Anisotropic Basis Functions for Geometric Representation

20 concepts10 papers1 proposal

When we try to represent a physical signal—whether it is a 2D image, a 3D surface, or a neural radiance field—we rely on mathematical building blocks called basis functions. Historically, standard tools like wavelets or Radial Basis Functions (RBFs) have been isotropic, meaning they are radially symmetric. They treat every direction equally, like a circular bubble.

While isotropic functions are excellent for representing smooth, diffuse signals, they fail fundamentally when encountering edges, contours, and sharp geometric transitions. To capture a sharp linear boundary using circular bubbles, you must pack together an infinite number of increasingly tiny circles. This causes two major problems: it destroys sparsity (requiring too many coefficients), and it introduces blurring or ringing artifacts across the boundary.

The core insight of this research direction is to replace isotropic bubbles with anisotropic basis functions. By allowing these building blocks to stretch, rotate, bend, and scale directionally, they can align precisely with the underlying geometry of the signal.


The Core Intuition: Stretching and Shrinking

To understand why anisotropy works, consider a sharp edge in an image or a sharp corner on a 3D mesh. An isotropic function smooths over this feature because its influence extends equally in all directions.

An anisotropic basis function solves this by decoupling its scale across different axes. As explored by Dinh et al. in the context of 3D surface reconstruction, we can use ellipsoidal basis functions that adapt to local geometry. By stretching the function along the direction of an edge, we maintain smooth continuity along the feature. Simultaneously, by shrinking its support across (perpendicular to) the edge, we prevent its mathematical influence from bleeding over the boundary. This directional control over smoothing allows the representation to maintain a sharp break in continuity rather than a rounded, blurred transition.

Isotropic (Circular):        Anisotropic (Ellipsoidal):
     _ .

- . _                    _ .
- - - . _
   /           \                 (     Edge    )  <-- Stretched along the edge
  |    Edge     |                 ` - - - - - `   <-- Shrunk across the edge
   \           /
     ` - . _ . `
(Bleeds over the boundary)     (Preserves sharp transition)

In modern machine learning, this same intuition is being integrated into implicit neural representations. Xu et al. introduced this capability to neural fields by utilizing RBFs with learnable covariance matrices. Instead of relying on rigid coordinate grids, their method allows Gaussian kernels to dynamically flatten into thin sheets or needles, efficiently capturing fine, high-frequency structures that fixed-grid methods struggle to represent.


Adapting to Curvature: Bending and Geometric Flows

Real-world edges are rarely perfectly straight lines; they curve, twist, and form complex contours. To represent these efficiently, anisotropic basis functions must go beyond simple rotation and elongation—they must adapt to the flow of the data.

One way to achieve this is through explicit geometric warping. Dossal et al. and Duval et al. demonstrate that warping transform bases along a local "geometric flow" optimizes representation sparsity. By aligning the dictionary atoms with the specific orientation and curvature of image contours, the representation minimizes the number of significant coefficients required to encode complex structures.

Alternatively, this geometric adaptation can be built directly into the dictionary generation rules. In work on very low bit-rate image coding, Peotta et al. and Peotta et al. designed redundant dictionaries where the generating functions undergo translation, rotation, anisotropic scaling, and a specialized bending operation. This bending allows the dictionary elements to curve along with the image contours, enabling an extremely sparse and visually accurate representation of edges even under severe compression constraints.


Fixed vs. Dynamic Anisotropy

The literature splits into two main paradigms for implementing anisotropy: fixed mathematical frames and dynamically optimized functions.

1. Fixed Geometric Frames (Curvelets and Contourlets)

Instead of adapting to a specific image on the fly, one can construct a massive, multi-scale dictionary of predefined anisotropic shapes. A foundational breakthrough in this domain was the introduction of curvelets by Candès et al.. Curvelets are highly directional, needle-like elements that obey an anisotropic scaling law: $\text{width} \propto \text{length}^2$. The authors proved that this specific parabolic scaling achieves near-optimal minimax rates for recovering objects with sharp edges from noisy, incomplete data—fundamentally outperforming classical isotropic wavelets.

Because these fixed frames excel at capturing smooth curves and directional edges, they are highly effective for analyzing physical structures with complex geometries, such as the human face. For instance, Mian et al. utilized the Discrete Contourlet Transform, and Zhou et al. utilized the Curvelet Transform to extract facial features. Both groups demonstrated that because facial geometry (like the jawline or eyes) is inherently anisotropic, directional geometric transforms provide a much more compact and descriptive representation of identity than standard isotropic methods.

2. Dynamically Optimized Frameworks

Rather than selecting from a predefined dictionary, continuous optimization frameworks evolve the parameters of the basis functions to fit the data. Slabaugh et al. formulated a variational framework for image segmentation where every parameter of an RBF—including its position, weight, and anisotropic covariance—is governed by differential equations. By allowing the basis functions to dynamically stretch, rotate, and migrate toward boundaries, the parametric contour can adapt its eccentricity to match sharp corners and complex boundaries without smoothing them over.


Why It Works: The Sparsity Argument

Ultimately, the success of anisotropic basis functions comes down to sparsity. In signal processing, a representation is sparse if it can approximate a signal using only a few non-zero coefficients.

When an isotropic transform encounters a curve, it requires a cascade of coefficients across multiple scales to correct for the geometric mismatch. An anisotropic transform, by contrast, can capture a segment of that same curve with a single, well-aligned, elongated atom. By matching the intrinsic dimensionality and directionality of edges, anisotropic basis functions compress the energy of geometric features into a minimal set of coefficients, unlocking near-optimal performance in denoising, compression, and spatial reconstruction.

Key Foundational Work

Reconstructing Surfaces Using Anisotropic Basis Functions (2001) — contributed 3 concept elements to this direction.
Recovering edges in ill-posed inverse problems: optimality of curvelet frames (2002) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_anisotropic_basis_functions_53e4fdbb,
  title = {anisotropic basis functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anisotropic-basis-functions/rdr_53e4fdbba7a44040a62319c1d88701ff},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_anisotropic_basis_functions_53e4fdbb_v1,
  title = {anisotropic basis functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/anisotropic-basis-functions/rdr_53e4fdbba7a44040a62319c1d88701ff/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake