{"id":"c4451625-f3b6-429b-b2a6-b725ccdda783","arxiv_id":"2412.00954","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A generalized samplet construction for arbitrary Banach-space functionals, with vanishing moments and a localization decay estimate, unifying RKHS, Tausch-White, and operator-adapted wavelet settings.","lead":"This paper builds a general mathematical framework for constructing localized multiscale bases, called samplets, that work with measurements from arbitrary linear functionals on a Banach space, not just point samples. The framework uses frames and spectral clustering to create bases with vanishing moments, useful for compressing scattered data and approximating functions in numerical analysis.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 11's proof needs unstated C∞-module/restriction structure and equates a cutoff infimum with the stated seminorm; the localization guarantee is not established as written.","rationale":"The reader's weakest-assumption analysis correctly identifies the missing C∞-module/restriction structure behind Theorem 11. My pass confirms that concern and adds a second, independent defect in the same proof: the infimum over cutoffs is not the same as the infimum over all admissible extensions, so even in Sobolev spaces the displayed chain does not imply the stated bound. Both defects concern the central claim, because Theorem 11 is the main quantitative guarantee of the generalized samplet construction. The rest of the framework—the algebraic construction via isometries, the QR-based vanishing moments, and the three examples—appears coherent; in the examples the missing structure (cutoff multiplication and restriction) is present, which is why the construction is plausible. The paper should be accepted conditionally on supplying the missing hypotheses and repairing the final step of Theorem 11. This does not change the reader's CONDITIONAL verdict.","tokens_in":15998,"tokens_out":17240,"duration_ms":185417,"concrete_test":"Independently re-derive the last inequality of Theorem 11 in B=H^1(R) with K=[−2,2] and P={0}. Take v∈H^1(R) with v=0 on K but non-zero outside K, e.g., a smooth bump supported in [3,4]. Compute ‖v‖_{H^1,K}=0 while inf_{χ=1 near K} ‖χv‖_{H^1} is generally positive; this shows the cutoff infimum is not equal to the stated seminorm. Then check whether the paper supplies any argument replacing χ(v−p) by an optimal local extension; if not, the proof of Theorem 11 fails as written and an explicit extension/trace hypothesis must be added before the localization bound can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative heart of the paper is Theorem 11, and its proof has two unjustified steps. First, the proof writes (ψ_i,v)=(ψ_i,χ(v−p)) and invokes a seminorm ‖v−p‖_{B,supp ψ_i} whose definition uses restrictions to compact sets. For this to be meaningful, B must support multiplication by C0∞ cutoffs and have a local restriction/trace structure. The stated hypothesis B′⊂E′(Ω) only says that the elements of B′ are compactly supported distributions; together with the existence of an ℓ2-frame it does not imply that arbitrary elements of B—a reflexive, topologically Hilbert Banach space—are closed under smooth pointwise multiplication or admit restrictions to compact sets. The remark that E(Ω) dense in B is not enough for this either. Second, even when such module/restriction structure is available, the proof actually establishes |(ψ_i,v)|≤√B_N inf_{p∈P} inf_{χ=1 near K} ‖χ(v−p)‖_B. Since every χ(v−p) is one admissible extension in the definition of ‖v−p‖_{B,K}, this double infimum is generally larger than ‖v−p‖_{B,K}; the final line of the proof asserts the reverse inequality without an argument. Closing the gap needs an explicit extension/trace theorem or a modified definition of the local seminorm. As written, the localization estimate—the main quantitative reason the samplet construction is useful—is not proven for the full class of Banach spaces claimed, and the proof also has a missing step in the standard Sobolev setting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a generalization of samplets—localized discrete signed measures with vanishing moments—to functionals in Banach spaces. The authors assume the functionals form an ℓ2-frame or a Riesz basis, construct a multiresolution hierarchy via spectral clustering of functional supports, and compute an orthogonal two-scale transform by QR-decomposing a moment matrix. The main quantitative result, Theorem 11, states that samplet coefficients satisfy |(ψ_i,v)| ≤ √B_N inf_{p∈P} ‖v−p‖_{B,supp(ψ_i)}. Three examples are presented: point-evaluation samplets in reproducing kernel Hilbert spaces, Tausch–White-type wavelets in H^1_0, and operator-adapted wavelets.","tokens_in":16292,"tokens_out":14640,"duration_ms":134125,"significance":"If the localization theorem were established in the claimed generality, the paper would unify several existing wavelet constructions (classical samplets, Tausch–White multiwavelets, operator-adapted wavelets) within a single Banach-space framework. The construction itself is elegant and internally consistent: the QR-based moment annihilation deterministically produces vanishing moments, the two-scale transform is isometric by construction, and the use of ℓ2-frames gives a clean abstract setting. However, Theorem 11 is the quantitative heart of the paper, and its proof has a load-bearing gap and rests on an ill-posed hypothesis. The paper does not, as written, prove the localization guarantee for the full class of Banach spaces claimed.","major_comments":[{"comment":"The proof of Theorem 11 relies on operations—restriction of elements of B to compact sets and pointwise multiplication by C∞_0 cutoffs—that are not defined or bounded under the stated hypotheses. The seminorm ‖·‖_{B,K} is defined via ṽ|_K = v, which presupposes a trace/restriction theory for B, and the proof replaces v by χ(v−p), which requires B to be a module over C∞_0. The hypothesis B' ⊂ E'(Ω) only constrains the dual space; it does not imply that elements of B are functions or that smooth cutoffs act continuously on B. Consequently, the statement of Theorem 11 is not well-posed and its proof is not valid in the claimed generality.","section":"§4.3 (Theorem 11)"},{"comment":"Even when the necessary module and restriction structure is available, the proof establishes only |(ψ_i,v)| ≤ √B_N inf_{p∈P} inf_{χ=1 near K} ‖χ(v−p)‖_B. For each fixed p, the set {χ(v−p)} is a subset of the feasible set in the definition of ‖v−p‖_{B,K}, so inf_χ ‖χ(v−p)‖_B is generally larger than ‖v−p‖_{B,K}. The final line of the proof asserts the reverse inequality without argument; closing this gap requires an explicit extension or trace theorem, or a redefinition of the seminorm as an infimum over admissible cutoffs. This is load-bearing because Corollary 13 and the claimed coefficient localization depend on Theorem 11.","section":"§4.3, proof of Theorem 11"},{"comment":"The sentence 'This assumption is, for example, satisfied if E(Ω) is a dense subspace of B' is incorrect: density of C∞(Ω) in B yields only an embedding of B' into the space of all distributions D'(Ω), not into the space of compactly supported distributions E'(Ω). In fact, the global hypothesis B' ⊂ E'(Ω) is not satisfied by standard Sobolev spaces on bounded domains; for instance, the functional v ↦ ∫_Ω v dx is not compactly supported. This matters because the examples in §4.4 are of this type. The paper should either weaken the standing assumption to compact support of the finite functional set f_N or revise the examples and the abstract theorem accordingly.","section":"§4.1 (standing assumption)"}],"minor_comments":[{"comment":"The upper frame bound is denoted by B, which conflicts with the Banach space B used throughout the paper; consider renaming one of them (e.g., use C for the frame bound) to avoid confusion.","section":"§2.1, Definition 1"},{"comment":"The step |(u_i, T⋆_N χ(v−p))| ≤ ‖T⋆_N χ(v−p)‖ implicitly uses ‖u_i‖_{ℓ2} = 1, which holds because U is orthogonal; this should be stated explicitly for completeness.","section":"§4.3, proof of Theorem 11"},{"comment":"The suggestion d(f_i,f_j) = dist(supp(f_i),supp(f_j)) is only a pseudo-metric when supports can overlap; this should be noted before invoking spectral clustering results that typically assume a metric or at least symmetric pairwise similarities.","section":"§4.1, similarity measure"},{"comment":"The statement that the resulting basis ~Ψ = U^⊺[κ̃_{x1},...,κ̃_{xN}]^⊺ is 'exactly' the dual embedded samplet basis of [7] would benefit from a precise specification of the matrix notation and the normalization conventions used in [7]; as written, the identification is not immediate.","section":"§4.4, Example 1"}],"recommendation":"major_revision","confidential_remarks":"The paper draws heavily on the authors' own prior work ([25,7]) for the samplet concept and construction; while this is not disqualifying, the referee report should ensure the manuscript is sufficiently self-contained for readers outside that immediate line of work. The main issue is not the construction but the proof of the central localization estimate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take before you spend time on this. The construction is a genuine generalization of samplets from point evaluations to general functionals that form ℓ2-frames or Riesz bases in Banach spaces. The algebra is sound: use the analysis operator's image, build a finite frame, do a QR on the moment matrix to get vanishing moments, then compose with an isometry. The three examples — RKHS point evaluation, Tausch–White wavelets, operator-adapted wavelets — are correctly identified as special cases. The spectral clustering step is a sensible replacement for the Euclidean clustering in [25]. If the paper only claimed the construction and the algebraic vanishing-moment property, I would be comfortable.