{"id":"17b22932-70c3-49a2-80b7-fa66225e6438","arxiv_id":"2507.12878","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A Bayesian framework that models the impulse response as a stochastic process with mean and fluctuation terms recovers linear time-invariant systems from single noisy pairs and tracks smoothly varying systems using Gaussian process priors.","lead":"This paper treats a system's impulse response as a random process, splitting it into an average part and a fluctuation part, and uses Bayesian neural networks and Gaussian processes to estimate time-varying systems from single noisy input-output pairs. It offers a general way to attach uncertainty bars to system identification, useful for seismology, wireless channels, and control, and it claims better results than classical averaging in simulated ambient noise tomography.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LTV tracking claim rests on a smoothness prior matched to the synthetic ground truth; without a sensitivity analysis the demonstration is conditional on the prior's timescale.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern, and I agree. The central claim has three components: the h = mu + E posterior representation, the LTI single-pair estimator, and the LTV tracker. The representation is mathematically standard and the paper explicitly disclaims novelty; the single-pair experiment is partially supported by a 100-replication coverage check and, while the 'single observation' phrasing is generous because one pair contains many output samples for only 16 unknown taps, that is not the main vulnerability. The LTV tracker is the hardest and most novel demonstration, and its success depends on a GP prior whose RBF length scale (16) is chosen for a ground truth generated by smooth interpolation. Without a sensitivity analysis, the reported tracking could simply reflect the prior's smoothness being imposed on data generated to be smooth. This is a conditionable issue rather than a fatal one: the paper acknowledges the prior dependence, and a targeted experiment sweeping the true variation timescale or the prior length scale would settle whether the method degrades gracefully. The ANT error-floor claim also lacks repeated-seed error bars, but it is a secondary demonstration and the paper frames it as an end-to-end pipeline comparison. I therefore keep the reader's conditional verdict.","tokens_in":17531,"tokens_out":6560,"duration_ms":77616,"concrete_test":"Re-run the Section 4.3 LTV experiment with the same input, noise, windows, and GP settings, but generate ground-truth interpolation weights alpha_i[n] as sinusoids with periods T in {64, 32, 16, 8} samples instead of the original smooth schedule, and report the normalized mean-squared error between the stitched posterior mean and the true h[n,k] for each T. If the error rises sharply as T approaches or falls below the GP length scale 16, the demonstration is prior-matched; if the error stays low across all T, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest practical claim is the LTV tracking demonstration in Section 4.3. The ground truth is generated by smooth interpolation between three FIRs (Eq. 18, Section 4.3.1), and the estimator is regularized by an RBF GP prior with length scale 16 over 32-sample windows (Appendix A). The rate at which the interpolation weights alpha_i[n] vary is never reported, and no sensitivity analysis over the prior length scale or over the true temporal variation rate is given. If the true impulse response varies on a timescale comparable to or shorter than the prior length scale, the posterior mean will be over-smoothed and the stitched estimate will fail to track. The paper's own Section 5 concedes that LTV regression quality is contingent on prior choice, but it gives no quantitative bound on when the method works. Because the experiment's success may be a consequence of matching the prior to the generative model, the claim 'can track a continuously varying LTV impulse response' is not established beyond the smooth, prior-matched regime. The paper is transparent about this limitation, but it remains the load-bearing soft spot of the empirical case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a Bayesian framework for identifying linear time-varying (LTV) systems in which the impulse response h(t,τ) itself is modeled as a stochastic process and decomposed as h = μ + E, with μ the posterior mean and E zero-mean posterior fluctuations. It derives first- and second-order output statistics under this representation, defines the Linear Time-Invariant in Expectation (LTIE) class as the case where the posterior mean is time-invariant, and then presents three synthetic experiments: single-observation LTI impulse-response regression with a Bayesian convolutional layer, simulated ambient noise tomography (ANT) comparing a Bayesian mean impulse response with classical CCF stacking, and LTV tracking using an amortized variational estimator with a Gaussian-process prior. The paper is explicit throughout that the experiments are demonstrations of the posterior representation under deliberate model misspecification, not evidence that the physical data-generating mechanism is stochastic.","tokens_in":17806,"tokens_out":6071,"duration_ms":73766,"significance":"If the empirical claims are supported, the framework would provide a useful unified posterior representation for uncertainty-aware system identification, with practical variational estimators that propagate uncertainty from the impulse response to derived quantities such as denoised signals, transfer functions, and cross-correlation functions. The paper has clear strengths: the moment derivations in Eqs. (4), (12), and (13) are correct under the stated integrability and measurability assumptions; the single-observation experiment includes a 100-replication coverage check; the authors honestly frame the experiments as misspecification tests and distinguish epistemic from intrinsic uncertainty; and the code and data are made publicly available. However, the empirical case rests on single synthetic datasets for the two headline applications, and the LTV experiment relies on a prior whose influence in the optimized objective is very weak and whose timescale is matched to the generative model without sensitivity analysis. These issues do not undermine the theoretical framework, but they currently limit the strength of the claimed practical demonstrations.","major_comments":[{"comment":"The LTV experiment claims that a structured GP prior regularizes an otherwise ill-posed regression, but the optimized objective places almost no weight on that prior. In Eq. (19), the loss is L = E_q ||y_window - g_hat_window||^2 + β·KL(q || GP_prior), and Appendix A reports β = 5×10^-5. With an observation-noise standard deviation of 0.2 in the generative model, a proper ELBO would have a data-weight coefficient of 1/(2σ^2) = 12.5 and a KL coefficient of 1; here the KL term is smaller than the data term by roughly five orders of magnitude. The apparent temporal smoothness may therefore come from the CNN encoder, the overlapping-window architecture, or the stitching procedure rather than from the GP prior. The paper should include an ablation or sensitivity analysis over β (e.g., β = 1 or an intermediate value) and ideally over the prior length scale, to substantiate the claim that the GP prior is doing the regularization work.","section":"Sec. 4.3.2, Eq. (19); Appendix A"},{"comment":"The LTV tracking demonstration lacks sensitivity analysis with respect to the temporal variation rate of the ground truth. The ground-truth response is generated by smooth interpolation among three FIRs via weights α_i[n] in Eq. (18), while the GP prior uses an RBF kernel with amplitude 2.0 and length scale 16 over 32-sample windows. The rate at which α_i[n] varies is not reported, and no experiment varies either the true variation rate or the prior length scale. As the paper acknowledges in Section 5, LTV regression quality is contingent on prior choice, but without any quantitative study of this contingency the claim that the method 'can track a continuously varying LTV impulse response' is established only for a smooth, prior-matched regime. Adding experiments with faster-varying ground truths and with mismatched prior length scales would directly address this load-bearing limitation.","section":"Sec. 4.3.1 and Appendix A"},{"comment":"The claim of a lower error floor than classical CCF stacking in the ANT experiment is based on a single simulated dataset. The reported mean target-curve errors (e.g., 0.225 versus 0.245 in the full-precision sweep, 0.241 versus 0.269 with 1-bit quantization) and minimum errors are point estimates over the pair-number sweep with no repeated-seed variability, confidence intervals, or error bars. Given that the differences are relatively small, it is plausible that they lie within seed-to-seed variation. The paper should either run multiple independent seeds/dataset realizations and report mean ± standard deviation or provide a statistical test. In addition, the 1-bit comparison changes both quantization and the Bayesian batch size (2000 to 4000), which confounds the comparison.","section":"Sec. 4.2.2, Figure C.7"}],"minor_comments":[{"comment":"The symbol 'B' is used to denote 'defined as' without an explicit explanation; please define it at first use or replace it with a standard notation such as ':='.","section":"Sec. 3.2, Eq. (14)"},{"comment":"The statement that 'the 291 