{"id":"5f8fe104-0ce1-4dd1-ac67-4b932bc9d634","arxiv_id":"2502.06414","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Scaled random hives with GUE boundary conditions converge in probability to a unique continuum hive whose value at a point v is the supremum of a functional over asymptotic height functions of lozenge tilings.","lead":"Random hives with GUE boundary conditions are shown to converge, in probability after scaling, to a single continuum hive, with the limiting value at each point given by a variational formula over lozenge tilings. The result upgrades an earlier subsequential concentration result and connects the randomized Horn problem to the surface tension of random tilings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 9.1's covariance bound does not follow from the quoted Theorem A.5, so the upper bound for Theorem 8.5 is unproved.","rationale":"The reader's verdict is CONDITIONAL and identifies the variance bound (Proposition 7.1 / Theorem A.5) as the weakest assumption. I agree that this is the most load-bearing concern: the upper-bound half of the main theorem is used for every point v, and the proof of the crucial variance decay is condensed into a two-line reduction to an unpublished one-dimensional result. Concretely, Theorem A.5 bounds the variance of a single linear combination of consecutive interlacing gaps along one row; the operator norm of the full patch covariance matrix requires cross-row control that is neither stated nor derived. A secondary, also genuine, gap is Lemma 8.11's nondegeneracy assumption, which is absent from the theorem statement and leaves boundary points v unproved; however, this affects only a boundary set and could be repaired by an approximation argument or by restricting the statement. The variance issue affects the validity of the upper bound for all v, so it is the primary concern. The recommended verdict remains CONDITIONAL: the paper should not be rejected, but it needs a complete proof of Lemma 9.1 or a different variance argument, and the boundary case of Lemma 8.11 must be addressed.","tokens_in":44450,"tokens_out":12266,"duration_ms":106119,"concrete_test":"Independently derive Lemma 9.1 from the GUE minor process determinantal kernel (or from Tao's Theorem A.5 plus Theorem A.4): for a vector a supported on two distinct rows of a (2m+1) patch, compute Cov(sum_l a_l gtilde^{(r)}_{i+l}, sum_l b_l gtilde^{(s)}_{j+l}) with r != s and show it is O((m / log^C m)) ||a|| ||b||. If no such bound is obtained, Lemma 9.1 is unsupported. As a numerical cross-check, simulate an n=5000 GUE minor process, take patches of size m = 5, 10, 20, 40 centered at fixed bulk indices, compute the sample covariance matrix of the interlacing gaps, and estimate its operator norm; if the operator norm does not scale like O(m^2 / log^C m) (or at least O(m^{2-delta})), Proposition 7.1's conclusion fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The upper bound argument in Section 8.4 hinges on Proposition 7.1, which bounds the variance of the maximum lozenge weight over an m-scale patch by O(m^4 / log^C m). Its proof reduces to Lemma 9.1, which asserts that the covariance matrix M_m of the O(m^2) interlacing gaps in a (2m+1) patch has operator norm O(m^2 / log^C m), with the justification: 'This follows from Theorem A.5, taking the parameter A in it to be sufficiently large.' Theorem A.5, however, is a one-dimensional statement: for fixed m, it bounds the variance of sum_{l=1}^m a_l gtilde_{i+l} for consecutive interlacing gaps along a single level. It does not control covariances between gaps in different rows of the patch, nor does it yield a uniform operator-norm bound for the full 4m^2 x 4m^2 covariance matrix. The trivial operator norm of M_m is O(m^2); the claimed polylogarithmic saving is precisely what makes the fluctuation term in Claim 4 be o_epsilon(n^2). Without an argument showing that cross-row correlations decay at the same rate, Proposition 7.1 is not established. Since Lemma 8.7's upper bound feeds directly into Theorem 8.5 for every limit point v, this is a load-bearing gap in the proof of the main theorem.