{"id":"0a0a7cc5-c95a-49d6-8823-fff43e12caca","arxiv_id":"2608.13110","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"A deterministic tensor network contraction computes exact probabilities for all of IBM's doped Clifford sampling outputs in 37.3 minutes on 32 H100 nodes.","lead":"This paper reports exact classical simulation of IBM's 70-qubit, 70-layer doped Clifford random circuit sampling experiment using a deterministic tensor network contraction, completing all 2051 published output batches in 37.3 minutes on 32 H100 nodes. The simulation's log-XEB estimate of 0.350 is compatible with IBM's reported fidelity lower bound, undercutting the claim that this instance is classically hard.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The d=70 numerical claims rest on complex64 arithmetic validated only to d=56; a complex128 spot-check would settle the main residual risk.","rationale":"The width theorem and Ratcatcher certification are well supported; the execution details are plausible and the XEB compatibility provides internal evidence for the circuit description. The sole load-bearing gap is the unvalidated precision at the actual operating depth. A complex128 spot-check is cheap relative to the reported 37-minute run and would either close the gap or expose a real problem. Conditional acceptance on that check is proportionate; the reader's ACCEPT is close, and this condition does not challenge the theoretical contribution.","tokens_in":18382,"tokens_out":28886,"duration_ms":267376,"concrete_test":"Recompute a random subset (e.g., 100) of the 2051 amplitude batches at d=70 in complex128 using the same contraction path on the same eight-H100 nodes (512 GiB largest tensor fits in 640 GB aggregate), and compare per-batch TV and the resulting log-XEB against the complex64 values. If the log-XEB shift is below ~0.01 (one-third of the reported SE), the complex64 result is secure; if it exceeds the SE, the reported 0.35034 is not established.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central numerical claim—exact probabilities for all 2051 published batches—depends on complex64 arithmetic at d=70, where the largest intermediate has 2^35 entries (§5.1, §5.2). The paper validates complex64 against complex128 only up to d=56, where the intermediate has 2^28 entries, reporting TV < 9×10^-7. The d=70 path has seven additional doublings of tensor size and roughly 4,900 contraction steps; float32 rounding could in principle accumulate to a scale comparable to the reported log-XEB standard error (0.02689), which would change the headline diagnostic. The paper labels the calculation 'exact' only in the no-truncation sense and explicitly flags finite precision, but it never measures the actual finite-precision error at the operating point. This is the one unresolved assumption that could invalidate the numerical execution without affecting the width theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Manabe, Gu, and Pan present a deterministic temporal-boundary tensor network contraction algorithm for open-boundary, one-dimensional brickwork circuits whose entangling gates have operator Schmidt rank 2. For an n-qubit circuit of depth d, the algorithm evaluates an exact amplitude with contraction width ceil(d/2) and cost O(nd 2^ceil(d/2)), and the authors use Ratcatcher to certify that no smaller width exists for the n=70 instances tested. A batched variant with k open output bits retains the same width for 1<=k<=floor(d/4). The paper applies the method to IBM's 70-qubit, 70-layer doped Clifford circuit with 468 T gates, computing all 2051 published-output amplitude batches in 37.3 minutes on 32 eight-H100 nodes. The resulting probabilities give log-XEB = 0.35034 with a 95% confidence interval [0.29763,0.40305], which the authors interpret as numerically compatible with IBM's fidelity lower bound under Porter-Thomas and scrambled-noise assumptions. The paper also analyzes how slicing, entangling-gate bond dimension, and periodic boundary conditions affect simulation cost, producing a classical-simulatability map for circuit design.","tokens_in":18495,"tokens_out":16384,"duration_ms":150467,"significance":"If the results hold, this is a significant advance in classical simulation of a specific doped Clifford RCS experiment. The width theorem is parameter-free and machine-verifiable via the Ratcatcher certification, and the independence of the dense contraction cost from T-gate count and placement is a clean structural insight. The reported execution is a concrete resource comparison