{"id":"153c915e-24e8-4b38-8001-5dd2177a8f36","arxiv_id":"2607.18745","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Choosing a pre-quantization contraction gauge by minimizing a derived product-error identity gives certified optimal diagonal folds and empirically validated gains for rotations and grouping, tested on quantized matrix products in an image classifier.","lead":"This paper studies how to pick a mathematical transformation applied to two matrices before rounding their entries to low-precision numbers, so that their rounded product stays close to the true product. It gives exact formulas for rounding error under a randomized rounding model and a fast, provably optimal way to choose rescaling factors; tests on a small image classifier match the theory.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Certified optimality holds only under the dither model; the RTN transfer that powers the headline gains rests on one small classifier, so the applied claim is not yet general.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: Theorem 4.2 certifies global optimality for the dither range law, not for deterministic RTN; the paper's headline error reductions are measured under RTN on one small classifier. The paper is honest about this gap (Remark 3.2, Section 11.1, Section 12) and provides a helpful explicit counterexample to universal transfer in Remark 9.3. My stress test did not find an internal inconsistency in the GP or identity theorems, and I agree with the reader's conditional posture: the mathematical core appears correct, the RTN transfer is plausible but not yet established beyond one compact vision model, and the artifact being private at review time further supports a condition on reproducibility and broader validation. Since the reader already set the verdict to CONDITIONAL, no adjustment is needed. I would therefore keep the verdict unchanged rather than escalate to REJECT or relax to ACCEPT.","tokens_in":43912,"tokens_out":18485,"duration_ms":159907,"concrete_test":"Run the Section 10.9 protocol on linear layers of a pretrained transformer with real token activations -- e.g., OPT-125M or Llama-2-7B -- at 4-bit RTN: calibration-split fitting of the GP fold and SmoothQuant-style grid, disjoint held-out activations, and per-product squared-relative-Frobenius error plus within-product Spearman rho between dither prediction and RTN error. If the median rho falls substantially below 0.9 or the GP fold no longer beats the grid on a clear majority of products, the RTN-transfer claim is not general and the applied conclusions should be scoped to the evaluated classifier family. If rho stays near 0.92+ and the GP wins on most products, the transfer concern is substantially settled for realistic transformer data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mathematical core is sound: Theorem 3.3 is an exact expansion under independent zero-mean entrywise errors, and Theorem 4.2 is a valid GP for the dither variance law v_G = c R_G^2, with scale-gauge compactification handled explicitly. The load-bearing gap is the transfer from that dither model to deterministic RTN. Section 11.1 and Remark 3.2 concede that RTN has input-dependent bias and inter-coordinate correlations, and Remark 9.3 gives an explicit pair (A = [1 1 1 0], B = I_4, H_4) on which the dither objective ranks H_4 above I_4 while INT8 RTN ranks I_4 above H_4. Thus ranking by the certified dither objective is not guaranteed to order realized RTN error, even at fixed quantization groups.\n\nThe only empirical bridge is Section 10.9: twelve internal products from one trained three-block width-64 ViT-like classifier on 8x8 handwritten digits. The median within-product Spearman correlations are 0.937 (8-bit) and 0.918 (4-bit), and the GP fold beats the SmoothQuant grid in geometric mean on these products. That is a single architecture, a single tiny image domain, and no language-model or larger-vision validation; Section 12 itself lists pretrained-language-model validation as an extension. Because the central applied claim -- that the GP fold gives 18-20% held-out RTN error reduction and is a reliable common objective under RTN -- depends on this transfer, the unvalidated transfer is the most load-bearing weakness. It does not undermine the exact identities or the dither-model certification, but it does mean the headline real-world claims are conditional on empirical transfer that could fail on other models or data regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies low-precision computation of C=AB when both factors are quantized. It derives an exact identity (Theorem 3.3) for the expected squared Frobenius product error under independent, zero-mean entrywise errors, valid for non-overloading subtractive dither and for independent stochastic rounding. Using the product-preserving equivalence AB=(AT)(T^{-1}B), it introduces contraction-gauge preconditioning. For positive diagonal gauges (folds), Theorem 4.2 shows that the domain-shared fold selection problem is a geometric program with a convex log-domain formulation, and Proposition 4.3 gives an LP test for identity-fold optimality. For rotations, partial rotations, and hierarchies, the paper derives computable selection statistics and upper bounds (Theorems 6.1, 7.1, 8.1–8.2). Quantizer-side results include asymmetric bit allocation (Theorem 9.1) and a clipped product-error identity (Theorem 9.4). The empirical section tests the dither-model predictions against deterministic RTN on synthetic