{"id":"787e9a45-2fa5-4dcb-b8d1-6d827c4d8f08","arxiv_id":"2510.13703","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified Le Cam-style asymptotic efficiency theory, including convolution and local asymptotic minimax theorems, is established for parameters taking values in Riemannian manifolds, and the influence-function calculus carries over.","lead":"This paper builds a statistical efficiency theory—the study of the best possible estimation error—for parameters that live on curved spaces called Riemannian manifolds instead of ordinary flat space. It supplies the geometric vocabulary and proves the key optimality theorems, then shows the theory reproduces and unifies known efficiency results for Fréchet means and single-index models.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's semiparametric convolution is misstated: the proof (eq. 62) yields N(0,V_x) with V_x=E[IF⊗IF], not N(0,G_P^{-1}); G_P is only defined for a submodel, so the central theorem is internally inconsistent as written.","rationale":"The reader's weakest_assumption points to uniform control of geometric expansions as the most fragile premise. On inspection, this is not the most load-bearing concern: for the fixed local perturbations θ_{n,h} = Exp_θ(h/√n) used in Theorem 3.1, the geodesics shrink to a point, so smoothness plus the existence of a normal neighborhood at each point gives the needed o(1) remainders without global bounded curvature or injectivity radius. Uniformity over h in compact sets, as needed for Theorem 3.3, can be supplied by continuity of the curvature and the prior's compact support; the paper's Assumption C.2 is plausible. The single-index application's hidden Gaussian error assumption is real and serious, but it is an application-level flaw rather than a flaw in the central theory. The most concrete and central issue is Theorem 4.1's misstatement: the proof and the statement disagree about the covariance of the Gaussian convolution factor. This is an internal inconsistency in the very theorem that defines the semiparametric efficiency bound, and it needs correction (either restate as N(0,V_χ) or redefine G_P as the efficient information matrix). Since the proof itself appears to establish the correct V_χ result, the appropriate verdict remains CONDITIONAL pending the correction, so the reader's verdict is unchanged.","tokens_in":40547,"tokens_out":20595,"duration_ms":170178,"concrete_test":"Re-derive the Gaussian covariance in Theorem 4.1 from the characteristic-function argument in Appendix D.1. In particular, verify that equation (62) implies the Gaussian factor has covariance V_χ = E[IF_χ⊗IF_χ]. Then take a concrete semiparametric model (e.g., mean estimation with unknown error distribution) and a one-dimensional submodel whose score s is not proportional to the efficient influence function IF. Compute G_P = E[s²] and V_χ = E[IF²]; if G_P^{-1} ≠ V_χ (which will be the case whenever E[IF·s]² < E[s²]E[IF²]), Theorem 4.1's statement as written is false. This check settles whether the theorem must be corrected to N(0,V_χ).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central semiparametric claim, Theorem 4.1, states that for any regular estimator of a differentiable functional χ(P) on a manifold, the limiting law decomposes as L_P = N(0, G_P^{-1}) ∗ Δ_P. But G_P was defined in Definition 4.1 as E[s⊗s] for a one-dimensional submodel; it is not a quantity attached to the full semiparametric model. The proof in Appendix D.1, specifically equations (59)–(62), derives instead a convolution with the Gaussian factor N(0, V_χ), where V_χ = E[IF_χ⊗IF_χ] is the covariance of the efficient influence operator. Equality between G_P^{-1} and V_χ holds only if the score s used to define G_P is the efficient score S_eff = V_χ^{-1}IF_χ, which is neither assumed nor stated. For a generic submodel, G_P^{-1} = 1/E[s²] while V_χ ≥ (E[IF·s])²/E[s²], with equality only when IF ∝ s. Therefore the theorem as written is either ill-posed (G_P not defined for the semiparametric model) or false if G_P is read as the submodel information. This is not a mere typo: it concerns the central efficiency bound that the framework claims to establish.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops an asymptotic efficiency theory for statistical models whose parameter of interest lies on a Riemannian manifold. It introduces manifold analogues of DQM/LAN, regular estimators, the Hájek–Le Cam convolution theorem, and the local asymptotic minimax theorem, and extends these to differentiable functionals over semiparametric models. The framework is then applied to two examples: the Fréchet mean, where the sample Fréchet mean is shown to attain the semiparametric bound E[IF⊗IF], and the single-index model, where the coefficient bound is derived from the manifold calculus.","tokens_in":40812,"tokens_out":5697,"duration_ms":49962,"significance":"If the main results are correct, this would be a substantial unification: it