{"id":"4bc40e45-66b1-47ea-8220-6b79b076f3e6","arxiv_id":"2603.18699","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"New rational 48-multiplication 4×4 matrix product achieves error-growth exponent 2.386 and a 12.09 n^2.79 straight-line program.","lead":"This paper gives a new rational algorithm that multiplies two 4×4 matrices using 48 non-commutative multiplications, with better worst-case rounding-error growth than earlier fast algorithms. It also provides an explicit program whose leading cost is about 12.09 n^2.79 operations, of interest to numerical linear algebra and high-performance computing.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's LRP identity is unverified; if false, all claims collapse","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the exactness of the LRP decomposition is asserted, not demonstrated. This is the single most serious issue because the entire paper's contribution—an 'accurate' algorithm—presupposes that the algorithm computes the exact product in exact arithmetic; the error analysis then only bounds floating-point perturbations. Without the decomposition, the growth factor, complexity, and experiments are all meaningless. The abstract/body inconsistencies reinforce this reading but are secondary; they would be resolved by correcting the text. I agree with the reader's conditional verdict: the paper should be accepted only after the decomposition identity is verified, either by a proof sketch or by an independent computational check such as the one proposed. The concrete test is decisive and easy to run, so the condition is reasonable.","tokens_in":34760,"tokens_out":3133,"duration_ms":36029,"concrete_test":"Use exact rational arithmetic (e.g., SageMath or Mathematica) to verify the tensor identity directly from Table 1: for each output index (i,j) with 1≤i,j≤4, expand P·((L·vec(A))⊙(R·vec(B))) as a bilinear form in the entries of A and B, and confirm that the coefficient of A_{i,k}B_{k,j} is 1 for each k, and all other coefficients are 0. Equivalently, check the 16×16 coefficient matrix equals the standard matrix multiplication tensor. If any monomial has a nonzero coefficient not of this form, the decomposition is incorrect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is a new exact algorithm for ⟨4×4×4:48⟩ with growth factor γ∞,2 = 27.314 and exponent 2.386. Every numerical claim—Table 2, the SLP operation counts, and the benchmarks—depends on the matrices L, R, P in Table 1 actually computing the 4×4 matrix product exactly. Yet the manuscript never proves or independently checks the identity P·((L·vec(A))⊙(R·vec(B))) = vec(AB). Section 2.2 simply states the variant is 'defined by the lrp representation given in Table 1'; Section 3 offers Listings 1–4 said to be 'obtained from' those matrices, but again no verification is provided. The only external support is a pointer to PLinOpt data files in Appendix B. If any entry in L, R, or P is mistyped, or if the listings do not realize the intended linear forms, then the reported γ, exponent, operation count, and experimental rankings are all vacuous—they describe an algorithm that may not exist. This is an internal gap, not a matter of convention. The abstract/body numeric discrepancies (2.335 vs 2.386; 316/32 vs 387/32) further signal that the manuscript was edited without a full re-check of the central object, making the missing verification more salient.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new rational non-commutative algorithm for multiplying 4×4 matrices with 48 non-scalar multiplications, represented as an exact (L, R, P) tensor decomposition (Table 1) and as a straight-line program (Listings 1–4). The central claims are that this variant achieves growth factor γ∞,2 = 27.314, error-bound exponent log_4 γ∞,2 ≈ 2.386 for the (∞,2)-norm, and a leading constant 387/32 n^{2+log_4 3} in its complexity bound. The paper compares these values with prior ⟨2×2×2:7⟩ and ⟨4×4×4:48⟩ algorithms, gives an alternative-basis variant, and reports numerical experiments suggesting improved max-norm accuracy.","tokens_in":35062,"tokens_out":6016,"duration_ms":56616,"significance":"If the decomposition and all reported constants are correct, the paper gives the most accurate currently known sub-cubic matrix multiplication scheme in the (∞,2)-norm setting, improving the asymptotic error exponent from about 2.628 to 2.386 while maintaining a reasonable operation count. The work is constructive: it provides explicit