{"id":"580dd506-ff32-4aef-b2e8-3f4c0a5b35f5","arxiv_id":"2610.28627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Rafael Pass gives a constructive proof that the matrix multiplication exponent omega is at most 9/4 using polynomial packings and ordinary block recursion.","lead":"An explicit proof claims that two n by n matrices can be multiplied using about n^2.25 operations, much faster than the current best exponent near 2.371. The proof is self-contained and works over ordinary integers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Section 4.1's one-round recurrence sets t=floor((3L-1)/(c-1)) and asserts (c-1)t≥3L-1; this is false, so the second packing step and the bound L'≤ρL+1 fail for generic L. A ceiling and ρ=1+3/(c-1) repair the recurrence with the same 9/4 asymptotics, so the verdict stays conditional.","rationale":"The reader and I isolate the same weak point: Section 4.1's one-round argument. The paper's central theorem cannot be accepted on the written proof because every recursive round passes through the false inequality (c-1)t≥3L-1, and the counterexamples are small and explicit. I do not see a separate obstruction in the polynomial-packing lemmas or in Bini's exactification; the paper gives enough proof detail there to make the main claim plausible pending repair. The suggested correction is natural and preserves the advertised 9/4, which supports a conditional rather than a reject verdict. Since the reader already assigned CONDITIONAL and identified this same assumption, my stress-test pass does not change the verdict. The reader's illustrative numbers for c=4, L=4 contain a small arithmetic slip (the floor is 3, not 1), but the structural point and the proposed ceiling repair are correct.","tokens_in":18313,"tokens_out":19161,"duration_ms":112743,"concrete_test":"Instantiate the one round of Section 4.1 at c=4, L=3 exactly as written with t=floor(8/3)=2 and L'=5, and check whether the second packing precondition ct≥L'+2L-1 holds (8≥10 fails), confirming the concern. Then rerun with t=ceil(8/3)=3 and ρ=1+3/(c-1): all packing preconditions hold, L+t≤ρL+1 becomes 6≤6, and log(64ρ^{6d})/log R remains 9/4+O(1/c), showing the repair is sufficient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The finite identity in Theorem 4.1 is obtained by iterating the one-round reduction in Section 4.1, so that reduction is load-bearing. The paper sets t=floor((3L-1)/(c-1)), L'=L+t, and claims the packing preconditions hold both at (a,b)=(L,L) and at (L',L), since (c-1)t≥3L-1. For the second application, Lemma 2.4 and Theorem 2.9 with a=L', b=L require ct ≥ a+2b-1 = (L+t)+2L-1 = 3L-1+t, which is exactly (c-1)t ≥ 3L-1. A floor cannot give this: (c-1)floor(K/(c-1)) ≤ K, with strict inequality unless c-1 divides K. For c=4, L=3, t=2 and ct=8, while the second precondition demands 3L-1+t=10; for c=4, L=2, even the first precondition ct≥5 fails because ct=4. The claimed length bound L'≤(1+3/c)L+1 is also false in general: c=4, L=10 gives t=9, L'=19 > 18.5. The defect is confined to this recurrence; the packing lemmas and Bini conversion are not implicated. Using t=ceil((3L-1)/(c-1)) and ρ=1+3/(c-1) restores (c-1)t≥3L-1 and L'≤ρL+1, and the exponent bound remains 9/4+O(1/c), so the flaw is repairable. But as written, Theorem 4.1 does not follow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper claims a self-contained constructive proof that the matrix multiplication exponent satisfies omega <= 9/4. It develops a theory of polynomial packings: a same-left primal packing that computes c+1 full polynomial products and one middle product from c source products modulo a formal parameter epsilon, together with a dual packing with reversed roles. These univariate gadgets are combined across 2d variables to amplify the compression, then converted from same-left targets to independent left inputs by an interpolation argument, and finally turned through role exchanges into reductions of the row, inner, and column dimensions simultaneously. Section 4 iterates a one-round recurrence to construct, for every tau > 9/4, a finite rational Strassen-style identity AB = sum_nu l_nu(A) m_nu(B) W_nu with fewer than v^tau scalar products; ordinary block recursion over this identity yields the integer-arithmetic algorithm stated in Theorem 1.1. The paper includes proofs of composition, transposition, and Bini's approximate-to-exact conversion.","tokens_in":18874,"tokens_out":20349,"duration_ms":120370,"significance":"Assuming the construction is correct, the result is significant: it would give an explicit identity and a recursive algorithm for omega <= 9/4, a bound previously claimed only through non-constructive asymptotic spectral arguments. The paper is unusually self-contained: it proves the closure properties it uses, and it reproves (as the c=1 case) the OpenAI gadget rather than relying on it. The exponent 9/4 is not fitted; it emerges