{"id":"bdd1b108-04dd-4d05-b91d-0d85a4e09cab","arxiv_id":"1909.00410","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A CLT for Fréchet means on closed Riemannian manifolds follows from Bhattacharya-Lin's omnibus theorem once the cut locus is topologically stable and avoids the measure, and the stability condition is essential.","lead":"This paper proves a central limit theorem for Fréchet means on closed Riemannian manifolds, under the assumption that the cut locus of the mean carries no probability mass and that the cut locus is topologically stable. It also shows that a weaker cut-locus condition used in prior work is not sufficient in non-compact manifolds, via a counterexample on a flat cylinder.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem A is conditionally valid, but the abstract's 'CLT for closed Riemannian manifolds' overstates the unconditional reach: (A6) is neither derived nor characterized, and it is the real analytic bottleneck, not (A5).","rationale":"The reader's CONDITIONAL verdict is reasonable but assigns the risk to (A5). On closed manifolds, (B)+(C) appears to imply (A4) and (A5): pick an open measure-zero neighborhood U of Cut(qo); compactness and closedness of the cut-locus graph (Thm 3.7) give a uniform positive distance between a small ball around qo and Cut(p) for p in M∖U, hence uniform smoothness of the squared-distance Hessian and its modulus of continuity. So the A5 concern is less load-bearing than stated. The actual unresolved analytic hypothesis is (A6), the non-singularity of Λ, which is not a consequence of uniqueness, topological stability, or the cut-locus measure-zero condition; smeary examples on compact manifolds (cf. Remark 4.4) show it is a genuine hypothesis. The counterexample's assertion of uniqueness for every 0≤α≤1 is overstated; the proof establishes only small α, which is sufficient for the counterexample and should be corrected. These points support keeping a CONDITIONAL verdict, with the requested revision focused on (A6) as the remaining geometric bottleneck and on correcting the α-overstatement in Counterexample 4.3.","tokens_in":14284,"tokens_out":32348,"duration_ms":310583,"concrete_test":"Construct a closed-manifold probability measure satisfying (A1), (B), (C) and (A2b) but with singular Λ, e.g. a rotationally symmetric measure on S^2 supported on a lower-dimensional subsphere chosen so the Hessian at the mean has a zero eigenvalue; simulate the empirical Fréchet mean for increasing n and test whether √n convergence to a Gaussian holds. If the limit is non-normal or the scaling is not √n, (A6) is a real extra hypothesis and the paper should state the CLT as conditional on it. If every such symmetric configuration forces nonsingular Λ, the scope concern dissolves.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is Theorem A, obtained by verifying (A2b) in the Omnibus CLT. The proof of (B)+(C) ⇒ (A2b) (Lemma 4.2) is sound, and Corollary 3.8 gives (B) on closed manifolds. The load-bearing weakness is that Theorem A still inherits (A4)–(A6) as unexplained hypotheses; Section 4.2 states: 'It is not clear under which geometric assumptions conditions (A4)–(A6) hold.' This limits the advertised closed-manifold CLT. In fact, for closed M the reader's A5 worry can be resolved: (B)+(C) supplies an open U around Cut(qo) with µ(U)=0, and the closedness of the cut-locus graph (Thm 3.7) makes the singular set in B(qo,r)×(M∖U) empty with positive distance, so the Hessian of h(·,p) is uniformly bounded and Lipschitz on a small coordinate ball; (A4) and (A5) follow. The genuinely unaddressed condition is (A6), nonsingularity of Λ. (A6) is necessary for a √n-normal limit and can fail for smeary means on compact manifolds (Remark 4.4 gestures at this). Thus the theorem is a correct conditional corollary, but the abstract's 'CLT for closed Riemannian manifolds' overstates its unconditional reach until the geometric content of (A6) is supplied.