Adaptive Basis Learning (ABLE)

• Adaptive Basis Learning (ABLE) is a dynamic representation method that learns its basis functions from data rather than relying on fixed bases.
• It is applied in diverse fields including operator learning, PDE solving, image deblurring, and reinforcement learning to capture localized and multiscale features.
• ABLE methods enhance parameter efficiency and computational performance by adapting basis geometry, scaling, and active subspace selection to the problem at hand.

Adaptive Basis Learning (ABLE) denotes a family of representation-learning methods in which the basis used to express a target function, operator, kernel field, or value function is not fixed a priori, but is instead learned, adapted, sparsified, or re-parameterized from data. In the PDE literature, the acronym “ABLE” is used explicitly for a learned adaptive spectral representation that replaces fixed Fourier layers in neural operators (Zhao et al., 11 May 2026). Closely related ideas appear across a broader body of work: localized finite-element-style bases for operator learning (Zhang et al., 30 Oct 2025), image-adaptive blur-kernel decompositions (Carbajal et al., 2021), task-supervised basis layers for functional data (Yao et al., 2021), environment-dependent atomic orbital rescaling (Khan et al., 2024), adaptive bases in actor-critic reinforcement learning (Castro et al., 2010), sparse adaptive reduced dictionaries for multiscale flow (Wang et al., 2022), adaptive basis-inspired neural blocks for PDE solvers (Li et al., 2024), and adaptive radial-basis edge functions in Kolmogorov–Arnold networks (Chiu et al., 12 Jan 2026). Across these works, the central shift is from learning only coefficients in a prescribed representation to learning the representation itself.

1. Conceptual scope and historical development

ABLE is best understood as a research direction rather than a single algorithm. The common template is an expansion of the form “target object = coefficients $\times$ basis,” with adaptation applied to the basis functions, their geometry, their active subset, or the mapping from inputs to basis parameters. In this sense, the explicit PDE framework called ABLE (Zhao et al., 11 May 2026) is part of a broader lineage that includes online adaptive basis functions for value approximation in reinforcement learning (Castro et al., 2010), POD enhancement through learned nonlinear snapshot correction (Lyu et al., 2019), supervised functional basis learning (Yao et al., 2021), and fixed-dictionary adaptive subspace selection in reduced-order multiscale models (Wang et al., 2022).

Taken together, these works suggest several distinct meanings of “adaptive.” In some settings, the basis is shared across a problem family after training but its geometry is learned globally, as in localized finite-element hats with trainable centers and support widths for operator learning (Zhang et al., 30 Oct 2025). In others, the basis is recomputed for each instance, as in image-adaptive global blur kernels combined with per-pixel mixing coefficients (Carbajal et al., 2021). A third pattern keeps the dictionary fixed but learns a sparse active subspace, as in AMS-Net’s thresholding, pruning, and greedy basis reactivation over precomputed multiscale or POD bases (Wang et al., 2022). A fourth pattern does not add new basis families at all, but learns continuous modifications of standard ones, such as shell-wise Gaussian exponent rescaling in atomic basis sets (Khan et al., 2024).

This diversity is important because “learning the basis” does not necessarily mean unconstrained dictionary discovery. Many ABLE-style methods restrict adaptation to a structured family: continuous piecewise linear hats (Zhang et al., 30 Oct 2025), convex combinations of normalized blur kernels (Carbajal et al., 2021), micro-network probes over $[0,1]$ (Yao et al., 2021), radial basis functions with trainable centers and widths (Chiu et al., 12 Jan 2026), or Fourier modes modulated by a learned ancillary density to form a Parseval frame (Zhao et al., 11 May 2026). The resulting literature is therefore unified more by where adaptivity enters the representation than by any single architecture.

2. Formal structure of learned basis representations

A canonical ABLE formulation appears in operator learning. FERN starts from the standard neural-operator view

$G: U \rightarrow V,$

with output represented as

$G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$

so that learning decomposes into learning coefficient functionals $c_k$ and basis functions $\phi_k$ (Zhang et al., 30 Oct 2025). This coefficient–basis factorization is the clearest abstract template for ABLE.

