Fourier solution to the Wave Equation

[Math 619 home]

Contents

The complex exponential

Many of the computations involving Fourier series are simpler when we use the complex exponential instead of sine and cosine, so in this section we will consider complex valued solutions to the wave equation.

Euler’s Formulas

By definition

eiθ=cosθ+isinθfor all θR.(1)e^{i\theta} = \cos \theta + i\sin\theta\quad\text{for all }\theta\in\R. \tag{1}

This implies

eiθ=cosθisinθfor all θRe^{-i\theta} = \cos \theta - i\sin\theta\quad\text{for all }\theta\in\R

and hence

cosθ=eiθ+eiθ2,sinθ=eiθeiθ2i(2)\cos\theta = \frac{e^{i\theta} + e^{-i\theta}}{2}, \qquad \sin\theta = \frac{e^{i\theta} - e^{-i\theta}}{2i} \tag{2}

Derivative of the complex exponential

Euler’s definition of eiθe^{i\theta} implies

deiθdθ=ieiθ.(3)\frac{de^{i\theta}}{d\theta} = i e^{i\theta}. \tag{3}

In the language of Linear Algebra, eiθe^{i\theta} is an eigenvector of the linear transformation fdfdθf\mapsto \frac{df}{d\theta} with eigenvalue ii.

Trigonometric polynomials

A function of the form

f(x)=A0+n=1N(Ancosnx+Bnsinnx)f(x) = A_0 + \sum_{n=1}^N \bigl(A_n\cos nx + B_n\sin nx\bigr)

where An,BnCA_n, B_n\in \mathbb{C} are constants is called a trigonometric polynomial. Using Euler’s formulas we can rewrite a trigonometric polynomial in terms of complex exponentials:

f(x)=A0+n=1N(Ancosnx+Bnsinnx)=A0+n=1NAniBn2einx+An+iBn2einx=n=NNf^neinx,\begin{aligned} f(x) &= A_0 + \sum_{n=1}^N \bigl(A_n\cos nx + B_n\sin nx\bigr) \\ &=A_0 +\sum_{n=1}^N \frac{A_n-iB_n}{2}e^{inx} + \frac{A_n+iB_n}{2} e^{-inx} \\ &=\sum_{n=-N}^N \hat f_n e^{inx}, \end{aligned}

provided we define

f^n={(AniBn)/2n<0(An+iBn)/2n>0A0n=0\hat f_n = \begin{cases} (A_n-iB_n)/2 & n\lt 0 \\ (A_n+iB_n)/2 & n\gt 0 \\ A_0 & n=0 \end{cases}

Finite string, Fourier solution

The wave equation is linear. If u,v:R2Ru, v:\R^2\to\R are solutions to the wave equation,

then w(x,t)=au(x,t)+bv(x,t)w(x, t) = au(x, t)+bv(x, t) is also a solution for any choice of a,bRa, b\in\R. This is true both for classical solutions and for weak solutions; i.e. if u,vu,v are classical solutions, then w=au+bvw=au+bv is a classical solution, and if u,vu,v are weak solutions, then w=au+bvw=au+bv is a weak solution.

Solutions that are periodic in space. Instead of considering arbitrary solutions to the wave equation, we look for solutions for which

u(x+2π,t)=u(x,t)u(x+2\pi, t) = u(x, t)

holds for all x,tx, t. By direct substitution in the equation we can verify that each of the functions

einteinx,einteinxe^{int}e^{inx}, \qquad e^{-int}e^{inx}

are classical solutions of the wave equation, and that they are 2π2\pi–periodic in the xx variable. It follows that for any choice of a^n,b^nC\hat a_n, \hat b_n\in\mathbb{C} the linear combination

u(x,t)=n=NNa^neinteinx+b^neinteinx(4)u(x, t) = \sum_{n=-N}^N \hat a_ne^{int}e^{inx} + \hat b_ne^{-int}e^{inx}\tag{4}

is again a solution of the wave equation. If we have infinitely many coefficients a^n,b^n\hat a_n, \hat b_n then we could try to show that the series converges and that its limit is a weak or classical solution of the wave equation.

Sometimes it is convenient to rewrite (4) by applying Euler’s formula e±int=cosnt±isinnte^{\pm int} = \cos nt \pm i\sin nt, with result

u(x,t)=n=NNu^ncos(nt)einx+v^nsin(nt)einx(5)u(x, t) = \sum_{n=-N}^N \hat u_n\cos (nt) e^{inx} + \hat v_n\sin(nt)e^{inx} \tag{5}

where u^n=a^n+b^n\hat u_n = \hat a_n+\hat b_n and v^n=i(a^nb^n)\hat v_n = i(\hat a_n-\hat b_n).