from asymptotic parameter choices. However, the one-round recurrence in Section 4.1 contains a false inequality on which the entire identity construction depends. The error is localized and has an obvious repair, but until the recurrence is corrected and the proof re-verified, the central theorem is not established.","major_comments":[{"comment":"The proof sets t = floor((3L-1)/(c-1)) and asserts that the packing preconditions ct >= a+2b-1 hold at (a,b)=(L,L) and at (L',L) because (c-1)t >= 3L-1. This inequality is false for a floor: (c-1)t <= 3L-1, with equality only when c-1 divides 3L-1. For example, c=4 and L=1 gives t=0 and ct=0 < 2, so even the first application of Theorem 2.9 is invalid; for c=4 and L=3, the second application demands ct >= 3L-1+t = 10 while ct=8. Since the one-round reduction is the mechanism that grows the length parameter and shrinks the matrix side, the iteration producing Theorem 4.1 does not follow as written. Replacing the floor by a ceiling restores (c-1)t >= 3L-1 and hence both packing preconditions, so the defect is repairable, but it is load-bearing.","section":"Section 4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The displayed length bound L' <= rho L + 1 is not justified by the displayed parameter choices. With t = floor((3L-1)/(c-1)) and rho = 1 + 3/c, the example c=4, L=10 gives t=9 and L'=19, while (1+3/c)L+1 = 18.5. The bound that is actually needed, and that does hold, is L' <= (1 + 3/(c-1))L + 1, obtained with t = ceil((3L-1)/(c-1)). With these corrected choices the subsequent exponent calculation log(64 rho^{6d})/log R = 9/4 + O(1/c) is unchanged, so the issue is local; but the authors need to state the corrected pair explicitly and re-verify the iteration.","section":"Section 4.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The word 'nexy' should be 'next'.","section":"Section 1.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The expression for rho is typeset ambiguously; it should be stated explicitly as rho = 1 + 3/(c-1) once the ceiling repair is made.","section":"Section 4.1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"With d = c^3, the error term in log(w/u) is actually O(log c), not O(c); this is harmless for the stated asymptotic conclusion but should be corrected.","section":"Equation (5)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The claim that polynomial multiplication of length L can be computed with O(L) scalar products via evaluation and interpolation is correct for exact rational arithmetic, but the evaluation points may require large representation; a sentence clarifying that the complexity count is algebraic would improve readability.","section":"Lemma 3.3","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Read this one. It is the first genuinely constructive shot at ω ≤ 9/4: an explicit rational identity plus integer-arithmetic recursion, standing on ordinary algebra rather than the OpenAI asymptotic-spectrum machinery. The multivariate packing amplification across 2d variables and the interpolation-based left-input separation (Lemma 2.8) are new, and the proof is self-contained — the three-for-one gadget is reproved as the c=1 case, and composition, transposition, and Bini conversion are proved in the text, not cited away. I checked the exponent count: d=c^3 gives log(w/u)=4c^2+O(c), so the per-round growth log(64ρ^{6d})/log R lands at 9/4+O(1/c). There is no fitting anywhere; the 9/4 falls out of the algebra. The galactic-constant caveat is stated honestly, and the AI-use disclosure is unusually candid, including a footnote attributing the separation encoding idea to Codex.\n\nThe soft spot is real, and it is exactly where the reader put it. In the proof of Theorem 4.1, t is set to floor((3L−1)/(c−1)) and the text claims (c−1)t ≥ 3L−1. A floor gives the opposite inequality, with equality only when c−1 divides 3L−1. Example: c=4, L=3 gives t=2, and the second packing application needs ct ≥ L′+2L−1 = 3L−1+t = 10, but ct=8. So the precondition of Theorem 2.9 fails at (L′,L), and the round, as written, is broken. This is load-bearing — Theorem 4.1's identity comes from iterating this round. The stress-test's repair is right: replace the floor with a ceiling; then (c−1)t ≥ 3L−1 holds and L′ ≤ (1+3/(c−1))L + O(1) preserves the same asymptotics. It is a one-line fix, and the architecture is intact.\n\nThe length-bound complaint is secondary. The rho in the text is garbled: if it means 1+3/(c−1), the floor version of L′ ≤ ρL+1 is fine; if it means 1+3/c, it fails. Either way, the floor inequality is the issue that matters.\n\nMinor caveat: no code, no machine-checked proof, and the identity is shown computable rather than exhibited. Acceptable for this genre, but it means the argument rests on human-checkable algebra.