7 Kundur FFT

Complexity of Filtering and the FFT
Complexity of Filtering in the Time-Domain
Digital Filtering in the Time Domain

Let x(n) and h(n) be real signals.
Let the support of x(n) be n = 0, 1, . . . , N 1. We are interested in
determining y (n) for n = 0, 1, . . . , N 1.

Dr. Deepa Kundur
y (n) = x(n) h(n)
X
=
x(k)h(n k)
University of Toronto
k=
N1
X
x(k)h(n k)
n = 0, 1, . . . , N 1
k=0
Dr. Deepa Kundur (University of Toronto)
1 / 34
Digital Filtering in the Time Domain
k=0
Is O(N 2 ) high?
I
N real multiplications
N 1 real additions
For all n (n=0, 1, . . . , N-1):

I
2 / 34
Complexity of Digital Filtering in the Time Domain
Complexity of doing a brute-force convolution is given by:

I For fixed n:
N1
X
y (n) =
x(k)h(n k)
Yes.
Idea: Maybe filtering in the frequency domain can reduce

complexity.
NN = N 2 = O(N 2 ) real multiplications

(N 1)N = N(N 1) = O(N 2 ) real additions
3 / 34
4 / 34
DFT
Discrete Fourier Transform (DFT)
Fourier Duality
Frequency analysis of discrete-time signals is conveniently

performed on a DSP.
Therefore, both time-domain and frequency-domain signals must
be discrete.
Time Domain
sinc
rectangle
Frequency Domain
rectangle
sinc
sinc2
triangle
triangle
sinc2
ringing
truncation
truncation
ringing
discrete
periodic
periodic
discrete
continuous
aperiodic
aperiodic
continuous
sampling
x(t) x(n)
X () X ( 2k
N ) or X (k)
DFT
sampling
What happens when we sample in the frequency domain?
5 / 34
DFT
DFT Intuition
among others . . .
6 / 34
DFT
DFT Intuition
Example
aperiodic + dst in time cts + periodic in freq

sampling
periodic + dst in freq-domain
periodic + dst in time-domain
periodic + dst in time dst + periodic in freq
DTFS
1
DTFS
periodic + discrete periodic + discrete
-7
-6
-5
-4
-3
-2
-1
-7
-6
DFT
x(n)
-5
-4
-3
-2
-1
X(k)
DFT
one period of dst samples one period of dst samples

n = 0, 1, . . . , N 1
k = 0, 1, . . . , N 1
note: signal examples are artificial
7 / 34
8 / 34
DFT
Frequency Domain Sampling

I
Frequency Domain Sampling
Recall, sampling in time results in a periodic repetition in

frequency.
x(n) = xa (t)|t=nT
1 X
2
X () =
Xa ( +
k)
T k=
T
Similarly, sampling in frequency results in periodic repetition in

time.
xp (n) =
x(n + lN)
... -1
X (k) = X ()|= 2 k
N
l=
...
9 / 34
DFT
...
Frequency Domain Sampling and Reconstruction
10 / 34
DFT
Therefore,
I
x(n) X ()
Q: Can we reconstruct x(n) from the samples of X ()?

I
xp (n) X (k)
Can we reconstruct x(n) from xp (n)?
A: Maybe.
"
xp (n) =
N-1
Note: N is directly proportional to the sampling rate of ; there are N samples

per 2
DFT
#
x(n + lN)
l=
Implications:
I
The samples of X () can be used to reconstruct xp (n).
11 / 34
12 / 34
DFT
DFT
N =4
x(n)
xp(n)
no temporal aliasing
-7
-6
-5
-4
-3
-2
-1
-7
-6
-5
-4
-3
-2
-1
x(n) can be recovered from xp (n) if there is no overlap when

taking the periodic repetition.
If x(n) is finite duration and non-zero in the interval
0 n L 1, then
x(n) = xp (n),
=x(n)
x(n)
time-domain aliasing
-6
-5
-4
-3
-2
-1
-7
-6
-5
-4
-3
-2
-1
when N L
xp(n)
I
If N < L then, x(n) cannot be recovered from xp (n).

