{"id":"4c96a41c-1f88-4f0f-9b6d-4f0e4c2882dd","arxiv_id":"2502.01995","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Fréchet regression in CAT(K) spaces is claimed to achieve Euclidean convergence rates, but the proofs contain unmatched tail exponents and an unsupported empirical-process bound, so the central rate claim fails as written.","lead":"This paper claims to give a unified mathematical analysis of Fréchet regression in curved metric spaces called CAT(K) spaces, including existence, uniqueness, concentration rates, and angle stability. The main rate theorem is not supported by the paper's own proof, and the experiments do not actually use a hyperbolic distance, so the headline guarantees should not be taken at face value.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.11's proof has a variance-scaling inconsistency: it derives E[d²] ≤ E[Δ]/α = O((nh^d)^{-1/2}), not O(1/(nh^d)), and Proposition 3.8 integrates an ε^4 tail as though it were ε^2.","rationale":"I read the paper in good faith and attempted to see whether Theorem 3.11 could be salvaged despite the defects noted by the reader. The existence and uniqueness results are standard, and the bias term in Theorem 3.11 is plausibly controlled by β-Hölder continuity. However, the variance control is not a matter of a missing citation or a minor typo: the proof's own inequality chain produces a variance of order (n h_n^d)^{-1/2} in squared error, whereas the theorem claims 1/(n h_n^d). This is an internal inconsistency, not merely a disagreement with the existing literature. The reader's weakest assumption correctly identifies the unproved uniform empirical-process bound in Theorem 3.7 as a gap; I partially agree, but I think the more load-bearing defect is that even granting a valid concentration inequality, the argument in Theorem 3.11's proof does not deliver its stated rate. Proposition 3.8's misuse of the ε^4 tail is a concrete example of the same class of error. I therefore support the reader's REJECT verdict, with no change in recommendation.","tokens_in":33181,"tokens_out":5742,"duration_ms":55454,"concrete_test":"Perform the following closed-form check on the proof of Theorem 3.11. Start from the paper's own inequality α(K,D) d²(μ̂*_n(x), μ̃*_n(x)) ≤ Δ_n(x) and take expectations. Insert the proof's asserted bound E[Δ_n(x)] = O((n h_n^d)^{-1/2}). The result is E[d²] = O((n h_n^d)^{-1/2}), which contradicts the theorem's claimed O(1/(n h_n^d)) variance term. Separately, integrate the Theorem 3.7 tail P[d(μ̂_n, μ) > ε] ≤ c_1 exp(-c_2 n ε^4) using E[d²] = ∫₀^∞ 2ε P[d > ε] dε; this gives n^{-1/2}, not n^{-1}. If both computations reproduce the mismatch, the proof of the headline rate fails as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is Theorem 3.11, the rate O(1/(n h_n^d) + h_n^{2β}) for nonparametric Fréchet regression in CAT(K) spaces. The proof of that theorem is internally inconsistent. The argument uses strong geodesic convexity to obtain α(K,D) d²(μ̂*_n(x), μ̃*_n(x)) ≤ Δ_n(x), and then takes expectations to conclude E[d²(μ̂*_n(x), μ̃*_n(x))] ≤ E[Δ_n(x)]/α(K,D). The proof then asserts E[Δ_n(x)] = O((n h_n^d)^{-1/2}), which yields a variance component O((n h_n^d)^{-1/2}), not the O(1/(n h_n^d)) claimed in (10). This is not a missing constant; it is a missing factor of (n h_n^d)^{1/2}, so the stated squared-error rate does not follow from the proof. A second, linked inconsistency appears in Proposition 3.8: it says it integrates the exponential tail of Theorem 3.7, but Theorem 3.7 has tail exponent ε^4, because the exponent is n(α(K,D) ε²)²/(8D²). Integrating exp(-c n ε^4) against p ε^{p-1} yields E[d^p] = O(n^{-p/4}), not O(n^{-p/2}) as Proposition 3.8 claims. For p=2, this is n^{-1/2}, not n^{-1}. Thus even if Theorem 3.7's uniform bound were made rigorous, the downstream moment and variance estimates would remain wrong. The proof of Theorem 3.7 is itself not closed: the uniform sup bound is asserted from 'standard references' with no citation, the net radius δ is never defined, and no Lipschitz or chaining argument connects a δ-net bound to sup over M. But the decisive issue is that the central rate theorem's own variance analysis contradicts its stated conclusion.