\n\nThe problem is Theorem 11, the paper's quantitative payoff. The proof has two gaps, and the first compounds the second. It multiplies v ∈ B by a C∞₀ cutoff χ and restricts to supports. That requires B to be closed under pointwise multiplication by smooth cutoffs and to have a restriction/trace to compact sets. The stated assumption B′ ⊂ E′(Ω) does not give you that. The remark that E(Ω) is dense in B is not enough. So the argument already assumes a C∞-module structure that is not in the hypotheses.\n\nEven granting that structure, the proof establishes |(ψᵢ,v)| ≤ √B_N inf_p inf_χ ‖χ(v−p)‖_B, with χ = 1 near supp ψᵢ. But the seminorm in the theorem is ‖v−p‖_{B,K} = inf_{ṽ : ṽ|_K = (v−p)|_K} ‖ṽ‖_B. Since χ(v−p) is one admissible candidate, the inner infimum is larger than or equal to the seminorm, not smaller. The last line of the proof asserts equality without an argument. So Theorem 11 is not proven as stated; the fix is not purely cosmetic. You need either a different definition of the local seminorm (e.g., infimum over functions equal on an open neighborhood of K) plus a trace/extension theorem, or additional hypotheses that make the cutoff infimum equal to the seminorm.\n\nI don't think this is fatal to the whole paper — the construction stands on its own, and the localization result is very likely true in the Sobolev/RKHS examples with a bit more care. But as written, the main quantitative claim is unsupported. The paper deserves a serious referee, but the referee should ask for a rewritten Theorem 11 with explicit hypotheses and a correct proof. I would not cite Theorem 11 in its current form. It is a worthwhile read for people working on multiresolution methods for unstructured data; just don't take the abstract as established.","headline":"The construction is a real generalization and the algebra is clean, but Theorem 11's localization estimate is not proven as stated; the proof needs extra module/restriction hypotheses and a step that equates the cutoff infimum with the local seminorm is unjustified.","tokens_in":16817,"tokens_out":6319,"would_cite":false,"duration_ms":55383,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["42C15","46B15","42C40"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper extends samplet bases with vanishing moments from point evaluations to frames of functionals in Banach spaces, and proves a localization bound for the resulting coefficients.","keywords":["samplets","Banach frames","Riesz bases","vanishing moments","spectral clustering","multiresolution analysis","localization estimate","compactly supported distributions"],"falsifier":"Find a Banach space $B$ that has an $\\ell^2$-frame and $B'\\hookrightarrow\\mathcal{E}'(\\Omega)$ but is not closed under multiplication by $C_0^\\infty$ cutoff functions, and exhibit $v\\in B$ and a samplet $\\psi_i$ for which $\\inf_{p\\in\\mathcal{P}}\\|v-p\\|_{B,\\operatorname{supp}(\\psi_i)}$ is undefined while $(\\psi_i,v)$ is finite; this would show the localization theorem needs an extra hypothesis beyond the stated assumptions.","tokens_in":15752,"feed_emoji":"🧮","tokens_out":10567,"duration_ms":93977,"temperature":0.7,"pith_summary":"The paper sets out to remove the main restriction of samplet constructions, which until now worked only for point-evaluation measurements (Dirac deltas) on a data set. It claims that the same wavelet-style bases—localized functionals with vanishing moments—can be built from any finite collection of linear measurements on a Banach space, provided those measurements form an ℓ2-frame or a Riesz basis and the dual space consists of compactly supported distributions. Because the analysis operator is injective in these cases, the samplet basis is obtained by applying an isometry to the image of the analysis operator, and the multilevel hierarchy comes from spectral clustering of the functionals' supports. The paper's main theoretical output is a localization bound: the coefficient of a generalized samplet against a function v is controlled by how well v can be approximated, on the samplet's support, by the chosen primitives. If the claim holds, scattered-data compression, feature detection, and operator-adapted wavelet construction become instances of one Banach-space framework.","feed_headline":"Samplets go beyond point samples to Banach-space frames","feed_subtitle":"A unified construction yields vanishing-moment bases for frames of functionals, with coefficient localization.","key_machinery":"The central object is the finite analysis operator $T_N^\\star: B\\to\\mathbb{R}^N$ with entries $(T_N^\\star v)_i=(f_i,v)_{B'\\times B}$, whose image carries the Gram structure of the functionals. The construction is carried by Lemma 6: applying any isometry $U$ to the image of $T_N^\\star$ turns the frame $\\{f_i\\}$ into a new sequence $\\{\\psi_i\\}$ that is a Parseval frame—or, for Riesz bases and unitary $U$, an orthonormal basis—with respect to the inner product induced by the frame operator. Vanishing moments are enforced by choosing $U$ through a QR decomposition of the moment matrix $M=[(p_m,\\phi_l)]$ relative to a chosen primitive space $\\mathcal{P}$, so that $(\\psi_i,p)=0$ for all $p\\in\\mathcal{P}$. The multilevel hierarchy is obtained by spectral bisection of a similarity graph built from the supports of the functionals, using the Fiedler vector of the graph Laplacian. The localization proof then cuts a test function $v$ with a smooth bump function $\\chi$ supported near $\\operatorname{supp}(\\psi_i)$, which is legitimate because $B'$ embeds into the compactly supported distributions $\\mathcal{E}'(\\Omega)$.","core_discovery":"On the paper's own terms, the discovery is that none of the essential samplet ingredients—localized basis functionals, vanishing moments, multilevel hierarchy, coefficient decay—depends on point evaluations. For a finite set $f_N:=\\{f_1,\\dots,f_N\\}\\subset B'$ that is an $\\ell^2$-frame for its span (or a subset of a Riesz basis), with $B'\\hookrightarrow \\mathcal{E}'(\\Omega)$, the authors construct a cluster tree by spectral bisection of a similarity graph defined through the supports of the $f_i$. They then define scaling functionals and samplets via a two-scale transform whose matrix is an isometry on the image of the analysis operator $T_N^\\star$; choosing that isometry by a QR decomposition of the moment matrix makes the samplets annihilate a finite-dimensional space $\\mathcal{P}$ of primitives. The resulting sequences are Parseval frames, and orthonormal bases when the underlying functionals form a Riesz basis, with respect to the Hilbert topology induced by the frame operator. Theorem 11 states the localization estimate $|(\\psi_i,v)_{B'\\times B}|\\le \\sqrt{B_N}\\,\\inf_{p\\in\\mathcal{P}}\\|v-p\\|_{B,\\operatorname{supp}(\\psi_i)}$; for polynomial primitives and Sobolev spaces it becomes the familiar diameter-dependent wavelet decay estimate.","pith_inferences":["Because the hierarchy only needs a similarity matrix between functional supports, one could replace the geometric distance with learned or data-driven similarities, producing samplets adapted to a data manifold rather than to physical space.","The framework suggests a modular numerical pipeline—measurement functionals, support graph, spectral clustering, QR moment matrix—that could be implemented as a black-box multiresolution transform for unstructured linear measurements.","The localization theorem implicitly asks that $B$ be closed under multiplication by $C_0^\\infty$ cutoff functions; spaces lacking this module property, beyond the stated dual-embedding assumption, would need an additional hypothesis for the bound to hold.","Choosing the primitives $\\mathcal{P}$ to be operator-dependent spaces, such as ranges of powers of an elliptic operator, would make samplet coefficients decay according to the operator's smoothness, linking the construction directly to numerical homogenization and optimal recovery."],"forward_implications":["Generalized samplets can be built from local averages, derivative evaluations, boundary fluxes, or other linear measurements, so data compression and feature detection are no longer limited to point clouds.","In Sobolev spaces with polynomial primitives, the localization bound becomes the classical wavelet estimate $|(\\psi_i,v)|\\le C(\\operatorname{diam}\\operatorname{supp}\\psi_i)^{m-k}\\|v\\|_{W^{m,p}}$, via the Bramble–Hilbert lemma.","When the functionals form a Riesz basis, the samplet transform is an orthonormal change of basis; when they only form a frame, it still