This action might not be possible to undo. Are you sure you want to continue?
ITS APPLICATIONS
MajorProject Report
by
Alex John Koshy B050326EC
Nidhin Chandran A K B050160EC
Subin B B050173EC
Vinay N K B050032EC
Under the guidance of
Dr. G. Abhilash
In Partial Fulﬁllment of the Requirements
for the Degree of
Bachelor of Technology
DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING
NATIONAL INSTITUTE OF TECHNOLOGY, CALICUT
Kerala, India
April 2009
i
NATIONAL INSTITUTE OF TECHNOLOGY CALICUT
DEPARTMENT OF ELECTRONICS AND COMMUNICATION
ENGINEERING
CERTIFICATE
This is to certify that this report titled FRACTIONAL FOURIER TRANS
FORM AND ITS APPLICATIONS is a bona ﬁde record of the majorproject
done by Alex John Koshy (Roll No. B050326EC), Nidhin Chandran A K
(Roll No. B050160EC), Subin B (Roll No. B050173EC) and Vinay N K
(Roll No. B050032EC), in partial fulﬁllment of the requirements for the award
of Degree of Bachelor of Technology in Electronics and Communication Engineering
from National Institute of Technology, Calicut.
Dr. G. Abhilash Dr. Lillykutty Jacob
(Project Advisor) Professor and Head
Assistant Professor
29 April 2009
NIT Calicut
ii
ACKNOWLEDGEMENT
We would like to thank Dr. G.Abhilash, Assistant Professor, Department of Elec
tronics and Communication Engineering for his guidance and inspiration in helping
us complete this project. We are also grateful to Dr. Lillykutty Jacob, Professor
and Head, Department of Electronics and Communication Engineering for providing
us with this opportunity to work on our project and also for permitting access to the
required facilities. We would also like to thank the lab staﬀ for their technical support
and providing us assistance. We also thank our batch mates who had supported us
and provided us with greatly appreciated technical and nontechnical aid throughout
our project.
Alex John Koshy
Nidhin Chandran A K
Subin B
Vinay N K
iii
Abstract
Signals can be viewed from diﬀerent perspectives using diﬀerent transforms. The
Fourier transform which allows us to observe the signal in terms of diﬀerent fre
quency components is widely used in signal processing and communication. Frac
tional Fourier Transform (FrFT) is a generalized Fourier transform which allows us
to take transforms of fractional order also.
Theory of FrFT is developed from Fourier transform. For computation we move from
continuous to discrete domain and theory of Discrete Fractional Fourier Transform
(DFrFT) is discussed. A time eﬃcient algorithm for DFrFT is also discussed. Theory
of designing optimal ﬁlter using FrFT is developed from ﬁrst principles. Optimal
ﬁlters are designed for some special cases and their performance is analyzed. Some
applications other than ﬁltering is also discussed.
iv
Contents
Abstract iii
1 Introduction 1
1.1 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 TimeFrequency Plane . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2 Fractional Fourier Transform 4
2.1 Fractional Fourier Transform from Fourier Transform . . . . . . . . . 4
2.1.1 Orthogonality of Eigenfunctions . . . . . . . . . . . . . . . . . 7
2.1.2 Orthonormality of Eigenfunctions . . . . . . . . . . . . . . . . 8
2.1.3 Completeness of Eigenfunctions . . . . . . . . . . . . . . . . . 9
2.2 Equation for Fractional Fourier Transform . . . . . . . . . . . . . . . 11
3 Discrete Fractional Fourier Transform 13
3.1 TimeEﬃcient Algorithm for ﬁnding FrFT . . . . . . . . . . . . . . . 13
3.2 Implementing TimeEﬃcient Algorithm in Matlab . . . . . . . . . . . 15
3.2.1 Matlab Code for FrFT . . . . . . . . . . . . . . . . . . . . . . 15
3.2.2 Fractional Fourier Transform of Delta function . . . . . . . . . 17
3.2.3 Fractional Fourier Transform of Sine function . . . . . . . . . 19
3.2.4 Fractional Fourier Transform of rectangular function . . . . . 21
4 Filtering in Fractional Fourier Domain 23
4.1 Filtering in Fractional Fourier Domain . . . . . . . . . . . . . . . . . 23
v
4.2 Optimal Filter Design . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2.1 Example 1 : Chirp signal contaminated with White Gaussian
Noise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1.1 Observations . . . . . . . . . . . . . . . . . . . . . . 32
4.2.1.2 Matlab Code . . . . . . . . . . . . . . . . . . . . . . 32
4.2.2 Example 2 : Square pulse in Linear FM noise. . . . . . . . . . 34
4.2.2.1 Observation . . . . . . . . . . . . . . . . . . . . . . . 35
4.2.2.2 Matlab Code . . . . . . . . . . . . . . . . . . . . . . 35
4.3 Other Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.1 FrFT for Compression . . . . . . . . . . . . . . . . . . . . . . 40
4.3.2 Multipath Channel Estimation Using FrFT . . . . . . . . . . . 40
4.3.2.1 Matlab Code . . . . . . . . . . . . . . . . . . . . . . 41
4.3.3 FrFT for measuring the acceleration of a moving object in radial
direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5 Critical Evaluation and Conclusion 46
5.1 Critical Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Bibliography 48
vi
List of Figures
3.1 (a)Delta Function Input (b)FrFT at 22.5
0
(c)FrFT at 45
0
. . . . . . . . 17
3.2 (a) FrFT of Delta Function at 67.5
0
(b)FrFT of Delta Function at 90
0
18
3.3 (a) Sinusoidal Input, (b) FrFT at 22.5
0
for Sinusoidal Input (c)FrFT at
45
0
for Sinusoidal Input . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 (a) FrFT at 67.5
0
for Sinusoidal Input (b) FrFT at 90
0
for Sinusoidal
Input (c) Inverse FrFT . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.5 (a) Rectangular Function Input (b) FrFT at 22.5
0
for Rectangular Func
tion Input (c) FrFT at 45
0
for Rectangular Function Input (d) FrFT at
67.5
0
for Rectangular Function Input . . . . . . . . . . . . . . . . . . . 21
3.6 (a) FrFT of Rect at 90
0
, (b) FrFT  Magniﬁed view of (a), (c) Inverse
FrFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.1 (a) Gaussian Signal (b) Chirp noise . . . . . . . . . . . . . . . . . . . . 24
4.2 (a) Gaussian Signal with chirp noise (b) FrFT of a at 72
0
(c) Part of
FrFT corresponding to chirp . . . . . . . . . . . . . . . . . . . . . . . . 25
4.3 (a) After windowing out FrFT corresponding to chirp (b) Chirp ex
tracted by inverse FrFT (c) Gaussian extracted by inverse FrFT of (a) 26
4.4 (a) Chirp Signal (b)WGN added to Chirp Signal SNR = 6dB (c) Frac
tional Fourier Transform of the signal at −82.5
0
s . . . . . . . . . . . . 30
4.5 (a)Optimum Multiplicative Filter (b)Extracted Chirp Signal SNR =
13dB (c) Variation of output noise power with fractional order . . . . . 31
4.6 (a) Square wave + Chirp noise, SNR = 20dB (b) FrFT at −76.5
0
. . . 35
4.7 (a) Optimal Filter (b) After denoising, SNR=6 dB (c) Error After De
modulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
vii
4.8 (a) Transmitted chirp signal (b) Received signalMulti Path, WGN SNR
= 6 dB (c) Signal after FrFT for coeﬃcients 1, 0.8, 0.6 Angle = −80
0
42
1
Chapter 1
Introduction
1.1 Transforms
Transforms play vital roles in the analysis of signals and systems. Transforms help in
ﬁnding out the hidden properties of a signal, which are unrecognizable from the time
domain representation of the signal. The choice of the transform depends on the type
of the signal and the application. Fourier Transform, Laplace Transform, Wavelet
Transform etc are some of the widely used transform in the ﬁeld of communication,
signal processing and system design. Among these transforms Fourier transform is
most fundamental one.
1.2 Fourier Transform
The relation between output y(t) and input x(t) for a Linear Time Invariant system
is of the form
m
k=0
a
k
d
k
dt
k
y(t) =
n
k=0
b
k
d
k
dt
k
x(t) (1.1)
We can see that if the input is of the form e
jωt
, output will follow the form k(ω)e
jωt
,
where k(ω) is a complex value which is independent of time and depends only on ω.
We consider e
jωt
as the component corresponding to frequency ω in the signal. In
order to understand how a LTI system works on a particular signal we have to get an
idea about the diﬀerent frequency components present in the signal. It can be found
2
as
¸
x(t), e
jωt
_
=
1
√
2π
_
+∞
−∞
x(t)e
−jωt
dt = X(ω) (1.2)
X(ω) is the Fourier transform of the signal x(t).
We can obtain the signal from its Fourier transform as follows.
x(t) =
1
√
2π
_
+∞
−∞
X(ω)e
jωt
dω (1.3)
1.3 TimeFrequency Plane
For convenience, we deﬁne Fourier transform as,
F [x(t)] =
1
√
2π
_
+∞
−∞
x(τ)e
−jτt
dτ (1.4)
Consider the following property of Fourier transform.
F[x(t)] = X(t) (1.5)
FF[x(t)] = F
2
[x(t)] = x(−t) (1.6)
FFF[x(t)] = F
3
[x(t)] = X(−t) (1.7)
FFFF[x(t)] = F
4
[x(t)] = x(t) (1.8)
Taking Fourier transform 
• once  gives the representation of signal in the frequency domain.
• twice  gives representation of signal in the time domain with time axis reversed.
• three times  gives the representation of signal in the frequency domain with
frequency axis reversed.
• four times  retains the signal in the time domain itself.
3
This observation can be easily visualized using the idea of timefrequency plane.
Consider a plane with time as horizontal axis. As F
2
gives the representation along
inverted time axis which is at an angle of 180
0
with the original time axis, F should
give the representation along 90
0
rotated axis. i.e., Frequency axis is the vertical
axis here. Each time we take the Fourier transform, we get the representation of the
signal along an axis which is at an angle of 90
0
with the original axis. Now the curious
question is, is it possible to get the Fourier transform for intermediate angles? We
will see the answer in the next chapter.
4
Chapter 2
Fractional Fourier Transform
Fourier transform allows us to rotate the axis of representation of signal in TF plane
by integral multiples of 90
0
only. If we can generalize this rotation operation for
intermediate angles also, then we will be able to use it for analyzing a wider class of
signals. Let F
p
represent taking Fourier transform p times. We call p as the order of
the Fourier transform operation.
If,
• p = 1, then rotation is by
π
2
• p = 2, then rotation is by π
i.e., rotation obtained for p
th
order is p(
π
2
). Fourier transform allows p to take only
integer values. Fourier transform which is generalized so that p can take any fractional
value is called Fractional Fourier Transform (FrFT).
2.1 Fractional Fourier Transform from Fourier Trans
form
Consider a cascaded structure of LTI systems of identical response H(ω). If the
input is of the form e
jωt
, output y(t) will be [H(ω)]
N
e
jωt
, where N is the number of
systems in cascade. Suppose we are making a strange assumption that, there are only
1.5 systems. Then we may have to write y(t) = [H(ω)]
1.5
e
jωt
. We are not discussing
5
whether such a system is possible or not, but this assumption gives us a clue for
proceeding to Fractional Fourier transform from Fourier transform.[1][7]
Let ϕ(t) be a function that satisﬁes the relation
F[ϕ(t)] = ρ(t) = γϕ(t) (2.1)
Then,
F
2
[ϕ(t)] = γ
2
ϕ(t) (2.2)
F
n
[ϕ(t)] = γ
n
ϕ(t) (2.3)
We can generalize this relation for fractional orders also, as, for fractional order p,
F
p
[ϕ(t)] = γ
p
ϕ(t) (2.4)
For any function that satisﬁes equation 2.1, we can ﬁnd the FrFT using this method.