against IBM's hardware experiment, and the public release of amplitudes and contraction paths supports reproducibility. The main unresolved risk is the use of complex64 arithmetic at the d=70 operating point, which is validated only up to d=56; a spot-check at d=70 would make the numerical claim fully secure.","major_comments":[{"comment":"The d=70 production calculation uses complex64 arithmetic, but the complex64-vs-complex128 validation in Fig. 7 (right panel) is only reported through d=56. The d=70 operating point has a 2^35-entry largest intermediate and roughly 4,900 contraction steps, and the paper does not report any complex128 comparison at depths above 56 nor a forward-error bound. Because the headline numerical results (log-XEB and its 95% interval) are computed from these d=70 amplitudes, the finite-precision error at the operating point is an unverified quantity that directly bears on the numerical claim. A complex128 spot-check for a subset of the 2051 batches at d=70 would settle this risk; the 512 GiB complex128 payload of the largest intermediate still fits on eight 80 GB H100 GPUs.","section":"§5.1–5.2, Fig. 7, Table 1"}],"minor_comments":[{"comment":"The abstract's phrase '256 times smaller than IBM's estimation' is ambiguous because IBM's memory-constrained cotengra estimate (2^30 scalars) is actually smaller than the paper's 2^35-entry intermediate; the qualifications in Section 1 about differing tensorization and memory-accounting conventions should be carried into the abstract.","section":"Abstract and §1"},{"comment":"The proof of Proposition 3.1 is quite terse; it analyzes two local configurations but does not state the full sweep order or an inductive invariant. Since the width bound is a central claim, an explicit inductive proof (or a more detailed pseudocode-level invariant) would make the argument easier to verify.","section":"§3.1, Proposition 3.1"},{"comment":"The proofs of Proposition 3.1 and Proposition 4.1 use the word 'rank' to mean the number of tensor legs, not the multilinear rank; this is clear in context, but a one-line definition at first use would prevent confusion for readers who interpret 'rank' differently.","section":"§3.1 and §4.3"},{"comment":"The text reports a makespan of 37 min 16 s while the abstract and Table 1 round to 37.3 minutes; the rounding convention should be stated or the numbers made consistent.","section":"§5.2 and Table 1"},{"comment":"The right panel of Figure 7 is labeled 'FP32 error,' while the text describes the total variation distance between complex64 and complex128 conditional distributions; relabeling the axis as 'TV distance (complex64 vs complex128)' would align the figure with the text.","section":"Figure 7"},{"comment":"Releasing the contraction schedules and amplitudes on Zenodo is helpful; making the simulator code available as well would further strengthen reproducibility of the numerical results.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The central algorithm and width theorem appear sound, and the reported execution is impressive. The only substantive gap I see is the missing numerical validation at d=70 (complex64 versus complex128), which is directly load-bearing for the log-XEB claim. If the authors supply a spot-check at the operating point or a rigorous error bound, I would be happy to support acceptance. The paper fits the scope of a broad quantum information journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this paper gives a deterministic tensor-network contraction of IBM's 70-qubit, 70-layer doped Clifford experiment with width ceil(d/2)=35, runs all 2051 published amplitude batches in 37.3 minutes on 32 eight-H100 nodes, and reports a log-XEB of 0.350 with a 95% interval that brackets IBM's fidelity lower bound. I think the central claim is right.\n\nWhat is actually new is the constructive width theorem. Proposition 3.1 proves that for open-boundary brickwork circuits with operator-Schmidt-rank-2 entangling gates and n>=4, 7<=d<=2n, the temporal-boundary sweep gives an unsliced path of width ceil(d/2). The proof is simple: absorb one-qubit gates, decompose each CZ into two rank-3 tensors, sweep the boundary state spatially. It is self-contained, not fitted, and the Ratcatcher certification that no smaller width exists for the tested instances is real evidence. The batched extension with k open output indices at no width increase is also a solid, useful contribution, and the design-sensitivity analysis for bond dimension, ring boundary, and slicing is honest and informative.