{"id":"114e758b-4c2c-4747-8e31-a7a175c29fb3","arxiv_id":"2411.16431","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A tensor-network branch-and-bound solver is slower and slightly less accurate than Ising machines on large random Pegasus and Zephyr spin glasses, but beats them on planted-instance energy.","lead":"This paper tests a tensor-network based solver on the same Ising spin-glass problems that run on D-Wave quantum annealers, comparing it against the annealer and a classical solver called Simulated Bifurcation. For large random problems the tensor-network solver is two orders of magnitude slower and finds slightly worse solutions, but on planted problems it finds better energies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Contraction-failure attribution is tied to a known randomized-SVD instability; the claimed generality of TN limitations remains conditional.","rationale":"The reader's weakest_assumption points directly at the boundary MPS contraction accuracy, and Section VI confirms that the randomized-SVD instability undermines this premise at the beta values and bond dimensions used for the large benchmarks. This is the most load-bearing part of the central claim because the entire 'limitations' conclusion rests on the contraction quality. I agree with the reader's identification. However, I do not think this requires changing the verdict: the paper is transparent about the instability, provides reproducible code and data, and its empirical comparison is a valid snapshot of a specific state-of-the-art implementation. The concern is that the title and abstract generalize from this implementation to 'tensor network approaches' broadly, which is not fully warranted given that a fixable numerical artifact is the identified failure mechanism. The concrete test would settle whether the limitation is intrinsic or implementation-specific. Since the paper already acknowledges this caveat and the empirical findings are not in question, the ACCEPT verdict remains appropriate, possibly with a tightened wording in the abstract/title.","tokens_in":27216,"tokens_out":8936,"duration_ms":89409,"concrete_test":"Re-run the Pegasus 5376-spin class I benchmark (20 instances, beta=0.5, M=1024) with boundary MPS bond dimension chi=16 and a stabilized randomized SVD (e.g., adding 2-3 power iterations or using a dense SVD on the projected subspace), on an 80GB GPU. Compare median dE and time-to-solution against the published chi=8 results and the Ising-machine baselines. If the median dE drops below 0.1% or reaches parity, the reported limitation is an artifact of the contraction subroutine; if the gap persists, the conclusion is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's strongest_claim states that the TN solver is 0.1%-1% worse than Ising machines on large Pegasus instances and attributes failure to approximate contraction. The load-bearing premise is that boundary MPS with bond dimension chi~8-10, built via randomized SVD and 1-site variational optimization, provides conditionals accurate enough for branch-and-bound. Section VI demonstrates exactly where this premise fails: at beta>=1 and D>=8, randomized SVD gives different outcomes across repeated runs, indicating local optima that 1-site variational updates cannot escape. This is a numerical-instability artifact of the randomized SVD, not a fundamental limitation of tensor networks. The paper itself notes that 'any improvement in the randomized SVD performance should translate to improved quality of the results.' Therefore, the demonstration that 'tensor network approaches' are limited is conditional on a specific, admittedly unstable contraction subroutine and on memory-limited chi. A more stable boundary-MPS update (e.g., randomized SVD with power iterations or block Lanczos) and larger chi could reduce or eliminate the reported gap. The central empirical snapshot remains valid, but the broader title-level conclusion about limitations of TN approaches is not established as a general property beyond this implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a tensor-network (TN) based branch-and-bound algorithm for the low-energy spectrum of Ising spin glasses on the Pegasus and Zephyr graphs of current D-Wave processors. The method groups spins into large unit cells (24 or 16 spins), builds an exact PEPS representation of the Boltzmann distribution, and uses boundary-MPS approximate contraction with a zipper scheme and randomized SVD to compute conditional probabilities; an optional loopy belief propagation stage reduces the local dimension. The solver is benchmarked against a D-Wave quantum annealer and the Simulated Bifurcation Machine on four instance classes, with time-to-solution curves for energy error and for a diversity metric based on maximum cliques in Hamming-distance graphs. For large i.i.d. instances (5376 spins on Pegasus), the TN solver reaches energies 0.1%-1% above the best Ising-machine solutions while taking two orders of magnitude longer, which the authors attribute to approximate contraction failures; for tile-planted instances the TN solver is about a factor of 3 more