Generalized Gaussian RBF Moments

• The paper introduces a Hermite-based stabilization scheme that transforms Gaussian kernels into a well-conditioned modal representation via generalized moment coefficients.
• It develops radial gausslet constructions enforcing near-origin moment conditions with localized analytic functions, enhancing operator diagonal approximations.
• The framework unifies HermiteGF expansions with tensor-product anisotropic generalizations and moment corrections to improve numerical accuracy in Gaussian-based approximations.

Generalized Gaussian radial basis moments denote moment-like quantities attached to Gaussian kernels or Gaussian-derived radial bases when those objects are recast into forms that emphasize stability, localization, or operator compression. In one formulation, a Gaussian radial basis function $\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big)$ is expanded in a Hermite basis, and the center-dependent coefficients of that expansion can be interpreted as generalized moments of the kernel (Yurova et al., 2017). In another, radial gausslets are constructed so that low-order centered moments behave like a quadrature rule on the half-line, with explicit near-origin moment repair by adding localized analytic functions (White, 23 Mar 2026). The resulting concept is not a single canonical invariant; rather, it is a family of moment constructions that organize Gaussian radial bases into numerically stable or operator-diagonal forms.

1. Conceptual scope and principal meanings

In the supplied literature, the phrase is associated with two distinct technical settings. The first is Gaussian radial basis function stabilization: Gaussian kernels are represented in a Hermite or Hagedorn basis, and the corresponding expansion coefficients serve as generalized moments that encode how each kernel contributes to each stable mode. The second is radial gausslet construction: localized orthonormal basis functions on $r\ge 0$ are engineered to satisfy centered-moment conditions closely enough that coordinate and Coulomb operators admit diagonal or integral-diagonal approximations (Yurova et al., 2017).

A concise comparison is useful.

Setting	Moment object	Functional role
HermiteGF Gaussian RBFs	Hermite/Hagedorn expansion coefficients	Stable representation of Gaussian kernels
Radial gausslets	Centered-moment conditions and defect $D$	Diagonal or integral-diagonal operator approximation
Gaussian mixture moment varieties	Truncated moments and cumulants	Algebraic parameterization of distributions

The unifying structure is that “moments” are not merely raw statistical moments of a Gaussian density. In the HermiteGF setting they are basis coefficients of a kernel expansion; in radial gausslets they are reproduction properties of localized functions; in algebraic statistics they are projective coordinates of a distribution family. This suggests that the term is best understood operationally: the moments are whatever coefficients or constraints make Gaussian-based representations tractable in the target formalism.

2. Hermite-generated moments for Gaussian radial basis functions

For one-dimensional Gaussian RBFs, the starting point is

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$

The stabilization framework in "Stable evaluation of Gaussian radial basis functions using Hermite polynomials" uses the physicists’ Hermite generating function

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$

with the substitutions

$$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$

This yields an exact Hermite-based representation of the Gaussian kernel, and truncation at $n=M-1$ gives the practical approximation (Yurova et al., 2017).

Within that construction, the HermiteGF coefficients are interpreted as generalized Gaussian radial basis moments. The supplied description characterizes them in three steps: the Gaussian centered at $y$ is represented in a Hermite basis; each coefficient is a weighted algebraic moment in the center variable $y$; and the family of coefficients across $n$ plays the role of the “moments” of the Gaussian with respect to the Hermite basis. In matrix form, these quantities are exactly the entries of the basis-transform matrix $r\ge 0$0.

The significance of this interpretation is structural. Instead of working directly with nearly linearly dependent Gaussian kernels, the method works with a stable Hermite basis plus a coefficient matrix that carries the center dependence. The moments are therefore not auxiliary diagnostics; they are the mechanism by which the original Gaussian RBF basis is converted into a stable modal representation.

