Introduction to the Fractional Fourier Transform and its Applications

Haldun M. Ozaktas and M. Alper Kutay

Bilkent University
Department of Electrical Engineering
TR-06533 Bilkent, Ankara, Turkey
David Mendlovic Tel-Aviv University, Faculty of Engineering,
69978 Tel-Aviv, Israel

Contents
1 Introduction
2 Notation and denitions
3 Fundamental properties
4 Common transform pairs
5 Eigenvalues and eigenfunctions
6 Operational properties
7 Relation to the Wigner distribution
8 Fractional Fourier Domains
9 Dierential equations
10 Hyperdierential form
11 Digital simulation of the transform
12 Applications to wave and beam propagation
12.1
12.2
12.3
12.4
12.5
12.6
12.7

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .
Quadratic-phase systems as fractional Fourier transforms .
Propagation in quadratic graded-index media . . . . . . . .
Fresnel diraction . . . . . . . . . . . . . . . . . . . . . . .
Fourier optical systems . . . . . . . . . . . . . . . . . . . . .
Optical implementation of the fractional Fourier transform
Gaussian beam propagation . . . . . . . . . . . . . . . . . .

13 Applications to signal and image processing

14 Acknowledgments

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

3
5
6
8
8
11
14
17
18
19
19
21
21
23
25
25
26
28
29

31
36

1 Introduction
The purpose of this chapter is to provide a self-complete introduction to the fractional Fourier transform for those who wish to obtain an understanding of the essentials without having to work through
the hundreds of papers which have appeared in the last few years. A general introduction will be
followed by the denition of the transform and a discussion of its fundamental and operational properties. Of central importance is the relationship of the transform to the Wigner distribution and other
phase-space distributions (also known as time-frequency or space-frequency representations). We will
concentrate on two main application areas which have so far received the most attention: wave and
beam propagation and signal processing.
The fractional Fourier transform is a generalization of the ordinary Fourier transform with an order
parameter a. Mathematically, the ath order fractional Fourier transform is the ath power of the Fourier
transform operator. The a = 1st order fractional transform is the ordinary Fourier transform. With
the development of the fractional Fourier transform and related concepts, we see that the ordinary
frequency domain is merely a special case of a continuum of fractional Fourier domains, and arrive at a
richer and more general theory of alternate signal representations, all of which are elegantly related to
phase-space distributions. Every property and application of the common Fourier transform becomes
a special case of that of the fractional transform. In every area in which Fourier transforms and
frequency domain concepts are used, there exists the potential for generalization and improvement by
using the fractional transform. For instance, the well-known result stating that the far-eld diraction
pattern of an aperture is in the form of the Fourier transform of the aperture can be generalized to
state that at closer distances, one observes the fractional Fourier transform of the aperture. The theory
of optimal Wiener ltering in the ordinary Fourier domain can be generalized to optimal ltering in
fractional domains, resulting in smaller mean-square errors at practically no additional cost.
In essence, the ath order fractional Fourier transform interpolates between a function f (u) and its
Fourier transform F (). The 0th order transform is simply the function itself, whereas the 1st order
transform is its Fourier transform. The 0:5th transform is something in between, such that the same
operation that takes us from the original function to its 0:5th transform will take us from its 0:5th
transform to its ordinary Fourier transform. More generally, index additivity is satised: The a2 th
transform of the a1 th transform is equal to the (a2 + a1 )th transform. The ?1th transform is the
inverse Fourier transform, and the ?ath transform is the inverse of the ath transform.
Scattered early papers related to the fractional Fourier transform include Wiener 1929, Condon
1937, Bargmann 1961, de Bruijn 1973. Of importance are two separate streams of mathematical papers
which appeared throughout the eighties (Namias 1980, McBride and Kerr 1987 and Mustard 1987a,
1989, 1991, 1996). However, the number of publications exploded only after the introduction of the
transform to the optics and signal processing communities (Lohmann 1993; Ozaktas and Mendlovic
1993a,b; Mendlovic and Ozaktas 1993; Ozaktas and others 1994a; Alieva and others 1994; Almeida
1994). Not all of these authors were aware of each other or building on the work of those preceding
them, nor is the transform always immediately recognizable in some of these works.
The fractional Fourier transform (or essentially equivalent transforms) appear in many contexts,
although it has not always been recognized as being the fractional power of the Fourier transform
and thus referred to as the fractional Fourier transform. For instance, the Green's function of the
quantum-mechanical harmonic oscillator is the kernel of the fractional Fourier transform. Also, the
fractional Fourier transform is a special case of the more general linear canonical transform (see Wolf
1979 for an introduction and references). This transform has been studied in many contexts, but again
3

the particular special case which is the fractional Fourier transform has usually not been recognized
as such.
The above does not represent a complete list of known historical references. For a more complete
list and also a more comprehensive treatment of the fractional Fourier transform and its relation to
phase-space distributions, we refer the reader to a forthcoming book on the subject by the authors
(Wiley, 1999). We expect further scattered historical references not known to us to be revealed in time.
Given the multitude of contexts in which essentially equivalent or closely related integral transforms
appear, it is probably not possible to attribute its invention to a particular set of authors. These many
contexts in which it was reinvented time after time in dierent guises is testimony to the elegance and
ubiquity of the transform.
Given the widespread use of the ordinary Fourier transform in science and engineering, it is important to recognize this integral transform as the fractional power of the Fourier transform. Indeed,
it has been this recognition which has inspired most of the many recent applications. Replacing the
ordinary Fourier transform with the fractional Fourier transform (which is more general and includes
the ordinary Fourier transform as its special case) adds an additional degree of freedom to the problem,
represented by the order parameter a. This in turn may allow either a more general formulation of the
problem (as in the diraction from an aperture example) or improvements based on the possibility of
optimizing over a (as in the optimal Wiener ltering example).
The fractional Fourier transform has been found to have several applications in the area known
as analog optical information processing, or Fourier optics. This transform allows a reformulation
of this area in a way much more general than that found in standard texts on the subject. It has
also led to generalizations of the notions of space (or time) and frequency domains, which are central
concepts in signal processing, leading to many applications in this area. More generally, the transform
may be expected to have an impact in the form of deeper understanding or new applications in every
area in which the Fourier transform plays a signicant role, and to take its place among the standard
mathematical tools of physics and engineering.
More specically, some applications which have already been investigated or suggested include
diraction theory (Alieva and others 1994; Gori and others 1994; Pellat-Finet 1994; Pellat-Finet
1995; Ozaktas and Mendlovic 1995; Abe and Sheridan 1995a; Alonso and Forbes 1997; Ozaktas
and Erden 1997), optical beam propagation and spherical mirror resonators (lasers) (Ozaktas and
Mendlovic 1994; Erden and Ozaktas 1997; Ozaktas and Erden 1997), propagation in graded index
media (Ozaktas and Mendlovic 1993a, b; Mendlovic and Ozaktas 1993; Mendlovic and others 1994a;
Alieva and Agullo-Lopez 1995, Abe and Sheridan 1995b; Gomez-Reino, Bao and Perez 1996), Fourier
optics (Bernardo and Soares 1994a, b; Pellat-Finet and Bonnet 1994; Ozaktas and Mendlovic 1995;
Ozaktas and Mendlovic 1996), statistical optics (Erden and others 1996a, b), optical systems design
(Dorsch 1995; Dorsch and Lohmann 1995; Lohmann 1995), quantum optics (Yurke and others 1990;
Aytur and Ozaktas 1995), radar, phase retrieval (Raymer and others 1994a, b; McAlister and others
1995), tomography (Beck and others 1993; Smithey and others 1993; Lohmann and Soer 1994; Wood
and Barry 1994a, b), signal detection, correlation, and pattern recognition (Mendlovic and others
1995d; Alieva and Agullo-Lopez 1995; Garca and others 1996; Lohmann and others 1996b; Bitran and
others 1996; Mendlovic and others 1995a; Mendlovic and others 1998), space- or time-variant ltering
(Ozaktas and others 1994a; Granieri and others 1995; Mendlovic and others 1996b; Ozaktas 1996;
Zalevsky and Mendlovic 1996; Mendlovic and others 1996b; Kutay and others 1997), signal recovery,
restoration, and enhancement (Lohmann and others 1996a; Erden and others 1997a, b; Ozaktas and
4

others 1997; Kutay and Ozaktas 1998; Kutay and others 1998a, b), multiplexing and data compression
(Ozaktas and others 1994a), study of space- or time-frequency distributions (Almeida 1994; Fonollosa
and Nikias 1994; Lohmann and Soer 1994; Ozaktas and others 1994a; Mendlovic and others 1995;
Dragoman 1996; Mendlovic and others 1996a; Ozaktas and others 1996a), and solution of dierential
equations (Namias 1980; McBride and Kerr 1987). We believe that these are only a fraction of the
possible applications. We hope that this chapter will make possible the discovery of new applications
by introducing the subject to new audiences.

2 Notation and denitions

The ath order fractional Fourier transform of the function f (u) will most often be denoted by fa (u) or
equivalently F a f (u). When there is possibility of confusion, we may more explicitly write F a [f (u)].
The transform is dened as a linear integral transform with kernel Ka (u; u0 ):
Z

fa(u) = F a [f (u)] = Ka (u; u0 )f (u0 ) du0 :

(1)

The kernel will be given explicitly below. All integrals are from minus to plus innity unless otherwise
stated. We prefer to use the same dummy variable u both for the original function in the space (or
time) domain, and its fractional Fourier transform. This is in contrast to the conventional practice
associated with the ordinary Fourier transform, where a dierent symbol, say , denotes the argument
of the Fourier transform F ():
Z

F () = f (u)e?i2u du;

(2)

f (u) = F ()ei2u d:

(3)

But these can be rewritten as

F (u) = f (u0 )e?i2uu0 du0 ;

(4)

f (u) = F (u0 )ei2u0 u du0 :

(5)

When it is desirable to distinguish the argument of the transformed function from that of the original function, we will let ua denote the argument of the ath order fractional Fourier transform:
fa(ua ) = (F a [f (u)])(ua ). With this convention, u0 corresponds to u, the space (or time) coordinate, u1 corresponds to the spatial (or temporal) frequency coordinate , and u2 = ?u0 , u3 = ?u1 .
Finally, we will agree to always interpret u as a dimensionless variable.
We will refer to F a [
], or simply F a , as the ath order fractional Fourier transform operator. This
operator transforms a function f (u) into its fractional Fourier transform fa (u). We will restrict
ourselves to the case where the order parameter a is a real number. f is a nite energy signal and
f (u) is a nite energy function which are well behaved in the sense usually presumed in physical
applications. In quantum mechanics f is the abstract state vector jf i and f (u) = hujf i is the urepresentation of f . Likewise, fa (u) = hua jf i is the ua -representation, which we will also refer to as the
representation of f in the ath order fractional Fourier domain. In thisRcontext jf (u)j2 is interpreted as
a probability distribution so that the energy of the function En[f ] = jf (u)j2 du = hf jf i corresponds
to its integrated probability and is thus equal to 1. In signal processing and optics, the energy can
5

take on any nite value but is conserved if attenuation or amplication mechanisms do not exist. (We
will also deal with sets of signals and functions whose energies are not nite (delta functions and
harmonic functions), but these will not correspond to physically realizable functions, but rather serve
as intermediaries in our formulations.)
We now dene the ath order fractional Fourier transform fa (u) through the following linear integral
transform:
Z

fa (u) = Ka (u; u0 )f (u0) du0 ;

(6)

Ka(u; u0 ) = A exp i(cot  u2 ? 2 csc  uu0 + cot  u02 ) ;

where

(7)
 = a
2;
p
A = 1 ? i cot  :
(8)
The square root is dened such that the argument of the result lies in the interval (?=2; =2]. The
kernel is not strictly dened when a is an even integer. However, it is possible to show that as a
approaches an even integer, the kernel behaves like a delta function under the integral sign. Thus,
consistent with the limiting behavior of the above kernel for values of a approaching even integers
(further discussed below), we dene K4j (u; u0 ) = (u ? u0 ) and K4j 2 (u; u0 ) = (u + u0 ), where j is an
arbitrary integer. Generally speaking, the fractional Fourier transform of f (u) exists under the same
conditions under which its Fourier transform exists (McBride and Kerr 1987, Almeida 1994).

