CHAPTER 2 LITERATURE SURVEY - Shodhgangashodhganga.inflibnet.ac.in/bitstream/10603/9391/2/11....
-
Upload
truongtuong -
Category
Documents
-
view
234 -
download
0
Transcript of CHAPTER 2 LITERATURE SURVEY - Shodhgangashodhganga.inflibnet.ac.in/bitstream/10603/9391/2/11....
[13]
CHAPTER 2 LITERATURE
SURVEY
HE literature review summarizes, interprets, and evaluates existing "literature" (or
published material) in order to establish current knowledge of a subject. The purpose for
doing so relates to ongoing research to develop that knowledge. The literature review may resolve a
controversy, establish the need for additional research, and define a topic of inquiry.
2.1 LCT AND ITS TYPES
The linear canonical transformation (LCT) is a family of integral transforms that generalizes
many classical transforms. It has four parameters and one constraint, so it is a 3-dimensional family, and
can be visualized as the action of the special linear group on the time–frequency plane [73, 84, 126, and
127]. It was first introduced in 1970s [128, 144]. Many operations, such as the Fresnel transform [71],
Fourier transform (FT), fractional Fourier transform (FRFT) [76, 151], and scaling operations are the
special cases of the LCT.
The linear canonical transform of a signal 푓(푡) with parameter set (푎, 푏, 푐,푑) is defined [134] as -
퐿( , , , )[푓(푡)] = 퐹( , , , )(푢) = 푒
∫ 푒 ∞ 푒
푓(푡) 푑푡 ∀ 푏 ≠ 0
√푑 푒 푓(푑 푢) ∀ 푏 = 0 (2.1.1)
Where, 퐿( , , , )[ ] is the LCT operator, 퐹( , , , )(푢) is the LCT of signal 푓(푡) and the four parameters
(푎, 푏, 푐,푑) associated with LCT should satisfy the following relation-
푎 푑 − 푏 푐 = 1 (2.1.2)
For the different values of parameter set (푎, 푏, 푐,푑), LCT transforms into special transform technique
which has different application areas. For –
푎 푏푐 푑 = 1 휆푧
0 1 (2.1.3)
Where, 휆 is the wavelength and 푧 is the distance, the LCT converts into Fresnel transform. The
Fresnel transform is the operation that describes the monochromic light propagating through the free
space [71]. The 2-Dimensional expression of Fresnel transform is given as –
T
[14]
퐹( , , , )(푢) =
∫ ∫ 푒 ( ) ( ) 푓(푥,푦) 푑푥 푑푦∞∞ (2.1.4)
Where, 푓(푥,푦) is the distribution of the monochromatic light, 휆 is its wavelength, and 푧 is the
propagation distance.
Similarly, fractional Fourier transform (FRFT) is a type of LCT when the parameter set
(푎, 푏, 푐,푑) of LCT is changing as -
푎 푏푐 푑 = 푐표푠(훼) 푠푖푛(훼)
− 푠푖푛(훼) 푐표푠(훼) (2.1.5)
Where, ‘α’ being the angle parameter related with FRFT, defined in time-frequency plane. Also,
LCT converts into Fourier transform (FT), when –
푎 푏푐 푑 = 0 1
− 1 0 (2.1.6)
Analyzing (2.1.5) and (2.1.6), it is evident that FRFT converts into FT when 훼 = (휋 2⁄ ).
Therefore, FRFT may be considered as a generalization of Fourier transform (FT), which was initially
one of the most frequently used tool in signal processing [120].
2.2 MATHEMATICAL PROPERTIES OF FRFT
The definition of FRFT (1.3.13 – 1.3.15), shows that for angle ‘α’ (not a multiple of π) the
computation of FRFT corresponds to following steps [76]:
Multiply by a chirp
Fourier transform with its argument scaled by ‘cosec(α)’
Multiply with another chirp
Multiply again by a complex amplitude factor
The utility of any transform technique is based on the set of mathematical properties possessed by
it because these properties of transform are actually responsible for the application area of the respective
transform technique. For any transform, the closed form expression of its inverse should exist in order to
satisfy unitary property which is a necessary condition for a transform. Similarly, the modulation,
shifting, integration and differentiation properties are facilitating the transform to be used in various
[15]
application areas. The convolution property for a transform is required at most which fits the transform in
filtering and amplification areas. Subsequently, if the transform is having the correlation theorem then it
may be used in radar communication, missile tracking and navigation purpose. Hence, a set of properties
is needed for any transform to have some realizable applications. This aspect of the FRFT is going to be
covered in following sections.
2.2.1 Properties Satisfied by the Kernel of FRFT
The kernel of FRFT ‘K (t, u)’ satisfies the following properties [10, 76] –
K (t, u) = K (u, t) (2.2.1)
K (t, u) = K∗ (t, u) (2.2.2)
K (−t, u) = K (t,−u) (2.2.3)
∫ K (t, u) K (u, z) du = K (t, z) (2.2.4)
∫ K (t, u) K∗ (t, v) dt = δ(u− v) (2.2.5)
Where the asterisk ‘∗’ represent the complex conjugate. The first three properties represented by
(2.2.1 – 2.2.3) can be easily derived from the definition of FRFT. The derivation of property (2.2.4) is
based on the definition of kernel of the FRFT and property (2.2.5) can be derived by considering the LHS
of identity (2.2.5)-
∫ K (t, u) K∗ (t, v) dt (2.2.6)
Using (2.2.1), K (t, u) is changed as K (u, t), and (2.2.6) converts into -
∫ K (u, t) K∗ (t, v) dt (2.2.7)
Now using (2.2.2), the conjugate of kernel is replaced and (2.2.7) becomes-
∫ K (u, t) K (t, v) dt (2.2.8)
And finally using (2.2.4) and definition of FRFT for α = 0, (2.2.8) can be written as
[16]
∫ K (u, t) K (t, v) dt = K ( )(u, v) = δ(u− v) (2.2.9)
As mentioned above, in comparison of FRFT, STFT and WD as a signal processing tool in time-
frequency plane, FRFT comes out as the winner based on its advantage over STFT in terms of
mathematical properties and clarity and over WD based on the cross-term interference, which is totally
removed in the FRFT. Motivated by this superiority of FRFT for analyzing the non-stationary signals,
next section is dedicated to the development of FRFT as a tool in time-frequency plane.
2.2.2 FRFT in Time-Frequency Plane
The expression of FRFT is defined for entire time-frequency plane. In the definition, the value of
angle parameter ‘α’, governs the rotation of the signal to be transformed in time-frequency plane. Given
that ‘F’ is Fourier transform operator, ‘I’ is the identity operator and ‘Fα’ is fractional Fourier transform
operator, the FRFT posses the following important properties in time-frequency plane [76, 121]-
Identity operator (Zero rotation): F0 = I (2.2.10)
Fourier operator (π/2 rotation): Fπ/2 = F (2.2.11)
Reflection operator (π rotation): Fπ = -I (2.2.12)
Inverse Fourier operator (3π/2 rotation): F3π/2 = F-1 (2.2.13)
Identity operator (2 π rotation): F0 = I (2.2.14)
Angle additivity property: Fα F β = Fα+β (2.2.15)
Inverse FRFT: Fα-1 = F- α (2.2.16)
Property (2.2.10), (2.2.12) and (2.2.14) holds by definition of FRFT (1.3.13 – 1.3.15). Property
(2.2.11) is a consequence of the fact that, for α = π/2, the kernel of FRFT ‘K (t, u)’ coincides with the
kernel of Fourier transform and property (2.2.13) is a consequence of the fact that for α = - (π/2),
‘K (t, u)’ coincides with the kernel of inverse Fourier transform. Property (2.2.15) can be easily derived
by using property (2.2.4) as-
Let the FRFT of a signal x(t) is defined as-
푦(푢) = 푋 (푢) = ∫ x(t) K (t, u) dt (2.2.17)
Then FRFT of 푦(푢) will be given as-
[17]
푌 (푧) = ∫ 퐾 (푢, 푧) 푦(푢) 푑푢 =∫ 퐾 (푢, 푧) ∫ x(t) K (t, u) dt 푑푢 (2.2.18)
By changing the order of integration and rearranging the terms-
푌 (푧) = ∫ x(t) ∫ K (t, u) 퐾 (푢, 푧) 푑푢 dt (2.2.19)
And finally using the property (2.2.4) of the kernel of FRFT-
푌 (푧) = ∫ x(t) K (t, z) 푑푡 = X (z) (2.2.20)
This proof shows that two successive FRFT of a signal by angles α and β will be same as a single
FRFT of the same signal by angle α + β.