The initial value problem. Suppose we want to find a solution to

utt=uxx, for all x,tu(x+2π,t)=u(x,t), for all x,tu(x,0)=f(x),ut(x,0)=g(x) for all x\begin{aligned} &u_{tt}=u_{xx}, \text{ for all }x,t\\ &u(x+2\pi, t)=u(x, t), \text{ for all }x, t\\ &u(x, 0) = f(x),\quad u_t(x, 0) = g(x) \text{ for all }x \end{aligned}

by looking for a function u(x,t)u(x, t) that is given by (5). The function already satisfies the PDE, and it is 2π2\pi–periodic in the xx variable, so we only have to choose the coefficients u^n,v^n\hat u_n, \hat v_n so that uu satisfies the initial conditions. It follows from (5) that

f(x)=u(x,0)=nu^neinx,g(x)=ut(x,0)=nnv^neinx.f(x) = u(x, 0) = \sum_n \hat u_ne^{inx}, \qquad g(x) = u_t(x, 0) = \sum_n -n\hat v_n e^{inx}.

If ff and gg are trigonometric polynomials given by

f(x)=n=NNf^neinx,g(x)=n=NNg^neinx,f(x) = \sum_{n=-N}^N \hat f_ne^{inx},\qquad g(x) = \sum_{n=-N}^N \hat g_ne^{inx},\qquad

then we should choose u^n,v^n\hat u_n, \hat v_n so that

u^n=f^n,v^n=g^nn.\hat u_n = \hat f_n,\qquad \hat v_n = \frac{\hat g_n}{n}.

Thus the solution is

u(x,t)=n=NN{f^ncos(nt)+g^nnsin(nt)}  einx.(6)u(x, t) = \sum_{n=-N}^N \left\{\hat f_n\cos(nt)+\frac{\hat g_n}{n}\sin(nt)\right\}\; e^{inx}. \tag{6}

How to find the coefficients f^n,g^n\hat f_n, \hat g_n if we know ff and gg.

If f(x)=NNf^neinxf(x)=\sum_{-N}^N \hat f_n e^{inx} then Fourier multiplied the equation with eikxe^{-ikx} and integrated from 00 to 2π2\pi:

02πf(x)eikxdx=02πn=NNf^nei(nk)xdx=n=NNf^n02πei(nk)xdx=2πf^k,\int_0^{2\pi} f(x)e^{-ikx}dx= \int_0^{2\pi}\sum_{n=-N}^N \hat f_n e^{i(n-k)x}dx= \sum_{n=-N}^N \hat f_n \int_0^{2\pi} e^{i(n-k)x}dx= 2\pi \hat f_k,

because when knk\neq n

02πei(nk)xdx=[ei(nk)xi(nk)]x=0π=0\int_0^{2\pi} e^{i(n-k)x}dx= \left[\frac{e^{i(n-k)x}}{i(n-k)}\right]_{x=0}^\pi =0

while when k=nk=n we get

02πei(nk)xdx=02π1  dx=2π.\int_0^{2\pi} e^{i(n-k)x}dx= \int_0^{2\pi} 1\;dx=2\pi.

Thus, if we know the function ff then its coefficients are given by

f^k=12π02πf(x)eikxdx(7)\hat f_k = \frac{1}{2\pi}\int_0^{2\pi} f(x)e^{-ikx}dx \tag{7}

Including infinitely many terms

If we completely ignore questions about convergence, then we could let NN\to\infty and claim that if the initial functions ff and gg are given by

f(x)=f^neinx,g(x)=g^neinx,(8)f(x) = \sum_{-\infty}^\infty \hat f_n e^{inx}, \qquad g(x) = \sum_{-\infty}^\infty \hat g_n e^{inx},\tag{8}

then the 2π2\pi–periodic solution to the wave equation with initial values u(x,0)=f(x)u(x, 0)=f(x) and ut(x,0)=g(x)u_t(x, 0)=g(x) is given by

u(x,t)={f^ncos(nt)+g^nnsin(nt)}einx(9)u(x, t) = \sum_{-\infty}^\infty \bigl\{\hat f_n \cos(nt) + \frac{\hat g_n}{n}\sin(nt)\bigr\} e^{inx} \tag{9}

So if you can write the initial function and time derivative as a Fourier series (8) then we have a solution given in (9). This leaves us with a few questions

Fourier’s remarkable answer to the first question was: every 2π2\pi–periodic function ff has a Fourier expansion and the coefficients can be found by computing the integrals in (7).

