Skip to content

Fast Fourier Transform

Discrete Fourier Transform (DFT)

Given an 1D vector \(v\in\mathbb R^m\), the 1D DFT is \(\mathcal F\cdot v\) where \(\mathcal F\in\mathbb R^{m\times m}\) is defined as

\[F_{j,k} = \omega^{j\times k}, 0\leq j,k\leq m - 1\]

where \(\omega\) is a complex number whose \(m\)'s power \(\omega^m = 1\)

\[\omega = e^{2\pi\frac{i}{m}} = \cos(\frac{2\pi}{m}) + i\sin(\frac{2\pi}{m})\]

2D DFT of an \(m\times m\) matrix \(V\) is \(\mathcal F V \mathcal F\), which is equivalent to do 1D DFT on all columns independently, and then all the rows.

Applications

FFT is often used in frequency filtering (convolution theorem), compressions (same idea), and interestingly, solving Poisson's equations.

The Poisson's problems can be written as solving \(L_1X + XL_1 = B\) and note that 2D FFT works similar to Eigen-decomposition, actually we have that \(L_1 = F \Lambda F^{-1}\) being a eigen decomposition where

\[F_{j,k} = \sqrt{\frac{2}{m+1}} \sin(\pi\frac{jk}{m+1}), \Lambda_j = 2(1 - \cos(\pi \frac{j}{m+1}))\]

we can substitute \(L_1 = F\Lambda F^T\) and transform the RHS \(B\) by \(B = FB'F^T\), and then we can solve the whole thing as

\[\begin{align*} F \Lambda F^{-1} X + X F D F^T &= FB'F^T\\ F(\Lambda (F^TXF) + (F^TXF)D)F^T &= FB'F^T\\ DX' + X'D &= B' &X'_{jk} = \frac{B'_{jk}}{\Lambda_j + \Lambda_k}\\ X &= FX'F^T \end{align*}\]

Thus, the total cost is 2 2D FFT , \(m^2\) adds, and \(m^2\) divisions. all of which in \(O(m^2\log m)\)

Serial Algorithms

For each entry of FFT \(F\cdot v\),

\[(F\cdot v)_j = \sum_{k=0}^{m-1} F_{j,k} v_k = V(\omega^j), V(x) = \sum_{k=0}^{m-1} x^k v_k\]

Therefore, FFT is the same as evaluating a degree \(m-1\) polynomial \(V(x)\) at \(m\) different points.

Divide and Conquer FFT

Note that we can decompose \(V\) into even and odd components

\[\begin{align*} V(x) &= \sum_{k=0}^{m-1} x^k v_k\\ &= (v_0 + x^2 v_2 + x^4 v_4 + \cdots) + x(v_1 + x^2 v_3 + x^4 v_5 + \cdots)\\ &= V_{even}(x^2) + V_{odd}(x^2) \end{align*}\]

Each \(V_{even}\) and \(V_{odd}\) has degree \(\frac{m}{2}-1\), and we note that \((\omega^{j+\frac{m}{2}})^2 = \omega^{2j}\omega^m = (\omega^j)^2\), hence we are only evaluating \(m/2\) different points.

Therefore, we have the algorithm

FFT_DnC
"""
Assume that m is power of 2
"""
def FFT(v, omega, m):
    if m == 1:
        return [v[0]]
    v_even = FFT(v[0::2], omega ** 2, m / 2)
    v_odd = FFT(v[1::2], omega ** 2, m / 2)
    omega_precomp = [omega ** i for i in range(m/2)]
    return concat(
        v_even + omega_precomp * v_odd,
        v_even - omega_precomp * v_odd,
    )

The time complexity, by Master's theorem, is \(T(m) = 2T(m/2) + O(m) \in O(m\log m)\).

Parallel FFT

Consider the stack tree of the recursive calls, it has \(m\) leaves and \(\log m\) depth. Therefore, instead of doing top-down recursion, we can do bottom-up computations, which gives a natural way for parallelizing the algorithm.

Then, consider the recursive call, note that the divide is done by odd/even, or the last digit of the binary. Now, consider \(m=2^4\), the stack tree looks like

XXX
XX0             XX1
X00     X10     X01     X11
000 100 010 110 001 101 011 111

Note that from left to right, the leaves are ordered by reversed bit order. Therefore, what we can do is to compute FFT at each level, rewrite the elements \(v_i = (F\cdot v)_{\text{bitreverse(i)}}\).

Then, for \(p\) processors, the computation takes \(2m\log m / p\), for communications, we need \(\log(p)\) stages and \(m\log(p)/p\) words.