{"id":"c0cbecdf-df6e-4c31-8c81-a3369231e27a","arxiv_id":"2607.14460","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Gaussian data in the proportional limit, the spectral-norm error of the sample covariance converges to γ̂√φ1/(√φ1−√α), with γ̂ solving an equation in the covariance spectrum.","lead":"This paper claims to compute the exact large-data limit of how far the sample covariance matrix gets from the true covariance in spectral norm, for Gaussian data. The result turns a rough scaling law into a number you can plug into sample-size decisions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (30) is false as stated: the unconstrained minimization over γx omits the dual-feasibility domain, and a d=2 instance gives infimum 0 instead of L(c)=√(7/12); this invalidates the closed-form upper bound and hence Theorem 11's proof.","rationale":"The paper's central claim is an exact asymptotic formula for the expected spectral norm of Σ̂−Σ for Gaussian data with arbitrary covariance. In good faith, the claim is plausible: the isotropic case recovers the known Wishart edge, and the finite-d simulations in Figs. 2–3 match the formula for one spectrum. However, the written proof of the covering upper bound passes through a nonconvex Lagrangian-duality step that is not justified and, in its stated form, is demonstrably wrong. The concrete d=2 instance shows that Eq. (30) — the closed-form value of L(c) — has an infimum of 0 when the minimization is taken over the natural real domain, whereas the true L(c) is √(7/12). The discrepancy disappears if one adds the implicit constraint γx+s_i²≤0 for all i (equivalently, Q⪰0), but this domain restriction is absent from the text and from the saddle-point equations that define γ̂x. All subsequent computations — the derivative conditions (35)–(42), the definition of φ1,φ2,φ3 in (43)–(45), the condition (46), and the final formula (49)/(151) — depend on this equivalence. Thus the proof has a genuine hole at its most load-bearing point. This is not an ad-hominem or a disagreement with consensus; it is an internal inconsistency in the derivation as written. The lower-bound side, including the bilinear-quadratic mechanism and Theorem 8, is less problematic on its own, but it cannot certify the upper-bound formula if the upper bound itself is built on an invalid duality claim. The recommended verdict remains CONDITIONAL, matching the reader's assessment: the numerical evidence and the isotropic limit strongly suggest the final formula is correct, but the derivation has a central gap that must be closed before this can be accepted as a theorem. The stress-test does not move the verdict; it identifies the precise repair needed: state the dual-feasibility domain and provide an S-procedure/Slater argument for the equality in (26)/(30), or replace the step with a valid asymptotic argument.","tokens_in":32171,"tokens_out":33870,"duration_ms":299841,"concrete_test":"For d=2, S=diag(1,2), c=3/2, g=(1,0), compute the exact maximum L(c)=max{x1 : x1²+x2²=1, x1²+4x2²=9/4} and compare with inf_{γx} √((γx+9/4)/(γx+1)) over real γx with nonnegative radicand. The exact value is √(7/12)≈0.7638; the unconstrained infimum is 0, attained at γx=−9/4. This single computation falsifies Eq. (30) as written and demonstrates the missing dual-feasibility constraint (γx+s_i²≤0 for all i in this case); after imposing it, the values match, but the proof never states or justifies that restriction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3's derivation of the random-dual upper bound hinges on Eq. (26), which asserts strong duality for the nonconvex QCQP L(c)=max_{||x||=1,||Sx||=c} g^T Sx, and on the closed form (30). The step is not just an unproven Slater/S-procedure condition: as written, (30) is incorrect because the minimization over γx is taken over the wrong domain. Take d=2, S=diag(1,2), c²=9/4, g=(1,0). The exact primal is max x₁ subject to x₁²+x₂²=1 and x₁²+4x₂²=9/4, which gives L(c)=√(7/12)=0.7638. The righthand side of (30) for this instance is min_γx √((γx+9/4)/(γx+1)) over real γx with a nonnegative radicand. The infimum is 0, attained at γx=−9/4; it is far from L(c). The correct value only emerges if one imposes the dual-feasibility constraint γx+s_i²≤0 for all i (here γx≤−4), which is never stated. Without that constraint, the subsequent saddle-point calculus, the definition of γ̂x in (46), and the closed-form δ_u in (49) rest on an invalid equivalence. Since Theorem 11's exact limit (151) is obtained by combining this upper bound with the lower-bound matching theorem, the central claim is left without a valid proof. The isotropic limit (59) and the reported simulations provide empirical support, but they do not repair the algebraic gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the spectral norm error ||Σ̂−Σ||₂ of the sample covariance for centered Gaussian vectors with covariance Σ in the proportional regime n/d→α. It develops a Random Duality