{"id":"a290de7d-f66c-4062-aa51-cda5f72e943f","arxiv_id":"2511.10535","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"As N → ∞, the empirical eigenvalue law of Brownian motion on GL(N,C) converges almost surely to the Brown measure of a free multiplicative Brownian motion, settling Biane's 1997 conjecture.","lead":"Random complex matrices that drift by Brownian motion spread their eigenvalues into a deterministic cloud as the matrix size grows; this paper proves the cloud is exactly the one predicted by free probability, resolving a 1997 conjecture of Philippe Biane. The proof introduces a sharp small-time approximation of the group-valued diffusion by its flat Lie-algebra counterpart.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1.7 rests on the unverified k^4 exponent in the Step-2 bookkeeping of Theorem 3.5; if the number of terms or t-power grows faster, the N^{2/3} concentration bounds collapse and with them the Hermitization proof of Theorem 1.6.","rationale":"The reader's weakest assumption identifies exactly the application of Lemma 2.25 to the n-dependent polynomial Q_n and the Step-2 bookkeeping that produces (tC_{\\varrho,T})^{2k} exp(C_{\\varrho,T} k^4/N^2). My stress-test concurs: this is the single most load-bearing point in the paper. Everything downstream—Theorem 1.7's N^{2/3} concentration, Proposition 4.1's small-singular-value bounds, and the Hermitization argument in Section 6—depends on the exact k^4 exponent. If the bookkeeping yields a faster k-growth (e.g., k^5/N^2), the final probability bound would carry an unabsorbable exp(N^{4/3}) factor and the proof of eigenvalue convergence would fail. This is a correctness risk, not a matter of disagreement with the literature or with the free-probability consensus. I also credit the paper's independent support: Lemma 1.9 is proved from prior work, Proposition 3.3 gives a clean L^2 approximation, and the numerical figures are presented as supportive evidence rather than proof. The concern is specifically that the condensed Step-2 counting is not sufficiently checkable in the preprint, and the quoted Lemma 2.25 is not uniform in the polynomial family used. Since the reader already asked for exactly this referee check and recommended CONDITIONAL, my read does not change the verdict.","tokens_in":108349,"tokens_out":14392,"duration_ms":133420,"concrete_test":"Have an independent expert (or a computer-algebra derivation) reconstruct Step 2 of Theorem 3.5 from Lemma 2.25, applying L^{T_i} to the explicit polynomial Q_n and tracking the number of terms, the t-power, and the k-dependence through i iterations. Concretely, fix k=2 and k=3, and compute symbolically the i=1 and i=2 cases: verify (a) the number of generated terms is at most (2k)^{4i}(\\lfloor 2^n t\\rfloor)^{2i}, and (b) every term satisfies 2r_1+r_2 \\ge 4k-4i with the claimed t^{2k-2i} factor after Hölder. If the count scales like k^{5i} or the t-exponent is k-ci with c<1, the k^4/N^2 exponent in (1.14) fails; if the small-i cases match and the induction is written out, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central eigenvalue convergence (Theorem 1.6) depends on the small-singular-value control of Proposition 4.1, which in turn depends on the N^{2/3} small-time concentration of Theorem 1.7, which is derived from the moment estimate (1.14) in Theorem 3.5. The proof of Theorem 3.5 imports Lemma 2.25, an asymptotic expansion from Parraud [74], and applies it to the sequence of polynomials Q_n = |B_n^N(t) - I - W^N(\\lfloor 2^n t\\rfloor/2^n)|^{2k}, whose degree grows with n. Lemma 2.25 is stated for fixed polynomials and is not uniform in n; that uniformity must be supplied by the Step-2 counting argument on pp. 34–37. The argument is condensed: it asserts that L^{T_i}\\cdots L^{T_1}(Q_n) has at most 2^{4i}(2k)^{4i}(\\lfloor 2^n t\\rfloor)^{2i} terms, states the key exponent inequality 2r_1+r_2 \\ge 4k-4i without proof, and compresses several differentiations into 'one also has similar formulas.' The exact k^4 in exp(C_{\\varrho,T} k^4/N^2) is load-bearing: choosing k \\asymp N^{2/3} gives exp(C N^{2/3}), which can be absorbed into the constant c_{\\varrho,T}. If the true count grows like (2k)^{5i} or the t-power is k-ci with c<1, the optimized exponent becomes k^5/N^2 or worse, yielding exp(C N^{4/3}) and invalidating Theorem 1.7, Proposition 4.1, and the Hermitization in §6. I am not asserting the bound is false; I am asserting that the text does not provide a checkable proof of the crucial k^4 exponent, and no independent verification or machine-checked proof is supplied. Lemma 2.25 itself is published in [74], but the novel application to n-dependent polynomials is where the risk lies.