3 Fundamental properties
We rst examine the case when a is equal to an integer j . We note that by denition F 4j and F 4j 2
correspond to the identity operator I and the parity operator P respectively (that is, f4j (u) = f (u)
and f4j 2(u) = f (?u)). For a = 1 we nd  = =2, A = 1, and

f1(u) =

1
exp(?i2uu0 )f (u0 ) du0 :
?1

(9)

We see that f1 (u) is equal to the ordinary Fourier transform of f (u), which was previously denoted
by the conventional upper case F (u). Likewise, it is possible to see that f?1 (u) is the ordinary inverse
Fourier transform of f (u). Our denition of the fractional Fourier transform is consistent with dening
integer powers of the Fourier transform through repeated application (that is, F 2 = FF , F 3 = FF 2 ,
and so on). Since  = a=2 appears in equation 6 only in the argument of trigonometric functions,
the denition is periodic in a (or ) with period 4 (or 2). Thus it is sucient to limit attention to
the interval a 2 [?2; 2). These facts can be restated in operator notation:

F0 = I;
F1 = F;
F2 = P;
F 3 = FP = PF ;
F4 = F0 = I;
F 4j+a = F 4j0 +a;
6

(10)
(11)
(12)
(13)
(14)
(15)

where j; j 0 are arbitrary integers.

Let us now examine the behavior of the kernel for small jaj > 0:
?isgn()=4
02
(16)
Ka(u; u0 ) = e p
jj exp[i(u ? u ) =]:
Now, using the well known limit
r
1 eiu2 =c ;
?
i=
4
(u) = clim
e
(17)
!0
c
the kernel is seen to approach (u ? u0 ) as a approaches 0. Thus dening the kernel Ka (u; u0 ) to
be precisely (u ? u0 ) at a = 0 maintains continuity of the transform with respect to a. A similar
discussion is possible when a approaches other integer multiples of 2. A more rigorous discussion of
continuity with respect to a may be found in McBride and Kerr (1987).
We now discuss the index additivity property:
F a1 F a2 f (u) = F a1 +a2 f (u) = F a2 F a1 f (u);
(18)
or in operator notation
F a1 F a2 = F a1 +a2 = F a2 F a1 :
(19)
This can be proved by repeated application of equation 6, and amounts to showing
Z
Ka2 (u; u00 )Ka1 (u00 ; u0 ) du00 = Ka1 +a2 (u; u0 )
(20)
by direct integration, which can be accomplished by using standard Gaussian integrals. We do not
present the details of this proof, since this property will follow much more simply from certain properties of the transform to be discussed.
The index additivity property is of central importance. Indeed, without it, we could hardly think
of F a as being the ath power of F (more will be said on this later). For instance, the 0:2nd fractional
Fourier transform of the 0:5th transform is the 0:7th Fourier transform. Repeated application leads to
statements such as, for instance, the 1:3th transform of the 2:1st transform of the 1:4th transform is the
4:8th transform (which is the same as the 0:8th transform). Transforms of dierent orders commute
with each other so that their order can be freely interchanged. From the index additivity property, we
deduce that the inverse of the ath order fractional Fourier transform operator (F a )?1 is simply equal
to the operator F ?a (because F ?a F a = I ). This can also be shown by directly demonstrating that
Z

Ka(u; u00 )K?a (u00 ; u0 ) du = (u ? u0 );

(21)

so that Ka?1 (u; u0 ) = K?a (u; u0 ). Thus we see that we can freely manipulate the order parameter a as
if it denoted a power of the Fourier transform operator F .
Fractional Fourier transforms constitute a one-parameter family of transforms. This family is a
subfamily of the more general family of linear canonical transforms which have three parameters (Wolf
1979). As all linear canonical transforms do, fractional Fourier transforms satisfy the associativity
property and they are unitary, as we can directly see by examining the kernel of the inverse transform
obtained by replacing a with ?a:
Ka?1 (u; u0 ) = K?a (u; u0 ) = Ka (u; u0 ) = Ka (u0 ; u):
(22)
The kernel Ka (u; u0 ) is symmetric and unitary, but not Hermitian. Unitarity implies that the fractional
Fourier transform can be interpreted as a transformation from one representation to another, and that
inner products and norms are not changed under the transformation.
7

f (u)
(u ? )
exp(i2u)
2
exp(iu )
2
exp[i (u + 2u)]
2
exp(?u )

fa (u)
2
A
exp[
i
(
u
cot

? 2u csc  + 2 cot )]

p1 + i tan
2
 exp[?i(u tan  ? 2u sec  + 2 tan )]
p
2
(1 + i tan )=(1 +  tan ) exp[iu ( ? tan )=(1 +  tan )]
p
2
2
(1 + i tan )=(1 +  tan ) expfi [u ( ? tan ) + 2u sec  ?  tan ]=[1 +  tan ]g
p
2
2
2
2
2
2
2
2
(1 ? i cot )=( ? i cot ) exp[iu cot ( ? 1)=( + cot )] exp[?u  csc =( + cot )]

Table 1: The functions on the right are the fractional Fourier transforms of the functions on the left.
j is an arbitrary integer, and  and  are real constants. For certain isolated values of a, the above
expressions should be interpreted in the limiting sense (equation 17). In the last pair,  > 0 is required
for convergence.

4 Common transform pairs

Table 4 gives the fractional Fourier transforms of a number of functions for which the integral appearing
in equation 6 can be evaluated analytically (often using standard Gaussian integrals). More will be
said on the fractional Fourier transforms of chirp functions exp[i(u2 + 2u)] after we discuss the
Wigner rotation property of the transform.
Greater insight can be obtained by considering some numerically obtained illustrations. Indeed, the
fractional Fourier transforms of many common functions do not have simple closed form expressions.
These may be obtained numerically using the algorithm discussed in section 11. We know that when
a = 0 we have the original function and when a = 1 we have its ordinary Fourier transform. As a
varies from 0 to 1, the transform evolves smoothly from the original function to the ordinary Fourier
transform. Figure 1 and 2 shows the evolution of the rect(u) function into the sinc(u)  (sin u)=(u)
function. Figure 3 shows the real parts of the fractional Fourier transforms of the Dirac delta function
(u ? 1). We note that for orders close to zero, the transform of the delta function is highly oscillatory,
and thus will approximately behave like the delta function under the integral sign, averaging out to
zero whatever function it happens to multiply.
p
Finally, we give the fractional Fourier transform of the quadratic-phase function f (u) = exp(?i=4) 1=rc exp(
with complex radius rc :
s
c tan  
 eiu2  1r?cr+tan
 ;
(23)
fa (u) = 11+?ri cot
cot 
c

provided =(rc)  0, which is also the condition for the original function f (u) to have nite energy.
From this result we conclude that the complex radius rc0 of the transformed function is
:
rc0 = 1r?c +r tan
(24)
c tan 
This result is useful in beam propagation problems since the original function f (u) represents a
Gaussian beam with complex radius rc .

5 Eigenvalues and eigenfunctions

The eigenvalues and eigenfunctions of the ordinary Fourier transform are well known (although seldom
discussed in introductory texts). They are the Hermite-Gaussian functions n (u), commonly known
as the eigensolutions of the harmonic oscillator in quantum mechanics, or the modes of propagation
of quadratic graded-index media in optics. The eigenvalues may be expressed as exp(?in=2) and are
8

a=0

a = 1/3

1.5

1.5

0.5

0.5

0
4

0
(a)

0
4

a = 2/3
1.5

0.5

0.5

0
(c)

a=1

1.5

0
4

0
(b)

0
4

0
(d)

Figure 1: Magnitudes of the fractional Fourier transforms of the rectangle function, I.

given by 1; ?i; ?1; i; 1; ?i; : : : for n = 0; 1; 2; 3; 4; 5; : : : . Thus the eigenvalue equation for the ordinary
Fourier transform may be written as

F n(u) = e?in=2 n(u);

(25)

where the Hermite-Gaussian functions are more explicitly given by

p
2
n (u) = An Hn ( 2 u)e?u ;
p
An = 21=4 = 2nn! ;

(26)
(27)

for n = 0; 1; 2; 3; 4; 5; : : : . Here Hn(u) are the Hermite polynomials. The particular scale factors which
appear in this equation are a direct consequence of the way we have dened the Fourier transform
with 2 in the exponent.
The ath order fractional Fourier transform shares the same eigenfunctions as the Fourier transform,
but its eigenvalues are the ath power of the eigenvalues of the ordinary Fourier transform:

F a n (u) = e?ian=2 n(u):

(28)

This result can be established directly from equation 6 by induction. First, we can show that 0 (u)
and 1 (u) are eigenfunctions with eigenvalues 1 and exp(?ia=2), by evaluating the resulting standard
complex exponential integrals. Then, by using standard recurrence relations for the Hermite-Gaussian
functions, it is possible to assume that the result-to-be-shown holds for n ? 1 and n, and show that
as a consequence it holds for n + 1. This completes the induction.
The above outlined demonstration of the fact that Hermite-Gaussian functions are eigenfunctions
of the fractional Fourier transform as dened by equation 6, reduces to the well known fact that
9

1.5
1
0.5
0
1
10
0.5
a

0
0 10

Figure 2: Magnitudes of the fractional Fourier transforms of the rectangle function, II.
Hermite-Gaussian functions are eigenfunctions of the ordinary Fourier transform when a = 1, since
equation 6 reduces to the denition of the ordinary Fourier transform and since e?ian=2 reduces to
e?in=2 when a = 1.
Readers familiar with functions FN (A) of an operator (or matrix) A with eigenvalues n will know
that in general FN (A) will have the same eigenfunctions as A and that its eigenvalues will be FN (n ).
The above eigenvalue equation is particularly satisfying in this light since F a as we have dened it, is
indeed seen to correspond to the ath power of the Fourier transform operator (FN (
) = (
)a ). However,
it should be noted that the denition of the ath power function is ambiguous, and our denition of the
fractional Fourier transform through equation 6 is associated with a particular way of resolving the
ambiguity associated with the ath power function (equation 28). Other denitions of the transform
also deserving to be called the fractional power of the Fourier transform are possible. The particular
denition we are considering is the one that has been most studied and that has led to the greatest
number of interesting applications. We are convinced it has a special place among other possible
denitions.
Knowledge of the complete set of eigenvalues and eigenfunctions of a linear operator is sucient to
completely characterize the operator. In fact, in some works, the fractional Fourier transform has been
dened through its eigenvalue equation (Namias 1980, Ozaktas and Mendlovic 1993a, b; Mendlovic
and Ozaktas 1993). To nd the fractional transform of a given function f (u) from knowledge of
the eigenfunctions and eigenvalues only, we rst expand the function as a linear superposition of the
eigenfunctions of the fractional Fourier transform (which are known to constitute a complete set):

f (u) =

1
X
n=0

Cn n(u)

10

(29)

a=0

a = 1/3
1.5

1.5

1
0.5

1
0
0.5

0.5

1
0

0
(a)

1.5

a = 2/3
1.5

0.5

0.5

0.5

0.5

1
2

0
(c)

a=1

1.5

1.5

0
(b)

1.5

0
(d)

Figure 3: Real parts of the fractional Fourier transforms of a delta function.

Z

Cn = n (u)f (u) du:

Applying F a on both sides of equation 29 and using equation 28 one obtains

F a f (u) =

1
X

n=0

e?ian=2 Cn n(u) =

Z X
1

n=0

e?ian=2 n (u) n (u0 )f (u0 ) du0 :

(30)
(31)

Upon comparison with equation 6, the kernel Ka (u; u0 ) is identied as

Ka (u; u0 ) =

1
X

n=0

e?ian=2 n (u) n (u0 ):

(32)

This is the spectral decomposition of the kernel of the fractional Fourier transform. The kernel given
in equation 32 can be shown to be identical to that given in equation 6 directly by using an identity
known as Mehler's formula:
"
1 ein
0 ei ? e2i (u2 + u02 ) #
X
1
2
uu
0
:
(33)
n p Hn(u)Hn (u ) = p p1 ? exp(2i) exp
1 ? e2i
n=0 2 n! 
Several properties of the fractional Fourier transform immediately follow from equation 28. In
particular the special cases a = 0, a = 1, and the index additivity property are deduced easily. (The
latter can be shown by applying F a0 to both sides of equation 28.)

