1
$\begingroup$

I try to understand in which sense the Fourier transform is the continuous analogue of the discrete Fourier transform. I know, there are many books and many questions on this site concerning this topic, but almost all of them are written from an engineer's point of view. I'm a mathematician, and as such, I'm not familiar with the application of Fourier analysis in engineering. So, I would like to handle the following objects as mathematical objects without any physical interpretation.


Let me first introduce the relevant objects: Let $n\in\mathbb N$ and $\omega\in\mathbb C$ be a primitive $n$-th root of unity.

It's easy to verify that the columns of $$\tilde V_\omega:=\frac 1{\sqrt n}\underbrace{\begin{pmatrix} \omega^{0\cdot 0}&\cdots&\omega^{0\cdot (n-1)}\\\vdots&\ddots&\vdots\\\omega^{(n-1)\cdot 0}&\cdots&\omega^{(n-1)\cdot (n-1)} \end{pmatrix}}_{=:V_\omega}$$ are an orthonormal basis of $\mathbb C^n$.

Now, $$\operatorname{DFT}_\omega:\mathbb C^n\to\mathbb C^n\;,\;\;\;z\mapsto\left(f_z\left(\omega^0\right),\ldots,f_z\left(\omega^{n-1}\right)\right)^T$$ with $$f_z(w):=\sum_{k=0}^{n-1}z_kw^k\;\;\;\text{for }z=\left(z_0,\ldots,z_{n-1}\right)^T\text{ and }w\in\mathbb C$$ is called discrete Fourier transform.

Again, it's easy to verify that $V_\omega$ is the transformation matrix of $\operatorname{DFT}_\omega$, i.e. $$\operatorname{DFT}_\omega z=V_\omega z\color{blue}{\stackrel{\text{def}}=\sqrt n\tilde V_\omega z}\;\;\;\text{for all }z\in\mathbb C^n\;,$$ and is invertible with $$V_\omega^{-1}=\frac 1nV_{\omega^{-1}}$$ and hence $\operatorname{DFT}_\omega$ is an isomorphism. We can easily prove that $$\langle\operatorname{DFT}_\omega z,\operatorname{DFT}_\omega w\rangle=n\langle z,w\rangle\;\;\;\text{for all }z,w\in\mathbb C^n$$ (using that $\tilde V_\omega$ is unitary).


Now, let $a,b\in\mathbb R$ with $a<b$ and $D:=(a,b)$.

It's a well-known fact that $$\phi_k(x):=\frac 1{\sqrt{b-a}}e^{2\pi{\rm i}\frac k{b-a}x}\;\;\;\text{for }k\in\mathbb Z\text{ and }x\in D$$ is an orthonormal basis of $L^2(D;\mathbb C)$ and hence $$u=\sum_{k\in\mathbb Z}\underbrace{\langle u,\phi_k\rangle_{L^2(D;\mathbb C)}}_{=:u_k}\phi_k\;\;\;\text{for all }u\in L^2(D;\mathbb C)\;.$$

Now, $$\sqrt{b-a}\sum_{k\in\mathbb Z}\tilde u_k\phi_k$$ with $$\tilde u_k:=\frac 1{\sqrt{b-a}}u_k\;\;\;\text{for }k\in\mathbb Z$$ is called Fourier series of $u\in L^2(D;\mathbb C)$.

Let $$\mathcal F:L^2(D;\mathbb C)\to L^2(\mathbb Z;\mathbb C)\;,\;\;\;u\mapsto(\tilde u_k)_{k\in\mathbb Z}\;.$$ We can show that $$\langle\mathcal Fu,\mathcal Fv\rangle_{L^2(D;\mathbb C)}=(b-a)\langle(\tilde u_k)_{k\in\mathbb Z},(\tilde v_k)_{k\in\mathbb Z}\rangle_{L^2(\mathbb Z;\mathbb C)}\;\;\;\text{for all }u,v\in L^2(D;\mathbb C)\;.$$


Is $\mathcal F$ what we call the Fourier transform? How are $\operatorname{DFT}_\omega$ and $\mathcal F$ related?

I would like to conclude that the discrete Fourier transform is an isomorphic isometry $(\mathbb C^n,\langle\;\cdot\;,\;\cdot\;\rangle)\to(\mathbb C^n,n\langle\;\cdot\;,\;\cdot\;\rangle)$ and that an analogue statement holds true for the Fourier transform on $(L^2(D;\mathbb C),\langle\;\cdot\;,\;\cdot\;\rangle_{L^2(D;\mathbb C)})$.

