{"id":"86e6540b-68b6-4cfe-a7fc-9f34b117fc4b","arxiv_id":"2502.02186","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For Gaussian matrices with arbitrary variance profiles, the expected ℓ_p to ℓ_q norm is comparable, up to constants depending only on p and q, to the sum of the largest row and column norms plus the expected maximum entry.","lead":"A 2025 paper proves a 2017 conjecture on the expected operator norms of Gaussian matrices with arbitrary variance profiles. The result gives dimension-free two-sided estimates for all p in [1,2] and q in [2,∞].","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified.","rationale":"The reader's weakest_assumption identifies the Gaussian nature of the entries as load-bearing because the proof uses Gaussian concentration, hypercontractivity, and Slepian–Fernique comparisons. That is a description of the theorem's scope, not a weakness of the central claim: the paper's conjecture is explicitly Gaussian, and the authors transparently discuss the different behavior for Rademacher entries (Remark 9). In my stress-test I focused on the internal structure of the proof: Propositions 15 and 16 imply Theorem 2 via Proposition 18; Proposition 15 is split into the ranges outside and inside [2,3]^2 with an exponent-reduction procedure (Proposition 27) and degree-dependent bounds (Propositions 28, 29, 34). I checked the key inequalities: the hypercontractivity step in Lemma 24, the covariance estimates in Lemma 36, the block-diagonal norm identity in Lemma 19, and the concentration arguments all have the correct direction and constants. The single most deferred step is Lemma 36's use of a comparison from van Handel's paper; this is a published result and the paper's own reductions around it are consistent. I found no post-hoc assumptions, no circular reasoning, and no algebraic sign error affecting the headline equivalence. Therefore the reader's ACCEPT verdict stands unchanged.","tokens_in":41106,"tokens_out":27284,"duration_ms":253540,"concrete_test":"Verify algebraically the covariance comparison E|Gv−Gu|^2 ≤ E|Zv−Zu|^2 in Lemma 36 for a small variance profile (e.g., the 2×2 profile with a single edge, and a two-edge path graph). If the inequality fails on any such profile, the Section 5 argument and with it the p=q=2 case would need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the proof carefully and found no internal gap in the argument. The Gaussian assumption is the theorem's intended scope rather than an unstated weakness: the paper explicitly proves the Gaussian and Weibull (r≤2) cases and notes in Remark 9 that the Bernoulli limit behaves differently. The most deferred step is Lemma 36's Slepian–Fernique comparison, which is quoted from [18, proof of Theorem 4.1] rather than reproved; Proposition 35 and hence the p*,q∈[2,3] case including p=q=2 rest on it. However, the comparison is a published, standard result and the surrounding reductions are coherent, so I do not treat this as an actual error. The main theorem's two-sided bound is supported by Propositions 15 and 16, whose exponent reduction, block decompositions, and Gaussian hypercontractivity steps are internally consistent; constants are honestly reported as non-optimal in p,q.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the expected ℓ_p^n → ℓ_q^m operator norms of Gaussian matrices with a deterministic variance profile (a_ij^2). The main result, Theorem 2, proves that for every 1 ≤ p ≤ 2 ≤ q ≤ ∞ and every deterministic matrix A, E||G_A||_{p→q} is comparable, up to constants depending only on p and q, to max_i ||(a_ij)_j||_{p*} + max_j ||(a_ij)_i||_q + E max_{i,j} |a_ij g_ij|. This confirms Conjecture 1 of Guédon–Hinrichs–Litvak–Prochno (2017), previously known only for p=1 or q=∞ and for the spectral case p=q=2. The proof is split into two upper estimates (Propositions 15 and 16) that are combined through a deterministic block decomposition (Proposition 18). Along the way the authors