{"id":"56a7c6d2-1f0d-4437-bc71-bc4e4d785a57","arxiv_id":"1908.03305","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The authors propose a Cramér-von Mises type test of independence built on the difference between joint and product recurrence rates, and prove its asymptotic behavior and consistency.","lead":"The paper introduces a new statistical test for independence between two random elements based on recurrence rates, the frequency with which pairs of sample points fall within a distance threshold. The test shows competitive or higher power than popular tests like distance covariance in many simulations, and it comes with proofs of its asymptotic null distribution and consistency.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 2's bound on H_n is algebraically wrong: the I4 coefficient is negative, so a concrete sample gives H_n<0; hence the proof of Theorem 3 is invalid as written.","rationale":"The reader's weakest assumption (Remark 3's distance-independence limitation) is a genuine and acknowledged scope restriction, and I agree it justifies caution. But the single most load-bearing technical problem is the false inequality in Lemma 2: it underpins the weak-convergence theorem for T_n under H0, and the mistake is visible in the algebra. The issue is repairable, so a conditional acceptance is appropriate rather than rejection. If the corrected bound fails, the claimed asymptotic null distribution would be unsupported, which would be more serious.","tokens_in":20674,"tokens_out":33384,"duration_ms":348634,"concrete_test":"Independently re-derive Lemma 2 algebraically (or run a 4-point simulation): with n=4 and only d(X1,X2)<r and d(Y3,Y4)<s, compute H_n; if H_n is negative, the displayed bound fails. Then check whether the corrected |H_n| bound is enough to conclude E_n-E'_n → 0 uniformly at rate 1/√n, which would determine whether Theorem 3's conclusion survives the fix.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Lemma 2 (Section 2.1) asserts E_n = E'_n - H_n with 0 ≤ H_n ≤ 4/√n, and Theorem 3 uses this to replace E_n by E'_n. The proof's decomposition of the double sum over ordered pairs is correct, but the coefficient of the I4 term is computed with the wrong sign: 1/N^2 - 1/(n)_4 = (1/N)(1/N - 1/((n-2)(n-3))) < 0 for n>2, whereas the display after (23) treats it as positive. A concrete configuration with n=4, d(X1,X2)<r, d(Y3,Y4)<s and no other recurrences gives RRX=RRY=1/6, RRXY=0, so E_n = 2(0-1/36)=-1/18; the I4 average in E'_n is 4/24=1/6, so E'_n=-1/3 and H_n=E'_n-E_n=-5/18<0. Thus the stated nonnegativity is false. The proof of Theorem 3, which is the basis for the null distribution of T_n, is therefore invalid as written; a corrected absolute bound |H_n|≤C/√n would likely restore the argument, but that correction is absent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new test of independence between two random elements X and Y taking values in metric spaces. The test statistic T_n is a Cramér-von Mises-type functional of the U-process E_n(r,s) = √n(RR^{X,Y}_n(r,s) − RR^X_n(r) RR^Y_n(s)), where the RR terms are empirical recurrence rates based on indicators 1{d(X_i,X_j)<r} and 1{d(Y_i,Y_j)<s}. By integrating the squared process against a distribution G on (0,∞)^2, the test avoids choosing a threshold radius. The paper claims: under H0, T_n converges weakly to the squared norm of a centered Gaussian process (Theorem 3); under contiguous alternatives the process acquires a deterministic drift (Proposition 1 and the subsequent assertion); and under alternatives for which the pair-distance variables d(X_1,X_2) and d(Y_1,Y_2) are dependent, T_n diverges to infinity in probability (Theorem 4). Simulation studies compare the test with HHG, distance covariance, and HSIC for scalar, vector, and time-series data.","tokens_in":2102,"tokens_out":1997,"duration_ms":144103,"significance":"If the theoretical claims are correct, the test is an attractive addition to the independence-testing literature: it applies to general metric spaces, avoids threshold selection by integrating over all radii, has a computationally explicit statistic (Section 3.4), and the reported power is competitive with or better than several popular tests in many of the studied alternatives. The paper also ships a permutation procedure for calibration, which is practically useful. However, the proof of the null-distribution theorem has a concrete algebraic error (Lemma 2) that as written invalidates the transfer from the approximating U-process to the actual process; the consistency proof omits a required uniform-integration step; and the contiguity result is only sketched. These are load-bearing issues, although all appear repairable without changing the