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proves that, for every ρ>0 and complex ζ with |ζ|<ρ, and for any sequence of random initial conditions B_0^N that are a.s. invertible, satisfy the uniform spectral-gap condition (1.13), and converge a.s. in *-distribution to b_0, the empirical eigenvalue law of B_0^N B^N(t) converges weakly almost surely to the Brown measure μ_{b_0 b(t)} of the free multiplicative Brownian motion. The proof combines: (i) a Donsker-type product approximation with uniform L^2 error O(2^{-n/2}) (Proposition 3.3); (ii) a moment estimate (1.14) giving N^{2/3} small-time concentration of B^N(t)-I-W^N(t) (Theorem 1.7); (iii) small-singular-value estimates for shifted products (Proposition 4.1); and (iv) the Bordenave–Chafaï Hermitization lemma (Lemma 6.1). A complementary analytic Wegner estimate is proved in Section 5 for identity initial condition.","tokens_in":108607,"tokens_out":15251,"duration_ms":162690,"significance":"If the technical estimates hold, this is a major advance: it resolves Biane's 1997 conjecture and gives the first eigenvalue-level (not merely trace-level) convergence for a curved non-normal matrix diffusion, with an explicit deterministic limit. The overall architecture is coherent and genuinely novel, particularly the small-time affine approximation theorem, which is of independent interest. The limit measure is not fitted: it is independently defined by free stochastic calculus. The paper also carefully separates the main Hermitization route from the alternative analytic Section 5. My concern is concentrated in one load-bearing place: the proof of the moment estimate (1.14), whose exponent in k is not made checkable.","major_comments":[{"comment":"The bound (1.14) and hence Theorem 1.7, Proposition 4.1, and the Hermitization in Section 6 all rest on the Step 2 claim that (L^{T_i}...L^{T_1})(Q_n) is a linear combination of at most 2^{4i}(2k)^{4i}(floor(2^n t))^{2i} terms with the stated coefficients, and on the inequality 2r_1+r_2 ≥ 4k−4i. The text asserts both without proof: the first after 'Consequently' and the second as 'then necessarily'. This is not a cosmetic omission. The optimized choice k∼N^{2/3} means the k^4 in exp(C k^4/N^2) is exactly what makes the Markov argument yield N^{2/3} concentration; if the true count were k^{5i}, the optimization would give exp(C N^{4/3}) and the proof would collapse. I am not asserting the estimate is false, but the manuscript does not supply a checkable proof. Please either turn this into a stated combinatorial lemma with proof, or give the full bookkeeping. In addition, the sentence 'oth","section":"Section 2.3, Lemma 2.25 and its use in (3.4)"},{"comment":"Lemma 2.25 is imported from [74] and the sets J_n are explicitly not defined in the paper ('their construction is quite lengthy and will not be used in this paper'). The lemma is stated for a fixed finite d-tuple of variables, but in Theorem 3.5 it is applied to Q_n, a polynomial in O(2^n) variables. What is needed is a version whose constants are uniform as d,n grow, or an explicit reduction to the fixed-d statement. This uniformity is part of the same load-bearing Step 1/Step 2 estimate of (1.14), so it should be addressed as part of the proof, not merely by referring to [74].","section":"Section 2.3, Lemma 2.25 and its use in (3.4)"}],"minor_comments":[{"comment":"The first paragraph says 'combine Corollary 5.5 with Lemma 5.5'; it should be Lemma 5.6. Also, Section 6 contains a typo 