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies the augmented hives arising from the spectra of two independent scaled GUE matrices, the model that encodes the Horn probability measure for the sum of two random Hermitian matrices. The main theorem (Theorem 8.5) states that for such a hive h_n, at any vertex v_n with v_n/n → v, the normalized expectation n^{-2} E_n h_n(v_n) converges to sup_{f^♯ ∈ AHT^∞_v} S_v(f^♯), where S_v is an explicit functional of asymptotic height functions on the infinite excavation hexagon 7^∞_v; Corollary 8.6 upgrades this to convergence in probability. This would settle the uniqueness of the continuum limit left open by the author's earlier joint work with Sheffield and Tao [30], which established only subsequential convergence and vanishing variance. The proof scheme is to express the hive value as the maximum lozenge weight on an excavation hexagon, then bound this maximum from above (Sections 8.2–8.4) via a quantitative-differentiation decomposition into dyadic triangles, local variance estimates (Proposition 7.1), and the surface tension σ_⋄, and from below (Section 8.5) by explicit patching of height functions of prescribed tilt near a candidate maximizer, with separate treatment of the equator and the hexagon weight in Lemmas 8.1 and 8.2. The argument draws on fixed-index universality results for the GUE minor process quoted from Tao's preprint [40] (Theorems A.3–A.5) and on Fefferman's quantitative differentiation theorem, whose proof is reproduced in Appendix B.","tokens_in":44676,"tokens_out":29957,"duration_ms":249522,"significance":"If Theorem 8.5 is proved, it is a significant advance for the theory of random hives and for the randomized Horn problem: it converts the subsequential compactness of [30] into genuine convergence in probability to a single continuum limit, and it identifies the limit pointwise by a variational formula that is explicit (a supremum of integrals of the surface tension σ_⋄, itself defined from fixed-index GUE minor patches and lozenge tilings) and falsifiable by the existing simulations of [11,12]. The paper has real structural strengths: the two-sided bounds are organized around quantitative differentiation (Fefferman's theorem, proved in Appendix B), the fixed-index convergence of the GUE minor process is supported by a self-contained argument in Appendix A.4, and the surface-tension calculus of Section 11 (subadditivity, convexity, continuity, and ψ-independence) is developed in unusual detail. The paper also makes a parameter-free statement in the end: the auxiliary cutoff ψ(m) = (log m)^{-1/10} of Definition 14 is shown in Lemma 11.4 not to affect the limiting value.","major_comments":[{"comment":"The bound ∥M∥_op = O(m²/log^C m) for the 4m²×4m² covariance matrix of interlacing gaps in a (2m+1)×(2m+1) patch is the load-bearing estimate for the upper bound, but the proof ('This follows from Theorem A.5, taking the parameter A in it to be sufficiently large') is ineffective. Theorem A.5 is one-dimensional and stated for fixed m: it bounds the variance of Σ_{l=1}^m a_l g̃_{i+l} for consecutive interlacing gaps along a single level of the minor process, and its right-hand side is a constant for fixed m, with no growing-m content. It therefore cannot control even a single row of the patch whose length is 2m+1, and it says nothing about covariances between gaps in different rows, which make up most of the 4m²×4m² matrix; the cross-row structure is precisely what a spectral-norm bound must control. The trivial bound on ∥M∥_op is O(m²), and without the polylogarithmic saving the variance estimate in the proof of Claim 4 (§8.4) gives O(n²) instead of o_ε(n²), so Lemma 8.7 and Lemma 8.9 are not established. The same Proposition 7.1 is used in the proof of Lemma 11.10, so the gap also propagates into Lemma 11.11, Proposition 7.5 and Proposition 8.4. I agree with the stress-test assessment: a genuinely two-dimensional covariance bound for fixed-index interlacing gaps (for example, from the determinantal kernel of the GUE minor process or from correlation decay in the bead model of [4]) must be supplied; the quoted Theorem A.5 cannot yield it.","section":"§9, Lemma 9.1, used by Prop. 7.1 and Claim 4"},{"comment":"Theorem 8.5 is asserted for every vertex limit v ∈ T, but the lower-bound Lemma 8.11 is proved only under the