produces a Parseval frame with respect to the frame-induced inner product.","The multilevel hierarchy is obtained from support similarities alone, so the construction applies to functionals supported on complicated shapes without a structured grid."],"supporting_citations":[{"why":"Provides the original samplet construction for point evaluations that this article generalizes to Banach-space functionals.","marker":"[25]"},{"why":"Introduces frames and the analysis/synthesis operator formalism on which the construction rests.","marker":"[21]"},{"why":"Establishes the Banach-frame setting with square-summable coefficients and the canonical dual frame used throughout.","marker":"[5]"},{"why":"Supplies the QR-decomposition-of-moment-matrix technique that enforces vanishing moments.","marker":"[1]"},{"why":"Provides the spectral clustering method and Fiedler-vector bisection used to build the multilevel hierarchy.","marker":"[41]"},{"why":"Gives the distribution identity $(\\psi_i,v)=(\\psi_i,\\chi v)$ for compactly supported $\\psi_i$ that underpins the localization proof.","marker":"[34]"},{"why":"Provides the Tausch–White multiwavelet construction recovered as an example of the generalized framework.","marker":"[39]"},{"why":"Provides the operator-adapted wavelet and optimal recovery spline setting used as a third example.","marker":"[32]"}],"fun_headline_variants":["Samplets go Banach: functionals instead of points","Generalized samplets: frames over Banach spaces","Samplets for Banach frames: no point evaluations","From points to functionals: samplets in Banach spaces","Banach-space samplets built from functionals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole construction assumes the dual space $B'$ embeds into compactly supported distributions on a domain and, in the localization proof, that elements of $B$ can be restricted to compact sets and multiplied by smooth cutoff functions; if a Banach space satisfying the stated hypotheses lacks that multiplicative and restriction structure, the main quantitative estimate does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Samplets go Banach: functionals instead of points","Generalized samplets: frames over Banach spaces","Samplets for Banach frames: no point evaluations","From points to functionals: samplets in Banach spaces","Banach-space samplets built from functionals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00062,"raw_usage":{"total_tokens":2935,"prompt_tokens":1066,"completion_tokens":1869,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":1791}},"tokens_in":682,"tokens_out":1869,"duration_ms":12787,"temperature":1.0,"reasoning_tokens":1791,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:50:27.429451+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a Banach space $B$ that has an $\\ell^2$-frame and $B'\\hookrightarrow\\mathcal{E}'(\\Omega)$ but is not closed under multiplication by $C_0^\\infty$ cutoff functions, and exhibit $v\\in B$ and a samplet $\\psi_i$ for which $\\inf_{p\\in\\mathcal{P}}\\|v-p\\|_{B,\\operatorname{supp}(\\psi_i)}$ is undefined while $(\\psi_i,v)$ is finite; this would show the localization theorem needs an extra hypothesis beyond the stated assumptions.","supporting_citations":[{"cited_title":"Harbrecht and M","cited_arxiv_id":null,"evidence_quote":"Provides the original samplet construction for point evaluations that this article generalizes to Banach-space functionals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces frames and the analysis/synthesis operator formalism on which the construction rests."},{"cited_title":"Balazs and H","cited_arxiv_id":null,"evidence_quote":"Establishes the Banach-frame setting with square-summable coefficients and the canonical dual frame used throughout."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the QR-decomposition-of-moment-matrix technique that enforces vanishing moments."},{"cited_title":"Von Luxburg","cited_arxiv_id":null,"evidence_quote":"Provides the spectral clustering method and Fiedler-vector bisection used to build the multilevel hierarchy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the distribution identity $(\\psi_i,v)=(\\psi_i,\\chi v)$ for compactly supported $\\psi_i$ that underpins the localization proof."},{"cited_title":"Tausch and J","cited_arxiv_id":null,"evidence_quote":"Provides the Tausch–White multiwavelet construction recovered as an example of the generalized framework."},{"cited_title":"Owhadi and C","cited_arxiv_id":null,"evidence_quote":"Provides the operator-adapted wavelet and optimal recovery spline setting used as a third example."}],"review_version":1}