Let ϕ
i
(t) be a set of functions such that,
F
p
[ϕ
i
(t)] = ρ
i
(t) = γ
p
i
ϕ
i
(t) (2.5)
And q(t) is a function of the form
q(t) =
i
a
i
ϕ
i
(t) (2.6)
Now if we take Fourier Transform,
F[q(t)] = F
_
i
a
i
ϕ
i
(t)
_
=
i
a
i
F [ϕ
i
(t)] =
i
a
i
γ
i
ϕ
i
(t) (2.7)
Generalizing this for order p, we get
F
p
[q(t)] = F
p
_
i
a
i
ϕ
i
(t)
_
=
i
a
i
F
p
[ϕ
i
(t)] =
i
a
i
γ
i
p
ϕ
i
(t) (2.8)
6
Now our aim is to obtain such a set of functions ϕ
i
(t). It is clear that each ϕ
i
(t) is
an eigenfunction of Fourier transform operator. i.e, ϕ
i
(t) is invariant under Fourier
transform. It is well known that the Gaussian function, e
−t
2
2
, is a function which is
invariant under Fourier transform.
We deﬁne two operators Φ and Ψ such that,
Φ =
d
dt
and Ψ = t (2.9)
Now,
FΦ[x(t)] = F
_
d
dt
[x(t)]
_
= jtF[x(t)] = jΨF[x(t)] (2.10)
Similarly,
FΨ[x(t)] = F[tx(t)] = j
d
dt
[F[x(t)]] = jΦF[x(t)] (2.11)
On subtracting (2.11) from (2.10), we get
F(Φ −Ψ)[x(t)] = −j(Φ −Ψ)F[x(t)] (2.12)
Suppose that a function ϕ(t) satisﬁes the relation,
F[ϕ(t)] = ρ(t) = γϕ(t) (2.13)
Then,
F(Φ−Ψ)[ϕ(t)] = −j(Φ−Ψ)F[ϕ(t)] = −j(Φ−Ψ)γϕ(t) = −jγ(Φ−Ψ)[ϕ(t)] (2.14)
i.e., We get a new eigenfunction ϕ
new
(t) and a new eigenvalue γ
new
given by,
ϕ
new
(t) = (Φ −Ψ)[ϕ(t)] (2.15)
γ
new
= −jγ (2.16)
7
This shows us how to generate a class of eigenfunctions from a single eigenfunction.
Let us start from Gaussian
ϕ
o
(t) = e
−t
2
2
= e
t
2
2
e
−t
2
(2.17)
We know,
F[ϕ
0
(t)] = ϕ
0
(t) (2.18)
Using the above generating equation, we get,
ϕ
1
(t) = (Φ −Ψ)[ϕ
0
(t)] (2.19)
⇒ ϕ
1
(t) =
_
d
dt
−t
_
ϕ
0
(t) (2.20)
⇒ ϕ
1
(t) =
_
d
dt
−t
_
e
t
2
2
e
−t
2
(2.21)
⇒ ϕ
1
(t) = e
t
2
2
_
d
1
dt
1
e
−t
2
_
(2.22)
Similarly we get,
ϕ
2
(t) = e
t
2
2
_
d
2
dt
2
e
−t
2
_
(2.23)
In general,
ϕ
n
(t) = e
t
2
2
_
d
n
dt
n
e
−t
2
_
(2.24)
2.1.1 Orthogonality of Eigenfunctions
Theorem 1 : {ϕ
i
(t)} forms an orthogonal system.
Proof : We denote
ϕ
n
(t) = e
t
2
2
d
n
dt
n
e
−t
2
. (2.25)
The inner product
ϕ
n
, ϕ
m
=
_
∞
−∞
e
t
2 d
n
dt
n
e
−t
2 d
m
dt
m
e
−t
2
dt (2.26)
8
can be evaluated using integration by parts, which gives
ϕ
n
, ϕ
m
=
_
e
t
2 d
n
dt
n
e
−t
2 d
m−1
dt
m−1
e
−t
2
_
∞
−∞
−
_
∞
−∞
d
dt
_
e
t
2 d
n
dt
n
e
−t
2
_
d
m−1
dt
m−1
e
−t
2
dt (2.27)
and hence all the terms under the diﬀerential sign contain the factor e
−t
2
. Since for
any k ∈ N, we have
t
k
e
−t
2
→ 0 as t → ∞ (2.28)
the ﬁrst term in equation 2.27 vanishes. Therefore, repeated integration by parts
leads to
ϕ
n
, ϕ
m
= 0 for n = m (2.29)
2.1.2 Orthonormality of Eigenfunctions
To obtain an orthonormal system we evaluate the norm
ϕ
n
2
=
_
∞
−∞
e
−t
2
_
d
n
dt
n
e
−t
2
_
2
dt. (2.30)
Integration by parts n times yields
ϕ
n
2
= (−1)
n
_
∞
−∞
e
−t
2 d
n
dt
n
_
d
n
dt
n
e
−t
2
_
dt. (2.31)
Since
d
n
dt
n
e
−t
2
is a polynomial of degree n, direct diﬀerentiation gives,
e
t
2 d
n
dt
n
e
−t
2
= (−2t)
n
+ .... (2.32)
and
d
n
dt
n
_
d
n
dt
n
e
−t
2
_
=
d
n
dt
n
((−2t)
n
+ ....) = (−1)
n
2
n
n!. (2.33)
Consequently,
ϕ
n
2
= 2
n
n!
_
∞
−∞
e
−t
2
dt = 2
n
n!
√
π (2.34)
9
Thus the functions
n
(t) =
1
_
2
n
n!
√
π
e
−t
2
2
d
n
dt
n
e
−t
2
(2.35)
form an orthonormal system in L
2
(R).
2.1.3 Completeness of Eigenfunctions
Theorem 2 : {ϕ
i
(t)} is complete in L
2
(R).
Proof :
We deﬁne a generating function G(z, t) for ϕ
i
(t) as
G(z, t) =
∞
i=0
ϕ
i
(t)
z
i
i!
(2.36)
_
d
dt
−t
_
G(z, t) =
∞
i=0
ϕ
i+1
(t)
z
i
i!
(2.37)
d
dz
G(z, t) =
∞
i=0
ϕ
i+1
(t)
z
i
i!
(2.38)
_
d
dt
−t
_
G(z, t) =
d
dz
G(z, t) (2.39)
Multiplying the above equation with e
−t
2
/2
on both sides and rearranging,
d
dt
_
e
−t
2
2
G(z, t)
_
=
d
dz
_
e
−t
2
2
G(z, t)
_
(2.40)
Let
H(z, t) = e
−t
2
2
G(z, t) (2.41)
Then,
d
dt
H(z, t) =
d
dz
H(z, t) (2.42)
Let H(z, t) be of the form H(z, t) = g(z)h(t)
d
dt
g(z)h(t) =
d
dz
g(z)h(t) (2.43)
10
⇒
g
(z)
g(z)
=
h
(t)
h(t)
= k. (2.44)
Where k is a constant.
⇒ g(z) = e
kz
and h(t) = e
kt
(2.45)
⇒ H(z, t) = e
k(z+t)
(2.46)
Notice that, the solution of H(z, t) can be in general of the form
H(z, t) = e
k(z+t)
n
(2.47)
G(z, t) = e
t
2
2
H(z, t) = e
t
2
2
e
k(z+t)
n
(2.48)
For z = 0, G(z, t) = ϕ
0
(t) = e
−t
2
2
⇒ k = −1 and n = 2
Final generating function is
G(z, t) = e
t
2
2
e
−(z+t)
2
(2.49)
Putting z = −
1
2
y gives a new generating function,
¯
G(y, t) = e
−
1
2
(y−t)
2
e
y
2
4
= e
y
2
4
ϕ
0
(y −t) (2.50)
Let there exist a function p(t) in L
2
(R) which is orthogonal to ϕ
i
(t), ∀ i ∈ Z.
ϕ
i
(t), p(t) = 0 (2.51)
_
¯
G(y, t), p(t)
_
= 0 ∀ y ∈ R (2.52)
_
∞
−∞
¯
G(y, t)p(t)dt = 0 (2.53)
_
∞
−∞
ϕ
0
(y −t)e
y
2
4
p(t)dt = 0 (2.54)
11
_
∞
−∞
ϕ
0
(y −t)p(t)dt = 0 (2.55)
p(y) ∗ ϕ
0
(y) = 0 (2.56)
Taking Fourier transform,
P(ω)ϕ
0
(ω) = 0 (2.57)
But, ϕ(ω) = 0
⇒ p(ω) = 0 (2.58)
⇒ p(y) = 0 (2.59)
i.e., It is not possible to ﬁnd a nonzero function in L
2
(R) which is orthogonal to
ϕ
i
(t), ∀ i ∈ Z. That is, {ϕ
i
(t), i ∈ Z} is dense in L
2
(R).
2.2 Equation for Fractional Fourier Transform
Any signal g(t) in L
2
(R) can be represented as the linear combination of
n
(t).
Mathematically,
g(t) =
∞
n=0
a
n
n
(t) (2.60)
Where,
a
n
= g(t),
n
(t) (2.61)
a
n
can be written as
a
n
=
_
+∞
−∞
g(t)
n
(t)dt (2.62)
Taking Fourier transform on both sides,
F[g(t)](τ) =
∞
n=0
a
n
F[
n
(t)](τ) (2.63)
That gives,
F[g(t)](τ) =
∞
n=0
a
n
γ
n
n
(τ) (2.64)
12
For a fractional order p,
F
p
[g(t)](τ) =
∞
n=0
a
n
γ
p
n
n
(τ). (2.65)
Substituting for a
n
from equation 2.62 to the above equation,
F
p
[g(t)](τ) =
∞
n=0
__
+∞
−∞
g(t)
n
(t)dt
_
γ
p
n
n
(τ) (2.66)
Interchanging summation and integration,
F
p
[g(t)](τ) =
_
+∞
−∞
g(t)
_
∞
n=0
γ
p
n
n
(t)
n
(τ)
_
dt (2.67)
By simplifying the term in the bracket, we get
F
α
[g(t)](τ) =
_
+∞
−∞
g(t)
_
(1 −j cot α)
2π
e
jπ(t
2
+τ
2
)
2
cotα−jπtτ csc α
dt (2.68)
Where α = p
π
2
, α = nΠ, n =0,1,2...
Taking the constant term outside of integration,
F
α
[g(t)](τ) =
_
(1 −j cot α)
2π
_
+∞
−∞
g(t)e
jπ(t
2
+τ
2
)
2
cotα−jπtτ csc α
dt (2.69)
The Equation 2.69 [?] can be used for ﬁnding the Fractional Fourier Transform and
Inverse Fractional Fourier transform for any g(t). We obtain the inverse Fractional
Fourier Transform by reversing the sign of α.
13
Chapter 3
Discrete Fractional Fourier
Transform
We have seen the equation of continuous time FrFT. In order to implement this in a
practical application we have to get a discrete time version of these equations [4][8].
While discretising these equations we are eﬀectively discretising the kernel itself. Once
we obtain the DFrFT equations, we can use them in DSP applications similar to as
that of DFT.
3.1 TimeEﬃcient Algorithm for ﬁnding FrFT
If we rewrite the equation 2.69 using variables x and y for t and τ respectively, the
power term in the exponential can be simpliﬁed as follows,
(x
2
+ y
2
)
2
cotα −xy csc α = −
x
2
2
tan
α
2
+
(x −y)
2
2
csc α −
y
2
2
tan
α
2
(3.1)
The process of ﬁnding out the Fractional Fourier transform now consists of three steps
[3].
