{"id":"80f569e6-77b2-4ddd-ad52-39417e8cdeb7","arxiv_id":"2411.13080","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A new class of rank-based, kernel and optimal-transport measures of multivariate association that are distribution-free under independence, exactly characterize independence and functional dependence, and have a uniform null CLT.","lead":"The authors define a family of dependence measures between two random vectors that is exactly distribution-free under independence, meaning the test statistic's null distribution does not depend on the unknown marginals. The measures are built from optimal-transport ranks and kernel embeddings, and they hit 0 exactly for independent variables and 1 exactly when one variable is a function of the other.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Consistency proof in Theorem 3.1(b) has an index-set error: the E[Z_n^2] decomposition into (I)+(II)+(III) double-counts overlapping edge pairs, so the proof is incomplete as written.","rationale":"The reader identified the Hölder condition (S1) as the weakest assumption, and that is a real limitation. However, the more immediate threat to the central claim is the proof of Theorem 3.1(b) itself: the displayed decomposition of E[Z_n^2] is not a partition of index configurations. The condition k≠ℓ and i≠j does not exclude k=j or ℓ=i, so (III) includes same-edge reverse terms and shared-vertex paths that belong in (I) and (II). The proof's later approximation arguments replace conditional expectations by products and are only valid for disjoint pairs. This is not an objection to the result's truth; it is a concrete formal gap in the written proof. Because Theorem 3.1(b) underpins the consistency claim and the consistency of the proposed test, it should be fixed before the paper is accepted. The paper is otherwise mathematically coherent: the distribution-free argument under independence is straightforward exchangeability, Theorem 3.2's 0/1 characterization follows from the kernel-embedding theorem in [22], and the CLT proof leans on [22] but is plausible. The authors' own Remark 3.1 also concedes that (S1) is a proof-technique restriction, which supports a conditional rather than unconditional acceptance. The missing disjointness condition in (III) is likely fixable, so I do not recommend rejection; the appropriate verdict remains CONDITIONAL, meaning the reader's verdict is unchanged.","tokens_in":23914,"tokens_out":22912,"duration_ms":238244,"concrete_test":"Set n=2, d1=d2=1, H_n chosen so the empirical rank graph has exactly one edge. Compute E[Z_n^2] directly from the definition of \\hatη_rank in (3.1). Then evaluate the sum (I)+(II)+(III) exactly as written in the proof. If the two disagree (as the algebra indicates, (I)+(II)+(III)=1.5·E[K(Y1,Y2)^2] vs E[Z_n^2]=E[K(Y1,Y2)^2]), the decomposition is incorrect. Then redo the proof with (III) restricted to edges {i,k} and {j,ℓ} disjoint and check whether the limits (5.2)–(5.5) and the bound (5.3) hold; this determines whether the gap is a fixable typo or a substantive obstruction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Theorem 3.1(b), Section 5.1, E[Z_n^2] is decomposed as (I)+(II)+(III). The term (III) is defined by sums over i≠j, k≠ℓ with (k,i),(ℓ,j)∈E, but it does not require the two edges {i,k} and {j,ℓ} to be disjoint. Consequently, reverse-edge terms and two-edge paths are included in (III) even though they are already counted in (I) or (II). A minimal check with n=2 and a single-edge graph gives (I)+(II)+(III)=1.5·E[K(Y1,Y2)^2] while E[Z_n^2]=E[K(Y1,Y2)^2]. The subsequent limits (5.2)–(5.5) and the error bound (5.3) are only justified when the two edges are disjoint. Thus the proof of consistency, which is the basis for the claimed consistent estimation and for consistency of the independence test, is not valid as printed. This gap is separate from the Hölder condition (S1); even granting (S1), the decomposition must be repaired.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a class of distribution-free measures of association between random vectors X in R^{d1} and Y in R^{d2}, constructed by applying the graph-based kernel estimator of Deb et al. [22] to optimal-transport multivariate ranks of the observations. The empirical measure \\hat\\eta_n^{rank} is shown to have a pivotal distribution under independence (Theorem 3.1(a)), to converge to a population measure \\eta_K^{rank} (Theorem 3.1(b)) that is 0 iff X and Y are independent and 1 iff Y is a measurable function of X (Theorem 3.2), and to satisfy a uniform CLT under the null (Theorem 3.3). For d2=1 and a particular kernel, \\eta_K^{rank} is shown to coincide with the Azadkia-Chatterjee/Chatterjee coefficient (Proposition 3.1). The proofs are collected in Section 5.","tokens_in":24175,"tokens_out":21040,"duration_ms":186992,"significance":"The proposed construction is conceptually appealing: OT-based multivariate ranks deliver exact distribution-freeness, while the RKHS/graph framework provides interpretable population limits and consistency. If the results hold, this is, to my knowledge, the first class of multivariate