{"id":"4568143c-6c68-4876-9379-1053603b06c9","arxiv_id":"1908.07758","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A curvature-constrained variant of SCSA, called C-SCSA, denoises pulse-shaped and general signals while preserving peak shape, without requiring peak location or noise-interval priors.","lead":"This paper combines the Semi-Classical Signal Analysis denoising method with a curvature penalty, creating C-SCSA, which smooths pulse-shaped signals without needing to know peak positions in advance. In tests on synthetic peaks, ECG records with real noise, and standard test signals, C-SCSA usually matches or beats Savitzky-Golay, wavelet, and EMD baselines while preserving peak height and width.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The curvature penalty in C-SCSA is applied to y_h, while Proposition 1 only justifies noise-driven curvature growth for the input yδ; the theoretical link is missing and peak preservation is therefore unsupported.","rationale":"The reader's verdict CONDITIONAL is well-calibrated. I re-read the derivation around Eqs. (7)-(9) and the cost in Eq. (10). Proposition 1 is internally consistent (up to a sign typo in the appendix, where 'E{k} =−' should be positive) and does show that the expected curvature of first differences of the input grows with noise variance. But the cost function's penalty is applied to y_h, so the proposition does not license the algorithm. This is the load-bearing gap: without it, the method is an ad hoc smoothing heuristic with an attractive name. The empirical comparisons are broad, and the ECG table plus the general-signal figures give some support, but the lack of code, oracle parameter selection, and the unexplained positivity issue further weaken the guarantee. A conditional accept remains appropriate: the paper should either supply the missing link or explicitly reframe the curvature penalty as a heuristic. My concrete test would either close the gap (if E{k(y_h)} is monotone in σ² and h_J tracks h_MSE) or expose it. I do not see a reason to move to reject, because the empirical results, though possibly overfitted, are plausible and the method may still work in practice.","tokens_in":11406,"tokens_out":8224,"duration_ms":79226,"concrete_test":"On the single-Gaussian-peak setup of §4.2, for noise levels 1%, 5%, and 12%, scan h over a fine grid and compute J̄(h)=Σ[yδ-y_h]² + μΣk(y_h), recording the minimizer h_J and the MSE-optimal h_MSE. Then recompute the penalty with the input's curvature, J_in(h)=Σ[yδ-y_h]² + μΣk(yδ) (whose curvature term is h-independent), and find h_in. If h_J differs from h_MSE by more than one grid step, or if peak-height preservation at h_J is worse than at h_MSE, the cost function is not implementing the theorem's intent. Independently, re-derive E{k(y_h)} with y_h in place of yδ under the same noise model; if it is not monotone increasing in σ², the theoretical bridge collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is the curvature constraint in Eq. (10), J̄ = Σ[yδ(t_i)-y_h(t_i)]² + μ Σk(t_i), where k(t_i) is the curvature of the reconstructed signal y_h (Eq. 6). The theoretical support, Proposition 1 (Eqs. 7-9), computes the expected curvature k_m from adjacent differences x_m=y_{m+1}-y_m and w_m=y_m-y_{m-1} of the noisy input yδ, under a zero-mean jointly Gaussian model that is only plausible in flat regions. The paper then asserts that since E{k_m} increases with noise variance σ², one should penalize k in the cost. This is a non-sequitur: the monotonicity is derived for the input's noisy differences, not for the curvature of the SCSA reconstruction y_h. Penalizing y_h's curvature does not follow from, and can even oppose, the theorem—peaks are high-curvature features, so the penalty may suppress the very peaks the method claims to preserve. No lemma transfers the monotonicity from yδ to y_h. Consequently, the 'guarantee' of peak preservation is an empirical observation, not a proven property, and the reported success could depend on the particular μ/h protocol rather than on the curvature constraint. The ECG results are also suspect because SCSA is defined for real positive potentials (Section 2.1) while ECG records contain negative deflections; the paper never states how this is handled. Together these gaps make the central claim conditional at best.