6 Operational properties
Various operational properties of the transform are listed in table 2 (Namias 1980, McBride and Kerr
1987, Mendlovic and Ozaktas 1993, Almeida 1994). Most of these are most readily derived or veried
11

f (u)
fa (u)
f (?u)
f
a (?u)
p
?
1
?
2
f (ku)
jkj (1 ? i cot )=(1 ? ik cot ) exp[iu2 cot (1 ? cos2 0 = cos2 )] fa (u sin 0 =k sin )
2
f (u ? )
exp(i sin  cos ) exp(?i2u sin )fa (u ?  cos )
2
exp(i2u)f (u)
exp(?i sin  cos ) exp(i2u cos )fa (u ?  sin )
?1
uf (u)
cos  ufa (u) ? sin  (i2 ) dfa (u)=du
?
1
?1
(i2 ) df (u)=du
sin  ufa (u) + cos  (i2 ) dfa (u)=du
Ru
Ru
0
0
2
0
02
0
sec  exp(?iu tan )
u0 f (u ) du
u0 fa (u ) exp(iu tan ) du
0

Table 2: Operational properties of the fractional Fourier transform.  is an arbitrary real number, k
is a real number (k 6= 0; 1), and n is an integer. 0 = arctan(k2 tan ), where 0 is taken to be in
the same quadrant as .
by using equation 6 or the symmetry properties of the kernel.
Operations satisfying the rst property are referred to as even operations, so that the fractional
Fourier transform is an even operation. This property also implies
F a [Evf (u)] = Evfa(u);
(34)
a
F [Odf (u)] = Odfa(u);
(35)
which in turn imply that the transform of an even function is always even and the transform of
an odd function is always odd. Similar facts can be stated in operator form: All even operators,
and in particular the fractional Fourier transform operator commute with the parity operator P
(F a P = PF a ), and satisfy F a = PF a P . The eigenfunctions of even operations can always be chosen
to be of denite (even or odd) parity (the Hermite-Gaussian functions satisfy this property).
The second property is the generalization of the ordinary Fourier transform property stating that
the Fourier transform of f (ku) is jkj?1 F (=k). Notice that the fractional Fourier transform of f (ku)
cannot be expressed a scaled version of fa (u) for the same order a. Rather, the fractional Fourier
transform of f (ku) turns out to be a scaled and chirp-modulated version of fa0 (u) where a0 6= a is a
dierent order.
Now we turn our attention to the fth and sixth properties. The fractional Fourier transform
of uf (u) is equal to a linear combination of ufa (u) and dfa (u)=du. The coecients of this linear
combination are cos  and ? sin . When a = 1, this reduces to the corresponding ordinary Fourier
transform property. Similar comments apply to the fractional Fourier transform of df (u)=du. The
essence of these properties are most easily grasped if we express them in pure operator form. Let us
dene the coordinate multiplication operator U and dierentiation operator D through their eects in
the space domain:
U f (u) = uf (u);
(36)
(37)
Df (u) = i21 dfdu(u) :
These are simply dimensionless versions of the position and momentum operators of quantum mechanics and might have been written as
hujU f i = uhujf i;
(38)
d hujf i:
(39)
hujDf i = i21 du
12

We may dene in the same spirit, operators Ua and Da which have the same eect on fa (u), the ath
order fractional Fourier transform of f (u):

Ua fa(ua ) = (Ua f )a(ua ) = ua fa(ua );

a (ua ) ;
Dafa(ua ) = (Da f )a(ua ) = i21 dfdu
a

(40)
(41)

where we have explicitly written ua to avoid confusion. The eect of these operators is to coordinate
multiply and dierentiate the fractional Fourier transform of f (u), rather than f (u) itself. Now, with
these denitions, the fth and sixth properties of table 2 can be written as

F a [U f (u)] = cos  (Uaf )a (ua ) ? sin  (Da f )a(ua );

(42)
a
F [Df (u)] = sin  (Uaf )a (ua ) + cos  (Da f )a(ua ):
(43)
(Ua f )a (ua ) is simply the ua representation of Ua f , which we also refer to as the representation of Ua f
in the ath fractional Fourier domain. In the notation of quantum mechanics, (Ua f )a (ua ) would have
been written as hua jUa f i. Similar comments apply to (Da f )a (ua ). The two preceding equations can

be written in abstract operator form as

#
"
#
"
#"
U = cos  ? sin  Ua ;
(44)

= a
2;
Da
D
sin  cos 
We see that the coordinate multiplication and dierentiation operators corresponding to order a are
related to those in the ordinary space (or time) domain by a simple rotation matrix.
The commutator [U ; D]  UD ? DU is well known to be equal to i=2. By using equation 44 we
can easily derive the commutator (Aytur and Ozaktas 1995, Ozaktas and Aytur 1995)


[Ua ; Ua0 ] = 2i sin(0 ? ):
(45)
Knowing the commutator of two operators allows one to deduce an uncertainty relation between the
two representations associated with those operators. In particular, the above commutation relation
leads to (Aytur and Ozaktas 1995, Ozaktas and Aytur 1995)
2 (0 ? )
(46)
U2 a U2 a0  sin 16
2 :
Here Ua is the standard deviation of jfa (ua )j2 and Ua0 is the standard deviation of jfa0 (ua0 )j2 .
The translation and phase shift operators can also be expressed in operator notation. Let T ( )
denote the operator which takes f (u) to f (u ?  ) and let P ( ) denote the operator which takes
f (u) to exp(i2u)f (u), all in the ordinary space domain. We may also dene Ta () and Pa () as the
operators which have the same eect on the ath order fractional Fourier transforms: Ta ( ) takes fa (ua )
to fa (ua ?  ) and Pa ( ) takes fa (ua ) to exp(i2ua )fa (ua ). Then the third and fourth properties of
table 2 can be expressed as (Aytur and Ozaktas 1995, Ozaktas and Aytur 1995)

F a [T ()f (u)] = ei2 sin  cos [Pa (? sin )Ta ( cos )f ]a (ua )
= e?i2 sin  cos  [Ta ( cos )Pa (? sin )f ]a (ua );
F a [P ()f (u)] = e?i2 sin  cos [Pa ( cos )Ta( sin )f ]a (ua )
= ei2 sin  cos  [Ta ( sin )Pa ( cos )f ]a (ua ):
13

(47)
(48)
(49)

Here again the notation [Af ]a (ua ), where A is some operator, denotes the ua representation of Af ,
which would be written as hua jAf i in quantum mechanics. In operator form

T () = ei2 sin  cos Pa (? sin )Ta( cos ) = e?i2 sin  cos Ta( cos )Pa (? sin );
P () = e?i2 sin  cos Pa ( cos )Ta ( sin ) = ei2 sin  cos Ta( sin )Pa ( cos ):

(50)
(51)

We again see that the eect of translation is a combination of translation (by cos ) and phase multiplication (by sin ) of the fractional Fourier transform. A similar comment applies to the eect of
phase multiplication. When a = 1, these results reduce to the corresponding well known properties of
the ordinary Fourier transform.
The fractional Fourier transform does not have a convolution or multiplication property of comparable simplicity to that of the ordinary Fourier transform.

7 Relation to the Wigner distribution

The direct and simple relationship of the fractional Fourier transform to the Wigner distribution, as
well as to certain other phase-space distributions is perhaps its most important and elegant property
(Mustard 1989, 1996; Lohmann 1993; Almeida 1993, 1994; Mendlovic and others 1994a; Ozaktas and
others 1994a).
Here we will dene and brie y discuss some of the most important properties of the Wigner
distribution. The Wigner distribution Wf (u; ) of a function f (u) is dened as
Z

Wf (u; ) = f (u + u0 =2)f  (u ? u0 =2)e?2iu0 du0:

(52)

Wf (u; ) can also be expressed in terms of F (), or indeed as a function of any fractional transform
of f (u). Some of its most important properties are
Z

jf (u)j2 = W (u; ) d;

Z
2
jF ()j = W (u; ) du;

(54)

En[f (u)] = W (u; ) du d:

(55)

(53)

Roughly speaking, W (x;  ) can be interpreted as a function that indicates the distribution of the signal
energy over space and frequency. The Wigner distribution of F (u) (the Fourier transform of f (u)), is a
ninety degree rotated version of the Wigner distribution of f (u). More on the Wigner distribution and
other such distributions and representations may be found in Mecklenbrauker (1980a, b, c), Hlawatsch
and Boudreaux-Bartels (1992), Cohen (1989, 1995).
Now, if Wf (u; ) denotes the Wigner distribution of f (u), then the Wigner distribution of the ath
fractional Fourier transform of f (u), denoted by Wfa (u; ), is given by
(56)
Wfa (u; ) = Wf (u cos  ?  sin ; u sin  +  cos );
so that the Wigner distribution of Wfa (u; ) is obtained from Wf (u; ) by rotating it clockwise by
an angle . Let us dene R to be the operator which rotates a function of (u; ) by angle  in the
conventional counterclockwise direction. Then we can write
(57)
Wfa (u; ) = R? Wf (u; ):
14

This elegant and fundamental property underlies an important number of the applications of the
fractional Fourier transform. In fact, some authors have dened the transform as that operation
which corresponds to rotation of the Wigner distribution of a function (Lohmann 1993).
Equation 56 can be derived directly from equation 6 and the denition of the Wigner distribution
given by equation 52 (Ozaktas and others 1994a). The derivation is somewhat lengthy but straightforward. A similar derivation is given by Mustard (1989, 1996) and by Almeida (1994). Lohmann
(1993) shows the reverse, starting from the rotation property and arriving at equation 6.
An at least equally important form of this result follows easily (Mustard 1989, 1996; Lohmann
and Soer 1994; Ozaktas and others 1994a). Let us recall equations 53 and 54 which state that the
integral projection of Wf (u; ) onto the u axis is the magnitude square of the u-domain representation
of the signal and that the integral projection of Wf (u; ) onto the  axis is the magnitude square of
the -domain representation of the signal. Now, let us rewrite the rst of these equations for fa (u),
the ath order fractional Fourier transform of f (u):
Z

Wfa (u; ) d = jfa (u)j2 :

(58)

Since Wfa (u; ) is simply Wf (u; ) clockwise rotated by angle , the integral projection of Wfa (u; )
onto the u axis is identical to the integral projection of Wf (u; ) onto an axis making angle  with
the u axis. This new axis making angle  = a=2 with the u axis is referred to as the ua axis. Let
RAD denote the Radon transform operator, which maps a two-dimensional function of (u; ), to its
integral projection onto an axis making angle  with the u axis (Bracewell 1995). Thus the above can
be written as
RADWf (u; ) = jfa (u)j2 :
(59)
In conclusion, the integral projection of the Wigner distribution of a function onto the ua axis is
equal to the magnitude square of the ath order fractional Fourier transform of the function (gure 4).
Equations 53 and 54 are special cases with a = 0 and a = 1. Wood and Barry discussed what they
referred to as the \Radon-Wigner transform" without realizing its relation to the fractional Fourier
transform (Wood and Barry 1994a, b). The above discussion demonstrates that the Radon-Wigner
transform is simply the magnitude squared of the fractional Fourier transform.
The results of this section, and in particular equation 59, continue to hold when Wf (u; ) is the
Wigner distribution of a random process since the expectation value operation can move inside the
Radon transform and rotation operators:

Wfa (u; ) = R? Wf (u; );

RDN Wf (u; ) = jfa (u)j2 :

(60)
(61)

Here the overbars denote ensemble averages.

The Wigner distribution is not the only time-frequency representation satisfying the rotation property (equation 57). The ambiguity function also satises this property because the ambiguity function
is the two-dimensional Fourier transform of the Wigner distribution, and the two-dimensional Fourier
transform of the rotated version of a function is the rotated version of the two-dimensional Fourier
transform of the original function (Ozaktas and others 1994a, Almeida 1994). Almeida (1994) showed
that the rotation property also holds for the spectrogram. It has been further shown that the rotation
property generalizes to certain other time-frequency distributions belonging to the so-called Cohen
class, whose members can be obtained from the Wigner distribution by convolving it with a kernel
15

Wf ( u, )
u
ua

fa (u)

Figure 4: Oblique integral projections of the Wigner distribution.

characterizing that distribution. The distributions for which the rotation property holds are those
which have a rotationally symmetric kernel (Ozaktas and others 1996a).
Thus, fractional Fourier transformation corresponds to rotation of many phase-space representations. This not only conrms the important role this transform plays in the study of such representations, but also supports the notion of referring to the axis making angle  = a=2 with the u axis as the
ath fractional Fourier domain. Despite this generalization, the only distribution which satises a relation of the form of equation 59 is the Wigner distribution (Mustard 1989). Cohen (1989) argues that
there is nothing special about the Wigner distribution among other members of the Cohen class, since
all members of this class, including the Wigner distribution, are derivable from each other through
convolution relations. However, the fact that only the Wigner distribution satises equation 59 has
led Mustard to suggest that the Wigner distribution is a specially distinguished member of the Cohen
class (Mustard 1989, 1996).
As an instructive application of the Wigner rotation property, we discuss the fractional Fourier
transforms of chirp functions. The transform of the chirp function exp[i(u2 + 2u)] was given in
table 4 as a rather complicated expression. Phase space oers a much more transparent picture. The
Wigner distribution of the chirp function exp[i(u2 +2u)] is Wf (u; ) = ( ? u ?  ) which is simply
a line delta in phase space along the line  = u +  making angle arctan  with the u axis (gure 5).
The Wigner distribution of exp(i2u) is Wf (u; ) = ( ?  ) and is seen to be a special case in which
the line delta is horizontal. Similarly, the Wigner distribution of (u ?  ) is Wf (u; ) = (u ?  ) is
also seen to be a special case in which the line delta is vertical. (That the chirp indeed behaves like
a delta function as ;  ! 1 can be shown by using equation 17.) Thus, since harmonic functions
and delta functions can be considered degenerate or limiting cases of chirp functions, it is possible to
16

make the general statement that the fractional Fourier transform of a chirp function is always another
chirp function. In phase space, we observe this as the rotation of line deltas into other line deltas.

