Chapter 8

Advanced FFT Topics and

Applications

Contents
8.1 Zoom Spectral Analysis . . . . . . . . . . . . . . . . . 8-3
8.2 Averaged Periodogram Spectral Estimation . . . . . . 8-9
8.2.1 Periodogram Properties . . . . . . . . . . . . . . 8-10
8.2.2 Averaged Periodograms . . . . . . . . . . . . . 8-12
8.3 Quantization Effects in DFT Computations . . . . . . 8-18
8.3.1 Quantization Errors in Direct Computation of the
DFT . . . . . . . . . . . . . . . . . . . . . . . . 8-18
8.3.2 Quantization Errors in the Radix-2 FFT . . . . . 8-21

8-1
CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

8-2 ECE 5650/4650 Modern DSP

 8.1. ZOOM SPECTRAL ANALYSIS

In this chapter we will begin by introducing a technique for im-

proving frequency resolution in an FFT based spectral analysis sys-
tem. The second topic will be averaged periodogram spectral esti-
mation. The third topic will be a brief look at quantization effects in
the computation of the DFT (FFT).

8.1 Zoom Spectral Analysis

To improve the frequency resolution in a spectral analysis system in
general requires increasing the DFT length N . Given a fixed band of
frequencies must be spectral analyzed, an efficient means of obtain-
ing high resolution is to subdivide the band into smaller bands (most
likely contiguous bands). High frequency resolution can be obtained
in each subband without the DFT length becoming too large. The
block diagram of a zoom spectral analysis system is shown below.

xI [ n ]
LPF
sc ( t ) x(t) x[n]
π
ω c = ----- M yI [ n ]
M
Y[k]
Complex
H aa ( jΩ ) C/D Decimators
cos ( ω o n ) FFT
LPF
B0
π M yQ [ n ]
–1 ω c = -----
M
T = fs
xQ [ n ] High Resolution
sin ( ω o n ) Spectrum Processor
...

Y[k ]
High Resolution
Spectrum Processor Bj

ωj

ECE 5650/4650 Modern DSP 8-3

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 The DFT analysis of a particular subband centered at o (fo or

!o) begins by frequency translating the subband to a spectral
location centered about  D 0 using a complex demodula-
tor. The required complex demodulation may be performed in
either continuous- or discrete-time.

F DFT
 Suppose x.t / ! X.j / and xŒn ! X Œk, then

j o t F
x.t /e ! X.j  C j o/
j 2ko =N DFT
xŒne ! X Œk C ko; 0  k  N 1

Note that in either case the output of the complex demodulator

is a complex signal (i.e. xŒn D xI Œn C jxQ Œn/

 Following complex demodulation the signal is lowpass filtered

and then decimated (downsampled) by a factor M

 The lowpass filter (in practice a digital filter) prior to decima-

tion is needed to prevent aliasing when the effective sampling
rate is reduced by M . Assuming an ideal lowpass filter, the
lowpass cutoff frequency, !c , must be less than =M

 On the following two pages appropriate frequency domain wave-

forms within the zoom spectrum analyzer are given

8-4 ECE 5650/4650 Modern DSP

 8.1. ZOOM SPECTRAL ANALYSIS

X ( jΩ )
Frequency Region
of Interest
Ω
N-point FFT –π ⁄ T –Ω o Ωo π⁄T Spectrum After
Gives ∆f = 1/(NT) jω Sampling
X(e )

ω
–ω o ωo
Frequency Translated
j ( ω – ωo ) Spectrum
X( e )

ω
BW
jω
H LP ( e )
ω
–π ⁄ M π⁄M
jω
X LP ( e )
ω
–2 π –π ⁄ M π⁄M 2π
jω
Y(e )

ω
–M π –π π Mπ
BW N-point FFT
Gives ∆f = 1/(MNT)
jω
Y(e )
f s ⁄ M Hz

–π π ω
fs fs f, Hz
– -------- --------
2M 2M

Frequency spectra from the input through the decimator

output

ECE 5650/4650 Modern DSP 8-5

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 With the sampling rate reduced by M , the frequency band an-

alyzed by an N -point DFT has total width 1=.M T / D fs =M
Hz (assuming the initial sampling rate is fs samples/sec)

 The minimum frequency spacing has been reduced from fs =N

to fs =.MN /

 To analyze a different subband all we need to do is change the

value of fo .ko/

 In the study of windows it was pointed out that to increase

frequency resolution the window length must increase. For the
zoom transform discussed above the signal capture- or collect-
time, is identical to using an MN -point DFT.