For example, the “sawtooth function” is given by

f(x)=πx2 for 0<x<2πf(x) = \frac{\pi-x}{2} \text{ for }0\lt x\lt 2\pi

Fourier computed the coefficients in its expansion and claimed

f(x)=n=1sinnxn=sinx+12sin2x+13sin3x+f(x) = \sum_{n=1}^\infty \frac{\sin nx}{n} = \sin x + \tfrac12 \sin 2x + \tfrac13 \sin 3x + \cdots

Wikipedia offers a one paragraph history of Fourier’s claims.

Digression: convergence of Fourier series

There is not enough space in this aside on Fourier series to do the topic justice. If you want to read more, the book Fourier Analysis by Stein and Shakarchi (Princeton lectures in analysis) offers a good introduction. Alternatively, T.W.Körner’s book Fourier Analysis (yes, same title) from Cambridge University press is very readable.

Questions: was Fourier right?

Let ff be a 2π2\pi–periodic function. Does the Fourier series for ff converge, and does it converge to ff? To be more precise, define the partial sums of the series

SNf(x)=NNf^neinx,f^n=12π02πeinxf(x)  dx.S_Nf(x) = \sum_{-N}^N \hat f_n e^{ i n x},\qquad \hat f_n = \frac{1}{2\pi}\int_0^{2\pi} e^{-inx}f(x)\; dx.

Under what conditions can we guarantee that limNSNf(x)=f(x)\lim_{N\to\infty} S_Nf(x) = f(x)? Is the converge uniform? What other notions of convergence might apply here?

The Dirichlet kernel. For any 2π2\pi–periodic function ff the partial sums of the Fourier series of ff are given by

SNf(x)=02πDN(xξ)f(ξ)  dξ,whereDN(x)=12πsin(N+12)xsinx2S_Nf(x) = \int_0^{2\pi} D_N(x-\xi) f(\xi)\; d\xi,\quad\text{where}\quad D_N(x) = \frac{1}{2\pi} \frac{\sin(N+\frac12)x}{\sin \frac{x}{2}}

The function DND_N is called the Dirichlet kernel. It satisfies

02πDN(xξ)  dξ=1 for all xR.\int_0^{2\pi} D_N(x-\xi)\;d\xi = 1 \text{ for all }x\in\R.

Proof.

SNf(x)=12πn=NN02πein(xξ)f(ξ)  dξ=12π02πn=NNein(xξ)f(ξ)  dξ=02πDN(xξ)f(ξ)  dξ S_Nf(x) = \frac{1}{2\pi}\sum_{n=-N}^N \int_0^{2\pi} e^{in(x-\xi)}f(\xi)\;d\xi = \frac{1}{2\pi}\int_0^{2\pi} \sum_{n=-N}^N e^{in(x-\xi)}f(\xi)\;d\xi =\int_0^{2\pi} D_N(x-\xi) f(\xi)\; d\xi

where

DN(x)=12πNNeinxD_N(x) = \frac{1}{2\pi}\sum_{-N}^N e^{inx}

We can compute the sum by using the formula for geometric sums:

NNeinx=eiNx+ei(N1)x++ei(N1)x+eiNx=eiNx{1+eix+e2ix++e2iNx}=eiNxe(2N+1)ix1eix1=ei(N+12)xei(N+12)xeix/2eix/2=sin(N+12)xsinx2\begin{aligned} \sum_{-N}^N e^{inx} &= e^{-iNx}+e^{-i(N-1)x}+\cdots+e^{i(N-1)x}+e^{iNx}\\ &= e^{-iNx} \bigl\{1+e^{ix} + e^{2ix} + \cdots + e^{2iNx}\bigr\} \\ &=e^{-iNx} \frac{e^{(2N+1)ix} - 1}{e^{ix}-1}\\ &=\frac{e^{i(N+\frac{1}{2})x} - e^{-i(N+\frac{1}{2})x}}{e^{ix/2}-e^{-ix/2}}\\ &=\frac{\sin(N+\frac{1}{2})x}{\sin \frac{x}{2}} \end{aligned}

Q.E.D.

Riemann–Lebesgue Lemma. If f:[a,b]Rf:[a,b]\to\R is Riemann integrable, then

limλabsin(λx)f(x)  dx=limλabcos(λx)f(x)  dx=0.\lim_{\lambda\to\infty} \int_a^b \sin(\lambda x) f(x)\; dx = \lim_{\lambda\to\infty} \int_a^b \cos(\lambda x) f(x)\; dx =0.

Proof. We show that sin(λx)f(x)dx0\int \sin(\lambda x)f(x)dx\to0, the proof that cos(λx)f(x)dx0\int \cos(\lambda x)f(x)dx\to0 being nearly identical.

