Chapter 12

The Mixed-Radix and Split-Radix FFTs

There are two kinds of mixed-radix FFT algorithms. The rst kind refers to a situation arising naturally when a radix-q algorithm, where q = 2m > 2, is applied to an input series consisting of N = 2k q s equally spaced points, where 1 k < m. In this case, out of necessity, k steps of radix-2 algorithm are applied either at the beginning or at the end of the transform, while the rest of the transform is carried out by s steps of the radix-q algorithm. For example, if N = 22m+1 = 2 4m , the mixed-radix algorithm combines one step of the radix-2 algorithm and m steps of the radix-4 algorithm.1 The second kind of mixed-radix algorithms in the literature refers to those specialized for a composite N = N0 N1 Nk . Dierent algorithms may be used depending on whether the factors satisfy certain restrictions. The FFT algorithms for composite N will be treated in Chapter 15.

12.2

The Split-Radix DIT FFTs

After one has studied the radix-2 and radix-4 FFT algorithms in Chapters 3 and 11, it is interesting to see that the computing cost of the FFT algorithm can be further reduced by combining the two in a split-radix algorithm. The split-radix approach was rst proposed by Duhamel and Hollmann in 1984 [39]. There are again DIT versions and DIF versions of the algorithm, depending on whether the input time series or the output frequency series is decimated. The split-radix DIT algorithm is derived from (3.1), which denes the discrete
1 It is of historical interest to note that a program for N = 22m+1 was written by Gentleman and Sande [47] in 1966, where they claimed a doubling of eciency by this approach. However, Singleton observed in [83] that when computing with all the data stored in memory, a good radix-2 program was nearly as ecient as a radix-4 plus one step of radix-2 program and was simpler.

2000 by CRC Press LLC

Fourier transform of a complex time series:

N 1

Xr =
=0
N 2

r x N , 1

r = 0, 1, . . . , N 1,
N 4

N 4

(12.1)

=
k=0
N 2

x2k N
1

r (2k)

+
k=0

x4k+1 N
N 4

r (4k+1)

+
k=0

x4k+3 N
N 4

r (4k+3)

=
k=0

x2k N

r (2k)

+ N
k=0

x4k+1 N

r (4k)

+ N

3r k=0

x4k+3 N

r (4k)

By decimating the time N/2 1}, the set {zk|zk k N/4 1}, the three 2 4 N = N and N = N
2 4

series into three sets, namely the set {yk |yk = x2k , 0 k = x4k+1 , 0 k N/4 1}, and the set {hk|hk = x4k+3 , 0 subproblems are dened after the appropriate twiddle factors are identied.

(12.2)

N 2

N 2

Yr =
k=0

x2k N

r (2k)

=
k=0

x2k N

2 rk

N 2

=
k=0

yk rk N ,
2

r = 0, 1, . . . , N/2 1.

(12.3)
N 2

N 4

Zr =
k=0

x4k+1 N

r (4k)

=
k=0

x4k+1 N

4 rk

N 4

=
k=0

zk rk N ,
4

r = 0, 1, . . . , N/4 1.

(12.4)
N 2

N 4

Hr =
k=0

x4k+3 N

r (4k)

=
k=0

x4k+3 N

4 rk

N 4

=
k=0

hk rk N ,
4

r = 0, 1, . . . , N/4 1.

After these three subproblems are each (recursively) solved by the split-radix algorithm, the solution to the original problem of size N can be obtained according to (12.1) for r = 0, 1, . . . , N 1. Because Yr+k N = Yr for 0 r N/2 1, Zr+k N = Zr for 2 4 0 r N/4 1, and Hr+k N = Hr for 0 r N/4 1, equation (12.1) can be
4

2000 by CRC Press LLC

rewritten in terms of Yr , Yr+ N , Zr , and Hr for 0 r N/4 1 as shown below.

r 3r Zr + N Hr Xr = Yr + N

(12.5)

r 3r = Yr + N Zr + N Hr ,

0r

N 1, 4 N 1, 4

(12.6)

Xr+ N = Yr+ N + N
4 4 4

r+ N 4

3(r + N 4 ) Zr + N Hr

r 3r = Yr+ N j N Zr N Hr ,

0r

(12.7)

Xr+ N = Yr + N
2

r+ N 2

3(r + N 2 ) Zr + N Hr

r 3r = Yr N Zr + N Hr ,

0r

N 1, 4 N 1. 4

(12.8)

Xr+ 3N = Yr+ N + N
4 4 4

N r + 34

N 3(r + 34 ) Zr + N Hr

r 3r = Yr+ N + j N Zr N Hr ,

0r

The computation represented by (12.5), (12.6), (12.7), and (12.8) is referred to as an unsymmetric DIT buttery computation in the literature as shown in Figure 12.1. Figure 12.1 The split-radix DIT FFT butteries.