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies nonparametric Fréchet regression for metric-space-valued responses, focusing on complete CAT(K) spaces. It claims existence, uniqueness, and stability results for Fréchet means; an exponential concentration inequality for sample Fréchet means; L^p moment bounds; pointwise consistency of kernel-based Fréchet regression; and a sup-norm mean-squared-error rate O(1/(n h_n^d) + h_n^{2β}) for β-Hölder regression functions. The theoretical statements are in Section 3 with proofs in Appendix B. The empirical section compares Fréchet regression on spherical versus hyperbolic representations of spherical data and reports lower MSE for the hyperbolic mapping.","tokens_in":33668,"tokens_out":6165,"duration_ms":61229,"significance":"If the main results were valid, the paper would provide a useful unification: classical Euclidean nonparametric rates would transfer to bounded-curvature metric spaces, with constants depending only on K and the diameter, and the angle-stability and jet-expansion results would add geometric insight. The paper is clearly organized and makes concrete, falsifiable claims, and it includes Python code and dataset details for the experiments. However, the central proof chain has internal inconsistencies: the main concentration bound is asserted with undefined quantities and no derivation, the L^p proposition integrates a tail with the wrong exponent, and the variance term in the central rate theorem is computed at the square-root scale but reported at the linear scale. As a result, the paper's headline convergence-rate result is not established by the supplied arguments.","major_comments":[{"comment":"The core of the proof is an unstated uniform empirical-process bound. After a fixed-z Hoeffding inequality, the proof asserts P[sup_{z in M} |F_n(z)-F(z)| >= t] <= c'_1 exp(-c'_2 n t^2) with c'_1 = 2(alpha(K,D)D/delta)^m and c'_2 = alpha(K,D)/(8D^2), attributed to \"standard references in manifold-valued statistics\" without a citation. The net radius delta is never defined, the dimension m is not specified for a general CAT(K) space, and no chaining or Lipschitz argument connects a delta-net bound to the supremum over M. Since all downstream consistency and rate results depend on this bound, Theorem 3.7 is not proven as stated.","section":"B.2, Theorem 3.7"},{"comment":"The tail exponent in Theorem 3.7 is exp(-n(alpha(K,D) epsilon^2)^2/(8D^2)) = exp(-O(n epsilon^4)), but the proof of Proposition 3.8 integrates exp(-c_2 n epsilon^2) and concludes E[d^p(muhat_n, mu)] = O(n^{-p/2}). Integrating the epsilon^4 tail gives O(n^{-p/4}), so for p=2 it gives O(n^{-1/2}), not O(n^{-1}). Proposition 3.8's claimed L^p rate is therefore not a consequence of Theorem 3.7; the proof of Proposition 3.8 restates a tail bound that contradicts the theorem it cites.","section":"B.2, Theorem 3.7 vs. Proposition 3.8"},{"comment":"The proof derives E[d^2(muhat*_n(x), mutilde*_n(x))] <= E[Delta_n(x)]/alpha(K,D), and then explicitly states E[Delta_n(x)] = O((n h_n^d)^{-1/2}). The proof concludes a variance component of size O((n h_n^d)^{-1/2}), while the theorem's displayed rate (10) claims O(1/(n h_n^d)). No argument is supplied to convert the square-root bound into the claimed linear-in-1/(n h_n^d) bound. This is a missing factor of (n h_n^d)^{1/2}, not a harmless constant mismatch, so the stated squared-error rate does not follow from the proof.","section":"B.2, Theorem 3.11"},{"comment":"The proof also relies on the assertion that a straightforward Hoeffding/Bennett-type argument gives E[Delta_n(x)] = O((n h_n^d)^{-1/2}) even though muhat*_n(x) appears inside the empirical process term and depends on the whole sample. No Efron-Stein, bounded-differences, or U-statistic calculation is given. Even if the preceding concentration theorem were correct and the tail-exponent mismatch were fixed, this step would still need a rigorous derivation before the bias-variance decomposition in Theorem 3.11 is established.","section":"B.2, Theorem 3.11"}],"minor_comments":[{"comment":"Equation (6) writes E[f(x) | X = x], but f is a function on M, so the expression should be E[f(Y) | X = x] or an equivalent conditional expectation with respect to the response variable.","section":"Section 3.2, Assumption 3.9"},{"comment":"In the displayed definition of Delta_n(x), the same term d^2(Y_i, muhat*_n(x)) appears in both summands; the second summand should involve d^2(Y_i, mutilde*_n(x)) or otherwise the expression does not match the two terms in the preceding display.","section":"B.2, Theorem 3.11"},{"comment":"The proof expands {d(p, m_n) - d(y, p)}^2 as d(p, m_n)^2 - 2d(p, m_n) + d^2(y, p); the linear term should be -2d(p, m_n)d(y, p). The displayed algebra and the subsequent bound need