accurate in energy. The paper includes a stability analysis (Sec. VI) identifying randomized-SVD instability as a key failure mode, and releases the Julia code and data.","tokens_in":27476,"tokens_out":7665,"duration_ms":70481,"significance":"If the conclusions are accepted, the paper provides a valuable, carefully controlled benchmark of a nontrivial TN heuristic on realistic quantum-annealing geometries, and it identifies a concrete numerical bottleneck (randomized-SVD boundary MPS) that limits this class of algorithms. Strengths include: medians over 20 instances with bootstrapped error bars, sensitivity analysis over beta and bond dimension, a clear discussion of the diversity metric and its limitations, and public release of the implementation and data. The paper also demonstrates a useful positive result on planted instances. Nevertheless, the scope of the 'limitations' claim is narrower than the title suggests, because the evidence is tied to a specific contraction subroutine and to a memory-limited bond dimension.","major_comments":[{"comment":"The title and abstract generalize the demonstrated limitations to 'tensor network approaches', but the evidence in Sec. VI and Sec. VII supports a limitation of the present branch-and-bound TN implementation: the authors show that randomized truncated SVD gets stuck in local optima for D>=8 at large beta, and they explicitly state that 'any improvement in the randomized SVD performance should translate to improved quality of the results.' The benchmark does not compare against other TN optimization strategies such as tropical tensor networks (Ref. [36]) or hyperoptimized approximate contraction (Ref. [35]), so the title/abstract overclaim. I recommend rewording to 'a tensor-network approach' or 'current boundary-MPS contraction techniques' and qualifying the attribution in the abstract, e.g., 'We provide evidence that these gaps are caused by instabilities in the approximate contraction subroutine used here.'","section":"Title/Abstract and Sec. VI"},{"comment":"The time-to-solution comparison for the D-Wave annealer excludes programming and access overhead. The caption of Fig. 10 estimates QA time as annealing time plus 100 us, but Appendix E reports a typical programming time of roughly 14.2 ms for the Advantage system and a QPU delay of 21 us per sample. Since the paper's central quantitative claim is that the TN solver is 'two orders of magnitude slower' than the Ising machines, the QA timing curve should include programming overhead (standard in time-to-solution studies) or the authors should provide a sensitivity analysis showing that the conclusion is unchanged. Without this, the relative standing of QA in the timing plots is not directly comparable to SBM and TN.","section":"Sec. V.C, Fig. 10, and Appendix E"},{"comment":"The causal attribution of the performance gap to 'approximate contraction failures' is only partially separated from branch-and-bound failures. Figure 12(a) shows that for beta < 0.25 the branch-and-bound procedure fails independently of the contraction quality, and for larger beta the failure correlates with randomized-SVD instability. However, the paper does not perform a controlled experiment that isolates the contraction error from the pruning error of the branch-and-bound, e.g., by comparing against exact contraction on small lattices or by using a deterministic full-rank SVD where feasible (as is done only for square lattices in Appendix B). The statement in the abstract that the results are attributed to approximate contraction failures is therefore an interpretation rather than a demonstrated fact. I suggest either adding such a control or softening the attribution.","section":"Sec. VI and abstract"}],"minor_comments":[{"comment":"The algorithm is described as 'deterministic', but Sec. VI shows that randomized SVD makes repeated runs produce different outcomes; please qualify the determinism claim, e.g., 'deterministic apart from the randomized SVD subroutine'.","section":"Abstract and Sec. III"},{"comment":"The quantity described as 'max(min(Schmidt value))' is confusing; it appears to be the maximum over cuts of the smallest retained singular value, and should be stated explicitly in the caption.","section":"Fig. 12(b)"},{"comment":"The text says 'We further show the influence of beta on the quality of the results in the following section and in Appendix C'; Sec. VI focuses on stability and success rate rather than the time-to-solution quality, so please adjust the cross-reference.","section":"Sec. V.C"},{"comment":"The factor 2 in the denominator of dE is non-standard; please justify it or replace with the more conventional (E - E_best)/|E_best|.","section":"Eq. (14)"},{"comment":"The caption refers to 'class II, III and VI' but the fourth class is 'IV'; please correct the typo.","section":"Appendix A, Fig. A1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of a condensed-matter or interdisciplinary journal, and the code and data release are commendable. The main risk is that the broad title creates an expectation of a general impossibility result, which the paper does not and cannot