AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June.

2006

Why is the Linear Canonical Transform so

little known?
Adrian Stern

Department of Electro Optical Engineering, Ben Gurion University of the Negev, Beer-Sheva 84105,
Israel, Phone: 972-8-6461840, Email: stern@bgu.ac.il

Abstract. The linear canonical transform (LCT), is the name of a parameterized continuum of
transforms which include, as particular cases, the most widely used linear transforms and
operators in engineering and physics such as the Fourier transform, fractional Fourier transform
(FRFT), Fresnel transform (FRST), time scaling, chirping, and others. Therefore the LCT
provides a unified framework for studying the behavior of many practical transforms and system
responses in optics and engineering in general. From the system-engineering point of view the
LCT provides a powerful tool for design and analysis of the characteristics of optical systems.
Despite this fact only few authors take advantage of the powerful and general LCT theory for
analysis and design of optical systems. In this paper we review some important properties about
the continuous LCT and we present some new results regarding the discretization and
computation of the LCT.
Keywords: linear canonical transform, fractional Fourier transform, Fresnel transform, Laplace
transform, sampling, Wigner distribution, time-frequency representation, phase space.
PACS: 02.30.Uu, 02.30.Nw, 42.40.Lx, 42.25.Bs

1. INTRODUCTION
The title of this paper is borrowed from Ref. [1] questioning why is the Fresnel
transform so little known. Here we repeat the question for a more general and more
widely used transform: the linear canonical transform (LCT). This paper tries to
convince that LCT deserves special consideration and should be better and widely
known. The main motivation is the fact that LCT provides a general mathematical tool
with very broad applicability in many fields of science and engineering. Although it is
not very much known, its special cases are widely used in various fields, often under
different names. Therefore, understanding the LCT may help to gain more insights on
its special cases and to carry over knowledge gained from one subject to others.
The LCT is a four-parameter (a,b,c,d) class of linear integral transform [2]-[5]. It
provides a canonical formalism for the response of a very large class of physical
systems. LCTs have been reinvented or reconsidered by many authors under many
different names at different times in different contexts [3], a fact that implies universal
importance of the transform. The LCT is also known as the Generalized Fresnel
Transform [6] or ABCD transform [7], Collins formula [8] generalized Huygesn
integrals [9] and Moshinsky and Queesne integrals [2], and is a special case of the
Special Affine Fourier Transform [10].
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

The LCT with parameters {a,b,c,d}of a function f(x), is defined :

∞
~M
f ( y ) = O M { f ( x)}( y ) = ∫ f ( x)C ( y, x )dx,
M
−∞
(1)
1 2b (ax 2 −2 xy + dy 2 )
j
CM ( y, x ) = e b≠0
j 2πb
where M ≡  
a b
 and det(M)=ad-bc=1. For the case of b=0 the LCT is defined as
c d 
the limit of (1) with b → 0 , yielding [4]:
j
~M cdy 2
f ( y ) = O M { f ( x )}( y ) = d e 2 f (dy ), b = 0 . (2)
Please note that since det(M)=1 only three transform parameters are free. Therefore
the LCT can be defined also as a three parameter transform [3]. However here we will
adopt the four parameter definition since the matrix M formalism is particularly
helpful in defining properties of LCT (see Sec. 2) and for understanding the operation
of the LCT on the phase space (see Sec. 3).
It is easy to verify that the LCT with parameters {a,b,c,d}={cosθ,sinθ,-sinθ, cosθ}
reduces to FRFT [3],[5] which, in the specific case θ=π/2, becomes the Fourier
transform. With parameters {a,b,c,d}={1,b,0,1} the LCT reduces to the Fresnel
transform [3],[5]. Multiplication by Gaussian or chirp function is obtained with an
{a,b,c,d}={1,0,c,1}. Scaling operator can be viewed as a special case of the LCT with
{a,b,c,d}={d-1,0,0,d}.
In this paper we restrict ourselves to the class of LCT with real parameters
{a,b,c,d}. In such cases the LCT operator is unitary in L2(R). With the price of
giving up on unitarity, other important transforms can be found as special cases of
LCT. One such example is the Laplace transform obtained with {a,b,c,d}={0,j,j,0}.
The LCT was found useful in many fields in engineering. The LCT describes the
response of any quadratic-phase system [11] and of cascade combinations of such
systems. In optics any system that is implemented using any arbitrary number of thin
lenses and propagation through free space under Frenel approximation, or through
sections of graded-index media, belong to the class of quadratic-phase systems. The
propagation of fields in first order optics can be described efficiently by the LCT. The
LCT can also be applied in the theory of chirp laser pulses [12]. In Radar theory, LCT
can be used in describing pulse compression and chirp-component detection. The LCT
can be used in image processing for pattern recognition, image deblurring, and
watermarking. In acoustics, LCT can be used to describe bat-chirp type signals and
propagation. In communication theory LCT, was used to design multichannel
communication systems and may be used to cope with multi-path problems.
The outline of this paper is as follows. In section 2 we summarize briefly some
important properties of the continuous parameter (time/space) LCT (CPLCT). In
section 3 we show the interpretation of the LCT in joint parameter space (e.g. phase
space or time-frequency space). In section 4 the sampling theorem for LCT is
presented and in section 5 the discrete LCT is derived. The material is essentially
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