overall approach. The structural limitation identified in Remark 3, that the test only detects dependence between the distance variables, is real and should be prominently acknowledged.","major_comments":[{"comment":"Lemma 2's assertion that 0 ≤ H_n(r,s) ≤ 4/√n is false. In the decomposition displayed after equation (23), the coefficient of the I_4 term is 1/N^2 − 1/(N(n−2)(n−3)), which is negative for every n>2; the manuscript's equation (24) treats this coefficient as positive. A concrete configuration with n=4, d(X_1,X_2)<r, d(Y_3,Y_4)<s, and no other recurrences gives RR^X_n=RR^Y_n=1/6, RR^{X,Y}_n=0, hence E_n = −1/18; the corresponding E'_n equals −1/3, so H_n = E'_n − E_n = −5/18 < 0. Thus the stated nonnegativity is false. Since Theorem 3 uses Lemma 2 to replace E_n by E'_n, the proof of Theorem 3 is invalid as written. The likely repair is to replace the nonnegativity claim with an absolute bound |H_n| ≤ C/√n, which follows from the same decomposition because all terms are bounded by constants times n^{−1/2}; but this correction is absent from the manuscript.","section":"Section 2.1, Lemma 2"},{"comment":"The consistency proof uses the inequality T_n ≥ (n/2)∫µ² dG − n∫(RR^{X,Y}_n − RR^X_n RR^Y_n − µ)² dG and then asserts that the second term is negligible as n→∞. No argument is given for this negligibility. The difference process involves products of U-statistics, and while pointwise convergence of RR^{X,Y}_n, RR^X_n, and RR^Y_n suggests the integrated squared difference should be o_P(1), proving this requires either a Donsker-type tightness result in L²(G) or a dominated-convergence argument for n∫(·)²dG. Without this step, the claim T_n →_P ∞ does not follow from the displayed inequality. The gap is load-bearing for the consistency result, though it appears fixable with a standard empirical-process argument.","section":"Section 2.1, Theorem 4 (proof)"},{"comment":"The weak convergence under H_n of {E_n(r,s)} to {E(r,s) + δµ(r,s)} is asserted after the sentence \"With a little more work, using the Le Cam third lemma,\" but no proof is provided. This is one of the three main theoretical claims of the paper, alongside Theorem 3 and Theorem 4. The assertion is not a trivial corollary: one must verify the contiguity conditions, handle the U-statistic remainder H_n under the triangular array of densities f^{(n)}_{X,Y}, and justify the weak convergence of the process in the space where the functional T_n is continuous. The omission of these details makes the contiguous-alternative result unverifiable as written. At minimum, a full proof or a precise reference with the required verification should be supplied.","section":"Section 2.2, contiguous alternatives"},{"comment":"In the \"Four independent clouds\" row, which is a null case (X and Y are independent), the column labeled N(1,1) reports power 0.512 for n=30 and n=50, while every other column in that row reports 0.046–0.057. At nominal level 0.05, a size of 0.512 is inconsistent with the paper's claim that the test has correct level under H0. Either there is a simulation error, a mislabeling, or a calibration problem specific to this weight function and sample size. The authors should explain or correct this entry; as printed, it undermines confidence in the numerical comparisons.","section":"Section 4.1, Tables 1 and 2"}],"minor_comments":[{"comment":"Theorem 3 as stated asserts weak convergence of {E_n(r,s) − E(E_n(r,s))} without assuming H0, but the proof explicitly begins with \"If H0 is true\". The theorem statement should include the independence assumption, since the claimed limit process and covariance structure are derived under H0.","section":"Section 2.1, Theorem 3 (statement)"},{"comment":"Remark 1 says the test statistic T_n is the norm of the Gaussian process, but T_n is defined with an un-squared distance from the definition: T_n = n∫(RR^{X,Y}_n − RR^X_n RR^Y_n)² dG, which converges to the squared norm ||E||². The remark should say \"squared norm\" or adjust the notation consistently.","section":"Section 2.1, Remark 1"},{"comment":"The intermediate algebra in the proof of Lemma 1 contains several apparent typos, such as \"p^{(3)}_X(r∧r′)p_Y(s)\" where the context suggests \"p^{(3)}_X(r∧r′)p_Y(s)p_Y(s′)\" and analogous expressions for p^{(3)}_Y. The final covariance formula (4) is plausible, but the intermediate display is difficult to follow and should be corrected.","section":"Section 2.1, Lemma 1 (proof)"},{"comment":"The unbiasedness argument for the permutation estimator is sketchy: the claim that each B_i is Binomial(m,1/n!) assumes that the values T_n under different permutations are distinct with probability one and that the permutation distribution is uniform over all n! permutations, but