Fourier Analysis

and
Holographic Representations of 1D and 2D Signals

G.A. G IRALDI1 AND B.F. M OUTINHO1,2

AND
D.M.L. DE C ARVALHO3 AND J.C. DE O LIVEIRA1

1 LNCCNational Laboratory for Scientific Computing -

arXiv:cs.CV/0603116 v2 3 Apr 2006

Av. Getulio Vargas, 333, 25651-070, Petropolis, RJ, Brazil

{pssr,gilson}@lncc.br
2 Catholic University of Petropolis

Engineering Department - Petropolis, RJ, Brazil

brunofig@lncc.br
3 Military Institute of Engineering

Rio de Janeiro, RJ, Brazil

dulio@centroin.com.br

Abstract. In this paper, we focus on Fourier analysis and holographic transforms for
signal representation. For instance, in the case of image processing, the holographic rep-
resentation has the property that an arbitrary portion of the transformed image enables
reconstruction of the whole image with details missing. We focus on holographic repre-
sentation defined through the Fourier Transforms. Thus, We firstly review some results in
Fourier transform and Fourier series. Next, we review the Discrete Holographic Fourier
Transform (DHFT) for image representation. Then, we describe the contributions of our
work. We show a simple scheme for progressive transmission based on the DHFT. Next,
we propose the Continuous Holographic Fourier Transform (CHFT) and discuss some
theoretical aspects of it for 1D signals.

1 Introduction
The techniques for signal representation composes a class of methods used to characterize the information therein
[12, 16]. In the area of digital image processing, a classical approach for image representation is based on unitary
matrices that define the so called image transforms. The Discrete Fourier Transform (DFT) is the most known
example in this class.
Stochastic models describe an image as a member of an ensemble, which can be characterized by its mean
and covariance functions. This allows development of algorithms that are useful for an entire class or an ensemble
of images rather than for a single image. Fall in this category the covariance models, the 1 D models (an image
is considered a one-dimensional random signal that appears at the output of a raster scanner) and 2 D models
(Causal, Semicausal and Noncausal) [16].
Multiresolution representations, such Pyramid [17] and Wavelets [20], compose another important class in
this field. Finally, for image understanding models, an image can be considered as a composition of several objects
detected by segmentation and computer vision techniques [12, 16].
In this paper we focus on the first category of image representation methods. We firstly review some results
in Fourier transform and Fourier series and show some connections between them (section 2). Next, in section 3,
we describe the discrete Holographic Fourier Transform (DHFT) proposed in [5], which is based on the DFT. The
holographic representation of an image has the property that an arbitrary portion of the holographic representation
enables reconstruction of the whole image with details missing. This is an interesting property that motivates
works using holographic representations for image compression [2], multimedia systems [11], watermarking (see
[5] and references therein) and representations of line images [13]. Besides the holographic representation based
on the DFT there are also approaches that use subsampling methods [4].
In sections 4 and 5, we show our contributions in this field. We propose the Continuous Holographic Fourier
Transform (section 4) and discuss some theoretical and numerical aspect of it. Then, in section 5, we develop a
simple scheme for progressive transmission of images through the DHFT. Section 6 presents our conclusions and
some comments. The Appendix reviews interesting results in discrete Fourier analysis. An extended version of
this work can be found in [1]

2 Fourier Analysis Revised

In this section we review some important results about continuous Fourier Analysis. A special attention will be
given to the connections between the Fourier Transform and Fourier Series representation of a signal. Thus, let
us begin with usual definitions in this area. For simplicity, we restrict our discussion to the one dimensional
case. Generalizations to higher-dimensional cases are straightforward. In this section our presentation follows the
reference [7].
Throughout this text, all functions f are defined on the real line, with values in the complex domain C,
f : C. Besides, they are assumed to be piecewise continuous and having the integral:
Z
|f (x)|p dx, (1)

finite. The set of such functions is denoted by LP (). In this set, the p-norm and the inner product are well
defined by:
Z 1/p
p
kf kp = |f (x)| dx , 1 p < , (2)

Z
hf, gi = f (x) g (x)dx, f, g L2 () . (3)

Definition 1: Given a function f L1 () , its Fourier Transform is defined by:

Z
b
f () = (f ) () = f (x) exp (2jx) dx. (4)

Definition 2: Given the Fourier transform fb, the transformed function f can be obtained by the Inverse Fourier
Transform, given as follows:
  Z
f (x) = 1 fb (x) = fb() exp (2jx) d. (5)

Definition 3: Let f and g be functions in L1 (). Then the convolution of f and g is also an L1 () function
h which is defined by:
Z
h (x) = (f g) (x) = f (x y) g (y) dy. (6)

Theorem 1: Any function f L2 (L/2, L/2) has a Fourier series representation given by:
+ 
X n
f (x) = an exp 2jx . (7)
n=
L

Dem: See [9].