 With the zoom transform fewer points are actually collected

(only one out of M samples is saved) and hence less compu-
tation is required, since only a narrow band of frequencies is
being analyzed. The signal is compressed and the frequency
axis in effect expands.

Example 8.1: Designing for 10 Hz Resolution

Given fs D 10 MHz find N and M to spectrum analyze a 100
kHz subband with f  10 Hz
fs
BW D ) M D 100
M
fs
f D ) N  10; 000 .N D 214 D 16; 384/
MN

8-6 ECE 5650/4650 Modern DSP

 8.1. ZOOM SPECTRAL ANALYSIS

Example 8.2: Using Octave Subbands 1

Parameter Requirement
A/D output frequency 6553.6 Hz obtained by divid-
ing down by 300 a 1.96608
MHz crystal oscillator
FFT outputs 5 octaves between 40 and
1280 Hz
FFT resolution 400 filters per octave yield-
ing resolutions of 1.6 Hz in
the 640-1280 Hz octave, 0.8
Hz in the 320-640 Hz octave,
and so on.
FFT size 512 points
Maximum passband ripple ˙0:5 dB
due to all analog and digital
filtering operations
Minimum rejection of sinu- -43 dB
soids aliased into the DFT
analysis band

D.F. Elliot and K.R. Rao, Fast Transforms Algorithms, Analysis,

1

Applications, Academic Press, Inc., 1982.

ECE 5650/4650 Modern DSP 8-7

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

6553.6 Hz
Anti-alias
Filter Input Signal x(t)
LPF
– j2π960n
exp ------------------------
F1(z) 1638.4 2:1
2:1 I

640-1280
3276.8 Hz 2:1 BPF I-Q LPF 
Q Octave
F2(z) Demod F1(z) 
to FFT
– j2π480n
LPF 1638.4 Hz exp ------------------------ 819.2 Hz
F1(z) 1638.4 2:1
2:1 I

320-640
1638.4 Hz 2:1 BPF I-Q LPF 
Octave
F2(z) Demod Q F1(z) 
to FFT
– j2π240n
LPF 819.2 Hz exp ------------------------ 409.6Hz
F1(z) 1638.4 2:1
2:1 I

160-320
819.2 Hz 2:1 BPF I-Q LPF 
Octave
F2(z) Demod Q F1(z) 
to FFT
– j2π120n
LPF 409.6 Hz exp ------------------------ 204.8 Hz
F1(z) 1638.4 2:1
2:1

I 80-160
406.6 Hz 2:1 BPF I-Q LPF 
Octave
F2(z) Demod Q F1(z) 
to FFT
LPF 204.8 Hz – j2π60n 102.4 Hz
exp ---------------------
F1(z) 1638.4 2:1 
2:1 I 40-80
204.8 Hz 2:1 BPF I-Q LPF 
Octave
F2(z) Demod Q F1(z)

to FFT
102.4 Hz 51.2 Hz

System block diagram for 40 – 1280 Hz coverage

8-8 ECE 5650/4650 Modern DSP

 8.2. AVERAGED PERIODOGRAM SPECTRAL ESTIMATION

8.2 Averaged Periodogram Spectral Es-

timation
For continuous-time signals the power spectrum of a wide-sense sta-
tionary random process xc .t / is defined as
Pxx .f / D FfRxx . /g
where Rxx . / is the autocorrelation function of xc .t /. For xc .t / in
general a complex signal, the autocorrelation is given by
Rxx . / D Efx.t C  /x .t /g
 The Fourier transform pair used here is
Z 1
X.f / D x.t /e j 2f t dt
Z 11

x.t/ D X.f /e j 2f t df

1

 An alternate definition for the power spectrum is

EfjXTL .f /j2g
Pxx .f / D lim
TL !1 TL
where XTL .f / D FfxTL .t /g and
(
x.t /; jtj  TL=2
xTL .t / D
0; otherwise