Similarly, the last property (2.2.16) can be derived by using the fact that FRFT of a signal by
angle ‘2π + α’ is same as the FRFT of the signal by angle ‘α’. By taking the FRFT of 푋 (푢) with an
angle ‘2π – α’ which will equal to the FRFT of 푋 (푢) with an angle ‘α’ and is given as-
∫ 푋 (푢) K (u, t) du =∫ 푋 (푢) K (u, t) du (2.2.21)
Using (2.2.2) -
∫ 푋 (푢) K (u, t) du = ∫ 푋 (푢) K∗ (u, t) du = x(t) (2.2.22)
Hence,
x(t) = ∫ 푋 (푢) K (u, t) du (2.2.23)
Therefore, the inverse FRFT of a transformed signal with angle ‘α’ can be calculated as the FRFT
of the transformed signal with angle ‘-α’.
Thus, FRFT may be thought as a method to express any signal on the basis by a set of
functions K (u, t), with ‘u’ acting as a parameter for spanning the set of basis functions [76]. The basis-
functions are chirp functions, orthonormal according to (2.2.5). For different value of ‘u’ the basis-
functions are only differ by a time-shift and phase factor that depends on u as -
K (t, u) = 푒 K (t − u sec α, 0) (2.2.24)
[18]
For any transform being interpretable as a rotation in time-frequency plane, it has to satisfy the
properties given by (2.2.10 – 2.2.16). Here, the properties like rotation in time-frequency plane, inverse
and angle additivity process are very well defined expressions thus making the FRFT an appropriate tool
[76]. The reason for being - if the FRFT is multiplied by 푒 with integer n, the properties of the
transform, i.e., given in (2.2.10 – 2.2.16) would still hold. In literature, the existence of any other
transform method which obeys this set of properties with such simple relationship with time-frequency
distribution has not been reported.
2.2.3 Properties of the FRFT
Various properties that a transform should uphold in order to make its application in various areas
are described in Table-2.1, such as: Translation, Modulation, Mixed product, Differentiation, Integration,
Multiplication with time variable, Division by time variable, Similarity / Parity, Scaling in time domain,
Linearity and Convolution.
The properties listed in Table-2.1 for FRFT are the extensions of the respective properties of FT.
Their proofs, except scaling property can be found in [2, 76, and 151]. Similarity/Parity property tells us
that if 푥(푡) is even, 푋 (푢) will also be even. Similarly, if 푥(푡) is odd, 푋 (푢) will also be odd. The proof
of scaling property is obtained by considering two signals 푥(푡) and 푦(푡), which are related as -
푦(푡) = 푥(푐푡) (2.2.25)
By taking the FRFT of 푦(푡) as -
푌 (푢) = ∫ 푥(푐푡) 푒 푑푡∞ (2.2.26)
Substituting, t = c t, (2.2.26) becomes -
푌 (푢) = | | ∫ 푥(푡) 푒 – 푑푡∞ (2.2.27)
After some manipulation (2.2.27) can be rearranged as -
푌 (푢) = | | ∫ 푥(푡) 푒 – 푑푡∞ (2.2.28)
[19]
Table 2.1- Properties of Fractional Fourier Transform
S.N. Properties Signal FRFT with angle α
1
Linearity
푎 푥(푡)
+ 푏 푦(푡)
푎푋 (푢) + 푏푌 (푢)
2
Scaling in time
domain
푥(푐푡) 1− 푗 cot훼푐 − 푗 cot 훼
푒 푋푢 sin훽푐 sin훼
,
for 훽 = tan (푐 tan훼)
3
Translation
푥(푡 − 휏) 푋 (푢 − 휏 csc훼) 푒 –
4
Modulation 푥(푡)푒 푋 (푢 − 푣 sin훼) 푒
5
Mixed product 푡푑 푥(푡)푑푡
− sin훼 (sin훼 − 푗푢 cos훼)푋 (푢) + 푢 cos 2훼 푋 (푢)
+ 푗sin 2훼
2푋 (푢)
6
Differentiation
푥 (푡)
푋 (푢) cos훼 + 푗 푢 푋 (푢) sin훼
7
Integration 푥(푡) 푑푡 sec훼 푒 푋 (푧) 푒 푑푧, ∀ 훼
≠ (푛 + 1)휋2
8 Multiplication
with ‘t’
푡 푥(푡)
푢 푋 (푢) cos훼 + 푗 푋 (푢) sin훼
9
Division by ‘t’
푥(푡)푡
−푗 sec훼 푒 푥(푧) 푒 푑푧, ∀ 훼
≠ 푛휋
10 Similarity /
Parity
푥(−푡)
푋 (−푢)
11 Parseval’s
Theorem 푥(푡)푦∗(푡) 푋 (푢) 푌∗(푢) 푑푢
[20]
Now by defining, 훽 = cot = tan (푐 tan훼), and rearranging terms,
푌 (푢) = ∫ 푥(푡) 푒 – 푑푡∞ (2.2.29)
After some step of manipulation (2.2.29) can be written as-
푌 (푢) =
푒 푋 (2.2.30)
The scaling property shows the effect of a change of units of the independent variable ‘푡’ in time
domain on the transformed quantity. In the classical Fourier transform, the effect of such a change is only
a corresponding change in units of the transform ‘frequency’ variable and a scaling of the amplitude of
the transformed quantity. Similarly, for the FRFT, the effects are a scaling of the transform ‘u’ variable
by, ′sin훽/푐 sin훼′, a complex amplitude scaling, a product by chirp, and most importantly, a change in
the angle at which transform is computed from 훼 to 훽 = cot (cot 훼 푐⁄ ) = tan (푐 tan훼). This effect
can be explained as - contracting the time axis in time-frequency plane has an analogous effect on
expansion of the frequency axis by the same factor. [76]. If both effects are analyzed simultaneously, as
FRFT works in time-frequency plane, then it will become clearer that the original angle at which the
transform has to be calculated will now rotate to a new angle. Rotation of domain ‘angle’ will be
counterclockwise if time axis is contracting and clockwise if time axis is expanding.
The Parseval’s theorem for FRFT [76] can easily be proved by expressing 푥(푡) as the inverse
fractional Fourier transform of 푋 (푢) as -
∫ 푥(푡)푦∗(푡)푑푡∞ = ∫ ∫ 푋 (푢) 푒 푑푢∞∞ 푦∗(푡)푑푡 (2.2.31)
After interchanging the order of integration and rearranging the terms -
∫ 푥(푡)푦∗(푡)푑푡∞ = ∫ 푋 (푢) ∫ 푦∗(푡) 푒 푑푡∞∞ 푑푢 (2.2.32)
Using (2.2.2) and the definition of FRFT, (2.2.32) can be written as -
∫ 푥(푡)푦∗(푡)푑푡∞ = ∫ 푋 (푢) 푌∗(푢) 푑푢∞ (2.2.33)
[21]
If both functions are assumed as same, then as a consequence of this equality is the energy-
preserving property of FRFT.
∫ |푥(푡)| 푑푡∞ = ∫ |푋 (푢)| 푑푢∞ (2.2.34)
Due to the energy-preserving property of the Fourier transform, the squared magnitude of the
Fourier transform of a signal |푥(휔)| is called as the energy spectrum of the signal and is interpreted as
the distribution of the signal’s energy among the different frequencies 푒 . Similarly, from the energy-
preserving property (2.2.34) of FRFT, |푋 (푢)| , the fractional energy spectrum of the signal 푥(푡) with
angle α can be interpreted as the distribution of the signal energy among the different chirps K α(t, u).