\n\nWho it is for: the matrix multiplication crowd, and anyone tracking the OpenAI 9/4 claim who wants a constructive witness. It deserves a serious referee; the claim is major and the defect is localized. My recommendation: send it to review with a specific request to check Section 4.1 recurrences, and expect a conditional accept with the ceiling fix. I would bring it to reading group.","headline":"First genuinely constructive route to ω ≤ 9/4 with a novel gadget architecture, stalled by a false floor inequality in Section 4.1 that a ceiling repairs while preserving the bound.","tokens_in":19372,"tokens_out":6466,"duration_ms":37867},"pith_extraction":{"msc":["68Q25","68Q17","15A69"],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"The paper claims a constructive proof that for every $\\gamma>0$ there is an explicit recursive algorithm multiplying $n\\times n$ integer matrices in $O(n^{9/4+\\gamma})$ arithmetic operations, establishing $\\omega \\le 9/4$.","keywords":["matrix multiplication exponent","omega <= 9/4","polynomial packing","bilinear complexity","Strassen identity","block recursion","approximate algorithm","integer arithmetic"],"falsifier":"Evaluate the Section 4.1 round at explicit parameters: for $c=4$ and $L=9$, $t = \\lfloor 26/3\\rfloor = 8$, $L' = 17$, while the claimed bound $(1+3/c)L+1 = 16.75$ falls short, and the stated justification $(c-1)t = 24 < 26 = 3L-1$ is also false even though the packing precondition $ct = 32 \\ge 26$ happens to hold; checking whether the corrected constants (ceiling, $\\rho = 1+3/(c-1)$) still give $\\log(64\\rho^{6d})/\\log R \\to 9/4$ settles whether the exponent survives.","tokens_in":19702,"feed_emoji":"🧮","tokens_out":10903,"duration_ms":58995,"temperature":0.7,"pith_summary":"The paper claims a proof that the matrix multiplication exponent satisfies $\\omega \\le 9/4$, by explicit construction rather than by the non-constructive spectral argument of a recent preprint claiming the same bound. The core object is a 'polynomial packing': a telescoping gadget that recovers $c+2$ polynomial products of matrix coefficients from only $c$ source products, whose saving is amplified across many variables, converted to independent left inputs, and turned into a fixed identity $AB = \\sum_\\nu \\ell_\\nu(A)m_\\nu(B)W_\\nu$ with $\\log_v r$ arbitrarily close to $9/4$. Ordinary block recursion on that identity then gives the claimed recursive integer algorithm with $O(n^{9/4+\\gamma})$ operations for every $\\gamma>0$. The construction is self-contained and runs on integer arithmetic. The critical one-round length-growth inequality in Section 4.1, however, appears to rely on a false bound, so the $9/4$ asymptotics likely survive a corrected round but not the proof exactly as written.","feed_headline":"Explicit identity pushes matrix multiplication to n^2.25","feed_subtitle":"A packing gadget plus block recursion promises omega <= 9/4 on integer matrices.","key_machinery":"The central device is the polynomial packing (PP; same-left SL-PP), a list of linear encoders, source polynomial products, and linear decoders whose composed map matches the target modulo a formal parameter $\\varepsilon$. The load-bearing identity is the telescoping sum $W = \\sum_j z^{t(j-1)} A E_j = AP_1 - z^{ct}AP_{c+1} + \\varepsilon z^b A Q$, in which all intermediate products cancel, the endpoints occupy disjoint coefficient ranges, and the shifted $\\varepsilon AQ$ fills the empty gap to yield a 'middle product', giving $c+2$ outputs from $c$ sources. Amplification comes from running primal packings in the first $d$ variables and dual packings in the last $d$ of $2d$ variables, gaining a multiplicative factor $((c+2)/c)^{2d}$, from separating left inputs via polynomial interpolation, and from a three-role exchange that reduces rows, inner dimension, and columns simultaneously. Bini's approximate-to-exact lemma converts the $\\varepsilon$-packing into the exact identity (6).","core_discovery":"On its own terms, the paper proves Theorem 1.1: for every $\\gamma>0$, an explicit recursive algorithm multiplies two $n\\times n$ integer matrices in $O(n^{9/4+\\gamma})$ integer arithmetic operations, so $\\omega \\le 9/4$. The route is elementary: a $(c+2)$-for-$c$ 'polynomial packing' modulo $\\varepsilon$ recovers $c+1$ full products and one middle product from $c$ source products via the telescoping sum $W = \\sum_j z^{t(j-1)}AE_j = AP_1 + \\varepsilon z^b AQ - z^{ct}AP_{c+1}$, where intermediate terms cancel and the shifted $\\varepsilon$-block fills an empty coefficient gap. Primal and dual packings are combined across $2d$ variables to amplify the output count by $((c+2)/c)^{2d}$, left inputs are separated by interpolation, all three matrix dimensions are reduced at once, and after finitely many rounds Bini's approximate-to-exact conversion yields an exact rational identity $AB = \\sum_\\nu \\ell_\\nu(A)m_\\nu(B)W_\\nu$ with $\\log_v r$ arbitrarily close to $9/4$; ordinary block recursion on this identity finishes the proof.","pith_inferences":["My