hypothesis that both trapezoids of the infinite excavation hexagon 7^∞_v contain Euclidean balls of positive radius, and this hypothesis is not stated in Theorem 8.5 or in Lemma 8.12, which appeals to Lemma 8.11 in the proof of the theorem. The hypothesis fails on the boundary of T: for v = (0, β) with 0 < β < 1 the vertices of Definition 6 give A = F and C = D, so both trapezoids collapse to line segments, and Claim 5's bound d(L_{T,ḡ}, ∂K) > cε², together with the inclusion HT^n_v(f^{♯,δ}, ε) ⊃ HT^n_v(ḡ_n, εC0) and the subsequent patching, have no geometric slack. Thus the pointwise formula is unproved for boundary points, which are included in the statement of the main theorem. The fix may be local (approximate boundary v by interior points and use continuity of both sides in v, or restrict Theorem 8.5 to the interior), but as written the theorem claims more than the proof establishes.","section":"Theorem 8.5, lower bound via §8.5, Lemma 8.11"}],"minor_comments":[{"comment":"The choice of the constant C0 is left implicit: the line that 2^{-k}C0 → ∞ as ε → 0 is only true if C0 is taken to grow with k (e.g., C0 > 2^{2k}, which is available since 2^k < C0^{1/2} is imposed); this dependence should be stated explicitly, since in the ε → 0 asymptotics the minimal allowed C0 is astronomically large.","section":"§8.4, proof of Claim 3"},{"comment":"The passage from the bound (8.6) for a fixed f^♯ and ε to the final display of Lemma 8.9 (limsup_{n→∞} E n^{-2} h_n(v_n) ≤ sup_{f^♯ ∈ AHT^∞_v} S_v(f^♯)) interchanges a supremum with an ε → 0 limit; this requires an explicit finite ε-covering of the compact space AHT^∞_v by tubes HT^n_v(f^♯_i, εn) with the o_ε(1) errors uniform over the finitely many centers.","section":"§8.4, passage from (8.6) to Lemma 8.9"},{"comment":"The last step identifies the integral against ∂f^{♯,δ} with S_{v,⋄}(f^♯) − o_ε(1), which uses δ = ε/10 → 0 together with convexity of σ_⋄(ρ,·) in the second argument (giving S_{v,⋄}(f^{♯,δ}) ≥ S_{v,⋄}(f^♯) − C_{|ρ|}δ); the text merely cites Lemma 8.3, which addresses mesh approximation rather than the δ-removal, so the step should be justified explicitly.","section":"§8.5, Lemma 8.11, last step"},{"comment":"The proof states that 'choosing m1 = ⌊m3/m0⌋, we see that m1m0(1 + C′ψ(m0)) = m3'; for arbitrary m3 this is not an identity, and the subadditivity inequality in Lemma 11.1 B is only stated for exact multiples with the specified border term. The rounding at both ends must be absorbed into the error terms before the argument feeds into Corollary 11.3 and Lemma 11.4.","section":"§11, proof of Lemma 11.2"},{"comment":"The claimed independence of σ_⋄(ρ, a_∞) from the choice of ψ is proved in a compressed paragraph: the constants in the patching inequality are unspecified, the divisibility assumptions on m1 and m2 are not stated, and the displayed condition 'm2ψ2(m2) > m1ψ1(m1)' appears to swap the roles of ψ1 and ψ2 relative to the contradiction in (11.3). Since this paragraph establishes the parameter-freeness of the limit claimed in the main theorem, it should be expanded into a complete argument.","section":"§11, final paragraph of Lemma 11.4"},{"comment":"The two fixed-index convergence statements are dimensionally inconsistent: the square-patch case is stated as taking values in R^{2m} with the vector 1_{2m}, while the line-segment case is stated in R^{2m+1} with 1_{2m+1}; the number of interlacing gaps in a (2m+1)×(2m+1) box should be fixed consistently, and the notation M|_{⌈ℓx⌉+2m} versus M|_{⌈ℓy⌉+l_m} should be defined with the same conventions.","section":"§5, Theorem 5.1 and Observation 2"},{"comment":"The normalization in §1.4 contains typos that make the scaling convention hard to verify: 'the the distribution of the eigenvalues of A + B are the pushforward', 'Cnσ−n2' (presumably C_n σ^{−n²}), and the expression 'A√σ2n' are all unclear as printed; these should be corrected.","section":"§1.4, normalization around (1.10)–(1.11)"}],"recommendation":"major_revision","confidential_remarks":"To the editor: I concur with the conditional assessment of the prior reader; the stress-test concern about Lemma 9.1 indeed lands, and I would add the boundary-vertex gap in Lemma 8.11. The author should be asked to provide a genuine two-dimensional covariance bound for fixed-index interlacing gaps and either to restrict Theorem 8.5 to interior v or to handle ∂T by an approximation argument. One further scholarly point: the proof depends on Tao's unpublished preprint [40] (Theorems A.3–A.5) and on [24, Theorem 1.4] for the convergence of the hexagon weight in Lemma 8.2; if the manuscript is accepted, the dependence on unpublished material should be resolved, for example by reproducing the needed statements with proofs or by using published versions. This is a provenance concern, not a correctness objection to the mathematical architecture."