Sec. 9.3 Decimation-in-Time FFT Algorithms 635 transform using the Goertzel algorithm, the number of real multiplications required is approximately N? and the number of real additions is approximately 2N?. While this is more efficient than the direct computation of the discrete Fourier transform, the amount of computation is still proportional to N?. In either the direct method or the Goertzel method we do not need to evaluate X{k] at all N values of k. Indeed, we can evaluate X[k] for any M values of k, with each DFT value being computed by a recursive system of the form of Figure 9.2 with appropriate coefficients. In this case, the total computation is proportional to NM. The Goertzel method and the direct method are attractive when M is small; however, as indicated previously, algorithms are available for which the computation is proportional to Nlog, N when N is a power of 2. Therefore, when M is less than log, N, either the Goertzel algorithm or the direct method may in fact be the most efficient method, but when all N values of X[k] are required, the decimation-in-time algorithms, to be considered next, are roughly (N/log, N) times more efficient than either the direct method or the Goertzel method. 9.3 DECIMATION-IN-TIME FFT ALGORITHMS In computing the DFT, dramatic efficiency results from decomposing the computation into successively smaller DFT computations. In this process, we exploit both the sym- metry and the periodicity of the complex exponential WA" = e~/@"/")k"_ Algorithms in which the decomposition is based on decomposing the sequence x{n] into successively smaller subsequences are called decimation-in-time algorithms. ‘The principle of the decimation-in-time algorithm is most conveniently illustrated by considering the special case of N an integer power of 2, ie., N = 2. Since N is an even integer, we can consider computing X [k] by separating x[n] into two (N/2)-point? sequences consisting of the even-numbered points in x[] and the odd-numbered points in x[nJ. With X[K] given by wa Xk = So xinlwek, k= 0,1,....N-1, (9.10) n= and separating x[n] into its even- and odd-numbered points, we obtain XK) = 2 oxlnl ek + SO atest, (9.11) neven mold or, with the substitution of variables n = 2r for n even and n = 2r +1 forn odd, (N2)-1 (N/D-1 XK = So xfar]wee + SO xfar + wy = od (9.12) wat wat = 2 xerwiyt+ we SO xa + 1ycwey. 0 0 3When discussing FFT algorithms, we use the words sample and point interchangeably to mean se- quence value. Also, we refer to a sequence of length N as an N-point sequence, and the DFT of a sequence of length JV will be called an N-point DFT.636 Computation of the Discrete Fourier Transform Chap. 9 But Wi = Wye, since PIRI) =e PHAN) = Wy, (9.13) Consequently, Eq. (9.12) can be rewritten as (nyt (wy X= SO x2r]Wih + WA SO x2 + 11K (0.4) r=0 = : = G[K]+ WEA[A, = k= 0,1,....N—1 Each of the sums in Eq. (9.14) is recognized as an (N/2)-point DFT, the first sum being the (N/2)-point DFT of the even-numbered points of the original sequence and the second being the (N/2)-point DFT of the odd-numbered points of the original sequence. Although the index kranges over N values, k = 0, 1,..., N—1, each of the sums must be computed only for k between 0 and (N/2)— 1, since G[k] and H[&] are each periodic in k with period N/2. After the two DFTs are computed, they are combined according to Eq. (9.14) toyield the N-point DFT X[k]. Figure 9.3 depicts this computation for N = 8. In this figure, we have used the signal flow graph conventions that were introduced in Chapter 6 for representing difference equations. That is, branches entering a node are summed to produce the node variable. When no coefficient is indicated, the branch transmittance is assumed to be unity. For other branches, the transmittance of a branch is an integer power of Ww. In Figure 9,3, two 4-point DFTs are computed, with G[&] designating the 4-point DFT of the even-numbered points and H/{k] designating the 4-point DFT of the odd- numbered points. X(0] is then obtained by multiplying [0] by Wf and adding the product to G[0]. X{1] is obtained by multiplying H{[1] by Wj, and adding that result to G[1]. Equation (9.14) states that, to compute X/[4], we should multiply H[4] by Wy and add the result of G[4]. However, since G[k] and H{A] are both periodic in k with oo) 10] > a x0) rpjory pein a DFT x[6]o-> x(t) xBlo - point Hil] we xijo> DET >» X[6] Figure 9.3 Flow graph of the AQ] ws decimation-in-time decomposition of an N-point DFT computation into two xo (7 (W/2}-point DFT computations AB] We (Wa)Sec. 9.3 Decimation-in-Time FFT Algorithms 637 period 4, H[4] = H{0] and G[4] = G[0]. Thus, X [4] is obtained