\n\nThe numerical results are internally consistent: single-GPU memory scaling matches the predicted 2^ceil(d/2), complex64 versus complex128 total variation distance stays below 9e-7 through d=56, and the measured 256-GiB intermediate matches the analytic prediction. The log-XEB analysis is careful to state the Porter-Thomas and scrambled-noise assumptions, and the authors explicitly note that transverse contraction is a standard tensor network idea. They do not oversell the technique.\n\nSoft spots, in proportion. The one that matters is the complex64 arithmetic at d=70. The validation against complex128 stops at d=56, where the intermediate is 2^28 entries; the full instance has 2^35 entries and about 4900 contraction steps. The stress-test note is right: rounding could in principle accumulate to something comparable to the reported standard error of 0.02689. The paper flags this as finite-precision but never measures it at the operating point. A complex128 spot-check on a few batches at d=70 would settle it. This does not threaten the width theorem, but it is the one unresolved risk to the numerical headline. A minor addition: no source code is released, only amplitudes and paths on Zenodo. That is a reproducibility miss for a simulation paper, though the algorithm is described well enough to reimplement.\n\nWho this is for: anyone working on classical simulation of random circuit sampling or on designing next-generation sampling experiments. It deserves a serious referee. The width theorem is checkable in a day; the numerical execution is expensive to reproduce but the data release helps.\n\nMy recommendation: accept after a complex128 spot-check at d=70 and a request for code or a stronger numerical-error bound. As it stands, it is a solid and important paper.","headline":"Clean, self-contained width-35 contraction of IBM's 70x70 doped Clifford circuit, run in 37 minutes; the one unresolved risk is float32 arithmetic at d=70.","tokens_in":19066,"tokens_out":2140,"would_cite":true,"duration_ms":20245,"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":"A deterministic tensor-network sweep computes exact amplitudes for IBM's 70-qubit doped Clifford sampling experiment, and the same width formula maps where this circuit family stays classically easy to simulate.","keywords":["random circuit sampling","doped Clifford circuits","tensor network contraction","contraction width","cross-entropy benchmarking","quantum advantage","GPU simulation","brickwork circuits"],"falsifier":"Evaluate one or more of the 2051 amplitude batches at $d=70$ using complex128 arithmetic on enough nodes to hold the 512 GiB tensor, or using an independent contraction path, and compare with the published complex64 probabilities: the authors' own $d\\leq 56$ validation gives a complex64-vs-complex128 total variation below $9\\times10^{-7}$, so any larger discrepancy at depth 70 would falsify the numerical claim.","tokens_in":18145,"feed_emoji":"⚛️","tokens_out":12253,"duration_ms":110433,"temperature":0.7,"pith_summary":"This paper tries to establish that IBM's headline 70-qubit, 70-layer 'doped Clifford' sampling experiment can be checked classically by a deterministic tensor-network contraction that exploits geometry rather than Clifford structure. The central claim is a width formula: for an open chain whose entangling gates have operator Schmidt rank 2, one exact amplitude (no bond truncation) costs contraction width $\\lceil d/2\\rceil$, independent of the values or placement of one-qubit gates and of the 468 $T$ gates. Applied to the IBM instance, the formula gives a width-35 path whose largest intermediate tensor holds $2^{35}$ complex64 entries (256 GiB), and the authors report evaluating all 2051 published-output amplitude batches in 37.3 minutes on 32 nodes with eight H100 GPUs each. The resulting log-XEB estimate is 0.35034 with a 95% interval [0.29763, 0.40305], which they present as numerically compatible with IBM's reported fidelity lower bound under the standard Porter\\textendash Thomas and scrambled-noise assumptions. The same machinery doubles as a design map, quantifying how entangler bond dimension, boundary geometry, depth, and memory capacity move the classical-simulation cost by orders of magnitude.","feed_headline":"37-minute GPU sweep reproduces IBM 70-qubit sampling outputs","feed_subtitle":"Exact tensor-network amplitudes match the experiment's fidelity bound and map where such circuits stay classically easy.","key_machinery":"The load-bearing object is the temporal-boundary sweep over the