{"id":"c3472b71-9dfa-4a9b-b405-3a3c023228be","arxiv_id":"2611.17064","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For row-normalized random matrices with heavy-tailed entries, the limiting eigenvalue distribution is a rotationally invariant law on the unit disk, interpolating between the circular law and a Dirac mass at zero.","lead":"This paper proves that the eigenvalues of a random matrix whose rows are normalized to length one follow a new family of probability laws, depending only on the tail index of the entries. The result provides a unified description of the spectrum of self-normalized matrices across the entire heavy-tailed regime, from the circular law to a point mass at zero.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The bounded density assumption (7) is used for eigenvalue convergence but its necessity is not tested; the paper does not provide a counterexample or a relaxation, leaving a gap between the singular-value result (no density) and the eigenvalue result.","rationale":"The reader's verdict identifies the bounded density assumption as the weakest assumption. I agree that this is the most load-bearing concern. The paper does not address the necessity of this assumption, nor does it provide evidence that it is not merely technical. The eigenvalue convergence is a central claim, and the assumption is a genuine restriction. The paper does not test its necessity. Therefore, I recommend a conditional verdict: accept the paper conditional on the authors either relaxing the assumption or providing a counterexample showing it is necessary. The concrete test I propose would help settle this. The paper is otherwise sound and well-written, but this gap is significant.","tokens_in":36827,"tokens_out":557,"duration_ms":5613,"concrete_test":"Construct a sequence of entry distributions with regularly varying tails of index alpha in (0,2) that do not have a bounded density (e.g., discrete heavy-tailed distributions, such as a lattice distribution with tail P(|X|>t) ~ t^{-alpha}). Simulate or analytically compute the empirical eigenvalue distribution of Y_n for moderate n (e.g., n=1000) and compare to the predicted mu^{SN}_alpha. If the eigenvalue distribution deviates significantly from the prediction, the bounded density assumption is necessary. If it matches, the assumption may be relaxed. This test would settle whether the concern lands.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central eigenvalue result (Theorem 2.1(iii)) requires the bounded density assumption (7). The singular-value convergence (Theorem 2.1(i)) does not. The authors use (7) to obtain the polynomial least singular value estimate (Proposition 3.14) and the logarithmic uniform integrability (Proposition 3.18). However, the paper does not demonstrate that (7) is necessary, nor does it provide a counterexample when (7) fails. The reader's verdict identifies this as the weakest assumption. The concern is that the eigenvalue convergence might hold under weaker conditions (e.g., discrete distributions) or might fail; the paper does not explore this. The bounded density assumption is a genuine restriction, and the paper does not test its necessity. This is a load-bearing concern because the eigenvalue result is a main claim, and the assumption is not obviously removable. The paper acknowledges the assumption but does not discuss its necessity or provide evidence that it is not merely technical.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper studies non-Hermitian random matrices Y_n obtained by normalizing each row of an i.i.d. matrix to Euclidean norm one. The entries have regularly varying tails of index α ∈ [0,2]. For 0<α<2, under an angular tail convergence condition, the authors prove convergence in probability of the empirical singular value measures of every shifted matrix Y_n - zI to deterministic laws characterized by the root resolvent of a column-normalized directed Poisson weighted infinite tree (PWIT). Under an additional bounded density assumption, they prove convergence of the empirical eigenvalue distribution to a deterministic, rotationally invariant measure μ^{SN}_α on the complex plane, supported on the closed unit disk with outer radius exactly one, no point atoms, and no mass on the unit circle. They also treat the endpoints: as α↑2, μ^{SN}_α converges to the circular law; as α↓0, to δ_0. For the normal domain of attraction (including finite variance), they show convergence to the Gaussian limits (circular law under bounded density). For α=0 with slowly varying tails, they show μ_{Y_n} ⇒ δ_0 almost surely. They also prove spectral