{"id":"0f9703a6-69ef-4389-a218-14a202b9417d","arxiv_id":"2504.15992","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a random symmetric matrix with Bernoulli or subgaussian entries, the paper proves a product-form joint lower-tail estimate for the least singular value at two separated bulk locations, and shows the bulk singular values are distinct with probability 1 minus exponentially small error.","lead":"For random symmetric matrices with independent entries, this paper proves a two-point estimate showing the least singular value behaves independently at two well-separated bulk locations, down to exponentially small scales. The result strengthens the known one-point bound and yields a quantitative lower bound on the gap between singular values, partially resolving a conjecture of Vu.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop 2.9's uniform one-location estimate is asserted without proof; all later decoupling and bootstrap steps depend on it.","rationale":"The reader's weakest_assumption is exactly the most load-bearing concern: an unproven uniform one-location estimate that is used throughout the proof. The paper's own text admits the proof is omitted, and the shift by λ changes the diagonal distribution, so the cited result [5] does not automatically apply. I do not see an internal contradiction that would warrant REJECT, but the explicit omission justifies the CONDITIONAL verdict: the central claim is plausible but not fully established from the written proof. Other deferred arguments (e.g., Theorem 1.3, Lemma 3.4, several net-construction details) are secondary; if Proposition 2.9 is supplied, the remaining gaps are more plausibly fillable. The novel two-location decoupling machinery is coherent and the strategy is credible, so the concern is about completeness rather than correctness of the overall architecture.","tokens_in":81616,"tokens_out":9217,"duration_ms":88974,"concrete_test":"Ask for a full proof of Proposition 2.9: carry out the proof of [5, Thm 1.1] for the shifted matrix A_n−λI, verifying that (i) the local semicircle law input of [5, §8] holds uniformly for λ∈[−(2−κ)√n,(2−κ)√n] with the same exponential error, and (ii) the event E3 of Lemma 2.1 (LCD ≥ e^{c3n} for every unit eigenvector of A_n−λI) holds with probability 1−e^{−Ω(n)} uniformly in λ, using the argument of [5, Lemma 4.1] with the shifted resolvent. If either step fails, Theorem 1.2 is not established; if both succeed, the gap is closed. A secondary check: replace Prop 2.9 by the pointwise estimate at the two fixed λ1,λ2 and track whether constants remain uniform in λ; this tests whether the uniform statement is actually necessary.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central two-point estimate (Thm 1.2) is built on the reduction in §5 and the bootstrap in §10, both of which invoke Proposition 2.9 to control the one-location tail P(σ_min(A_n−λI)≤δ n^{-1/2}) ≤ Cδ+e^{-cn} uniformly for all λ in the bulk. The paper explicitly says 'we omit the proof of this straightforward generalization' of [5], but [5] treats λ=0 for a mean-zero Wigner matrix. For λ≠0 the diagonal entries of A_n−λI have mean −λ, so the matrix is no longer mean-zero; the proof in [5] is not verbatim applicable. In particular, the eigenvector LCD event E3 in Lemma 2.1 is only justified in [5] for eigenvectors of A_n, not of A_n−λI, and the 'straightforward' modification is asserted. Since Proposition 2.9 is used at arbitrary bulk locations (e.g., to rule out the µ1µ2 large event), any non-uniformity or missing exponential error would prevent the product Cδ1δ2 from following. This is a load-bearing gap, not a stylistic omission.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims a two-point decoupling estimate for the least singular value of a symmetric random matrix: for bulk locations λ1, λ2 separated by at least a fixed macroscopic multiple of n^{1/2}, the probability that both σmin(A_n − λ_i I) are small is bounded by Cδ1δ2 plus an exponentially small error. A mesoscopic version is claimed with separation Δ n^{σ−1/2} and error e^{−c n^{σ/2}} under only subgaussian assumptions. From this the author derives that all singular values in a bulk interval are distinct with probability 1 − e^{−cn}, with minimal gap at least order n^{−3/2} up to a constant, giving a quantitative form of Vu's conjecture for a (1−κ)-fraction of the spectrum. The proof develops a two-location quasirandomness framework based on an inversion-of-randomness method, a decoupling of quadratic forms, and a bootstrap argument.","tokens_in":81793,"tokens_out":4408,"duration_ms":45046,"significance":"If