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes C-SCSA, a denoising method that augments semi-classical signal analysis (SCSA) with a curvature penalty. Section 3 defines the cost function Jbar = sum_i [y_delta(t_i) - y_h(t_i)]^2 + mu sum_i k(t_i), where k(t_i) is the curvature of the SCSA reconstruction y_h, and Proposition 1 computes the expected curvature of adjacent differences of the noisy input under a zero-mean jointly Gaussian model, showing that this expectation increases with the noise variance. The algorithm then scans the semi-classical parameter h to minimize Jbar. Section 4 evaluates C-SCSA against Savitzky-Golay, wavelet, and EMD-IT denoising on single- and multi-peak Gaussian signals, ECG records corrupted with real noise from the MIT-BIH database, and standard test signals (Doppler, piecewise-regular, blocks, bumps) at multiple noise levels and sampling frequencies. The paper claims that C-SCSA preserves peak height and width and outperforms or matches the baselines.","tokens_in":11715,"tokens_out":8333,"duration_ms":73213,"significance":"The paper addresses a practically relevant problem: denoising pulse-shaped signals without peak-localization priors. The C-SCSA cost function is a reasonable extension of SCSA, and the simulation protocol is detailed, including multiple signals, noise levels, sampling frequencies, and a real-noise ECG test against three baselines. Proposition 1 is a nontrivial calculation, and the appendix derivation is internally consistent. However, the central theoretical link from noise-driven curvature of the input to a penalty on the reconstruction's curvature is not established, and the parameter selection uses oracle knowledge of the clean signal. As presented, the claimed guarantee of peak preservation is not supported, so the significance is conditional on additional analysis and validation.","major_comments":[{"comment":"The theoretical support for the curvature penalty is derived for the adjacent differences x_m and w_m of the noisy input y_delta under a zero-mean jointly Gaussian model (Eq. 7), but the penalty in Eq. (10) is applied to the curvature k(t_i) of the SCSA reconstruction y_h. The paper does not provide a lemma or estimate showing that penalizing the reconstruction's curvature is equivalent to, or even correlated with, reducing the input curvature caused by noise. Because signal peaks have high curvature, the penalty can suppress the very features the method claims to preserve; therefore the statement in Section 5 that C-SCSA 'guarantees the peaks are well preserved' is not justified by Proposition 1. Either supply a transfer argument relating curvature of y_h to noise in y_delta, or reword the claim and support peak preservation empirically over a range of mu and h.","section":"Section 3, Eqs. (5)-(10) and Proposition 1"},{"comment":"The description 'we optimize each method's parameter at noise level 5%, where we iteratively optimize its parameter using the noisy signal and true signal' means that mu (and h) for C-SCSA, as well as baseline parameters, are chosen with the true clean signal. This is oracle tuning: it does not correspond to a deployable procedure when the clean signal is unknown, and it weakens the claim that C-SCSA does not need to locate signal peaks or know the noise level. The paper should either specify a fully data-driven selection rule for mu (the current rule involving max{|y_delta|} * sum k(i) * 10^nu leaves nu unspecified) or present the comparison as an oracle-parameter study and discuss behavior when parameters are selected without the clean signal.","section":"Section 4.2, parameter selection paragraph"},{"comment":"SCSA is defined for a real positive potential y(t) (Section 2.1 and Eq. 2), yet the ECG records shown in Fig. 5 have amplitudes ranging from about -200 to 400, i.e., they contain negative values. The paper never states how the potential is made positive before computing the Schrodinger spectrum. If a baseline shift is applied, it should be described and its effect on the reported SNR improvements quantified; without this, the eigenvalue problem and the reconstruction in Eq. (1) are not well defined for these signals.","section":"Section 4.3, ECG experiments"},{"comment":"The peak-preservation experiment is performed on a single simulated Gaussian peak, and no repeated noise realizations or error bars are reported for the peak-height and peak-width relative errors; the text elsewhere reports ensemble averages over 90 realizations. Consequently, the claimed superiority of C-SCSA in peak preservation could reflect a single noise draw rather than a stable advantage. Repeating the experiment over multiple noise realizations and reporting mean plus or minus standard deviation (or box plots) is necessary to support the central peak-preserving claim.","section":"Section 4.2, Fig. 3"}],"minor_comments":[{"comment":"The summation lower limit is written as 'i = i' and should be 'i = 1'.","section":"Eq. (10)"},{"comment":"The sign of the alpha * |SNR_yh| term in the alpha-SCSA cost appears inconsistent with minimization: if a larger SNR is desirable, adding a positive SNR term increases the cost for better reconstructions. Clarify whether the term should be negative or should be alpha / SNR_yh.","section":"Eq. (4)"},{"comment":"The caption states that the noise level ranges between 0.1% and 15% with interval 0.1%, but the horizontal axis shows values from 1 to 10, and the preceding paragraph describes a 1% to 12% range. Please make the range and units consistent.","section":"Fig. 4 caption"},{"comment":"The x-axis labels are inconsistent with the experiments: Fig. 6 labels the