{"id":"efbbf354-5305-4476-9596-e3e90b010f9d","arxiv_id":"2411.17087","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For three classes of limit processes, M-estimator limiting distributions are symmetric when the drift, the stochastic process, and the constraint cone are all even; for linear-process limits this evenness condition is also necessary under weak assumptions.","lead":"This paper gives conditions under which the limiting distribution of an M-estimator is symmetric around the target parameter, covering three known families of non-normal limits. Symmetry matters because it lets a recent confidence-interval method, HulC, produce valid inference in irregular problems where standard Wald or bootstrap methods fail.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.2 omits the a.s.-uniqueness hypothesis it needs, and Lemma 3.1's proof relies on a false strict-infimum identity even when the argmin is unique.","rationale":"The reader correctly identified uniqueness as a load-bearing condition, but the issue is more serious than the reader's statement suggests. Uniqueness is not merely needed to rule out asymmetric tie-breaking; it is absent from the statement of Theorem 3.2, making the theorem false as written. Moreover, the proof of Lemma 3.1 contains a concrete error even in the unique case: the event {W ∈ A} is not equal to {inf_A Z < inf_{A^c} Z} merely because the argmin is unique, since the infimum over the complement can equal the minimum without being attained. The deterministic example Z(u) = u^4 with A = {0} makes this precise and is within the paper's own framework. These are fixable issues: Lemma 3.1 can be repaired by observing that the argmin functional is equivariant under the reflection u ↦ -u, and Theorem 3.2 can be repaired by adding the a.s.-uniqueness hypothesis or by explicitly choosing the argmin via a symmetric measurable selection. Because the high-level symmetry result is likely correct under appropriate hypotheses and because the necessity arguments in Section 4 are substantially independent of this proof gap, the appropriate verdict remains CONDITIONAL rather than REJECT or ACCEPT. The reader's conditional verdict is therefore upheld, but the reasoning needs to be sharpened to flag the false identity in Lemma 3.1 and the missing hypothesis in Theorem 3.2.","tokens_in":23194,"tokens_out":18906,"duration_ms":196807,"concrete_test":"Verify Lemma 3.1's proof identity with Z(u) = u^4 on R and A = {0}: the unique minimizer is 0, so P(W ∈ A) = 1, but P(inf_A Z < inf_{A^c} Z) = P(0 < 0) = 0, showing the printed event identity is false. Separately, test Theorem 3.2 on the deterministic even process Z ≡ 0 on Θ* = R with the selection W ≡ 1: all stated hypotheses hold but the conclusion fails, confirming that an a.s.-uniqueness or symmetric-tie-breaking condition must be added.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central sufficiency result is not stated correctly. Theorem 3.2 concludes that W = argmin(D+S+X) is symmetric from evenness of D, X, S, but it does not assume the argmin is almost surely unique. Without uniqueness, 'W' is not a well-defined random variable unless a tie-breaking rule is specified, and a bad tie-break can break symmetry. For example, take R as the domain, D ≡ 0, S ≡ 0, X ≡ 0. Every real number is a minimizer, the process is even, and the constant selection W ≡ 1 is not symmetric. So Theorem 3.2 as written is false. This matters for the HulC application because Class II and III examples do not all come with established uniqueness; the paper only proves general uniqueness for parts of Class I (Theorem 4.3), and Example 2.7 explicitly conditions on uniqueness as an external assumption. There is also a proof problem under the uniqueness assumption: Lemma 3.1 asserts P(W ∈ A) = P(inf_A Z < inf_{A^c} Z) for every Borel A. This identity does not follow from a.s. uniqueness. For the deterministic even process Z(u) = u^4, the unique minimizer is W = 0; with A = {0}, the left side is 1 but the right side is P(0 < 0) = 0. The conclusion of Lemma 3.1 is true and can be proved by equivariance of the argmin map under u ↦ -u, but the printed proof is invalid. The paper therefore needs a corrected proof and an explicit uniqueness or symmetric-tie-breaking hypothesis in Theorem 