brings the classical linear-space machinery of efficiency theory to a class of nonlinear parameter spaces and reduces previously case-specific derivations to a common geometric language. A reassuring check is that the two application bounds are consistent with existing external results: the Fréchet-mean covariance matches Bhattacharya–Patrangenaru, and the single-index bound matches Kuchibhotla–Patra. The paper also provides a useful conceptual vocabulary (Table 1) and explicit influence-operator formulas. However, the central semiparametric theorem has a statement/proof mismatch, and some geometric uniformity conditions are not stated as primitive assumptions.","major_comments":[{"comment":"Theorem 4.1 states that the limiting law of any regular estimator satisfies L_P = N(0, G_P^{-1}) * Δ_P. But G_P is defined in Definition 4.1 only for a one-dimensional parametric submodel as G_P = E[s⊗s]; it is not attached to the full semiparametric model. The proof in Appendix D.1, equations (59)–(62), derives a convolution with N(0, V_χ) where V_χ = E[IF_χ⊗IF_χ]. These two covariance operators coincide only when the score s used in G_P is the efficient score s_χ = V_χ^{-1}IF_χ, which is neither stated in Theorem 4.1 nor part of Definition 4.1. If G_P is read as the information of an arbitrary one-dimensional submodel, the statement is false: G_P^{-1} is a scalar while V_χ is the efficient covariance. This is a load-bearing inconsistency in the paper’s central semiparametric claim and must be fixed by restating the theorem in terms of V_χ or the efficient score, and by aligning the not","section":"Section 4, Definition 4.1, Theorem 4.1, Appendix D.1"},{"comment":"The proof of Theorem 3.1 relies on uniform Taylor expansions of Exp and Exp^{-1} with O(‖h‖^3) remainders (Lemma A.6), and on the transported-residual drift being o(1) at the √n scale, as used in equations (19), (53), and (54). The assumptions stated in the paper — completeness, smoothness, and Assumption 3.2 — do not provide primitive sufficient conditions under which these expansions hold uniformly along local perturbation paths. In particular, unbounded sectional curvature or a shrinking injectivity radius near the support of the estimator would break the required uniformity, and the parameter path ψ(θ_{n,h}) itself is not explicitly assumed to avoid the cut locus of ψ(θ). Please add explicit geometric conditions (e.g., bounded curvature, positive injectivity radius, uniform non-cut-locus neighborhoods) or prove that the current assumptions are sufficient.","section":"Section 3.2, Lemma C.6, Appendix A.2"},{"comment":"The efficient score formula in (34) uses (Y - g_0(β_0^T X))/σ^2, which corresponds to ℓ'_{ε|X}(ε) = ε/σ^2, i.e., Gaussian conditional errors. Model (29), however, is stated only with E(ε|X)=0 and finite variance σ^2; no Gaussian assumption is made. For non-Gaussian errors the efficient score should contain ℓ'_{ε|X}(ε) rather than ε/σ^2, so equation (34) and Theorem 5.3 are only valid under an additional Gaussian-error assumption (or under a different derivation). This should be stated explicitly, or the general formula should be derived.","section":"Section 5.2, equation (34), Theorem 5.3"}],"minor_comments":[{"comment":"The parameter space B is written as {β: ‖β‖=1, β_1 ≥ 0}, which is a manifold with boundary (a closed hemisphere). The exponential-map formula (31) and the tangent-space description are for the smooth sphere, and paths starting at boundary points may leave B. Since identifiability typically requires β_1 > 0, the set should be stated as an open hemisphere, or the boundary case should be addressed.","section":"Section 5.2, equation (30)"},{"comment":"The notation G_P is used in Theorem 4.1 for the semiparametric covariance while Definition 4.1 defines it only for a one-dimensional submodel. Consider using V_χ or E[S_eff⊗S_eff]^-1 consistently, and distinguish it from the parametric G_θ.","section":"Section 4, equations (24)–(25)"},{"comment":"Assumption C.1 is stated after the proof of Lemma 3.3 begins, but it is used in the proof. It should be moved before the lemma statement or explicitly referenced as a standing assumption in the lemma.","section":"Appendix C.4, Lemma 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is ambitious and the application results provide useful external checks, but the semiparametric convolution theorem as stated is internally inconsistent with its proof. The fix appears fairly local — replace G_P^{-1} by the efficient covariance V_χ derived in the proof — so revision rather than rejection seems appropriate. The geometric uniformity and Gaussian-error issues also need to be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper genuinely does something new: it builds a Le Cam-style efficiency vocabulary for Riemannian parameter manifolds, states and proves convolution and LAM theorems for parametric models, and develops an influence-operator calculus. The two applications match existing known bounds (Fréchet mean, single-index model), which is the strongest evidence that the underlying calculus is right. Second, the semiparametric headline, Theorem 4.1, is misstated. The proof in Appendix D.1 derives a convolution with N(0, V_x) where V_x = E[IF⊗IF]; the theorem instead states N(0, G_P^{-1}), where G_P is the Fisher information of a one-dimensional submodel. Those two coincide only when the submodel score is proportional to the efficient influence operator, which is neither assumed nor generally true. So the theorem as written is either ill-posed at the semiparametric level or false. The applications don't rely on the faulty statement, but the main claim of Section 4 does.