'Lebsegue'.","section":"Section 5.3"},{"comment":"References [30] and [31] appear to be the same entry (same title, same journal, same volume and year). Please check the intended second paper.","section":"References"},{"comment":"The abstract states a constant C=C_ρ for t∈[0,1], while Theorem 1.7 states C=C(ρ,T) for t∈[0,T]. Please align the two statements.","section":"Abstract and Theorem 1.7"},{"comment":"At the beginning of Step 2, the displayed expression for Q_n looks like the polynomial B_n(t)-I-W_N, not |B_n(t)-I-W_N|^{2k}. Clarify the notation, even if the subsequent argument implicitly handles the 2k-th power.","section":"Section 3, Step 2"},{"comment":"The displayed definition of c_{ρ,T} after the Markov step is garbled in the manuscript; please rewrite the sup expression cleanly.","section":"Proof of Theorem 1.7"}],"recommendation":"major_revision","confidential_remarks":"The central theorem is very likely correct and important, and the overall strategy is convincing. However, the proof of the key small-time concentration estimate is not checkable in its present form: the combinatorial Step 2 in Theorem 3.5 is too condensed, and it relies on a uniformity statement for an imported expansion that is not stated. I would want to see that part expanded before publication. Since Lemma 2.25 comes from a paper by one of the authors, an independent check of the uniformity in n and in the number of variables is especially advisable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is the real thing. The paper proves almost-sure convergence of the empirical eigenvalue law of Brownian motion on GL(N,C) to the Brown measure of the free multiplicative Brownian motion, for the full two-parameter family of invariant metrics and general invertible initial conditions. That is exactly Biane's conjecture, not the weaker *-distribution convergence that previous work established. The gap between traces of polynomials and eigenvalues of non-normal matrices is real, and the authors are honest about it.\n\nWhat's new is the small-time approximation: the Euclidean increment I+W(t) captures the multiplicative Brownian motion B(t) up to an error with exponentially small probability, with a precise N^{2/3} exponent. That concentration bound is the engine that powers the Hermitization machinery, and it is a substantial technical contribution on its own. The overall architecture — *-distribution convergence, small-singular-value control, then Bordenave–Chafai style log-potential convergence — is coherent and standard in the good sense. No fitting, no invented entities; the limit is the Brown measure constructed by Brown's method. The simulations are presented as numerical evidence, not proof, which is appropriate.\n\nSoft spots: the proof of Theorem 3.5, the moment bound behind the N^{2/3} exponent, is genuinely compressed. Step 2 (pp. 34–37) asserts rather than demonstrates the key counting inequality 2r1+r2 ≥ 4k−4i and the resulting k^4 exponent; 'one also has similar formulas' covers a lot of differentiations. The stress-test note's worry that a k^5 exponent would break the whole theorem is not paranoid. The argument may well be right, but the text does not give a checkable proof of the constant bookkeeping. A referee who knows Parraud's asymptotic expansion and this kind of counting needs to verify this carefully. The other soft spot is small: the abstract says 'fully resolves' Biane's conjecture, but Biane's quoted phrase was 'limit in distribution,' which is trace-level; the eigenvalue-law reading is standard, but a precise citation would be cleaner.