{"id":"6b7cd186-5bfa-46b2-8473-09d7ea2310e3","arxiv_id":"1908.07151","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":1,"one_line_summary":"Trend filtering, a locally adaptive nonparametric smoother, is introduced to astronomy because it can denoise heterogeneous signals (smooth trends plus sharp features) better than linear smoothers.","lead":"This paper introduces trend filtering, a nonparametric denoising method, to astronomy for signals that mix smooth trends with sharp features. The method is claimed to outperform linear smoothers such as splines and Gaussian process regression on such signals while remaining fast and easy to tune.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's unconditional superiority claim goes beyond the minimax theorem: the guarantee holds only over L1 Sobolev classes in a worst-case, asymptotic sense, and real astronomical signals are not shown to lie in that class.","rationale":"Reading the paper in good faith, it is a competent introduction of an existing statistical method to astronomy, and the cited minimax theory is real. The problem is the gap between the abstract's unconditional wording and the actual content of the theorems. The theorems establish worst-case asymptotic optimality over L1 Sobolev classes, not universal finite-sample superiority on arbitrary heterogeneous signals. Section 2.1.2 explicitly disclaims mathematical precision in the central definition, and Section 2.2 provides only a single simulated example with no quantitative error statistics. This makes the reader's identified weakest assumption the key issue: the practical superiority claim depends on astronomical signals being well-represented by the function class for which trend filtering is minimax optimal. That premise is neither proved nor tested. The scalability claim is also imported from prior work, though it is a less central part of the paper's contribution. Since the paper should qualify the abstract and provide an astronomical-realism benchmark or a precise statement of the class of signals covered, the conditional verdict is appropriate. My analysis does not move the verdict; it reinforces the need for the stated conditions.","tokens_in":19146,"tokens_out":7316,"duration_ms":78356,"concrete_test":"Compile a benchmark library of high-SNR astronomical signals with known ground truth, such as BOSS quasar spectra and simulated supernova or exoplanet-transit light curves. For each signal, add realistic heteroskedastic Gaussian noise, then compare quadratic trend filtering tuned by SURE or 10-fold cross-validation against LOESS, smoothing splines, and Gaussian process regression tuned by the same cross-validation procedure on the same folds. Compute the distribution of MSPE ratios over at least 50 signals and multiple noise realizations, at n=10^4 and n=10^6. If tuned linear smoothers win on a substantial fraction of cases, or if trend filtering's advantage observed in Figures 1-2 disappears, the abstract's unconditional superiority claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is in Section 2.1.2, where 'spatially heterogeneous' is effectively identified with membership in the L1 Sobolev class F1,k (Eq. 14), and the definition is admitted to be 'intentionally mathematically imprecise.' The abstract then asserts without qualification that trend filtering is 'superior to any statistical estimator that is a linear combination of the observed data.' The cited results (Nemirovskii et al. 1985; Tibshirani 2014) are minimax statements: over F1,k, linear smoothers are suboptimal while trend filtering achieves the optimal worst-case rate (e.g., n^{-1/2} vs n^{-2/3} for k=0). These are asymptotic, worst-case guarantees over a specific function class. They do not imply lower MSPE than a well-tuned linear smoother for a given fixed signal, and they say nothing about signals outside F1,k. The empirical section uses a single hand-constructed signal and reports no quantitative error comparison or noise-averaged uncertainties. Thus the practical claim that trend filtering will be superior on astronomical light curves and spectra rests on an unverified premise: that such signals are well represented by F1,k and that rate-optimality translates into finite-sample advantage. A secondary imported claim, scalability to n>=10^7, is taken from Ramdas and Tibshirani (2016) rather than demonstrated here, but the superiority claim is the central issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces trend filtering to the astronomical community as a locally adaptive nonparametric denoiser for one-dimensional signals with spatially