Function Approximation๐
An interactive guide to Fourier basis, RBF, and neural networks.
1. Introduction to Function Approximation๐
Function approximation is the task of finding a simpler function that closely matches a complex target function. This is fundamental in machine learning, signal processing, and numerical analysis.
Key Idea: Represent any function f(x) as a combination of simpler "basis functions":
\[f(x) \approx \sum_i w_i \phi_i(x)\]
where \(\phi_i(x)\) are basis functions and \(w_i\) are weights/coefficients.
Why Function Approximation?
- Compression: Store complex functions using fewer parameters
- Generalization: Learn patterns from data and predict on new inputs
- Analysis: Understand function properties through basis decomposition
- Computation: Replace expensive function evaluations with fast approximations
2. Fourier Basis Approximation๐
Fourier approximation represents functions as sums of sine and cosine waves at different frequencies.
Fourier Series:
\[f(x) = a_0 + \sum_n [a_n\cos(nx) + b_n\sin(nx)]\]
Coefficients:
\[a_0 = \frac{1}{2\pi} \int f(x)dx\]
\[a_n = \frac{1}{\pi} \int f(x)\cos(nx)dx\]
\[b_n = \frac{1}{\pi} \int f(x)\sin(nx)dx\]
Interactive Fourier Approximation
5
Approximation
Individual Basis Functions
Key Properties
- Global basis: Each term affects the entire domain
- Frequency interpretation: Low frequencies = trends, high frequencies = details
- Best for: Periodic signals, smooth functions
- Gibbs phenomenon: Overshoot at discontinuities (~9%)
3. Radial Basis Functions (RBF)๐
RBF networks represent functions as weighted sums of "bumps" centered at different points.
RBF Approximation:
\[f(x) = \sum_i w_i \phi(||x - c_i||)\]
Common RBF Types:
- Gaussian: \(\phi(r) = \exp(-r^2/\sigma^2)\)
- Multiquadric: \(\phi(r) = \sqrt{r^2 + \sigma^2}\)
- Inverse multiquadric: \(\phi(r) = 1/\sqrt{r^2 + \sigma^2}\)
Interactive RBF Approximation
7
0.5
Approximation
Individual RBF Basis Functions
Key Properties
- Local basis: Each RBF has limited influence
- Spatial interpretation: Centers determine which regions are captured
- Best for: Scattered data, non-periodic functions, interpolation
- Meshless: Works in multiple dimensions without structured grid