{"id":"64247f4c-afe1-4aee-a1c4-601c7230229e","arxiv_id":"1909.02559","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A tensor-network-based QAOA simulator is shown to evaluate energies on 1000-vertex regular graphs, and QAOA energies are used to separate non-isomorphic graphs and expose graph structure.","lead":"This paper reports the Alibaba Cloud Quantum Development Platform's tensor-network simulator for QAOA, with benchmarks on regular graphs up to 1000 vertices and depth 5. It also uses QAOA energy values to distinguish non-isomorphic graphs and to reveal graph structure such as triangle counts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Graph-separation claim rests on an unproved conjecture and an unreproducible numerical sweep; needs high-precision reproduction.","rationale":"The paper's simulation benchmarks are plausible: the tensor-network lightcone method for sparse QAOA energy evaluation is standard, and the running times in Table 1 are consistent with known techniques, so I do not see a reason to doubt the platform's core capability. The small-cycle-free optimization section is honestly labeled as 'best function values found'; without global optimality certificates the word 'optimal' in the section title is stronger than the evidence, but this is a presentation issue more than a flaw in the simulation. The graph-structure-discovery section is where the strongest claim rests on the least supported foundation. Theorem 1 and Conjecture 1 are both cited to an unpublished companion paper [17], and the conjecture as quoted is inconsistent with Theorem 1 when p=1: on 3-regular graphs with equal triangle counts, depth-one energies coincide on every angle sequence, so the claimed probability-1 separation cannot hold at p=1. The empirical separation of all small cubic graphs is a finite numerical statement; without the angle-sampling protocol and precision analysis, rounding artifacts cannot be excluded. I therefore keep the reader's conditional verdict. This agrees partially with the reader's weakest_assumption, which also flagged Conjecture 1, though it additionally focused on global optimality of the Table 3 angles.","tokens_in":8958,"tokens_out":13117,"duration_ms":136353,"concrete_test":"Re-run the exhaustive depth-4 QAOA separation of all 41301 non-isomorphic 3-regular graphs on 18 nodes using a fresh random angle sequence and 128-bit or arbitrary-precision arithmetic, and record the minimum energy gap between non-isomorphic pairs. If any pair has a zero gap at all sampled angles, or if the minimum gap is at or below the 64-bit floating-point error of the original tensor-network contractions, the paper's 'full separation' claim is not established by the reported data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 5, the central graph-structure claim is that QAOA energies computed on AC-QDP separate all non-isomorphic 3-regular graphs up to size 18 (and other families), leading the authors to suggest that QAOA energies give a full characterization of isomorphism classes. Two load-bearing supports are missing. First, the only theoretical support, Conjecture 1 (cited to the unpublished [17]), is stated as: with probability 1 over all degree sequences in [0,2π]^{2d}, the energies of any two non-isomorphic n-node graphs differ. As stated, this conflicts with the paper's own Theorem 1 ([17]): for 3-regular graphs, depth-one QAOA values depend only on the number of triangles, so two non-isomorphic 3-regular graphs on the same n with the same triangle count have exactly equal depth-one energies for every angle sequence, making the probability 0, not 1. If the conjecture is meant only for sufficiently large d, the paper does not say so. Second, the reported exhaustive separation is not accompanied by the sampling protocol, the number of angle draws, or numerical tolerances; if the actual energy gaps between some non-isomorphic pairs are at or below floating-point precision, the separation claim could be an artifact of rounding rather than a true property of QAOA. Because the paper's most distinctive conclusion rests on an unproved and internally underspecified conjecture plus an unreproducible numeric sweep, the graph-separation claim is the least secure load-bearing element.