{"id":"3773ac15-cae7-4f77-a208-c7457fb16e26","arxiv_id":"2605.31054","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For orthogonally invariant positive semidefinite matrix perpetuities, the expected eigenvalue distribution has a power-law tail whose index solves E[A_11^η]=1, the largest eigenvalue dominates the tail, and as the dimension grows the distribution converges to a free perpetuity.","lead":"This paper proves that solutions of random matrix equations X = A X A^T + B have eigenvalue distributions with power-law tails, and that in the subcritical regime these eigenvalue distributions converge to a free-probability limit as the matrix size grows. The results give a spectral counterpart of Kesten's classical heavy-tail theorem.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"I read the paper as aiming to prove precise tail asymptotics for the expected ESD of orthogonally invariant PSD matrix perpetuities. The proof has a clear architecture: compression to a scalar perpetuity, moment control of the top two eigenvalues via the auxiliary h_k functions, and a final inversion to recover the largest-eigenvalue tail. I checked the main places where a hidden assumption could break the argument: (1) the compression identity in Lemma 4.13 works even with dependent A,B because it is reached through Lemma 4.12; (2) Lemma 5.7 is applied to the coefficient C=A^{1/2}, so its condition E[Tr(CC^T)^{2p}]<∞ becomes E[Tr(A)^η]<∞, which follows from E[A_{11}^η]<∞ by orthogonal invariance; (3) the tail comparison between X11 and U^2λ1(X) is carefully done and the use of the Converse Breiman lemma is legitimate since the Mellin non-vanishing condition is checked; (4) the λ2-tail negligibility is supported by the condition E[det(A^{[2]})^{η/2}]<1. The orthogonal-invariance/PSD restriction is substantial but explicit and is not a correctness flaw. The reader's weakest_assumption names this same structural restriction; I agree it is the main limitation, but I do not see it as a load-bearing objection that should change the verdict. The one step that would benefit from independent verification is the Converse Breiman inversion, since the paper relies on a somewhat delicate external theorem at that point.","tokens_in":27642,"tokens_out":57559,"duration_ms":500224,"concrete_test":"Independently re-derive the Converse Breiman step for Y=U_{11}^2, whose density is proportional to y^{-1/2}(1-y)^{(N-3)/2}, starting from P(Yλ1(X)>t) ∈ R_{-η} and the verified Mellin non-vanishing E[Y^{η+iθ}]≠0. Reproduce the inversion P(λ1(X)>t) ∼ P(Yλ1(X)>t)/E[Y^η] by a direct Mellin/Karamata computation and check that the resulting prefactor equals √π Γ(η+N/2)/(Γ(N/2)Γ(η+1/2)). Any discrepancy in this Gamma prefactor would invalidate the central constant in Theorem 5.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection identified. Theorem 5.1's argument is internally coherent: Lemma 4.13 reduces the corner entry to the scalar perpetuity X11 = A11X11 + B11; Lemma 5.7, applied with the coefficient C = A^{1/2} (so CC^T = A), supplies E[(λ1λ2)^{η/2}]<∞ from h1(η/2)<1 and h2(η)<1; Lemma 5.8 makes the λ2 tail negligible; and the Converse Breiman step is justified by the checked non-cancellation condition E[U^{2η+iθ}]≠0. The orthogonal-invariance/PSD restriction is a genuine scope limitation, not an internal gap. The least standard link is the Converse Breiman inversion (via [20, Thm 4.2]) that turns the tail of U^2λ1(X) into the tail of λ1(X); this is the only step I would want independently confirmed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies finite-dimensional matrix perpetuities, i.e. solutions of the affine fixed-point equation X = A X A^⊤ + B (or, in the positive semidefinite case, X = A^{1/2} X A^{1/2} + B), with emphasis on the expected empirical spectral distribution. Existence and uniqueness are first obtained by vectorization and classical vector-perpetuity theory under a negative top Lyapunov exponent. Under orthogonal invariance and positive semidefiniteness, the authors prove a compression identity: principal submatrices of a matrix perpetuity are again perpetuities driven by the corresponding compressed coefficients. This yields the main theorem (Theorem 5.1): under E[A_{11}^η] = 1 and regularity conditions, μ_X(t,∞) is asymptotically equal to (1/N) P(λ_max(X) > t), and to an explicit constant times P(X_{11} > t), where X_{11} is the