{"id":"92d454e8-f954-4a66-a769-cb2f528e1bd9","arxiv_id":"1908.04060","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The least singular value of R*XT + U*YV, after scaling by N, converges in distribution to 1 - e^{-r^2}, the same limit as for a complex Gaussian matrix.","lead":"For a random matrix built by adding two independently rotated diagonal matrices, the smallest singular value follows the same limiting distribution as for a random Gaussian matrix. The result extends a key universality theorem to a correlated ensemble that arises naturally as a sum of random matrices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Gaussian comparison rests on Proposition 7.5, an imported result whose original statement in [33] had a missing hypothesis; the corrected version is only sketched, so the main theorem inherits an unverified external dependency.","rationale":"The reader's formal weakest_assumption is Assumption (7), the free-convolution density condition near zero. I agree that this hypothesis is restrictive and difficult to verify, but it is an explicit condition of Theorem 2.1 rather than a gap in the argument, and the paper provides two sufficient conditions for it in Appendix C. The more serious correctness risk is Proposition 7.5: the paper itself flags in Remark 7.6 that the imported theorem from [33] had a missing hypothesis, and the corrected statement is only sketched. Because Theorem 7.7 and the final Gaussian comparison in Theorem 2.1 rest on Proposition 7.5, any hidden defect there would invalidate the main result. This is not an internal inconsistency; the surrounding Sections 4-7 appear coherent, and the local laws and eigenvector bounds are plausible. It is an external dependency that needs to be closed. The reader's rationale already mentions the reliance on [33, Theorem 3.2] as the main caveat, so my concern is partially aligned with the reader's reasoning, even though the formal weakest_assumption is different. The appropriate verdict remains CONDITIONAL: accept only after Proposition 7.5 is independently and fully verified.","tokens_in":32442,"tokens_out":6823,"duration_ms":75832,"concrete_test":"Independently write out the proof of Proposition 7.5 following [33] with Definition 7.4. Specifically, derive the scaling factor π ρ^t0(0) from the subordination relation m_{t0}(0) = m_V(w), w = -t0 m_{t0}(0). Using (7.17) at η = Im w ≍ t0, which is allowed because gN^σ ≤ t0, and assumption (2.12), prove |π ρ^t0(0) - 1| ≤ C N^{-c}. Then check that every estimate in the proof of [33, Theorem 3.2] that used the missing hypothesis is covered by (7.17) together with rigidity and level repulsion of V. If any step requires a condition not listed in Proposition 7.5, such as control of Im m_V at η < g or a lower bound on ρ^t0, then the statement needs another amendment before Theorem 2.1 can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the short-time Gaussian comparison Proposition 7.5, which is taken from [33, Theorem 3.2] after correcting a missing hypothesis. Remark 7.6 concedes that the original statement of [33, Theorem 3.2] omitted a hypothesis used in its proof, and the replacement Proposition 7.5 is only sketched: the proof is described as following [33] 'up to a scaling of the particles si by π ρ^t0(0)', with the verification that this scaling is 1 + O(N^{-c}) delegated to [32, Theorem 2.4]. This is the load-bearing uncertainty because Theorem 7.7, and hence Theorem 2.1, inherit Proposition 7.5 directly: if the corrected statement is still missing a condition, the comparison to the Gaussian least singular value is unsupported. The new regularity condition (7.17) is intended to ensure π ρ^t0(0) = 1 + O(N^{-c}), but the paper does not derive this from (7.17) under the parameter constraints gN^σ ≤ t0 ≤ N^{-σ}G^2. Since the original theorem was already flawed, an independent, fully written verification of Proposition 7.5 is the key correctness risk. Assumption (7) is restrictive, but it is an explicit hypothesis with sufficient conditions given in Lemmas C.4 and C.5; it is not where the proof gap concentrates.