Language Reference

IFFT Function

IFFT (f);

The IFFT function computes the inverse finite Fourier transform of a matrix f, where f is an $np \times 2$ numeric matrix.

The IFFT function expands a set of sine and cosine coefficients into a sequence equal to the sum of the coefficients times the sine and cosine functions. The argument f is an $np \times 2$ matrix; the value returned is an $n \times 1$ vector.

If the element in the last row and second column of f is exactly 0, then n is $2np-2$; otherwise, n is $2np-1$.

The inverse finite Fourier transform of a two column matrix $\mb{F}$, denoted by the vector $\mb{x}$, is

\[ x_ i = F_{1,1} + 2 \sum _{j=2}^{np} \left( F_{j,1} \cos \left( \frac{2 \pi }{n} (j-1)(i-1) \right) + F_{j,2} \sin \left( \frac{2 \pi }{n} (j-1)(i-1) \right) \right) + q_ i \]

for $i=1,\ldots ,n$, where $q_ i = (-1)^ i \mb{F}_{np,1}$ if n is even, or $q=0$ if n is odd.

For the most efficient use of the IFFT function, n should be a power of 2. If n is a power of 2, a fast Fourier transform is used (Singleton, 1969); otherwise, a Chirp-Z algorithm is used (Monro and Branch, 1977).

The expression IFFT(FFT(X)) returns n times $\mb{x}$, where n is the dimension of $\mb{x}$. If f is not the Fourier transform of a real sequence, then the vector generated by the IFFT function is not a true inverse Fourier transform. However, applications exist in which the FFT function and the IFFT function can be used for operations on multidimensional or complex data (Gentleman and Sande, 1966; Nussbaumer, 1982).

As an example, the convolution of two vectors $\mb{x}$ ($n \times 1$) and $\mb{y}$ ($m \times 1$) can be accomplished by using the following module: 

start conv(u,v);
/* w = conv(u,v) convolves vectors u and v.
 * Algebraically, convolution is the same operation as
 * multiplying the polynomials whose coefficients are the
 * elements of u and v. Straight convolution is too slow,
 * so use the FFT.
 *
 * Both of u and v are column vectors.
 */
   m = nrow(u);
   n = nrow(v);

   wn = m + n - 1;
   /* find p so that 2##(p-1) < wn <= 2##p */
   p = ceil( log(wn)/ log(2) );
   nice = 2##p;

   a = fft( u // j(nice-m,1,0) );
   b = fft( v // j(nice-n,1,0) );
   /* complex multiplication of a and b */
   wReal = a[,1]#b[,1] - a[,2]#b[,2];
   wImag = a[,1]#b[,2] + a[,2]#b[,1];
   w = wReal || wImag;
   z=ifft(w);
   z = z[1:wn,1] / nice;  /* take real part and first wn elements */
   return (z);
finish;

/* example of convolution of two waveforms */
TimeStep = 0.01;
t = T( do(0,8,TimeStep) );

Signal = j(nrow(t),1,5);
Signal[ loc(t>4) ] = -5;

ImpulseResponse = j(nrow(t),1,0);
ImpulseResponse[ loc(t<=2) ] = 3;

/* The time domain for this convolution is [0,16]
   with the same time step.
   For waveforms, rescale amplitude by the time step. */
y = conv(Signal,ImpulseResponse) * TimeStep;

Other applications of the FFT and IFFT functions include windowed spectral estimates and the inverse autocorrelation function.