Hermite Spectral Interpolation

• Hermite spectral interpolation is a high-order global method that constructs interpolants using Hermite polynomials or functions to match both function values and derivatives at selected nodes.
• It achieves root-exponential convergence under strip analyticity conditions and leverages stable computational techniques like the Golub–Welsch algorithm and barycentric formulations.
• The approach extends to periodic, manifold-valued, and meshfree settings, providing practical benefits for differentiation, quadrature, and error control.

Hermite spectral interpolation is a family of global high-order interpolation procedures in which the approximant is built from Hermite polynomials, Hermite functions, or Hermite-type derivative constraints. On the real line, it usually denotes interpolation in $\mathbb P_n$ or in the Hermite-function space $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$ at Gauss–Hermite nodes, namely the zeros of a Hermite polynomial. In a broader numerical-analysis literature, the same expression also covers periodic Fourier constructions and other spectral schemes that interpolate both function values and derivative values at a prescribed set of nodes (Wang et al., 2023, Denysiuk, 2021, Hinsberg et al., 2012).

1. Scope and terminology

On $\mathbb R$, Hermite spectral methods approximate a function $f:\mathbb R\to\mathbb C$ by a finite linear combination of Hermite polynomials or Hermite functions. Two standard constructions are distinguished: Hermite spectral projection, in which coefficients are chosen by orthogonality, and Hermite spectral interpolation, in which coefficients are chosen so the approximant matches $f$ at selected nodes, typically Gauss–Hermite nodes (Wang et al., 2023). In this usage, “Hermite” refers to the Hermite basis.

A second, equally established usage appears in periodic and collocation settings. There, Hermite spectral interpolation means that one interpolates both a function and some of its derivatives by means of global spectral bases such as trigonometric polynomials. The trigonometric Hermite spline construction on a uniform periodic grid, for example, matches $f^{(k)}(t_j)$ for $k=0,\dots,p$ and uses Fourier coefficients obtained by FFT, while cubic and higher-order Hermite interpolation in Fourier spectral codes matches values and derivatives at cell endpoints (Denysiuk, 2021, Hinsberg et al., 2012). In this usage, “Hermite” refers primarily to the data being interpolated.

This terminological breadth is important because it explains why the literature ranges from Hermite-function interpolation on unbounded domains to higher-order Hermite–Fejér schemes on Jacobi nodes, periodic trigonometric Hermite splines, and manifold-valued Hermite interpolation. The common structure is the imposition of value and derivative information within a high-order global or semi-global approximation framework.

2. Classical Hermite spectral interpolation on $\mathbb R$

For Hermite polynomials, the weight is $\omega(x)=e^{-x^2}$, and the $n$th Hermite polynomial is

$\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$0

They satisfy

$\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$1

The associated Hermite functions are

$\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$2

and they form an orthonormal basis of $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$3 (Wang et al., 2023).

In the polynomial formulation, one takes the zeros $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$4 of $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$5. The Hermite spectral interpolant $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$6 is the unique polynomial satisfying

$\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$7

These nodes are precisely the Gauss–Hermite nodes for $\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$8-point quadrature. In the Hermite-function formulation, one seeks

$\mathbb H_n=\mathrm{span}\{\psi_0,\dots,\psi_n\}$9

such that

$\mathbb R$0

Equivalently, $\mathbb R$1 is a polynomial of degree $\mathbb R$2 interpolating $\mathbb R$3 at the same nodes (Wang et al., 2023).

The interpolation problem can also be written in matrix form. If $\mathbb R$4 are Gauss–Hermite nodes and

$\mathbb R$5

then the forward Hermite transform maps coefficients $\mathbb R$6 to nodal values $\mathbb R$7, and the backward transform maps nodal data to coefficients through $\mathbb R$8. This transform viewpoint is the computational core of Hermite pseudospectral interpolation and collocation (Webb et al., 2 Apr 2026).

3. Strip analyticity, root-exponential convergence, and scaling

A central convergence regime is analyticity in an infinite strip

$\mathbb R$9

For Hermite polynomials, the standard assumptions are that $f:\mathbb R\to\mathbb C$0 is analytic in $f:\mathbb R\to\mathbb C$1 and satisfies $f:\mathbb R\to\mathbb C$2 as $f:\mathbb R\to\mathbb C$3 in the strip. For Hermite functions, the corresponding condition is $f:\mathbb R\to\mathbb C$4 (Wang et al., 2023).

Under these assumptions, Hermite spectral interpolation converges root-exponentially. For polynomial interpolation at Gauss–Hermite nodes,

$f:\mathbb R\to\mathbb C$5

while for Hermite-function interpolation,

$f:\mathbb R\to\mathbb C$6

The same paper derives contour integral representations for Hermite coefficients and for the interpolation remainder, and identifies these bounds as asymptotically sharp (Wang et al., 2023).

A later scaling theory recasts these results in a bandwidth language. For scaled generalized Hermite approximations with scaling parameter $f:\mathbb R\to\mathbb C$7, the effective spatial and frequency bandwidths are

$f:\mathbb R\to\mathbb C$8

The error is decomposed into a spatial truncation term outside $f:\mathbb R\to\mathbb C$9, a frequency truncation term outside $f$0, and an exponentially small spectral term. In this framework, choosing $f$1 is equivalent to balancing the spatial and frequency truncation errors. The same analysis explains root-exponential, algebraic, and pre-asymptotic behaviors, gives systematic guidance for choosing the optimal scaling factor, and shows that scaling can recover geometric convergence for some Gaussian-type targets and can double the algebraic order for some algebraically decaying functions (Hu et al., 3 Feb 2026).