3.2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies conditions under which the limiting distributions of M-estimators, represented as minimizers of processes Z(u)=D(u)+S(u)+X(u), are symmetric around zero. Symmetry of the limiting distribution is used to guarantee zero limiting median bias, which is the requirement for valid inference via the HulC method of Kuchibhotla et al. (2024). The paper gives a general sufficient condition (Theorem 3.2) based on evenness of the deterministic drift D, the constraint indicator X, and the stochastic process S, and verifies this condition for Pflug's three classes of limiting processes: linear stochastic processes, Gaussian processes, and generalized Poisson hyperplane processes. It then proves necessity of the drift/constraint evenness assumption (A1) for Class I limits in one dimension (Theorem 4.4) and in higher dimensions under stronger symmetry of Y (Theorem 4.7), using convex analysis and optimal transport uniqueness.","tokens_in":23523,"tokens_out":2660,"duration_ms":25744,"significance":"If the main results stand, the paper delivers a useful and nontrivial contribution: it identifies a simple, checkable sufficient condition for symmetry of non-normal M-estimator limits, and it shows in an important special case that the condition is also necessary. The verification of evenness for several classical irregular estimators (shorth, LMS, mode, LAD, bridge) and the explicit discussion of when (A1) fails (e.g., Example 2.7, mode estimation at the boundary tuning parameter) give practitioners concrete guidance for when HulC is or is not applicable. The necessity proofs through optimal transport are elegant and appear internally consistent. However, the central sufficiency theorem and its supporting lemma are not stated correctly, as detailed below.","major_comments":[{"comment":"The statement P(W in A) = P(inf_A Z < inf_{A^c} Z) is false even under the stated almost-sure uniqueness assumption. For the deterministic even process Z(u)=u^4 on R, the unique minimizer is W=0; taking A={0}, the left-hand side equals 1 while the right-hand side is P(0<0)=0. The conclusion W d= -W is nonetheless true and can be proved by the equivariance of the argmin map under u -> -u, but the printed proof is invalid as written.","section":"Lemma 3.1"},{"comment":"Theorem 3.2 omits the almost-sure uniqueness hypothesis that Lemma 3.1 requires. As stated, the theorem is false: take Omega=R, D(u)=0, S(u)=0, and X the indicator of R. Every real number is a minimizer, the process is even, and the deterministic selection W=1 is not symmetric around zero. The theorem should either assume that argmin(D+S+X) is almost surely unique or specify a symmetric tie-breaking rule that makes W a well-defined random variable; the current wording does not.","section":"Theorem 3.2"},{"comment":"The missing uniqueness hypothesis is not merely cosmetic: the paper proves uniqueness for parts of Class I in Theorem 4.3, but for Classes II and III it does not provide general uniqueness results, and Example 2.7 explicitly conditions on the limiting process attaining a unique minimum almost surely as an external assumption. Since Theorem 3.2 is invoked for these classes, the manuscript needs to state explicitly which uniqueness condition is being assumed in each application, or to restrict the sufficiency claim accordingly.","section":"Section 4.1 and Example 2.7"}],"minor_comments":[{"comment":"There is a typographical artifact in the last paragraph of the introduction: 'min36imizers' should read 'minimizers'.","section":"Section 1"},{"comment":"The spelling 'Rockafeller' is used inconsistently; the standard spelling is 'Rockafellar'.","section":"References"},{"comment":"The phrase 'W = argmin_{u in Omega} D(u)+S(u)+X(u)' should specify that Omega is the effective domain and that the argmin is taken almost surely unique, to avoid ambiguity with the counterexample above.","section":"Theorem 3.2"},{"comment":"In the proof of Theorem 4.4, the sentence 'the derivatives here exist almost surely from almost sure uniqueness of minimizers' would benefit from a brief justification or a reference, since differentiability of the conjugate at the relevant point is what is being used.