correction.","section":"B.1, Lemma 3.2"},{"comment":"The manuscript alternates between symbols such as muhat in Theorem 3.7 and muhat_n elsewhere, and between mu^*, mu^*(x), and mu*_n(x). Standardizing the notation for the sample Fréchet mean and the local population mean would improve readability.","section":"All sections"}],"recommendation":"reject","confidential_remarks":"The central rate theorem is not supported by the supplied proof, and the mismatch is not a localized or typographical issue: the variance term is derived at the square-root scale while the theorem claims the standard 1/(n h_n^d) scale, and the concentration tail used downstream has an inconsistent exponent. The authors would need to supply a genuine empirical-process bound for the supremum over M, correct the L^p integration, and rework the variance argument before the main claims could be evaluated. I recommend rejection rather than major revision, though a substantially revised version with complete proofs could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the headline results—the exponential concentration bound for Fréchet means in CAT(K) spaces and the nonparametric rate O(1/(nh^d) + h^{2β})—are not established. The proof of the rate theorem derives a variance term of order (nh^d)^{-1/2} and then claims O(1/(nh^d)); that's a missing factor, not a typo. The concentration theorem has tail exponent ε^4, while Proposition 3.8 integrates a tail with exponent ε^2, so the Lp bound is off by the same factor. Neither theorem is closed.\n\nWhat's actually new: the angle-stability section (3.3) is the most original part. The idea that Alexandrov angles at the conditional Fréchet mean vary Lipschitzly with the predictor under Wasserstein perturbations is worth pursuing. The local jet expansion is standard Riemannian Taylor material. The existence/uniqueness results in Section 3.1 are mostly restatements of known facts in CAT(K)/Hadamard spaces, and the paper cites the relevant literature (Petersen & Müller, Hein, Steinke & Hein). It is honest that proofs are sketches, but they are sketches of results that are either known or unproven.\n\nThe soft spots are proportional to how central they are. Theorem 3.7's uniform empirical-process bound is asserted via \"standard references\" with no citation, and δ is never defined. The bias–variance decomposition in Theorem 3.11 uses a Δ_n(x) whose expectation is claimed to be O((nh^d)^{-1/2}) without proof, and even granting that, the final rate does not follow. The experiments are too simple to validate curvature claims; stereographic projection to a \"hyperbolic\" plane is not a test of the theoretical results.\n\nWho is this for? A reader new to Fréchet regression might get a useful map of the CAT(K) landscape, but anyone needing reliable rates or concentration should look elsewhere. This paper does not deserve a serious referee; the main theorems have load-bearing gaps and internal contradictions. I'd send it back with a clear explanation rather than spend a full review cycle on it.","headline":"The paper's central rate theorem is not proven—its own proof derives a slower variance term—and the concentration bound is an unclosed sketch; the existence/uniqueness part is standard.","tokens_in":34168,"tokens_out":4503,"would_cite":false,"duration_ms":40927,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62R30","53C23","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Nonparametric Fréchet regression on bounded-diameter CAT(K) spaces is claimed to attain Euclidean-type convergence rates, with exponential concentration of sample Fréchet means as the supporting mechanism.","keywords":["Fréchet regression","CAT(K) spaces","comparison geometry","nonparametric regression","concentration inequalities","curvature bounds","manifold-valued data","kernel smoothing"],"falsifier":"A Monte Carlo check on a low-dimensional sphere with known $K>0$ and diameter $D < \\pi/(2\\sqrt{K})$ could settle the concentration claim: compare the observed tail $P(d(\\hat{\\mu}_n, \\mu) > \\epsilon)$ with the bound in Theorem 3.7. If the tail decays like $\\exp(-c n \\epsilon^2)$ rather than $\\exp(-c n \\epsilon^4)$, or if the prefactor $(\\alpha(K,D) D/\\delta)^m$ cannot be made finite with any explicit covering radius $\\delta$, then the exponential concentration step—and the rate theorem that integrates it—fails as stated.","tokens_in":32990,"feed_emoji":"📐","tokens_out":9839,"duration_ms":90819,"temperature":0.7,"pith_summary":"This paper aims to show