radius convergence to 1 for α∈(0,2] under symmetry and bounded density. The paper provides a general criterion (Theorem 4.4) for the outer support radius of rotationally invariant measures from fresh cavity recursions.","tokens_in":37156,"tokens_out":3634,"duration_ms":22307,"significance":"If the results hold, the paper makes a significant contribution to random matrix theory. It extends the PWIT framework of Bordenave, Caputo, and Chafaï to row-normalized (self-normalized) matrices, which are natural in statistics (sample correlation matrices) and exhibit distinct behavior from deterministically normalized heavy-tailed matrices. The family of limiting eigenvalue distributions μ^{SN}_α interpolates between the circular law and δ_0, providing a new one-parameter family of non-Hermitian spectral laws. The geometric properties (support radius exactly one, no atoms, no mass on circle) are nontrivial and derived via a general cavity recursion criterion. The spectral radius result (convergence to 1) is striking because it contrasts with the unbounded support for deterministically normalized heavy-tailed matrices. The paper is rigorous, with detailed proofs building on the objective method, Hermitization, and careful probabilistic estimates. The results are likely to be influential in heavy-tailed random matrix theory and related fields.","major_comments":[{"comment":"The bounded density assumption (7) is used for the eigenvalue convergence (Theorem 2.1(iii)) via the polynomial least singular value estimate (Proposition 3.14) and logarithmic uniform integrability (Proposition 3.18). The singular value convergence (Theorem 2.1(i)) does not require this assumption. The paper does not discuss whether the bounded density assumption is necessary for eigenvalue convergence, nor does it provide a counterexample when it fails. This is a genuine restriction, and the eigenvalue result is a main claim. While the authors acknowledge the assumption, they do not test its necessity or discuss possible relaxations. This is a load-bearing concern because the eigenvalue result is central. I recommend that the authors add a discussion of the role of the bounded density assumption, perhaps noting that it is used only for the least singular value estimate and that removing it would require a different approach. They could also mention whether discrete distributions (e.g., Rademacher) are covered or not. This does not require new results, but the paper should be transparent about the gap.","section":"§2.1, Theorem 2.1(iii) and §3.7–3.9","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The general criterion for the outer support radius (Theorem 4.4) is a key technical contribution. However, the proof relies on Proposition 4.2, which uses the smoothing transform and references [AM12, Lemma 6.3]. The paper should verify that all hypotheses of [AM12, Lemma 6.3] are satisfied in the context of the proof. In particular, the assumption (A1) of [AM12] requires P(T ∈ {0,1}^N) < 1. The authors argue that if A_1 = ... = A_N = 1 a.s., then (68) gives ϕ(s) ≤ ϕ(s)^N, which is impossible at s=1 since ϕ(1)=1/2 and N≥2. However, the case where some A_k are zero or one with positive probability but not all one should be addressed. The paper states that if (A1) fails, the inequality leads to a contradiction, but the reasoning is brief. Please expand the justification that the conditions of [AM12, Lemma 6.3] are met, or provide a self-contained argument.","section":"§4.2, Theorem 4.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The independent source averaging (90) states that Q has the same distribution as t + Σ_k π_k H_{+,k}, where (H_{+,k}) are i.i.d. copies of H_+ independent of the weight sequence (π_k). This is a crucial step for the cavity recursion. The justification relies on the fact that the edges contributing to Q - t are those with a_{uo} ≠ 0, i.e., edges from child u to parent o. The raw squared magnitudes form a Poisson cloud, and the normalizer d(o)^2 = Σ_k J_k. The claim that the descendant subtrees are independent of each other and of the root source cloud is plausible but should be argued more carefully. The paper states: 'Removing such a connecting edge leaves a fresh normalized subtree at the child: no normalizer inside that subtree contains the connecting coefficient.' However, the normalizer at the child (d(u)) does depend on the connecting coefficient because d(u)^2 includes the parent edge contribution (see (18)). When the edge is removed (cavity), the normalizer at u is recomputed without that edge. The paper asserts that the descendant subtrees are independent of the root source cloud and of each other. This independence is essential for the recursion. Please provide a more detailed justification or a reference to a similar construction in [BCC11b] or [BCCP17].","section":"§4.3, Lemma 4.8","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The sentence 'The mechanism behind the model changes substantially with α.' is vague. Consider rephrasing to briefly indicate how the behavior changes (e.g., sparsity, row norm concentration).","section":"§1, page 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The definition of a_n uses max{1, inf{...