{"id":"c244edad-2c86-4f2a-92f8-d110626c230e","arxiv_id":"2507.13480","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Samplet coefficient decay rates are fit to estimate local Hölder exponents for non-uniformly sampled multivariate signals.","lead":"A new algorithm uses samplet transforms to estimate local smoothness in scattered data, such as point clouds and irregularly sampled images, in near-linear time. The method works by measuring how quickly multiresolution coefficients shrink across scales.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2 assumes samplet coefficients decay at the exact rate α+d/2, but Theorem 3.1 supplies only an upper bound; a function in C^α(x0) can have faster-decaying coefficients, so the fitted exponent can overestimate α.","rationale":"The reader's weakest assumption is exactly the load-bearing point: Theorem 3.1 is one-sided, while Algorithm 2 treats the decay rate as exact. I agree with that reading and have supplied a concrete function that satisfies the theorem's hypotheses but produces faster-decaying samplet coefficients, so the fitted exponent disagrees with the true pointwise Hölder exponent. The counterexample is pathological, but it is within the stated domain of C^α(x0) and shows that the upper bound alone cannot justify the slope inference. The secondary diam(τ) ≤ R mismatch between the theorem's proof and the algorithm's use of all levels further weakens the connection, but the missing lower bound is the more fundamental gap. Since the theorem itself appears correct and the numerical demonstrations are internally consistent, the paper remains a useful heuristic contribution; however, the central algorithmic claim needs either a lower-bound theorem under additional assumptions or an explicit non-degeneracy condition, plus an empirical comparison with existing edge detectors. This does not change the reader's CONDITIONAL verdict.","tokens_in":15710,"tokens_out":19537,"duration_ms":241832,"concrete_test":"Run Algorithm 2 in 1D on f(x) = x·1_Q(x) + x^5 sampled on N = 2^22 uniform grid points in [0,1], using q+1 = 5 vanishing moments as in §5.1. Every grid point is rational, so the data vector is x_i + x_i^5; the terms of degree ≤4 are annihilated, so the leading non-vanishing term is x^5 and the fitted slope should be ≈5.5, implying α≈5. The true local Hölder exponent at x0=0 is 1 (f ∈ C^1(0), f ∉ C^{1+ε}(0)). If the algorithm returns α > 1, then the equality in (15) is not implied by Theorem 3.1; a matching lower bound or a non-degeneracy assumption is needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central inference of §4 is that the least-squares slope in (15) equals α+d/2. Theorem 3.1, however, gives only the upper bound |(σ_{j,k}, f)| ≲ diam(τ)^α √#τ; there is no matching lower bound or converse. Consequently the model e_j ≈ c(b_j)^{α+d/2} is not justified: a function in C^α(x0) may have samplet coefficients that decay faster than the bound, and the fitted slope then overestimates α. This is not a remote technicality. Let q be the vanishing-moment degree and consider f(x) = x·1_Q(x) + x^{q+1} on R with x0=0. As in Remark 2.2, f ∈ C^1(0) (take P(x)=x), while f ∉ C^{1+ε}(0), so its local Hölder exponent is 1. On any uniform floating-point grid all sample points are rational, so the input data equal x + x^{q+1}. With q vanishing moments, the polynomial part of degree ≤ q is annihilated and the coefficients decay like 2^{-j(q+1+1/2)} (Remark 3.2), yielding a fitted exponent α ≈ q+1 instead of 1. Thus the upper-bound theorem alone cannot support the equality used by Algorithm 2. A secondary issue: the proof of Theorem 3.1 explicitly assumes diam(τ) ≤ R, but the statement and Algorithm 2 use all levels, so coarse clusters containing unrelated singularities can also corrupt the fit.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a samplet-based method for detecting local Hölder regularity in non-uniformly sampled multivariate signals. It proves an upper decay bound for samplet coefficients of functions in Jaffard's microlocal spaces C^α(x0) (Theorem 3.1), derives corollaries for Hölder and Sobolev-Slobodeckij spaces, and proposes a near-linear-time algorithm (Algorithms 1 and 2) that fits the decay rate of samplet coefficient norms along tree branches to estimate the local Hölder exponent. The method is tested extensively on 1D, 2D, and 3D signals, including scattered point clouds and images.","tokens_in":16007,"tokens_out":2558,"duration_ms":29743,"significance":"If the algorithmic inference is justified, this is a valuable extension of wavelet-based edge detection to scattered and multivariate data, with an appealing near-linear practical cost. The upper bound in Theorem 3.1 is clean and correctly derived from the vanishing-moment property, and the numerical experiments cover a wide range of settings, including non-uniform point clouds. The central gap is that the algorithm relies on an equality between the fitted slope and α+d/2, whereas the paper only proves an upper