12.2.1

Analyzing the arithmetic cost

To determine the arithmetic cost of the split-radix DIT FFT algorithm, observe that r 3r N Zr and N Hr must be computed before the two partial sums can be formed. Since these two subproblems are each of size N/4, N/2 complex multiplications and N/2 complex additions are required in order to obtain the partial sums. Among the N/2 complex multiplications, there are four special cases which were already identied in the earlier discussion of the radix-4 algorithm: they are two cases of multiplication by 1 and two cases of multiplication by an odd power of 8 . Recall that the former two cases are trivial, and only four real operations rather than six would be used in the latter two cases. Thus, (3+3) (N/2 4)+4 2+2 (N/2) = 4N 16 nontrivial real operations are performed in the rst stage of buttery computation. In the second stage of the buttery computation, only N complex additions or 2N real operations are required. The total cost for a single split-radix step thus involves 6N 16 nontrivial real operations (ops).

2000 by CRC Press LLC

To set up the recurrence equation, the boundary conditions for both N = 4 and N = 2 are needed; when the size of a subproblem is reduced to 8, the three subsequent subproblems are of sizes 4, 2 and 2. As noted earlier, T (4) = 16 ops, and T (2) = 4 ops. The cost of the split-radix FFT algorithm (in terms of nontrivial ops) can be represented by the following recurrence: N N n T 2 + 2T 4 + 6N 16 if N = 4 > 4 , (12.9) T (N ) = 16 4 if N = 4 , if N = 2 .

Solving (12.9) (see Appendix B), one obtains the solution (12.10) T (N ) = 4N log2 N 6N + 8 .

12.3

The Split-Radix DIF FFTs

A split-radix DIF FFT algorithm can be derived by recursively applying both radix-2 and radix-4 DIF FFT algorithm to solve each subproblem resulting from decimating the output frequency series in a similar fashion. That is, the frequency series is recursively decimated into three subsets, i.e., the set denoted by Yk = X2k for 0 k N/2 1, the set denoted by Zk = X4k+1 for 0 k N/4 1, and the set denoted by Hk = X4k+3 for 0 k N/4 1 as shown below. The derivation begins with the discrete Fourier transform dened by (3.1). Using the results developed earlier for the radix-2 DIF algorithm in (3.11), one obtains
N 2

1 r x N +

N 1

Xr = (12.11) =
=0 =0
N 2

r x N =N 2 rN r N ,

x +x

+N 2

N 2

r = 0, 1, . . . , N 1.

By letting Yk = X2k , y = x + x
N 2

+N 2 1

, one subproblem of half the size is dened by x +x k N

Yk = X2k = (12.12) =
=0 =0
N 2

+N 2

y k N ,
2

k = 0, 1, . . . , N/2 1.

To construct the other two subproblems of size N/4, begin with the DFT denition in (3.1) and use the results developed earlier for the DIF radix-4 algorithm in (11.22).
N 1

Xr = (12.13) =
=0 =0
N 4

r x N , 1

r = 0, 1, . . . , N 1 ,

x +x

+N 4

r 4 +x

+N 2

2r 4 +x

N + 34

3r r N 4 .

2000 by CRC Press LLC

By substituting r = 4k + 1 and r = 4k + 3, one again obtains (12.14)

Zk = X4k+1 =
=0
N 4

x +x
1

+N 4

4k+1 4 +x

+N 2

2(4k+1)

+x

N + 34

3(4k+1)

(4k+1)

=
=0
N 4

x x
1

+N 2

j x

+N 4

N + 34

N k N
4

=
=0

z k N ,
4

k = 0, 1, . . . , N/4 1 .

(12.15)
N 4

Hk = X4k+3 =
=0
N 4

x +x
1

+N 4

4k+3 4 +x

+N 2

2(4k+3)

+x

N + 34

3(4k+3)

(4k+3)

=
=0
N 4

x x
1

+N 2

+j x

+N 4

N + 34

3 k N N
4

=
=0

h k N ,
4

k = 0, 1, . . . , N/4 1 .

To form these three subproblems using two stages of unsymmetric buttery computation, the computation of the partial sums is again rearranged to facilitate the buttery computation. (12.16) (12.17) (12.18) (12.19) y y = x +x
+N 4 +N 2

,
N + 34

0 ,

= x

+N 4

+x

N 1. 4 N 0 1. 4
+N 4

z =

x x
+N 4

+N 2

j x
N + 34

N + 34

N ,
3 N ,

h = j x

+ x x

+N 2

N 1. 4 N 1. 0 4

The computation represented by (12.16), (12.17), (12.18), and (12.19) again yields an unsymmetric DIF buttery computation as depicted in Figure 12.2.

2000 by CRC Press LLC

Figure 12.2 The split-radix DIF FFT butteries.

12.4

Notes and References

The split-radix FFT was originally developed by Duhamel and Hollmann [39] in 1984, and it was subsequently extended and implemented for complex, real and real-symmetric data by Duhamel in [38]. In 1986, Sorensen, Heideman, and Burrus presented an indexing scheme which eciently implemented the Duhamel-Hollmann split-radix FFT [86]. Both DIF and DIT Fortran programs were presented in [86]. The history of the ideas on the fast Fourier transforms from Gauss to the split-radix algorithm is presented in [41].

2000 by CRC Press LLC