the proof is completed, the result would be a significant advance in the quantitative invertibility of symmetric random matrices: it gives exponential-type distinctness of singular values and an n^{−3/2} gap bound, and it extends the one-location estimate of Campos et al. to a genuine two-location decoupling. The paper also contains a serious new technical apparatus for studying joint arithmetic structure of two random vectors associated with shifted Wigner matrices. The main caveat is that several load-bearing statements are asserted with proofs omitted or only sketched, so the significance currently rests on unverified technical claims rather than on a fully checkable argument.","major_comments":[{"comment":"The uniform one-location estimate P(σmin(A_n − λI) ≤ ε n^{−1/2}) ≤ cε + e^{−cn} for all λ in the bulk is stated without proof. The text says the proof of [5] generalizes straightforwardly, but [5] treats λ = 0 for a mean-zero Wigner matrix, while A_n − λI has nonzero diagonal mean for λ ≠ 0; in particular the eigenvector LCD event E3 in Lemma 2.1 is not established for A_n − λI by a verbatim application of [5]. This proposition is used in Proposition 5.3, Lemma 9.2, Lemma 10.2, and the bootstrap in Section 10, so the product form Cδ1δ2 in Theorem 1.2 depends on it. A complete proof, or a precise citation of a theorem that gives this uniform bulk estimate with the same exponential error, is required.","section":"Section 2.2, Proposition 2.9 (Eq. (2.12))"},{"comment":"The proof of Theorem 1.3 is omitted: the final line says 'The proof here is identical to the proof of Theorem 1.2... The details are omitted.' Lemma 11.2 and Lemma 11.3, which are the mesoscopic analogues of Lemma 8.1 and Lemma 9.1, are only sketched. Since Theorem 1.3 is a stated main result with a different parameter range (n^σ separation), a different error e^{−c n^{σ/2}}, and without the log-Sobolev assumption, this is not a cosmetic omission. In particular, the analogue of the quadratic-form decoupling at mesoscopic separation needs a careful check that all constants and error scales behave as claimed. Please provide full proofs of Lemmas 11.2 and 11.3 and a detailed derivation of Theorem 1.3 from them.","section":"Section 11, Theorem 1.3 and Lemmas 11.2, 11.3"},{"comment":"Several statements that are load-bearing for Theorem 2.6 are asserted without proof or with only a sketch. Lemma 3.3 (box covering of Λ_{ε,ε1}(c)) is stated with 'proof omitted'; Lemma 3.4 (random grid vectors have no rigid arithmetic structure) is stated with 'the proof is essentially the same as [18], Lemma 3.16' and details omitted; Lemma 3.13 (Fourier replacement) is stated with 'we omit the straightforward proof'; Proposition 3.9 is only sketched. These results feed directly into the net cardinality bounds and hence into the verification of the quasirandomness event E4 and Theorem 2.6. Without complete proofs or exact references that cover the two-vector statements, the central bootstrap cannot be verified. Please add the missing arguments.","section":"Section 3, Lemmas 3.3, 3.4, 3.13 and Proposition 3.9"}],"minor_comments":[{"comment":"The separation parameter is denoted κ in the abstract and ∆ in Theorem 1.2; please unify the notation to avoid confusion.","section":"Abstract and Theorem 1.2"},{"comment":"The phrase 'for any (n-dependent) constants a1, a2, D with a1 = 1' is awkward; a1 is a fixed coefficient, not an n-dependent constant. Please rephrase.","section":"Corollary 1.9"},{"comment":"There are numerous typographical issues, including 'F act 2.11' and undefined notation such as |sin(v_D, r_D)| before it is introduced. These should be cleaned up.","section":"Section 2.3"},{"comment":"In the proof, p is fixed to be 1.01, while Proposition 5.1 is stated for arbitrary p > 1. The dependence of the constants on p and on the choice p near 1 should be tracked explicitly.","section":"Lemma 10.2 proof"},{"comment":"The inequality |θ_I| ≥ (θ1θ2)^{1/2} appears to require an absolute value and a fixed normalization of θ; please clarify the normalization used in the two cases.","section":"Lemma 8.1 proof, Part 2/Part 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is promising and the claimed results are strong, but the number of omitted proofs and 'straightforward generalizations' is too large for a refereed journal. In particular, Proposition 2.9 and Theorem 1.3 are main results in the abstract, yet their proofs are either omitted or replaced by sketches. The author also merges an unpublished preprint [19] into the present manuscript; the novelty overlap with [18] and [19] should be clarified in the final version, but this is not the main blocking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this is a credible and significant extension of the Campos et al. one-point estimate to two separated bulk locations, with the right product form C δ1δ2 + e^{-cn}, and it delivers a quantitative n^{-3/2} minimal-gap statement for most of the bulk spectrum. The novelty is real: the inversion-of-randomness step is upgraded to vector pairs, with angle-dependent threshold functions, and the bootstrap in §10 is a genuine machine. Section 12 contains substantial proof detail for the conditioned inverse Littlewood–Offord theorem. If the main theorems are correct, this is a major within-field result.\n\nWhat is actually new: Theorem 1.2 is a two-point decoupling at macroscopic separation; Theorem 1.3 extends to mesoscopic separation with e^{-c n^{σ/2}} error without LSI. Corollaries 1.8 and 1.9 follow by covering arguments and are plausible.\n\nNow the soft spots, in proportion. The biggest is Proposition 2.9: the uniform one-location estimate for arbitrary bulk λ is asserted with 'we omit the proof of this straightforward generalization' of [5]. That is not a stylistic omission. For λ ≠ 0 the matrix A_n − λI has diagonal entries with mean −λ, so the mean-zero structure used in [5] is broken, and the local-law part does not transfer verbatim. The proof of Lemma 2.1, the reduction in §5, and the bootstrap in §10 all lean on this proposition. It is probably true—one expects P(σmin(A_n−λI) ≤ ε n^{-1/2}) ≲ ε + e^{-cn} uniformly in the bulk—but the paper does not show it, and the claim that the shift is harmless needs an actual argument or a citation to a published uniform version. Second, Theorem 1.3's proof is dismissed as 'identical' with details omitted; given the parameter changes, that is a lot to wave through. The rest of the proof is long and, as far as I can verify, carefully put together; the main architecture is coherent and the cited building blocks are real.\n\nWho it's for: specialists in random matrix theory, especially people working on least singular values and eigenvalue gaps. It deserves a serious referee: the potential payoff is large and the approach is new. But I would not accept it in the current form. Send it out with a request to supply the missing uniform one-point estimate and a detailed proof for the mesoscopic case, or to point to a verifiable source. My verdict: conditional accept after major revision.","headline":"A strong, genuinely new two-point decoupling result with a load-bearing gap where the uniform one-point estimate is asserted rather than proved; worth serious refereeing if the gap is filled.","tokens_in":82360,"tokens_out":5119,"would_cite":false,"duration_ms":52001,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","15B52"],"pacs":[],"model":"deepseek-v4-flash","headline":"Joint small singular values of a random symmetric matrix decouple into a product at separated bulk locations.","keywords":["random symmetric matrix","least singular value","singular value gaps","Wigner matrix","small ball probability","inverse Littlewood-Offord","decoupling","Vu's conjecture"],"falsifier":"Simulate symmetric Bernoulli matrices at $n\\approx 2000$ and, for many bulk pairs $\\lambda_1,\\lambda_2$ with $|\\lambda_1-\\lambda_2|\\geq\\sqrt{n}$, estimate $\\mathbb{P}(\\sigma_{\\min}(A_n-\\lambda_i I)\\leq\\delta_i n^{-1/2}, i=1,2)$ for $\\delta_1=\\delta_2=e^{-n^{1/4}}$; a value above $C\\delta_1\\delta_2+2e^{-cn}$ would falsify Theorem 1.2. Separately, to test the hinge, estimate the one-point probability at a fine grid of bulk locations; if for some $\\lambda$ it exceeds $C\\delta+e^{-cn}$, then the uniform Proposition 2.9 fails.","tokens_in":81368,"feed_emoji":"🎲","tokens_out":7981,"duration_ms":70690,"temperature":0.7,"pith_summary":"The paper's central claim is that, for a symmetric random matrix with independent mean-zero variance-one subgaussian entries, the event that the least singular value of $A_n-\\lambda_1 I$ is small and the same event at a separated bulk location $\\lambda_2$ are essentially independent: the joint probability is bounded by $C\\delta_1\\delta_2$ plus an exponentially small error. If correct, this decoupling lets the author prove that all singular values in $[\\kappa\\sqrt{n},(2-\\kappa)\\sqrt{n}]$ are distinct with probability $1-e^{-cn}$, and that the minimal gap between them is at least order $n^{-3/2}$ with high probability. A sympathetic reader would care because this is a strong quantitative resolution of Vu's conjecture on the absence of repeated singular