Other domains instantiate the same pattern with different objects. In non-uniform blur estimation, each per-pixel kernel is expressed as

$\mathbf{k}_i = \sum_{b=1}^B m_i^b \mathbf{k}^b,$

where the basis kernels $\mathbf{k}^b$ are global within an image and the coefficients $m_i^b$ vary per pixel, with both constrained to be nonnegative and normalized (Carbajal et al., 2021). In functional data analysis, a learned Basis Layer computes coefficients

$$c_i=\langle \beta_i, X\rangle = \int \beta_i(t)\,X(t)\,dt,$$

with each basis function parameterized by a micro neural network $[0,1]$0 and trained jointly with the predictor (Yao et al., 2021). In reinforcement learning, the critic is written as

$[0,1]$1

where $[0,1]$2 are linear value-function coefficients and $[0,1]$3 are basis-shaping parameters updated online on a slower timescale (Castro et al., 2010).

The explicit ABLE spectral framework replaces a fixed Fourier representation by a learned frame. With a learned ancillary density $[0,1]$4 satisfying

$[0,1]$5

it defines adaptive spectral atoms

$[0,1]$6

together with an analysis operator $[0,1]$7 and synthesis operator $[0,1]$8 that preserve invertibility and Parseval norm identities (Zhao et al., 11 May 2026). In Free-RBF-KAN, the learned basis is univariate and edge-local:

$[0,1]$9

with trainable coefficients $G: U \rightarrow V,$0, centers $G: U \rightarrow V,$1, and widths $G: U \rightarrow V,$2 (Chiu et al., 12 Jan 2026). In quantum chemistry, adaptation occurs at the level of Gaussian basis parameters rather than basis coefficients: primitive exponents $G: U \rightarrow V,$3 in a contracted Gaussian are scaled by environment-dependent factors $G: U \rightarrow V,$4, with optimal labels defined by

$G: U \rightarrow V,$5

(Khan et al., 2024).

These formulations differ in detail, but they all move part of the representational burden from fixed coordinates to learned coordinates in function space.

3. Modes of adaptation

Taken together, the literature suggests a useful taxonomy of ABLE mechanisms. One mode is family-level basis geometry learning. FERN learns a shared set of localized finite-element hat functions, parameterized by trainable centers $G: U \rightarrow V,$6 and support widths $G: U \rightarrow V,$7, and then predicts sample-specific coefficients through branch subnetworks (Zhang et al., 30 Oct 2025). The adaptation is global to a PDE family rather than per-sample remeshing.

A second mode is instance-adaptive basis construction. In non-uniform blur estimation, the network predicts a fresh set of image-specific basis kernels for each blurry image, while coefficient maps vary per pixel within that image (Carbajal et al., 2021). This is a two-level adaptive representation: image-adaptive global basis and pixel-adaptive local mixing.

A third mode is task-supervised functional probing. AdaFNN does not seek a reconstruction basis for the input process; rather, it learns a predictive subspace through basis functions $G: U \rightarrow V,$8 optimized directly by the supervised loss, optionally regularized by orthogonality and sparsity penalties (Yao et al., 2021). The paper is explicit that the learned bases need not identify the “true” underlying functions uniquely; they define a predictive representation.

A fourth mode is adaptive active-subspace selection over a fixed dictionary. AMS-Net assumes a large precomputed library and adapts only the active subset by soft-thresholding coefficient outputs, pruning coefficient-linked connections, and greedily reactivating omitted basis functions using residual correlations (Wang et al., 2022). This is ABLE in the sense of adaptive basis usage rather than adaptive basis synthesis.

A fifth mode is architecture growth through localized basis insertion. ABI-DNN uses basis-inspired blocks initialized from local centers and widths, then runs an AFEM-like “solve, estimate, mark, enhancement” loop. High-residual regions are clustered, and new BI-blocks are added around the resulting centroids with support size derived from cluster radius (Li et al., 2024). This is closest to adaptive local basis enrichment in the finite-element sense.

A sixth mode is parameter adaptation within a conventional basis family. Adaptive atomic basis sets keep the basis size fixed and learn only shell-wise radial scaling factors for selected Gaussian orbitals as functions of the local chemical environment (Khan et al., 2024). Free-RBF-KAN likewise stays within the RBF family while learning basis locations and smoothness parameters (Chiu et al., 12 Jan 2026). In both cases, the family is fixed but the geometry is not.

