantum-assisted variational Monte Carlo with linear circuit width and depth, constant measurement count, and polynomial storage, covering both amplitude and…","keywords":["neural quantum states","restricted Boltzmann machine","quantum-assisted variational Monte Carlo","surrogate distribution","Hamiltonian simulation","zero-variance extrapolation","ground-state energy","Markov chain Monte Carlo"],"falsifier":"Fix a driver Hamiltonian whose dominant local-energy contributions come from configurations that the $q$-largest heuristic misses or that lie outside the fitted $O(n^2)$ set; compute $\\mu_{\\langle H\\rangle}$ and its variance from Eq. (8) with the paper's constant-$N_s$ protocol as $n$ grows. If the estimator develops a bias that grows with $n$, or if the variance stops being compensated by the fixed measurement count, the polynomial-efficiency claim collapses.","tokens_in":44771,"feed_emoji":"⚛️","tokens_out":8571,"duration_ms":71955,"temperature":0.7,"pith_summary":"This paper claims that neural-network quantum states—restricted Boltzmann machines with complex-valued parameters in the concrete implementation—can be trained by a variational Monte Carlo loop whose quantum resources are polynomially efficient. The central move is to fit a simple quadratic surrogate distribution to the RBM's diagonal probability weights from O($n^{2}$) configurations, then use a short Trotterized quantum circuit as the Markov-chain proposal for sampling that surrogate. If the claim holds, ground states of spin models and molecular electronic Hamiltonians can be learned with linear circuit width and depth, constant measurement count (independent of n), no mid-circuit measurements, and polynomial storage, with amplitude and phase both learned from the same complex parameters. The paper validates the workflow on the XXZ spin chain and on LiH and H2O potential energy surfaces, reporting near-exact energies and chemical accuracy after zero-variance extrapolation.","feed_headline":"Quantum sampler trains neural quantum states at chemical accuracy","feed_subtitle":"Surrogate RBM sampling with short Trotterized circuits keeps width, depth, storage, and measurement count polynomial.","key_machinery":"The load-bearing object is the surrogate network $G_2$: a fully-connected Ising model over $n$ classical spins with distribution $\\phi(\\vec{v})\\propto \\exp\\bigl(-\\beta\\sum_i l_i(\\vec{X})v_i + \\sum_{i,j} J_{ij}(\\vec{X})v_i v_j\\bigr)$, fitted to the diagonal RBM distribution $\\rho_{vv}(\\vec{X})$ from $O(n^2)$ configurations. Theorem II.1 guarantees that any discrete distribution $P(\\vec{v})$ can be written as $Z\\,\\kappa(\\vec{v})\\,\\phi(\\vec{v})$ with $\\phi$ of user-chosen polynomial degree $k$, which is what lets the surrogate stand in for the full RBM. The sampling machinery is the quantum proposal: a Trotterized time evolution $U(\\tau,\\gamma)=e^{-i(\\gamma h_1+(1-\\gamma)h_2)\\tau}$ with $h_1$ generated by the surrogate's fields and couplings and $h_2$ a transverse-field mixer, used as the Metropolis-Hastings proposal distribution; its spectral gap decays roughly three times more slowly than classical local, uniform, or Haar-random proposals, giving faster mixing. A final piece is the kernel $\\kappa(\\vec{v}, \\vec{X})$, which carries the residual difference between surrogate and true distribution and is included in the local-energy estimate of Eq. (8).","core_discovery":"The paper's central discovery is that the intractable diagonal distribution $\\rho_{vv}(\\vec{X})$ of a complex-valued RBM ansatz can be factored as $\\kappa(\\vec{v},\\vec{X})\\,\\phi_{\\vec{v}}(\\vec{l}(\\vec{X}),\\vec{J}(\\vec{X}))$, where $\\phi$ is the distribution of a fully-connected quadratic Ising surrogate network and $\\kappa$ is a configuration-dependent prefactor; Theorem II.1 proves such a factorization exists for any discrete distribution over $n$-bit configurations. The authors give a constructive, data-driven recipe that fixes the surrogate parameters $(\\vec{l},\\vec{J})$ by weighted least-squares plus BFGS refinement on only $O(n^2)$ sampled configurations, choosing the $q$ largest-probability configurations plus random ones. Sampling from $\\phi$ is then done by Metropolis-Hastings with a quantum proposal $U(\\tau,\\gamma)=e^{-i(\\gamma h_1+(1-\\gamma)h_2)\\tau}$ built from the surrogate couplings $h_1=\\sum_i l_i \\sigma_z^{(i)}+\\sum_{i,j} J_{ij}\\sigma_z^{(i)}\\sigma_z^{(j)}$ and a transverse mixer $h_2=\\sum_i \\sigma_x^{(i)}$, implemented with Trotterization. The paper claims this yields linear circuit width $O(n)$, linear circuit depth $O(\\tau n)$, $O(N_s)$ queries independent of $n$, no mid-circuit measurements, $O(mn)$ storage, and analytical gradients that involve both amplitude and phase; benchmark results on XXZ, LiH, and H2O are reported to match exact diagonalization and CASSCI, with zero-variance extrapolation bringing errors to or below chemical accuracy.","pith_inferences":["The constant-measurement-count claim is only as strong as the surrogate fit: if the residual kernel $\\kappa(\\vec{v})$ varies strongly over configurations that carry large local energy, the variance of Eq. (8) could grow with $n$ and force $N_s$ to grow; a stress test on non-local (Jordan-Wigner) molecular Hamiltonians with fixed $N_s$ would settle this.","The surrogate construction is distribution-agnostic, so the same fitting-plus-quantum-proposal