unique windows were tiled twice for training' is unclear; please state the length of the input signal and explain how 291 follows from the window length and stride so that the setup is reproducible from the text alone.","section":"Appendix A, LTV experiment"},{"comment":"The ANT method is described at a high level ('deconvolve the source characteristics'), but the network architecture, the exact regression target, and the loss function used for the ANT experiment are not specified as fully as in the other experiments. Expanding this description would improve reproducibility despite the availability of code.","section":"Sec. 4.2.1 and Appendix C"},{"comment":"The sentence 'A proof is omitted' for the convergence of the averaged periodogram is unnecessary, since the statement is a direct consequence of the law of large numbers, but if kept it should be integrated into the surrounding argument rather than standing alone.","section":"Sec. 3.1, Eqs. (5)-(7)"}],"recommendation":"major_revision","confidential_remarks":"The theoretical framework is sound and the paper is transparent about its limitations, but the two most prominent empirical claims—the ANT error-floor advantage and the LTV tracking capability—currently rest on single synthetic datasets with either no repeated-seed statistics or a prior whose effect is not quantitatively demonstrated. These are fixable with additional experiments and sensitivity analyses, so I recommend major revision rather than rejection. The paper's contribution is more conceptual and demonstrative than a mathematically deep advance; this should be weighed when considering whether the scope fits the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this is a modest but honest paper. The conceptual core—splitting an impulse response into a posterior mean and a zero-mean fluctuation—is explicitly acknowledged as not new, and LTIE is just a constant-mean random channel. What the paper really contributes is a clean posterior interpretation and a set of practical variational estimators built on it.\n\nThe discrete-time moment calculations (Eqs. 4, 12, 13) are correct under standard integrability assumptions. The single-observation LTI experiment is the strongest part: a 100-replication coverage check shows conservative intervals that contract with more data, and the posterior propagation to CCFs and transfer functions is a clear practical win. The ANT comparison is a nice end-to-end demonstration, though it rests on one synthetic dataset.\n\nThe LTV tracking experiment is the load-bearing empirical claim and the softest. The prior uses an RBF length scale of 16 over 32-step windows, and the ground truth is generated by smooth interpolation between three FIRs. That is a prior matched to the target timescale. No sensitivity analysis over length scale or true variation rate is provided, and the paper itself concedes that LTV regression quality is contingent on prior choice. So the claim that the method 'can track a continuously varying LTV impulse response' is true only in the smooth, prior-matched regime. That should be stated more plainly, or better, probed with mismatched priors. Also, the ANT 'lower error floor' is reported from a single synthetic dataset without repeated-seed error bars; that is a smaller issue, since the pair sweep gives some idea, but a few independent runs would confirm the gap.\n\nThe paper is transparent about misspecification and about the approximate nature of the variational posterior. That is to its credit.\n\nThis deserves a serious referee. It's not a big advance, but it's solid, readable, and useful for practitioners who want a principled way to attach uncertainty to impulse-response estimates from very little data. The main things I'd ask the authors to fix are the LTV sensitivity analysis and a bit more seed averaging in ANT. I'd recommend conditional acceptance after revisions, but definitely send it out rather than desk-reject.","headline":"Modest but honest: the h = mu + E decomposition is old, the posterior interpretation and single-observation deconvolution experiment are useful, and the LTV tracking claim is real but contingent on a matched smoothness prior.","tokens_in":18317,"tokens_out":1803,"would_cite":true,"duration_ms":19697,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Treating the impulse response as a random process lets one estimator recover an LTI filter from a single noisy pair and track a smoothly varying LTV response.","keywords":["linear time-varying systems","system identification","Bayesian inference","Gaussian processes","Bayesian neural networks","variational inference","impulse response estimation","ambient noise tomography"],"falsifier":"Run the LTV estimator on ground-truth responses whose three base