{"id":"c4946d63-55b0-469e-8675-ea77f25d5b02","arxiv_id":"2606.27899","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Exact large-deviation characterization of minimax mean estimation error probability in symmetric moment classes via two-point Hellinger exponent, achieved by M-estimator from two-parameter convex program.","lead":"The paper derives an exact large-deviation rate for the worst-case probability that any estimator exceeds a fixed error margin when the data distribution has only a known bound on an even moment. A smart generalist might read it to see how convex optimization yields provably optimal robust estimators for heavy-tailed data.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption identifies the moment-class membership, which is the explicit modeling hypothesis rather than a potential gap in the derivation. The abstract-only limitation is noted but does not create a technical objection once the full text (containing the duality argument and explicit constructions) is taken as given. No load-bearing internal inconsistency or unsecured step is visible in the central claim.","tokens_in":1936,"tokens_out":390,"duration_ms":106417,"concrete_test":"For the bounded-variance case, extract the two multipliers from the convex program, form the corresponding estimating function ψ, and compute the exact finite-n error probability of the resulting monotone M-estimator under each of the three-atomic least-favorable distributions; check whether it is ≤ exp(−n r(Δ)) with equality in the large-n limit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the minimax β_n(Δ) admits the exact exponential rate r(Δ) given by the worst-case two-point Hellinger affinity over the shifted classes C_{±Δ}, with the upper bound β_n(Δ) ≤ exp(−n r(Δ)) attained for every finite n by an M-estimator whose estimating function is obtained from a two-parameter convex program via Lagrangian duality. The argument is internally consistent: the reduction of the infinite-dimensional search over monotone estimating functions to two multipliers is a standard convex-analytic device, the emergence of the sandwich envelope is a natural consequence, the explicit r(Δ) = ½ log(1 + Δ²/σ²) for ϕ(x) = x² recovers a known form, and the least-favorable distributions being at most three-atomic is compatible with moment-constrained extremal problems. The modeling assumption that P lies in C_0 is the problem statement, not an unsecured premise.","agreement_with_reader":"disagree"},"referee_report":{"model":"grok-4.3","summary":"The paper studies optimal estimation of the location parameter for distributions in a symmetric moment class C_0 (mean-zero with bounded even moment ∫ϕ dP ≤ B). It claims an exact large-deviation characterization of the minimax error probability β_n(Δ) in the fixed-margin regime: the exponential rate is the two-point Hellinger exponent r(Δ) = -log sup ∫√(dP_{-Δ} dP_Δ) over shifted classes C_{±Δ}, achieved non-asymptotically by β_n(Δ) ≤ exp(-n r(Δ)) via a monotone M-estimator synthesized from a two-parameter convex program. Lagrangian duality reduces the search over estimating functions to two multipliers yielding sandwich envelopes. For ϕ(x)=x² it recovers r(Δ)=½ log(1+Δ²/σ²); in the fixed-confidence regime it matches Catoni's √2 constant and Lee-Bhatt et al. constants, shown tight, with least-favorable distributions at most three-atomic.","tokens_in":2119,"tokens_out":458,"duration_ms":26440,"significance":"If the central derivations hold, the result is significant for providing the first exact non-asymptotic exponential rate for this problem class, together with an explicit convex-analytic construction of the optimal estimator. The reduction of the infinite-dimensional monotone-function search to two multipliers, the natural emergence of the sandwich envelope, and the recovery of sharp constants (including tightness for α-moments) are notable strengths. The three-atomic least-favorable distributions and the explicit r(Δ) for the variance case further strengthen the contribution.","major_comments":[],"minor_comments":[{"comment":"The abstract states that the estimator is 'synthesized from a two-parameter convex program' but does not specify the precise objective and constraint set of that program; a short display of the program (e.g., in §3) would improve readability.","section":null},{"comment":"Notation for the shifted classes C_{±Δ} is introduced only in the abstract; an explicit definition in the main text (near the statement of the main theorem) would avoid any ambiguity for readers.","section":null}],"recommendation":"accept","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their positive summary, significance assessment, and recommendation to accept the manuscript.","responses":[],"tokens_in":1535,"tokens_out":38,"duration_ms":12217,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main result is an exact large-deviation rate for the fixed-margin error probability β_n(Δ) under symmetric moment constraints. The rate equals the two-point Hellinger exponent over the shifted classes, and a monotone M-estimator built from a two-parameter convex program achieves the bound for every finite n.\n\nLagrangian duality collapses the search over estimating functions to two multipliers that define the optimal envelopes. This explains the sandwich shape that earlier papers introduced by hand. For bounded variance the exponent matches the known form, and in the fixed-confidence regime the same construction recovers the sharp constants of Catoni and of Lee-Bhatt, proving they are tight. The least-favorable distributions having at most three atoms follows naturally from the moment constraints.