Among the properties of the Fourier Transform, the following ones have a special place for signal processing
techniques, according to our discussion presented in [1].
 Property 1: If the derivative df /dx of f exists and is in L1 (), then:
  
df
() = fb () = j fb. (8)
dx
Property 2: fb 0, as or .
For the proofs, see [7], pp. 25.
Property 3 (Convolution Theorem): The Fourier transform of the convolution of two functions is the product
of their Fourier transforms, that is:

h (x) = (f g) (x) b
h () = fb() gb () . (9)
Dem: See [9].
An important aspect for signal (and image) processing is the relationship between the coefficients an in the se-
ries (7) and the Fourier transform defined by the integral (4), in the case of a function f 2
 L (L/2,nL/2).

In or-

der to find this relationship, let us remember that, from the orthogonality of the functions exp 2jx L , n Z ,
it is easy to show what the coefficients an can be obtained by the expression:
Z 
1 L/2
n
an = f (x) exp 2jx dx. (10)
L L/2 L
Now, we must observe that this expression is equivalent to:
Z
1 b n  1  n
f = f (x) exp 2jx dx, (11)
L L L L
once f L2 (L/2, L/2) .

Thus,

1 b n 
an = f . (12)
L L
This result shows the connections between Fourier series and the Fourier Transform. It is used in [1] to
discussion Fourier analysis in te context of signal processing.

3 Discrete Holographic Fourier Transform

Optical holography technology uses interference and diffraction of light to record and reproduce 3D information
of an optical wavefront [3, 19]. This technology has been applied for 3D displays [18] and data storage devices
[15]. From an arbitrary portion of an optical hologram that encodes a scene, a low quality version of the entire
scene can be reconstructed. The quality of the recovered signal depends only on the size of the hologram portion
used, but not of the place from where it was taken [8]. The holographic representation of images tries to mimic
such property through image transforms. In this section, we summarize the results presented in [5].
The Discrete Holographic Fourier Transform (DHFT) of an image I = I (x, y) is defined as [5]:
h i
H (u, v) = 1 I (x, y) e2jP (x,y) , (13)

where P (x, y) is a random phase image so that E [P (x, y) P (x, y)] = 0 for (x, y) 6= (x, y) and P (x, y) is a
random variable uniformly distributed over [1, 1] . By holographic representation, we mean that from an image
portion H c (u, v) cropped from anywhere in the complex image H (u, v) , we can, by 2D Fourier transformation,
get a version of I (x, y) so that the degradation is proportional to the size of H c (u, v). Clearly, if H (u, v) is
available we get back I (x, y) as the amplitude of its 2D Fourier transform, and the same image recovery process
will be applied over cropped parts of H (u, v). Certainly, we can take a portion of an unitary transform and
perform the same process. However the result will be strong dependent from the place from where the portion
was cropped.
 In order to demonstrate that the image transform defined in expression (13) has the holographic property, let
us follow the development found in [5] and consider the 1D discrete version of the above proposed representation
method. The one-dimensional DHFT is:
M
X 1
1 2
H (u) = I (k) ej2P (k) ej M uk , (14)
k=0
M
where {P (k)} is a set of independent and identically distributed random numbers uniformly distributed over
[0, 1]. We will represent the process of cropping a portion of H (u) by multiplying it with a window function
W (u), given by:

1, u [a, a + (L 1)] ,
W (u) = (15)
0, u
/ [a, a + (L 1)] ,
where a {0, 1, ..., M L} for simplicity. Now, we shall consider the question: what can be recovered from
H c (u) = H (u) W (u) by the one-dimensional Fourier transform. To answer this question we call IW (r) the
obtained signal. Thus:
M
X 1
1 2
Iw (r) = H (u) W (u) ej M ur =
u=0
M
M 1 M 1
!
X X
j2P (k) 1 j 2 uk 1 2
I (k) e e M W (u) ej M ur =
u=0 k=0
M M
M 1 M 1
!
1 X X 2
I (k) ej2P (k) ej M u(kr) W (u) =
M
u=0 k=0
M 1 M 1
!
1 X X j 2 u(rk)
e M W (u) I (k) ej2P (k) . (16)
M u=0
k=0
Let us define g (r, k) as follows:

M 1 a+(L1)
X 2 X 2
g (r, k) = ej M u(rk) W (u) = ej M u(rk) . (17)
u=0 u=a

Thus, by changing variable ( = u a), we find:

L1
X
2 2
g (r, k) = ej M (rk)a ej M (rk) . (18)
=0
But, from the Property A2, in the Appendix:
L1
X sincLx
ej2xu = Lej(L1)x . (19)
u=0
sincx
Through expressions (19) and (18) we can rewrite equation (16) as:
M 1 L(rk)
1 X j 2 (rk)a j(L1) (rk) Lsinc M
IW (r) = e M e M
(rk)
I (k) ej2P (k) .
M sinc
k=0 M
Thus, we finally have:
M
Lsinc L(rk)
X 1
2 2
Iw (r) = ej M ra ej M ka ej(L1) M r ej(L1) M k M
I (k) ej2P (k) =
k=0 M sinc (rk)
M
 h i M 1 L(rk)
!
j 2
(L1)
a+ 2 r X Lsinc
=e M
I (k) e j2 [P (k)+ M
k
(a+ L1
2 )] M
=
(rk)
k=0 M sinc M
h i M 1 L(rk)
!
X
j 2 a+ (L1) r j2 Pa (k) LsincM
=e M 2
I (k) e (rk)
. (20)
k=0 M sinc M
Let us define:

