Riemann–Lebesgue Lemma

What happens to Fourier coefficients as oscillation speeds up?

A rapidly wiggling wave averages away the function's bumps. Within one wavelength the function looks nearly constant. So as the frequency grows, every coefficient dies to zero.

For example, shake a 🧡 string very fast. The wiggle blurs into a faint band, the ups and downs cancelling to leave only the middle.

an=1Ο€βˆ«βˆ’Ο€Ο€f(x)cos⁑(nx) dxβ†’0a_n = \frac1\pi\int_{-\pi}^{\pi} f(x)\cos(nx)\,dx \to 0
bn=1Ο€βˆ«βˆ’Ο€Ο€f(x)sin⁑(nx) dxβ†’0b_n = \frac1\pi\int_{-\pi}^{\pi} f(x)\sin(nx)\,dx \to 0

Examples & Exercises

Compute directly the Fourier coefficients bn=1Ο€βˆ«02Ο€f(x)sin⁑(nx) dxb_n = \frac{1}{\pi}\int_0^{2\pi} f(x)\sin(nx)\,dx of f=1[0,Ο€]f = 1_{[0, \pi]} (on [0,2Ο€][0, 2\pi]), find ana_n as well, and prove completely that an,bnβ†’0a_n, b_n \to 0.

bn=1βˆ’(βˆ’1)nnΟ€β€…β€Šβ†’nβ†’βˆžβ€…β€Š0,an=0b_n = \frac{1 - (-1)^n}{n\pi} \;\xrightarrow[n \to \infty]{}\; 0, \qquad a_n = 0
Step mode, n=12n = 12, ∣b∣|b| stems: odd terms sink as ∣bn∣=2nΟ€|b_n| = \frac{2}{n\pi} and even terms are exactly 0 β€” ∣b11∣=211Ο€β‰ˆ0.0579|b_{11}| = \frac{2}{11\pi} \approx 0.0579.
Show proof
  1. Since ff is 1 on [0,Ο€][0, \pi] and 0 on (Ο€,2Ο€](\pi, 2\pi], bn=1Ο€βˆ«0Ο€sin⁑(nx) dxb_n = \frac{1}{\pi}\int_0^{\pi} \sin(nx)\,dx. With ∫sin⁑(nx) dx=βˆ’cos⁑(nx)/n\int \sin(nx)\,dx = -\cos(nx)/n, we get bn=1Ο€[βˆ’cos⁑(nx)n]0Ο€=1βˆ’cos⁑(nΟ€)nΟ€b_n = \frac{1}{\pi}\left[-\frac{\cos(nx)}{n}\right]_0^{\pi} = \frac{1 - \cos(n\pi)}{n\pi}, and cos⁑(nΟ€)=(βˆ’1)n\cos(n\pi) = (-1)^n gives bn=1βˆ’(βˆ’1)nnΟ€b_n = \frac{1 - (-1)^n}{n\pi}.

  2. Since ∣1βˆ’(βˆ’1)nβˆ£β‰€2|1 - (-1)^n| \le 2, ∣bnβˆ£β‰€2nΟ€β†’0|b_n| \le \frac{2}{n\pi} \to 0 as nβ†’βˆžn \to \infty.

    Meanwhile an=1Ο€βˆ«0Ο€cos⁑(nx) dx=sin⁑(nΟ€)nΟ€=0a_n = \frac{1}{\pi}\int_0^{\pi} \cos(nx)\,dx = \frac{\sin(n\pi)}{n\pi} = 0 β€” exactly 0 for every nβ‰₯1n \ge 1.

    Hence an→0a_n \to 0 and bn→0b_n \to 0 are both proved.

  3. The scene's step mode reproduces this formula by numerical quadrature: the odd stems ride the 2/(nΟ€)2/(n\pi) curve downward while the even stems sit at 0 up to quadrature error. A discontinuous function's coefficients go at rate 1/n1/n β€” slower than for smooth functions (compare the tent mode).

Prove the Riemann–Lebesgue lemma for a step function s=βˆ‘k=1mck 1[xkβˆ’1,xk)s = \sum_{k=1}^{m} c_k\,1_{[x_{k-1}, x_k)} (on a partition of [0,2Ο€][0, 2\pi]): an(s),bn(s)β†’0a_n(s), b_n(s) \to 0.

∣1Ο€βˆ«xkβˆ’1xksin⁑(nx) dxβˆ£β‰€2nΟ€β€…β€Šβ†’nβ†’βˆžβ€…β€Š0\left|\frac{1}{\pi}\int_{x_{k-1}}^{x_k} \sin(nx)\,dx\right| \le \frac{2}{n\pi} \;\xrightarrow[n \to \infty]{}\; 0
Step mode, n=24n = 24: each of the finitely many intervals contributes at most 2/(nΟ€)2/(n\pi), so the total ∣bn∣|b_n| goes to 0 β€” the numerical version of the general step-function proof.
Show proof
  1. By linearity of the integral, bn(s)=βˆ‘k=1mck bn(1[xkβˆ’1,xk))b_n(s) = \sum_{k=1}^{m} c_k\,b_n(1_{[x_{k-1}, x_k)}). Each interval indicator contributes bn(1[xkβˆ’1,xk))=1Ο€βˆ«xkβˆ’1xksin⁑(nx) dx=cos⁑(nxkβˆ’1)βˆ’cos⁑(nxk)nΟ€b_n(1_{[x_{k-1}, x_k)}) = \frac{1}{\pi}\int_{x_{k-1}}^{x_k} \sin(nx)\,dx = \frac{\cos(nx_{k-1}) - \cos(nx_k)}{n\pi}.

