{"id":"fef94863-379b-4e54-be02-05f43c181e09","arxiv_id":"2412.02974","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A tensor-network-based MCMC algorithm using stochastic projectors removes the systematic error of finite bond dimension truncation and shows exponential variance reduction on the 2D Ising model.","lead":"This paper proposes a new way to combine Markov chain Monte Carlo sampling with tensor network representations, replacing fixed low-rank approximations with random projectors that are sampled during the simulation. The goal is to remove the bias usually introduced by tensor network truncation while keeping the computational benefits, with demonstrations on the two-dimensional Ising model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies the scale factors that make sampled projectors average to the identity; Eq. (3) is asserted rather than derived, so the unbiasedness claim—and hence the exponential-variance and sign-problem claims—rest on an unverified construction.","rationale":"The reader's weakest_assumption identifies exactly the most load-bearing point of the argument. I checked whether Appendix D supplies the missing construction: it defines weights and marginals and gives an overflow-free dynamic-programming sampler, but it never gives the scaled stochastic projector and never proves Eq. (3). The footnote references scale factors but keeps them implicit. This is not a fatal mathematical error—the 1/q_i construction would close the gap—but it is a genuine omitted derivation in the core identity. The numerical demonstrations are consistent with the claim but cannot replace the identity: a small bias could be hidden inside the statistical error of the d=6 data, and the deterministic TRG curves used for comparison also carry controlled but nonnegligible bias. Therefore the paper should be conditionally accepted pending the explicit construction and the small numerical check; the reader's verdict remains appropriate.","tokens_in":15843,"tokens_out":11453,"duration_ms":120866,"concrete_test":"Standalone numerical verification of Eq. (3): fix small r and d (e.g., r=4, d=2 and r=6, d=3), choose random dual-orthonormal bases, set weights w_i via Eq. (D1), compute marginals q_i from Eqs. (D3)/(D14), define P(S)=Σ_{i∈S} η_i ξ_i^*/q_i, and average P(S) over all C(r,d) subsets (or 10^6 draws). If the mean is not I_r to machine precision, the scale factors are wrong. Then run the same check inside the 4×4 Ising TRG initialization and compare the mean of the stochastic contraction with the exact partition function; a mismatch beyond statistical error refutes unbiasedness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method eliminates systematic error is true only if the stochastic projectors satisfy E[P(θ)] = I_r. For the Appendix D sampling, where a subset S is drawn with probability p(S) ∝ ∏_{i∈S} w_i, the way to make this identity hold is P(θ) = Σ_{i∈S(θ)} η_i ξ_i^* / q_i, with q_i = P(i ∈ S) the marginal inclusion probability; then E[P] = Σ_i η_i ξ_i^* = I_r by the dual-basis completeness of Eq. (B10). Nowhere does the paper state this. Eq. (3) is asserted with a citation to Ferris, Appendix D computes q_i and gives the dynamic-programming sampler, and footnote [70] admits scale factors but never defines them. As printed, Eq. (3) even has a normalization inconsistency: if p(θ) is a probability, the average should be Σ p(θ) W_R W_L^*, not (1/n_c) Σ p(θ) W_R W_L^*. Without the explicit P(θ), the unbiased estimator is not defined; a reader implementing from the text cannot tell whether the scale factors are 1/q_i, 1/w_i, or something else. If they are not 1/q_i, the average of each inserted projector is a non-identity matrix, low-rank truncation bias is not removed, and the agreement in Fig. 2 is not evidence for the method. Since the variance-reduction and sign-problem results are measured with respect to this same estimator, they inherit the gap. This is the load-bearing weak point of the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Markov chain Monte Carlo method over stochastic projectors inserted in tensor network contractions. The key idea is to replace deterministic low-rank projectors in the tensor renormalization group by randomly sampled projectors whose average is the identity, so that the unbiasedness of the partition-function estimator removes the systematic error of finite bond-dimension truncation. The method is demonstrated on the two-dimensional Ising model at zero field, where d=6 results