{"id":"5cffeb9a-78e9-4b47-97df-49988417980c","arxiv_id":"2412.05608","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Graph-based sign and runs tests built on randomized spherical variants achieve exact distribution-free nulls in all dimensions and are consistent for high-dimensional alternatives under stated conditions.","lead":"This paper introduces new nonparametric tests for whether a high-dimensional data cloud is spherically symmetric about a specified center. The tests are exact distribution-free, meaning their statistical behavior is the same in every dimension, and are designed to work even when the number of variables greatly exceeds the sample size.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The consistency theorems are proven for the exact shortest covering path P, but the implemented Prim heuristic P0 is only validated by a small n=5 simulation; the actual test's power under H1 is therefore unproven.","rationale":"The reader's weakest assumption identifies exactly the gap I find most load-bearing: the theory proves consistency for the exact shortest covering path, while the implementation uses a Prim heuristic that is only checked for n=5 on one example. I also checked the more foundational pieces of the paper for internal problems. The exchangeability argument in Theorem 2.1 is sound for the exact argmin, and the same group-invariance argument extends the distribution-free null property to any reasonable deterministic heuristic such as P0. Condition (3.1) and the tightness reasoning in the proof of Theorem 3.2 are plausible, and the consistency proof for the exact path is coherent. The Earthquakes truncation is disclosed and post hoc, but it does not affect the central methodological claim as strongly as the heuristic mismatch. Since the concern is real but the paper already acknowledges it and gives partial empirical support, the existing CONDITIONAL verdict is appropriate: not a rejection, but not a clean accept until the heuristic gap is resolved either by a proof that P0 recovers the same sign/rank statistics under the stated alternatives or by a more substantial exact-vs-heuristic comparison.","tokens_in":40630,"tokens_out":5376,"duration_ms":58981,"concrete_test":"Enumerate all 2^{n-1} n! covering paths for n=7 (322,560 paths) and, if feasible, n=8, for B=100 independent datasets from the spiked-covariance alternative of Example 3.3 (covariance diag(d,1,...,1)) at d=8,32,128,512,1024. For each dataset compute the exact shortest path P by brute force and the heuristic path P0 from Section 2.3, then record TS, TR and rejection decisions at level 0.05. If P0 and P give the same sign and runs statistics on all or almost all replicates at high d, the heuristic gap is empirically benign in this regime. If instead P0 yields different statistics with non-negligible frequency, or if the rejection rate using P0 fails to approach 1 while the exact-path rejection rate does, the consistency claim for the implemented test is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central distribution-free null result (Theorem 2.1 and its analog for the modified cost in Theorem 4.1) is secure for the implemented heuristic P0: because the augmented pairs (Xi, Xi') are exchangeable under H0, any deterministic algorithm equivariant under permutations and coordinate-wise sign flips yields S uniform on {0,1}^n, R uniform on Sn, and independence. The load-bearing gap is on the alternative side. Theorems 3.2, 4.1(b), and 4.2 establish that the exact argmin P of (2.2)/(4.1) has S converging to 1_n under conditions such as (3.1) or (4.2). Section 2.3 explicitly replaces P by a Prim-based heuristic P0 and concedes it may be suboptimal. The paper only checks, for n=5, that TS and TR computed along P and P0 coincide in a specific normal example, with differences shrinking as d grows. No theorem shows that P0 inherits the sign/rank degeneracy of P under the HDLSS or HDHSS alternatives covered by the consistency results. Since the tests actually shipped and simulated use P0, the claimed consistency applies to a statistic that is not the one evaluated in Section 5 or used by practitioners. This is not an internal contradiction, but it is a substantial missing link between theory and implementation. A suboptimal P0 could, in principle, select paths containing many Xi' even when the exact path is all-X, breaking the argument that TS reaches n and TR reaches 1 with probability tending to 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes new tests of spherical symmetry for multivariate distributions based on data augmentation and a graph-based notion of string signs and string ranks. A shortest covering path is constructed on the augmented set of original observations and their spherically symmetric variants, and test statistics are built from the sign and rank sequences along that path. The central theoretical result, Theorem 2.1, states that under the null hypothesis of spherical symmetry the string signs are uniform on {0,1}^n, the string ranks are uniform on the permutation group, and the two are independent, irrespective of the dimension. Consistency