A version-1 arXiv preprint proposes a mathematical way to build multiscale framelet transforms from scattered, progressively refined samples on compact manifolds. Its central guarantee is controlled norm distortion: for a chosen tolerance 0 < η < 1, weighted p-norms on the dyadic polynomial spaces P_(2^j) stay between 1 − η and 1 + η.
The construction is deterministic and uses nested point sets with controlled spacing and cardinality order. These Marcinkiewicz–Zygmund, or MZ, measures attach positive weights to the points; at a fixed polynomial degree n, the reported tolerance falls at least like C_(M,p) n 2^-j for sufficiently large j, so the MZ constants approach 1 as the sets are refined.
Where the extra inversion comes in
Framelets are the paper’s multiscale building blocks. In the fully discrete procedure, a filter bank analyzes the samples, a one-pass synthesis puts them back together, and canonical reconstruction inverts the frame operator with conjugate gradients, an iterative solver. Under MZ sampling, the one-pass route is generally not exact.
With p = 2 and the stated generator assumptions, the semi-discrete system is η-nearly tight: its frame bounds are 1 − η and 1 + η. Adding s nested refinement levels is stated to reduce the theoretical frame-bound deviation to C_M 2^-s, making the hierarchy asymptotically tight.
The same deviation also sets a bound on the inversion problem. At a finite level, the coefficient frame operator has a condition number—a measure of how difficult the numerical solve is—of at most (1 + ε)/(1 − ε), so a smaller deviation means a better-conditioned conjugate-gradient reconstruction. The approximation result separates finite-space approximation error from coefficient perturbation: canonical reconstruction has perturbation factor 1/√(1 − ε), and when coefficient error e = 0, every polynomial in P_(2^L) is exactly reconstructed.
What the examples showed
The numerical tests used a sphere and a flat torus. In a Wendland sphere signal, one-pass reconstruction error fell as sampling density increased, while canonical conjugate-gradient error was several orders smaller. Its eventual scale was compatible with a reconstruction tolerance τ_rec = 10^-9 after conditioning and fast-transform error.
A second sphere example used an Earth texture with cutoffs 128, 64 and 32, corresponding to 41,603, 10,563 and 2,723 nodes. The projection error was 1.5885 × 10^-1. Relative to the projected signal, one-pass error was 3.4203 × 10^-2, compared with 6.4390 × 10^-10 for canonical reconstruction. Relative to the original image, the error was 1.5885 × 10^-1 and the signal-to-noise ratio was 15.888 dB; the larger error reflected projection rather than solver error.
In an oscillatory torus example, levels 6, 5 and 4 used cutoffs 64, 32 and 16, an oversampling factor of 2.5, jitter of 0.25 and a fixed random seed of 5. Projection error was 4.8773 × 10^-11; one-pass synthesis error was 1.2697 × 10^-2, compared with 2.7717 × 10^-10 after canonical inversion. The paper describes these as observed MZ-type-rule results, not theoretical frame bounds.
A separate torus denoising test found lower relative error and higher signal-to-noise ratio with canonical reconstruction than with one-pass synthesis at noise levels σ = 0.10, 0.20 and 0.30. At σ = 0.20, the noisy input, one-pass and canonical relative errors were 9.0461 × 10^-2, 3.7380 × 10^-2 and 3.5799 × 10^-2. Their corresponding signal-to-noise ratios were 20.871, 28.547 and 28.922 dB.
What the result does not settle
The sphere rules were quasi-uniform MZ-type sampling rules built from nested farthest-point prefixes and partition-type weights, explicitly rather than spherical designs or exact quadrature. The examples therefore illustrate controlled non-exact sampling, not a head-to-head demonstration of superiority over exact quadrature.
The mathematical guarantees are conditional on compact-manifold geometry, generator support and partition assumptions, and the specified MZ constructions. Theoretical constants are largely symbolic rather than numerically instantiated in the examples, while oversampling and tighter tolerances add node, computation and storage costs. Canonical reconstruction also depends on conjugate-gradient stopping criteria and fast-transform accuracy.
Within that scope, the contribution is a chain of mathematical controls linking sampling distortion, frame tightness, solver conditioning and reconstruction error. Whether the trade-off holds across other manifolds, sampling patterns and larger implementations—and how it compares with other non-exact schemes—remains open.
Paper data and sources
Original title: Nearly tight framelet systems from nested Marcinkiewicz--Zygmund measures on compact Riemannian manifolds
Authors: Hao-Ning Wu, Xiaosheng Zhuang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text