{"id":"525efe1a-16c7-4c2e-af36-3b6bd90b6f07","arxiv_id":"2505.14364","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"The authors show that Bayesian active selection of measurement delays, guided by prior spectral knowledge, can outperform conventional Fourier-transform spectroscopy, with Nyquist sampling recovered as the ignorant limit.","lead":"This paper proposes an information-driven way to sample optical spectra, replacing fixed, equally spaced measurement delays with delays chosen one at a time to maximize expected information gain given prior knowledge. It demonstrates faster convergence and real-time uncertainty quantification in simulations of blood-plasma spectroscopy and in two experimental setups: vortex beam characterization and a compact RGB-camera hyperspectral imager.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'never below FTS' floor is not established: the described prior-selection mechanism (Eqs. S248-S249) is biased against the uninformed baseline and contains no actual fallback rule, so a misspecified informative prior can dominate the softmax average and degrade performance.","rationale":"Reading the manuscript in good faith, the core Gaussian update machinery is standard and the three demonstrations, especially the vortex uncertainty propagation, are valuable; the paper does not need the universal floor to be useful. But the abstract and Section I deliberately make the stronger claim that performance never falls below conventional sampling, and that claim is load-bearing for the paper's framing as a 'guarantee'. The reader's conditional verdict targeted exactly the unproven fallback; my stress test sharpens it: the described evidence weighting is not merely unproven but plausibly biased in the wrong direction, because Bayesian evidence strongly penalizes the high-variance prior that represents FTS, and the softmax does not implement a fallback. A misspecified tight prior can dominate model averaging while adaptively steering future measurements away from where they are needed. This is not an internal inconsistency in the Gaussian updates, but a gap between the stated guarantee and the proposed selection mechanism. I therefore keep the reader's CONDITIONAL verdict; acceptance should require either a proof that the weighted estimate dominates FTS under prior misspecification, or softening the claim to 'matches FTS when priors are correct and improves performance in favorable cases'. A concrete misspecification benchmark, as proposed, would settle whether the fallback works.","tokens_in":17433,"tokens_out":4718,"duration_ms":53423,"concrete_test":"Run a controlled misspecification experiment on the molecular-fingerprinting pipeline: generate test spectra from a distribution not in the prior ensemble (e.g., Lorentzian absorption lines with positions and widths drawn from a disjoint range), and use as the sole informative prior a Gaussian fit whose mean is shifted by at least two prior standard deviations in a diagnostic band. Execute the BAS loop in Section I and S2.4 exactly as described, record the softmax weights (Eq. S249) after every measurement, and compare final reconstruction SNR/RMSE to uniform Nyquist FTS with identical noise and measurement count. If any such misspecified prior yields BAS worse than FTS, or if the uninformed model's weight never becomes dominant, the 'never below' claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"To support the headline claim that BAS 'never falls below conventional sampling', the algorithm must detect an unhelpful prior and fall back to the uninformed Nyquist/FTS strategy. The only mechanism actually specified is softmax weighting of candidate prior models by their cumulative log evidence (Eq. S249), where each evidence term includes the model-complexity penalty -1/2 log det(Sigma_F + R Sigma_S R^T) (Eq. S248). This mechanism is structurally the wrong tool for the claimed fallback. For the 'uninformed' model Sigma_0 = alpha I used to represent ignorance (Eq. S221), the determinant term grows unboundedly with alpha, so the flat prior receives vanishing weight in any finite-data comparison; the more ignorant the baseline is, the less it is favored. In contrast, a tight but wrong prior can accumulate moderate predictive scores on early measurements and dominate the softmax before the mismatch becomes evident. Because the adaptive delay selection is driven by the currently weighted posterior, the sampling policy itself is then steered by the misspecified model, so re-weighting after the fact cannot recover the lost measurements. No threshold, decision rule, or temperature regime is specified that would implement 'fall back to baseline if priors fail'; the guarantee is therefore asserted, not derived. The synthetic blood-plasma benchmark additionally tests only spectra drawn from the same Gaussian family used to construct the priors, so it does not exercise the misspecification regime the guarantee depends on.