{"id":"60a21c08-a5e4-4044-8be9-097f9006c4a6","arxiv_id":"2607.06785","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For random A with independent mean-zero unit-variance entries of bounded fourth moments and anti-concentration, s_min(MA) is of order 1/||M^{-1}||_HS with high probability for any fixed invertible M.","lead":"A random matrix A with independent entries stays well-conditioned after left-multiplication by any fixed invertible matrix M: its smallest singular value is typically of size 1 over the Hilbert-Schmidt norm of M inverse. This extends classical invertibility theory to products and to matrices with unequal row or column variances.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The Reader correctly isolates the uniform fourth-moment bound as the weakest (and only non-standard) assumption, notes that the authors themselves highlight it as the sole place it is used, and judges the overall argument sound under the stated hypotheses. My re-examination of the compressible reduction (Proposition 3.2), the distance formula (Lemma 4.3), the projection comparison (Theorem 4.4), the anti-concentration input (Lemma 4.5 / Corollary 4.6), and the final union bound confirms that the logic holds once E[Π_j] ≺ C/n Id is granted. Because the paper does not claim a fourth-moment-free result, the limitation does not undermine the theorem as stated. Gaussian upper and lower bounds (Propositions 5.3–5.4) and the stable-rank example (Proposition 5.5) further corroborate that ||M^{-1}||_HS is the correct scale. Consequently the Reader’s ACCEPT / high-confidence assessment stands; no adjustment is warranted.","tokens_in":19156,"tokens_out":680,"duration_ms":6783,"concrete_test":"Independently re-derive the upper bound of Theorem 4.4 (the only direction used for Theorem 1.1) from the recursive relation A_{m+1} ≺ (1 - (1-η_{n-m})/(n-m)) A_m, verifying that the product ∏_{d=d_0}^n (1 - (1-η_d)/d) is O(1/n) with the stated η_d coming only from fourth-moment Chebyshev; if the product fails to be O(1/n) under E[a_{ij}^4] ≤ K, the projection comparison (and thus the incompressible estimate) collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 1.1) rests on the compressible/incompressible decomposition, with the novel ingredient being the Loewner comparison E[Π_j] ≺ (1/n) Id in Theorem 4.4. That comparison uses fourth-moment bounds via Chebyshev on quadratic forms (Var(∥ξ_m∥_2^{2} | G_m) ≤ (K+2)d) and related moment calculations for both the upper and lower bounds. The authors flag this limitation explicitly (Introduction and the Remark after Theorem 4.4) and do not claim the result without it. The remainder of the argument (net estimates for Comp, anti-concentration via [10] for the numerator of dist(MY_j, MH_j), and the averaging of the denominator via Theorem 4.4) is standard and appears correctly assembled. Gaussian sharpness (Section 5) independently confirms the scale is optimal in expectation. No hidden circularity, unstated assumption, or gap that would invalidate the stated theorem under its hypotheses was found.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies the smallest singular value of the product MA, where A is an n\times n random matrix with independent mean-zero, variance-one entries that have uniformly bounded fourth moments and a uniform anti-concentration bound L(a_ij,1)≤κ, and M is a fixed invertible deterministic matrix. The main result (Theorem 1.1) asserts that there exist constants C,c>0 depending only on K and κ such that P(s_min(MA)≤ε/∥M^{-1}∥_HS)≤Cε+e^{-cn} for every ε≥0. The argument follows the classical compressible/incompressible decomposition: the compressible contribution reduces to a cited bound for A itself, while the incompressible contribution is reduced, via a distance formula and a standard large-coordinate lemma, to controlling the random projection Π_j onto the orthogonal complement of all but the j-th column. The novel technical ingredient is Theorem 4.4, which establishes the Loewner comparison c/n Id ⪫ E[Π_j] ⪫ C/n Id by sequential revelation of columns and fourth-moment calculations; this yields an averaged bound on the denominator ∥M^{-\top}z_j∥_2. Anti-concentration of the numerator is taken from existing invertibility machinery. In the Gaussian case the authors further prove matching upper bounds (Theorem 5.1, Proposition 5.3) showing that E[s_min(MA)]≍1/∥M^{-1}∥_HS, together with a lower-tail estimate and an example illustrating the necessity of a stable-rank residual term.","tokens_in":19368,"tokens_out":1166,"duration_ms":12671,"significance":"The result supplies the natural scale that replaces the classical 1/√n when a fixed invertible matrix multiplies a random matrix with independent entries. Even the diagonal case already covers row- or column-scaled models whose variances need not be identical, and the Gaussian sharpness confirms that the Hilbert–Schmidt scale is optimal in expectation. The proof is modular and largely self-contained once standard black-box invertibility and anti-concentration statements are granted; the new projection comparison (Theorem 4.4) is of independent interest. The authors candidly flag the fourth-moment hypothesis as the sole obstacle to broader entry distributions, which is a clear and honest limitation rather than a hidden gap. Overall the paper advances the non-asymptotic theory of products of random and deterministic matrices in a clean and usable form.","major_comments":[],"minor_comments":[{"comment":"Introduction, paragraph after Theorem 1.1: the remark that results such as [3,10] treat more general entry distributions is helpful; a one-sentence pointer to precisely which moment or anti-concentration assumptions those works use would make the comparison sharper.","section":null},{"comment":"Section 2.1: the reduction for finitely many small n via s_min(MA)≥s_min(M)s_min(A) loses a √n factor relative to the HS scale; while harmless for fixed n, a brief remark that the constants may then depend on n (or that one simply absorbs them into C) would avoid any ambiguity.","section":null},{"comment":"Theorem 4.4, upper-bound iteration: the choice of d_0 involving (2+2√((K+2)(K+3)))^4 is correct but opaque; a short parenthetical explaining that it forces η_d≤1/2 would improve readability.","section":null},{"comment":"Lemma 4.5 and Corollary 4.6: the constants C_{4.5},c_{4.5} are taken from [10]; stating explicitly that they depend only on K and κ (already true) and that the intersection with R_j does not degrade the rate would make the dependence transparent.","section":null},{"comment":"Section 5, Proposition 5.5: the construction is clear, but the final comparison 1/∥M^{-1}∥_HS ≤ a_{5.5}/(n√r) could be written with the same a_{5.5} that appears in the definition of M^{-1}, to avoid a momentary notational mismatch.","section":null},{"comment":"Typographical: abstract and title use both “V ALUE” and “MA TRICES” with spaces; these should be cleaned for the published version. Occasional double spaces and line-break artifacts (e.g., “W ashington”) appear in the author addresses.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is technically solid and the fourth-moment limitation is already disclosed by the authors; I see no reason to request a major revision. The result is a clean and useful extension of the Rudelson–Vershynin theory and fits well in a probability journal of this calibre. The only soft concern is that the novelty is concentrated in the projection estimate of Theorem 4.4; once that estimate is granted, the rest of the argument is largely standard assembly. That is not a defect, but the editor may wish to weigh it against competing submissions of comparable length."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: for independent mean-zero variance-one entries with uniform fourth moments and a uniform anti-concentration bound, s_min(MA) sits at scale 1/||M^{-1}||_HS with the usual Cε + exp(-cn) lower tail. That is the natural replacement for 1/√n, and it immediately covers row- or column-scaled models. The Gaussian section shows the scale is sharp in expectation, and they even give a stable-rank example explaining when high-probability upper bounds need an extra term.\n\nWhat is actually new is Theorem 4.4: the random projection Π_j onto the orthogonal complement of all-but-one columns satisfies E[Π_j] ≍ (1/n) Id in Loewner order. That averaged control on the denominator of the distance formula is what lets the incompressible argument go through for general fixed invertible M. Everything else is the standard compressible/incompressible program: Comp reduces to a cited lemma from Livshyts, the numerator uses anti-concentration from Livshyts–Tikhomirov–Vershynin, and small-n cases are absorbed by the classical bound. The proof is modular and carefully written; the sequential-revelation argument for the projections is the technical heart and looks solid under the stated hypotheses.\n\nThe only soft spot that matters is the uniform fourth-moment assumption. It is used exactly where they say (variance bounds on the quadratic forms that appear in the projection comparison), and they are honest that they could not remove it. That is a genuine limitation relative to some recent heavy-tailed invertibility results, but it is not a gap in the theorem as stated. No circularity, no hidden parameters, citations look appropriate.