{"id":"40a0fbdc-42f3-48f2-9363-5411a46b4496","arxiv_id":"2411.17821","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Approximate classical tensor-network simulations of the quantum proposal can reproduce the scaling advantage of quantum-enhanced Monte Carlo on small spin glasses, pointing to a quantum-inspired classical algorithm.","lead":"This paper tests the quantum-enhanced Monte Carlo algorithm on small spin glasses, maps out the best settings for its quantum proposal step, and shows that a classical tensor-network simulation can approximate the quantum proposal closely enough to keep the same scaling advantage. It suggests that the benefit of this Monte Carlo scheme may be available on classical computers, not only on quantum hardware.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MPS scaling advantage in Fig. 9 is inferred from a 7-point fit dominated by n<=5, where chi=4 is exact; the truly approximate n>=6 region may not support k=0.23.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the scaling exponents are extracted from small-system fits (n=3..9) with no error bars or larger-size validation. My stress-test sharpens this. For the central quantum-inspired claim, the chi=4 MPS is exact for n<=5, so those points cannot test the 'unconverged' regime; the approximate n>=6 points are the only evidence for the claimed advantage, and Fig. 8 shows the symmetry error growing linearly in n, making the extrapolation fragile. The paper itself acknowledges the need for larger-scale verification (Secs. III A and V B 2). I do not see an internal inconsistency or a more fundamental flaw: the MPS proposal with the Phi-corrected acceptance is a valid MCMC proposal, the spectral gaps are computed faithfully, and the order-of-magnitude crossover analysis in Sec. V B 3 is honest about prefactors. The concern is therefore about robustness of the quantitative extrapolation, not about the conceptual construction. The proposed n=10,11 computation would directly test whether the approximate-region scaling continues; the cheap restricted-fit diagnostic is a first step. Because the reader's CONDITIONAL verdict already reflects this uncertainty, my assessment leaves the verdict unchanged.","tokens_in":22287,"tokens_out":12141,"duration_ms":111126,"concrete_test":"Extend the exact spectral-gap calculation for the MPS chi=4 proposal to n=10 and n=11 on the same 100 disorder instances, constructing the full proposal matrix Q via 2^n TEBD evolutions and diagonalizing the resulting transition matrix P, then refit k over n=3..11 and over n>=6 only, reporting bootstrap confidence intervals. If the n>=6 fit (or the n=10,11 points) gives an upper 95% CI for k above about 0.7, the claimed scaling advantage is not supported; if k remains about 0.2 in the approximate region, the extrapolation holds. A cheap first diagnostic is to refit Fig. 9 restricted to n=6..9 and compare the slope and its uncertainty with k_c about 1.0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that an MPS with chi=4 retains the QEMC scaling advantage (k=0.23 vs. classical k about 1.0, Fig. 9) rests on an exponential fit over n=3..9. For chi=4, the MPS is exact for n<=5 (since 2^(n/2) <= 4), so the first three fit points are identical to the exact Trotterized quantum proposal and cannot test the approximate ('unconverged') regime. Only n=6..9 are truly approximate, and the fitted slope is heavily influenced by the exact-region points. This is not cosmetic: Fig. 8 shows the symmetry-breaking ratio Phi deviating more strongly with n (sigma(log2 Phi) grows linearly), with order-of-magnitude violations already at n=9; there is no evidence that the Phi-corrected acceptance maintains the exponential-gap closing rate once n clearly exceeds the exact region. The same fragility affects the t0.01 linear-scaling claim (Fig. 2b) used to argue that circuit depth does not spoil the speedup. The authors themselves acknowledge that larger-scale verification is required, but as it stands the numerical support for the headline 'quantum-inspired scaling advantage' is a 7-point fit with no confidence intervals and with the approximate regime contributing only four points.