\n\nThis is a serious paper for the random matrix/free probability crowd. I'd bring it to reading group, and I'd cite it as soon as it's validated. Send it to referees who know the GUE-moment technology; the recommendation is accept after the Section 3 estimate is pinned down.","headline":"Strong paper that likely closes Biane's conjecture at the eigenvalue level, powered by a genuinely new small-time concentration estimate; the key moment bound in Section 3 is condensed and needs referee scrutiny before acceptance.","tokens_in":109352,"tokens_out":1922,"would_cite":true,"duration_ms":25134,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","46L54","15B52"],"pacs":[],"model":"deepseek-v4-flash","headline":"For every fixed time, the eigenvalue distribution of Brownian motion on GL(N,C) converges almost surely to the Brown measure of the free multiplicative Brownian motion.","keywords":["Brownian motion on GL(N,C)","free multiplicative Brownian motion","Brown measure","eigenvalue convergence","Hermitization","non-normal random matrices","singular value concentration","unitarily invariant diffusions"],"falsifier":"Directly test the small-time tail bound for small N: for N=8 and t=0.1, estimate P(||B(t)−I−W(t)|| ≥ δ) and check whether it is bounded by (Ct/δ)^{N^{2/3}} with a numerical constant C. Alternatively, simulate B_0^N B^N(t) for N=2000 with a fixed invertible initial condition and a large time t, and compare the empirical eigenvalue histogram with the analytic boundary and density of the Brown measure of b0b(t); a persistent mismatch away from the boundary would contradict Theorem 1.6 as stated.","tokens_in":108020,"feed_emoji":"🎲","tokens_out":4839,"duration_ms":54172,"temperature":0.7,"pith_summary":"This paper claims that, for every fixed time t, the random eigenvalues of a broad family of Brownian motions on the complex general linear group GL(N,C) — not just traces of polynomial functions, but the full empirical eigenvalue density — converge almost surely to a fixed, explicitly identifiable probability measure. That limit is the Brown measure of the free multiplicative Brownian motion b0b(t), a free-probability object built from freely independent semicircular Brownian motions. The result covers all unitarily invariant (\"elliptic\") Brownian motions on GL(N,C), parametrized by ϱ>0 and ζ with |ζ|<ϱ, and any uniformly bounded invertible initial condition with its own large-N limit. It closes a gap left open since 1997, where convergence of trace moments was known but convergence of eigenvalues — much harder for non-normal matrices — was not. The proof's engine is a new quantitative statement: for small times, the multiplicative Brownian motion is very close in operator norm to a single additive Gaussian increment, with failure probability at most (Ct/δ)^{N^{2/3}}.","feed_headline":"Brownian motion on GL(N,C) eigenvalues have a deterministic limit","feed_subtitle":"For every fixed time, the random eigenvalue cloud converges to the Brown measure of a free stochastic process, settling a 1997 conjecture.","key_machinery":"The small-time affine approximation of Theorem 1.7 — B(t) ≈ I+W(t) with tail (Ct/δ)^{N^{2/3}} — is the load-bearing mechanism. It is proved by first replacing B(t) by a dyadic product of i.i.d. Gaussian increments (a Donsker-type random walk), proving an N-independent L^2 convergence rate for the approximation, and then using an iteration of Gaussian integration by parts (Schwinger–Dyson equations) and a borrowed asymptotic moment expansion to compare the product to a single increment. The N^{2/3} exponent arises from a trace moment bound E tr_N |B(t)−I−W(t)|^{2k} ≤ (tC)^{2k} exp(C k^4/N^2), choosing k ~ N^{2/3}. This one mechanism converts the curved, non-normal diffusion into an additive G","core_discovery":"The central claim is Theorem 1.6: let B_{ϱ,ζ}^N be any nondegenerate, unitarily invariant Brownian motion on GL(N,C), let B_0^N be independent, invertible, uniformly bounded, and converging almost surely in *-distribution to b0, and let b(t) be the free multiplicative Brownian motion. Then almost surely the empirical eigenvalue measure of B_0^N B^N(t) converges weakly to the Brown measure of b0b(t). The proof follows Girko's Hermitization: convergence of traces gives convergence of singular-value measures of shifted matrices, and the new ingredient is enough control of the small singular values to make the log-integral pass from smooth test functions to log. This control comes from Theorem 1","pith_inferences":["A natural extension is to probe the edge of the limiting eigenvalue cloud at scales N^{-2/3} rather than N^{-1/2}: the N^{2/3} concentration exponent suggests that outlier or boundary-rigidity thresholds for non-normal multiplicative processes may differ from the Hermitian random matrix scale.","The small-time affine approximation is a template that should transfer to other unitarily invariant matrix diffusions and matrix random walks: the essential requirement is an N-independent dyadic L^2 approximation rate, after which Gaussian anti-concentration tools can be imported into multiplicative problems.","If the moment-expansion lemma that supplies the k^4/N^2 control were made fully self-contained, it would likely improve the mesoscopic scale N^{-2/11} in the paper's alternative Wegner estimate and sharpen the constants in the Hermitization argument.","The theorem identifies a deterministic limit but does not compute its density for general non-unitary initial conditions; the paper's stated support conjecture becomes directly testable by simulating the same initial conditions at large N and comparing the cloud boundary with the predicted conformal image region."],"forward_implications":["The scattered eigenvalue cloud of B_0^N B^N(t) has, for every fixed time, a deterministic limiting density that can in principle be computed by solving the known PDE formulas for the Brown measure of the free multiplicative Brownian motion.","The same Hermitization machinery yields quantitative singular-value control at mesoscopic scales: the smallest singular values of B(t)−zI are bounded below by polynomial-in-N quantities with high probability, for every z in the plane.","The result validates finite-N eigenvalue simulations as evidence about free multiplicative Brownian motion: for N=2000 the eigenvalue scatter tracks the analytic support boundaries of the Brown measure.","It completes the large-N spectral theory of the Segal–Bargmann–Hall construction: both trace-level and eigenvalue-level convergence now hold for the heat kernel measures on GL(N,C)."],"fun_headline_variants":["1997 Biane conjecture resolved: Brownian eigenvalues converge almost surely","Eigenvalue law of Brownian motion on GL(N) is deterministic, 1997 conjecture settled","Free Brownian motion gives the limit eigenvalue law for GL(N) Brownian motion","Almost sure convergence of Brownian eigenvalues on GL(N) to a free Brown measure"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole proof leans on an asymptotic expansion of Gaussian-matrix moments quoted without proof from earlier work; if that expansion, applied to |B_n^N(t)−I−W^N(⌊2^n t⌋/2^n)|^{2k}, fails to yield the claimed k^4/N^2 error term, the small-time concentration theorem — and with it the eigenvalue convergence — collapses; the initial condition also requires a uniform spectral gap K^{-1} ≤ σ_min(B_0^N) ≤ σ_max(B_0^N) ≤ K.","fun_headline_variants_meta":{"raw":{"variants":["1997 Biane conjecture resolved: Brownian eigenvalues converge almost surely","Eigenvalue law of Brownian motion on GL(N) is deterministic, 1997 conjecture settled","Free Brownian motion gives the limit eigenvalue law for GL(N) Brownian motion","Almost sure convergence of Brownian eigenvalues on GL(N) to a free Brown measure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000714,"raw_usage":{"total_tokens":3097,"prompt_tokens":840,"completion_tokens":2257,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2169}},"tokens_in":584,"tokens_out":2257,"duration_ms":17372,"temperature":1.0,"reasoning_tokens":2169,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:24:21.380270+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Directly test the small-time tail bound for small N: for N=8 and t=0.1, estimate P(||B(t)−I−W(t)|| ≥ δ) and check whether it is bounded by (Ct/δ)^{N^{2/3}} with a numerical constant C. Alternatively, simulate B_0^N B^N(t) for N=2000 with a fixed invertible initial condition and a large time t, and compare the empirical eigenvalue histogram with the analytic boundary and density of the Brown measure of b0b(t); a persistent mismatch away from the boundary would contradict Theorem 1.6 as stated.","supporting_citations":[],"review_version":1}