that nonparametric Fréchet regression—regression when the response lies in a metric space—keeps the classical Euclidean rates of kernel smoothing whenever the space satisfies the CAT(K) curvature condition. The central result states that for a complete CAT(K) space of diameter at most $D$, a $\\beta$-Hölder regression function, and standard kernel weights, the mean squared error obeys $\\sup_{x \\in X_0} \\mathbb{E}[d^2(\\hat{\\mu}^*_n(x), \\mu^*(x))] = O(1/(n h_n^d) + h_n^{2\\beta})$, the usual bias-variance trade-off. The paper also claims an exponential concentration inequality for sample Fréchet means, $L^p$ convergence at rate $n^{-p/2}$, and angle-stability results linking curvature to the local directional geometry of the estimator. If these claims hold, curvature bounds—not manifold smoothness—are what govern the statistical difficulty, with non-positive curvature requiring no diameter restriction.","feed_headline":"Fréchet regression matches Euclidean rates on curved spaces","feed_subtitle":"Kernel smoothing on CAT(K) spaces concentrates exponentially and keeps the classical bias–variance trade-off, the paper argues.","key_machinery":"The load-bearing object is the strong geodesic convexity constant $\\alpha(K,D)$ of the squared distance function: in a CAT(K) space of diameter $D$, the Fréchet functional satisfies $F(z) - F(\\mu) \\geq \\alpha(K,D) d^2(z,\\mu)$, with $\\alpha = 1/2$ for $K \\leq 0$ and a positive curvature-dependent constant for $K > 0$ when $D < \\pi/(2\\sqrt{K})$. This inequality turns statistical error into a functional gap, which is then controlled by a uniform empirical-process tail bound for $\\sup_{z \\in M} |F_n(z) - F(z)|$. The secondary machinery is comparison-geometry angle technology: comparison triangles in the constant-curvature model space and Alexandrov angles, which supply the angle-stability lemmas and the local jet expansion of the Fréchet functional.","core_discovery":"The paper's discovery is a set of comparison-geometry guarantees for Fréchet regression in CAT(K) spaces. It proves existence and uniqueness of conditional Fréchet means under diameter constraints, gives an exponential concentration bound for the sample Fréchet mean with constant $\\alpha(K,D)$ coming from strong geodesic convexity, and establishes pointwise almost-sure consistency for kernel-weighted estimators. The rate theorem, stated as Theorem 3.11, asserts that in a complete CAT(K) space of diameter at most $D$, with a $\\beta$-Hölder regression function and standard kernel weights, $\\sup_{x \\in X_0} \\mathbb{E}[d^2(\\hat{\\mu}^*_n(x), \\mu^*(x))] = O(1/(n h_n^d) + h_n^{2\\beta})$, matching Euclidean nonparametric rates. The proof runs through a bias-variance decomposition with a local population measure and uses strong convexity of the Fréchet functional to convert distance error into functional gap; a separate set of Alexandrov-angle lemmas shows that angles at the conditional Fréchet mean vary Lipschitzly with Wasserstein perturbations of the conditional measures.","pith_inferences":["A testable extension the paper leaves implicit: stereographic projection into hyperbolic space acts as a variance-stabilizing transform for spherical responses, analogous to the log transform, and one could formally characterize when it reduces Fréchet regression risk under heteroscedastic angular noise.","Because the rate proof relies only on the weight LLN condition and strong geodesic convexity, similar comparison-geometry arguments should transfer to local-polynomial or random-forest-weighted Fréchet regression.","The appendix's $\\epsilon$-approximate CAT(K) results suggest the whole theory can be made robust: with comparison error $\\epsilon$, convexity degrades by $O(\\epsilon D)$ and uniqueness weakens to a $O(\\sqrt{\\epsilon})$ neighborhood, extending the guarantees to nearly-CAT data."],"forward_implications":["Bandwidth selection for Fréchet regression can follow the Euclidean recipe: balance $1/(n h_n^d)$ against $h_n^{2\\beta}$.","Negative-curvature CAT(K) spaces give convexity constant $1/2$ with no diameter restriction, so Hadamard-type spaces are the friendliest setting for the theory.","Positive-curvature spaces need support diameter below $\\pi/(2\\sqrt{K})$; beyond that, uniqueness and the rates can break down.","The exponential concentration and $L^p$ rate $n^{-p/2}$ imply sample Fréchet means are as efficient as Euclidean averages in the bounded-diameter regime.","Angle stability means directional features