4. Major application families

PDE learning is the most prominent ABLE application area. FERN uses localized trainable FEM hats for operator learning on seven PDE families, including Allen–Cahn, Cahn–Hilliard, Fokker–Planck, aggregation–diffusion, Keller–Segel, KdV, and viscous Burgers, with the stated motivation that many of these solutions exhibit shocks, bumps, sharp gradients, or rapid decay (Zhang et al., 30 Oct 2025). The explicit ABLE spectral framework targets Burgers, Darcy flow, and Navier–Stokes by replacing fixed Fourier layers with learned adaptive spectral branches (Zhao et al., 11 May 2026). ABI-DNN addresses PDEs with localized phenomena such as steep peaks, corner singularities, and Burgers shock formation by inserting basis-inspired blocks where residuals are large (Li et al., 2024). AMS-Net addresses two-phase multiscale flow by predicting sparse coefficient trajectories in GMsFEM or POD spaces rather than full fields (Wang et al., 2022). A hybrid predecessor appears in residual-diffusivity computation, where SRGAN is trained on snapshot pairs at two diffusivities and POD is then applied to the corrected snapshots to obtain a more transferable reduced basis (Lyu et al., 2019).

Inverse problems and imaging provide a distinct ABLE pattern. The adaptive blur-kernel model estimates a dense spatially varying motion-blur field through a small set of image-specific global basis kernels and dense coefficient maps, preserving nonnegativity and unit-mass constraints by convex-combination structure (Carbajal et al., 2021). The same paper emphasizes that the representation cost drops from $G: U \rightarrow V,$9 to $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$0, while still supporting a physically grounded forward model and non-blind deblurring via a non-uniform Richardson–Lucy solver.

Functional data analysis supplies a supervised statistical version of ABLE. AdaFNN learns basis functions over the domain variable $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$1 and projects discretely observed curves onto these learned functions before a conventional downstream MLP (Yao et al., 2021). The paper explicitly contrasts this with Fourier, B-spline, and FPCA expansions chosen without access to the prediction target.

Electronic-structure computation offers a chemistry-specific form. Adaptive atomic basis sets personalize shell-wise Gaussian scaling factors to each atom’s local environment before SCF, using KRR trained on labels from Hartree–Fock energy minimization over QM9 molecules (Khan et al., 2024). The method does not increase the number of basis functions; it changes only their radial extent.

Reinforcement learning provides an early formulation of adaptive bases. In actor-critic methods, the critic basis is allowed to change online via parameters $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$2, and the paper develops updates under three objectives: approximation square error, Bellman residual, and projected Bellman residual (Castro et al., 2010). This predates the acronym ABLE but fits its central idea closely.

5. Theory, constraints, and computational structure

The theoretical picture is heterogeneous. Some papers provide strong structural guarantees. The ABLE spectral framework proves that the adaptive family $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$3 is a Parseval tight frame when the learned density satisfies the normalization condition, so the transform is norm-preserving and invertible on its image (Zhao et al., 11 May 2026). It also proves that FNO is a special case obtained from a trivial ancillary space and constant density, and therefore that ABLE strictly generalizes fixed-basis Fourier neural operators. Free-RBF-KAN proves universality for RBF-KANs under continuous non-polynomial kernels and specializes the result to Gaussian RBFs, so adaptive center/width learning does not weaken approximation power (Chiu et al., 12 Jan 2026). AdaFNN proves a consistency-style approximation theorem for targets of the form $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$4, where $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$5 is a finite collection of linear functionals of the input function, and also states a small expected generalization gap under compactness, Lipschitz, and SGD-style assumptions (Yao et al., 2021). In reinforcement learning, convergence is established with multiple-timescale stochastic approximation: fast critic updates, slower actor updates, and slowest basis updates, with almost sure convergence to local stationary points under the stated assumptions (Castro et al., 2010).

Other works are theoretically motivated but less formally closed. FERN cites an adaptive FEM approximation estimate,

$G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$6

as motivation for localized adaptive approximation, but it does not prove a new convergence theorem for the learned adaptive basis itself (Zhang et al., 30 Oct 2025). Its main technical claim is architectural: the 1D hat function

$G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$7

is represented exactly by ReLUs, so basis construction incurs “no approximation error” (Zhang et al., 30 Oct 2025). ABI-DNN likewise derives basis-inspired blocks from FEM hats and uses residual-driven enhancement, but does not supply a dedicated convergence theorem for the adaptive procedure (Li et al., 2024). The residual-diffusivity work is explicitly empirical: SRGAN improves snapshot geometry and therefore POD basis quality, but no generalization theorem is given (Lyu et al., 2019).

A recurring computational theme is that ABLE methods often preserve the efficiency of a structured baseline while making the representation adaptive. ABLE spectral layers maintain FFT-based complexity $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$8 with small adaptive branch count $G(u)(x) \approx \sum_{k=1}^N c_k(u)\,\phi_k(x),$9 (Zhao et al., 11 May 2026). FERN replaces deep trunk networks with shallow exact ReLU assemblies of local hats, reducing basis-parameter counts to $c_k$0 (Zhang et al., 30 Oct 2025). Adaptive atomic basis sets preserve basis size and therefore essentially preserve SCF cost, with timing differences reported as negligible (Khan et al., 2024). Free-RBF-KAN keeps the forward form of fixed RBF-KAN at inference time, so training-time adaptivity does not increase eventual forward complexity relative to a fixed RBF basis (Chiu et al., 12 Jan 2026).