The sampling theorem of FRFT plays an important role in communication systems and signal
processing. Sampling is a process of converting a continuous-time band limited signal to obtain its
discrete-time version. Sampling theorem of FT states that “a bandlimited signal can be reconstructed
exactly if it is sampled by a rate at least twice the maximum frequency component in it”. For discrete
processing of non-stationary signals the FRFT domain sampling theorem is needed. In literature many
attempts were made to establish sampling theorem [7, 26, 122, 147, and 156]. Tao et al [122] presented a
theorem for FRFT which was based on the product theorem of FRFT and derived sampling theorem with
reconstruction formula, which explains how to sample a continuous-time signal to obtain its discrete-time
version for band limited signals in the fractional Fourier domain.
The FRFT of a signal 푥(푡) satisfies the following condition for 0 ≤ Ω ≤ Ω -
푋 (푢) = 0, 푤ℎ푒푛 |푢| > Ω or |푢| < Ω (2.2.35)
Then 푥(푡) is defined as a bandlimited signal in FRFT domain with bandwidth given below -
Ω = Ω −Ω (2.2.36)
For 푥(푡), the sampling theorem is defined for an appropriate value of 푁, for 1 ≤ 푁 ≤ 푖푛푡[Ω Ω⁄ ] as –
Ω α ≤ Ω ≤ Ω α (2.2.37)
Where, sampling rate Ω is given as Ω = 2π T⁄ for sampling interval ′T ′.
The uncertainty principle says that a function and its Fourier transform cannot be simultaneously
"small". The origin of such type of statements lies in quantum mechanics - for a quantum particle it is
[22]
impossible to determine its position and momentum simultaneously. In the classical time-frequency
domain, the product of time spread and frequency spread cannot be small unlimitedly. Their tradeoff
relationship is called the uncertainty principle, i.e., for one given real absolutely integratable signal 푥(푡)
and its Fourier transform 푋(푢), the product has the lowest bound and is given as –
∆푡 ∆푢 = ∫ |(푡 − 푡 ) 푥(푡)| ∞∞ 푑푡 ∫ |(푢 − 푢 ) 푋(푢)| ∞
∞ 푑푢 ≥ 14 (2.2.38)
Where,
푡 = ∫ 푡 | 푥(푡)| ∞∞ 푑푡 (2.2.39)
And
푢 = ∫ 푢 | 푋(푢)| ∞∞ 푑푢 (2.2.40)
For ∆푡 is the time spread and ∆푢 is the frequency spread.
It has been shown in FRFT domain that uncertainty principle has lower bounds than those of
traditional cases, which provides the possible selections for improved resolution analysis. Therefore, there
is a great need to extend the traditional entropic uncertainty principle to the FRFT domains as well. Many
articles are available in literature on uncertainty principle of FRFT [142, 154, and 155]. Guanlei et al
[154] had derived logarithmic, Heisenberg’s and short-time uncertainty principle associated with FRFT.
Subsequently, Guanlei et al [155] gave two entropic uncertainty principle based on fractional Fourier
transform, in which one was Shannon’s entropy uncertainty principle and the other was Renyi’s entropy
uncertainty principle. Shinde et al [142] introduced an uncertainty principle for real signals in the
fractional Fourier domain, in which a lower bound on the uncertainty product of signal representations in
two FRFT domains for real signals was obtained, and it was shown that a Gaussian signal achieves the
lower bound.
For a unit energy real-valued signal 푥(푡) and its fractional Fourier transform, 푋 (푢), the
uncertainty principle can be defined as –
∆푢 ∆푢 ≥ ∆푡 cos훼 cos훽 + ∆
+ { ( )} (2.2.41)
Where,
∆푢 = ∫ |(푢 − 푢 ) 푋 (푢)| ∞∞ 푑푢 (2.2.42)
[23]
And equality is achieved when
푥(푡) = 푒푥푝 −
(2.2.43)
Where, 휎, is an arbitrary real constant.
2.2.4 Existing Convolution & Product Theorem for FRFT
Studies reported in above sections reveals that almost all the properties of FRFT are already
derived except convolution theorem, product theorem and correlation theorem. The usefulness of
convolution theorem can be best explained by its application in filtering. Since, filtering can be performed
in both ways i.e., time domain filtering and frequency domain filtering. Simultaneously, if the
computational complexity is a basis parameter then it can be shown that under different input conditions
one type of filtering has advantage over other and vice-versa [30]. . For non-stationary signals and noise,
filtering of the signal from the noise can be performed by designing a filter in FRFT domain with this
optimum angle parameter value [121]. FRFT domain convolution and product theorems were extensively
investigated previously and summarized as - definitions given by Almeida [77], Zayed [6], Deng Bing et
al [28], and Deyun Wei et al [35].
2.2.5 Existing Correlation Theorems and Related Theories
The fractional domain correlation theorem was analyzed and reported recently in many articles
[43, 103, 109, 123, 124, 125, and 161]. Firstly, Akay et al [103] had defined the fractional correlation
integral by using fractional shift operator, but an exclusive identity for evaluating the correlation was not
presented. Similarly, if the FRFT of correlation integral is determined then it results in a very complex
expression. Subsequently, Tao et al [123] had tried to formulate the correlation theorem especially for
FRFT, but in their definition the correlation integral contains only one chirp. It suggests that once FRFT
of this correlation integral will be carried out, the transformed quantity will not satisfy the variable
dependability constraint. Cottone [43] used fractional spectral moments to represents the fractional
correlation and power spectral density but the correlation theorem for FRFT was not documented
concisely. Torres [125] had documented another correlation theorem for the FRFT by using fractional
translational operator.
[24]
By definition, the FRFT is a subclass of integral transformations characterized by quadratic
complex exponential kernels. These complex exponential kernels often introduce very fast oscillations.
Hence, it is not possible to evaluate these transformations by direct numerical integration since these fast
oscillations require excessively large sampling rates. A possible approach is to decompose these integral
trans- formations into sub-operations. This results in greater time of computation, larger numerical
inaccuracies, and the need for more memory. Also with the increasing trend of analyzing the signal in
discrete form to utilize the advantages of digital systems over analog systems, the need of discretization
of FRFT, i.e. definition or algorithms to obtain discrete FRFT (DFRFT), arises rapidly. Next section is
dedicated to the analysis of existing DFRFT algorithms.
2.3 DISCRETE FRACTIONAL FOURIER TRANSFORM (DFRFT)
Many researchers have tried to come up with some method of evaluating DFRFT. On this
journey, many methods and algorithms had come into existence. But none got acclamation due to
deficiency of either not satisfying some of the prime properties that it’s continuous type posses or non-
existence of a closed form expression. In this section, all the approaches available in the literature are
included and divided into six classes on the basis of their methodology of evaluation.
2.3.1 Sampling Type DFRFT
The sampling theorem for the FRFT of band-limited and time-limited signals is follows from
Shannon’s sampling theorem [5, 7, 26, 74, 122, 147, and 156]. With this approach, the simplest way to
derive the DFRFT is sampling the continuous FRFT and computing it directly from the samples, but by
sampling the continuous FRFT directly the resultant discrete transform obtained will lose many important
properties. The most serious problem with this type DFRFT will be non-compliance of unitary and
reversibility properties. In addition, the DFRFT obtained by direct sampling of the FRFT lacks closed-
form expressions and is non-additive, which affects its application domain. In order to maintain some of
the FRFT properties, a type of DFRFT was derived as a special case of the continuous FRFT as reported
in [105]. Specifically, the input function was assumed as a periodic, equally spaced impulse train. Since
this type of DFRFT is a special case of continuous FRFT, many properties of the FRFT exist and have the
fast algorithm. However, this type of DFRFT cannot be defined for all values of α due to various imposed
constraints. At the same time, Ozaktas et al [55] had proposed two innovative approaches for obtaining
[25]
the DFRFT through the sampling of the FRFT. Both methods were based on the idea of manipulating the
expression of the FRFT by sampling appropriately.