  2. Since ∣cos⁑(nxkβˆ’1)βˆ’cos⁑(nxk)βˆ£β‰€2|\cos(nx_{k-1}) - \cos(nx_k)| \le 2, ∣bn(1[xkβˆ’1,xk))βˆ£β‰€2nΟ€\left|b_n(1_{[x_{k-1}, x_k)})\right| \le \frac{2}{n\pi} β€” a uniform bound independent of the interval's position.

    Hence ∣bn(s)βˆ£β‰€βˆ‘k=1m∣ckβˆ£β€‰2nΟ€=Cn|b_n(s)| \le \sum_{k=1}^{m} |c_k|\,\frac{2}{n\pi} = \frac{C}{n} (with C=2Ο€βˆ‘βˆ£ck∣C = \frac{2}{\pi}\sum |c_k| independent of nn), which goes to 0 as nβ†’βˆžn \to \infty.

    The same computation with cos⁑\cos in place of sin⁑\sin sends an(s)β†’0a_n(s) \to 0.

  3. The point is finiteness: with only finitely many intervals, adding up the O(1/n)O(1/n) contributions still gives O(1/n)O(1/n). This is the platform the density argument to continuous functions (problem 3 below) stands on.

First prove that a continuous ff on [0,2Ο€][0, 2\pi] can be uniformly approximated by step functions (via uniform continuity), then use it to prove bn(f)β†’0b_n(f) \to 0 (and an(f)β†’0a_n(f) \to 0) completely by an Ξ΅/3\varepsilon/3 argument.

∣bn(f)βˆ£β‰€βˆ£bn(fβˆ’s)∣+∣bn(s)βˆ£β‰€2Ξ΅+∣bn(s)βˆ£β€…β€Šβ†’nβ†’βˆžβ€…β€Š2Ξ΅|b_n(f)| \le |b_n(f - s)| + |b_n(s)| \le 2\varepsilon + |b_n(s)| \;\xrightarrow[n \to \infty]{}\; 2\varepsilon
Tent mode, n=24n = 24: continuous functions are uniformly approximable by step functions, so ∣anβˆ£β†’0|a_n| \to 0 too β€” here sinking even faster as ∣an∣=4/(Ο€2n2)|a_n| = 4/(\pi^2 n^2) (odd nn).
Show proof
  1. Lemma (uniform approximation by step functions): if ff is continuous on [0,2Ο€][0, 2\pi] it is uniformly continuous, so for every Ξ΅>0\varepsilon > 0 there is Ξ΄>0\delta > 0 with ∣xβˆ’y∣<Ξ΄β‡’βˆ£f(x)βˆ’f(y)∣<Ξ΅|x - y| < \delta \Rightarrow |f(x) - f(y)| < \varepsilon.

    Split [0,2Ο€][0, 2\pi] by a partition 0=x0<x1<β‹―<xm=2Ο€0 = x_0 < x_1 < \cdots < x_m = 2\pi (each width <Ξ΄{} < \delta), with intervals Ik=[xkβˆ’1,xk)I_k = [x_{k-1}, x_k) for k<mk < m and the last interval Im=[xmβˆ’1,2Ο€]I_m = [x_{m-1}, 2\pi] closed at the right endpoint, so the partition covers all of [0,2Ο€][0, 2\pi] including x=2Ο€x = 2\pi.

    Define the step function ss taking the constant value f(xkβˆ’1)f(x_{k-1}) (left endpoint) on each IkI_k (with s(2Ο€)=f(xmβˆ’1)s(2\pi) = f(x_{m-1}) as well).

    For x∈Ikx \in I_k, ∣xβˆ’xkβˆ’1∣<Ξ΄|x - x_{k-1}| < \delta, so ∣f(x)βˆ’s(x)∣<Ξ΅|f(x) - s(x)| < \varepsilon β€” this holds at every point of [0,2Ο€][0, 2\pi], i.e. βˆ₯fβˆ’sβˆ₯∞<Ξ΅\|f - s\|_\infty < \varepsilon.

    Lemma proved.

  2. Fix Ξ΅>0\varepsilon > 0 and pick a step function ss with βˆ₯fβˆ’sβˆ₯∞<Ξ΅\|f - s\|_\infty < \varepsilon by the lemma.

    By the triangle inequality ∣bn(f)βˆ£β‰€βˆ£bn(fβˆ’s)∣+∣bn(s)∣|b_n(f)| \le |b_n(f - s)| + |b_n(s)|.

    The first term satisfies ∣bn(fβˆ’s)βˆ£β‰€1Ο€βˆ«02Ο€βˆ£fβˆ’sβˆ£β€‰βˆ£sin⁑(nx)βˆ£β€‰dx≀1π Ρ⋅2Ο€=2Ξ΅|b_n(f - s)| \le \frac{1}{\pi}\int_0^{2\pi} |f - s|\,|\sin(nx)|\,dx \le \frac{1}{\pi}\,\varepsilon \cdot 2\pi = 2\varepsilon β€” independent of nn.

  3. The second term ∣bn(s)βˆ£β†’0|b_n(s)| \to 0 as nβ†’βˆžn \to \infty by problem 2.

    So choose NN large enough that ∣bn(s)∣<Ξ΅|b_n(s)| < \varepsilon for nβ‰₯Nn \ge N; for those nn, ∣bn(f)∣<2Ξ΅+Ξ΅=3Ξ΅|b_n(f)| < 2\varepsilon + \varepsilon = 3\varepsilon.

    Since Ξ΅\varepsilon is arbitrary, bn(f)β†’0b_n(f) \to 0.

    The same argument with ∣cos⁑(nx)βˆ£β‰€1|\cos(nx)| \le 1 gives an(f)β†’0a_n(f) \to 0.

    Every step is complete β€” nothing is left at idea level.

9. Series of Functions & Fourier