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper revisits the quantum-enhanced Markov chain Monte Carlo (QEMC) algorithm of Layden et al. (Nature 619, 282) and investigates three questions: the optimal choice of the mixing Hamiltonian strength and total evolution time; alternative proposal circuits (time-dependent schedules and a symmetric QAOA ansatz); and a classical quantum-inspired version in which the proposal step is approximately generated by a matrix product state (MPS) simulation. The central claim is that an MPS with small bond dimension (chi = 4), below the value required for exact state representation, reproduces the quantum spectral-gap scaling exponent k = 0.23 up to n = 9, matching the exact Trotterized quantum proposal (k = 0.24) and giving a large advantage over classical local (k ~ 1.0) and uniform (k ~ 0.96) updates. The authors also report an optimal gamma near the quantum phase transition, a linear growth of the time-to-target t_0.01 with n, and an optimal Trotter step dt = 0.8, and they provide cost estimates for the MPS proposal.","tokens_in":22610,"tokens_out":7568,"duration_ms":65396,"significance":"If the scaling claims hold, the paper would be significant in two ways: it identifies favorable operating regimes for QEMC, and it suggests that the QEMC scaling advantage does not require exact quantum evolution, opening the door to quantum-inspired classical samplers. The paper is also careful in several respects: it uses exact spectral-gap diagonalization for all reported gaps (the gold standard for mixing-time comparisons at the sizes considered), it reports averages over 100 disorder instances, and it includes a transparent cost model for the MPS proposal in Appendix F and Table I, with an explicit threshold formula. The work is framed as a proof of principle, and the authors repeatedly acknowledge the small-system limitation.","major_comments":[{"comment":"The load-bearing claim that an MPS with chi = 4 maintains the quantum scaling exponent k = 0.23 rests on an exponential fit over n = 3 to 9, but chi = 4 is exact for n <= 5 (the maximum bond dimension is 2^{floor(n/2)} = 4 for n <= 5). Only n = 6, 7, 8, 9 actually probe the approximate, truncated regime, and the fit includes no confidence intervals or error bars. With four approximate points, the fitted exponent cannot robustly distinguish k = 0.23 from substantially larger values, especially because Fig. 8 shows that the symmetry-breaking ratio Phi deviates more strongly with n for fixed chi. Please report fit uncertainties, refit excluding the n <= 5 exact points, and, if possible, extend to n = 10-12 using the sampling-based or iterative methods mentioned in the outlook.","section":"Sec. V.B.2, Fig. 9"},{"comment":"The role of the symmetry-breaking ratio Phi in the spectral-gap calculation of Fig. 9 is ambiguous. The text in Sec. V.B.1 states that for chi < 2^{n/2} one must include the ratio Q(s|s')/Q(s'|s) explicitly in the acceptance step, doubling the computation. However, Fig. 9 and its caption do not state whether the reported spectral gaps were obtained with this corrected acceptance rule or with the naive symmetric acceptance of Eq. (8). If the naive rule was used, the Markov chain does not satisfy detailed balance with respect to pi, and the computed 'spectral gap' is not the convergence rate to the target distribution. If the corrected rule was used, the authors should state this explicitly and explain how Phi was evaluated for the full transition matrix, since the proposal matrix is then no longer symmetric and the cost of constructing P changes.","section":"Sec. V.B.1 and V.B.2"},{"comment":"The parameters gamma = 0.45 and t = 12 are selected by a grid search on the same n = 3 to 9 instances used to evaluate performance, so the comparison with the randomized strategy in Fig. 3 is in-sample. The monotonic decrease of gamma_opt with n in Fig. 1(c) means there is no evidence that a fixed gamma = 0.45 remains optimal at larger sizes; the flattening at n = 9 is only three points. Similarly, the linear scaling of t_0.01 in Fig. 2(b) is based on seven points with no error bars, and this linearity is used in Sec. II.D to argue that the circuit depth does not spoil the asymptotic scaling advantage. Please provide out-of-sample tests (e.g., hold out instances or re-optimize on one half and test on the other) and report fit statistics, or at least explicitly quantify the sensitivity of the scaling conclusions to the parameter-choice procedure.","section":"Sec. III.A/B, Figs. 1-3, Eq. (12)"}],"minor_comments":[{"comment":"The displayed equality e^{-2iH_Z dt} [U_2nd]^T e^{+2iH_Z dt} = [U_2nd]^T is not generally correct, since H_Z is a non-trivial diagonal matrix and does not commute with the full second-order Trotterized unitary. The symmetry of the second-order Trotterized proposal follows from each Trotter factor being symmetric; please correct the derivation or remove the misleading intermediate expression.","section":"Appendix C, Eq. (C1)"},{"comment":"There is a typo in the sentence 'Q(s'|s) ≠ Q(s'|s) for chi < 2^{n/2}'; the two quantities should be Q(s'|s) and Q(s|s').","section":"Sec. V.B.1"},{"comment":"The phrase 'requirements fo practical quantum advantage' should be 'requirements for practical quantum advantage'.","section":"Sec. II"},{"comment":"The labels 'avg' and 'opt' in the inset of panels (b) and (c) are not defined in the caption; please add a definition or use a legend.","section":"Fig. 1"},{"comment":"The number of free parameters in the Bayesian optimization schedule (five equidistant interpolation points) appears only in Appendix B; stating this in the main text would help the reader assess the expressiveness of the optimized schedules in Fig. 12.","section":"Sec. IV.A, Appendix B"},{"comment":"The memory and time estimates assume a local spin dimension d = 2; this assumption is implicit and should be stated in the table caption or in the text around Eq. (20).","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The referee agrees with the reader's conditional assessment. The strongest issue is the symmetry-correction ambiguity in the MPS quantum-inspired results: if Fig. 9 uses the uncorrected acceptance, the spectral gaps reported there do not correspond to sampling from the target Boltzmann distribution, which would invalidate the central claim in its current form. The second issue is statistical: the chi = 4 scaling exponent is dominated by n <= 5 exact points. These are fixable with additional analysis and, ideally, larger-system data, so I recommend major revision rather than rejection. The paper is generally well written and the cost-model appendix is a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth a read for one reason: it shows, on small fully connected spin glasses, that a classical MPS simulation of the QEMC proposal can reproduce the quantum spectral-gap scaling with a tiny bond dimension (chi=4, k=0.23, versus classical k~1.0). That is a genuinely new idea, and it changes the conversation about QEMC from 'needs quantum hardware' to 'maybe a classical emulator is enough.' The paper also does a careful parameter study: optimal gamma sits near the phase transition, the total evolution time appears to grow roughly linearly in n, and the optimal Trotter step converges back to dt=0.8.\n\nThe honest caveats are real. The scaling exponents come from exponential fits over n=3..9, typically seven points, with no error bars in the key figures. gamma, t, and dt are selected on the same benchmark data, so there is mild in-sample fitting. The disorder distribution is never specified in the text, and no code or data are shipped. On the MPS result specifically, the stress-test concern holds up: for chi=4 the MPS is exact for n<=5, so the first three fit points are identical to the exact Trotterized proposal. The approximate regime contributes only n=6..9, and without confidence intervals the k=0.23 claim is not as strong as the figure suggests. The t0.01 linear-scaling evidence is similarly a single-threshold measure and could be fragile.\n\nThat said, the authors do not oversell. They call the MPS result a proof of principle and state explicitly that larger-scale verification is needed. The central conceptual claim—that exact quantum dynamics may not be required for the QEMC speedup—is supported as a proof of principle. I would not treat the quantitative scaling advantage as established, but it is a legitimate and publishable hypothesis.\n\nThis paper deserves a serious referee. The revision should add error bars, specify the disorder model, ship code, and ideally show the MPS fit restricted to the genuinely approximate regime. If those are addressed, it is a solid contribution to the QEMC and quantum-inspired sampling literature. I would bring it to a reading group focused on Monte Carlo methods, and I would cite it as a proof-of-principle reference if I wrote about QEMC.","headline":"The MPS quantum-inspired proposal is a genuinely new idea, but the scaling advantage rests on small-system fits with no error bars and deserves a conditional refereeing, not a desk reject.","tokens_in":23128,"tokens_out":2791,"would_cite":true,"duration_ms":24066,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P68","82B80","65C05"],"pacs":["03.67.Ax","05.10.Ln","75.10.Nr"],"model":"deepseek-v4-flash","headline":"Quantum-enhanced Monte Carlo may not need a quantum computer: approximate tensor-network proposals can preserve its scaling advantage over classical samplers.","keywords":["quantum-enhanced