Decaying Anisotropic Radial Basis Functions

• DARBFs are decaying anisotropic radial basis functions that employ metrics like the Mahalanobis distance with parameters ξ and β to control directional decay.
• They provide a flexible framework for interpolation and reconstruction, yielding improved convergence and reduced memory usage in applications like 3D splatting and PDE collocation.
• DARBF methods modify traditional RBFs using tailored decay profiles (e.g., Gaussian, cosine) to adapt performance for complex geometries in ice-sheet modeling and tissue scaffold design.

Decaying anisotropic radial basis functions (DARBFs) are radial basis functions whose value is a non-negative decaying profile of an anisotropic distance, most explicitly the Mahalanobis distance in DARB-Splatting, where each primitive is centered at a mean $\boldsymbol{\mu}\in\mathbb{R}^3$ and shaped by a covariance $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$ (Arunan et al., 21 Jan 2025). In the broader anisotropic-RBF literature, closely related constructions appear through scaled norms, full positive-definite matrices, geometry-defined segment distances, and coordinatewise directional scales (Cheng et al., 2017, Yurova et al., 2017, Liu et al., 2016, Xia et al., 7 Apr 2026). The term “DARBF” is therefore narrower than the general category of anisotropic RBFs: it denotes anisotropic kernels with explicit falloff, whereas several antecedent methods introduce anisotropy without adopting the DARBF name, and some anisotropic kernels remain globally supported or non-decaying in the compact-support sense (Arunan et al., 21 Jan 2025, Liu et al., 2016).

1. Definition and scope

The standard isotropic RBF interpolant reviewed in the scaffold-design literature has the form

$f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$

with interpolation matrix entries $\phi_{ji}=\phi(\|x_j-x_i\|)$ and evaluation

$s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$

In that setting, anisotropy is introduced by replacing the Euclidean distance with an application-specific distance $\|\cdot\|_L$, yielding

$s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$

and the linear system is updated accordingly (Liu et al., 2016).

DARB-Splatting makes the definition substantially more explicit by identifying DARBFs as “non-negative functions of the Mahalanobis distance.” Its basic anisotropic distance is

$$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$

and a DARBF is a decaying profile $w(\mathbf{x})=\phi(d_M;\xi,\beta)$, with $\xi>0$ controlling spread and $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$0 controlling main-lobe roll-off (Arunan et al., 21 Jan 2025). The Gaussian then becomes a special case of a larger class rather than the defining kernel.

A persistent point of terminology is that not every anisotropic RBF is a DARBF. The tissue-engineering ARBF paper does not introduce or name a “Decaying Anisotropic Radial Basis Function” as a separate formal method; it presents an anisotropic radial basis function interpolation method in which anisotropy is induced by redefining distance, and any decay comes from the chosen scalar kernel after that substitution (Liu et al., 2016). Likewise, the adaptive spatiotemporal paper uses anisotropic Gaussian RBFs with direction-dependent scales and adaptive centers, but it does not present an explicit DARBF framework (Xia et al., 7 Apr 2026).

2. Mechanisms for introducing anisotropy

A central theme across the literature is that anisotropy is typically introduced through the distance argument rather than through an entirely new approximation architecture. In the continental ice-sheet work, the isotropic norm is replaced by a scaled norm

$\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$1

which in the 2D flowline setting becomes

$\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$2

This is equivalent to using a scaling matrix $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$3 or metric tensor $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$4, even though the paper writes the construction directly in norm form (Cheng et al., 2017).

The Gaussian-stabilization paper treats anisotropy in a more general matrix form through kernels of the form

$\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$5

with $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$6 positive definite. This accommodates full anisotropy, including rotated anisotropy, rather than only axis-aligned scaling. The paper emphasizes that tensor-product HermiteGF stabilization corresponds to diagonal cases, whereas the Hagedorn generating-function expansion addresses the full $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$7-based setting (Yurova et al., 2017).

The scaffold-design paper introduces anisotropy by a different route: it measures distances from points to line segments and between line segments derived from mesh primitives. For a point $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$8 and a line segment $\boldsymbol{\Sigma}\in\mathbb{R}^{3\times 3}$9, the distance is defined casewise, using zero on the segment, perpendicular distance in the acute-triangle case, and $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$0 in the obtuse-triangle case. For two segments $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$1 and $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$2, the paper uses

$f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$3

The paper states that this changes the effective support domain from circular or spherical to elliptical or hyper-elliptical and thereby controls pore-opening direction (Liu et al., 2016).

