{"id":"d997aa62-f50e-45bd-89aa-aa611d6febce","arxiv_id":"2608.11084","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Sharp dimension-free Frobenius-norm concentration inequalities for sample moment tensors, with matching two-sided Gaussian estimates and an even/odd parity effect.","lead":"This paper proves sharp bounds on the error of estimating a high-order moment tensor from a finite sample, for both sub-Gaussian and Gaussian data. The bounds do not depend on dimension and reveal a parity effect: even and odd tensor orders behave differently for sub-Gaussian data.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"External black box [10, Prop. 3.3] is the load-bearing step for Theorem 2.3; if its strict-slack coupling theorem is false or weaker than stated, the induction collapses, though a weaker log-concavity version would suffice.","rationale":"I read the full proof of Theorem 2.3 and the surrounding arguments. The internal mathematics is consistent: the reduction to finite support, the conditional expansion (4.4), the bounds on Gaussian chaos in Proposition 2.5, the Stein-kernel recursion in Proposition 2.6, and the sharpness example all check out. The only place where the argument depends on something not proved in the paper is the strict-slack coupling theorem [10, Proposition 3.3]. The reader flagged exactly this as the weakest assumption. I agree it is the load-bearing external input, but it is a normal citation dependency: the theorem is recent, precisely cited, and used in a structurally explicit way. Moreover, even if [10] provides only log-concave conditional laws rather than the exponential-tilt form, the proof can be patched by a convexity argument, so the substantive risk is limited to the theorem being false. Since there is no evidence of that, and the rest of the paper is sound, I see no reason to change the reader's ACCEPT verdict.","tokens_in":35036,"tokens_out":21542,"duration_ms":177955,"concrete_test":"Extract the exact statement of [10, Proposition 3.3] from arXiv:2605.10908 and verify (i) it applies to finitely supported Z with strict slack, and (ii) the conditional laws are either exponential tilts as in (4.1) or at least log-concave w.r.t. N(0,Σ). If only (ii), redo Section 4.1 Step 2 with W(r)=V(r-z)+⟨z,Σ^{-1}r⟩ and check that Proposition 2.6 still yields (4.3); if the proposition is weaker or false, the induction in Theorem 2.3 fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 (Step 2, after eq. (4.1)) imports [10, Proposition 3.3] to obtain, for finitely supported Z with Z ⪯_cx (1-ε)G_Σ, a martingale coupling whose conditional densities are exponential tilts f_i(y)=exp(a_i+⟨b_i,y⟩)/D(y) w.r.t. N(0,Σ). This specific form is used to show ∇² V_i(y) ⪰ Σ^{-1} for the conditional law of R=G_Σ-Z, yielding the a.s. bound ‖M_r(Z)‖_F ≤ U_r ρ_r(Σ) via Proposition 2.6. That bound is then the backbone of the induction in Theorem 2.3 and, ultimately, the sub-Gaussian weak-moment estimate Lemma 3.5 and Theorem 2.1. If [10, Prop. 3.3] only guarantees log-concave densities w.r.t. N(0,Σ) (without the softmax denominator), the explicit Hessian computation would need replacement, but the convexity argument still gives ∇²V_i ⪰ Σ^{-1}, so the proof survives. The real risk is therefore only the correctness/exact strength of the cited theorem. No other internal gap was found; the remaining external input [18] is a comparison theorem used only to enter the convex-domination framework.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper establishes sharp dimension-free Frobenius-norm concentration inequalities for centered sample moment tensors (1/N)Σ_i X_i^{⊗p} − E[X^{⊗p}] of i.i.d. centered sub-Gaussian vectors in R^d. Theorem 2.1 gives an L_q upper bound with three explicit scales, Tr(Σ)^{p/2}/√N, ρ_p(Σ)√((q∧N)/N), and ‖Σ‖^{p/2}q^{p/2}/N, where ρ_p(Σ) depends on the parity of p; for Gaussian X it gives matching two-sided bounds with a different intermediate scale ρ_{p,G}(Σ). Example 3.6 shows that all three sub-Gaussian terms are necessary, and the paper identifies a parity effect in which the Gaussian and sub-Gaussian intermediate scales coincide for odd p but differ for even p. The second main result, Theorem 2.3, proves a dimension-free polynomial moment bound under Gaussian convex domination, using a structured martingale coupling from [10] together with new moment-tensor estimates for uniformly log-concave