reprove the p=q=2 case without spectral-theoretic tools. The paper also derives a boundedness criterion for infinite Gaussian matrices, moment and tail bounds, and an extension to symmetric Weibull entries with shape parameter r ≤ 2, with a remark explaining why r > 2, in particular the Bernoulli case, behaves differently.","tokens_in":41229,"tokens_out":6810,"duration_ms":61563,"significance":"The result, if correct, resolves a well-known conjecture in the theory of structured random matrices and gives the first sharp two-sided bounds for these operator norms outside the extremal cases. One notable strength is the proof architecture: the high-level reduction (Theorem 2 from Propositions 15, 16, and 18) is clean and each proposition is proved in detail; constants are reported honestly, including the explicit dependence on p and q in Remark 3. The paper is also careful about scope: it proves the Gaussian and Weibull (r ≤ 2) cases and explicitly flags in Remark 9 that the Bernoulli limit diverges from the stated form, which is a falsifiable and testable prediction. The main proof uses standard probabilistic tools (Gaussian concentration, hypercontractivity, Slepian–Fernique comparison); the only deferred ingredient is the comparison lemma (Lemma 36), which is quoted from a published proof in [18], but this is a standard reference rather than a gap. I see no internal inconsistency and no circularity: the p=q=2 case is reproved from the new machinery, not assumed.","major_comments":[],"minor_comments":[{"comment":"The exponent in the logarithmic factor is typeset confusingly (\"Logγ3 − ln((γ1p∗) ∨ (γ2q)))\"); it should be written unambiguously, for example as Log^{γ3 / (-ln((γ1 p_*) ∨ (γ2 q)))} or with an explicit lower bound on k0, so that the dependence on the parameters is clear to the reader.","section":"Section 3.2, Proposition 27"},{"comment":"The notation \"γ q∗\" in the bound \"√γq∗ Logm\" is ambiguous; the proof uses the sparsity bound |{i : |s_i| ≥ γ^{-1}}| ≤ γ^{q_*}, so the statement should display γ^{q_*} as an exponent to avoid confusion with a product γ·q_*.","section":"Section 3.1, Proposition 20"},{"comment":"The arXiv source contains numerous OCR artifacts (e.g., \"/BD\" for indicators, \"/greaterorsimilar\" for ≳, and broken spacing in the title and abstract); these should be cleaned in the final version, since they distract from the otherwise clear presentation.","section":"Throughout"}],"recommendation":"accept","confidential_remarks":"The paper is a strong contribution that resolves a conjecture of current interest. The reader's report and my own reading agree that the proof is coherent; the comparison lemma dependence is standard and explicitly cited. The proof is long but well-organized, with a clear reduction to Propositions 15, 16, and 18. I have no concerns about novelty, citation practice, or fit with the journal's scope, and I recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper settles a 2017 conjecture: for deterministic A and 1 ≤ p ≤ 2 ≤ q ≤ ∞, the expected ℓ_p→ℓ_q norm of (a_ij g_ij) is two-sided comparable, up to p,q-dependent constants, to the sum of the three natural terms. That is the headline, and it is a genuine result, not a parameter scan. The novelty is real: the full range p* , q ≥ 2 was open, with only the endpoints p=1 or q=∞ and the spectral case p=q=2 known. The proof of the p=q=2 case without spectral theory is a nice extra, and the dimension-free two-sided bounds imply moment and tail estimates that go beyond the expectation statement.\n\nThe paper is well organized and the authors are honest about the limits of their constants: they do not claim the optimal dependence on p and q, and they explicitly flag where the constant blows up. The main theorem is deduced cleanly from Propositions 15 and 16 via Proposition 18, and the proofs of those propositions are detailed. I did not machine-check every algebraic step, but the logic is coherent and the dependencies are transparent.