The adaptive spatiotemporal work implements anisotropy through coordinatewise scale vectors: $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$4 where $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$5 are time-dependent directional scales and $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$6 denotes the Hadamard product. A faithful equivalent matrix interpretation is diagonal scaling, but the paper itself formulates anisotropy through heterogeneous per-coordinate widths rather than a full metric tensor (Xia et al., 7 Apr 2026).

3. Decay profiles and kernel families

Decay behavior is not uniform across anisotropic RBF constructions. Some kernels decay exponentially, some algebraically, some are compact only after main-lobe restriction, and some do not decay to zero at all. The literature explicitly distinguishes these cases, which is why DARBFs should not be conflated with anisotropic RBFs in general (Arunan et al., 21 Jan 2025, Cheng et al., 2017, Liu et al., 2016).

Kernel family	Decay/support behavior	Representative context
Modified Gaussian	Exponential decay; Gaussian is a special case of the DARBF family	DARB-Splatting (Arunan et al., 21 Jan 2025)
Modified half cosine	Compact main lobe after support restriction	DARB-Splatting (Arunan et al., 21 Jan 2025)
Modified raised cosine	Compact when the main lobe is isolated	DARB-Splatting (Arunan et al., 21 Jan 2025)
Modified sinc (modulus)	Oscillatory with decaying envelope; truncated in practice	DARB-Splatting (Arunan et al., 21 Jan 2025)
IMQ / IQ	Algebraic decay; globally supported	Ice-sheet and scaffold contexts (Cheng et al., 2017, Liu et al., 2016)
MQ	Globally supported; does not decay to zero	Ice-sheet and scaffold contexts (Cheng et al., 2017, Liu et al., 2016)
TPS	Globally supported and non-decaying in the compact-support sense	Scaffold design (Liu et al., 2016)
Wendland weights	Compactly supported anisotropic patch weights, not the main trial basis	RBF-PUM for ice sheets (Cheng et al., 2017)

In DARB-Splatting, the kernel family is parameterized directly in terms of $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$7, $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$8, and $f(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|),$9. The paper lists the following forms: $\phi_{ji}=\phi(\|x_j-x_i\|)$0

$\phi_{ji}=\phi(\|x_j-x_i\|)$1

with domain restrictions used where necessary to isolate the main lobe and preserve bounded localized support (Arunan et al., 21 Jan 2025).

The ice-sheet paper employs smooth globally supported kernels

$\phi_{ji}=\phi(\|x_j-x_i\|)$2

and notes that the anisotropic Gaussian is a clear example of a decaying anisotropic RBF in the exponential sense, while anisotropic IMQ and IQ provide slower algebraic attenuation (Cheng et al., 2017). The scaffold paper makes the same structural point in application-specific terms: the decay or falloff behavior comes entirely from the chosen scalar $\phi_{ji}=\phi(\|x_j-x_i\|)$3 after substituting the anisotropic distance $\phi_{ji}=\phi(\|x_j-x_i\|)$4, not from a separately derived directional decay law (Liu et al., 2016).

4. Numerical formulations and computational treatment

DARBF-related methods appear in several numerical regimes: interpolation, meshfree PDE collocation, partition-of-unity localization, stable flat-kernel evaluation, and adaptive spatiotemporal parameter evolution.

In porous scaffold design, the pipeline begins from a volumetric mesh, assigns scalar values $\phi_{ji}=\phi(\|x_j-x_i\|)$5 to mesh vertices and $\phi_{ji}=\phi(\|x_j-x_i\|)$6 to edge, face, tile, or sub-volume centers, solves the ARBF interpolation system, evaluates the interpolant on voxels, and extracts an iso-surface. The algorithm is summarized as loadMesh();, ARBFInterpolation();, marchCube();, exportResult();, with ARBFInterpolation() consisting of assignMeshValues();, solveCoefficients();, and applyCoefficientsToInterpolation();. The paper repeatedly emphasizes that the iso-value is a major design control parameter and that pore sizes increase as the iso-values increases (Liu et al., 2016).

For large-aspect-ratio PDEs, the continental ice-sheet work uses Kansa-type collocation and an anisotropic RBF partition of unity method: $\phi_{ji}=\phi(\|x_j-x_i\|)$7 Patch functions satisfy Shepard normalization,

$\phi_{ji}=\phi(\|x_j-x_i\|)$8

with compactly supported Wendland weights scaled anisotropically on each patch. The paper uses $\phi_{ji}=\phi(\|x_j-x_i\|)$9 overlap, defines

$s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$0

and reports that $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$1 works well for ISMIP-HOM B while $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$2 works well for the continental ice cap. It also recommends tuning $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$3 so that the interpolation matrix condition number stays around $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$4, estimated with MATLAB’s condest, and advocates a nodal-value formulation through $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$5 rather than solving directly for the coefficients in flat or ill-conditioned regimes (Cheng et al., 2017).