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main theorem is real progress: it upgrades the subsequential convergence of [30] to an actual limit and identifies that limit as the supremum of a functional over asymptotic height functions. That was an open problem, and the natural proof strategy—quantitative differentiation plus the lozenge-tiling surface tension—is applied with care. The paper reproduces Fefferman's theorem and Tao's proof of Theorem 5.1 in appendices, and the upper-semicontinuity argument in Section 8.2 is sound. Sections 8 and 11 are structured honestly, and the auxiliary function ψ(m) is handled properly: Lemma 11.4 shows the limit is independent of it.\n\nThat said, I would not call the theorem proven as stated. There are two repair-level gaps. First, Lemma 8.11 assumes both trapezoids of the infinite hexagon contain Euclidean balls of positive radius, while Theorem 8.5 is stated for every vertex v. On the boundary of the triangle one trapezoid degenerates, and the lower bound does not follow; the author needs either an approximation argument for boundary vertices or an amended statement. Second, Lemma 9.1 is justified by a one-line citation to Theorem A.5, and the citation does not do the work. A.5 is a single-row, fixed-m variance bound; it says nothing about covariances between gaps in different rows of a growing patch, and it does not give the O(m^2/log^C m) operator-norm bound for the full 4m^2 × 4m^2 covariance matrix. Proposition 7.1 and Claim 4, and hence the whole upper bound, rest on exactly this. The bound may well be true—the bead model should give it—but as written it is unsupported. This is the kind of thing a careful referee would catch, and it is fixable, but it is real.\n\nA smaller issue: key inputs (Tao's preprint [40], personal communications) are not publicly available. The appendices mitigate this, but verification is harder than it should be. The citation pattern is otherwise clean; building on [30] is appropriate.\n\nWho gets value: people working on the randomized Horn problem, random hives, or variational principles for tilings. I would send it to peer review without hesitation, with explicit referee instructions to check Lemma 9.1 and the boundary cases. Expect a major revision.","headline":"Real progress on an open limit problem; two repair-level gaps (boundary cases and a covariance bound) keep the theorem from being proven as stated.","tokens_in":45255,"tokens_out":6288,"would_cite":true,"duration_ms":53686,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","60F05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that random hives with independent GUE boundary conditions converge in probability to a single continuum hive, with the pointwise limit given by a variational formula over asymptotic height functions.","keywords":["random hives","GUE random matrices","Horn problem","lozenge tilings","asymptotic height functions","surface tension","continuum limit"],"falsifier":"Fix an interior point $v$, simulate two independent scaled GUE minor processes up to large $n$, compute $h_n(v_n)$ as the maximum lozenge weight of the excavation hexagon via the octahedron recurrence, and compare $n^{-2}h_n(v_n)$ with the predicted supremum $\\sup_{f^{\\sharp} \\in AHT^{\\infty}_v} S_v(f^{\\sharp})$. If the rescaled values do not converge, or converge to a number different from that supremum for some $v$, the central claim is wrong.","tokens_in":44167,"feed_emoji":"📐","tokens_out":14820,"duration_ms":114167,"temperature":0.7,"pith_summary":"Random hives encode the possible spectra of sums of Hermitian matrices through discrete