Please keep this in mind, when you formulate an answer. I suppose $\mathcal F$ is not the mapping I'm looking for, cause its nature seems to be different $\operatorname{DFT}_\omega$. I guess the mapping should be $(L^2(D;\mathbb C)\to (L^2(D;\mathbb C)$ where the image is equipped with an inner product of the form $c\langle\;\cdot\;,\;\cdot\;\rangle_{L^2(D;\mathbb C)}$ for some $c\in\mathbb C$.

$\endgroup$
  • $\begingroup$ It might be worth to note that $$\omega:=e^{-{\rm i}\frac{2\pi}n}$$ is somehow the unique primitive $n$-th root of unity. It's worth to note, cause the form of $\omega$ is similar to the form of the $\phi_k$. $\endgroup$ – 0xbadf00d Feb 28 '16 at 11:46
  • $\begingroup$ google.com/url?sa=t&source=web&rct=j&url=http://… Here you go. A signal that is discrete in one domain is continuous in the other. $\endgroup$ – D J Sims Feb 28 '16 at 11:48
  • $\begingroup$ start by showing how the Fourier series is the limiting case when $N \to \infty$ of the discrete Fourier transform on $\mathbb{C}^N$ $\endgroup$ – reuns Feb 28 '16 at 12:15
  • $\begingroup$ @Mustang Your link doesn't contain anything new (something which is not part of the statements in the question), does it? I suppose you mean the following: The "continuous signal" $u\in L^2(D;\mathbb C)$ is mapped to the "discrete signal" $(\tilde u_k)_{k\in\mathbb Z}$ by $\mathcal F$. Is that what you've meant? $\endgroup$ – 0xbadf00d Feb 28 '16 at 18:06
  • $\begingroup$ @user1952009 I've no idea how to start. $\endgroup$ – 0xbadf00d Feb 28 '16 at 19:10
1
$\begingroup$

The operator $\frac{1}{i}\frac{d}{dt}$ considered on the domain of absolutely continuous functions $f \in L^2[-\pi,\pi]$ for which $f'\in L^2$ and $f(0)=f(2\pi)$ is selfadjoint, and the map $U : L^2[-\pi,\pi]\rightarrow\ell^2(\mathbb{Z})$ defined by $$ Uf = \{ \hat{f}(n) \}_{n=-\infty}^{\infty} $$ is a unitary map such that $$ U\frac{1}{i}\frac{d}{dx}f = \{n\hat{f}(n)\}=M\{\hat{f}(n)\}, $$ where $M$ is the multiplication operator on $\ell^2(\mathbb{Z})$ $$ M \{ a_n \} = \{ n a_n \}. $$ So the discrete Fourier transform is a unitary map that turns differentiation into multiplication.

The operator $\frac{1}{i}\frac{d}{dt}$ considered on the domain consisting of all absolutely continuous $f\in L^2(\mathbb{R})$ for which $f'\in L^2(\mathbb{R})$ is selfadjoint, and the Fourier transform $\mathcal{F} : L^2(\mathbb{R})\rightarrow L^2(\mathbb{R})$ is a unitary map such that $$ \mathcal{F}\frac{1}{i}\frac{d}{dx}f=s(\mathcal{F}f)(s)=M\mathcal{F}f $$ where $M$ is the multiplication function on $L^2(\mathbb{R})$ given by $$ (Mf)(s) = sf(s). $$ So the Fourier continuous Fourier transform is a unitary map that turns differentiation into multiplication.

Matrix diagonalization works the same way on a finite-dimensional Hilbert space $H$. If $A$ is selfadjoint with an orthonormal basis of eigenvectors $\{ e_n \}$ with corresponding eigenvalues $\{ \lambda_n \}$, then the map $$ U : x \in H \mapsto \{ (x,e_n) \}_{n=1}^{N}\in\mathbb{C}^{N} $$ is a unitary map that turns $A$ into multiplication on $\ell^2(\{1,2,\cdots,N\})=\mathbb{C}^{N}$: $$ UAx = \{ \lambda_n (x,e_n)\} = M\{ (x,e_n) \} $$ where $M$ is the multiplication operator on $\ell^2(\{1,2,3,\cdots,N\})$ given by $$ M\{ a_n \}_{n=1}^{N} = \{ \lambda_n a_n \}_{n=1}^{N}. $$

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.