{"id":"fed108f2-44bc-40dd-a1cf-abf07ff743b7","arxiv_id":"2411.14010","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Finite-sample posterior accuracy of local Whittle likelihoods in time-varying AR models is assessed; all three bias corrections help, with dynamic Whittle slightly ahead of block Whittle.","lead":"This paper tests how well fast frequency-domain likelihood approximations reproduce the exact Bayesian posterior in time-varying autoregressive models. It finds that standard local Whittle approximations are improved by tapering, prewhitening, or boundary correction, but that tapering can distort variance estimates unless rescaled.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (36) states the Q full conditional as IW(S0+φ_{1:M}^Tφ_{1:M}), but model (31) evolves θ_j=θ_{j-1}+η_j, so the sufficient statistic should be Σ(θ_j−θ_{j-1})(θ_j−θ_{j-1})^T. If implemented as written, the gold-standard posterior and all perturbation comparisons are invalid.","rationale":"The reader's weakest-assumption concern about missing PGAS diagnostics is reasonable, but it is not the most load-bearing issue. The Eq. (36) full conditional is a checkable analytic internal inconsistency: if the implemented sampler follows the paper, the time-domain 'gold standard' is not the target posterior even in the limit of infinite PGAS draws, and the perturbation measure (11) is contaminated for every method and experiment. This would directly undermine the paper's strongest claims that the dynamic Whittle is competitive with the time-domain posterior and that tapering, prewhitening, and boundary correction improve accuracy. There are independent positives: the comparison design is sensible, three simulation experiments cover relevant regimes, and the reported effects are internally coherent. The missing code and data, plus the apparent typo in Eq. (28), mean the issue cannot be resolved from the manuscript alone. I therefore keep the reader's CONDITIONAL verdict: the paper is potentially a solid contribution, but acceptance should require code or a correction confirming that the Q update in the time-domain sampler uses squared θ increments. If the typo is confirmed and the code is correct, the central conclusions may stand; if not, the simulation evidence would need to be redone.","tokens_in":19740,"tokens_out":8252,"duration_ms":86435,"concrete_test":"Obtain the code, or re-implement the Gibbs sampler in Section 4.3, and inspect the update for Q in the time-domain posterior. Verify that the scale matrix is S0 + Σ_{j=1}^M(θ_j−θ_{j-1})(θ_j−θ_{j-1})^T. If instead it matches Eq. (36), rerun Experiment 1 with the corrected update and recompute the perturbation measure (11) for the raw and boundary-corrected dynamic Whittle likelihoods; if the rankings or magnitudes change materially, the conclusions in Section 5 are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an empirical ranking of spectral likelihood approximations against the time-domain posterior, so that comparison is only as good as the time-domain sampler. In Section 4.3, the full conditional for the state-evolution covariance Q is stated as Q | φ_{0:M}, σ², I ∼ Inv-Wishart(M+ν0, S0+φ_{1:M}^Tφ_{1:M}). This is internally inconsistent with the model in (31): the state equation is θ_j = θ_{j-1}+η_j, with η_j∼N(0,Q) and φ_j=g(θ_j). Conditionally on the latent path θ_{0:M}, the data are independent of Q, and Q enters only through the Gaussian increments η_j=θ_j−θ_{j-1}. The correct full conditional is therefore Inv-Wishart(M+ν0, S0+Σ_{j=1}^M(θ_j−θ_{j-1})(θ_j−θ_{j-1})^T). Using levels of the AR coefficients, or omitting the increments altogether, would bias Q and thereby distort the θ trajectories sampled in the time-domain posterior. Since no code is provided, the reader cannot tell whether (36) is a typographical error or the implemented update. This concern is independent of the PGAS convergence issue: even with infinitely many particle draws, the sampler would target the wrong stationary distribution if (36) is used. The boundary-correction periodogram in (28) also appears to be a typo, which further underlines the need for the code before the simulation results can be taken at face value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper compares two spectral-domain approximate likelihoods, the block Whittle and the dynamic Whittle, for Bayesian inference in time-varying autoregressive (tvAR) models. The accuracy of the approximate posteriors is measured relative to the exact time-domain posterior using a relative RMSE efficiency criterion and a quantile-based perturbation measure. Three modifications (tapering, prewhitening, boundary