\n\nSoft spots, in proportion: the most deferred step is Lemma 36, the Slepian–Fernique comparison quoted from van Handel's paper. Proposition 35, and hence the p* , q ∈ [2,3] case including p=q=2, rests on it. That is a published standard tool, so it is not an actual gap, but it is the part a referee should look at most carefully. Also, the Gaussian assumption is load-bearing, not a weakness: the authors rightly show the Weibull (r≤2) extension and note in Remark 9 that Bernoulli entries behave differently. The constants in the upper bound are non-optimal in p,q; the authors say so and even conjecture the sharp form. That is honest, not a flaw.\n\nWho is this for? Anyone working on random matrices with variance profiles, operator norms of structured Gaussian matrices, or dimension-free estimates in Banach space theory. It deserves a serious referee: the proof is intricate but the structure is sound, and the conjecture resolution is important enough to justify careful checking. I would accept it for peer review and recommend moving forward after the standard verification of the quoted Slepian–Fernique step.","headline":"Resolves the Guédon–Hinrichs–Litvak–Prochno conjecture for Gaussian matrices with arbitrary variance profile, and the proof holds together.","tokens_in":41787,"tokens_out":822,"would_cite":true,"duration_ms":10678,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","46B09"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any variance profile, the expected $\\ell_p\\to\\ell_q$ norm of a Gaussian matrix equals, up to $p,q$-only constants, the sum of the largest row $\\ell_{p^*}$-norm, the largest column $\\ell_q$-norm, and the expected largest Gaussian entry.","keywords":["Gaussian random matrices","operator norms","ell_p to ell_q","variance profile","two-sided bounds","hypercontractivity","random matrix theory","conjecture confirmation"],"falsifier":"For fixed $p=4/3$, $q=4$ and the variance profile $a_{ij}=(i+j)^{-1}$ on $m\\times n$ grids of increasing size, estimate $\\mathbb{E}\\|G_A\\|_{p\\to q}$ by Monte Carlo; the theorem predicts the ratio to $\\max_i\\|(a_{ij})_j\\|_{p^*}+\\max_j\\|(a_{ij})_i\\|_q+\\mathbb{E}\\max_{i,j}|a_{ij}g_{ij}|$ stays within a $p,q$-dependent constant, so a systematically growing ratio would refute the claim.","tokens_in":40906,"feed_emoji":"🎲","tokens_out":8292,"duration_ms":69281,"temperature":0.7,"pith_summary":"The paper establishes a complete two-sided estimate for the expected operator norm of a Gaussian random matrix whose entry variances are arbitrary (a 'variance profile'): acting from $\\ell_p$ to $\\ell_q$ with $1\\le p\\le 2\\le q\\le \\infty$, the norm is determined, up to constants depending only on $p$ and $q$, by the three quantities: the maximum $\\ell_{p^*}$ norm of a row, the maximum $\\ell_q$ norm of a column, and the expected maximum of the Gaussian entries themselves. This confirms a conjecture of Guédon, Hinrichs, Litvak, and Prochno from 2017, previously known only in the cases $p=1$, $q=\\infty$, and $p=q=2$. The proof supplies a new, non-spectral argument for the already-known $p=q=2$ spectral case. The result matters because matching two-sided bounds of this type automatically yield moment and tail estimates, and characterize when an infinite Gaussian matrix is almost surely a bounded operator between $\\ell_p$ and $\\ell_q$.","feed_headline":"Gaussian random matrix norms reduce to three explicit quantities","feed_subtitle":"The expected lp-to-lq operator norm is governed by three terms, settling a 2017 conjecture.","key_machinery":"The argument runs on three main mechanisms. First, the target expression is shown to be equivalent to several other forms: the expected maximum of Gaussian entries is comparable to an inf-sup quantity $\\max_{k}\\inf_{|I|=k}\\sqrt{\\log k}\\,\\max_{(i,j)\\notin I}|a_{ij}|$, and to the expected $\\ell_{p^*}$ and $\\ell_q$ norms of the rows and columns of $G_A$ (Theorem 2). Second, the