Step 1  Multiplication by chirp in time domain
g
1
(x) =
_
(1 −j cot α)
2π
g(x)e
−jπ
x
2
2
tan
α
2
(3.2)
14
Step 2  Convolution with the chirp
g
2
(y) =
_
∞
−∞
g
1
(x)e
jπ
(y−x)
2
2
csc α
dx (3.3)
Step 3  Multiplication with chirp in transformed domain
G(y) = g
3
(y) = g
2
(y)e
−jπ(
y
2
2
tan
α
2
)
(3.4)
We should be able to calculate the samples of transform G(y), from the samples of
signal g(x). We will assume that the timedomain representation of our signal is
approximately conﬁned to the interval [∆
t
2
, ∆
t
2
] and that its frequencydomain rep
resentation is conﬁned to the interval [∆
f
2
, ∆
f
2
]. With this statement, we mean that
a suﬃciently large percentage of the signal energy is conﬁned to these intervals. For a
given class of functions, this can be ensured by choosing ∆t and ∆f suﬃciently large.
We then deﬁne the timebandwidth product N = ∆t∆f, which is always greater than
unity because of the uncertainty relation.
Let us now introduce the scaling parameter s with the dimension of time and intro
duce scaled coordinates x =
t
s
and U = fs. With these new coordinates, the time
and frequency domain representations will be conﬁned to intervals of length ∆(
t
s
) and
∆(fs). Let us choose s =
_
(
∆t
∆f
) so that the lengths of both intervals are now equal
to the dimensionless quantity
_
(
∆f
∆t
), which we will denote by ∆x. In the newly
deﬁned coordinates, our signal can be represented in both domains with N = ∆x
2
samples spaced
1
∆x
.
For 0.5 ≤ p ≤ 1.5,  tan(α/2) < 1. The maximum frequency in the chirp will go up
to ∆x; this can be sampled at frequency 2∆x.
For implementation in discrete domain the following are the steps to be followed [3].
15
Step 1:
g
1
_
m
2∆x
_
=
_
(1 −j cot α)
2π
g
_
m
2∆x
_
e
−jπ
2
tan
α
2
(
m
2∆x
)
2
, −N ≤ n ≤ N (3.5)
Step 2:
g
2
_
n
2∆x
_
=
_
∞
−∞
g
1
_
m
2∆x
_
e
jπ
(n−m)
2
8∆x
2
csc α
dx, −N ≤ n ≤ N (3.6)
Step 3:
G
_
n
2∆x
_
= g
3
_
n
2∆x
_
= g
2
_
n
2∆x
_
e
−jπ
2
tan
α
2
(
n
2∆x
)
2
−N ≤ n ≤ N (3.7)
Prior to the ﬁrst step, input should be upsampled (by factor 2) and after third step
result should be downsampled(by factor 2). For fractional orders p < 0.5, take
inverse Fourier transform and use the above algorithm with order 1 + p.
Steps 1 and 3 have computational complexities of order n. Step 2 is a convolution
operation which can be computed with computation complexity of order nlog(n) by
using FFT.
3.2 Implementing TimeEﬃcient Algorithm in Mat
lab
The algorithm stated above was implemented in Matlab and Fractional Fourier Trans
forms of some common signals were found out for diﬀerent fractional orders. The
Matlab code for fractional Fourier Transform and the transforms of some common
signals at various fractional orders are presented.
3.2.1 Matlab Code for FrFT
function frft2=frft(f,a)
16
N=length(f);
%interpolation at the begining
f=interpft(f,2*N);
xaxis=(N/2:N/21);
%...............variables for chirp.....................
alpha = a*pi/2;
talpha2 = tan(alpha/2);
sinalpha=sin(alpha);
%..................defining chirp.......................
chirp = exp(i/N*pi/4*talpha2*(N:N1)’.^2);
%...............multiplying with chirp..................
for p=1:2*N1
f(p)= f(p)*chirp(p);
end
%................convoluting with chirp.................
c = pi/N/sinalpha/4;
chirp2=exp(i*c*((2*N):2*N1)’.^2);
Frft = conv(chirp2,f);
%............post multiplying with chirp................
Frft = chirp.*Frft(2*N:4*N1);
%............multipling by the gain term...............
Frft = Frft(1:2*N)*exp(i*pi*sign(sinalpha)/4
+i*alpha/2)/sqrt(abs(sinalpha))/2/sqrt(N);
%..........down sampling Frft to N terms..............
17
% length(frft)
frft2=Frft(1:2:2*N);
3.2.2 Fractional Fourier Transform of Delta function
Fractional Fourier Transform of Delta function for angles 22.5
0
, 45
0
, 67.5
0
, 90
0
are as
shown in ﬁgure 3.1 and 3.2
.
Figure 3.1: (a)Delta Function Input (b)FrFT at 22.5
0
(c)FrFT at 45
0
18
Figure 3.2: (a) FrFT of Delta Function at 67.5
0
(b)FrFT of Delta Function at 90
0
19
3.2.3 Fractional Fourier Transform of Sine function
Fractional Fourier Transform of Sinusoidal Signal for angles 22.5
0
, 45
0
, 67.5
0
, 90
0
are
as shown in ﬁgure 3.3 and 3.4.
Figure 3.3: (a) Sinusoidal Input, (b) FrFT at 22.5
0
for Sinusoidal Input (c)FrFT at
45
0
for Sinusoidal Input
20
Figure 3.4: (a) FrFT at 67.5
0
for Sinusoidal Input (b) FrFT at 90
0
for Sinusoidal
Input (c) Inverse FrFT
21
3.2.4 Fractional Fourier Transform of rectangular function
Fractional Fourier Transform of rectangular pulse for angles 22.5
0
, 45
0
, 67.5
0
, 90
0
are
as shown in ﬁgure 3.5 and 3.6.
Figure 3.5: (a) Rectangular Function Input (b) FrFT at 22.5
0
for Rectangular Func
tion Input (c) FrFT at 45
0
for Rectangular Function Input (d) FrFT at 67.5
0
for
Rectangular Function Input
22
Figure 3.6: (a) FrFT of Rect at 90
0
, (b) FrFT  Magniﬁed view of (a), (c) Inverse
FrFT
23
Chapter 4
Filtering in Fractional Fourier
Domain
Consider the case of two sine waves of diﬀerent frequencies. They cannot be separated
in time domain using a multiplicative ﬁlter as they are completely overlapped in time
domain. But they are easily separable in frequency domain. Similarly, consider two
impulses that are shifted in time. They are inseparable using a multiplicative ﬁlter
in frequency domain. But can be separated in time domain. It gives us an intuition
that there should be signals which are separable at some particular fractional domain
but are overlapped in time and frequency domains. Such signals can be ﬁltered using
FrFT.
4.1 Filtering in Fractional Fourier Domain
Our scheme of FrFT ﬁltering is as follows.
1. Analyze the signal and obtain the fractional domain at which signals get sepa
rated.
2. Take the FrFT of the signal.
3. Remove the interfering noise part using a multiplicative window.
4. Take inverse FrFT
24
The above scheme is applied in the case of Gaussian window contaminated with chirp
noise. Signals after each processing are shown. Notice that there are disturbances at
either side of the processing window because of the ﬁnite time duration of the window
we considered. See Fig. 4.1, Fig. 4.2 and Fig. 4.3
Figure 4.1: (a) Gaussian Signal (b) Chirp noise
25
Figure 4.2: (a) Gaussian Signal with chirp noise (b) FrFT of a at 72
0
(c) Part of FrFT
corresponding to chirp
26
Figure 4.3: (a) After windowing out FrFT corresponding to chirp (b) Chirp extracted
by inverse FrFT (c) Gaussian extracted by inverse FrFT of (a)
27
4.2 Optimal Filter Design
Consider a case where signals that we consider are overlapped in all the fractional
domains. In this case we have to choose a window which gives maximum separation
between the signals we consider [9]. Our modiﬁed scheme of FrFT ﬁltering will be as
follows.
1. Take the FrFT of the ﬁlter in all Fractional domains.
2. Obtain the optimum multiplicative ﬁlter in all the fractional domains.
3. Eﬀect ﬁltering in all the fractional domains with corresponding window.
4. Calculate SNR obtained in each order.
5. Select the fractional order that gives highest SNR.
6. Take the inverse FrFT corresponding to that order which maximizes SNR.
It is impractical to consider all the inﬁnitesimally separated fractional orders. So we
discretize the fractional orders and searching is carried out over that set only. A trade
oﬀ is possible between processing time and optimality of fractional order. Another
method is to go for a coarse and ﬁne searching.
Let x(t) be the transmitted signal and y(t) be the received signal after the eﬀect of
noise n(t), such that,
y(t) = x(t) + n(t) (4.1)
We use a multiplicative ﬁlter g(t), such that we get an estimate of x(t), from y(t).
´ x(t) = g(t).y(t) (4.2)
Our aim is to ﬁnd out an optimal ﬁlter g(t) which minimizes J, where J is,
J = E
__
∞
−∞
¸
¸
(x(t) −g(t).y(t))
2
¸
¸
dt
_
(4.3)
28
At any instant, Let x be the transmitted value and y be the received value, assuming
channel delay to be zero.
y = x + n (4.4)
Even if x is a constant value, y will be a set of values because of n. Let x be the
ensemble of x.
i.e., x = [x
1
, x
2
, ..., x
n
]
T
such that, x
1
= x
2
= ..... = x
n
= x = x(t)
t=t
0
The ensemble of received values,
y = [y
1
, y
2
, ..., y
n
]
T
(4.5)
Now, ´ x, the estimate of x will be,
´ x = g.y, such that g = g(t)
t=t
0
.
Our aim is to minimize the norm of j = xg.y, which happens when j ⊥ y. i.e.,
(x −g.y)
T
.y = 0
x
T
.y −g.y
T
.y = 0
g = (x
T
.y)/(y
T
.y) (4.6)
g(t)
t=t
0
=
R
xy
R
yy
t=t
0
(4.7)
We can generalize the result to fractional domains also as,
g
a
(t
a
)
t
a
=t
a
0
=
R
x
a
y
a
R
y
a
y
a
t
a
=t
a
0
(4.8)
For each a value we ﬁnd out a J value and we select the value of a which minimizes J.
29
4.2.1 Example 1 : Chirp signal contaminated with White
Gaussian Noise.
Consider the case of a Gaussian pulse contaminated with chirp noise. Gaussian will
remain Gaussian in all the fractional domains, while chirp forms an impulse at a
particular fractional order. Now we can separate the impulse corresponding to the
chirp with a rectangular window. Taking inverse FrFT, we get both signals separated.
See Fig.4.4 and Fig.4.5.
Here input is chirp x(t) and output is y(t).
y(t) = x(t) + n(t) (4.9)
In general,
y
a
(t) = x
a
(t) + n
a
(t) (4.10)
We have seen that the optimum ﬁlter in fractional domain is of the form
g
opt
(t
a
) =
R
x
y(t
1
, t
2
)
R
y
y(t
1
, t
2
)

t
1
=t
2
=t
a
(4.11)
g
opt
(t
a
) =
E[x
a
(t
a
)y
∗
a
(t
a
)]
E[y
a
(t
a
)y
∗
a
(t
a
)]
(4.12)
White noise will remain white in all the fractional domains.
E[n
a
(t
a
)] = 0 (4.13)
Expanding y
a
(t
a
) , we get
g
opt
(t
a
) =
[x
a
(t
a
)x
∗
a
(t
a
)]
[x
a
(t
a
)x
∗
a
(t
a
)] + E[n
a
(t
a
)n
∗
a
(t
a
)]
(4.14)
Where E[n
a
(t
a
)n
∗
a
(t
a
)] is the average noise power.
Multiplicative ﬁltering is carried out for diﬀerent values of a and the one gives maxi
mum SNR is selected for ﬁltering.
30
Figure 4.4: (a) Chirp Signal (b)WGN added to Chirp Signal SNR = 6dB (c) Frac
tional Fourier Transform of the signal at −82.5
0
s
31
Figure 4.5: (a)Optimum Multiplicative Filter (b)Extracted Chirp Signal SNR = 13dB
(c) Variation of output noise power with fractional order
32
4.2.1.1 Observations
• Rectangular window is taken for simplicity. It need not be the window that
maximizes the separation.
• Error is mainly concentrated at the ends of the frame taken. This is because
of ﬁnite time width of the DFrFT window and also because of the rectangular
window ﬁltering.