establish. I would encourage the editor to request the title/abstract revision outlined in the major comments, but I do not see any issue that would warrant rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, honest benchmark paper. It extends the TN branch-and-bound method from Chimera graphs to Pegasus and Zephyr, solves the large-unit-cell contraction problem with sparse tensors and GPUs, and compares against D-Wave and SBM with medians over 20 instances, bootstrapped errors, and publicly available code and data. The headline result – for random i.i.d. problems above ~5000 spins, TN is 0.1–1% worse than the Ising machines and about 100x slower – is supported by the data. The planted-instance result, where TN gets a factor of 3 closer to the ground state, is a useful counterpoint and shows the method has a niche.\n\nThe paper's greatest strength is its failure analysis. Section VI diagnoses the breakdown: at large beta and bond dimension >=8, the randomized truncated SVD used in boundary MPS compression produces local optima that 1-site variational sweeps can't escape. The authors show repeated runs give different outcomes and make the link to the contraction subroutine explicit. That is the right way to report a negative result.\n\nThe soft spot is exactly the one implied by that diagnosis. The title says 'limitations of tensor network approaches,' but the evidence is about a particular implementation constraint: randomized SVD instability plus memory-limited chi around 10. The authors themselves note that improving the SVD would improve results, so the broad claim is stronger than the demonstrated result. A serious referee should ask for a qualified title and a sentence in the abstract saying the limitation is for this class of approximate contraction with feasible resources. This is a framing issue, not a fatal flaw.\n\nMinor issues: the timing comparison excludes D-Wave programming and access overhead, which the authors mention in discussion but not in the main figures; beta and chi are tuned per graph, so the comparison is not strictly head-to-head; and the diversity metric is computed from the union of all solvers' outputs, which is standard but slightly favors solvers that produce many samples. None of these change the main conclusions.\n\nVerdict: worth refereeing. The benchmark is careful, reproducible, and the negative result is useful to the quantum-inspired optimization community. The self-identified contraction instability is a real contribution to the tensor-network literature. I'd send it to review and ask the authors to scope the title.","headline":"A careful, reproducible benchmark showing TN branch-and-bound loses to Ising machines on large random Pegasus/Zephyr problems but wins on planted ones; the title overreaches slightly, but the evidence is solid.","tokens_in":27996,"tokens_out":2831,"would_cite":true,"duration_ms":26326,"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 shows that a deterministic tensor-network branch-and-bound solver, applied to the quasi-two-dimensional graphs of current quantum annealers, reaches energies 0.1% to 1% above the best solutions of a D-Wave quantum annealer and…","keywords":["tensor networks","Ising spin glasses","branch and bound","quantum annealing","simulated bifurcation machine","Pegasus graph","boundary matrix product states","randomized SVD"],"falsifier":"Run the 216-spin Pegasus class I benchmark with $\\beta = 3$ and bond dimension $\\chi = 8$ using the same randomized-SVD zipper: the paper predicts repeated runs with different random seeds will disagree about whether the ground state is found, with success rates well below 100%. If every run reliably found the ground state while the contraction spectrum still showed steeply decaying singular values, the attribution of failure to randomized-SVD instability would be refuted.","tokens_in":27063,"feed_emoji":"🧲","tokens_out":12265,"duration_ms":94794,"temperature":0.7,"pith_summary":"The paper tries to establish where deterministic tensor-network (TN) methods stand as solvers for Ising spin-glass optimization on the quasi-two-dimensional graphs of current quantum annealers. It builds a branch-and-bound algorithm that groups spins into Potts-style clusters and ranks partial configurations using conditional probabilities obtained from approximate contractions of a PEPS tensor network. Benchmarked against a D-Wave quantum annealer and the classical Simulated Bifurcation Machine (SBM) on Pegasus and Zephyr graphs, the TN solver finds the best energies on small instances but on i.i.d. problems with over 5,000 spins lands 0.1% to 1% above the best Ising-machine solutions while running about two orders of magnitude slower. The paper attributes this gap to failures of the approximate tensor-network contraction, specifically instabilities of the randomized SVD used to build boundary matrix product states. It also reports a regime where TN wins: on embedded tile-planting instances it reaches about 0.1% from the planted ground state, a factor of 3 closer than the Ising machines.","feed_headline":"Tensor networks trail Ising machines on big spin-glass