association measures combining all three properties (zero/one characterization, consistent distribution-free estimation, and a null CLT). The connection with Chatterjee's coefficient is a nice sanity check. The paper is honest about limitations, e.g., Condition (S1) being a proof-technique assumption. However, a load-bearing gap in the consistency proof needs to be addressed before the results can be considered established.","major_comments":[{"comment":"The set defining (III) only imposes i≠j and k≠ℓ, so it contains reverse-edge terms (k=j, ℓ=i) and two-edge path terms (i=ℓ or k=j) that are already counted in (I) or (II). For n=2 and a single-edge graph, (I)+(II)+(III)=1.5·E[K(Y1,Y2)^2] while E[Z_n^2]=E[K(Y1,Y2)^2]. The later limits (5.2) and the error bound (5.3) require factorizing conditional expectations, which is valid only when the two edges are disjoint. Since the proof of consistency (Theorem 3.1(b)) and hence the consistency of the independence test rest on this step, the gap is load-bearing. The likely repair is to add the distinctness condition on i,j,k,ℓ in (III) and in (~III), but the corrected proof must be supplied.","section":"Section 5.1, decomposition of E[Z_n^2] into (I)+(II)+(III)"},{"comment":"The sentence 'the last inequality follows once again from Proposition 5.1' is not justified as written. Proposition 5.1 gives n^{-1}∑_i ||\\hat R^X_n(X_i)-R^X(X_i)||^2 → 0 almost surely, but (5.4) requires control of the β-Hölder edge sum ∑_i d_i^{-1}∑_{k∼i} ||R^X(X_i)-R^X(X_k)||^β in terms of the corresponding sum with \\hat R^X_n. An additional argument (e.g., triangle inequality plus (a+b+c)^β ≤ a^β+b^β+c^β for 0<β≤1) is needed to show the replacement error is o(1) under (S2).","section":"Section 5.1, Eq. (5.4)"}],"minor_comments":[{"comment":"The subscript 'RX(Xl)' should read 'RX(X_\\ell)'.","section":"Section 5.1, definition of (III)"},{"comment":"The symbols ~Y1 and ~Y2 are introduced without definition; please align the notation with the earlier (Y', ~Y').","section":"Section 5.1, after Eq. (5.2)"},{"comment":"The step 'The conclusion follows directly from [22, Theorem 4.1]' should be expanded to state why the uniformity over ~G ∈ Jθ is covered and why the rank-transformed observations satisfy the conditions of that theorem.","section":"Theorem 3.3, proof of (5.13)"},{"comment":"The theorems should clarify whether the sets H^{d1}_n and H^{d2}_n are fixed or random; several steps (e.g., the determinism of D_n and the ~g_j's) assume fixed design points, while (S3) only requires weak convergence.","section":"Section 3.1 and Remark 3.4"}],"recommendation":"major_revision","confidential_remarks":"The index-set issue in Section 5.1 appears repairable (likely a missing disjointness condition in (III)), but as printed it invalidates the proof of the consistency theorem. The paper relies heavily on the authors' own prior work [22,23]; this is acceptable, but the key steps should be sufficiently self-contained. I recommend major revision rather than rejection because the central idea is sound and the gap is local."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a genuinely nice idea and mostly delivers what it claims, but the proof of Theorem 3.1(b) has a concrete gap in the second-moment computation. The gap is repairable, but it is real.\n\nWhat is new and good: the construction is simple once you see it. Applying OT-based multivariate ranks to the authors' earlier kernel/graph measure gives an estimator that is exactly distribution-free under independence, has a well-defined population limit with the 0/1 independence/functional-dependence characterization, and recovers Chatterjee / Azadkia-Chatterjee at the population level for a particular kernel. The distribution-free property is not trivial because OT ranks are exchangeable; the proof is short and correct. The 0/1 characterization also goes through cleanly by inverting the rank map. The uniform null CLT is valuable, and the variance estimator is explicit and pivotal.\n\nThe main issue is in Section 5.1. The decomposition E[Z_n^2] = (I)+(II)+(III) is not an identity. (III) as written sums over all ordered pairs of edges (k,i) and (l,j) without requiring the edges to be disjoint. Same-edge pairs and adjacent-edge pairs are therefore counted in (III) as well as in (I) and (II). With n=2 and a single edge, (I)+(II)+(III) gives 1.5 E[K^2] while E[Z_n^2] = E[K^2]. The limits (5.2)-(5.5) are only justified for the disjoint part, so the proof of part (a) is incomplete as printed. I think the fix is straightforward: split (III) into disjoint and overlapping parts and show the overlapping part is negligible using (3.4)-(3.5). For k-NN and MST graphs the overlapping contributions are O(1/n k^2) or O(1/n). But the authors need to write this out.\n\nTwo lesser issues. Condition (S1) is a genuine smoothness restriction on the conditional kernel mean embedding; the authors admit in Remark 3.1 that it is only needed for their proof technique. That is honest, but it narrows the scope of the consistency result. The proof also leans on the authors' unpublished preprint [22] for two key ingredients (Theorem 2.1 and Theorem 4.1). That is acceptable if [22] is public and the results are cited precisely, but it makes independent verification harder.