4.2.1.2 Matlab Code
%chirp in wgn.
%white noise remain white in all fractional domains.
clear variables
ch_len = 4096;
ch_k = .0001;
T = 1*ch_len/2+1:ch_len/2
%generating chirp
chirp = zeros(ch_len,1);
for m=1:ch_len
chirp(m) = 1*exp( j*((mch_len/2)^2 )*ch_k);
end
plot(T,real(chirp));
plt = 0;
%generating wgn
pow =4.3;
nos = wgn(ch_len,1,10*log10(pow));
rcv = nos+chirp;
plot(T,real(rcv));
33
plt =0;
%calculating SNR
snr = 10*log10(sum(chirp .*conj(chirp))/sum(nos.*conj(nos)))
m=1;
for ang = 90:1:70
rcv_tr = frft(rcv,ang/90); % taking FrFT
plot(T,abs(rcv_tr));
plt = 0;
chirp_tr = frft(chirp,ang/90);
plot(T,abs(chirp_tr));
plt = 0;
rxx = abs(chirp_tr).^2;
plot(T,rxx);
plt=0;
g = rxx ./ (rxx+pow); %multiplicative filter
plot(T,g);
plt = 0;
rcvflt_tr = g .* rcv_tr; %applying filter
plot(T,abs(rcvflt_tr));
plt =0;
rcvflt = frft(rcvflt_tr, 1*ang/90);% taking inverse FrFT
plot(T,real(rcvflt));
plt = 0;
34
err = chirp  rcvflt;
plot(T,abs(err));
plt =0;
snr_flt = 10*log10(sum(chirp .*conj(chirp))/sum(err.*conj(err)))
err_str(m) = sum(err .* conj(err))/ch_len;
m=m+1;
[ang err_str(m1)]
end
plot(90:1:70,err_str);y(t)=x(t)+0.8x(t200T_s )+0.6x(t450T_s)
4.2.2 Example 2 : Square pulse in Linear FM noise.
Linear FM noise is basically chirp signals. Linear FM signals can be used as wideband
interference signals. Chirp signal will transform to impulse at a particular fractional
Fourier domain and can be easily separated from message signal [6].
See Fig.4.6 and Fig.4.7
It involves two steps,
• Identifying the proper fractional domain.
• Removing the “peak” portions corresponding to chirp using an optimal multi
plicative ﬁlter.
Proper fractional domain can be identiﬁed by carrying out a search on diﬀerent frac
tional domains for maximum peaking. A suitable function that gives the amount of
peaking in each fractional domain is
J =
_
X
a
(t
a
)
h
dt
a
, h > 2 (4.15)
35
Now using an optimal multiplicative ﬁlter at a fractional order that maximizes J,
we can ﬁlter out the chirp signal from message. Signal is to be taken as frames of
typically about 50% overlap in order to reduce the eﬀect of windowing.
Figure 4.6: (a) Square wave + Chirp noise, SNR = 20dB (b) FrFT at −76.5
0
4.2.2.1 Observation
Noise after processing is mainly concentrated at the ends of the segment. This is
due to the ﬁnite length of the segment considered. As the pulse used is having lower
frequency components, whenever the part of the chirp with low frequency comes,
there is a higher chance of bit error.
4.2.2.2 Matlab Code
%Square pulse in chirp noise
clear variables
sym_len = 10;
%generating random data sequence
data =pn(7,1);
no_sym = length(data);
36
Figure 4.7: (a) Optimal Filter (b) After denoising, SNR=6 dB (c) Error After De
modulation
37
%modulating with square pulse
strt = 1;
for k = 1:no_sym
for m = 1:sym_len
trn(strt+m) = 2*data(k)1;
end
strt = strt +sym_len;
end
stem(data);
plot(trn);
tot_len = no_sym * sym_len;
chirp = zeros(1,tot_len);
chirp_rate = .0006;
%generating chirp
for k = 1:tot_len+1
chirp(k) = 10*exp(j * (k1)*(k1)*chirp_rate);
end
plot(real(chirp));
ch_pow = sum(abs(chirp) .* abs(chirp));
rcv=trn+chirp;
sgn_pow = sum(abs(trn) .* abs(trn));
38
%snr calculation before processing
snr = 10 * log10(sgn_pow/ch_pow)
plot(real(rcv));
%finding out the optimum angle
for ang = 90:1:90
chirp_fr = frft(chirp, ang/90);
trn_fr = frft(trn, ang/90);
rcv_fr = frft(rcv, ang/90);
j = (sum((abs(rcv_fr)).^3)/tot_len);
plot(abs(chirp_fr));
plot(abs(trn_fr));
plot(abs(rcv_fr));
[ang, j]
end
trn_fr = frft(trn,76.5/90);
plot(abs(trn_fr));
rcv_fr = frft(rcv,76.5/90);
plot(abs(rcv_fr));
chirp_frd = frft(chirp,76.5/90);
plot(abs(chirp_frd));
%calculating the filter window
39
g = (abs( chirp_frd)).^2;
g = ones(length(g),1)./(1+g);
plot(g);
rcv_fr = g.*rcv_fr;
plot(abs(rcv_fr));
rcv_dns = frft(rcv_fr,76.5/90);
plot(real(rcv_dns));
%calculating error power
err_pow = sum(abs(rcv_dnstrn’).*abs(rcv_dnstrn’));
SNR1 = 10 * log10(sgn_pow/err_pow)
%demodulating
strt = 1;
for k = 1:no_sym
acc =0;
for m = 1:sym_len
acc = acc+rcv_dns(strt+m);
end
if(acc>0)
data_rcv(k) =1;
else
data_rcv(k) = 0;
end
strt = strt +sym_len;
40
end
%finding bit error
err = abs(data  data_rcv);
stem(err);
4.3 Other Applications
4.3.1 FrFT for Compression
Transforms are widely used for signal compression applications. Consider the case
of a speech data sample. After applying Fourier transform, most of the energy will
be concentrated on very few frequency components. The fraction of energy in other
frequency components will be very small. After a thresholding all such low energy
components will be removed. That is, most of the coeﬃcients will be zero and the
resulting data can be easily compressed.
If we use FrFT instead of Fourier transform, we can search for a fractional order,
which gives highest energy compaction for a given signal. Thus compression can be
optimized. For signals with highly non stationary spectral characteristics, signal has
to be divided into ﬁnite length segments and fractional order optimization has to be
applied on each segment.
We observed that for speech and music, Fourier domain gives the best performance
for compression compared to other fractional domains.
4.3.2 Multipath Channel Estimation Using FrFT
Channel characteristics of a static multipath channel can be measured using FrFT. A
chirp signal is transmitted from the transmitter side and received signal is analyzed
41
using FrFT. Chirp signals gives sharp peaking in fractional domains decided by chirp
parameters. Consider the time shift property of FrFT.
F
α
[x(t −τ)] (u) = F
α
[x(t)] (u −τ cos α)e
j
τ
2
2
sin αcos α−juτ sin α
(4.16)
If we consider the magnitude,
F
α
[x(t −τ)] (u) = F
α
[x(t)] (u −τ cos α) (4.17)
Received signal will have many chirp pulses, each with diﬀerent delay and diﬀerent
gain. Each impulse corresponding to diﬀerent copies of transmitted chirp signal will
be shifted in fractional domain by τcos(α) where τ is the time shift between the
chirps and α is the fractional angle at which the transmitted chirp gives peaking.
The multipath model used is
y(t) = x(t) + 0.8x(t −200T
s
) + 0.6x(t −450T
s
) (4.18)
4.3.2.1 Matlab Code
%chirp filtering in a multipath channel+wgn
clear variables;
chirp_len = 600;
chirp = zeros(1,chirp_len);
chirp_rate = .0005;
%generating chirp
for k = 1:chirp_len
chirp(k) = exp(j * (k1)*(k1)*chirp_rate);
end
42
plot(real(chirp));
plt = 1;
tot_len = 1200;
rcv = zeros(1,tot_len);
%generating a multipath model output
Figure 4.8: (a) Transmitted chirp signal (b) Received signalMulti Path, WGN SNR
= 6 dB (c) Signal after FrFT for coeﬃcients 1, 0.8, 0.6 Angle = −80
0
43
for k = 1:chirp_len
rcv(k) = rcv(k) + 1 * chirp(k);
rcv(200+k) = rcv(200+k) + .8 * chirp(k);
rcv(450+k) = rcv(450+k) + .6 * chirp(k);
end
plot(real(rcv));
plt = 1;
rcv_pow = sum(abs(rcv) .* abs(rcv));
%generating wgn
n = wgn(1,tot_len,2.6);
plot(n);
plt =1;
nos_pow = sum(abs(n) .* abs(n));
SNR = 10 * log(rcv_pow / nos_pow)
rcv = rcv + n;
plot(real(rcv));
plt = 1;
%search for optimum angle
for ord = 90:1:90
rcv_fr = frft(rcv,ord/90);
plot(abs(rcv_fr));
plt =1;
ord
44
end
rcv_fr = frft(rcv, 80/90);
plot(abs(rcv_fr));
plt=0;
peak = 0;
magn_rcv_fr = abs(rcv_fr);
peak = max(magn_rcv_fr);
threshold = .4* peak;
for k =1:tot_len
if magn_rcv_fr(k) < threshold
magn_rcv_fr(k) = 0;
rcv_fr(k) = 0;
end
end
plot(abs(rcv_fr));
plt=0;
val = 0;
rcv_fr_est = zeros(1,tot_len);
rcv_fr_est(485:495) = rcv_fr(485:495);
plot(abs(rcv_fr_est));
plt = 0;
45
rcv_est = frft(rcv_fr_est, 80/90)’;
plot (real(rcv_est));
plt = 0;
4.3.3 FrFT for measuring the acceleration of a moving object
in radial direction
When illuminated by a constant frequency sinusoid, the reﬂection from a radially
accelerating object will be a chirp. A search in diﬀerent fractional domains is carried
out for obtaining the peak. Let α
max
be the angle corresponding to maximum peaking
[5].
Transmitted signal
S
t
(t) = e
j2πf
0
t
(4.19)
Received signal
S
r
(t) = e
j2πf
0
t+j2π
2v
λ
0
t+π
2a
λ
0
t
2
+Φ
0
(4.20)
The estimation of radial acceleration is
a
est
= −
λ
0
f
s
2T
cot α
max
(4.21)
46
Chapter 5
Critical Evaluation and Conclusion
5.1 Critical Evaluation
Based on the study and analysis of the simulation results, we conclude that:
1. FrFT gives one more degree of freedom while designing signal processing tools
compared to Fourier transform.
2. Signal analysis in timefrequency plane is easy with the help of FrFT.
3. FrFT is computationally eﬃcient and has the same order of complexity as that
of Fourier transform.
4. FrFT based computations are to be done framewise, which will result in errors
due to windowing and ﬁnite time duration. It can be reduced to a great extent
by taking frames with 50 percent overlap.
5. Most of the time, optimum order for computation will be unknown, as a result,
a search over a range of fractional orders should be carried out, which is a time
consuming task.
6. FrFT performs excellently if any of the signals that we consider is chirplike.
7. Signals of nonstationary spectral characteristics can be analyzed using FrFT
with superior performance compared to Fourier transform.
47
8. For speech and music processing, advantage that we get by using FrFT is min
imal.
9. Signals after FrFT processing will contain complex part, which should be con
sidered for any further processing.
5.2 Conclusion
We can conclude that FrFT is a more general method for signal processing and sys
tem design. FrFT based systems can replace the current frequency domain systems.
FrFT can have bigger roles in ﬁelds like radar and sonar where chirp signals are very
common. There is large scope of research in FrFT in the ﬁelds of spread spectrum
communication, signal watermarking and encryption, cognitive radio and so on. FrFT
can also be used for the design of faster optical signal processing systems.
48
Bibliography
[1] V. Namias, “The Fractional Order Fourier Transform And Its Application To
Quantum Mechanics,” J. Inst. Math. Applicat., Vol. 25, Pp. 241265, 1980.
