{"id":"00f79945-d262-40c7-bf73-ba7c2a7dc4c3","arxiv_id":"2602.20896","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new omnibus uniformity test on S^{p-1} is derived from a Stein characterization, shown to be a Sobolev test, with closed-form null and alternative asymptotic distributions.","lead":"The paper introduces a new test for uniformity on the hypersphere based on a Laplace–Beltrami Stein characterization with exponential test functions. It derives explicit asymptotic distributions and shows the test can be tuned to gain power against specific alternatives.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Consistency theorem only proved for L² densities; Proposition 1.1's statement exceeds its proof, so omnibus consistency for rough absolutely continuous alternatives is unsupported.","rationale":"The reader's weakest assumption correctly identified the q∈L² premise as load-bearing. I agree that the proof as written does not support consistency for non-L² absolutely continuous alternatives, and Remark 3.1 overclaims 'all absolutely continuous non-uniform distributions' given that Theorem 3.3 assumes q∈L². However, the underlying characterization appears to be true for all probability measures via a different argument (the addition theorem and positivity of the Gegenbauer coefficients), so this is a proof gap rather than a demonstrated falsehood. The second 'fragile premise' flagged by the reader—sufficiency of the exponential class for a single λ—is actually robust: z=0 implies the spherical mgf is constant, and the harmonic coefficients of that constant force all nonconstant moments of P to vanish. The oracle-tuning issue in the numerical section is a separate weakness of the empirical comparison, not of the central theoretical claim; I do not base the verdict on it. Since the reader already returned CONDITIONAL with moderate confidence, and my concern reinforces the need for either an extension of the proof or a qualified statement of scope, the verdict remains CONDITIONAL (no change).","tokens_in":29378,"tokens_out":14984,"duration_ms":146082,"concrete_test":"Independently derive Proposition 1.1 for arbitrary probability measures: (i) show z(s)=0 for all s∈S^{p−1} iff M_X(λs) is constant on S^{p−1}; (ii) use the addition theorem C_k^{(p−2)/2}(s·x)=γ_{k,p}Σ_{r=1}^{d_{k,p}}Y_{r,k}(s)Y_{r,k}(x) and the positivity of m_{k,p}(λ) to conclude ∫Y_{r,k}dP=0 for all k≥1; (iii) apply Stone–Weierstrass to conclude P is uniform. If step (ii) fails for some k or for some non-L² density, restrict Theorem 3.3 and Remark 3.1 to q∈L²; if it succeeds, the proof gap is cosmetic and the consistency claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central consistency claim rests on Proposition 1.1 and Theorem 3.3. Proposition 1.1 is stated for any absolutely continuous random vector, but its proof in Appendix A begins by assuming q∈L²(S^{p−1}) and uses the L² spherical-harmonic expansion (10). Lemma 3.1 and Theorem 3.3 carry the same q∈L² assumption. Many absolutely continuous non-uniform laws on S^{p−1} are integrable but not square-integrable (e.g., q(x)∝||x−e_1||^{−β} with (p−1)/2 < β < p−1); the stated theory does not cover these. If the characterization in (2) failed for such a q, the test would not be consistent against it. The gap is real but likely repairable: z(s)=∆_{S}M_X(λs)=0 implies M_X(λs) is constant on S^{p−1}, and the addition theorem together with strict positivity of m_{k,p}(λ) forces all nonconstant harmonic moments of P to vanish, a Stone–Weierstrass argument that does not require q∈L². The paper does not give this argument, so the stated omnibus consistency claim is formally unsupported for non-L² densities.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new test of uniformity on the hypersphere S^{p-1} based on the Laplace–Beltrami Stein operator applied to the exponential test functions e^{λt·x}. The statistic is T_n(λ)=||n^{-1/2}∑_j Δ_{S^{p-1}}e^{λt·X_j}||^2_{L^2(S^{p-1})}. The authors derive an explicit Gegenbauer/spherical-harmonic expansion of T_n(λ), show that it belongs to the class of Sobolev tests, and obtain: the null limit as a weighted series of independent chi-square variables (Theorem 3.2), an almost-sure limit under fixed alternatives (Theorem 3.3), and asymptotic normality of the centered statistic (Theorem 3.5). They also analyze the limits λ→0 and λ→∞, connect the construction to the dKSD test, propose a data-driven tuning procedure, and report extensive Monte Carlo power comparisons.","tokens_in":29699,"tokens_out":14752,"duration_ms":146069,"significance":"If the results hold, the paper is a valuable contribution to directional statistics. It adds a new member to the Sobolev class of