{"id":"f772e2ac-74ce-44f4-809e-e7140e8b9a7b","arxiv_id":"2505.15803","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A single wavelet-denoising procedure estimates a time-varying signal's latest point at near-optimal rates, powers one-call ERM under drift, and implies minimax-optimal TV denoising.","lead":"This paper shows that a classic signal-processing tool, wavelet soft-thresholding, can estimate the latest value of a drifting time series with near-optimal error guarantees, without knowing how fast the signal is drifting. The same estimator also gives faster training under temporal distribution shift and reveals new optimal methods for total-variation denoising.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's general-W guarantee is the main soft spot; Haar Theorem 2 and TV Theorem 11 appear intact, so the conditional verdict stands.","rationale":"The reader's weakest assumption correctly identifies the fragile premise: Lemma 1's proof counts log2 n wavelet coefficients affecting theta_hat_1, a property that is true for Haar but not for general wavelet bases. This is a genuine, precise gap in the paper, and it is load-bearing for the claimed generality to higher-order wavelets, for Corollary 3, and for the general-W statement of Theorem 9. However, it does not threaten the strongest central claims as framed by the reader. Theorem 2 is Haar-specific and uses Lemma 14, whose indexing is explicit and correct for Haar; the constant manipulations in its proof, while containing minor notational slips, can be read consistently with the dyadic set S(r). Theorem 11 is a black-box reduction from a pointwise bound of Theorem 2/Corollary 3 type to TV-denoising rates; the bin-partition argument is valid, and the rates n^{1/3} C^{2/3} sigma^{4/3} and n^{2/3} C^{1/3} sigma^{2/3} match the minimax benchmarks. Thus the central estimation and TV-denoising contributions survive the general-W concern. I also note two further issues that do not change the verdict but reinforce CONDITIONAL: Appendix D's upper-bound proof for Theorem 7 drops a square in the Pinsker step (the displayed beta/(2 ln 2) should be beta^2/(2 ln 2)), and the proof of Theorem 9 replaces a union-bound threshold of order sqrt(d log n) with sqrt(d log log n) in line (a), which is not justified by Sauer's lemma. Both are addressable. Since the reader's verdict already is CONDITIONAL and my concern supports that assessment without moving it further, the verdict remains UNCHANGED.","tokens_in":26234,"tokens_out":21150,"duration_ms":183126,"concrete_test":"For n = 8 (and also n = 16 for CDJV), construct the orthonormal Daubechies-2 (DB2) wavelet transform matrix W, and compute the index set I = {i : W_{i,n} != 0} for the last coordinate (theta_1). Count |I| and compare it with log2 n. Repeat for the CDJV boundary-wavelet matrix used in Corollary 3. If |I| > log2 n, the union bound in the proof of Lemma 1 is invalid for general W as stated, and Corollary 3 and the general-W form of Theorem 9 require an enlarged threshold; if |I| = log2 n for every such W, the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing unresolved issue is the generality claimed in Lemma 1 (Appendix F). The proof asserts that 'due to the dyadic MRA structure of the wavelet basis, only log2 n wavelet coefficients are going to affect the value of the estimate theta_hat_1', then union-bounds exactly log2 n noise variables and sets lambda = 2 sigma sqrt(2 log(log n / delta)). Lemma 1 is stated for an arbitrary orthonormal wavelet transform matrix W. For Haar wavelets the coefficient count log2 n is correct, but for Daubechies and CDJV bases with more than one vanishing moment, a constant (filter-length-dependent) number of wavelets per scale overlap the boundary point n, so |I| = c log2 n with c > 1. The union bound would then require lambda = 2 sigma sqrt(2 log(c log2 n / delta)), and the constant 6 in the lemma would need to be adjusted. As stated, the general-W bound is not established. This propagates to Corollary 3 (the CDJV fail-safe bound) and to the general-W form of Theorem 9, and it also undermines the paper's claim that using higher-order wavelets such as DB8 carries a proven pointwise guarantee. It does not, however, invalidate Theorem 2 or Theorem 11: those proofs rely only on the Haar-specific Lemma 14 and the exact Haar coefficient count. The Haar-based central claims appear internally consistent, with only minor typographical slips (e.g., the 'minimum' versus 'maximum' wording in the proof of Theorem 2 and the use of all t <= r* rather than dyadic t in the same proof, which can be repaired by restricting to S(r*)).