5
$\begingroup$

Can someone help me in a step-by-step derivation for the Fourier Transform of a line ? It appears to be simple but still cannot figure out. I know what is the end result but I am unable to figure out the intermediate steps. I tried to use the Shifting and the Similarity theorems but I am not getting the final result.

We have a line $Ax+By+C=0$ which can be called a line impulse $\delta(Ax+By+C)$ where $\delta$ is the delta function.

We would like to derive the Fourier transform $\mathcal{F}$ of this line impulse which is mentioned in $[1]$ to be:

$$\mathcal{F}(\delta(Ax+By+C))(u,v) = e^{i2 \pi C (Au+Bv)} \delta(-Bu+Av)$$

where $u,v$ are the frequency components in two dimensions.

$[1]$: Line Segment Sampling with Blue-Noise Properties By Xin Sun et al 2013

$\endgroup$
1
$\begingroup$

The result seems to be only valid for $A^2 + B^2= 1$.

first approach (brute force):

Let us simply start with the definition $$\mathcal{F}(\delta(Ax+By+C))(u,v) = \int dx\! \int dy\, \delta (A x + B y +C) e^{-2\pi i (x u + y v)}. \tag{F}$$

In a first step, we perform the integration over $y$. Note that $$\delta (A x + By +C) = \delta ( B (y + C/B +A x /B)) = \frac{1}{|B|} \delta(y + C/B +A x /B)$$ (as $\delta(\alpha x) = |\alpha|^{-1} \delta(x)$) and therefore $$\begin{align}\int dy\,\delta (A x + B y +C) e^{-2\pi i (x u + y v)} &= \frac1{|B|} \exp[-2\pi i ( x u -(C/B +A x /B) v)]\\ &= \frac1{|B|} \exp[ -2\pi i x (u - A v/B) + 2\pi i C v/B].\end{align} $$

We are left with the integral over $x$. Using the standard result $$\int\! dx\, e^{-2\pi i x \alpha} = \delta(\alpha),$$ we obtain $$\mathcal{F}(\delta(Ax+By+C))(u,v) = \frac1{|B|} e^{2\pi i C v/B} \delta(u - A v/B).$$ We have $|B|^{-1} \delta(u - A v/B) = \delta(B u - A v)$. Because of the $\delta$-function, we can use $ B u = Av$ in the prefactor. Thus, we can "symmetries" ($g$ is an arbitrary function) $$\frac{1}{|B|} g(v/B) \delta(u- Av /B) = g(v/B) \delta(B u -A v)=g[(v/B)(A^2 +B^2) /(A^2 + B^2)]\delta(B u -A v)=g[(A (Av/B) + B v) /(A^2 + B^2)]\delta(B u -A v) = g[(A u + B v) /(A^2 + B^2)] \delta(B u -A v). $$

We obtain the final result $$ \mathcal{F}(\delta(Ax+By+C))(u,v) = \exp[2\pi i C (A u + B v) /(A^2 + B^2)] \delta(B u - A v). \tag{res}$$

second approach:

To make maximal use of the $\delta$-function, we would like to introduce a new variable $s= Ax + By + C$ in (F). We want to use a second variable whose coordinate axis are perpendicular, so we choose $t= ( Ay -B x)/(A^2+B^2)$. The factor is chosen such as to make the Jacobian for the variable change from $(x,y)$ to $(s,t)$ unity.

A simple calculation shows that $$ x u + y v =(s -C) (A u + B v)/(A^2+B^2) + t (-Bu + Av) . $$ So we have $$\mathcal{F}(\delta(Ax+By+C))(u,v) = \int \!dt \int\! ds\, \delta(s) \exp\{-2 \pi i[(s -C) (A u + B v)/(A^2+B^2) + t (-Bu + Av)] \}.$$

The integral over $s$ is readily performed and we have $$\mathcal{F}(\delta(Ax+By+C))(u,v) = \int \!dt \exp\{2 \pi i C (A u + B v) /(A^2+B^2) - 2\pi i t (-Bu + Av) ] \}.$$