for specific heat and magnetization squared agree with exact transfer-matrix data, and on the Yang-Lee zero field z=-1, where the average sign is claimed to improve systematically with increasing bond dimension cutoff d. The paper also reports an exponential decrease of the asymptotic variance with d and states that this yields an exponential speedup at O(d^5 N log N) cost per sweep.","tokens_in":16194,"tokens_out":5837,"duration_ms":55024,"significance":"If the central unbiasedness claim can be made rigorous, the idea is valuable because it offers a systematic way to eliminate the dominant systematic error of tensor renormalization methods while retaining their efficiency. The paper ships a useful algorithmic ingredient in Appendix D, where an overflow-free dynamic programming sampler for the marginal inclusion probabilities is developed, and Appendix A gives a clear demonstration of the exponential variance problem of naive importance sampling. The numerical benchmarks against exact transfer-matrix data are appropriate and the code-independent description of the computational graph is a strength. However, the manuscript's central identity, Eq. (3), is not fully specified, and the quantitative claims of exponential variance reduction and sign-problem resolution rest on limited numerical evidence. These issues are significant because the paper's headline claims depend directly on them.","major_comments":[{"comment":"The paper's central unbiasedness condition, Eq. (3), is not actually defined. The average of a function of a random variable θ with probability p(θ) is Σ_θ p(θ) f(θ), not (1/n_c) Σ_θ p(θ) f(θ), so the normalization in Eq. (3) is internally inconsistent. More importantly, the stochastic projector P(θ) = W_R(θ) W_L(θ)^* is never explicitly written: footnote [70] admits that scale factors are introduced, but the text and Appendix D do not state how the sampled set S = {k_1,...,k_d} is turned into an operator. A reader implementing the algorithm cannot tell whether the correct choice is P(S) = Σ_{i∈S} w_i η_i ξ_i^*, P(S) = Σ_{i∈S} (1/w_i) η_i ξ_i^*, or P(S) = Σ_{i∈S} η_i ξ_i^*/q_i, where q_i is the marginal inclusion probability of Eq. (D3). Only the last choice, with q_i defined by Eq. (D3), together with the completeness of the augmented dual basis in Eq. (B13), yields E[P] = I_r and hence unbiasedness. Because the scale factors are load-bearing for the elimination of systematic error, they must be specified explicitly and the identity verified.","section":"Eq. (3) and Appendix D"},{"comment":"The paper does not define the estimators used for physical quantities. It states that impurity tensors are used and that projectors 'become the identity operators after taking the random average and do not depend on external variables,' but it never writes the MCMC estimator for, e.g., the specific heat. In particular, for the negative-weight case of Fig. 3 the relevant estimator is a ratio of expectations with a reweighting factor, and the unbiasedness of such ratio estimators is not automatic at finite MCMC sample size. Without the explicit estimator formulas and the justification that they are unbiased (or at least consistent with controlled bias), the comparison with exact transfer-matrix results in Fig. 2 does not by itself validate the observable estimates.","section":"Physical quantities / impurity tensor method"},{"comment":"The headline claim of exponential variance reduction is supported by data of limited scope. The asymptotic variance is reported at a single system size (16×16), over a modest range of d (4 to 14), and without error bars on the variance estimates; it is also not reported how many independent MCMC runs or how many total samples were used for each point. Since the inset is the only quantitative evidence for 'exponential acceleration,' the paper should provide error bars on the variance, confirm the trend over a wider range of d and N, and state the computational cost actually incurred. As written, the exponential claim is not established.","section":"Inset of Fig. 2"},{"comment":"The sign-problem demonstration is incomplete. The figure shows only the average sign as a function of temperature for N = 32×32 and d = 2,3,4,6, without statistical errors or a comparison at different system sizes. The conclusion that the sign problem is 'prevented with polynomial computational time' requires showing that the total computational cost needed to reach a given statistical accuracy grows only polynomially with N and d; the