In the first method, the DFRFT of a signal 푥(푡) can be obtained by, multiplying the signal with a
chirp function as -
푥 (푡) = 푒 푥(푡) (2.3.1)
For this chirp multiplication, the bandwidth and time-bandwidth product of 푥 (푡) can be as large
as twice of 푥(푡). Thus, the sampling of 푥 (푡) will be done at intervals of 12 ∆푥 . If the samples of
푥(푡) are spaced at 1∆푥 , then this need to be interpolate before multiply it by the samples of the chirp
function to obtain the desired samples of 푥 (푡).
The next step was to convolve the signal 푥 (푡) with another chirp function. To perform this
convolution, the chirp signal was replaced by its band-limited version because 푥 (푡) assumed as a band-
limited signal.
푥 (푡) = 퐴 ∫ 푥(휏) 푒 ( ) 푑휏 (2.3.2)
The band-limited version of chirp function can be expressed in the form of Fresnel integral.
Finally, again performing a chirp multiplication, the DFRFT of signal 푥(푡) can be obtained as -
ℑ[x(t)] = 푋 = 푒 푥 (푡) (2.3.3)
The convolution operation in the above equation can be achieved by sampling 푥(푡), and
performing the convolution using the fast Fourier transform (FFT). Overall, the procedure starts with N
samples spaced at 1∆푥 , which uniquely characterize the function 푥(푡), and returns the same for 푋 .
For x and 푋 denote column vectors with N elements containing the samples of 푥(푡) and its DFRFT
respectively. The above procedure can be given in a matrix notation as -
푋 = 퐷 Λ 퐻 Λ 퐽 푥 (2.3.4)
Here, 퐷 and 퐽 are matrices representing the decimation and interpolation operations, Λ is a
diagonal matrix that corresponds to chirp multiplication, and 퐻 corresponds to the convolution operation.
The above method had allowed obtaining the samples of the 푎 transform in terms of the samples of the
original function.
[26]
First method includes the analysis of one step in terms of Fresnel integral, which require large
number of computation time particularly in inverse DFRFT. To remove this discrepancy, a second
method was introduced by Ozaktas et al [55]. In this method one assumption was taken that the Wigner
distribution of function 푥(푡) could be zero outside a circle of diameter ∆푢 centered at the origin and by
limiting the order ‘a’ to the interval 0.5 ≤ |푎| ≤ 1.5. Simultaneously, the amount of vertical shear in
Wigner space resulting from the chirp modulation is bounded by ∆푢/2. The FRFT can be written as –
ℑ[x(t)] = 푋 (푢) = 퐴 푒 ∫ 푒 푒 푥(푡) 푑푡 (2.3.5)
Thus, 푒 푥(푡) can be written by using Shannon’s interpolation formula as –
푒 푥(푡) = ∑ 푒 ∆ 푥 푛2∆푢 푠푖푛푐 2∆푢 푡 − 푛
2∆푢 (2.3.6)
Where, 푁 = (∆푢) , putting (2.3.6) in (2.3.5) and interchanging the order of integration and summation -
푋 (푢) = 퐴 푒 푒 ∆ 푥푛
2∆푢푒 푠푖푛푐 2∆푢 푡 −
푛2∆푢
푑푡
(2.3.7)
And after some algebraic manipulation, (2.3.7) can be restructured as-
푋∆
=∆
푒 ( )( ∆⁄ ) ∑ 푒 ∆ 푒 ( )∆ 푥
∆ (2.3.8)
This summation can be interpreted as the convolution of 푒 ∆ and the chirp modulated
function 푒 ( )∆ 푥
∆. This convolution can be computed by fast Fourier transform
(FFT). And the overall methodology can be written in matrix form as-
푋 = 퐷 퐾 퐽 푥 (2.3.9)
Where, the 퐾 matrix was given as -
퐾 (푚,푛) =∆
푒 ∆ ( ∆ ) ∆ (2.3.10)
[27]
In the same article, Ozaktas et al [55] gave alternate methods to calculate the DFRFT by
introducing scaling process. All these methods resulted in the same time complexity. However, in
practice, these alternate methods are not preferable because they require coordinate scaling. Scaling is not
an advantageous process if the original data has already been sampled. This is because scaling requires
additional interpolations, etc., which will require additional computation.
However, in both the cases, the Wigner distribution of the signal was confined to a circle of
diameter ∆푢 around the origin. Thus, there might be several discrete fractional Fourier transform
matrices, some of which were more elegantly expressible than the others, which gave the same result
within the accuracy of this approximation. All these matrices would yield results that were in increasingly
better agreement as the signal energy contained in the circle was increasingly closer to the total energy.
2.3.2 Linear Combination Type DFRFT
In [22, 23, 24, and 41], the discrete fractional Fourier transform was derived by using the linear
combination of identity operation (퐹 ), discrete Fourier transform 퐹 ⁄ , time inverse operation (퐹 ) ,
and inverse discrete Fourier transform 퐹 ⁄ . In [24], the concept of DFRFT was given by linear
combination of Lagrange interpolation polynomial of degree 3. For any function 푥(푣)
푥(푣) = ∑ 휙 (푣) 푔(휆 ) (2.3.11)
Where,
휆 = 푒 , 푓표푟 0 ≤ 푘 ≤ 3 (2.3.12)
And, 휙 (푣) is the Lagrange interpolation polynomial of degree 3 taking value ‘one’ at 푣 = 휆
and ‘zero’ otherwise. It follows that -
휙 (푣) = ∑ 푢 푒 / (2.3.13)
Finally by taking the principle nth - root of eigenvalue 휆 , the fractional transform can be given as -
퐹 = ∑ 퐹 훼 (푛) (2.3.14)
Where,
[28]
훼 (푛) = ∑ 푒 ( ) / , 푓표푟 0 ≤ 푖 ≤ 3 (2.3.15)
Subsequently, in [22, 23], it had also been shown that the operator defined by (2.5.15) is unitary,
satisfies angle additivity and angle multiplicity properties and the operator is periodic with a fundamental
period of four. One analogous approach of this method was presented in [41], in which the DFRFT was
calculated by first chirping the signal in time domain then taking its Fourier transform and finally chirping
again in frequency domain. Depending on the character of the chirped signal, the embedded FT can be
one of the two known classes in discrete case- discrete-time and periodic discrete-time.
For discrete-time DFRFT, the expression for DFRFT of signal 푥(푡) was given as-
푋 (푓) = ∑ 푇 푥(푘푇) 퐾 (푓,푘푇) (2.3.16)
And its inverse transform as -
푥(푘푇) = ∫ 푋 (푓) 퐾∗(푓,푘푇) 푑푓 (2.3.17)
Where, 퐹 =
and 퐾 (푡,푢) is the kernel of FRFT. Similarly, for periodic discrete-time
signal 푥(푡) with period 푇 the expression for DFRFT was given as -
푋 (푛푓) = ∑ 푇 푥(푘푇) 퐾 (푛퐹,푘푇) (2.3.18)
And its inverse transform as -
푥(푘푇) = ∑ 퐹 푋 (푛퐹) 퐾∗(푛퐹,푘푇) (2.3.19)
Where, 푇 = 푁푇; 퐹 = 푁퐹 =
In this type of DFRFT, the transform matrix is orthogonal, and satisfying the additivity property
along with the reversibility property. However, the main problem with this method that the transform
result is not matched to the continuous FRFT. Besides, it performed very similarly to the original Fourier
transform or the identity operation but loses the important characteristic of fractionalization.
[29]
2.3.3 Eigenvector Decomposition Type DFRFT
The authors had derived another type of discrete fractional Fourier transform as reported in [25,
27, 67, 130, 131, 135, 136, 137, and 139], by searching the eigenvectors and eigenvalues of the DFT
matrix followed by computing the fractional power of the DFT matrix based on Hermite function. This
type of DFRFT worked very similarly to the continuous FRFT and fulfills the properties of orthogonality,
additivity, and reversibility. The fractional power of matrix could be calculated from its eigen-
decomposition and the power of eigenvalues. Unfortunately, there exist two types of ambiguity in
deciding the fractional power of the DFT kernel matrix.
i. Ambiguity in Deciding the Fractional Powers of Eigenvalues- The square roots of unity are ‘1’
and –‘1’ from elementary mathematics. This indicates that there exists root ambiguity in deciding
the fractional power of eigenvalues.
ii. Ambiguity in Deciding the Eigenvectors of the DFT Kernel Matrix- The DFT eigenvectors
constitute four major eigen-subspaces, therefore, the choices for the DFT eigenvectors to
construct the DFRFT kernel are multiple and not unique.
