{"id":"314711a6-70fd-4969-9cd6-30e595e790ef","arxiv_id":"1908.02634","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A temporally smoothed wavelet periodogram for multivariate point processes is shown to be a multi-wavelet periodogram, asymptotically Wishart, and used to build a stationarity test.","lead":"This paper develops a wavelet-based spectral estimator for multivariate point processes, such as neural spike trains, and derives its asymptotic distribution. The framework enables tests for non-stationarity and time-scale measures of coherence between event streams.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Weighted eigen-wavelet sum is not Wishart: Theorem 2's 'asymptotic' law is an equivalent-dof approximation, not a limiting distribution.","rationale":"I read the paper's central claim as the asymptotic Wishart calibration of the temporally smoothed wavelet periodogram, which drives the coherence density and the stationarity test. For that claim to hold, the eigen-wavelet expansion must turn a weighted sum of asymptotically independent complex Wishart(1) matrices into a complex Wishart distribution. It does not: only equal weights give an exact Wishart, and the weights here are fixed eigenvalues of the smoothing kernel. Matching first two moments gives the reported n but not the full law, so Theorem 2 is at best a Satterthwaite approximation. This is an internal argument gap, not a disagreement with external consensus, and it is checkable by a one-dimensional distribution comparison. The reader's Assumption 4 typo is genuine and affects circularity, but fixing it would not repair the weighted-sum step. The contribution may be salvageable by reframing Theorem 2 and Corollary 1 as equivalent-degree-of-freedom approximations supported by error analysis and the existing simulations, so I keep the conditional verdict but with a different mandatory revision.","tokens_in":21444,"tokens_out":11980,"duration_ms":144462,"concrete_test":"For the Morlet kernel with κ=20, compute the leading M eigenweights η_1,...,η_M (M large enough that the tail of Ση² is negligible). Let X_l be iid Exp(1) and compare the distribution of S=Σ η_l X_l with the claimed (1/n)Gamma(n,1), where n=1/Ση_l². Report the 0.95 upper-tail quantile of each and the probability assigned to S exceeding the claimed-law quantile. If the relative quantile error or tail-probability mismatch exceeds a few percent, Theorem 2's 'asymptotically Wishart' statement is refuted; the same conclusion then applies to the coherence density in Corollary 1. A complementary finite-T bivariate Poisson simulation with large T can confirm which law the empirical coherence quantiles track.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The Assumption 4 typo is real but secondary. Even after correcting it to the intended condition ∫ψ(t)²dt=0, the proof of Theorem 2 does not establish the stated Wishart limit. The argument uses eigen-wavelet expansion Ω = Σ_l η_l v_l v_l^H. If the v_l are asymptotically independent CN_p(0,S), then each v_l v_l^H is WC_p(1,S), but a weighted sum of independent complex Wisharts is a complex Wishart only when all weights are equal. Matching mean and covariance fixes n = 1/Ση_l², which is the Satterthwaite/Box equivalent-degrees-of-freedom approximation, not an exact asymptotic distribution. Because the eigenvalues η_l are fixed as T→∞, the discrepancy between the weighted-sum law and the claimed (1/n)WC_p(n,S) law does not vanish. For example, with p=1 and weights (0.9, 0.1), the MGF of Ση_l Exp(1) at t=1 is 11.1, whereas the claimed (1/n)χ²_{2n} MGF is about 8.0. Corollary 1's coherence density inherits this mismatch. The citation to Walden (2000, p.776) covers unweighted multitaper averaging; unequal eigenvalues require either a different argument or an explicit approximation statement with error control.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a temporally smoothed wavelet periodogram for multivariate point processes, expresses it as a weighted sum of eigen-wavelet periodograms, and derives asymptotic distributional results. Under stationarity, Theorem 1 claims the wavelet transform is asymptotically circular complex normal, and Theorem 2 claims the smoothed periodogram is asymptotically a scaled complex Wishart with degrees of freedom n = 1/Σ_l η_l² determined by the eigenvalues of the smoothing kernel. From Theorem 2 the authors derive the distribution of wavelet coherence (Corollary 1) and construct a likelihood-ratio test for stationarity across scales (Theorem 3). The paper includes simulation support, a real neural spike-train analysis, and a GitHub implementation. The