[2] A. C. Mcbride and F. H. Kerr, “OnNamias’ Fractional Fourier Transforms,” IMA
J. Appl. Math., Vol. 39, Pp. 159175, 1987.
[3] Haldun M. Ozaktas, Orhan Ankan, M. Alper Kutay and Gozde Bozdak, “Digi
tal Computation Of The Fractional Fourier Transform,” IEEE Transactions on
Signal Processing, Vol.44, No.9, September 1996.
[4] SooChang Pei, MinHung Yeh and ChienCheng Tseng, “Discrete Fractional
Fourier Transform Based On Orthogonal Projections,” IEEE Transactions on
Signal Processing, Vol.47, No.5, May 1999.
[5] Wenchao Du, Xueqiang Gao and Guohong Wang, “Using Frft To Estimate
Target Radial Acceleration”, Proceedings of the 2007 International Conference
on Wavelet Analysis and Pattern Recognition, Beijing, China, 24 Nov. 2007.
[6] Qi Lin, Tao Ran and Zhou SiYong, “Rejection of Linear FM Interference in DSSS
System Based on Fractional Fourier Transform,” Journal of Beijing Institute of
Technology, Vol.14, No.2, 2005.
[7] Luis B. Almeida, “The Fractional Fourier Transform and TimeFrequency Repre
sentations,” IEEE Transactions on Signal Processing, Vol.42, No.11, November
1994.
49
[8] Haldun M. Ozaktas, Orhan Ankan, M. Alper Kutay and Gozde Bozdaki, “Digital
Computation of the Fractional Fourier Transform,” IEEE Transactions on Signal
Processing, Vol.44, No.9,September 1996.
[9] M. Alper Kutay, Haldun M. Ozaktas, Orhan Arikan and Levent Onural, “Op
timal Filtering in Fractional Fourier Domains,” IEEE Transactions on Signal
Processing, Vol.45, No.5, May 1997.
i NATIONAL INSTITUTE OF TECHNOLOGY CALICUT
DEPARTMENT OF ELECTRONICS AND COMMUNICATION ENGINEERING
CERTIFICATE
This is to certify that this report titled FRACTIONAL FOURIER TRANSFORM AND ITS APPLICATIONS is a bona ﬁde record of the majorproject done by Alex John Koshy (Roll No. B050326EC), Nidhin Chandran A K (Roll No. B050160EC), Subin B (Roll No. B050173EC) and Vinay N K (Roll No. B050032EC), in partial fulﬁllment of the requirements for the award of Degree of Bachelor of Technology in Electronics and Communication Engineering from National Institute of Technology, Calicut.
Dr. G. Abhilash (Project Advisor) Assistant Professor
Dr. Lillykutty Jacob Professor and Head
29 April 2009 NIT Calicut
ii
ACKNOWLEDGEMENT
We would like to thank Dr. G.Abhilash, Assistant Professor, Department of Electronics and Communication Engineering for his guidance and inspiration in helping us complete this project. We are also grateful to Dr. Lillykutty Jacob, Professor and Head, Department of Electronics and Communication Engineering for providing us with this opportunity to work on our project and also for permitting access to the required facilities. We would also like to thank the lab staﬀ for their technical support and providing us assistance. We also thank our batch mates who had supported us and provided us with greatly appreciated technical and nontechnical aid throughout our project.
Alex John Koshy Nidhin Chandran A K Subin B Vinay N K
iii
Abstract
Signals can be viewed from diﬀerent perspectives using diﬀerent transforms. The Fourier transform which allows us to observe the signal in terms of diﬀerent frequency components is widely used in signal processing and communication. Fractional Fourier Transform (FrFT) is a generalized Fourier transform which allows us to take transforms of fractional order also. Theory of FrFT is developed from Fourier transform. For computation we move from continuous to discrete domain and theory of Discrete Fractional Fourier Transform (DFrFT) is discussed. A time eﬃcient algorithm for DFrFT is also discussed. Theory of designing optimal ﬁlter using FrFT is developed from ﬁrst principles. Optimal ﬁlters are designed for some special cases and their performance is analyzed. Some applications other than ﬁltering is also discussed.
iv
Contents
Abstract 1 Introduction 1.1 1.2 1.3 Transforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . TimeFrequency Plane . . . . . . . . . . . . . . . . . . . . . . . . . . iii 1 1 1 2 4 4 7 8 9 11 13 13 15 15 17 19 21 23 23
2 Fractional Fourier Transform 2.1 Fractional Fourier Transform from Fourier Transform . . . . . . . . . 2.1.1 2.1.2 2.1.3 2.2 Orthogonality of Eigenfunctions . . . . . . . . . . . . . . . . . Orthonormality of Eigenfunctions . . . . . . . . . . . . . . . . Completeness of Eigenfunctions . . . . . . . . . . . . . . . . .
Equation for Fractional Fourier Transform . . . . . . . . . . . . . . .
3 Discrete Fractional Fourier Transform 3.1 3.2 TimeEﬃcient Algorithm for ﬁnding FrFT . . . . . . . . . . . . . . . Implementing TimeEﬃcient Algorithm in Matlab . . . . . . . . . . . 3.2.1 3.2.2 3.2.3 3.2.4 Matlab Code for FrFT . . . . . . . . . . . . . . . . . . . . . . Fractional Fourier Transform of Delta function . . . . . . . . . Fractional Fourier Transform of Sine function . . . . . . . . . Fractional Fourier Transform of rectangular function . . . . .
4 Filtering in Fractional Fourier Domain 4.1 Filtering in Fractional Fourier Domain . . . . . . . . . . . . . . . . .
.1 4. . . . .v 4. . . . . . . .2 4. . . .2 Observation .2.3. . . . . . . . . . . . 45 46 46 47 48 5 Critical Evaluation and Conclusion 5. . . . . . . . . . . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . .2 Critical Evaluation . . . . . .3. . . . . . . . . . . . . . 29 32 32 34 35 35 40 40 40 41 27 Example 2 : Square pulse in Linear FM noise. . . . . . . FrFT for measuring the acceleration of a moving object in radial direction . . . .2. .1 4. .3 Matlab Code . . . .3 Other Applications . . . . . . . . . . . Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Matlab Code . . .1 Example 1 : Chirp signal contaminated with White Gaussian Noise. . 4. . . . . . . . . . . . .2. . . . . . . . . . . . .2 Observations . . . . . . .2. .2. . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . 4. . . Matlab Code . .1. . . . .3. . . . . . . . . . . . . .1 4. . . . .2. . . . . . . . . .2. . . . .1 5. . . . . . . Conclusion . . . . . . . . . . Multipath Channel Estimation Using FrFT . . . .1.2. . . . . . . .2. . . . . 4. . . . . 4. . . .3. . . . . .2 FrFT for Compression . . . . . . . . . . .1 4. .2 Optimal Filter Design . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .5 (a)Optimum Multiplicative Filter (b)Extracted Chirp Signal SNR = 13dB (c) Variation of output noise power with fractional order . . . . .1 3. . . . . . . . .50 for Rectangular Function Input (c) FrFT at 450 for Rectangular Function Input (d) FrFT at 67. . . . . .5 (a) Rectangular Function Input (b) FrFT at 22. .50 s . . . . . . . . . . . . . . . SNR=6 dB (c) Error After Demodulation . . . . . . . .2 3. . . . . . . . . . . . . . . . .7 (a) Square wave + Chirp noise. . . . . . . . . . . . .6 4. . . . . . . . .2 . . . (a) FrFT of Delta Function at 67. . . . . . 3. (b) FrFT . (c) Inverse FrFT 4. . . . .50 for Sinusoidal Input (b) FrFT at 900 for Sinusoidal Input (c) Inverse FrFT . . . . (b) FrFT at 22. . . .1 4.3 (a)Delta Function Input (b)FrFT at 22. . 36 . . . . . . . .4 (a) Chirp Signal (b)WGN added to Chirp Signal SNR = 6dB (c) Fractional Fourier Transform of the signal at −82. . . . . . . 22 24 21 20 19 17 18 (a) Gaussian Signal (b) Chirp noise . . .50 for Sinusoidal Input (c)FrFT at 450 for Sinusoidal Input .4 (a) FrFT at 67. .50 (c)FrFT at 450 .6 (a) FrFT of Rect at 900 . . . . . . . . . . . . . . . . . 31 35 4.vi List of Figures 3. . . (a) Optimal Filter (b) After denoising. . . .50 . . 3. . . . . . . . . . . . . . . .3 (a) After windowing out FrFT corresponding to chirp (b) Chirp extracted by inverse FrFT (c) Gaussian extracted by inverse FrFT of (a) 26 4.50 (b)FrFT of Delta Function at 900 (a) Sinusoidal Input. . . 25 4.50 for Rectangular Function Input . . . . . . . . . . . SNR = 20dB (b) FrFT at −76. . . . . . . . .Magniﬁed view of (a). 30 4. . . . . . . . . . . . . . . . . . . . . . 3. (a) Gaussian Signal with chirp noise (b) FrFT of a at 720 (c) Part of FrFT corresponding to chirp .
0. WGN SNR = 6 dB (c) Signal after FrFT for coeﬃcients 1.vii 4.6 Angle = −800 42 .8.8 (a) Transmitted chirp signal (b) Received signalMulti Path. 0.
output will follow the form k(ω)ejωt . where k(ω) is a complex value which is independent of time and depends only on ω. We consider ejωt as the component corresponding to frequency ω in the signal. which are unrecognizable from the time domain representation of the signal. The choice of the transform depends on the type of the signal and the application. 1. Among these transforms Fourier transform is most fundamental one. In order to understand how a LTI system works on a particular signal we have to get an idea about the diﬀerent frequency components present in the signal. It can be found . Fourier Transform. Laplace Transform.1 Transforms Transforms play vital roles in the analysis of signals and systems.1 Chapter 1 Introduction 1.2 Fourier Transform The relation between output y(t) and input x(t) for a Linear Time Invariant system is of the form dk ak k y(t) = dt k=0 m n bk k=0 dk x(t) dtk (1. Transforms help in ﬁnding out the hidden properties of a signal. Wavelet Transform etc are some of the widely used transform in the ﬁeld of communication.1) We can see that if the input is of the form ejωt . signal processing and system design.
4) Consider the following property of Fourier transform.3 TimeFrequency Plane For convenience. ejωt = √ 2π +∞ x(t)e−jωt dt = X(ω) −∞ (1. • three times . • four times . (1. 1 x(t) = √ 2π +∞ X(ω)ejωt dω −∞ (1.gives representation of signal in the time domain with time axis reversed.gives the representation of signal in the frequency domain with frequency axis reversed. • twice .gives the representation of signal in the frequency domain. F[x(t)] = X(t) FF[x(t)] = F 2 [x(t)] = x(−t) FFF[x(t)] = F 3 [x(t)] = X(−t) FFFF[x(t)] = F 4 [x(t)] = x(t) Taking Fourier transform • once .retains the signal in the time domain itself. 1 F [x(t)] = √ 2π +∞ x(τ )e−jτ t dτ −∞ (1.8) . we deﬁne Fourier transform as.7) (1.3) 1.2 as 1 x(t).6) (1. We can obtain the signal from its Fourier transform as follows.5) (1.2) X(ω) is the Fourier transform of the signal x(t).
Consider a plane with time as horizontal axis. is it possible to get the Fourier transform for intermediate angles? We will see the answer in the next chapter. Now the curious question is. F should give the representation along 900 rotated axis. we get the representation of the signal along an axis which is at an angle of 900 with the original axis. Frequency axis is the vertical axis here. . Each time we take the Fourier transform.3 This observation can be easily visualized using the idea of timefrequency plane. i.. As F 2 gives the representation along inverted time axis which is at an angle of 1800 with the original time axis.e.
