{"id":"d088ec1a-f658-46dd-a82c-b2d9c3898512","arxiv_id":"2608.03987","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A graph rewrite maps complex tensor networks to real ones with proven arithmetic overhead at most 3x and measured speedups on real-only NPUs.","lead":"This paper rewrites any complex tensor network into a real-valued network, so quantum circuit simulations can run on hardware that only accelerates real-number matrix multiplication. A proven cost bound (at most 3x the real arithmetic, often less) plus NPU benchmarks suggest this could be a practical default for TPU/NPU quantum simulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wall-clock speedups are confounded by execution-path differences; the all-real control shows 0.99–1.46x baseline variation at multiplication parity.","rationale":"I read the paper as making two connected claims: (1) a provable arithmetic cost law for realified contraction, and (2) a demonstrated wall-clock advantage on a real-only accelerator. I checked the proof of Theorem 4 in some detail: the rank-3 decomposition of C in Eq. (6) reproduces both slices exactly; the projections onto u_k, v_k, w_k involve only entries in {0,±1} and hence are additions/sign flips, so the 'three multiplications per merge' count is not hiding small multiplication terms; ride and pass costs are correct; the memory bound follows from carrying at most one green leg per intermediate. I therefore do not see a load-bearing flaw in the algebraic core. The load-bearing uncertainty is instead the causal attribution of the measured wall-clock speedups to the rewrite. Section 3.2's all-real control is the right check, and it already shows a 0.99–1.46 spread at exact multiplication parity, with median 1.16. Since the paper does not decompose this spread, the reported median 1.42x structured speedup could be substantially execution-path rather than arithmetic. The reader identified exactly this weakest assumption, and I agree. The appropriate verdict remains CONDITIONAL: the cost law is likely sound, but the hardware speedup claims need a better-controlled baseline before they can be taken at face value.","tokens_in":21398,"tokens_out":11346,"duration_ms":140618,"concrete_test":"Re-run the nine all-real control cells of §3.2 with a 'network-4M' executor: expand each complex merge into four rank-1 contractions using the same graph-rewrite engine as network-3M, so the only difference from network-3M is the multiplication count (4 vs 3 per merge), not execution scaffolding. If the m=r=0 median t4M/tnet-3M remains ≈1.16, the baseline difference is engine-path and the structured speedups should be adjusted accordingly; if it drops to ≈1.0, the all-real spread is a small-sample artifact and the speedup claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central hardware claim—network-3M beats GEMM-4M by median 1.7x (random) and 1.4x (structured)—is attributed to the realification rewrite, but §3.2's own all-real control undermines this attribution. On nine device-clean all-real cells (m=r=0), every executor performs identical multiplications, yet t4M/tnet-3M ranges 0.99–1.46 with median 1.16. Thus a substantial fraction of the structured median speedup could be an execution-path artifact (kernel launch counts, materialization, data movement) rather than the arithmetic saving. The paper explicitly states that the excess of measured speedups over the multiplication-count reference is not decomposed into launch, materialization, and data-movement contributions, and it excludes non-'device-clean' cells without reporting exclusion counts. If the GEMM-4M runner is systematically slower even at multiplication parity, the reported speedups overstate the method's benefit. The cost law itself appears sound; the vulnerability is the causal link from the algebraic rewrite to the measured wall-clock wins.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a network-level realification rewrite for complex tensor-network contraction on real-only matrix accelerators. Each complex tensor is augmented with a dimension-2 'green' leg encoding real and imaginary parts; pairwise contractions are classified as merge (two green operands), ride (one green operand), or pass (no green operand), and the structure tensor C realizes complex multiplication. The central theoretical contribution is Theorem 4, which states that, on any binary contraction tree, the realified contraction has arithmetic overhead 1+2m+r in real multiplications, where m and r are the merge and ride volume fractions, with 1 ≤ overhead ≤ 3 and every intermediate at most doubled in element count. The paper also reports contraction-volume audits on 67 circuits, order-transfer experiments showing convert-only is within 5e-4 of full reoptimization on 66 of 67 circuits, and Ascend 910 wall-clock comparisons in which network-3M beats GEMM-4M and GEMM-3M on twelve random circuits and on 52 of 55 device-clean structured cells. Appendices provide algebraic verification of the spider-fusion rules, a landscape theory for order-transfer flatness, a reverse-mode differentiation rule, precision gates, and a reproducibility/data-availability statement.","tokens_in":21636,"tokens_out":10027,"duration_ms":114474,"significance":"If the cost law and hardware results hold, this is a practically useful and theoretically clean method: the overhead formula is derived from a rank-3 tensor decomposition with no fitted parameters, it is audited across the full real-to-complex range, and the memory bound is simple and