values, for any fixed fraction of the bulk spectrum away from the edge, and it supplies a two-point version of the recent one-point least-singular-value result. The mechanism is a new two-location 'inversion of randomness' argument replacing the Cauchy interlacing tricks that had blocked earlier attempts.","feed_headline":"Bulk singular values of random symmetric matrices stay distinct","feed_subtitle":"Two separated bulk locations decouple; tiny singular values multiply, yielding n^{-3/2} minimal gaps.","key_machinery":"The argument's central object is a two-vector version of the essential least common denominator (LCD), defined for a pair of vectors $(A_n-\\lambda_1 I)^{-1}\\tilde{X}$ and $(A_n-\\lambda_2 I)^{-1}\\tilde{X}$ restricted to a random subset of coordinates; it measures how close any linear combination can come to a lattice point. Around this the paper builds a two-location inversion-of-randomness scheme: a zeroed-out matrix $M_n$ with independent blocks replaces the dependent symmetric matrix, threshold functions $\\tau_{L,\\varepsilon_1}$ stratify vector pairs by their anti-concentration, and a double-counting argument bounds the size of the corresponding nets. The second pillar is spectral decoupling: the local semicircle law (or the super-exponential concentration from a log-Sobolev inequality) shows that at separated locations the $k$-th singular value at one location dominates the correlated singular value at the other, so small-ball estimates at the two locations multiply. A bootstrap lemma then iterates the product exponent up to the optimal $\\delta_1\\delta_2$.","core_discovery":"Theorem 1.2 states that for locations $\\lambda_1,\\lambda_2$ in the bulk $[-(2-\\kappa)\\sqrt{n},(2-\\kappa)\\sqrt{n}]$ with $|\\lambda_1-\\lambda_2|\\geq \\Delta\\sqrt{n}$, one has $\\mathbb{P}(\\sigma_{\\min}(A_n-\\lambda_i I)\\leq \\delta_i n^{-1/2}, i=1,2)\\leq C\\delta_1\\delta_2+2e^{-cn}$, under a finite log-Sobolev constant. Theorem 1.3 achieves the same product structure for mesoscopic separation $|\\lambda_1-\\lambda_2|\\geq \\Delta n^{\\sigma-1/2}$ with error $e^{-c n^{\\sigma/2}}$ and no log-Sobolev assumption. As corollaries, the singular values in $[\\kappa\\sqrt{n},(2-\\kappa)\\sqrt{n}]$ are all distinct with probability $1-e^{-cn}$, and the minimal gap among them satisfies $\\mathbb{P}(\\text{gap}\\leq \\varepsilon n^{-3/2})\\leq C\\varepsilon+e^{-cn}$ (or $C\\varepsilon+e^{-c n^{\\sigma/2}}$ in the mesoscopic case). This is what the paper means by a strong quantitative form of Vu's conjecture up to a $(1-\\kappa)$-fraction of the spectrum.","pith_inferences":["Extension: numerically, one expects the joint tail $\\mathbb{P}(\\sigma_{\\min}(A_n-\\lambda_1 I)\\leq\\delta_1 n^{-1/2},\\sigma_{\\min}(A_n-\\lambda_2 I)\\leq\\delta_2 n^{-1/2})$ to track $C\\delta_1\\delta_2$ at separations much larger than $n^{-1/2}$; Monte Carlo at very small $\\delta$ would be a direct check.","Extension: Proposition 2.9 is the hinge; testing one-point probabilities over a fine grid of bulk locations separates the paper's theorem from its assumed uniform generalization.","Extension: the $n^{-3/2}$ gap scale suggests that $n^{3/2}$ times the minimal bulk gap should converge in distribution; the paper does not assert this, but its bound is consistent with it."],"forward_implications":["For symmetric Bernoulli matrices, with probability $1-e^{-cn}$ all singular values in $[\\kappa\\sqrt{n},(2-\\kappa)\\sqrt{n}]$ are distinct; the proof is direct from Theorem 1.2 and a covering argument.","The minimal gap between these bulk singular values is at least $n^{-3/2}$ up to constants: the probability of a gap below $\\varepsilon n^{-3/2}$ is at most $C\\varepsilon+e^{-cn}$.","Extreme small-ball events of the least singular value at two separated bulk locations factor, up to a constant and an exponentially small error, exactly as if the two locations were independent.","Linear statistics of distant eigenvalues are controlled: the probability that two distant bulk eigenvalues satisfy $a_1 x_1+a_2 x_2=D$ within $\\varepsilon n^{-3/2}$ is $O(\\varepsilon)$ plus an exponential error.","For mesoscopic separations $|\\lambda_1-\\lambda_2|\\geq \\Delta n^{\\sigma-1/2}$, the same decoupling holds with exponential error $e^{-c n^{\\sigma/2}}$ for any subgaussian entry distribution without log-Sobolev."],"supporting_citations":[{"why":"Supplies the one-point least singular value estimate for symmetric Wigner matrices that the paper extends to a joint two-point estimate and assumes uniformly in Proposition 2.9.","marker":"[5]"},{"why":"Supplies the