{"id":"043b81fa-459f-4e28-8169-2890e973abed","arxiv_id":"2505.11948","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A single spectral formula for correlated matrix products reduces to both the Marchenko-Pastur law and the elliptic law, and predicts nonmonotonic memory-network stability as a function of stored pattern count.","lead":"The paper derives a closed-form eigenvalue density for products of two correlated random matrices, a single formula whose limits include both the Marchenko-Pastur law and the elliptic law. It then uses the spectrum to show that hetero-associative memory networks can lose and regain stability as the number of stored patterns grows.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stated elliptic-law limit in Eqs. (12)/(54) is over-normalized by a factor 1/(1-τ^4), so the paper's advertised unification with the elliptic law is incorrect as written.","rationale":"I considered the replica-exchange step flagged by the reader as the weakest assumption. It is indeed unproved, but it is a standard technique and Eq. (9) already has nontrivial numerical support in Fig. 2 and correct limiting checks (τ=0 gives the product-Ginibre density; τ→1 gives Marchenko-Pastur). The most concrete, directly falsifiable flaw is the miscomputed elliptic-law limit: as printed, Eq. (12)/(54) violates normalization and therefore cannot be the elliptic law. This is a load-bearing part of the paper's central unification claim, and it must be corrected. Since the underlying Eq. (9) correctly yields 1/[π(1-τ^4)] when expanded about the actual support center, the issue is local and fixable, so a conditional verdict remains appropriate rather than acceptance or rejection.","tokens_in":15375,"tokens_out":26882,"duration_ms":272679,"concrete_test":"Recompute Eq. (54) by taking α→∞ directly in the two-dimensional density Eq. (9), using the correct shift X=τ(β+1/β). The asymptotic constant is 1/[π(1-τ^4)], and ∫ρ over the support ellipse equals 1; this exposes the missing factor (1-τ^4) in the printed Eq. (12)/(54). A finite-α simulation (e.g., N=2000, τ=0.5, α=400) will confirm the corrected prefactor and reject the printed one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim requires Eq. (9) to reduce to the elliptic law as α→∞. The printed limit does not. Taking α→∞ in the full two-dimensional density (not the y-marginal used in Eq. (49)): at the support center X=τ(β+1/β), the denominator in Eq. (9) is β(1+τ^2)/2+O(1), so ρ_b(ω+τβ) tends to 1/[π(1-τ^4)] inside ((x/(1+τ^2))^2+(y/(1-τ^2))^2)<1. Eq. (12)/(54) instead gives 1/[π(1-τ^4)^2]; integrating over the support ellipse of area π(1-τ^4) yields 1/(1-τ^4)>1 for τ≠0. Hence the stated 'elliptic law' is not a probability density, and the claimed unification fails as written. The underlying Eq. (9) appears to give the correct limit if the prefactor is fixed, so this is a concrete internal inconsistency in a central claim, independent of the replica-exchange assumption.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the ensemble J = UV^T/sqrt(NM) with correlated Gaussian N x M factors U,V, and derives the limiting bulk spectral density rho_b(omega) shown in Eq. (9). The authors argue that this single formula unifies the Marchenko-Pastur law (tau -> 1) and the elliptic law (alpha -> infinity), and they apply the support edge to determine the linear stability of a hetero-associative memory network, yielding a nonmonotonic stability condition in the pattern number M. The main formula is checked against direct numerical diagonalization, and the stability boundary is compared to simulations of the network dynamics.","tokens_in":15616,"tokens_out":20897,"duration_ms":193392,"significance":"If Eq. (9) is correct, this is a valuable contribution: it gives a closed-form spectral density for a correlated product ensemble that interpolates between two cornerstone laws of random matrix theory, and it makes a falsifiable prediction for the stability of a recurrent neural network storing correlated key-value pairs. The numerical tests of Eq. (9) are genuine and independent, and the derivation in the supplement is coherent enough to reproduce the known product-Ginibre and Marchenko-Pastur limits. However, the advertised elliptic-law limit contains a normalization error as printed, and the replica-decoupling step that underpins the derivation is asserted rather than proved for this