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Bayesian Autocorrelation Spectroscopy (BAS), an adaptive framework for Fourier transform spectroscopy and related linear inverse problems. After each measurement, a Gaussian posterior over the spectrum is updated through Eq. (3), and the next interferometric delay is chosen to maximize the expected information gain of Eq. (4). The authors argue that under an uninformed prior this criterion reduces to D-optimality and recovers Nyquist sampling as the optimal fixed design, while with informative priors BAS can sample adaptively and outperform conventional FTS. The claims are supported by three demonstrations: reconstruction of synthetic blood-plasma FTIR spectra, spectral phase retrieval of an optical vortex with propagated uncertainties, and hyperspectral imaging from an RGB camera combined with an interferometric delay stage.","tokens_in":17743,"tokens_out":5776,"duration_ms":57745,"significance":"If the central performance guarantee were established, the paper would provide a useful unification of classical sampling theory with Bayesian experimental design, with practical potential for faster spectroscopic measurements and multiplexed imaging. The linear-Gaussian update equations and information-gain criterion are standard and correctly presented; the use of existing D-optimality results to recover equispaced sampling for trigonometric models is legitimate; and the covariance propagation through the vortex phase-retrieval chain is a genuine methodological contribution. However, the headline claim that BAS 'never falls below conventional sampling' is not proved, and the main quantitative benchmark is weakened by the fact that the synthetic test spectra are drawn from the same Gaussian model family used to build the priors. These issues are load-bearing for the abstract and Section I, although they do not invalidate the underlying framework.","major_comments":[{"comment":"The claim that BAS 'at least matches' FTS even with unhelpful priors is asserted rather than derived. The only specified prior-selection mechanism is the softmax weighting of cumulative log evidence in Eq. (S249), where each evidence contribution includes the complexity penalty (1/2) log det(Sigma_F + R Sigma_S R^T) in Eq. (S248). For the uninformed baseline Sigma_0 = alpha I of Eq. (S221), this determinant term grows with alpha, so the more diffuse the baseline is, the less weight it receives; the mechanism is structurally biased against the model that is supposed to represent FTS. No threshold, decision rule, or temperature regime is specified that would implement a 'fall back to baseline' behavior, and because the adaptive delay schedule is steered by the currently weighted posterior, a misspecified informative prior can dominate the sampling policy before the evidence reweights it. The guarantee therefore needs either a theorem with explicit misspecification assumptions, an explicit fallback algorithm with a correctness argument, or a substantial reformulation of the claim to a statement about typical rather than guaranteed performance.","section":"Section I, last paragraph; S2.4.2, Eqs. (S248)-(S249)"},{"comment":"The quantitative demonstration that BAS outperforms FTS is conducted entirely on synthetic spectra generated by fitting multivariate Gaussian distributions to the L4L cohort, with the same training-set means and covariances serving as prior models. The test draws therefore come from the same distribution family used to construct the priors, which is a within-model validation loop and does not exercise the misspecification regime in which the 'never below FTS' guarantee matters. A convincing demonstration would need held-out real spectra or a deliberate misspecification experiment, such as priors derived from one subgroup evaluated on another, and the reported gains should be presented as applying to the well-specified case rather than as a general clinical-performance statement.","section":"S1, 'Synthetic data on molecular fingerprinting'; Fig. 1"},{"comment":"The evidence computation as written is ambiguous. Equation (S247) correctly defines the model evidence as a product of one-step-ahead predictive densities p(F_i | F_1,...,F_{i-1}, M_k), but Eq. (S248) is stated with mu_S and Sigma_S as 'the moments of the prior.' If these are the original prior moments rather than the posterior moments after F_1,...,F_{i-1}, the product is not the marginal likelihood and the softmax weights in Eq. (S249) are not Bayesian model evidence. If they are instead meant to be the current posterior moments, that should be stated explicitly and the sequential update shown. This distinction matters because the claimed robustness to prior misspecification relies on the evidence weighting being a valid predictive measure.","section":"S2.4.2, Eqs. (S247)-(S248)"}],"minor_comments":[{"comment":"The main text describes gamma as a 'Bayesian update weight,' but in the vector case gamma is a matrix, specifically Sigma_old R^T (R Sigma_old R^T + sigma_add^2)^(-1); the notation should be reconciled between the main text and the supplement.","section":"Eq. (3) and S2.3.1"},{"comment":"The phrase 'starting from prior set of prior set with zero mean' contains a duplicated phrase and should read 'starting from a prior set with zero mean.'","section":"S1, Vortex and RGB paragraphs"},{"comment":"The curve labeled 'informed Nyquist sampling' should be defined explicitly as fixed equispaced delays with an informative prior, since the distinction between adaptive