Multiresolution Samplet Framework

• The multiresolution samplet framework is a wavelet-inspired method that creates hierarchical, sparse representations from scattered data in arbitrary dimensions.
• It uses balanced spatial clustering and local QR decompositions to construct orthonormal bases with vanishing moments, ensuring efficient fast transforms and compressed matrix algebra.
• The framework supports reliable numerical linear algebra, learning, and signal processing on diverse domains including Euclidean spaces, manifolds, and graphs.

The multiresolution samplet framework is a wavelet-like construction generalized to scattered data sites in arbitrary dimensions, extended to various data types—including Euclidean, manifold, and graph domains. It provides a means for hierarchical, localized, and sparse representation of functions, operator kernels, and signals on nonuniform discrete sets by constructing orthonormal bases of signed measures (“samplets”) with spatial localization, vanishing moments for polynomial annihilation, and multiresolution nesting. For numerical linear algebra, learning, and signal processing on scattered, high-dimensional, or graph-based data, the framework enables compressed matrix algebra, fast transforms, and structured sparsity with rigorous approximation properties.

1. Theoretical Foundations and Multiresolution Structure

Samplets are constructed as discrete signed measures on a finite data set $$X = \{x_1, ..., x_N\} \subset \Omega \subset \mathbb{R}^d$$ (or, more generally, vertices on a graph). The ambient space is the span of Dirac measures,

$$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$

with the inner product $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$.

A hierarchy of nested subspaces

$$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$

is constructed using balanced spatial clustering of $X$, typically via dyadic (binary) partitioning. At each level $j$, the subspace $\mathcal{X}'_j$ is spanned by scaling distributions $\Phi_j = \{\phi_{j,k}\}$, and its orthogonal complement in $\mathcal{X}'_{j+1}$ is spanned by detail distributions (“samplets”) $\Sigma_j = \{\sigma_{j,k}\}$: $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$0 The global samplet basis

$$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$1

is an orthonormal basis for $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$2.

Each samplet $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$3 exhibits

• Vanishing moments: $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$4 for all polynomials $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$5 of degree $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$6,
• Spatial locality: $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$7 is contained in a spatial cluster at scale $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$8,
• $$\mathcal{X}' = \mathrm{span}\{\delta_{x_i}\}_{i=1}^N$$9 stability: $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$0, where $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$1.

The two-scale relation is implemented via local QR decomposition of the moment matrix on each cluster, ensuring refinement, orthogonality, and vanishing moments.

This basis construction generalizes to graphs by first partitioning the vertex set into clusters, embedding each cluster into a Euclidean patch (usually via Isomap or other manifold learning approaches), performing the samplet construction in the embedded domain, and pulling back the basis to the original graph (Elefante et al., 25 Jul 2025).

2. Computational Algorithms: Fast Transforms and S-formatted Matrix Algebra

A “fast samplet transform” enables a linear-time analysis and synthesis approach for expressing a discrete measure or data vector in the samplet basis (Avesani et al., 17 Jul 2025). The transform, both forward and inverse, can be computed in $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$2 operations after preprocessing, exploiting the binary tree structure.

Kernel matrices $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$3 associated to a Calderón–Zygmund-type kernel $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$4 are represented in the samplet basis as $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$5, which is only quasi-sparse. Imposing a fixed block sparsity pattern, called the S-format, is achieved by retaining blocks $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$6 only if their clusters are within a prescribed scale-dependent distance: $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$7 The S-format matrix $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$8 then has $$(\sum_i u_i \delta_{x_i}, \sum_j v_j \delta_{x_j}) = \sum_{i=1}^N u_i v_i$$9 nonzeros, and construction leverages fast multipole methods and $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$0-matrix-style acceleration (Harbrecht et al., 2022).

Arithmetic operations in S-format (addition, multiplication) are performed on the fixed sparsity pattern, with complexity $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$1 and $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$2, respectively. Matrix inversion and selected inversion are supported, enabling efficient computation of select entries of the inverse in $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$3 (for $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$4) or $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$5 (for $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$6), and often nearly linear time for practical data (Harbrecht et al., 2022).

3. Approximation Rates, Coefficient Decay, and Regularity Detection

Samplet coefficients $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$7 exhibit decay rates governed by the pointwise or microlocal regularity of the underlying function $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$8. For $$\mathcal{X}'_0 \subset \mathcal{X}'_1 \subset \cdots \subset \mathcal{X}'_J = \mathcal{X}'$$9 and $X$0,

$X$1

For balanced binary trees in $X$2, this gives

$X$3

This decay yields optimal $X$4-term approximation rates: $X$5 A direct implication is that samplets enable local smoothness detection (regularity estimation) on irregular and high-dimensional data, using the distribution of significant coefficients (Avesani et al., 17 Jul 2025).