varying smoothness. The authors argue in Section 2 that linear smoothers are suboptimal for such signals, citing minimax results of Nemirovskii et al. and Tibshirani, and present a simulation comparing trend filtering with LOESS, smoothing splines, and Gaussian process regression at fixed effective degrees of freedom. Section 3 provides a tutorial-style description of trend filtering, its relationship to splines, software, hyperparameter selection via SURE and cross-validation, bootstrap uncertainty quantification, and a relaxed variant. The paper's concrete contributions are the astronomical exposition, the SURE selection recipe, three bootstrap algorithms, and supplementary R code.","tokens_in":19447,"tokens_out":5480,"duration_ms":54173,"significance":"If the claims are taken with appropriate qualifications, the paper addresses a genuine need: many astronomical time series and spectra combine smooth continua with sharp features, and standard linear smoothers are not locally adaptive. Trend filtering is a principled alternative with a minimax optimality guarantee over L1 Sobolev classes, flexible sampling requirements, a single tuning parameter, and scalable algorithms. The paper also ships code and clear recipes. However, the manuscript currently overstates the practical meaning of the minimax results and does not quantitatively demonstrate the finite-sample advantage it claims; these are fixable within the scope of the paper.","major_comments":[{"comment":"The abstract's claim that 'when the underlying signal is spatially heterogeneous, trend filtering is superior to any statistical estimator that is a linear combination of the observed data' is stated without qualification. The supporting theory in Section 2.1.2 is a minimax result over the L1 Sobolev class F1,k in Eq. (14): trend filtering achieves the optimal worst-case rate, while linear smoothers do not. This is a worst-case, asymptotic statement over a specific function class, not a guarantee of lower error for every heterogeneous signal. The manuscript itself says the definition of 'spatially heterogeneous' is 'intentionally mathematically imprecise' and notes that L1 Sobolev classes contain spatially homogeneous signals as well, so the abstract overstates the theorem. Please rephrase the abstract to say 'minimax optimal over L1 Sobolev classes (signals with derivatives of bounded variation)' and treat the representativeness of real astronomical signals as a stated, testable premise.","section":"Abstract; Section 2.1.2"},{"comment":"The empirical comparison is visual only. The text reports that trend filtering with 55 effective degrees of freedom is sufficient while LOESS, smoothing splines, and Gaussian process regression require 192 to recover the localized features, but no quantitative error metric (e.g., mean squared prediction error) is reported, no averaging over independent noise realizations is shown, and no uncertainty is attached to the comparison. This single hand-constructed signal is offered as evidence of 'clear practical consequences' for the theoretical gap. The argument would be much stronger with a table of error metrics computed over repeated simulations and, ideally, over several signal shapes.","section":"Section 2.2; Figures 1 and 2"},{"comment":"The scalability claim that trend filtering handles n ≳ 10^7 within a few minutes appears in the abstract and in Section 3.2, but it is a summary of Ramdas and Tibshirani (2016) and is not reproduced or benchmarked here. Because practical scalability is one of the paper's advertised properties (P3 in Section 1), either provide a small timing benchmark in this paper or explicitly mark the abstract statement as a claim from the cited literature.","section":"Section 3.2; Abstract"}],"minor_comments":[{"comment":"The text refers to 'Table 3.4' for software links, but the table in Section 3.4 is numbered Table 2; please correct the cross-reference.","section":"Section 4"},{"comment":"The caption says the constant trend filtering estimate is discontinuous at the knots but is 'interpolate[d] here for visual purposes'; this should be clarified in the main text so that readers do not mistake the plotted curve for the actual piecewise-constant estimator.","section":"Figure 3 caption"},{"comment":"The binomial coefficient in the variable-knot regression splines row is printed as n!/(n!(n−p)!) but should be n!/(p!