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies pointwise estimation of the latest ground truth theta_1 from noisy observations y_n, ..., y_1 under temporal distribution shift. The proposed Algorithm 1 applies a wavelet soft-thresholding estimator and, for the Haar system, is claimed to recover the optimal pointwise error bound previously achieved by adaptive window selection (Theorem 2). A general sparsity-based bound is stated as Lemma 1 and used to derive a CDJV fail-safe bound (Corollary 3) and excess-risk guarantees for binary classification with a single ERM call (Theorems 9 and 10). The paper also proves that any estimator satisfying the pointwise bound, run iteratively, is minimax optimal for total-variation denoising (Theorem 11), and it reports synthetic and real-data experiments in Section 6.","tokens_in":26503,"tokens_out":11288,"duration_ms":98648,"significance":"If the theorems were fully established as stated, the paper would make several useful contributions: it shows that a fixed, drift-independent wavelet-shrinkage routine matches the previous best adaptive-window pointwise guarantees; it provides a black-box reduction from pointwise bounds to minimax TV-denoising rates; and it reduces the ERM-oracle complexity of distribution-shift classification to a single call. The Haar-based Theorem 2 and the TV-denoising Theorem 11 are the strongest and most self-contained parts, and the explicit threshold lambda with no fitted constants in the theory is a strength. The main unresolved issue is the generality claimed in Lemma 1, which is load-bearing for Corollary 3 and Theorem 9 but is only proved for the Haar coefficient count, and the missing union bound in the proof of Theorem 11.","major_comments":[{"comment":"The proof is not established for the stated generality. The assertion that 'due to the dyadic MRA structure of the wavelet basis, only log2 n wavelet coefficients are going to affect the value of the estimate theta_hat_1' is correct for the Haar system but not for a general orthonormal wavelet transform matrix W. For Daubechies or CDJV bases with more than one vanishing moment, a filter-length-dependent constant number of wavelets per scale overlap the boundary point, so |I| = c log2 n with c > 1. The union bound and the threshold lambda = 2 sigma sqrt(2 log(log n / delta)) are calibrated to |I| = log2 n; with |I| = c log2 n, the displayed union bound does not give the claimed 1 - delta event. This propagates to Corollary 3 and Theorem 9, and to the paper's claim that higher-order wavelets such as DB8 carry proven pointwise guarantees. The fix is local: either state Lemma 1 only for Haar, or replace log2 n by c log2 n and adjust the threshold (and verify the probability calculation) for the chosen wavelet family.","section":"Appendix F, Lemma 1"},{"comment":"The probability calculation leading to Eq. (5) is not correct as written. Solving 2 log2 n exp(-t^2/(2 sigma^2)) = delta gives t = sigma sqrt(2 log(2 log2 n / delta)), not sigma sqrt(2 log(log n / delta)). The subsequent threshold lambda is twice the displayed right-hand side, so the final high-probability statement may still hold for the parameter ranges used in the paper, but the proof as written does not derive the claimed 1 - delta guarantee. Please redo the union bound with explicit constants and state the exact event used by the three-case argument.","section":"Appendix F, Lemma 1, Eq. (5)"},{"comment":"The proof applies the pointwise bound of Theorem 2 or Corollary 3 separately to each prefix j in [n] and then sums the resulting errors, but it does not take a union bound over the n different high-probability events. As written, the claimed 'with probability at least 1 - delta' for the simultaneous bound over all time points is not justified. This is repairable by running the pointwise guarantee with failure probability delta/n and absorbing the extra log factor into the O-tilde notation; please make this explicit.","section":"Section 5, Theorem 11 proof"}],"minor_comments":[{"comment":"The definition of U_tilde(r) reads max_{t in S(r)} |TV(theta_{t:1}) - theta_1|, but TV(theta_{t:1}) is already a scalar total variation, so the expression |TV(theta_{t:1}) - theta_1| is dimensionally inconsistent; presumably the intended quantity is max_{t in S(r)} TV(theta_{t:1}) or a similar variational term. Please correct the definition.","section":"Corollary 3"},{"comment":"The proof switches between 'maximum value among the ordered set S' and 'when the minimum in the set S is attained', and between max over t <= r* and max over t in S(r*). These should be made consistent with the dyadic set S(r) defined in the statement; the intended argument is clear but the notation should be fixed.","section":"Appendix F, proof of Theorem 2"},{"comment":"The theorem states that the log-loss is bounded by 'monotonic non-decreasing functions on beta' without giving explicit formulas for LB(beta) and UB(beta). Please state