3. Flat-limit stabilization and the basis-transform viewpoint

The motivation for generalized Gaussian radial basis moments in the HermiteGF setting is the flat-limit difficulty. As $r\ge 0$1, Gaussian RBFs become very wide and increasingly linearly dependent on bounded domains, so the collocation or interpolation matrix becomes severely ill-conditioned even though the interpolant itself can become highly accurate. The supplied material identifies this as the RBF paradox: the best approximation regime is often the worst numerically (Yurova et al., 2017).

The HermiteGF method rewrites evaluation schematically as

$r\ge 0$2

Here the ill-conditioning is moved into $r\ge 0$3, which depends only on centers and parameters, not on the evaluation point $r\ge 0$4. The evaluation basis $r\ge 0$5 can be computed stably, especially via Hermite-function recurrences. The supplied account gives four reasons this helps: the dangerous flat-limit dependence is extracted into a structured factor; the Hermite basis is orthogonal-like and well-behaved on a suitably centered interval; diagonal factors decay rapidly like $r\ge 0$6, so high modes become negligible; and the remaining matrix products can be organized to avoid catastrophic cancellation.

The parameter $r\ge 0$7 controls the spatial scale on which the Hermite polynomials are evaluated. The supplied guidance is explicit: too small $r\ge 0$8 makes basis functions too similar and worsens conditioning; too large $r\ge 0$9 makes Hermite polynomials grow large and risks overflow or cancellation; empirically, $D$0 is often good on $D$1 (Yurova et al., 2017).

This framework also clarifies the relation to earlier stabilization schemes. Compared with the Chebyshev-QR method of Fornberg, Larsson, and Flyer, the HermiteGF construction uses a direct Hermite generating-function expansion, does not require the complicated parameterization of the Chebyshev-QR approach, and extends naturally to tensor-product multivariate grids. Compared with Fasshauer and McCourt’s Gauss-QR, it is similar in spirit but uses $D$2 as a geometrically transparent scaling parameter rather than the extra tuning parameter $D$3; the supplied description states that there is no direct parameter correspondence matching both the exponential width and Hermite scaling simultaneously.

4. Tensor-product and anisotropic generalizations

For separable multivariate Gaussians,

$D$4

the HermiteGF expansion is applied independently in each dimension. On a tensor grid, the interpolant takes a Kronecker-product form,

$D$5

and after eliminating the center-dependent factors against the collocation inverse,

$D$6

The supplied text emphasizes that this is computationally attractive because it avoids forming the full $D$7-sized matrix: only one-dimensional matrices are stored and inverted, giving a memory-sparse method (Yurova et al., 2017).

For anisotropic Gaussians,

$D$8

the generalization uses the Hagedorn generating function framework. The multivariate expansion is written as

$D$9

with

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$0

In this anisotropic setting, the generalized moments are the multivariate coefficients $$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$1 multiplying tensor Hermite polynomials. The supplied account notes that this gives a stable basis representation for anisotropic kernels, though it is computationally more expensive than the tensor-product isotropic method.

A plausible implication is that the moment viewpoint scales particularly well when the Gaussian structure is separable or nearly separable, since then the generalized moments inherit Kronecker or tensor-product organization.

5. Radial gausslets and moment correction on the half-line

A second major realization of generalized Gaussian radial basis moments appears in radial gausslets. Standard gausslets are one-dimensional orthonormal localized functions

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$2

with orthonormality

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$3

Their defining moment property is

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$4

so that for smooth $$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$5,

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$6

This implies a diagonal coordinate operator,

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$7

and accurate diagonal approximations such as $$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$8 or the integral-diagonal form

$$\phi_y(x)=\exp\!\big(-\varepsilon^2(x-y)^2\big), \qquad \varepsilon>0.$$9

For two-electron interactions, the same structure yields

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$0

collapsing the tensor from four indices to two (White, 23 Mar 2026).

The radial case differs because the basis lives on the half-line and must respect the three-dimensional metric. The construction therefore uses the reduced radial function

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$1

with normalization

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$2

and enforces the physical boundary condition $e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$3. Simply truncating full-line gausslets by $e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$4 loses orthogonality and completeness near the origin, so boundary gausslets are built by restoring a small number of “tail” functions from the negative side and then diagonalizing

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$5

For the strictly radial basis, the authors then remove the one degree of freedom that carries the value at $e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$6, so the remaining basis functions satisfy $e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$7.

