Robust Hermite Expansion Methods

• Robust Hermite expansion is a family of modified Hermite representations that preserves key invariants through scaling, filtering, and projection to ensure stability in complex systems.
• It is applied in numerical analysis, kinetic theory, stochastic processes, and approximation theory to control truncation errors and maintain conservation laws.
• The methodology enhances sparse recovery in generalized polynomial chaos and improves robustness in lattice Boltzmann and density estimation models via adaptive basis transformations.

Searching arXiv for recent and foundational papers on robust Hermite expansions across numerical analysis, kinetic theory, stochastic processes, and approximation theory. arxiv_search(query="Hermite expansion robust approximation Hermite functions lattice Boltzmann generalized polynomial chaos stable transforms", max_results=10) Robust Hermite expansion denotes a family of Hermite-based representations in which the polynomial or function basis is modified, truncated, scaled, rotated, regularized, or projected so that the resulting approximation remains stable under the structural constraints of the problem. In different literatures, robustness means different things: correct recovery of macroscopic conservation laws in kinetic models, controlled truncation error for almost time- and band-limited functions, stable forward and inverse transforms at Gauss–Hermite nodes, preservation of sparsity under Gaussian rotations, or enforcement of non-negativity and normalization for density and copula constructions (Coelho et al., 2013, Jaming et al., 2014, Webb et al., 2 Apr 2026, Shiraya et al., 2023).

1. Foundations and core objects

The common algebraic core is the Hermite system. In one standard normalization, the physicists’ Hermite polynomials are

$H_n(x)=e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$

and the normalized Hermite functions are

$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$

The family $(h_n)_{n\ge 0}$ is an orthonormal basis of $L^2(\mathbb{R})$, with parity determined by $n$, and it diagonalizes the unitary Fourier transform through

$\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$

These properties make Hermite expansions simultaneously natural for Gaussian weights, harmonic-oscillator spectral theory, and time–frequency analysis (Jaming et al., 2014).

In multivariate Gaussian settings, the basis is adapted to the covariance structure. For $X\sim \mathcal{N}(\mu,\Sigma)$, multivariate Hermite polynomials are defined by the generating function

$$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$

For general $\Sigma$, the associated system is complete in $L^2(\mu_\Sigma)$, but only weakly orthogonal: Hermite polynomials of different total degree are orthogonal, जबकि polynomials of the same degree are generally coupled through a non-diagonal Gram block. This distinction is the algebraic source of many robustification procedures, since it replaces scalar coefficient formulas by blockwise projection systems (Rahman, 2017).

A second foundational distinction concerns anisotropy. Several kinetic constructions replace the scalar-temperature Gaussian weight by one built from a full symmetric positive definite temperature tensor $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$0. The generalized Hermite basis then aligns the expansion with the intrinsic anisotropy of the state rather than with coordinate axes, which is crucial for hyperbolicity and coordinate-invariant closure (Fan et al., 2014).

2. Recurring mechanisms of robustness

Across the literature, robustness is achieved by a small number of recurring devices: scaling of the basis, filtering of kernels, orthogonal–diagonal factorization, rotations preserving Gaussianity, convex projection onto admissible sets, and closure at a sufficiently high Hermite order.

Domain	Robustness mechanism	Representative paper
Nearly time-/band-limited approximation	scaled Hermite basis and explicit projection error bounds	(Jaming et al., 2014)
Local approximation from sampled data	filtered kernels and Marcinkiewicz–Zygmund quadrature	(Mhaskar, 2016)
Discrete Hermite transforms	factorization $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$1 with $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$2 orthogonal	(Webb et al., 2 Apr 2026)
Sparse polynomial chaos	iterative orthogonal rotations $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$3	(Yang et al., 2015)
Density and copula correction	projection enforcing non-negativity and normalization	(Shiraya et al., 2023)
Kinetic moment closures	anisotropic basis and globally hyperbolic regularization	(Fan et al., 2014)

These mechanisms are mathematically different, but they share a common objective: the Hermite representation should respect the dominant invariants of the target problem. In approximation theory, the invariants are $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$4-orthogonality and locality; in stochastic expansions they are Gaussian measure and sparsity structure; in kinetic theory they are mass, momentum, energy, and hyperbolicity; in density estimation they are positivity and unit mass.