ties can occur and the double limit in n and m is not addressed. A more careful statement of the permutation test's exactness or asymptotic validity would improve the presentation.","section":"Section 3.2, permutation p-value"},{"comment":"The same letter φ is used both for the standard normal density and for the normal CDF in Section 3.1 (equations involving φ^{-1}(F_X(X))), which is confusing; the text should distinguish the density from the cumulative distribution function.","section":"General notation"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a genuinely useful problem and the test statistic is computationally attractive, but the theoretical core is not yet in a publishable state: Lemma 2's bound is demonstrably false, Theorem 4's proof has a missing convergence argument, and the contiguity result is only a sketch. These are repairable in principle, but each requires real work. I would also ask the editor to verify the simulation anomaly in the \"Four independent clouds\" rows of Tables 1 and 2 before the paper is further considered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the statistic is genuinely new—a Cramér–von Mises functional of a U-process built from recurrence rates—and the simulation study shows it is competitive with HHG, DCOV, and HSIC on several alternatives. The paper deserves a serious referee, but the referee should be sent back for a substantive revision: Lemma 2 is wrong as stated.\n\nI checked the stress-test note against the proof. The decomposition of the double sum in (23) is fine, but the coefficient of the four-distinct-index term in the expression for H_n is computed with the wrong sign. In fact the coefficient should be negative for n>2, and the numerical example with n=4 gives H_n = -5/18, contradicting the claimed 0 ≤ H_n ≤ 4/√n. Since Lemma 2 is the bridge between E_n and the U-process E'_n that feeds Theorem 3, the proof of the null distribution is invalid as written. This looks fixable: an absolute bound |H_n| ≤ C/√n would likely restore the argument, and the constant 4 is not the issue. But the current text cannot stand.\n\nOther soft spots are smaller. The contiguous-alternative limit is asserted after 'a little more work' without proof; Proposition 1 only gives the mean drift. The data-dependent choice of the weight G in Section 3.3 is outside the theory, which assumes a fixed G. And Table 2 has an obvious size anomaly: the 'Four independent clouds' row should be approximately 0.05 under the null, but the N(1,1) column reports 0.512. That is almost certainly a simulation or labeling error, but it needs an explanation.\n\nWhat the paper does well: the statistic is not a reparameterization of distance covariance or HHG; it uses all thresholds through integration, and the consistency theorem is correctly stated for alternatives where the distance variables are dependent. The authors are also honest in Remark 3 that their test cannot detect certain dependent pairs whose distances are independent—a real structural limitation, not a hidden one.\n\nBottom line: this is a serious paper with a fixable but load-bearing proof gap. I would not cite it as-is, but I would send it to a competent referee and ask for a revision that corrects Lemma 2, supplies the contiguous-alternative argument, and resolves the Table 2 anomaly.","headline":"A promising new independence test with strong simulations, but the proof of the null distribution has a concrete algebraic error that needs fixing before the results can be trusted.","tokens_in":124,"tokens_out":4656,"would_cite":false,"duration_ms":54485,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H15","62H20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A test that compares joint and marginal recurrence rates detects dependence between random elements whenever dependence leaves a trace in the pairwise distances.","keywords":["independence test","recurrence rates","U-process","Cramér-von Mises statistic","metric spaces","distance-based independence","contiguous alternatives","time series"],"falsifier":"Fix a dependent distribution with dependent pair-distances, for example bivariate normal with nonzero correlation, and simulate $T_n$ for n = 50, 100, 200, 400, 800 using a weight with positive density. The consistency theorem predicts $T_n \\to \\infty$ in probability; if the statistic stays stochastically bounded for such a distribution, Theorem 4 is wrong. A complementary check is to estimate the surface $\\mu(r,s)$ for any candidate alternative: if it is identically zero while X and Y are dependent, the theorem does not apply and the test should show no power.","tokens_in":20493,"feed_emoji":"📊","tokens_out":12803,"duration_ms":134848,"temperature":0.7,"pith_summary":"This