Figure 5: Wigner distribution of a chirp function.

8 Fractional Fourier Domains

Equations 57 and 59 immediately lead to the interpretation of oblique axes in phase space as fractional
Fourier domains. Just as the projection of the Wigner distribution onto the space domain gives
the magnitude square of the space-domain representation of the signal, and the projection of the
Wigner distribution onto the frequency domain gives the magnitude square of the frequency-domain
representation of the signal (equations 53 and 54), the projection on the axis making angle  = a=2
with the u axis gives the magnitude square of the ath fractional Fourier-domain representation of the
signal (equation 59). When we need to be explicit we will use the variable ua as the coordinate variable
in the ath domain, so that the representation of the signal f in the ath order fractional Fourier domain
will be written as fa (ua ). We immediately recognize that the 0th and 1st domains are the ordinary
space and frequency domains and that the 2nd and 3rd domains correspond to the negated space and
frequency domains (u0 = u, u1 = , u2 = ?u, u3 = ?). The representation of the signal in the
a0 th domain is related to its representation in the ath domain through an (a0 ? a)th order fractional
Fourier transformation:
Z
(62)
fa0 (ua0 ) = Ka0 ?a (ua0 ; ua )fa (ua ) dua :
When (a0 ? a) is an integer, this corresponds to a forward or inverse Fourier integral.
The notion of fractional Fourier domains as oblique axes in space-frequency is also conrmed by
the operational formula presented earlier. We had seen that a translation of u0 corresponded to
17

cos  much of the shift and a sin -much phase shift in the ath domain. Likewise, we had seen that
multiplication by u corresponded to cos  much of the multiplication and a sin -much dierentiation.

9 Dierential equations
Here we will see that the fractional Fourier transform fa (u) of a function f0 (u) is the solution of a
dierential equation, where f0 (u) may be interpreted as the initial condition of the equation. This
is the quantum-mechanical harmonic oscillator dierential equation and also the equation governing
optical propagation in quadratic-graded index media. (In the former case the order parameter a
corresponds to time and in the latter case it corresponds to the coordinate along the direction of
propagation.) (Agarwal and Simon 1994; Ozaktas and Mendlovic 1995) In fact, in some sources the
solution is written in the form of an integral transform whose kernel is sometimes referred to as the
harmonic oscillator Green's function, without the authors knowing that this is the fractional Fourier
transform. (To be precise, we must note that the dierential equations governing these physical
phenomena dier slightly from the equation we discuss, but this small dierence is inconsequential.)
The dierential equation is
"

a (u) = ? 1 @ 2 + u2 ? 1 f (u);

i 2 @f@a
4 @u2
2 a

(63)

with the initial condition f0 (u) = f (u). The solution fa (u) of the equation is the ath order fractional
Fourier transform of f (u), as can be shown by direct substitution of equation 6.
Alternatively, and more instructively, we may take an eigenvalue equation approach. Substituting
the form fa (u) = exp(?ia)f0 (u) in equation 63, we obtain

d2 f0(u) + 42  1 + 2 ? u2  f (u) = 0:

(64)
0
du2
2 2
Comparing this equation with the standard equation
d2 f (u) + 42  2n + 1 ? u2  f (u) = 0;
(65)
du2
2
whose solutions are well known as the Hermite-Gaussian functions n (u), we conclude that the nth
Hermite-Gaussian function is a solution of equation 64 when  = n = n=2. It is now possible
to write down arbitrary solutions of the equation as a linear superposition of these eigenfunctions
(modes). If the initial condition f0 (u) is n (u), the solution fa (u) is exp(?ian=2) n (u); this is what
it means to be an eigenfunction. Given an arbitrary initial condition f0 (u), we can expand it in terms
of the Hermite-Gaussian functions n (u) as
f0 (u) =
Cn =

1
X

Cn n (u);

(66)

n (u)f0 (u) du:

(67)

n=0

Since equation 63 is linear, the solution corresponding to this initial condition is readily obtained as

fa(u) =

1
X
n=0

Ane?ian=2 n(u);
18

(68)

from which one can obtain

fa (u) = Ka (u; u0 )f0 (u0 ) du0 ;

Ka (u; u0 ) =

1
X

n=0

(69)

e?ian=2 n(u) n (u0 );

exactly as in equation 32.

The dierential equation in question can also be written in terms of the operators U and D as
follows:
a (u) ;
(70)
(U 2 + D2 )fa(u) ? 12 fa(u) = i @@f
(a=2)
or
a (u) :
Hfa(u) = i @@f
(71)
(a=2)

Readers with a background in quantum mechanics will readily recognize the Hamiltonian H = (U 2 +
D2) ? 1=2 which characterizes the harmonic oscillator. This Hamiltonian is domain-invariant, which
means that (Ua2 + Da2 ) is the same operator regardless of the value of a, as can be readily shown
by using equation 44. This rotational invariance of the Hamiltonian ties in with the Wigner rotation
property. (The extra ?1=2 represents the inconsequential discrepancy mentioned earlier.)
More on the relationship of the fractional Fourier transform to dierential equations and their
solutions is found in Namias (1980) and McBride and Kerr (1987).

10 Hyperdierential form

The hyperdierential form of the fractional Fourier transform operator is given by (Namias 1980;
Mustard 1987a, c).

F a = e?i(a=2)H ;


H =  D2 + U 2 ? 21 :

Applied to a function f (u) we may write

"

(72)

!#


a
d2 + u2 ? 1 f (u):
?i 2 ? 41 du
(73)
2
2
This hyperdierential representation may be considered the formal solution of the dierential equation
written in the form of equation 71. The operator F a given in equation 72 generates fa (u) for all values
of a from f0(u) = f (u). The index additivity property and the special case a = 0 immediately follow
from the exponential form exp(?iH).

F a f (u) = exp

11 Digital simulation of the transform

Here we brie y discuss how the fractional Fourier transform may be computed on a digital computer,
referring the reader to Ozaktas and others (1996b) for further details.
The dening equation (equation 6) can be put in the form
Z

fa (u) = A ei cot u2 e?i2 csc uu0 ei cot u02 f (u0) du0 :
19

(74)

We assume that the representations fa (ua ) of the signal f in all fractional Fourier domains are approximately conned to the interval [?
u=2;
u=2] (that is, a suciently large percentage of the signal
energy is conned to these intervals). This assumption is equivalent to assuming that the Wigner distribution of f (u) is approximately conned within a circle of diameter
u (by virtue of equation 59).
Again, this means that a suciently large percentage of the energy of the signal is contained in that
circle. We can ensure that this assumption is valid for any signal by choosing
u suciently large.
Under this assumption, and initially limiting the order a to the interval 0:5  jaj  1:5, the modulated function eiu02 f (u0 ) may be assumed to be approximately band-limited to 
u in the frequency
domain. Thus eiu02 f (u0 ) can be represented by Shannon's interpolation formula

eiu02 f (u0 ) =

NX
?1






ei cot ( 2
nu )2 f 2
n u sinc 2
u u0 ? 2
n u ;
n=?N

(75)

where N = (
u)2 . The summation goes from ?N to N ? 1 since f (u0 ) is assumed to be zero outside
[?
u=2;
u=2]. By using equation 75 and equation 74, and changing the order of integration and
summation we obtain



 Z

?1
n )2
n
n
2 NX
0
i
cot

(
i
cot
u
0
?
i
2

csc
uu
2

u
f 2
u e
e
fa(u) = Ae
sinc 2
u u ? 2
u du0 : (76)
n=?N
By recognizing the integral to be equal to (1=2
u)e?i2 csc u 2
nu rect(csc u=2
u), we can write



NX
?1
A
2 ?i2 csc u n i cot ( n )2
n

i
cot
u
2

u
2

u
e
e
(77)
e
f 2
u ;
fa (u) = 2
u
n=?N
since rect(csc u=2
u) = 1 in the interval juj 
u=2. Then, the samples of fa (u) are given by
NX
?1 icot ( m )2 ?2 csc  mn +cot ( n )2   n 
A
m
2
u
2
u f
(2
u)2
e
fa( 2
u ) = 2
u
(78)
2
u ;
n=?N
which is a nite summation allowing us to obtain the samples of the fractional transform fa (u) in
terms of the samples of the original function f (u). Direct computation of equation 78 would require
O(N 2 ) operations. A fast (O(N log N )) algorithm can be obtained by putting equation 78 into the
following form:


?1
m?n )2 i(cot ?csc )( n )2
m )2 NX
A
n
m

i
csc

(
i
(cot

?
csc

)(
2

u
2

u
2

u
e
e
f 2
u : (79)
fa( 2
u ) = 2
u e
n=?N

We now recognize that the summation is the convolution of ei csc ( 2
nu )2 and the chirp modulated
function f (
). The convolution can be computed in O(N log N ) time by using the fast Fourier transform (FFT). The output samples are then obtained by a nal chirp modulation. Hence the overall
complexity is O(N log N ).
We had limited ourselves to 0:5  jaj  1:5 in deriving the above algorithm. Using the index
additivity property of the fractional Fourier transform we can extend this range to all values of a
easily. For instance, for the range 0  a  0:5, we can write
F a = F a?1+1 = F a?1 F 1 :
(80)
Since 0:5  ja ? 1j  1, we can use the above algorithm in conjunction with the ordinary Fourier
transform to compute fa (u). The overall complexity remains at O(N log N ).
20

12 Applications to wave and beam propagation

A considerable number of papers have been written on the application of the fractional Fourier transform to wave and beam propagation problems, mostly in an optical context. Our presentation will
also be phrased in the notation and terminology of optics. Nevertheless, the reader should have no
diculty translating the results to other propagation, diraction, and scattering phenomena which
are mathematically equivalent or similar.
Whenever we can express the result of an optical problem (such as Fraunhofer diraction) in terms
of a Fourier transform, we tend to think of this as a simple and elegant result. This is justied by the
fact that the Fourier transform has many simple and useful properties which make it attractive to work
with. The Fourier transform and image occur at certain privileged planes in an optical system. Often
all our intuition about what happens in between these planes is that the amplitude distribution is
given by a complicated integral. We will see below that the distribution of light at intermediate planes
can be expressed in terms of the fractional Fourier transform (which also has several useful properties
and operational formulas). Thus the fractional Fourier transform completes in a very natural way the
study of optical systems often called \Fourier optics."
Fourier optical systems can be analyzed using geometrical optics, Fresnel integrals (spherical wave
expansions), plane wave expansions, Hermite-Gaussian beam expansions, and as we will discuss, fractional Fourier transforms. The several approaches prove useful in dierent situations and provide
dierent viewpoints which complement each other. The fractional Fourier transform approach is appealing in that it describes the continuous evolution of the wave as it propagates through the system.