Lsinc L(rk)
M
L (r k) := . (21)
M sinc (rk)
M
Thus, we have:
h i M 1
!
j 2 a+ (L1) X
r j2 Pa (k)
IW (r) = e M 2
I (k) e L (r k) . (22)
k=0

Firstly, we should verify the result when L = M , and consequently, a = 0. In this case, expression (20)
becomes:
M 1 h i
X j2 P (k)+ k(M
j 2 M 1
M( 2 )
1)
r
IW (r) = e I (k) e 2M
L (r k) . (23)
k=0

Besides, from the definition of L we can see that, when L = M , we have:


1, r k = 0,
L (r k) = (24)
0, r k 6= 0,
because 0 r, k M 1. Finally, by using the result (24) in expression (23) we find:
M 1 h i
2 M 1 X k(M 1)
j2 P (k)+ 2M
IW (r) = ej M ( 2 )r I (k) e L (r k) =
k=0
h i
r(M 1)
j 2 M 1
M( 2 )
r j2 P (r)+ 2M
e I (r) e =

2 M +1 M +1
= ej M ( 2 )r ej 2
M ( 2 ) I (r) e
r 2P (r)
=

= I (r) ej2P (r) . (25)

So, as expected, the whole image is recovered as the amplitude of the above result. In order to know how good
will be the approximation of the original image generated when L < M , we can perform a statistical analysis of
the recovered complex image.
Following [5], we will assume that I (k) = I0 constant:
h i M 1
!
j 2 a+ (L1)
r X L1
I (k) ej2[P (k)+ M (a+ 2 )] (r k)
k
I (r) = e M
W
2
L
k=0

M 1 h i
X (kr)
j2 P (k)+ M (a+ L1
2 )
= I0 e L (r k) =
k=0

M
X 1

= I0 ej2Pa (kr) L (r k) .
k=0
 Henceforth:

|IW (r)|2 = IW (r) IW (r) =

M 1
! M 1
!
X X
2 j2 Pa (kr) j2 Pa (lr)
= |I0 | e L (r k) e L (r l) =
k=0 l=0

MX1 XM
X 1

= |I0 |2 2L (r k) + ej2Pa (kr)L (r k) ej2Pa (lr) L (r l) =

k=0 k6=l k,l=0

MX1 XM
X 1

= |I0 |2 2L (r k) + ej2[Pa (kr)Pa (lr)] L (r k) L (r l) .

k=0 k6=l k,l=0

To answer how good an estimate is |IW (r)| for the value I0 , it is used in [5] the following result:
Lemma 1: Let 0 , 1 , ..., M 1 be independent and identically distributed random numbers uniformly chosen
from [0, 1], and the random variable:
M
X 1
V = ej2k (k) ,
k=0

where (k) is a sequence of real numbers. Thus, we have the following statistics for V :

 
M 1
X
E (V ) = 0, E V 2  = 2 (k) , (26)
k=0
v
u
 2 
uX M
X 1

V  =t 2 (k) 2 (l) . (27)
k6=l k,l=0

MP
1
As IW (r) is of the form ej2k (k), it follows from this Lemma that:
k=0
h i M
X 1
E |IW (r)|2 = |I0 |2 2L (r k) ,
k=0
v
h i u
2
uX M
2t
X 1
|IW (r)| = |I0 | 2L (r k) 2L (r l).
k6=l k,l=0

If L = M , expression (24) implies that:

h i h i
E |IW (r)|2 = |I0 |2 , |IW (r)|2 = 0,

as expected from the result (25).

Supposing L < M , it follows:

h i M
X 1
E |IW (r)|2 = |I0 |2 2L (r k) =
k=0

M
L2 sinc2 L(rk)
X 1
= |I0 |2 M
=
k=0 M 2 sinc2 (rk)
M
 M
sinc2 L(rk)
X 1
L2 2 M
= |I0 | . (28)
M2 sinc2 (rk)
k=0 M
Simplifying the summation:
M
sinc2 L(rk)
X 1
M M
= ,
k=0 sinc2 (rk)
M
L

so, returning to expression (28) we find:

h i M 1
2 L2 2
X sinc2 L(rk) L
E |IW (r)| = 2 I0 M
(rk)
= I0 2 ,
M sinc2 M
k=0 M
h i L
E |IW (r)|2 = I0 2 .
M
q
L
So the image recovered is an approximation of the original image multiplied by the factor M .
In the previous demonstrations we recovered the original image from a window in the holographic domain
extended by zeros. We consider now recovering the image from the window without extending it. The Fourier
transform of the cropped sequence H(u), u [a, a + (L 1)], is:
L1
X 1 2
IT (r) = F T [H (u)] = H (k + a) ej L rk . (29)
k=0
L

Defining the extended sequence H(u), u [0, M 1] and its Fourier transform as below:

H (k + a) , k [0, L 1] ,
H(k) =
0, k [L, M L] ,

h i MX1
1 2
IW S (v) = F T H(u) = H(u) ej M vu =
u=0
M
L1
X 1 2
= H (k + a) ej M vk . (30)
k=0
M
Comparing the two expressions (29) and (30) it can be straightforward noticed that:
r  
L L L 2L (L 1) L
IW S (v) = IT v , v = 0, , , ..., . (31)
M M M M M
We must also observe that:
a+L1
X 1 2
Iw (r) = H (u) ej M ur . (32)
u=a M
Hence, by changing k = u a, we have
L1
X 1 2
Iw (r) = H (k + a) ej M (k+a)r =
k=0
M
L1
!
X 1 2 2 2
= H (k + a) ej M kr ej M ar = IW S (r) ej M ar . (33)
k=0
M
 2
IW S (r) = IW (r) ej M ar . (34)
Substituting this result in (31), we have:
  r
L M 2
IT v = IW (v) ej M av . (35)
M L
L
Considering r = M v, follows:
r  
M M 2
IT (r) = IW r ej L ar . (36)
L L
Since the image is the absolute value of the holographic representation, it follows:
  r 
M  M 
|IT (r)| = IW r  . (37)
L  L
q 

Which proves that, being |IT (r)| a subsampling of M  M 
L IW L r , it is a good estimate of I0 (multiplying
q
L
|IT (r)| by M ), in case of I (k) constant.

4 Continuous Holographic Fourier Transform

In this section we extend some developments of section 3 for continuous signals. Therefore, we start with a
definition for the Continuous Fourier Holographic Transform (CHFT):
Z
H () = f (x) exp (2jx) exp (2jP (x)) dx, (38)

where P (x) is a random phase. Following the development of section 3, we will also consider a window in
the Fourier domain and check the quality of the recovered signal. Before this, we shall arrange the terms of the
integral in the following form:
Z
H () = exp (2jx) [f (x) exp (2jP (x))]dx. (39)

We can think about the Fourier transform of the signal h (x) = f (x) exp (2jP (x)) , which is well defined
if f L2 . Now, we use a window function (low-pass filter) given by expression:

1, k k,
W () = (40)
0, otherwise.
The filtering of signal h by W can be represented, in the Fourier domain by:

G () = W () H () , (41)
where G () is the Fourier transform of the output signal and H is the Fourier transform of h, that is, the CHFT
just defined in expression (38). The convolution theorem applied to expression (41) renders:
Z
g (x) = h (y) w (x y) dy, (42)

which is equivalent to:

Z
g (x) = f (y) exp (2jP (y)) w (x y) dy. (43)

 A simple numerical scheme to compute expression (43) offers an interesting application of Lemma 1. In fact,
if we approximate that integral by the Riemann sum:

N
X
g (x) [f (yn ) w (x yn ) y] exp (2jP (yn )) , (44)
n=0

we can , from a straightforward application of Lemma 1, estimate the f (x) through:

N
!1/2
X 2
f (x) [f (yn ) w (x yn ) y] . (45)
n=0

This expression is used in [1] to check the quality of the recovered signal.

4.1 CHFT Analysis

The first point to be demonstrated is the fact that we can recover the input signal f (x) if we take the inverse
Fourier transform of the CHT given b y:
Z
1
F (H ()) = H () exp (2jx) d. (46)

If we substitute the CHT H () by expression (38) we obtain:

Z Z 
1
F (H ()) = f (y) exp (2jy) exp (2jP (y)) dy exp (2jx) d. (47)



Using development of [9] we shall introduce the kernels exp 2 /n2 in order to deal with convergence
problems for this integral. Therefore, we get:

Z Z   
2
fn (x) = f (y) exp (2jy) exp (2jP (y)) dy exp (2jx) exp 2 d. (48)
n

Now, we can change the order of integration and re-write this expression as:
Z Z   
2
fn (x) = f (y) exp (2jP (y)) exp [2j (x y)] exp 2 d dy. (49)
n
The inner integral has analytical solution named by kn (x y) :
Z  
2
kn (x y) = exp [2j (x y)] exp 2 d, (50)
n
which gives:

Z  
2
kn (x y) = cos[2j(x y)] exp 2 d + (51)
n
Z  
2
j sin[2j(x y)] exp 2 d. (52)
n

From the fact that sine is an odd function and exp is an even one, it is easy to shown that the imaginary part
will be zero and the real part of the last expression can be computed as:
Z  
2
kn (x y) = 2 cos[2(x y)] exp 2 d. (53)
0 n
 This integral has analytical solution that can be obtained by expression [10]:
Z r  2
2

1 b
cos bx exp ax dx = exp . (54)
0 2 a 4a
Henceforth, the application of this expression gives:
" #
[2 (x y)]2 n2
kn (x y) = n exp . (55)
4
Therefore, returning to equation (49) we obtain:
Z
fn (x) = f (y) exp (2jP (y)) kn (x y) dy. (56)

It can be proved that (see [9] for details):

lim fn (x) = f (x) exp (2jP (x)) . (57)

n

But, we have also:

Z  
2
fn (x) = H () exp (2jx) exp 2 d. (58)
n
It is also possible to prove that:
Z   Z
2
lim H () exp (2jx) exp 2 d = H () exp (2jx) d. (59)
n n

From (57) and (59) it follows that:

Z
f (x) exp (2jP (x)) = H () exp (2jx) d, (60)

and so, as expected the whole signal is recovered as the amplitude of the inverse Fourier transform of the CHT.