reading of Section 4.1 is that the false length bound is a repairable gap: replacing the floor by a ceiling and $\\rho = 1+3/c$ by $1+3/(c-1)$ preserves the asymptotic $9/4$, so the plausible outcome is a corrected proof rather than a dead claim, and a careful reader should demand the corrected round before trusting the conclusion.","If the packing chain is sound, tuning $c$, $d$, and the terminal evaluation bound may push the constant toward or below $9/4$ or yield rectangular multiplication bounds, since the $(c+2)/c$ gain and the $d+1$ shape-selection loss are the only free parameters.","The exactness of the final identity rests on the terminal bound $O(L^{6d})$ matrix products and on the degree $D$ in Bini's conversion; computing these constants for small $c,d$ would give a cheap empirical test of the whole construction.","A natural testable extension is to run the $(c+2)$-for-$c$ packing with the packing precondition $ct \\ge a+2b-1$ enforced exactly (using a ceiling), and measure the resulting exponent for small $d$ to see how close to $9/4$ the finite identity actually lands."],"forward_implications":["For every $\\gamma>0$ the construction yields an explicit recursive algorithm computing $n\\times n$ integer matrix products in $O(n^{9/4+\\gamma})$ operations, so $\\omega \\le 9/4$.","The bound is certified by a concrete finite identity $AB = \\sum_\\nu \\ell_\\nu(A)m_\\nu(B)W_\\nu$ with $\\log_v r < 9/4 + \\gamma$, making the exponent checkable by direct verification of the identity rather than by a spectral existence argument.","The algorithm runs on integer arithmetic with bit cost $O(n^{9/4+\\gamma}M(B + \\log n))$, unlike the non-constructive complex-scalar statement it answers.","Because the constants are 'galactic' (enormous, as the authors state), the result is a structural bound, not a practical algorithm."],"supporting_citations":[{"why":"Supplies the three-for-one polynomial degeneration (Lemma 4.2) that the paper generalizes to a (c+2)-for-c packing, and the non-constructive omega <= 9/4 claim this work makes constructive.","marker":"[Ope26a]"},{"why":"Provides the block-recursion scheme and the original omega <= log_2 7 bound that the final recursive step mirrors.","marker":"[Str69]"},{"why":"The approximate-to-exact (Bini) lemma that converts the epsilon-packing into the exact rational identity (6).","marker":"[Bin80]"},{"why":"Supplies the closure properties (composition, combining indices), evaluation-interpolation for the terminal bound, and the transposition conventions used throughout.","marker":"[BCS97]"},{"why":"The Reed-Muller-style separation encoding (Lemma 2.8) that gives each target its own left input.","marker":"[Sud02]"}],"fun_headline_variants":["Polynomial packings yield explicit omega <= 9/4 proof","Constructive identity sets matrix multiply exponent at 9/4","Explicit packing trick proves omega ≤ 9/4 for matrices","Elementary block recursion shows omega ≤ 9/4 constructively"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The one-round reduction in Section 4.1 assumes that $t = \\lfloor (3L-1)/(c-1)\\rfloor$ satisfies $(c-1)t \\ge 3L-1$ and that $L+t \\le (1+3/c)L+1$; the floor gives $(c-1)t \\le 3L-1$, and for large $L$ the length bound fails (for $c=4$, $L=9$, $L'=17 > 16.75 = (1+3/4)\\cdot 9+1$), so the written proof rests on a false inequality, although a ceiling and $\\rho = 1+3/(c-1)$ would restore the asymptotic $9/4$.","fun_headline_variants_meta":{"raw":{"variants":["Polynomial packings yield explicit omega <= 9/4 proof","Constructive identity sets matrix multiply exponent at 9/4","Explicit packing trick proves omega ≤ 9/4 for matrices","Elementary block recursion shows omega ≤ 9/4 constructively"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"cost":0.00021924,"is_byok":false,"cost_details":{"upstream_inference_cost":0.00021924,"upstream_inference_prompt_cost":0.00012656,"upstream_inference_completions_cost":9.268e-05},"total_tokens":1235,"prompt_tokens":904,"completion_tokens":331,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":0,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":256}},"tokens_in":904,"tokens_out":331,"cost_source":"openrouter_usage","duration_ms":2673,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":256,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-21T04:24:13.150745+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Evaluate the Section 4.1 round at explicit parameters: for $c=4$ and $L=9$, $t = \\lfloor 26/3\\rfloor = 8$, $L' = 17$, while the claimed bound $(1+3/c)L+1 = 16.75$ falls short, and the stated justification $(c-1)t = 24 < 26 = 3L-1$ is also false even though the packing precondition $ct = 32 \\ge 26$ happens to hold; checking whether the corrected constants (ceiling, $\\rho = 1+3/(c-1)$) still give $\\log(64\\rho^{6d})/\\log R \\to 9/4$ settles whether the exponent survives.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}