concave functions on a triangle. This paper treats hives whose boundary data on two sides are the eigenvalues of two independent scaled GUE matrices and whose third side carries the Vandermonde weight from the randomized Horn problem. It proves that, after rescaling by $n^{-2}$, the value at any fixed relative location converges in probability to a single deterministic limit, not merely along a subsequence. The limit is identified as $\\sup_{f^{\\sharp} \\in AHT^{\\infty}_v} S_v(f^{\\sharp})$, a variational supremum over asymptotic height functions of lozenge tilings of a scaled excavation hexagon. If correct, this gives a single explicit continuum limit for random hives and settles the convergence question left open by the earlier concentration result.","feed_headline":"GUE random hives converge to one continuum limit","feed_subtitle":"At every point the scaled hive value matches a variational supremum over lozenge-tiling height functions.","key_machinery":"The load-bearing construction is the excavation hexagon $\\mathcal{H}^n_v$ together with the octahedron recurrence: the hive value $h_n(v)$ is exactly the maximum weight of a lozenge tiling of this hexagon, with lozenge weights read from two independent GUE minor processes. The limiting variational data are the asymptotic height function pairs in $AHT^{\\infty}_v$ and the functional $S_v = S_{v,\\diamond}+S_{v,\\Delta}+S_{v,7}$, whose bulk term integrates a surface tension $\\sigma_\\diamond(\\rho,\\partial f)$ over the hexagon. The upper bound decomposes each height function into dyadic nearly-linear patches using quantitative differentiation and controls the fluctuation of each patch with the patch-variance estimate, while the lower bound patches local optimal matchings together, so the upper and lower estimates meet at the supremum.","core_discovery":"Let $\\lambda_n$ and $\\mu_n$ be the eigenvalues of two independent scaled GUE matrices, and let $a_n$ be sampled from the normalized Lebesgue measure on the augmented hive polytope with those boundary data. The main theorem states that if $v_n/n \\to v$, then $\\lim_{n\\to\\infty} n^{-2} \\mathbb{E}_n h_n(v_n) = \\sup_{f^{\\sharp} \\in AHT^{\\infty}_v} S_v(f^{\\sharp})$, and as a corollary $n^{-2}h_n(v_n)$ converges in probability to the same value. The content is that the continuum hive is unique: the earlier subsequential limit and vanishing variance are upgraded to an exact limit, and the limiting value is the value of a specific variational problem over pairs of asymptotic height functions on the excavation hexagon.","pith_inferences":["The same proof structure should yield a unique continuum limit for any ensemble of boundary data whose minor process is log-concave and has the analogous interlacing-gap covariance decay, not only GUE.","The pointwise variational formula is strong evidence that the rate-function minimizer in the large-deviation principle for augmented hives is unique, since any minimizer must attain this supremum at every point.","One testable extension is to compute the left-hand side numerically by direct simulation of the octahedron recurrence and compare it with an independent numerical evaluation of the variational supremum; agreement at moderate $n$ would support the formula, and disagreement would pinpoint the step to examine."],"forward_implications":["At each fixed relative location $v$, the scaled hive value $n^{-2}h_n(v_n)$ converges in probability to the explicit number $\\sup_{f^{\\sharp} \\in AHT^{\\infty}_v} S_v(f^{\\sharp})$.","The earlier concentration result is upgraded from vanishing variance and subsequential convergence to full convergence to a single continuum hive.","The randomized Horn problem for two independent scaled GUE matrices acquires a well-defined asymptotic law, so numerical studies can compare against a unique limit.","The variational formula gives a computable route to the limiting value, since $S_v$ is an integral of a surface tension over the bulk plus explicit equator and hexagon terms."],"supporting_citations":[{"why":"Establishes the vanishing variance and subsequential convergence for scaled GUE hives that this paper upgrades to a full limit.","marker":"[30]"},{"why":"Supplies the covariance decay bound for interlacing