ensemble; these issues prevent the current version from fully establishing the headline unification claim.","major_comments":[{"comment":"The displayed limit to the elliptic law is over-normalized. Taking beta -> infinity in Eq. (9) at the interior point x=0,y=0 of the shifted support gives beta / [2 pi (1-tau^2)] * 1 / sqrt( ((1-tau^2) beta / 2)^2 + (tau beta)^2 ) -> 1/[pi(1-tau^4)], not 1/[pi(1-tau^4)^2]. Since the support ellipse in Eq. (12) has area pi(1-tau^4), the printed constant integrates to 1/(1-tau^4) > 1 for tau != 0; the stated 'elliptic law' is therefore not a probability density. This is a load-bearing error because the unification with the elliptic law is a central advertised result. Please correct the prefactor in Eq. (12) and Eq. (54) and re-check any text or figures that rely on the numerical value of this constant.","section":"Eq. (12) and Supplementary Eq. (54)"},{"comment":"The derivation of Eq. (9) depends on interchanging the ensemble average and the logarithm, which the text justifies by asserting that 'replicas decouple in this limit' and citing references [8,15,16,38-42]. This is a genuinely load-bearing step for the correlated product ensemble defined by Eq. (2); if replica decoupling fails for the U-V correlation structure, Eq. (9) and the subsequent stability edge Eq. (14) do not follow. The manuscript should supply the replica calculation for this ensemble or state precisely which hypotheses of the cited results are satisfied here.","section":"Supplementary A.1, Eqs. (16)-(29)"}],"minor_comments":[{"comment":"The marginalization formula appears to contain a spurious factor (1+tau^2). Direct integration of Eq. (45) gives beta/[pi(1-tau^2)] * tanh^{-1}(Y/sqrt(k^2+x^2+Y^2)), not beta(1+tau^2)/[pi(1-tau^2)] times the same factor. The later Jacobian in Eq. (50) appears to compensate for this, but the printed formula should be corrected to make the derivation reproducible.","section":"Supplementary B, Eq. (49)"},{"comment":"The sentence 'the density function is symmetric about the origin, whereas its boundary contour is determined by an ellipse that is not necessarily centered at the origin' is confusing: the algebraic expression is centrally symmetric, but since the support is a shifted ellipse, the full density including its support is not symmetric about the origin. Please reword.","section":"Main text near Eq. (9)"},{"comment":"The phrase 'the ellipse is shifted by tau beta from the origin' is confusing in the context of Eq. (12), where the displayed support is centered at the origin; the shift refers to the original coordinates of Eq. (9). Please clarify the coordinate convention.","section":"Eq. (12) and Fig. 3B caption"},{"comment":"Reference [37] is cited as 'in preparation'; if the paper relies on this work, please provide a preprint or remove the citation.","section":"Reference [37]"}],"recommendation":"major_revision","confidential_remarks":"The prefactor error in Eq. (12)/(54) is a concrete algebraic slip in a headline claim; with the corrected constant, Eq. (9) would give the announced elliptic-law limit, so I view this as fixable within the scope of a revision rather than a reason to reject. The replica-decoupling issue is more fundamental and should be addressed head-on, even if only by a careful statement of the applicable known results. The numerical figures do not appear to test the prefactor quantitatively, which is likely why the error was not caught."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core spectral formula is the genuine article: a closed-form bulk density for J = UV^T/sqrt(NM) with correlated Gaussian U, V, interpolating between the product-Ginibre and Marchenko-Pastur limits. Second, the advertised elliptic-law reduction is wrong as printed. Taking alpha->infinity in Eq. (9) gives a constant 1/[pi(1-tau^4)] on the support ellipse, not 1/[pi(1-tau^4)^2] as in Eqs. (12) and (54); the printed density integrates to 1/(1-tau^4) > 1 for tau != 0, so it is not a probability density. The same slip appears twice, so it is likely algebra rather than a flaw in Eq. (9), but it undermines the 'unification' claim until fixed.