correction) are tested in three simulation experiments and in an application to weekly egg prices with stochastic volatility. The reported findings are that both approximate likelihoods perform reasonably, the dynamic Whittle has a slight edge, and all three modifications improve accuracy, with caveats about tapering-induced variance distortions.","tokens_in":20052,"tokens_out":8728,"duration_ms":87093,"significance":"The question addressed is relevant: frequency-domain likelihood approximations are widely used for long nonstationary time series, and most existing theory concerns the MLE rather than the posterior. The simulation design is sensible, with several DGP regimes, deliberately untuned priors, and comparisons against a time-domain gold-standard posterior. The paper is also explicit about limitations, such as the rescaling step for tapered segments and the inflated variance from boundary correction in Section 6. If the numerical results are correct, the paper offers useful practical guidance on when local Whittle posteriors can be trusted and which corrections help. The main caveats are that the reference posterior is itself a particle-MCMC approximation, that two key equations appear to contain errors, and that no code or Monte Carlo uncertainty measures are provided, so the quantitative rankings cannot currently be verified.","major_comments":[{"comment":"The stated full conditional for Q is inconsistent with the state equation in (31). Under theta_j = theta_{j-1} + eta_j with eta_j ~ N(0,Q), conditional on theta_{0:M}, Q appears only through the increments, so the sufficient statistic is sum_{j=1}^M (theta_j - theta_{j-1})(theta_j - theta_{j-1})^T, not phi_{1:M}^T phi_{1:M}. If Algorithm 1 uses Eq. (36) as written, the PGAS sampler targets the wrong posterior, and because this posterior is the gold standard in the perturbation measure (11), all main comparisons are invalid. Since no code is provided, it is impossible to tell whether this is a typo or the implemented update; this must be resolved and the affected experiments rerun or explicitly justified.","section":"Section 4.3, Eq. (36)"},{"comment":"The complete periodogram is defined as tilde I_T(omega) = (1/(2 pi T)) hat J(omega; f) J(omega). This is not the periodogram of the complete DFT tilde J(omega) = J(omega) + hat J(omega; f) defined in (25); the squared modulus with the appropriate conjugation and cross-terms is missing. The boundary-correction results in Sections 5 and 6 rely on this quantity, so if Eq. (28) is implemented literally, the BC results are not valid. Please correct the formula and clarify the exact estimator used in the simulations.","section":"Section 3.3, Eq. (28)"},{"comment":"The gold-standard time-domain posterior is produced by PGAS, but no details of the particle filter are given: number of particles, resampling scheme, effective sample size, or convergence diagnostics for the 100 replications. The perturbation measure treats this posterior as exact; a biased or unconverged particle approximation would contaminate every reported comparison. Please report these diagnostics or otherwise validate the reference sampler, for example by comparing runs with different particle counts.","section":"Sections 4.2-4.3, Algorithm 1 and Eq. (11)"},{"comment":"The efficiency and perturbation values are reported as point estimates over 100 datasets with no measure of Monte Carlo uncertainty. Several conclusions in the text are rankings between close values, for example dynamic versus block Whittle and TA versus PW versus BC; without standard errors, confidence intervals, or per-replicate distributions, it is unclear whether these differences are meaningful. Please add uncertainty quantification to the reported comparisons.","section":"Section 5, Figures 5, 8, 10"}],"minor_comments":[{"comment":"The symbol phi is used both for the stable AR parameters (phi_j = g(theta_j)) and for the lag polynomial phi_p(L); in Eq. (36) the notation phi_{1:M} is also undefined as a matrix. Please disambiguate these uses.","section":"Notation throughout"},{"comment":"The manuscript does not state whether code and data are available. Given the simulation-based nature of the paper and the apparent errors in Eqs. (28) and (36), providing the code would be essential for reproducibility and for verifying that the reported results are not artifacts of the implementation.","section":"Reproducibility"},{"comment":"The text repeatedly refers to 'this chapter'; if the manuscript is intended as a journal submission, please use consistent wording such as 'this paper' or 'this article'.","section":"Abstract and Section 1"},{"comment":"The time-domain likelihood used as the gold standard for the stochastic volatility extension is not fully specified; please state the complete observation and state model, including how the time-varying variance enters the likelihood, so that the comparison is unambiguous.