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes the Alibaba Cloud Quantum Development Platform (AC-QDP), a tensor-network-based classical simulator for evaluating QAOA energy functions. It benchmarks single energy-query times against Cirq, Qiskit, and qTorch on random regular graphs, reports optimization results for small-cycle-free regular graphs using several heuristic optimizers, and proposes using QAOA energies as a graph invariant for distinguishing non-isomorphic graphs. The main empirical claims are that AC-QDP can evaluate 1000-vertex regular graphs up to depth 5, that the platform found good (titled 'best') angle sequences for tree-like regular graphs, and that for all tested small graphs QAOA energies at modest depth separate non-isomorphic graphs in several families. The theoretical grounding for the graph-separation claims is cited to an unpublished companion paper [17].","tokens_in":9220,"tokens_out":4960,"duration_ms":53091,"significance":"If the benchmark numbers are reproducible, AC-QDP is a useful practical tool for QAOA studies at sizes beyond state-vector simulation; the comparison against three external simulators is an appropriate and non-circular evaluation. The graph-structure discovery results are potentially interesting, but as presented they rest on an unproved conjecture and on finite numerical sweeps, so the manuscript's strongest conclusion ('QAOA energies give a full characterization of isomorphism classes') is not established. The paper would be strengthened by clearly separating engineering results from conjectural statements and by supplying the missing experimental and theoretical support.","major_comments":[{"comment":"The abstract states that AC-QDP can 'simulate thousands of qubits for up to 4 layers', but the benchmark data in Table 1 report only n=1000 vertices, and for d=4 and d=5 the depth does not even reach p=4. If the 'thousands' claim is supported by additional experiments or by the earlier papers [8,20], it should be stated explicitly; otherwise the abstract should be corrected. In addition, Fig. 1 and Table 1 report average query times without error bars, standard deviations, or per-instance variation, although the text notes large variation for n=30 and n=50; please include the spread or the raw data so the comparison is reproducible.","section":"Abstract and Section 3.2"},{"comment":"As stated, Conjecture 1 conflicts with Theorem 1 of the same section. For 3-regular graphs, Theorem 1 says that depth-one QAOA values depend only on n and the number of triangles, so any two non-isomorphic 3-regular graphs with the same triangle count have identical depth-one energies for every angle sequence, making the probability of a difference 0, not 1. If the conjecture is meant only for sufficiently large p, or for generic (not all) degree sequences, that qualification is missing. Since this conjecture is the only theoretical support for the graph-separation algorithm, it must be stated precisely and proved, or explicitly labeled as an open problem, before the separation claims can be accepted.","section":"Section 5, Conjecture 1 and Theorem 1"},{"comment":"The paper calls the angle sequences in Table 3 'best' and says the platform 'solved' the small-cycle-free cases, but Table 2's own heading says 'Best function values found', and the methods used (dlib, differential evolution, FOURIER, grid search plus local refinement) are heuristics with no optimality certificate. The claim of optimal QAOA sequences is therefore not established. Moreover, the sentence 'However, with the increase of the number of angles, optimization .' at the end of Section 4 is incomplete; it appears to acknowledge a limitation and must be completed, since an unfinished sentence cannot be assessed.","section":"Section 4, Tables 2 and 3"},{"comment":"The exhaustive-separation claims (the rows 'All 4060 non-iso 3-regular graphs' and 'All 41301 non-iso 3-regular graphs on 18 nodes') are reported without the sampling protocol: the number of random angle draws per graph pair, the numerical tolerance for declaring two energies different, the floating-point precision used, and the specific depth at which each family is separated. If some non-isomorphic pairs have energy gaps at or below machine precision, the separation could be an artifact of rounding rather than a property of QAOA. Please specify the protocol, ideally with arbitrary-precision checks for near-degenerate pairs, and clarify that the 4060 row refers to 16-vertex graphs, which is not stated in the table. The broader statement that QAOA energies give 'a full characterization of isomorphism classes' goes beyond the evidence, since Conjecture 1 is unproved and the tests cover only finitely many small graphs.","section":"Section 5, graph separation protocol"}],"minor_comments":[{"comment":"There are several typographical errors, including 'Altough' (Section 3.1), 'whre' (Section 3.1), 'tensot entwork' (Section 3.1), 'chanllenging' (Sections 3.1 and 4), and 'supressed' (Section 3.2).","section":"Throughout"},{"comment":"The caption says 'For the cased = 2' but