12.1 Introduction
Optical systems involving an arbitrary sequence of thin lenses separated by arbitrary sections of free
space (under the Fresnel approximation) belong to the class of quadratic-phase systems. Mathematically, quadratic-phase systems are equivalent to linear canonical transforms (Wolf 1979). Systems
containing arbitrary sections of quadratic graded-index media also belong to this class. The class
of Fourier optical systems (or rst-order optical systems) consist of arbitrary thin lters sandwiched
between arbitrary quadratic-phase systems. Members of the class of quadratic-phase systems are characterized by linear transformations of the form (Bastiaans 1978, 1979a, 1979b, 1989, 1991; Nazarathy
and Shamir 1982; Ozaktas and Mendlovic 1995)
Z

fout (u) = h(u; u0 )fin (u0 ) du0 ;

h(u; u0 ) = K 0 exp[i(u2 ? 2uu0 + u02 )];

(81)

where K 0 is a complex constant, and ,  , and are real constants. Comparing this equation with
equation 6, we see that fractional Fourier transforms are a special case of quadratic-phase systems.
Until this point, all variables have been considered to be dimensionless and were denoted by u,
, etc. and all functions and kernels took dimensionless arguments and were denoted by f (u), g(u),
h(u; u0 ), etc. In optical applications, we will often employ variables with the dimensions of length or
inverse length which we will denote by x, p etc. Functions and kernels taking such arguments will be
distinguished as f^(u), g^(u), h^ (u; u0 ), etc. With these conventions, equation 81 can be rewritten as
Z

f^out (x) = h^ (x; x0 )f^in(x0) dx0 ;

21



^h(x; x0 ) = K exp i2 (x2 ? 2xx0 + x02 ) ;
s

(82)

where K = K 0 =s, s is a constant with the dimension of length, and f^out (x)  fout (x=s), etc. The
choice of s essentially corresponds to the choice of units. We will assume s is specied once and for
all throughout our analysis.
The kernels associated with a thin lens with focal length f , and free-space propagation over a
distance d are given respectively by (Saleh and Teich 1991)
h^ lens (x; x0 ) = Klens (x ? x0 ) exp(?ix2 =f );
(83)
0
0
2
h^ space (x; x ) = Kspace exp[i(x ? x ) =d]:
(84)

 is the wavelength of light in free space and Klens and Kspace are constants. We continue to work with
one-dimensional notation for simplicity, although most optical systems are two dimensional. These
kernels are special cases of the kernel given in equation 81. It is possible to prove that any arbitrary
concatenation of kernels of these form will result in a kernel of the form given in equation 82.
Apart from the constant factor K which has no eect on the resulting spatial distribution, a member
of the class of quadratic-phase systems is completely specied by the three parameters ,  , and
(equation 81). Alternatively, such a system can also be completely specied by the transformation
matrix (Bastiaans 1989; Nazarathy and Shamir 1982; Ozaktas and Mendlovic 1995)
"

"

s2 = ;
T  CA DB  ?=s2 +  =s =
2  =

(85)

with AD ? BC = 1. Here again s is the scale factor relating our dimensional and dimensionless variables. If several systems each characterized by such a matrix are cascaded, the matrix characterizing
the overall system can be found by multiplying the matrices of the several systems. The matrix dened
above also corresponds to the well known ray matrix employed in ray optical analysis. At a certain
plane perpendicular to the optical axis, a ray can be characterized by its distance from the optical
axis x and its paraxial angle of inclination . We will dene the ray vector as [x p]T where p  =.
Then, the ray vector at the output is related to the ray vector at the input by
"

"

xout = A B
pout
C D

#"

xin :
pin

(86)

The matrices corresponding to a thin lens and a section of free space are given respectively by
"

and

"

A B =
1 0 ;
C D
?1=f 1
"

"

(87)

A B = 1 d :
(88)
C D
0 1
Quadratic graded-index media exhibit a parabolic refractive index prole n(x) about the optical axis,
characterized by the two parameters n0 and  as follows:
n2 (x) = n20 (1 ? (x=)2 ):
22

(89)

The matrix corresponding to quadratic graded-index media is given by

"

"

A B =
cos[(d=d0 )=2] (=n0 ) sin[(d=d0 )=2] ;
?
1
C D
?(=n0 ) sin[(d=d0 )=2]
cos[(d=d0 )=2]

(90)

where d is the length of the medium and d0 = =2. This matrix can be derived by a simple application
of the ray equation (Saleh and Teich 1991).
The space spanned by the coordinates x and p also constitutes a phase space which directly
corresponds to the space-frequency plane on which the Wigner distribution was dened. A particular
ray characterized by its phase space vector [x p]T will be mapped to another according to its ABCD
matrix given above. If we consider a bundle of rays constituting a region in this phase space, this
region will likewise be transformed according to the same matrix. For instance, let us consider the
rectangular bundle consisting of rays whose intercepts lie between ?x0 and x0 and whose inclinations
lie between ?p0 = ?0  and p0 = 0  (gure 6a). If this bundle passes through a lens, it will be
transformed according to equation 87 into the bundle shown in part b. If this bundle passes through a
section of free space, it will be transformed according to equation 88 into the bundle shown in part c.
More generally, for arbitrary ABCD, it will be transformed into a bundle of the general form shown
in part d.
This mapping of phase space regions can also be posed in terms of the Wigner distribution. It is
known that if f^out is the linear canonical transform of f^in with parameters ABCD, then the Wigner
distribution of f^out is related to that of f^in by the relation (Bastiaans 1979b)
W^ out (xout ; pout ) = W^ in (xin ; pin ):
(91)
where xout ; pout and xin ; pin are related according to equation 86. This is simply a generalization of the
Wigner rotation property discussed in section 7. In particular, since we know from this property that
fractional Fourier transformation corresponds to rotation in phase space, it follows that the ABCD
matrix for fractional Fourier transforms should be the rotation matrix.

12.2 Quadratic-phase systems as fractional Fourier transforms

As is evident by comparing equations 6 and 81, the one-parameter class of fractional Fourier transforms
is a subclass of the class of three-parameter quadratic-phase systems. If we allow an additional
magnication parameter M and a phase curvature parameter 1=R, the family of fractional Fourier
transforms will now also have three-parameters and can be put in one-to-one correspondence with the
family of quadratic-phase systems. The kernel of this three-parameter transform may be written as
Z

f^out (x) = h^ (x; x0 )f^in(x0 ) dx0

!#
"
2
0
x
xx
i
0
2
0
2
h^ (x; x ) = K exp(ix =R) exp s2 M 2 cot  ? 2 M csc  + x cot  ;

(92)

which is in the form of a quadratic-phase system. (The pure mathematical form given by equation 6 is recovered by setting x=s = u, M = 1, R = 1.) This kernel maps a function f (x=s) into
K 0 exp(ix2 =R)fa (x=sM ), where fa(u) is the ath order fractional Fourier transform of f (u). Here
 = a=2 as before and M > 0 is referred to as the magnication associated with the transform and
R is the radius of the spherical surface on which the perfect fractional Fourier transform is observed.
23

po +xo / f

po

xo

xo

-xo

-xo
-po

-po -xo / f

(a)

(b)

po
-xo-po f
xo+ pof
-po

(d)

(c)

Figure 6: Eect of quadratic-phase systems in phase space.

When R = 1, the quadratic-phase term disappears and the perfect fractional Fourier transform is
observed on a planar surface.
The above family of kernels is in one-to-one correspondence with the family of kernels given in
equation 81. The parameters ,  , and are recognized to be related to the parameters , M , and R
through the relations
 = cot =M 2 + s2 =R;
(93)
 = csc =M;
(94)
= cot :
(95)
Alternatively, the ABCD parameters are related to , M , and R through the relations
"
#
"
#
A B =
M cos 
s2 M sin  ;
(96)
C D
? sin =s2 M + M cos =R cos =M + s2M sin =R
which can be inverted to yield
(97)
tan  = s12 B
A;
q
(98)
M = A2 + (B=s2 )2 ;
B=A
1 = 1
C:
+
(99)
4
2
2
2
R s A + (B=s ) A
24

The above result essentially means that any quadratic-phase system can be interpreted as a magnied
fractional Fourier transform, perhaps with a residual phase curvature. Since a relatively large class
of optical systems can be modeled as quadratic-phase systems, these systems can also be interpreted
as fractional Fourier transforms (Ozaktas and Mendlovic 1996). We will rst consider two elementary
examples: propagation in quadratic graded-index media and diraction in free space, and then treat
the more general case of arbitrary composition of thin lenses and sections of free space.

12.3 Propagation in quadratic graded-index media

Quadratic graded-index media have a natural and direct relationship with the fractional Fourier transform. Light is simply fractional Fourier transformed as it propagates through such media. The refractive index distribution for such media was already given in equation 89.
The ABCD matrix for graded-index media was given as equation 90. Comparing this with equation 96, we immediately conclude that propagation through a section of graded-index media results
in a fractional Fourier transform of order a = 2d=  d=d0 , provided the scale parameter s is chosen
such that s2 = =n0 . Agreeing on this choice of s, there is no magnication (M = 1) and no residual
phase curvature (R = 1). Recalling the comment at the end of subsection 12.1, we conclude that as
light propagates through quadratic graded-index media, its Wigner distribution rotates. Quadratic
graded-index media realize fractional Fourier transforms in their purest and simplest form. If the
distribution of light at the input plane is given by f (x=s), then the distribution of light at the output plane is proportional to fa (x=s), where the transform order a increases linearly with distance of
propagation.
The same result can be arrived at by starting from the Helmholtz equation for quadratic gradedindex media, nding its modes (which are the Hermite-Gaussian functions), and constructing arbitrary
solutions as linear superpositions of these modes. This approach may be found in Mendlovic and
Ozaktas (1993) and Ozaktas and Mendlovic (1993a, b).

12.4 Fresnel diraction

Although quadratic graded-index media is perfectly matched to the fractional Fourier transform, it
is of interest to discuss the more basic problem of diraction from a planar screen with complex
amplitude transmittance t^(x). The complex amplitude distribution g^(x) of light in a diraction plane
at distance d is given by the Fresnel integral (equation 84):
Z

f^out (x) = h^ (x; x0 )f^in(x0 ) dx0

h^ space (x; x0 ) = Kspace exp[i(x ? x0 )2 =d];

(100)

assuming illumination of the screen by a uniform plane wave. Now, with t(x=s)  t^(x), it is possible to
cast this
integral in the form of the integral of equation 92 by identifying  = a=2 = arctan(d=s2 ),
p
M = 1 + (d=s)2 , R = (s4 + 2 d2 )=d. (The same results can be arrived at by comparing
equation 96 with equation 88, or by specializing equations 97, 98, and 99 to A = D = 1, B = d,
C = 0.)
At a distance d from the diracting object or aperture, we observe the ath order fractional Fourier
transform of the object on a spherical reference surface with radius R. The transform is magnied
by M . As d is increased from 0 to 1, the order a of the fractional transform increases according to
a = (2=) arctan(d=s2 ) from 0 to 1 (gure 7). Letting d ! 1, we obtain a = 1, M = d=s2 / d,
25

and R = d, which we readily associate with the Fraunhofer diraction pattern, which is nothing but
the Fourier transform of the diracting screen. Note that in this limit, the magnication and radius
of curvature are both proportional to the distance d.
1
0.9
0.8
0.7
0.6

0.5
0.4
0.3
0.2
0.1
0
0

10

12

14

16

18

20

d/s

Figure 7: a = (2=) arctan(d=s2 ) as a function of d=s2 .

Thus we see that the propagation of light along the +z direction can be viewed as a process
of continual fractional Fourier transformation. As light propagates, its distribution evolves through
fractional transforms of increasing orders. The fact that the far-eld diraction pattern is the Fourier
transform of the diracting object is one of the central results of diraction theory. We have shown
that the eld at closer distances are given by fractional Fourier transforms of the diracting object.
More generally, there exists a fractional Fourier transform relation between the amplitude distribution of light on two spherical reference surfaces of given radii and separation (Ozaktas and Mendlovic
1994, 1995). It is possible to determine the order and scale parameters associated with this fractional
transform given the radii and separation of the surfaces. Alternatively, given the desired order and
scale parameters, it is possible to determine the necessary radii and separation.

12.5 Fourier optical systems

We will now see that any quadratic-phase system can always be interpreted as a fractional Fourier
transforming system. Thus the fractional Fourier transform can describe all systems composed of an
arbitrary number of lenses separated by arbitrary distances, whereas imaging and Fourier transforming
systems are only special cases.
All of the essential ingredients have already been provided and it merely remains to state the
result. Any quadratic-phase system (equation 82) can be characterized by its ABCD parameters. As
discussed earlier, the kernel in equation 82 can be cast in the form of the kernel in equation 92 by
identifying  = a=2, M , and R according to equations 97, 98, and 99. Since any quadratic-phase
system can be interpreted as a fractional Fourier transform, and since any optical system consisting
26

of an arbitrary concatenation of lenses and sections of free space can be modeled as a quadratic-phase
system, it follows that such an optical system can also be interpreted in terms of the fractional Fourier
transform.

0.2 m

(a.)

0.16 m

0.2 m

0.28 m

0.4 m

0.2 m

2
1
0
1
2
0

(b.)

0.2

0.4

0.6

0.8

1
z (m)

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

1
z (m)

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

1
z (m)

1.2

1.4

1.6

1.8

0.2

0.4

0.6

0.8

1
z (m)

1.2

1.4

1.6

1.8

3
2
1
0
0

(c.)

3
2.5

2
1.5
1
0.5
0
0
(d.)

1 / R (1/m)

5
0

Figure 8: Evolution of a(z ), M (z ), 1=R(z ) as functions of z (Ozaktas and Erden 1997).