6. Empirical behavior, misconceptions, and open problems

Across the literature, the clearest empirical advantage of ABLE appears when the target exhibits localized, spatially heterogeneous, or multiscale structure. FERN repeatedly reports that learned local hats are especially well matched to shocks, sharp gradients, bumps, interfaces, and rapidly decaying tails, while remaining competitive on smoother problems (Zhang et al., 30 Oct 2025). The explicit spectral ABLE paper reports its largest gains in regimes with sharp gradients and multiscale behavior, including low-viscosity Burgers and turbulent Navier–Stokes (Zhao et al., 11 May 2026). ABI-DNN is motivated by the same failure mode for standard PINNs and shows that residual-guided localized enhancement reduces error near peaks, singularities, and shock-like regions (Li et al., 2024). In blur estimation, the adaptive basis is effective because many per-pixel kernels are correlated but image-dependent, so a small instance-specific basis captures dense non-uniform blur fields efficiently (Carbajal et al., 2021).

A second empirical pattern is parameter efficiency. FERN reports near-DeepONet accuracy with dramatically fewer basis parameters, for example 80 parameters for 40 bases or 160 for 80 bases, versus tens of thousands in deep trunk networks (Zhang et al., 30 Oct 2025). Free-RBF-KAN is positioned similarly: fixed RBF-KAN improves speed but loses accuracy relative to spline KAN, and adaptive centers plus adaptive smoothness recover much of that gap (Chiu et al., 12 Jan 2026). Adaptive atomic basis sets report improved energetics in up to $c_k$1 of cases on 30,000 QM9 molecules, with effectively unchanged runtime because only exponent values are modified (Khan et al., 2024). AdaFNN emphasizes that it often achieves better predictive performance than raw data, B-splines, or FPCA using only a few learned bases (Yao et al., 2021).

Several recurrent misconceptions are corrected by this literature. First, ABLE does not always mean learning an entirely new unrestricted dictionary. In many methods, the family is fixed and only centers, widths, scales, or supports are learned (Zhang et al., 30 Oct 2025, Chiu et al., 12 Jan 2026, Khan et al., 2024). Second, adaptivity can be dataset-level rather than per-instance: FERN learns a basis shared across a PDE family after training, whereas the blur model predicts a new basis for each image (Zhang et al., 30 Oct 2025, Carbajal et al., 2021). Third, interpretability does not imply identifiability. AdaFNN explicitly notes that the true predictive signal may appear as a linear combination of learned bases rather than one learned basis function matching one latent ground-truth component (Yao et al., 2021). Fourth, adaptive basis selection over a fixed dictionary, as in AMS-Net, is a genuine ABLE strategy even when the basis atoms themselves are not regenerated (Wang et al., 2022).

Open issues remain substantial. FERN is demonstrated only in 1D spatial settings and does not formalize its training loss or constraints on support-width positivity (Zhang et al., 30 Oct 2025). ABI-DNN relies on heuristic residual indicators, clustering, and cluster-radius-to-support rules, and its multidimensional localization remains coordinatewise rather than a genuinely compactly supported multivariate basis (Li et al., 2024). Adaptive atomic basis learning is currently demonstrated only for Hartree–Fock, small-molecule organic chemistry, and restricted radial scaling of selected shells (Khan et al., 2024). Free-RBF-KAN provides universality but not approximation rates or a stability analysis for adaptive center/width learning (Chiu et al., 12 Jan 2026). The spectral ABLE framework is developed on regular grids with FFT structure, and extension to irregular geometries and more expensive cross-branch interactions remains an active question (Zhao et al., 11 May 2026).

In this broader sense, ABLE is less a single method than a representational principle: when the dominant bottleneck is the mismatch between a fixed basis and the geometry of the target family, learning the basis itself can be more effective than increasing model depth, widening coefficient networks, or refining only the output map. The current literature shows that this principle can be instantiated as exact activation-level basis construction, supervised learned probes, adaptive spectral frames, environment-dependent basis parameterization, sparse adaptive active-set selection, and residual-driven local enrichment, depending on the structure of the underlying problem (Zhang et al., 30 Oct 2025, Yao et al., 2021, Zhao et al., 11 May 2026).