problems and on twelve internal products from a trained three-block ViT-like image classifier, reporting median within-product rank correlations of 0.937 (8-bit) and 0.918 (4-bit) and geometric-mean held-out error reductions of 18.0% and 20.5% relative to the identity fold.","tokens_in":44328,"tokens_out":11174,"duration_ms":108058,"significance":"The mathematical core is a genuine contribution: an exact finite-dimensional accounting of simultaneous quantization errors, a globally optimal formulation for shared diagonal folds within the stated dither model, and a clean LP test for identity-fold optimality. The proofs in the appendix are careful, and the controlled experiments are designed to isolate each mechanism. The main value of the paper, if it holds, is a common objective and decision statistics for comparing gauge families under stochastic quantization. However, the headline applied claims concern deterministic RTN, and those claims rest on an empirical transfer from the dither model that is validated on only a single small image classifier. The paper is honest about this gap in Remark 9.3, Section 11.1, and Section 12, but the abstract and conclusion present the RTN transfer as a working feature of the framework. The soundness of the core theorems is not in question; the breadth of the applied claim is.","major_comments":[{"comment":"The certified results (Theorem 4.2 and Proposition 4.3) are proved for the dither variance law v_G = c R_G^2. Deterministic RTN is explicitly outside the model: Section 11.1 notes input-dependent bias and inter-coordinate correlations, and Remark 9.3 gives an explicit pair (A=[1 1 1 0], B=I_4, H_4) on which the dither objective ranks H_4 above I_4 while INT8 RTN ranks I_4 above H_4. Therefore the abstract's statement that the GP fold provides a 'common objective for evaluating reusable transform candidates under RTN' is an empirical hypothesis, not a consequence of the certified optimum. The only transfer evidence is Section 10.9: twelve internal products from one trained three-block width-64 ViT-like classifier on 8x8 digit images. This is too narrow to support the general applied claim. Please either broaden the validation (e.g., pretrained language models, as already listed in Section","section":"Remark 9.3; Section 11.1; Section 10.9"},{"comment":"The 'globally optimal shared fold' claim is global only after imposing a compactification such as finite box bounds h_low <= h_k <= h_high (or an alternative scale-gauge constraint). The experiments report numerically fitted GP folds but do not state the box bounds used, nor whether the resulting solution lies in the interior of the box. Without this information, the reader cannot verify that the deployed fold is the global minimizer of eq. (9) rather than a boundary solution of an arbitrary compactification. Since the headline comparisons against identity and SmoothQuant use these solved folds, the numerical setup needs to be specified, and boundary activity should be checked, for the 'globally optimal' claim to be meaningful in the experimental sections.","section":"Theorem 4.2; Sections 10.2 and 10.9"}],"minor_comments":[{"comment":"The median Spearman correlations are given as 0.937 and 0.918 in the text but are rounded to 0.94 and 0.92 in the figure caption. Please make the values consistent.","section":"Figure 11d and Section 10.9"},{"comment":"The artifact is described as complete with checksums and one-command regeneration, but the repository is private pending clearance. The reproducibility claim cannot be verified by reviewers. Please make the code public before final acceptance, or clearly mark the artifact as unavailable and provide the full numerical outputs in the supplement.","section":"Appendix F"},{"comment":"The log-magnitude regularizer tau_log is a free parameter and is fixed in advance in the experiments. Please report the chosen value and, ideally, a sensitivity analysis or a selection rule, since Theorem 5.2's transfer to raw spread requires the condition tau_log <= epsilon a_min,k.","section":"Sections 4.2 and 10.4"},{"comment":"The reuse descriptors n_gauge and n_opp are central to the paper, but the subscript formatting is inconsistent in the text (e.g., 'nopp' vs 'n opp'). Please use a consistent notation, e.g., n_opp, in Table 1 and in all figures that reference copy counts.","section":"Table 1"},{"comment":"The identity-to-GP modeled objective ratio of 161.2x is a ratio of modeled dither objectives, not of realized RTN errors. The text does say this, but the figure could mislead readers; consider making this distinction more prominent in the panel label or caption.","section":"Section 10.2"}],"recommendation":"major_revision","confidential_remarks":"The mathematical core is solid and likely publishable. The decisive issue is the gap between the certified dither-model results and the empirical RTN transfer that supports the headline gains. If the authors can broaden the RTN validation to at least one additional architecture or provide a principled condition for when the dither ranking transfers, and make the artifact public, I would support acceptance. As it stands, the applied claims are over-broad relative to the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper is worth knowing for two reasons. The exact expected squared product-error identity (Theorem 3.3) is clean and genuinely useful, and the shared-fold geometric program (Theorem 4.2), together with the identity-fold LP test, is a real algorithmic contribution. The math is careful, and the authors are honest about what is exact and what is surrogate.