2
$\begingroup$

In d'Halluin et al. (2005) (http://imajna.oxfordjournals.org/content/25/1/87.short) the authors claim that the correlation integral $$ I(x) = \int_{-\infty}^{\infty} V(x+y) f(y) dy $$ can be approximated with FFT methods by $$ I_{k} = IFFT((FFT(V)_{k})(FFT(f))_{k}^{*}) $$ where $(.)^{*}$ denotes the complex conjugate.

I have tried to program this in Matlab with $V = \max(90 e^{y} - K,0)$ and $f$ being the probability density function of the normal distribution with $\mu = 0.1$ and $\sigma = 0.2$. Analytically I get the result $I(0) = 8.783$ but in whatever way I try the FFT approximation the results are nowhere close to that.

My math background is really weak and I don't have any intuition or knowledge regarding the FFT method. I calculated a set of values (powers of 2) for $V()$ and $f()$, did both element-wise and matrix multiplications of the FFTs of the two vectors and checked the real part of the IFFT, but I never get the desired result.

What am I doing wrong? Can someone tell me how I get to the result?

Here's the code of one of the ways I did it:

S = 90;
K = 100;           
mu = 0.1;
sig = 0.2;
AA = 2.5;
N = 255;             
x = (linspace(-AA,AA,N+1))';

V = max(S.*exp(x) - K,0);
ffff = normpdf(x,mu,sig);

I_k = real(ifft(fft(V).*conj(fft(ffff))));
$\endgroup$
4
  • 1
    $\begingroup$ We can't know what you did wrong, your code is not even in the post. Have you checked the Convolution theorem? $\endgroup$
    – nicoguaro
    Jun 14, 2016 at 15:14
  • $\begingroup$ okay, I added the code, but that's just one of the possibilities I have tried $\endgroup$
    – C. S.
    Jun 14, 2016 at 15:33
  • $\begingroup$ In your description, what is the index $k$ in the second formula? $\endgroup$ Jun 14, 2016 at 18:41
  • $\begingroup$ I'm not sure, but I guess it's the index for the discretized $x$, $x_{k}$ $\endgroup$
    – C. S.
    Jun 14, 2016 at 18:43

1 Answer 1

1
$\begingroup$

Even though this is a strict requirement, the authors explicitly mention that $x$ is equally spaced in the space of $log(S)$. They also generate it this way in the pseudocode. In that case $\bar{V}(log S) = V(S)$. I think you do not satisfy that. And in the paper, it is also the case that $\bar{f}$ is a probability density. This is just to warn in advance.

Their resolution to interpolation along non-equally spaced intervals comes at a later stage, where they explicitly interpolate the discrete values of the result onto the original $S$ grid.

So, theoretically, the equation is an altered cross-correlation, where $f$ is not a complex conjugate. Then in that case, the formula is correct (not an approximation) - with periodic border treatment of course.

$\endgroup$
1
  • $\begingroup$ I don't understand. Both the analytic and the FFT integral are in the same space and the $x$ for the FFT are equidistant. $\endgroup$
    – C. S.
    Jun 15, 2016 at 13:16

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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