recipe could be applied to autoregressive or feed-forward neural quantum states, not just RBMs; the paper proves generality but only demonstrates it on RBMs.","The $q$-largest configuration selection is a heuristic, and the paper's own Appendix D acknowledges adversarial instances where it fails; one can probe this by constructing RBM parameter sets with two well-separated probability peaks and checking whether the fitted surrogate keeps the second peak accurate.","Since ZVE corrects only the energy estimate, not the trained parameters, claims about the quality of the final state rest on the variance diagnostic; an independent fidelity check against the exact state would be a stronger validation than energy alone."],"forward_implications":["Training an RBM neural quantum state with quantum-assisted sampling needs $O(n)$ qubits, $O(\\tau n)$ depth per Trotter layer, and $O(N_s)$ queries with $N_s$ independent of $n$, so the end-to-end resource count is polynomial in system size.","Because the RBM parameters are complex-valued and the surrogate maps them holistically, both amplitude and phase of the target state are learned without a separate classical phase-preprocessing step, enlarging the trial space over earlier amplitude-only quantum training.","Quantum proposals built from the surrogate Hamiltonian show a spectral gap that decays about three times slower than classical local, uniform, or Haar-random proposals, which translates into shorter mixing time and roughly five-fold smaller $\\ell^2$ sampling error in the paper's $n=8$ tests.","On the XXZ model, relative ground-state energy errors remain below $5\\times10^{-3}$ across antiferromagnetic, XY, and ferromagnetic phases when zero-variance extrapolation is applied.","The same workflow reproduces CASSCI ground-state energies for LiH and H2O over stretched bonds and distorted angles, with errors at or below the chemical-accuracy threshold ($10^{-3}$ a.u.) for most points."],"supporting_citations":[{"why":"Defines the RBM neural quantum state and the variational Monte Carlo workflow that this paper extends.","marker":"[1]"},{"why":"Proves RBMs are universal approximators for discrete probability distributions, grounding the surrogate representation claim.","marker":"[20]"},{"why":"Is the previous quantum-assisted RBM training protocol that this algorithm claims to surpass, especially in storage and phase handling.","marker":"[28]"},{"why":"Introduces the earlier quantum circuit for sampling RBM distributions with mid-circuit measurements and repeat-until-success, the key comparison for circuit width and measurement cost.","marker":"[43]"},{"why":"Supplies the Trotter-Suzuki product-formula error bounds used to size $N_{\\text{trot}}$ for the quantum proposal circuit.","marker":"[61]"},{"why":"Is the Metropolis-Hastings update rule, Eq. (9), on which the sampling loop is built.","marker":"[82]"},{"why":"Provides the spectral-gap and mixing-time analysis used to compare quantum and classical proposals.","marker":"[89]"},{"why":"Is the zero-variance extrapolation technique used to reach chemical accuracy from the converged RBM energies.","marker":"[95]"}],"fun_headline_variants":["Polynomially efficient quantum VMC trains NQS to chemical accuracy","Linear-depth quantum sampler for neural quantum states","Quantum-assisted MCMC: polynomial resources, no mid-circuit reads","Surrogate Ising sampling on quantum circuits matches CASSCI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quadratic surrogate distribution $\\phi$ fitted from $O(n^2)$ configurations faithfully reproduces the RBM diagonal distribution over the configurations that dominate the energy and gradient estimates, with the residual captured by $\\kappa(\\vec{v})$; the paper supplies no error bound for the fit, and Appendix D concedes that the $q$-largest-configuration selection is heuristic and can fail on specially curated instances.","fun_headline_variants_meta":{"raw":{"variants":["Polynomially efficient quantum VMC trains NQS to chemical accuracy","Linear-depth quantum sampler for neural quantum states","Quantum-assisted MCMC: polynomial resources, no mid-circuit reads","Surrogate Ising sampling on quantum circuits matches CASSCI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000391,"raw_usage":{"total_tokens":2133,"prompt_tokens":1097,"completion_tokens":1036,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":713,"completion_tokens_details":{"reasoning_tokens":968}},"tokens_in":713,"tokens_out":1036,"duration_ms":9116,"temperature":1.0,"reasoning_tokens":968,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:07:48.355080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a driver Hamiltonian whose dominant local-energy contributions come from configurations that the $q$-largest heuristic misses or that lie outside the fitted $O(n^2)$ set; compute $\\mu_{\\langle H\\rangle}$ and its variance from Eq. (8) with the paper's constant-$N_s$ protocol as $n$ grows. If the estimator develops a bias that grows with $n$, or if the variance stops being compensated by the fixed measurement count, the polynomial-efficiency claim collapses.","supporting_citations":[{"cited_title":"Sajjan, S","cited_arxiv_id":null,"evidence_quote":"Is the previous quantum-assisted RBM training protocol that this algorithm claims to surpass, especially in storage and phase handling."}],"review_version":1}