rational matrices, a concrete SLP, public data files via PLinOpt, and reproducible operation counts. These are real strengths. However, the central algebraic identity is not verified in the manuscript, and the abstract disagrees with the body on both headline numbers; the contribution cannot be accepted in its present form without resolving these issues.","major_comments":[{"comment":"The central claim is that the L, R, P matrices in Table 1 form an exact rank-48 decomposition of the 4×4 matrix multiplication tensor, and that Listings 1–4 realize the same linear forms. The paper never proves or independently checks either identity. Section 2.2 only says the variant is ‘defined by the lrp representation given in Table 1’; Section 3 only says the SLPs are ‘obtained from’ those matrices. Appendix B points to PLinOpt data files, but a reader cannot verify correctness from the manuscript alone. Since every numerical result (γ∞,2, Table 2, Eq. (6), Section 4) depends on this decomposition being exact, the missing verification is load-bearing. Please include a machine-checkable certificate — for example a short script that checks P·((L·vec(A))⊙(R·vec(B))) = vec(AB) for all basis matrices, or a hash/transcript of such a verification.","section":"§2.2, Table 1, Listings 1–4, Appendix B"},{"comment":"The abstract reports an error-bound exponent of log_4 γ∞,2 ≈ 2.335 and a leading constant of 316/32 n^{2+log_4 3}; the body reports log_4 γ∞,2 ≈ 2.386 (Section 1 and Table 2) and 387/32 n^{2+log_4 3} (Section 1 and Eq. (6)). These are numerically very different claims. The operation counts in Section 3 (104+88+129 additions and 16+18 shifts = 355 operations) give 1 + 355/(48−16) = 387/32, so the abstract’s 316/32 is not consistent with the presented SLP. The exponent 2.386 is consistent with γ∞,2 = 27.314 in Table 2, while 2.335 would correspond to a different γ. This discrepancy must be resolved and the correct values used consistently in the final version.","section":"Abstract vs. §1, §2.3, Eq. (6)"},{"comment":"The experiments report max-norm output error but do not specify the input norm, while the theoretical ranking is made with (p,q)=(∞,2). The text says the ranking ‘follows the one indicated by the second column of Table 2 (for (p,q)=(∞,2))’ and then says ‘regardless of any norm on the input (parameter q)’. This is not a formal comparison. Please state precisely what is measured (e.g., normalized random matrices, fixed q, number of trials) and how that relates to the (∞,2) bound. Otherwise the experimental section is only anecdotal, and the text should not claim it confirms the theoretical ranking.","section":"§4 / Table 2"}],"minor_comments":[{"comment":"The y-axis has tick values but no label; the caption says ‘normal distribution’ but does not specify the distribution parameters, sample size, or whether multiple runs were averaged. Please add these details.","section":"Figure 1"},{"comment":"The text in Section 2.2 uses ‘lpr representation’ while the rest of the paper uses ‘LRP’ or ‘lrp’. Please standardize the terminology.","section":"§2.2"},{"comment":"Table 2 labels the row for the previous rational ⟨4×4×4:48⟩ as ‘DPS25’ but the companion text in Section 1 attributes the rational variant to reference [8]. Please check that the citation in Table 2 is the intended one, and distinguish clearly between [8] (the 2025 arXiv report) and [9] (the journal article).","section":"References"},{"comment":"The alternative-basis matrices Lalt, Ralt, Palt and the cobasis matrices are extremely large and are not accompanied by a verification script. If the data files in Appendix B are the only source, this should be stated explicitly and the identity L = Lalt·Lcob, R = Ralt·Rcob, P = Pcob·Palt should be checked by the same certificate as the main decomposition.","section":"Appendix A and Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"This is a short, computer-assisted paper whose main contribution could be verified very quickly with a small script. The unverified central identity and the abstract/body numerical contradictions are substantive but fixable. I would not reject on the basis of the missing proof alone, provided the authors supply a machine-checkable verification of the LRP identity and reconcile the two sets of headline numbers. If those are resolved, the paper would be a useful, incremental contribution to the accuracy analysis of fast matrix multiplication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, it does appear to set a new best for the (∞,2) accuracy exponent among sub-cubic matrix multiplication algorithms, with log_4 γ∞,2 ≈ 2.386 and a concrete straight-line program. Second, the paper never verifies the central L, R, P decomposition, and the abstract disagrees with the body on both headline numbers. That combination makes me treat the results as plausible but not yet established.