The Gaussian-stabilization paper addresses the classical flat-limit ill-conditioning of Gaussian RBFs by changing basis. Its preferred practical formula for $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$6 is

$s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$7

which bypasses direct solution with the original Gaussian collocation matrix. The paper recommends evaluating the HermiteGF basis through Hermite functions and states that $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$8 works well in most cases. For anisotropic Gaussians, it provides a Hagedorn generating-function expansion for kernels of the form $s(x)=\sum_{i=1}^N w_i \phi(\|x-x_i\|).$9, thereby giving a stabilization template for full-matrix anisotropic Gaussian kernels (Yurova et al., 2017).

The adaptive spatiotemporal paper replaces static coefficients and shape parameters by dynamical variables: $\|\cdot\|_L$0 It differentiates the adaptive RBF ansatz in time, introduces a model-induced operator $\|\cdot\|_L$1, and trains a neural ODE to minimize

$\|\cdot\|_L$2

Implementation is organized around initialization of the RBF approximation at $\|\cdot\|_L$3, ODE integration with odeint in torchdiffeq, random sampling of interior and boundary points, and gradient-based training in PyTorch (Xia et al., 7 Apr 2026).

5. Applications and empirical behavior

The clearest explicit DARBF application is splatting-based 3D reconstruction. DARB-Splatting preserves the 3D Gaussian Splatting parameterization, replaces the Gaussian footprint by a DARBF footprint, and approximates the 3D-to-2D covariance conversion with a kernel-specific scalar correction factor $\|\cdot\|_L$4. For half-cosine-squared, the paper reports $\|\cdot\|_L$5; for support matching it derives $\|\cdot\|_L$6 for the half-cosine case and reports $\|\cdot\|_L$7 for a best-performing raised cosine configuration. The abstract states “up to 34% faster convergence during training and a 45% reduction in memory consumption across various DARB reconstruction kernels, while maintaining comparable PSNR, SSIM, and LPIPS results.” In the reported 30K-iteration averages on Mip-NeRF360, 3DGS-updated achieved SSIM 0.813, PSNR 27.45, LPIPS 0.218, training time 19m06s, and 633 MB; raised cosine achieved SSIM 0.813, PSNR 27.45, LPIPS 0.214, 19m36s, and 645 MB; half cosine achieved SSIM 0.790, PSNR 27.04, LPIPS 0.247, 16m15s, and 524 MB. The paper also reports a strong correction-factor ablation: on Mip-NeRF360 average, raised cosine with $\|\cdot\|_L$8 gave PSNR 27.45, SSIM 0.81, LPIPS 0.21, whereas without $\|\cdot\|_L$9 it gave PSNR 26.69, SSIM 0.78, LPIPS 0.25 (Arunan et al., 21 Jan 2025).

In ice-sheet simulation, anisotropy is used to resolve extreme aspect ratios that defeat isotropic RBFs. The continental ice-cap experiment spans $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$0 km horizontally and up to $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$1 km vertically, with aspect ratio around $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$2. The paper states that isotropic RBFs failed completely: in several tests with different node layouts and shape parameters, “in none of the tests we succeeded to obtained a stable solution.” By contrast, anisotropic RBF-PUM produced qualitatively correct horizontal and vertical velocities on Cartesian and Halton nodes. For the glacier-scale ISMIP-HOM B benchmark, the finest comparison reported error $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$3, $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$4, time $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$5 s for FEM; error $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$6, $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$7, time $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$8 s for global RBF; and error $s'(x)=\sum_{i=1}^N w'_i \phi(\|x-x_i\|_L),$9, $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$0, time $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$1 s for RBF-PUM. For the continental problem, convergence rates were approximately $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$2 on Cartesian nodes and $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$3 on Halton nodes (Cheng et al., 2017).

In tissue engineering, anisotropy is used to control internal pore orientation rather than merely local resolution. The paper states that “openings in desired scaffold architecture should start from sub-volume center and grows towards pores on structure surface,” and its figures show that in a single tetrahedron the scaffold opening grows from the tetrahedron center toward the four triangle centers. The same method produces scaffolds from a single hexahedron, eight hexahedra forming a cube, and four hexahedra forming a rod, while disturbed meshes generate disturbed porous structures responsive to local geometric perturbation. The paper also reports that isotropic 2D RBF interpolation exhibits visible “circle-shape artifacts,” whereas ARBF removes these artifacts by using line-segment-aware distances (Liu et al., 2016).