FIRs switch abruptly within a few steps, faster than the Gaussian-process length scale of 16, and compare the stitched posterior mean against ground truth; strong tracking degradation would show the result is carried by the smoothness prior. A complementary check is to apply the single-observation estimator to a known FIR that is much longer than the 16-tap prior support and test whether the nominal 90–95 percent credible intervals remain conservative.","tokens_in":17299,"feed_emoji":"🎛️","tokens_out":9793,"duration_ms":101485,"temperature":0.7,"pith_summary":"The paper argues that identifying a linear time-varying system from input-output data should be treated as Bayesian inference over the system's impulse response, modeled as a random process rather than a fixed unknown function. Its central move is the decomposition $h = \\mu + E$, where $\\mu$ is the posterior mean response and $E$ is a zero-mean fluctuation process, and the claim is that this one representation covers both genuinely random channels and epistemic uncertainty about a deterministic channel. On that basis the paper defines a new system class, Linear Time-Invariant in Expectation (LTIE), in which the mean response is constant while fluctuations may have arbitrary covariance structure. It then demonstrates practical variational estimators: a Bayesian convolutional layer infers a 16-tap FIR from a single noisy input-output pair, a Gaussian-process-prior model with an amortized CNN encoder tracks a continuously varying LTV response, and the same machinery reaches a lower error floor than classical cross-correlation stacking in a simulated ambient-noise tomography setting. The experiments are framed as validations of usefulness under deliberate model misspecification, not as evidence that real channels are literally stochastic.","feed_headline":"A noisy pair can identify an LTI system and track a drifting one","feed_subtitle":"The same stochastic impulse-response representation quantifies uncertainty and beats classical CCF stacking in simulated tomography.","key_machinery":"The load-bearing object is the decomposition of the impulse response into a posterior mean and a zero-mean fluctuation process, $h = \\mu + E$, together with the identity $\\mathrm{Var}[g[n]|D] = f[n]^\\top \\Sigma[n] f[n]^*$ that carries uncertainty from the taps to the output. This identity is what makes predictive uncertainty depend on the probing input and what lets posterior samples of the filter be pushed forward to derived quantities like the denoised signal, the transfer function, and the cross-correlation function. For the LTV experiment the additional mechanism is a truncated Gaussian-process prior over each tap with a squared-exponential kernel in time, block-diagonal across taps, used as a KL regularizer on a Bayesian latent impulse-response layer whose parameters are produced by a deterministic CNN encoder; stitching the windowed posterior means gives the global time-varying estimate.","core_discovery":"The paper's central claim is that representing the impulse response as a stochastic process through $h(t,\\tau) = \\mu(t,\\tau) + E(t,\\tau)$ yields a posterior representation that unifies intrinsic channel variability and epistemic uncertainty: when inference favors time-invariant responses, a deterministic unknown LTI system induces an LTIE posterior law whose mean is time-invariant, and the same law can also describe a genuinely random channel. The output variance is then signal-dependent, since $\\mathrm{Var}[g[n]|D] = f[n]^\\top \\Sigma[n] f[n]^*$ with posterior tap covariance $\\Sigma[n]$, so uncertainty vanishes when the input is silent. The paper further shows that this representation supports practical estimators: a mean-field Bayesian convolutional layer with a zero-mean Gaussian prior recovers a 16-tap FIR from one noisy earthquake-signal pair and propagates credible intervals to denoised signals, transfer functions, and cross-correlations; a Gaussian-process prior over 32-step windows with an RBF kernel over time allows an amortized CNN encoder to track a smoothly interpolated LTV response from a single time series; and in a 15,000-pair simulated ambient-noise tomography sweep the Bayesian mean impulse response reaches a lower mean target-curve error than classical CCF stacking. Throughout, the paper states that these experiments are demonstrations of the utility of the posterior representation under misspecification, not direct confirmation that the stochastic impulse-response model is the physical data-generating mechanism.","pith_inferences":["A natural next step the paper leaves untested is using the posterior spread as a stopping rule: in the ANT sweep the Bayesian error floor saturates earlier, so the posterior variance could decide