scalar perpetuity A_{11}X_{11}+B_{11} and the tail constant is given by Goldie's theorem. In the subcritical regime τ(A)<1, the expected ESD is shown to converge weakly to the law of the corresponding free perpetuity. The results are illustrated by matrix Beta prime perpetuities, where explicit limiting spectral distributions are available.","tokens_in":27868,"tokens_out":12451,"duration_ms":100051,"significance":"If correct, Theorem 5.1 is a finite-dimensional spectral Kesten theorem: it gives precise power-law tail asymptotics for the expected empirical spectral distribution of a matrix perpetuity, identifies the largest eigenvalue as the sole contributor to the tail, and connects the tail constant to the classical scalar perpetuity. The compression identity for principal submatrices under orthogonal invariance is a new structural tool, and the weak convergence to free perpetuities in the subcritical regime is a meaningful bridge between random matrix theory and free probability. The proof is detailed and internally coherent: the scalar reduction is justified by Lemma 4.13, the moment bounds on λ1λ2 are supplied by Lemmas 5.4–5.7, the tail of X_{11} is handled by Goldie's theorem, and the Converse Breiman step is explicitly checked. The paper is honest about the orthogonal-invariance and positive-semidefiniteness restrictions, and the matrix Beta prime example provides explicit, falsifiable constants.","major_comments":[],"minor_comments":[{"comment":"The theorem is stated for the recursion X = A^{1/2} X A^{1/2} + B, while Sections 3–4 and the auxiliary lemmas in Section 5.1 are written for X = A X A^T + B. The proof is coherent only after replacing the coefficient by C = A^{1/2} throughout, so that C C^T = A. This substitution is never stated explicitly. Please add a sentence at the beginning of Section 5 clarifying that all lemmas from Section 4 are applied with C = A^{1/2}; equivalently, the notation in Theorem 5.1 refers to the invariant matrix C C^T.","section":"Theorem 5.1 / Section 5"},{"comment":"The application of Lemma 5.5 to the m-step recursion X = Π_m X Π_m^T + C_m is compressed. In particular, the condition h1(p)<1 for the original coefficient implies the corresponding condition h_{Π_m}(p)=h1(p)^m<1 for the coefficient Π_m; this is not spelled out. A short justification would make the argument easier to verify.","section":"Lemma 5.7, proof around Eq. (5.20)–(5.22)"},{"comment":"The Converse Breiman step is the least standard ingredient. The non-cancellation condition E[U^{2η+2iθ}]≠0 is checked by the displayed Gamma-function formula, and the application appears valid. For the reader, please state explicitly the exact form of [20, Theorem 4.2] used, and note that it supplies both the regular variation of P(λ1(X)>t) and the constant 1/E[U^{2η}], not merely the asymptotic comparability.","section":"Theorem 5.1, final step (Converse Breiman)"},{"comment":"The remark that the critical regime τ(A)=1 is not covered is useful. It might be worth adding a sentence that the tail exponents of the finite-dimensional perpetuities are not expected to match the free tail exponent, since the paper already notes the discontinuity of tail behavior under weak convergence.","section":"Section 6, end"}],"recommendation":"accept","confidential_remarks":"I found no blocking issue. The main theorem is strong and the proof chain is coherent. The only step I would ask an independent referee or the authors to double-check is the Converse Breiman application, but in my reading it is sound. The orthogonal-invariance/PSD scope is genuine and clearly stated, not a hidden gap."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine advance and the main theorem checks out. The paper opens the spectral analysis of matrix perpetuities X = A X A^T + B, which nobody has done, and the two headline results are a finite-dimensional spectral Kesten theorem and a weak convergence to free perpetuities in the subcritical regime.