problems","feed_subtitle":"On 5,000-spin instances, tensor-network energies sit 0.1–1% higher and runtimes are ~100x slower.","key_machinery":"The machinery is a generalized Potts representation of the Ising Hamiltonian on a square lattice of clusters (24 spins per cluster for Pegasus, 16 for Zephyr), whose Boltzmann distribution at inverse temperature $\\beta$ is encoded exactly as a PEPS tensor network with site tensors, bond tensors, and mediating tensors that carry diagonal interactions. A branch-and-bound sweep adds one cluster at a time, keeping the $M$ most probable partial configurations; the marginal probabilities $p(x_{n+1}|\\partial x_n)$ are obtained from an approximate contraction of the PEPS via boundary matrix product states with bond dimension $\\chi \\approx 8$ to $10$, built by a variational zipper that uses randomized truncated SVD followed by 1-site variational overlap optimization. The paper's central diagnostic is that this approximate contraction, not the branch-and-bound search, is the load-bearing part: at larger $\\beta$ and $\\chi \\geq 8$ the randomized SVD lands in local optima that the variational refinement cannot escape, producing different outcomes on repeated runs and collapsing ground-state success rates.","core_discovery":"On its own terms, the central claim is that a deterministic TN-based branch-and-bound search can serve as a competitive low-energy solver for spin glasses on Pegasus and Zephyr graphs, but its practical ceiling is set by approximate contraction quality rather than by the search strategy. For random i.i.d. instances with more than 5,000 spins, the TN solver's best energies are 0.1% to 1% above the best energies found by the quantum annealer or SBM, and the TN solver is roughly two orders of magnitude slower; the paper traces this gap to approximate boundary-MPS contractions built with randomized truncated SVD, which become unstable at the inverse temperatures needed for the branch-and-bound resolution. For embedded tile-planting instances, the same algorithm reaches about 0.1% above the planted ground state, outperforming both Ising machines by a factor of about 3 in energy. On sampling diversity, all three solvers find a few mutually distant low-energy states, but the deterministic TN approach returns at most a handful of independent solutions, while the two randomized Ising machines can output sets of thousands.","pith_inferences":["A direct test of the paper's diagnosis would be to swap the randomized SVD for a more robust truncated decomposition at the same bond dimension and check whether ground-state success on the 216-spin Pegasus instances is restored at $\\beta \\geq 3$; if it is not, the attribution of failure to randomized-SVD instability would need revision.","The observed correlation between graph connectivity and contraction hardness, with the square lattice easiest and Zephyr hardest, suggests that boundary-MPS accuracy degrades roughly as the number of diagonal and long-range couplings per unit cell increases, and a quantitative study of truncation error versus diagonal-coupling density would make this concrete.","Because the solver never reaches the discarded-probability bound that would certify a ground state, TN optimization on these geometries is heuristic in practice; making it certifying would require exact or nearly exact contractions, which are #P-hard, so the certification gap is likely structural rather than a matter of tuning.","The diversity results suggest a natural hybrid: use the TN solver to enumerate a small set of high-quality deterministic solutions and a randomized sampler to populate the rest of the low-energy manifold, combining TN's energy precision on structured instances with the samplers' coverage."],"forward_implications":["Improving the stability of boundary-MPS contraction, for example by replacing randomized SVD with a more robust truncated decomposition or by adding gauge transformations, should directly improve the TN solver's energies on Pegasus and Zephyr problems, since the paper identifies contraction failure, not search width, as the bottleneck.","On tile-planted instances, the TN approach's factor-of-3 energy advantage over both Ising machines suggests that planted-structure problems are a favorable use case for deterministic tensor-network solvers even on current annealer geometries.","Because the branch-and-bound width $M$ caps the number of independent low-energy states the TN solver can return, applications that need thousands of diverse near-optimal solutions will continue to prefer randomized samplers such as the quantum annealer and SBM.","The same TN machinery applies to generalized Potts models and king's-graph problems, so the solver is not limited to Ising spin glasses; its failures on denser graphs also serve as a stress test for approximate TN contraction algorithms more broadly.","With runtimes two orders of magnitude longer than the Ising machines, tensor-network approaches will need either contraction-speed breakthroughs or a restriction to small or structured instances to be practically competitive in time-to-solution."],"supporting_citations":[{"why":"Supplies