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the overall empirical strategy is appropriate. The main concern is verification: two load-bearing equations appear to contain typos, and the gold-standard posterior is not diagnosed. These issues are fixable with code, corrected equations, and additional diagnostics, so I would not reject at this stage, but the current manuscript cannot be accepted as is."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-designed finite-sample comparison of spectral likelihoods for Bayesian tvAR inference, with a genuinely useful ranking, but the version I read has a likely error in the Gibbs update for Q that, if real, poisons the gold-standard comparison, and there is no code to check.\n\nThe genuinely new bits are the direct comparison of block vs dynamic Whittle posterior accuracy (earlier work focused on MLE asymptotics), the finding that all three modifications—tapering, prewhitening, boundary correction—meaningfully improve both likelihoods, and the caution that tapering can distort stochastic volatility inference on short segments. The simulation design is sensible: three DGPs covering a smooth tvAR, a stationary AR, and a near-unit-root path, with a quantile-based perturbation measure against the exact time-domain posterior. That part deserves credit.\n\nThe soft spots are fixable but real. Equation (36) states the full conditional for Q as Inv-Wishart(M+ν0, S0 + φ_{1:M}^T φ_{1:M}), yet the model in (31) puts the random walk on θ, not on φ=g(θ). The sufficient statistic should be a sum of squared increments of θ. If that equation reflects the implementation, the Gibbs sampler targets the wrong stationary distribution and every perturbation comparison against the 'gold standard' is contaminated. If it's a typo, it's a confusing one, and the absence of code makes it impossible to tell. The complete periodogram in (28) also looks wrong—it should involve the modulus of the complete DFT, not the product of the predictive and ordinary DFTs. The paper reports no Monte Carlo error bars on the efficiency and perturbation measures, no PGAS diagnostics (particle count, ESS), and no code or data. The decision to exclude the debiased and generalized Whittle likelihoods is defensible, but it limits the scope; the reader cannot verify the rankings for the excluded methods.\n\nOn balance, the paper deserves serious peer review: the question is important, the experimental design is careful, and most conclusions are plausible. But I would not take the numerical results at face value until the authors correct or clarify the Q update, fix (28), and share the code. The central empirical claims could stand, but the referee should ask for those before publication.","headline":"Useful and well-designed simulation study, but a likely bug in the Gibbs update for Q and a missing code release keep me from trusting the gold-standard comparison.","tokens_in":20613,"tokens_out":4116,"would_cite":false,"duration_ms":37243,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62M15","62F15","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Fast frequency-domain Whittle-type likelihoods can closely reproduce exact Bayesian posteriors for time-varying autoregressive models, with the dynamic Whittle variant performing best.","keywords":["Whittle likelihood","frequency domain likelihood","locally stationary processes","time-varying autoregressive models","Bayesian inference","particle MCMC","periodogram","spectral density"],"falsifier":"Recompute the perturbation measure for Experiment 1 with a time-domain posterior obtained from a dense-particle PGAS run (or from a Kalman-filter-based sampler where the tvAR is linear in the states) using several times more particles, and check whether the ranking of the spectral likelihoods and the size of the reported improvements are stable.","tokens_in":19483,"feed_emoji":"📈","tokens_out":4802,"duration_ms":44044,"temperature":0.7,"pith_summary":"The paper asks whether fast spectral-domain likelihood approximations can stand in for the exact time-domain likelihood in Bayesian inference for time-varying parameter models. Exact Gaussian likelihoods cost $O(T^3)$ and become prohibitive for long series, while Whittle-type approximations cost $O(T\\log T)$ and are natural when models are specified by their spectral density. The paper compares the block Whittle and dynamic Whittle