\n\nOne thing the paper does not do is any simulation; the authors explicitly defer it. For a new measure of association that is a real omission, though not a correctness issue.\n\nWho this is for: theoretical nonparametrics people working on dependence measures, OT ranks, and graph-based inference. It deserves a serious referee. I would send it out despite the proof gap, with instructions to check the repair of (III).","headline":"A clever and mostly correct construction of distribution-free OT-rank-based association measures, but the consistency proof has a real index-set error that should be fixable.","tokens_in":24682,"tokens_out":13093,"would_cite":true,"duration_ms":110636,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62H20","60F05","60D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Replacing observations with optimal-transport multivariate ranks makes a kernel graph-based association measure exactly distribution-free, with 0 exactly under independence and 1 exactly when one variable is a measurable function of the…","keywords":["measure of association","maximum mean discrepancy","multivariate ranks","optimal transport","distribution-free test","geometric graphs","reproducing kernel Hilbert space","uniform central limit theorem"],"falsifier":"Under independence, fix small $n$, the grids, the kernel, and the graph, and enumerate the exact distribution of $\\hat\\eta_n^{\\mathrm{rank}}$ for two very different marginal pairs; Theorem 3.1(a) predicts identical distributions, so any discrepancy would disprove the pivotal claim.","tokens_in":23694,"feed_emoji":"📊","tokens_out":9733,"duration_ms":118844,"temperature":0.7,"pith_summary":"This paper proposes a class of association measures between random vectors $X$ and $Y$ that combines reproducing kernel Hilbert spaces, geometric graphs, and optimal-transport multivariate ranks. The key claim is that the sample measure $\\hat\\eta_n^{\\mathrm{rank}}$ is exactly distribution-free when $X$ and $Y$ are independent, so it yields an exact finite-sample test of independence in any dimension. The same statistic consistently estimates a population limit $\\eta_K^{\\mathrm{rank}}$ that is $0$ exactly under independence and $1$ exactly when $Y$ is a measurable function of $X$. If correct, this gives practitioners a single nonparametric number that both measures dependence strength and tests for independence without permutations or asymptotic calibration.","feed_headline":"Optimal transport ranks give distribution-free dependence tests","feed_subtitle":"New association measure: 0 exactly under independence, 1 exactly under functional dependence.","key_machinery":"The load-bearing object is the empirical multivariate rank map, defined by optimally transporting the sample to a fixed set $H_n$ of $n$ uniform-like points in $[0,1]^d$ using the Brenier-McCann optimal transport map. Because the observations are exchangeable, the rank vectors are uniformly distributed over permutations of $H_n$, which is exactly what makes the statistic pivotal under independence. The kernel $K$ measures similarity of rank-transformed $Y$-values, and the geometric graph on the rank-transformed $X$-values estimates the conditional expectation $E[K(R_Y(Y'), R_Y(\\tilde Y'))\\mid X']$. The statistic is the rank analogue of Spearman correlation: compute the kernel association on ranks rather than raw data.","core_discovery":"The central discovery is that replacing $X_i$ and $Y_i$ by their empirical multivariate ranks, defined as the optimal-transport map to a fixed uniform-like grid, makes the kernel-based geometric-graph estimator $\\hat\\eta_n$ distribution-free while preserving consistency and interpretability. Under the null $X \\perp\\!\\!\\perp Y$, the ranks are a uniform random permutation of fixed points, so $\\hat\\eta_n^{\\mathrm{rank}}$ has a pivotal distribution. The paper proves that $\\hat\\eta_n^{\\mathrm{rank}}$ converges to $\\eta_K^{\\mathrm{rank}}$, establishes that $\\eta_K^{\\mathrm{rank}}\\in[0,1]$ with equality to $0$ iff independence and to $1$ iff $Y=g(X)$ almost surely, and gives a uniform central limit theorem under the null over a large class of geometric graphs. The authors further claim that, to their knowledge, this is the only class of procedures that simultaneously has all of these properties.","pith_inferences":["Because the pivotal property relies only on permutation uniformity of the ranks, the same construction should yield valid exact tests when $H_n$ is any deterministic low-discrepancy sequence; the paper only details uniform-grid choices, but the mechanism is general.","The uniform CLT opens the door to data-driven tuning of the graph, such as choosing $k$ by a pilot estimate of dependence, without invalidating the test; the paper mentions this possibility but does not develop a concrete procedure.","Symmetrizing by $\\max(\\hat\\eta_n^{\\mathrm{rank}}(X,Y), \\hat\\eta_n^{\\mathrm{rank}}(Y,X))$ gives a distribution-free measure