Theory (RDT) framework: a Gaussian-process upper bound for Eξ(c), an explicit saddle-point evaluation of the dual, and a new bilinear-quadratic lower-bound mechanism. The main claim, Theorem 11 (Eq. 151), is that the limiting expected spectral norm equals δ(α)=lim_d γ̂_x √φ₁(γ̂_x)/(√φ₁(γ̂_x)−√α), where γ̂_x solves the fixed-point equation (46) and φ₁ is defined in (43). For Σ=I the formula reduces to 2/√α+1/α, matching the known Wishart edge. The paper also reports numerical simulations for d up to a few thousand that agree well with the formula.","tokens_in":32559,"tokens_out":7936,"duration_ms":80745,"significance":"If the result is correct, this is a valuable precise characterization of sample-covariance error beyond scaling laws, with a deterministic, parameter-free formula and a proof strategy (RDT + replica lower bounds) that could potentially be exported to other problems. The claimed reduction to the known Wishart edge is a useful consistency check, and the simulations are encouraging. However, the proof as written has several load-bearing gaps, most seriously an incorrect unconstrained dual evaluation (Eq. 30) and an unproved strong-duality step (Eq. 26). These issues directly affect the central claim, so the paper cannot be accepted in its present form.","major_comments":[{"comment":"The closed-form evaluation of L(c) is false as stated. Eq. (30) minimizes over γ_x without any dual-feasibility domain. Take d=2, S=diag(1,2), c²=9/4, g=(1,0). The primal problem is max x₁ subject to x₁²+x₂²=1 and x₁²+4x₂²=9/4, whose value is √(7/12)=0.7638. The RHS of (30) is inf_{γ_x} √((γ_x+9/4)/(γ_x+1)) over real γ_x with nonnegative radicand, which is 0, attained at γ_x=−9/4. Equality would only hold under an additional restriction such as γ_x+s_i²≤0 (here γ_x≤−4), which is never stated. Since Theorems 3 and 11 inherit this step, the upper-bound proof and the exact-limit proof are not valid as written.","section":"Section 3.3, Eq. (30)"},{"comment":"Equation (26) asserts strong duality between the QCQP L(c)=max_{‖x‖=1,‖Sx‖=c} g^T Sx and its Lagrangian dual. The constraints are two nonconvex quadratic equalities, and no Slater-type, S-procedure, or coercivity argument is provided. For such problems a positive duality gap is possible, and every subsequent closed-form expression for L(c), including (30) and the final limit (151), depends on this swap. The paper needs a rigorous justification of (26) or a different derivation of L(c).","section":"Section 3.3, Eq. (26)"},{"comment":"The paper repeatedly uses concentration and interchanges of E with max_c and with lim_d. Eq. (23) writes Eλ_n(Σ̂−Σ)=max_c((Eξ)²−c²) based on the assertion in Remark 1 that all objects 'trivially concentrate.' Uniform concentration over the compact c-domain is a nontrivial ingredient that is not proved. Likewise, (31)–(33) pass limits through the max over c and min over γ_x. These interchanges are load-bearing for the upper bound and hence for Theorem 11; a rigorous treatment with quantitative tail bounds is needed.","section":"Remark 1; Eqs. (23), (33)"},{"comment":"The lower-bound matching step relies on the flatness implication imported from [120,121], and the verification of condition (118) is not fully rigorous. Theorem 8's proof states 'from (39), one also has γ̃_x ≤ 2c²' but (39) actually gives γ̃_x≤0 and γ̃_x+2c²≤0, i.e., γ̃_x≤−2c²; the subsequent claim γ̃_x≠−c² requires this corrected inequality and additional justification. More importantly, condition (118) is a global inequality over t∈(0,1) and q∈(−1,1), and the stationary-point contradiction in Theorem 8 does not address all possible boundary or infimum cases. The numerical check in Figure 1 for one spectrum is suggestive but does not constitute a proof for general Σ.","section":"Section 3.4.2 and Theorem 8"}],"minor_comments":[{"comment":"The paper uses m and n interchangeably in several places (e.g., the proof of Theorem 1 and eq. (64)), and 'y∈S^m' appears where S^n is intended.","section":"General notation"},{"comment":"There is a typo '1‘/2c²' in the expression for EG_u(X^(a1))G_u(X^(a2)); the correct factor should be 1/(2c²).","section":"Eq. (64)"},{"comment":"The simulation section does not report the number of Monte-Carlo trials, error bars, or the exact simulation protocol. This makes it hard to assess the claimed 'excellent agreement.'","section":"Section 3.6, Figures 2–3"},{"comment":"The parenthetical 'from (39), one also has γ̃_x ≤ 2c²' appears to be a sign typo; (39) implies γ̃_x ≤ −2c². Please correct and clarify the implication for γ̃_x ≠ −c².","section":"Theorem 8 proof"},{"comment":"The statement 'We tested quite a few ensembles and always obtained that (118) holds' is informal. If this is only numerical evidence, it should be clearly labeled as such and not used as a substitute for a proof.","section":"Section 3.4.2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper proposes an interesting approach and a plausible final formula, but the current proof is not reliable: the counterexample to Eq. (30) and the unproved strong-duality step undermine the central derivation. I believe these gaps may be repairable within the manuscript's scope, so I recommend major revision rather than rejection. I would also ask the editor to encourage the authors to state precisely all domain restrictions on the dual variables and to provide a rigorous concentration argument for the max-over-c interchange."