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper examines the hypotheses of Bhattacharya and Lin's Omnibus Central Limit Theorem for Fréchet means in the Riemannian setting. The main result, Theorem A, states that if a complete Riemannian manifold has a unique Fréchet mean q_o, the cut locus is topologically stable, a neighborhood of Cut(q_o) carries zero probability, and the analytic conditions (A4)-(A6) hold, then the normalized Fréchet sample means converge in law to a Gaussian with covariance Λ^{-1}CΛ^{-1}. The paper proves that on closed manifolds the cut locus is always topologically stable (Corollary 3.8) and that topological stability plus a measure-zero cut-locus neighborhood implies the differentiability condition (A2b) of the OCLT (Lemma 4.2). It also gives a counterexample on the flat cylinder, intended to show that the measure-zero condition (C) alone does not imply (A2b) on noncompact manifolds.","tokens_in":14514,"tokens_out":14438,"duration_ms":133665,"significance":"If the counterexample is made fully rigorous, the paper provides a clean geometric sufficient condition for the key differentiability hypothesis of the OCLT and correctly identifies compactness as a sufficient condition for cut-locus stability. The proof of Lemma 4.2 is direct, uses no fitted parameters, and there is no circularity: Theorem A is obtained as a corollary of the external OCLT. The geometric material in Section 3, especially Theorem 3.7 and Corollary 3.8, is standard and appears sound. The main limitation is that the analytic hypotheses (A4)-(A6) are left as unexplained assumptions; the paper itself states in Section 4.2 that it is not clear under which geometric assumptions they hold. This limits the advertised 'CLT for closed Riemannian manifolds' to a conditional statement.","major_comments":[{"comment":"The theorem is conditional on (A4)-(A6), and Section 4.2 explicitly states: 'It is not clear under which geometric assumptions conditions (A4)–(A6) hold.' The abstract, however, advertises 'a Central Limit Theorem for closed Riemannian manifolds' without mentioning these uncharacterized analytic hypotheses. In particular, (A6), the nonsingularity of the Hessian matrix Λ, is necessary for a √n-normal limit and can fail for smeary means. As written, the paper's new geometric input replaces only (A2b), not the analytic bottleneck (A6). The authors should either prove or characterize (A4)-(A6) under the geometric hypotheses, at least for closed manifolds, or explicitly qualify the abstract and introduction so that the conditional nature of the CLT is not obscured.","section":"Section 4.2 and Theorem A"},{"comment":"The counterexample is load-bearing because it is the sole evidence that condition (C) alone does not imply (A2b), hence motivating the need for topological stability. As written, the proof is not rigorous. The text asserts that for every α∈[0,1] the measure µ_α has the unique Fréchet mean iy_ν, but the subsequent argument only establishes this for α below an explicit threshold. More seriously, the displayed inequality chain near the end appears to use the upper bound (4.6) for F̃_νε(w_α) as if it were a lower bound, and it is not shown that F^C_1(w_α) ≥ F^C_1(iy_ν). The counterexample should be rewritten with a complete verification: specify a concrete admissible α, prove uniqueness of the Fréchet mean for that α, and prove that for a positive µ_α-measure set of p the function h(·,p) is not C² on the chosen chart V. Without this, the necessity claim is not established.","section":"Counterexample 4.3"}],"minor_comments":[{"comment":"The abstract says 'closed Riemannian manifolds' while Theorem A is stated for a complete Riemannian manifold and uses closedness only through Corollary 3.8. Please make the roles of completeness and closedness consistent in the abstract, theorem statement, and introduction.","section":"Abstract and Theorem A"},{"comment":"The notation Cut(B(p,r)) is used in Definition 3.6 before being formally defined as the union of the cut loci of points in B(p,r). Please insert the definition explicitly at the point of first use.","section":"Definition 3.6"},{"comment":"The text 'Let M be a complete Riemannian manifold with Fréchet mean qo for the volume measure' should presumably read 'for the probability measure µ', since the paper is not about the volume measure. Please correct this and similarly clarify that 'Cut(U) has measure zero' refers to µ-measure.","section":"Section 4.2, (RA2b)"},{"comment":"The concluding sentence says 'we get for every v∈V that v↦h(v,p) = d²(φ^{-1}(v),p) is C² at v', but the quantifier over p is missing. The correct statement is: for µ-a.e. p∈M, the function v↦h(v,p) is C² at every v∈V.","section":"Lemma 4.2 proof"},{"comment":"The description of the neighborhood U(Cut(iy0)) involves conditions such as 'π − x < 1/|y| < π' and 'π + x < 1/|y| < π', which as written are hard to parse and do not obviously define an open neighborhood of the cut locus. Please rewrite these inequalities with clear absolute values and verify openness.","section":"Example 3.9"},{"comment":"Reference [11] lists the second author as 