by multiplying [0] by W and adding the result to G[0]. As shown in Figure 9.3, the values X{5], X [6], and X[7] are obtained similarly. With the computation restructured according to Eq. (9.14), we can compare the number of multiplications and additions required with those required for a direct com- putation of the DFT. Previously we saw that, for direct computation without exploiting symmetry, N? complex multiplications and additions were required. By comparison, Eg. (9.14) requires the computation of two (N/2)-point DFTs, which in turn requires 2(.N/2)? complex multiplications and approximately 2(N/2)* complex additions if we do the (N/2)-point DFTs by the direct method. Then the two (N/2)-point DFTs must be combined, requiring N complex muitiplications, corresponding to multiplying the second sum by Wf, and N complex additions, corresponding to adding the product obtained to the first sum. Consequently, the computation of Eq. (9.14) for all values of k requires at most N + 2(N/2)* or N + N?/2 complex multiplications and complex additions. It is easy to verify that for N > 2, the total N + N?/2 will be less than N?. Equation (9.14) corresponds to breaking the original N-point computation into two (N/2)-point DFT computations. If N/2 is even, as it is when N is equal to a power of 2, then we can consider computing each of the (N/2)-point DFTs in Eq. (9.14) by breaking each of the sums in that equation into two (N/4)-point DFTs, which would then be combined to yield the (N/2)-point DFTs. Thus, G[A] in Eq. (9.14) would be represented as (way (npay=1 (npay-t Gid= XS slrlwis= Yo sewak+ SY eetruwes, 0.15) = = = or (N/4)-1 (N/A)-1 GK = S> sl2€]Wiis+ Whe D> el2e+ 1) Wii, (9.16) = = Similarly, H[k] would be represented as yt (jay AUK) = D> AE Wihg + Why D2 nize + 1)W ih (9.17) - = Consequently, the (N/2)-point DFT G[A] can be obtained by combining the (N/4)-point DFTs of the sequences g[2¢] and g[2€ + 1]. Similarly, the (N/2)-point DFT H[K] can be obtained by combining the (N/4)-point DFTs of the sequences h[2¢) and h[2¢ + 1] Thus, if the 4-point DFTs in Figure 9.3 are computed according to Eqs. (9.16) and (9.17), then that computation would be carried out as indicated in Figure 9.4. Inserting the computation of Figure 9.4 into the flow graph of Figure 9.3, we obtain the complete flow graph of Figure 9.5, where we have expressed the coefficients in terms of powers of Wy rather than powers of Wyo, using the fact that Wy = W3. For the 8-point DFT that we have been using as an illustration, the computation has been reduced to a computation of 2-point DFTs. For example, the 2-point DFT of the sequence consisting of x[0] and x[4] is depicted in Figure 9.6. With the computation ‘For simplicity, we shall assume that Nis large, so that (1V ~ 1) can be approximated by N.638 Computation of the Discrete Fourier Transform Chap. 9 slo point eal Det apo Figure 9.6 Flow graph of the point decimation-in-time decomposition of an 4 (N/2)-point DFT computation into two xf6}o—>{_PFT (W/4)-point DFT computations (N = 8), xo xt} N {Point DFT xo x{elo point aloes prt xo xis} X~ point Point pr xBloe xt N-point 3 - point DET Figure 9.5 Result of substituting the structure of Figure 9.4 into Figure 9.3. x14) = Figure 9.6 Flow graph of a 2-point Wr=WH"=-1 DFT. one of Figure 9.6 inserted in the flow graph of Figure 9.5, we obtain the complete flow graph for computation of the 8-point DFT, as shows in Figure 9.7. For the more general case, but with N still a power of 2, we would proceed by decomposing the (N/4)-point transforms in Eqs. (9.16) and (9.17) into (N/8)-point transforms and continue until we were left with only 2-point transforms. This requires v = log, N stages of computation. Previously, we found that in the original decomposi- tion of an N-point transform into two (N/2)-point transforms, the number of complex multiplications and additions required was N + 2(N/2)?. When the (N/2)-point trans- forms are decomposed into (N/4)-point transforms, the factor of (NV/2)? is replaced by N/2 + 2(N/4), so the overall computation then requires N + N + 4(N/4)? complex multiplications and additions. If N = 2", this can be done at most v = log, N times, so that after carrying out this decomposition as many times as possible, the number of, complex multiplications and additions is equal to Nv = Nlog, N.Sec. 9.3. Decimation-in-Time FFT Algorithms 639 Figure 9.7 Flow graph of complete decimation-in-time decomposition of an 8-point DFT computation. Figure 9.8 Flow graph of basic butterfly computation in Figure 9.7. ‘The flow graph of Figure 9.7 displays the operations explicitly. By counting branches with transmittances of the form