Monte Carlo","quantum-inspired algorithms","matrix product states","spectral gap scaling","spin glasses","Markov chain Monte Carlo","Trotterization","quantum advantage"],"falsifier":"Run the quantum-inspired proposal with a heavily compressed tensor network (a compression level that is exact only up to five spins) for $n=10$ to $14$ and measure the spectral-gap or autocorrelation exponent; if the exponent rises from roughly 0.23 toward 0.96–1.0, or if the time to reach a fixed gap grows faster than linearly, the claimed advantage is refuted.","tokens_in":22047,"feed_emoji":"⚛️","tokens_out":8751,"duration_ms":74081,"temperature":0.7,"pith_summary":"This paper studies quantum-enhanced Monte Carlo (QEMC), an algorithm in which a short Hamiltonian evolution proposes spin configurations that a classical Metropolis step accepts. It tries to establish two things: that QEMC has a stable optimal working point, with a mixing strength near the model's phase transition and a total evolution time that grows only linearly with system size, and that the proposal step can be moved from quantum hardware to a classical approximate simulator. The central numerical finding is that a matrix-product-state simulator with bond dimension $\\chi=4$, although far too small to represent the exact quantum state for $n>5$, reproduces the ideal spectral-gap scaling exponent $k\\approx0.23$ up to $n=9$, matching the exact quantum value $k=0.24$. If this persists at larger sizes, the empirical speedup of QEMC would not require exact quantum dynamics or quantum hardware, turning it into a quantum-inspired sampling method.","feed_headline":"Bond-dimension-4 simulator matches quantum proposal scaling","feed_subtitle":"Classical matrix-product-state proposals reproduce the spectral-gap exponent of the quantum-enhanced sampler up to n=9.","key_machinery":"The load-bearing mechanism is the proposal matrix $Q(s|s_0)=|\\langle s|U|s_0\\rangle|^2$ generated by time evolution under $H=(1-\\gamma)\\alpha H_c+\\gamma H_{\\mathrm{mix}}$, with a symmetric $U$ so that detailed balance reduces to the classical Metropolis rule. The paper replaces the exact unitary with a Trotterized time evolution (at step $dt\\approx0.8$) implemented on a matrix product state—a compressed tensor-network representation of a quantum state—via time-evolving block decimation, truncating to bond dimension $\\chi$. The central quantity is the spectral gap $\\delta$ of the full transition matrix $P(s_0\\to s)=Q(s|s_0)A(s|s_0)$; its exponential closing rate $k$ is the figure of merit that lets a small-$\\chi$ approximate state match the quantum proposal's scaling.","core_discovery":"On the paper's own terms, the central discovery is that the efficiency of QEMC survives approximate classical emulation of its proposal dynamics. It shows that the spectral gap $\\delta$ of the resulting Markov chain closes as $\\delta \\propto 2^{-kn}$ with $k\\approx0.23$ for a matrix-product-state proposal at bond dimension $\\chi=4$ (and $\\chi=8$), statistically indistinguishable from the exact continuous-time quantum proposal ($k=0.24$), while uniform and local classical proposals close much faster ($k\\approx0.96$-$1.0$). Because the acceptance step is classical and exact, approximation errors alter only the proposal quality, never the equilibrium distribution. The paper also finds a clear optimal proposal regime: a fixed mixing weight $\\gamma\\approx0.45$, slightly below the finite-size critical value $\\gamma_c\\approx0.50$, with total evolution time $t=12$, slightly outperforms the original randomized parameter choice, and the time to reach a fixed gap $\\delta=0.01$ grows roughly linearly with $n$, suggesting circuit depth is not an exponential bottleneck.","pith_inferences":["If the $\\chi=4$ scaling survives at larger $n$, the practical race shifts from building noisy quantum hardware to engineering a classical tensor-network or neural-network proposal with a competitive prefactor.","The same robustness argument suggests other cheap approximate quantum dynamics—e.g., neural quantum states or tree tensor networks with sublinear connectivity—could also preserve part of the speedup, which the paper lists but does not test.","A direct test would be to run the matrix-product-state proposal with $\\chi=4$ on systems $n=10$-$14$ using Markov-chain autocorrelation estimates rather than exact diagonalization; the paper's exponential matrix construction limits this regime.","The optimal $\\gamma$ sitting just below $\\gamma_c$ hints that the proposal's power comes from enhanced fluctuations