uniformity tests with explicit, easily computable coefficients; closed-form null and alternative asymptotic distributions; a tuning parameter that interpolates between Rayleigh-type and extreme-angle-type behavior; and a clean spectral relation to the dKSD test. The derivations are careful and internally consistent: the coefficient identities from Bessel integrals, the Funk–Hecke steps, the Karhunen–Loève expansion, and the variance formulas all check out. The main caveats concern the scope of the consistency theorem (square-integrable densities only) and the implementability of the adaptive tuning procedure; both are repairable but currently weaken the strength of some claims.","major_comments":[{"comment":"The consistency claim is formally proved only for densities q∈L^2(S^{p-1}), while Proposition 1.1 and Remark 3.1 assert characterization and consistency for all absolutely continuous distributions. The proof of Proposition 1.1 begins by assuming q∈L^2 and uses the L^2 harmonic expansion (10); Lemma 3.1 and Theorem 3.3 carry the same q∈L^2 assumption. Many absolutely continuous non-uniform laws on S^{p-1} have densities that are integrable but not square-integrable (e.g., q(x)∝||x−e_1||^{−β} for (p−1)/2<β<p−1). Thus Remark 3.1's statement that Theorem 3.3 implies consistency against all absolutely continuous non-uniform distributions is unsupported as written. The gap is likely repairable: z(s)=Δ_S M_X(λs)=0 implies M_X(λs) is constant on S^{p-1}, and strict positivity of m_{k,p}(λ) together with the addition theorem should force all nonconstant harmonic moments of P to vanish, without re","section":"Proposition 1.1, Theorem 3.3, Remark 3.1, Appendix A"},{"comment":"The adaptive tuning procedure is evaluated largely with an oracle version. The statistic T_n(λ̃) defined in (20) uses 10,000 independent draws from the candidate alternative density to estimate E[A_k]; this information is unavailable in practice. The paper mentions cross-validation as the practical alternative, but the cross-validated statistic T_{n,20}(λ) reported in Table 1 is not covered by the theoretical null distribution in Theorem 3.2, and no size calibration or asymptotic theory is given for it. Consequently, the headline power comparisons in Tables 2–4 for T_n(λ̃) are oracle-based and do not by themselves support the abstract's claim of a data-driven strategy. The authors should either (i) present the cross-validated procedure as the main tuned test with a valid null calibration, or (ii) clearly label T_n(λ̃) as an oracle benchmark and present the implementable variant separatel","section":"Sections 5.2–5.3, Tables 1–4"}],"minor_comments":[{"comment":"The statement contains a duplicated 'S^{p-1} S^{p-1}': 'random vector X on S^{p-1} S^{p-1} with density q'.","section":"Theorem 3.5"},{"comment":"In the displayed covariance kernel, the second term is written as '−Δ_{S^{p-1}}M_X(λs)Δ_{S^{p-1}}M_X(λt)'; this should be z(s)z(t) with the operator acting appropriately on s and t. The current notation is ambiguous.","section":"Theorem 3.4"},{"comment":"There is a typo: 'explicitly derive the coefficients βk in closed form for for the von Mises–Fisher' — delete the duplicated 'for'.","section":"Example 3.1"},{"comment":"The grid for λ̃ is {i/10 : i=1,...,300}, but Figure 5 plots λ up to 20. It would help to state the grid or range used in Figure 5.","section":"Section 5.2"},{"comment":"The representation of general Sobolev tests as L^2-Stein tests assumes the coefficients b_{k,p} are such that the defining series for f_t converges in L^2(S^{p-1}); this integrability condition is not stated.","section":"Section 4.2"}],"recommendation":"major_revision","confidential_remarks":"The main mathematical structure is sound and the paper is likely publishable after revision. The L^2-density gap is not a sign of a wrong result; a measure-theoretic argument using strict positivity of the Gegenbauer coefficients should close it. The more delicate issue is the oracle-based tuning evaluation: the authors need to clearly separate the fixed-λ test (fully supported by theory) from the adaptive procedure (currently supported only by simulations with oracle tuning). I would be willing to accept a revised version that either closes these gaps or weakens the claims accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core contribution here is real: a new Stein-type test of uniformity on S^{p-1} built on the exponential class, with explicit Gegenbauer coefficients, a clean Karhunen–Loève expansion giving the null limit as a weighted chi-square series, and fixed-alternative asymptotics. I checked several of the coefficient identities and the variance formulas; they are internally consistent. The harmonic decomposition of the dKSD statistic is a useful byproduct, and the limit behavior as lambda goes to 0 or infinity (Rayleigh and Cai et al. extremes) is a nice touch. This is careful, reproducible mathematics.