is claimed in the HDLSS regime under condition (3.1) for the original sign/runs tests and under condition (4.2) for the modified tests, and in the HDHSS regime under additional assumptions. The paper also presents a sample-splitting extension for an unknown center of symmetry and reports simulations and a real-data analysis comparing the proposed tests with optimal-transport and density-based competitors.","tokens_in":40966,"tokens_out":4646,"duration_ms":46265,"significance":"If the main claims hold, the paper makes a useful contribution: it provides genuinely distribution-free tests of spherical symmetry that work even when the dimension is much larger than the sample size, and it gives a new way to convert a high-dimensional symmetry problem into univariate sign and runs problems. The null distribution result in Theorem 2.1 is elegant and appears correct, and the observation that the test statistics are exactly distribution-free regardless of dimension is valuable in an area where many high-dimensional tests are only asymptotically calibrated. The paper also includes R code in supplementary material, and the consistency theorems are stated under explicit conditions rather than being left entirely to heuristics. The main weakness is that the consistency theorems are proved for the exact shortest covering path P, while the implemented algorithm is a Prim-based heuristic P0; the paper only provides a small simulation study linking the two. This leaves the alternative-side theory of the actual implemented test incomplete.","major_comments":[{"comment":"The consistency theorems are proved for the exact shortest covering path P defined by the minimizer in (2.2) and (4.1), but the tests implemented and simulated in Section 5 and Section 6 use the Prim-based heuristic P0 described in Section 2.3. The paper explicitly concedes that P0 may be suboptimal and supports the identification of P0 with P only by the n=5 simulation in Figure 4. No theorem shows that P0 inherits the key degeneracy property S → 1_n (or T_M^S → n and T_M^R → 1) under the alternatives covered by conditions (3.1) and (4.2). Since a suboptimal path could in principle contain many spherically symmetric variants X'_i even when the exact path is all-X, the claimed consistency of the tests actually shipped is not established. This is a load-bearing gap between theory and implementation.","section":"§2.3 and §3.1, Theorem 3.2 / Theorem 4.1(b) / Theorem 4.2"},{"comment":"The claimed HDHSS consistency of the sign and runs tests is conditional on inequalities that are not derived from explicit distributional assumptions. For the sign test, the text states that under H1, p_S is 'expected to be higher' than 0.5, and consistency is asserted when p_S > 0.5; for the runs test, consistency is asserted when p_R < 0.5. No theorem establishes p_S > 0.5 or p_R < 0.5 for any concrete class of non-spherical alternatives beyond the already-proven degeneracies under condition (3.1), which are again stated for the exact P. Similarly, Theorem 4.4 assumes the relevant limits of E[TS/n] or E[~TS/n] are away from 0.5 without giving sufficient conditions on P. Thus the HDHSS consistency statement in the abstract is stronger than what is proved.","section":"§3.2, Theorem 3.4 and following paragraph; §3.2 Theorem 3.6"},{"comment":"The sample-splitting procedure for the unknown-center case is presented as a generalization with the exact distribution-free property and asymptotic properties 'similar' to those in Sections 3 and 4, but no proofs are given for these claims. Lemma 6.1 establishes only that spherical symmetry of the differenced distribution is equivalent to spherical symmetry of the original distribution; it does not establish the distribution-free property of the resulting test or its consistency under the HDLSS/HDHSS alternatives. Since this section is advertised in the introduction and abstract, the missing formal support should be either supplied or explicitly left as a conjecture.","section":"§6, sample-splitting extension"}],"minor_comments":[{"comment":"There are several typos: 'string' is written as 'sting' in 'signs and string ranks' and 'sting ranks'; 'δ denotes the the indicator function' has a doubled article.","section":"§2.2"},{"comment":"The sentence begins 'In thus section' instead of 'In this section'.","section":"§3.2, first sentence"},{"comment":"The phrase 'where we do have no signals from the diagonal part' appears in the discussion of Example 3.1; the intended meaning is clear but the wording should be corrected.","section":"§4, before Theorem 4.1"},{"comment":"The caption of Figure 2 says 'over 100 simulations', but the text of §2.2 says the experiment was repeated 1000 times; these numbers should be reconciled.","section":"Figure 2 caption vs. §2.2 text"},{"comment":"The text contains a typo: 'he DT test' instead of 'the DT test'.","section":"§5.1"},{"comment":"The notation in condition (3.1) writes the probability with 'P[...] → 1 for all M > 0'; since the quantity inside depends on d, it would be clearer to write the limit as d → ∞ explicitly.","section":"§3.1, Theorem 