\n\nWhat is actually new: the principal-submatrix compression identity (Cor 4.5) is clean and appears new; it lets the authors prove that principal submatrices of a matrix perpetuity are themselves perpetuities of the same form. That is what reduces the (1,1) entry to a scalar perpetuity X11 = A11 X11 + B11, and Theorem 5.1 then gives precise power-law tails for the expected ESD: the tail is governed by the largest eigenvalue, and the prefactor is explicit in terms of E[U^{2η}] and the Goldie constant. The proof is coherent. The moment lemmas are real work, the use of Goldie's theorem is legitimate, and the Converse Breiman step is carefully checked with the non-cancellation condition. I don't see a circularity problem: the tail theorem relies on classical vector perpetuity results, and the self-citation to [5] is ordinary reliance for the existence of the free limit.\n\nSoft spots. The orthogonal invariance plus positive semidefiniteness assumption in Theorem 5.1 is genuinely restrictive; the whole compression argument depends on it, and it is not just a technical convenience. The paper is transparent about that, but it does mean the result covers a specific symmetry class of models. The weak convergence theorem is only subcritical (τ(A)<1), and the authors note the critical case is open. The least standard step, as the stress test says, is the Converse Breiman inversion that turns the tail of U^2 λ1(X) into the tail of λ1(X); that relies on Jacobsen–Mikosch–Rosiński–Samorodnitsky and the non-vanishing condition. On reading, the application is legitimate, but it is the one place I'd want a referee to look closely.\n\nThe matrix Beta prime example is a nice illustration, with explicit limiting spectral densities via a Coulomb gas argument. Not everything there is new, but it is useful.\n\nBottom line: this deserves a serious referee. The core theorem is likely correct, the exposition is careful, and the restrictions are stated openly. I would accept it with the usual request for clarity on the Converse Breiman step and for an explicit statement of how restrictive the orthogonal invariance is.","headline":"Solid new results on ESDs of matrix perpetuities under orthogonal invariance; the spectral Kesten theorem holds up, with the orthogonal-invariance restriction being the main scope limit.","tokens_in":28317,"tokens_out":3346,"would_cite":true,"duration_ms":27451,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","60H25","46L54","60G70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Matrix perpetuities have power-law eigenvalue tails governed by the largest eigenvalue.","keywords":["matrix perpetuity","empirical spectral distribution","tail asymptotics","Kesten theorem","orthogonal invariance","free perpetuity","beta prime distribution","random matrices"],"falsifier":"Simulate a matrix Beta prime perpetuity with parameters near the critical case, where E[det(A^[2])^{η/2}] approaches 1, and compare the empirical eigenvalue tail with the formula; specifically, check whether P(λ_2>t)/P(λ_1>t) tends to 0 and whether the tail prefactor matches √π Γ(η+N/2)/(Γ(N/2)Γ(η+1/2)) times the X_11 tail. A mismatch there would pinpoint a failure of the principal-compression reduction.","tokens_in":27560,"feed_emoji":"📊","tokens_out":6987,"duration_ms":60608,"temperature":0.7,"pith_summary":"The paper establishes a spectral analogue of the classical Kesten theorem for matrix perpetuities, solutions of X = A X A^T + B. Under orthogonal invariance and positive semidefiniteness, the expected empirical spectral distribution has a power-law tail, that tail is asymptotically equal to (1/N) times the tail of the largest eigenvalue, and the prefactor is explicitly expressed through the tail of the scalar perpetuity satisfied by the top-left entry. The central reduction is that principal submatrices of a matrix perpetuity are themselves lower-dimensional perpetuities, so the whole spectral tail is controlled by a one-dimensional recursion. This matters because such recursions model heavy-tailed stochastic algorithms, and the result gives a parameter-free prediction of how multiplicative noise and additive perturbations shape the eigenvalue distribution. In the subcritical regime, the expected ESD converges weakly to the corresponding free perpetuity as dimension grows.","feed_headline":"Eigenvalue tails of matrix perpetuities follow power laws","feed_subtitle":"A symmetry reduction ties the whole spectrum to the top-left entry and gives explicit constants.","key_machinery":"The engine is the principal-compression identity for symmetric multiplicative convolution: if B is orthogonally invariant, then for every k the k×k leading principal submatrix of A B A^T has the same law as ((AA^T)^[k])^{1/2} B^[k] ((AA^T)^[k])^{1/2}. Iterating along the product M_n = Π_n Π_n^T shows that principal