5 Progressive Transmission
In this section we propose a progressive transmission approach based on the DHFT. This is performed by dividing
the image (in the holographic domain) in several equally sized portions. Then, we take the recovered signal of a
portion and incrementally add those ones corresponding to the other pieces.
Formally, given W1 and W2 two window functions similar to expression (15), we have:
M
X 1
1 2
Iw1 (r) = H (u) W1 (u) ej M ur , (61)
u=0
M
M
X 1
1 2
Iw2 (r) = H (u) W2 (u) ej M ur , (62)
u=0
M
where H (u) is the DHFT defined in equation (13). Thus, the linearity of the Fourier transform allows to write:
M
X 1
1 2
Iw1 (r) + Iw2 (r) = H (u) (W1 (u) + W2 (u)) ej M ur , (63)
u=0
M
and so:

Iw1 (r) + Iw2 (r) = Iw1 +w2 (r) . (64)

 Observe that if W1 , W2 are disjoint windows, W1 + W2 is another window function. Consequently, there is
not changes in the spectrum of the signal inside the window.
Thus, following expression (61)-(62), each recovered portion (Iw1 (r) and Iw2 (r) above) can be considered
as a packet to be transmitted, whose arrival order does not matter.

6 Conclusions
In this paper we discuss some results concerning to Fourier analysis and Holographic representation of images.
The approach described in section 5 must be analyzed from the viewpoint of progressive transmission ap-
proaches [14, 6]. Certainly, the need of double precision representation and the known instability of Fourier
analysis [7] may be several limitations for such method, if compared with other ones [6, 14]. Probably, the main
areas of application for the Holographic transform would be in distributed words and watermarking methods. Be-
sides a more complete theoretical analysis for CHFT must offered in order to quantify how good the estimate in
expression (43) is for the input signal f .

References
[1] G.A. Giraldi, B.F. Moutinho D.M.L. de Carvalho, and J.C. de Oliveira. Fourier analysis
for linear filtering and holographic representations of 1d and 2d signals. Technical report,
National Laboratory for Scientific Computing, Department of Compututer Science.Web address:
http://virtual01.lncc.br/giraldi/TechReport/holo-image.pdf.

[2] On holographic transform compression of images. In ICPR 00: Proceedings of the International Conference
on Pattern Recognition (ICPR00)-Volume 3, page 3236, Washington, DC, USA, 2000. IEEE Computer
Society.

[3] Tibor Balogh, Tams Forgcs, Tibor Agocs, Olivier Balet, Eric Bouvier, Fabio Bettio, Enrico Gobbetti, and
Gianluigi Zanetti. A scalable hardware and software system for the holographic display of interactive graph-
ics applications. In EUROGRAPHICS 2005 Short Papers Proceedings, Conference Held in Dublin, Ireland,
August 2005, 2005.

[4] A. M. Bruckstein, R. J. Holt, and A. N. Npngetravali. Holographic image representations: the subsampling
method. In ICIP 97: Proceedings of the 1997 International Conference on Image Processing (ICIP 97)
3-Volume Set-Volume 1, page 177, Washington, DC, USA, 1997. IEEE Computer Society.

[5] A.M. Bruckstein, R.J. Holt, and A.N. Netravali. Holographic representation of images. IEEE Transactions
on Image Processing, 7(11):15831597, 1998.

[6] Chin-Chen Chang and Yi-Long Liu. Progressive image transmission using singular value decomposition.
EUROGRAPHICS Multimedia Workshop, 2004.

[7] C.K. Chui. An Introduction to Wavelets. New York Academic Press, 1992.

[8] Maurice Francon. Optical Image Formation And Processing. Academic Press Inc., New York, New York,
1979.

[9] Djairo G. de Figueiredo. Analise de Fourier e Equacoes Diferenciais Parciais. Edgard Blcher, 1977.

[10] Murray R. Spiegel e John Liu. Colecao Schaum Manual de Formulas e Tabelas Matematicas. Bookman,
2004.png

[11] L. C. Ferri. Holographic images in multimedia information systems. Technical report, Fraunhofer Institute
for Computer Graphics, www.inigraphics.net/ini-ccc/projects/holographie/paper1.pdf, 1996.

[12] R.C. Gonzalez. Digital Image Processing. Reading Addison-Wesley, 1992.

[13] Nickolas D. Gorsky. Experiments with handwriting recognition using holographic representation of line
images. Pattern Recognition Letters, 15(9):853859, 1994.

[14] See References in:. IEEE Transactions on Image Processing, 8(1), 1999.

[15] G. W. Burr H. Coufal H. Guenther J. A. Hoffnagle C. M. Jefferson B. Marcus R. M. Macfarlane R. M. Shelby

J. Ashley, M.-P. Bernal and G. T. Sincerbox. Holographic data storage. IBM Jornal of Research and Devel-
opment, 44(3):4653, 2000.

[16] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989.

