Skip to content

Inner Product Space and Fourier Series

Inner Products Spaces

An inner product ,:VR\langle \cdot, \cdot\rangle :V\rightarrow \mathbb R is a function that satisfies

  • positive definite x,x>0\langle x,x\rangle > 0 and x,x=0\langle x,x\rangle =0 IFF x0x\equiv0
  • symmetry x,y=y,x\langle x,y\rangle =\langle y,x\rangle
  • bilinear ax+cy,z=ax,z+cy,z\langle ax+cy,z\rangle =a\langle x,z\rangle +c\langle y,z\rangle

Example 1

On Rn,xy=1nxiyi\mathbb R^n, x\cdot y = \sum_1^n x_i y_i

  • xx=nxi20x\cdot x =\sum^n x_i^2 \geq 0
  • xy=xiyi=yixi=yxx\cdot y = \sum x_i y_i = \sum y_i x_i = y\cdot x
  • (ax+cy)z=(axi+cyi)zi=axizi+cyizi=axz+cyz(ax+cy)\cdot z = \sum (ax_i + cy_i) z_i = a\sum x_i z_i + c\sum y_i z_i = ax\cdot z + cy\cdot z

Example 2

Let An×nA_{n\times n} be a positive definite symmetric matrix, i.e. all its eigenvalues are strictly positive. Then xyA=xAyT\langle x\cdot y\rangle _A = x A y^T is an inner product on Rn\mathbb R^n

Note that such AA can be diagonalized into A=PDPTA=PDP^T

  • x,xA=xAxT=xPDPxT=vDvT=vk2λk0\langle x,x\rangle _A = xAx^T = xPDPx^T = vDv^T = \sum v_k^2 \lambda_k \rangle 0 (since AA is definite, λk0\lambda_k \rangle 0)
  • x,x=0\langle x,x\rangle =0 IFF vk2λk=0vk=0xP=0\sum v_k^2 \lambda_k = 0\Rightarrow v_k = 0 \Rightarrow xP = 0, since PP is orthogonal, x=0x=0

Thrm 1. Cauchy Schwarz inequality

For x,y(V,,),x,yx,x2y,y2x,y\in (V,\langle ,\rangle ), |\langle x,y\rangle |\leq \langle x,x\rangle ^2\langle y,y\rangle ^2

proof. By positive definite, for tRt\in\mathbb R, xty,xty0\langle x-ty, x-ty\rangle \geq 0.
Take t=x,yy,yt = \frac{\langle x,y\rangle }{y,y},

x,x2tx,y+t2y,y=x,x2x,y2y,y+x,y2y,y=x,xx,y2x,y0    x,y2x,xy,y\begin{align*} \langle x,x\rangle - 2t |\langle x,y\rangle | + t^2 \langle y,y\rangle &= \langle x,x\rangle - 2\frac{\langle x,y\rangle ^2}{\langle y,y\rangle} + \frac{\langle x,y\rangle ^2}{\langle y,y\rangle }\\ =&\langle x,x\rangle - \frac{\langle x,y\rangle ^2}{\langle x,y\rangle } \geq 0\\ \implies &\langle x,y\rangle ^2 \leq \langle x,x\rangle \langle y,y\rangle\\ \end{align*}

Example 3

Over C[a,b]C[a,b], f,gL2=abfgdx\langle f,g\rangle _{L^2} = \int_a^b f g dx

  • f,fL2=abf2dx0\langle f,f\rangle _{L^2} = \int_a^b f^2 dx \geq 0
  • f,fL2=0\langle f,f\rangle _{L^2}=0 IFF abf2dx=0\int_a^b |f|^2 dx = 0 IFF f(x)=0f(x)=0 because ff is continuous.

Assume f(x0)>0f(x_0) > 0, then by continuity, take Iδ(x0):=[x0δ,x0+δ],xIδ(x0).f(x)>ϵ>0I_\delta(x_0):=[x_0-\delta, x_0+\delta], \forall x\in I_\delta(x_0). f(x) > \epsilon > 0 so that abf2dx>Iδ(x0)f(x)2dx>ϵ2\int_a^b |f|^2 dx > \int_{I_\delta(x_0)} |f(x)|^2 dx > \epsilon^2 contradicts with f,fL2=0\langle f,f\rangle _{L^2}=0 (other details to be filled)