\n\nWhat's genuinely new: the specific L, R, P triple in Table 1. It's not in the authors' earlier papers [8] or [9], and it gives γ∞,2 = 27.314, down from 38.163, and γ2,2 = 109.26, down from 139.906. The operation counts in Section 3 are internally consistent: 104 additions for L, 88 + 16 shifts for R, 129 + 18 shifts for P, total 355, leading to the stated constant 387/32 n^{2+log4 3}. The γ values in Table 2 are consistent with Eq. (4). The paper also points to PLinOpt data files for reproducibility. That's real work.\n\nNow the soft spots, in proportion. The big one is the unverified identity. The manuscript asserts that the matrices in Table 1 and the listings of Section 3 compute the 4×4 product over Z[1/2], but gives no symbolic check, no short proof, and only a pointer to external data files. If any entry in L, R, or P is mistyped, the growth factor, the exponent, the SLP counts, and the experiments all describe a phantom. This is load-bearing, not a technicality. The stress-test note is right about that.\n\nThe second issue is the abstract/body discrepancy: the abstract says 2.335 and 316/32, the body says 2.386 and 387/32. The body numbers are internally consistent with the stated γ and operation counts, so the abstract looks like a leftover from an earlier run, but it undermines confidence in the manuscript's editorial care. Third, Figure 1 has no error bars and it's not clear how many trials, but that's minor for a quick benchmark.\n\nThe reader's conditional verdict is fair. The central argument is believable and the arithmetic checks out, but the missing verification is a real gap, not a manufactured one. If I were refereeing, I'd ask the authors to add a short symbolic verification of the L,R,P identity (or a script that checks it) and to reconcile the abstract with the body. The paper deserves serious peer review: the claim is important enough in its niche and it's checkable. I'd engage with it, but only after those fixes.\n\nRecommendation: send to a referee, but condition acceptance on the verification and the abstract cleanup.","headline":"Plausible new best accuracy exponent for sub-cubic 4×4×4:48, but the central L,R,P identity is unverified and the abstract contradicts the body on both headline numbers.","tokens_in":35549,"tokens_out":4598,"would_cite":false,"duration_ms":43429,"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 new 4×4 matrix multiplication scheme uses 48 multiplications and achieves an error-bound exponent of about 2.386, the sharpest among compared sub-cubic algorithms.","keywords":["4×4 matrix multiplication","48 multiplications","non-commutative algorithm","numerical accuracy","growth factor","error-bound exponent","straight-line program","sub-cubic complexity"],"falsifier":"Symbolically verify the identity for arbitrary 4×4 matrices, or evaluate both sides at random integer matrices over, say, the rationals or a finite field; any mismatch disproves the algorithm. Separately recompute γ∞,2 from the published L, R, P and confirm that log_4(27.314) ≈ 2.386.","tokens_in":34645,"feed_emoji":"🧮","tokens_out":4412,"duration_ms":45566,"temperature":0.7,"pith_summary":"The paper presents a rational, non-commutative algorithm for multiplying 4×4 matrices with 48 multiplications, valid over any ring containing an inverse of 2. Its central claim is improved numerical accuracy: the (∞,2)-growth factor is 27.314, giving an error-bound exponent log_4(27.314) ≈ 2.386, lower than Strassen's, Winograd's, and earlier 48-multiplication variants. The authors also give a concrete straight-line program with a leading complexity constant of about 12.09 n^2.792, plus an alternative-basis version with constant 8 n^2.792. If the decomposition is exact, this becomes the most accurate currently known sub-cubic matrix multiplication scheme in the 2-norm input setting.","feed_headline":"New 48-product 4×4 matmul cuts error drift to exponent 2.386","feed_subtitle":"Rational decomposition beats Strassen, Winograd and earlier 4×4 fast schemes on max-norm accuracy — at sub-cubic cost.","key_machinery":"The LRP representation: a triple of matrices L, R, P such that vec(AB) = P·((L·vec A) ⊙ (R·vec B)), where ⊙ is the Hadamard product and the 48 intermediate products are the non-commutative multiplications. Accuracy is governed by the (p,q)-growth factor γ_{p,q} = ||(Σ_i ||L_i,*||_{q*} ||R_i,*||_{q*} |P_{j,i}|)_j||_p; its logarithm in base k gives the recursive error exponent. The new variant was selected using a smoothed relaxation γ_2 ≥ γ_{2,2}, which the paper argues better predicts max-norm behavior in practice.","core_discovery":"We exhibit an LRP decomposition of the 4×4×4 matrix-multiplication tensor using 48 rank-one terms, over Z[1/2], whose (∞,2)-growth factor is γ∞,2 = 27.314. This lowers the recursive error-bound exponent from roughly 2.628 (previous rational 4×4×4:48) and 2.577 (best 2×2×2:7 variant) to about 2.386. A straight-line realization uses 355 additions/binary shifts plus 48 multiplications, with leading term (387/32)n^{2+log_4 3}; an alternative-basis variant achieves leading constant 8. Experiments on normally distributed floats show max-norm error close to the conventional cubic algorithm and better than Strassen, Winograd, and the prior 48-multiplication scheme.","pith_inferences":["The 'most accurate sub-cubic algorithm' claim is norm-dependent: for max-norm on the inputs, Strassen's algorithm still holds the best bound; the improvement here applies to the 2-norm input regime.","The abstract quotes the exponent as about 2.335 while the body and Table 2 give 2.386 for the same quantity; the manuscript does not resolve this discrepancy, so comparisons should verify the intended value.","The smoothed relaxation γ_2 used for search is not the true γ_{2,2}; similar relaxations could be applied to other tensor decompositions, and the gap between relaxation and exact growth factor is an open handle for further accuracy tuning.","Because the coefficients are rational with powers of 1/2, the algorithm is likely to be usable in exact and modular arithmetic, though coefficient growth during recursion would need separate testing."],"forward_implications":["Recursive use of this 4×4 block gives an error bound growing like n^2.386, the best among the compared sub-cubic algorithms when inputs are measured in 2-norm.","The concrete straight-line program lets implementers reproduce the algorithm without re-deriving the decomposition.","The claimed accuracy carries over to any ring containing 1/2, including exact rational arithmetic and finite fields of odd characteristic.","The alternative-basis variant reaches a leading complexity constant of 8 n^2.792 while keeping similar accuracy, improving the practical constant over the main variant.","The experiments indicate max-norm output error close to that of the classical cubic algorithm, despite using only 48 scalar products per 4×4 block."],"fun_headline_variants":["4x4 matmul with 48 products: record accuracy, low error drift","Better than Strassen: new 48-multiplication 4x4 algorithm","Rational 4x4 matmul hits exponent 2.386, beating prior fast schemes","New 48-mult 4x4 matmul: accuracy near cubic, cost sub-cubic","Sharpest fast 4x4 matrix multiply yet: error exponent 2.386"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The printed L, R, P tables are asserted to exactly satisfy P·((L·vec A) ⊙ (R·vec B)) = vec(AB) for every 4×4 pair A, B, but the paper gives no algebraic proof of this identity in the text.","fun_headline_variants_meta":{"raw":{"variants":["4x4 matmul with 48 products: record accuracy, low error drift","Better than Strassen: new 48-multiplication 4x4 algorithm","Rational 4x4 matmul hits exponent 2.386, beating prior fast schemes","New 48-mult 4x4 matmul: accuracy near cubic, cost sub-cubic","Sharpest fast 4x4 matrix multiply yet: error exponent 2.386"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":2792,"prompt_tokens":703,"completion_tokens":2089,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":1976}},"tokens_in":447,"tokens_out":2089,"duration_ms":14828,"temperature":1.0,"reasoning_tokens":1976,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T17:51:29.080845+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Symbolically verify the identity for arbitrary 4×4 matrices, or evaluate both sides at random integer matrices over, say, the rationals or a finite field; any mismatch disproves the algorithm. Separately recompute γ∞,2 from the published L, R, P and confirm that log_4(27.314) ≈ 2.386.","supporting_citations":[],"review_version":1}