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper has an appealing and likely correct target: an exact large-limit value for the operator-norm error of the sample covariance for Gaussian data with arbitrary covariance. Second, the written proof of that target is not valid at a load-bearing point. I would not treat Theorem 11 as proven on the strength of this text.\n\nThe genuinely new content is formula (151), which goes beyond the scaling-order results of Koltchinskii–Lounici and gives a precise limit for non-isotropic Σ. The bilinear-quadratic lower-bound mechanism (Theorems 4–8) is a real new idea, and the authors are honest that the upper bound (22) was already in Han’s work. The isotropic reduction to the Wishart edge 2/√α + 1/α is a nice sanity check, and the simulations in Figures 2–3 show agreement for d in the thousands. That is real evidence the formula is right.\n\nThe problem is the upper-bound chain. Equation (26) asserts strong duality for a nonconvex QCQP, and the subsequent minimization in (30) omits the dual-feasibility domain. This is not just a missing Slater-type argument; the stress-test example is concrete and correct. Take S = diag(1,2), c² = 9/4, g = (1,0). The primal L(c) is √(7/12) ≈ 0.764. The right-hand side of (30) is min over γx of √((γx + 9/4)/(γx + 1)) . Over real γx with nonnegative radicand, this infimum is 0, attained at γx = −9/4. Even if you impose the natural dual-feasibility constraint γx + sᵢ² ≥ 0, the infimum is 1, still not the primal value. So (30) is false as stated, and Theorem 3’s upper bound collapses. Since Theorem 11 explicitly combines that upper bound with the lower-bound matching, the central theorem is left without a valid proof.\n\nThe lower-bound side is also imported rather than proved: condition (118) is verified numerically and by a stationary-point argument in Theorem 8, but the flatness implication from [120,121] is invoked without proof. The concentration remarks are informal. These are not minor quibbles anywhere—they are the core of the proof.\n\nWho gets value from this? People working on high-dimensional covariance estimation and random matrix theory will care about formula (151) and the lower-bound mechanism. But the paper currently reads as a claim with insufficient proof. I would not cite it as a theorem yet.\n\nRecommendation: send it to a serious referee. The result is important enough and the approach novel enough that referee time is warranted, but the author should be pushed to either fix the duality step or state the result as a conjecture with the gap clearly flagged.","headline":"A plausible exact limit for sample covariance spectral norm error, but the upper-bound derivation contains a concrete duality gap that invalidates the proof as written.","tokens_in":33071,"tokens_out":7806,"would_cite":false,"duration_ms":68754,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","60B20","62E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"The sample covariance error has an exact limiting formula in high dimensions.","keywords":["sample covariance matrix","spectral norm error","random duality theory","Gaussian comparison","proportional high-dimensional asymptotics","effective rank","bilinear-quadratic comparison","sample size tradeoff"],"falsifier":"Choose a deterministic diagonal covariance with a non-trivial spectrum (e.g., s_i equally spaced in [0.5,1]), compute δ(α) from (151) for several α, and compare against high-precision Monte Carlo estimates of E∥Σ̂−Σ∥₂ for d and n around 10,000. A mismatch beyond the expected concentration scale would falsify the claimed equality. A more targeted check evaluates the two-replica condition (118) numerically for a spectrum with two separated eigenvalue clusters; if the inequality reverses, the lower bound no longer matches the upper bound.","tokens_in":31985,"feed_emoji":"📊","tokens_out":8042,"duration_ms":81253,"temperature":0.7,"pith_summary":"The paper aims to move past scaling-order estimates and determine the exact expected spectral-norm error of the sample covariance matrix for centered Gaussian data. In the proportional regime where n/d tends to a fixed α, it claims the error E∥Σ̂−Σ∥₂ converges to a closed-form expression derived from the true covariance spectrum. The expression reproduces the known isotropic edge when Σ = I, and numerical simulations already track the prediction at dimensions in the thousands. A correct formula of this kind would let practitioners compute the precise