{"id":"a73120fb-490c-4f44-b47d-1d996211807c","arxiv_id":"2602.02245","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Sampling from 2D isometric tensor network states is made efficient by a row-wise QR sweep and MPO-MPS contraction, with single-sample and top-K variants.","lead":"The paper introduces two algorithms for sampling from two-dimensional isometric tensor network states: one that draws a single independent sample, and one that greedily finds K high-probability bitstrings. Efficient sampling from 2D quantum states is the bottleneck for quantum Monte Carlo, tensor-network calculations, and quantum-advantage verification.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing renormalization after approximate row contraction leaves sequential conditionals subnormalized, so stored p_LL is not a true probability and the sampling distribution is not well-defined under truncation.","rationale":"The reader's weakest assumption pinpoints the same load-bearing issue: the approximate MPO-MPS contraction in step 5 of Algorithms 4/6 is not accompanied by a renormalization step, so the sequential conditional probabilities used for sampling can be subnormalized. I focused on the precise consequence: the trace used in Algorithm 3 produces a valid probability mass function only when the state being sampled is normalized; after a truncated contraction, the row representing the conditional state is not guaranteed normalized. This breaks the claim that the algorithm draws samples from the isoTNS distribution (or even from the approximate T_χ) and means the returned p_LL is not the sample's true probability. The paper's error analysis in §3.2 bounds the global approximation error but does not establish that the sequential conditionals are mutually consistent or normalized. The proposed test directly checks normalization of the returned probabilities and isolates the effect of renormalization, which is a concrete and minimal fix. The concern does not invalidate the exact no-truncation regime or the overall algorithmic framework, but it does require a correction or explicit renormalization before the truncated-regime sampling claim is fully supported, so the existing CONDITIONAL verdict remains appropriate.","tokens_in":18061,"tokens_out":8380,"duration_ms":80914,"concrete_test":"Using the released code, run Algorithm 4 on a 3×3 random isoTNS with χ=4 (so truncation is active). Enumerate all 2^9 configurations; for each, record the returned p_LL and compute S = Σ p_LL. (i) If S ≠ 1, the stored probabilities are not a normalized distribution. (ii) Implement a variant that, after each MPO-MPS contraction in step 5, divides the approximate row by its computed norm (or rescales the center B), and repeat the enumeration. If the variant yields S = 1 and a KL divergence to the exact distribution at the truncation floor while the original has S < 1 and a larger KL, the missing renormalization is the cause. Also check the first-site conditional of row 2: Σ_σ Tr[B^σ(B^σ)*] should equal 1 for a valid sampler.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Algorithm 4, after sampling row i, the next row R_{i+1} is produced by approximate MPO-MPS contraction (zip-up) with maximum bond dimension χ (step 5, eq. (32)). The sampled row R_i has norm squared equal to the probability of the sampled row configuration, not 1; the exact contraction of R_i with the next row inherits that norm. Zip-up truncation does not preserve this norm, and the paper never states that R_{i+1} is renormalized. Algorithm 3 then computes each conditional p(σ_ij|...) = Tr[B^σ_ij (B^σ_ij)*] from the orthogonality center B_ij. This trace only defines a normalized distribution if the row (the state of the remaining rows conditioned on prior samples) has unit norm. If ∥R_{i+1}∥ ≠ 1, the computed conditionals are subnormalized: their sum over σ_ij equals ∥R_{i+1}∥², not 1. Consequently the product p_LL of sequentially drawn p_ij is not the probability of the sample under either the exact state or the approximate state T_χ, and 's_ij ← independent sample from p(...)' is not a sample from a probability distribution unless an implicit normalization occurs that is nowhere specified. The global Frobenius bound ∥T_χ−T∥_F ≤ ε in §3.2 does not resolve this: even if T_χ is close to T, the sequential procedure can sample from a different, subnormalized distribution. The same flaw affects Algorithm 6's top-K probabilities.