paper proposes an independence test for random elements X and Y that uses only the distances between sample points, so it can be applied to variables, vectors, functions, and time series in any metric space. The test compares the joint recurrence rate, the fraction of pairs that are close in both coordinates, with the product of the two marginal recurrence rates, which are equal under independence. The statistic is a Cramér-von Mises functional that integrates the squared difference over all thresholds r and s with respect to a weight distribution, so no single radius has to be chosen. Under the null the statistic converges to the squared norm of a centered Gaussian process, and under alternatives for which the pair-distances $d(X_1,X_2)$ and $d(Y_1,Y_2)$ are not independent it diverges to infinity, making the test consistent. The paper also derives the limiting drift under contiguous alternatives and reports simulations where the test matches or outperforms several distance-based tests.","feed_headline":"New test catches dependence from pairwise distances alone","feed_subtitle":"It compares joint and marginal recurrence rates across all distance thresholds, so no radius needs tuning.","key_machinery":"The engine is the recurrence rate: for a threshold r, the fraction of pairs of observations whose distance is below r. The test looks at these rates for X, for Y, and for the pair (X,Y) simultaneously, and the key process is the normalized difference $E_n(r,s)$. To get asymptotic theory, the paper approximates $E_n$ by a four-index U-process $E'_n(r,s)$, a U-process being a normalized average of a kernel over all ordered 4-tuples of distinct sample indices, with the approximation error bounded by $4/\\sqrt{n}$. This approximation puts the process under a U-process weak-convergence theorem that yields the Gaussian null limit. The consistency argument is carried by the deterministic discrepancy function $\\mu(r,s)=P(d(X_1,X_2)<r,d(Y_1,Y_2)<s)-P(d(X_1,X_2)<r)P(d(Y_1,Y_2)<s)$: positivity of the weight density plus any point with $\\mu^2>0$ forces $T_n$ to diverge. The weight G, typically a product of Gaussian densities centered at the average distances, is what lets the statistic average over all thresholds instead of picking one.","core_discovery":"The central claim is that independence can be tested by contrasting the joint recurrence rate with the product of the marginal recurrence rates. With $E_n(r,s)=\\sqrt{n}(\\mathrm{RR}^{X,Y}_n(r,s)-\\mathrm{RR}^X_n(r)\\mathrm{RR}^Y_n(s))$, the proposed statistic is $T_n = n\\int_0^\\infty\\int_0^\\infty E_n(r,s)^2\\,dG(r,s)$. The paper proves two complementary limit statements. Theorem 3 shows that when X and Y are independent and the distributions of the distances are continuous, the process $\\{E_n(r,s)\\}$ converges weakly to a centered Gaussian process $\\{\\mathcal{E}(r,s)\\}$, so $T_n$ converges in distribution to $\\|\\mathcal{E}\\|^2_{L^2(dG)}$. Theorem 4 shows that if $d(X_1,X_2)$ and $d(Y_1,Y_2)$ are continuous and not independent, and the weight function has density positive everywhere, then $T_n \\to \\infty$ in probability; this is the consistency result, with multivariate normal pairs as a corollary. Under contiguous alternatives the limit process becomes $\\{\\mathcal{E}(r,s)+\\delta\\mu(r,s)\\}$, adding a deterministic drift that gives nontrivial local power. The paper notes that some dependent distributions produce independent pair-distances, and for those the consistency argument does not apply.","pith_inferences":["The theorem's condition is about the pair-distance variables, not about (X,Y) itself; effectively the test's null domain is 'the pair distances are independent,' so dependence that lives only in angles or absolute locations is invisible to the statistic.","An adaptive weight function could in principle concentrate mass where $\\mu(r,s)$ is large for a suspected alternative; the paper fixes G for the theory, so data-dependent G would need new arguments but is a natural practical extension.","For time series, the permutation scheme removes the contemporaneous X-Y pairing while preserving each series' internal order, suggesting the test can target cross-dependence even when the marginal series are serially dependent; the paper's asymptotic proofs are for i.i.d. pairs, so the time-series case rests on this permutation calibration plus simulation.","A quick diagnostic for a new application is to estimate $\\mu(r,s)$ before relying on the test: if the estimate is flat at zero while X and Y are believed dependent, the consistency theorem offers no power guarantee and another test should be considered."],"forward_implications":["If the paper is correct, any dependence between