Trigonometry
Trigonometric Polynomials and Approximation
Цели урока
- Explain why the Fejer kernel gives uniform convergence while the Dirichlet kernel does not
- Apply the Weierstrass theorem on trigonometric approximation
- Compute the best approximation rate via Jackson's theorem
- Connect function smoothness to compressibility
Предварительные знания
- Fourier Series
- L^2 spaces
- Functional analysis
Why do some functions compress well and others do not - and how can this be measured precisely?
- FM synthesis: instrument timbres as trigonometric polynomials (Yamaha DX7)
- Digital filters: window functions vs. the Gibbs effect in EEG/ECG processing
- Image compression: smooth regions compress better than sharp edges
- Spectral PDE methods: exponential convergence for smooth data
History: from Weierstrass to Jackson
Weierstrass in 1885 proved that continuous functions can be uniformly approximated by algebraic polynomials. The trigonometric analogue followed. Du Bois-Reymond in 1873 constructed a continuous function with a divergent Fourier series, destroying hopes for universal pointwise convergence. Fejer in 1904 proved Cesaro means converge uniformly - a constructive bypass. Dunham Jackson's 1912 dissertation established quantitative rates: approximation speed is governed by smoothness.
Trigonometric Polynomials and Weierstrass Theorem
The Yamaha DX7 synthesizer reproduces 96 simultaneous tones by approximating instrument timbres with trigonometric polynomials of degree 32 - indistinguishable from a live piano for 99% of listeners. The mathematics behind this: the Weierstrass theorem guarantees that any continuous 2pi-periodic function can be approximated by a trigonometric polynomial to any desired accuracy.
Cesaro sums sigma_N f converge uniformly even where partial Fourier sums S_N f diverge. Du Bois-Reymond (1873) constructed a continuous function with a divergent Fourier series at a point.
What is the fundamental difference between the Fejer kernel and the Dirichlet kernel?
F_N(x) >= 0 eliminates the Gibbs phenomenon. Partial sums with D_N can diverge even for continuous f (du Bois-Reymond). The nonnegativity of F_N guarantees uniform convergence of sigma_N f.
Dirichlet Kernel and the Gibbs Phenomenon
The Dirichlet kernel is the convolution kernel through which partial Fourier sums are expressed. Its oscillations and the growing central peak explain why the Gibbs phenomenon is unavoidable with partial sums. EEG signals in real-time processing require windowing functions precisely because of the Gibbs effect.
Window functions (Hann, Blackman, Hamming) replace abrupt Fourier truncation with smooth tapering, reducing side lobes at the cost of widening the main lobe.
Why does the growth of ||D_N||_L1 as log(N) matter for convergence theory?
By the Banach-Steinhaus theorem: if sup_N ||S_N|| = infinity, then there exists f in C with sup_N ||S_N f||_inf = infinity. This is why uniform convergence of Fourier series for all continuous functions is impossible.
Jackson's Theorem and Approximation Rates
Jackson's theorem (1912) gives a quantitative bound: how accurately a C^k function is approximated by a trigonometric polynomial of degree N. This is the key theorem for understanding why smooth signals compress well: the higher the bounded derivatives, the fewer coefficients are needed.
| Function class | Rate E_N(f) | Example |
|---|---|---|
| Continuous f in C | O(omega(f, 1/N)) -> 0 | Square wave |
| Lipschitz f in Lip(alpha) | O(N^{-alpha}), 0 < alpha <= 1 | Fractional derivative |
| f in C^k | O(N^{-k}) | Polynomial * exp(-x^2) |
| Analytic f | O(e^{-alpha N}) | sin(x), cos(x), exp(cos(x)) |
What does Jackson's theorem guarantee for f in C^k?
Jackson: for f in C^k with bounded k-th smoothness modulus, E_N(f) <= C_k omega_k(f, 1/N) / N^k = O(N^{-k}). Smoother functions give faster approximation rate.
Connections to other topics
Trigonometric approximation theory connects functional analysis to compression practice
- Weierstrass theorem — Related topic
- Fejer and Dirichlet kernels — Related topic
- Spectral methods — Related topic
- Digital filters — Related topic
Итоги
- The Dirichlet kernel D_N generates S_N f via convolution; its L^1-norm grows as log(N), allowing divergent Fourier series
- The Fejer kernel F_N >= 0 generates uniformly convergent Cesaro sums sigma_N f for any continuous f
- Weierstrass theorem: any continuous 2*pi-periodic function is uniformly approximatable by trigonometric polynomials
- Jackson's theorem: for f in C^k, the best approximation E_N(f) = O(N^{-k}); analytic functions give exponential decay
Вопросы для размышления
- Why do Cesaro sums converge where partial Fourier sums can diverge?
- How does Jackson's theorem explain the quality difference when compressing smooth vs. discontinuous image regions?
- What happens to the approximation rate when only the first derivative has a jump?
Связанные уроки
- trig-21 — Trigonometric polynomials are finite Fourier partial sums
- trig-23 — Jackson's theorem quantifies trigonometric approximation rates
- trig-25-dct — DCT uses trigonometric polynomials as basis