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces two sampling algorithms for two-dimensional isometric tensor network states (isoTNS). Algorithm 4 draws an independent configuration by conditioning row-by-row, using QR moves within each row and an approximate MPO-MPS (zip-up) contraction between rows; Algorithm 6 is the corresponding top-K greedy variant. The authors claim polynomial costs O(L^2 d^2 chi^6) and O(L^2 K^2 d^2 chi^6), exactness when row contractions are performed without truncation, and validate on GHZ, W, and random states.","tokens_in":18486,"tokens_out":12050,"duration_ms":118479,"significance":"If the algorithms are correct, they provide the first efficient classical sampling and top-K search for 2D isoTNS, a class relevant to simulation of 2D quantum systems. Strengths: the zero-truncation limit is exact; numerical benchmarks compare against analytic GHZ/W distributions and state-vector calculations; code is publicly available; no constants are fitted. The main risk is in the truncated regime: the paper's approximation control is heuristic and, as detailed below, the algorithm as written does not state a normalization step that is required for the returned probabilities to be probabilities.","major_comments":[{"comment":"After the approximate MPO-MPS contraction in step 5 of Algorithm 4, the row R_{i+1} is not renormalized. Algorithm 3 then uses p(σ_ij|...) = Tr[B^{σ_ij}(B^{σ_ij})*] as a conditional distribution. This is only a normalized distribution if the row has unit norm; otherwise the probabilities sum to ||R_{i+1}||^2, not 1. Thus the stored p_LL is not the Born probability of the sampled configuration under either T or T_χ. The Frobenius bound in §3.2 does not repair this: even if T_χ is close to T, sequential sampling from subnormalized conditionals is biased. The authors should explicitly normalize R_{i+1} (or otherwise define the conditioning step) and state how the returned probability is computed.","section":"§3.1, Algorithm 3 step 2; Algorithm 4 step 5"},{"comment":"The error analysis treats the algorithm as sampling from a fixed approximate tensor T_χ with ||T_χ - T||_F ≤ ε. But step 5 of Algorithms 4 and 6 compresses the contraction of the already sampled row with the next row; the compressed row depends on the sample history. The procedure therefore does not, in general, correspond to sampling from a single global T_χ, and a global Frobenius bound on T_χ does not control the distance between the algorithm's output distribution and |T|^2. Either a definition of T_χ compatible with the sequential procedure (e.g., fixed compression applied before conditioning) or a separate analysis of the history-dependent approximation is needed for the claimed 'associated probability' to be meaningful.","section":"§3.2"}],"minor_comments":[{"comment":"The pseudocode applies the QR/contract step at j=L even though Q_{i(L+1)} does not exist. The surrounding text suggests the final site should be left as S_{iL}; the loop condition should be j<L inside the i<L block.","section":"Algorithm 3 / Algorithm 5"},{"comment":"Caption says '10000 samples' but the text in §4.2 says 10^5 samples for the χ=8 empirical distribution. Please make these consistent.","section":"Fig. 2"},{"comment":"The text refers to 'the blue line' for the median KL divergence; specifying the color in the caption or text would improve readability.","section":"§4.2"},{"comment":"'generalize existing 1D tensor networks sampling algorithms' appears twice with slightly different wording; one occurrence should be edited.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The normalization issue is genuine and fixable in a revision; the authors should add an explicit normalization/renormalization step and clarify what probability is returned. The second major comment about the history-dependent approximation is more conceptual and may require new numerical or theoretical work; if the authors can convincingly frame the output as an approximate importance weight, that would also be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the row-wise QR sweep that moves the orthogonality center across a 2D isoTNS row, plus the top-K extension that propagates a K-index through rows. That combination is not in the MPS/TT or PEPS sampling literature, and the cost bound O(L^2 d^2 chi^6) is real and better than generic PEPS sampling. The exact zero-truncation regime is properly validated: GHZ, W, and random 3x3 states match analytic or state-vector distributions, and the code is public. No constants are fitted to the target distribution, and the self-citations are used only for input-state preparation. This deserves credit.