X and Y that makes the pair-distance variables dependent is eventually detected: $T_n$ grows without bound as $n$ increases.","Under $H_0$ the asymptotic null distribution is available, and for finite samples the paper's permutation scheme estimates p-values by breaking the X-Y pairing, so the test can be calibrated without simulating the limiting Gaussian process.","Because the statistic uses only distances, the same test applies to random vectors, function-valued data, and discrete or continuous time series; the multivariate normal case is covered by Corollary 1.","A sup-version $T'_n=\\sqrt{n}\\sup_{r,s>0}|E_n(r,s)|$ is allowed by the same arguments, giving a user an alternative that avoids choosing a weight function.","Under contiguous alternatives the limiting power is governed by the drift $\\delta\\mu(r,s)$; in the simulation study the test shows higher power than several distance-based competitors in most of the settings considered."],"supporting_citations":[{"why":"Supplies the U-process weak convergence theorem that gives the Gaussian limit of the test process under H0.","marker":"[1]"},{"why":"Gives the strong law for U-statistics that makes the recurrence rates converge almost surely to their probabilities.","marker":"[14]"},{"why":"Introduces recurrence plots, the source of the recurrence-rate idea the test is built on.","marker":"[9]"},{"why":"Provides the HHG test used as one of the main simulation baselines in the power study.","marker":"[13]"},{"why":"Provides distance covariance, the DCOV baseline the new test is compared against.","marker":"[25]"},{"why":"Provides the HSIC baseline used in the power comparisons.","marker":"[12]"},{"why":"Establishes the contiguity conditions used to derive the limiting drift under contiguous alternatives.","marker":"[8]"}],"fun_headline_variants":["No tuning needed: independence test from recurrence rates","All thresholds at once: recurrence-rate independence test","Tuning-free independence test via joint vs marginal recurrence","Independence test from pairwise distances, no radius to set"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"All the consistency guarantees rest on the assumption that dependence between X and Y shows up as dependence between the pairwise distances $d(X_1,X_2)$ and $d(Y_1,Y_2)$; the paper itself exhibits dependent variables whose distances are independent, and for those distributions the test has no asymptotic power.","fun_headline_variants_meta":{"raw":{"variants":["No tuning needed: independence test from recurrence rates","All thresholds at once: recurrence-rate independence test","Tuning-free independence test via joint vs marginal recurrence","Independence test from pairwise distances, no radius to set"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1645,"prompt_tokens":929,"completion_tokens":716,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":654}},"tokens_in":545,"tokens_out":716,"duration_ms":7437,"temperature":1.0,"reasoning_tokens":654,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:17:58.531011+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a dependent distribution with dependent pair-distances, for example bivariate normal with nonzero correlation, and simulate $T_n$ for n = 50, 100, 200, 400, 800 using a weight with positive density. The consistency theorem predicts $T_n \\to \\infty$ in probability; if the statistic stays stochastically bounded for such a distribution, Theorem 4 is wrong. A complementary check is to estimate the surface $\\mu(r,s)$ for any candidate alternative: if it is identically zero while X and Y are dependent, the theorem does not apply and the test should show no power.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the U-process weak convergence theorem that gives the Gaussian limit of the test process under H0."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the strong law for U-statistics that makes the recurrence rates converge almost surely to their probabilities."},{"cited_title":"P, Oliffson Kamphorst, S, Ruelle, D","cited_arxiv_id":null,"evidence_quote":"Introduces recurrence plots, the source of the recurrence-rate idea the test is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HHG test used as one of the main simulation baselines in the power study."},{"cited_title":"J., Rizzo, M","cited_arxiv_id":null,"evidence_quote":"Provides distance covariance, the DCOV baseline the new test is compared against."},{"cited_title":"and Sch¨ olkopf, B","cited_arxiv_id":null,"evidence_quote":"Provides the HSIC baseline used in the power comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the contiguity conditions used to derive the limiting drift under contiguous alternatives."}],"review_version":1}