distributions (Proposition 2.6). The proof of Theorem 2.1 reduces strong moments to a square-function term and a weak-moment term via Lemma 3.1, then applies Latała’s estimates and the polynomial bound.","tokens_in":35291,"tokens_out":17694,"duration_ms":154934,"significance":"If the external inputs hold, this is a substantial contribution: it gives the sharp dimension-free Frobenius-norm concentration theory for sample moment tensors, extending covariance results such as [3,15] to all tensor orders and to all L_q moments, and it introduces a genuinely new Hilbert-valued polynomial moment bound under convex domination. The proof is carefully organized and mostly self-contained: the square-function/weak-moment reduction, the Latała estimates, the Gaussian chaos decomposition, and the Stein-kernel recursion for uniformly log-concave distributions are all spelled out in detail. There are no fitted parameters or ad-hoc assumptions internal to the paper, and the lower-bound construction is explicit. The main caveat, which drives my recommendation, is the heavy reliance on the recent preprint [10, Proposition 3.3]; this is the load-bearing step for Theorem 2.3, and the manuscript does not reproduce or prove it.","major_comments":[{"comment":"The proof uses [10, Proposition 3.3] in the strong form that the conditional densities of G_Σ given Z are exponential tilts f_i(y)=exp(a_i+⟨b_i,y⟩)/D(y), and the subsequent computation of ∇²V_i = Σ^{-1}+∇²log D(y+z_i) ⪰ Σ^{-1} depends on this softmax structure. However, the paper’s own statement of [10, Proposition 3.3] in Section 2.2 only asserts a martingale coupling with log-concave conditional densities with respect to N(0,Σ). If the stronger statement is indeed proved in [10], the displayed computation is correct; if only the weaker log-concavity statement is available, the softmax form is unjustified. This issue is load-bearing because the a.s. bound ‖M_r(Z)‖_F ≤ U_rρ_r(Σ) in (4.3) feeds directly into the induction in Theorem 2.3 and ultimately into Lemma 3.5 and Theorem 2.1. Please quote the exact statement of [10, Proposition 3.3], or rewrite Step 2 under the weaker log-concavity hypothesis, which would still give ∇²V_i ⪰ Σ^{-1} by convexity of the negative log-density.","section":"Section 4.1, Step 2 (after (4.1))"},{"comment":"Theorem 2.3 also depends on [10, Lemma 3.1] for the finite-support approximation and on [18, Theorem 1.1] for the convex-domination comparison used to enter the framework. These are external results from recent preprints and are plausible, but they are central enough that the manuscript should state the exact versions used, with enough context for the reader to verify that the hypotheses are met. In particular, the finite-support approximation step needs [10, Lemma 3.1] to produce Y_n ⪯cx G with Y_n ⇒ Σ^{-1/2}Z; this should be stated precisely.","section":"Section 2.2 and Section 4.1"}],"minor_comments":[{"comment":"The phrase ‘max_{r≥1} r^{-1} log r = 1/e’ should read ‘max_{r≥1} (log r)/r = 1/e’ for clarity.","section":"Section 3.2, Step 2 of Proposition 3.2"},{"comment":"The Rademacher vector ε in Example 3.6 and the strict-slack parameter ε_n in Section 4.1 use the same symbol; renaming one of them would avoid confusion.","section":"Example 3.6 and Section 4.1"},{"comment":"In the discussion after Example 3.6, the sentence describing the Gaussian fourth-moment identity states that taking A to be the orthogonal projector onto a top eigendirection gives equality; this is correct, but it might help to write the identity explicitly for the projector to make the equality transparent.","section":"Section 3.4"},{"comment":"The sharpness assertion for Theorem 2.3 is correct for the noncentered Gaussian polynomial bound, but the phrase ‘the terms defining Ψ_p form a geometric progression’ is only true up to constants depending on p because q^{j/2} and the covariance factors are allowed to vary independently; the displayed ≍p comparison is what is actually needed and is proved in Lemma 3.5.","section":"Section 2.2, Remark 2.4"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the dependence on [10, Proposition 3.3], which is a load-bearing external result from a very recent preprint. If the editors require the paper to be verifiable without consulting that preprint, the authors should be asked to include a precise statement and proof of the coupling theorem, or to