A concrete example will be useful. Figure 8a shows a system consisting of several lenses, whose
focal lengths have been indicated in meters. The input plane is taken as z = 0. The output plane is
variable, ranging from z = 0 to z = 2 m. Two rays have been drawn through the system. Let A(z ),
B (z), C (z), D(z) denote the ABCD parameters of the section of the system occupying the interval
[0; z ], which can be readily calculated using the matrices for lenses and sections of free space and the
concatenation property. Also let [x(z ) p(z )]T denote the ray vector at z . Then,
"
#
"
#"
#
x(z) = A(z) B (z) x(0) :
(101)
p(z)
C (z) D(z)
p(0)
27

We further let (z ) = a(z )=2, M (z ), R(z ) represent the order, magnication, and phase curvature of
the fractional Fourier transform observed at z . These can be determined again by using equations 97,
98, and 99.
The fractional transform order a(z ), the scale parameter M (z ), and the radius R(z ) of the spherical
surface on which the perfect transform is observed, are plotted as functions of z in gure 8 (Ozaktas
and Erden 1997). Letting j denote an arbitrary integer, when a = 4j we observe an erect image, when
a = 4j + 2 we observe an inverted image, when a = 4j + 1 we observe the common Fourier transform,
and when a = 4j ? 1 we observe an inverted Fourier transform (which is the same as an inverse Fourier
transform). The reader should study the behavior of the two rays in conjunction with the graphs in
gure 8. At z = 0:4 we obtain a conventional Fourier transform (a = 1) as a result of the conventional
2f system occupying the interval [0; 0:4]. An inverted image (a = 2) is observed at z  0:65. We see
that M < 1 and R > 0, as conrmed by an examination of the rays. (The ray represented by the solid
line crosses the z = 0:65 plane at a negative value (implying an inverted image) smaller than unity
in magnitude (implying M < 1), with a slope indicating divergence (implying R > 0).) An inverted
Fourier transform (a = 3) is observed at z  1:2, almost coincident with the lens at that location.
An erect image (a = 4) is observed at z  1:4, immediately after the lens at that location. The
eld curvature 1=R of this image has a very small negative value and the magnication M is slightly
smaller than 1. The imaging systems discussed in Bernardo and Soares (1994b) provide additional
useful examples which the reader may wish to study in a similar manner. (Ozaktas and Erden 1997)
Fourier optical systems consist of an arbitrary number of thin lters sandwiched between arbitrary quadratic-phase systems. It readily follows that any Fourier optical system can be modeled as
lters sandwiched between fractional Fourier transform stages, or as repeated ltering in consecutive
fractional Fourier domains (section 13). (Ozaktas and Mendlovic 1996)

12.6 Optical implementation of the fractional Fourier transform

Here we mention a number of systems which map f^in(x) = f (x=s) into f^out (x) / fa (x=s).
Conceptually simplest is to use a section of quadratic graded-index media of length d = (a=2) =
ad0 with s2 = =n0 (subsection 12.3).
In practice, systems consisting of bulk lenses may be preferred. Two such systems were rst
presented by Lohmann (1993). We present these systems without derivation, referring the reader to
Lohmann (1993) and Ozaktas and Mendlovic (1995) for details. The rst system consists of a section
of free-space of length d followed by a lens of focal length f followed by a second section of free-space
of length d. To obtain an ath order fractional Fourier transform with scale parameter s, we must
choose d and f according to
2
(102)
d = s tan(=2);
2
(103)
f = s csc :

The second systems consists of a lens of focal length f followed by a section of free-space of length d
followed by a second lens of focal length f . This time d and f must be chosen according to
2
(104)
d = s sin ;
2
f = s cot(=2):
(105)
28

More general systems can easily be obtained by using the general formulation presented in Ozaktas
and Mendlovic (1995), Mendlovic and others (1995b, c), Jiang (1995), Liu and others (1995), Sahin
and others (1995), Ozaktas and Erden (1997).

12.7 Gaussian beam propagation

We have already seen that the propagation of light can be viewed as a process of continuous fractional
Fourier transformation. In this section we will discuss the same facts, but this time in terms of
Hermite-Gaussian beam expansions, rather than Fresnel integrals or plane wave expansions. We will
further see that the order of the fractional Fourier transform is proportional to the Gouy phase shift
accumulated by the beam as it propagates.
Let f^(x; 0) denote the complex amplitude distribution at the plane z = 0. We can expand this
function in terms of the Hermite-Gaussian functions:
 
1
^f (x; 0) = X C^n 1 n x ;
(106)
s
s
n
=0
 
Z
1
^
^
Cn = f (x; 0) s n xs dx:
(107)
We can interpret the function s?1 n (x=s) as the amplitude distribution of a one-dimensional nth
order Hermite-Gaussian beam at its waist. Then, it becomes an easy matter to write the amplitude
distribution f^(x; z ) at an arbitrary plane, since we know how each of the Hermite-Gaussian components
propagates (Saleh and Teich 1991)
#

"

1
X



ikx2 ? i(n + 1=2) (z) :
^f (x; z ) = C^n 1 n x
(108)
exp
ikz
+
m(z)s
m(z)s
2r(z )
n=0
In this equation
m(z)  p w(z)=s, where w(z) = w(0)[1 + (z=z0 )2 ]1=2 is the beam radius. Thus
p
m(0) =  w(0)=s, where w(0) is the waist radius. The Rayleigh range z0 is related to s by the
relation s2 = z0 . We also have k = 2=, where  is the wavelength. r(z ) = z [1 + (z0 =z )2 ] is the
radius of curvature of the wavefronts and  (z ) = arctan(z=z0 ) is the Gouy phase shift (Saleh and Teich
1991).
Equation 108 can be written in a considerably simple manner in terms of the fractional Fourier
transform. Let us dene functions with normalized arguments such that f^(x; z ) = f (x=s; z=s), etc.
Then the amplitude distribution at any plane is given by

f^(x; z) = m1(z) e[ikz?i (z)=2] e

where

ikx2
2r(z)

F a(z) f (u; 0)

a(z) = 2  (z):



x 
m(z)s

(109)
(110)

In equation 109, the fractional Fourier transform is taken with respect to u, and f (u; 0) = f^(su; 0).
Rewriting
(111)
(z) = 2 a(z) =  (z);
we see that the \angular order"  of the fractional Fourier transform in question is simply equal to
the Gouy phase shift accumulated in propagating from z = 0 to z . As z ! 1, we see that  (z ) ! =2
29

and a(z ) ! 1, corresponding to the ordinary Fourier transform. This is the same result discussed in
subsection 12.4.
This result can be generalized for propagation between two spherical references surfaces with
arbitrary radii (Ozaktas and Mendlovic 1994). Let the radius of the surface at z = z1 be denoted by
R1 and that of the surface at z = z2 be denoted by R2 . The radii are positive if the surface is convex
to the right. Then, there exists a fractional Fourier transform between these two surfaces whose order
is given by
(112)
a =  (z2 )=?2 (z1 ) :

It is well known that if a certain relation between R1 , R2 and z2 ? z1 holds, one obtains an ordinary
Fourier transform relation between two spherical surfaces. What we have shown is that, for other
values of the parameters, we obtain a fractional Fourier transform relation. Given any two spherical
surfaces, what we need to do to nd the order a of the fractional Fourier transform relation existing
between them, is to nd the Rayleigh range and waist location of a Gaussian beam that would \t"
into these surfaces, and then calculate a from equation 112.
We may also think of a complex amplitude distribution \riding" on a Gaussian beam wavefront.
The spatial dependence of the wavefront as the wave propagates is like a carrier dening spherical surfaces, on top of which the complex amplitude distribution rides, being fractional Fourier transformed
in the process.
Since laser resonators commonly consist of two spherical mirrors, it becomes possible to characterize
such resonators in terms of a fractional order parameter, again obtained from equation 112. The
well known stability (or connement) condition for spherical mirror resonators can be stated in a
particularly simple form in terms of the parameter a: As long as a is real, we have a stable resonator.
(In our discussion we have implicitly assumed that a and the Rayleigh range z0 are real, which means
that we have implicitly assumed stable resonators.) Unstable resonators are described by values of a
which are not real. Further details may be found in Ozaktas and Mendlovic (1994).
In addition to the relation between the fractional order parameter (z ) and the Gouy phase shift
 (z), the reader might also have noticed the similarity between the behavior of M (z), and R(z), and
the common parameters of Gaussian beams, namely the beam diameter w(z ), and the wavefront radius
of curvature r(z ). Indeed, readers well familiar with the propagation of Gaussian beams will have no
diculty interpreting the evolution of R(z ) and M (z ) in gure 8 as the wavefront radius and diameter
of a Gaussian beam.
In considering systems such as that in gure 8, we will use  (z ) to denote the accumulated Gouy
phase shift with respect to the input plane at z = 0, rather than the conventional Gouy phase shift
with respect to the last waist of the beam (Erden and Ozaktas 1997). Essentially, the accumulated
Gouy phase shift of a Gaussian beam passing through an optical system, is dened as the phase
accumulated by the beam in excess of the phase accumulated by a plane wave passing through the
same system.
In Ozaktas and Erden (1997) we have determined how the expressions for  (z ), w(z ), and r(z )
are related to the expressions for (z ), M (z ) and R(z ) (given in equations 97, 98, and 99). The
main result can be stated as follows: \Let the output of an arbitrary system consisting of lenses and
sections of free space be interpreted as a fractional Fourier transform of the input of order (z ) with
scale factor M (z ) observed on a spherical surface of radius R(z ). Let a Gaussian beam whose waist
is located at z = 0 with waist diameter w0 exhibit an accumulated Gouy phase shift  (z ), beam
diameter w(z ), and wavefront radius of curvature r(z ) at the output of the same system. If the unit
30

s appearing in equation 92 and 96 is related to w0 as s = p w0; then (z) =  (z), M (z) = w(z)=w0 ,
and R(z ) = r(z )." (Ozaktas and Erden 1997)
The reader is referred to Ozaktas and Erden (1997) for further details.

13 Applications to signal and image processing

The fractional Fourier transform has found many applications in optical and digital signal and image
processing, where the ordinary Fourier transform has traditionally played an important role. Here we
satisfy ourselves by considering a number of basic concepts and simple application examples.
In many signal processing applications, signals which we wish to recover are degraded by a known
distortion and/or by noise. Then the problem is to reduce or eliminate these degradations. Appropriate solutions to such problems depend on the observation model and the objectives, as well as the
prior knowledge available about the desired signal, degradation process and noise. A commonly used
observation model is
Z
fobs(u) = h(u; u0 )f (u0 ) du0 + n(u);
(113)
where h(u; u0 ) is the kernel of the linear system that degrades the desired signal f (u), and n(u) is an
additive noise term. The problem is to nd an estimation operator represented by the kernel g(u; u0 ),
such that the estimated signal
Z
fest (u) = g(u; u0 )fobs(u0 ) du0
(114)
minimizes the mean square error dened as:
Z

e = jfest(u) ? f (u)j2 du;

2

(115)

where the overline denotes an ensemble average. The classical Wiener lter provides a solution to the
above problem when the degradation is time-invariant and the input and noise processes are stationary.
The Wiener lter is time-invariant, and can thus be expressed as a convolution and implemented eectively with a multiplicative lter in the conventional Fourier domain with the fast Fourier transform
algorithm (gure 9a). For an arbitrary degradation model or non-stationary processes, the classical
Wiener lter often cannot provide a satisfactory result. In this case, the optimum recovery operator
is in general time-varying and has no fast implementation.
The dual of ltering in the ordinary Fourier domain is ltering in the space- or time-domain
(gure 9b). This operation simply corresponds to multiplying the original function with a mask
function. Filtering in the ordinary space or Fourier domains can be generalized to ltering in the ath
order fractional Fourier domain (gure 9c.) (Mendlovic and others 1996, Ozaktas 1996, Zalevsky and
Mendlovic 1996, Kutay and others 1997). For a = 1 this reduces to the ordinary multiplicative Fourier
domain lter and for a = 0 it reduces to space domain multiplicative ltering.
To understand the basic motivation for ltering in fractional Fourier domains, consider gure 10,
where the Wigner distributions of a desired signal and an undesired distortion are superimposed. We
observe that they overlap in both the 0th and 1st domains, but they do not overlap in the 0:5th
domain (consider the projections onto the u0 = u, u1 = , and u0:5 axes). Although we can not
eliminate the distortions in the space or frequency domains, we can eliminate them easily by using a
simple amplitude mask in the 0:5th domain.
31

-1

fest

fobs
(a)

fobs

fest
(b)

-a

fobs

fest
(c)

Figure 9: (a) Filtering in the Fourier domain. (b) Filtering in the space (or time) domain. (c) Filtering
in the ath order fractional Fourier domain.
We now discuss the optimal ltering problem mathematically. The estimated (ltered) signal fest
is expressed as (gure 9c)
fest(u) = Tsinglefobs(u);
(116)
Tsingle = F ?a g F a ;
(117)
where F a is the ath order fractional Fourier transform operator, g denotes the operator corresponding
to multiplication by the lter function g(u) and Tsingle is the operator representing the overall ltering
conguration. According to equation 117, we rst take the ath order fractional Fourier transform of
the observed signal fobs(u), then multiply the transformed signal with the lter g(u) and take the
inverse ath order fractional Fourier transform of the resulting signal to obtain our estimate. Since the
fractional Fourier transform has ecient digital and optical implementations, the cost of fractional
Fourier domain ltering is approximately the same as the cost of ordinary Fourier domain ltering.
With the above form of the estimation operator, the problem is to nd the optimum multiplicative
lter function gopt (u) that minimizes the mean-square error dened in equation 115.
For a given transform order a, gopt (u) can be found analytically using the orthogonality principle
or the calculus of variations (Kutay and others 1997):
R R
)K?a (ua ; u0 )Rf fobs (u; u0 ) du0 du ;
(118)
gopt (u) = R R KKa(u(ua; ;uu)K
a a
?a (ua ; u0 )Rfobs fobs (u; u0 ) du0 du
32


