{"id":"8133ef23-ab5a-450f-982e-aab4b30a954e","arxiv_id":"2607.18758","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The mean-square rounding error of a reduction tree T is, to leading order, u²ν(τ²Λ₁(T)+μ²Λ₂(T)), where Λ₁ and Λ₂ are tree statistics.","lead":"This paper shows that the typical rounding error of a floating-point sum depends on the order of additions through two tree statistics: total leaf depth and sum of squared subtree sizes. It derives closed-form error formulas, finds optimal summation orders, and matches measured errors within a few percent on tested configurations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Constant-ν approximation can in principle overturn Λ-based optimal-schedule rankings; paper scopes its claims explicitly, so no verdict change.","rationale":"The mathematical derivation is sound: the common-ancestor kernel identity (Prop. 2.4) and the i.i.d. compression (Cor. 2.5) are correct, and the optimization results follow from the closed-form statistics. The GEMM contraction (Prop. 4.3) also checks out. The paper's experiments are unusually careful — exact residual evaluation, bitwise tree validation, independent calibration for GEMM, and honest reporting of post-hoc stochastic-rounding diagnostics and failure regimes. The main fragility, the constant-ν assumption, is explicitly identified and empirically probed; its known failure modes (coefficient drift, cross terms, stagnation) are quantified and disclosed. The three-stage schedule test's inconclusive ranking is a limitation, not an inconsistency, because the paper does not claim the fine-grained ranking was experimentally resolved. Given the paper's conditional framing and its extensive validation within the stated regimes, the reader's ACCEPT verdict with moderate confidence remains appropriate. No basis was found to lower the verdict.","tokens_in":22254,"tokens_out":13858,"duration_ms":113987,"concrete_test":"Using the supplementary per-trial and local-error data, compute νeff separately for the three k=2^14 schedules (doubling-gap, geometric, perturbed) and form corrected predictions u²νeffΛ2 for each. Check whether the ordering by Λ2 alone matches the ordering by u²νeffΛ2 and by measured RMS. If the corrected ordering differs, the schedule-optimality claim depends on per-tree coefficient variation and should be presented as model-relative; if it matches, the constant-ν optimality is robust in this regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Eq. 1) is an exact consequence of the constant-ν model (Corollary 2.3) but only an empirical approximation under round-to-nearest, where the local variance is not νu²x². The paper's own data show νeff drifting from 0.217 to 0.190 with tree size (Appendix 1.2), cross terms ranging from −10.2% to +5.2% of MSE (Appendix 1.3), and complete stagnation/bias failure in low-precision positive sums (Section 6.2). The optimality results (Props. 3.1–3.4, 3.2) minimize Λ1 and Λ2 for a single tree-independent ν. If νeff varies across trees by more than the relative differences in Λ1/Λ2, the tree minimizing actual MSE need not be the Λ-optimal tree. The one direct test of a predicted fine-grained ranking — the three-stage schedule comparison (Section 6.5) — yields paired bootstrap intervals containing unity, so the doubling-gap versus geometric ranking is not statistically confirmed. This does not invalidate the conditional claim; the paper explicitly labels its results as 'under the constant-ν model' and lists stagnation/bias as limitations. But it does mean the optimal-schedule theorems are model-relative rather than established for round-to-nearest arithmetic in all regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops a second-moment theory for floating-point reduction trees. Under conditionally unbiased rounding with local variance νu²x² (the constant-ν model), the authors derive an exact recurrence for the accumulated error and show that the leading tree-dependent cost for a fixed input vector p is the quadratic form pᵀK_Tp, where K_T is the common-ancestor kernel counting shared internal ancestors. For i.i.d. inputs this cost reduces to τ²Λ₁(T)+μ²Λ₂(T), with Λ₁ total leaf depth and Λ₂ sum of squared internal subtree sizes. The paper then proves optimality results: balanced/sequential trees are the Λ₁ extrema; two-stage sequential blocking has RMS exponent 3/4 at block size √k; fixed-stage hierarchies have geometric schedules optimal for Λ₁ and a 'doubling-gap' schedule optimal for the polynomial order of Λ₂; and Huffman coding minimizes variance-weighted depth for heterogeneous centered inputs. The framework is extended to GEMM via a Hadamard-product contraction ⟨K_T,G⟩. Extensive experiments under