adapt the argument to the weaker log-concavity version they themselves state in Section 2.2. Apart from this, the proofs are detailed and the results appear correct and significant. I recommend major revision to resolve this dependence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Cutting to the chase: this paper is a real contribution, not a rehash. It resolves the dimension-free Frobenius concentration question for sample moment tensors over sub-Gaussian vectors, gives matching two-sided Gaussian bounds, and identifies a parity effect that actually changes the covariance scale at even orders. The technical core is Theorem 2.3, a Hilbert-valued polynomial moment bound under Gaussian convex domination. The proof is assembled cleanly: reduction lemmas, square-function and weak-moment steps, chaos decomposition, Stein-kernel recursion, and an induction that closes with the right scales.\n\nCredit where due: Example 3.6 is a well-constructed sharpness example. It separates a Gaussian block from a random-radius Rademacher block, and that separation is exactly what makes the even-order gap visible. The authors also state plainly what is new and what is imported. The comparison with prior work on p=2 is careful; they don't claim more than (2.1) gives.\n\nSoft spots. The main theorem leans on two recent preprints, [18] and [10]. The stress-test note is right: [10, Prop. 3.3] is load-bearing in the induction for Theorem 2.3. If that structural coupling theorem fails, expansion (4.4) and the whole induction collapse. But the note also correctly observes that the proof likely survives with a weaker log-concavity version; the explicit softmax form is only needed for the Hessian computation, and convexity of the log-density already gives the needed lower bound. So this is an external dependency to verify, not an internal gap. It does mean the paper's strongest claims are conditional on results that have not yet passed full peer review. Minor point: the discussion of [15] is a bit compressed; I had to re-check their assumptions to see why they don't cover the full sub-Gaussian class. Not a flaw, just a speed bump.\n\nWho should read this? People working on covariance and moment concentration in high dimensions. It deserves a serious referee. I'd send it out.","headline":"Sharp, well-argued paper on Frobenius concentration for sample moment tensors; the main risk is an unproven external coupling result, but the architecture is sound.","tokens_in":35844,"tokens_out":2419,"would_cite":true,"duration_ms":28170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E15","62H12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves sharp, dimension-free Frobenius-norm concentration bounds for sample moment tensors, optimal over sub-Gaussian laws and two-sided for Gaussians, with a parity effect at even tensor orders.","keywords":["sample moment tensors","Frobenius norm concentration","sub-Gaussian random vectors","dimension-free bounds","Gaussian chaos","convex order","parity effect","Hilbert-valued polynomials"],"falsifier":"Compute exactly the $L^2$ weak moment $\\sup_{\\|A\\|_F=1}\\mathbb{E}\\langle A, Y^{\\otimes 2}-\\Lambda\\rangle_F^2$ for $Y=R\\Lambda^{1/2}\\varepsilon$ with $R\\in\\{0,\\sqrt{2}\\}$ and $\\varepsilon\\in\\{\\pm1\\}^d$ having independent symmetric Bernoulli entries. The sharpness claim requires this value to equal $\\|\\Lambda\\|_F^2$, whereas for a Gaussian of covariance $\\Lambda$ the value is $\\|\\Lambda\\|^2$; any smaller value would falsify the asserted even-order intermediate scale.","tokens_in":34805,"feed_emoji":"📊","tokens_out":13416,"duration_ms":111210,"temperature":0.7,"pith_summary":"The paper proves sharp, dimension-free bounds on how far the empirical $p$-th moment tensor of an i.i.d. centered sub-Gaussian sample can drift from its expectation, measured in Frobenius norm. The deviation is governed by three additive scales: a trace scale $\\mathrm{Tr}(\\Sigma)^{p/2}/\\sqrt{N}$, an intermediate covariance scale $\\rho_p(\\Sigma)\\sqrt{(q\\wedge N)/N}$, and a large-deviation scale $\\|\\Sigma\\|^{p/2}q^{p/2}/N$; all three terms are shown necessary over the sub-Gaussian class. For Gaussian data the same three terms give matching lower and upper bounds for every covariance $\\Sigma$. The paper uncovers a parity effect: at odd tensor orders the Gaussian and sub-Gaussian intermediate scales coincide, while at even