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the least singular value of M = R* X T + U* Y V, where X and Y are deterministic diagonal matrices and R, T, U, V are independent Haar-distributed unitary matrices. The main result, Theorem 2.1, states that under assumptions (1)–(7) of Section 2.2, the rescaled least singular value N λ1(M) satisfies P(N λ1(M) ≤ r) = 1 - e^{-r^2} + O(N^{-c}) uniformly in r ≥ 0, with c > 0 an absolute constant. The proof follows the dynamical method of Che and Landon [32]: a carefully chosen flow on the unitary group produces a system of SDEs for the singular values of the symmetrized 2N × 2N Hermitian matrix; a local law (Section 4) and eigenvector estimates (Section 5) control the drift; well-posedness is proved in Section 6; and Section 7 compares the SDE to a symmetrized Dyson Brownian motion, invoking a short-time universality result from [33]. The final comparison to the Gaussian least singular value uses the explicit distribution of the latter.","tokens_in":32636,"tokens_out":3988,"duration_ms":42302,"significance":"If the central claim is established, the paper gives a substantial extension of least-singular-value universality from independent-entry ensembles to a model with strongly correlated entries, with a polynomially explicit error rate that is uniform in r. The proof is detailed and largely self-contained except for the imported short-time universality statement: it includes a local law, a well-posedness analysis, and explicit rigidity and coupling estimates. The authors are transparent about the restrictiveness of Assumption (7), and they provide two sufficient conditions in Appendix C. The main significance is therefore conditional on the unresolved status of Proposition 7.5, which is the load-bearing external input of the paper.","major_comments":[{"comment":"The proof of Proposition 7.5 is not supplied in the manuscript. Remark 7.6 explicitly concedes that the original statement of [33, Theorem 3.2] omitted a necessary hypothesis used in its proof, and that the corrected statement is justified only by saying that one may follow the old proof up to a scaling of the particles by π ρ̂_{t0}(0), with the verification of this scaling delegated to [32, Theorem 2.4]. Since Theorem 7.7 and hence Theorem 2.1 inherit Proposition 7.5 directly, this missing proof is load-bearing. The reader has no way to check that the corrected hypothesis (7.17) is sufficient and that no further missing condition remains in the old argument. A complete, self-contained proof of Proposition 7.5, or a reference to a published version containing the full corrected proof, is required.","section":"Section 7, Proposition 7.5 and Remark 7.6"},{"comment":"The paper asserts, but does not demonstrate, that the regularity condition (7.17) implies the matching condition π ρ̂_{t0}(0) = 1 + O(N^{-c}) under the parameter constraints g N^σ ≤ t0 ≤ N^{-σ} G^2. This matching is the reason the particle system can be coupled to the Gaussian reference process at time t0, so the derivation is not a routine detail. The interaction between the choices of g, G, σ, ω0, ω1 and the assumption (2.12) should be written out explicitly, including how the constants in (7.17) enter the error rate. Without this verification, the short-time comparison step is incomplete.","section":"Section 7, Eqs. (7.17)–(7.18)"},{"comment":"The statement that the hypotheses of Proposition 7.5 are verified with overwhelming probability for the singular values of M(0) by Corollary 5.9 is made without specifying the parameters g and G, the precise interval of energies, or the constants involved in Definition 7.4. Since Corollary 5.9 is proved under the bulk condition (2.12) and with a specific rate, the choice of g and G must be compatible with the constraints in (7.18). This compatibility should be stated explicitly rather than left implicit.","section":"Section 7, verification of (g,G)-regularity"}],"minor_comments":[{"comment":"The name \"Stieltjes\" is misspelled as \"Stietjes\" in at least two places (near Eq. (2.6) and Section 4.4).","section":"Sections 2.2 and 4.4"},{"comment":"In the proof of Lemma 4.8, after deriving the second equation of (4.41), the text says \"To prove the second equation\"; this should read \"To prove the first equation\" (or \"the remaining equation\").","section":"Section 4.3, proof of Lemma 4.8"},{"comment":"The definition of y*_k as an infimum over an equality condition is clean for continuous µ2, but for measures with atoms the condition should be a weak inequality (with the usual quantile convention) to avoid ambiguity; this is a minor notational point.","section":"Section 2.2, Eq. (2.10)"},{"comment":"The phrase \"noting that (dˆB_ij)(dˆB_kl) = δ_il δ_jk\" omits the differential dt in the quadratic covariation; the computation is standard, but the notation is imprecise.","section":"Appendix A, near Eq. (A.9)"},{"comment":"The real-case formula 1 - e^{-r^2/2 - r} is stated without a reference for