If we can generalize this rotation operation for intermediate angles also.5 ejωt . rotation obtained for pth order is p( π ). then rotation is by π 2 • p = 2. We are not discussing . then rotation is by π i. there are only 1.5 systems. Let F p represent taking Fourier transform p times.. Fourier transform allows p to take only 2 integer values. • p = 1. We call p as the order of the Fourier transform operation. Fourier transform which is generalized so that p can take any fractional value is called Fractional Fourier Transform (FrFT). If. If the input is of the form ejωt .4 Chapter 2 Fractional Fourier Transform Fourier transform allows us to rotate the axis of representation of signal in TF plane by integral multiples of 900 only. where N is the number of systems in cascade. Suppose we are making a strange assumption that.e. output y(t) will be [H(ω)]N ejωt . Then we may have to write y(t) = [H(ω)]1. then we will be able to use it for analyzing a wider class of signals.1 Fractional Fourier Transform from Fourier Transform Consider a cascaded structure of LTI systems of identical response H(ω). 2.
1.5 whether such a system is possible or not. but this assumption gives us a clue for proceeding to Fractional Fourier transform from Fourier transform.7) Generalizing this for order p. Let ϕi (t) be a set of functions such that. F[q(t)] = F i ai ϕi (t) = i ai F [ϕi (t)] = i ai γi ϕi (t) (2.6) Now if we take Fourier Transform. as.3) (2. we can ﬁnd the FrFT using this method.2) (2.8) . F p [ϕ(t)] = γ p ϕ(t) (2. F 2 [ϕ(t)] = γ 2 ϕ(t) F n [ϕ(t)] = γ n ϕ(t) (2.4) For any function that satisﬁes equation 2. F p [ϕi (t)] = ρi (t) = γip ϕi (t) And q(t) is a function of the form q(t) = i (2.[1][7] Let ϕ(t) be a function that satisﬁes the relation F[ϕ(t)] = ρ(t) = γϕ(t) Then.5) ai ϕi (t) (2. we get F p [q(t)] = F p i ai ϕi (t) = i ai F p [ϕi (t)] = i ai γi p ϕi (t) (2.1) We can generalize this relation for fractional orders also. for fractional order p.
11) d [x(t)] = jtF[x(t)] = jΨF[x(t)] dt (2. i. We get a new eigenfunction ϕnew (t) and a new eigenvalue γnew given by.12) d [F[x(t)]] = jΦF[x(t)] dt (2. ϕnew (t) = (Φ − Ψ)[ϕ(t)] γnew = −jγ (2. Φ= Now.13) (2. It is well known that the Gaussian function. FΨ[x(t)] = F[tx(t)] = j On subtracting (2.e. ϕi (t) is invariant under Fourier transform. is a function which is We deﬁne two operators Φ and Ψ such that. e invariant under Fourier transform. It is clear that each ϕi (t) is an eigenfunction of Fourier transform operator.11) from (2. F[ϕ(t)] = ρ(t) = γϕ(t) Then.6 Now our aim is to obtain such a set of functions ϕi (t).14) i.10). F(Φ − Ψ)[ϕ(t)] = −j(Φ − Ψ)F[ϕ(t)] = −j(Φ − Ψ)γϕ(t) = −jγ(Φ − Ψ)[ϕ(t)] (2.15) (2.e.. FΦ[x(t)] = F Similarly. we get F(Φ − Ψ)[x(t)] = −j(Φ − Ψ)F[x(t)] Suppose that a function ϕ(t) satisﬁes the relation.10) d dt and Ψ = t (2.16) (2. −t2 2 .9) .
25) ϕn . ϕ2 (t) = e 2 In general.18) (2. Proof : We denote ϕn (t) = e 2 The inner product ∞ t2 dn −t2 e .21) ⇒ ϕ1 (t) = e 2 Similarly we get.26) .24) 2. ϕm = −∞ et 2 dn −t2 dm −t2 e e dt dtn dtm (2.19) (2. F[ϕ0 (t)] = ϕ0 (t) Using the above generating equation. we get.1 Orthogonality of Eigenfunctions Theorem 1 : {ϕi (t)} forms an orthogonal system. ϕn (t) = e t2 2 t2 t2 d1 −t2 e dt1 d2 −t2 e dt2 dn −t2 e dtn (2.7 This shows us how to generate a class of eigenfunctions from a single eigenfunction.23) (2.1. ϕ1 (t) = (Φ − Ψ)[ϕ0 (t)] ⇒ ϕ1 (t) = ⇒ ϕ1 (t) = d − t ϕ0 (t) dt t2 d 2 − t e 2 e−t dt −t2 2 = e 2 e−t t2 2 (2. Let us start from Gaussian ϕo (t) = e We know.20) (2.17) (2.22) (2. dtn (2.
. et 2 dn −t2 e = (−2t)n + .27 vanishes..29) 2. dtn dtn (2.32) and dn dn dn −t2 e = n ((−2t)n + .) = (−1)n 2n n!. ϕm dn −t2 dm−1 −t2 = e e e dtn dtm−1 t2 ∞ ∞ − −∞ −∞ d t2 dn −t2 dm−1 −t2 e e e dt dt dtn dtm−1 2 (2.27) and hence all the terms under the diﬀerential sign contain the factor e−t .8 can be evaluated using integration by parts. dtn (2. Since for any k ∈ N..2 Orthonormality of Eigenfunctions To obtain an orthonormal system we evaluate the norm ∞ ϕn 2 = −∞ e −t2 dn −t2 e dtn 2 dt.30) Integration by parts n times yields ∞ ϕn dn −t2 e dtn 2 = (−1) n −∞ e−t 2 dn dn −t2 e dt. we have tk e−t → 0 as t → ∞ 2 (2.. (2. ϕm = 0 f or n = m (2.. dtn dtn dt Consequently. direct diﬀerentiation gives. ∞ (2.33) ϕn 2 = 2 n! −∞ n √ 2 e−t dt = 2n n! π (2.28) the ﬁrst term in equation 2.1.34) . repeated integration by parts leads to ϕn . which gives ϕn ..31) Since is a polynomial of degree n. Therefore.
43) (2. t) = G(z. t) = e 2 G(z.37) (2. t) (2.35) form an orthonormal system in L2 (R).41) .1. t) = d − t G(z. t) for ϕi (t) as ∞ G(z.9 Thus the functions n (t) = √ e 2n n! π 1 −t2 2 dn −t2 e dtn (2. t) dt dz Let H(z. t) dt dz Let H(z.42) −t2 2 G(z.39) i=0 ∞ ϕi+1 (t) i=0 d d − t G(z. t) = e Then. d d H(z. (2.36) (2.3 Completeness of Eigenfunctions Theorem 2 : {ϕi (t)} is complete in L2 (R).40) −t2 −t2 d d e 2 G(z. 2. t) = g(z)h(t) d d g(z)h(t) = g(z)h(t) dt dz (2. t) = dt d G(z. t) dt dz Multiplying the above equation with e−t 2 /2 on both sides and rearranging. t) be of the form H(z.38) (2. t) = H(z. Proof : We deﬁne a generating function G(z. t) = dz i=0 ∞ ϕi (t) ϕi+1 (t) zi i! zi i! zi i! (2.
t) = e 2 ek(z+t) For z = 0.54) G(y. G(z. t) = ek(z+t) t2 n (2.53) (2. t)p(t)dt = 0 −∞ ∞ −∞ ϕ0 (y − t)e 4 p(t)dt = 0 y2 .45) (2. t) = e− 2 (y−t) e 4 = e 4 ϕ0 (y − t) Let there exist a function p(t) in L2 (R) which is orthogonal to ϕi (t). ϕi (t). ⇒ g(z) h(t) Where k is a constant. t) = e 2 H(z.51) (2.46) ⇒ H(z.10 h (t) g (z) = = k. ⇒ g(z) = ekz and h(t) = ekt (2.52) (2. the solution of H(z. ∀ i ∈ Z.50) (2. t) = ϕ0 (t) = e −t2 2 t2 (2.48) ⇒ k = −1 and n = 2 Final generating function is G(z. 2 G(y.47) n G(z. t). t) = e 2 e−(z+t) Putting z = − 1 y gives a new generating function. p(t) = 0 G(y. t) can be in general of the form H(z. p(t) = 0 ∀ y ∈ R ∞ 1 2 y2 y2 t2 2 (2.49) (2.44) (2. t) = ek(z+t) Notice that.
2. ∀ i ∈ Z.11 ∞ ϕ0 (y − t)p(t)dt = 0 −∞ (2.63) That gives.59) i.57) (2.55) (2. ∞ F[g(t)](τ ) = n=0 an F[ n (t)](τ ) (2. i ∈ Z} is dense in L2 (R). That is. It is not possible to ﬁnd a nonzero function in L2 (R) which is orthogonal to ϕi (t). an = g(t). P (ω)ϕ0 (ω) = 0 But.60) Where.e. ϕ(ω) = 0 ⇒ p(ω) = 0 ⇒ p(y) = 0 (2.62) Taking Fourier transform on both sides.2 Equation for Fractional Fourier Transform n (t). Any signal g(t) in L2 (R) can be represented as the linear combination of Mathematically. {ϕi (t)..61) +∞ g(t) n (t)dt (2.64) .56) p(y) ∗ ϕ0 (y) = 0 Taking Fourier transform. F[g(t)](τ ) = ∞ an γn n (τ ) n=0 (2. g(t) = n=0 ∞ an n (t) (2. an can be written as an = −∞ n (t) (2.58) (2.
69) The Equation 2. .12 For a fractional order p. ∞ +∞ p g(t) n (t)dt γn n (τ ) n=0 −∞ F [g(t)](τ ) = p (2. n =0.66) Interchanging summation and integration. We obtain the inverse Fractional Fourier Transform by reversing the sign of α. (1 − j cot α) 2π +∞ jπ(t2 +τ 2 ) cotα−jπtτ 2 Fα [g(t)](τ ) = g(t)e −∞ csc α dt (2.65) Substituting for an from equation 2. ∞ F [g(t)](τ ) = n=0 p p an γn n (τ ).2.62 to the above equation.. we get +∞ Fα [g(t)](τ ) = −∞ g(t) (1 − j cot α) jπ(t2 +τ 2 ) cotα−jπtτ csc α e 2 dt 2π (2. (2. α = nΠ.68) Where α = p π ..69 [?] can be used for ﬁnding the Fractional Fourier Transform and Inverse Fractional Fourier transform for any g(t). +∞ ∞ p p γn n (t) n (τ ) dt n=0 F [g(t)](τ ) = −∞ g(t) (2.67) By simplifying the term in the bracket.1. 2 Taking the constant term outside of integration.
Once we obtain the DFrFT equations. (x2 + y 2 ) x2 α (x − y)2 y2 α cotα − xy csc α = − tan + csc α − tan 2 2 2 2 2 2 (3.Multiplication by chirp in time domain x2 α (1 − j cot α) g(x)e−jπ 2 tan 2 2π g1 (x) = (3. we can use them in DSP applications similar to as that of DFT.1) The process of ﬁnding out the Fractional Fourier transform now consists of three steps [3].13 Chapter 3 Discrete Fractional Fourier Transform We have seen the equation of continuous time FrFT. Step 1 .69 using variables x and y for t and τ respectively. the power term in the exponential can be simpliﬁed as follows. In order to implement this in a practical application we have to get a discrete time version of these equations [4][8]. While discretising these equations we are eﬀectively discretising the kernel itself. 3.1 TimeEﬃcient Algorithm for ﬁnding FrFT If we rewrite the equation 2.2) .