of mutual dependence that is $1$ iff either variable is a measurable function of the other; the paper notes this but does not analyze its limit distribution.","One could use other reference measures than uniform on the cube, such as a Gaussian or Student-$t$ distribution, to emphasize tail regions; the paper says proofs carry over for compactly supported references but leaves the unbounded-support case open."],"forward_implications":["For any $d_1,d_2\\ge 1$, one can test mutual independence at an exact level without permutations, resampling, or asymptotic critical values, because the null distribution of $\\hat\\eta_n^{\\mathrm{rank}}$ is known once $n$, the grid, the kernel, and the graph are fixed.","The population measure gives a single interpretable number: $0$ means independence, $1$ means $Y$ is a measurable function of $X$, and intermediate values compare strength of association.","Consistency holds for estimators based on $k$-nearest-neighbor graphs and minimum spanning trees, so the method inherits the flexibility of graph-based dependence estimation.","The uniform CLT permits data-dependent graph choices, such as a nearest-neighbor radius $k$ that grows logarithmically with $n$, while retaining Gaussian limiting behavior.","For univariate $Y$ with the kernel $K(y_1,y_2)=|y_1|+|y_2|-|y_1-y_2|$, the population limit reduces to the previously proposed scalar regression-dependence coefficient, placing that measure as a special case of this family."],"supporting_citations":[{"why":"Defines the kernel measure $\\eta_K$ and the geometric-graph estimator $\\hat\\eta_n$ that the rank version modifies.","marker":"[22]"},{"why":"Supplies the consistency of empirical OT ranks to population ranks and their pivotal permutation distribution.","marker":"[23]"},{"why":"Gives the scalar regression-dependence coefficient recovered as a special case in Proposition 3.1.","marker":"[16]"},{"why":"Gives the conditional-dependence coefficient whose population limit coincides with $\\eta_K^{\\mathrm{rank}}$ for $d_2=1$.","marker":"[4]"},{"why":"Introduced the univariate copula-based measure of regression dependence with properties (P1)-(P3).","marker":"[25]"},{"why":"Provides the Brenier-McCann theorem used to define population and empirical OT rank maps.","marker":"[47]"},{"why":"Supplies the multivariate rank methodology and algorithmic background for empirical OT ranks.","marker":"[59]"},{"why":"Supplies the distance covariance kernel that yields the special-case coefficient in Proposition 3.1.","marker":"[67]"}],"fun_headline_variants":["Optimal transport ranks yield distribution-free independence tests","Dependence measure: 0 iff independent, 1 iff functional","Distribution-free association via optimal transport ranks","Exact finite-sample tests from rank-based association measure"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The consistency proof needs a Hölder (power-law smoothness) condition on the conditional kernel inner product after rank transformation; the authors say this condition is used only by their proof technique and expect it can be relaxed.","fun_headline_variants_meta":{"raw":{"variants":["Optimal transport ranks yield distribution-free independence tests","Dependence measure: 0 iff independent, 1 iff functional","Distribution-free association via optimal transport ranks","Exact finite-sample tests from rank-based association measure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2382,"prompt_tokens":962,"completion_tokens":1420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1358}},"tokens_in":578,"tokens_out":1420,"duration_ms":12831,"temperature":1.0,"reasoning_tokens":1358,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:51:07.037298+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Under independence, fix small $n$, the grids, the kernel, and the graph, and enumerate the exact distribution of $\\hat\\eta_n^{\\mathrm{rank}}$ for two very different marginal pairs; Theorem 3.1(a) predicts identical distributions, so any discrepancy would disprove the pivotal claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the consistency of empirical OT ranks to population ranks and their pivotal permutation distribution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the scalar regression-dependence coefficient recovered as a special case in Proposition 3.1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the conditional-dependence coefficient whose population limit coincides with $\\eta_K^{\\mathrm{rank}}$ for $d_2=1$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced the univariate copula-based measure of regression dependence with properties (P1)-(P3)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Brenier-McCann theorem used to define population and empirical OT rank maps."},{"cited_title":"Drton, and F","cited_arxiv_id":null,"evidence_quote":"Supplies the multivariate rank methodology and algorithmic background for empirical OT ranks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the distance covariance kernel that yields the special-case coefficient in Proposition 3.1."}],"review_version":1}