Because of the above mentioned ambiguities, several DFRFT kernel matrices were possible
which could obey the rotational properties. The idea for developing the DFRFT is to find the discrete
form of the following -
퐾 (푡,푢) = ∑ 푒 퐻 (푡) 퐻 (푢) (2.3.20)
Where, 훼 indicates the rotation angle associated with FRFT of signal, 퐻 (푡) represents the nth -
order normalized Hermite function with unit variance. The nth - order normalized Hermite function with
variance 휎 is defined as-
퐻 (푡) = ! √
ℎ 푒 (2.3.21)
Where, ℎ (. ) represents the nth - order Hermite polynomial. However, the normalized Hermite
functions with unitary variance 퐻 (. ) indicate the eigenfunction of the FRFT. With the help of Hermite
functions, the FRFT of a signal can be computed [136] as -
푋 (푢) = ∫ 푥(푡) 퐾 (푡,푢) 푑푡 =∑ 퐻 (푢)푒 ∫ 푥(푡) 퐻 (푡) 푑푡 (2.3.22)
[30]
The FRFT can be interpreted as a weighting summation of Hermite functions, as given in
(2.3.22). The weighting coefficients are obtained by multiplying the phase term 푒 and the inner
product of the input signal with the corresponding Hermite functions [136]. The development of this
DFRFT was based on the eigen-decomposition of the DFT kernel. The eigenvalues of DFT kernel ‘F’ can
be given as {1,−푗,−1, 푗} and corresponding to each eigenvalue one eigenvector can be evaluated. The
continuous FRFT has a Hermite function with unitary variance as its eigenfunction. The corresponding
eigenfunction property for the DFT can be given as -
퐹 ⁄ [푢⏞ ] = 푒 푢⏞ (2.3.23)
Where, 푢⏞ represents the eigenvector of DFT corresponding to the nth-order discrete Hermite
functions. In order to retain the eigenfunction properties, the unit variance Hermite function is sampled
with a period 푇 = 2휋 푁⁄ . In the case of continuous FRFT, the terms of the Hermite functions are
summed up from order zero to infinity. However, for the discrete case, only eigenvectors for the DFT
Hermite eigenvectors can be added. The selection of the DFT Hermite eigenvectors is usually made from
low to high orders, due to small approximation error of the low DFT Hermite eigenvectors [136]. Finally,
the transform kernel of DFRFT can be defined as -
퐹 ⁄ = 푈⏞ 퐷 ⁄ 푈⏞ =∑ 푒 푢⏞ 푢⏞ , ∀ 푁 = 4푚 + 1∑ 푒 푢⏞ 푢⏞ + 푒 푢⏞ 푢⏞ , ∀푁 = 4푚
(2.3.24)
Where, 푈⏞ = 푢⏞ 푢⏞ … …푢⏞ for N as odd and 푈⏞ = 푢⏞ 푢⏞ … …푢⏞ 푢⏞ for N is even, 푢⏞
represents the normalized eigenvector corresponding to the 푘 - order discrete Hermite function and 퐷
can be defined as follows-
For, N odd -
퐷 ⁄ =푒
푒 0
0⋱
푒 ( )
(2.3.25)
And, for N even -
퐷 ⁄ =푒
⋱ 0
0푒 ( )
푒
(2.3.26)
[31]
The convergence for the eigenvectors obtained from these matrixes is not so fast for the high-
order Hermite functions. To ensure orthogonality of the DFT Hermite eigenvectors 푢⏞ , the Gram–
Schmidt algorithm (GSA) or the orthogonal procrustes algorithm (OPA) can be used [136]. The GSA
minimizes the errors between the samples of the Hermite functions and the orthogonal DFT Hermite
eigenvectors. On the other hand, the OPA minimizes the total errors between those samples. The main
difference between the approach proposed in [136] and similar approaches proposed in [130, 131, and
135] is found in the eigenvectors obtained. The eigenvectors obtained by method reported in [136] were
discrete Mathieu function, although the Mathieu functions can converge to Hermite functions as obtained
by methods reported in [130, 131, and 135]. Given the number of points 푁 and the rotation angle 훼, the
DFT Hermite eigenvectors can be computed followed by determining the eigenvalues of DFRFT. The
computation of the DFRFT can be implemented only by a transform kernel matrix multiplication. The
complexity of computing the DFRFT is 푂(푁 ) same as in the DFT case. By adjusting the rotational
angles, the method for implementing the DFRFT of a signal 푥 can be given as -
퐹 ⁄ = 푈⏞ 퐷 ⁄ 푈⏞ 푥 = ∑ 푎 푒 푢⏞ (2.3.27)
The coefficients 푎 ′푠 are the inner products of signal and eigenvectors, and they can be computed
in advance. If rotation angle changed then only the diagonal matrix 퐷 ⁄ need to be recomputed.
Additionally, the authors [136] investigated the relationship between the FRFT and the DFRFT and were
found that for a sampling period equal to 2휋 푁⁄ , the DFRFT performs a circular rotation of the signal in
the time–frequency plane. However, the DFRFT becomes an elliptical rotation in the continuous time–
frequency plane for sampling periods different from 2휋 푁⁄ [136]. Therefore, for these elliptical
rotations an angle modification and a post-phase compensation in the DFRFT have been required to
obtain results similar to the continuous FRFT [136]. This approach has been extended to the so-called
multiple-parameter discrete fractional Fourier transform (MPDFRFT) [67, 137]. In fact, the MPDFRFT
maintains all desired properties and reduces to the DFRFT when all of its order parameters are same.
A similar approach to [136] has been proposed in [27]. However, authors in [27] believe that the
discrete time counterparts of the continuous time Hermite–Gaussians maintained the same properties
because these discrete time counterparts exhibited better approximations than the other proposed
approaches [25]. In order to resolve this issue about the approximation of Hermite–Gaussian functions, a
nearly tri-diagonal commuting matrix of the DFT and a corresponding version of the DFRFT were
proposed in [139]. Most of the eigenvectors of this proposed nearly tri-diagonal matrix result in a good
[32]
approximation of the continuous Hermite–Gaussian functions by providing a smaller approximation error
in comparison to the previous approaches.