'Eltzner, Benjamin; Huckemann, Stephan F. Huckemann', which duplicates the surname. Please correct the author list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's sufficient-condition result is sound and likely publishable after revision: Corollary 3.8 and Lemma 4.2 give a clean geometric route to condition (A2b). The main risk is the counterexample, which is central to the claimed necessity of topological stability and currently contains incorrect or unsupported inequalities. I would not reject if the counterexample is repaired, but I would not accept it in the present form. The abstract should also be toned down to reflect that (A4)-(A6) remain uncharacterized analytic hypotheses."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the genuinely new and valuable part here is the geometry — topological stability of the cut locus on closed manifolds, and a counterexample showing that condition (C) from the Omnibus CLT does not imply (A2b) on non-compact manifolds. The CLT itself is a corollary of Bhattacharya–Lin, and its advertised reach is constrained by an uncharacterized analytic hypothesis.\n\nThe paper's main geometric results are clean. Definition 3.6, Theorem 3.7, and Corollary 3.8 are standard but carefully argued, and they fill a real gap in the literature. The flat cylinder counterexample is intricate and, as far as I can tell, does what it claims: it exhibits a measure satisfying (C) but not (A2b), so the Omnibus CLT cannot be applied with (C) alone. That is a genuine and useful clarification.\n\nWhere the paper finds its limit is in the CLT itself. Theorem A inherits conditions (A4)–(A6) from the Omnibus CLT, and the authors admit in Section 4.2 that it is not clear under which geometric assumptions those hold. I agree with the stress-test note that (A4) and (A5) can be recovered on closed manifolds from topological stability plus the measure-zero condition, but (A6) — non-singularity of the expected Hessian — is the real bottleneck. It can fail for smeary means, as the paper itself hints in Remark 4.4. So the abstract's phrase 'a Central Limit Theorem for closed Riemannian manifolds' is stronger than the theorem actually delivers; what is proven is a conditional CLT under an uncharacterized regularity condition. That should be stated much more clearly.\n\nA smaller issue: in Counterexample 4.3, the text says uniqueness of the Fréchet mean holds for every α ∈ [0,1], but the proof establishes it only for sufficiently small α. The authors should correct that overstatement. It does not damage the counterexample, which only needs small α.\n\nWho should read this: geometric statisticians and anyone working on Fréchet means. The cut-locus stability results are worth citing. The CLT is a useful clarification of the Omnibus CLT, but not a breakthrough. I would send it to peer review, with a request to fix the α-overstatement and soften the abstract. It deserves serious refereeing.","headline":"Genuinely new cut-locus stability results, but the CLT is only a conditional corollary and the abstract oversells its reach.","tokens_in":15159,"tokens_out":3603,"would_cite":true,"duration_ms":29421,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["53C20","60F05","62E20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves a central limit theorem for Fréchet means on closed Riemannian manifolds, using the topological stability of the cut locus to fill a gap in the Omnibus CLT.","keywords":["Fréchet mean","cut locus","central limit theorem","Riemannian manifold","topological stability","metric stability","intrinsic mean","Omnibus CLT"],"falsifier":"Look for a closed Riemannian manifold and a measure with unique Fréchet mean, $\\mu(\\mathrm{Cut}(q_o))=0$, and (A4) and (A6) satisfied, but with the Hessian of the squared distance failing the local $L^1$ smoothness condition (A5); if the empirical means are then not asymptotically Gaussian, Theorem A's hypothesis set is exactly right, while if they are still Gaussian, (A5) is stronger than needed.","tokens_in":14005,"feed_emoji":"📐","tokens_out":10947,"duration_ms":101053,"temperature":0.7,"pith_summary":"On a closed Riemannian manifold—compact and without boundary—the sample Fréchet mean, the point minimizing the empirical expected squared distance, obeys a $\\sqrt{n}$ central limit theorem with Gaussian limit $N(0,\\Lambda^{-1}C\\Lambda^{-1})$. The paper proves this by showing that a purely geometric condition, topological stability of the cut locus, is what turns \"the cut locus carries no probability mass\" into differentiability of the