[17] J. M. Jolion and A. Montanvert. The adaptive pyramid: A framework for 2d image analysis. CVGIP: Image
Understanding, 55(3):339348, May 1992.

[18] L. B. Lesem, P. M. Hirsch, and Jr. J. A. Jordan. Scientific applications: Comppnguter synthesis of holograms
for 3-d display. Commun. ACM, 11(10):661674, 1968.

[19] Ting-Chung Poon. Digital Holography and Three-Dimensional Display: Principles and Applications.
Springer, 2006.

[20] Y. Zeevi and R. Coifman. Signal and Image Representation in Combined Spaces (Wavelet Analysis and Its
Applications). Academic Press, 1998. png

png

7 Appendix
In this section we review interesting properties in discrete Fourier analysis.
Property A1: The circular convolution theorem states that the DFT of the circular convolution of two se-
quences is equal to the product of their DFTs, that is, if:
N
X 1
x2 (n) = h (n k)c x1 (k) , 0 n 1, (65)
k=0

where h (n k)c is the a circular kernel (see [16]) then:

DF T [x2 (n)] = DF T [h (n)] DF T [x1 (n)] . (66)

Expression (66) can be written as:
N
X 1 N
X 1 N
X 1
2 2 2
x2 (k) ej N vk = h(k)ej N vk x1 (k) ej N vk .
k=0 k=0 k=0

If we divide this expression by N we get:
N
X 1 N
X 1 N
X 1
1 2 1 2 2
x2 (k) ej N vk = h(k) ej N vk x1 (k) ej N vk .
k=0
N k=0
N k=0

Therefore, we get the same relation for the Unitary Discrete Fourier Transform:

U DF T [x2 (n)] = U DF T [h (n)] DF T [x1 (n)] . (67)

Now, if we define

png1, k = a,
x1 (k) =
0, k 6= a,
we have that, from expression (65), x2 (n) = h (n a)c , i.e., h (n) is a circular shift of x2 (n). Hence
N
X 1
2
U DF T [x2 (n)] = U DF T [h (n)] x1 (k) ej N vk ,
k=0

which implies
2
U DF T [x2 (n)] = U DF T [h (n)] ej N va .png
Property A2. Let us consider the summation:
L1
X L1
X
u
ej2xu = ej2x . (68)
u=0 u=0

By remembering the geometric series:

L1
X
S= au = 1 + a + a2 + + aL1 ,
u=0

we know that:
1 aL
S= .
1a
If we apply this result in expression (68) we get:
L1
L

X
j2xu 1 ej2x 1 ej2xL ejxL ejxL ejxL
e = = = , (69)
1 ej2x 1 ej2x ejx (ejx ejx )
u=0

But we know that:

ejx ejx
sin (x) = ,
2j

ejxL ejxL
sin (Lx) = .
2j
Therefore, expression (69) becomes:
L1
X ejxL 2j sin (Lx) sin (Lx)
ej2xu = = ej(L1)x . (70)
ej 2j sin (x) sin (x)
u=0

Hence, if x = 0 we get:
L1
X
ej2xu = L.
u=0
sin(x)
We can also re-write expression (70) by using the sinc function, which is defined by: sinc (x) = x .
Thus, that expression becomes:
L1
X Lx sinc (Lx) sinc (Lx)
ej2xu = ej(L1)x = Lej(L1)x .
u=0
x sinc (x) sinc (x)
 Fourier Analysis for Linear Filtering
and
Holographic Representations of 1D and 2D Signals

G.A. G IRALDI1 AND B.F. M OUTINHO 1,2

AND
D.M.L. DE C ARVALHO 3 AND J.C. DE O LIVEIRA 1

1 LNCCNational Laboratory for Scientific Computing -

Av. Getulio Vargas, 333, 25651-070, Petropolis, RJ, Brazil
{pssr,gilson}@lncc.br
2 Catholic University of Petropolis

Engineering Department - Petropolis, RJ, Brazil

brunofig@lncc.br
3 Military Institute of Engineering

Rio de Janeiro, RJ, Brazil

dulio@centroin.com.br

Abstract. In this paper, we focus on Fourier analysis and holographic transforms for
signal representation. For instance, in the case of image processing, the holographic rep-
resentation has the property that an arbitrary portion of the transformed image enables
reconstruction of the whole image with details missing. We focus on holographic repre-
sentation defined through the Fourier Transforms. Thus, we firstly consider the Fourier
analysis in the context of signal processing. Next, we review the Discrete Holographic
Fourier Transform (DHFT) for image representation. Then, we describe the contributions
of our work. We show a simple scheme for progressive transmission based on the DHFT.
Next, we propose the Continuous Holographic Fourier Transform (CHFT) and discuss
some theoretical aspects of it for 1D signals. Finally, some testes are presented in the
experimental results.

