🔢 mathematics

Data-dependent approximation through RBF

This paper proposes a data-dependent Radial Basis Function (RBF) interpolation method that adaptively varies the shape parameter based on smoothness indicators to force kernel functions toward Kronecker delta behavior near discontinuities, thereby significantly reducing spurious oscillations while preserving accuracy and matrix conditioning in smooth regions.

Original authors: José Kuruc, David Levin, Pep Mulet, Juan Ruiz-Álvarez, Dionisio F. Yáñez

Published 2026-03-25

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: José Kuruc, David Levin, Pep Mulet, Juan Ruiz-Álvarez, Dionisio F. Yáñez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to draw a perfect map of a landscape based on a few scattered landmarks. This is what mathematicians call interpolation: filling in the blanks between known points to guess what the whole picture looks like.

For decades, scientists have used a powerful tool called Radial Basis Functions (RBFs) to do this. Think of RBFs as "smart rubber bands." If you pull on a rubber band at one point, it stretches out and influences the shape of the whole map. By adjusting how "stiff" or "loose" these rubber bands are (a setting called the shape parameter), you can make the map very smooth and accurate.

The Problem: The "Gibbs" Ghost

However, there's a big problem when your landscape has a sudden, sharp cliff or a jagged edge (a discontinuity).

If you try to stretch a smooth rubber band over a sharp cliff, it doesn't just stop; it gets confused. It tries to smooth out the cliff, but in doing so, it creates a wild, wiggly mess right next to the edge. In math, this is called the Gibbs phenomenon. It's like trying to draw a straight line with a shaky hand; the closer you get to the sharp corner, the more your hand trembles, creating ugly, spiky oscillations that ruin the picture.

The Solution: The "Smart Rubber Band"

The authors of this paper, a team from Spain and Israel, came up with a clever trick to fix this. They decided to make the rubber bands data-dependent.

Here is how their new method works, using a simple analogy:

The Smooth Zones (The Meadow):
In areas where the landscape is flat and smooth (like a meadow), the rubber bands behave normally. They stretch out gently, connecting the dots perfectly to create a smooth, accurate curve.
The Discontinuity (The Cliff):
When the computer detects a sharp cliff (a jump in the data), it doesn't try to stretch a rubber band over it. Instead, it instantly tightens the rubber band until it becomes infinitely stiff.
- The Analogy: Imagine a rubber band that, when it senses a cliff, shrinks down until it's just a tiny, tight knot right on top of the landmark. It stops trying to reach its neighbors.
- The Result: This "knot" acts like a delta function (a mathematical spike). It stops the wiggles from spreading.
The Cleanup:
Once the computer realizes a rubber band has shrunk into a tiny knot because it's near a cliff, it simply cuts that knot out of the final drawing. It says, "Okay, we don't need this specific rubber band to help draw the smooth parts; it's just causing trouble near the cliff."

By removing these "tight knots" near the cliff, the final map no longer has those ugly, spiky wiggles. The cliff remains sharp, and the rest of the map stays smooth.

How Do They Know Where the Cliff Is?

You might ask, "How does the computer know where the cliff is if it hasn't seen the whole picture yet?"

The authors use Smoothness Indicators. Think of this as a "shakiness detector."

For grid data (like a chessboard): They look at three points in a row. If the middle point is way higher or lower than its neighbors, the "shakiness" score goes up.
For scattered data (like raindrops): They use a more complex math trick (Least Squares) to see if the points around a spot are behaving like a smooth surface or a chaotic mess.

If the "shakiness" score is high, the computer knows, "Ah, there's a cliff here!" and it triggers the "tighten the rubber band" protocol.

Why Is This Important?

It's Automatic: You don't need to tell the computer where the cliffs are. It finds them itself.
It's Stable: A common fear with these tricks is that the math might break or become too messy to solve. The authors proved that their method keeps the math stable and solvable, just like the old method.
It Works Everywhere: They tested it on 1D lines (like a stock market graph) and 2D surfaces (like a 3D terrain map), and it worked beautifully for both.

The Bottom Line

Imagine you are a sculptor trying to carve a statue with a smooth face and a sharp, jagged beard. The old tools would either smooth out the beard (losing the detail) or make the face wobble (ruining the smoothness).

This new method is like a smart chisel that knows exactly when to be gentle and when to be sharp. It automatically tightens its grip near the jagged parts to stop the vibrations, ensuring the final statue is both sharp where it needs to be and smooth everywhere else.

1. Problem Statement

Radial Basis Function (RBF) interpolation is a powerful technique for approximating multivariate functions. However, when applied to functions containing jump discontinuities, classical RBF methods (which often use global or high-order linear approximations) suffer from the Gibbs phenomenon. This results in severe, non-physical oscillations near the discontinuities that degrade the quality of the approximation and can propagate far from the jump.

Existing solutions often require prior knowledge of the discontinuity locations (edge detection) or rely on variably scaled RBFs that introduce discontinuities into the interpolant itself. The authors aim to develop a method that:

Automatically detects discontinuities without prior knowledge.
Reduces oscillations near jumps.
Maintains high accuracy and stability in smooth regions.
Preserves the invertibility and favorable conditioning of the interpolation matrix.

2. Methodology

The proposed approach introduces a data-dependent modification of the RBF shape parameter ( $\varepsilon$ ). The core strategy involves adaptively increasing the shape parameter near discontinuities to force the basis functions to behave like local delta functions, effectively isolating the influence of data points at the jump.