Let ϵ>0\epsilon\gt0 be given. Then, by definition of Riemann integrability of the function ff, there exists a partition a=x0<x1<x2<<xm1<xm=ba=x_0\lt x_1\lt x_2\lt \cdots\lt x_{m-1}\lt x_m=b, and numbers mi<Mim_i\lt M_i such that

Define the step function

s(x)=mk for all x(xk1,xk] and for all k=1,2,,ms(x) = m_k \text{ for all $x\in (x_{k-1}, x_k]$ and for all $k=1, 2, \ldots, m$}

Then s(x)f(x)s(x)\leq f(x) and 0ab(f(x)s(x))  dx<ϵ/(2(ba))0\leq\int_a^b (f(x)-s(x))\;dx \lt \epsilon/\bigl(2(b-a)\bigr).

To show that absin(λx)f(x)  dx0\int_a^b \sin(\lambda x) f(x)\; dx\to0 we rewrite the integral as follows

absin(λx)f(x)  dx=absin(λx)s(x)  dx+absin(λx)(f(x)s(x))  dx=A+B.\int_a^b \sin(\lambda x) f(x)\; dx =\int_a^b \sin(\lambda x) s(x)\; dx + \int_a^b \sin(\lambda x) \bigl(f(x)-s(x)\bigr)\; dx = A+B.

We can compute the first term AA explicitly:

A=absin(λx)s(x)  dx=k=1mxk1xksin(λx)mkdx=k=1mmk[cos(λx)λ]xk1xk=1λk=1mmk[cosλxk1cosλxk] A= \int_a^b \sin(\lambda x) s(x)\; dx = \sum_{k=1}^m \int_{x_{k-1}}^{x_k} \sin(\lambda x) m_kdx = \sum_{k=1}^m m_k\left[\frac{-\cos(\lambda x)}{\lambda}\right]_{x_{k-1}}^{x_k} =\frac {1}{\lambda}\sum_{k=1}^m m_k\left[\cos\lambda x_{k-1} - \cos\lambda x_{k}\right]

which implies

A2λk=1mmk.\left\vert A\right\vert \leq \frac{2}{\lambda} \sum_{k=1}^m |m_k| .

Therefore, if we define λϵ=ϵ4mk\lambda_\epsilon = \dfrac{\epsilon}{4\sum m_k} then A<ϵ/2|A|\lt \epsilon/2 holds for all λ>λϵ\lambda\gt\lambda_\epsilon.

Next we show that BB is also small. No matter what λ\lambda is, we always have

B=absin(λx)(f(x)s(x))  dxabsin(λx)f(x)s(x)  dx(ba)ϵ2(ba)=ϵ2.|B|=\left\vert \int_a^b \sin(\lambda x) \bigl(f(x)-s(x)\bigr)\; dx \right\vert \leq \int_a^b \left\vert \sin(\lambda x) \right\vert \big|f(x)-s(x)\big|\; dx \leq (b-a)\cdot \frac{\epsilon}{2(b-a)} = \frac\epsilon 2.

Therefore we get A+BA+B<ϵ/2+ϵ/2=ϵ|A+B| \leq |A|+|B| \lt \epsilon/2 + \epsilon/2 = \epsilon for all λ>λϵ\lambda \gt \lambda_\epsilon.

Q.E.D.

Definition. A function f:RRf:\R\to\R is called Lipschitz continuous at xRx\in\R if there exists a C>0C\gt 0 such that

f(ξ)f(x)Cxξ for all ξR.|f(\xi)-f(x)|\leq C|x-\xi| \text{ for all }\xi\in\R.

Convergence of a Fourier series at a point of Lipschitz continuity

Assume that f:RRf:\R\to\R is 2π2\pi–periodic, Riemann integrable, and that ff is Lipschitz continuous at aa. Then

f(a)=kZf^keika, i.e. limNk=NNf^keika=f(a)f(a) = \sum_{k\in \Z} \hat f_k e^{ika}, \text{ i.e. } \lim_{N\to\infty} \sum_{k=-N}^N \hat f_k e^{ika} = f(a)

Proof. It follows from 02πDN(ξa)dξ=1\int_0^{2\pi} D_N(\xi-a)d\xi=1 that f(a)=02πDN(ξa)f(a)dξf(a) = \int_0^{2\pi} D_N(\xi-a)f(a)d\xi. Therefore

SNf(a)f(a)=02πDN(ξa)(f(ξ)f(a))  dξ.S_Nf(a)- f(a) = \int_0^{2\pi} D_N(\xi - a) \bigl(f(\xi)-f(a)\bigr) \; d\xi.