round-to-nearest, using exact residuals and bitwise-validated tree extraction, support the topology ordering and absolute MSE levels within a few percent in tested regimes, while honestly documenting failure regimes (low-precision positive sums, stagnation, bias, and ν_eff drift).","tokens_in":22611,"tokens_out":21462,"duration_ms":153582,"significance":"The principal contribution is a clean, exact structural decomposition of the expected squared partial-sum cost in terms of the tree-only statistics Λ₁,Λ₂ and the input second-moment matrix. The kernel formulation is elegant and unifies summation, correlated inputs, and GEMM. The optimization results are model-relative but internally consistent, and the paper is unusually explicit about the empirical status of the constant-ν approximation. The experimental section is a model of reproducible numerical investigation: exact residual strategies per precision, bootstrap intervals, bitwise validation of the reconstructed NumPy tree, independent calibration of ν from node-local errors, and transparent reporting of inconclusive schedule comparisons and post-hoc stochastic-rounding checks. If the results hold, the paper provides a practical tool for predicting and comparing the typical error of reduction geometries.","major_comments":[],"minor_comments":[{"comment":"The phrase 'the divisible case nb=k/b' is ambiguous: use n_b = k/b and define n_b as the number of blocks, with b the block size, so that the closed-form asymptotics in the following table are unambiguous.","section":"Section 2.4"},{"comment":"The notation 'k^{3/4}=2c_p/c_a' is confusing because k already denotes the reduction size. Rename the crossover size (e.g., κ or k_*) and state that the crossover satisfies k_* = 2c_p/c_a.","section":"Appendix 1.1, Eq. (18)"},{"comment":"The construction 'combine pairwise blocks of size g=k^{2-2α}' is degenerate at the endpoint α=1 (g=1). Please state separately that the sequential tree attains α=1.","section":"Section 3.1"},{"comment":"In the row for stochastic rounding, 'stated grids 24–600' is vague. Please point to the exact grid definition (e.g., the Figure 9 caption) or list the sizes directly in the table.","section":"Table 2"},{"comment":"The doubling-gap schedule is the Λ₂-optimal fixed-stage schedule under the constant-ν model; the direct experimental comparison (geometric vs. doubling-gap) is statistically inconclusive, with paired bootstrap intervals containing unity. The paper already discloses this in Section 6.5, but a one-sentence reminder at the end of Section 3.4 would help readers avoid over-reading the theorem as an empirical guarantee for round-to-nearest arithmetic.","section":"Section 3.4 / Section 6.5"},{"comment":"The kernel matrices displayed in the bottom panel are dense and hard to read in preprint form. Consider larger spacing, separate panel titles, or highlighting the diagonal and all-entry sums to improve legibility.","section":"Figure 1"}],"recommendation":"accept","confidential_remarks":"The manuscript is a strong, well-executed contribution. The conditional nature of the constant-ν model, the failure regimes, and the inconclusive schedule-ranking test are all disclosed honestly and with sufficient precision. The central derivations are sound, and the experimental validation is unusually thorough. Remaining issues are presentation-level only. I recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a serious look. The paper gives a genuinely new handle on how reduction-tree geometry affects mean-square rounding error: the common-ancestor kernel K_T and its i.i.d. compression to Λ1 and Λ2. The exact recurrence under conditionally unbiased rounding is clean, and the optimization results — the k^{3/4} two-stage blocking, the geometric versus doubling-gap schedules, the Huffman import for variance-weighted depth — are well executed and clearly scoped. The GEMM extension via G=(A^T A)∘(BB^T) is a nice contraction, and the independent calibration work is about as careful as I have seen in this literature.\n\nThe soft spot is the constant-ν model itself. Under round-to-nearest it is an empirical approximation, not an exact consequence. The paper's own data show νeff drifting from 0.217 to 0.190, cross terms contributing 5–10% of MSE, and complete breakdown in low-precision positive sums due to stagnation and bias. The three-stage schedule ranking is not statistically confirmed: the paired bootstrap intervals contain unity. So the optimal-schedule theorems are model-relative; outside the calibrated regime they may not minimize actual MSE. That is a real limitation. But it is not a hidden one. The paper labels its claims 'under the constant-ν model', maps where the model fails, and the structural predictions — topology ordering, exponents, schedule rankings — are parameter-free functions of the tree statistics. The absolute level uses a calibrated ν, and the authors say so plainly.