the functions explicitly or provide their closed forms in the appendix, since the monotonicity claim alone does not quantify the rate of degradation.","section":"Theorem 7"}],"recommendation":"major_revision","confidential_remarks":"The paper's central Haar-based results and the TV-denoising reduction appear sound and are worth publishing once the general-W statement of Lemma 1 is either repaired or narrowed. If the authors choose to restrict Lemma 1 to Haar, the higher-order wavelet claims (DB8, CDJV) should be reframed as empirical rather than proven, and the statements of Corollary 3 and Theorem 9 should be adjusted accordingly. The missing union bound in Theorem 11 is a simple technical fix. The Section 3 analysis of distribution shift is comparatively weak and somewhat disconnected, but it is not the basis for my recommendation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest read: the Haar-wavelet core is real, and the TV-denoising reduction is the most interesting thing in the paper. The general-wavelet story overreaches.\n\nWhat's new and good. Theorem 2 shows a fixed wavelet soft-thresholding routine, with no drift knowledge, matches the optimal pointwise error bound from Mazzetto and Upfal. That is a genuinely simpler route to a known rate. Theorem 11 — any estimator with that pointwise guarantee is automatically minimax optimal for TV denoising — is a nice black-box observation, and it likely makes several existing algorithms newly certifiable. The single-ERM-call classification bound is a real computational improvement over the O(log n)-call procedure, and the surrogate-loss differentiability point is sensible.\n\nWhere it gets soft. Lemma 1 is stated for an arbitrary orthonormal wavelet transform, but the proof relies on \"only log2 n wavelet coefficients affect θ1\". That is exactly true for Haar. For Daubechies or CDJV bases with more than one vanishing moment, a constant multiple of wavelets per scale overlap the boundary point, so your union bound is over c log n noise variables, not log n. The threshold and constants in Lemma 1, and everything built on it (Corollary 3, the general-W form of Theorem 9) are not established as stated. The authors actually admit they have no construction where the general bound beats Haar, so the advertised \"faster rates via higher-order wavelets\" is an empirical hope, not a proven statement.\n\nSection 3 also has an algebra slip. In the proof of Theorem 7, after applying Pinsker to the marginal KL, the β should be β^2. With the corrected factor the upper bound β KL(·) − β^2·(1/2ln2)(...)² is not obviously monotone in β on [0,1], which weakens Corollary 8's monotone-degradation claim.\n\nThe TV-denoising novelty claim ignores the authors' own 2021 reduction of TV denoising to adaptive online learning. The specific statement may be new, but \"not uncovered before\" is too strong without citation and discussion.\n\nThese are fixable. The Haar-based Theorem 2 and Theorem 11 appear internally consistent; I don't see a fatal flaw in the central argument. But the paper as written overstates what is proven for general wavelets, and the Section 3 analysis needs correction.\n\nVerdict: conditionally worthwhile. A serious referee should engage, but a major revision is needed. If Lemma 1 is restricted to Haar or fixed for each specific basis, and Section 3 is cleaned up, this could be a solid paper. As is, it deserves peer review but not acceptance on the current text.","headline":"Haar-wavelet core is real, but the general-wavelet claims are not proven; referee it for the TV-denoising reduction.","tokens_in":27097,"tokens_out":4098,"would_cite":true,"duration_ms":33939,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","42C40","62C20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A fixed wavelet soft-thresholding estimator, with no knowledge of the drift, matches the optimal pointwise error bound for estimating the latest value of a noisy non-stationary sequence, and iterating it yields minimax-optimal…","keywords":["temporal distribution shift","wavelet soft-thresholding","pointwise estimation error","adaptive window selection","total-variation denoising","binary classification under drift","Haar wavelet","excess risk"],"falsifier":"Compute the last column of a compactly supported orthonormal wavelet matrix with more than one vanishing moment (not Haar) for $n=16$ and $n=64$; the number of entries with magnitude above numerical precision grows as a constant multiple of $\\log_2 n$, exceeding the proof's asserted count. Then simulate the Lemma 1 guarantee: draw $n$ i.i.d. $\\sigma$-sub-Gaussian noise samples on a signal with zero true wavelet coefficients, set $\\lambda=2\\sigma\\sqrt{2\\log(\\log n/\\delta)}$ with $\\delta=0.05$, and measure the empirical coverage