benefit of adding samples instead of relying on order-of-magnitude bounds.","feed_headline":"Sample covariance error has an exact limit formula","feed_subtitle":"The spectral-norm gap becomes a closed-form function of the covariance spectrum, replacing qualitative scaling bounds.","key_machinery":"Random duality theory (RDT), the paper's central device, rewrites the error as a maximum over the sphere of a Gaussian process — the random primal ξ(c) — and bounds it from above and below by two Gaussian comparison processes. The upper bound uses a linear 'random dual' L(c); the lower bound introduces a bilinear-quadratic process B(c). A two-replica argument on the overlap q of two copies of the maximization problem is then invoked to show the two bounds coincide. The final formula is organized by the spectral sum φ₁(γ) = (1/d) Σ sᵢ⁴/(γ+sᵢ²)² and the optimal parameter γ̂_x defined by equation (46).","core_discovery":"The paper's central claim is that for centered Gaussian vectors with covariance Σ = U S² Uᵀ, in the limit n/d → α, the expected spectral norm of the sample covariance error converges to the closed-form quantity δ(α) = lim_d γ̂_x √φ₁(γ̂_x) / (√φ₁(γ̂_x) − √α), where φ₁(γ) = (1/d) Σᵢ sᵢ⁴/(γ+sᵢ²)² and γ̂_x solves the stationarity condition lim_d (φ₁(γ̂_x) − φ₂(γ̂_x)φ₃(γ̂_x)) = 0. The author derives this value by sandwiching the error between an RDT upper bound and a new bilinear-quadratic lower bound, then showing the two match in the large-d limit. In the isotropic case Σ = I the formula reduces to the familiar edge 2/√α + 1/α.","pith_inferences":["One immediate editorial extension: differentiating the closed form with respect to α gives the marginal value of an extra sample, a quantity the paper does not compute explicitly.","Remark 4 of the paper already notes that for random covariances with a spectral density the empirical sums can be replaced by integrals; replacing them yields a fully analytic prediction for such priors.","A useful stress test is a two-cluster or near-degenerate spectrum, where the two-replica condition (118) rests on numerical verification and the contradiction proof in Theorem 8 is least transparent; agreement there would strengthen confidence in the formula's generality.","The matching upper/lower template suggests the same RDT sandwich could produce exact error formulas for related problems such as spiked covariance estimation or covariance estimation under missing data, though the paper leaves those extensions for future work."],"forward_implications":["If the formula holds, the exact error for any Gaussian covariance spectrum can be obtained by solving the scalar equation (46), without Monte Carlo simulation.","It converts the qualitative effective-rank scaling of earlier work into an exact large-d limit, allowing precise statements about how doubling or tripling n changes the error.","The isotropic reduction to 2/√α + 1/α ties the result directly to classical random matrix edges.","Because the framework is built generically, the same upper/lower comparison strategy is intended to carry over to other covariance error metrics and structured covariance models."],"fun_headline_variants":["Covariance error gets exact large-dim limit","Sample covariance error: closed-form limit","RDT nails spectral norm of sample covariance","Exact limit for sample covariance error","Precise spectral norm error for covariance"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing step is the unproved strong-duality swap in equation (26), which equates a non-convex spherical quadratic maximization with its Lagrangian min-max, together with the two-replica condition (82)/(118) that the paper verifies numerically and proves only by contradiction; if either fails, the closed-form limits (48) and (151) collapse.","fun_headline_variants_meta":{"raw":{"variants":["Covariance error gets exact large-dim limit","Sample covariance error: closed-form limit","RDT nails spectral norm of sample covariance","Exact limit for sample covariance error","Precise spectral norm error for covariance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000159,"raw_usage":{"total_tokens":1062,"prompt_tokens":735,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":479,"tokens_out":327,"duration_ms":3726,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:02:33.056080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a deterministic diagonal covariance with a non-trivial spectrum (e.g., s_i equally spaced in [0.5,1]), compute δ(α) from (151) for several α, and compare against high-precision Monte Carlo estimates of E∥Σ̂−Σ∥₂ for d and n around 10,000. A mismatch beyond the expected concentration scale would falsify the claimed equality. A more targeted check evaluates the two-replica condition (118) numerically for a spectrum with two separated eigenvalue clusters; if the inequality reverses, the lower bound no longer matches the upper bound.","supporting_citations":[],"review_version":1}