\n\nWhat is actually new: the two-parameter correlated-product density for finite alpha and general tau. The tau=0 case is a known product-Ginibre spectrum and tau->1 gives Marchenko-Pastur, but the interpolating form, including the off-center ellipse, is not in the cited literature. The derivation is coherent: a saddle-point treatment of the potential, and the numerical histograms (N=5000, several alpha) match Eq. (9) with no fitted parameters. The stability application is clean too: the right edge of the support gives lambda = g*tau*(beta + 1/beta) + g*(1+tau^2) - 1, and the nonmonotonic dependence on pattern count is confirmed by simulation. That part held up.\n\nSoft spots, in proportion. The elliptic-limit prefactor error is the main one; it is concrete and central, but small to repair. The replica decoupling step in Supplementary A.1 is asserted with citations rather than proved for this ensemble; for a physics paper I would let it pass, since the numerics independently back Eq. (9), but the authors should state clearly that this is an assumption, not a theorem. The 'universal ensemble' framing overreaches: this is one specific Gaussian ensemble, not a universality class. And there is no code or data release; minor, since the simulations are simple, but a check of the normalization in Fig. 3B would have caught the elliptic-limit issue.\n\nWho this is for: random-matrix people working on correlated non-Hermitian ensembles, and the associative-memory / linear-attention community will find the stability criterion worth reading. It deserves a serious referee, not a desk reject. My recommendation: send to review, and require the prefactor correction, a toned-down universality claim, and one sentence acknowledging the replica assumption.","headline":"Eq. (9) is a genuinely useful closed-form density for a two-parameter correlated product ensemble, and the stability story is clean, but the printed alpha->infinity 'elliptic law' limit is over-normalized by a factor 1/(1-tau^4), which is a real error in the paper's central unification claim, though a fixable one.","tokens_in":16124,"tokens_out":5298,"would_cite":true,"duration_ms":49381,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","15B52","82C32"],"pacs":[],"model":"deepseek-v4-flash","headline":"For products of correlated Gaussian matrices, the paper derives the exact spectral density, unifies the Marchenko-Pastur and elliptic laws, and predicts nonmonotonic stability in hetero-associative memory networks.","keywords":["random matrix theory","Marchenko-Pastur law","elliptic law","spectral density","correlated random matrices","hetero-associative memory","linear attention","neural network stability"],"falsifier":"Draw one large realization of the ensemble, for example $N=4000$, $M=2000$, and $\\tau=0.8$, and compare the empirical eigenvalue histogram with Eq. (9) inside the predicted ellipse and with zero outside; the claimed $1/\\sqrt{k^2+x^2+y^2}$ profile, with $k=(1-\\tau^2)(\\beta-1/\\beta)/2$, and the sharp elliptical cutoff are the direct check. In parallel, measure the largest real-part eigenvalue for several $\\beta$ and compare it with $\\tau(\\beta+1/\\beta)+1+\\tau^2$ to test the nonmonotonic stability prediction.","tokens_in":15166,"feed_emoji":"🧠","tokens_out":16370,"duration_ms":142858,"temperature":0.7,"pith_summary":"The paper establishes a single random-matrix ensemble whose bulk eigenvalue density interpolates between two foundational limiting laws: the Marchenko-Pastur law and the elliptic law. The ensemble is the product $J=UV^\\top/\\sqrt{NM}$ of two $N\\times M$ Gaussian matrices whose corresponding entries are correlated with coefficient $\\tau$, and the paper derives the exact density in the limit $N,M\\to\\infty$ with $M/N=\\alpha$. The explicit formula recovers the Marchenko-Pastur law as $\\tau\\to 1$ and the elliptic law as $\\alpha\\to\\infty$. Because such products are the connectivity matrices of hetero-associative memories, the result also predicts that a recurrent network storing correlated key-value pairs has a stability threshold that depends nonmonotonically on the number of stored patterns.","feed_headline":"One formula unifies the Marchenko-Pastur and elliptic laws","feed_subtitle":"The same law predicts associative-memory networks can lose and regain stability as patterns accumulate.","key_machinery":"The central object is the correlated product ensemble $J=UV^\\top/\\sqrt{NM}$ with the joint Gaussian law of Eq. (2). The argument is carried by the potential function $\\Phi(\\omega)=-\\frac{1}{N}\\langle\\log\\det((\\omega^*I-J^\\top)(\\omega I-J))\\rangle$, whose derivative with respect to $\\omega$ gives the disorder-averaged resolvent, and whose $\\omega^*$ derivative gives the density. Evaluating the averaged determinant by a saddle-point approximation yields two branches: inside an ellipse the saddle-point solution has $\\sigma\\sim\\sqrt{\\varepsilon}$, and the Green's function develops $\\omega^*$-dependence that produces the explicit density; outside, the Green's function depends only on $\\omega$, so the density vanishes. The two parameters $\\tau$ and $\\alpha$ interpolate between the classical laws: $\\tau\\to1$ degenerates the ellipse onto the real axis and yields Marchenko-Pastur, while $\\alpha\\to\\infty$ flattens the density into the elliptic law.","core_discovery":"The central discovery is a closed-form bulk spectral density for $J=UV^\\top/\\sqrt{NM}$, where $U$ and $V$ are $N\\times M$ standard Gaussian matrices with $\\langle U_{ij}V_{ij}\\rangle=\\tau$. Writing $\\omega=x+iy$, $\\alpha=M/N$, and $\\beta=\\sqrt{\\alpha}$, the density is $$ \\rho_b(\\omega)=\\frac{\\$\\beta$}{2\\pi(1-\\$tau^{2}$)}\\left[\\left(\\frac{1-\\$tau^{2}$}{2}\\left(\\$\\beta$-\\frac{1}{\\$\\beta$}\\right)\\right)^2+$x^{2}$+$y^{2}$\\right]^{-1/2} $$ inside the ellipse $$ \\left(\\frac{x-\\tau(\\$\\beta$+1/\\$\\beta$)}{1+\\$tau^{2}$}\\right)^2+\\left(\\frac{y}{1-\\$tau^{2}$}\\right)^2<1, $$ and zero outside. When $\\alpha<1$, the remaining $N-M$ eigenvalues are exactly zero, producing an additional delta peak. The paper derives this density from the disorder-averaged log-determinant potential through a saddle-point approximation, validates it against numerical diagonalization, and reads off the spectral edge to obtain a linear-stability condition for a recurrent tanh network.","pith_inferences":["In a linear-attention layer, the effective weight matrix has the same product-of-correlated-factors form as $J$, so Eq. (9) predicts the bulk spectrum of finite-width attention; this could be tested on the weight matrices of trained attention models.","Near $\\alpha=1$, the squared offset $k^2$ in the density vanishes, so the spectrum becomes strongly concentrated near the origin; in a recurrent network this implies slow transients in memory retrieval around the capacity sweet spot, a consequence the paper does not develop.","Because the sign of $\\tau$ reverses the direction of nonmonotonic stability, anti-correlated key-value pairs should make the network unstable for intermediate memory loads but stable at small and large loads; this is a numerical prediction for small recurrent circuits."],"forward_implications":["At finite size the spectrum of a single matrix already follows the predicted elliptical bulk, so the ensemble is self-averaging and Eq. (9) is a direct numerical prediction.","When $\\alpha=M/N<1$, the matrix has exactly $N-M$ zero eigenvalues; the bulk formula accounts for the remaining $M$ eigenvalues, and its integral is $\\min(1,\\alpha)$.","Taking $\\tau\\to1$ collapses the support onto the real axis and recovers the Marchenko-Pastur density, while $\\alpha\\to\\infty$ recovers the elliptic law; the Wigner semicircle and circular laws follow from these limits as further special cases.","For the recurrent network in Eq. (13), the largest real part in the spectrum is $\\tau(\\beta+1/\\beta)+1+\\tau^2$, so the fixed point is stable when $g\\tau(\\beta+1/\\beta)+g(1+\\tau^2)-1<0$; for fixed $\\tau>0$ this holds only in a finite range of $\\beta$, and for $\\tau<0$ it holds outside a finite range.","If $g<1/4$, for every correlation $\\tau$ there exists some number of stored patterns that makes the trivial fixed point stable."],"supporting_citations":[{"why":"Defines the Marchenko-Pastur law that Eq. (9) must reproduce as $\\tau\\to1$; supplies the limiting density used as the comparison target.","marker":"[5]"},{"why":"Defines the elliptic law recovered in the $\\alpha\\to\\infty$ limit; provides the reference spectrum for correlated non-symmetric Gaussian matrices.","marker":"[7]"},{"why":"Supplies the potential-function and log-determinant saddle-point machinery, and the elliptic-law result for non-symmetric matrices, on which the derivation is built.","marker":"[8]"},{"why":"Establishes the resolvent method for eigenvalue spectra of random neural-network matrices; the paper's derivation follows this route.","marker":"[15]"},{"why":"Applies the same spectral-stability reasoning to complex ecosystems, a direct precedent for the