The adaptive spatiotemporal Gaussian framework supplies a further DARBF-adjacent application class in multidimensional PDEs and integrodifferential equations. In Example 1, adaptive RBFs significantly outperform non-adaptive RBFs with fixed centers and scales; as the dimension increases from $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$4 to $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$5, the error rises moderately but remains below $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$6 for $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$7. In Example 2, non-adaptive RBF error stagnates once $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$8, whereas adaptive basis movement and reshaping continue to improve accuracy. In Example 3, the method uses only $$d_M=\left[(\mathbf{x-\mu})^T\mathbf{\Sigma}^{-1}(\mathbf{x-\mu})\right]^{1/2},$$9 basis functions to capture the expected rotational swarming pattern in a 4D Vlasov-type integrodifferential equation on $w(\mathbf{x})=\phi(d_M;\xi,\beta)$0, while imposing the lower bound $w(\mathbf{x})=\phi(d_M;\xi,\beta)$1 componentwise to preserve vanishing behavior at infinity (Xia et al., 7 Apr 2026).

6. Limitations, misconceptions, and open directions

A recurrent misconception is that DARBFs are simply any anisotropic radial basis functions. The cited literature does not support that equivalence. The scaffold-design paper is explicit that it presents an ARBF method with geometry-defined anisotropic distance, not a dedicated DARBF framework; it does not define an anisotropic matrix norm, a metric tensor, a compact-support construction, or an independently derived anisotropic decay law (Liu et al., 2016). The adaptive spatiotemporal paper likewise uses anisotropic Gaussian widths but does not introduce explicit DARBF terminology, full anisotropy matrices, rotated anisotropy, compactly supported anisotropic kernels, or conditioning theory for anisotropic shape parameters (Xia et al., 7 Apr 2026).

A second misconception is that anisotropy implies decay. The literature repeatedly contradicts that. In the scaffold and ice-sheet papers, MQ does not decay to zero, TPS is non-decaying in the compact-support sense, and the main globally supported kernels remain globally supported after anisotropic scaling unless additional localization is introduced (Liu et al., 2016, Cheng et al., 2017). Even in the ice-sheet RBF-PUM framework, compact support enters through Wendland partition weights rather than through compactly supported anisotropic trial kernels (Cheng et al., 2017).

In DARB-Splatting, the principal mathematical limitation is projection. Unlike Gaussian splats, generalized DARBFs do not inherit exact closed-form 3D-to-2D integration. The method works by approximating the Gaussian shortcut through a kernel-specific empirical correction factor $w(\mathbf{x})=\phi(d_M;\xi,\beta)$2, so the image-space covariance conversion is approximate rather than exact. The paper also notes that oscillatory kernels require main-lobe restriction to avoid ring artifacts, that each kernel needs its own $w(\mathbf{x})=\phi(d_M;\xi,\beta)$3, support restriction, and $w(\mathbf{x})=\phi(d_M;\xi,\beta)$4 scaling, and that the broader characterization of which DARBFs are best for 3D reconstruction remains incomplete. A cautious interpretation is that the work is more a proof of feasibility and usefulness than a final theory of optimal non-Gaussian splatting kernels. The quantitative memory headline also requires care: the abstract states a 45% reduction in memory consumption across various DARB reconstruction kernels, whereas the detailed synthesis notes that the paper text most consistently claims up to 15% memory reduction in the abstract and teaser, while supplementary tables show larger per-scene reductions for some kernels (Arunan et al., 21 Jan 2025).

Conditioning and parameter selection remain major unresolved issues for decaying anisotropic kernels. The ice-sheet paper states that no analytical procedure is available for choosing the shape parameter and uses conditioning-based heuristics instead; the Gaussian-stabilization paper is motivated precisely by the severe ill-conditioning of flat Gaussian interpolation and develops a stabilized Hermite/Hagedorn basis to address it (Cheng et al., 2017, Yurova et al., 2017). This suggests that any mature DARBF theory must treat anisotropy, decay, localization, and numerical stability as coupled design variables rather than independent choices.

A plausible implication of the current literature is that DARBFs are best understood not as a single fixed kernel family but as a design principle: combine a direction-sensitive metric or geometry-aware distance with a decaying profile whose support, smoothness, and numerical treatment are tailored to the application. The existing record already spans Mahalanobis splatting kernels, aspect-ratio-aware PDE collocation, geometry-defined scaffold interpolation, stabilized anisotropic Gaussians, and adaptive direction-scaled Gaussian bases. What remains incomplete is a unified theory covering full anisotropy matrices, compactly supported anisotropic trial kernels, directional decay laws beyond Gaussian or inverse-multiquadric forms, and systematic criteria for kernel selection across these regimes (Arunan et al., 21 Jan 2025, Cheng et al., 2017, Xia et al., 7 Apr 2026).