Prior Art and Related Work
Purpose: an honest record of where the EBF idea sits in the existing literature, so that claims made about this library are accurate and so that future work can build on what has already been done elsewhere.
Summary in one line: the core idea — replacing the scalar Euclidean radius of an RBF with a per-node positive-definite matrix — is not novel and has carried the name "elliptical basis function" since at least 2000. What is distinctive here is the specific combination of training scheme, regularization, and basis-function generality, plus the absence of any comparable open-source Python implementation.
Scope caveat: this survey is the result of a targeted literature search, not an exhaustive prior-art review. It is adequate for setting honest expectations in documentation. It is not adequate as a patentability opinion — that would require systematic Google Scholar sweeps and a patent-database search.
1. The core concept is established prior art
1.1 Elliptical basis function networks (the name is taken)
Mak & Kung introduced full per-node covariance matrices into RBF networks and called the result elliptical basis function networks:
M. W. Mak and S. Y. Kung, "Estimation of elliptical basis function parameters by the EM algorithm with application to speaker verification," IEEE Transactions on Neural Networks, vol. 11, no. 4, pp. 961–969, 2000. https://ieeexplore.ieee.org/document/857775/
A companion comparative study exists as well ("Elliptical basis function networks and radial basis function networks for speaker verification: a comparative study"). This line of work was pursued through the speaker-verification and speech-processing literature for years.
Relationship to this library: the model equation is the same generalization. The differences are in how the ellipsoids are obtained. Mak & Kung fit them with EM (the network is effectively a Gaussian mixture model with a discriminative output layer); this library learns them by end-to-end gradient descent on the regression loss.
1.2 Anisotropic RBF interpolation (numerical analysis)
G. Casciola, D. Lazzaro, L. B. Montefusco, S. Morigi, "Shape preserving surface reconstruction using locally anisotropic radial basis function interpolants," Computers & Mathematics with Applications, vol. 51, no. 8, pp. 1185–1198, 2006. https://www.sciencedirect.com/science/article/pii/S0898122106000721
Replaces the Euclidean norm with a local metric per center to preserve edges, flat regions, and corners when reconstructing surfaces from unorganized point sets.
Relationship to this library: their metric is derived analytically from the local distribution of points (local PCA / density), not learned by optimization. This is the closest prior work in spirit — anisotropy for scattered-data interpolation quality — but the mechanism for choosing the anisotropy is entirely different.
1.3 Ellipsoidal RBF neural networks (jointly optimized)
Y. Hu et al., "Molecular Sparse Representation by 3D Ellipsoid Radial Basis Function Neural Networks via L1 Regularization," J. Chem. Inf. Model., 2020. https://arxiv.org/abs/2005.05307
"Sparse Ellipsoidal Radial Basis Function Network for Point Cloud Surface Representation," 2025. https://arxiv.org/html/2505.02350v1
Both jointly optimize weights, centers, shapes, and orientations of ellipsoidal RBFs — the same set of free parameters this library trains.
Relationship to this library: closest prior work on the training axis. Differences: they optimize with quasi-Newton methods (sOWL-QN) rather than Adam/autodiff; they target implicit surface / density-field fitting rather than general scattered-data regression; and they use L1 sparsity as the structural regularizer where this library uses a node-spread penalty (see §2.1).
1.4 Mahalanobis-distance RBF networks
"Using a Mahalanobis-Like Distance to Train Radial Basis Neural Networks," IWANN 2005. https://link.springer.com/chapter/10.1007/11494669_32
Evolves a generalized Mahalanobis-structured metric for RBF neurons using a genetic algorithm. A broader scattered literature exists on learning RBF metrics, some of it using exactly the Cholesky parameterization described in §1.6.
1.5 Compact elliptical basis functions
R. Southern, "Compact elliptical basis functions for surface reconstruction," Technical Report TR-NCCA-2011-01, Bournemouth University, 2011. https://eprints.bournemouth.ac.uk/17797/
Surface reconstruction from point clouds using compactly-supported EBFs, built top-down. Confirms the term "elliptical basis function" was in use in graphics as well as in speech.
1.6 The LLᵀ parameterization is a standard device
Guaranteeing positive-definiteness by optimizing an unconstrained triangular factor rather than the matrix itself is a long-standing, widely-used trick (Cholesky parameterization). It appears throughout covariance estimation, Gaussian process hyperparameter learning, and variational inference. ADR-001 should be read as a sound engineering choice, not an invention.