NOISE
ua

SIGNAL

Figure 10: Filtering in fractional Fourier domains as observed in the space- (or time-) frequency plane.
where the stochastic auto- and cross-correlation functions Rf fobs (u; u0 ) and Rfobs fobs (u; u0 ) can be
computed from the correlation functions Rff (u; u0 ) and Rnn(u; u0 ) (which are assumed to be known).
Fractional Fourier domain ltering is particularly advantageous when the distortion or noise is of a
chirped nature. Such situations are encountered in many real-life applications. For instance, a major
problem in the reconstruction from holograms is the elimination of twin-image noise. Since this noise is
essentially a modulated chirp signal, it can be dealt with fractional Fourier domain ltering. Another
example is the correction of the eects of point or line defects found on lenses or lters in optical
systems, which appear at the output plane in the form of chirp artifacts. Another application arises
in synthetic aperture radar which employs chirps as transmitted pulses, so that the measurements are
related to the terrain re ectivity function through a chirp convolution. This process results in chirp
type disturbances caused by moving objects in the terrain, which should be removed if high resolution
imaging is to be achieved. Fractional Fourier domain ltering has also been applied to restoration of
images blurred by camera motion or atmospheric turbulence. (Kutay and others 1997)
Further generalizations of the concept of ltering in fractional Fourier domains have been referred
to as multi-stage (repeated) and multi-channel (parallel) ltering in fractional Fourier domains (Erden
1997, Erden and others 1997a, b, Ozaktas and others 1997, Kutay and others 1998a, b). These systems
consist of M single-stage fractional Fourier domain stages in series or in parallel (gure 11a ,b). M = 1
correspond to single-stage ltering in both cases. In the multi-stage system shown in gure 11a, the
input is rst transformed into the a1 th domain where it is multiplied by a lter g1 (u). The result is then
transformed back into the original domain and the same process is repeated M times consecutively.
(Notice that this amounts to sequentially visiting the domains a1 , a2 ? a1 , a3 ? a2 , etc. and applying
a lter in each.) On the other hand, the multi-channel lter structure consists of M single-stage
blocks in parallel (gure 11b). For each channel k, the input is transformed to the ak th domain,
multiplied with a lter gk (u), and then transformed back. Let gj denote the operator corresponding
33

g2

g1
a1

a2 - a1

gM

......... FaM - aM-1

-a M

fobs

fest
(a)

g1
a1

-a 1

g2
a2

-a 2

F
.........

fobs

fest

gM
aM

-a M

Figure 11: (a) Multi-stage ltering in fractional Fourier domains. (b) Multi-channel ltering in fractional Fourier domains.
to multiplication by the lter function gj (u). Then, the outputs fest ser (u) and fest par (u) of the serial
and parallel congurations are related to the input fobs (u) according to the relations


fest ser (u) = F ?aM gM : : : F a2 ?a1 g1 F a1 fobs(u) = Tserfobs(u);

(119)

F ?ak gk F ak fobs(u) = Tparfobs(u);

(120)

fest par(u) =

"

M
X

k=1

where F aj represents the aj th order fractional Fourier transform operator and Tms, Tmc the operators
representing the overall ltering congurations.
As M is increased both the cost and exibility of the systems increase. The digital implementation
of these systems takes O(MN log N ) time and their optical implementation requires an M -stage or M channel optical system, each of whose stages or channels which should have space-bandwidth product
N . The increase in exibility as M increases will often translate into a reduction of the estimation
error. Thus we can trade o between cost and accuracy by choosing an appropriate number of stages
or channels.
34

(a)

(b)

(c)

(d)

(e)

(f)

Figure 12: Image restoration with the fractional Fourier transform.

As a simple example, we consider restoration of images blurred by a non-constant velocity (spacevariant) moving camera. Figure 12a shows the original image and gure 12b shows the blurred image.
Figure 12c shows the restoration possible by using ordinary Fourier domain ltering and gure 12d
shows that possible by single-stage ltering. In this case, the optimal domain was a = 0:7, resulting
in a mean-square error of 5%. Figure 12e and gure 12f show the restored images obtained by using
multi-channel and multi-stage ltering congurations with M = 5. We see that the two latter options
oer the best performance.
A further extension of these concepts is to combine the serial and parallel ltering congurations
in an arbitrary manner to obtain generalized ltering congurations or circuits (gure 13) (Kutay and
35

others 1998a, b).

fin

fout

Figure 13: Filter circuits. Each block corresponds to single-stage ltering.

In the above discussion, we have posed the multi-stage and multi-channel congurations as lter
structures for optimal image estimation. They can also be used for cost-ecient synthesis of desired
linear systems, transformations, or mappings, including geometric distortion compensators, beam
shapers and synthesizers, as well as linear recovery operators. In this approach, given a general linear
system H characterized by the kernel h(u; u0 ) which we wish to implement, we try to nd the optimal
orders ak and lter coecients gk , such that the overall linear operator Tsingle , Tser or Tpar (as given by
equation 117, equation 119 or equation 120), is as close as possible to H, according to some specied
criteria (such as minimum Froebenius norm of the dierence of the kernels).
The optical and digital implementations of general linear systems are costly. Using the above
mentioned approach, it is possible to approximate these systems by multi-stage or multi-channel
ltering operations in fractional Fourier domains, which are much cheaper to implement. This would
allow signicant savings in cost with little or no decrease in performance. Further discussion of this
approach in a signal processing context may be found in Erden (1997), Erden and Ozaktas (1998),
Ozaktas and others (1997), Kutay and others (1998a). We believe that this approach will nd further
applications in many other contexts.
Finally, we note that optimal ltering and image restoration is only one of the many signal processing applications explored. Correlation and pattern recognition applications have also received a
considerable amount of interest. We refer the reader to the chapter by Mendlovic and others (1998)
and also to the papers Mendlovic and others (1995d); Alieva and Agullo-Lopez (1995); Garca and
others (1996); Lohmann and others (1996b); Bitran and others (1996); Mendlovic and others (1995a).

14 Acknowledgments
It is a pleasure to acknowledge the benet of interactions with Adolf W. Lohmann. We also acknowledge the contributions of M. Fatih Erden to various parts of this chapter.
36

Parts of this report previously appeared in Ozaktas and Mendlovic 1994, 1995, and Ozaktas and
Erden 1997.

References
[1] Abe, S., and Sheridan, J. T. (1995a). Almost-Fourier and almost-Fresnel transformations. Optics
Communications, 113:385{388.
[2] Abe, S., and Sheridan, J. T. (1995b). Comment on `The fractional Fourier transform in optical
propagation problems.' J. Modern Optics, 42:2373{2378.
[3] Alieva, T., Lopez, V., Agullo-Lopez, F., and Almeida, L. B. (1994). The fractional Fourier transform in optical propagation problems. J. Modern Optics, 41:1037{1044.
[4] Alieva, T., and Agullo-Lopez, F. (1995). Reconstruction of the optical correlation function in
a quadratic refractive index medium. Optics Communications, 114:161{169, 1995. Erratum in
118:657.
[5] Almeida, L. B. (1994). The fractional Fourier transform and time-frequency representations. IEEE
Trans. Signal Process., 42:3084{3091.
[6] Alonso, M. A., and Forbes, G. W. (1997). Uniform asymptotic expansions for wave propagators
via fractional transformations. Submitted.
[7] Aytur, O. and Ozaktas, H. M. (1995). Non-orthogonal domains in phase space of quantum optics
and their relation to fractional Fourier transforms. Optics Communications, 120:166-170.
[8] Bargmann, V. (1961). On a Hilbert space of analytic functions and an associated integral transform, Part I. Comm Pure and Applied Mathematics, 14:187{214.
[9] Bastiaans, M. J. (1978). The Wigner distribution applied to optical signals and systems. Optics
Communications, 25:26-30.
[10] Bastiaans, M. J. (1979). The Wigner distribution function and Hamilton's characteristics of an
geometric-optical system. Optics Communications, 30:321-326.
[11] Bastiaans, M. J. (1979). Wigner distribution function and its application to rst-order optics. J.
Optical Society of America A, 69:1710{1716.
[12] Bastiaans, M. J. (1989). Propagation laws for the second-order moments of the Wigner distribution function in rst-order optical systems. Optik, 82:173{181.
[13] Bastiaans, M. J. (1991). Second-order moments of the Wigner distribution function in rst-order
optical systems. Optik, 88:163{168.
[14] Beck, M., Raymer, M. G., Walmsley, I. A., and Kong, V. (1993). Chronocyclic tomography for
measuring the amplitude and phase structure of optical pulses. Optics Letters, 18:2041{2043.
[15] Bernardo, L. M. and Soares, O. D. D. (1994a). Fractional Fourier transforms and optical systems.
Optics Communications, 110:517{522.
37

[16] Bernardo, L. M. and Soares, O. D. D. (1994b). Fractional Fourier transforms and imaging. J.
Optical Society of America A, 11:2622{2626.
[17] Bitran, Y., Zalevsky, Z., Mendlovic, D., and Dorsch, R. G. (1996). Fractional correlation operation: Performance analysis. Applied Optics, 35:297{303.
[18] De Bruijn, N. G. (1973). A theory of generalized functions, with applications to Wigner distribution and Weyl correspondence. Nieuw Archief voor Wiskunde, 21:205{280.
[19] Bracewell, R. N. (1995).1 Two-Dimensional Imaging. Prentice-Hall, Englewood Clis, New Jersey,
1995.
[20] Claasen, T. A. C. M. and Mecklenbrauker, W. F. G. (1980a). The Wigner distribution|a tool for
time-frequency signal analysis. Part I: continuous-time signals. Philips J. Reseasrch, 35:217{250.
[21] Claasen, T. A. C. M. and Mecklenbrauker, W. F. G. (1980b). The Wigner distribution|a tool
for time-frequency signal analysis. Part II: discrete-time signals. Philips J. Research, 35:276{300.
[22] Claasen, T. A. C. M. and Mecklenbrauker, W. F. G. (1980c). The Wigner distribution|a tool
for time-frequency signal analysis. Part III: relations with other time-frequency signal transformations. Philips J. Research, 35:372{389.
[23] Cohen, L. (1989). Time-frequency distributions|a review. Proceedings of IEEE, 77:941{981.
[24] Cohen L. (1995). Time-Frequency Analysis. Prentice-Hall, Englewood Clis, New Jersey.
[25] Condon, E. U. (1937). Immersion of the Fourier transform in a continuous group of functional
transformations. Proc National Academy of Sciences, 23:158{164.
[26] Dorsch, R. G. (1995). Fractional Fourier transformer of variable order based on a modular lens
system. Applied Optics, 34:6016{6020.
[27] Dorsch, R. G. and Lohmann, A. W. (1995). Fractional Fourier transform used for a lens design
problem. Applied Optics, 34:4111{4112.
[28] Dragoman, D. (1996). Fractional Wigner distribution function. J. Optical Society of America A,
13:474{478.
[29] Erden, M. F., Ozaktas, H. M., and Mendlovic, D. (1996a). Propagation of mutual intensity
expressed in terms of the fractional Fourier transform. J. Optical Society of America A, 13:1068{
1071.
[30] Erden, M. F., Ozaktas, H. M., and Mendlovic, D. (1996b). Synthesis of mutual intensity distributions using the fractional Fourier transform. Optics Communications, 125:288{301.
[31] Erden, M. F., Ozaktas, H. M., Sahin, A., and Mendlovic, D. (1997a). Design of dynamically
adjustable anamorphic fractional Fourier transformer. Optics Communications, 136:52-60.
[32] Erden, M. F., Ozaktas, H. M. (1997). Accumulated Gouy phase shift in Gaussian beam propagation through rst-order optical systems, J. Optical Society of America B, 14:2190-2194.
38