bound; this undermines the theoretical support for the estimated exponents until a matching lower bound or characterization is supplied.","major_comments":[{"comment":"The model e_j ≈ c (b_j)^{α+d/2} assumes that samplet coefficients decay at the exact rate α+d/2, but Theorem 3.1 provides only the upper bound |(σ_{j,k}, f)| ≲ diam(τ)^α √#τ. There is no matching lower bound or converse theorem showing that the rate is attained. A concrete failure mode is a function such as f(x) = x·1_Q(x) + x^{q+1} at x0=0 (a variant of Remark 2.2): on any floating-point grid the data equal x + x^{q+1}, and with q vanishing moments the polynomial part is annihilated, so the coefficients decay like 2^{-j(q+1+1/2)} and the fitted exponent is near q+1 rather than the true local Hölder exponent 1. The paper must either prove a lower bound or characterize the decay rate, or clearly reframe the slope-fitting step as a heuristic whose accuracy is only empirically validated.","section":"§4, Eq. (15)"},{"comment":"The proof begins by assuming 'diam(τ) ≤ R', where R is the radius in the definition of C^α(x0), but this hypothesis is absent from the theorem statement. Algorithm 2 uses coefficients from all levels of the tree, including coarse clusters for which diam(τ) may exceed R and which may contain multiple singularities. This gap means the theoretical justification for applying the decay bound to the full branch is incomplete; the statement should include the condition diam(τ) ≤ R, or the algorithm should restrict the fit to clusters satisfying it.","section":"Theorem 3.1, proof"}],"minor_comments":[{"comment":"The statement contains 'x0 ∈ Ω ⊆ Ω', which appears to be a typo; it should likely read 'x0 ∈ Ω′ ⊆ Ω'.","section":"Corollary 3.5"},{"comment":"In the last sentence, 'the las claim' should be 'the last claim'.","section":"Lemma 3.4, proof"},{"comment":"The lines 'er ← [ ]' and 'br ← [ ]' appear after the function definition and read as global initializations; please clarify the intended scope of these variables (e.g., by moving them into the main routine or annotating them as global).","section":"Algorithm 1"},{"comment":"For the 3D experiments, the entry '28 × 28 × 28' seems small compared with the other settings; please clarify whether these are grid dimensions, cluster-tree depths, or another quantity.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of math.NA and the experimental work is substantial, but the central algorithmic claim should be supported by a matching lower bound or explicitly reclassified as a heuristic. The current gap between Theorem 3.1 and the slope-fitting step is load-bearing and should be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful part of this paper is Theorem 3.1: for f in C^α(x0), samplet coefficients on clusters containing x0 decay like diam(τ)^α√#τ, assuming vanishing moments of degree at least floor(α) and diam(τ)≤R (the latter appears in the proof but not in the statement). That is a genuine extension of Jaffard's wavelet result to samplets, and the proof is short and clean. The corollaries for Hölder and Sobolev-Slobodeckij spaces follow naturally. The heavy self-citation to samplet papers is appropriate because samplets are the object under study, not a way to pad references.\n\nThe soft spot is the step from Theorem 3.1 to Algorithm 2. The theorem is an upper bound. The algorithm fits log e_j against log b_j and reads α+d/2 from the slope. That equality requires a matching lower bound or at least a converse, which is not provided. The stress-test note is right. Take f(x)=x·1_Q(x)+x^{q+1} at x0=0 with q vanishing moments. On any floating-point grid all samples are rational, so the data equal x+x^{q+1}; the linear part is annihilated, and the fitted Hölder exponent comes out q+1 instead of the true local exponent 1. This is not a remote pathology; it shows the method reports the smoothness it can see after annihilation, not the actual pointwise regularity. Also, the failure to track diam(τ)≤R in the algorithm means coarse clusters can violate the hypothesis and corrupt the fit.\n\nWhat the paper does well: the numerics are extensive, runtimes are reported, and the method is fast and training-free. But there are no comparisons against existing edge detectors or kernel methods, so the claimed practical advantage is not demonstrated, and there are no error bars.