.4) We should be able to calculate the samples of transform G(y). For a given class of functions. With this statement. we mean that 2 2 a suﬃciently large percentage of the signal energy is conﬁned to these intervals. which is always greater than unity because of the uncertainty relation. the time t and frequency domain representations will be conﬁned to intervals of length ∆( s ) and ∆(f s).Multiplication with chirp in transformed domain G(y) = g3 (y) = g2 (y)e−jπ( 2 y2 tan α ) 2 (3. this can be ensured by choosing ∆t and ∆f suﬃciently large. which we will denote by ∆x. ∆ 2 ] and that its frequencydomain rep resentation is conﬁned to the interval [∆ f .5 ≤ p ≤ 1. In the newly ∆t deﬁned coordinates.3) Step 3 . For implementation in discrete domain the following are the steps to be followed [3]. The maximum frequency in the chirp will go up to ∆x.5. We then deﬁne the timebandwidth product N = ∆t∆f . Let us choose s = ∆t ( ∆f ) so that the lengths of both intervals are now equal to the dimensionless quantity ( ∆f ). With these new coordinates. this can be sampled at frequency 2∆x.Convolution with the chirp ∞ g2 (y) = −∞ g1 (x)ejπ (y−x)2 2 csc α dx (3. our signal can be represented in both domains with N = ∆x2 samples spaced 1 . ∆x For 0. We will assume that the timedomain representation of our signal is t t approximately conﬁned to the interval [∆ 2 .14 Step 2 . Let us now introduce the scaling parameter s with the dimension of time and introduce scaled coordinates x = t s and U = f s.  tan(α/2) < 1. from the samples of signal g(x). ∆ f ].
input should be upsampled (by factor 2) and after third step result should be downsampled(by factor 2).5.15 Step 1: m = 2∆x −jπ α m 2 (1 − j cot α) m g e 2 tan 2 ( 2∆x ) .7) Prior to the ﬁrst step. 2∆x −N ≤ n ≤ N (3. 3. 2π 2∆x g1 Step 2: −N ≤ n ≤ N (3.2. Step 2 is a convolution operation which can be computed with computation complexity of order nlog(n) by using FFT. take inverse Fourier transform and use the above algorithm with order 1 + p.6) −jπ α n n n n 2 = g3 = g2 e 2 tan 2 ( 2∆x ) 2∆x 2∆x 2∆x −N ≤n≤N (3. For fractional orders p < 0. 3. The Matlab code for fractional Fourier Transform and the transforms of some common signals at various fractional orders are presented.5) g2 Step 3: G n = 2∆x ∞ g1 −∞ (n−m)2 m ejπ 8∆x2 csc α dx.a) . Steps 1 and 3 have computational complexities of order n.1 Matlab Code for FrFT function frft2=frft(f.2 Implementing TimeEﬃcient Algorithm in Matlab The algorithm stated above was implemented in Matlab and Fractional Fourier Transforms of some common signals were found out for diﬀerent fractional orders.
......2*N)....... Frft = Frft(1:2*N)*exp(i*pi*sign(sinalpha)/4 +i*alpha/2)/sqrt(abs(sinalpha))/2/sqrt(N).....variables for chirp................... %interpolation at the beginingf=interpft(f....convoluting with chirp.................. end %..16 N=length(f).... %....... %...post multiplying with chirp....... Frft = conv(chirp2............ c = pi/N/sinalpha/4....*Frft(2*N:4*N1)............ alpha = a*pi/2. Frft = chirp.. .... sinalpha=sin(alpha)........... chirp2=exp(i*c*((2*N):2*N1)’.... %....multiplying with chirp........ for p=1:2*N1 f(p)= f(p)*chirp(p)... %..... %...............f)...........................^2).. chirp = exp(i/N*pi/4*talpha2*(N:N1)’.. %.....^2)... talpha2 = tan(alpha/2)..................defining chirp....multipling by the gain term....down sampling Frft to N terms........ xaxis=(N/2:N/21).
50 . Figure 3.2 . 3.50 (c)FrFT at 450 .17 % length(frft) frft2=Frft(1:2:2*N). 67.2 Fractional Fourier Transform of Delta function Fractional Fourier Transform of Delta function for angles 22.1: (a)Delta Function Input (b)FrFT at 22.1 and 3. 900 are as shown in ﬁgure 3.50 .2. 450 .
2: (a) FrFT of Delta Function at 67.50 (b)FrFT of Delta Function at 900 .18 Figure 3.
900 are as shown in ﬁgure 3.4.50 .50 for Sinusoidal Input (c)FrFT at 450 for Sinusoidal Input . 67.2.3 Fractional Fourier Transform of Sine function Fractional Fourier Transform of Sinusoidal Signal for angles 22.50 . Figure 3.3: (a) Sinusoidal Input.19 3.3 and 3. (b) FrFT at 22. 450 .
20 Figure 3.4: (a) FrFT at 67.50 for Sinusoidal Input (b) FrFT at 900 for Sinusoidal Input (c) Inverse FrFT .
50 for Rectangular Function Input (c) FrFT at 450 for Rectangular Function Input (d) FrFT at 67. 67. 450 .6.5: (a) Rectangular Function Input (b) FrFT at 22. Figure 3.2.50 .5 and 3.50 .50 for Rectangular Function Input .4 Fractional Fourier Transform of rectangular function Fractional Fourier Transform of rectangular pulse for angles 22. 900 are as shown in ﬁgure 3.21 3.
(c) Inverse FrFT .6: (a) FrFT of Rect at 900 . (b) FrFT .Magniﬁed view of (a).22 Figure 3.
2.1 Filtering in Fractional Fourier Domain Our scheme of FrFT ﬁltering is as follows. 4. Take inverse FrFT .23 Chapter 4 Filtering in Fractional Fourier Domain Consider the case of two sine waves of diﬀerent frequencies. Take the FrFT of the signal. consider two impulses that are shifted in time. Remove the interfering noise part using a multiplicative window. Analyze the signal and obtain the fractional domain at which signals get separated. Similarly. They are inseparable using a multiplicative ﬁlter in frequency domain. But can be separated in time domain. They cannot be separated in time domain using a multiplicative ﬁlter as they are completely overlapped in time domain. Such signals can be ﬁltered using FrFT. 1. It gives us an intuition that there should be signals which are separable at some particular fractional domain but are overlapped in time and frequency domains. 4. 3. But they are easily separable in frequency domain.
Notice that there are disturbances at either side of the processing window because of the ﬁnite time duration of the window we considered. 4.1: (a) Gaussian Signal (b) Chirp noise .2 and Fig.1. 4. Fig. See Fig. Signals after each processing are shown.3 Figure 4.24 The above scheme is applied in the case of Gaussian window contaminated with chirp noise. 4.
2: (a) Gaussian Signal with chirp noise (b) FrFT of a at 720 (c) Part of FrFT corresponding to chirp .25 Figure 4.
3: (a) After windowing out FrFT corresponding to chirp (b) Chirp extracted by inverse FrFT (c) Gaussian extracted by inverse FrFT of (a) .26 Figure 4.
x(t) = g(t). Take the FrFT of the ﬁlter in all Fractional domains. from y(t). ∞ (4. Obtain the optimum multiplicative ﬁlter in all the fractional domains. 3. such that. Calculate SNR obtained in each order. In this case we have to choose a window which gives maximum separation between the signals we consider [9].27 4. So we discretize the fractional orders and searching is carried out over that set only. where J is. Select the fractional order that gives highest SNR.2 Optimal Filter Design Consider a case where signals that we consider are overlapped in all the fractional domains. Another method is to go for a coarse and ﬁne searching. It is impractical to consider all the inﬁnitesimally separated fractional orders.2) J =E −∞ (x(t) − g(t). Eﬀect ﬁltering in all the fractional domains with corresponding window. Take the inverse FrFT corresponding to that order which maximizes SNR. A trade oﬀ is possible between processing time and optimality of fractional order. Our modiﬁed scheme of FrFT ﬁltering will be as follows.y(t) Our aim is to ﬁnd out an optimal ﬁlter g(t) which minimizes J. 4. 5. y(t) = x(t) + n(t) (4.1) We use a multiplicative ﬁlter g(t). 1. 6. such that we get an estimate of x(t).y(t))2 dt (4. 2. Let x(t) be the transmitted signal and y(t) be the received signal after the eﬀect of noise n(t).3) .
...5) We can generalize the result to fractional domains also as. (x − g. = xn = x = x(t) t=t0 The ensemble of received values. yn ]T Now. y2 ...y) g(t) t=t0 = Rxy Ryy t=t0 (4.e. x. x2 .y.y T . which happens when j ⊥ y. such that g = g(t) t=t0 ...e. ga (ta ) ta =ta = 0 Rxa ya Rya ya ta =ta (4.y − g.. the estimate of x will be.7) (4..y.y = 0 g = (xT . x = g. Let x be the ensemble of x. i.y)/(y T . Our aim is to minimize the norm of j = xg. i. y =x+n (4. y = [y1 . x = [x1 ..8) 0 For each a value we ﬁnd out a J value and we select the value of a which minimizes J. . Let x be the transmitted value and y be the received value. assuming channel delay to be zero.6) (4.4) Even if x is a constant value..y)T . .y = 0 xT .. x1 = x2 = . y will be a set of values because of n.28 At any instant.. xn ]T such that.
1 Example 1 : Chirp signal contaminated with White Gaussian Noise.5.29 4. t2 ) t =t =t Ry y(t1 .13) Where E[na (ta )n∗ (ta )] is the average noise power. See Fig. Here input is chirp x(t) and output is y(t).11) gopt (ta ) = (4.4. we get both signals separated.10) (4.14) (4.2.9) (4.4 and Fig. Taking inverse FrFT. y(t) = x(t) + n(t) In general. ya (t) = xa (t) + na (t) We have seen that the optimum ﬁlter in fractional domain is of the form gopt (ta ) = Rx y(t1 . Now we can separate the impulse corresponding to the chirp with a rectangular window. we get gopt (ta ) = [xa (ta )x∗ (ta )] a [xa (ta )x∗ (ta )] + E[na (ta )n∗ (ta )] a a (4. Gaussian will remain Gaussian in all the fractional domains. E[na (ta )] = 0 Expanding ya (ta ) .12) White noise will remain white in all the fractional domains.4. . t2 ) 1 2 a ∗ E[xa (ta )ya (ta )] ∗ E[ya (ta )ya (ta )] (4. Consider the case of a Gaussian pulse contaminated with chirp noise. while chirp forms an impulse at a particular fractional order. a Multiplicative ﬁltering is carried out for diﬀerent values of a and the one gives maximum SNR is selected for ﬁltering.
30 Figure 4.50 s .4: (a) Chirp Signal (b)WGN added to Chirp Signal SNR = 6dB (c) Fractional Fourier Transform of the signal at −82.
5: (a)Optimum Multiplicative Filter (b)Extracted Chirp Signal SNR = 13dB (c) Variation of output noise power with fractional order .31 Figure 4.
It need not be the window that maximizes the separation.2. 4. ch_k = .1 Observations • Rectangular window is taken for simplicity. end plot(T. plt = 0.10*log10(pow)). %generating wgnpow =4.2 Matlab Code %chirp in wgn. T = 1*ch_len/2+1:ch_len/2 %generating chirpchirp = zeros(ch_len. plot(T.32 4.3.1.1. • Error is mainly concentrated at the ends of the frame taken.real(chirp)). rcv = nos+chirp.0001.1). clear variables ch_len = 4096. nos = wgn(ch_len. . This is because of ﬁnite time width of the DFrFT window and also because of the rectangular window ﬁltering.1.real(rcv)).2. for m=1:ch_len chirp(m) = 1*exp( j*((mch_len/2)^2 )*ch_k). %white noise remain white in all fractional domains.
g). rcvflt = frft(rcvflt_tr. 1*ang/90).abs(rcv_tr)). plt = 0.*conj(chirp))/sum(nos. rcvflt_tr = g .real(rcvflt)).abs(chirp_tr)).ang/90). .rxx).* rcv_tr. plt = 0. rxx = abs(chirp_tr). plt=0. plt = 0. plt = 0. %multiplicative filter plot(T. for ang = 90:1:70 rcv_tr = frft(rcv. %applying filter plot(T. plot(T. % taking FrFT plot(T./ (rxx+pow).% taking inverse FrFT plot(T.33 plt =0.*conj(nos))) m=1. plot(T.abs(rcvflt_tr)). g = rxx . %calculating SNRsnr = 10*log10(sum(chirp . plt =0. chirp_tr = frft(chirp.ang/90).^2.