3. Approximation theory, localization, and function spaces

For almost time- and band-limited functions, robustness is expressed by explicit error control under Hermite projection. If $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$5 is $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$6-almost time-limited to $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$7 and $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$8-almost band-limited to $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$9, and if $(h_n)_{n\ge 0}$0, then the local projection error satisfies

$(h_n)_{n\ge 0}$1

The scaled projection $(h_n)_{n\ge 0}$2 improves the geometric factor to $(h_n)_{n\ge 0}$3, and the projections are contractions in $(h_n)_{n\ge 0}$4. This is the canonical approximation-theoretic meaning of a robust Hermite expansion: leakage and truncation errors are separated explicitly, and scaling can be used to match concentration in physical space (Jaming et al., 2014).

A complementary construction uses smooth low-pass filters on Hermite coefficients. The filtered kernel

$(h_n)_{n\ge 0}$5

satisfies off-diagonal localization bounds of the form

$(h_n)_{n\ge 0}$6

which yield wavelet-like block operators $(h_n)_{n\ge 0}$7. The local behavior of $(h_n)_{n\ge 0}$8 characterizes local Besov smoothness, and analogous expansions can be built from point values at arbitrary nodes through Marcinkiewicz–Zygmund quadrature measures (Mhaskar, 2016).

At the level of function spaces, the robust endpoint is the Fréchet space of functions satisfying, for every $(h_n)_{n\ge 0}$9,

$L^2(\mathbb{R})$0

The corresponding Hermite characterization is exact: $L^2(\mathbb{R})$1 satisfies these near-optimal Gaussian bounds if and only if its Hermite coefficients obey

$L^2(\mathbb{R})$2

More generally, weighted variants are governed by the Young conjugate $L^2(\mathbb{R})$3 of $L^2(\mathbb{R})$4, giving coefficient bounds $L^2(\mathbb{R})$5. Here robustness means stability at the Hardy-type Gaussian threshold and Fourier invariance at the level of the entire Hermite coefficient sequence (Neyt et al., 2024).

For generalized Hermite processes in the $L^2(\mathbb{R})$6-th Wiener chaos, robustness appears as almost sure uniform convergence of wavelet-type random series. The approximation process $L^2(\mathbb{R})$7 satisfies

$L^2(\mathbb{R})$8

with $L^2(\mathbb{R})$9. This solves the long-open extension of the Meyer–Sellan–Taqqu and Pipiras constructions from fractional Brownian motion and the Rosenblatt process to arbitrary Hermite order $n$0 (Ayache et al., 2023).

4. Stochastic inference and uncertainty quantification

In generalized polynomial chaos, robustness is often synonymous with sparsity preservation and correct treatment of dependence. For i.i.d. Gaussian inputs $n$1, a Hermite gPC approximation

$n$2

can be made substantially sparser by an orthogonal rotation $n$3. The rotation is chosen from the eigendecomposition

$n$4

so that dominant variability is aligned with low-index coordinates. Because Gaussianity is rotationally invariant, the Hermite basis in $n$5 has the same orthonormal structure, while the coefficient vector may become much sparser. The method integrates naturally with basis pursuit denoising, weighted $n$6, and OMP, and the reported experiments show effective performance up to $n$7, typically with $n$8–$n$9 rotation steps (Yang et al., 2015).

When Gaussian inputs are dependent rather than independent, robustness requires leaving the tensor-product setting altogether. The generalized Wiener–Hermite expansion under $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$0 remains complete, but coefficient recovery is governed by a coupled linear system $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$1 on each total-degree block because the Gram matrix is only block-diagonal under weak orthogonality. The variance formula correspondingly contains additional same-degree cross terms that vanish only in the independent case. This is a structural rather than numerical notion of robustness: the expansion converges to the correct $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$2-limit without imposing false independence assumptions (Rahman, 2017).