 An estimator for the power spectrum is the periodogram

1
Ixx .f / D jXTL .f /j2
TL
which has units of Watts/Hz

ECE 5650/4650 Modern DSP 8-9

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 The discrete-time equivalent to XTL .f / is the L-point DFT

(FFT) V Œk, 0  k  L 1 and the associated periodogram
2k
Ixx Œk D Ixx .!k /; !k D
L
T
D jV Œkj2; 0  k  L 1
L
where the factor T =L has been included to give Ixx Œk units of
power spectral density (i.e. Watts/ Hz)
Note: Ixx Œk is defined in terms of vŒn D xŒnwŒn where, as
shown in notes chapter 7, wŒn selects an L sample segment of
xŒn

 For analysis purposes it may be easier to deal with the contin-

uous frequency quantities V .e j! / and Ixx .!/

 For a nonuniform window, wŒn, we define the modified peri-

odogram as
T
Ixx Œk D jV Œkj2
LU
where U is a bias normalization factor depending upon the
window function

8.2.1 Periodogram Properties

Since xŒn is a random process the periodogram is also a random
process in ! or index k. The quality of the periodogram estimator
can be determined from the mean and variance of Ixx .!/. Ideally
we desire an estimator that is unbiased and consistent. An estimator
is unbiased if the mean is the quantity we are trying to estimate.

8-10 ECE 5650/4650 Modern DSP

 8.2. AVERAGED PERIODOGRAM SPECTRAL ESTIMATION

An estimator is consistent if as the number of samples used in the

estimate increases the variance goes to zero.

 The expected value of the periodogram can be shown to be


T
Z
EfI.!/g D Pxx . /Cww .e j.! /
/ d
2LU 

where Pxx .!/ is the true power spectrum and Cww .e j! / is the
magnitude squared of the Fourier transformed window func-
tion, wŒn (i.e. Cww .e j! / D jW .e j! /j2)

 As L increases W .e j! / approaches a delta function, and EfIxx .!/g

becomes Pxx .!/ provided U is properly chosen

 By setting the area under U1 jW .e j! /j2 on Π; / equal to one

we can find the proper value for U

L 1
1X
U D .wŒn/2
L nD0

 The variance of Ixx .!/ is difficult to calculate in general. In

Jenkins and Watts2 it is shown that for a wide range of condi-
2
tions the variance is on the order of Pxx .!/

 The conclusion is that although the periodogram is asymptoti-

cally unbiased, the estimator is inconsistent
2
G.M. Jenkins and D.G. Watts, Spectral Analysis and Its Applica-
tions, Holden Day, San Francisco, 1968

ECE 5650/4650 Modern DSP 8-11

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

8.2.2 Averaged Periodograms

To reduce the variance of the periodogram spectral estimate we may
average periodograms computed over K, possibly overlapping, data
segments. Assume the available data sequence is xŒn, 0  n 
Q 1.

 Divide xŒn into K segments each of length L

xr Œn D xŒrR C nwŒn; 0  n  L 1

For R < L the segments overlap, and if R D L they are

contiguous

Data Record of
x[n] Length Q

... K Nonoverlapping
1 2 3 K Data Segments

1
K Overlapping
R Data Segments
2
...
L ...
K

Data record segmentation

 The number of segments K is the largest integer that gives

.K 1/R C .L 1/  Q 1

 The rth segment periodogram is given by

r T
Ixx .!/ D jXr .e j! /j2
LU
8-12 ECE 5650/4650 Modern DSP
 8.2. AVERAGED PERIODOGRAM SPECTRAL ESTIMATION

 The averaged periodogram is defined by

K
X1
1
INxx .!/ D r
Ixx .!/
K rD0

 Since each periodogram is windowed with wŒn the mean of

the averaged periodogram is also biased

 If as an approximation we assume independence between pe-

riodogram samples, then the variance of the averaged peri-
odogram is
1 2
varŒIOxx .!/ ' Pxx .!/
K
Thus the variance will approach zero as K increases

