Hyperinterpolation: Discrete L2 Projection

• Hyperinterpolation is a quadrature-based discretization of the L2 projection onto finite-dimensional polynomial spaces, replacing continuous inner products with cubature sums.
• It guarantees projection properties under exact cubature and maintains stability through modified techniques like Marcinkiewicz–Zygmund inequalities and regularized least-squares.
• Variants such as filtered, spectral multiplier, and thresholded hyperinterpolation enhance localization and robustness, expanding applications from spherical harmonics to sparse recovery.

Hyperinterpolation is a quadrature-based discretization of the $L^2$-orthogonal projection onto a finite-dimensional polynomial or spectral space. In its classical form, it replaces the projection coefficients $\langle f,p_\alpha\rangle$ by cubature sums over sampled nodes, thereby producing a fully discrete approximant that retains projection-like structure when the cubature is sufficiently exact. The construction originated in Sloan’s 1995 framework and has since expanded from spherical polynomials to general regions, compact manifolds, filtered and regularized variants, and more recent formulations that bypass exact cubature through Marcinkiewicz–Zygmund inequalities, Sobolev discrepancy, and quasi-Monte Carlo sampling (An et al., 6 Oct 2025, Wu, 18 May 2026).

1. Formal definition and projection viewpoint

Let $\Omega$ be a compact region or manifold with finite positive measure $d\omega$, and let $\Pi_n$ denote a finite-dimensional polynomial or spectral subspace with orthonormal basis $\{\phi_k\}$. The exact $L^2$-projection onto $\Pi_n$ is

$P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$

Hyperinterpolation replaces these continuous coefficients by discrete inner products computed from a positive-weight quadrature or cubature rule $\{(x_j,w_j)\}$: $\langle f,p_\alpha\rangle$0 Under exactness of degree $\langle f,p_\alpha\rangle$1, this reproduces $\langle f,p_\alpha\rangle$2 exactly, because products of basis functions of degree at most $\langle f,p_\alpha\rangle$3 lie in degree at most $\langle f,p_\alpha\rangle$4 (An et al., 6 Oct 2025, An et al., 2022).

On the sphere $\langle f,p_\alpha\rangle$5, the basis is typically given by spherical harmonics $\langle f,p_\alpha\rangle$6, and the projection kernel is

$\langle f,p_\alpha\rangle$7

The hyperinterpolant then takes the kernel form

$\langle f,p_\alpha\rangle$8

which is the discrete analogue of the orthogonal projector $\langle f,p_\alpha\rangle$9 (Wu, 18 May 2026, An et al., 17 Jan 2026).

Classically, hyperinterpolation is also equivalent to a weighted discrete least-squares problem. Writing the approximant as $\Omega$0, the coefficient vector solves

$\Omega$1

and exactness implies $\Omega$2, so the minimizer is given directly by the discrete Fourier coefficients. This equivalence is central to later regularized variants, because $\Omega$3, $\Omega$4, and mixed penalties then decouple coefficientwise under the same exactness hypothesis (An et al., 2022, An et al., 2020).

2. Classical theory: exact cubature, stability, and operator growth

The classical theory assumes positive weights and cubature exactness on $\Omega$5. Under that hypothesis, hyperinterpolation inherits the projection property on $\Omega$6 and satisfies the standard $\Omega$7 estimates

$\Omega$8

where $\Omega$9 and $d\omega$0 (An et al., 6 Oct 2025, An et al., 2020).

On the sphere, the operator norm in $d\omega$1 behaves polynomially with degree. The survey literature reports

$d\omega$2

and on $d\omega$3,

$d\omega$4

These estimates are the analogue of Lebesgue-constant growth for hyperinterpolation and quantify the uniform-norm amplification induced by the discrete projector (An et al., 17 Jan 2026).

On the unit ball $d\omega$5 with Gegenbauer weight

$d\omega$6

the Lebesgue constant of degree-$d\omega$7 hyperinterpolation grows like

$d\omega$8

when $d\omega$9 is a positive half-integer. In the unweighted disk $\Pi_n$0, this specializes to linear growth in $\Pi_n$1, improving earlier $\Pi_n$2 bounds for specific cubature constructions (Wade, 2012).

A recurring structural point is the role of exact cubature in guaranteeing algebraic reproduction. Exactness of degree $\Pi_n$3 is sufficient, but often expensive. This tension drives much of the modern literature: either one weakens exactness while preserving stability through norm equivalences, or one modifies the projection kernel to gain localization and robustness.

3. Relaxed exactness, Marcinkiewicz–Zygmund conditions, and unfettered hyperinterpolation

One line of development relaxes the degree-$\Pi_n$4 cubature requirement to degree $\Pi_n$5, with $\Pi_n$6, supplemented by a Marcinkiewicz–Zygmund (MZ) inequality on $\Pi_n$7: $\Pi_n$8 In that setting, hyperinterpolation no longer reproduces all of $\Pi_n$9; it reproduces $\{\phi_k\}$0, remains $\{\phi_k\}$1-stable with constants independent of $\{\phi_k\}$2, and its approximation error is governed by $\{\phi_k\}$3 rather than $\{\phi_k\}$4. This trades convergence speed for reduced cubature cost and a larger admissible family of quadrature rules (An et al., 2022).

On the sphere, the exactness assumption can be bypassed entirely. The unfettered hyperinterpolation framework keeps the same operator

$\{\phi_k\}$5

but replaces exactness by MZ frame bounds on $\{\phi_k\}$6. The resulting $\{\phi_k\}$7-error estimate has two terms: $\{\phi_k\}$8 where $\{\phi_k\}$9 is a best uniform approximant in $L^2$0. The first term is the classical approximation component; the second is the compensation for loss of exactness (An et al., 2022).

This formulation substantially enlarges the admissible node families. Equal-area points, Coulomb-energy points, Fekete points, random samples, and spherical $L^2$1-designs all fit naturally into the analysis when their MZ constants are controlled. In quasi-Monte Carlo regimes, refined $L^2$2-type behavior becomes available for Sobolev classes, and the same geometric viewpoint also enters product-integration schemes for weakly singular integral equations on $L^2$3, where the MZ property ties solver accuracy to point-set geometry and mesh norm (An et al., 2022, An et al., 2024).

A common misconception is that exact polynomial reproduction is the essential feature of hyperinterpolation. The modern MZ-based theory shows instead that stability and approximation can survive without exact reproduction, provided the discrete $L^2$4 structure of the trial space remains controlled.

4. Spectral multiplier hyperinterpolation and Sobolev discrepancy

A second, more analytic generalization replaces the sharp spectral projector by a general spectral multiplier. On $L^2$5, with Laplace–Beltrami eigenvalues $L^2$6, one considers

$L^2$7

where

$L^2$8

Sampling with a measure $L^2$9 yields the discrete operator

$\Pi_n$0

The key identity is that the discretization error is the action of the same spectral operator on the discrepancy distribution $\Pi_n$1 (Wu, 18 May 2026).

The weak cubature assumption is no longer algebraic exactness but a Sobolev discrepancy estimate

$\Pi_n$2

with $\Pi_n$3 the mesh norm. Combined with an approximation property

$\Pi_n$4

and a smoothing bound

$\Pi_n$5

this gives

$\Pi_n$6

Balancing the two terms by $\Pi_n$7 yields the rate

$\Pi_n$8

The same framework extends to quasi-Monte Carlo designs, where $\Pi_n$9 for equal weights (Wu, 18 May 2026).

This theory applies not only to the sharp spectral projection $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$0, but also to compactly supported smooth filters, Bessel potential operators, and heat kernel operators. The sharp projector recovers classical hyperinterpolation, but it lacks strong spatial localization, so uniform $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$1-stability generally fails because Lebesgue constants grow with $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$2. By contrast, sufficiently localized multipliers satisfy kernel bounds such as

$P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$3

or Gaussian heat-kernel bounds, and then one obtains uniform $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$4-stability under quasi-uniform nodes, geometric weight bounds, and the bandlimit constraint $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$5 (Wu, 18 May 2026).