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The central flaw is confined to the statement and proof of Lemma 3.1 and Theorem 3.2, and it is correctable within the manuscript's scope: the sufficiency result survives once uniqueness (or a symmetric tie-breaking rule) is imposed and the proof is replaced by the equivariance argument. The necessity results in Sections 4.2-4.3 appear sound under their stated assumptions. I would not reject on this basis, but the current text cannot be published as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my read. The paper's core claim—that for Pflug's three classes you can certify symmetry of the limiting M-estimator distribution by checking evenness of D, symmetry of the stochastic process, and symmetry of the constraint cone—is sound in spirit and mostly sound in execution. The new content is real: the bridge-estimator dichotomy (A1 holds for all theta0 only for mu<1), the evenness criteria for Gaussian and Poisson hyperplane processes, and the optimal-transport-based necessity results for Class I. The necessity arguments are the most interesting part, and I did not find a hole in them. The paper should be engaged with.\n\nBut there are two correctable defects that need fixing before I'd call it done. Theorem 3.2 as stated omits a.s. uniqueness of the argmin. Without it W is not a random variable unless you pick a tie-break, and a bad tie-break breaks symmetry: D=S=X=0 on R with the constant selection W≡1 is a counterexample. The paper's own Section 4 acknowledges that uniqueness is needed, so this is a statement issue, not a fatal one. Lemma 3.1's proof is also wrong as printed: it asserts P(W∈A)=P(inf_A Z < inf_{A^c}Z) for any Borel A, which fails even for uniquely minimized Z(u)=u^4 with A={0}. The lemma's conclusion is true—it follows from equivariance of the argmin under u↦-u—but the proof given needs to be rewritten.\n\nA smaller issue: Theorem 4.2's proof is simply omitted, and neither the reader nor I could fully check the optimal transport steps in the supplement. These deserve to be spelled out or referenced more carefully. The self-citations to Kuchibhotla et al. are motivational and not a problem.\n\nOverall, the central arguments hold up. The paper delivers a practical, checkable criterion for when HulC inference is valid, and the limitations are stated honestly. It deserves a serious referee, and with the fixes above it should be acceptable.","headline":"Fix the statement of Theorem 3.2 and the proof of Lemma 3.1; the symmetry results are otherwise credible and useful.","tokens_in":24033,"tokens_out":2864,"would_cite":true,"duration_ms":25810,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62F25","62G20","60F17"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that an M-estimator's limiting distribution is symmetric around zero whenever the constraint set, the deterministic drift, and the stochastic process are all even in law, and proves for linear-process limits that…","keywords":["M-estimators","limiting distributions","symmetry","median bias","HulC","even stochastic process","tangent cone","cube-root asymptotics"],"falsifier":"Choose a one-dimensional linear-process limit $Z(u)=D(u)+uY$ with $Y$ standard normal and a superlinear convex $D$ that is not even, and compute the distribution of $W=\\arg\\min Z(u)$; Theorem 4.7 predicts this distribution is asymmetric whenever $D$ is not even, so an observed symmetric distribution in this setup would refute the necessity claim.","tokens_in":22973,"feed_emoji":"⚖️","tokens_out":14708,"duration_ms":119652,"temperature":0.7,"pith_summary":"The paper asks when the limiting distribution of an M-estimator—the scaled error of a minimizer of an empirical loss—is symmetric about zero. Its main sufficient condition is that the constraint cone is symmetric, the deterministic drift $D$ is even, and the mean-zero stochastic process $S$ has the same law as its reflection, in which case the argmin $W$ of $D+S+X$ satisfies $W \\stackrel{d}{=} -W$. For linear-process limits, with a symmetric, absolutely continuous, full-support random vector $Y$, the condition is also necessary, not just sufficient. Symmetric limits make under- and over-estimation equally likely, so the limiting median bias is zero—exactly what the HulC confidence-interval construction (a method that combines independent estimators) needs—and