\n\nThe soft spot is the one the stress-test flags, and I think it is real. In Algorithm 4, after sampling a row, the next row is produced by approximate MPO-MPS contraction (zip-up) with no stated renormalization. Algorithm 3 then computes conditionals as Tr[B^sigma B^sigma*], which is only a probability distribution if the row has unit norm. Zip-up truncation does not preserve that norm, so under truncation the conditionals are subnormalized and the stored p_LL is not the probability of the sample—neither under the exact state nor under the truncated T_chi. The global Frobenius bound in Section 3.2 does not resolve this; it bounds closeness of states, not normalization of sequential conditionals. The same issue carries into Algorithm 6. This is not a fatal flaw in the exact regime, but it is a load-bearing gap for the paper's general claim.\n\nA few smaller things: the error analysis is heuristic, as the authors admit, and the truncated-regime experiments (Figure 2, Figure 4) do not separate bias from lack of normalization versus ordinary truncation error. Also, the text says “sampling one isoTNS column” where it means row, a minor typo. The core construction is sound and the fix looks straightforward: renormalize each contracted row, or explicitly sample from the normalized conditional distribution, and report repeated-trial error statistics in the truncated regime.\n\nWho gets value from this: anyone working on isoTNS algorithms, METTS, uncertainty quantification, or classical verification of near-term quantum experiments. It is a solid extension rather than a conceptual leap, but it is the first sampling algorithm for this tensor-network class and should go to peer review. A serious referee should ask for the normalization fix and for truncated-regime error statistics, not for a rewrite.","headline":"First practical sampling algorithms for 2D isoTNS, with clean row-wise QR sweeps and public code; the main gap is an under-specified normalization step in the truncated regime that makes the stored probabilities not true probabilities.","tokens_in":18917,"tokens_out":4682,"would_cite":true,"duration_ms":51675,"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":"This paper introduces the first practical sampling algorithms for two-dimensional isometric tensor network states, extending one-dimensional matrix-product-state sampling to 2D lattices.","keywords":["tensor networks","isometric tensor network states","sampling","Born distribution","matrix product states","greedy search","MPO-MPS contraction","quantum simulation"],"falsifier":"On a small random state (e.g., 4×4), compute the exact Born distribution by direct state-vector simulation, then run the independent sampler with a truncated bond dimension. If the empirical distribution does not converge to the exact distribution as the number of samples grows—or if the average of the returned probabilities p_LL differs systematically from 1—the truncation step is biasing the sampler. This directly tests whether the heuristic Frobenius bound translates into an accurate sampling distribution.","tokens_in":17979,"feed_emoji":"🎲","tokens_out":6770,"duration_ms":59200,"temperature":0.7,"pith_summary":"This paper introduces the first practical sampling algorithms for two-dimensional isometric tensor network states (isoTNS), extending the well-established one-dimensional matrix-product-state sampling to 2D lattices. The first algorithm draws an independent spin configuration from the Born distribution together with its exact probability; the second greedily finds K high-probability configurations. Both run in time polynomial in system size and bond dimension, with the dominant cost O(L^2 d^2 χ^6) per sample or O(L^2 K^2 d^2 χ^6) for the top-K variant. When the intermediate row contractions are performed without truncation, sampling is exact; when truncation is used, the paper shows that errors propagate linearly from the wavefunction into the sampled probabilities. This matters because sampling is a core subroutine in quantum Monte Carlo methods, distribution estimation, and verification of quantum devices.","feed_headline":"Two new algorithms sample 2D isoTNS in polynomial time","feed_subtitle":"Independent samples and top-K searches now run in polynomial time, opening 2D tensor networks to Monte Carlo use.","key_machinery":"The central object is the isometric tensor network state (isoTNS) in canonical form: a 2D tensor network in which every tensor except the orthogonality center is an isometry, so that norms and marginals become trivial to compute. The algorithm's workhorse is a row-sampling routine that reads the marginal from the center tensor, projects out the sampled index, and uses a QR decomposition to move the center one step to