By Cauchy Schwartz inequality,

fgf2g2    f,gL2fL2gL2    fL2([a,b])    fL1([a,b])\begin{align*} &\int fg\leq \sqrt{\int f^2}\sqrt{\int g^2}\\ \implies &\langle f,g\rangle _{L^2} \leq \|f\|_{L^2}\|g\|_{L^2}\\ \implies &f\in L^2([a,b])\\ \implies &f\in L^1([a,b])\\ \end{align*}

Therefore, every inner product gives rise to a norm xv:=x,x\|x\|_v := \sqrt{\langle x,x\rangle }

Def'n. Hilbert space

a complete inner product space is called Hilbert space

Completeness over inner product space A inner product space is complete if every Cauchy sequence xnx_n has a limit in the space. i.e.

xnxV2=xnx,xnx0\|x_n - x\|_V^2 = \langle x_n - x, x_n - x\rangle \rightarrow 0
L2([a,b]):={f:[a,b]R:abf2}L^2([a,b]):= \{f:[a,b]\rightarrow \mathbb R: \int_a^b |f|^2 \langle \infty\}

Remark The space S:=(C[a,b],L2)S:=(C[a,b], \|\cdot\|_{L^2}) is not complete, but L2([a,b])L^2([a,b]) is the completion of SS. which means taking any Cauchy sequence, then it has a limit in L2L^2. In another words, continuous functions can approximate any square integrable function in L2L^2-norm.

Thrm 2. Monotone Convergence Theorem

Take nonnegative fn(x)0f_n(x)\geq 0, if f1(x)f2(x)f_1(x)\leq f_2(x) \leq \cdots then point-wise limit fn(x)f(x)f_n(x)\rightarrow f(x) i.e. fn(x)f(x)0|f_n(x)-f(x)|\rightarrow 0

Corollary By MCT, limfn=limfn=f\lim \int f_n = \int \lim f_n = \int f

Thrm 3. Dominated/Integral Convergence Theorem

If {hn}\{h_n\} satisfy

  • hn(x)B(x)|h_n(x)|\leq B(x), which BB is an integrable function
  • hn(x)point-wiseh(x)h_n(x)\rightarrow^{\text{point-wise}} h(x)

Then, limfn=limfn=f\lim \int f_n = \int \lim f_n = \int f

Thrm 4. Completeness

The space (L2,L2)(L^2, \|\cdot\|_{L^2}) is complete

proof. Take Cauchy sequence fnL2([a,b])f_n \in L^2([a,b]). By Cauchy, take subsequence fn,kf_{n,k} s.t. fn,kfn,k+12k\|f_{n,k} - f_{n,k+1}\| \leq 2^{-k}.
Let f(x):=fn,1(x)+(fn,k+1fn,k(x))f(x):=f_{n,1}(x) + \sum (f_{n, k+1} - f_{n,k}(x)). WTS fL2(i)f\in L^2 (i), fn,kL2f(ii)f_{n,k}\rightarrow^{L^2} f (ii).
Define g:=fn,1+fn,k+1fn,kg:= |f_n,1| + \sum|f_{n,k+1}- f_{n,k}|, SM(g)=fn,1+Mfn,k+1fn,kS_M(g) = |f_n,1| + \sum^M|f_{n,k+1}- f_{n,k}|, then fg,SM(f)SM(g)f\leq g, S_M(f) \leq S_M(g)
Apply MCT to {SM(g)},0fn,1fn,2fn,1...,S1(g)S2(g)S3(g)\{S_M(g)\}, 0\leq |f_{n,1}| \leq |f_{n,2}-f_{n,1}|\leq ..., S_1(g) \leq S_2(g) \leq S_3(g)
Then, WTS SM(g)g(x)S_M(g)\rightarrow g(x), which suffices to show that the sequence gg converges, a.k.a. gL2g\in L^2

SM(g)L2fn,1L2+Mfn,k+1fn,kfn,1+M2ktri. ineq.<\begin{align*} \|S_M(g)\|_{L^2} &\leq \|f_{n,1}\|_{}L^2 + \sum^M \|f_{n,k+1} - f_{n,k}\| \\ &\leq \|f_{n,1}\| + \sum^M 2^{-k} &\text{tri. ineq.}\\ &<\infty \end{align*}

Thus we have that

g2dx=limMSM(g)2dx=limnSM(g)2dxMCTlimMfn,1+M2kfn,1+2<\begin{align*} \int |g^2| dx &= \int \lim_{M\rightarrow \infty} |S_M(g)^2|dx\\ &=\lim_{n\rightarrow\infty} \int |S_M(g)|^2 dx &\text{MCT}\\ &\leq \lim_{M\rightarrow \infty} \|f_{n,1}\| + \sum^M 2^{-k} \leq \|f_{n,1}\| + 2 \\ &< \infty \end{align*}