1 Introduction
The techniques for signal representation composes a class of methods used to characterize the information therein
[11, 15]. In the area of digital image processing, a classical approach for image representation is based on unitary
matrices that define the so called image transforms. The Discrete Fourier Transform (DFT) is the most known
example in this class.
Stochastic models describe an image as a member of an ensemble, which can be characterized by its mean
and covariance functions. This allows development of algorithms that are useful for an entire class or an ensemble
of images rather than for a single image. Fall in this category the covariance models, the 1 D models (an image
is considered a one-dimensional random signal that appears at the output of a raster scanner) and 2 D models
(Causal, Semicausal and Noncausal) [15].
Multiresolution representations, such Pyramid [16] and Wavelets [19], compose another important class in
this field. Finally, for image understanding models, an image can be considered as a composition of several objects
detected by segmentation and computer vision techniques [11, 15].
In this paper we focus on the first category of image representation methods. We firstly review some results
in Fourier transform and Fourier series and show some connections between them (section 2). Next, in section 4,
we describe the discrete Holographic Fourier Transform (DHFT) proposed in [4], which is based on the DFT. The
holographic representation of an image has the property that an arbitrary portion of the holographic representation
enables reconstruction of the whole image with details missing. This is an interesting property that motivates
works using holographic representations for image compression [1], multimedia systems [10], watermarking (see
[4] and references therein) and representations of line images [12]. Besides the holographic representation based
on the DFT there are also approaches that use subsampling methods [3].
In sections 5 and 6, we show our contributions in this field. We propose the Continuous Holographic Fourier
Transform (section 5) and discuss some theoretical and numerical aspect of it. Then, in section 6, we develop a
simple scheme for progressive transmission of images through the DHFT. Finally, some testes with the progressive
transmission and the CHFT are presented on section 7. Section 8 presents our conclusions and some comments.
The Appendix reviews interesting results in discrete Fourier analysis.

2 Fourier Analysis Revised

In this section we review some important results about continuous Fourier Analysis. A special attention will be
given to the connections between the Fourier Transform and Fourier Series representation of a signal. Thus, let
us begin with usual definitions in this area. For simplicity, we restrict our discussion to the one dimensional
case. Generalizations to higher-dimensional cases are straightforward. In this section our presentation follows the
reference [6].
Throughout this text, all functions f are defined on the real line, with values in the complex domain C,
f : < C. Besides, they are assumed to be piecewise continuous and having the integral:
Z
|f (x)|p dx, (1)

finite. The set of such functions is denoted by L P (<). In this set, the p-norm and the inner product are well
defined by:
Z 1/p
p
kf kp = |f (x)| dx , 1 p < , (2)

Z
hf, gi = f (x) g (x)dx, f, g L2 (<) . (3)

Definition 1: Given a function f L1 (<) , its Fourier Transform is defined by:

Z
fb() = (zf ) () = f (x) exp (2jx) dx. (4)

Definition 2: Given the Fourier transform fb, the transformed function f can be obtained by the Inverse Fourier
Transform, given as follows:
  Z
f (x) = z1 fb (x) = fb() exp (2jx) d. (5)

Definition 3: Let f and g be functions in L 1 (<). Then the convolution of f and g is also an L 1 (<) function
h which is defined by:
Z
h (x) = (f g) (x) = f (x y) g (y) dy. (6)

Theorem 1: Any function f L2 (L/2, L/2) has a Fourier series representation given by:
+
X  n
f (x) = an exp 2jx . (7)
n=
L

Dem: See [8].

Among the properties of the Fourier Transform, the following ones have a special place for signal processing
techniques as we shall see later (section 3).
Property 1: If the derivative df /dx of f exists and is in L 1 (<), then:
   
df
z () = fb0 () = j fb. (8)
dx

Property 2: fb 0, as or .
For the proofs, see [6], pp. 25.
Property 3 (Convolution Theorem): The Fourier transform of the convolution of two functions is the product
of their Fourier transforms, that is:

h (x) = (f g) (x) b
h () = fb() gb () . (9)
Dem: See [8].
An important aspect for signal (and image) processing is the relationship between the coefficients a n in the se-
ries (7) and the Fourier transform defined by the integral (4), in the case of a function f 2
 L (L/2,nL/2).

In or-

der to find this relationship, let us remember that, from the orthogonality of the functions exp 2jx L , n Z ,
it is easy to show what the coefficients a n can be obtained by the expression:
Z 
1 L/2
n
an = f (x) exp 2jx dx. (10)
L L/2 L
Now, we must observe that this expression is equivalent to:
Z
1 b n  1  n
f = f (x) exp 2jx dx, (11)
L L L L
once f L2 (L/2, L/2) .

Thus,

1 b n 
an =
f . (12)
L L
This result shows the connections between Fourier series and the Fourier Transform. It will be used in the
following discussion.

3 Fourier Analysis and Filtering Methods

In this section we analyze usual linear filtering techniques named low-pass, high-pass and band-pass methods. In
the space domain, these methods can be mathematically described through convolutions between the input signal
f (x) and the filter kernel, named W (x) next. Therefore, from Property 3, these processes are represented by the
product of the Fourier transform of the kernel of the filter and the Fourier transform of the input signal. Such dual
viewpoint is very much useful for the analysis of linear filters. Besides, the expression (12) can be also explored
in such analysis as we shall see next.