gaps of the GUE minor process at fixed index that underpins the patch-variance estimate.","marker":"[40]"},{"why":"Provides the variational principle and approximation lemma for asymptotic height functions used in the upper-semicontinuity argument.","marker":"[5]"},{"why":"Supplies the height-function framework, asymmetric distance, and extension criteria for lozenge tilings used throughout.","marker":"[13]"},{"why":"Gives the convergence of the GUE minor process to the bead model in a fixed-size square, which defines the local surface tension.","marker":"[1]"},{"why":"Provides the mesoscopic central limit theorem used to show convergence of the hexagon-weight term in the variational formula.","marker":"[24]"},{"why":"Proves the octahedron recurrence relation that identifies the maximum lozenge-tiling weight with the hive value.","marker":"[36]"}],"fun_headline_variants":["Random hives with GUE edges settle to one limit","Unique continuum hive emerges from GUE random hives","GUE hives converge to variational tiling supremum","One limit: random hives meet lozenge tilings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The upper bound requires that the maximum lozenge weight on a patch of side $m$ has variance at most $O(m^4/\\log^C m)$ for every fixed $C$; the paper obtains this from a covariance decay bound for interlacing gaps of the GUE minor process quoted from an unpublished preprint, together with log-concavity. If that patch-level variance does not decay, the variational upper bound on the expected hive value fails.","fun_headline_variants_meta":{"raw":{"variants":["Random hives with GUE edges settle to one limit","Unique continuum hive emerges from GUE random hives","GUE hives converge to variational tiling supremum","One limit: random hives meet lozenge tilings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00015,"raw_usage":{"total_tokens":1161,"prompt_tokens":873,"completion_tokens":288,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":221}},"tokens_in":489,"tokens_out":288,"duration_ms":3301,"temperature":1.0,"reasoning_tokens":221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:30:52.381555+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix an interior point $v$, simulate two independent scaled GUE minor processes up to large $n$, compute $h_n(v_n)$ as the maximum lozenge weight of the excavation hexagon via the octahedron recurrence, and compare $n^{-2}h_n(v_n)$ with the predicted supremum $\\sup_{f^{\\sharp} \\in AHT^{\\infty}_v} S_v(f^{\\sharp})$. If the rescaled values do not converge, or converge to a number different from that supremum for some $v$, the central claim is wrong.","supporting_citations":[{"cited_title":"Sums of GUE matrices and concentration of hives from correlation decay of eigengaps","cited_arxiv_id":null,"evidence_quote":"Establishes the vanishing variance and subsequential convergence for scaled GUE hives that this paper upgrades to a full limit."},{"cited_title":"On the distribution of eigenvalues of GUE and its minors at fixed index","cited_arxiv_id":null,"evidence_quote":"Supplies the covariance decay bound for interlacing gaps of the GUE minor process at fixed index that underpins the patch-variance estimate."},{"cited_title":"A variational principle for domino tilings","cited_arxiv_id":null,"evidence_quote":"Provides the variational principle and approximation lemma for asymptotic height functions used in the upper-semicontinuity argument."},{"cited_title":"Lectures on Random Lozenge Tilings","cited_arxiv_id":null,"evidence_quote":"Supplies the height-function framework, asymmetric distance, and extension criteria for lozenge tilings used throughout."},{"cited_title":"The Dyson Brownian Minor Process","cited_arxiv_id":null,"evidence_quote":"Gives the convergence of the GUE minor process to the bead model in a fixed-size square, which defines the local surface tension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the mesoscopic central limit theorem used to show convergence of the hexagon-weight term in the variational formula."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves the octahedron recurrence relation that identifies the maximum lozenge-tiling weight with the hive value."}],"review_version":1}