\n\nOne minor caveat: the stochastic-rounding diagnostics have a whiff of post-hoc selection (two initial misses, then follow-up runs), though the paper discloses this. It would be easy to overstate the confirmatory power there.\n\nWho gets value from this? Numerical analysts and HPC practitioners who care about typical rounding error, schedule choice, or mixed-precision reductions. The paper deserves a serious referee, not a desk reject, and with reasonable revisions it would be a solid contribution.","headline":"Honest, careful second-moment theory for reduction-tree rounding; the constant-ν simplification is the main soft spot and the paper says so itself.","tokens_in":23097,"tokens_out":2242,"would_cite":true,"duration_ms":20980,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65G50"],"pacs":[],"model":"deepseek-v4-flash","headline":"Ordinary floating-point rounding error in a sum is set by two tree statistics: total leaf depth and the sum of squared internal-subtree sizes.","keywords":["floating-point arithmetic","rounding error","second-moment analysis","reduction tree","mean-square error","common-ancestor kernel","summation order","GEMM error analysis"],"falsifier":"Measure exact mean-square error for centered i.i.d. sums on one fixed balanced binary64 tree across k = 2^10 to 2^26; if the effective coefficient νeff = MSE/(u²τ²Λ1) drifts by more than about 20% over that range, the constant-ν model is not a predictive law for round-to-nearest summation.","tokens_in":22144,"feed_emoji":"🧮","tokens_out":9293,"duration_ms":81214,"temperature":0.7,"pith_summary":"The paper develops a mean-square-error (MSE) theory for floating-point summation on a fixed binary reduction tree, claiming that the tree-dependent part of typical rounding error is not set by operation count but by shared ancestry. Under a constant-ν model of rounding, the expected squared error for i.i.d. inputs is u²ν(τ²Λ1 + μ²Λ2), where Λ1 is total leaf depth, Λ2 is the sum of squared internal-subtree sizes, μ is the input mean, and τ² is the input variance. That two-statistic decomposition lets the authors prove which tree topologies and schedules are optimal: balanced and sequential trees are the centered extrema, two-stage blocking achieves a k^{3/4} root-mean-square exponent, fixed-stage hierarchies prefer geometric schedules for centered inputs and a halving-gap schedule for noncentered inputs, and variance-weighted prefix codes minimize error under heterogeneous variances. The authors validate the leading formula under round-to-nearest across four precisions, recover topology orderings, and predict absolute GEMM error within 3%, while reporting that positive low-precision sums stagnate and become biased, where the model no longer applies. A reader should care because the theory turns reduction geometry into two cheaply computable scalars, letting implementations choose summation order for performance without losing a tree-specific accuracy prediction.","feed_headline":"Rounding error follows two tree statistics, not worst-case bounds","feed_subtitle":"A mean-square formula ranks summation trees by predicted error and fixes optimal block sizes, validated across four precisions.","key_machinery":"The common-ancestor kernel K_T — the matrix whose (i,j) entry counts the internal nodes that are ancestors of both leaves i and j — is the central object. It turns the tree-dependent second-moment cost Σ_v q_v² into the quadratic form p^T K_T p, so the expected cost for random inputs is the Frobenius contraction ⟨K_T, M⟩ = tr(K_T^T M). For i.i.d. inputs, that contraction reduces to the two scalar statistics Λ1 = tr K_T (total leaf depth) and Λ2 = 1^T K_T 1 (sum of squared internal-subtree sizes), which separate variance-driven from mean-driven error. All optimality results flow from minimizing these statistics: external path length arguments for Λ1, the prefix-code/entropy bound for variance","core_discovery":"The paper's central claim is that the tree-dependent part of the mean-square rounding error of a fixed binary reduction tree is not determined by the number of operations but by shared ancestry. For conditionally unbiased rounding, the error obeys an exact recurrence whose tree-dependent cost is Σ_v q_v², the sum of squared exact partial sums; this equals p^T K_T p, where K_T's (i,j) entry counts internal ancestors shared by leaves i and j. When inputs are i.i.d. with mean μ and variance τ², the expected cost collapses to τ²Λ1 + μ²Λ2, with Λ1 the total leaf depth and Λ2 the sum of squared internal-subtree