the exact distribution of the least singular value of a real Gaussian matrix; citing the precise source, for example [72, Theorem 1.3] as indicated in the text above the theorem, would help the reader.","section":"Appendix B, Theorem B.1"}],"recommendation":"major_revision","confidential_remarks":"The decisive issue is the unproved Proposition 7.5. The authors' own Remark 7.6 concedes that the original theorem in [33] had a missing hypothesis, and the replacement proof is only a sketch with the key scaling verification delegated to [32]. Because Theorem 2.1 inherits this step, I would require a fully written proof of Proposition 7.5 (or a pointer to a published corrected version) before accepting the paper. If the authors can supply that argument, the rest of the paper is carefully executed and the result is likely to be correct. The paper may also deserve a note to the authors that the realistic scope, given the restrictive assumptions (5)–(7), should be summarized in the introduction so that readers do not overstate the applicability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Know this about the paper: it proves a new and likely correct universality theorem for the least singular value of a correlated non-Hermitian ensemble, but the load-bearing short-time comparison, Proposition 7.5, is a corrected-and-sketchy version of the authors' own earlier theorem. If that step holds, the rest of the proof is in good shape.\n\nThe model is M = R*XT + U*YV with independent Haar unitaries and deterministic diagonal X,Y. Previous work on this model only controlled the spectrum on the macroscopic scale; the limiting law of Nλ1(M) is genuinely new. The proof is substantial and carefully written: the local law via free-convolution stability is neat, the eigenvector estimates are done with explicit rates, and the real-versus-complex hard-edge discussion (the orthogonal model has no repulsion between λ1 and λ-1 and picks up a hard edge) is both informative and honest. The authors also flag that Assumption (7), the normalization ρ(0)=1/π, is hard to check, and they give two useful sufficient conditions.\n\nNow the soft spot, which the reader's report and the stress-test note both locate correctly. Proposition 7.5 is the comparison between the singular values of the perturbed matrix and those of a Gaussian at short times. The statement of [33, Theorem 3.2] had a missing hypothesis; the authors correct it by adding (7.17), a strong regularity condition on the initial density. That is a plausible fix, and Remark 7.6 explains why it is needed. But the proof of the corrected statement is not actually written out: the argument is 'one may follow the proof ... up to a scaling of the particles by π ρ^t0(0)', and the claim that this scaling is 1+O(N^{-c}) is delegated to [32, Theorem 2.4]. The current paper does not show, under the parameter constraints gN^σ ≤ t0 ≤ N^{-σ}G^2, that the hypotheses of [32, Theorem 2.4] are met. Since Theorem 7.7 and then Theorem 2.1 inherit this step directly, the main theorem is conditional on an external dependency that is not fully verified. This is not a fatal flaw—the dependency is narrow, the corrected statement is believable, and the authors are upfront about the original error—but it is a real gap.\n\nWho should read it: random matrix theorists working on edge universality for dependent or non-Hermitian ensembles. It deserves a serious referee at a strong journal, not a desk reject. I'd send it out with an explicit request that the referee verify Prop 7.5 in detail, ideally by providing the missing scaling argument. I would not personally rely on the main theorem as a black box until that happens, but the paper is worth engaging with now.","headline":"A new and likely correct least-singular-value universality theorem, but the proof leans on a sketched, corrected comparison (Prop 7.5) that should be the referee's focus.","tokens_in":33262,"tokens_out":5822,"would_cite":false,"duration_ms":54127,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","15B52","46L54"],"pacs":[],"model":"deepseek-v4-flash","headline":"For sums of two Haar-conjugated diagonal matrices, the smallest singular value has the same limiting distribution as for a Gaussian matrix.","keywords":["least singular value","universality","sum of random matrices","Haar unitary","free convolution","Brownian motion flow","local law","hard edge"],"falsifier":"Take a concrete pair of deterministic spectra satisfying assumptions (1)–(6) but for which the free-convolution density at zero