upper bound is proved by splitting the norm into a part controlled by a standard net argument with a $\\sqrt{\\log(mn)}$ error (Proposition 15) and a part controlled by a block-diagonal decomposition with a weighted-supremum error (Proposition 16); a combinatorial lemma (Proposition 18) then shows that these two bounds force the logarithmic factor to disappear. Third, in the most delicate regime $p^*,q\\in[2,3]$, the proof introduces an interpolation norm $\\|x\\|_{2,q,\\lambda}$ on $\\mathbb{R}^m$ whose key property—hypercontractivity of Gaussian variables (Lemma 24)—allows a multiscale induction; the Slepian–Fernique comparison (Lemma 36) bounds the sup over small-coordinate vectors. The Gaussian assumption enters precisely at these three points: hypercontractivity, Gaussian concentration, and comparison.","core_discovery":"The central claim is Theorem 2: for every deterministic $m\\times n$ matrix $A=(a_{ij})$ and every $1\\le p\\le 2\\le q\\le \\infty$, $\\mathbb{E}\\|(a_{ij}g_{ij})_{i,j}:\\ell_p^n\\to\\ell_q^m\\|$ is comparable, with constants depending only on $p$ and $q$, to $\\max_i\\|(a_{ij})_j\\|_{p^*}+\\max_j\\|(a_{ij})_i\\|_q+\\mathbb{E}\\max_{i,j}|a_{ij}g_{ij}|$, where $g_{ij}$ are iid standard Gaussians and $p^*$ is the Hölder conjugate of $p$. Equivalent forms replace the expected maximum by an inf-sup over deleting $k$ rows and columns with a $\\sqrt{\\log k}$ factor, and by expected row/column maxima of the randomized matrix. The authors obtain the upper bound by combining a dimension-dependent bound with a logarithmic factor (Proposition 15) and a block-decomposition bound (Proposition 16), then eliminating the logarithm through a graph-theoretic exponent reduction. In the range $p^*,q\\in[2,3]$ the proof uses Gaussian hypercontractivity in an interpolated norm and a Slepian–Fernique comparison; the Gaussian character of the entries is used essentially throughout.","pith_inferences":["The paper leaves open the sharp dependence on $p$ and $q$; for the all-ones matrix the correct growth is $\\sqrt{p^*\\wedge\\log n}+\\sqrt{q\\wedge\\log m}$, but the constants in Theorem 2 grow like $(p^*\\vee q)^{13/2}$ or $(p^*\\vee q)^{5/2}$. One can test the refined conjecture numerically for small $p^*,q$ by computing the ratio for the all-ones matrix.","The Bernoulli case ($r=\\infty$) is explicitly excluded, and the paper conjectures a different three-term formula with a hard-to-compute third term; if that conjecture is true, the Gaussian-versus-Bernoulli gap is exactly the distinction between $\\sqrt{\\log k}$ and $\\log k$ scaling for the large-entries term.","The graph-theoretic exponent-reduction technique (Propositions 27–29) is not specific to Gaussian entries; it may extend to other log-concave or sub-exponential profiles, with the hypercontractivity step replaced by the appropriate moment bound.","Corollary 5 suggests a probabilistic proof that almost-sure boundedness of Gaussian operators between $\\ell_p$ spaces has a deterministic characterization; this could be relevant for the theory of random linear operators on sequence spaces."],"forward_implications":["Corollary 10: the two-sided bound for the expectation lifts to all moments and to Gaussian tail bounds, so for every $\\rho\\ge1$, $(\\mathbb{E}\\|G_A\\|_{p\\to q}^{\\rho})^{1/\\rho}$ is comparable to the same three-term expression plus $\\sqrt{\\rho}\\,\\max_{i,j}|a_{ij}|$.","Corollary 5: an infinite Gaussian matrix $(a_{ij}g_{ij})$ defines a bounded operator on $\\ell_p\\to\\ell_q$ almost surely if and only if the three quantities are finite; this is a sharp, checkable condition.","Corollaries 7 and 8: the same three-term formula holds for Gaussian mixtures and for independent symmetric Weibull entries with shape parameter $r\\le2$ (with $\\sqrt{\\log k}$ replaced by $(\\log k)^{1/r}$); for the all-ones matrix this recovers the known Weibull estimates.","Corollary 13: a general upper bound for independent centered entries with