3.2, displayed condition"}],"recommendation":"major_revision","confidential_remarks":"The core null-distribution result is sound and is the clear strength of the paper. The main obstacle to acceptance is the missing link between the exact shortest covering path P used in the theorems and the heuristic P0 used in the implementation: without a proof that P0 inherits the alternative-side degeneracy, the stated consistency claims do not cover the actual test. This seems fixable within the manuscript's scope, for example by proving a lemma that under condition (3.1) or (4.2) P0 also selects the all-X path with probability tending to one, or by explicitly re-stating the consistency results as conditional on the exact P and marking the P0 behavior as empirical. I do not see internal circularity in the main theorem; the missing pieces are genuinely missing proofs rather than hidden assumptions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things you should know. First, the core null result is real and important: string signs and string ranks computed along the shortest covering path on the augmented data have exactly the same null distribution as univariate signs and ranks, regardless of dimension. That is a genuine first, and it is proved correctly. Second, the consistency theorems are all for the exact shortest covering path, while the algorithm actually implemented and simulated is a Prim-based heuristic. The paper explicitly says the heuristic may be suboptimal, so this is an honest limitation, but it is a real gap: no theorem shows the heuristic inherits the sign-degeneracy under the alternatives covered by Theorems 3.2, 4.1(b), and 4.2. Only a small n=5 simulation for one normal example is given.\n\nThe distribution-free null property is secure even for the heuristic, as the authors note: under H0, any deterministic algorithm equivariant under permutations and sign flips yields uniform signs and ranks. So the gap is confined to the alternative side. Still, since practitioners will use the heuristic, the claimed consistency of the actual test statistic is unproven. That is the main reason I would not call this a clean accept.\n\nWhat is genuinely new: the string signs/ranks along a covering path, and the exact dimension-free null distributions. The paper also gives a thoughtful treatment of the HDLSS geometry, explains why inner-product-based tests fail under the sphericity condition, and proposes a modified cost to catch scale alternatives. The simulations are broad, and the Earthquakes analysis is unusual in that they disclose and then address a possible data-curation bias by truncation. That is good practice.\n\nThe weak points, in proportion: (1) the heuristic gap above, which is substantial; (2) Theorems 3.4 and 3.6 are variance bounds rather than full consistency, though the paper clearly states the extra conditions needed; (3) the Bonferroni-based modified tests are simple but may be conservative, which the paper acknowledges.\n\nWho is this for: any statistician working on nonparametric multivariate tests, especially in HDLSS settings. It deserves a serious referee; the idea is novel and the null result alone is enough to justify careful reading. I would recommend sending it out, with instructions to the authors to close or at least quantify the heuristic gap—perhaps a worst-case comparison or significantly larger simulations across more alternatives.","headline":"The exact distribution-free null result is secure; the open gap is whether the implemented Prim heuristic inherits the consistency theorems proven for the exact path.","tokens_in":41446,"tokens_out":2824,"would_cite":true,"duration_ms":26539,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62H15","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Spherical symmetry can be tested distribution-free in any dimension by reading string signs and ranks off a shortest covering path in an augmented data set.","keywords":["spherical symmetry","distribution-free tests","data augmentation","string signs and ranks","runs test","sign test","high-dimensional asymptotics","HDLSS"],"falsifier":"For a fixed small sample size such as $n=6$ or $n=8$, simulate many datasets from an alternative satisfying condition (3.1) at large $d$ (for example, a spiked covariance model), enumerate all covering paths to find the exact shortest path $P$, and compare the sign and runs statistics on $P$ with those on the Prim heuristic path $P_0$; if $P_0$ gives $T_S < n$ or $T_R > 1$ with non-negligible frequency while $P$ gives $T_S = n$ and $T_R = 1$, the consistency theorem does not apply to the implemented test.","tokens_in":40431,"feed_emoji":"📊","tokens_out":10252,"duration_ms":85659,"temperature":0.7,"pith_summary":"The paper constructs exact distribution-free tests of spherical symmetry that work even when the dimension far exceeds the sample size. On an augmented data set containing each observation and its spherically symmetric variant, a shortest covering path selects one of the two copies per observation; the resulting string signs and string ranks have a null distribution that is uniform on all sign