the right along the row, keeping the row's bond dimension under control. Between rows, an MPO-MPS contraction reforms the next row with a capped bond dimension; this is the only step that can introduce error. This machinery turns the global Born distribution into a product of lo","core_discovery":"The central claim is that a 2D isoTNS can be sampled by sweeping rows from top to bottom, conditioning on the indices already drawn. Because all tensors except the orthogonality center are isometries, each marginal distribution at the current site is just the diagonal of the center tensor contracted with itself; a sampled index is projected out, and a QR decomposition shifts the orthogonality center to the next site in the row. After a row is fully sampled, the row is contracted into the next row using an approximate MPO-MPS multiplication (zip-up) with bond dimension capped at χ, which is the only approximation in the whole procedure. The paper proves the cost bounds and demonstrates numeri","pith_inferences":["Editorial: Because the independent sampler is exact when contractions are exact, it could serve as a drop-in sampling subroutine for thermal and Monte Carlo tensor-network algorithms on 2D lattices, where no comparable exact sampler existed.","Editorial: The top-K search's greedy nature means it can miss probability mass concentrated in multiple well-separated regions; a beam-search variant that keeps diverse partial strings might recover more of the distribution at modest extra cost.","Editorial: A simple internal-consistency check for the truncated algorithm is to sum the stored probabilities p_LL over many independent samples; a systematic shortfall would directly reveal the subnormalization introduced by truncation."],"forward_implications":["Independent samples from a 2D isoTNS can be drawn in O(L^2 d^2 χ^6) time, polynomial in the system size—previously no such algorithm existed.","The greedy top-K variant returns K high-probability bitstrings in O(L^2 K^2 d^2 χ^6) time, enabling identification of the most likely measurement outcomes.","If row contractions are performed exactly, both algorithms are perfect samplers, meaning the empirical distribution converges to the true Born distribution at the usual Monte Carlo rate.","When truncation is used, the error in the sampled probability distribution is first-order in the truncation error, so small truncation errors do not blow up.","The algorithms apply to any isoTNS, including states prepared by unitary circuits, and generalize directly to three-dimensional lattices."],"fun_headline_variants":["Two algorithms for sampling 2D tensor networks now run in polynomial time","Polynomial-time sampling for 2D isometric tensor networks","Independent and top-K sampling for 2D isoTNS now polynomial","2D tensor network sampling: two algorithms, polynomial time"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The algorithm's correctness depends on the assumption that after truncating each row contraction, the resulting row tensor still gives a faithful, properly normalized conditional distribution for the remaining sites; the paper bounds this only heuristically in the Frobenius norm and does not include a renormalization step, so in the truncated regime the stored probabilities can be subnormalized.","fun_headline_variants_meta":{"raw":{"variants":["Two algorithms for sampling 2D tensor networks now run in polynomial time","Polynomial-time sampling for 2D isometric tensor networks","Independent and top-K sampling for 2D isoTNS now polynomial","2D tensor network sampling: two algorithms, polynomial time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000689,"raw_usage":{"total_tokens":2919,"prompt_tokens":666,"completion_tokens":2253,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":410,"completion_tokens_details":{"reasoning_tokens":2179}},"tokens_in":410,"tokens_out":2253,"duration_ms":14484,"temperature":1.0,"reasoning_tokens":2179,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:23:17.994903+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small random state (e.g., 4×4), compute the exact Born distribution by direct state-vector simulation, then run the independent sampler with a truncated bond dimension. If the empirical distribution does not converge to the exact distribution as the number of samples grows—or if the average of the returned probabilities p_LL differs systematically from 1—the truncation step is biasing the sampler. This directly tests whether the heuristic Frobenius bound translates into an accurate sampling distribution.","supporting_citations":[],"review_version":1}