stability application in Eq. (14).","marker":"[16]"},{"why":"Introduces correlation-matrix memories, giving the hetero-associative-memory interpretation of $J=UV^\\top/\\sqrt{NM}$.","marker":"[19]"},{"why":"Supplies the auto- and hetero-associative memory network model, including the $\\tau=1$ symmetric case, that motivates the ensemble.","marker":"[20]"},{"why":"Founds the attention architecture that the paper identifies with linear attention, grounding the claim that the ensemble is a linear-attention connectivity matrix.","marker":"[34]"}],"fun_headline_variants":["Unified spectrum law predicts memory networks lose and regain stability","Unified matrix law reveals nonmonotonic memory stability","One formula explains random matrix spectra and memory stability","Correlated random matrices yield one law for memory stability"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, for this correlated product ensemble, the ensemble average of the log-determinant can be replaced by the log of the averaged determinant in the large-system limit; that interchange, taken from earlier random-matrix work, is not proved for the correlated case here, and if it fails the explicit density and the stability threshold derived from it do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Unified spectrum law predicts memory networks lose and regain stability","Unified matrix law reveals nonmonotonic memory stability","One formula explains random matrix spectra and memory stability","Correlated random matrices yield one law for memory stability"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001201,"raw_usage":{"total_tokens":4971,"prompt_tokens":988,"completion_tokens":3983,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":3919}},"tokens_in":604,"tokens_out":3983,"duration_ms":26273,"temperature":1.0,"reasoning_tokens":3919,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:46:36.749869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Draw one large realization of the ensemble, for example $N=4000$, $M=2000$, and $\\tau=0.8$, and compare the empirical eigenvalue histogram with Eq. (9) inside the predicted ellipse and with zero outside; the claimed $1/\\sqrt{k^2+x^2+y^2}$ profile, with $k=(1-\\tau^2)(\\beta-1/\\beta)/2$, and the sharp elliptical cutoff are the direct check. In parallel, measure the largest real-part eigenvalue for several $\\beta$ and compare it with $\\tau(\\beta+1/\\beta)+1+\\tau^2$ to test the nonmonotonic stability prediction.","supporting_citations":[{"cited_title":"DISTRIBUTION OF EIGENVALUES FOR SOME SETS OF RANDOM MA- TRICES","cited_arxiv_id":null,"evidence_quote":"Defines the Marchenko-Pastur law that Eq. (9) must reproduce as $\\tau\\to1$; supplies the limiting density used as the comparison target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the elliptic law recovered in the $\\alpha\\to\\infty$ limit; provides the reference spectrum for correlated non-symmetric Gaussian matrices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the potential-function and log-determinant saddle-point machinery, and the elliptic-law result for non-symmetric matrices, on which the derivation is built."},{"cited_title":"A random matrix perspective on mixtures of nonlinear- ities in high dimensions","cited_arxiv_id":null,"evidence_quote":"Establishes the resolvent method for eigenvalue spectra of random neural-network matrices; the paper's derivation follows this route."},{"cited_title":"Characteristics of random nets of ana - log neuron-like elements","cited_arxiv_id":null,"evidence_quote":"Applies the same spectral-stability reasoning to complex ecosystems, a direct precedent for the stability application in Eq. (14)."},{"cited_title":"Baron and Tobias Galla","cited_arxiv_id":null,"evidence_quote":"Introduces correlation-matrix memories, giving the hetero-associative-memory interpretation of $J=UV^\\top/\\sqrt{NM}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the auto- and hetero-associative memory network model, including the $\\tau=1$ symmetric case, that motivates the ensemble."},{"cited_title":"Sta- tistical mechanics of temporal association in neural net- works with transmission delays","cited_arxiv_id":null,"evidence_quote":"Founds the attention architecture that the paper identifies with linear attention, grounding the claim that the ensemble is a linear-attention connectivity matrix."}],"review_version":1}