x-axis 'Noise Level (%)' while the text indicates noise levels 7%-15%, and Fig. 7 panels show both 'Noise Level (%)' and 'Number of Samples' on the x-axis although the text describes a sampling-frequency experiment. Correct all axis labels, legends, and figure cross-references.","section":"Figs. 6 and 7 captions"},{"comment":"Both panels (b) and (c) in the caption are described as 'C-SCSA denoising', but the surrounding text states that alpha-SCSA failed on the piecewise-regular signal. Clarify which panel shows alpha-SCSA and which shows C-SCSA.","section":"Fig. 8 caption"},{"comment":"The text cites 'Fig. 7(a) and (d)' and 'Fig. 8(a)(b)(c)(d)' in a way that does not match the figures' content; all figure cross-references should be checked and updated.","section":"Section 4.4, cross-references"},{"comment":"The Savitzky-Golay parameters differ between Fig. 3 (window 29, degree 4) and Fig. 4 (filter length 17, order 4); the manuscript should explain why different settings are used in the two experiments or use a single configuration.","section":"Section 4.2, Savitzky-Golay settings"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision. The manuscript is within the scope of the journal, and the authors' prior SCSA work is appropriately cited, but the central contribution needs a genuine argument linking the curvature of the reconstruction to the noise in the input. The oracle parameter selection should be disclosed in the evaluation section, and the ECG positivity issue must be addressed. I see no conflict of interest."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful incremental paper, not a breakthrough. The new bit is C-SCSA—using a curvature penalty to pick the semi-classical parameter h, instead of the peak-localization and noise-region priors that α-SCSA needed. That is a real practical improvement for pulse-shaped biomedical signals, and the paper backs it with comparisons against Savitzky-Golay, wavelet, and EMD on Gaussian peaks, ECG with real MIT-BIH noise, and several benchmark signals at different sampling rates. The derivation of Proposition 1 checks out: under the zero-mean jointly Gaussian model for adjacent differences, the expected curvature of the noisy input increases with noise variance. The citation pattern is fine too—the authors lean on their own prior SCSA work, which is expected when the whole method builds on it.\n\nThe main soft spot is exactly the one flagged in the stress test. The theorem is about curvature of the noisy input yδ, while the penalty in Eq. (10) is applied to the SCSA reconstruction y_h. Penalizing y_h's curvature does not follow from the noise-driven increase in yδ's curvature. Peaks are high-curvature features, so a curvature penalty can in principle suppress them; the \"guarantee\" of peak preservation is not established by the proposition. The simulation evidence suggests the method preserves peaks in the tested cases, but that is an empirical claim, and the paper should say so. This is a genuine gap, but not fatal to the paper's practical value: the penalty can be read as a standard smoothness prior, and the empirical results are what they are.\n\nThe other concerns are smaller. Parameters for C-SCSA and the baselines are tuned at 5% noise using the true clean signal, then fixed across noise levels; that is acceptable for a controlled comparison, but the reported advantage is an upper bound on what users should expect. There are no error bars on the peak-preservation curves in Fig. 3, and no code, so reproducibility is limited. SCSA is introduced for positive potentials, and ECG records contain negative deflections; the paper never says how this is handled. That last point needs to be addressed in revision.\n\nWho is this for? Someone working on SCSA or peak-preserving denoising of MRS/ECG/pulse signals will find the algorithm and comparison useful. It deserves a serious referee, but with the expectation of major revision: reframe the theoretical claim, address the positivity issue, release code or at least detailed pseudo-code, and report variance over noise realizations.","headline":"C-SCSA is a plausible incremental extension of the authors' SCSA denoiser, but the curvature-penalty justification does not actually transfer from the noisy input to the reconstruction, so the peak-preservation claim should be read as an empirical observation, not a proven guarantee.","tokens_in":12236,"tokens_out":4515,"would_cite":false,"duration_ms":43262,"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":"A curvature penalty lets Schrödinger-operator denoising preserve pulse peaks without requiring peak locations or noise-region priors.","keywords":["signal denoising","Schrödinger operator","semi-classical signal analysis","curvature constraint","peak preservation","eigenvalue spectrum","ECG denoising","semiclassical parameter"],"falsifier":"Generate a signal with sharp true spikes whose curvature is comparable to or larger than the curvature of the added noise, run C-SCSA