main novelty is the eigen-wavelet representation and the attempt to extend multitaper-style Wishart asymptotics to point-process wavelet spectra.","tokens_in":21719,"tokens_out":6156,"duration_ms":70642,"significance":"If Theorem 2 and its corollaries were correct, the paper would provide a practically useful and theoretically motivated inferential framework for wavelet coherence and stationarity testing in point-process data, extending earlier work by Cohen and Walden from time series to point processes. The eigen-wavelet representation in Section 3 is elegant and Proposition 2 is a clean and useful result. The authors also provide simulations, a real-data example, and reproducible code, which are strengths. However, the central distributional claim is not valid as stated: the weighted sum of independent complex Wisharts that arises in the eigen-wavelet representation is not itself a scaled complex Wishart when the eigenvalue weights are unequal, and the weights are fixed as T→∞. Since Corollary 1 and Theorem 3 inherit this claim, the paper's main inferential results do not follow. The manuscript is therefore not publishable in its current form, although the underlying ideas may be salvageable if reframed as an approximation with explicit error control.","major_comments":[{"comment":"The proof of Theorem 2 does not establish the stated Wishart limit. From Eq. (S5) and Lemma 3, the asymptotics give Ω(T)(ã,b̃) ≈ Σ_l η_l v_l v_l^H with the v_l asymptotically independent CN_p(0,S). Each term v_l v_l^H is WC_p(1,S), but a weighted sum of independent complex Wisharts with unequal weights is not a scaled WC_p(n,S); a WC_p(n,S) distribution corresponds to an unweighted average of n i.i.d. rank-one terms. Choosing n = 1/Σ_l η_l² matches the mean and a certain covariance adjustment, but this is the Satterthwaite/Box equivalent-degrees-of-freedom approximation, not an exact asymptotic distribution. The mismatch does not vanish as T→∞ because the eigenvalues η_l are fixed. For example, for p=1 with weights (0.9,0.1), the MGF of Σ η_l Exp(1) at t=1 is 1/{(1−0.9)(1−0.1)} ≈ 11.1, while the claimed (1/n)WC_p(n,S) MGF is (1 − (1/n)t)^{-n} ≈ 8.0 at t=1. The citation to Walden (2000, p. 776) concerns unweighted multitaper averaging and does not cover unequal eigenvalues. This issue directly invalidates Corollary 1 and the distributional statements used in Theorem 3 and Proposition 7.","section":"Section 4.2, Theorem 2 and its proof in Supplementary Material Section 1"},{"comment":"Assumption 4 as printed is internally inconsistent: it states that ψ is orthogonal to its complex conjugate, i.e. ∫ ψ(t)ψ*(t)dt = 0, but Assumption 1 requires ∫|ψ(t)|²dt = 1, so this condition is impossible. The proofs actually require the different condition ∫ ψ(t)ψ(t)dt = 0; see Lemma 2, where ψ1 = ψ*(T) and ψ2 = ψ(T) are used to conclude ∫ ψ*(T)(t)ψ*(T)(t)dt = 0. This stronger condition is not satisfied exactly by the Morlet wavelet used throughout the paper. If it fails, cov{w(T), w*(T)} does not vanish asymptotically, so the limit is not circular complex normal and the complex Wishart claim in Theorem 2 does not follow. The authors should either restrict to wavelets for which the intended condition holds exactly, or quantify the effect of its violation on the limiting distribution.","section":"Section 4.1, Assumption 4; also Lemma 2 and Theorem 1 proof in Supplementary Material"},{"comment":"The statement that the eigen-wavelet coefficients v_l are asymptotically independent is only sketched. Lemma 2 bounds a second-order cumulant between two wavelet transforms with orthogonal wavelets, but asymptotic joint normality of the infinite family {v_l} requires control of all higher-order joint cumulants, or a suitable infinite-dimensional CLT. The phrase 'Lemma 2 combined with a similar argument to Theorem 1' does not supply this argument. This matters for Theorem 2 because the weighted-sum law in the preceding comment presupposes independence of the v_l; it also matters for Proposition 5, where independence of the periodograms at different dyadic locations is asserted from non-overlap of kernels. An explicit proof, or a reference that covers joint asymptotic normality of wavelet coefficients in point-process settings, is needed.","section":"Theorem 2 proof and Proposition 5 proof, Supplementary Material Section 1"}],"minor_comments":[{"comment":"The main text states 0 < c < 1/2, while the supplementary statement of Theorem 3 says 0 < c < 1. Since β = min{c, 1/2 − c} is used, the condition c < 1/2 is required; the supplementary statement should be corrected for consistency.","section":"Main text Theorem 3 