\n\nThe soft spots are two. First, the main power tables use lambda-tilde chosen from 10,000 independent draws from the alternative distribution. That is oracle information, and the comparison against competitors who do not receive it is not fair. The paper does acknowledge this and includes a 20-fold cross-validated version in Table 1, but the headline story is still built on the oracle numbers. The authors should present the cross-validated tuning as the primary evidence, or clearly separate \"potential power\" from achievable power.\n\nSecond, the omnibus consistency claim is stated more broadly than the proof supports. Proposition 1.1 is stated for any absolutely continuous random vector, but the proof begins by assuming q is in L^2(S^{p-1}) and uses the L^2 spherical-harmonic expansion. The same assumption appears in Lemma 3.1 and Theorem 3.3. Remark 3.1 then concludes consistency against all absolutely continuous non-uniform distributions, but as written that only follows for square-integrable densities. The stress-test example q(x) proportional to ||x-e_1||^{-beta} with (p-1)/2 < beta < p-1 is a legitimate counterexample to the stated claim. The gap looks repairable — the positivity of the Gegenbauer coefficients plus a Stone–Weierstrass moment argument should do it — but the paper does not supply that argument. This is a real but bounded flaw.\n\nFor a specialist in directional statistics, the paper is worth a serious referee. The theoretical machinery is new and largely correct; the characterization gap and the oracle-tuning presentation are fixable in revision. I would not desk-reject this, and I would expect a competent referee to engage with the substance rather than the packaging.","headline":"A solid, genuinely new Sobolev-class uniformity test with careful asymptotics; the empirical power story leans on oracle tuning and the claimed omnibus consistency exceeds the L2-density proof.","tokens_in":30138,"tokens_out":2061,"would_cite":true,"duration_ms":23727,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H11","62G10","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a new uniformity test on the hypersphere, proves it is consistent against all absolutely continuous non-uniform distributions, and derives its explicit asymptotic distribution.","keywords":["directional statistics","uniformity test","Stein characterization","Laplace-Beltrami operator","spherical harmonics","Gegenbauer polynomials","Sobolev tests","tuning parameter"],"falsifier":"Simulate U(0,1)-distributed samples from the uniform law on S², compute T_n(λ) for λ = 1 and n = 1000 over 100,000 replicates, and compare the empirical 95th percentile to the closed-form null series Σ_k c_{k,3}(1) γ_{k,3} χ²_{2k+1} truncated at k = 100. A systematic mismatch beyond Monte Carlo error would refute the claimed null distribution.","tokens_in":29309,"feed_emoji":"🌐","tokens_out":4447,"duration_ms":46741,"temperature":0.7,"pith_summary":"The paper builds a test of uniformity on the sphere S^{p-1} from a Stein characterization tied to the Laplace–Beltrami operator. The test checks whether the empirical average of Δ(e^{λ t·X}) vanishes in L² over all directions t, and the authors prove that only the uniform law makes the population version zero. They show the statistic has a closed-form Gegenbauer expansion, belongs to the classical Sobolev class of tests, and converges under the null to an explicit weighted sum of chi-squares. Under fixed alternatives the statistic grows like n times a positive constant, so the test detects every absolutely continuous non-uniform distribution. A tuning parameter λ shifts sensitivity from low-order to high-order features, with data-driven selection that improves power against concentrated and multimodal alternatives.","feed_headline":"A Stein test that catches every non-uniform law on the sphere","feed_subtitle":"Statistic built from the sphere's Laplace operator has explicit null distribution plus a tunable parameter for power.","key_machinery":"The central object is the statistic T_n(λ) = ‖ n^{-1/2} Σ_j Δ_{S^{p-1}} e^{λ t·X_j} ‖²_{L²(S^{p-1})}. The mechanism that carries the argument is the eigenfunction relation for the Laplace–Beltrami operator: Gegenbauer (and Chebyshev) polynomials are eigenfunctions, so the exponential test function e^{λ t·x} has an explicit harmonic expansion with coefficients involving modified Bessel functions. This turns the statistic into a double sum over Gegenbauer polynomials of