likelihoods against the time-domain posterior in three simulation experiments on time-varying AR processes, measuring both estimation efficiency and the quantile-wise perturbation of the approximate posterior. It finds both approximations perform relatively well, with a consistent edge for the dynamic Whittle, and that tapering, prewhitening, and boundary correction each substantially improve accuracy. It also documents a defect: tapering can distort variance on short segments and, with overlapping windows, produce spurious volatility clusters.","feed_headline":"Fast spectral likelihoods nearly match exact Bayesian posteriors","feed_subtitle":"Dynamic Whittle with tapering, prewhitening, or boundary correction tracks the time-domain posterior in TV-AR models.","key_machinery":"The central objects are two approximate likelihoods built from local periodogram ordinates treated as independent exponential observations. The block Whittle likelihood (Dahlhaus, 1997) computes periodograms over sliding, partially overlapping windows of length $N$ and treats all frequencies in each window as conditionally independent. The dynamic Whittle likelihood (Tang et al., 2023) uses the moving local periodogram of Häfner and Kirch (2017): each time point contributes exactly one frequency, cycling through the positive Fourier frequencies across time, which keeps neighboring observations nearly independent while preserving full time resolution. The three modifications act on the segment periodogram: tapering multiplies each segment by a data taper to reduce leakage; prewhitening filters with a fitted AR model and rescales the residual periodogram; boundary correction augments the DFT with a predictive DFT from the best linear predictor, removing boundary mismatch. The posterior under either likelihood is sampled by a Gibbs sampler with a particle Gibbs with ancestor sampling (PGAS) step, and accuracy is scored by relative efficiency of posterior medians and by the quantile perturbation measure in (11).","core_discovery":"The central claim is that the posterior distribution from local Whittle likelihoods is a serviceable approximation of the time-domain posterior for time-varying AR processes, and that the approximation is dramatically improved by three established modifications: tapering, prewhitening, and boundary correction. In three simulation experiments with 100 replications of length $T=1500$, the dynamic Whittle likelihood with any of the modifications attains near-time-domain efficiency and small quantile perturbation, and the modified versions are much less sensitive to segment length than the raw versions. The dynamic Whittle posterior also tends to have better calibrated interval widths, while the block Whittle posterior is often too tight. In the near-unit-root experiment the modifications help less, likely because the process spends little time near the instability boundary. On egg prices with stochastic volatility, raw and prewhitened Whittle posteriors track the time-domain posterior for the AR parameters, while tapered versions need per-segment variance rescaling to avoid spurious volatility clusters; boundary correction without tapering performs well on the variance path.","pith_inferences":["If the modified dynamic Whittle posterior tracks the time-domain posterior in tvAR models, a natural extension is to trust the same approximation for locally stationary models whose time-domain likelihood is intractable, where no gold-standard comparison is possible.","The quantile-perturbation measure could serve as a general diagnostic for any approximate likelihood, not just spectral ones, by benchmarking against a small-scale exact posterior before scaling up.","The tapering-induced volatility-cluster artifact suggests a testable extension: comparing tapered and untapered posteriors for the log-variance path could flag when segment overlap is creating spurious dynamics.","Boundary correction without tapering performed best on the egg-price variance path, hinting that tapering is not always the right default when the process has heteroscedasticity."],"forward_implications":["For long time-varying AR series, the modified dynamic Whittle likelihood gives posteriors close to the exact time-domain posterior at $O(T\\log T)$ cost.","Any of the three modifications makes both likelihood approximations substantially less sensitive to the segment length and step size.","Because periodogram ordinates are asymptotically independent, the modified likelihoods still support subsampling MCMC, so the accuracy gains do not forfeit the main speed-up that motivates spectral likelihoods.","Tapering should be used with