vs. Supplementary Material Theorem 3"},{"comment":"In the proof of Proposition 1, the sentence 'Property (iii) is immediate from Proposition 1 and the fact that ψ(t) itself obeys the admissibility condition' appears to contain a typo: it should cite Proposition 2, not Proposition 1.","section":"Supplementary Material, proof of Proposition 1"},{"comment":"The paper states that ψ(t) is used to denote both the original and the approximating wavelet with support (−α/2, α/2). Since several theoretical results rely on the approximating support, the dependence of the asymptotics on α should be made explicit in the statements of Proposition 4 and Theorem 2, rather than only in the informal discussion.","section":"Section 2.2, approximating wavelet"}],"recommendation":"reject","confidential_remarks":"The central theorem is not just insufficiently proved but false as stated: the weighted eigen-wavelet sum does not converge to the claimed Wishart distribution. I would encourage the authors to reframe the contribution as an equivalent-degrees-of-freedom approximation with a quantifiable error, or to derive the exact asymptotic distribution of the weighted sum and base the coherence and stationarity tests on that distribution. The paper has strong expository qualities and a useful computational toolkit, so a substantially revised version could be viable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely nice. Rewriting the temporally smoothed wavelet periodogram as a weighted sum of eigen-wavelet periodograms (S5) is elegant, and it opens a practical route to computation via the Nyström method. The paper is also honest about the machinery: the proofs rely on Brillinger's cumulant conditions, the simulations are serious, and there is code. That is real value.\n\nThe problem is Theorem 2. The authors claim the smoothed periodogram is asymptotically (1/n)WC_p(n, S) with n = 1/Ση_l². The proof says this follows from asymptotic independence of the eigen-wavelet coefficients, in analogy to Walden (2000). But it doesn't. For independent v_l ~ CN(0,S), each v_l v_l^H is WC_p(1,S), but Ση_l v_l v_l^H is a weighted sum of Wisharts. A weighted sum of independent Wisharts is only Wishart when all weights are equal. Matching the mean and covariance fixes n = 1/Ση_l², but that is Satterthwaite-style moment matching, not an asymptotic law. The η_l are fixed by the kernel as T→∞, so the discrepancy between the true distribution of the weighted sum and the claimed Wishart does not vanish. The stated asymptotic distribution is therefore not a limit. Corollary 1's coherence density and the likelihood-ratio stationarity test inherit this mismatch.\n\nThis is not a nitpick. The Wishart result is load-bearing for the testing framework. The simulations still look reasonable, because the eigenvalues decay quickly and the moment-matched approximation can be decent, but that doesn't make the theorem correct as written.\n\nThe other issues are smaller. Assumption 4 as printed requires ∫ψψ* = 0, which is impossible since ||ψ||=1; the intended condition is presumably ∫ψψ = 0. Easy typo, but it appears in the statement of a key assumption. The proof of Theorem 2 is also a sketch: the independence step is asserted through Lemma 2 and a hand-wave to Walden. And calling p=0.032 at one of several scales \"strong evidence\" ignores the multiple comparisons; minor but should be softened.\n\nWho is this for? People analyzing multivariate event data, especially neural spike trains, will find the framework useful in principle. The paper deserves a serious referee, but not because the theorem is correct as stated. It deserves one because the framework is good enough to fix. The authors should be pushed to either prove a real limit (which I doubt is possible) or, much more realistically, present the Wishart law as an explicit approximation with error bounds, and recalibrate the stationarity test and coherence thresholds accordingly. That is a substantial revision, but the core contribution can survive it. Send it to peer review, but the referee should insist on this change before acceptance.","headline":"The eigen-wavelet representation is a real advance, but the asymptotic Wishart result in Theorem 2 is a moment-matched approximation, not a genuine limiting distribution.","tokens_in":830,"tokens_out":924,"would_cite":false,"duration_ms":43431,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M15","60G55","42C40","62H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"A temporally smoothed wavelet periodogram of a multivariate point process is asymptotically a scaled complex Wishart matrix, giving closed-form