2.3.4 Closed Form Type DFRFT
Pei et al [132] had derived a different type of DFRFT which was neat in concept and reversible
but lacks the additivity property. Due to the orientation of practical usage, two types of DFRFT were
derived which are differing in parameterization. The parameters of the first type have directly linked to
the continuous FRFT and suits to the applications of computing the continuous FRFT. On the other hand,
second type has the simpler parameters set and allows more elegant expression for the operator kernels
and more suitable for other applications of DFRFT, such as the filter design, pattern recognition and
phase retrieval. Considering the expression of continuous FRFT -
퐹 (푢) = 푒 ∫ 푒 푒 푓(푡) 푑푡 (2.3.28)
And to derive the DFRFT, sampling of 푓(푡) and 퐹 (푢) with the interval ∆푡 and ∆푢 need to be performed
as –
푦(푛) = 푓(푛 ∗ ∆푡) & 푌 (푚) = 퐹 (푚 ∗ ∆푢) (2.3.29)
Where, 푛 = −푁, … … ,푁 and 푚 = −푀, … … ,푀. By using this sampling, FRFT can be converted [132] as-
푌 (푚) = ∆푡 푒 ∆
∑ 푒 ∆ ∆ 푒 ∆
푦(푛) (2.3.30)
The above can be written as -
푌 (푚) = ∑ 퐹 (푚,푛)푦(푛) (2.3.31)
Where,
퐹 (푚,푛) = ∆푡 푒 ∆
푒 ∆ ∆ 푒 ∆
(2.3.32)
And 푌 (푚) can be reversible only if the inverse transform of 퐹 (푚,푛), ∀ 푀 ≥ 푁 would be
Hermitian (conjugate and transpose), given [132] as-
[33]
푦(푛) = ∑ 퐹∗(푚,푛) 푌 (푚) 푓표푟 푀 ≥ 푁 (2.3.33)
From (2.3.31) and (2.3.33),
푦(푛) = ∑ ∑ 퐹∗(푚,푛) 퐹 (푚,푘)푦(푘) (2.3.34)
After some manipulation -
푦(푛) = ∆ | | ∑ ∑ 푒
∆
푒 ( ) ∆ ∆ 푦(푘) (2.3.35)
The summation for 푚 in (2.3.35) should equal to 훿(푛 − 푘), then -
∆푡 ∆푢 =
(2.3.36)
Where, |푆| is some integer prime to (2 푚 + 1), in this case (2.3.32) becomes -
퐹 (푚,푛) = ∆푡 푒 ∆
푒 푒
∆
(2.3.37)
And,
∑ ∑ 퐹∗(푚,푛) 퐹 (푚,푘)푦(푘) = | | ∆푡 푦(푛) (2.3.38)
After normalization, the transform matrix [132] can be given as -
퐹 (푚,푛) = | | ( ) 푒 ∆
푒 ( ) 푒
∆
(2.3.39)
The definition of DFRFT is further divided into two parts [132]. The first part was defined for ′sin훼 >
0′ as -
푌 (푚) = 푒 ∆
∑ 푒 푒
∆
푦(푛) (2.3.40)
And second part was defined for ′sin 훼 < 0 ′, as -
푌 (푚) = 푒 ∆
∑ 푒 푒
∆
푦(푛) (2.3.41)
[34]
This definition of DFRFT possesses reversibility property and the periodicity property -
푌 (−푚) = 푌 (푚) & 푌 (푚) = 푌 (푚) (2.3.42)
The DFRFT of type-1 has a very important advantage in terms of its efficiency in calculating and
implementing the DFRFT [132]. Due to two chirp multiplications and one FFT required for the
implementation of this type of DFRFT, the total number of the multiplication operations required was
‘2P+(P/2) log2P’, for length of the output as P=2M+1.
Subsequently, another type of DFRFT, named as type-2, was introduced by modifying the
definitions obtained for discrete affine Fourier transform (DAFT) and presented as -
푌 (푀) = 푒
∑ 푒± 푒
푦(푛) ∀ 푀 ≥ 푁 (2.3.43)
For the DFRFT of type-2, the parameter ‘p’ is used to control the variation of the chirp in
frequency domain [132]. The DFRFT of type-2 also needed ‘2P + (P/2) log2P’ multiplication operations.
It does not follow the additivity property but observes the convertibility operation.
These definitions of DFRFT are better to previous ones, i.e. sampling type, linear combination
type, and eigen-function type, as it require less number of multiplications, having less complexity. Also it
has a closed form expression and having rich in properties.
2.3.5 Weighted Summation Type DFRFT
Yeh [92] had developed a new method for DFRFT computation. The idea of the developed
method was to compute the DFRFT at any angle by a weighted summation of the DFRFT’s with the
special angles. In this method DFRFT computation with odd point of length can be realized by the
weighted summation of the DFRFTs in special angles. The special angles were multiples of 2π / N for the
odd case and multiples of 2π / (N+1) for even length. Regardless of even- or odd-length cases, the
weighting coefficients were obtained from an IDFT operation. The N-point IDFT was needed for odd
length, and the (N + 1) - point was computed for even length.
For discrete signal ′푥′ having odd length ′푁′, the DFRFT of ′푥′ for rotation angle ‘훼’ can be
computed as -
푋 = ∑ 퐵 , 푋 , ∀ 훽 = 2휋 푁⁄ (2.3.44)
[35]
Where, the weighting coefficients 퐵 , was given as -
퐵 , = 퐼퐷퐹푇 푒 ∀ 푘 = 0,1, … . ,푁 − 1 = ∑ 푒 푒 ( ⁄ ) (2.3.45)
The weighted summation, 퐵 , has a closed form solution -
퐵 , = ( )( )
( ) , ∀ 훼 ≠ 푘훽 훿(푛 − 푘), ∀ 훼 = 푘훽
(2.3.46)
Similarly, for discrete signal ′푥′ having even length N, the DFRFT of ′푥′ for rotation angle ‘훼’ can
be computed by the following expression -
푋 = ∑ 퐵 , 푋 , ∀ 훽 = 2휋 (푁 + 1)⁄ (2.3.47)
The weighting coefficients, 퐵 , can be computed as-
퐵 , = 퐼퐷퐹푇 푒 ∀ 푘 = 0,1, … . ,푁 = ∑ 푒 푒 ( ( )⁄ ) (2.3.48)
Similar to the odd length, the weighted summation, 퐵 , for even case has also a closed form solution -
퐵 , = ( )
( ) ∀ 훼 ≠ 푘훽 훿(푛 − 푘) ∀ 훼 = 푘훽
(2.3.49)
The eigenvector decomposition methods require N2 storages to store the transform kernel. If the
computing angle is changed, the transform kernel is also changed and needs to be recomputed. In this
algorithm, the DFRFT of the specified angle was stored. The DFRFT at any angle can be obtained by
weighted summation of these specified DFRFTs. The weighted computation in this algorithm still takes
푂(푁 ) multiplications; therefore, its computation load is still 푂(푁 ). In [92], two implementation
methods for the DFRFT computation algorithm were introduced. One was called the parallel method, and
the other cascade method. Similar to the parallel method, the special angle and numbers of terms of
cascade forms are also different for the even and odd cases. Both the implementation methods can have
the advantages over each other for different application. The parallel method was suitable for the signal,
whose DFRFT’s with special angles were already known. Hence, the computation of the DFRFT will
become only a linear combination of the DFRFT’s in special angles. And, the cascade method means that
the computation of the DFRFT can be realized from the DFRFT with only one specified angle. If the
[36]
DFRFT with the specified angle can be computed efficiently, the computation of the DFRFT will become
efficient. Therefore, it has been found suitable for VLSI implementation [92].
2.3.6 Random Type DFRFT
Pei et al [138] had introduced a modified method to calculate the DFRFT based on the random
DFT eigenvectors and eigenvalues. In this method a new commuting matrix with random DFT
eigenvectors was constructed. Then a random discrete fractional Fourier transform (RDFRFT) kernel
matrix with random DFT eigenvectors and eigenvalues was developed. The magnitude and phase of the
transformed output are found to be random in RDFRFT. This can be considered as a special feature of
RDFRFT. Previously, the DFT was randomized to define the discrete random Fourier transform (DRFT)
in [159] by taking random powers of eigenvalues of the DFT matrix. However, the eigenvectors of the
DRFT were not random and computed using the same method as proposed in [135]. Although, the phase of
the DRFT was random its magnitude was not random, as shown in [159]. The eigen-decomposition of DFT
matrix ‘F’ was given as -
퐹 = ∑ 휆 푒 푒 (2.3.50)
The DFRFT with one order parameter ‘a’ was defined by [137] -
퐹 = ∑ 휆 푒 푒 (2.3.51)
The definition of RDFRFT was based on the definition of multiple-parameter DFRFT
(MPDFRFT) [137], and given as -
퐹⏞ = ∑ 휆 푒 푒 (2.3.52)
Where, the 1 × 푁 parameter vector 푎⏞ was defined as -
푎⏞ = [푎 ,푎 , … … . . ,푎 ] (2.3.53)
This order parameter of the MPDFRFT can be taken as independent random numbers to define an
MPDFRFT with the following random eigenvalues [138], 휆 ,푘 = 0, 1, … . ,푁 − 1, and random
eigenvalues were given as-
[37]
휆 =
⎩⎪⎨
⎪⎧ 푒 , 푓표푟 휆 = 1
푒 ⁄ , 푓표푟 휆 = −푗푒 , 푓표푟 휆 = −1
푒 ⁄ , 푓표푟 휆 = 푗
(2.3.54)
The generation of RDFRFT is given in the following section. In this generation of RDFRFT, the
DFT-commuting matrix with orthonormal random DFT eigenvectors is used.