\n\nThe argument looks internally consistent and the modeling assumptions are just the problem statement. The citation pattern is appropriate because the paper focuses on matching and tightening existing constants rather than unrelated claims.\n\nA minor question is whether the duality step requires extra regularity on ϕ beyond being even; the abstract indicates it works for the classes considered, including slowly varying cases where optimality holds only to leading order. That is a small caveat rather than a load-bearing issue.\n\nThe paper is for readers in robust statistics who care about exact rates and explicit constructions under weak moment assumptions. It deserves a serious referee because the claims are sharp enough to be checked and the method is reproducible in principle.","headline":"The paper pins down an exact non-asymptotic minimax rate for heavy-tailed mean estimation by reducing the problem to a two-parameter convex program via duality.","tokens_in":2603,"tokens_out":360,"would_cite":true,"duration_ms":37057,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A two-parameter convex program yields a monotone M-estimator that attains the exact minimax exponential rate for mean estimation under moment bounds.","keywords":["heavy-tailed mean estimation","convex optimization","M-estimator","Hellinger affinity","minimax estimation","large deviations","moment constraints","robust statistics"],"falsifier":"A concrete distribution in C_0 together with an n and Δ for which the constructed M-estimator has error probability strictly larger than exp(−n r(Δ)) would show that the claimed rate is not achieved.","tokens_in":2833,"feed_emoji":"📈","tokens_out":773,"duration_ms":29234,"temperature":0.7,"pith_summary":"The paper determines the smallest worst-case probability β_n(Δ) that any estimator can guarantee of exceeding a fixed error margin Δ when the underlying distribution belongs to a symmetric moment class. This probability decays at an exact exponential rate given by the two-point Hellinger affinity between the worst-case distributions shifted to means ±Δ. The rate holds non-asymptotically and is achieved by an M-estimator constructed from a convex program whose duality reduces the infinite-dimensional problem to two multipliers that define optimal estimating functions. For bounded variance the rate takes the explicit form ½ log(1 + Δ²/σ²), and the same construction recovers the sharp constants of Catoni and of Lee-Bhatt et al. in the shrinking-margin regime.","feed_headline":"Convex program attains exact rate for heavy-tailed mean estimation","feed_subtitle":"The monotone M-estimator matches the two-point Hellinger exponent over shifted moment classes and recovers known sharp constants.","key_machinery":"The two-parameter convex program whose Lagrangian duality collapses the search over estimating functions to a pair of multipliers that generate the optimal sandwich-shaped envelopes.","core_discovery":"The minimax error probability β_n(Δ) satisfies β_n(Δ) ≤ exp(−n r(Δ)) where r(Δ) equals −log of the supremum of the Hellinger affinity ∫√(dP_{−Δ} dP_Δ) taken over all pairs of distributions in the shifted moment class C_{±Δ}; equality in the exponential rate is attained for every finite n by the monotone M-estimator synthesized from the two-parameter convex program, whose least-favorable distributions are supported on at most three atoms.","pith_inferences":["The duality reduction to two multipliers may simplify computation of the estimator in practice for given data.","Similar convex programs could be derived for other location or scale estimation tasks under moment constraints.","The three-atom structure of the worst-case distributions suggests that discretizing the moment class may preserve optimality.","The approach supplies a template for proving tightness of other ad-hoc robust estimators that rely on envelope functions."],"forward_implications":["The bound β_n(Δ) ≤ exp(−n r(Δ)) holds for every finite sample size, not merely in the large-n limit.","When ϕ(x) = x² the exponent simplifies to ½ log(1 + Δ²/σ²).","As the target β tends to zero the estimator attains the sharp constant √2 for bounded variance.","It attains the constant L(α) for bounded α-moments with α ∈ (1,2).","The least-favorable distributions remain supported on at most three atoms for every concrete class examined."],"fun_headline_variants":["Convex program yields exact Hellinger rate for means","M-estimator attains two-point Hellinger exponent","Duality pins minimax rate in moment classes","Monotone estimator matches affinity bound exactly","Three-atom laws achieve optimal heavy-tail rate"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The unknown distribution lies in the symmetric moment class C_0 of mean-zero laws satisfying a bound on the integral of a fixed even function ϕ.","fun_headline_variants_meta":{"raw":{"variants":["Convex program yields exact Hellinger rate for means","M-estimator attains two-point Hellinger exponent","Duality pins minimax rate in moment classes","Monotone estimator matches affinity bound exactly","Three-atom laws achieve optimal heavy-tail rate"]},"model":"grok-4.3","cost_usd":0.004856,"raw_usage":{"total_tokens":2486,"prompt_tokens":872,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":48562000,"prompt_tokens_details":{"text_tokens":872,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1546,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":872,"tokens_out":68,"duration_ms":18519,"temperature":1.0,"reasoning_tokens":1546,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T02:41:02.269490+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A concrete distribution in C_0 together with an n and Δ for which the constructed M-estimator has error probability strictly larger than exp(−n r(Δ)) would show that the claimed rate is not achieved.","supporting_citations":[],"review_version":1}