Trigonometry
Trigonometric Approximation
Цели урока
- Compute and bound the best trigonometric approximation E_N(f) via Jackson's theorem
- Explain how Chebyshev nodes eliminate Runge's phenomenon
- Compare Fejer sums to partial Fourier sums and explain why Fejer converges uniformly
- Connect smoothness class to compression efficiency
Предварительные знания
- Fourier Series
- L^2 spaces
- Continuity and uniform convergence
How many terms of a trigonometric polynomial are needed to approximate a smooth function to 6 decimal places - and why do smooth functions compress better?
- Core Audio (Apple): sinc-based resampling at 44100 to 48000 Hz uses trigonometric interpolation
- Chebyshev approximation in libm: mathematical functions computed via polynomial minimax
- EEG/ECG: windowed Fourier analysis avoids Gibbs artifacts in biomedical signal processing
- Spectral PDE methods: exponential convergence for smooth initial data
History: from Weierstrass to Jackson
Weierstrass proved in 1885 that continuous functions can be uniformly approximated by algebraic polynomials. The trigonometric analogue followed. Du Bois-Reymond in 1873 destroyed the hope for universal pointwise convergence of Fourier series. Fejer in 1904 showed Cesaro means converge uniformly - a constructive proof of Weierstrass. Dunham Jackson's 1912 dissertation gave quantitative rates: approximation speed is controlled by smoothness.
Best Trigonometric Approximation and Jackson's Theorem
Dunham Jackson in 1912 proved a quantitative approximation theorem: smooth functions are approximated by trigonometric polynomials at rate O(n^{-k}) for C^k functions. Apple uses trigonometric interpolation in Core Audio: resampling audio from 44,100 to 48,000 Hz processes 10^6 samples in 12 ms.
Decay rate of best approximation by function class: C^0 gives E_N = O(omega(f, 1/N)); Lipschitz alpha gives O(N^{-alpha}); C^k gives O(N^{-k}); analytic functions give O(e^{-alpha N}).
What does Jackson's theorem guarantee for f in C^k?
For f in C^k: E_N(f) <= C_k * omega_k(f, 1/N) / N^k = O(N^{-k}). Higher smoothness gives faster decay. Analytic functions give exponential decay.
Chebyshev Nodes and Runge's Phenomenon
Polynomial interpolation at equally spaced nodes diverges for many smooth functions - Runge's phenomenon. The classic example is f(x) = 1/(1 + 25x^2): with 20 uniform nodes the error explodes to over 1000. Chebyshev nodes solve this completely. LibreOffice and MATLAB use Chebyshev-based techniques internally for function evaluation.
Why do Chebyshev nodes eliminate Runge's phenomenon?
The node polynomial omega(x) = prod(x - x_k) has min-max norm 1/2^n for Chebyshev nodes - the smallest possible among all monic degree-(n+1) polynomials. This controls interpolation error and keeps the Lebesgue constant small.
Fejer Sums and Uniform Convergence
The Fejer kernel is nonnegative - and that single property makes all the difference. Where partial Fourier sums can diverge (du Bois-Reymond's 1873 example of a continuous function with divergent Fourier series), Fejer sums always converge uniformly. EEG signal processing uses windowed Fourier analysis precisely to avoid Gibbs artifacts.
Why do Fejer sums converge uniformly for continuous f when partial Fourier sums may not?
D_N has oscillating sign and ||D_N||_L1 grows as log(N), allowing divergent behavior. F_N = average of D_0,...,D_{N-1} is nonnegative. Convolution with a nonneg. approximate identity converges uniformly for any continuous f.
Connections to other topics
Trigonometric approximation theory links functional analysis to practical signal processing
- Weierstrass theorem — Related topic
- Fejer and Dirichlet kernels — Related topic
- Chebyshev polynomials — Related topic
- Spectral methods for PDEs — Related topic
Итоги
- Jackson's theorem: for f in C^k, the best trigonometric approximation E_N(f) = O(N^{-k}); analytic functions give exponential decay
- Fejer sums sigma_N f converge uniformly for any continuous 2*pi-periodic f; partial Fourier sums S_N f can diverge (du Bois-Reymond)
- Chebyshev nodes eliminate Runge's phenomenon by minimizing the node polynomial norm; Lebesgue constant grows as O(log n) vs O(2^n/n)
- Smoothness class determines compression efficiency: C^k signals need O(N^{-k}) more coefficients per unit accuracy than analytic signals
Вопросы для размышления
- Why does the Fejer kernel's nonnegativity guarantee uniform convergence while the Dirichlet kernel fails?
- How does Jackson's theorem explain the difference in compression quality for smooth vs. discontinuous image regions?
- What happens to the approximation rate at the boundary: f has a jump in only the first derivative?
Связанные уроки
- trig-21 — Trigonometric polynomials are finite Fourier partial sums
- trig-26-trig-poly — Dirichlet and Fejer kernels connect to approximation theory
- trig-25-dct — DCT exploits rapid coefficient decay of smooth functions