 In summary if both K and L are made large as Q ! 1, then

both the bias and variance can approach zero. In practice Q
is finite so there is a tradeoff between bias (resolution) and
variance

Example 8.3: Two Sinusoids in Noise

Estimate the power spectrum of the signal
A
xŒn D A cosŒ2.n=8/ C cosŒ2.3n=8/ C wŒn
2
where wŒn is a white Gaussian noise sequence with zero mean and
variance w2 . The sinusoid signal-to-noise ratio (SNR) is defined as

A2
SNR D 2
2w
ECE 5650/4650 Modern DSP 8-13
CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

Note that the SNR of the sinusoid with amplitude A=2 is 6 dB lower
in power than the sinusoid with amplitude A.

 For a single sinusoid in white noise a wide-sense stationary

random process model is

zŒn D A cos.!on C  / C wŒn

where  is an independent random variable uniform on Œ0; 2/

 It can be shown that the corresponding power spectrum of zŒn

is
A2  
ı.! !o/ C ı.! C !o/ C w2

Pzz .!/ D
2
– The only difference here being a factor of 2 in just the
spectral line components

 Assuming we wish to display the above power spectrum using

frequency f in Hz, and T D 1 s, we can write
A2 
fo/ C ı.f C fo/ C w2

Pzz .f / D ı.f
4
– The only difference here being a factor of 2 in just the
spectral line components

 For the first set of example results let A D 1, SNR = 0 dB, and
L D 256

 We then consider K D 1, 10, and 50

8-14 ECE 5650/4650 Modern DSP

 8.2. AVERAGED PERIODOGRAM SPECTRAL ESTIMATION

20
L = 256, K = 1, SNR = 0 dB
PSD in dB
10

−10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

20
L = 256, K = 10, SNR = 0 dB
15
PSD in dB

10

−5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
20
L = 256, K = 50, SNR = 0 dB
15
PSD in dB

10
~21 dB
5

−5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalized Frequency /(2
)

Averaged periodogram of two sinusoids in white noise at

SNR = 0 dB
 As K increases the variance is clearly reduced

 The theoretical power spectrum should have a noise spectral

density of
12 1
w2 D D or -3.01 dB
2
100=10 2
ECE 5650/4650 Modern DSP 8-15
CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 The spectral line at f D 1=8 should in theory be a delta func-

tion with area equal to 12=4 or -6.02 dB

 The spectral lines in the averaged periodogram estimate are

both well above 0 dB! why?

 As in a true analog spectrum analyzer we cannot calibrate the

display to correctly show both a continuous spectral density
and the power in a spectral line

 The height of the spectral peak at f D 1=8 is actually the

signal power divided by the resolution bandwidth or FFT bin
width, which in this case is f D fs =L D 1=256 Hz

 The spectral peak in dB should be 10 log10ŒA2=4C10 log10ŒL D

6:02 C 24:08 D 18:06 dB

 The difference between the spectral peak and the noise spectral
density is about 18.06 - (-3.01) = 21.07 dB

 As another special case consider A D 1, L D 1024 and SNR

= -20 dB

8-16 ECE 5650/4650 Modern DSP

 8.2. AVERAGED PERIODOGRAM SPECTRAL ESTIMATION

30
L = 1024, K = 1, SNR = -20 dB
25
PSD in dB
20

15

10

5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

30
L = 1024, K = 10, SNR = -20 dB
PSD in dB

25

20

15

10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

30
L = 1024, K = 50, SNR = -20 dB
25
PSD in dB

~-7 dB
20

15

10
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Normalized Frequency /(2
)

Averaged periodogram of two sinusoids in white noise at

SNR = -20 dB

 It seems an impossible task to discern a sinusoid at a -20 dB

SNR

 The increased spectral resolution makes this possible

ECE 5650/4650 Modern DSP 8-17

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 We expect the noise spectral density to be at