4. Derivatives, quadrature, and superconvergence

Hermite interpolation is especially important because it interacts well with differentiation and quadrature. For Hermite polynomial projections,

$f$2

and for Hermite-function projections,

$f$3

Gauss–Hermite quadrature satisfies

$f$4

so differentiation and quadrature inherit the same strip-controlled root-exponential structure (Wang et al., 2023).

A more refined phenomenon is superconvergence. For interpolation in $f$5 at zeros of $f$6, the first derivative converges faster at the zeros of $f$7, and the second derivative converges faster at the zeros of $f$8. The global error bounds satisfy

$f$9

while at superconvergence points the first- and second-derivative errors gain an additional factor of order $f^{(k)}(t_j)$0 (Wang et al., 21 Jul 2025).

The same structure extends to Hermite spectral collocation for model differential equations. In the harmonic-oscillator-type and Helmholtz-type examples studied there, the collocation error $f^{(k)}(t_j)$1 is proportional to $f^{(k)}(t_j)$2. Consequently, function values superconverge at zeros of $f^{(k)}(t_j)$3, first derivatives at zeros of $f^{(k)}(t_j)$4, and second derivatives at zeros of $f^{(k)}(t_j)$5 (Wang et al., 21 Jul 2025).

5. Stable computation and barycentric structure

Large Hermite interpolants are numerically delicate because direct evaluation of Hermite recurrences or transform matrices encounters overflow, underflow, and loss of orthogonality. A recent stable transform algorithm factors the Hermite transform matrix as

$f^{(k)}(t_j)$6

where $f^{(k)}(t_j)$7 is orthogonal and $f^{(k)}(t_j)$8 is diagonal with entries

$f^{(k)}(t_j)$9

The factorization is computed from the eigendecomposition of the Hermite Jacobi matrix by the Golub–Welsch algorithm. This isolates ill-conditioning in a diagonal scaling, makes both forward and inverse transforms stable, and avoids the overflow/underflow problems of the weight matrix $k=0,\dots,p$0. Numerical experiments reported there show that the direct method loses accuracy drastically around $k=0,\dots,p$1, whereas the Golub–Welsch-based method remains stable and accurate well beyond $k=0,\dots,p$2 (Webb et al., 2 Apr 2026).

A second line of work concerns barycentric formulations of general Hermite interpolation. For distinct nodes $k=0,\dots,p$3 with multiplicities $k=0,\dots,p$4, the Hermite interpolant is the unique polynomial of degree $k=0,\dots,p$5, $k=0,\dots,p$6, satisfying $k=0,\dots,p$7. The barycentric Hermite formula introduces local weight vectors $k=0,\dots,p$8, and a key practical result is that if an additional derivative is prescribed at one interpolation point, the barycentric coefficients can be updated using only $k=0,\dots,p$9 operations. The same paper reports very good numerical stability even when derivatives of high order are involved (Sadiq et al., 2011).

For higher-order Hermite–Fejér interpolation at Gauss–Jacobi or Jacobi–Gauss–Lobatto pointsystems, stability is achieved through the second barycentric formula. The exponentially increasing common factor in the barycentric weights cancels, so the simplified barycentric weights can be computed efficiently and the implementation cost is linear in the number of grids. This produces fast higher-order Hermite–Fejér interpolation with explicit convergence rates on Jacobi pointsystems (Xiang et al., 2014).

6. Periodic, Jacobi, and geometric extensions

In periodic settings, Hermite spectral interpolation is realized by trigonometric Hermite splines. On the uniform grid

$\mathbb R$0

the goal is to construct a $\mathbb R$1-periodic function $\mathbb R$2 such that

$\mathbb R$3

The construction uses aliasing identities such as $\mathbb R$4, so the Hermite conditions decouple into small linear systems per Fourier mode: $\mathbb R$5 systems for $\mathbb R$6 and $\mathbb R$7 systems for $\mathbb R$8. FFT is used to compute the coefficients of the trigonometric interpolants of $\mathbb R$9 and its derivatives, and the small modewise systems depend only on the grid, so they can be inverted once and reused (Denysiuk, 2021).

Hermite ideas also extend beyond Hermite bases. Higher-order Hermite–Fejér interpolation on Gauss–Jacobi and Jacobi–Gauss–Lobatto nodes uses derivative data up to order $\omega(x)=e^{-x^2}$0 at each node and admits stable second-barycentric implementations with $\omega(x)=e^{-x^2}$1 preprocessing and $\omega(x)=e^{-x^2}$2 pointwise evaluation. For analytic functions in a Bernstein ellipse, the classical Hermite–Fejér interpolant on Gauss–Jacobi nodes enjoys geometric convergence with explicit bounds (Xiang et al., 2014).

At the geometric end of the spectrum, Hermite interpolation has been generalized to manifold-valued functions $\omega(x)=e^{-x^2}$3. One construction defines the interpolant by weighted Riemannian barycenters and converts derivative interpolation into linear constraints on the derivatives of scalar weight functions; this approach is intrinsic and requires no vector transport. An alternative performs classical Hermite interpolation in a single tangent space after transporting sample values and derivatives there; this is straightforward but depends on the selected base point (Zimmermann et al., 2022).

Related local and meshfree variants reinforce the same pattern. Modified Hermite radial basis interpolation incorporates both function and gradient information and combines Gaussian kernels with polynomial factors to improve conditioning and accuracy, while low-degree Hermite osculatory spline spaces on refined partitions admit normalized B-spline-like bases and quasi-interpolation operators (Fashamiha et al., 21 Feb 2025, Boushabi et al., 2024). Together these developments show that Hermite spectral interpolation is not a single algorithm but a broad approximation paradigm in which high-order global or semi-global representations are constrained by value and derivative data.