vanishes; if the edge mechanism claimed here is right, $N\\lambda_1$ should develop a hard edge or a different edge scale instead of the soft-edge law $1-e^{-r^2}$. For spectra that do satisfy assumption (7), compute the empirical CDF of $N\\lambda_1$ at a fixed $r$ for increasing $N$ and check that the discrepancy from $1-e^{-r^2}$ decays like a power of $N$.","tokens_in":32141,"feed_emoji":"🎲","tokens_out":12034,"duration_ms":104942,"temperature":0.7,"pith_summary":"This paper proves that the smallest singular value of the random matrix sum $M = R^* X T + U^* Y V$, where $R,T,U,V$ are independent Haar-distributed unitary matrices and $X,Y$ are deterministic diagonal matrices, has the same limiting distribution as the smallest singular value of a matrix of i.i.d. complex Gaussian entries. Concretely, for every $r \\ge 0$, $\\mathbb{P}(N \\lambda_1(M) \\le r) = 1 - e^{-r^2} + O(N^{-c})$ with an absolute constant $c>0$ uniform in $r$. The interest is that $M$ is strongly correlated, with no independent entries, so universality of the least singular value is extended from entrywise-independent models to a natural unitary-invariant sum model. The paper also shows the real orthogonal version is genuinely different, with limiting law $1 - e^{-r^2/2 - r}$ and a hard edge at zero.","feed_headline":"Random unitary sums obey the Gaussian least singular value law","feed_subtitle":"This correlated matrix sum has the same least-singular-value law as a Gaussian matrix, up to polynomial error.","key_machinery":"The load-bearing object is the $2N\\times 2N$ Hermitian symmetrization of the matrix $\\widehat M(t)$, whose $2N$ eigenvalues are the singular values of $\\widehat M(t)$ and their negatives. The paper feeds this matrix into a specially chosen unitary Brownian flow (equations (3.2)–(3.5)) so that the eigenvalue flow is an SDE of the form $d\\lambda_i = (2N)^{-1/2}dB_i + (2N)^{-1}\\sum_{j\\ne i}(1-\\gamma_{ij})(\\lambda_i-\\lambda_j)^{-1}dt + R_i$, a symmetrized Brownian motion flow with a remainder controlled by eigenvector delocalization. The second pillar is the system (4.51) of Stieltjes-transform equations whose solution is the free convolution $\\mu_X^{\\mathrm{sym}}\\boxplus\\mu_Y^{\\mathrm{sym}}$; a stability analysis of this system (Propositions 4.10–4.11) yields the local law for the Green's function and the eigenvector estimates that bound the $\\gamma_{ij}$ and $R_i$. Short-time universality for the symmetrized flow (Proposition 7.5, with a missing regularity hypothesis supplied) then couples the singular values to those of a Gaussian matrix, and the explicit Gaussian law finishes the proof.","core_discovery":"The central claim is that universality of the least singular value holds for the ensemble $M = R^*XT + U^*YV$, with deterministic diagonal $X,Y$ satisfying assumptions (1)–(7) on their empirical measures. In this regime the rescaled least singular value $N\\lambda_1(M)$ converges in distribution to the complex Gaussian law, and the convergence is quantitative: the maximum error over all $r\\ge 0$ is $O(N^{-c})$. The paper identifies the normalization $\\rho(0)=1/\\pi$ in assumption (7), where $\\rho$ is the density of the free convolution of the two limiting symmetrized spectra, as the condition that places the edge of the spectrum at exactly the Gaussian scale; without it the constant in the limiting law would change. The proof is dynamical: a unitary Brownian flow preserving the law of $M$ is shown to drive the singular values by a symmetrized Brownian-type eigenvalue flow, and short-time relaxation to the Gaussian process yields the theorem.","pith_inferences":["Editorial inference: the same symmetrization strategy should extend to sums of $k$ independent Haar-conjugated deterministic matrices, with the main new check being the density condition near zero for the $k$-fold free convolution.","Editorial inference: the real/complex contrast suggests the edge law is governed by whether the symmetrized flow has repulsion between the smallest positive and negative singular values, so other symmetry classes would plausibly produce their own explicit laws.","Editorial inference: because the edge scale is fixed by the single number $\\rho(0)=1/\\pi$, one can rescale $X$ and $Y$ to tune that value and empirically observe the edge scale change, giving a direct numerical probe of the theorem's mechanism.","Editorial inference: the quantitative bound implies an immediate invertibility estimate for this correlated ensemble, which could be useful in algorithms involving