at noise levels from 1% to 12%, and measure peak height and width relative errors. If the errors are as large as those of Savitzky-Golay filtering, the claim that C-SCSA guarantees peak preservation fails. Alternatively, compare the curvature of the chosen $y_h$ to the curvature of the true signal: if the reconstructed curvature is systematically lower for signals with genuinely high-curvature features, the curvature penalty is over-smoothing beyond noise removal.","tokens_in":11210,"feed_emoji":"📉","tokens_out":5853,"duration_ms":150316,"temperature":0.7,"pith_summary":"This paper tries to establish that a denoising method built on the Schrödinger operator's eigenspectrum can be made general by adding a curvature constraint. Earlier SCSA denoising required knowing where the signal peaks are and where noise dominates, which restricted it to specialized applications such as magnetic resonance spectroscopy. The proposed C-SCSA replaces that prior knowledge with a penalty on the wiggliness of the reconstructed signal, and the reported experiments indicate that it preserves peak height and width better than Savitzky-Golay, wavelet, and EMD baselines while delivering competitive signal-to-noise ratios. If correct, this offers a peak-preserving denoising strategy for pulse-shaped and slowly varying biomedical signals that needs no manual peak annotation.","feed_headline":"Curvature penalty tunes Schrödinger denoiser without peak priors","feed_subtitle":"A Schrödinger-operator denoiser scans a curvature penalty to keep pulse peaks intact.","key_machinery":"The central object is the Schrödinger operator $-h^2 \\frac{d^2\\psi}{dt^2} - y(t)\\psi = \\lambda\\psi$, with the noisy signal acting as the potential. Its negative eigenvalues $\\lambda_{nh} = -\\kappa_{nh}^2$ and the associated $L^2$-normalized eigenfunctions $\\psi_{nh}$ are used to reconstruct the signal as $y_h(t) = 4h \\sum_{n=1}^{N_h} \\kappa_{nh} \\psi_{nh}^2(t)$. The semiclassical parameter $h$ controls how many eigenfunctions are retained, and the curvature constraint $k(t) = |y_h''(t)|/(1 + y_h'(t)^2)^{3/2}$ is the mechanism that selects $h$. The load-bearing identity is Proposition 1's formula, which gives the expected curvature of adjacent differences as an integral that grows with the noise variance, thereby connecting larger noise to larger curvature and motivating the penalty term in the cost function.","core_discovery":"The central claim is that selecting the semiclassical parameter $h$ by minimizing a cost function that combines fidelity to the noisy signal with a curvature penalty on the reconstruction yields a denoiser that removes noise without flattening peaks. Specifically, the paper defines $J = \\|y_\\delta - y_h\\|_2^2 + \\mu \\int |k(t)|\\,dt$, where $k(t)$ is the curvature of the reconstructed signal, and then scans $h$ to minimize this cost. The supporting theoretical result, Proposition 1, derives an expression for the expected curvature of adjacent differences in the noisy signal and shows that this expectation increases with the noise variance; the paper uses this to justify penalizing curvature as a noise-reduction mechanism. Numerically, C-SCSA reports lower peak-height and peak-width relative errors than the comparison methods on single Gaussian peaks, and higher or comparable post-denoising SNR on simulated multi-peak signals, real-noise ECG records, and several standard test signals, especially when the signal is regular and slowly varying.","pith_inferences":["The curvature-based rule suggests a general principle for spectral denoising: choose the decomposition threshold by minimizing fidelity plus a geometric regularity term, which could also be applied to wavelet coefficient selection or EMD interval thresholding.","Proposition 1's relation between adjacent-difference curvature and noise variance could be inverted to estimate the noise level from a single noisy record, allowing the smoothing parameter $\\mu$ to be set automatically rather than by an order-of-magnitude scan.","The method's dependence on reconstruction curvature implies it may over-smooth signals whose true features have curvature comparable to noise, such as very sharp spikes or edges; testing on such signals would reveal the boundary of the peak-preservation claim.","Viewing the curvature penalty as a form of total-variation or mean-curvature flow could connect C-SCSA to nonlinear diffusion denoising and potentially yield convergence guarantees that the paper does not provide."],"forward_implications":["C-SCSA can be applied to pulse-shaped signals without needing to locate peaks or identify noise-dominant intervals, unlike the earlier α-SCSA method.","On single Gaussian peaks with 1–12% Gaussian noise, C-SCSA reports lower peak-height and peak-width relative errors than EMD-IT, Savitzky-Golay, and wavelet denoising.","On multi-peak simulated signals and real-noise ECG records from the MIT-BIH database, C-SCSA achieves SNR after denoising that is competitive with or