coherence distributions and a chi-square test for non-stationarity.","keywords":["wavelet periodogram","multivariate point process","Hawkes process","multitaper","Wishart distribution","wavelet coherence","stationarity test","neural spike train"],"falsifier":"Run a stationary bivariate Hawkes or Poisson process through the estimator and compare the empirical distribution of the smoothed periodogram, or the coherence, against the claimed $(1/n)\\mathrm{WC}_p$ law with $n$ taken from the kernel eigenvalues; a statistically significant deviation, or a failure of the empirical covariance of $w^{(T)}$ and its conjugate to decay to zero as $T$ grows, would refute Theorem 2.","tokens_in":21194,"feed_emoji":"🧠","tokens_out":9853,"duration_ms":98587,"temperature":0.7,"pith_summary":"The paper tries to put wavelet spectral analysis of multivariate point processes on the same statistical footing that Fourier and multitaper methods enjoy for stationary time series. It shows that a wavelet periodogram of event data, smoothed across time with a kernel, is exactly a weighted average of periodograms formed from an orthonormal eigen-wavelet system, and that under stationarity this average is asymptotically a scaled complex Wishart matrix. That distribution gives closed-form thresholds for wavelet coherence and a likelihood-ratio test, with $\\chi^2$ asymptotics, for whether the process is stationary at a given scale. The development matters because it makes statements about time-varying correlation in neural spike trains testable from a single trial.","feed_headline":"Spike-train wavelet spectra: Wishart law and χ² stationarity test","feed_subtitle":"Smoothing makes the periodogram a multi-wavelet average, giving closed-form coherence and stationarity tests.","key_machinery":"The load-bearing object is the Hermitian kernel $K(s,t)=\\int h_\\kappa(u)\\psi(s-u)\\psi^*(t-u)\\,du$ formed from the wavelet $\\psi$ and smoothing window $h$, together with its Mercer eigen-decomposition into eigen-wavelets $\\{\\phi_l\\}$ and eigenvalues $\\{\\eta_l\\}$. The identity $\\sum_l \\eta_l |\\Phi_l(f)|^2 = |\\Psi(f)|^2$ preserves the wavelet's frequency response across the channel bank. This machinery converts temporal smoothing into a finite, in practice, number of approximately independent wavelet channels, so that the smoothed periodogram becomes a weighted sum of Wishart matrices; the weights $\\eta_l$ directly set the effective degrees of freedom $n=1/\\sum_l \\eta_l^2$.","core_discovery":"On the paper's own terms, the central discovery is a distributional equivalence: for a $p$-dimensional stationary point process with spectral density matrix $S(f)$, the temporally smoothed wavelet periodogram $\\Omega^{(T)}(\\tilde a,\\tilde b)$ is asymptotically $(1/n)\\,\\mathrm{WC}_p\\{n, S(f_{\\tilde a})\\}$ as $T\\to\\infty$, where $n=1/\\sum_l \\eta_l^2$ is computed from the eigenvalues of the kernel $K(s,t)=\\int h_\\kappa(u)\\psi(s-u)\\psi^*(t-u)\\,du$. The proof runs through the Mercer expansion $K(s,t)=\\sum_l \\eta_l \\phi_l(s)\\phi_l^*(t)$, which makes the smoothed periodogram a weighted sum of eigen-wavelet periodograms, each asymptotically complex normal by a cumulant argument under the paper's mixing condition. The Wishart law then yields a closed-form density for temporally smoothed wavelet coherence (Corollary 1) and, through a likelihood-ratio construction, an asymptotic $\\chi^2$ test for stationarity at dyadic scales (Theorem 3).","pith_inferences":["Extension: the identity $\\Omega=\\sum_l \\eta_l v_l v_l^H$ suggests that the same $n=1/\\sum_l\\eta_l^2$ calibration should apply to any non-negative definite, trace-one smoothing kernel, not only the rectangular-window examples; testing a different $h$ would be a direct check.","Extension: the paper tests each dyadic scale separately, but the rejection information across scales could be combined into a change-point localizer for when the dependence structure changes, something the authors do not attempt.","Extension: because the Morlet wavelet is not exactly analytic, the circular-complex-normal step holds only approximately; a finite-sample study of how far the Wishart approximation drifts with the size of $\\int\\psi(t)^2\\,dt$ would quantify the practical gap."],"forward_implications":["Wavelet coherence of point processes can be thresholded without bootstrap or permutation: the 95% contour follows from Corollary 1 with $n$ taken from the kernel eigenvalues.","The likelihood-ratio statistic gives a scale-by-scale stationarity test with known $\\chi^2$ degrees of freedom, and the scale-specific tests combine into a global test.","The degrees of