and fixed-informed sampling is central to the message.","section":"Fig. 1B"},{"comment":"The function g(z) is presented without derivation; a short derivation or a citation would help the reader verify the claimed maximum at z approximately 1.164 pi.","section":"S2.2.1.5, Eq. (S224)"},{"comment":"The phrase 'performance never falls below conventional sampling' is a strong universal guarantee; the main text supports it only with the sentence 'we can guarantee,' which is not backed by a theorem or algorithm. The abstract should be tempered to what is actually proved or validated empirically.","section":"Abstract and Section I"}],"recommendation":"major_revision","confidential_remarks":"The paper is overclaimed relative to what is proved: the 'never below FTS' guarantee is the central selling point, and the current manuscript does not supply the required fallback mechanism or misspecification analysis. The synthetic benchmark being generated from the same Gaussian family as the priors should be disclosed more prominently. The underlying framework and the uncertainty-propagation work are potentially publishable, but the headline claims need to be brought into line with the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe paper applies Bayesian experimental design to Fourier-transform spectroscopy: build a Gaussian prior over the spectrum, pick delays to maximize expected information gain, update sequentially, and weight multiple candidate priors by their evidence. The central claim is that this \"never falls below\" conventional Nyquist-sampled FTS. The framework is standard linear-Gaussian inference, and the theoretical sections are largely a clean re-derivation of known results (D-optimality implies equispaced sampling; myopic information gain converges to Nyquist after a few steps). That's not original, but it's well done and useful as a unifying perspective. The demonstrations—synthetic blood-plasma FTIR, real vortex phase reconstruction with covariance propagation, and an RGB-camera hyperspectral imager—are interesting, and the uncertainty propagation through the vortex analysis chain is a genuine plus. The paper is clearly written and the math appears correct.\n\nThe soft spot is the headline guarantee. The mechanism for falling back to the uninformed baseline is described only vaguely (\"depending on the result, we can either combine predictions or fall back\"). The actual model-weighting scheme in S2.4.2 (softmax over cumulative log-evidence, Eq. S249) does not implement fallback; in fact the evidence penalty for the uninformed prior (large alpha identity covariance) grows with alpha, so the uninformed model receives vanishing weight even when it is the only honest model. A misspecified informed prior can dominate early and steer the adaptive sampling policy, after which re-weighting cannot recover. So the \"never below FTS\" claim is asserted, not derived, and the stress-test note is right. The synthetic blood-plasma benchmark is generated from the same fitted Gaussian family used as priors, so it doesn't exercise the misspecification regime the guarantee needs. No code or data are released, so the figures cannot be independently checked.\n\nNone of this kills the paper. The core methodology is sound and could be useful, especially the uncertainty propagation and the RGB fusion. But the abstract's overreach will undermine it in review. The authors should either prove a rigorous fallback bound (or a regret bound for a genuinely robust selection rule) or substantially soften the claim to \"we never observed worse performance in our benchmarks.\" They should also add a real-data benchmark where the prior is intentionally misspecified, and ideally release the data and code.\n\nI'd send this to peer review—the ideas are worth engaging, and the referee can push for the missing guarantee and the real-data test. The novelty is within-subfield rather than paradigm-level, but that's fine. The citation pattern is honest and relevant.","headline":"Standard Bayesian experimental design applied to FTS, with a genuine but unproven 'never below baseline' guarantee and a synthetic benchmark that doesn't stress the failure mode.","tokens_in":18294,"tokens_out":2754,"would_cite":false,"duration_ms":28005,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62K05","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Adaptive Bayesian spectroscopy can beat fixed Nyquist-rate measurement.","keywords":["Bayesian Autocorrelation Spectroscopy","adaptive sampling","information gain","Fourier transform spectroscopy","Nyquist sampling","prior knowledge","uncertainty quantification","hyperspectral imaging"],"falsifier":"Run BAS on spectra drawn from a distribution deliberately different from all candidate priors, with an overconfident prior that initially predicts early measurements well, and compare the cumulative reconstruction error to an uninformed Nyquist-sampled baseline; if the weighted BAS estimate remains below the baseline after enough measurements for the evidence to switch, the asserted floor fails. A second direct test is to run the molecular-fingerprinting pipeline on real (not synthetic) blood-plasma interferograms and compare adaptive BAS against standard