4. Multiresolution Samplet Pursuit and Sparsity-Driven Learning

Embedding samplets via the Riesz map into an RKHS (defined by a reproducing kernel $X$6) provides a dictionary for sparsity-driven and compressed learning: $X$7 The expansion

$X$8

covers a substantially larger class of sparse functions compared to expansions in kernel translates $X$9 (Baroli et al., 2023).

The $j$0-regularized problem in the samplet domain,

$j$1

enables efficient estimation and induces sparsity aligned with signal multiresolution structure. Fast Iterative Shrinkage-Thresholding (FISTA) and a semismooth Newton (MRSSN) method efficiently solve these large-scale problems, with MRSSN achieving dramatic active-set sparsity and large speedups relative to single-scale approaches (Baroli et al., 2023). Benchmarking reveals orders-of-magnitude improvements in computational performance (factor $j$220 in run-time) and compression efficiency for piecewise smooth and multiscale signals.

5. Infinite Data Limit, Multiwavelet Structure, and Generalization

Analysis in the probabilistic (infinite data) regime demonstrates that samplet constructions converge to multiwavelet systems with local polynomial densities (“broken polynomials”) as the number of sample sites increases (Giacchi et al., 2 Apr 2026). For congruent (uniform dyadic) spatial partitions, this yields classical multiwavelets with scale/partition-invariant filter coefficients, matching Alpert-type constructions for compactly supported multiwavelets.

A key feature is vanishing-moment flexibility: by choosing appropriate primitive function sets (e.g., total degree, tensor product, or anisotropic polynomial indices), the construction supports non-tensor-product, anisotropic, or domain-adapted multiresolution analyses. Numerically, samplet weights stabilize rapidly and exhibit Monte Carlo ($j$3) or quasi-Monte Carlo ($j$4) convergence, depending on data sampling (Giacchi et al., 2 Apr 2026).

6. Extensions to Graph Signals and Manifold Domains

The samplet framework extends to discrete signal analysis on weighted graphs $j$5. By partitioning $j$6 into patches (METIS K-way clustering), embedding each patch into a local Euclidean space (via landmark Isomap), constructing and pulling back the samplet basis, one achieves an orthonormal, multiresolution transform with vanishing moments defined with respect to local graph-polynomial charts (Elefante et al., 25 Jul 2025).

For smooth or microlocal-smooth graph signals, samplet coefficients decay as

$j$7

yielding geometric decay rates and optimal $j$8-term approximation. Empirical studies show compression ratios for manifold-based graph signals improved by one to two orders of magnitude over Haar graph wavelets, with $j$9 transform complexity.

7. Applications and Demonstrated Impact

The multiresolution samplet algebra underpins fast, scalable algorithms for kernel-based learning and inference, including spatial statistics and Gaussian process regression. Samplet-compressed kernel matrices enable near-linear complexity for matrix assembly, S-format algebra, and (selected) inversion, all with end-to-end error control (Harbrecht et al., 2022). Posterior mean and covariance evaluation for Gaussian processes can be carried out efficiently even for large data sets (e.g., $\mathcal{X}'_j$0).

In sparse learning, samplet basis pursuit via $\mathcal{X}'_j$1 methods directly exploits the multiresolution sparsity structure, giving both algorithmic scalability and optimal error/compression for piecewise smooth signals. Large-scale applications such as space-time environmental data analysis, high-resolution surface reconstruction from nonuniform point clouds, and feature-aware manifold-based signal compression have demonstrated the framework’s flexibility and computational tractability over previous methods (Harbrecht et al., 2022, Baroli et al., 2023, Avesani et al., 17 Jul 2025, Elefante et al., 25 Jul 2025).

---

In sum, the multiresolution samplet framework provides a theoretically justified, algorithmically efficient, and widely extensible approach to hierarchical analysis and learning on scattered data, kernel operators, and signals on both Euclidean and non-Euclidean domains, tightly integrating multiresolution structure, vanishing-moment theory, and computational sparsity (Harbrecht et al., 2022, Baroli et al., 2023, Avesani et al., 17 Jul 2025, Giacchi et al., 2 Apr 2026, Elefante et al., 25 Jul 2025).