}}. For n=1, a_1=1. This is fine, but the text says 'For n≥2, define the tail quantile a_n = ...'. Then a_1=1 is defined separately. This is a minor inconsistency; consider defining for all n≥1 uniformly.","section":"§2.1, equation (6)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation d(v)^2 is defined, but later d(v) is used as the normalizer. It might be clearer to denote the normalizer as d(v) (positive square root) and its square as d(v)^2. The text uses d(v)^2 for the sum and then d(v) for the normalizer in (19). This is consistent but could be explicitly noted.","section":"§3.1, after equation (18)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The gauge covariance lemma is stated for operators on the same rooted vertex set. The proof uses U = diag(U_+, U_-). The condition that U_+ and U_- are diagonal unitaries preserving D_fin(V) is given. However, the application in Proposition 3.9 uses U_+ = U_- = U (phase removal) and U_+ = V_ϑ, U_- = e^{iϑ} V_ϑ (rotation). In the rotation case, U_- is not unitary? Actually e^{iϑ} V_ϑ is unitary if V_ϑ is unitary. So it's fine. But the condition that U_+ and U_- preserve D_fin(V) holds. The lemma is clear.","section":"§3.5, Lemma 3.8","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof uses the smoothing transform and references [AM12, Lemma 6.3]. The notation N_T in [AM12] is the number of positive T_k. Here T_k = A_k for k≤N and 0 otherwise, so N_T = N a.s. The condition (A2) of [AM12] is E N_T > 1, which holds since N≥2. The condition (A3a) is inf_{θ∈[0,1]} m(θ) ≤ 1, which is negated by (59). The paper should explicitly state that (A1) of [AM12] is P(T ∈ {0,1}^N) < 1. The authors argue that if A_1 = ... = A_N = 1 a.s., then a contradiction arises. But what if some A_k are 0 or 1 with positive probability but not all 1? The condition (A1) could still fail if the support of T is contained in {0,1}^N. The paper should address this case or note that it is excluded by the strict positivity of A_k (since A_k > 0 a.s., the only way to be in {0,1}^N is to be exactly 1, which is handled). Actually A_k > 0 a.s., so the only possible values in {0,1} are 1. So the support is not contained in {0,1}^N unless all A_k = 1 a.s. So (A1) holds automatically. This reasoning should be made explicit.","section":"§4.2, Proposition 4.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The formula (94) gives F_{α,t}(r) = E[r^2/(r^2 + P_o Q_o)]. This is derived from the root block expression. The text says 'The first equality also shows that r φ'_{α,t}(r) = μ_{α,t}{ |z| ≤ r }.' This is correct. However, the notation F_{α,t}(r) is used for the distribution function of the regularized measure μ_{α,t}. It might be helpful to define F_{α,t} explicitly.","section":"§4.4, Lemma 4.9","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof uses Lemma 4.13 to bound E[(Q/q)^{-s}]. The constant C_{α,r_1,s} is claimed to be independent of r∈[r_1,1) and sufficiently small t. The dependence on r_1 is through δ_* = r_1^4 - χ_α > 0. This is fine. However, the step 'P(H_+/q ≥ 1/4) ≥ δ_*/16' uses (109) and Chebyshev? Actually it uses: EV ≥ 1/2, EV ≤ 1/4 + P(V ≥ 1/4)^{1/2} (Var V)^{1/2}? Wait: EV = E[V] = E[H_+/q]. The bound EV ≤ 1/4 + P(V ≥ 1/4)^{1/2} (E[V^2])^{1/2} by Cauchy-Schwarz? Actually: EV = E[V 1_{V<1/4}] + E[V 1_{V≥1/4}] ≤ 1/4 + (E[V^2])^{1/2} P(V≥1/4)^{1/2}. Then rearranging gives P(V≥1/4) ≥ (EV - 1/4)^2 / E[V^2]. Using EV ≥ 1/2 and E[V^2] ≤ 1/(r^4 - χ_α) from (109) (since V = H_+/q ≤ Q/(r^2 q) and E[(Q/q)^2] ≤ 1/(1 - χ_α/r^4) from (101)? Actually (101) gives E[(Q/q)^2] ≤ 1/(1 - χ_α/r^4). But V = H_+/q ≤ Q/(r^2 q), so E[V^2] ≤ (1/r^4) E[(Q/q)^2] ≤ 1/(r^4 - χ_α). So indeed E[V^2] ≤ 1/(r^4 - χ_α). Then P(V≥1/4) ≥ (1/2 - 1/4)^2 / (1/(r^4 - χ_α)) = (1/4)^2 (r^4 - χ_α) = (r^4 - χ_α)/16. Since r ≥ r_1, r^4 - χ_α ≥ r_1^4 - χ_α = δ_*. So P(V≥1/4) ≥ δ_*/16. This is correct. The derivation could be made clearer.","section":"§4.7, Proposition 4.14","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The proof uses Hoeffding's inequality for bounded random