strings and permutations, independent of the dimension. Consequently the sign and runs statistics have the same null distributions as their classical univariate counterparts, so cutoffs come from standard tables. Under high-dimensional asymptotics the paper proves consistency against alternatives such as spiked covariance models, and a modified cost function extends this to alternatives that differ only in the scales of the coordinates. A sample-splitting variant handles the case of an unknown center of symmetry.","feed_headline":"One path yields exact spherical-symmetry tests in any dimension","feed_subtitle":"A sign and runs statistic from a covering path matches univariate null tables in every dimension, even when d exceeds n.","key_machinery":"The load-bearing object is the shortest covering path on the augmented complete graph: given $X_1,\\dots,X_n$ and independent spherical copies $X_1',\\dots,X_n'$, the path minimizes the sum of edge costs $\\theta(Z_i,Z_j)=\\exp\\{-(Z_i^\\top Z_j/d)^2\\}$ (or $\\tilde\\theta$ for scale alternatives) while visiting exactly one member of each pair $\\{X_i,X_i'\\}$. The binary choice at each pair is the string sign $S_i$; the position along the path is the string rank $R_i$. Under the null the pairs are exchangeable, which forces $(S,R)$ to be uniform on $\\{0,1\\}^n \\times S_n$ and independent; this uniformity is the mechanism that transfers the classical univariate distribution-free property to arbitrary dimension.","core_discovery":"The central claim is that spherical symmetry of a $d$-dimensional distribution can be tested by comparing each observation $X_i$ with an independent spherically symmetric copy $X_i' = \\|X_i\\|U_i$, $U_i \\sim \\mathrm{Unif}(S^{d-1})$, in a graph whose edge costs are $\\theta(Z_i,Z_j)=\\exp\\{-(Z_i^\\top Z_j/d)^2\\}$. The shortest covering path through the $2n$ augmented points chooses exactly one of $\\{X_i, X_i'\\}$ for each $i$; the choices form the string sign vector $S$ and the positions form the string rank vector $R$. Theorem 2.1 shows that under the null, $S \\sim \\mathrm{Unif}(\\{0,1\\}^n)$, $R \\sim \\mathrm{Unif}(S_n)$, and $S$ and $R$ are independent, so $T_S = \\sum_i S_i$ and $T_R = 1 + \\sum_{i=1}^{n-1} I\\{S_{\\pi_i} \\ne S_{\\pi_{i+1}}\\}$ have exact null distributions equal to the classical univariate sign and runs tests in every dimension. Under condition (3.1) the sign vector converges to all ones in high dimension, giving consistency of sign and runs tests in the HDLSS regime; with a modified cost $\\tilde\\theta(Z_i,Z_j)=\\exp\\{-\\frac1d \\sum_q Z_{iq}^2 Z_{jq}^2\\}$, detection of scale alternatives is proved, and the two are combined into modified tests consistent under either signal. The unknown-center case is handled by applying the tests to the differences $X_i - X_{n/2+i}$, which retain the exact distribution-free property.","pith_inferences":["The same exchangeability argument would produce exact distribution-free tests for other composite nulls in which each observation has a family of alternatives definable by a cost on an augmented graph; for example, tests of elliptical symmetry after a robust whitening step, though standardization would likely destroy exactness and need calibration.","The practical gap between the exact path $P$ and the Prim heuristic $P_0$ is the main unproved step; a natural extension is to prove or test that $P_0$ tracks $P$ under condition (3.1), or to use exact enumeration for small $n$ as a gold-standard benchmark.","Alternatives with scalar covariance but non-spherical coordinates (e.g., i.i.d. Laplace coordinates) are not covered by either consistency theorem; the paper shows empirical power for two such examples, suggesting a possible regime where the tests are consistent but the proof would need new conditions beyond (3.1) and (4.2).","Pairwise differencing for unknown center halves the effective sample size; using overlapping differences could improve power but would break exact independence, an extension worth exploring."],"forward_implications":["A practitioner can test spherical symmetry when $d \\gg n$ using ordinary univariate sign-test and runs-test tables; no permutation resampling or dimension-dependent critical values are needed.","In the high-dimension, low-sample-size regime, the sign statistic converges to $n$ and the runs statistic to $1$ in probability for alternatives satisfying condition (3.1), so both tests are consistent once $n$ is large enough that $2^n > 1/\\alpha$ for the sign test.","In the high-dimension, high-sample-size regime, the null distributions of $T_S$ and $T_R$ are asymptotically normal with variance $1/4$, still independent of dimension.","The modified cost function $\\tilde\\theta$ yields consistent tests against alternatives where only the diagonal scales of the covariance differ, and the Bonferroni-combined modified tests are consistent under either the inner-product condition or the scale condition.","When the center of symmetry is unknown, replacing observations by pairwise differences $X_i - X_{n/2+i}$ gives an exact distribution-free test and avoids the inflated Type I error