of $|\\hat\\theta_1-\\theta_1|\\le\\sum_{i\\in I}6|W_{i,n}|(|\\beta_i|\\wedge\\lambda)$ over many replicates; coverage clearly below $1-\\delta$ would refute the general-basis claim.","tokens_in":25987,"feed_emoji":"📉","tokens_out":14586,"duration_ms":115113,"temperature":0.7,"pith_summary":"When only noisy observations of a time-varying sequence are available, estimating the most recent value usually requires choosing an averaging window that balances bias against variance, and the right choice depends on unknown drift. This paper claims that a single fixed routine—wavelet soft-thresholding with Haar wavelets, with the threshold set by the noise level and the failure probability—achieves the same optimal pointwise error as the earlier adaptive-window algorithm, without ever choosing a window. The error is controlled by $U(r^*)$, the best among dyadic windows of the maximum local drift versus the standard error $\\sigma/\\sqrt{r}$. The same estimator yields excess-risk guarantees for binary classification under temporal shift with a single call to an empirical risk minimization oracle, and, run iteratively, achieves minimax-optimal squared and absolute risks for total-variation denoising. If correct, this turns a drift-adaptive estimator into a sparsity-adaptive denoiser with one fixed thresholding rule.","feed_headline":"Wavelet soft-thresholding, no drift tuning, hits optimal error","feed_subtitle":"One fixed estimator matches adaptive-window guarantees and is minimax for TV denoising.","key_machinery":"The mechanism is the orthonormal wavelet transform $W$ paired with coordinate-wise soft-thresholding $T_\\lambda$. Algorithm 1 maps the reverse-time observation vector to empirical wavelet coefficients $\\tilde\\beta=Wy$, shrinks each coefficient toward zero by $\\lambda$, reconstructs $\\hat\\theta=W^T T_\\lambda(\\tilde\\beta)$, and returns the last coordinate. Since the estimate of $\\theta_1$ is a linear combination of wavelet coefficients with weights $W_{i,n}$, only coefficients whose wavelets overlap the boundary point matter; for Haar wavelets this is a logarithmic number, so a union bound over these noise terms calibrates $\\lambda=2\\sigma\\sqrt{2\\log(\\log n/\\delta)}$. The true coefficients $\\beta_i$ are large only where the sequence has structure—constant blocks for Haar, piecewise-polynomial trends for higher-order systems—so the bound $\\sum_{i\\in I}6|W_{i,n}|(|\\beta_i|\\wedge\\lambda)$ is a sparsity-adaptive bias-variance tradeoff with no window selection. The TV-denoising result uses the fact that the pointwise bound implies, inside any interval of bounded variation, an error of order $\\sigma/\\sqrt{k}$, and summing these errors over a partition chosen by variation budget yields the aggregate risk rates.","core_discovery":"The central claim is that pointwise estimation under temporal distribution shift is governed by the sparsity of the ground-truth sequence in a wavelet basis, and that soft-thresholding empirical wavelet coefficients realizes the optimal bias-variance tradeoff automatically. For the Haar system, Theorem 2 gives, with probability at least $1-\\delta$, $|\\hat\\theta_1-\\theta_1| \\le \\kappa U(r^*)$, where $U(r^*)$ minimizes over $r$ the larger of the maximum deviation of dyadic block averages from $\\theta_1$ and $\\sigma/\\sqrt{r}$; this is the same variational bound previously obtained by algorithms that explicitly pick a window, and it holds without knowing the drift level. Lemma 1 gives the coordinate-wise bound $\\sum_{i\\in I} 6|W_{i,n}|(|\\beta_i|\\wedge\\lambda)$, showing that only wavelet coefficients whose basis elements overlap the latest time point matter, with the threshold $\\lambda$ set by the maximum of a logarithmic number of noise coefficients. These estimates feed a binary classification result with excess-risk control under temporal shift that needs one ERM oracle call and preserves differentiability for surrogate losses, and Theorem 11, which states that any estimator with the Theorem 2 guarantee, applied at every time point, is minimax optimal for total-variation denoising up to logarithmic factors.","pith_inferences":["Editorial inference: The sparsity-based view suggests the same soft-thresholding routine could serve as a drop-in estimator for other boundary-anchored prediction tasks, such as online forecasting, by choosing a wavelet basis matched to the expected trend order; the paper does not develop this.","Editorial inference: Replacing the exact count $\\log_2 n$ in Lemma 1 with a basis-dependent constant would likely leave the Haar-based Theorem 2 and the TV-denoising implications intact, since those results do not depend on the general-basis claim.","Editorial inference: A practical testable extension is to select the wavelet basis and threshold by