care, or combined with per-segment variance rescaling, whenever the model includes a time-varying variance, since tapering can induce spurious volatility clusters.","The dynamic Whittle posterior is better calibrated than the block Whittle posterior, which tends to be overconfident in its interval widths."],"supporting_citations":[{"why":"Defines the Whittle log-likelihood, the starting point for all spectral approximations studied.","marker":"(Whittle, 1953)"},{"why":"Introduces the block Whittle likelihood and proves consistency and efficiency of its MLE.","marker":"(Dahlhaus, 1997)"},{"why":"Introduces the dynamic Whittle likelihood and proves posterior consistency for locally stationary processes.","marker":"(Tang et al., 2023)"},{"why":"Motivates tapering to correct periodogram leakage and shows small-sample efficiency gains.","marker":"(Dahlhaus, 1988)"},{"why":"Source of the prewhitening modification that flattens the spectrum before periodogram computation.","marker":"(Priestly, 1981)"},{"why":"Proposes boundary correction via the complete DFT and the predictive DFT, the third modification studied.","marker":"(Subba Rao and Yang, 2021)"},{"why":"Defines the moving local periodogram that underlies the dynamic Whittle likelihood.","marker":"(Häfner and Kirch, 2017)"},{"why":"Provides the particle Gibbs with ancestor sampling algorithm used to draw the gold-standard time-domain posterior.","marker":"(Lindsten et al., 2014)"},{"why":"Supplies finite-sample efficiency results for the stationary Whittle MLE that the paper extends to the time-varying setting.","marker":"(Contreras-Cristán et al., 2006)"}],"fun_headline_variants":["Spectral likelihoods with tweaks nearly equal exact Bayes","Modified Whittle posterior near-exact for TV-AR","Fast spectral likelihoods with corrections near-exact","Near-exact posteriors from tweaked Whittle likelihoods"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison's gold standard, the time-domain posterior sampled with particle Gibbs with ancestor sampling, is itself a Monte Carlo approximation, and the paper does not report the number of particles or convergence diagnostics, so a biased or unconverged particle run would contaminate the perturbation measure and could change the ranking of the spectral likelihoods.","fun_headline_variants_meta":{"raw":{"variants":["Spectral likelihoods with tweaks nearly equal exact Bayes","Modified Whittle posterior near-exact for TV-AR","Fast spectral likelihoods with corrections near-exact","Near-exact posteriors from tweaked Whittle likelihoods"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001693,"raw_usage":{"total_tokens":6689,"prompt_tokens":912,"completion_tokens":5777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":5711}},"tokens_in":528,"tokens_out":5777,"duration_ms":37822,"temperature":1.0,"reasoning_tokens":5711,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:38:33.886399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the perturbation measure for Experiment 1 with a time-domain posterior obtained from a dense-particle PGAS run (or from a Kalman-filter-based sampler where the tvAR is linear in the states) using several times more particles, and check whether the ranking of the spectral likelihoods and the size of the reported improvements are stable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Whittle log-likelihood, the starting point for all spectral approximations studied."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the block Whittle likelihood and proves consistency and efficiency of its MLE."},{"cited_title":"Bayesian nonparametric spectral analysis of locally stationary processes","cited_arxiv_id":"2303.11561","evidence_quote":"Introduces the dynamic Whittle likelihood and proves posterior consistency for locally stationary processes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates tapering to correct periodogram leakage and shows small-sample efficiency gains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the prewhitening modification that flattens the spectrum before periodogram computation."},{"cited_title":"and Yang, J","cited_arxiv_id":null,"evidence_quote":"Proposes boundary correction via the complete DFT and the predictive DFT, the third modification studied."},{"cited_title":"I., and Schon, T","cited_arxiv_id":null,"evidence_quote":"Provides the particle Gibbs with ancestor sampling algorithm used to draw the gold-standard time-domain posterior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies finite-sample efficiency results for the stationary Whittle MLE that the paper extends to the time-varying setting."}],"review_version":1}