average sign alone is not sufficient, because the cost also depends on the autocorrelation time of the Markov chain and on the variance of the estimator. A concrete scaling analysis is needed before the abstract's claim about the sign problem can be accepted.","section":"Fig. 3 and sign problem"}],"minor_comments":[{"comment":"The heading 'T ensor Renormalization Group' contains a typo, and the caption of Fig. 6 spells the authors of Ref. [39] as 'Leven-Nave' rather than 'Levin-Nave'.","section":"Appendix C and Fig. 6"},{"comment":"The number of Monte Carlo steps is printed as '2 14' and '2 11'; these should be formatted as 2^14 and 2^11.","section":"Main text near Fig. 2"},{"comment":"Equation (D14) introduces square roots in the denominator without derivation; since this is the first place the overflow-free marginal probability is defined, a short derivation or a reference to the identity that justifies this expression would help.","section":"Eq. (D14)"},{"comment":"The legend for the horizontal lines is ambiguous ('those by the standard MH method (horizontal lines)'); the lines should be labeled per quantity.","section":"Inset of Fig. 2"},{"comment":"Reference [70] appears only as a footnote and is not listed in the reference list; it should either be moved into the numbered reference list or converted to a proper footnote.","section":"Footnote [70] and references"},{"comment":"The paper notes that the optimal value of ω should be studied in the future, but for reproducibility it would be good to report the sensitivity of the results to ω, even if only in an appendix.","section":"Appendix D and hyperparameter ω"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is written by an established researcher in the field, and the idea is interesting. However, the central missing piece is the explicit construction of the stochastic projectors; the current text simply cites Ferris. The editor may want to request the definition of scale factors and a short proof of E[P] = I_r as a condition for acceptance. The numerical evidence for the exponential variance and sign-problem claims also needs to be strengthened, as outlined in the major comments. The paper fits the journal's scope but in its present form it is not fully reproducible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes replacing the low-rank projectors in TRG with stochastic ones sampled by MCMC, and claims this removes the finite-d bond-dimension bias while cutting variance exponentially with d. The core idea is genuinely worth a look, but the manuscript as written doesn't pin down the estimator well enough to support those claims. The unbiasedness rests on Eq. (3), which cites Ferris and then introduces a scale factor in a footnote—without saying what that scale factor is. A reader cannot tell whether the inserted tensor is sum_{i in S} eta_i xi_i^*/q_i, sum_{i in S} eta_i xi_i^*, or something else. The 1/n_c in Eq. (3) is also off if p(theta) is a probability. These are not cosmetic: if the scale factors are wrong, the average of each inserted projector is not the identity, and the 'elimination of systematic error' simply doesn't happen.\n\nWhat is new here is the specific construction that makes MCMC viable: freezing the projector set and weights from an initial deterministic TRG, so the proposal is independent across projectors, and using a computational graph to keep a sweep at O(d^5 N log N). The impurity tensor adaptation is a sensible way to get observables. That is a real step beyond Ferris's independent importance sampling, which the paper correctly identifies as variance-exploding.\n\nThe numerical evidence is promising but thin. The 16x16 Ising results at d=6 match the exact transfer-matrix values, and the variance appears to drop quickly with d. But the inset has no error bars on the variance estimates, and 'exponential' is inferred from a few points. The sign-problem demonstration is limited to N=32x32 and d=2,3,4,6; that is not enough to establish that the sign problem is 'prevented' at larger sizes. The paper's own Appendix D gives the marginal probabilities q_k, which are exactly what you need for the 1/q_i scale factor, but the connection is never made.