We have Cauchy seuqnce fnL2f_n \in L^2, WTS fnfL20\|f_n -f\|_{L^2}\rightarrow 0

(i) Since fg|f|\leq g and gL2,f2g2<    fL2g\in L^2, \int |f|^2 \leq \int |g|^2 < \infty\implies f\in L^2

(ii) fnfL22=fnf2dx0\|f_n - f\|_{L^2}^2 = \int |f_n - f|^2 dx\rightarrow 0.

Let hk:=fn,kf2h_k:=|f_{n,k}- f|^2, then

hkSk(g)+g2gL2h_k \leq |S_k(g)| + |g| \leq 2|g|\in L^2

so that fn,kp.w.ff_{n,k}\rightarrow^{\text{p.w.}} f since gL2g\in L^2 so limkfn,k(x)=f(x)\lim_{k\rightarrow \infty} f_{n,k}(x)=f(x) exists
BY DCT, limkhk2=limkhk2=0\lim_{k\rightarrow \infty} \int |h_k|^2 = \int \lim_{k\rightarrow \infty} |h_k|^2 =0,

Therefore, fn,kL2ff_{n,k}\rightarrow^{L^2}f

Then, expansion fn,kf_{n,k} to fnf_n. i.e. fnL2ff_n \rightarrow^{L^2} f.
By Cauchy sequence of fnf_n, if the sub-sequence converges to some limit, so that the whole sequence will do so.

Fourier Series

Def'n. Orthogonal

Given x,y in(V,,)x,y\ in (V,\langle\cdot, \cdot\rangle), x,y are orthogonal if x,y=0\langle x,y\rangle = 0

Example

xy=cos(θ)xyx\circ y = \cos(\theta) |x||y|, so when θ=π/2,xy=0\theta = \pi /2, x\circ y = 0

Def'n. Orthogonal set

A collection of {ek}\{e_k\} is called orthonormal if ek,em=I(k=m)\langle e_k,e_m\rangle = \mathbb I(k=m)

Thrm 4.

Every Hilbert space has orthonormal basis, i.e. xH.x=x,ekek\forall x\in H. x = \sum^\infty \langle x,e_k\rangle e_k

Example

Take ek=(0,...,0,1,0)e_k = (0,...,0,1,0) where 11 is at kkth position, {ek}\{e_k\} is orthonormal set.

proof. ekem=n(ek)i(em)i=1I(k=m)e_k \cdot e_m = \sum^n (e_k)_i (e_m)_i = 1\mathbb I(k=m)

Example

{1,2cos(nθ),2sin(nθ)}\{1, \sqrt2\cos(n\theta), \sqrt2\sin(n\theta)\} is orthogonal for (C([π,π]),L2)(C([-\pi, \pi]), \|\cdot\|_{L^2})

proof. We have the followings hold

2πππcos(nθ)cos(kθ)=1I(n=k)2πππsin(nθ)sin(kθ)=1I(n=k)2πππsin(nθ)cos(kθ)=0\begin{align*} &\frac{\sqrt2}{\pi}\int_{-\pi}^\pi \cos(n\theta)\cos(k\theta) = 1\mathbb I(n=k)\\ &\frac{\sqrt2}{\pi}\int_{-\pi}^\pi \sin(n\theta)\sin(k\theta) = 1\mathbb I(n=k)\\ &\frac{\sqrt2}{\pi}\int_{-\pi}^\pi \sin(n\theta)\cos(k\theta) = 0 \end{align*}

Def'n. Fourier Series

For function fL2([a,b])f\in L^2([a,b]), the Fourier series is defined to be

f(x):=a0+ancos(nπx)L+bnsin(nπx)Lf(x):= a_0 + \sum a_n \frac{\cos(n\pi x)}{L} + b_n \frac{\sin(n\pi x)}{L}

where

an=f,cos(nπx)L)L2,bn=f,sin(nπx)LL2a_n = \langle f, \frac{\cos(n\pi x)}{L})\rangle _{L^2}, b_n = \langle f, \frac{\sin(n\pi x)}{L}\rangle _{L^2}

Thrm. Carleson Hunt Theorem

If fL2([L,L])f\in L^2([-L,L]), then ff equals its Fourier series almost everywhere.
almost everywhere outside a measure zero set