freedom grow linearly with the smoothing width $\\kappa$ for a rectangular window (Proposition 4), so users can trade time resolution against test power explicitly.","Real-valued wavelets such as the Mexican hat inherit the framework with a real Wishart limit and a Beta null coherence distribution, so the same calibration applies there.","In applications to neural spike trains, the test can flag changes in cross-stream correlation over short single-trial recordings."],"supporting_citations":[{"why":"Provides the mixing and cumulant conditions, and the cumulant lemma, that give the asymptotic complex normality of the wavelet transform.","marker":"Brillinger (1972)"},{"why":"Supplies the eigenfunction expansion that represents the smoothed periodogram as a weighted multi-wavelet average.","marker":"Mercer (1909)"},{"why":"Gives the multivariate multitaper argument by which the weighted sum of independent Wishart matrices yields a scaled Wishart limit.","marker":"Walden (2000)"},{"why":"Gives the distribution of coherence implied by a complex Wishart matrix, used in Corollary 1.","marker":"Goodman (1963)"},{"why":"Defines the spectral density matrix of mutually exciting point processes used as the key simulation and application model.","marker":"Hawkes (1971)"},{"why":"Establishes the spectral density matrix of a point process as the Fourier transform of its covariance density, the target of estimation.","marker":"Bartlett (1963)"},{"why":"Defines the continuous wavelet transform of a point process as a stochastic integral over event times.","marker":"Brillinger (1996)"}],"fun_headline_variants":["Smoothed wavelet periodogram: asymptotically Wishart for point processes","Multivariate point-process wavelet spectra: Wishart and chi-square tests","Detecting non-stationarity in point processes via smoothed wavelet spectra","Closed-form wavelet coherence and stationarity test for spike trains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument's load-bearing premise is that the complex wavelet used is analytic, meaning it has no response at negative frequencies; the paper states this as orthogonality to its own complex conjugate, a condition that cannot hold for a unit-norm wavelet, and the Morlet wavelet used in the applications only satisfies the intended condition approximately.","fun_headline_variants_meta":{"raw":{"variants":["Smoothed wavelet periodogram: asymptotically Wishart for point processes","Multivariate point-process wavelet spectra: Wishart and chi-square tests","Detecting non-stationarity in point processes via smoothed wavelet spectra","Closed-form wavelet coherence and stationarity test for spike trains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001002,"raw_usage":{"total_tokens":4225,"prompt_tokens":919,"completion_tokens":3306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":3232}},"tokens_in":535,"tokens_out":3306,"duration_ms":26324,"temperature":1.0,"reasoning_tokens":3232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:38:57.480587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a stationary bivariate Hawkes or Poisson process through the estimator and compare the empirical distribution of the smoothed periodogram, or the coherence, against the claimed $(1/n)\\mathrm{WC}_p$ law with $n$ taken from the kernel eigenvalues; a statistically significant deviation, or a failure of the empirical covariance of $w^{(T)}$ and its conjugate to decay to zero as $T$ grows, would refute Theorem 2.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the mixing and cumulant conditions, and the cumulant lemma, that give the asymptotic complex normality of the wavelet transform."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the eigenfunction expansion that represents the smoothed periodogram as a weighted multi-wavelet average."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the multivariate multitaper argument by which the weighted sum of independent Wishart matrices yields a scaled Wishart limit."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the distribution of coherence implied by a complex Wishart matrix, used in Corollary 1."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the spectral density matrix of mutually exciting point processes used as the key simulation and application model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the spectral density matrix of a point process as the Fourier transform of its covariance density, the target of estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the continuous wavelet transform of a point process as a stochastic integral over event times."}],"review_version":1}