12
w2 D D 50 or 17.0 dB
2
10 20=10
 The spectral peak in dB should be 10 log10Œ1=4C10 log10Œ1024 D
6:02 C 30:10 D 24:08 dB
 The difference between the spectral peak and the noise spectral
density is about 24.08 - (17) = 7.08 dB

8.3 Quantization Effects in DFT Com-

putations
In this section we will study the effects of finite precision arithmetic
in the direct computation of the DFT and also in a radix-2 FFT. The
source of error initially considered is due to rounding after multi-
plication. For a B C 1 bit word length (one sign bit) multiplication
produces a 2B C 1 bit word which we assume is rounded back to
B C 1 bits. A simplified analysis will be performed by assuming
that at each point in the algorithm where coefficient multiplication
occurs we place an additive white noise source.

8.3.1 Quantization Errors in Direct Computation

of the DFT
For the sequence xŒn, 0  n  N 1 the DFT is defined as
N
X1
X Œk D xŒnWNk n; k D 0; 1; : : : ; N 1
nD0

8-18 ECE 5650/4650 Modern DSP

 8.3. QUANTIZATION EFFECTS IN DFT COMPUTATIONS

j.2=N /
where as defined earlier WN D e .

 If we assume that xŒn is a complex valued sequence, then the

product xŒnWNk n requires four real multiplications

 The total complex error due to a twiddle factor multiplication

is denoted eŒn; k
e [ 0, k ]
0
WN = 1
x[0 ]
e [ 1, k ]
k
WN
x[1 ]
e [ 2, k ]
2k
WN
x[2 ]
...

e [ N – 1, k ]
( N – 1 )k
WN
x[N – 1]
X′ [ k ] = X [ k ] + F [ k ]

Linear noise model roundoff noise contributions in the kth

point of an N -point DFT

 As a result of each multiplication/rounding operation an error

sequence ei Œn; k, i D 1, 2, 3, 4 is generated

 The following assumptions are made about each ei Œn; k

– The errors are uniform over the interval Œ , , where

 D 2 B 1 assuming B C 1 bit signed fractions. The
variance is 2 2B =12

ECE 5650/4650 Modern DSP 8-19

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

– The error sequences are mutually uncorrelated from each

other over the 4N multiplies required per DFT point
– The errors are uncorrelated with the input signal and hence
also the output transform

 The complex error has zero mean and variance

2 2B 1 2B
varfeŒn; kg D 4 D 2
12 3

 The quantization error variance associated with each DFT point

is just
N
F2 D 2 2B
3
 To prevent overflow we must limit the input dynamic range. To
avoid overflow in addition suppose that we have jX Œkj < 1,
then
N
X1
jX Œkj  jxŒnj < N; k D 0; 1; : : : ; N 1
nD0

Thus in the worst case we must scale each input by N to avoid

overflow (assuming initially that jxŒnj < 1).

 Further suppose that the input is a white sequence uniform on

Œ 1=N; 1=N 

 The variance of the input is

.2=N /2 1
x2 D D
12 3N 2
8-20 ECE 5650/4650 Modern DSP
 8.3. QUANTIZATION EFFECTS IN DFT COMPUTATIONS

 The variance of the DFT output X Œk is just

1
X2 D Nx2 D
3N
since the twiddle factors have magnitude of one

 Finally the output SNR is

X2 22B
2
D 2
F N

Example 8.4: 1024-Point DFT

In the computation of a 1024-point DFT an SNR of at least 30
dB is required. Find how many bits of precision are required.
 Using the results presented above
30 D 10 log10 2.2B 20/
 

D 3:01.2B 20/ ) B  15

8.3.2 Quantization Errors in the Radix-2 FFT

The radix-2 FFT algorithm requires significantly fewer multiplica-
tions than the direct DFT. From this fact one might initially conclude
that there is less quantization error as well. This is not true unless
stage-by-stage scaling is used.
Consider the equations for a decimation-in-time butterfly
XmŒp D Xm 1Œp C WNr Xm 1Œq
XmŒq D Xm 1Œp WNr Xm 1Œq
The linear noise model for this butterfly is shown below.