tutorial. However, the presentation is original, and it includes a number of new results
regarding the discretization of the LCT (sections 4 and 5).

2. SOME PROPERTIES OF THE CONTINUOUS PARAMETER

LCT
In this section we summarize some important properties of the CPLCT.
The LCT defined in (1)-(2) obeys the additivity property:
O M1 O M 2 = O M1M 2 . (3)
From (3) we see that the inverse LCT (ILCT) is obtained by an LCT with
 d − b −1
M −1 =   , so that the integration kernel in (1) satisfies C M ( y , x) = C M −1 ( y, x) .
− c a 
Following this property, together with the definition (1), it is easy to verify that if M is
real then the inverse transform kernel is the Hermitian conjugate of the original
transform kernel:
C M−1 ( y, x) = C M* ( x, y ) . (4)
Equation (4) implies that the LCT with real parameters is unitary. Associated with the
unitarity property is the generalized Parseval’s relation:
~M
∫ f ( y )g~ ( y ) dy = ∫ f (x )g (x ) dx
M * *
(5)
~M
where f ( y ) and g~ M ( y ) are the LCTs of two arbitrary functions f(x) and g(x). In the
specific case f(x) = g(x) (5) reduces to Parseval’s relation (energy conservation):
~M
f ( y ) = f ( x ) , where ⋅ denotes the regular norm of order 2.
( )
The LCT obeys the associative law; i.e., C M1 C M 2 C M 3 = C M1 C M 2 C M 3 . This implies ( )
that a cascade of quadratic phase systems can be arbitrary rearranged.
Some useful properties of the LCT are summarized in Table 1.

3. GEOMETRIC INTERPRETATION
Useful insight on the operation of the LCT can be gained by examining its effect on
the Wigner distribution. The Wigner distribution of f(x) is defined by [3],[5],[11]:
∞
1  ξ  ξ
W f ( x, ω x ) = ∫ f  x + 2  f  x − 2 e
− jξω x
dξ
2π −∞ . (6)
~M
It can be shown [3],[5], [13] that the Wigner distribution of f ( y ) is related to that of
f(x) by:
W ~f M ( y , ω y ) = W f [dy − bω y ,−cy + aω y ] . (7)
The meaning of (7) is that the LCT performs a homogeneous linear mapping in the
Wigner domain:
 y   a b  x 
  =    . (8)
ω 
  
y c d  ω y 
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

TABLE 1. Properties of LCT. The sign ‘*’ in property 8 denotes convolution and “ ⊗ ” in property
9 denotes correlation.
f(x) ~M
f ( y)
1 f(-x) ~M
f (− y )
2 f ∗ (x) ~ ∗ M −1
f ( y)
3 f (x / k)  a b/k 2 
~
k f M ' ( y / k ) M ' =  2 

 ck d 
4 f ( x − x0 ) jcx0 y − 2 acx0 ~ M
j 2

e e f ( y − ax 0 )
5 e jω 0 x f ( x ) − dbω 2 ~
j

e jdω0 y e 2 0 f M ( y − bω 0 )
6 x n f (x) n
 d  ~M
dy + jb  f ( y )
 dy 
7 d n f ( x) n
j −n  d  ~M
dx n − cy − ja  f ( y )
 dy 
8 { f ∗ g}(x) ~M − c (ax
j
2
− 2 xy )
∫ f ( y − ax) g ( x)e 2 dx

9 {f ⊗ g }(x) ~M − c (ax 2 + 2 xy )
j

∫ f ( y + ax ) g * ( x ) e 2
dx

Examples of the linear mapping (8) performed by some special case of LCT are
depicted in Fig.1. The mapping (8) preserves the area of the support in Wigner domain
implying the conservations of degrees of freedom under LCT, which is consistent with
the fact that the LCT is invertible.
Finally it is noted that the above described geometric effect of the LCT in the (x,ω)
domain holds not only on Wigner distribution but on almost any Cohen class
representation [7].

