Hermite Spectral Collocation Method

• Hermite spectral collocation is a global pseudospectral method that uses Hermite polynomials and functions to approximate solutions on unbounded domains with Gaussian decay.
• It employs zeros of Hermite polynomials as nodes, enabling efficient differentiation matrix construction and achieving exponential convergence with superconvergence properties.
• The method is applied to Schrödinger-type, Fokker–Planck, and fractional PDEs, with extensions to high-dimensional and semi-infinite problems via mapping and sparse grid strategies.

Hermite spectral collocation method denotes a class of global or mapped pseudospectral discretizations in which the approximation is built from Hermite polynomials or Hermite functions and the governing equation is enforced at Hermite nodes, typically the zeros of a Hermite polynomial. Its canonical setting is an unbounded domain, most often $\mathbb{R}$, where the Gaussian structure of the Hermite basis encodes decay at infinity and makes the method especially suitable for Schrödinger-type operators, Fokker–Planck-type equations, and other problems with rapidly decaying solutions (Wang et al., 21 Jul 2025).

1. Basis functions, approximation spaces, and node sets

The algebraic foundation is the family of Hermite polynomials

$H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$

which satisfy the weighted orthogonality

$$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$

A closely related orthonormal basis on $L^2(\mathbb{R})$ is furnished by Hermite functions, obtained by absorbing the Gaussian weight into the basis itself (Wang et al., 21 Jul 2025).

For collocation and interpolation on $\mathbb{R}$, one works in finite-dimensional Hermite spaces such as

$$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$

or, in alternate notation,

$$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$

The nodes are chosen as the real zeros of $H_{n+1}$, equivalently of $\psi_{n+1}$, and inherit the symmetry $x_j=-x_{n-j}$ (Wang et al., 21 Jul 2025). In the collocation interpretation used for kinetic problems, the velocity variable is represented by Lagrange cardinal functions built on the zeros of $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$0, so that differentiation and quadrature are both tied to Gauss–Hermite structure (Fatone et al., 2018).

Two Hermite-type bases are prominent in unbounded-domain collocation for nonlocal operators: the normalized Hermite functions $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$1 and the over-scaled functions $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$2. Both are paired with Gauss–Hermite collocation points, but they lead to markedly different conditioning properties (Tang et al., 2018).

2. Collocation formulation and differentiation structure

In its simplest global form, Hermite spectral interpolation seeks $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$3 such that

$H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$4

at the zeros of $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$5. The same nodal set underlies spectral collocation for differential equations: one approximates the solution by a Hermite expansion and enforces the differential operator pointwise at the Hermite nodes (Wang et al., 21 Jul 2025).

For second-order model problems on $H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$6, two standard collocation formulations are

$H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$7

and

$H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$8

Using cardinal functions

$H_n(x)=(-1)^n e^{x^2}\frac{d^n}{dx^n}e^{-x^2},$9

the collocation ansatz yields matrix systems of the form

$$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$0

or

$$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$1

where $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$2 is the Hermite second-derivative matrix and $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$3 is the diagonal contribution from the potential term (Wang et al., 21 Jul 2025).

For fractional PDEs on $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$4, the principal technical task is the construction of differentiation matrices for $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$5. In that setting, the method consists of expanding the solution with proper global basis functions and imposing collocation conditions on the Gauss-Hermite points; the corresponding differentiation matrices are computed recursively, using explicit formulas involving confluent hypergeometric functions and scaling relations for the fractional Laplacian (Tang et al., 2018).

A distinct but related collocation realization appears in semi-Lagrangian Vlasov–Poisson discretization. There, the velocity dependence is represented by Lagrange cardinal functions at Gauss–Hermite nodes, derivatives in $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$6 are produced by Hermite differentiation matrices, and velocity integrals are approximated by Gauss–Hermite quadrature (Fatone et al., 2018). This retains the collocational character but applies it to phase-space transport rather than directly to a spatial boundary-value problem.