This removal breaks the exact COMX theorem locally. The supplied text diagnoses the resulting moment defect by comparing the X-eigenvalue center $e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$8 with

$e^{2st-t^2}=\sum_{n\ge 0}\frac{t^n}{n!}h_n(s),$9

and defines

$$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$0

The odd-even construction improves this defect: odd combinations

$$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$1

automatically satisfy $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$2, while even combinations

$$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$3

are used in limited number to restore even-polynomial completeness. Even so, the supplied result is $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$4.

The decisive repair is to add narrow boundary-compatible functions

$$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$5

called $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$6-Gaussians, and optimize the widths $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$7 using Nelder–Mead to reduce $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$8. Adding one or two such functions dramatically improves moment matching; with two optimized $$t=\frac{\varepsilon^2 y}{\gamma},\qquad s=\gamma x.$$9-Gaussians, the reported value is

$n=M-1$0

compared with

$n=M-1$1

for the unaugmented radial basis (White, 23 Mar 2026). The supplied text identifies this as the main “generalized Gaussian radial basis moment” idea in practice: near-origin moment-fidelity is not achieved exactly by the boundary projection, but it can be made extremely accurate by explicit moment fitting with a few local analytic basis additions.

The same paper combines this with the coordinate map

$n=M-1$2

and distorted physical-space basis functions

$n=M-1$3

which preserve orthonormality and concentrate resolution near the nucleus. Reported numerical behavior includes microhartree accuracy for helium with fewer than 20 radial functions, about $n=M-1$4 accuracy with around 30 functions, agreement to roughly 10 digits with MRChem for unrestricted Hartree–Fock energies of Li through Ne, and an extrapolated residual error of about $n=M-1$5 for correlated helium relative to the complete-basis-set limit.

6. Related moment formalisms and terminological boundaries

The literature also contains nearby but distinct uses of Gaussian moments. In "Moment Varieties of Gaussian Mixtures," the central objects are truncated moments of Gaussian distributions and mixtures, organized as projective coordinates in a moment variety $n=M-1$6, with mixtures represented by secant varieties $n=M-1$7 (Améndola et al., 2015). In that setting, moments and cumulants are coordinate systems for algebraic statistics, not radial-basis-function coefficients. The paper explicitly states that it does not develop a radial-basis-function theory in the machine-learning sense; the closest “basis” viewpoint is the use of moments and cumulants as coordinate systems, with cumulants providing a Cremona-type linearization through

$n=M-1$8

A separate terminological overlap appears in Dynamic Boltzmann Machines for time-series prediction. There, the phrase “generalized Gaussian distribution” refers to a distribution with an additional shape parameter important to approximate heavy-tailed distributions, and the paper’s abstract motivates extensions of Gaussian DyBM beyond fixed variance (Raymond et al., 2017). That usage concerns temporal first-order and second-order moments and generalized Gaussian distributions in financial data, not radial-basis moments.

A common misconception is therefore to treat all “Gaussian moments” as interchangeable. The supplied record does not support that. In the HermiteGF RBF setting, generalized moments are expansion coefficients of Gaussian kernels in a stable basis; in radial gausslets, they are centered-moment conditions that enable diagonal approximations; in algebraic statistics, they are truncated distributional moments and cumulants; and in DyBM, they are moments of a predictive distribution whose variance or shape may vary. The common thread is Gaussian structure, but the mathematical role of “moment” is model-dependent.

These distinctions matter because they delimit what is being stabilized, approximated, or inferred. Generalized Gaussian radial basis moments are most naturally understood as the moment structures that arise when Gaussian radial bases are forced to satisfy two demanding requirements simultaneously: numerical stability in flat or anisotropic kernel regimes, and high-fidelity low-order moment behavior in radial operator calculations.