FTS.","tokens_in":17213,"feed_emoji":"🔬","tokens_out":6857,"duration_ms":66409,"temperature":0.7,"pith_summary":"The paper claims that choosing measurement points to maximize expected information gain is the true principle behind sampling, and that the Nyquist-Shannon rule is merely the special case of this principle when no prior knowledge exists. It introduces Bayesian Autocorrelation Spectroscopy (BAS), which sequentially picks interferometer delays in Fourier transform spectroscopy by evaluating how much each candidate measurement would reduce uncertainty in a Gaussian prior over the spectrum. According to the paper, BAS never performs worse than conventional fixed-interval Fourier transform spectroscopy, because if its informed priors fail to predict the data it can fall back to the uninformed baseline; when priors are informative it reaches the same spectral accuracy with far fewer measurements. The authors demonstrate accelerated reconstruction in medical blood-plasma fingerprinting, wavelength-resolved optical vortex characterization, and hyperspectral imaging with a compact RGB-camera interferometer. A reader should care because the claim implies existing spectrometers could become faster and uncertainty-quantified by software changes alone, without new optics.","feed_headline":"Adaptive Bayesian spectroscopy beats fixed Nyquist sampling","feed_subtitle":"Choosing each delay by information gain yields faster spectra and calibrated error bars, never worse than classical sampling.","key_machinery":"The mechanism is BAS, a sequential Bayesian inference scheme that chooses interferometer delays by expected information gain. Its mathematical core is the Gaussian prior-to-posterior update, a rank-1 covariance reduction per measurement, with expected information gain $\\frac{1}{2}\\log\\left|I+R\\Sigma R^T/(\\sigma^2)\\right|$ for each candidate delay. Under a uniform large-variance prior, maximizing this gain is D-optimal experimental design, and the paper cites the classical theorem that equispaced samples are D-optimal for trigonometric models, which is how Nyquist sampling emerges as a limiting case. The other load-bearing mechanism is the sequential model evidence weighting, which scores each candidate prior by how well it predicts each new measurement, so that informed priors are followed when they work and discarded when they do not.","core_discovery":"Bayesian Autocorrelation Spectroscopy treats spectral reconstruction as sequential Bayesian inference: each measurement at delay $\\tau$ is a noisy linear projection $F(\\tau)=\\int S(\\omega)[1+\\cos(\\omega\\tau)]d\\omega$ of the spectrum, and a Gaussian prior over spectral coefficients is updated by Bayes' rule (Eq. 3). The next delay is chosen to maximize the information gain $\\frac{1}{2}\\log(|\\Sigma_{\\mathrm{prior}}|/|\\Sigma_{\\mathrm{posterior}}|)$ (Eq. 4), which reduces to D-optimality and hence to equidistant Nyquist sampling when the prior is a large isotropic covariance. When several candidate priors are available, BAS computes sequential Bayesian evidence for each, weights them by a softmax over log-evidence, and can revert to the uninformed model if informed priors predict the data poorly. On synthetic blood-plasma infrared spectra modeled from real clinical data, adaptive BAS reaches higher signal-to-noise ratios with fewer measurements than Nyquist sampling; on vortex beams it produces wavelength-resolved phase with fully propagated covariances; and combined with an RGB filter array and a liquid-crystal retarder it reconstructs hyperspectral data from a handful of delay steps.","pith_inferences":["A natural next step would be to use the same evidence machinery to choose between sensor modalities (which color channel, which delay, which detector) as well as within a modality, turning the RGB-plus-retarder demo into a general active-sensor fusion rule.","The 'never below FTS' guarantee is asserted from simulations rather than proven; a rigorous version would need a non-asymptotic bound on the softmax model selection, including the temperature parameter.","If the framework transfers as claimed, any instrument whose observations are linear projections of a field, such as optical coherence tomography, X-ray ptychography, or NMR, could adopt the same adaptive scheme using a covariance prior.","The blood-plasma gain is plausibly upper-bounded by the fact that the test spectra were generated from the same multivariate Gaussian fits that define the priors; prospective real FTIR data would show how much of the gain survives distribution shift."],"forward_implications":["Standard Fourier transform spectroscopy is mathematically a limiting case of BAS, so upgrading an existing FTIR instrument to adaptive selection requires software, not new hardware, to reach the same spectrum faster.","When a clinically relevant spectral prior is available, adaptive BAS needs substantially fewer delay measurements to reach a given SNR, which shortens measurement time for high-throughput blood analysis.","Because BAS returns a full spectral covariance matrix, downstream analyses such as vortex phase retrieval can propagate correlated uncertainties in closed form, replacing point estimates with calibrated error bars.","Combining coarse RGB color