3. Convergence, error bounds, and superconvergence

For analytic data on $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$7, Hermite spectral interpolation exhibits exponential convergence. Under analyticity in the strip

$$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$8

Wang and Zhang derive a global max-norm error estimate with decay factor $$\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}\,dx=\gamma_n\delta_{n,m},\qquad \gamma_n=2^n n!\sqrt{\pi}.$$9; derivative errors carry polynomial prefactors, roughly $L^2(\mathbb{R})$0 for the $L^2(\mathbb{R})$1-th derivative (Wang et al., 21 Jul 2025).

The central refinement is superconvergence. For Hermite interpolation in $L^2(\mathbb{R})$2, the first derivative superconverges at the zeros $L^2(\mathbb{R})$3 of $L^2(\mathbb{R})$4, while the second derivative superconverges at the zeros $L^2(\mathbb{R})$5 of $L^2(\mathbb{R})$6. Globally, the first-derivative error at $L^2(\mathbb{R})$7 improves from $L^2(\mathbb{R})$8 to at least $L^2(\mathbb{R})$9, and in the central region $\mathbb{R}$0 to $\mathbb{R}$1. For the second derivative, the error at $\mathbb{R}$2 improves from $\mathbb{R}$3 to $\mathbb{R}$4 (Wang et al., 21 Jul 2025).

The same pattern carries into collocation for differential equations. If the exact solution belongs to $\mathbb{R}$5, then the collocation error has the exact form

$\mathbb{R}$6

so the function error vanishes at the collocation nodes, the derivative error superconverges at the zeros of $\mathbb{R}$7, and the second-derivative error superconverges at the zeros of $\mathbb{R}$8 (Wang et al., 21 Jul 2025). An important contrast with Chebyshev interpolation is that, for Hermite interpolation, the number of superconvergence points increases with derivative order (Wang et al., 21 Jul 2025).

For unbounded-domain fractional PDEs, the dominant practical issue is not only approximation error but conditioning. Numerical evidence shows that the differentiation matrices associated with the normalized Hermite basis have algebraically growing condition numbers, whereas the over-scaled basis leads to much worse conditioning; scaling factors strongly influence accuracy and convergence speed (Tang et al., 2018).

4. Unbounded and semi-infinite domains

The native domain of the Hermite basis is $\mathbb{R}$9, and several formulations work directly there. In Wang and Zhang’s analysis, no mapping from a finite interval is used; the Gaussian factor in the Hermite functions provides the decay needed to approximate solutions vanishing at infinity (Wang et al., 21 Jul 2025). This direct formulation is particularly natural when the operator itself contains $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$0, as in harmonic-oscillator or Schrödinger-type problems (Wang et al., 21 Jul 2025).

For semi-infinite domains $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$1, two mapping strategies recur. In the Thomas–Fermi setting, Hermite functions on $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$2 are composed with

$$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$3

so that transformed basis functions $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$4 can be collocated at mapped Hermite nodes $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$5 (Bayatbabolghani et al., 2016). In Lane–Emden and Volterra-type problems, a logarithmic–hyperbolic map

$$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$6

transfers $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$7 to $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$8, after which transformed Hermite functions are used in a collocation framework with boundary behavior built into the ansatz (Parand et al., 2010, Parand et al., 2010).

These mapped formulations address two recurrent obstacles of semi-infinite problems: singular behavior at the origin and boundary conditions at infinity. In the Lane–Emden formulation, the approximate solution is modified to incorporate $$\mathbb{H}_n=\operatorname{span}\{\psi_0,\psi_1,\dots,\psi_n\},$$9 and $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$0 directly through

$$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$1

so that the transformed Hermite part does not have to enforce derivative data at the singular point itself (Parand et al., 2010). In the Thomas–Fermi formulation, a lifting function

$$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$2

is combined with the transformed Hermite expansion, and the residual is set to zero at transformed Hermite collocation nodes, producing a nonlinear algebraic system solved by Newton’s method (Bayatbabolghani et al., 2016).