Substitute ξ=a+s\xi=a+s, and use the definition of the Dirichlet kernel:

SNf(a)f(a)=ππDN(s)(f(a+s)f(a))  ds=12πππsin((N+12)s)  f(a+s)f(a)sins/2  ds=ππsin((N+12)s)  g(s)  ds\begin{aligned} S_Nf(a)- f(a) &= \int_{-\pi}^{\pi} D_N(s) \bigl(f(a+s)-f(a)\bigr) \; ds\\ &= \frac{1}{2\pi} \int_{-\pi}^{\pi} \sin\bigl((N+\tfrac12)s\bigr)\; \frac{f(a+s)-f(a)}{\sin s/2} \; ds\\ &=\int_{-\pi}^\pi \sin\bigl((N+\tfrac12)s\bigr) \; g(s)\;ds \end{aligned}

where

g(s)=deff(a+s)f(a)2πsins/2.g(s)\stackrel{\sf def}{=} \frac{f(a+s)-f(a)}{2\pi\sin s/2} .

This function is bounded because

g(s)Cs2πsin(s/2)C2 for all s(π,π),s0.|g(s)|\leq \frac{C|s|}{2\pi|\sin (s/2)|} \leq \frac C2 \text{ for all } s\in (-\pi, \pi), s\neq0.

Here we have used that sinxπ2x\sin x\geq \frac{\pi}{2}x for all x[0,π2]x\in[0, \frac\pi2].
using the boundedness of gg and the fact that ff is Riemann integrable, it is now a somewhat lengthy exercise in real analysis to show that gg is also Riemann integrable on [π,π][-\pi, \pi]. We can therefore invoke the Riemann–Lebesgue lemma and conclude that there is a λϵ\lambda_\epsilon such that for all λ>λϵ\lambda\gt\lambda_\epsilon one has

ππsin(λs)g(s)ds<ϵ.\left\vert\int_{-\pi}^{\pi} \sin (\lambda s) g(s) ds\right\vert \lt \epsilon.

and therefore SNf(a)f(a)<ϵ|S_Nf(a)- f(a)| \lt \epsilon for all λ>λϵ\lambda\gt \lambda_\epsilon, i.e. for all N>λϵ12N>\lambda_\epsilon-\frac12.

Q.E.D.

Theorem. If ff is a 2π2\pi–periodic C2C^2 function then its Fourier coefficients f^n\hat f_n satisfy

f^nfn2 for all n0.|\hat f_n| \leq \frac{\Vert f''\Vert_\infty}{n^2} \qquad \text{ for all }n\neq 0.

The Fourier series f(x)=f^neinxf(x) = \sum_{-\infty}^\infty \hat f_n e^{ i nx} converges uniformly, and absolutely.

Proof. Integration by parts in the definition of f^n\hat f_n gives us

2πf^n=02πeinxf(x)  dx=[einxinf(x)]02π1in02πeinxf(x)  dx=1in02πeinxf(x)  dx.2\pi\hat f_n = \int_0^{2\pi} e^{-inx}f(x)\; dx = \left[\frac{e^{-inx}}{-in}f(x)\right]_0^{2\pi} - \frac{1}{-in}\int_0^{2\pi} e^{-inx}f'(x)\; dx = \frac{1}{in}\int_0^{2\pi} e^{-inx}f'(x)\; dx.

Integrating by parts again we get

2πf^n=1n202πeinxf(x)  dx.2\pi\hat f_n= \frac{-1}{n^2}\int_0^{2\pi} e^{-inx}f''(x)\; dx.

This implies

2πf^n=1n202πeinxf(x)  dx1n202πeinxf(x)  dx1n202πfdx=2πfn2.2\pi|\hat f_n| =\left|\frac{-1}{n^2}\int_0^{2\pi} e^{-inx}f''(x)\; dx\right| \leq \frac{1}{n^2}\int_0^{2\pi} \left|e^{-inx}f''(x)\right|\; dx \leq \frac{1}{n^2}\int_0^{2\pi} \|f''\|_\infty dx = 2\pi\frac{\|f''\|_\infty}{n^2}.

The terms in the Fourier series for ff are bounded by

f^neinx=f^nfn2(n0).\left|\hat f_n e^{inx}\right| =|\hat f_n|\leq \frac{\|f''\|_\infty}{n^2}\qquad (n\neq0).

Therefore, by the Weierstrass M-test, the series f^neinx\sum_{-\infty}^\infty \hat f_n e^{ i nx} converges uniformly, and absolutely.

Since ff is C1C^1 it is also Hölder continuous at all xRx\in\R. Therefore Theorem 3.5 implies that f^neinx=f(x)\sum_{-\infty}^\infty \hat f_n e^{ i nx} = f(x) for all xRx\in\R.

Q.E.D.

Fourier series and the L2L^2 Inner Product

An inner product

Let Rper\mathcal R_{\rm per} be the set of Riemann integrable 2π2\pi–periodic functions, and let Cper\mathcal C_{\rm per} be the subset of continuous 2π2\pi–periodic functions. For any two functions f,gRperf, g\in \mathcal R_{\rm per} we define

f,g=02πf(x)g(x)  dx.\langle f, g\rangle = \int_0^{2\pi} f(x)\overline{g(x)}\; dx\,.

