{"id":"2d5c5583-1802-4ceb-8676-a871f2e17b4a","arxiv_id":"2603.04365","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A comparison theorem shows extreme eigenvalues of independent random matrix sums are dominated by a moment-matched Gaussian matrix, resolving the lower-distortion half of the Nelson-Nguyen conjecture.","lead":"This paper proves that the largest eigenvalue of a sum of independent random matrices is controlled by the largest eigenvalue of a matched Gaussian matrix, up to explicit error terms. The result sharpens existing matrix concentration bounds and yields a new proof of a 2013 conjectured injectivity property for sparse random dimension reduction maps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader's weakest_assumption was that Stahl's theorem (Fact 4.3) might fail, but Stahl's theorem is a proved theorem, so this is not a correctness risk for the central claim. The central comparison theorem's proof is internally consistent: the one-sided upper bound on λmax(W_i−EW_i) is exactly what makes the trace exponential expectation finite, the Lindeberg exchange in Prop 4.6 is valid, and the subsequent Gaussian concentration and tail derivations are routine and correct. The two errors identified by the reader (Prop 4.11's direction and Conjecture 3.5's scaling) are real but do not bear on Theorem 1.1; they are presentation errors in peripheral statements. Therefore I have no significant objection to the central claim, and I would keep the reader's CONDITIONAL verdict unchanged: the condition is that the typos be fixed, not that any mathematical step in Theorem 1.1 be reconsidered.","tokens_in":33311,"tokens_out":42866,"duration_ms":378860,"concrete_test":"Verify the algebraic step in Prop 4.10: for u=√(2s)+s/3, confirm that u−(1+u)ln(1+u)≤−s for all s≥0; if this fails, the final tail bound in Theorem 1.1 would not follow from the stated Bennett bound. Also check that the next revision corrects Prop 4.11's inequality direction and Conjecture 3.5's k scaling to k=O(α^{-2}d).","verdict_should_be":"UNCHANGED","load_bearing_attack":"I find no load-bearing flaw in Theorem 1.1. The proof's reliance on Stahl's theorem (Fact 4.3) is not a genuine vulnerability: Stahl's theorem is a proved result, and the one-sided hypothesis λmax(W_i−EW_i)≤R_+ ensures the trace mgf is finite even when summands are unbounded below. I checked the Lindeberg exchange (Prop 4.6), the telescoping sum (Prop 4.8), Gaussian concentration (Prop 4.9), and the tail derivations (Prop 4.10); the inequality directions and constants are consistent. The manuscript does contain two genuine errors, but they are typographical and peripheral: Proposition 4.11 states its tail bound with the wrong inequality direction, and Conjecture 3.5's embedding-dimension scaling k≤C α^{-2} log d is impossible (for an injection one needs k≥d) and disagrees with Theorem 3.6. Neither error affects the central comparison theorem, whose proof appears sound.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a nonasymptotic comparison theorem for the extreme eigenvalues of a sum of independent random self-adjoint matrices. The main result, Theorem 1.1, states that if Y = Σ W_i with E W_i = 0 and λmax(W_i) ≤ R_+, then the expected maximum eigenvalue of Y is bounded by that of a Gaussian proxy Z ∼ normal(EY, Var Y) plus explicit error terms depending on R_+, the Gaussian fluctuation φ(Z), and the weak variance σ_*^2(Z); an analogous one-sided tail bound is also proved. The proof uses Lindeberg's exchange method together with Stahl's theorem on the trace exponential to compare trace mgfs, then Gaussian concentration to extract eigenvalue bounds. Corollaries cover the minimum eigenvalue, the spectral norm of rectangular sums, and unbounded summands via truncation. Applications are given to random regular graphs, the random Pauli model, sample covariance matrices, and, as the headline application, a SparseStack dimension-reduction map for which the paper claims to prove the lower-distortion part of the Nelson–Nguyen conjecture.","tokens_in":33502,"tokens_out":8320,"duration_ms":79918,"significance":"Theorem 1.1 is a substantial refinement of existing matrix concentration and