responses with a few interferometric delays gives a compact hyperspectral imager that keeps full spatial resolution, a capability that does not exist for fixed Nyquist-style RGB sampling alone."],"supporting_citations":[{"why":"Supplies the D-optimality theorem that equispaced sampling maximizes the Fisher information determinant for trigonometric models, the step that makes Nyquist sampling emerge as the uninformed optimum.","marker":"[7]"},{"why":"Provides the real infrared blood-plasma clinical data from which the synthetic molecular-fingerprinting spectra and their priors are modeled.","marker":"[17]"},{"why":"Supplies the time-frequency logon cells used to argue that a band- and time-limited signal has roughly 2BT independent degrees of freedom.","marker":"[6]"},{"why":"Gives the prolate-spheroidal eigenanalysis showing that about 2BT eigenvalues dominate for time- and band-limited signals.","marker":"[40]"},{"why":"Establishes that such signals can be approximated by 2BT basis terms, justifying the finite trigonometric model on which BAS updates run.","marker":"[42]"},{"why":"Supplies the regularity result that the posterior inherits the smoothness of the prior, used to justify structural RBF priors as well-posed regularizers.","marker":"[45]"},{"why":"Motivates the whole approach by showing that prior structure can reduce the number of samples needed below the Nyquist rate.","marker":"[3]"},{"why":"Provides the classical bandlimited sampling theorem that defines the baseline regime BAS generalizes.","marker":"[1]"}],"fun_headline_variants":["Adaptive Bayesian spectroscopy beats Nyquist sampling","Information-optimal measurement: adaptive beats fixed sampling","Bayesian adaptive sampling: richer spectra with fewer delays","Adaptive measurement outshines fixed sampling in spectroscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The performance floor (BAS never worse than FTS) rests on the unproven assumption that the softmax-weighted evidence across the small, hand-chosen set of prior models can detect a misleading prior before it has corrupted the estimate; an overconfident prior that earns high early weights could push the method below the uninformed baseline.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive Bayesian spectroscopy beats Nyquist sampling","Information-optimal measurement: adaptive beats fixed sampling","Bayesian adaptive sampling: richer spectra with fewer delays","Adaptive measurement outshines fixed sampling in spectroscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1517,"prompt_tokens":911,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":546}},"tokens_in":527,"tokens_out":606,"duration_ms":7129,"temperature":1.0,"reasoning_tokens":546,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:35:33.272869+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BAS on spectra drawn from a distribution deliberately different from all candidate priors, with an overconfident prior that initially predicts early measurements well, and compare the cumulative reconstruction error to an uninformed Nyquist-sampled baseline; if the weighted BAS estimate remains below the baseline after enough measurements for the evidence to switch, the asserted floor fails. A second direct test is to run the molecular-fingerprinting pipeline on real (not synthetic) blood-plasma interferograms and compare adaptive BAS against standard FTS.","supporting_citations":[{"cited_title":"Pukelsheim,Optimal Design of Experiments(Society for Industrial and Applied Mathematics, Philadelphia, Pa, 2006)","cited_arxiv_id":null,"evidence_quote":"Supplies the D-optimality theorem that equispaced sampling maximizes the Fisher information determinant for trigonometric models, the step that makes Nyquist sampling emerge as the uninformed optimum."},{"cited_title":"Huber, K","cited_arxiv_id":null,"evidence_quote":"Provides the real infrared blood-plasma clinical data from which the synthetic molecular-fingerprinting spectra and their priors are modeled."},{"cited_title":"Gabor, Journal of the Institution of Electrical Engineers-part III: radio and communication engineering 93, 429 (1946)","cited_arxiv_id":null,"evidence_quote":"Supplies the time-frequency logon cells used to argue that a band- and time-limited signal has roughly 2BT independent degrees of freedom."},{"cited_title":"Slepian, Proceedings of the IEEE64, 292 (1976)","cited_arxiv_id":null,"evidence_quote":"Gives the prolate-spheroidal eigenanalysis showing that about 2BT eigenvalues dominate for time- and band-limited signals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that such signals can be approximated by 2BT basis terms, justifying the finite trigonometric model on which BAS updates run."},{"cited_title":"Latz, SIAM Review65, 831 (2023), publisher: Society for Industrial and Applied Mathematics","cited_arxiv_id":null,"evidence_quote":"Supplies the regularity result that the posterior inherits the smoothness of the prior, used to justify structural RBF priors as well-posed regularizers."},{"cited_title":"Shannon, Proceedings of the IRE37, 10 (1949)","cited_arxiv_id":null,"evidence_quote":"Provides the classical bandlimited sampling theorem that defines the baseline regime BAS generalizes."}],"review_version":1}