the surrounding text says the figure shows degree 3 cases; this is likely a typo and should be corrected.","section":"Figure 1 caption"},{"comment":"Theorem 1 and Conjecture 1 are attributed to [17], which is described as 'in preparation'. Relying on an unpublished manuscript for the central theoretical statements makes the paper difficult to verify; either include the proofs in an appendix or state the results as conjectures with the evidence provided.","section":"References"},{"comment":"The angle sequences are reported to only four decimal places. Since the optimization landscape can be flat or have narrow features, please either provide more digits in an appendix or state that the full-precision sequences are available from the platform.","section":"Table 3"},{"comment":"The random-walk observations in Fig. 3 are qualitative; no error bars, number of walk steps, or statistical measures of 'distance' versus 'QAOA separation' are given. This section should be framed as an exploratory observation.","section":"Section 5, random walk"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on an unpublished companion paper [17] by one of the coauthors, and the conflict between Conjecture 1 and Theorem 1 is a serious internal consistency issue that must be resolved. The engineering benchmark results are plausible and useful, but the graph-structure discovery claims need to be substantially rewritten, with either proofs or clear conjecture labels. The proprietary nature of AC-QDP and the absence of a detailed sampling protocol may also make the numerical claims difficult to reproduce independently."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is an engineering report with a real kernel of new results. The tensor-network QAOA simulator is benchmarked against Cirq, Qiskit, and qTorch on regular graphs up to 1000 vertices, which is a genuinely useful data point for anyone who needs classical QAOA energy evaluations. The small-cycle-free angle tables for degrees 3–7 are also practical, even if they are clearly heuristic rather than certified optima. The graph-separation experiments are the most interesting part, and the empirical separation of all 3-regular graphs up to 18 nodes is a striking demonstration if it holds up.\n\nThe soft spots are real but not fatal. The abstract says 'thousands of qubits' while the data show 1000 vertices; that is an overstatement, but a small one, and the rest of the text is more careful. Calling the angles in Table 2 'optimal' is loose—they are the best found values, with no optimality certificate. The running times in Table 1 have no error bars, which matters because Figure 1 shows high variance in some regimes.\n\nThe larger concern is the graph-structure claim. Conjecture 1, cited to the unpublished [17], says that non-isomorphic graphs give different QAOA energies with probability 1 over angle sequences. As the stress-test note points out, this conflicts with Theorem 1 (also from [17]): for 3-regular graphs at depth 1, energies depend only on triangle count, so two such graphs with the same triangle count have equal energies for every angle sequence. The conjecture might be intended only for sufficiently large depth, but the paper does not say that. The exhaustive separation sweep also lacks the sampling protocol, number of angle draws, and tolerances, so the claim is not independently reproducible as reported. These are fixable but need to be addressed.\n\nI came away convinced the core benchmark is honest: they checked against external simulators, so the platform's capability is not circular. The paper is a reasonable demonstration of a practical tool, but the flashy conclusion about graph isomorphism should be read with caution until [17] appears and the numerics are documented.\n\nWho is this for? Someone working on QAOA simulation, variational quantum algorithms, or graph-similarity heuristics will get value from the benchmark data and the angle tables. It is worth a serious referee pass, but not as-is; the abstract, the 'optimal' language, and the Conjecture 1/Theorem 1 tension all need revision. If I were an editor, I would send it to review, with the expectation that the authors tighten the claims and make the separation experiments reproducible.","headline":"Solid QAOA simulation engineering with useful benchmark data, but the graph-isomorphism separation claim leans on an unproved conjecture and the abstract overstates scale.","tokens_in":9766,"tokens_out":1415,"would_cite":false,"duration_ms":17297,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A classical tensor-network engine evaluates QAOA energies for 1,000-qubit graphs.","keywords":["quantum approximate