\n\nWhat is genuinely good: the parametric convolution theorem and LAM theorem for general manifolds appear to be new, and the external consistency checks (Bhattacharya-Patrangenaru for Fréchet means, Kuchibhotla-Patra for single-index) are reassuring. The geometric regularity premise is the real soft spot: the proofs need uniform control of curvature and cut-locus structure along perturbation paths, but the paper only assumes completeness and almost-sure avoidance of the cut locus. Primitive conditions like bounded curvature or a positive injectivity radius should be stated explicitly.\n\nSmaller issues: the single-index example introduces 1/σ^2 in the efficient score (equation 34), silently assuming Gaussian errors, while the model only states E(ε|X)=0. Assumption C.2, on which the LAM theorem leans, is asserted to hold under continuity but verification is one sentence. A few regularity steps in Theorem 5.1 and Appendix B.1 are compressed.\n\nOverall, the framework is real and fixable. The misstated Theorem 4.1 should be corrected or the proof aligned, and the geometric assumptions made explicit. If you work in non-Euclidean statistics, this is worth engaging with now; it deserves peer review with a strong request for revision.","headline":"Real framework contribution, but the semiparametric headline theorem is misstated as written and needs fixing before the paper is usable.","tokens_in":41386,"tokens_out":2816,"would_cite":false,"duration_ms":25547,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62B05","62F12","62G20","62R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the classical efficiency bound—inverse Fisher information transported through the functional derivative—holds for regular estimators of manifold-valued parameters, and applies it to Fréchet means and single-index coeff","keywords":["asymptotic efficiency theory","semiparametric theory","Riemannian manifolds","convolution theorem","local asymptotic minimax","Fréchet mean","single-index model","influence operator"],"falsifier":"On a manifold with a point whose sectional curvature grows without bound near the support of the estimator, simulate a regular estimator (e.g., Fréchet mean on a family of tori with shrinking injectivity radius) and check whether the limiting distribution of √n Exp^{-1}_{ψ(θ)} ψ̂n is a convolution with N(0, ψ̇G⁻¹ψ̇*) or whether the Gaussian component is altered; equivalently, verify numerically whether the equality in Lemma A.6 (∇Exp^{-1}_{μ} μ_h = id + (1/6)R_μ(h,·)h + O(‖h‖³)) holds uniformly in h along the path h/√n.","tokens_in":40255,"feed_emoji":"📐","tokens_out":5547,"duration_ms":43686,"temperature":0.7,"pith_summary":"This paper extends Le Cam's asymptotic efficiency theory from normed linear spaces to Riemannian manifolds. The central result is a convolution theorem: the limiting law of any regular estimator of a manifold-valued parameter decomposes as a Gaussian with covariance ψ̇(θ)G⁻¹ψ̇(θ)* convolved with an independent nuisance measure, and no regular estimator can beat this bound. The same machinery yields a local asymptotic minimax theorem and a working calculus of influence operators for manifold-valued parameters. These results unify previously case-by-case efficiency bounds: the sample Fréchet mean attains the semiparametric bound, and the single-index coefficient bound follows from the exponential map on the unit hemisphere. The paper matters because modern datasets—covariance matrices, networks, shapes—live on nonlinear spaces, and this provides a single standard for optimal estimation there.","feed_headline":"Manifold-valued estimators obey the classical efficiency bound","feed_subtitle":"Fréchet mean and single-index coefficient estimators now have provably optimal variance.","key_machinery":"The central objects are the exponential map Expθ and its inverse logarithmic map Exp^{-1}μ, which replace addition and subtraction on a manifold, together with parallel transport Πψ(θ)ψ(θ') along the distance-minimizing geodesic, which moves residuals between tangent spaces and defines regular estimators. The key identity is the influence-operator equation ψ̇(θ) = E[IFψ·S(θ)], and the convolution decomposition Lψθ = N(0, ψ̇(θ)G⁻¹θψ̇(θ)*) * Δψ(θ). The workhorse technical fact is the second-order