Gaussian universality results. Its one-sided hypothesis λmax(W_i − EW_i) ≤ R_+ is much weaker than uniform norm control, and the logarithmic factors appear only in the error terms, so the bound is always competitive with the matrix Bernstein inequality and with the Brailovskaya–van Handel comparison results. The proof is self-contained modulo standard external facts, and it is parameter-free: no constants are fitted, and the Gaussian proxy statistics are computed from the input model rather than calibrated to the target. If the central theorem is correct, the applications—especially the SparseStack injectivity result—are notable contributions. The manuscript is clearly written and the proof strategy is transparent, with explicit constants and a clean five-step structure.","major_comments":[{"comment":"As stated, the conjecture is impossible: an embedding from C^n to C^k with k < d cannot be injective on a d-dimensional subspace, yet the conjecture claims k ≤ C α^{-2} log d. This also contradicts Theorem 3.6, which requires k ≥ 16 α^{-2}(d ∨ log(d/p)). Consequently, the abstract's claim that the paper gives 'the first complete proof' of the Nelson–Nguyen conjecture is not accurate as written. The intended conjecture almost certainly has a dimension factor (e.g., k ≤ C α^{-2} d log d), and Theorem 3.6 may well be a stronger statement, but the text must be corrected and the claim rephrased to say precisely which version of the conjecture is being proved.","section":"Section 3.4.3, Conjecture 3.5"}],"minor_comments":[{"comment":"The displayed tail bound has the inequality direction reversed: it states P{λmax(Y+Δ) ≤ μ + …} ≤ P{M>R} + d e^{-s}, but the proof and Corollary 1.2 establish an upper tail bound of the form P{λmax(Y+Δ) ≥ μ + …} ≤ P{M>R} + d e^{-s}. This is a clear typographical error and should be fixed.","section":"Section 4.10, Proposition 4.11"},{"comment":"In the GUE invariance line, the right-hand side should be U^* X_gue U rather than U^* X_goe U. The current text appears to be a copying error.","section":"Section 2.5.2"},{"comment":"The comparison with [BH24b] cites 'Thm. 3.8' for the random regular graph bound, while elsewhere the relevant result is referred to as Cor. 2.7. Please ensure the reference is consistent.","section":"Section 3.1.1"},{"comment":"In the last paragraph of the proof, the condition 'requiring that 3√k ≤ N' and the earlier theorem assumption k ≤ N^2/9 are consistent, but the wording '3√k' is ambiguous; it would be clearer to state √k ≤ N/3 or k ≤ N^2/9.","section":"Section 3.2.1"}],"recommendation":"minor_revision","confidential_remarks":"The central theorem and its proof appear sound; I found no load-bearing flaw in the derivation of Theorem 1.1. The two substantive issues are local: the Nelson–Nguyen conjecture is misstated (with an impossible k-scaling), and Proposition 4.11 has a reversed inequality sign. Both are easily corrected without changing the main argument. The paper is a strong contribution to matrix concentration and deserves publication after these corrections."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main comparison theorem is a genuine advance: it gives a one-sided Gaussian comparison for the maximum eigenvalue with error terms controlled by R_+ φ(Z) + σ_*^2(Z) and only log d factors, strictly improving the upper bound from Brailovskaya–van Handel in many settings. The proof is transparent, uses Stahl's theorem in a new way via Lindeberg exchange, and I checked the central inequality chain — it holds up. The SparseStack injectivity result (Theorem 3.6) is the first proof of the lower-distortion half of the Nelson–Nguyen conjecture with the conjectured column sparsity ζ = O(α^{-1} log d) and embedding dimension k = O(α^{-2} d). That alone makes the paper significant.\n\nWhat the paper does well: it states clearly what is new vs. the companion paper and BH24b, gives credible applications across graph theory, quantum information, covariance estimation, and numerical linear algebra, and the applications actually use the one-sided control rather than paying for two-sided assumptions. The companion paper is cited appropriately, and the Stahl-based arguments are parameter-free, not fitted.