the original tensor-network branch-and-bound method on Chimera graphs that the paper extends to Pegasus and Zephyr geometries.","marker":"[31]"},{"why":"Provides the boundary-MPS and variational-optimization machinery used for approximate PEPS contraction.","marker":"[24]"},{"why":"Supplies the randomized truncated SVD algorithm used to compress boundary MPS, which the paper identifies as the source of contraction instability.","marker":"[43]"},{"why":"Supplies the randomized linear algebra background and the caveat that randomized SVD can get stuck in local optima, which the paper invokes in its diagnosis.","marker":"[44]"},{"why":"Provides the PEPS representation of a classical partition function used to encode the Boltzmann distribution over Potts variables.","marker":"[29]"},{"why":"Establishes that exact PEPS contraction is #P-hard, which the paper cites to justify approximate contraction and explain why TN optimization is heuristic.","marker":"[30]"},{"why":"Defines the Simulated Bifurcation Machine that serves as the classical Ising-machine baseline in the benchmarks.","marker":"[22]"},{"why":"Provides the modified, discretized SBM variant used in the benchmarks and the basis for its parallel GPU implementation.","marker":"[23]"},{"why":"Supplies the class III (CBFM-P) instance distribution on Pegasus graphs and the quantum-annealing benchmark that the paper extends with SBM and TN results.","marker":"[51]"},{"why":"Supplies the class IV tile-planting instances with known ground states used to measure the TN approach's factor-of-3 energy advantage over the Ising machines.","marker":"[52]"}],"fun_headline_variants":["Tensor nets trail quantum annealers on big spin-glass problems","On 5,000-spin glasses, tensor networks are 100x slower and 0.1–1% off","TN search lags Ising machines on random spin-glass sampling diversity","Tensor networks win on planted tile glasses but lose on random ones","Deterministic TN solver: few low-energy states, Ising machines get thousands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole search rests on the premise that a boundary matrix product state of bond dimension around 8 to 10, obtained via randomized SVD and a 1-site variational sweep, gives conditional probabilities accurate enough that the branch-and-bound procedure ranks partial configurations in the right order; the paper shows this premise failing at larger inverse temperatures and bond dimensions.","fun_headline_variants_meta":{"raw":{"variants":["Tensor nets trail quantum annealers on big spin-glass problems","On 5,000-spin glasses, tensor networks are 100x slower and 0.1–1% off","TN search lags Ising machines on random spin-glass sampling diversity","Tensor networks win on planted tile glasses but lose on random ones","Deterministic TN solver: few low-energy states, Ising machines get thousands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1618,"prompt_tokens":1123,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":739,"completion_tokens_details":{"reasoning_tokens":392}},"tokens_in":739,"tokens_out":495,"duration_ms":5570,"temperature":1.0,"reasoning_tokens":392,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:07:43.968678+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the 216-spin Pegasus class I benchmark with $\\beta = 3$ and bond dimension $\\chi = 8$ using the same randomized-SVD zipper: the paper predicts repeated runs with different random seeds will disagree about whether the ground state is found, with success rates well below 100%. If every run reliably found the ground state while the contraction spectrum still showed steeply decaying singular values, the attribution of failure to randomized-SVD instability would be refuted.","supporting_citations":[{"cited_title":"Nishino, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the original tensor-network branch-and-bound method on Chimera graphs that the paper extends to Pegasus and Zephyr geometries."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PEPS representation of a classical partition function used to encode the Boltzmann distribution over Potts variables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that exact PEPS contraction is #P-hard, which the paper cites to justify approximate contraction and explain why TN optimization is heuristic."},{"cited_title":"Bando, Y","cited_arxiv_id":null,"evidence_quote":"Defines the Simulated Bifurcation Machine that serves as the classical Ising-machine baseline in the benchmarks."},{"cited_title":"Kashimata, M","cited_arxiv_id":null,"evidence_quote":"Provides the modified, discretized SBM variant used in the benchmarks and the basis for its parallel GPU implementation."},{"cited_title":"Yedidia, W","cited_arxiv_id":null,"evidence_quote":"Supplies the class III (CBFM-P) instance distribution on Pegasus graphs and the quantum-annealing benchmark that the paper extends with SBM and TN results."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the class IV tile-planting instances with known ground states used to measure the TN approach's factor-of-3 energy advantage over the Ising machines."}],"review_version":1}