\n\nWhat is new: the identity is an elementary expansion but exact under independent zero-mean entrywise errors, covering subtractive dither and independent stochastic rounding. AB=(AT)(T^{-1}B) is classical, but formulating the diagonal shared fold as a GP and the identity-fold optimality test as an LP is new. The log-profile clustering guarantee and the n_opp copy-count framing also add something. The paper does well: assumptions are stated explicitly, proofs go to the appendix, and places where results are upper bounds rather than exact optima are flagged.\n\nThe soft spots are where the applied claims outrun the evidence. The certified optima hold for the dither variance law v = c R^2, not for deterministic round-to-nearest. RTN has input-dependent bias and inter-coordinate correlations, and Remark 9.3 gives an explicit pair where the dither objective ranks H_4 above I_4 but INT8 RTN ranks I_4 first. So the GP's 'global optimum' is for a surrogate. The bridge to RTN is empirical: one three-block width-64 ViT-like classifier on 8x8 digits, twelve products, at 8- and 4-bit. The median within-product Spearman correlations (0.937, 0.918) are respectable, but it is one tiny architecture, no language model, no larger vision model. Section 12 itself lists LLM validation as an extension. The artifact repository is private at review time, even though the reproducibility section is unusually thorough.\n\nNone of this undermines the identity or the GP itself. It does mean the headline held-out RTN error reductions (18–20%) are conditional until the transfer is shown on more models. The paper would be stronger if the pinned repository were public and if at least one larger-scale validation of the RTN ranking were added before the generalization claims.\n\nBottom line: the theoretical core deserves a serious referee and likely publication; the empirical claims need to be scaled or carefully hedged. I would cite Theorem 3.3 and the GP result in my own work. Worth a look if you care about quantization for inference.\n\nRecommendation: send it to peer review, with a request for broader empirical validation or a more tightly scoped claim.","headline":"The exact dither-model identity and the shared-fold GP are solid and genuinely useful; the RTN-transfer claims rest on one tiny classifier and need broader validation.","tokens_in":44790,"tokens_out":2172,"would_cite":true,"duration_ms":20179,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A60","65G50","90C25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantizing both factors of a matrix product before multiplying introduces error; this paper derives the exact expected size of that error and shows that the best rescaling of the shared dimension — a 'fold' — is the solution of a provably s","keywords":["quantized matrix multiplication","product-error identity","contraction gauge","diagonal folds","geometric programming","subtractive dither","round-to-nearest","transform selection"],"falsifier":"Reproduce the twelve-product selection protocol on a different architecture, such as a pretrained language model with real token activations, and measure the median within-product Spearman correlation between dither-predicted and round-to-nearest product errors; if it falls well below 0.9, or if the GP fold is beaten by a grid-tuned baseline on a majority of products, the surrogate-to-round-to-nearest transfer that carries the practical claims fails. A faster check is the paper's characteristic-function diagnostic Ξ: activations whose Ξ sits far above the uniform-residue null are warning signs","tokens_in":43855,"feed_emoji":"🧮","tokens_out":11648,"duration_ms":92841,"temperature":0.7,"pith_summary":"This paper tries to put low-precision matrix multiplication on a single quantitative footing: it derives an exact formula for the expected squared error of a quantized product, then shows that one whole family of error-reducing tricks — rescaling the shared dimension before quantization (a 'fold') — can be chosen by convex optimization with a certified global optimum. The formula matters because the error of a quantized product is not the entrywise error of either factor but that error propagated through the other factor's energy: an outlier in one operand is harmless exactly when the opposite factor has little energy in that coordinate. The identity holds exactly under subtractive dither and independent stochastic rounding; for the hardware-standard deterministic round-to-nearest it is an approximation, so the paper measures how well its predictions transfer. On twelve products of a trained three-block image classifier, the dither-model predictions rank measured round-to-nearest errors with median within-product correlation 0.937 at 8 bits and 0.918 at 4 bits, and the optimized fold cuts held-out product error by 18.0% and 20.5% relative to the identity fold.","feed_headline":"Cut matrix-product error 20% with a provably optimal rescaling","feed_subtitle":"Exact error accounting and convex optimization choose the rescaling; a trained classifier confirms gains at 8 and 4 bits.","key_machinery":"The load-bearing object is the product-error identity (Theorem 3.3): an exact finite-dimensional formula expressing the expected squared error of the quantized product as a function of the entrywise variance fields v^A and v^B, weighted by opposite-factor row and column energies, plus a bilinear cross term for simultaneous errors in both