Then \langle \dots \rangle has the following properties:

Theorem. If fCperf\in \mathcal C_{\rm per} and f,f=0\langle f, f\rangle = 0 then f=0f=0.

On the other hand there exist functions fRperf\in\mathcal R_{\rm per} with f0f\neq 0 and f,f=0\langle f, f\rangle = 0.

Proof. Arguing by contradiction we suppose f,f=0\langle f, f\rangle = 0 and f(x0)0f(x_0)\neq 0 for some x0[0,2π)x_0\in [0, 2\pi). Since ff is continuous, there is an δ>0\delta\gt0 such that f(x)12f(x0)|f(x)|\geq \frac12 |f(x_0)| for all x(x0δ,x0+δ)x\in (x_0-\delta, x_0+\delta). This implies

0=f,f=02πf(x)2dxx0δx0+δf(x)2dx(12f(x0))2(2δ)>0.0=\langle f, f\rangle = \int_0^{2\pi} |f(x)|^2dx \geq \int_{x_0-\delta}^{x_0+\delta} |f(x)|^2dx \geq \bigl(\tfrac12 |f(x_0)|\bigr)^2 \cdot (2\delta) >0.

This is a contradiction, so we conclude that f(x0)=0f(x_0)=0 for all x0x_0. \qquad

To prove the second part of the theorem, consider the function ff for which f(x)=0f(x)=0 at all xRx\in\R except x=nπx=n\pi (nZn\in\Z) where f(nπ)=1f(n\pi)=1. Then ff is Riemann integrable and 2π2\pi–periodic, and f,f=0\langle f, f\rangle = 0 even though f0f\neq 0.

Q.E.D.

Theorem. Cper\mathcal C_{\rm per} with f,g\langle f, g \rangle defined as above is a complex vector space with an inner product.

See the analysis appendix for a review of complex inner products.

We define the L2L^2 norm of fRperf\in\mathcal R_{\rm per} to be f2\|f\|_2 where

f22=f,f=02πf(x)2  dx.\|f\|_2^2 = \langle f, f\rangle = \int_0^{2\pi} |f(x)|^2\;dx.

The L2L^2 and the supremum norm. The L2L^2 norm and the supremum norm f=supxf(x)\|f\|_\infty = \sup_x|f(x)| are generally different. They always satisfy

f22πf for any fRper.\|f\|_2 \leq \sqrt{2\pi} \|f\|_\infty \text{ for any }f\in\mathcal R_{\rm per}.

This follows from the following short computation

f22=02πf(x)2dx02πf2dx=2πf2.\|f\|_2^2 = \int_0^{2\pi} |f(x)|^2dx \leq \int_0^{2\pi}\|f\|_\infty^2dx=2\pi\|f\|_\infty^2.

Fourier coefficients as inner products. Let ek(x)=eikxe_k(x)= e^{ikx}. Then for any kZk\in\Z and any lkl\neq k one has

ek2=2π, and ek,el=0, i.e. ekel.\|e_k\|^2 = 2\pi, \text{ and } \langle e_k, e_l\rangle = 0, \text{ i.e. }e_k\perp e_l.

If f=f^NeN++f^NeNf=\hat f_{-N}e_{-N}+\cdots+\hat f_Ne_N then we have

f,ek=f^kek2=2πf^k.\langle f, e_k\rangle = \hat f_k \|e_k\|^2 = 2\pi \hat f_k.

The kthk^{\rm th} Fourier coefficient of fRperf\in\mathcal R_{\rm per} is given by

f^k=12πf,ek\hat f_k = \frac{1}{2\pi} \langle f, e_k\rangle

The NthN^{\rm th} partial sum of the Fourier series of fRperf\in\mathcal R_{\rm per} is given by

SNf=12πk=NNf,ekekS_Nf = \frac{1}{2\pi}\sum_{k=-N}^N \langle f, e_k\rangle e_k

Bessel’s inequality. For any fRperf\in\mathcal R_{\rm per} we have

SNf22+fSNf22=f22.\|S_Nf\|_2 ^2 + \|f-S_Nf\|_2^2 = \|f\|_2^2.

In particular SNf2f2\|S_Nf\|_2\leq \|f\|_2 always holds.