inversion-of-randomness technique and exponential singularity bounds for symmetric matrices, upgraded here to vector pairs.","marker":"[6]"},{"why":"Provides the compressible/incompressible vector decomposition, the distance interpretation of the least singular value, and inverse Littlewood-Offord estimates used throughout.","marker":"[29]"},{"why":"Gives the local semicircle law in the bulk, controlling eigenvalue counts used to show that separated locations have non-overlapping relevant singular values.","marker":"[11]"},{"why":"Supplies the super-exponential concentration of the empirical spectral measure under a log-Sobolev inequality, used to obtain the $e^{-cn}$ error in Lemma 7.4.","marker":"[17]"},{"why":"States Vu's conjecture on distinct singular values; Corollary 1.8 is its strong quantitative bulk form.","marker":"[35]"},{"why":"Provides no-gaps delocalization of eigenvectors, used to find coordinates where both least-singular-vector components are of size about $n^{-1/2}$ in the distance reduction.","marker":"[27]"}],"fun_headline_variants":["Bulk singular values distinct in random symmetric matrices","Random symmetric matrix: bulk singular values distinct, gap n^{-3/2}","All bulk singular values unique, gap n^{-3/2} in random symmetric matrix","Bulk singular values distinct with high probability, gap n^{-3/2}","Decoupled extremes yield distinct bulk singular values in random matrices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Proposition 2.9: the one-location least-singular-value bound of [5] extends to every bulk location $\\lambda$ uniformly with the same exponential error $e^{-cn}$; the paper gives this as a stated but omitted 'straightforward generalization', and both the two-point reduction and the log-factor-removal bootstrap use it at arbitrary bulk locations.","fun_headline_variants_meta":{"raw":{"variants":["Bulk singular values distinct in random symmetric matrices","Random symmetric matrix: bulk singular values distinct, gap n^{-3/2}","All bulk singular values unique, gap n^{-3/2} in random symmetric matrix","Bulk singular values distinct with high probability, gap n^{-3/2}","Decoupled extremes yield distinct bulk singular values in random matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001209,"raw_usage":{"total_tokens":5072,"prompt_tokens":1129,"completion_tokens":3943,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":745,"completion_tokens_details":{"reasoning_tokens":3848}},"tokens_in":745,"tokens_out":3943,"duration_ms":24064,"temperature":1.0,"reasoning_tokens":3848,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:12:53.690559+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate symmetric Bernoulli matrices at $n\\approx 2000$ and, for many bulk pairs $\\lambda_1,\\lambda_2$ with $|\\lambda_1-\\lambda_2|\\geq\\sqrt{n}$, estimate $\\mathbb{P}(\\sigma_{\\min}(A_n-\\lambda_i I)\\leq\\delta_i n^{-1/2}, i=1,2)$ for $\\delta_1=\\delta_2=e^{-n^{1/4}}$; a value above $C\\delta_1\\delta_2+2e^{-cn}$ would falsify Theorem 1.2. Separately, to test the hinge, estimate the one-point probability at a fine grid of bulk locations; if for some $\\lambda$ it exceeds $C\\delta+e^{-cn}$, then the uniform Proposition 2.9 fails.","supporting_citations":[{"cited_title":"The least singular value of a rando m symmetric matrix","cited_arxiv_id":null,"evidence_quote":"Supplies the one-point least singular value estimate for symmetric Wigner matrices that the paper extends to a joint two-point estimate and assumes uniformly in Proposition 2.9."},{"cited_title":"The singularity probability of a r andom symmetric matrix is exponentially small","cited_arxiv_id":null,"evidence_quote":"Supplies the inversion-of-randomness technique and exponential singularity bounds for symmetric matrices, upgraded here to vector pairs."},{"cited_title":"Universality of Wigner random matrices: a survey of recent results","cited_arxiv_id":null,"evidence_quote":"Gives the local semicircle law in the bulk, controlling eigenvalue counts used to show that separated locations have non-overlapping relevant singular values."},{"cited_title":"Large deviations for the largest eigenvalue of Rademacher matrices","cited_arxiv_id":null,"evidence_quote":"Supplies the super-exponential concentration of the empirical spectral measure under a log-Sobolev inequality, used to obtain the $e^{-cn}$ error in Lemma 7.4."},{"cited_title":"Recent progress in combinatorial random mat rix theory","cited_arxiv_id":null,"evidence_quote":"States Vu's conjecture on distinct singular values; Corollary 1.8 is its strong quantitative bulk form."}],"review_version":1}