PEPS-like tensor network obtained by replacing each CZ gate (operator Schmidt rank 2) with two rank-3 tensors joined by a 2-dimensional bond, then absorbing all one-qubit gates into neighboring tensors. The circuit becomes a rectangular planar network of temporal width $\\lceil d/2\\rceil$ and spatial length $n$. Algorithm 1 sweeps a 'boundary tensor' along the spatial direction, reversing time on alternate rows, and contracts pairs of tensors whenever a direct absorption would exceed rank $\\lceil d/2\\rceil$; Proposition 3.1 proves this keeps every intermediate at width $\\lceil d/2\\rceil$. A batched variant (Algorithm 2) leaves $k \\leq \\lfloor d/4\\rfloor$ output indices open and returns $2^k$ amplitudes at the same width, which is what turns amplitudes into samples. Ratcatcher supplies the optimality witness on tested instances, and a hyper-optimized path search provides the comparison baseline and the slicing-overhead measurements.","core_discovery":"The paper's discovery is that an open-boundary, one-dimensional brickwork circuit whose entangling gates have operator Schmidt rank 2 possesses an exact, deterministic contraction path with width $\\lceil d/2\\rceil$ for depth $d$, and that this path is width-minimal on every tested instance, as certified by the Ratcatcher carving-width algorithm. For the IBM circuit with $n=d=70$, this yields a width-35 schedule, a 256 GiB largest-intermediate tensor, and a measured 37.3-minute makespan on 32 eight-H100 nodes for all 2051 published amplitude batches. The authors use these exact probabilities to compute log-XEB = 0.35034 with 95% interval [0.29763, 0.40305], which they interpret, under the Porter\\textendash Thomas and scrambled-noise assumptions, as numerically compatible with IBM's fidelity lower bound of 0.284. They further show that the favorable width is a geometric property: increasing the entangler bond dimension from 2 to 4, or closing the chain into a ring, roughly doubles the width, while slicing one bit below the natural width incurs a very large overhead. Consequently, they present the method both as an independent diagnostic of the experimental output and as a quantitative tool for designing future doped Clifford sampling experiments.","pith_inferences":["If the width formula holds beyond the tested range, the cliff for this circuit family sits near depth 86 for the open $\\chi=2$ geometry under 1024-H100 memory assumptions; any 'hard' instance in this family should be checked against that ceiling before invoking quantum advantage.","The independence of the schedule from $T$-gate count suggests that, for open one-dimensional chains, the practical hardness of doped Clifford circuits comes from geometry and entangler rank rather than from magic; closing the chain or using rank-4 entanglers is the natural hardening route, though it may also break the efficient spacetime-code verification the experiment relies on.","A direct testable extension is to run the same sweep on other rank-2 entanglers, such as CNOT or iSWAP-like gates, to see whether the width formula survives when the entangler is not diagonal; the paper's construction suggests it should as long as the network remains an open brickwork.","The projected 10.6-minute classical sampling time for the fidelity-weighted workload, compared with the experiment's 16.1-minute sampling run, suggests this specific 70-qubit instance sits close to parity on the sampling task itself with a 32-node classical cluster; whether that counts as 'advantage' depends on like-for-like resource and fidelity accounting."],"forward_implications":["For any open-boundary, bond-dimension-2 brickwork circuit with $n\\geq 4$ and $7\\leq d\\leq 2n$, one exact amplitude can be evaluated with contraction width $\\lceil d/2\\rceil$, time $O(nd2^{\\lceil d/2\\rceil})$, and space $O(2^{\\lceil d/2\\rceil})$; changing single-qubit gates or adding and moving $T$ gates does not change these counts.","The 2051 published output bitstrings of IBM's 70-qubit experiment are now individually calibrated: their exact ideal probabilities give log-XEB = 0.35034 with 95% interval [0.29763, 0.40305], an independent check of the reported fidelity scale.","For the IBM instance, the fidelity-weighted classical sampling workload (583 exact contractions for fidelity 0.284) projects to about 10.6 minutes on 32 eight-H100 nodes, while the measured full verification run took 37.3 minutes.","Memory, not arithmetic alone, is the practical bottleneck: slicing the width from 35 to 34 already creates many independent subtasks with