ECE 5650/4650 Modern DSP 8-21

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

Xm – 1 [ p ] Xm [ p ]

r
WN
Xm – 1 [ q ] Xm [ q ]
–1
e [ m, q ]

Decimation-in-time butterfly with roundoff noise source

eŒm; q

 Each butterfly noise source is assumed to be zero mean and

white with variance

1 2b
varfeŒn; kg D 2 D B2
3

 Each butterfly noise source is transmitted to the butterfly out-

puts with unity gain, since all the flow graph transmission co-
efficients have unity magnitude

 Note that in the radix-2 FFT each output point (node) is con-
nected directly or indirectly to N 1 butterflies. The 7 butterfly
connections for the point X Œ0 when N D 8 are shown below.

8-22 ECE 5650/4650 Modern DSP

 8.3. QUANTIZATION EFFECTS IN DFT COMPUTATIONS

0
WN
–1 0
WN
–1
0
WN
–1 0
WN
–1
0
WN

–1 0
WN

0
–1
WN

–1

N D 8 butterfly connections for computing X Œ0

 Each butterfly noise source propagates to each output node

with variance B2 . The variance due to quantization noise is
thus
F2 D .N 1/B2 ' NB2 for N large
Note this is the same result as obtained for the direct DFT

 To insure there is no overflow we can again scale each input

point by N

 An improved scaling method can be devised by first consider-

ing the overflow constraints at each butterfly:

max.jXm 1Œpj; jXm 1Œqj/  max.jXmŒpj; jXmŒqj/

ECE 5650/4650 Modern DSP 8-23

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

and

max.jXmŒpj; jXmŒqj/  2 max.jXm 1Œpj; jXm 1Œqj/

Note that the magnitude is nondecreasing from stage to stage.

 Assuming that the FFT outputs must be less than unity and not-
ing that the maximum magnitude increase from stage to stage
is at most a factor of 2, overflow scaling can be accomplished
by:

– Requiring jxŒnj < 1

– Incorporating an attenuation factor of 2 into each butterfly
(overall the scaling is still 1=N )

1 e [ m, p ]
---
2
Xm – 1 [ p ] Xm [ p ]

1 r
--- W N
2
Xm – 1 [ q ] Xm [ q ]
–1
e [ m, q ]

Radix-2 butterfly with overflow scaling and roundoff noise

 The noise variance that each butterfly contributes to the total

output noise, is still B2

 The total noise variance contributed by all N 1 butterflies

will be less than the 1=N scaling case since each noise source
passes through various levels of attenuation before arriving at
an output node

8-24 ECE 5650/4650 Modern DSP

 8.3. QUANTIZATION EFFECTS IN DFT COMPUTATIONS

 There are v butterfly stages, where N D 2v

 A general formula which takes into account the 1/2 scale fac-
tors is
8 9
ˆ
ˆ >
>
<N 1 v 1 N 1 v 2
ˆ     >
=
2 2
F D B C C


C „ƒ‚…1
ˆ 2 4 4 4 >
vth stage
ˆ
ˆ„ ƒ‚ … „ ƒ‚ … >
>
: ;
1st stage 2nd stage
(  )
v 1  v 2
1 1
D B2 C C


C 1
2 2
v 1
X
D B2 .0:5/m
mD0
  v 
1 2
D 2B2 1 ' 2B2 D 2 2B
2 3
 If we assume that the input is a white sequence uniform on
Œ 1; 1. The variance of the input is
1
x2 D
3
 The variance of the DFT output X Œk is still X2 D 1=.3N /
since the twiddle factors have magnitude of one. Hence the
output SNR is
X2 22B
D
F2 2N

Example 8.5: 1024-Point Radix-2 FFT

In the computation of a 1024-point DFT an SNR of at least 30
dB is required. Find how many bits of precision are required.

ECE 5650/4650 Modern DSP 8-25

CHAPTER 8. ADVANCED FFT TOPICS AND APPLICATIONS

 Using the results presented above

 .2B 11/
30 D 10 log10 2
D 3:01.2B 11/ ) B  11

 This compares with B  15 required for the direct DFT or

FFT with all the scaling performed on the input

8-26 ECE 5650/4650 Modern DSP