On computing the Discrete Fourier Transform

Abstract

New algorithms for computing the Discrete Fourier Transform of n points are described. For n in the range of a few tens to a few thousands these algorithms use substantially fewer multiplications than the best algorithm previously known, and about the same number of additions.Computing the Discrete Fourier Transform (DFT) of n points:n-1 2vt Aj = wiiai, j ... .,n -1, w= e[1] =_0 has many applications in scientific and engineering calculation. In 1965 Cooley and Tukey (1) described an algorithm for computing DFT in n/2 log2 n/2 complex multiplications and n log2 n complex additions, when n = 2s is a power of 2. In this note we will describe a new algorithm for computing the DFT. For n in the range of a few tens to a few thousands this algorithm uses substantially fewer multiplications than that described by Cooley and Tukey, and about the same number of additions as theirs. Theoretical background In ref. 2 we considered the following problem: Let Pn = Un + E2 %laiui be a polynomial with coefficients in a field G, let Rn = ZL; xjui and Sn = En=%lyjui be two polynomials with indeterminate coefficients. What is the minimum number of multiplications needed to compute the coefficients of Tp = Rn'Sn mod Pn when multiplications by an element g e C are not counted? Let Tp denote the set of coefficients. It was proved that:THEOREM. If Pn = JV Qi1" where Qj is irreducible (over C) and (Qi, Qj) = 1 for i # j, then the minimum number of multiplications needed to compute Tp is 2n -k.Let p be a prime and consider computing the DFT of p elements. The difficult part of the computation is that of computing Ak =~J~j~d ,=ia,, k = 1,2,..., p-1, w = exp(2iri)/p. The exponent of w is k-j mod p, and since the group of non-zero integers with operation of multiplication modulo p is isomorphic to Zp-j it follows that there exists a permutation H of 11,2, . . , p-1 such that the matrix whose (i,j) element is wvr(i)-(j) is cyclic. Therefore computing A r) = Er=jlwr(i)r(i) aT(j) j = 1,2,... p-1 is the same as computing the coefficients of P-1 p-l E wT)u1'_) a (l) + E:a, p a l-ui) mod uP'-1.[2] According to the theorem this can be done in 2(p-1)-k multiplications, where k is the number of irreducible factors (over the rationals) of uP-1-1.Similar results are obtained when n = pr is a power of a prime. In this case we use the result that the group of integers relatively prime to p with group operation of multiplication modulo pr is Z(p-l)pr for p 5 2 and is Z2 X Z2r-2 for p =2. Summary of resultsAll known algorithms for computing Tp in the minimum number of multiplications require a large number of additions when P has large irreducible factors. Therefore the algorithms for DFT as outlined in the previous section are not of practical interest unless the number of points n is small. Table 1 summarizes the number of multiplications and additions used to compute DFT for small n. We will need later to consider multiplication by w0 = 1 as a multiplication, and we therefore indicate these multiplications as well in the table.For all these al...

Power-Aware Architecting for data-dominated applications

“…Consequently, the memory requirements do not increase while performing an FFT. Other FFTs like the Winograd FFT [84] do not have this property and require more and more memory space.…”

VLSI Implementation of Digital Fourier Transforms.

“…New techniques were developed for the construction of large scale FF1 processors which are geared toward the use of VLSI. These techniques employ the traditional Cooley-Tukey version of the FFT [4] as well as the prime-factor a.gorithm of Good [5] and elements of the Winograd Fourier Transform [6].…”

The design of low complexity low power pipelined short length Winograd Fourier transforms

“…Winograd Fourier Transform (WFT) [1] algorithm is highly preferable in designs that involves Discrete Fourier Transform (DFT). Twiddle factor multiplication is not required for WFT, which in turn reduces the number of real multipliers needed.…”