(n−p)!).","section":"Table 1"},{"comment":"The radial basis functions φ_j in the simulation are not specified; please state their exact form, width, and location so the simulation is reproducible.","section":"Equation (16)"}],"recommendation":"major_revision","confidential_remarks":"This is a useful methods tutorial for the astronomical community, and the theoretical foundation is externally well established. The main editorial risk is the abstract's unqualified superiority claim, which goes beyond the cited minimax theorems; the proposed revisions should make the scope explicit and add quantitative empirical support. The paper fits MNRAS's scope and does not appear to misrepresent prior work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this is a well-written tutorial introducing trend filtering to astronomy, not a new statistical discovery. The minimax results are Tibshirani's and Nemirovskii's, and the authors are transparent about that. For an astronomy audience, the paper has real value—it explains why linear smoothers fail on signals with sharp localized features, gives clear recommendations on hyperparameter selection via SURE and cross-validation, and ships R code for tuning and bootstrap uncertainty. That is genuine service to the community.\n\nThe paper does well at what it sets out to do. The derivation of trend filtering as an L1 analog of smoothing splines is clear. The falling-factorial basis and the generalized lasso formulation are laid out correctly. The simulation with equal effective degrees of freedom makes the oversmoothing-vs-overfitting tradeoff visually concrete. I especially like the explicit advice to use quantile-based bootstrap bands because the estimator's sampling distribution is non-Gaussian. That is the kind of practical detail that saves people from mistakes.\n\nSoft spots, in proportion. The abstract overstates the result. It says trend filtering is 'superior to any statistical estimator that is a linear combination of the observed data' without saying 'on L1 Sobolev classes, in a minimax sense.' That matters: the guarantee is asymptotic worst-case over a function class, and real astronomical signals have not been shown to lie in that class. The paper itself admits the definition of 'spatially heterogeneous' is intentionally imprecise. So the practical superiority on real light curves and spectra is plausible but unproven here. The empirical section is also only visual—one hand-built signal, no error bars, no quantitative MSPE comparison. That is fine for a tutorial but not evidence for a universal claim. The n≥10^7 speed claim is cited from Ramdas and Tibshirani, not reproduced; the paper does not misrepresent it as new, but it does repeat it confidently.\n\nThe relaxed trend filtering variant is presented as new but is only a sketched heuristic. That is okay, but it is not evaluated, and the reader should not treat it as validated. Minor.\n\nAll in, the central claim is sound as a transfer of known theory. It is not circular to cite the minimax results; the authors are honest about provenance. Would I referee this? Yes. It deserves a serious referee, mostly to sand down the abstract and strengthen the empirical section. The type of reader who benefits is an astronomer who wants a principled default denoiser and is tired of hand-tuning kernel widths. I would probably use this paper's code and cite it when applying trend filtering to survey data.","headline":"A solid, honest methods-transfer paper that would land better if the abstract matched the theorem: trend filtering is minimax-optimal over L1 Sobolev classes, not universally superior to linear smoothers.","tokens_in":19973,"tokens_out":1842,"would_cite":true,"duration_ms":19374,"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":"Trend filtering is minimax optimal for spatially heterogeneous signals and beats every linear smoother.","keywords":["trend filtering","spatially heterogeneous signals","minimax optimality","linear smoothers","nonparametric regression","time-domain astronomy","astronomical spectroscopy","L1 Sobolev class"],"falsifier":"Simulate worst-case mean-squared error over a well-specified L1 Sobolev ball of heterogeneous signals, comparing trend filtering against the best linear smoother tuned with oracle knowledge of the signal; if any linear smoother matches trend filtering's $n^{-2/3}$ rate for $k=0$, the paper's central optimality claim is false.","tokens_in":18982,"feed_emoji":"🔭","tokens_out":7964,"duration_ms":75521,"temperature":0.7,"pith_summary":"Trend filtering is a nonparametric method for denoising one-dimensional signals whose smoothness changes across the domain, such as a light curve with a smooth baseline interrupted by transient dips, or a spectrum with a smooth continuum plus sharp emission and absorption lines. The paper