For a 푁 × 푁 real random Matrix ‘A’ being ‘K’ symmetric and generated by 퐴 = 퐾 퐴 퐾, where K
represents circular reversal matrix, given by -
퐾 = 퐹 = 퐹 = 1 00 퐽 (2.3.55)
Where, 퐽 being the (푁 − 1) × (푁 − 1) reversal matrix, having nonzero entries ‘ones’ placed
on the anti-diagonal locations. Taking the K-symmetric part ‘E’ of the random matrix ‘A’ as 퐵 =
{(퐴 + 퐾 퐴 퐾) 2⁄ }. With the help of this matrix ‘a random matrix ‘C’ was generated as 퐶 = (퐵 + 퐵 ) 2⁄ ,
where ‘T’ denotes the matrix transpose operation, and finally a DFT commuting matrix ‘D’ will be
formed [138] as 퐷 = 퐶 + 퐹 퐶 퐹 . Subsequently, with the random DFT-commuting matrix ‘D’ and the
MPDFRFT, the random-DFRFT (RDFRFT) with the 1 × 푁 parameter vector 푎⏞ = [푎 ,푎 , … … . . ,푎 ]
of a signal ‘x’ can be determined [138] as-
푋 ⏞, = 퐹⏞ ∗ 푥 = ∑ 휆 푟 푟 ∗ 푥 (2.3.56)
Where, 푟 represents the orthonormal random DFT eigenvectors computed from ‘D’ and 휆
represents the DFT eigenvalue corresponding to 푟 , given by (2.3.54).
The RDFRFT kernel matrix 퐹⏞ satisfied various properties of the FRFT like - additivity, FT
convertibility and angle additivity. The computational complexity of the -point RDFRFT was 푂(푁 )
same as that of the MPDFRFT, but RDFRFT has an advantage over MPDFRFT in terms of one more free
parameter associated with it as parameter vector 푎⏞. The magnitude and phase responses of the
transformed output, with the RDFRFT, are random because the eigenvectors and eigenvalues of the
RDFRFT matrix are random [138] as shown in (2.3.56). This aspect establishes its advantageous role in
security areas like - cryptography, encryption and watermarking.
[38]
Although, the existence of one well accepted closed-form definition of DFRFT is not available, in
spite there are many definitions with their pros and cons are available. A comparison of all the proposed
classes of DFRFT based on some properties and complexity is shown in Table-2.2. In Table-2.2, YES is
included against the method which posses given property \ parameter otherwise NO is mentioned.
Table-2.2: Comparison of Different Types of DFRFT Algorithms
Properties /Parameters
Sampling Type
Linear Combination
Type
Eigenvector Decomposition
Type
Closed Form Type
Weighted Summation
Type
Random Type
Reversibility NO YES YES YES YES YES Additivity NO YES YES NO YES YES Periodicity YES YES YES YES YES YES Symmetry YES NO YES YES YES YES Angle Additivity
NO NO YES YES YES YES
Closed form expression
YES YES YES YES YES YES
Similarity with FRFT
YES NO YES YES NO YES
Complexity O(N log2N) O(N log2N) O(N2) O(N log2N) O(N2) O(N2) Constraints Many Many Many Fewer Fewer Very less
From this comparative study, it can be recapitulated that RDFRFT provides a better solution for
computation of the DFRFT. The computation of RDFRFT requires more time than sampling type, linear
combination type and closed form type, however it satisfies more properties than any other method with
very less number of imposed constraints. Previously, IEEE has conferred to use the Eigenvectors
decomposition-type DFRFT, as the most appropriate definition of DFRFT. In the year 2006, Pei et al
[137] had established the superiority of their method (MPDFRFT) over the earlier reported methods.
Later on in the year 2009, the MPDFRFT was again modified by Pei et al [138] as RDFRFT. It was also
established that RDFRFT has one more degree of freedom, while satisfying the same set of properties as
satisfied by MPDFRFT with equal computational complexity. This makes the RDFRFT as a better choice
in security applications.
2.4 ELECTROMAGNETIC PROPAGATION
A signal changes its form from guided wave to unguided wave through the antenna in wireless
communication. Till date, the majority of reported work analyzed the nature of the signal propagating
[39]
through the channel in time and frequency domains only, whereas the travelling electromagnetic (EM)
wave which is actually propagating through the channel, have both time and spatial dependencies. Very
little work has been reported in which this aspect of EM wave is being analyzed, but the time dependency
of frequency and spatial dependency of propagation constant associated with the EM wave have not been
included, which will affect the performance of the channel in practical conditions.
2.4.1 Dual-Domain Transform Based on FT
The EM wave has been analyzed with the help dual-domain transform (DDT-FT) [36], by
transforming the travelling wave expression twice, first from time domain to frequency domain and
second in spatial coordinate domain to propagation constant domain, by considering transform method as
Fourier transform in both cases. Although, the application of FRFT was investigated in the area of
electromagnetic wave propagation [52, 53, and 110] but, the time dependency of frequency components
and spatial variable dependency of propagation constant has not been considered.
2.5 RADAR SIGNAL PROCESSING
Cross-ambiguity function is used to estimate the delay and Doppler parameters for a particular
target. The Fourier transform ambiguity function is normally used in pulse-Doppler radar for detecting
targets despite the phenomenon known as Doppler smearing which limits the performance of this method
[14]. Many techniques have been introduced for the analysis of linear frequency modulated (LFM) signals
for delay and Doppler estimation of target [15, 47, 148]. Subsequently, the ambiguity function is defined
for both narrowband and wideband, where the discrimination between narrowband and wideband is done
on the basis of comparison of the received signal bandwidth to its center frequency [18, 76]. The FRFT
has been used extensively in the analysis of LFM signals for mainly Doppler estimation of moving targets
[45, 111, and 115]. If only the delay and Doppler estimation is calculated by all these methods then the
computation is more complex as either it requires the computation of cross-ambiguity function of
transmitted and received echo or by superimposing some other technique with FRFT.
2.6 APPLICATION OF FRFT IN 1-D SIGNAL PROCESSING
The FRFT has found numerous applications in signal processing and image processing. Signal
processing areas includes - filtering, de-noising, interference suppression, radar signal processing,
[40]
electromagnetic wave propagation, orthogonal frequency division multiplexing systems and wireless
communication systems. The utility of FRFT can be best explained in the processing of non-stationary
signals. Therefore, filtering and beamforming application of FRFT is considered here.
2.6.1 Filtering
To pass the signal or attenuating the noise, filtering is a useful signal processing technique. The
application areas of filtering are - signal restoration, signal reconstruction and signal synthesis. These
processes are different in nature, but all requires filtering of an incoming signal to produce the desired
signal. Both time domain filtering and frequency domain filtering are utilized and superiority of one
above other depends on the computational complexity required under different input conditions [30]. For
non-stationary signals and noise, the time and frequency domain filtering both fails because the signal and
noise may have their respective Wigner distribution overlapping to each other in time and frequency
domains. In this case, fractional Fourier based filtering can provide a better solution, where for a rotated
domain corresponding to an optimum value of angle parameter in time-frequency plane, the Wigner
distribution of signal and noise may be separated and filtering of the signal from the noise can be
performed by designing a filter in the FRFT domain corresponding to this optimum angle. Numerous
works has been carried in past on the filtering in the FRFT domain [49, 50, 78, 86, 90, 91, and 145]. An
adaptive filtering scheme in fractional Fourier domain was introduced in [82], which had shown that
fractional Fourier domain adaptive filtering scheme provides less error in the case of linear frequency
modulated (LFM) signals.