around the conditional Fréchet mean inherit Lipschitz continuity in the predictor, which matters for shape and directional statistics."],"supporting_citations":[{"why":"Defines Fréchet regression for random objects and the estimator whose rates this paper extends to CAT(K) spaces.","marker":"Petersen & Müller (2019)"},{"why":"Supplies the CAT(K) space definitions, comparison geometry, and geodesic convexity facts used throughout the proofs.","marker":"Bridson & Haefliger (2013)"},{"why":"Provides the nonpositive-curvature theory behind the K≤0 existence and uniqueness lemmas.","marker":"Ballmann (1995)"},{"why":"Introduces the kernel regression weights whose law-of-large-numbers condition Assumption 3.9 is patterned on.","marker":"Nadaraya (1964)"},{"why":"Introduces the smoothing regression estimator used as the template for the nonparametric Fréchet estimator.","marker":"Watson (1964)"},{"why":"Gives empirical-process and rate results for nonparametric regression with metric-space-valued output that Theorem 3.11 builds on.","marker":"Hein (2009)"},{"why":"Provides uniform convergence results for local Fréchet regression, the closest prior rate theorem this work generalizes.","marker":"Chen & Müller (2022)"},{"why":"Supplies the computational Fréchet regression implementation with kernel weights used in the experiments.","marker":"Davis et al. (2010)"}],"fun_headline_variants":["Fréchet regression attains Euclidean rates on CAT(K) spaces","Curved-space regression matches flat-space speed","Exponential concentration for Fréchet means on manifolds","Comparison geometry delivers sharp Fréchet regression rates"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the empirical Fréchet functional converges uniformly over the whole space with a specific exponential tail; the paper asserts this bound, with constants depending on an undefined net radius, rather than proving it.","fun_headline_variants_meta":{"raw":{"variants":["Fréchet regression attains Euclidean rates on CAT(K) spaces","Curved-space regression matches flat-space speed","Exponential concentration for Fréchet means on manifolds","Comparison geometry delivers sharp Fréchet regression rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002231,"raw_usage":{"total_tokens":8619,"prompt_tokens":923,"completion_tokens":7696,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":7644}},"tokens_in":539,"tokens_out":7696,"duration_ms":53468,"temperature":1.0,"reasoning_tokens":7644,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:44:32.756702+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A Monte Carlo check on a low-dimensional sphere with known $K>0$ and diameter $D < \\pi/(2\\sqrt{K})$ could settle the concentration claim: compare the observed tail $P(d(\\hat{\\mu}_n, \\mu) > \\epsilon)$ with the bound in Theorem 3.7. If the tail decays like $\\exp(-c n \\epsilon^2)$ rather than $\\exp(-c n \\epsilon^4)$, or if the prefactor $(\\alpha(K,D) D/\\delta)^m$ cannot be made finite with any explicit covering radius $\\delta$, then the exponential concentration step—and the rate theorem that integrates it—fails as stated.","supporting_citations":[{"cited_title":"and M \\\"u ller, H.-G","cited_arxiv_id":null,"evidence_quote":"Defines Fréchet regression for random objects and the estimator whose rates this paper extends to CAT(K) spaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CAT(K) space definitions, comparison geometry, and geodesic convexity facts used throughout the proofs."},{"cited_title":"Lectures on spaces of nonpositive curvature, volume 25","cited_arxiv_id":null,"evidence_quote":"Provides the nonpositive-curvature theory behind the K≤0 existence and uniqueness lemmas."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the smoothing regression estimator used as the template for the nonparametric Fréchet estimator."},{"cited_title":"Robust nonparametric regression with metric-space valued output","cited_arxiv_id":null,"evidence_quote":"Gives empirical-process and rate results for nonparametric regression with metric-space-valued output that Theorem 3.11 builds on."},{"cited_title":"and M \\\"u ller, H.-G","cited_arxiv_id":null,"evidence_quote":"Provides uniform convergence results for local Fréchet regression, the closest prior rate theorem this work generalizes."},{"cited_title":"C., Fletcher, P","cited_arxiv_id":null,"evidence_quote":"Supplies the computational Fréchet regression implementation with kernel weights used in the experiments."}],"review_version":1}