sums of unitarily conjugated data matrices."],"forward_implications":["For every $r\\ge 0$, the probability that $N\\lambda_1(M)\\le r$ is $1-e^{-r^2}+O(N^{-c})$, so the least singular value is $O(1/N)$ with a universal Gaussian tail.","The limiting law is independent of the detailed entry correlations and depends on the diagonal data $X,Y$ only through the free-convolution density at zero, up to the imposed normalization.","The ensemble is invertible with overwhelming probability, with quantitative control on the smallest singular value matching the i.i.d. Gaussian case.","The real orthogonal analogue has a different edge: $\\mathbb{P}(N\\lambda_1(M)\\le r)=1-e^{-r^2/2-r}+O(N^{-c})$, showing the universality class depends on the base field.","The proof corrects and strengthens the short-time universality statement for the symmetrized flow used in prior work, adding a regularity hypothesis that the earlier statement omitted."],"supporting_citations":[{"why":"Supplies the unitary-flow construction and the dynamical strategy that this paper adapts to the non-Hermitian sum model.","marker":"[32]"},{"why":"Provides the short-time universality theorem for the symmetrized eigenvalue flow; Proposition 7.5 states the corrected version used here as the main input.","marker":"[33]"},{"why":"Gives the local single ring theorem used to control Green's functions and prove the strong law at small energies.","marker":"[15]"},{"why":"Gives the exact Gaussian law $1-e^{-r^2}$ for the least singular value, which is the target distribution.","marker":"[39]"},{"why":"Establishes least-singular-value universality for i.i.d. entry models and supplies the quantitative Gaussian formula used for the real case.","marker":"[72]"}],"fun_headline_variants":["Unitary sums share Gaussian least singular law","Gaussian law for least singular value of unitary sums","Random unitary mixtures mimic Gaussian edge statistics","Universal least singular value for unitary sums","Sum of random unitaries obeys Gaussian tail law"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole proof rests on the assumption that the deterministic limiting spectrum of the sum has, near its edge at zero, a density that is bounded away from zero and normalized to exactly $1/\\pi$ at zero; the authors state this condition is difficult to verify in general and supply only partial sufficient conditions.","fun_headline_variants_meta":{"raw":{"variants":["Unitary sums share Gaussian least singular law","Gaussian law for least singular value of unitary sums","Random unitary mixtures mimic Gaussian edge statistics","Universal least singular value for unitary sums","Sum of random unitaries obeys Gaussian tail law"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1256,"prompt_tokens":851,"completion_tokens":405,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":467,"completion_tokens_details":{"reasoning_tokens":337}},"tokens_in":467,"tokens_out":405,"duration_ms":5022,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:53:37.811666+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a concrete pair of deterministic spectra satisfying assumptions (1)–(6) but for which the free-convolution density at zero vanishes; if the edge mechanism claimed here is right, $N\\lambda_1$ should develop a hard edge or a different edge scale instead of the soft-edge law $1-e^{-r^2}$. For spectra that do satisfy assumption (7), compute the empirical CDF of $N\\lambda_1$ at a fixed $r$ for increasing $N$ and check that the discrepancy from $1-e^{-r^2}$ decays like a power of $N$.","supporting_citations":[{"cited_title":"Che and B","cited_arxiv_id":null,"evidence_quote":"Supplies the unitary-flow construction and the dynamical strategy that this paper adapts to the non-Hermitian sum model."},{"cited_title":"Che and P","cited_arxiv_id":null,"evidence_quote":"Provides the short-time universality theorem for the symmetrized eigenvalue flow; Proposition 7.5 states the corrected version used here as the main input."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the local single ring theorem used to control Green's functions and prove the strong law at small energies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the exact Gaussian law $1-e^{-r^2}$ for the least singular value, which is the target distribution."},{"cited_title":"Tao and V","cited_arxiv_id":null,"evidence_quote":"Establishes least-singular-value universality for i.i.d. entry models and supplies the quantitative Gaussian formula used for the real case."}],"review_version":1}