I'm trying to compute the analytical FT of a bidimensional exponential decay as follows:

$f_E(\vec{x}) = \exp \left( - \sqrt{ (\vec{x} - \vec{\mu})^T \Sigma^{-1} (\vec{x} - \vec{\mu}) } \right)$

with $\Sigma^{-1} \in \mathbb{R}^{2 \times 2}$ a symmetric positive-definite matrix, with the following code:

A = {{a, b}, {b, c}}
x = {x1, x2}
μ = {μ1, μ2}
z = {z1, z2}
FourierTransform[Exp[-Sqrt[(x - μ).A.(x - μ)]], x, z]

But the notebook's kernel keeps running and dies after a while. I also tried using Assumptions to make explicit that coefficients involved are real and A is positive-definite with no luck. Also tried

M = {{d, e}, {f, g}}
FourierTransform[Exp[-Norm[M.(x - μ)]], x, z]

I'm new to Mathematica so I don't know if there's an error with the code, a conceptual (mathematical) mistake or if it is just related to computing resources.


I don't know why Mathematica does not get it right. At least, it can compute

FourierTransform[Exp[-Sqrt[x.x]], x, z]

1/(1 + z1^2 + z2^2)^(3/2)

From there we can continue on paper:

Let $\varPhi \colon \mathbb{R}^n \to \mathbb{R}^n$ be a diffeomorpism of class at least $C^1$ and denote by $\mathcal{F}$ the Fourier transform. By the transformation formular for integrals (and by certain denseness and continuous extension arguments) one obtains

$$\begin{aligned}\mathcal{F}(f \circ \varPhi)(\xi) &= C \int_{\mathbb{R}^n} f(\varPhi(x)) \, \exp(- \mathrm{i} \, \langle\xi , x \rangle) \, \mathrm{d} x\\&= C \int_{\mathbb{R}^n} f(y) \, \exp(- \mathrm{i} \, \langle\xi , \varPhi^{-1}(y) \rangle) \, |\det(D\varPhi(y))|^{-1}\, \mathrm{d} y\end{aligned}$$

with a constant $C$ that depends on some convention; the precise value won't matter in the end.

In case of an affine mapping $\varPhi(x) = L \, (x - \mu) = L\, x - L \, \mu$ with $L^T \, L =A$, we obtain $|\det(D\varPhi(y))| = \sqrt{\det(A)}$. Thus, assuming that $A$ is positive definite, we obtain

$$\begin{aligned}\mathcal{F}(f \circ \varPhi)(\xi) &= C\,\det(A)^{-\frac{1}{2}} \int_{\mathbb{R}^n} f(y) \, \exp(- \mathrm{i} \, \langle\xi , L^{-1}\,y + \mu \rangle) \, \mathrm{d} y\\&= C \, \det(A)^{-\frac{1}{2}} \, \exp(- \mathrm{i} \, \langle\xi , \mu \rangle)\int_{\mathbb{R}^n} f(y) \, \exp(- \mathrm{i} \, \langle L^{-T}\xi , y\rangle) \, \mathrm{d} y\\&= \det(A)^{-\frac{1}{2}} \, \exp(- \mathrm{i} \, \langle\xi , \mu \rangle) \,\mathcal{F}(f)(L^{-T}\xi).\end{aligned}.$$

With $f(x) = \exp(-\sqrt{\langle x, x\rangle})$ and $$\mathcal{F}(f)(\xi) = (1+\langle \xi, \xi\rangle)^{-\frac{3}{2}},$$ we should obtain $$\mathcal{F}(f \circ \varPhi)(\xi) = \det(A)^{-\frac{1}{2}} \, \exp(- \mathrm{i} \, \langle\xi , \mu \rangle) \, (1+\langle \xi, A^{-1}\xi\rangle)^{-\frac{3}{2}}.$$

But of course, I could have made a mistake. Please check thoroughly.

  • $\begingroup$ I have gone over this a few times now and it checks out, there's just a little typo, inside the sqrt of f(x) should be <x,x>. Thanks! $\endgroup$ – vlizana Feb 21 '19 at 21:01
  • $\begingroup$ You're welcome! I am glad that this was helpful. $\endgroup$ – Henrik Schumacher Feb 21 '19 at 21:03

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.