The integral over $t$ leads to a $\delta$-function $$\int\!dt \exp\{-2 \pi i t (-Bu + Av) \} = \delta(-Bu +Av).$$ And we end up with the result (res) quoted above.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Thanks a lot @Fabian for providing a step by step derivation. I have just one question in the burte-force appraoch. You mentioned that we can use Bu=Av in the prefactor, I don't really understand this step. Can you explain how you symmetrize here ? $\endgroup$ – sinner Jan 12 '16 at 2:14
  • $\begingroup$ Because of the $\delta$-function, we can replace $Bu$ with $Av$ everywhere in the factor multiplying the $\delta$ function. I added a few steps in the derivation for clarification. $\endgroup$ – Fabian Jan 12 '16 at 21:19
  • $\begingroup$ Is this result a line ? $\endgroup$ – haleh Apr 17 '17 at 13:24
0
$\begingroup$

Here I'll give an answer for arbitrary dimension.

For $\vec{x},\vec{w}_1\in \mathbb{R}^n$, what is the Fourier transform along the line $\vec{w}_1\cdot\vec{x} + c = 0$? Assume $\vec{w}_1$ is normalized.

Using the dirac delta to define the line in $\mathbb{R}^n$ we therefore must calculate,

\begin{align} \mathcal{F}[\delta(\vec{w}_1\cdot\vec{x} + c)](\vec{\nu}) = \int d\vec{x} \, \delta(\vec{w}_1\cdot\vec{x} + c) \exp{[-2\pi i \vec{x}\cdot\vec{\nu}]},\tag{1} \end{align} where $\vec{\nu}$ are the fourier space variables.

Let $\{\vec{w}_2,...,\vec{w}_n\}$ be the orthonormal basis that spans $\mathrm{Null}[\vec{w}_1]$. You can determine this for any dimension with the Wolfram command NullSpace, e.g. for 3D.

Define a change of variables by, \begin{align} s_1 =& \vec{w}_1\cdot\vec{x} + c\\ s_2 =& \vec{w}_2\cdot\vec{x}\\ &...\\ s_n =& \vec{w}_n\cdot\vec{x}. \end{align}

Written more precisely, \begin{align} \vec{s} =& W \vec{x} + \hat{s}_1 c\\ I =& W W^T. \tag{2} \end{align} The second line comes from the orthonmality of the basis of choosen to span the null space of $\vec{w}_1$. Note: The determinant of an orthogonal matrix satisfying (2) is $1$, and represents a rotation matrix, and therefore the Jacobian of the transform is $1$.

With these new variables equation (1) becomes, \begin{align} \mathcal{F}[\delta(s_1)](\vec{\nu}) = \int d\vec{s} \, \delta(s_1) \exp{[-2\pi i W^T(\vec{s}-\hat{s}_1 c)\cdot\vec{\nu}]}. \end{align} Now we carry out the integral over $s_1$ by using $\vec{s} = \sum_{i=1}^n s_i \hat{s}_i$, \begin{align} \mathcal{F}[\delta(s_1)](\vec{\nu}) =& \int ds_n ... \int ds_1 \, \delta(s_1)\exp{[-2\pi i W^T(\sum_{i=1}^n s_i \hat{s}_i-\hat{s}_1 c)\cdot\vec{\nu}]}\\ =& \exp{[2\pi i c(W^T\hat{s}_1)\cdot\vec{\nu}]} \int ds_n ... \int ds_2 \, \exp{[-2\pi i \sum_{i=2}^n s_i (W^T\hat{s}_i)\cdot\vec{\nu}]}. \end{align} Performing the remaining integrations using the dirac delta definition $\delta(y) = \int dp\, \exp[-2\pi i p y]$, we get, \begin{align} \mathcal{F}[\delta(s_1)](\vec{\nu}) =& \exp{[2\pi i c(W^T\hat{s}_1)\cdot\vec{\nu}]} \prod_{i=2}^n \delta(W^T\hat{s}_i\cdot\vec{\nu}). \end{align}

In order to evaluate $W^T\hat{s}_i$, note the basis vectors in the rotated space are related to the original basis vectors via the same rotation matrix, i.e. $[\hat{s}_1, ..., \hat{s}_n] = W [\hat{x}_1, ..., \hat{x}_n]$. Therefore $W^T\hat{s}_i = \hat{x}_i = \hat{x}_i (\vec{w}_1\wedge\vec{w}_1 + ... + \vec{w}_n\wedge\vec{w}_n)$.

| cite | improve this answer | |
$\endgroup$

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.