4. SAMPLING OF THE LCT

In this and the following section we consider only LCT with b ≠ 0 . This is the
more interesting case. As can be seen in equation (2) the LCT with b=0 basically
scales the parameter axis.
Sampling is central in almost any domain because it provides the link between the
continuous physical signals and the discrete time domain. Sampling of LCTed signals
was investigated in [14] and [15]. In [14] is demonstrated that if a function f(x) has a
~
compact support such that f(x)=0 for x > Lx 2 , then its LCT f M ( y ) can be exactly
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

~
reconstructed from its sampled version fTMy ( y ) at points yn=nTy, n ∈ Z by the
formula:
~M
f ( y) =
2π M 
ωy
x −1 ~
{ 
O rect ( )O M f TMy ( y ) ( x)( y )
Lx
} (9)
 
if the following sampling condition is fulfilled.
2π b
Ty ≤ . (10)
Lx
One may verify that for LCT with parameters {0,1,-1,0} implementing a FT, condition
(10) reduces to the Nyquist sampling criterion and (9) performs a reconstruction by
~
applying a low-pass filter in the reciprocal domain of f M ( y ) . Alternatively,
~M
f ( y ) can be reconstructed using the interpolation formula [15]:

f ( y ) = e 2 b ∑ f M (nT y )
~M j
dy 2 ~ [ ( )]
sin π Ty − n − j d (nTy )2
y

n∉Z ( )
π Ty − n
y
e 2b . (11)

It can be shown that previously developed sampling theorems for FT (Shannon

sampling theorem), FRST and FRFT are special cases of the LCT described above
[14],[15].

ωx ωy

x y

(a)
(b)
ωy
ωy

y y

(c) (d)

Fig. 1 Examples of the effect of the LCT on the support of WD: (a) FRST, (b)
FRFT, (d) FT.
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

5. DISCRETIZATION OF THE LCT

~
Let us consider the sampled LCT signal f M ( y ) obtained by a multiplication of f(x)
by an impulse comb with intervals Ty:
∞
f Ty ( y ) = f M ( y ) ∑ δ ( y − kT y )
~M ~
(12)
k = −∞
By using the well known
Poisson formula for Dirac’s comb
∞ ω y ∞ jkω y y 2π
∑ δ ( y − nT y ) = ∑ e ,where ω y = , together with property 5 in Table 1, it
n = −∞ 2π k = −∞ Ty
~
is easy to show that the ILCT of fTMy ( y ) gives [14]

{~f }
∞
( y )C M−1 (kT y , x ) =
~M
( y ) (x ) = ∑f
−1
f px ( x) = O M M
Ty Ty
k = −∞

ω y ~ M ωy
jkω y 
{~f }
∞ ∞
( y ) (x + nbω y ) =
( )2
f ( y) ∑ e ( x ) = ∑e
−1 janω y x ( j / 2 ) ab nω y −1
OM 
y
e OM M

 2π n = −∞  2π n = −∞
∞
f (x + nbω y ).
1 ( )2
∑e
janω y x ( j / 2 ) ab nω y
e
Ty n = −∞

(13)
2π b
By denoting Px = , (13) can be written as:
Ty
∞ a a a a
Px 1 − j 2b x 2 ∞ j 2b ( x − mPx )2
( mPx )2
f ( x + mPx ) =
f ( x − mPx ) .
j mPx x j
f p x ( x) =
2π b
m = −∞
∑e Ty
b
e ∑ee
m = −∞
2b