Jacobi-field expansion of Exp and Exp^{-1} (Lemma A.6), which shows curvature enters the finite-sample Cramér–Rao bound only through terms that vanish at √n scale.","core_discovery":"The central discovery is that the Hájék–Le Cam convolution theorem and the local asymptotic minimax theorem hold for parameters valued in a complete Riemannian manifold, with the same Gaussian component ψ̇(θ)G⁻¹θψ̇(θ)* as in linear spaces. The proof works by defining regular estimators through parallel transport of √n-scaled residuals across tangent spaces along the unique distance-minimizing geodesic, and by controlling the curvature remainders in the Taylor expansions of the exponential and logarithmic maps. As a consequence, the sample Fréchet mean is semiparametrically efficient with influence operator IFμ0 = {E[∇Exp^{-1}μ0 X]}⁻¹ Exp^{-1}μ0 X (Theorem 5.2), and the single-index regressio","pith_inferences":["The framework suggests that curvature affects only finite-sample Cramér–Rao bounds and vanishes asymptotically, implying that large-sample inference on curved spaces can safely ignore curvature—but also that non-asymptotic or higher-order efficiency will depend on curvature and should be studied separately.","The same vocabulary could be pushed to infinite-dimensional parameter manifolds (e.g., Wasserstein space of distributions), but the paper's reliance on finite-dimensional tangent-space isomorphism suggests new tools would be needed; a testable extension is whether the convolution theorem survives for Fréchet means in Wasserstein space with non-unique geodesics.","A reader can empirically check the regularity condition behind the theorem by computing bootstrap distributions of parallel-transported residuals: if the law of √n Πψ(θ)ψ(θ_{n,h}) Exp^{-1}_{ψ(θ_{n,h})} ψ̂n stabilizes in h, the estimator is regular and the bound applies."],"forward_implications":["Any regular estimator of a manifold-valued parameter has an asymptotic variance lower bound equal to the transported inverse Fisher information; a Hodges-type superefficient estimator is excluded by regularity in the manifold sense.","The sample Fréchet mean is asymptotically efficient among regular estimators and attains the semiparametric bound V = E[IF⊗²], including under missing-at-random mechanisms where the influence operator generalizes the classical MAR mean influence function.","Single-index model coefficients, constrained to a unit hemisphere, obtain a semiparametric efficiency bound without crafting ad hoc parametric submodels; the exponential map automatically produces valid submodel paths.","The calculus of influence operators—solving E[IF·S] = ψ̇(θ) and projecting onto the tangent space—is licit for manifold-valued parameters, so practitioners can construct one-step or debiased estimators on manifolds.","Local asymptotic minimax holds, so the bound applies to all estimator sequences, not just regular ones, with a curvature-dependent term that disappears in the large-n limit."],"fun_headline_variants":["Efficiency theory now covers Riemannian manifold parameters","Fréchet mean and single-index estimators achieve optimal variance","Hajek-Le Cam bounds hold for estimators on manifolds","Optimality proven for estimators on curved spaces"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assumes that along local perturbation paths the curvature of the parameter manifold is controlled well enough that the second-order Taylor expansions of the exponential and logarithmic maps hold uniformly, so the parallel-transported residuals drift by o(1) at the √n scale; if the manifold has unbounded curvature or the estimator approaches the cut locus, this uniformity can fail and the convolution bound may collapse.","fun_headline_variants_meta":{"raw":{"variants":["Efficiency theory now covers Riemannian manifold parameters","Fréchet mean and single-index estimators achieve optimal variance","Hajek-Le Cam bounds hold for estimators on manifolds","Optimality proven for estimators on curved spaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00085,"raw_usage":{"total_tokens":3569,"prompt_tokens":816,"completion_tokens":2753,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2698}},"tokens_in":560,"tokens_out":2753,"duration_ms":18676,"temperature":1.0,"reasoning_tokens":2698,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T09:43:58.986967+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a manifold with a point whose sectional curvature grows without bound near the support of the estimator, simulate a regular estimator (e.g., Fréchet mean on a family of tori with shrinking injectivity radius) and check whether the limiting distribution of √n Exp^{-1}_{ψ(θ)} ψ̂n is a convolution with N(0, ψ̇G⁻¹ψ̇*) or whether the Gaussian component is altered; equivalently, verify numerically whether the equality in Lemma A.6 (∇Exp^{-1}_{μ} μ_h = id + (1/6)R_μ(h,·)h + O(‖h‖³)) holds uniformly in h along the path h/√n.","supporting_citations":[],"review_version":1}