factors. It converts every design decision — transform, sharing pattern, grouping, bit allocation, clipping, rounding rule — into one scalar objective. The second mechanism is the contraction-gauge equivalence AB = (AT)(T⁻¹B): a diagonal gauge (a 'fold') makes the shared-domain objective a geometric program, convex in log coordinates, so the optimum is certi","core_discovery":"The paper's central claim is an exact identity for the expected squared error of a quantized matrix product under independent zero-mean entrywise noise (exact for subtractive dither and stochastic rounding, approximate for round-to-nearest): the error equals each operand's quantization variances charged against the opposite factor's row/column energies, plus a bilinear simultaneous-error term. On top of this identity the paper proves that choosing a diagonal contraction gauge — a per-coordinate rescaling (A,B)↦(AD,D⁻¹B) that preserves the product — over a shared output domain is a geometric program with a convex log-domain form and certified global optimum, and that a linear program decides","pith_inferences":["Editorial extension: the certified optima apply to the dither variance law v = c·range²; an immediate testable upgrade would feed the same geometric program with variance fields empirically calibrated from round-to-nearest residuals, which would make the 'global optimum' certified for the actual deployment quantizer.","Editorial extension: the weighted-norm generalization (Corollary 3.4) means the same selection machinery can target residual- or energy-weighted objectives in iterative solvers; the paper points at this application but does not demonstrate it.","Editorial extension: the bit-allocation rule offers a cheap audit for existing quantized inference stacks — compute P_A/P_B per layer on real activations and compare the deployed split with ½ log₂(P_A/P_B); where they disagree, the paper's experiments suggest up to ~1.7× error is on the table."],"forward_implications":["Diagonal fold selection for a shared output domain stops being a heuristic: the log-domain geometric program gives a certified global optimum of the range-law objective, and the linear-program test decides exactly when the identity fold is already optimal.","One product-error objective ranks candidates across transform families, so comparisons like flat rotation versus hierarchy reduce to evaluating a single statistic such as the slice-energy overlap g Σ p_r q_r.","A fixed bit budget between the two operands should be split asymmetrically as b*_A − b*_B = ½ log₂(P_A/P_B); the paper's experiments show the predicted 9/7 split beats the symmetric 8/8 split by 1.7× in squared relative error.","Transfer of the dither model to deterministic RTN is measurable: on a trained classifier the dither predictions rank realized errors with median within-product correlation above 0.9, and the GP fold beats both the identity fold and a grid-tuned scaling baseline on ten of twelve products at each precision, lowering composed logit MSE by 15.4% and 26.4%.","Transform reuse is made explicit through the quantized-copy count n_opp: a shared gauge costs one transformed copy of the opposite factor, block-specific gauges cost up to one per block, and the measured copy/error trade-off on heavy-tailed operands shows three extra copies buy 24.6% lower error."],"fun_headline_variants":["Exact error identity picks rescaling that cuts matmul error 20%","Provably optimal rescaling shrinks quantized matmul error by 20%","Matmul error cut 20% by provably optimal contraction-gauge","Exact expected-error formula yields 20% matmul error cut","Contraction-gauge tuning cuts matmul error 20% with proven optimum"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The certified optima — the GP fold, the identity-fold test, the selection statistics — are exact for the stochastic dither model with variance v = c·range²; transferring that ranking to deterministic round-to-nearest, which has input-dependent bias and correlated errors, is verified empirically on one small classifier, and if the transfer fails on other models the 'global optimum' would be optimizing a surrogate.","fun_headline_variants_meta":{"raw":{"variants":["Exact error identity picks rescaling that cuts matmul error 20%","Provably optimal rescaling shrinks quantized matmul error by 20%","Matmul error cut 20% by provably optimal contraction-gauge","Exact expected-error formula yields 20% matmul error cut","Contraction-gauge tuning cuts matmul error 20% with proven optimum"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1649,"prompt_tokens":873,"completion_tokens":776,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":677}},"tokens_in":617,"tokens_out":776,"duration_ms":9737,"temperature":1.0,"reasoning_tokens":677,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:29:28.511639+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the twelve-product selection protocol on a different architecture, such as a pretrained language model with real token activations, and measure the median within-product Spearman correlation between dither-predicted and round-to-nearest product errors; if it falls well below 0.9, or if the GP fold is beaten by a grid-tuned baseline on a majority of products, the surrogate-to-round-to-nearest transfer that carries the practical claims fails. A faster check is the paper's characteristic-function diagnostic Ξ: activations whose Ξ sits far above the uniform-residue null are warning signs","supporting_citations":[],"review_version":1}