*conj(err))) err_str(m) = sum(err .2.7 It involves two steps. plt =0. Proper fractional domain can be identiﬁed by carrying out a search on diﬀerent fractional domains for maximum peaking. A suitable function that gives the amount of peaking in each fractional domain is J= Xa (ta )h dta .4. Linear FM noise is basically chirp signals.y(t)=x(t)+0. • Identifying the proper fractional domain.15) . snr_flt = 10*log10(sum(chirp . plot(T.abs(err)).34 err = chirp .4.6 and Fig.err_str).6x(t450T_s) 4. • Removing the “peak” portions corresponding to chirp using an optimal multiplicative ﬁlter. See Fig.*conj(chirp))/sum(err. [ang err_str(m1)] end plot(90:1:70. m=m+1.rcvflt. Linear FM signals can be used as wideband interference signals.8x(t200T_s )+0. Chirp signal will transform to impulse at a particular fractional Fourier domain and can be easily separated from message signal [6].* conj(err))/ch_len. h > 2 (4.2 Example 2 : Square pulse in Linear FM noise.
1).35 Now using an optimal multiplicative ﬁlter at a fractional order that maximizes J. 4. .2.50 4. Figure 4.2. Signal is to be taken as frames of typically about 50% overlap in order to reduce the eﬀect of windowing.2. we can ﬁlter out the chirp signal from message.2 Matlab Code %Square pulse in chirp noise clear variables sym_len = 10. whenever the part of the chirp with low frequency comes. there is a higher chance of bit error. As the pulse used is having lower frequency components.1 Observation Noise after processing is mainly concentrated at the ends of the segment. no_sym = length(data).2. %generating random data sequencedata =pn(7.6: (a) Square wave + Chirp noise. This is due to the ﬁnite length of the segment considered. SNR = 20dB (b) FrFT at −76.
7: (a) Optimal Filter (b) After denoising.36 Figure 4. SNR=6 dB (c) Error After Demodulation .
* abs(chirp)). . for k = 1:no_sym for m = 1:sym_len trn(strt+m) = 2*data(k)1.* abs(trn)). end stem(data). sgn_pow = sum(abs(trn) . end plot(real(chirp)). rcv=trn+chirp. tot_len = no_sym * sym_len.0006. ch_pow = sum(abs(chirp) .tot_len). chirp_rate = . %generating chirpfor k = 1:tot_len+1 chirp(k) = 10*exp(j * (k1)*(k1)*chirp_rate). chirp = zeros(1. plot(trn). end strt = strt +sym_len.37 %modulating with square pulsestrt = 1.
5/90).5/90). %calculating the filter window . j] end trn_fr = frft(trn. plot(abs(chirp_frd)). plot(abs(chirp_fr)). ang/90). plot(abs(rcv_fr)).76. plot(abs(rcv_fr)). rcv_fr = frft(rcv. plot(abs(trn_fr)). ang/90). rcv_fr = frft(rcv.^3)/tot_len). chirp_frd = frft(chirp.76. %finding out the optimum anglefor ang = 90:1:90 chirp_fr = frft(chirp. j = (sum((abs(rcv_fr)). [ang.5/90). plot(abs(trn_fr)).38 %snr calculation before processingsnr = 10 * log10(sgn_pow/ch_pow) plot(real(rcv)).76. trn_fr = frft(trn. ang/90).
plot(g).76.*rcv_fr. for m = 1:sym_len acc = acc+rcv_dns(strt+m). else data_rcv(k) = 0. end strt = strt +sym_len. rcv_dns = frft(rcv_fr. SNR1 = 10 * log10(sgn_pow/err_pow) %demodulatingstrt = 1./(1+g). for k = 1:no_sym acc =0.*abs(rcv_dnstrn’)). plot(abs(rcv_fr)).5/90).1).39 g = (abs( chirp_frd)). . plot(real(rcv_dns)). g = ones(length(g).^2. end if(acc>0) data_rcv(k) =1. %calculating error power err_pow = sum(abs(rcv_dnstrn’). rcv_fr = g.
we can search for a fractional order.data_rcv). That is.40 end %finding bit errorerr = abs(data . The fraction of energy in other frequency components will be very small.3. For signals with highly non stationary spectral characteristics. most of the coeﬃcients will be zero and the resulting data can be easily compressed. Fourier domain gives the best performance for compression compared to other fractional domains. After a thresholding all such low energy components will be removed. signal has to be divided into ﬁnite length segments and fractional order optimization has to be applied on each segment.3. A chirp signal is transmitted from the transmitter side and received signal is analyzed . 4. Consider the case of a speech data sample. If we use FrFT instead of Fourier transform. We observed that for speech and music. most of the energy will be concentrated on very few frequency components.2 Multipath Channel Estimation Using FrFT Channel characteristics of a static multipath channel can be measured using FrFT. which gives highest energy compaction for a given signal.1 Other Applications FrFT for Compression Transforms are widely used for signal compression applications. After applying Fourier transform.3 4. 4. Thus compression can be optimized. stem(err).
3.6x(t − 450Ts ) (4. chirp_len = 600. Fα [x(t − τ )] (u) = Fα [x(t)] (u − τ cos α) (4. The multipath model used is y(t) = x(t) + 0. Fα [x(t − τ )] (u) = Fα [x(t)] (u − τ cos α)ej If we consider the magnitude. Chirp signals gives sharp peaking in fractional domains decided by chirp parameters. end .16) Received signal will have many chirp pulses. %generating chirpfor k = 1:chirp_len chirp(k) = exp(j * (k1)*(k1)*chirp_rate). each with diﬀerent delay and diﬀerent gain.0005.18) 4. chirp = zeros(1.chirp_len). chirp_rate = .1 Matlab Code %chirp filtering in a multipath channel+wgn clear variables. Consider the time shift property of FrFT. Each impulse corresponding to diﬀerent copies of transmitted chirp signal will be shifted in fractional domain by τ cos(α) where τ is the time shift between the chirps and α is the fractional angle at which the transmitted chirp gives peaking.17) τ2 2 sin α cos α−juτ sin α (4.2.8x(t − 200Ts ) + 0.41 using FrFT.
8.8: (a) Transmitted chirp signal (b) Received signalMulti Path.6 Angle = −800 .tot_len). plt = 1. 0. 0.42 plot(real(chirp)). rcv = zeros(1. %generating a multipath model output Figure 4. tot_len = 1200. WGN SNR = 6 dB (c) Signal after FrFT for coeﬃcients 1.
ord/90). nos_pow = sum(abs(n) . plt = 1. %generating wgnn = wgn(1.* abs(n)). end plot(real(rcv)). ord . plt =1. rcv(450+k) = rcv(450+k) + . plot(real(rcv)). rcv_pow = sum(abs(rcv) . plt = 1.8 * chirp(k). plot(n).6).6 * chirp(k). rcv(200+k) = rcv(200+k) + . SNR = 10 * log(rcv_pow / nos_pow) rcv = rcv + n. plt =1.2.tot_len. %search for optimum angle for ord = 90:1:90 rcv_fr = frft(rcv.43 for k = 1:chirp_len rcv(k) = rcv(k) + 1 * chirp(k).* abs(rcv)). plot(abs(rcv_fr)).
plt=0. . for k =1:tot_len if magn_rcv_fr(k) < threshold magn_rcv_fr(k) = 0. rcv_fr_est(485:495) = rcv_fr(485:495). 80/90). peak = max(magn_rcv_fr). end end plot(abs(rcv_fr)). magn_rcv_fr = abs(rcv_fr). val = 0. plt=0. plot(abs(rcv_fr_est)). plt = 0. rcv_fr(k) = 0. peak = 0.tot_len).4* peak. plot(abs(rcv_fr)). rcv_fr_est = zeros(1. threshold = .44 end rcv_fr = frft(rcv.
21) . A search in diﬀerent fractional domains is carried out for obtaining the peak. plot (real(rcv_est)).19) (4.3. 4.45 rcv_est = frft(rcv_fr_est. plt = 0. the reﬂection from a radially accelerating object will be a chirp. Transmitted signal St (t) = ej2πf0 t Received signal Sr (t) = e 2v 2a j2πf0 t+j2π λ t+π λ t2 +Φ0 0 0 (4.20) The estimation of radial acceleration is aest = − λ0 fs cot αmax 2T (4.3 FrFT for measuring the acceleration of a moving object in radial direction When illuminated by a constant frequency sinusoid. 80/90)’. Let αmax be the angle corresponding to maximum peaking [5].
5. 4. FrFT is computationally eﬃcient and has the same order of complexity as that of Fourier transform. It can be reduced to a great extent by taking frames with 50 percent overlap. 3. Most of the time. 6. 2. FrFT based computations are to be done framewise. which is a time consuming task. Signals of nonstationary spectral characteristics can be analyzed using FrFT with superior performance compared to Fourier transform.1 Critical Evaluation Based on the study and analysis of the simulation results. 7.46 Chapter 5 Critical Evaluation and Conclusion 5. optimum order for computation will be unknown. . FrFT gives one more degree of freedom while designing signal processing tools compared to Fourier transform. FrFT performs excellently if any of the signals that we consider is chirplike. a search over a range of fractional orders should be carried out. as a result. which will result in errors due to windowing and ﬁnite time duration. Signal analysis in timefrequency plane is easy with the help of FrFT. we conclude that: 1.
47 8. . FrFT based systems can replace the current frequency domain systems. 9. cognitive radio and so on. Signals after FrFT processing will contain complex part. advantage that we get by using FrFT is minimal. There is large scope of research in FrFT in the ﬁelds of spread spectrum communication. FrFT can have bigger roles in ﬁelds like radar and sonar where chirp signals are very common. For speech and music processing. which should be considered for any further processing. FrFT can also be used for the design of faster optical signal processing systems. 5.2 Conclusion We can conclude that FrFT is a more general method for signal processing and system design. signal watermarking and encryption.
Vol. Kerr. [5] Wenchao Du. Vol. “The Fractional Order Fourier Transform And Its Application To Quantum Mechanics. 2005.42.” IMA J.5. 1987.48 Bibliography [1] V. Mcbride and F. Applicat. May 1999. Pp.. Tao Ran and Zhou SiYong. [2] A.” J. 1980. M.” IEEE Transactions on Signal Processing. Alper Kutay and Gozde Bozdak. 24 Nov. 159175. Vol. September 1996. Math. Vol. No. . No.” IEEE Transactions on Signal Processing. “Digital Computation Of The Fractional Fourier Transform. MinHung Yeh and ChienCheng Tseng. Vol. Proceedings of the 2007 International Conference on Wavelet Analysis and Pattern Recognition. 39. No. [4] SooChang Pei. 241265. [3] Haldun M. “The Fractional Fourier Transform and TimeFrequency Representations. Orhan Ankan. November 1994.47.2.. [6] Qi Lin. Pp.44. China. “Rejection of Linear FM Interference in DSSS System Based on Fractional Fourier Transform. Almeida. [7] Luis B. C. “Discrete Fractional Fourier Transform Based On Orthogonal Projections.” IEEE Transactions on Signal Processing. Ozaktas. Beijing. “Using Frft To Estimate Target Radial Acceleration”. “OnNamias’ Fractional Fourier Transforms. H. Math. 25. Vol. Xueqiang Gao and Guohong Wang.” Journal of Beijing Institute of Technology.14. Appl. Inst.11.9. No. 2007. Namias.
May 1997. “Digital Computation of the Fractional Fourier Transform. Vol. Orhan Arikan and Levent Onural. “Optimal Filtering in Fractional Fourier Domains.” IEEE Transactions on Signal Processing. Orhan Ankan.5. M.44.49 [8] Haldun M. Ozaktas.45. Ozaktas. No. Haldun M.” IEEE Transactions on Signal Processing.September 1996. Vol. Alper Kutay and Gozde Bozdaki. Alper Kutay.9. [9] M. . No.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue reading from where you left off, or restart the preview.