variables. The summands are 2(1 - max_j |Y_{ij}|), which are bounded between 0 and 2. However, the rows are independent, so the sum is over i.i.d. variables. Hoeffding's inequality gives P(|Δ_n^2 - EΔ_n^2| > u) ≤ 2 exp(-2 n u^2 / (2^2))? Actually for variables in [0,2], the bound is 2 exp(-2 n u^2 / (2^2)) = 2 exp(-n u^2/2). The paper writes 2 exp(-n u^2/2), which is correct. Good.","section":"§5.2, Proposition 5.7","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The inequality (122) uses the rank of B^k. The proof uses the Schatten Hölder inequality (120) and the Rotfel'd–Thompson inequality (119). The exponent p = 2/k ≤ 1, so (119) applies. The bound for ∥B^k∥_{p,n}^p uses that B^k is a signed one-hot matrix with rank d = rank(B^k). The inequality ∑_{j=1}^d s_j(B^k)^{2/k} ≤ d^{1-1/k} (∑ s_j(B^k)^2)^{1/k} is by Hölder with exponents (k/(k-1), k). Since ∑ s_j(B^k)^2 = ∥B^k∥_F^2 = n (since each row has one nonzero entry of modulus 1, so Frobenius norm squared = n). So the bound is (d/n)^{1-1/k}. This is correct.","section":"§5.2, Lemma 4.19","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper is well-written and the proofs are detailed. However, there are occasional typos and minor grammatical issues. For example, in the abstract: 'These also describe the limiting eigenvalue distributions of Y_n for real entries with slowly varying tails and for centered entries with bounded density in the normal domain of attraction, respectively.' The word 'respectively' is unclear because there are two cases but the sentence structure is ambiguous. Consider rephrasing.","section":"General","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Reference [FHH26] is to an arXiv preprint (2608.28169). The paper relies on results from this preprint for the upper bound on spectral radius. Since [FHH26] is not yet published, the authors should ensure that the results are correctly cited and that the dependence is clear. This is acceptable for a preprint, but the referee should note that the paper's conclusions depend on another work that may not be fully peer-reviewed yet.","section":"References","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Johannes,\n\nThis is a serious piece of work. The paper extends the Bordenave–Caputo–Chafaï framework for heavy-tailed non-Hermitian matrices to the row-normalized (self-normalized) setting. The main results are the limiting singular value and eigenvalue distributions for Y_n, where each row is normalized to Euclidean length one. The eigenvalue measure μ^{SN}_α is rotationally invariant, supported on the closed unit disk, has outer radius exactly one, no atoms, and no mass on the circle. It interpolates between δ_0 as α↓0 and the circular law as α↑2. The spectral radius converges to 1 in probability for α∈(0,2] under symmetry and bounded density.\n\nWhat is new: the L² normalization changes the PWIT structure significantly from the L¹ normalization of BCCP17, and the analysis of the resulting cavity equations is nontrivial. The general criterion for the outer support radius (Theorem 4.4) is a clean contribution. The paper is well-written, the proofs are detailed, and the authors are transparent about assumptions.\n\nSoft spots: The bounded density assumption (7) is required for the eigenvalue convergence (Theorem 2.1(iii)) but not for the singular value convergence (Theorem 2.1(i)). The stress-test note is correct: the paper does not test necessity or provide a counterexample when (7) fails. The authors acknowledge the assumption but do not discuss whether it is merely technical. This is a genuine restriction, though not a fatal one—the paper is honest about it, and the singular value results stand without it. The gap between the singular-value result (no density) and the eigenvalue result is real, but the paper does not claim otherwise.\n\nWho this is for: random matrix theorists working on heavy-tailed models, non-Hermitian spectra, and self-normalized statistics. The paper deserves a serious referee; the mathematics is rigorous and the results are significant. I would cite it.\n\nRecommendation: send to peer review. The bounded-density issue is a limitation worth flagging in a review, but it does not undermine the core contribution.","headline":"Solid extension of BCC to row-normalized heavy-tailed matrices; the bounded-density gap for eigenvalues is real but the paper is honest about it.","tokens_in":37623,"tokens_out":530,"duration_ms":6524},"pith_extraction":null,"created_at":"2026-11-13T10:59:19.016990+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":null,"supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}