\n\nWho is this for? Researchers working on scattered-data multiresolution and local regularity estimation. It deserves a serious referee because the decay theorem is correct and the algorithm is plausible, but the referee should push for either a lower bound under extra assumptions or a revised algorithm that honestly reports an upper-bound estimate, plus a test on the pathology above. I would accept it for peer review, expecting substantial revision; the core theorem is worth publishing, but the current framing overclaims.","headline":"A clean samplet decay theorem wrapped around a slope-fitting algorithm whose central equality the theorem does not support; the numerics look good, but the theory overclaims.","tokens_in":16554,"tokens_out":2966,"would_cite":true,"duration_ms":33255,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["42C40","65T60"],"pacs":[],"model":"deepseek-v4-flash","headline":"Samplet coefficient decay rates determine local Hölder exponents on scattered, non-uniformly sampled signals.","keywords":["samplets","local Hölder exponents","microlocal spaces","scattered data","edge detection","multiresolution analysis","non-uniform sampling","pointwise regularity"],"falsifier":"Take a known test signal $f(x) = |x-x_0|^\\alpha$ sampled on scattered sites, run the branch slope fit for many values of $\\alpha$ and $N$, and compare each recovered exponent to the true $\\alpha$; if the fitted slope systematically deviates as $\\alpha$ grows or as coarse clusters dominate the fit, the identification fails. A second check is to compute the coefficients for a $C^\\infty$ bump, for which the fit will assign a large exponent, testing only the upper bound rather than the equality model.","tokens_in":15479,"feed_emoji":"📉","tokens_out":6902,"duration_ms":75783,"temperature":0.7,"pith_summary":"This paper claims that the local Hölder exponent of a signal sampled at scattered, non-uniformly spaced points can be read off from how fast its samplet coefficients decay along a multiresolution tree. Samplets are discrete signed measures, localized at data sites and built to annihilate low-degree polynomials; the paper proves that for a function in the microlocal space $C^\\alpha(x_0)$, every samplet coefficient on a cluster containing $x_0$ satisfies $|(\\sigma_{j,k}, f)| \\lesssim \\operatorname{diam}(\\tau)^\\alpha \\sqrt{\\#\\tau}$. The proof turns this decay into a near-linear-time algorithm: walk down each branch of the $2^d$-tree, fit a line to log coefficient norm versus log cluster diameter, and read the slope as $\\alpha + d/2$. If correct, this gives a deterministic, parameter-lean way to segment images, point clouds, and volumetric data by local smoothness, without structured grids or trained networks. The numerical experiments in one, two, and three dimensions support the claim across jumps, corners, and smooth regions.","feed_headline":"Decay slopes read local Hölder exponents from scattered points","feed_subtitle":"Near-linear-time slope fit on samplet coefficients separates smooth regions from singularities in 1D, 2D, and 3D.","key_machinery":"The load-bearing object is the samplet: a localized signed measure $\\sigma_{j,k} = \\sum_\\ell \\omega_\\ell \\delta_{x_\\ell}$ with $\\ell^2$-normalized weights, built from a hierarchical $2^d$-tree so that it has vanishing moments against polynomials of total degree $q$ and is orthonormal within each detail space. The fast samplet transform changes the data between Dirac-$\\delta$ coordinates and samplet coordinates in linear time once the tree exists. The paper's main identity is the coefficient estimate $|(\\sigma_{j,k}, f)| \\lesssim \\operatorname{diam}(\\tau)^\\alpha \\sqrt{\\#\\tau}$, obtained by subtracting the approximating polynomial from $f$ and using Cauchy-Schwarz on the $\\ell^1$ norm of the weights; the same estimate, specialized to balanced trees, gives the decay rate that the slope-fitting algorithm reads.","core_discovery":"The central discovery is a decay-to-regularity correspondence for samplets that mirrors the classical wavelet result. For any cluster $\\tau$ containing $x_0$, if $f \\in C^\\alpha(x_0)$ and the samplets have vanishing moments of degree at least $\\lfloor\\alpha\\rfloor$, then $|(\\sigma_{j,k}, f)| \\lesssim \\operatorname{diam}(\\tau)^\\alpha \\sqrt{\\#\\tau}$; on a balanced $2^d$-tree this becomes $\\lesssim \\sqrt{N}\\, d^\\alpha\\, 2^{-j(\\alpha+d/2)}$. The paper then proposes to treat this bound as an equality model $e_j \\approx c\\, (b_j)^{\\alpha+d/2}$ on each branch and recover $\\alpha+d/2$ by least-squares regression of $\\log e_j$ against $\\log b_j$. This turns regularity detection into a slope estimate and assigns a microlocal class $C^\\alpha$ to every leaf cluster. Corollaries extend the decay estimate to H\\\"older and Sobolev-Slobodeckij classes, and experiments on grid and scattered data, images, the sphere, and point-cloud surfaces demonstrate the resulting edge and singularity charts.","pith_inferences":["Because Theorem 3.1 is only an upper bound, the slope fit can overestimate $\\alpha$ for functions that are smoother at $x_0$ than their formal microlocal exponent; a matching lower-bound or two-sided estimate would make the recovered exponent exact.","The same slope machinery should extend to any metric-measure setting where a hierarchical partition with diameter decay is available, even beyond the sphere and surface examples treated in the paper.","Thresholding the recovered slopes yields a multiscale edge detector whose scale sensitivity is controlled by the vanishing-moment order $q$; choosing $q$ adaptively could separate jumps from higher-order