better than Savitzky-Golay and EMD-IT, and comparable to the wavelet method.","On piecewise-regular, Doppler, blocks, bumps, and sing test signals, C-SCSA outperforms EMD-IT and Savitzky-Golay for regular and slowly varying cases, and remains effective across sampling frequencies from 256 to 2048 samples.","Because the cost function operates on the whole signal, C-SCSA extends SCSA-based denoising beyond magnetic resonance spectroscopy to a broader class of biomedical and spectral signals."],"supporting_citations":[{"why":"Provides the original α-SCSA denoising method and its cost function that requires peak localization and noise-region knowledge, which C-SCSA extends.","marker":"[1]"},{"why":"Introduces the semi-classical signal analysis reconstruction formula and the Schrödinger operator framework used throughout.","marker":"[2]"},{"why":"Supplies the smoothing-spline style penalty inspiration for adding a curvature term to the cost function.","marker":"[8]"},{"why":"Supports the assumption that curvature-based penalties respond to noise in homogeneous regions, grounding Proposition 1's noise model.","marker":"[9]"},{"why":"Defines the Savitzky-Golay filter, one of the main baseline methods compared against C-SCSA.","marker":"[12]"},{"why":"Provides the PhysioNet database and the real noise records used in the ECG denoising experiments.","marker":"[17]"},{"why":"Supplies the EMD interval-thresholding denoising method used as a baseline comparison.","marker":"[18]"}],"fun_headline_variants":["Curvature cost tunes Schrödinger denoiser for pulse peaks","C-SCSA: curvature-constrained Schrödinger denoising","Schrödinger denoiser adds curvature check to keep peaks","Curvature penalty sharpens Schrödinger-based denoising"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof shows that adjacent differences in the noisy input have larger expected curvature when noise increases, but the algorithm actually penalizes the curvature of the reconstructed signal; the paper does not prove that these two curvatures are equivalent or that minimizing the latter removes the former.","fun_headline_variants_meta":{"raw":{"variants":["Curvature cost tunes Schrödinger denoiser for pulse peaks","C-SCSA: curvature-constrained Schrödinger denoising","Schrödinger denoiser adds curvature check to keep peaks","Curvature penalty sharpens Schrödinger-based denoising"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1351,"prompt_tokens":876,"completion_tokens":475,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":492,"completion_tokens_details":{"reasoning_tokens":403}},"tokens_in":492,"tokens_out":475,"duration_ms":439546,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:57:55.383042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a signal with sharp true spikes whose curvature is comparable to or larger than the curvature of the added noise, run C-SCSA at noise levels from 1% to 12%, and measure peak height and width relative errors. If the errors are as large as those of Savitzky-Golay filtering, the claim that C-SCSA guarantees peak preservation fails. Alternatively, compare the curvature of the chosen $y_h$ to the curvature of the true signal: if the reconstructed curvature is systematically lower for signals with genuinely high-curvature features, the curvature penalty is over-smoothing beyond noise removal.","supporting_citations":[{"cited_title":"Spectral data de-noising using semi- classical signal analysis: application to localized MRS","cited_arxiv_id":null,"evidence_quote":"Provides the original α-SCSA denoising method and its cost function that requires peak localization and noise-region knowledge, which C-SCSA extends."},{"cited_title":"Semi-classical signal analysis","cited_arxiv_id":null,"evidence_quote":"Introduces the semi-classical signal analysis reconstruction formula and the Schrödinger operator framework used throughout."},{"cited_title":"Hastie, R","cited_arxiv_id":null,"evidence_quote":"Supplies the smoothing-spline style penalty inspiration for adding a curvature term to the cost function."},{"cited_title":"Denoising an image by denoising its curvature image","cited_arxiv_id":null,"evidence_quote":"Supports the assumption that curvature-based penalties respond to noise in homogeneous regions, grounding Proposition 1's noise model."},{"cited_title":"Smoothing and Differentiation of Data by Simplified Least Squares Procedures","cited_arxiv_id":null,"evidence_quote":"Defines the Savitzky-Golay filter, one of the main baseline methods compared against C-SCSA."},{"cited_title":"PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals","cited_arxiv_id":null,"evidence_quote":"Provides the PhysioNet database and the real noise records used in the ECG denoising experiments."},{"cited_title":"Development of EMD-Based Denoising Methods Inspired by Wavelet Thresholding,","cited_arxiv_id":null,"evidence_quote":"Supplies the EMD interval-thresholding denoising method used as a baseline comparison."}],"review_version":1}