how many receiver pairs are enough.","The signal-dependent variance formula predicts a falsifiable contrast with additive-noise models: probing with an input that has a spectral null should shrink the posterior output variance in that band even when epistemic uncertainty about the corresponding taps remains, and this could be measured directly.","The LTV experiment could be extended by learning the GP length scale or using a spectral mixture kernel; the paper lists richer priors as future work but does not test whether tracking survives when the true variation speed is unknown."],"forward_implications":["If the central claim is right, system identification from a single noisy input-output pair is a tractable variational inference task, with posterior intervals replacing point estimates of the filter.","The same posterior law can be read either as intrinsic channel randomness or as epistemic uncertainty about a deterministic system, so an LTIE posterior is the natural object to report when only time-invariant behavior is assumed.","Because output uncertainty is quadratic in the input and vanishes without excitation, predictions from the framework are self-signaling about which parts of the system are actually probed.","The Bayesian mean impulse response reaching a lower error floor than classical CCF stacking in simulated tomography suggests the method can work without spectral whitening and can retain spectral-amplitude information.","Smoothly varying LTV responses can be tracked from one time series if the temporal variation is slow relative to the Gaussian-process prior length scale."],"supporting_citations":[{"why":"Supplies the stochastic characterization of randomly time-variant channels (WSSUS) that the $h = \\mu + E$ decomposition generalizes.","marker":"Bello, 1963"},{"why":"Supplies the kernel-based Bayesian identification of linear systems that grounds the prior-over-impulse-response approach.","marker":"Pillonetto and Ljung, 2023"},{"why":"Provides the Gaussian-process regression formulation for nonparametric LTV identification that the structured GP prior extends.","marker":"Hallemans et al., 2020"},{"why":"Provides the Earthquakes time series used as the known input signal in the single-observation and LTV experiments.","marker":"Dau et al., 2019"},{"why":"Gives the Flipout estimator that makes gradient estimation for the Bayesian convolutional layer tractable.","marker":"Wen et al., 2018"},{"why":"Defines the classical CCF stacking workflow used as the baseline in the simulated ambient-noise tomography comparison.","marker":"Bensen et al., 2007"},{"why":"Supplies the spectral formulation used to convert estimated impulse responses into phase-velocity curves for evaluation.","marker":"Ekström et al., 2009"},{"why":"Supplies the reparameterization trick that enables ELBO optimization in the variational estimators.","marker":"Kingma and Welling, 2013"}],"fun_headline_variants":["Single noisy pair yields LTI identification and LTV tracking","Bayesian impulse response beats CCF with just one noisy pair","Single-pair Bayesian identification: LTI and LTV in one framework","Stochastic impulse response: identify LTI from one pair, track drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the true impulse response is smooth at the temporal scale encoded by the Gaussian-process prior (RBF kernel amplitude 2.0 and length scale 16 over 32-step windows), because single-observation local inference is otherwise severely ill-posed.","fun_headline_variants_meta":{"raw":{"variants":["Single noisy pair yields LTI identification and LTV tracking","Bayesian impulse response beats CCF with just one noisy pair","Single-pair Bayesian identification: LTI and LTV in one framework","Stochastic impulse response: identify LTI from one pair, track drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000801,"raw_usage":{"total_tokens":3575,"prompt_tokens":1053,"completion_tokens":2522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":2448}},"tokens_in":669,"tokens_out":2522,"duration_ms":19425,"temperature":1.0,"reasoning_tokens":2448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:37:00.460984+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the LTV estimator on ground-truth responses whose three base FIRs switch abruptly within a few steps, faster than the Gaussian-process length scale of 16, and compare the stitched posterior mean against ground truth; strong tracking degradation would show the result is carried by the smoothness prior. A complementary check is to apply the single-observation estimator to a known FIR that is much longer than the 16-tap prior support and test whether the nominal 90–95 percent credible intervals remain conservative.","supporting_citations":[],"review_version":1}