tight. The paper is unusually transparent about several limitations, including the un-decomposed wall-clock excess and the adversarial green-sparse family in Appendix A. The algebraic core is a nice compilation of known *-algebra realification into a contraction-level method, and the proof of Theorem 4 is short and verifiable. The empirical audits and reproducible data repository (pinned toolchain, archived plans) are strengths. The main weakness is the causal attribution of the measured wall-clock speedups to the arithmetic saving; the paper's own all-real control shows a substantial execution-path effect, and the device-clean exclusion is not fully documented in the text.","major_comments":[{"comment":"The all-real control cells execute identical multiplication counts in every executor (m=r=0), yet t4M/tnet-3M ranges 0.99–1.46 with median 1.16. This is a large execution-path speedup that is folded into every measured ratio. The paper states that the excess over the multiplication-count reference is not decomposed, but the structured median speedup of 1.42x and random median 1.68x are then not cleanly attributable to the realification arithmetic. Since the abstract and conclusions advertise these wall-clock wins, please provide a corrected analysis: report speedups net of a matched all-real baseline, or at least per-family all-real controls and a conservative lower bound on the arithmetic-specific speedup. The sign test (52 of 55) is uninformative on this point because network-3M also wins at multiplication parity.","section":"§3.2 (Table 3 and surrounding text)"},{"comment":"The 'device-clean' structured set excludes cells with a documented device-software fallback, but the number and identity of excluded cells are not reported in the manuscript. Without this, the 52-of-55 win record and the median speedups could be sensitive to exclusion. Please report the exclusion count, the reasons, and whether any non-win cells were excluded. The reader should not have to consult the data repository to assess the benchmark claim.","section":"§3.2, Table 3 caption"},{"comment":"The all-real control is only reported for structured cells. The random circuits (Table 2) have no multiplication-parity control, so the same execution-path confound is unconstrained for the median 1.68x random speedup. Since random circuits are the headline hardware result, either add a parity control for at least a subset of random cells or qualify the random speedups as total wall-clock ratios rather than arithmetic-derived speedups.","section":"§3.2, random-circuit rows"}],"minor_comments":[{"comment":"The heading 'F orward rule' contains a typographical space; should be 'Forward rule'.","section":"Appendix E"},{"comment":"The conclusion says the method saves 'up to 4x in the all-real limit' against 'flat 4x per-step GEMM-4M lowering', but Section 3.2 defines GEMM-4M with a structural-real shortcut that would give 1x in the all-real limit. Make explicit that the 4x comparison is against the naive 4M lowering without the structural-real shortcut.","section":"§4 / §3.2"},{"comment":"The text says a horizontal rule separates the nine core networks from the three extension circuits, but the rule is not visible in the table as typeset. Add a visible separator or explain the grouping in the caption.","section":"Table 1"},{"comment":"The 'four algebraic rules' are introduced as Eq. (4) plus the cascade rule; the enumeration is clear but could be labeled (R1)–(R4) to make the subsequent references easier to follow.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The theoretical contribution (Theorem 4 and the audits) is sound and publishable. The hardware section is the main weakness: the all-real control shows that the measured speedups are partly an execution-path artifact, and the device-clean exclusion is under-documented. I would not reject on the current evidence, but the hardware claims need recalibration and additional reporting before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a solid, useful paper. The core contribution—a contraction-tree-aligned realification of complex tensor networks with the cost law overhead = 1+2m+r—is correct and cleanly proved. I checked Theorem 4: it follows from the rank-3 decomposition of the structure tensor, and the volume-weighting argument is tight. The audits over 67 circuits have no fitted parameters, and the table values match the formula to displayed precision. The order-transfer flatness is a nice empirical find, and the appendix theorem that pins down when flatness holds is a real piece of theory.\n\nThe hardware validation is also better than most. All three executors share the same plan, inputs, and software stack; the f32 accuracy gates are sensible; and the paper explicitly says the multiplication count does not bound wall-clock time. That said, the speedup numbers are softer than they look. The nine all-real control cells run at multiplication parity, yet t4M/tnet-3M ranges 0.99–1.46, median 1.16. Since the structured median speedup is only 1.42x, a meaningful fraction of that could be execution-path difference—kernel launch counts, materialization, data movement—rather than the arithmetic saving. The control cuts both ways: it suggests network-3M has a real structural benefit even at parity, which is part of the method, but it also means you can't cleanly attribute the structured speedups to the arithmetic reduction. The paper owns this by saying the excess is not decomposed, but it remains the main weakness for anyone who wants to port the result to another accelerator. A proper decomposition would strengthen the paper a lot.