argues that for such spatially heterogeneous signals, trend filtering is minimax optimal over the natural L1 Sobolev signal class, while every linear smoother—kernel smoothers, LOESS, smoothing splines, and Gaussian process regression—is provably suboptimal. Concretely, for piecewise constant signals, trend filtering achieves worst-case error rate $n^{-2/3}$, and a linear smoother would need $n^{4/3}$ observations to match it. The paper also shows the estimate can be computed by a specialized convex optimization algorithm that handles $n \\gtrsim 10^7$ points in minutes, with a single hyperparameter chosen data-drivenly. If correct, astronomers gain a practical default denoiser for time series and spectra that combines the adaptivity of wavelets without requiring equally spaced, power-of-two samples.","feed_headline":"Trend filtering beats every linear smoother at denoising bumpy signals","feed_subtitle":"Recovers supernova dips and spectral lines without oversmoothing, and scales to 10 million points.","key_machinery":"The central object is the trend filtering estimator, defined as the solution to $\\min_{m} \\sum_i (f(t_i)-m(t_i))^2 w_i + \\gamma \\sum_i |\\Delta^{(k+1)} m(t_i)| \\Delta t$, where $\\Delta^{(k+1)}$ is the discrete $(k+1)$-st derivative; this is an L1 analog of the smoothing spline problem and an $\\ell^1$ relaxation of variable-knot regression splines. Equivalently, in the falling factorial basis it is a lasso with a single tuning parameter $\\gamma$. The $\\ell^1$ penalty drives most knot coefficients to zero, giving automatic knot selection and local adaptivity, while the generalized lasso structure (identity predictor, banded penalty) permits an ADMM algorithm with near-linear practical complexity. The hyperparameter is selected by Stein's unbiased risk estimate for fixed-input error or 10-fold cross-validation for random-input error, using the exact degrees of freedom formula $\\mathrm{df}(\\hat{f}_0)=E[\\text{number of knots}]+k+1$.","core_discovery":"The central claim is that trend filtering simultaneously meets four requirements no classical method meets: statistical optimality for spatially heterogeneous signals, practical analysis assumptions (unequally spaced inputs, heteroskedastic noise), scalable computation, and a one-dimensional hyperparameter space with automatic selection. The estimator fits a kth-order piecewise polynomial with knots chosen automatically by an $\\ell^1$ penalty on discrete $(k+1)$-st derivatives, i.e. a total-variation penalized least squares problem. The paper's theoretical case relies on results that linear smoothers are uniformly suboptimal over L1 Sobolev classes, while trend filtering attains the minimax rate, e.g. $n^{-2/3}$ for $k=0$ versus at best $n^{-1/2}$ for linear smoothers. The empirical comparison fixes all methods at 55 effective degrees of freedom and shows trend filtering recovering both the smooth trend and sharp localized features, whereas LOESS, smoothing splines, and Gaussian process regression oversmooth the features; the linear smoothers need 192 degrees of freedom to match, and then overfit elsewhere.","pith_inferences":["The same two-stage relaxation idea used here—select knots with an $\\ell^1$ penalty, then refit by weighted least squares—could be applied to other $\\ell^1$-penalized estimators in astronomy to reduce shrinkage bias while keeping automatic knot selection.","If real astronomical signals deviate from L1 Sobolev classes, for instance by being fractal or containing multi-scale noise, the minimax guarantee is not directly binding; a natural extension is to test trend filtering against wavelet denoising on unequally sampled real data, where wavelet optimality is void.","Because trend filtering is nonlinear, its sampling distribution is non-Gaussian even for Gaussian noise; the paper's bootstrap recommendations imply that reporting only Gaussian error bars for denoised light curves will understate uncertainty, and quantile bands should be adopted widely.","The $n^{4/3}$ sample-size equivalence could inform survey design: to detect sharp spectral features at a given significance with linear methods one needs substantially more exposure time, whereas with trend filtering shorter exposures may suffice."],"forward_implications":["Denoising light curves with trend filtering should recover transient events such as supernova rises and exoplanet transits without the oversmoothing and ringing that fixed-bandwidth linear smoothers produce.","Spectroscopic continuum and line flux estimates can be obtained in one fit, since the estimator separates smooth continuum from