a data-dependent criterion such as cross-validation or risk estimation, rather than the practitioner heuristic the paper describes.","Editorial inference: The black-box TV-denoising result suggests a converse direction: minimax-optimal denoising algorithms specialized to the final coordinate may admit pointwise bounds of the Theorem 2 form, which would enlarge the known class of drift-adaptive estimators."],"forward_implications":["A binary classifier can be trained under temporal shift with a single call to an empirical risk minimization oracle, instead of a logarithmic number of calls, while keeping surrogate-loss objectives differentiable almost everywhere.","Any algorithm satisfying the pointwise bound of Theorem 2 or Corollary 3 is automatically minimax optimal for total-variation denoising, with squared risk $\\tilde{O}(n^{1/3}C^{2/3}\\sigma^{4/3})$ and absolute risk $\\tilde{O}(n^{2/3}C^{1/3}\\sigma^{2/3})$.","Estimating the latest value requires no prior knowledge of the drift level: the threshold depends only on the noise level, the failure probability, and the horizon $n$.","Higher-order wavelet bases can give sharper estimation errors than Haar when the ground truth has smooth trends, and CDJV wavelets provide a fail-safe guarantee comparable to the Haar bound.","Under the paper's idealized training assumption, adding more dissimilar samples to the training distribution degrades the log-loss monotonically, with matching upper and lower bounds."],"supporting_citations":[{"why":"Establishes the optimal adaptive-window pointwise bound that this work recovers via Haar-wavelet soft-thresholding, and supplies the O(log n)-ERM-call baseline.","marker":"Mazzetto and Upfal (2023)"},{"why":"Introduces the soft-thresholding denoising procedure used as Algorithm 1.","marker":"Donoho (1995)"},{"why":"Presents an adaptive-window model-selection algorithm satisfying the same pointwise bound form and serves as a comparison baseline.","marker":"Han et al. (2024)"},{"why":"Constructs the CDJV boundary wavelets used in Corollary 3 to give a fail-safe error bound.","marker":"Cohen et al. (1993)"},{"why":"Gives the minimax rates for total-variation denoising that Theorem 11 matches up to logarithmic factors.","marker":"Donoho et al. (1996)"},{"why":"Provides the wavelet-shrinkage machinery and the Besov-embedding fact that relate total variation to wavelet coefficients.","marker":"Donoho et al. (1998)"},{"why":"Supplies the wavelet and Besov-norm background used in the CDJV argument and threshold calibration.","marker":"Johnstone (2017)"}],"fun_headline_variants":["Wavelet soft-thresholding nails drift without tuning","One wavelet estimator matches adaptive windows","Soft-threshold wavelets: optimal without drift level","Wavelet sparsity unlocks minimax drift estimation","Minimax wavelets without knowing drift"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, for an arbitrary orthonormal wavelet transform, only about $\\log_2 n$ wavelet coefficients affect the estimate of the last value $\\hat\\theta_1$, because this count sets the threshold $\\lambda$ and the probability calibration; the premise holds for Haar wavelets but not for other compactly supported orthonormal systems, so the general-transform guarantees in Lemma 1 and their downstream uses rest on an unproved assertion.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet soft-thresholding nails drift without tuning","One wavelet estimator matches adaptive windows","Soft-threshold wavelets: optimal without drift level","Wavelet sparsity unlocks minimax drift estimation","Minimax wavelets without knowing drift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000749,"raw_usage":{"total_tokens":3369,"prompt_tokens":1012,"completion_tokens":2357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":2288}},"tokens_in":628,"tokens_out":2357,"duration_ms":15625,"temperature":1.0,"reasoning_tokens":2288,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:12:40.374221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the last column of a compactly supported orthonormal wavelet matrix with more than one vanishing moment (not Haar) for $n=16$ and $n=64$; the number of entries with magnitude above numerical precision grows as a constant multiple of $\\log_2 n$, exceeding the proof's asserted count. Then simulate the Lemma 1 guarantee: draw $n$ i.i.d. $\\sigma$-sub-Gaussian noise samples on a signal with zero true wavelet coefficients, set $\\lambda=2\\sigma\\sqrt{2\\log(\\log n/\\delta)}$ with $\\delta=0.05$, and measure the empirical coverage of $|\\hat\\theta_1-\\theta_1|\\le\\sum_{i\\in I}6|W_{i,n}|(|\\beta_i|\\wedge\\lambda)$ over many replicates; coverage clearly below $1-\\delta$ would refute the general-basis claim.","supporting_citations":[],"review_version":1}