1.7 The strongest modern parallel: 3D Gaussian Splatting
3D Gaussian Splatting (2023 onward) is mathematically the closest living relative of this algorithm, despite coming from an unrelated field:
- thousands of primitives, each with its own learned anisotropic covariance
- covariance factored (rotation × diagonal scale) specifically to guarantee positive-semi-definiteness under gradient descent
- all parameters trained end-to-end by gradient descent on a reconstruction loss
See also DARB-Splatting, which explicitly generalizes splatting to decaying anisotropic radial basis functions: https://arxiv.org/pdf/2501.12369
Relationship to this library: independent convergence on the same core construction. Useful as validation that the approach is sound, and as a source of ideas — the splatting literature has done substantial work on initialization, densification/pruning of primitives, and fast evaluation that may transfer to node placement here.
2. What appears to be distinctive here
No direct precedent was found for the following. These are the defensible claims.
2.1 Metric-aware node-spread regularization (ADR-002)
The var_weight · 1/var(dist_nodes) term, where pairwise node distances are measured in the
learned non-Euclidean metric rather than in Euclidean space.
This has no analog in the surveyed work because the problem it solves does not arise there:
| Prior work | Why it doesn't need this |
|---|---|
| Mak & Kung (EM) | Centers come from clustering; collapse is structurally impossible |
| Casciola et al. | Centers are the data points; not free parameters |
| ERBFNN (2020/2025) | Uses L1 sparsity to control structure instead |
| Gaussian Splatting | Uses explicit densify/prune heuristics instead |
The dual role documented in ALGORITHM.md — collapse prevention and the primary smoothness
control, via the optimizer's compensating reduction of the distance scale — also appears to be
an original observation.
2.2 Arbitrary basis functions over the Mahalanobis radius
The EBF literature is effectively Gaussian-only, because it descends from GMM/EM formulations
where the per-node matrix has a probabilistic interpretation as a covariance. This library
drops that interpretation and treats Aᵢ as a pure metric, which frees it to apply any basis
function — including conditionally positive-definite growing families such as
multiquadric (the default) and thin_plate — on top of a learned per-node ellipsoid.
The anisotropic-interpolation literature (§1.2) does use non-Gaussian bases, but with analytically-derived rather than learned metrics.
2.3 End-to-end gradient training for general engineering regression
Joint autodiff training of centers, ellipsoids, amplitudes, and a global linear trend, aimed at scattered regression on noisy physical measurements — combined with robust losses (Huber/Tukey, ADR-013/ADR-014), adaptive thresholds, and early stopping. The individual ingredients are all known; the assembly targeted at engineering surface fitting is not something the surveyed work does.
2.4 The tooling gap is real
No mainstream Python package offers learned anisotropic RBF interpolation:
| Package | Anisotropy support |
|---|---|
scipy.interpolate.RBFInterpolator |
None — isotropic only |
scipy.interpolate.Rbf (legacy) |
None — isotropic only |
treverhines/RBF |
None — isotropic only |
The published academic implementations of §1.3 are research code tied to their specific
domains. A fit / predict library for engineers, with per-node learned ellipsoids, robust
losses, save/load, and visualization, does not otherwise exist as far as this search found.
3. How to describe this project honestly
Do not claim: a novel method, a new mathematical concept, or the invention of elliptical basis functions.
Do claim: a novel combination — metric-aware node-spread regularization, arbitrary basis function families over learned per-node ellipsoids, and robust end-to-end gradient training — delivered as (as far as is known) the first practical open-source Python implementation of learned anisotropic RBF interpolation for engineering data.
Historical note, for the record: the core algorithm here was developed independently around 2019–2020 without knowledge of the literature above. Independent rediscovery of an idea that several separate fields converged on is a point in the idea's favour.
4. Open leads for a deeper search
If a rigorous prior-art review is ever needed:
- Google Scholar: "metric learning radial basis function", "anisotropic kernel interpolation", "adaptive Mahalanobis kernel regression"
- Forward-citation search on Mak & Kung (2000) and Casciola et al. (2006)
- USPTO / EPO patent search — several near-miss patents surfaced on RBF-with-covariance for signal separation and anomaly detection
- The kriging / geostatistics literature on anisotropic variograms, which solves a closely related problem with different vocabulary and was not surveyed here