[33] Erden, M. F. (1997). Repeated Filtering in Consecutive Fractional Fourier Domains. Ph.D. Thesis,
Bilkent University, Ankara.
[34] Erden, M. F., Kutay, M. A., and Ozaktas, H. M. (1997b). Repeated ltering in consecutive
fractional Fourier domains and its application to signal restoration. Sub. to IEEE Trans. Signal
Procesing.
[35] Erden, M. F. and Ozaktas, H. M. (1998). Synthesis of general linear systems with repeated
ltering in consecutive fractional Fourier domains. To appear in J. Optical Society of America A.
[36] Fonollosa, J. R. and Nikias, C. L. (1994). A new positive time-frequency distribution. In Proc.
1994 Int. Conf. Acoustics, Speech, and Signal Processing, IEEE, Piscataway, New Jersey, IV-301{
IV-304.
[37] Garca, J., Mendlovic, D., Zalevsky, Z., and Lohmann, L. (1996). Space-variant simultaneous
detection of several objects by the use of multiple anamorphic fractional-Fourier-transform lters.
Applied Optics, 35:3945{3952.
[38] Gori, F., Santarsiero, M., and Bagini, V. (1994). Fractional Fourier transform and Fresnel transform. Atti Foundaz Georgio Ronchi.
[39] Granieri, S., Trabocchi, O., and Sicre, E. E. (1995). Fractional Fourier transform applied to
spatial ltering in the Fresnel domain. Optics Communications, 119:275{278.
[40] Hlawatsch, F. and Boudreaux-Bartels, G. F. (1992), Linear and quadratic time-frequency signal
representations. IEEE Signal Processing Magazine, April:21{67.
[41] Jiang, Z. (1995). Scaling laws and simultaneous optical implementation of various order fractional
Fourier transforms. Optics Letters, 20:2408{2410.
[42] Kutay, M. A., Ozaktas, H. M., Arkan, O., and Onural, L. (1997). Optimal Filtering in Fractional
Fourier Domains. IEEE Trans Signal Process., 45: 1129-1143.
[43] Kutay, M. A. and Ozaktas, H. M. (1998). Optimal image restoration with the fractional Fourier
transform. To appear in J. Optical Society of America A.
 and Candan, C. (1998a).
[44] Kutay, M. A., Erden, M. F., Ozaktas, H. M., Arkan, O., Guleryuz, O.
Space-Bandwidth ecient realizations of linear systems. Submitted to Optics Letters.
[45] Kutay, M. A., Arkan, O., Candan, C., Guleryuz, O ., Erden, M. F., Ozaktas, H. M. (1998b). CostEcient approximation of linear systems with multi-channel fractional Fourier domain ltering.
Submitted to IEEE Signal Process. Letters.
[46] Liu, S., Xu, J. , Zhang, Y. , Chen, L., and Li, C. (1995). General optical implementation of
fractional Fourier transforms. Optics Letters, 20:1053{1055.
[47] Lohmann, A. W. (1993). Image rotation, Wigner rotation, and the fractional order Fourier transform. J. Optical Society of America A, 10:2181{2186.
[48] Lohmann, A. W. and Soer, B. H (1994). Relationships between the Radon-Wigner and fractional
Fourier transforms. J. Optical Society of America A, 11:1798{1801.
39

[49] Lohmann, A. W. (1995). A fake zoom lens for fractional Fourier experiments. Optics Communications, 115:437{443.
[50] Lohmann, A. W., Mendlovic, D., Zalevsky, Z., and Dorsch, R. G. (1996a). Some important
fractional transformations for signal processing. Optics Communications, 125:18{20.
[51] Lohmann, A. W., Zalevsky, Z., and Mendlovic, D. (1996b). Synthesis of pattern recognition lters
for fractional Fourier processing. Optics Communications, 128:199{204.
[52] McAlister, D. F., Beck, M., Clarke, L., Meyer, A., and Raymer, M. G. (1995). Optical phaseretrieval by phase-space tomography and fractional-order Fourier transforms. Optics Letters,
20:1181{1183.
[53] McBride, A. C. and Kerr, F. H. (1987). On Namias's fractional Fourier transform. IMA J. Applied
Mathematics, 39:159{175.
[54] Mecklenbrauker, W. F. G. (1993). The Wigner Distribution: Theory and Applications in Signal
Processing. W. F. G. Mecklenbrauker, editor. Elsevier, Amsterdam.
[55] Mendlovic, D. and Ozaktas, H. M. (1993). Fractional Fourier transformations and their optical
implementation, I. J. Optical Society of America A, 10:1875{1881.
[56] Mendlovic, D., Ozaktas, H. M., and Lohmann, A. W. (1994a). Graded-index bers, Wignerdistribution functions, and the fractional Fourier transform. Applied Optics, 33:6188{6193.
[57] Mendlovic, D., Ozaktas, H. M., and Lohmann, A. W. (1994b). Self Fourier functions and fractional
Fourier transforms. Optics Communications, 105:36{38.
[58] Mendlovic, D., Bitran, Y., Dorsch, R. G., and Lohmann, A. W. (1995a). Optical fractional
correlation: experimental results. Applied Optics, 34:1665{1670.
[59] Mendlovic, D., Bitran, Y., Dorsch, R. G., Ferreira, C., Garcia, J., and Ozaktaz, H. M. (1995b).
Anamorphic fractional Fourier transform: optical implementation and applications. Applied Optics, 34:7451{7456.
[60] Mendlovic, D., Zalevsky, Z., Konforti, N., Dorsch, R. G., and Lohmann, A. W. (1995c). Incoherent
fractional Fourier transform and its optical implementation. Applied Optics, 34:7615{7620.
[61] Mendlovic, D., Ozaktas, H. M., and Lohmann, A. W. (1995d). Fractional correlation. Applied
Optics, 34:303{309.
[62] Mendlovic, D., Dorsch, R. G., Lohmann, A. W., Zalevsky, Z., and Ferreira, C. (1996a). Optical
illustration of a varied fractional Fourier-transform order and the Radon-Wigner display. Applied
Optics, 35:3925{3929.
[63] Mendlovic, D., Zalevsky, Z., Lohmann, A. W., and Dorsch, R. G. (1996b). Signal spatial-ltering
using the localized fractional Fourier transform. Optics Communications, 126:14{18.
[64] Mendlovic, D., Zalevsky, Z., and Ozaktas, H. M. (1998). The applications of the fractional Fourier
transform to optical pattern recognition. In Optical Pattern Recognition, Academic Press.
40

[65] Mihovilovic, D. and Bracewell, R. N. (1991). Adaptive chirplet representation of signals on timefrequency plane. Electronics Letters, 27:1159{1161.
[66] Moshinsky, M. and Quesne, C. (1971). Linear canonical transformations and their unitary representations. J Mathematics Physics, 12:1772{1780,.
[67] Moshinsky, M., Seligman, T. H., and Wolf, K. B. (1972). Canonical transformations and the
radial oscillator and Coulomb problem. J Mathematics Physics, 13:901{907.
[68] Mustard, D. A. (1987). Lie group imbeddings of the Fourier transform. School of Mathematics
Preprint AM87/13, The University of New South Wales, Kensington, Australia.
[69] Mustard, D. A. (1987). The fractional Fourier transform and a new uncertainty principle. School
of Mathematics Preprint AM87/14, The University of New South Wales, Kensington, Australia,
1987.
[70] Mustard, D. A. (1989). The fractional Fourier transform and the Wigner distribution. School of
Mathematics Preprint AM89/6, The University of New South Wales, Kensington, Australia.
[71] Mustard, D. A. (1991). Uncertainty principles invariant under the fractional Fourier transform.
J Australian Mathematical Society B, 33:180{191.
[72] Mustard, D. A. (1996). The fractional Fourier transform and the Wigner distribution. J Australian
Mathematical Society B, 38:209{219.
[73] Mustard, D. A. (1997). Fractional convolution. To appear in J Australian Mathematical Society
B.
[74] Namias, V. (1980). The fractional order Fourier transform and its application to quantum mechanics. J Inst Maths Applics, 25:241{265.
[75] Nazarathy, M. and Shamir, J. (1982). `First-order optics|a canonical operator representation:
lossless systems. J. Opt. Soc. Am., 72:356{364.
[76] Ozaktas, H. M. and Mendlovic, D. (1993a). Fourier transforms of fractional order and their optical
interpretation. Optics Communications, 101:163{169.
[77] Ozaktas, H. M. and Mendlovic, D. (1993b). Fractional Fourier transformations and their optical
implementation, II. J. Optical Society of America A, 10:2522{2531.
[78] Ozaktas, H. M., Barshan, B. , Mendlovic, D., and Onural, L. (1994a). Convolution, ltering, and
multiplexing in fractional Fourier domains and their relation to chirp and wavelet transforms. J.
Optical Society of America A, 11:547{559.
[79] Ozaktas, H. M., Barshan, B. and Mendlovic, D. (1994b). Convolution and Filtering in Fractional
Fourier Domains. Optical Review, 1:15{16.
[80] Ozaktas, H. M. and Mendlovic, D. (1994) Fractional Fourier transform as a tool for analyzing
beam propagation and spherical mirror resonators. Optics Letters, 19:1678{1680.
[81] Ozaktas, H. M. and Mendlovic, D. (1995) Fractional Fourier optics. J. Optical Society of America
A, 12:743{751.
41

[82] Ozaktas, H. M. and Aytur, O. (1995). Fractional Fourier domains. Signal Processing, 46:119{124.
[83] Ozaktas, H. M. (1996). Repeated fractional Fourier domain ltering is equivalent to repeated
time and frequency domain ltering. Signal Processing, 54:81-84.
[84] Ozaktas, H. M. and Mendlovic, D. (1996). Every Fourier optical system is equivalent to consecutive fractional-Fourier-domain ltering. Applied Optics, 35:3167{3170.
[85] Ozaktas, H. M., Erkaya, N., and Kutay, M. A. (1996a). Eect of fractional Fourier transformation
on time-frequency distributions belonging to the Cohen class. IEEE Signal Processing Lett, 3:40{
41.
[86] Ozaktas, H. M., Arikan, O., Kutay, M. A., and Bozdag, G. (1996b). Digital computation of the
fractional Fourier transform. IEEE Trans Signal Processing, 44:2141-2150.
[87] Ozaktas, H. M. and Erden, M. F. (1997). Relationships among ray optical, Gaussian beam, and
fractional Fourier transform descriptions of rst-order optical systems.Optics Communications
143:75-86.
[88] Ozaktas, H. M., Erden, M. F. and Kutay, M. A. (1997). Cost-Ecient Approximation of Linear
Systems with Repeated Filtering. Submitted to IEEE Signal Process. Letters.
[89] Pellat-Finet, P. (1994). Fresnel diraction and the fractional-order Fourier transform. Optics
Letters, 19:1388{1390.
[90] Pellat-Finet, P. and Bonnet, G. (1994). Fractional order Fourier transform and Fourier optics.
Optics Communications, 111:141{154.
[91] Raymer, M. G., Beck, M., and McAlister, D. F. (1994a). Complex wave-eld reconstruction using
phase-space tomography. Physical Review Letters, 72:1137{1140.
[92] Raymer, M. G., Beck M., and McAlister D. (1994b). Spatial and temporal optical eld reconstruction using phase-space tomography. In Quantum Optics VI, Springer, Berlin.
[93] Sahin, A., Ozaktas, H. M., and Mendlovic, D. (1995). Optical implementation of the twodimensional fractional Fourier transform with dierent orders in the two dimensions. Optics
Communications, 120:134{138.
[94] Saleh, B. E. A. and Teich, M. C. (1991). Fundamentals of Photonics. Wiley, New York.
[95] Seger, O. (1993). Model Building and Restoration with Applications in Confocal Microscopy, Ph.D.
thesis, Linkoping University, Sweden.
[96] Smithey, D. T., Beck, M., Raymer, M. G., and Faridani, A. (1993). Measurement of the Wigner
distribution and the density matrix of a light mode using optical homodyne tomography: application to squeezed states and the vacuum. Physical Review Letters, 70:1244{1247.
[97] Wiener, N. (1929). Hermitian Polynomials and Fourier Analysis.Journal of Mathematics Physics
MIT 18, pp 70-73.
[98] Wiener, N. (1933). The Fourier Integral and Certain of its Applications. Cambridge University
Press, Cambridge.
42

[99] Wolf, K. B. (1979). Construction and properties of canonical transforms. In Integral Transforms
in Science and Engineering, Plenum Press, New York, chapter 9.
[100] Wood, J. C. and Barry, D. T. (1994a). Tomographic time-frequency analysis and its application
toward time-varying ltering and adaptive kernel design for multicomponent linear-FM signals.
IEEE Trans Signal Processing, 42:2094{2104.
[101] Wood, J. C. and Barry, D. T. (1994b). Linear signal synthesis using the Radon-Wigner transform.
IEEE Trans Signal Processing, 42:2105{2111.
[102] Zalevsky, Z. and Mendlovic, D. (1996). Fractional Wiener lter. Applied Optics, 35:3930{3936.

43