Wx, we note that each stage has N complex muitiplications and N complex additions. Since there are log, N stages, we have a total of Nlog, Ncomplex multiplications and additions. This is the substantial computational savings that we have previously indicated was possible. For example, if N = 2" = 1024, then N? = 2 = 1,048,576, and Nlog, N = 10,240, a reduction of more than two orders of magnitude! ‘The computation in the flow graph of Figure 9.7 can be reduced further by exploit- ing the symmetry and periodicity of the coefficients W,. We first note that, in proceeding from one stage to the next in Figure 9.7, the basic computation is in the form of Fig- ure 9.8, ‘e., it involves obtaining a pair of values in one stage from a pair of values in the preceding stage, where the coefficients are always powers of Wy and the exponents are separated by N/2. Because of the shape of the flow graph, this elementary computation is called a butterfly. Since WAP =e iRRININE grin = (9.18) the factor Wx"? can be written as Wii? = WP Wh = Whe (9.19)x{o]° x{dJo x2]e x{6l° x{ljo 640 Computation of the Discrete Fourier Transform Chap. 9 With this observation, the butterfly computation of Figure 9.8 can be simplified to the form shown in Figure 9.9, which requires only one complex multiplication instead of two. Using the basic flow graph of Figure 9.9 as a replacement for butterflies of the form of Figure 9.8, we obtain from Figure 9.7 the flow graph of Figure 9.10. In particular, the number of complex multiplications has been reduced by a factor of 2 over the number in Figure 9.7. (m=—1)st rth stage stone we Figure 9.9 Flow graph of simplified oe butterfly computation requiring only one “1 complex multiplication. 9.3.1 In-Place Computations ‘The flow graph of Figure 9.10 describes an algorithm for the computation of the discrete Fourier transform. The essential features of the flow graph are the branches connecting the nodes and the transmittance of each of these branches. No matter how the nodes in the flow graph are rearranged, it will always represent the same computation, provided that the connections between the nodes and the transmittances of the connections are maintained. The particular form for the flow graph in Figure 9.10 arose out of deriving the algorithm by separating the original sequence into the even-numbered and odd-numbered points and then continuing to create smaller and smaller subsequences in the same way. An interesting by-product of this derivation is that this flow graph, in addition to describing an efficient procedure for computing the discrete Fourier transform, also suggests a useful way of storing the original data and storing the results of the computation in intermediate arrays. To see this, it is useful to note that according to Figure 9.10, each stage of the computation takes a set of N complex numbers and transforms them into another set of Figure 9.10 Flow graph of 8-point DFT using the butterfly computation of Figure 9.9,646 Computation of the Discrete Fourier Transform Chap. 9 access memory is not available is shown in Figure 9.16. This flow graph represents the decimation-in-time algorithm originally given by Singleton (1969). (See DSP Com- mittee, 1979, for a program using serial memory.) Note first that in this flow graph the input is in bit-reversed order and the output is in normal order. The important feature of the flow graph is that the geometry is identical for each stage: only the branch trans- mittances change from stage to stage. This makes it possible to access data sequentially. Suppose, for example that we have four separate disk files, and suppose that the first half of the input sequence (in bit-reversed order) is stored in one file and the second half is stored in a second file. Then the sequence can be accessed sequentially in files 1 and 2 and the results written sequentially on files 3 and 4, with the first half of the new array being written to file 3 and the second half to file 4. Then at the next stage of computation, files 3 and 4 are the input, and the output is written to files 1 and 2. This is repeated for each of the v stages. Such an algorithm could be useful in computing the DFT of extremely long sequences, 9.4 DECIMATION-IN-FREQUENCY FFT ALGORITHMS ‘The decimation-in-time FFT algorithms are all based on structuring the DFT compu- tation by forming smaller and smaller subsequences of the input sequence x{n]. Alter- natively, we can consider dividing the output sequence X(k] into smaller and smaller subsequences in the same manner. FFT algorithms based on this procedure are com- monly called decimation-in-frequency algorithms. To develop these FFT algorithms, let us again restrict the discussion to N'a power of 2 and consider computing separately the even-numbered frequency samples and the odd-numbered frequency samples. Since na X[K = Soxtnlwat, k= 01....N-1, (9.23) “= the even-numbered frequency samples are Net Xr] = So xtn]wee?, or .