Proof. We begin by showing that SNf(fSNf)S_Nf \perp (f-S_Nf). For any integer kk with kN|k|\leq N we have

ek,fSNf=ek,f12πl=Nnek,f,elel=ek,f12πl=Nnf,elek,el=ek,f12πl=Nnel,fek,elek,el=0 if kl=ek,f12πek,fek,ekek,ek=2π=0.\begin{aligned} \langle e_k, f-S_Nf\rangle &= \langle e_k, f\rangle - \frac{1}{2\pi}\sum_{l=-N}^n \langle e_k, \langle f, e_l\rangle e_l\rangle\\ &= \langle e_k, f\rangle - \frac{1}{2\pi}\sum_{l=-N}^n \overline{\langle f, e_l\rangle}\langle e_k, e_l\rangle\\ &= \langle e_k, f\rangle - \frac{1}{2\pi}\sum_{l=-N}^n \langle e_l, f\rangle\langle e_k, e_l\rangle &\qquad&\langle e_k, e_l\rangle=0 \text{ if }k\neq l \\ &= \langle e_k, f\rangle -\frac{1}{2\pi}\langle e_k, f\rangle\langle e_k, e_k\rangle &\qquad&\langle e_k, e_k\rangle=2\pi \\ &=0. \end{aligned}

Thus ekfSNfe_k\perp f-S_Nf for all k=N,,Nk=-N, \dots, N. Since SNfS_Nf is a linear combination of eN,,eNe_{-N}, \dots, e_N it follows that SNffSNfS_Nf\perp f-S_N f.

Bessel’s inequality now follows from Pythagoras: since SNffSNfS_Nf\perp f-S_N f we have

f22=fSNf+SNf22=fSNf22+SNf22.\|f\|_2^2 = \|f-S_Nf+S_Nf\|_2^2 = \|f-S_Nf\|_2^2 + \|S_Nf\|_2^2.

Q.E.D.

Partial sums as best approximations. If g=NNg^kekg=\sum_{-N}^N \hat g_k e_k is a trigonometric polynomial of degree at most NN then

fSNf2fg2,\|f-S_Nf\|_2 \leq \|f-g\|_2,

with strict inequality if gSNfg\neq S_Nf.

Proof. First note that SNfgS_Nf-g is a linear combination of eN,,eNe_{-N}, \dots, e_N, and therefore that SNfgSNfS_Nf-g \perp S_N-f. Then apply Pythagoras to fg=fSNf+SNfgf-g = f-S_Nf+S_Nf-g, to get

fg22=fSNf22+SNfg22.\|f-g\|_2^2 = \|f-S_Nf\|_2^2 + \|S_Nf-g\|_2^2.

This implies fSNf2fg2\|f-S_Nf\|_2 \leq \|f-g\|_2, and if fSNf2=fg2\|f-S_Nf\|_2 = \|f-g\|_2 then SNfg2=0\|S_Nf-g\|_2=0.
Since SNfgS_Nf-g is a trigonometric polynomial it is a continuous function. Therefore SNfg2=0\|S_Nf-g\|_2=0 implies that Sfg=0S_f-g=0, i.e. SNf=gS_Nf=g.

Q.E.D.

Convergence in L2L^2 of Fourier series. Let fRperf\in\mathcal R_{\rm per}. Then

limNSNff2=0\lim_{N\to\infty} \|S_Nf-f\|_2 = 0

and

f22=12πnZf^n2.\|f\|_2^2 = \frac{1}{2\pi} \sum_{n\in\Z} |\hat f_n|^2.

This statement is known as the Plancherel or Parseval identity.

Proof. Let ϵ>0\epsilon\gt0 be given. Since ff is Riemann integrable, a C2C^2 function gg exists that is 2π2\pi–periodic, and that satisfies

02πf(x)g(x)2dxϵ24, i.e. fg2<ϵ2.\int_0^{2\pi} |f(x)-g(x)|^2dx\leq \frac{\epsilon^2}{4}, \text{ i.e. } \|f-g\|_2\lt \frac{\epsilon}{2}.

(details in lecture). Since gg is C2C^2, we know that the Fourier series of gg converges uniformly to gg, i.e.

SNgg0.\|S_Ng-g\|_\infty \to 0.

For any given ϵ2>0\epsilon_2\gt0 we can therefore find an NϵN_{\epsilon} such that SNgg<ϵ22π||S_Ng-g\|_\infty \lt \frac\epsilon{2\sqrt{2\pi}} for all NNϵ2N\geq N_{\epsilon_2}. This implies

SNgg22π  SNgg<ϵ2.\|S_Ng-g\|_2\leq \sqrt{2\pi} \; \|S_Ng-g\|_\infty \lt \frac{\epsilon}{2}.