In online inference, the Gauss–Hermite expansion supports sequential density, CDF, and quantile estimation. With coefficients updated by

$\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$3

the static estimator has $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$4 cost per observation and $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$5 memory. The Exponentially Weighted Gauss–Hermite update

$\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$6

adds adaptivity to nonstationary streams. The paper derives analytic CDF formulas, consistency results, and quantile error bounds tied directly to the MISE of the density estimator, and reports competitive performance against EWSA, including stable tail behavior at $\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$7 on heavy-tailed FX return data (Stephanou et al., 2015).

A related ergodic construction estimates homogenized invariant densities from multiscale Langevin trajectories by truncating a Hermite expansion

$\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$8

The resulting estimator is robust in the precise homogenization sense that, with

$\widehat{h_n}(\omega)=(-i)^n h_n(\omega).$9

one has

$X\sim \mathcal{N}(\mu,\Sigma)$0

The decomposition into deterministic truncation/homogenization bias and statistical ergodic error makes the scale coupling explicit (Borodavka et al., 29 Oct 2025).

5. Kinetic theory, transport, and moment closures

In lattice Boltzmann theory for quantum gases, robustness means recovering the correct macroscopic mass, momentum, and energy equations together with the correct transport coefficients for Bose–Einstein and Fermi–Dirac equilibria. The key result is that the equilibrium distribution must be expanded to fourth order in Hermite polynomials. Truncation at third order leaves mass and momentum correct, but misses the isotropic $X\sim \mathcal{N}(\mu,\Sigma)$1 contribution in the energy moment $X\sim \mathcal{N}(\mu,\Sigma)$2, so the energy equation and thermal conductivity are wrong. Exact quadrature for the required moments demands eighth-order isotropy; in two dimensions, d2q37 provides it, whereas d2q17 does not. With the fourth-order expansion, the quantum BGK transport coefficients match the Uehling–Uhlenbeck result of Yang et al., including $X\sim \mathcal{N}(\mu,\Sigma)$3 and the $X\sim \mathcal{N}(\mu,\Sigma)$4 correction to the heat flux (Coelho et al., 2013).

A related fourth-order construction yields compressible lattice Boltzmann models with arbitrary specific heat ratio. The equilibrium is factorized into translational and rotational parts,

$X\sim \mathcal{N}(\mu,\Sigma)$5

and each part is expanded by Hermite polynomials under Gauss–Hermite quadrature. The translational discretization uses D2Q37, the rotational one D1Q7, both with the same scaling $X\sim \mathcal{N}(\mu,\Sigma)$6. The Chapman–Enskog analysis recovers the compressible Navier–Stokes equations with

$X\sim \mathcal{N}(\mu,\Sigma)$7

For the shock-tube test at $X\sim \mathcal{N}(\mu,\Sigma)$8, the reported relative $X\sim \mathcal{N}(\mu,\Sigma)$9 errors are $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$0 for $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$1, $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$2 for $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$3, $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$4 for $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$5, and $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$6 for $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$7 (Hu et al., 2016).

For the 1D–1V Vlasov–Poisson system, robustness is obtained by asymmetrically weighted Hermite functions with a dynamic shift $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$8 and scale $$G(t,x)=\exp\!\big(t^\top(x-\mu)-\tfrac12 t^\top \Sigma t\big) =\sum_{\alpha\in\mathbb{N}_0^d} H_\alpha(x;\Sigma)\,\frac{t^\alpha}{\alpha!}.$$9,

$\Sigma$0

together with an adaptive update driven by the numerical bulk velocity and thermal width. The remap between successive bases is a lower-triangular matrix $\Sigma$1 whose entries are available in closed form and which preserves total mass, momentum, and energy exactly. The fully discrete Fourier–Hermite scheme with implicit midpoint time stepping retains discrete conservation under periodic boundary conditions, and the adaptive method is reported to outperform the non-adaptive one in both accuracy and stability for manufactured solutions and the two-stream instability (Pagliantini et al., 2022).