\n\nAll of these issues are addressable. The right next step is a serious referee round asking for an explicit definition of the stochastic projector, a corrected Eq. (3), and more systematic tests, including scaling of the variance and sign with system size. I'd take it to review rather than desk-reject; the idea is strong enough to warrant the effort. If the estimator is fixed and the evidence holds up, this could be a useful method.","headline":"Promising idea with a serious gap: the stochastic projector estimator is never fully defined, and the sign-problem evidence is too thin to support the headline claims.","tokens_in":16678,"tokens_out":6312,"would_cite":true,"duration_ms":60394,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["82B80","65C05"],"pacs":["05.10.Ln","75.10.Hk"],"model":"deepseek-v4-flash","headline":"Replacing deterministic low-rank projectors with stochastic ones sampled by Markov chain Monte Carlo makes tensor-network partition-function estimates unbiased and makes statistical error fall exponentially with bond dimension.","keywords":["Markov chain Monte Carlo","tensor network","tensor renormalization group","stochastic projector","unbiased estimator","bond dimension","sign problem","Ising model"],"falsifier":"Compute, for a small random tensor network, the exact contraction and the MCMC contraction estimator at several $d$ values; if the ensemble average over many chains differs from the exact value by more than the estimated statistical error, Eq. (3) is violated. The same check can be done directly by Monte Carlo averaging $W_R W_L^*$ over the proposed projector sets and comparing to $I_r$.","tokens_in":15638,"feed_emoji":"🎲","tokens_out":6152,"duration_ms":56176,"temperature":0.7,"pith_summary":"Tensor network methods approximate the partition function by repeatedly truncating bond dimension with low-rank projectors, and that truncation is a source of systematic error. This paper argues the truncation can be converted into a sampling problem: choose stochastic projectors whose average is the identity, and run a Markov chain over projector configurations. The resulting Monte Carlo estimator is claimed to be unbiased, so the bond dimension cutoff no longer biases results. In two-dimensional Ising simulations, the statistical error decreases exponentially as the bond dimension cutoff grows, and the negative sign problem at an imaginary magnetic field is substantially mitigated. If the claim holds, it turns the bond dimension from a bias parameter into a precision knob for statistical inference.","feed_headline":"Stochastic projectors make tensor networks both fast and unbiased","feed_subtitle":"Sampling projectors whose average is the identity removes truncation bias and shrinks statistical error exponentially.","key_machinery":"The load-bearing object is the stochastic projector average identity, Eq. (3), where a projector $P=W_R W_L^*$, a low-rank truncation operator inserted at a bond, is replaced by a random rank-$d$ projector drawn from a distribution $p(\\theta)$ such that $\\langle W_R W_L^*\\rangle_\\theta=I_r$. A set of rank-1 projectors is weighted by powers of the singular values, Eq. (D1), and a dynamic-programming scheme, Eqs. (D6)–(D14), draws subsets of rank-1 projectors without overflow or underflow. A computational graph stores the contraction tree so that an MCMC sweep re-evaluates only affected ancestors, and impurity tensors provide physical quantities without projector derivatives. Together these pieces turn the systematic truncation error into a statistical sampling problem.","core_discovery":"The paper's central claim is that inserting projectors $P=W_R W_L^*$ that are sampled rather than chosen optimally removes the systematic error of low-rank tensor contraction while keeping the method usable. The random projectors are tuned so their ensemble average is the identity, $\\langle W_R W_L^*\\rangle_\\theta=I_r$, meaning that any rank-$d$ truncation is 'undone' on average and the partition function estimator $\\sum_{\\{\\theta_i\\}} g(\\theta_1,\\dots,\\theta_{N_p})p(\\theta_1)\\cdots p(\\theta_{N_p})$ is unbiased. Markov chain Monte Carlo samples the projector configurations, and a computational graph restricts each update to the ancestors of the changed projector, reducing a sweep to $O(d^5 N\\log N)$ work. On the $N=16\\times16$ Ising model the paper reports specific heat and magnetization squared consistent with exact transfer-matrix results at $d=6$, and an asymptotic variance that falls exponentially with $d$. At the Yang–Lee edge field with negative fugacity, the average sign rises toward unity as $d$ grows, which the paper reads as mitigation of the sign problem at polynomial cost.","pith_inferences":["A direct numerical test of Eq. (3) on random matrices at moderate $r,d$ would isolate whether