caused by centering in high dimensions."],"supporting_citations":[{"why":"Supplies the characterization that a random vector is spherically symmetric if and only if it has the same distribution as its radial part times an independent uniform direction, the starting point for data augmentation.","marker":"Fang et al., 1990"},{"why":"Provides the pairwise-distance distributional equality that characterizes spherical symmetry, motivating the use of squared inner products as the edge cost.","marker":"Maa et al., 1996"},{"why":"Prior data-augmentation test of spherical symmetry that the present work extends; also supplies the contiguity proposition used in the Pitman-efficiency argument.","marker":"Banerjee and Ghosh (2024)"},{"why":"Introduces the shortest covering path construction and its Prim-based heuristic, which the paper adapts to compute string signs and ranks.","marker":"Biswas et al., 2015"},{"why":"Prim's algorithm is the actual computational device used to construct the heuristic path P0 in Section 2.3.","marker":"Prim (1957)"},{"why":"Provides the sphericity condition and HDLSS geometric representation that underlie Theorem 3.1 and the covariance assumptions used in Corollary 1.","marker":"Jung and Marron (2009)"},{"why":"Gives the Beta distribution of a squared coordinate of a uniform random vector on the sphere, used in the proof of Theorem 3.2 to show the inner-product comparison is tight in high dimension.","marker":"Liang et al., 2008"},{"why":"Supplies the martingale central limit theorem used to derive the dimension-free asymptotic null distribution of the runs statistic in Theorem 3.5.","marker":"Brown (1971)"}],"fun_headline_variants":["Exact spherical symmetry tests for high-dimensional data","A path gives exact distribution-free spherical tests","Spherical symmetry tests that work when d exceeds n","New exact tests for spherical symmetry in any dimension","Graph-based signs and ranks yield exact spherical tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The consistency theorems are proved for the exact shortest covering path, but the implemented algorithm is a Prim-based heuristic that can return a suboptimal path; the paper assumes, without proof, that the heuristic yields the same sign and runs statistics under alternatives in the high-dimensional regimes, supported by a simulation with only five observations.","fun_headline_variants_meta":{"raw":{"variants":["Exact spherical symmetry tests for high-dimensional data","A path gives exact distribution-free spherical tests","Spherical symmetry tests that work when d exceeds n","New exact tests for spherical symmetry in any dimension","Graph-based signs and ranks yield exact spherical tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1899,"prompt_tokens":1096,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":732}},"tokens_in":712,"tokens_out":803,"duration_ms":5827,"temperature":1.0,"reasoning_tokens":732,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:33:39.455643+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a fixed small sample size such as $n=6$ or $n=8$, simulate many datasets from an alternative satisfying condition (3.1) at large $d$ (for example, a spiked covariance model), enumerate all covering paths to find the exact shortest path $P$, and compare the sign and runs statistics on $P$ with those on the Prim heuristic path $P_0$; if $P_0$ gives $T_S < n$ or $T_R > 1$ with non-negligible frequency while $P$ gives $T_S = n$ and $T_R = 1$, the consistency theorem does not apply to the implemented test.","supporting_citations":[{"cited_title":"T., Kotz, S., and Ng, K","cited_arxiv_id":null,"evidence_quote":"Supplies the characterization that a random vector is spherically symmetric if and only if it has the same distribution as its radial part times an independent uniform direction, the starting point for data augmentation."},{"cited_title":"K., and Bartoszy\\' n ski, R","cited_arxiv_id":null,"evidence_quote":"Provides the pairwise-distance distributional equality that characterizes spherical symmetry, motivating the use of squared inner products as the edge cost."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the shortest covering path construction and its Prim-based heuristic, which the paper adapts to compute string signs and ranks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prim's algorithm is the actual computational device used to construct the heuristic path P0 in Section 2.3."},{"cited_title":"and Marron, J","cited_arxiv_id":null,"evidence_quote":"Provides the sphericity condition and HDLSS geometric representation that underlie Theorem 3.1 and the covariance assumptions used in Corollary 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Beta distribution of a squared coordinate of a uniform random vector on the sphere, used in the proof of Theorem 3.2 to show the inner-product comparison is tight in high dimension."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the martingale central limit theorem used to derive the dimension-free asymptotic null distribution of the runs statistic in Theorem 3.5."}],"review_version":1}