\n\nThis is for people who work on non-asymptotic invertibility and want the correct scale for products or for heterogeneous row/column variances. It is not a conceptual breakthrough, but it is a clean, usable theorem with a new technical ingredient that is worth having. I would send it to a serious referee without hesitation; the math holds up under its hypotheses.","headline":"Clean, correct extension of Rudelson–Vershynin invertibility to MA, with the right scale 1/||M^{-1}||_HS and a genuine new projection comparison; fourth-moment bound is the only real soft spot and the authors flag it.","tokens_in":20010,"tokens_out":538,"would_cite":true,"duration_ms":57858,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","15B52"],"pacs":[],"model":"grok-4.5","headline":"Multiplying a random matrix by a fixed invertible matrix replaces the classical 1/√n scale for the smallest singular value by 1 over the Hilbert–Schmidt norm of the inverse.","keywords":["smallest singular value","random matrices","Hilbert–Schmidt norm","compressible and incompressible vectors","anti-concentration","product of random and deterministic matrices","stable rank"],"falsifier":"Produce a sequence of invertible matrices M_n and random matrices A_n whose entries satisfy only second-moment and anti-concentration hypotheses (no uniform fourth-moment bound) for which P(s_min(M_n A_n) ≤ ε / ||M_n^{-1}||_HS) fails to be O(ε) + exp(-c n).","tokens_in":20058,"feed_emoji":"📐","tokens_out":733,"duration_ms":75886,"temperature":0.7,"pith_summary":"When a random square matrix A with independent mean-zero unit-variance entries is multiplied on the left by a fixed invertible matrix M, the smallest singular value of the product MA is controlled by the Hilbert–Schmidt norm of M inverse, not by the usual 1/√n. The paper proves a lower-tail bound: the probability that this singular value falls below ε times that scale is at most a constant times ε plus an exponentially small term in n. The argument works under a uniform fourth-moment bound and a uniform anti-concentration condition on the entries. In the special case of Gaussian entries the same scale is shown to be sharp for the expectation, so E[s_min(MA)] is comparable to 1/||M^{-1}||_HS. The result therefore gives a natural non-asymptotic replacement for the classical Rudelson–Vershynin invertibility estimate when rows or columns are deterministically rescaled.","feed_headline":"Random-matrix invertibility scale set by HS norm of inverse","feed_subtitle":"Fixed M times random A: s_min(MA) lives at 1/||M^{-1}||_HS, with matching Gaussian expectation","key_machinery":"The Loewner comparison E[Π_j] ≼ (C/n) Id for the random orthogonal projection Π_j onto the orthogonal complement of all columns of A except the j-th; this average projection bound supplies the Hilbert–Schmidt factor that appears in the denominator of the distance formula for incompressible vectors.","core_discovery":"Under independent mean-zero unit-variance entries with uniform fourth-moment bound K and uniform Lévy concentration L(a_ij,1) ≤ κ < 1, there exist constants C, c depending only on K and κ such that for every fixed invertible M and every ε ≥ 0 one has P(s_min(MA) ≤ ε / ||M^{-1}||_HS) ≤ Cε + e^{-cn}. For Gaussian A the same scale is sharp in expectation: E[s_min(MA)] ≍ ||M^{-1}||_HS^{-1}.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["s_min(MA) controlled by HS norm of M inverse","Invertibility scale of MA set by ||M^{-1}||_HS","Random A: s_min(MA) lives at 1/||M^{-1}||_HS","HS of M inverse pins smallest singular value of MA","Product MA singular only below 1/||M^{-1}||_HS"],"cache_read_input_tokens":4352,"weakest_assumption_plain":"Every entry of the random matrix is assumed to have a fourth moment bounded by the same constant K; without that uniform bound the key projection comparison fails.","fun_headline_variants_meta":{"raw":{"variants":["s_min(MA) controlled by HS norm of M inverse","Invertibility scale of MA set by ||M^{-1}||_HS","Random A: s_min(MA) lives at 1/||M^{-1}||_HS","HS of M inverse pins smallest singular value of MA","Product MA singular only below 1/||M^{-1}||_HS"]},"model":"grok-4.5","effort":"low","cost_usd":0.005014,"raw_usage":{"total_tokens":1457,"prompt_tokens":836,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":50140000,"prompt_tokens_details":{"text_tokens":836,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":536,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":836,"tokens_out":85,"duration_ms":5318,"temperature":1.0,"reasoning_tokens":536,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T21:20:40.838626+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Produce a sequence of invertible matrices M_n and random matrices A_n whose entries satisfy only second-moment and anti-concentration hypotheses (no uniform fourth-moment bound) for which P(s_min(M_n A_n) ≤ ε / ||M_n^{-1}||_HS) fails to be O(ε) + exp(-c n).","supporting_citations":[],"review_version":1}