5. High-dimensional, fractional, and sparse formulations

In high dimensions, the principal obstacle is the tensor-product growth of the Hermite basis. For generalized Hermite functions

$$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$3

full grids lead to $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$4 degrees of freedom, whereas regular hyperbolic cross and optimized hyperbolic cross index sets reduce the approximation space dramatically (Luo et al., 2013). The regular hyperbolic cross has cardinality $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$5, while for optimized hyperbolic cross with $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$6 the cardinality is $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$7 (Luo et al., 2013).

Although the high-dimensional parabolic theory is presented in Galerkin form, the same generalized Hermite basis, hyperbolic cross index sets, and sparse Smolyak Gauss–Hermite grids directly inform collocation design. The paper’s approximation estimates show that, in weighted Korobov spaces, optimized hyperbolic cross Hermite approximations can retain dimension-independent convergence exponents, with dimensional dependence entering only through constants (Luo et al., 2013). This suggests that sparse-grid Hermite collocation is the natural high-dimensional analogue of the one-dimensional Gauss–Hermite collocation method.

For space-fractional PDEs on $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$8, Hermite collocation replaces local derivative matrices by differentiation matrices for $$\mathcal{H}_N=\operatorname{span}\{\widetilde H_0,\widetilde H_1,\dots,\widetilde H_N\}.$$9. Two Hermite-type bases are employed, the collocation points are Gauss–Hermite points, and the matrices are computed recursively; multi-term fractional PDEs are handled by summing the matrices for the individual fractional orders (Tang et al., 2018).

A further extension appears in kinetic equations, where Hermite collocation is applied in velocity space rather than in physical space. In the semi-Lagrangian Vlasov–Poisson method, the unbounded velocity domain is treated with Gauss–Hermite nodes and Hermite differentiation matrices, and the numerical behavior depends strongly on a scaling parameter in the Gaussian weight, indicating that adaptive scaling is a central unresolved issue for Hermite-based solvers on unbounded velocity domains (Fatone et al., 2018).

6. Applications, terminology, and limitations

Hermite spectral collocation is used for unbounded-domain ODEs and PDEs, including Schrödinger-type operators, fractional PDEs, Lane–Emden equations, Thomas–Fermi equations, Volterra population models, and kinetic equations in velocity space (Wang et al., 21 Jul 2025, Tang et al., 2018, Parand et al., 2010, Bayatbabolghani et al., 2016, Parand et al., 2010, Fatone et al., 2018). Across these settings, three implementation themes recur: collocation at Hermite or transformed Hermite nodes, basis scaling or mapping to match decay, and the reduction of the continuous problem to linear or nonlinear algebraic systems.

A common misconception is that every “Hermite collocation” method is a global spectral method on $H_{n+1}$0. In the Maxwell literature, however, “Hermite methods” can denote local tensor-product Hermite–Birkhoff schemes on Cartesian grids. In that formulation there is no global Hermite basis, no Hermite transform, and no Hermite quadrature; instead, one stores derivative data at grid vertices, reconstructs a local polynomial on each cell, and advances the solution explicitly with an energy-conserving update (Appelo et al., 2024). The relation to Hermite spectral collocation is structural rather than literal: both use Hermite interpolation data, but one is global and modal or nodal on Gauss–Hermite points, whereas the other is local and cell-based (Appelo et al., 2024).

The main limitations are also consistent across the literature. Conditioning can deteriorate rapidly, especially for over-scaled bases in fractional problems (Tang et al., 2018). On semi-infinite domains, performance depends on mapping and scaling parameters, which are usually tuned empirically (Bayatbabolghani et al., 2016, Parand et al., 2010). In high dimensions, full tensor Hermite spaces remain prohibitive without hyperbolic-cross or sparse-grid compression (Luo et al., 2013). For collocation in velocity space, a poor choice of Gaussian scaling can substantially degrade accuracy or even destabilize long-time computations, so adaptive scaling remains an important practical problem (Fatone et al., 2018).

Taken together, these developments define Hermite spectral collocation method as a family of Gaussian-adapted collocation techniques for unbounded or mapped-unbounded geometries, distinguished by Hermite node sets, sparse derivative structure in Hermite coordinates, and, in the most recent work, a detailed superconvergence theory that identifies where collocation and interpolation are most accurate (Wang et al., 21 Jul 2025).