orders sub-Gaussian fluctuations can be larger by a factor up to $\\sqrt{\\mathrm{rank}(\\Sigma)}$. The engine is a dimension-free moment bound for Hilbert-valued polynomials under Gaussian convex domination.","feed_headline":"Sharp rates pin down sample tensor deviations","feed_subtitle":"Optimal three-term rates, dimension-free; Gaussian case two-sided, with a parity gap at even orders.","key_machinery":"The central object is the parity-dependent covariance scale $\\rho_p(\\Sigma)$, together with the three-term decomposition of the deviation into trace, intermediate, and large-deviation scales. The argument is carried by the expansion $T[Z^p]=\\mathbb{E}[T(G_\\Sigma,\\ldots,G_\\Sigma)\\mid Z]-\\sum_{r=2}^p \\binom{p}{r} T[Z^{p-r},M_r(Z)]$, where $R=G_\\Sigma-Z$ arises from a structured martingale coupling and $M_r(Z)=\\mathbb{E}[R^{\\otimes r}\\mid Z]$ are conditional moment tensors. The coupling's conditional laws are log-concave tilts, making the conditional law of $R$ uniformly log-concave, so the paper's moment-tensor bound for uniformly log-concave distributions controls $\\|M_r(Z)\\|_F$ almost surely by $\\rho_r(\\Sigma)$; the lower-degree corrections are then absorbed by induction on the degree $p$. This converts a nonconvex polynomial moment estimate under sub-Gaussianity into a Gaussian polynomial term plus controlled lower-degree corrections.","core_discovery":"The paper's first main theorem states that for i.i.d. centered sub-Gaussian $X$ with covariance $\\Sigma$ and sub-Gaussian constant $K$, for every integer $p\\ge 2$ and every $q\\ge 1$, $$\\left(\\mathbb{E}\\left\\|\\frac{1}{N}\\sum_{i=1}^N $X_i^{{\\otimes p}}$-\\mathbb{E}$X^{{\\otimes p}}$\\right\\|_F^q\\right)^{1/q} \\lesssim_p K^p\\left(\\frac{\\mathrm{Tr}(\\Sigma)^{p/2}}{\\sqrt{N}}+\\rho_p(\\Sigma)\\sqrt{\\frac{q\\wedge N}{N}}+\\frac{\\|\\Sigma\\|^{p/2}$q^{{p/2}}$}{N}\\right),$$ where $\\rho_p(\\Sigma)=\\|\\Sigma\\|_F^{p/2}$ for even $p$ and $\\rho_p(\\Sigma)=\\|\\Sigma\\|_F^{(p-1)/2}\\|\\Sigma\\|^{1/2}$ for odd $p$. An explicit two-block example shows each of the three terms is necessary, up to constants depending only on $p$, over the sub-Gaussian class. When $X$ is Gaussian, the same three-term expression is a two-sided estimate for every covariance matrix, including singular ones. The parity effect is the statement that at even orders the sub-Gaussian intermediate coefficient can exceed the Gaussian one by $\\|\\Sigma\\|_F/\\|\\Sigma\\|$, which can be as large as $\\sqrt{\\mathrm{rank}(\\Sigma)}$. The second main theorem supplies the engine: a dimension-free moment bound for Hilbert-valued homogeneous polynomials evaluated at a random vector dominated in convex order by a Gaussian.","pith_inferences":["Beyond the paper's claims, the same square-function and weak-moment reduction suggests a uniform version: bounding $\\sup_{A\\in\\mathcal{A}}\\langle A, \\frac{1}{N}\\sum_{i=1}^N X_i^{\\otimes p}-\\mathbb{E}X^{\\otimes p}\\rangle_F$ over a tensor class $\\mathcal{A}$ at a metric-entropy rate, since the pointwise estimates are dimension-free.","Beyond the paper's claims, the parity effect implies that dimension-free Frobenius bounds for even-order moment statistics of sub-Gaussian data cannot beat the Frobenius scale without extra structural assumptions such as log-concavity, a caveat worth stating in statistical applications.","A testable extension would be to check whether the same three-term form, with the same $\\rho_p(\\Sigma)$, holds for independent but non-identically distributed sub-Gaussian vectors, using the same moment-estimate reductions."],"forward_implications":["For $p=2$, the bound recovers and sharpens Frobenius-norm sample covariance concentration: with probability at least $1-e^{-u}$, $\\|\\hat{\\Sigma}_N-\\Sigma\\|_F\\lesssim \\mathrm{Tr}(\\Sigma)/\\sqrt{N}+\\|\\Sigma\\|_F\\sqrt{(u\\wedge N)/N}+\\|\\Sigma\\|u/N$, with matching lower bounds over the sub-Gaussian class.","At even tensor orders, any procedure relying on Gaussian concentration underestimates the worst-case sub-Gaussian Frobenius error by up to $\\sqrt{\\mathrm{rank}(\\Sigma)}$, so the parity gap is intrinsic and not an artifact of the proof.","Taking $q=u$ and applying Markov's inequality gives high-probability versions of both the sub-Gaussian and Gaussian bounds for every confidence level $u\\ge 1$.","Theorem 2.3 provides a reusable dimension-free estimate: a degree-$p$ Hilbert-valued polynomial evaluated at a vector dominated in convex order by $\\mathcal{N}(0,\\Sigma)$ has $L^q$ norm bounded by a constant times $\\Psi_p(q,\\Sigma)\\|T\\|_F$.","The Gaussian two-sided estimates hold for every covariance matrix, including singular ones, so the rates are matched across all Gaussian designs."],"supporting_citations":[{"why":"Supplies the structured martingale coupling theorem with log-concave conditional laws that Theorem 2.3 imports as a black box.","marker":"[10]"},{"why":"Establishes sub-Gaussian comparison in convex order, linking condition (2.1) to Gaussian convex domination.","marker":"[18]"},{"why":"Provides a bounded integration-by-parts kernel for uniformly log-concave laws, used in Proposition 2.6 to bound conditional moment tensors.","marker":"[7]"},{"why":"Moment estimates for sums of nonnegative and centered random variables control the square-function and weak-moment terms.","marker":"[13]"},{"why":"Gaussian chaos decomposition and hypercontractivity underpin Proposition 2.5's Gaussian polynomial moment bounds.","marker":"[11]"},{"why":"Two-sided symmetric Bernoulli moment estimates prove the even-order intermediate lower bound in Example 3.6.","marker":"[14]"},{"why":"The Hilbert-space decoupling inequality is the basis of the square-function and weak-moment reduction in Lemma 3.1.","marker":"[19]"},{"why":"Its random-radius construction is adapted into an anisotropic example showing sharpness of the sub-Gaussian bound.","marker":"[15]"}],"fun_headline_variants":["Sharp tensor bounds reveal odd-even parity gap","Dimension-free sharpness for moment tensor deviations","Optimal rates for sample tensors, with parity twist","Two-sided Gaussian, sharp sub-Gaussian tensor bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof depends on an imported theorem that, whenever a finitely supported random vector is convex-dominated by a slightly scaled Gaussian, there exists a martingale coupling whose conditional laws are log-concave tilts; if that theorem fails, the remainder expansion and the induction collapse.","fun_headline_variants_meta":{"raw":{"variants":["Sharp tensor bounds reveal odd-even parity gap","Dimension-free sharpness for moment tensor deviations","Optimal rates for sample tensors, with parity twist","Two-sided Gaussian, sharp sub-Gaussian tensor bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000337,"raw_usage":{"total_tokens":1881,"prompt_tokens":980,"completion_tokens":901,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":841}},"tokens_in":596,"tokens_out":901,"duration_ms":7887,"temperature":1.0,"reasoning_tokens":841,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:37:11.375742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute exactly the $L^2$ weak moment $\\sup_{\\|A\\|_F=1}\\mathbb{E}\\langle A, Y^{\\otimes 2}-\\Lambda\\rangle_F^2$ for $Y=R\\Lambda^{1/2}\\varepsilon$ with $R\\in\\{0,\\sqrt{2}\\}$ and $\\varepsilon\\in\\{\\pm1\\}^d$ having independent symmetric Bernoulli entries. The sharpness claim requires this value to equal $\\|\\Lambda\\|_F^2$, whereas for a Gaussian of covariance $\\Lambda$ the value is $\\|\\Lambda\\|^2$; any smaller value would falsify the asserted even-order intermediate scale.","supporting_citations":[{"cited_title":"F athi,Stein kernels and moment maps, The Annals of Probability, 47 (2019), pp","cited_arxiv_id":null,"evidence_quote":"Provides a bounded integration-by-parts kernel for uniformly log-concave laws, used in Proposition 2.6 to bound conditional moment tensors."},{"cited_title":"Latała,Estimation of moments of sums of independent real random variables, The Annals of Probability, 25 (1997), pp","cited_arxiv_id":null,"evidence_quote":"Moment estimates for sums of nonnegative and centered random variables control the square-function and weak-moment terms."},{"cited_title":"Janson,Gaussian Hilbert Spaces, vol","cited_arxiv_id":null,"evidence_quote":"Gaussian chaos decomposition and hypercontractivity underpin Proposition 2.5's Gaussian polynomial moment bounds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Two-sided symmetric Bernoulli moment estimates prove the even-order intermediate lower bound in Example 3.6."},{"cited_title":"Puchkin, F","cited_arxiv_id":null,"evidence_quote":"Its random-radius construction is adapted into an anisotropic example showing sharpness of the sub-Gaussian bound."}],"review_version":1}