very large cost overhead, and the aggregate-memory frontier for this circuit family sits near depth $d\\approx 86$ with 1024 H100 GPUs.","Circuit design can tune hardness against this method: replacing CZ with a rank-4 entangler doubles the width, and closing the chain raises the width to $\\min(n,d)$, though both changes carry physical-gate and spacetime-code costs."],"supporting_citations":[{"why":"Defines the IBM doped Clifford sampling experiment, its circuit, postselection, and the reported fidelity lower bound that this paper's simulation is benchmarked against.","marker":"[1]"},{"why":"Supplies the standard tensor-network contraction model in which arithmetic cost and contraction width are measured.","marker":"[38]"},{"why":"Provide the temporal-boundary and transverse-evolution ideas that the deterministic sweep adapts to the brickwork geometry.","marker":"[39, 40]"},{"why":"Provide the Ratcatcher carving-width algorithm and its planar tensor-network connection used to certify that no smaller width exists on the tested instances.","marker":"[43–45]"},{"why":"Introduces frugal rejection sampling and the fidelity-weighted resource accounting used to convert exact amplitudes into sampling-cost projections.","marker":"[19]"},{"why":"Provide big-batch and batched tensor-network sampling methods that the paper extends to $k$ open output indices.","marker":"[41, 42]"},{"why":"Supplies the hyper-optimized tensor-network contraction path search used as the comparison baseline and for the slicing-overhead studies.","marker":"[48]"},{"why":"Supplies the distributed multi-GPU tensor contraction framework used to execute the 32-node production runs.","marker":"[53]"},{"why":"Supply the Porter–Thomas statistics and linear cross-entropy benchmarking conventions used to interpret log-XEB as a fidelity proxy.","marker":"[12, 13]"}],"fun_headline_variants":["Exact 37-min classical simulation of IBM's 70-qubit circuit","37-minute exact classical run matches IBM 70-qubit sampling","Quantum advantage? 70-qubit circuit classically exact in 37 minutes","Tensor network halves width: exact 37-min simulation of 70-qubit IBM","Exact classical simulation: 70-qubit IBM sampling in 37 minutes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole schedule depends on the IBM circuit being exactly an open one-dimensional chain of operator-Schmidt-rank-2 CZ gates, and on complex64 rounding at depth 70 staying as small as it was measured to be through depth 56; if any entangler had rank 4, the chain were closed, or rounding grew, the width or the amplitudes would no longer match the experiment.","fun_headline_variants_meta":{"raw":{"variants":["Exact 37-min classical simulation of IBM's 70-qubit circuit","37-minute exact classical run matches IBM 70-qubit sampling","Quantum advantage? 70-qubit circuit classically exact in 37 minutes","Tensor network halves width: exact 37-min simulation of 70-qubit IBM","Exact classical simulation: 70-qubit IBM sampling in 37 minutes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001105,"raw_usage":{"total_tokens":4708,"prompt_tokens":1146,"completion_tokens":3562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":762,"completion_tokens_details":{"reasoning_tokens":3463}},"tokens_in":762,"tokens_out":3562,"duration_ms":24996,"temperature":1.0,"reasoning_tokens":3463,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:34:55.777327+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate one or more of the 2051 amplitude batches at $d=70$ using complex128 arithmetic on enough nodes to hold the 512 GiB tensor, or using an independent contraction path, and compare with the published complex64 probabilities: the authors' own $d\\leq 56$ validation gives a complex64-vs-complex128 total variation below $9\\times10^{-7}$, so any larger discrepancy at depth 70 would falsify the numerical claim.","supporting_citations":[{"cited_title":"Sampling hard circuits with verifiably high fidelity","cited_arxiv_id":"2607.25941","evidence_quote":"Defines the IBM doped Clifford sampling experiment, its circuit, postselection, and the reported fidelity lower bound that this paper's simulation is benchmarked against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard tensor-network contraction model in which arithmetic cost and contraction width are measured."},{"cited_title":"Huang, F","cited_arxiv_id":null,"evidence_quote":"Introduces frugal rejection sampling and the fidelity-weighted resource accounting used to convert exact amplitudes into sampling-cost projections."}],"review_version":1}