no Gaussian assumption follows, giving a dimension-dependent bound in terms of $\\ell_{p^*}$ and $\\ell_q$ moments.","The $p=q=2$ spectral case now has a proof that avoids spectral theory and the trace method, replacing it with interpolation and comparison techniques."],"supporting_citations":[{"why":"States the conjecture and proves the cases $p=1$ and $q=\\infty$ that the paper extends.","marker":"[7]"},{"why":"Establishes the $p=q=2$ spectral case and supplies the dimension-free structure and block-decomposition ideas used in the proof.","marker":"[13]"},{"why":"Provides the Gaussian-process comparison and net arguments used in Proposition 35 and the exponent reduction.","marker":"[18]"},{"why":"Gives the equivalence of the various forms of the right-hand side and previous partial results for structured random matrices.","marker":"[1]"},{"why":"Supplies the graph-based analysis of how the logarithm is reduced, adapted here from Rademacher to Gaussian matrices.","marker":"[14]"},{"why":"Gives the sharp nonasymptotic bound that Proposition 15 generalizes from $p=q=2$ to all $p^*,q$.","marker":"[2]"}],"fun_headline_variants":["Gaussian matrix norms pinned by three simple terms","2017 conjecture on Gaussian matrix norms resolved","Three quantities govern Gaussian lp-to-lq norms","Gaussian matrix operator norms: explicit formula found"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof relies essentially on the entries being Gaussian: hypercontractivity, Gaussian concentration, and the Slepian–Fernique comparison each enter at a different stage, and the authors note in Remark 9 that replacing Gaussians by symmetric Bernoulli entries changes the behavior of the third term.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian matrix norms pinned by three simple terms","2017 conjecture on Gaussian matrix norms resolved","Three quantities govern Gaussian lp-to-lq norms","Gaussian matrix operator norms: explicit formula found"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000473,"raw_usage":{"total_tokens":2385,"prompt_tokens":1013,"completion_tokens":1372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1314}},"tokens_in":629,"tokens_out":1372,"duration_ms":10279,"temperature":1.0,"reasoning_tokens":1314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:01:00.436475+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For fixed $p=4/3$, $q=4$ and the variance profile $a_{ij}=(i+j)^{-1}$ on $m\\times n$ grids of increasing size, estimate $\\mathbb{E}\\|G_A\\|_{p\\to q}$ by Monte Carlo; the theorem predicts the ratio to $\\max_i\\|(a_{ij})_j\\|_{p^*}+\\max_j\\|(a_{ij})_i\\|_q+\\mathbb{E}\\max_{i,j}|a_{ij}g_{ij}|$ stays within a $p,q$-dependent constant, so a systematically growing ratio would refute the claim.","supporting_citations":[{"cited_title":"Guédon, A","cited_arxiv_id":null,"evidence_quote":"States the conjecture and proves the cases $p=1$ and $q=\\infty$ that the paper extends."},{"cited_title":"Latała, R","cited_arxiv_id":null,"evidence_quote":"Establishes the $p=q=2$ spectral case and supplies the dimension-free structure and block-decomposition ideas used in the proof."},{"cited_title":"van Handel, On the spectral norm of Gaussian random matrices , Trans","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian-process comparison and net arguments used in Proposition 35 and the exponent reduction."},{"cited_title":"Adamczak, J","cited_arxiv_id":null,"evidence_quote":"Gives the equivalence of the various forms of the right-hand side and previous partial results for structured random matrices."},{"cited_title":"On the spectral norm of Rademacher matrices","cited_arxiv_id":"2405.13656","evidence_quote":"Supplies the graph-based analysis of how the logarithm is reduced, adapted here from Rademacher to Gaussian matrices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the sharp nonasymptotic bound that Proposition 15 generalizes from $p=q=2$ to all $p^*,q$."}],"review_version":1}