Moment closures for the Boltzmann equation motivate a more structural use of generalized Hermite expansion. Replacing the isotropic Grad basis by an anisotropic basis built from the full temperature tensor $\Sigma$2 yields a modified 13-moment system in which equilibrium lies in the interior of the hyperbolicity region. For arbitrary order $\Sigma$3, the unregularized generalized Grad closure is still not globally hyperbolic, but an admissible regularization of only the $\Sigma$4 equations restores hyperbolicity on the full admissible set

$\Sigma$5

The directional eigenvalues take the explicit form

$\Sigma$6

where $\Sigma$7 is a root of the Hermite polynomial $\Sigma$8. This is one of the strongest meanings of robustness in the Hermite literature: local well-posedness is enforced at the level of the closure itself (Fan et al., 2014).

6. Applied and interdisciplinary extensions

Several applied papers use robust Hermite expansions precisely because they remain workable when more classical transforms or parametric models fail. For Lévy walks, the Hermite polynomial expansion is introduced as a complement to Fourier and Laplace methods, especially in time–space coupled and nonlinear cases. Expanding the turnover density $\Sigma$9 and propagator $L^2(\mu_\Sigma)$0 in the weighted Hermite basis produces explicit recursions for Laplace-domain coefficients, from which one can extract moments, first-passage statistics, and asymptotic regimes for models with $L^2(\mu_\Sigma)$1, $L^2(\mu_\Sigma)$2, $L^2(\mu_\Sigma)$3, or even $L^2(\mu_\Sigma)$4, where standard inverse-transform methods become intractable (Xu et al., 2019).

In Gaussian-kernel interpolation, the flat limit $L^2(\mu_\Sigma)$5 is numerically unstable for direct RBF solvers. A Hermite generating-function expansion isolates the ill-conditioning into diagonal scalings and yields a stabilized basis

$L^2(\mu_\Sigma)$6

together with an analytic truncation rule derived from a multi-index tail estimate and a multidimensional Mehler formula. This construction applies to both isotropic and fully anisotropic Gaussians and provides an automatic choice of truncation degree and series parameter $L^2(\mu_\Sigma)$7 (Kormann et al., 2019).

For copulas, a finite Hermite expansion of a joint density is not by itself admissible because it can become negative. The corrected-Hermite approach projects the truncated density factor onto the convex set defined by non-negativity, normalization, and selected moment constraints, using Dykstra’s algorithm. The corrected density then induces a valid copula after recomputing marginals. In the cross-FX application, a five-parameter 2D Hermite copula was used to fit EUR–JPY volatility smiles from EUR–USD and USD–JPY smiles, and daily updating of only the correlation parameter $L^2(\mu_\Sigma)$8 after monthly calibration of higher-order coefficients was reported as the most stable one-parameter update (Shiraya et al., 2023).

In low-resolution MIMO detection, second-order Hermite expansion yields the linear approximation

$L^2(\mu_\Sigma)$9

where $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$00 is the first-order Hermite kernel and $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$01 is the second-order kernel. This SOHE model leads to an enhanced LMMSE equalizer with symbol-level normalization,

$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$02

which was reported to deliver the most visible gains at $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$03–$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$04 bits and moderate array sizes, while the exact 1-bit arcsine-based model remains optimal in the strictly antisymmetric sign case (Liu et al., 2021).

In cosmology, the Wiener–Hermite expansion of dark-matter density and velocity fields is proven equivalent to the $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$05-expansion of renormalized perturbation theory. Matching the low-$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$06 one-loop SPT limit and the high-$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$07 exponential behavior produces an approximate nonlinear matter power spectrum with accuracy better than $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$08 or $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$09 on BAO scales $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$10 for $$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$11–$$h_n(x)=\alpha_n\,H_n(x)e^{-x^2/2},\qquad \alpha_n=\frac{1}{\pi^{1/4}\sqrt{2^n n!}}.$$12, while involving only single and double integrals (Sugiyama et al., 2012).

Taken together, these developments show that robust Hermite expansion is best understood as a methodological class rather than a single algorithm. Its unifying principle is that Hermite structure is retained only after the basis has been adapted to the governing geometry—Gaussian measure, anisotropic covariance, local support, quadrature grid, conservation law, or admissible density set—so that truncation and discretization preserve the analytical invariants that matter in the target problem.