\n\nSoft spots: two mechanical errors in the manuscript. Proposition 4.11 states the tail bound with the inequality in the wrong direction — it should be a lower bound on the probability that λmax is at least the threshold, not at most. And Conjecture 3.5 says k ≤ Const·α^{-2} log d, but any embedding needs k ≥ d and the paper's own Theorem 3.6 requires k ≳ α^{-2} d. These are genuine typos/statement errors, but neither affects Theorem 1.1 or the proof in Section 4. I also note the paper does not attempt two-sided bounds, so its statements are weaker than BH24b in that respect; the author acknowledges this, so it is not a hidden flaw.\n\nWho should read this: anyone working on matrix concentration, universality, or randomized linear algebra. The paper deserves a serious referee, despite the mechanical errors. My recommendation is to send to peer review, and require the author to correct the direction of Proposition 4.11 and the impossible scaling in Conjecture 3.5 before final acceptance.","headline":"Real new comparison theorem and a credible first proof of the lower-distortion half of Nelson-Nguyen, but two public-facing errors need correction before citable.","tokens_in":34043,"tokens_out":1990,"would_cite":true,"duration_ms":20742,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15B52","60B20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The top eigenvalue of any sum of independent random symmetric matrices is controlled by the top eigenvalue of a matching Gaussian matrix, plus explicit error terms.","keywords":["comparison theorem","random matrix","extreme eigenvalues","Gaussian proxy","Lindeberg method","trace exponential","matrix concentration","subspace embedding"],"falsifier":"Compute for a small explicit example, say a sum of independent Bernoulli rank-one matrices with matching Gaussian proxy, the empirical expectation E λmax(Y) and compare it with the theorem's right-hand side; a single violation for d ≥ 2 would refute Theorem 1.1. More directly, for a fixed matrix A and a sparse random W with λmax(W) ≤ R, evaluate the trace-exponential inequality E Tr e^{A+W} ≤ E Tr e^{A+gX}; a counterexample to this one-matrix step would break the entire chain.","tokens_in":1417,"feed_emoji":"🎲","tokens_out":1540,"duration_ms":78221,"temperature":0.7,"pith_summary":"The paper proves that the largest eigenvalue of a sum of independent random self-adjoint matrices is bounded above by the largest eigenvalue of a Gaussian matrix with matching mean and variance, plus small error terms that decay when no single summand dominates. The error terms depend only on a one-sided fluctuation bound for the summands, the dimension, and two summary statistics of the Gaussian proxy. This transfers the rich toolkit of Gaussian random matrix theory to arbitrary matrix sums, yielding sharper concentration bounds than previous matrix Bernstein and universality results. The method delivers the first complete proof of the injectivity half of a 2013 conjecture on sparse dimension reduction maps, together with new spectral bounds for random regular graphs, random Pauli matrices, and sample covariance matrices.","feed_headline":"Gaussian proxy controls top eigenvalue of random matrix sums","feed_subtitle":"New explicit bounds sharpen spectral estimates in graph theory, statistics, and sparse random embeddings.","key_machinery":"The central object is the trace exponential f(t) = Tr e^{A + tH} and Stahl's theorem (the former BMV conjecture), which represents it as the Laplace transform of a positive measure on the interval [λmin(H), λmax(H)]. This representation makes the even derivatives of the trace exponential positive and gives sharp bounds on the Taylor remainder, so that Lindeberg's method — exchanging one random summand for its matching Gaussian at a time — yields a one-sided comparison of trace moment generating functions. The summary statistics of the Gaussian proxy that carry the final bound are the matrix fluctuation φ(Z) = E λmax(Z - E Z) and the weak variance σ_*^2(Z), the maximal variance of the quadrat","core_discovery":"The central claim is a stochastic domination result: for an independent sum Y of self-adjoint matrices with