optimization algorithm","tensor network contraction","classical simulation of quantum circuits","graph isomorphism","small-cycle-free regular graphs","MAX-CUT","QAOA energy landscape","graph similarity"],"falsifier":"Run the energy-separation procedure on the 40-node hard graph pair described in the paper at depth 5 or higher over many random angle sequences; if their energies coincide for a non-negligible fraction of angle draws, the conjecture that QAOA energies separate all non-isomorphic graphs is false.","tokens_in":8766,"feed_emoji":"⚛️","tokens_out":10367,"duration_ms":100833,"temperature":0.7,"pith_summary":"This paper claims that a classical tensor-network simulator can evaluate the QAOA energy function for sparse graphs far beyond the reach of state-vector simulators: regular graphs with 1,000 vertices and up to depth 4 or 5. Because QAOA is a leading candidate for near-term quantum advantage, being able to test and optimize its angles classically at this scale turns it into a practically testable hybrid algorithm. The paper further claims that QAOA energies carry structural information about graphs: at depth one on 3-regular graphs the energy is determined by the number of triangles, and empirically the energy values separate all non-isomorphic 3-regular graphs up to 18 nodes, all strongly regular graphs up to 26 nodes, and hard-to-distinguish 20-node graph pairs. If these claims hold, QAOA energies are both a scalable classical benchmark and a promising graph invariant.","feed_headline":"Classical runs of QAOA reach 1,000-qubit graphs","feed_subtitle":"The same tensor-network engine separates all tested non-isomorphic graphs by their cut-energy signatures.","key_machinery":"The workhorse is the QAOA energy function $F_p(\\vec\\gamma,\\vec\\beta)$, the expectation of the cut Hamiltonian in the variational state. The platform evaluates each Hamiltonian term by contracting a tensor network built from the lightcone of that edge, exploiting the diagonality of the cost gates and the uniform-plus initial state to shrink the tensors. For small-cycle-free graphs, all edge lightcones are isomorphic to one tree whose size depends only on the degree $d$ and depth $p$, so the whole simulation collapses to a single template. For graph structure discovery, the machinery is the energy landscape: the vector of energies at several random angle sequences serves as a graph signature, and Theorem 1 identifies the depth-one invariant as the triangle count.","core_discovery":"On its own terms, the paper demonstrates that tensor-network contraction, applied one edge at a time within the lightcone of that edge, reduces QAOA energy evaluation to a set of tractable contractions, and that this reduction is powerful enough to handle regular graphs with 1,000 vertices at depth 4 (and degree 3 at depth 5), while remaining competitive with state-vector simulators at small sizes. The paper computes near-optimal angle sequences for small-cycle-free regular graphs, where the girth is at least 2p+2 and every edge's lightcone is the same tree, reporting the best values found for degrees 3 through 7. It then uses energy values at random angles as a graph signature: at depth 4 it separates all 4,060 non-isomorphic 3-regular graphs on 16 nodes and all 41,301 on 18 nodes, and at depth 3 it separates the strongly regular 26,10,3,4 family. The paper also states Theorem 1, attributed to a companion work, that depth-one QAOA energies of 3-regular graphs depend only on the degree sequence, the number of nodes, and the number of triangles; the observed clustering of the graph landscapes is decoded by triangle number.","pith_inferences":["The authors do not test whether the small-cycle-free angle tables remain near-optimal on graphs with short cycles; a direct test would compare their performance against fresh optimization on random regular graphs of the same degree.","If QAOA energies form a hierarchy of local-subgraph counts, then depth $p$ should distinguish graphs differing in counts of small subgraphs of size roughly $p$; comparing energy values against explicit subgraph counts on random graphs would test this.","The clustering observed in 3-regular graph landscapes suggests that energy vectors could group graphs by structural similarity, not just decide isomorphism; the random-walk result points toward a metric on graph space that a clustering algorithm could consume.","Because the 40-node hard pair only fails at depth 4, it sets a concrete stress test: a simulator reaching depth 5 on that pair would resolve whether the separation phenomenon extends there."],"forward_implications":["QAOA parameter optimization can be moved off the quantum