We now have for all NNϵN\geq N_\epsilon

SNff2=SNfSNg+SNgg2SNfSNg2+SNgg2=SN(fg)2+SNgg2fg2+SNgg2<ϵ2+ϵ2=ϵ.\begin{aligned} \|S_Nf-f\|_2 &=\|S_Nf-S_Ng+S_Ng-g\|_2 \\ &\leq\|S_Nf-S_Ng\|_2 + \|S_Ng-g\|_2 \\ &=\|S_N(f-g)\|_2 + \|S_Ng-g\|_2 \\ &\leq\|f-g\|_2 + \|S_Ng-g\|_2 \\ &\lt \frac{\epsilon}{2} + \frac{\epsilon}{2} = \epsilon. \end{aligned}

The Parseval—Plancherel identity now follows by letting NN\to\infty in f22=SNf22+fSNf22\|f\|_2^2=\|S_Nf\|_2^2+\|f-S_Nf\|_2^2.

Q.E.D.

Problems

Consider the solutions u:R2Ru:\R^2\to\R to the wave equation described in (4).

Find functions F,G:RRF, G:\R\to\R such that u(x,t)=F(x+t)+G(xt)u(x, t)= F(x+t)+G(x-t) for all x,tx, t. (Hint: don’t use the derivation of d’Alembert’s solution but instead use eiθeiϕ=ei(θ+ϕ)e^{i\theta}e^{i\phi} = e^{i(\theta+\phi)} and take a good look at (4).

Various questions about the Dirichlet kernel and Fourier coefficients

(a) Show that 02πDN(s)  ds=1\int_0^{2\pi} D_N(s)\; ds =1.
(b) The kthk^{\rm th} Fourier coefficient f^k\hat f_k of a 2π2\pi–periodic function f:RCf:\R\to\mathbb{C} is defined as an integral from x=0x=0 to x=2πx=2\pi. Show that f^k=12πππeikxf(x)  dx\hat f_k = \frac{1}{2\pi}\int_{-\pi}^\pi e^{-ikx} f(x)\;dx.

Let f(x)=sinx2f(x)=|\sin \frac x2|.

(a) Compute the Fourier series of ff (the integrals simplify if you write sinx2\sin \frac{x}{2} in terms of complex exponentials.)
(b) For which xRx\in\R does the Fourier series converge to ff?
(c) What do you get if you set x=0x=0 in the Fourier expansion of ff?

The “plucked string” function

Let \wp be the 2π2\pi–periodic function that satisfies (x)=x\wp(x)= x for x[0,π]x\in[0,\pi], and (x)=2πx\wp(x)=2\pi-x for x[π,2π]x\in[\pi,2\pi]

(a) Compute the Fourier series of \wp
(b) For which xRx\in\R does the Fourier series converge to \wp?
(c) What do you get if you set x=0x=0 in the Fourier expansion of \wp?

Solving ordinary differential equations with Fourier series

(a) Let f:RCf:\R\to\mathbb C be a 2π2\pi–periodic function that is C1C^1. Show that if g=fg=f' then g^k=ikf^k\hat g_k = ik\hat f_k.
(b) Let f:RCf:\R\to\mathbb C be a 2π2\pi–periodic function that is C2C^2. Show that if g=fg=f'' then g^k=k2f^k\hat g_k = -k^2\hat f_k.
(c) Show that there exist numbers mkCm_k\in\mathbb{C} such that the following holds for any C2C^2 2π2\pi–periodic function f:RCf:\R\to\mathbb C: if f4f=gf''-4f=g then f^k=mkg^k\hat f_k = m_k\hat g_k.
(d) Find the Fourier series of the 2π2\pi–periodic solution f:RCf:\R\to\mathbb C of f(x)4f(x)=sin(x)f''(x) - 4f(x) = \sin (x) (0x2π0\leq x\leq 2\pi).
(e) Find the Fourier series of the 2π2\pi–periodic solution f:RCf:\R\to\mathbb C of f(x)4f(x)=(x)f''(x) - 4f(x) = \wp(x) (0x2π0\leq x\leq 2\pi) where \wp is the “plucked string” function from the previous problem

Let f:[0,2π]Rf:[0,2\pi]\to\R be Riemann integrable, and let f^k\hat f_k be the kthk^{\rm th} Fourier coefficient of ff.

(a) Show that f^kf|\hat f_k|\leq \|f\|_\infty for all kZk\in\Z.
(b) Show that limk±f^k=0\lim_{k\to\pm\infty} \hat f_k =0.
(c) Suppose that ff is mm times continuously differentiable. Show that f^kf(m)km|\hat f_k| \leq \frac{\|f^{(m)}\|_\infty}{|k|^m} for all k0.k\neq 0.

Let z:RRz:\R\to\R be the 2π2\pi–periodic “saw tooth function” given by z(x)=πx2z(x)=\frac{\pi-x}{2} for 0<x<2π0\lt x\lt 2\pi.

Compute z2\|z\|_2 and the Fourier coefficients of zz. Which identity do you find by applying the Plancherel—Parseval’s identity?