two finite moments and a uniform upper bound R_+ on the centered maximum eigenvalues, the Gaussian proxy Z with the same first two moments satisfies E λmax(Y) ≤ E λmax(Z) + sqrt((R_+ φ(Z)/3 + σ_*^2(Z))·2 log d) + R_+ log d / 3, with an analogous tail inequality at level s. Here φ(Z) is the expected fluctuation of the Gaussian maximum and σ_*^2(Z) is the supremum over unit vectors of the variance of quadratic forms of Z. The proof proceeds by a Lindeberg exchange of summands, using Stahl's theorem to show the trace exponential is the Laplace transform of a positive measure and thereby c","pith_inferences":["Editorial inference: the same trace-exponential comparison should extend to matrix martingale difference sequences via Azuma-type exchangeable arguments, as the paper notes but does not develop; this would give Gaussian comparison bounds for adaptive sums.","Editorial inference: the one-sided nature of the bound is likely inherent: a matching two-sided comparison would need control of λmin of summands, not just λmax, and the paper's methods do not address the minimum singular value of rectangular matrices.","Editorial inference: a numerical test on small worst-case sums, such as Bernoulli rank-one summands, could reveal whether the log d factors in the error terms are necessary or an artifact of the proof.","Editorial inference: the Gaussian proxy's weak variance σ_*^2 is often much smaller than the square of the matrix fluctuation, so the theorem is strongest when fluctuations are spread across many directions; constructions with highly localized variance may exhibit the worst-case behavior."],"forward_implications":["Any accurate estimate for the maximum eigenvalue of a Gaussian matrix transfers to arbitrary independent matrix sums with matching first two moments, paying only sqrt(log d) error terms.","For Wigner matrices the bound gives E λmax(Y) ≤ 2√d + O(d^{1/4} √log d), with the sharp leading constant 2√d.","For Rademacher covariance matrices, the minimum-eigenvalue comparison reproduces the Bai–Yin first-order limit 1 - 2√ρ when n ≫ d.","For the random Pauli model with N = 2^n, k ≳ n^2 α^{-4} summands suffice to match GUE spectral edges to relative error α, improving prior n^3 and n^4 requirements.","It gives the first complete proof that a SparseStack matrix with column sparsity ζ ≍ α^{-1} log(d/p) and embedding dimension k ≍ α^{-2} d is injective on any d-dimensional subspace with high probability, confirming the lower-distortion half of a 2013 conjecture."],"fun_headline_variants":["Gaussian proxy controls extreme eigenvalues of random sums","New comparison theorem sharpens spectral bounds","Sparse random embedding injectivity finally proven","Gaussian comparison yields tighter spectral estimates"],"cache_read_input_tokens":35456,"weakest_assumption_plain":"The argument rests on Stahl's theorem — that the trace exponential along a line is the Laplace transform of a positive measure; if that positivity failed, the key Taylor-remainder control for the Lindeberg exchange would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian proxy controls extreme eigenvalues of random sums","New comparison theorem sharpens spectral bounds","Sparse random embedding injectivity finally proven","Gaussian comparison yields tighter spectral estimates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1269,"prompt_tokens":685,"completion_tokens":584,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":530}},"tokens_in":429,"tokens_out":584,"duration_ms":6467,"temperature":1.0,"reasoning_tokens":530,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T18:50:48.369361+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute for a small explicit example, say a sum of independent Bernoulli rank-one matrices with matching Gaussian proxy, the empirical expectation E λmax(Y) and compare it with the theorem's right-hand side; a single violation for d ≥ 2 would refute Theorem 1.1. More directly, for a fixed matrix A and a sparse random W with λmax(W) ≤ R, evaluate the trace-exponential inequality E Tr e^{A+W} ≤ E Tr e^{A+gX}; a counterexample to this one-matrix step would break the entire chain.","supporting_citations":[],"review_version":1}