(14)
~M
We can see that the ILCT of fTy ( y ) yields a chirp-periodical signal by forming
modulated shifts replicas of the original signal f(x). Each replica is multiplied be a
constant and linear phase term depending on the order of the replica m, Px and on the
~
ratio a/b. The relation between fTMy ( y ) and its ILCT f p x (x) is illustrated in Figs. 2 (c)
and (d).
Note that if f(x) has a compact support such that f(x)=0 for x > Lx 2 and if
1
L x ≤ Px then f ( x ) for x ≤ Px / 2 . Therefore, if L x ≤ Px , which is
f p x ( x) =
Tx
equivalent to the sampling condition (10), then f(x) can be precisely reconstructed
from the zero order replica of f p x (x) . This observation leads to the reconstruction
formula (9).
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

f(x) fM(y)
A

1
y
M
Arf{ f (y)}
x

y
(a) (b)
f Px ( x ) ~M
Px
f Ty ( y )
1/Ty
A

x y
Arg{ f P ( x ) } Ty
x
Arg{ ~fTM ( y ) }
y

x y

(c) (d)

~M
f Py ( y )
f Tx ( x ) Py
A/Tx

1
y
Arg{ ~f PM ( y ) }
y
x
Tx
y
(e) (f)

f PxTx ( x ) ~M
Px Py f PyTy ( y )
1 Ty/Tx

Tx x y
Ty
(g) (h)
Fig. 2. (a) The function f(x) and its ILCT (b), (c) the ILCT of the sampled LCT (d), (e ) the
sampled f(x) and its ILCT, (f) sampled and periodically replicated f(x) and its LCT (h)
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

From similar considerations leading to (14) we find that if we sample the function
f(x) at intervals Tx:
∞
f Tx ( x) = f ( x) ∑ δ (x − nTx ) (15)
n = −∞

then the LCT of the sampled version of fTx (x) is dual to (14), i.e.:
1 j 2b y 2 ∞ − j 2b ( y −lPy )2 ~ M
∞ d d

f ( y − lPy ). (16)
~M ~M
f Py ( y ) = ∑ f (nT x )C M ( y , nT x ) = e ∑e
n = −∞ Tx l = −∞

2π b ~
where Py = . The relation between the pair fTx (x) and f PM ( y ) is illustrated in
Tx
Figures 2(e) and (f). Thus we see that sampling in x domain causes chirp- periodicity
in y domain.
For numerical calculation of LCT pairs we need to consider sampling both x and y
domains. Combining the results in Eqs. (14) and (16) together with the results in
section 4, we see that a signal sampled at rate 1/Tx and chirp-periodic with a period Px
is a CPLCT periodic with a period Py defined on discrete values of y with interval Ty,
where
2π b 2π b
Ty = and Tx = . (17)
Px Py
This relation is illustrated in Figs. 2(g) and (h). Note that the number of samples in
P Px Py
each replica in the x domain [Fig. 2(g)] is N x = x = , which is the same as the
Tx 2π b
Py Px Py
number of samples in each replica in y domain; N y = = = N x . For future use
Ty 2π b
we define N = N x = N y = 2 N 1 + 1 .
The sampling- (chirp) periodical replication property shown in Fig. 2 is reminiscent
of the sampling-periodical replication of the Fourier pair. Therefore, it is natural to
define the discrete linear canonical transform (DLCT) pair in analogy with the
definition of the discrete Fourier transform:

~ M ,Tx ,Ty N N1 [a (nTx )2 −2 nkTxTy + d (kTy )2 ]

i

f [k ] = 1 1
∑ f x y [n]e 2b
T ,T N
, (18)
Tx j 2πb n = − N1
and
− [d ( nTx )2 − 2 nkTxT y + a (kTy )2 ]
N1 i
1 1 ~ M ,Tx ,Ty N
f
Tx ,T y N
[n] = ∑ f [k ]e 2b
. (19)
Ty j 2πb n = − N1
From Eqs. (14),(16) and Fig. 2 we see that if f(x)=0 for x > L x 2 and Lx<Px, and
~
if its CPLCT f M ( y ) is ε-concentrated in the range y ≤ Py / 2 , i.e., only a small
1
fraction ε of its energy is out this range, then f p x ( x) = f ( x ) for x > L x 2 and
Tx
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

~M 1 ~M
f py ( y) ≈ f ( y ) for y ≤ Py / 2 . This implies that if f(x) has compact support and if
Ty
the aliasing in y domain is negligible, then the DLCT pair defined by (18) and (19) is a
good approximation of the of the CPLCT pair at discrete values, that is:
f (kT y ) ≈ f x y [k ] ,
~M ~ M ,T ,T N
(20)
f (nTx ) ≈ f x y [n].
T ,T N
(21)
Thus, for compact signals and negligible aliasing in the reciprocal LCT domain,
(20) and (21) may be used for numerical calculation of the LCT defined in (1).
As a final note we point out that fast algorithms may be used to calculate the DLCT
defined in Eqs. (18) and (19). One such way is via the fast Fourier transform [13]. The
other is by applying a fast algorithm for transforms with quadratic kernels such as
developed in [16]. Although the fast LCT algorithm in [16] is developed for a slightly
different DLCT kernel, it can be easily adapted to the DLCT defined here.