2.6.2 Beam-forming
Beamforming is a signal processing technique used to control the directionality of the reception
or transmission of a signal on an array of sensors [19]. Beamforming is widely used for signal
enhancement, interference suppression, capturing signals of interest and direction of arrival [48, 141]. The
filtering weights at the sensors are chosen to achieve a certain goal. In literature, a number of beam-
forming algorithms are accessible like - linearly constrained minimum variance (LCMV) beamformer
[106], Capon beamformer [63, 157], and maximum signal to noise ratio (SNR) beamformer [119]. These
methods are widely applied in the fields of communication [62], radar or sonar target detection and
classification [60], and biomedical imaging [149], etc. In array signal processing, due to the need of
[41]
analyzing chirp signals, a FT beamformer should be replaced by FRFT based beamformer. Yetik et al
[61] had documented a minimum mean square error beamformer based on FRFT for additive white
Gaussian noise channel, and established that the FRFT domain beamformer gave better results than
spatial domain and frequency domain beamformers for all the three cases - stationary source, source
moving with constant velocity and accelerated source.. Subsequently, this work was explored for more
practical case i.e., for Rayleigh faded channel by Khanna et al [118], where the superiority of FRFT based
beamformer was shown over spatial and frequency domain beam-formers. The optimum FRFT order for
beamforming was obtained in [61, 116, 117, and 118], by comparing the respective covariance matrices
of the delayed signals. While in [93], the optimum order was decided by maximizing the second-order
central FRFT moment. This makes the desired chirp signal substantially concentrated whereas the noise
rejected considerably.
2.7 APPLICATION OF FRFT IN 2-D IMAGE PROCESSING
Similarly, FRFT has an extensive application in the image processing areas, which includes-
watermarking, compression and encryption. The angle parameter associated with FRFT empowers it to
provide one more key, making FRFT as a better tool in security areas.
2.7.1 Encryption
Encryption is the process of transforming information (plaintext) using an algorithm (cipher) to
produce a unreadable data to anyone except those possessing special knowledge, usually referred to as a
key. The result of the process is encrypted information. Numerous works has been carried out in the
literature for the encryption of image using fractional Fourier transform (FRFT). Alok et al [11] had
proposed a technique in which the Jigsaw transform was utilized with FRFT for encryption. Phase
retrieval method and phase masking were employed with FRFT to encrypt the image [20, 70].
Subsequently, in [37, 96], the authors had proposed a method for twin color images. For triple image
encryption a method was documented by Liu et al [160]. Motivated by development of the MPDFRFT, it
had been tried by researchers in encrypting the image [69, 72, and 102]. FRFT was used along with
wavelet transform by Chen et al [79] for encryption. Here, first the dual fractional Fourier-wavelet
domain was calculated by discrete fractional Fourier transform and wavelet decomposition then different
random phases was used in different wavelet sub-bands for encryption. Similarly, the FRFT with radial
Hilbert transform (RHT) was used by Joshi et al [95]. In this method, the RHT and image subtraction
[42]
techniques were first used to segregate image into two parts and each part was encrypted independently
using double random phase encoding in the FRFT domain.
2.7.2 Watermarking
Watermarking is the process of embedding information into a digital signal in a way that it
becomes difficult to remove. It is a process by which multimedia data is protected [16, 44, 58, 59, and 89]
against illegal recording, distribution and copyright protection. A watermark needs a transformation
domain for embedding it into digital media. This domain can be spatial domain [39, 101] as well as
frequency domain [68, 88, and 153]. In frequency domain techniques, watermark is embedding in the
spread of frequencies. Therefore, the signal energy distributed among signal frequencies (thus the
watermark) becomes undetectable. Djurovic et al. [57] introduced image watermarking in the fractional
Fourier domain, which is a combination of the spatial and frequency domains. The FRFT domain
watermarking maintains its advantages over spatial and frequency domain watermarking because it offers
two more degrees of freedom associated with the FRFT angles. Many other approaches of watermarking
in FRFT domain had been presented in literature [1, 40, 46, 87, 99, 104, 143, and 150]. A normal
distributed random sequence was used to modify the transform coefficients in the middle range of
magnitude of source image in [1]. Naveen [99] had used an encryption technique for modifying the
watermark image in FRFT domain before embedding it in the host image. A bio-inspired algorithm e.g.
ant colony optimization had also been utilized for watermarking in [46], whereas [87] included Fourier
coefficients for moment-based image analysis and the fractional Fourier transformation for watermark
embedding.
2.8 FINDINGS OF THE REVIEW
In this chapter, various time-frequency distribution tools, like – ambiguity function, Wigner
distribution, short-time Fourier transform and fractional Fourier transform (FRFT), has been analyzed and
it has been found that FRFT serves as a better tool as it has no cross-term. The mathematical expression
of FRFT and its importance in time-frequency plane is discussed extensively. Subsequently, various
properties of FRFT are presented along with their proof. In exploration of the properties of FRFT it has
been concluded that the definitions of convolution theorem, product theorem and correlation theorems for
the FRFT need to be examined more rigorously. Then, the discrete version of the FRFT (DFRFT) is
studied and it is observed that there are many methods of evaluating the DFRFT are available in the
[43]
literature. A comparison of all the methods has been performed by dividing them into six categories. And
it has been established that random type DFRFT (RDFRFT) is a better option for evaluating DFRFT.
RDFRFT is based on the eigenvalue decomposition method but it is different to other methods in terms of
choice of eigenvectors and eigenvalues, as in this method a random set of DFT eigenvectors and
eigenvalues is used. Due to its random nature this algorithm is advantageous in the security applications,
e.g. encryption watermarking, cryptography, steganography etc.
Also various application areas of the FRFT and its related properties are investigated. In this
journey, it is found that FRFT has a sound application in electromagnetic (EM) wave propagation. In
wireless communication the EM wave has to undergo a medium which has different portions having
different permittivity and permeability values in turn making the propagation constant associated with the
EM wave as distance (spatial coordinate) dependent also due to random motion of transmitter and/or
receiver a frequency becomes time dependent.
Second application area of the FRFT is investigated in the field of radar communication. The
motivation behind this application is straight forward, as in radar communication; the signal being
analyzed to extract information is actually a chirp signal. And FRFT has great strength in localizing the
chirp signals in time-frequency plane.
In the same manner, noise corrupted non-stationary signals can be filtered precisely using DFRFT
and FRFT based beamformer can be utilized for directing signal from an antenna array. In both the
application, FRFT found an advantage because of the same reason that it localizes chirp signal accurately
in time-frequency plane.
And finally, the utility of FRFT is searched in the area if image processing particularly,
encryption and watermarking. As FRFT offers one more degree of freedom, in terms of angle in time-
frequency plane at which the transform is calculated, so it can be used in security application because this
one degree of freedom serves as an extra key.
2.9 STATEMENT OF THE PROBLEM
The detailed studies on FRFT and its comparability with other integral transforms confirm that
the FRFT is not a completely or perfectly defined integral transform. For the completion of the domain of
FRFT and making it compatible with other integral transforms, efforts are made in defining the various
non-existent relationships as well as modifying the existent identities in a corrective manner. Such newly
[44]
introduced identities can be very well examined and tested with some new and existing applications to
confirm their perfectness and correctness in the domain of FRFT and integral transforms.
Therefore, following problems are undertaken in this course of study –
i. A convolution and product theorem for FRFT need to be defined - which converts into its FT
counterpart for angle 훼 = 휋 2⁄ , both side of the identity is dependent on corresponding variables
only and number of chirp multiplication is less as it introduces error in computation. Also this
theorem should satisfy various properties like; commutativity, associativity and distributivity that
is being satisfied by its FT counterpart.
ii. A cross - and auto - correlation theorem for FRFT is required because this theorem has extensive
application in power spectrum estimation. Then the defined identity for FRFT should satisfies the
basic properties that the correlation theorem of FT is satisfying and these required properties are –
commutativity, associativity, distributivity and evenness.
iii. A mathematical tool is needed which should be capable of analyzing the electromagnetic wave
propagation in a medium affecting the wave in making time dependent frequency components
and spatial dependent propagation constant.
iv. A method is required to estimate the range and Doppler shift (or velocity) associated with the
echo signal in radar communication which is easy and having less complexity.
v. Due to the advantageous role of FRFT in signal processing areas like – filtering and
beamforming, a FRFT domain filtering and beamforming need to be developed.
vi. FRFT offers one more degree of freedom in terms of angular parameter associated with it. For
this reason the role of FRFT in image security areas like- encryption and watermarking need to
investigate as FRFT provides one extra key.
The above mentioned problems are dealt with detailed analysis, appropriate applications and
simulation studies in the following Chapters.
*******************************************************