the paper extends valid inference to irregular, constrained, and cube-root-rate problems where normality and continuity-based methods do not apply.","feed_headline":"Three evenness conditions certify symmetric M-estimator limits","feed_subtitle":"For irregular estimators with non-normal limits, the same checks unlock valid median-bias-free confidence intervals.","key_machinery":"The load-bearing object is the limiting stochastic process $Z(u)=D(u)+S(u)+X(u)$ from display (5), where $X$ is $0$ on the tangent cone of the constraint set at the true parameter and $+\\infty$ outside it. A process is called even when it and its reflection have the same law; Lemma 3.1 and Theorem 3.2 transfer componentwise evenness plus almost-sure uniqueness of the argmin into $W \\stackrel{d}{=} -W$. For the necessity direction, the argument shifts to convex duality: the minimizer $W$ is identified as a subgradient of the convex conjugate of $D+X$ evaluated at $-Y$, and uniqueness of optimal transport maps forces two such subgradient selections to agree almost everywhere, which yields evenness of $D+X$. Epi-convergence supplies the bridge from finite-sample minimizers to the limiting argmin.","core_discovery":"The central claim is that the symmetry of a limiting argmin distribution is governed by the symmetry of the limit process's ingredients, not by the particular loss or estimator. Theorem 3.2 shows that if $Z(u)=D(u)+S(u)+X(u)$ and $Z(-u)$ have the same law and the argmin is almost surely unique, the random minimizer is symmetric about zero. Theorems 4.4 and 4.7 give the converse for the linear class $Z(u)=D(u)+\\langle u,Y\\rangle$ when $Y$ is symmetric, absolutely continuous, and dominates Lebesgue measure: then a symmetric $W$ forces $D+X$ to be even, so for Class I the paper obtains a complete characterisation. For Class II Gaussian-process limits and Class III Poisson-hyperplane limits, the same theorem supplies checkable sufficient conditions, which the paper verifies for shorth, least-median-of-squares, mode, bridge, LAD, and constrained regression examples.","pith_inferences":["Because Theorem 3.2 is stated for any even-in-law process, the same symmetry argument should apply to resampling or bootstrap versions of these limits; any bootstrap scheme that preserves componentwise evenness would inherit the symmetric law.","The paper leaves open whether a non-even drift outside the support of $Y$ can still produce a symmetric $W$; its own remarks suggest the true boundary is equality of two conjugates on the support of $Y$, which could be tested by constructing asymmetric $D$ that agree on that support.","Median unbiasedness holds under weaker conditions than full symmetry (Theorem 4.5), so inference procedures built on vanishing median bias rather than distributional symmetry could cover irregular cases where the symmetric-limit condition fails."],"forward_implications":["For cube-root-rate estimators such as shorth, least median of squares, and mode estimation with the tuning parameter below the boundary case, the Gaussian-process covariance is automatically even, so checking the drift and constraint cone decides symmetry of the limit.","For LAD regression, symmetry holds when the error distribution's point-mass increments just above and below zero balance in the limit; the paper's examples show asymmetric error increments lead to asymmetric limits and nonzero median bias.","For bridge and LASSO-type penalized regression, concave penalties ($\\mu<1$) give even drift at every true parameter, while convex penalties ($\\mu\\ge 1$) give even drift only at $\\theta_0=0$; uniform symmetric inference over the parameter space therefore cannot be certified for the convex case.","For constrained mean estimation, the tangent cone must be a symmetric set; when the true parameter lies on a boundary with an asymmetric cone, the limiting distribution is asymmetric even though the errors are Gaussian."],"supporting_citations":[{"why":"Classifies the three limiting-process families and supplies the covariance form (7) used in the Gaussian examples.","marker":"Pflug (1995)"},{"why":"Provides epi-convergence, conjugate duality, subdifferentials, and the variational analysis tools used throughout the proofs.","marker":"Rockafeller and Wets (1998)"},{"why":"Defines