I
-7
0n N 1
or equivalently
X () cannot be recovered from its samples

k
due
to
time-domain aliasing
X 2
N
=x(n)
13 / 34
DFT
xp (n) for all n

x(n)
where

x(n) =
and
(k) =
X
DTFT, DTFS and DFT
x(n) for all n
14 / 34
DFT
The Discrete Fourier Transform Pair
DTFT
DTFS
DFT
X () for all
DFT and inverse-DFT (IDFT):
X (k) for all k
X (k) =
(k)
X
xp (n)
0
for n = 0, . . . , N 1
otherwise
X (k)
0
for n = 0, . . . , N 1
otherwise
x(n) =
N1
X
n=0
N1
X
1
N
x(n)e j2k N ,
k = 0, 1, . . . , N 1
n
X (k)e j2k N ,
n = 0, 1, . . . , N 1
k=0
Note: we drop the notation from now on.
15 / 34
16 / 34
DFT
Important DFT Properties
Property
Notation:
Periodicity:
Linearity:
Time reversal
Circular time shift:
Circular frequency shift:
Complex conjugate:
Circular convolution:
Multiplication:
Parsevals theorem:
Complexity of the DFT (and IDFT)
Time Domain
x(n)
x(n) = x(n + N)
a1 x1 (n) + a2 x2 (n)
x(N n)
x((n l))N
x(n)e j2ln/N
x (n)
x1 (n) x2 (n)
x1 (n)x2 (n)
P
N1
n=0 x(n)y (n)
X (k) =
Frequency Domain
X (k)
X (k) = X (k + N)
a1 X1 (k) + a2 X2 (k)
X (N k)
X (k)e j2kl/N
X ((k l))N
X (N k)
X1 (k)X2 (k)
1
X1 (k) X2 (k)
N P
N1
1
k=0 X (k)Y (k)

N
x(n) =
X (k) =
17 / 34
X (k) =
n=0
N1
X
1
N
n = 0, 1, . . . , N 1
1
N
k = 0, 1, . . . , N 1
X (k)WNkn ,
n = 0, 1, . . . , N 1
k=0
18 / 34
DFT
Complexity of the DFT
k = 0, 1, . . . , N 1
X (k) =
N1
X
x(n)WNkn ,
k = 0, 1, . . . , N 1
n=0
Straightforward implementation of DFT to compute X (k) for

k = 0, 1, . . . , N 1 requires:
I N(N 1) complex additions
1 complex mult =
(aR + jaI )(bR + jbI ) = (aR bR aI bI ) + j(aR bI + aI bR )
= 4 real mult + 2 real add
4N 2 = O(N 2 ) real multiplications
x(n)WNkn ,
Straightforward implementation of DFT to compute X (k) for

k = 0, 1, . . . , N 1 requires:
I N 2 complex multiplications
X (k)e j2k N ,
n=0
N1
X
n=0
k = 0, 1, . . . , N 1
k=0
N1
X
DFT
x(n)WNkn ,
x(n)e j2k N ,

N1
X
N1
X
New notation: WN = e j N
x(n) =
DFT
19 / 34
1 complex add =
(aR + jaI )+(bR + jbI ) = (aR + bR ) + j(aI + bI ) = 2 real add
2N(N 1) + 2N 2 (from complex mult) real additions
= 2N(2N 1) = O(N 2 ) real additions.
20 / 34
DFT

I
How can we reduce complexity?

I
Is O(N 2 ) high?
Exploit symmetry of the complex exponential.

k+ N2
WN
I
FFT
k+ N2
Yes. A linear increase in the length of the DFT increases the

complexity by a power of two.
Given the multitude of applications where Fourier analysis is
employed (linear filtering, correlation analysis, spectrum
analysis), a method of efficient computation is needed.
LHS = WN
= WNk
= e j2
k+N/2
N
= e j2 N e j2
N/2
N
= e j2 N e j
k
= e j2 N (cos() + j sin())
k
= e j2 N (1)
k
= e j2 N = WNk = RHS
21 / 34
FFT

I
22 / 34
FFT
Radix-2 FFT: Decimation-in-time
How can we reduce complexity?