device for sparse instances: the angle tables for small-cycle-free graphs can seed or replace the optimization loop on large random regular graphs.","The energy-separation results give a classical, circuit-free certificate of non-isomorphism for entire families of graphs, including 41,301 3-regular graphs on 18 nodes.","Because depth-one energies of 3-regular graphs are a function of triangle count, graphs with different triangle counts are immediately separated at depth one, making triangle count a cheap subclassifier.","Benchmarking at 1,000 vertices gives future QAOA simulators and hardware demonstrations a concrete scale target: single energy queries at depth 4 take tens of seconds on this platform.","The random-walk experiment suggests that QAOA energy separation grows with graph distance, so the landscape can be used as a rough dissimilarity measure for slowly changing graph sequences."],"supporting_citations":[{"why":"Supplies the tensor-network classical simulation engine that AC-QDP builds on; the paper cites it for the platform's computational capability.","marker":"[8]"},{"why":"Supplies the large-scale classical simulation techniques, including diagonal-gate simplification, reused by the QAOA package.","marker":"[20]"},{"why":"Defines the QAOA ansatz, the energy function $F_p$, and the lightcone observation that the simulator exploits.","marker":"[10]"},{"why":"Establishes the tensor-network approach to QAOA simulation that this paper extends, and serves as a benchmark baseline.","marker":"[15]"},{"why":"Provides Conjecture 1 and Theorem 1, the theoretical grounding for using QAOA energies to separate non-isomorphic graphs and for the triangle-count characterization.","marker":"[17]"},{"why":"Supplies the FOURIER heuristic used in the search for optimal angle sequences on small-cycle-free graphs.","marker":"[21]"},{"why":"Proposes using small-cycle-free angle sequences directly to skip the optimization phase, motivating the optimization use case.","marker":"[16]"},{"why":"Supplies analytic QAOA energy results for special cases that the small-cycle-free numerical tables extend.","marker":"[19]"}],"fun_headline_variants":["QAOA on 1,000 qubits: tensor nets make it classical","Alibaba's quantum platform simulates QAOA past 1,000 qubits","Tensor-network QAOA scales to 1,000-qubit graphs","Graph structure discovery via QAOA energy signatures"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The graph-separation claims rest on an unproved conjecture: for any two different graphs, a randomly chosen set of QAOA angles will almost surely give different energy values; the paper checks only a finite list of graph families and angle draws.","fun_headline_variants_meta":{"raw":{"variants":["QAOA on 1,000 qubits: tensor nets make it classical","Alibaba's quantum platform simulates QAOA past 1,000 qubits","Tensor-network QAOA scales to 1,000-qubit graphs","Graph structure discovery via QAOA energy signatures"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000453,"raw_usage":{"total_tokens":2256,"prompt_tokens":902,"completion_tokens":1354,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":1279}},"tokens_in":518,"tokens_out":1354,"duration_ms":10204,"temperature":1.0,"reasoning_tokens":1279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:47:07.734539+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the energy-separation procedure on the 40-node hard graph pair described in the paper at depth 5 or higher over many random angle sequences; if their energies coincide for a non-negligible fraction of angle draws, the conjecture that QAOA energies separate all non-isomorphic graphs is false.","supporting_citations":[{"cited_title":"Alibaba Cloud Quantum Development Platform: Large-Scale Classical Simulation of Quantum Circuits","cited_arxiv_id":"1907.11217","evidence_quote":"Supplies the large-scale classical simulation techniques, including diagonal-gate simplification, reused by the QAOA package."},{"cited_title":"Schuyler Fried, N","cited_arxiv_id":null,"evidence_quote":"Establishes the tensor-network approach to QAOA simulation that this paper extends, and serves as a benchmark baseline."},{"cited_title":"Graph structure discovery with QAOA","cited_arxiv_id":null,"evidence_quote":"Provides Conjecture 1 and Theorem 1, the theoretical grounding for using QAOA energies to separate non-isomorphic graphs and for the triangle-count characterization."},{"cited_title":"Quantum approximate optimization algorithm for MaxCut: A fermionic view","cited_arxiv_id":null,"evidence_quote":"Supplies analytic QAOA energy results for special cases that the small-cycle-free numerical tables extend."}],"review_version":1}