singularities without changing the algorithm.","The residual scatter in the log-log regression could provide empirical confidence intervals for the inferred H\\\"older exponent, although the paper does not discuss uncertainty quantification."],"forward_implications":["For a signal in $C^\\alpha(x_0)$, every samplet coefficient on a cluster containing $x_0$ decays at least like $\\operatorname{diam}(\\tau)^\\alpha \\sqrt{\\#\\tau}$, provided the vanishing-moment order $q$ satisfies $q \\geq \\lfloor\\alpha\\rfloor$.","On balanced $2^d$-trees this means coefficient norms along a branch decay as $2^{-j(\\alpha+d/2)}$, so $\\alpha+d/2$ is exactly the slope in a log-log plot.","The same bound transfers to H\\\"older functions $C^{m,\\vartheta}$ and to Sobolev-Slobodeckij spaces $W^{s,p}$ with $s > d/p$, yielding decay rates $m+\\vartheta$ and $s-d/p-\\varepsilon$, respectively.","A depth-first traversal of the tree plus one least-squares fit per branch yields a local regularity chart in essentially linear time once the cluster tree is available.","The experiments assign $C^0$ to jumps, $C^1$ to corners and nondifferentiable points, and high exponents to smooth regions in one, two, and three dimensions on both gridded and scattered data."],"supporting_citations":[{"why":"Constructs samplets with vanishing moments and the fast samplet transform, supplying the basis and transform the algorithm runs on.","marker":"[11]"},{"why":"Defines the microlocal spaces $C^\\alpha(x_0)$ and the wavelet-coefficient decay characterization that Theorem 3.1 adapts to scattered data.","marker":"[15]"},{"why":"Provides the pointwise-regularity framework and wavelet decay results that motivate the slope-fitting approach.","marker":"[16]"},{"why":"Introduces the idea of reading local H\\\"older exponents from the decay of wavelet coefficients across multiscale maxima, the starting point the paper extends.","marker":"[23]"},{"why":"Gives the Sobolev embedding theorem used to transfer samplet decay estimates to Sobolev-Slobodeckij spaces.","marker":"[6]"},{"why":"Supplies the depth-first search used by Algorithm 1 to traverse the samplet tree in linear time.","marker":"[29]"}],"fun_headline_variants":["Samplet decay slopes reveal local Hölder exponents","Slope of samplet decay maps pointwise smoothness","Fast detector for local regularity from scattered data","Decay-rate fit yields pointwise Hölder charts","Samplet slopes read pointwise regularity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The algorithm assumes the decay rate in Theorem 3.1 is sharp enough that the slope of log coefficient norm versus log diameter equals $\\alpha+d/2$; the theorem only proves an upper bound, and for coarse clusters the required condition $\\operatorname{diam}(\\tau) \\leq R$ may fail.","fun_headline_variants_meta":{"raw":{"variants":["Samplet decay slopes reveal local Hölder exponents","Slope of samplet decay maps pointwise smoothness","Fast detector for local regularity from scattered data","Decay-rate fit yields pointwise Hölder charts","Samplet slopes read pointwise regularity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000781,"raw_usage":{"total_tokens":3462,"prompt_tokens":972,"completion_tokens":2490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2419}},"tokens_in":588,"tokens_out":2490,"duration_ms":20207,"temperature":1.0,"reasoning_tokens":2419,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:23:26.193598+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a known test signal $f(x) = |x-x_0|^\\alpha$ sampled on scattered sites, run the branch slope fit for many values of $\\alpha$ and $N$, and compare each recovered exponent to the true $\\alpha$; if the fitted slope systematically deviates as $\\alpha$ grows or as coarse clusters dominate the fit, the identification fails. A second check is to compute the coefficients for a $C^\\infty$ bump, for which the fit will assign a large exponent, testing only the upper bound rather than the equality model.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the microlocal spaces $C^\\alpha(x_0)$ and the wavelet-coefficient decay characterization that Theorem 3.1 adapts to scattered data."},{"cited_title":"Wavelet Methods for Pointwise Regularity and Local Oscillations of Functions, volume 587","cited_arxiv_id":null,"evidence_quote":"Provides the pointwise-regularity framework and wavelet decay results that motivate the slope-fitting approach."},{"cited_title":"Mallat and S","cited_arxiv_id":null,"evidence_quote":"Introduces the idea of reading local H\\\"older exponents from the decay of wavelet coefficients across multiscale maxima, the starting point the paper extends."},{"cited_title":"Demengel","cited_arxiv_id":null,"evidence_quote":"Gives the Sobolev embedding theorem used to transfer samplet decay estimates to Sobolev-Slobodeckij spaces."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the depth-first search used by Algorithm 1 to traverse the samplet tree in linear time."}],"review_version":1}