The conceptual significance is the separation of three roles: continuous approximation belongs to the multiplier, discretization error is measured through smoothing acting on discrepancy, and geometric quality enters only through Sobolev discrepancy. This shifts the theory away from exact polynomial cubature toward spectral approximation from scattered data.

5. Filtered, generalized, sparse, and regularized variants

Filtered hyperinterpolation modifies the high-frequency tail of the spectrum. In the spherical setting, a compactly supported smooth filter $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$6 with $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$7 on $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$8 and $P_n f=\sum_k \langle f,\phi_k\rangle \phi_k.$9 on $\{(x_j,w_j)\}$0 defines

$\{(x_j,w_j)\}$1

and the associated filtered operator has bounded $\{(x_j,w_j)\}$2-norms independent of $\{(x_j,w_j)\}$3 and localized kernels under $\{(x_j,w_j)\}$4 smoothness assumptions. Filtered hyperinterpolation is central in manifold learning and in modern $\{(x_j,w_j)\}$5-stable constructions on the sphere (Montúfar et al., 2020).

Generalized hyperinterpolation replaces the sharp reproducing kernel by a zonal kernel $\{(x_j,w_j)\}$6. In the algebraic study of the “hyperinterpolation-class” on $\{(x_j,w_j)\}$7, generalized hyperinterpolation is hyper self-adjoint and commutative with hyperinterpolation under the discrete inner product $\{(x_j,w_j)\}$8. That framework also defines hyper projection operators, hyper algebra, and ideal-type relations among thresholded variants (An et al., 2024).

Regularized forms arise from penalized discrete least squares. Lasso hyperinterpolation applies soft thresholding to the discrete coefficients,

$\{(x_j,w_j)\}$9

and is the unique solution of an $\langle f,p_\alpha\rangle$00-regularized weighted least-squares problem when $\langle f,p_\alpha\rangle$01. Its operator norm remains bounded independently of degree, and its $\langle f,p_\alpha\rangle$02-error can be smaller than that of classical hyperinterpolation when noise is large (An et al., 2020).

Hard thresholding hyperinterpolation replaces soft thresholding by a coefficientwise hard threshold. It is the unique minimizer of an $\langle f,p_\alpha\rangle$03-regularized weighted discrete least-squares problem, is idempotent, commutes with hyperinterpolation, and satisfies a discrete Pythagorean theorem with respect to the quadrature inner product (An et al., 2022).

Hybrid hyperinterpolation combines spectral filtering with soft thresholding. It is the unique minimizer of an $\langle f,p_\alpha\rangle$04-regularized discrete least-squares functional, its operator norm is reduced by a factor $\langle f,p_\alpha\rangle$05 relative to the classical hyperinterpolation bound, and numerical tests on spheres, disks, cubes, and unions of disks show improved robustness under Gaussian and impulse noise (An et al., 2023).

Recovery-thresholding hyperinterpolations extend this logic further by embedding hard, springback, and Newton thresholding into the hyperinterpolation coefficient domain for sparse signal reconstruction. Because exact quadrature yields $\langle f,p_\alpha\rangle$06, the resulting nonconvex regularized problems decouple into one-dimensional minimizations, which the Newton-thresholding scheme solves coefficientwise (An et al., 23 Jul 2025).

For singular or oscillatory factors $\langle f,p_\alpha\rangle$07, standard hyperinterpolation may require many quadrature points. Efficient hyperinterpolation addresses this by hyperinterpolating the smooth factor $\langle f,p_\alpha\rangle$08 and integrating $\langle f,p_\alpha\rangle$09 against the polynomial expansion through modified moments. The reported theory and experiments show that this product-integration viewpoint can outperform classical hyperinterpolation when the number of integration points is limited (An et al., 2022).