(N/2) 1, (9.24) = which can be expressed as nyt wet Xr] = SO xf warr + SO xin we" (9.25) = 7 With a substitution of variables in the second summation in Eq. (9.25), we obtain wart (wat : Xr] = So xfer SO xf + (ayy) (9.26) n= n=OSec.9.4 _Decimation-in-Frequency FFT Algorithms 647 Finally, because of the periodicity of 2”, wayne — wR WAN = wien, (9.27) and since Wi = Ww/2, Eq. (9.26) can be expressed as wat XR] = Yo Gln tale + (N/2D)WH, = 0,1,...,(N/2)- 1. (9.28) n=O Equation (9.28) is the (N/2)-point DFT of the (N/2)-point sequence obtained by adding the first half and the last half of the input sequence. Adding the two halves of the input sequence represents time aliasing, consistent with the fact that in computing only the even-numbered frequency samples, we are undersampling the Fourier transform of x{71). We can now consider obtaining the odd-numbered frequency points, given by wat Xr +1) = So xtaywye*?, 1 0,1,...,(N/2)= 1. (9.29) As before, we can rearrange Eq. (9.29) as way wa Xr iy= So xterra wget. (9.30) = non An alternative form for the second summation in Eq. (9.30) is S (2r+1) oat [n(N/2)2r+1) DS xtaywye? = SS xine (N/a) Wy Cr none = (N21 = WYRED Stn + (N/2) WA no way =- DO vt (Ny2yywrern, (931) n= where we have used the fact that Wi"”"" = 1 and Wi” = —1. Substituting Eq. (9.31) into Eq, (9.30) and combining the two summiations, we obtain (N/2)=1 Xr +1) = SO Caled = xl (N/2)) we, (9.32) =«(01 x(t] x2 =) x(4] x51 «(61 x(7] 648 Computation of the Discrete Fourier Transform Chap. 9 or, since Wy = Ww, (nzzjnt X[2r+ t= Yo (ln) — ale + (N/2))WHW Ki. c= (9.33) r=0,1,....(N/2)-1. Equation (9.33) is the (N/2)-point DFT of the sequence obtained by subtracting the second half of the input sequence from the first half and multiplying the resulting sequence by Wf. Thus, on the basis of Eqs, (9.28) and (9.33), with g[n] = x[n] + x[n-+N/2] and hn] = x{n}—x[n+ N/2], the DFT can be computed by first forming the sequences g[7] and h{n], then computing A{]W,,, and finally computing the (N/2)-point DFTs of these two sequences to obtain the even-numbered output points and the odd- numbered output points, respectively. The procedure suggested by Eqs. (9.28) and (9.33) is illustrated for the case of an 8-point DFT in Figure 9.17. Proceeding in a manner similar to that followed in deriving the decimation-in- time algorithm, we note that since N is a power of 2, N/2 is even; consequently, the (N/2)-point DFTs can be computed by computing the even-numbered and odd num- bered output points for those DFTs separately. As in the case of the procedure leading to Eqs. (9.28) and (9.33), this is accomplished by combining the first half and the last, half of the input points for each of the (N/2)-point DFTs and then computing (N/4)- point DFTs. The flow graph resulting from taking this step for the 8-point example is shown in Figure 9.18. For the 8-point example, the computation has now been reduced to the computation of 2-point DFTs, which are implemented by adding and subtracting the input points, as discussed previously. Thus, the 2-point DFTs in Figure 9.18 can be replaced by the computation shown in Figure 9.19, so the computation of the 8-point DFT can be accomplished by the algorithm depicted in Figure 9.20. By counting the arithmetic operations in Figure 9.20 and generalizing to N = 2", we see that the computation of Figure 9.20 requires (N/2) log, ‘V complex multiplica- tions and Nlog, N complex additions. Thus, the total number of computations is the same for the decimation-in-frequency and the decimation-in-time algorithms. 0 (01 aa NX point [OAPI DFT Jo x[4] p> X16] }-—° X[1] A point XB) 2 Figure 9.17 Flow graph of per [X15] decimation-in-frequency decomposition ‘of an N-point DFT computation ints two }-—o.x[7] _ (N/2)-point DFT computations ( N= 8)x(0] xt) [2] x13] fa] x{S] x{6] x7] Sec. 9.4 Decimation-in-Frequency FFT Algor ths x pin aR DF }—« x10] fo X14] <1 JB point 7 DFT Jo x12] [+0 x16] <1 8 point Wy DFT [0x11] fo x15] ST] — pon point X11) Xl fo x13] AT Xp] Xa x7 Figure 9.18 Flow graph of decimation-in-frequency decomposition of an 8-point DFT into four 2-point OFT computations. Figure 9.19 Flow graph of a typical 2-point OFT as required in the last stage of decimation-in-frequency decomposition. Figure 9.20 Flow graph of complete decimation-in-frequency decomposition of an 8-point DFT computation.