I
Exploit periodicity of the complex exponential.
LHS =
WNk+N
WNk+N
X (k) =
j2 k+N
N
=e
j2 Nk j2 N
N
j2 Nk
(cos(2) + j sin(2))
x(n)WNkn
n odd
(N/2)1
k(2m)
x(2m)WN
m=0
X
m=0
23 / 34
k(2m+1)
x(2m + 1)WN
m=0
(N/2)1
(N/2)1
= e j2 N = WNk = RHS
(N/2)1
= e j2 N (1)
k = 0, 1, . . . , N 1
x(n)WNkn +
n even
= e j2 N e j2
= e
x(n)WNkn
n=0
WNk
= e
N1
X
x(2m) WN2km +
| {z }
f1 (m)
X
m=0
x(2m + 1) WN2km WNk

| {z }
f2 (m)
24 / 34
FFT

Note:
WN2
=e
j 2
2
N
=e
2
j N/2
X (k) =
X
m=0
= WN/2
x(2m) WN2km +
| {z }
f1 (m)
| m=0
X
m=0
N
)
2
N
F2 (k) = F2 (k + )
2
x(2m + 1) WN2km WNk

| {z }
f2 (m)
(N/2)1
km
f1 (m)WN/2
{z
N
DFT
2
of f1 (m)
+WNk
= F1 (k) + WNk F2 (k),
X
| m=0
N
-DFTs:
2
F1 (k) = F1 (k +
(N/2)1
(N/2)1

Note: since F1 (k) and F2 (k) are
(N/2)1
FFT
we have,
km
f2 (m)WN/2
{z
N
DFT
2
of f2 (m)
X (k) = F1 (k) + WNk F2 (k)

N
N
N
k+ N
X (k + ) = F1 (k + ) + WN 2 F2 (k + )
2
2
2
= F1 (k) WNk F2 (k)
k = 0, 1, . . . , N 1
since WNk+ 2 = e j N (k+ 2 ) = e j N k e j N

25 / 34
FFT
= e j
2
N k
(1) = WNk
2 N
2
26 / 34
FFT

Repeating the decimation-in-time for f1 (n) and f2 (n), we obtain:
Therefore,
N
1
2
N
N
X (k + ) = F1 (k) WNk F2 (k) k = 0, 1, . . . , 1
2
2
X (k) = F1 (k) + WNk F2 (k) k = 0, 1, . . . ,
v11 (n)
= f1 (2n) n = 0, 1, . . . , N/4 1
v12 (n)
= f1 (2n + 1)
v21 (n)
= f2 (2n) n = 0, 1, . . . , N/4 1
v22 (n)
= f2 (2n + 1)
n = 0, 1, . . . , N/4 1
n = 0, 1, . . . , N/4 1
and
k
F1 (k) = V11 (k) + WN/2
V12 (k) k = 0, 1, . . . , N/4 1
F1 (k + N/4)
F2 (k)
F2 (k + N/4)
k
V11 (k) WN/2
V12 (k) k = 0, 1, . . . , N/4 1
k
= V21 (k) + WN/2
V22 (k) k = 0, 1, . . . , N/4 1
k
V21 (k) WN/2
V22 (k) k = 0, 1, . . . , N/4 1
consisting of N/4-DFTs.
27 / 34
28 / 34
FFT
FFT
For N = 8.
For N = 8.
x(0)
x(4)
x(2)
x(6)
Combine
2-point
DFTs
2-point
DFT
x(1)
x(5)
Combine
4-point
DFTs
2-point
DFT
x(3)
x(7)
Combine
2-point
DFTs
2-point
DFT
Stage 1
x(0)
2-point
DFT
x(4)
X(0)
X(1)
X(2)
X(3)
X(4)
X(5)
X(6)
X(7)
Stage 3
X(0)
X(1)
-1
0
W8
x(2)
X(2)
-1
2
x(6)
W8
0
W8
-1
x(5)
0
W8
-1
1
W8
0
W8
-1
-1
0
FFT
2
W8
0
-1
-1
3
W8
-1
X(5)
-1
X(6)
X(7)
FFT Complexity
W8
-1
W8
X(4)
W8
x(3)
29 / 34
X(3)
-1
x(1)
x(7)
W8
Stage 2
30 / 34
FFT
Convolution using FFT
A=a
To compute the convolution of x(n) (support: n = 0, 1, . . . , L 1)