submatrices of the perpetuity itself solve matrix perpetuities of smaller dimension, and for k=1 reduces to the scalar perpetuity X_11 = (AA^T)_11 X_11 + B_11. This reduction also identifies the top Lyapunov exponent as E[log(AA^T)_11], replacing the usual spectral-radius condition by the scalar moment condition E[A_11^η]=1.","core_discovery":"Theorem 5.1: for N≥2, if A,B are a.s. positive semidefinite and (A,B) orthogonally invariant, and η>0 solves E[A_11^η]=1 with moment and non-arithmeticity conditions, the unique solution X satisfies μ_X(t,∞)∼(1/N)P(λ_max(X)>t)∼(1/N)[√π Γ(η+N/2)/(Γ(N/2)Γ(η+1/2))]P(X_11>t), where X_11 solves X_11=(AA^T)_11 X_11 + B_11 and its tail has the explicit constant E[(A_11 X_11+B_11)^η − (A_11 X_11)^η]/(η E[A_11^η log A_11]). The expected eigenvalue spectrum is then a power law of index η, dominated in its tail by the largest eigenvalue, with all constants explicit in terms of the (1,1) entries and the dimension. In the subcritical regime τ(A)<1, the expected ESD converges weakly to the free perpetuity","pith_inferences":["The compression mechanism is not specific to real symmetric matrices; the same triangular-factor argument should hold for complex Hermitian models under unitary invariance, giving the same reduction to a scalar perpetuity for the (1,1) entry.","This suggests a concrete diagnostic for heavy-tailed stochastic optimization: fit the top-left entry of the update matrix, solve E[A_11^η]=1, and compare the predicted power-law exponent with the empirical eigenvalue tails.","The tail-bulk separation proved here — a free bulk plus Kesten-type extremes — is probably a general phenomenon for orthogonally invariant affine random-matrix recursions, and may hold beyond the perpetuity equation."],"forward_implications":["The expected empirical spectral distribution of a matrix perpetuity has a power-law tail t^{-η}, with η determined by the one-dimensional condition E[A_11^η]=1.","The tail is asymptotically (1/N) times the tail of the largest eigenvalue; the probability that the second eigenvalue exceeds t is negligible compared with the largest eigenvalue tail.","The prefactor connecting the spectral tail to P(X_11>t) is explicit: √π Γ(η+N/2)/(Γ(N/2)Γ(η+1/2)), times the scalar tail, and the scalar tail constant is given in closed form.","In the subcritical regime τ(A)<1, the expected ESD converges weakly to the free perpetuity distribution, so the bulk spectrum is asymptotically free while the extreme eigenvalues carry the heavy tail.","For matrix Beta prime perpetuities, the tail index is explicit: η_1 = β − (N−1)/2, with an explicit multiplicative constant."],"fun_headline_variants":["Matrix perpetuity spectra: power-law tails, explicit constants","One eigenvalue drives the tail of matrix perpetuities","Symmetry reduces spectrum tail to a scalar perpetuity","Matrix eigenvalues: power law tied to diagonal entry","Tail of matrix spectrum pinned by symmetry and dimension"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is orthogonal invariance of the pair (A,B) together with A,B ≥ 0; if the update law is not invariant under conjugate rotations, the principal-submatrix reduction to a scalar perpetuity fails and the tail index is no longer E[A_11^η]=1.","fun_headline_variants_meta":{"raw":{"variants":["Matrix perpetuity spectra: power-law tails, explicit constants","One eigenvalue drives the tail of matrix perpetuities","Symmetry reduces spectrum tail to a scalar perpetuity","Matrix eigenvalues: power law tied to diagonal entry","Tail of matrix spectrum pinned by symmetry and dimension"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1373,"prompt_tokens":842,"completion_tokens":531,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":457}},"tokens_in":586,"tokens_out":531,"duration_ms":5334,"temperature":1.0,"reasoning_tokens":457,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T02:10:37.358077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a matrix Beta prime perpetuity with parameters near the critical case, where E[det(A^[2])^{η/2}] approaches 1, and compare the empirical eigenvalue tail with the formula; specifically, check whether P(λ_2>t)/P(λ_1>t) tends to 0 and whether the tail prefactor matches √π Γ(η+N/2)/(Γ(N/2)Γ(η+1/2)) times the X_11 tail. A mismatch there would pinpoint a failure of the principal-compression reduction.","supporting_citations":[],"review_version":2}