squared-distance function almost everywhere, the hypothesis the Omnibus CLT needs. Topological stability holds automatically when the manifold is closed, but the paper constructs a flat-cylinder example where it fails and where the weaker zero-mass condition is genuinely insufficient. The upshot is a geometric clarification of when intrinsic Fréchet-means asymptotics are Gaussian, rather than a new analytic estimate: the remaining analytic hypotheses (A4)-(A6) are inherited unchanged, and the paper states that it is not clear which geometric assumptions imply them.","feed_headline":"Stable cut locus yields CLT for Fréchet means on closed manifolds","feed_subtitle":"On compact spaces the stability condition is automatic; away from compactness it is exactly what can fail.","key_machinery":"The cut locus $\\mathrm{Cut}(p)$ is the set of points reached from $p$ by more than one shortest geodesic, or by a shortest geodesic through a conjugate point; it is exactly where the squared-distance function loses smoothness. Topological stability requires that for every open neighborhood of $\\mathrm{Cut}(p)$, the cut loci of all points sufficiently close to $p$ lie inside that neighborhood. The proof chain is: topological stability plus condition (C) gives condition (C')—$\\mu(\\mathrm{Cut}(B(q_o,r)))=0$ for some ball—and Lemma 4.2 turns (C') into (A2b), so the Omnibus CLT applies. Corollary 3.8 supplies topological stability on closed manifolds via a continuity theorem for tangent cut loci under $C^\\infty$ convergence of Riemannian metrics. The cylinder example for noncompact manifolds supplies the negative foil that isolates why stability is indispensable.","core_discovery":"The central claim is Theorem A: for a complete Riemannian manifold $(M,g)$ and a probability measure $\\mu$ with unique Fréchet mean $q_o$, if the cut locus of $q_o$ is topologically stable, some neighborhood of it has $\\mu$-measure zero, and conditions (A4)-(A6) hold, then every measurable selection $q_o^n$ of empirical Fréchet means satisfies $\\sqrt{n}(\\varphi(q_o^n)-\\varphi(q_o)) \\xrightarrow{d} N(0,\\Lambda^{-1}C\\Lambda^{-1})$, where $\\varphi=\\exp_{q_o}^{-1}$ is a normal coordinate chart. The paper's contribution is to identify what is needed for condition (A2b), almost-everywhere twice differentiability of $v\\mapsto d^2(\\varphi^{-1}(v),p)$: topological stability of the cut locus plus the zero-mass neighborhood (C) yields the stronger condition (C'), that the cut locus of a whole ball around $q_o$ has measure zero, and Lemma 4.2 shows (C') implies (A2b). It then proves, as Corollary 3.8, that every closed manifold has a topologically stable cut locus, by showing that the tangent cut locus is closed under limits when metrics converge in $C^\\infty$. A counterexample on the flat cylinder shows that condition (C) alone does not imply (A2b) on noncompact manifolds, so the topological-stability hypothesis is doing real work.","pith_inferences":["Editorial inference: if condition (A5) ever fails on a closed manifold satisfying the measure and cut-locus hypotheses, Theorem A would not apply, so the paper's theorem is best read as a conditional reduction of the statistical problem to a regularity question about squared distance.","Editorial inference: the flat-cylinder construction with power-law mass near the cut locus suggests a boundary regime in which the $n^{1/2}$ Gaussian rate gives way to a slower, smeary limit; the paper's Remark 4.4 points in this direction without developing it.","Editorial inference: because topological stability is a property of the manifold alone, it is checkable a priori, unlike (A4)-(A6), which depend on the unknown mean and measure; identifying curvature or topological conditions that imply it on noncompact manifolds is the natural next step."],"forward_implications":["On every closed Riemannian manifold, sample Fréchet means are asymptotically Gaussian whenever the mean is unique, its cut locus has a zero-probability neighborhood, and (A4)-(A6) hold.","For noncompact manifolds, the zero-mass condition alone is not enough; a practitioner must check topological stability of the cut locus, and the flat cylinder shows the failure is not merely technical.","Combining topological stability and condition (C) into condition (C') gives a sufficient route to (A2b), so the Omnibus CLT becomes available exactly when cut-locus stability can be geometrically established.","The continuity theorem for the tangent cut locus stands on its own as a geometric convergence statement for cut loci under