A. Adaptive Shape Parameter

Instead of a global constant $\varepsilon$ , the authors define a point-dependent shape parameter $\tilde{\varepsilon}_i$ for each data point $x_i$ :
$\tilde{\varepsilon}_i = \frac{\varepsilon}{c + \psi(I_i)}$
Where:

$I_i$ is a smoothness indicator (a local measure of data smoothness).
$\psi(x)$ is a function that rounds $e^{-(Cx)^t}$ to the nearest integer.
$c, C, t$ are tuning constants ( $C=10, t=2, c=10^{-16}$ in experiments).

Behavior:

In smooth regions: $I_i$ is small ( $O(h^m)$ ), so $\psi(I_i) \approx 0$ . Thus, $\tilde{\varepsilon}_i \approx \varepsilon$ , preserving standard RBF behavior.
Near discontinuities: $I_i$ is large ( $O(1)$ ), causing $\psi(I_i)$ to become large. Consequently, $\tilde{\varepsilon}_i \to \infty$ . As $\varepsilon \to \infty$ , the RBF kernel $\phi(\tilde{\varepsilon}_i \|x - x_i\|)$ converges to a discrete delta function (1 at the center, 0 elsewhere).

B. Smoothness Indicators

To detect discontinuities, the method employs specific indicators based on the data type:

Grid-based data: Uses the squared undivided second-order differences (1D) or the squared five-point stencil Laplacian (2D).
Scattered (mesh-free) data: Uses a Least Squares (MLS) approximation of the Laplacian. The weights are computed to be exact for a polynomial space, and the result is scaled by the square of the local fill distance ( $h_{loc}^2$ ) to match the units of a second difference. The indicator is then squared.

C. The Interpolation Scheme

The final interpolant is defined as:
$\tilde{s}_{f,X}(x) = \sum_{i=1}^{N} \lambda_i \psi(I_i) \phi(\tilde{\varepsilon}_i \|x - x_i\|)$
Crucially, the term $\psi(I_i)$ acts as a switch. If a point is near a discontinuity, $\psi(I_i)$ becomes non-zero, but the kernel $\phi(\tilde{\varepsilon}_i \dots)$ becomes a delta function. The authors propose eliminating the "nails" (the localized spikes) generated by these points from the final reconstruction. Effectively, the interpolation is performed only on the smooth points, while the points near discontinuities are excluded from the summation to prevent oscillations.

3. Key Contributions & Theoretical Analysis

A. Invertibility of the Modified Matrix

A major theoretical concern is whether modifying the shape parameters (and thus changing columns of the interpolation matrix) destroys the matrix's invertibility.

Proof: The authors prove that if the original matrix $A$ is symmetric and positive definite, and columns corresponding to discontinuities are replaced by columns with a single non-zero diagonal entry (due to the delta-like behavior), the resulting matrix $B$ remains non-singular.
Mechanism: The determinant of the modified matrix is the product of the diagonal element and the determinant of the sub-matrix formed by the remaining smooth points.

B. Condition Number Analysis

The authors analyze the condition number of the new system. By rearranging the system into a block matrix form:
$M = \begin{bmatrix} \tilde{A} & 0 \\ C & I \end{bmatrix}$
Where $\tilde{A}$ is the interpolation matrix for smooth points and $I$ is the identity matrix for the "delta" points.

Result: The condition number of the new system is bounded and remains comparable to the condition number of the reduced system ( $\tilde{A}$ ) where discontinuity-adjacent points are excluded. This ensures numerical stability is not compromised.

4. Numerical Results

The method was tested on both 1D and 2D piecewise smooth functions using various kernels (Gaussian, Inverse Multiquadric, Wendland, Matérn) and data distributions (uniform grids and Halton sequences).

Oscillation Mitigation: In all experiments, the classical RBF method exhibited severe Gibbs oscillations near jumps. The data-dependent (DD-RBF) method significantly reduced these oscillations, effectively suppressing the Gibbs phenomenon.
Accuracy in Smooth Regions: The error norms ( $L_\infty$ ) in smooth regions remained essentially unchanged compared to the classical method, confirming that the adaptive mechanism does not degrade performance where the function is smooth.
Conditioning: The condition numbers of the DD-RBF matrices were comparable to, and often slightly lower than, the classical matrices.
Trade-off: The method introduces a mild "smearing" or smoothing of the discontinuity itself (the interpolant does not strictly pass through the data point at the jump), but this is a deliberate trade-off to eliminate the much larger oscillatory artifacts.

5. Significance

This work provides a robust, automatic solution to the Gibbs phenomenon in RBF interpolation without requiring prior knowledge of discontinuity locations.

Generality: It applies to various RBF kernels (both globally and compactly supported) and data types (gridded and scattered).
Stability: It resolves the theoretical concern that adaptive shape parameters might lead to singular or ill-conditioned systems.
Practicality: The method is computationally efficient, relying on local smoothness indicators and standard linear algebra operations, making it suitable for applications in surface reconstruction, image processing, and solving PDEs with discontinuous solutions.

In summary, the paper successfully bridges the gap between high-order RBF accuracy and the need for stability near discontinuities by transforming the interpolation problem into a locally adaptive, data-driven process that effectively filters out oscillatory components.