\n\nSecond soft spot: the benchmark set is restricted to \"device-clean\" cells, and the excluded cells are not enumerated or counted. That is a reproducibility gap. They should provide the full list with fallback flags.\n\nMinor: GEMM-3M is not hand-fused, so the comparison to network-3M conflates the rewrite with kernel fusion. They note this too. And the speedups are modest (1.1–1.9x) and stack-specific.\n\nOverall, the algebra and the cost law are the real contributions, and they hold up. The hardware claims are plausible but partially confounded. This paper deserves a serious referee; I'd send it out, asking for the exclusion list and ideally a wall-clock decomposition. The method is now on my radar, and I'd cite it for the cost law if I worked on real-only tensor contraction.","headline":"Solid and honest paper: the cost law is provably right, the hardware speedups are plausible but partly confounded by execution-path effects the authors acknowledge but don't decompose.","tokens_in":22140,"tokens_out":4122,"would_cite":true,"duration_ms":41200,"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":"Any complex tensor network—the workhorse of quantum circuit simulation—can be rewritten into an equivalent real tensor network whose contraction cost is at most 3x the real-skeleton cost, bounded by the tight volume-weighted law 1+2m+r.","keywords":["quantum circuit simulation","tensor network contraction","realification","complex-to-real rewrite","structure tensor","Gauss 3M","contraction order","real-valued accelerators"],"falsifier":"Run an all-real benchmark cell (m=r=0) where all three executors perform the identical number of multiplications; if the realified executor's wall-clock time is not close to the four-GEMM baseline's, executor-level effects—not the arithmetic rewrite—explain part of the speedup. A stronger test: hand-fuse a per-GEMM 3M kernel that avoids separate launches and combination passes; if it matches network-3M's wall-clock time on the same contraction plans, the claimed structural benefit of the network-level rewrite is not structural.","tokens_in":21306,"feed_emoji":"⚛️","tokens_out":11270,"duration_ms":113469,"temperature":0.7,"pith_summary":"Quantum circuit simulation via tensor-network contraction normally needs complex-number matrix multiplication, but the matrix engines in modern accelerators only execute real arithmetic, so complex products must be rebuilt in software at a flat 3x or 4x cost. This paper's central proposal is a graph rewrite—realification—that converts any complex tensor network into an equivalent real tensor network by giving each complex tensor a single dimension-2 index for its real and imaginary parts and routing products through a fixed symmetric structure tensor. The paper proves a tight cost law: relative to the same network contracted over the reals, the realified contraction costs 1+2m+r times as many real multiplications, where m and r are the volume fractions of two-complex and one-complex contractions, so the overhead lies between 1x (all-real) and 3x (all-complex) and every intermediate's element count at most doubles. It also shows that contraction orders optimized on the complex network transfer to the realified network almost unchanged, and that the rewrite makes reverse-mode differentiation a standard real-valued chain rule. Benchmarks on a real-only accelerator show the realified executor beating both the four-real-GEMM and three-real-GEMM complex lowerings on 64 of 67 circuits, with median speedups of about 1.7x over the four-GEMM baseline.","feed_headline":"Complex tensor networks go real with at most 3x arithmetic","feed_subtitle":"Proof and benchmarks: overhead 1+2m+r beats flat 4-real-GEMM lowering on real-only hardware.","key_machinery":"The structure tensor C—the 2x2x2 real tensor encoding the multiplication table of C as a real *-algebra—is the load-bearing object. Contracting the green legs of two realified operands through C (plus a sign flip Z) reproduces complex multiplication with exactly three real multiplications, the tensor form of Gauss's 3M algorithm; its permutation symmetry, conjugate covariance, unit rule, and cascade rule make the wiring a gauge choice, so any contraction tree works. The cost law is carried by classifying each pairwise contraction as merge, ride, or pass with volume fractions m and r, yielding overhead 1+2m+r and a 2x bound on every intermediate.","core_discovery":"The paper proves that any complex tensor network—the workhorse of quantum circuit simulation—can be rewritten as a real tensor network by a local graph rewrite, with no loss of contraction-order freedom. Each complex tensor gains a dimension-2 index for its real and imaginary parts; a fixed symmetric tensor C realizes complex multiplication in three real multiplications. Pairwise steps cost 3x (merge), 2x (ride), or 1x (pass), so by Theorem 4 total overhead is 1+2m+r, between 1x and 3x, with intermediates at most doubled. Complex-optimal contraction orders transfer within 5e-4 on 66 of 67 circuits; on a real-only NPU the rewrite beat four-real-GEMM and per-GEMM Gauss lowerings on 64 of 67 ci","pith_inferences":["The cost law should extend unchanged to any complex tensor network—expectation values of projected entangled pair states, complex linear algebra, and similar settings—so real-only hardware could serve a wider range of scientific computing; benchmarking that is a natural next step the paper does not take.","The placement-versus-count result suggests a practical compilation rule: on real-only hardware, schedule or compile complex gates into contiguous spatial clusters so contraction trees can route most volume through cheaper rides and passes.","The contrast with circuit-level realification—whose frozen ancilla worldline can inflate intermediates exponentially—points to a general design principle: auxiliary algebraic bookkeeping should be aligned with the contraction tree, not with gate order; testing this on other algebras such as quaternions or Clifford algebras would be a natural extension.","The paper's open conjecture that the worst-case conversion gap is strictly below 3x could be settled by explicit constructions; if the gap can reach 3x, green-sparse networks would need a lightweight green-aware polish rather than convert-only deployment."],"forward_implications":["Real-only matrix engines (NPUs, TPUs, GPU tensor cores) can execute complex tensor-network contraction with arithmetic overhead never above 3x the real skeleton, instead of a flat 4x textbook penalty.","Partially real circuits pay less: overhead is set by where complex gates sit in the contraction tree, not by their count—spatially clustered complex gates can cut overhead from roughly 2.9x to 2.0x at fixed complex-gate count.","Contraction-order optimizers written for complex networks can be reused without reoptimization: convert-only plans match full green-aware annealing to within 5e-4 on 66 of 67 circuits.","Real-valued objectives computed from realified networks differentiate via the ordinary real chain rule, removing the need for complex Wirtinger calculus in reverse-mode automatic differentiation of multilinear contraction.","The network-level rewrite, not the 3M arithmetic identity alone, is what converts the arithmetic saving into wall-clock speedup: a per-GEMM Gauss lowering stayed within a few percent of the 4M baseline, while the embedded rewrite gave median ~1.7x speedups."],"supporting_citations":[{"why":"supplies the *-algebra realification and the structure tensor C that the paper compiles into a contraction-level method.","marker":"[18, 19]"},{"why":"proves that three real multiplications are necessary for complex multiplication, making the 3x merge cost tight.","marker":"[11]"},{"why":"establishes the numerical stability of Gauss's 3M complex product, justifying its practical use in the rewrite.","marker":"[12]"},{"why":"provides the high-performance 3M/4M complex-GEMM implementations that are the baseline lowerings the paper improves on.","marker":"[13]"},{"why":"founded tensor-network contraction as a scalable method for quantum circuit simulation, the application domain.","marker":"[1]"},{"why":"supplies the twelve random circuit benchmark networks used for validation.","marker":"[2]"},{"why":"introduces circuit-level realification via an ancilla worldline, the contrast that motivates tree-aligned realification.","marker":"[14–17]"}],"fun_headline_variants":["Realified tensor networks: quantum circuits on real-only hardware","Quantum circuits on real matrix engines: ≤3x overhead","Rewrite maps complex tensor networks to real ones, ≤3x cost","Realification: complex quantum networks run on real GEMMs","Beat 4-real-GEMM: realified tensor networks ≤3x"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The measured wall-clock advantage assumes the compared executors differ only in their complex-multiplication strategy; if the baseline is slower for unrelated execution-path reasons, the speedup is not entirely due to the rewrite.","fun_headline_variants_meta":{"raw":{"variants":["Realified tensor networks: quantum circuits on real-only hardware","Quantum circuits on real matrix engines: ≤3x overhead","Rewrite maps complex tensor networks to real ones, ≤3x cost","Realification: complex quantum networks run on real GEMMs","Beat 4-real-GEMM: realified tensor networks ≤3x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000732,"raw_usage":{"total_tokens":3181,"prompt_tokens":881,"completion_tokens":2300,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":2212}},"tokens_in":625,"tokens_out":2300,"duration_ms":18821,"temperature":1.0,"reasoning_tokens":2212,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T04:29:22.236420+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an all-real benchmark cell (m=r=0) where all three executors perform the identical number of multiplications; if the realified executor's wall-clock time is not close to the four-GEMM baseline's, executor-level effects—not the arithmetic rewrite—explain part of the speedup. A stronger test: hand-fuse a per-GEMM 3M kernel that avoids separate launches and combination passes; if it matches network-3M's wall-clock time on the same contraction plans, the claimed structural benefit of the network-level rewrite is not structural.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes the numerical stability of Gauss's 3M complex product, justifying its practical use in the rewrite."},{"cited_title":"Van Zee and Tyler M","cited_arxiv_id":null,"evidence_quote":"provides the high-performance 3M/4M complex-GEMM implementations that are the baseline lowerings the paper improves on."},{"cited_title":"A flexible high-performance simulator for verifying and benchmarking quantum circuits implemented on real hardware.npj Quantum Information, 5:86, 2019","cited_arxiv_id":null,"evidence_quote":"supplies the twelve random circuit benchmark networks used for validation."}],"review_version":2}