localized absorption and emission features while accounting for heteroskedastic per-pixel noise.","Survey-scale data sets ($n \\gtrsim 10^7$) can be processed within minutes on a single machine, making the method practical for current and upcoming large surveys.","Users get a single hyperparameter, selected automatically, instead of Gaussian-process kernel parameters or locally varying bandwidths, removing a major source of manual tuning.","The minimax rate gap implies a sample-equivalence result: for $k=0$ heterogeneous signals, a linear smoother needs about $n^{4/3}$ observations to learn as fast as trend filtering, so trend filtering is especially valuable in low-signal or sparse regimes."],"supporting_citations":[{"why":"Introduces trend filtering and proves its minimax optimality over L1 Sobolev and bounded-variation signal classes, which is the paper's core statistical guarantee.","marker":"Tibshirani (2014)"},{"why":"Proves that linear smoothers are uniformly suboptimal over L1 Sobolev classes, establishing the baseline that trend filtering claims to beat.","marker":"Nemirovskii et al. (1985)"},{"why":"Companion result showing that linear estimators cannot achieve the minimax rate on spatially heterogeneous signal classes.","marker":"Nemirovskii (1985)"},{"why":"Characterizes L1 versus L2 Sobolev classes in terms of spatial heterogeneity and supplies wavelet optimality results that motivate the target signal class.","marker":"Donoho & Johnstone (1998)"},{"why":"Provides the specialized ADMM algorithm whose near-linear practical complexity underlies the paper's $n \\gtrsim 10^7$ scalability claim.","marker":"Ramdas & Tibshirani (2016)"},{"why":"Develops higher-order trend filtering as a generalized lasso, supplying the formulation and degrees-of-freedom theory used for hyperparameter selection.","marker":"Tibshirani & Taylor (2011)"}],"fun_headline_variants":["Trend filtering beats all linear smoothers for bumpy signals","Optimal denoising for astronomical signals with trend filtering","Trend filtering recovers spikes and dips scalably to 10M points","Minimax-optimal trend filtering outperforms linear smoothers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central superiority claim is guaranteed only for signals belonging to an L1 Sobolev class, which the paper equates with 'spatially heterogeneous' signals; if real light curves and spectra fall outside that class, the minimax theorem does not ensure trend filtering beats linear smoothers in practice.","fun_headline_variants_meta":{"raw":{"variants":["Trend filtering beats all linear smoothers for bumpy signals","Optimal denoising for astronomical signals with trend filtering","Trend filtering recovers spikes and dips scalably to 10M points","Minimax-optimal trend filtering outperforms linear smoothers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000865,"raw_usage":{"total_tokens":3773,"prompt_tokens":991,"completion_tokens":2782,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2711}},"tokens_in":607,"tokens_out":2782,"duration_ms":19823,"temperature":1.0,"reasoning_tokens":2711,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:23:43.291973+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate worst-case mean-squared error over a well-specified L1 Sobolev ball of heterogeneous signals, comparing trend filtering against the best linear smoother tuned with oracle knowledge of the signal; if any linear smoother matches trend filtering's $n^{-2/3}$ rate for $k=0$, the paper's central optimality claim is false.","supporting_citations":[{"cited_title":"J., 2014, The Annals of Statistics, 42, 285","cited_arxiv_id":null,"evidence_quote":"Introduces trend filtering and proves its minimax optimality over L1 Sobolev and bounded-variation signal classes, which is the paper's core statistical guarantee."},{"cited_title":"L., Johnstone I","cited_arxiv_id":null,"evidence_quote":"Characterizes L1 versus L2 Sobolev classes in terms of spatial heterogeneity and supplies wavelet optimality results that motivate the target signal class."},{"cited_title":"J., 2016, Journal of Computational and Graphical Statistics, 25, 839","cited_arxiv_id":null,"evidence_quote":"Provides the specialized ADMM algorithm whose near-linear practical complexity underlies the paper's $n \\gtrsim 10^7$ scalability claim."},{"cited_title":"J., Taylor J., 2011, The Annals of Statistics, 39, 1335","cited_arxiv_id":null,"evidence_quote":"Develops higher-order trend filtering as a generalized lasso, supplying the formulation and degrees-of-freedom theory used for hyperparameter selection."}],"review_version":1}