HulC and proves its validity from vanishing median bias, the inference goal motivating the symmetry results.","marker":"Kuchibhotla et al. (2024)"},{"why":"Derives the cube-root asymptotics used for the shorth, least-median-of-squares, and related examples.","marker":"Kim and Pollard (1990)"},{"why":"Derives the LAD regression limiting process whose drift evenness is analyzed in Example 2.2.","marker":"Knight (1999a)"},{"why":"Derives the bridge and LASSO limiting drift whose evenness properties drive the symmetry conclusions in Example 2.3.","marker":"Knight and Fu (2000)"},{"why":"Establishes constrained M-estimation limits through tangent cones, forming the Class I framework.","marker":"Geyer (1994)"},{"why":"Develops constrained local M-estimator asymptotics and tangential regularity, used for the constraint term X.","marker":"Shapiro (2000)"},{"why":"Supplies the one-dimensional optimal transport uniqueness theorem used in the proof of Theorem 4.4.","marker":"Ambrosio (2003)"},{"why":"Supplies the multidimensional optimal transport uniqueness result used in the proof of Theorem 4.7.","marker":"Villani (2021)"}],"fun_headline_variants":["Symmetry conditions unlock inference for non-normal M-estimator limits","Characterizing symmetric limit laws for irregular M-estimators","Evenness conditions certify symmetric M-estimator limits","Symmetry of argmin distributions via process symmetry","Complete symmetry test for linear-class M-estimator limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the normalized estimator converges in distribution to the almost surely unique minimizer of $D+S+X$, since uniqueness is what lets symmetry of the ingredients be transferred to the argmin, and for Classes II and III this uniqueness is assumed externally for the examples.","fun_headline_variants_meta":{"raw":{"variants":["Symmetry conditions unlock inference for non-normal M-estimator limits","Characterizing symmetric limit laws for irregular M-estimators","Evenness conditions certify symmetric M-estimator limits","Symmetry of argmin distributions via process symmetry","Complete symmetry test for linear-class M-estimator limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000596,"raw_usage":{"total_tokens":2751,"prompt_tokens":869,"completion_tokens":1882,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":1802}},"tokens_in":485,"tokens_out":1882,"duration_ms":13867,"temperature":1.0,"reasoning_tokens":1802,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:32:36.355550+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose a one-dimensional linear-process limit $Z(u)=D(u)+uY$ with $Y$ standard normal and a superlinear convex $D$ that is not even, and compute the distribution of $W=\\arg\\min Z(u)$; Theorem 4.7 predicts this distribution is asymmetric whenever $D$ is not even, so an observed symmetric distribution in this setup would refute the necessity claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Classifies the three limiting-process families and supplies the covariance form (7) used in the Gaussian examples."},{"cited_title":"and Wets, R","cited_arxiv_id":null,"evidence_quote":"Provides epi-convergence, conjugate duality, subdifferentials, and the variational analysis tools used throughout the proofs."},{"cited_title":"K., Balakrishnan, S., and Wasserman, L","cited_arxiv_id":null,"evidence_quote":"Defines HulC and proves its validity from vanishing median bias, the inference goal motivating the symmetry results."},{"cited_title":"and Pollard, D","cited_arxiv_id":null,"evidence_quote":"Derives the cube-root asymptotics used for the shorth, least-median-of-squares, and related examples."},{"cited_title":"and Fu, W","cited_arxiv_id":null,"evidence_quote":"Derives the bridge and LASSO limiting drift whose evenness properties drive the symmetry conclusions in Example 2.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes constrained M-estimation limits through tangent cones, forming the Class I framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Develops constrained local M-estimator asymptotics and tangential regularity, used for the constraint term X."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the one-dimensional optimal transport uniqueness theorem used in the proof of Theorem 4.4."}],"review_version":1}