6. Domains, computational realizations, applications, and limitations

Hyperinterpolation now exists on intervals, circles, disks, balls, cubes, spheres, spherical polygons, spherical triangles, unions of disks, compact manifolds, radial manifolds, and polyhedral elements (An et al., 6 Oct 2025, Sommariva, 2024, Sommariva et al., 5 Feb 2025). The underlying node sets include Gauss–Legendre and Clenshaw–Curtis quadratures on intervals, spherical $\langle f,p_\alpha\rangle$10-designs and Lebedev rules on $\langle f,p_\alpha\rangle$11, Xu-type cubatures on tensor-product domains, PI-type cubatures on general regions, and QMC lattice or polynomial lattice rules on high-dimensional cubes (An et al., 17 Jan 2026, An et al., 13 Jun 2025).

On spherical polygons, low-cardinality positive cubature formulas are built numerically through gnomonic projection, spherical triangulation, near-exact moment matching, and Carathéodory–Tchakaloff compression; in the reported Australia example, compression reduced an ADE-10 rule from $\langle f,p_\alpha\rangle$12 nodes to $\langle f,p_\alpha\rangle$13 nodes with moment error about $\langle f,p_\alpha\rangle$14 (Sommariva, 2024). On polyhedral elements, hyperinterpolation of the indicator function in a bounding box produces cubature rules exact on $\langle f,p_\alpha\rangle$15; the final weights are obtained by a single matrix–vector product, and the stability theory allows some negative weights without loss of boundedness (Sommariva et al., 5 Feb 2025).

Applications are correspondingly broad. Filtered hyperinterpolation on compact manifolds supports distributed learning from deterministic or random samples and attains optimal non-distributed order together with matching distributed rates under per-server size conditions (Montúfar et al., 2020). Lebedev-based hyperinterpolation underlies spectral exterior-calculus and hydrodynamic solvers on radial manifolds, where the reported convergence is spectral or super-algebraic depending on geometry and resolution (Gross et al., 2017, Gross et al., 2018). On $\langle f,p_\alpha\rangle$16, hyperinterpolation-driven product integration yields practical $\langle f,p_\alpha\rangle$17 error bounds for second-kind integral equations with weakly singular kernels (An et al., 2024). In signal processing, thresholded hyperinterpolation variants are used for sparse recovery and denoising under Gaussian and impulse noise (An et al., 23 Jul 2025).

Several limitations are persistent across the literature. Exact positive cubature of degree $\langle f,p_\alpha\rangle$18 is difficult in high dimension; relaxed exactness slows convergence from $\langle f,p_\alpha\rangle$19 to $\langle f,p_\alpha\rangle$20; sharp spectral projection lacks the localization required for uniform $\langle f,p_\alpha\rangle$21-stability; Bessel and heat-kernel approximations saturate at order $\langle f,p_\alpha\rangle$22; the Sobolev-discrepancy theory requires $\langle f,p_\alpha\rangle$23 and $\langle f,p_\alpha\rangle$24; hard-thresholding hyperinterpolation is basis-dependent and not symmetric in general; and QMC or MZ-based theories still require geometric control of nodes and weights (An et al., 2022, Wu, 18 May 2026, An et al., 2022, An et al., 13 Jun 2025).

Open problems identified in the cited literature include extending the discrepancy-based framework to other manifolds and measures, optimizing filters that combine strong localization with improved low-frequency consistency, developing adaptive and fast high-dimensional constructions, understanding sharper MZ constants for practical node families, and further integrating hyperinterpolation with sparse and low-rank approximation architectures (Wu, 18 May 2026, An et al., 13 Jun 2025, Che et al., 15 Oct 2025). These directions indicate that hyperinterpolation is no longer only a discrete projection method tied to exact cubature, but a broader approximation paradigm connecting spectral truncation, sampling geometry, regularization, and numerical integration.