X_i·X_j, reveals it as a Sobolev test, and yields closed-form series for the null distribution, the fixed-alternative limit, and the limiting covariance kernels.","core_discovery":"For an absolutely continuous random vector X on S^{p-1}, the identity E[Δ_{S^{p-1}} e^{λ t·X}] = 0 for every t ∈ S^{p-1} holds if and only if X is uniform. The paper therefore defines T_n(λ), the squared L² norm of the empirical version of that quantity, and proves its asymptotic theory: under H0 it converges to Σ_{k=1}∞ c_{k,p}(λ) γ_{k,p} χ²_{d_{k,p}}, and under fixed alternatives T_n(λ)/n converges almost surely to a positive constant, giving consistency. The construction is a Sobolev test with positive weights on every spherical-harmonic degree, and the tuning parameter λ interpolates between a low-order moment test and a maximum inner-product-type test.","pith_inferences":["Allowing λ to be complex would replace the moment-generating function with the characteristic function and the modified Bessel coefficients with ordinary Bessel coefficients; the paper notes this possibility, and it could yield a family of oscillatory kernels with different detection profiles for asymmetric alternatives.","The data-driven tuning rule is demonstrated with an oracle sample from the alternative; in practice a cross-validated version will generally achieve less than the oracle power, so the reported gains are an upper bound for what is attainable from data alone.","The construction relies only on having Laplace–Beltrami eigenfunctions, so it extends in principle to other compact manifolds with empty boundary, though the explicit Gegenbauer coefficients and Funk–Hecke simplifications would be lost and numerical eigenfunctions would be needed.","The λ→∞ limit reduces the statistic to a maximum over pairwise inner products, which suggests the test could be particularly strong against very concentrated or sparse alternatives, and connects to high-dimensional phenomena of random angles on spheres."],"forward_implications":["The test provides an omnibus, consistent procedure for uniformity on S^{p-1}: any absolutely continuous non-uniform distribution is eventually detected with probability one.","The closed-form null distribution allows asymptotic p-values and power curves to be computed without simulation, once the series is truncated at a sufficiently large order.","The tuning parameter λ gives a principled way to trade off detection of diffuse low-frequency deviations against concentrated or multimodal deviations, with a data-driven selection rule.","Because the statistic is a Sobolev test with positive weights at every degree, it connects the Stein-characterization route to the classical Beran–Giné framework of rotation-invariant tests.","For rotationally symmetric alternatives such as the von Mises–Fisher distribution, both the limiting signal τ and the asymptotic variance have explicit expressions, enabling direct power comparisons."],"fun_headline_variants":["Stein-based test catches all non-uniform laws on the sphere","New uniformity test on sphere with explicit null distribution","Tunable Sobolev test for spherical uniformity","Stein characterization yields powerful uniformity test","Laplace-Beltrami operator powers new sphere test"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The consistency and characterization arguments assume every alternative distribution has a square-integrable density with respect to the uniform measure on the sphere; if an alternative has atoms or a very rough density, the spherical-harmonic expansion and the proof that τ > 0 can fail.","fun_headline_variants_meta":{"raw":{"variants":["Stein-based test catches all non-uniform laws on the sphere","New uniformity test on sphere with explicit null distribution","Tunable Sobolev test for spherical uniformity","Stein characterization yields powerful uniformity test","Laplace-Beltrami operator powers new sphere test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000343,"raw_usage":{"total_tokens":1708,"prompt_tokens":712,"completion_tokens":996,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":922}},"tokens_in":456,"tokens_out":996,"duration_ms":7716,"temperature":1.0,"reasoning_tokens":922,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T21:11:23.710621+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate U(0,1)-distributed samples from the uniform law on S², compute T_n(λ) for λ = 1 and n = 1000 over 100,000 replicates, and compare the empirical 95th percentile to the closed-form null series Σ_k c_{k,3}(1) γ_{k,3} χ²_{2k+1} truncated at k = 100. A systematic mismatch beyond Monte Carlo error would refute the claimed null distribution.","supporting_citations":[],"review_version":1}