perturbation of the Riemannian metric."],"supporting_citations":[{"why":"Supplies the Omnibus CLT whose hypotheses Theorem A verifies; without it the main theorem has no engine.","marker":"[6]"},{"why":"Provides the classical cut-locus and distance-function facts, including smoothness off the cut locus and the Hessian formula, that Lemma 4.2 uses.","marker":"[26]"},{"why":"Establishes existence and strong consistency of Fréchet means on Riemannian manifolds, the baseline setting this paper refines.","marker":"[7]"},{"why":"Gives the earlier intrinsic Fréchet-mean CLT that Theorem A improves, positioning the paper's result against it.","marker":"[8]"},{"why":"Supplies the strong law of large numbers for empirical Fréchet means, used to verify condition (A3).","marker":"[29]"},{"why":"Shows the cut locus of a Fréchet mean has measure zero under mild assumptions, providing the measure-theoretic backdrop for condition (C).","marker":"[23]"},{"why":"Used to recall that cut loci are closed subsets, which is the substrate for the stability definitions.","marker":"[10]"}],"fun_headline_variants":["Cut locus stability: the key to Fréchet CLT on manifolds","Closed manifolds: cut locus stability always holds, so CLT follows","Noncompact manifolds: cut locus stability is the missing hypothesis for CLT","When the cut locus gets wild, Fréchet CLT on manifolds breaks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, at the true mean, the squared-distance function is smooth enough in the sense of conditions (A4)-(A6), especially a locally uniform $L^1$-smoothness of its Hessian; the paper does not say which manifolds and measures guarantee these, only that the cut-locus hypotheses do not cover them.","fun_headline_variants_meta":{"raw":{"variants":["Cut locus stability: the key to Fréchet CLT on manifolds","Closed manifolds: cut locus stability always holds, so CLT follows","Noncompact manifolds: cut locus stability is the missing hypothesis for CLT","When the cut locus gets wild, Fréchet CLT on manifolds breaks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000944,"raw_usage":{"total_tokens":4017,"prompt_tokens":916,"completion_tokens":3101,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":3018}},"tokens_in":532,"tokens_out":3101,"duration_ms":21672,"temperature":1.0,"reasoning_tokens":3018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:53:44.280807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Look for a closed Riemannian manifold and a measure with unique Fréchet mean, $\\mu(\\mathrm{Cut}(q_o))=0$, and (A4) and (A6) satisfied, but with the Hessian of the squared distance failing the local $L^1$ smoothness condition (A5); if the empirical means are then not asymptotically Gaussian, Theorem A's hypothesis set is exactly right, while if they are still Gaussian, (A5) is stronger than needed.","supporting_citations":[{"cited_title":"Omnibus CLTs for Fr´ echet means and nonparametric inference on non-Euclidean spaces","cited_arxiv_id":null,"evidence_quote":"Supplies the Omnibus CLT whose hypotheses Theorem A verifies; without it the main theorem has no engine."},{"cited_title":"Riemannian geometry","cited_arxiv_id":null,"evidence_quote":"Provides the classical cut-locus and distance-function facts, including smoothness off the cut locus and the Hessian formula, that Lemma 4.2 uses."},{"cited_title":"Large sample the ory of intrinsic and extrinsic sample means on manifolds","cited_arxiv_id":null,"evidence_quote":"Establishes existence and strong consistency of Fréchet means on Riemannian manifolds, the baseline setting this paper refines."},{"cited_title":"Large sample the ory of intrinsic and extrinsic sample means on manifolds","cited_arxiv_id":null,"evidence_quote":"Gives the earlier intrinsic Fréchet-mean CLT that Theorem A improves, positioning the paper's result against it."},{"cited_title":"On expected figures and a strong law of l arge numbers for random elements in quasi-metric spaces","cited_arxiv_id":null,"evidence_quote":"Supplies the strong law of large numbers for empirical Fréchet means, used to verify condition (A3)."},{"cited_title":"On the measure of the cut loc us of a Fr´ echet mean","cited_arxiv_id":null,"evidence_quote":"Shows the cut locus of a Fréchet mean has measure zero under mild assumptions, providing the measure-theoretic backdrop for condition (C)."},{"cited_title":"Riemannian geometry","cited_arxiv_id":null,"evidence_quote":"Used to recall that cut loci are closed subsets, which is the substrate for the stability definitions."}],"review_version":1}