6. CONCLUSIONS
In this paper we point out the usefulness of the LCT and give a sketch of its theory.
We have summarized important properties of the CPLCT and presented new results
regarding the discrestization of the LCT.
We imply that more broad awareness of the LCT and its properties could be of
great benefit, as it provides a unified mathematical tool of many commonly used tools.
As such, it gives a more general perspective on many specific mathematical tools and
a better insight to them. Moreover, it helps to carry over knowledge from one field to
other and avoids rediscovering the same theories in different fields just because
different terminology is used. For example, theories of Fresnel transform studied for
optical field propagation, theories developed for quadrature-phase filtering used in
signal processing and for pulse compression, theories developed for chirp pulse
propagation in radar, in optical communication theory and for echo bat propagation,
and theory of the imaginary Gauss-Weierstrass transform studied in mathematics,
physics and heat theory, were developed separately although they may be unified as an
LCT with parameters (1,b,0,1). So, after pointing out the benefits of LCT we return to
the question in the title: why is the LCT still so little known? One possible reason is
that the LCT can be expressed in terms of the Fourier transform; it can be written as
suitable filtering operator on the Fourier transform of some suitable modification of
the signal. But isn’t the preference of such a point of view biased from our general
tendency to analyze automatically every problem through Fourier glasses, with which
are we used to? Also, one may argue that expressing the LCT in terms of operations
on Fourier transform is in concordance with the general approach with which
complicated things are understood by decomposing them into elementary ones. Indeed
this approach is generally useful. But sometimes the opposite approach may be more
advantageous; i.e., identifying simpler things (e.g. Fourier transform, Fresnel
transform etc.) as a subclass of some more complicated one (e.g. LCT). For instance a
chair may be better defined as a type of furniture used to sit on rather than as a certain
combination of wood pieces, screws and nails.
AIP Conference Proc., 5’th Int. Workshop on Information Optics, Vol. 860, pp. 225-234, June. 2006

REFERENCES
1. F. Gori, ‘‘Why is the Fresnel transform so little known?’’ in Current Trends in Optics, J. C. Dainty, ed.
London, Academic, 1994, pp.140-148.
2. M. Moshinsky and C. Quesne, J. Math. Phys., 12( 8), 1772–1783 (1971).
3. H. M. Ozaktas, M. A. Kutay, and Z. Zalevsky, The Fractional Fourier Transform With Applications in
Optics and Signal Processing. New York: Wiley (2000).
4. K. R. Wolf, Integral transforms in science and engineering, NY, Plenum Press, 1979, chps. 9,10.
5. S.C. Pei, and J. J. Ding, IEEE Trans. on Signal Proc, 49 (8),1638-1655 (2001).
6. D. F. V. James and G. S. Agarwal, Opt. Comm.., 126, 207–212 (1996).
7. L. M. Bernardo, Opt. Eng., 35,(3), 732–740 (1996).
8. S. A. Collins, J. Opt. Soc. Am. A, 60, 1168-1174 (1970).
9. M. Nazarathy and J. Shamir, ‘‘First-order optics—a canonical operator representation: lossless systems,’’
J. Opt. Soc. Am. 72, pp. 356–364, 1982.
10. S. Abe, J.T. Sheridan, Opt. Lett., 19, 1801-1803 (1994).
11. M. J. Bastiaans, J. Opt, Soc, Am. 69, 1710-1716 (1979).
12. E. Siegman, Lasers, California, Mill Valley, 1986.
13. M. Hennelly and J. T. Sheridan,” J. Opt, Soc, Am. A., 22, 917-927, (2005).
14. A. Stern, Signal Processing (publication of EURASIP) , 86 (7), 1421-1425 (2006).
15. A. Stern, “Sampling of compact signals in the offset linear canonical domain” submitted
to publication
16. M. Hennelly and J. T. Sheridan, J. Opt, Soc, Am. A., 22, 928–937 (2005).