near the spin-glass transition; if so, temperature or field schedules tuned to that critical region could further improve scaling."],"forward_implications":["A fixed working point ($\\gamma\\approx0.45$, $t=12$) is at least as good as the original randomized proposal, so the method may not require instance-by-instance fine-tuning.","The roughly linear growth of $t_{0.01}$ with $n$ means the total evolution time—and hence circuit depth—does not erase the polynomial speedup.","The optimal Trotter step coincides with $dt=0.8$, the value used in the original hardware experiment, so the hardware-friendly setting is also computationally near-optimal.","Approximation in the proposal step cannot corrupt the sampler: because acceptance is exact, the equilibrium distribution is protected, and only mixing efficiency can degrade.","A bond dimension $\\chi=4$ matrix product state, despite being inexact for $n>5$, preserves the quantum scaling exponent $k\\approx0.23$ up to $n=9$, implying the speedup may be reproducible classically."],"supporting_citations":[{"why":"Supplies the original QEMC algorithm, its randomized proposal parameters, and the empirical scaling baseline ($k=0.264$ vs. $k_c=0.94$) that this paper reanalyzes and compares against.","marker":"[11]"},{"why":"Introduces the wavefunction-collapse sampling idea that QEMC formalizes; provides the conceptual background for the proposal mechanism.","marker":"[10]"},{"why":"Provides the crossover-size and runtime-prefactor analysis used here to formulate conditions for a practical speedup.","marker":"[8]"},{"why":"Argues that QEMC requires fine-tuned quenches; the paper's fixed-parameter and randomized-strategy comparisons respond to this concern.","marker":"[13]"},{"why":"Proposes a symmetric QAOA-style proposal circuit that this paper benchmarks against the Hamiltonian-evolution proposal.","marker":"[14]"},{"why":"Supplies the matrix-product-state formalism used to build the approximate classical proposal.","marker":"[59]"},{"why":"Provides the time-evolving block decimation algorithm used to evolve the matrix-product-state proposals.","marker":"[66]"}],"fun_headline_variants":["Classical MPS proposal matches quantum Monte Carlo scaling","Tensor-network emulation preserves quantum sampling advantage","Bond-dim-4 classical proposal matches quantum scaling","Quantum-inspired Monte Carlo: classical emulation works"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the assumption that scaling trends measured for systems of 3 to 9 spins—especially the linear growth of the needed evolution time and the preserved speedup of a heavily compressed tensor-network proposal—continue to hold at larger sizes.","fun_headline_variants_meta":{"raw":{"variants":["Classical MPS proposal matches quantum Monte Carlo scaling","Tensor-network emulation preserves quantum sampling advantage","Bond-dim-4 classical proposal matches quantum scaling","Quantum-inspired Monte Carlo: classical emulation works"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000731,"raw_usage":{"total_tokens":3251,"prompt_tokens":902,"completion_tokens":2349,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":2289}},"tokens_in":518,"tokens_out":2349,"duration_ms":17335,"temperature":1.0,"reasoning_tokens":2289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:48:39.527389+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the quantum-inspired proposal with a heavily compressed tensor network (a compression level that is exact only up to five spins) for $n=10$ to $14$ and measure the spectral-gap or autocorrelation exponent; if the exponent rises from roughly 0.23 toward 0.96–1.0, or if the time to reach a fixed gap grows faster than linearly, the claimed advantage is refuted.","supporting_citations":[{"cited_title":"Lemieux, B","cited_arxiv_id":null,"evidence_quote":"Supplies the original QEMC algorithm, its randomized proposal parameters, and the empirical scaling baseline ($k=0.264$ vs. $k_c=0.94$) that this paper reanalyzes and compares against."},{"cited_title":"Mazzola, Sampling, rates, and reaction currents through reverse stochastic quantization on quantum com- puters, Physical Review A 104, 022431 (2021)","cited_arxiv_id":null,"evidence_quote":"Argues that QEMC requires fine-tuned quenches; the paper's fixed-parameter and randomized-strategy comparisons respond to this concern."},{"cited_title":"Layden, G","cited_arxiv_id":null,"evidence_quote":"Proposes a symmetric QAOA-style proposal circuit that this paper benchmarks against the Hamiltonian-evolution proposal."}],"review_version":1}