the Appendix D sampling weights realize the identity average exactly; this is a cheap falsification check.","The hyperparameter $\\omega$ controlling the weight of singular values is left free; tuning it per system could further reduce the asymptotic variance beyond the exponential-in-$d$ scaling shown.","The same sampler could estimate free-energy differences or response functions at nearby couplings by reweighting the stored projector configurations, which the paper does not discuss.","For fermionic or real-time networks with complex weights, the sign improvement is likely representation-dependent, so the practical range of system sizes for which the average sign stays large is a quantitative question for future work."],"forward_implications":["Bond dimension cutoff becomes a statistical convergence parameter: increasing $d$ reduces Monte Carlo variance exponentially instead of removing a deterministic bias.","Physical observables obtained from impurity tensors inherit the unbiasedness, so specific heat, magnetization, and similar quantities can be matched to exact results at modest $d$.","Systems with negative or complex weights, where the sign problem is severe, can be simulated with average sign improving systematically as $d$ grows, at polynomial cost.","The projector formulation of TRG carries over to other tensor network algorithms, so the same MCMC treatment can be built on TEBD, ATRG, BTRG, CATN, or HOTRG.","Because the projectors average to the identity, the impurity-tensor method no longer needs projector derivatives, removing an extra source of systematic error present in ordinary tensor renormalization."],"supporting_citations":[{"why":"supplies the stochastic-projector sampling scheme and the identity condition that makes the estimator unbiased","marker":"[69]"},{"why":"defines the Levin–Nave tensor renormalization group that the paper recasts in projector form","marker":"[39]"},{"why":"give the general MCMC background and variance formula used to define asymptotic variance","marker":"[1, 2]"},{"why":"provide the Metropolis–Hastings update used to sample projector configurations","marker":"[73, 74]"},{"why":"present sequential Monte Carlo with resampling as the alternative variance-control method motivating the MCMC choice","marker":"[71, 72]"},{"why":"supplies the dynamic programming scheme used to sample subsets of rank-1 projectors","marker":"[84]"},{"why":"supports the impurity tensor method for computing physical quantities within the estimator","marker":"[75–77]"},{"why":"fixes the critical temperature at which the exponential variance reduction is demonstrated","marker":"[78]"},{"why":"defines the imaginary-field/negative-fugacity test case used to demonstrate sign-problem mitigation","marker":"[79]"}],"fun_headline_variants":["Unbiased tensor MCMC via sampled projectors","Stochastic projectors kill truncation bias in MCMC","Exponential error reduction in tensor MCMC","Sampled projectors beat sign problem in tensor MCMC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument rests on the random projectors averaging to exactly the identity; if the sampling probabilities and scale factors are not perfectly matched, the systematic error from low-rank truncation is not fully removed.","fun_headline_variants_meta":{"raw":{"variants":["Unbiased tensor MCMC via sampled projectors","Stochastic projectors kill truncation bias in MCMC","Exponential error reduction in tensor MCMC","Sampled projectors beat sign problem in tensor MCMC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1323,"prompt_tokens":946,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":562,"tokens_out":377,"duration_ms":3732,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:54:21.056137+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for a small random tensor network, the exact contraction and the MCMC contraction estimator at several $d$ values; if the ensemble average over many chains differs from the exact value by more than the estimated statistical error, Eq. (3) is violated. The same check can be done directly by Monte Carlo averaging $W_R W_L^*$ over the proposed projector sets and comparing to $I_r$.","supporting_citations":[{"cited_title":"Levin and C","cited_arxiv_id":null,"evidence_quote":"defines the Levin–Nave tensor renormalization group that the paper recasts in projector form"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the dynamic programming scheme used to sample subsets of rank-1 projectors"}],"review_version":1}