sizes; under the constant-ν model the predicted RMS is u√(ν(τ²Λ1 + μ²Λ2)). This two-sta","pith_inferences":["Because the paper's dominance criterion uses only the ratio |μ|/τ and the tree statistics, a runtime could compute Λ1 and Λ2 during the reduction and automatically raise precision when the mean-to-deviation ratio crosses the threshold; the paper gives the criterion but stops short of implementing the policy.","The common-ancestor contraction should extend to other multi-way bilinear reductions — stencil computations, tensor contractions, batched inner products — whenever inputs carry a known second-moment matrix; testing it on correlated gradient reductions in low-precision training would be a direct next experiment.","The reported drift of the effective coefficient (0.217 down to 0.190) is consistent with the paper's own significand-based explanation, suggesting that a two-parameter model with a z^{-2} local-variance law might remove most of the configuration dependence that the constant-ν model leaves behind.","The measured cross-term covariances (5–10% of MSE) are small but not negligible; a correction term built from them could extend equation (1) to regimes where constant-ν currently breaks down, such as positive low-precision sums with stagnation and bias."],"forward_implications":["For centered i.i.d. inputs the predicted RMS error is uτ√(νΛ1), so tree topology alone orders the error: pairwise is the low extreme, sequential the high extreme, with blocked trees in between; the paper measures this ordering in binary64, binary32, binary16, and bfloat16.","Optimal two-stage sequential blocking balances inner and outer accumulation: b ~ √k gives a k^{3/4} RMS exponent for centered inputs, and b ~ k^{2/3} gives a 7/6 exponent for noncentered inputs; fixed-stage hierarchies are geometric in the centered regime and follow a halving-gap schedule in the noncentered regime.","For independent centered inputs with unequal variances, the error-minimizing free assignment of leaves to tree positions is a variance-weighted prefix code, and the optimal cost lies between the entropy lower bound and the entropy-plus-one upper bound; measured penalties for pairwise and natural-order sequential trees are 1.79–3.37× and 55.5–106× the optimum, respectively.","In GEMM, the same kernel contraction ⟨K_T, G⟩ with G = (A^T A) ∘ (B B^T) controls the tree-dependent error; for centered random matrices, the RMS forward error relative to the absolute-value scale is Θ(u) for sequential FMA, Θ(uk^{-1/4}) for √k-blocked, and Θ(u√(log k/k)) for pairwise accumulation, with predictions matching measurements within 3% on the tested grid.","Because Λ1 and Λ2 require just one postorder traversal, any reduction routine can report its predicted typical error alongside its result; the paper demonstrates the workflow by extracting the actual tree of an array library's sum and predicting its measured RMS within 7% across sizes 27 to 2^20."],"fun_headline_variants":["Sum rounding error? Count shared ancestors","Exact MSE for summation trees: shared ancestry governs","Rounding error obeys tree depth and subtree size","New theory ranks summation trees by exact mean-square error","Forget worst-case: rounding error follows two tree stats"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is the constant-ν model: every rounding step is conditionally unbiased with local variance exactly νu²x² for a single tree-independent coefficient; under round-to-nearest this is an approximation, and the paper's own measurements show ν drifting with size (0.217 to 0.190) and stagnation/bias dominating positive low-precision sums.","fun_headline_variants_meta":{"raw":{"variants":["Sum rounding error? Count shared ancestors","Exact MSE for summation trees: shared ancestry governs","Rounding error obeys tree depth and subtree size","New theory ranks summation trees by exact mean-square error","Forget worst-case: rounding error follows two tree stats"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":2953,"prompt_tokens":924,"completion_tokens":2029,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":668,"completion_tokens_details":{"reasoning_tokens":1966}},"tokens_in":668,"tokens_out":2029,"duration_ms":15079,"temperature":1.0,"reasoning_tokens":1966,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T14:26:27.197952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure exact mean-square error for centered i.i.d. sums on one fixed balanced binary64 tree across k = 2^10 to 2^26; if the effective coefficient νeff = MSE/(u²τ²Λ1) drifts by more than about 20% over that range, the constant-ν model is not a predictive law for round-to-nearest summation.","supporting_citations":[],"review_version":1}