and h(n) (support: n = 0, 1, . . . , M 1):
+ W Nr b
1. Assign N to be the smallest power of 2 such that N = 2r M + L 1.
r
WN
-1
A = a - W Nr b
2. Zero pad both x(n) and h(n) to have support n = 0, 1, . . . , N 1.

3. Take the N-FFT of x(n) to give X (k), k = 0, 1, . . . , N 1.
Each butterfly requires:

I
I
5. Produce Y (k) = X (k) H(k), k = 0, 1, . . . , N 1.
In total, there are:

I
I
4. Take the N-FFT of h(n) to give H(k), k = 0, 1, . . . , N 1.
one complex multiplication

two complex additions
N
2
butterflies per stage

log N stages
6. Take the N-IFFT of Y (k) to give y (n), n = 0, 1, . . . , N 1.
Order of the overall DFT computation is: O(N log N)
31 / 34
32 / 34
FFT
Convolution using FFT
FFT
Complexity of Convolution using FFT
To compute the convolution of x(n) (support: n = 0, 1, . . . , L 1)

and h(n) (support: n = 0, 1, . . . , M 1):
1. Assign N to be the smallest power of 2 such that N = 2r M + L 1.
2. Zero pad both x(n) and h(n) to have support n = 0, 1, . . . , N 1.
O(1)
Therefore, the overall complexity of conducting convolution via the

FFT is:
3. Take the N-FFT of x(n) to give X (k), k = 0, 1, . . . , N 1.
O(N log N)
4. Take the N-FFT of h(n) to give H(k), k = 0, 1, . . . , N 1.

O(N log N)
which is lower than O(N 2 ) through direction computation of

convolution in the time-domain.
5. Produce Y (k) = X (k) H(k), k = 0, 1, . . . , N 1.

O(N)
6. Take the N-IFFT of Y (k) to give y (n), n = 0, 1, . . . , N 1.

O(N log N)
33 / 34
34 / 34

7 Kundur FFT

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

7 Kundur FFT

Загружено:

Авторское право:

Доступные форматы

Complexity of Filtering and the FFT

Complexity of Filtering in the Time-Domain

Digital Filtering in the Time Domain

Complexity of Filtering and the FFT

y (n) = x(n) h(n)

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering in the Time-Domain

Complexity of Filtering and the FFT

Digital Filtering in the Time Domain

Complexity of Filtering in the Time-Domain

For all n (n=0, 1, . . . , N-1):

Complexity of Digital Filtering in the Time Domain

Complexity of doing a brute-force convolution is given by:

Complexity of Filtering and the FFT

Idea: Maybe filtering in the frequency domain can reduce

NN = N 2 = O(N 2 ) real multiplications

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Discrete Fourier Transform (DFT)

Frequency analysis of discrete-time signals is conveniently

What happens when we sample in the frequency domain?

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

aperiodic + dst in time cts + periodic in freq

periodic + dst in freq-domain

periodic + dst in time-domain

periodic + dst in time dst + periodic in freq

periodic + discrete periodic + discrete

one period of dst samples one period of dst samples

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

note: signal examples are artificial

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Frequency Domain Sampling

Frequency Domain Sampling

Recall, sampling in time results in a periodic repetition in

Similarly, sampling in frequency results in periodic repetition in

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Dr. Deepa Kundur (University of Toronto)

Complexity of Filtering and the FFT

Complexity of Filtering and the FFT

Frequency Domain Sampling and Reconstruction

Frequency Domain Sampling and Reconstruction

Q: Can we reconstruct x(n) from the samples of X ()?

Can we reconstruct x(n) from xp (n)?

Note: N is directly proportional to the sampling rate of ; there are N samples

Complexity of Filtering and the FFT

The samples of X () can be used to reconstruct xp (n).