3.1 Low-Pass Filtering

To simplify the explanation of the ideas, we are going to consider the following example.
Example 1: Let f L2 (L/2, L/2) defined as follows:

0, L/2 x < 0
f (x) = (13)
A, 0 x < L/2
The Fourier transform of this function is:
Z L/2
fb() = f (x) exp (2jx) dx =
L/2
 Z L/2
A
A exp (2jx) dx = (1 exp [jL]) .
0 2j
Henceforth:
   2
 b 2 A
f () = [2 2 cos (L)] . (14)
2j
But, we know that:

1 cos (y) = 2 sin2 (y/2) .

Thus replacing this expression in equation (14) we find:
   
 b 2 4A2 2 L
f () = sin = (15)
(2)2 2
      L 
 b   A   .
f () =   sin  (16)
2
We can verify Property 2 as well as observe that the indetermination 0/0 for = 0 can be easily resolved by
observing that, the LHopitals rule gives:
  AL
 
lim fb() =
0 2
The Figure 1 shows the graphical representation of the function f and its Fourier transform just computed in
(16).

(a) (b)

b expression (16).
Figure 1: (a) Step function f , expression (13). (b) Graphical representation of | f|,

Now, let us consider the Fourier series representation for f , given by expression (7), and the expression (12)
which links the coefficients an and the Fourier transform. If we truncate the Fourier series, that is, if we take the
new series:
+K
X  n
fe(x) = an exp 2jx , K > 0, (17)
L
n=K

than, from the viewpoint of the Fourier transform, the expression (12) shows that we are cutting off the frequencies
such that || > K/L. Following the usual nomenclature in signal processing, we say that we cut the highest
frequencies. The effect of such procedure can be better visualized through a numerical example. Therefore, we
take the function given by expression (13), but using L = 1 and A = 1. Figure 2 shows the f, e expression (17),
for some values of K.
We can observe that the function fe is a smoothed version of the original one. Thus, the process of cutting off
high frequencies in the Fourier domain () is equivalent to smooth the function in the space domain (x) .
 (a) K=3 (b) K=5

(c) K=10 (d) K=20

(e) K=100 (f) K=500

Figure 2: Function fe, expression (17), for some values of K and L = 1.

 In another way, we can analyze the same cut-off process from the viewpoint of the Fourier transform only.
To perform this task, let us consider the window function, in the Fourier domain, given by:

c () = 1, || K/L,
W (18)
0, otherwise.
To cut-off frequencies higher than K/L it is just a matter of taking the function:

gb () = fb() W
c () , (19)
c is a low-pass filter in the frequency domain with cut-off frequency given by = K/L. From the
that is, W
Convolution Theorem we can write that:

g (x) = (f W ) (x) , (20)

where the functions g, f, W are the inverse Fourier transforms of gb, fb and W
c , respectively. The function W can
be obtained by using the Inverse Fourier Transform defined by expression (5). A simple algebra shows that, if
L = 1, we get:

sin(2Kx)
W (x) = . (21)
x

Figure 3: Low-Pass filter kernel W defined in expression (21), for K = 8.

The Figure 3 pictures this kernel for K = 8. The Figure 4 shows the function g for some specific values of
K. We observe the same behavior of Figure 2, that is, the function g is a smoothed version of the function f .
The function W is the kernel of a low-pass filter, or, the impulse response of the filter [15]. In the above
analysis we went from the frequency domain to the space domain, that is, we firstly design a filter W c in the
frequency (or Fourier) domain. Then, we find out that filter in the space domain, through the Convolution Theorem
and the Inverse Fourier Transform.
Now, we are going to follow the opposite direction: A low-pass filter will be firstly designed in the space
domain. Then, we will analyze that filter in the frequency space. The most usual low-pass filter class is the
Gaussian one, given by:
 
1 x2
G (x) = exp , > 0. (22)
2 4
The following properties can be demonstrated (see [6], pp.28-29, for details):
Z +
G (x) dx = 1, (23)

 (a) K=3 (b) K=5

(c) K=8 (d) K=15

(e) K=30 (f) K=150

Figure 4: Function g, expression (20), for some values of K.



c () = exp 2 .
G (24)
The Figure 5 shows G , for = 1, 1/4 and 1/16.

Figure 5: Gaussian function G , expression (22), for some values of .

Now, we take the signal given by expression (13) and consider the effect of the convolution of this function
with the Gaussian kernel. Let us take L = 1, A = 1 in expression (13) and some specific values of as follows.
The Figure 6 shows the result.
Again, we observe that the result is a smoothed version of f . For consistency, we expect to observe a cut-off
process of high frequencies in the Fourier domain. In fact, the Convolution Theorem and the Gaussian Kernel
shape, given by expression (24), offer a straightforward way to confirm this behavior. In fact, once Gc () 0
for high values of and the same is true for the Fourier transform of any input signal f L it is clear that high
2

frequencies will be cut off in the filtered signal gb, given by gb () = fb() G
c () .

3.2 High-Pass Filtering

Now, our aim is to design a convolution kernel that cuts-off low frequencies. An interesting way to perform this
task is to take the Fourier Transform of the Gaussian Kernel, given by expression (24), and to apply the Property
1. Thus, we will obtain:
  
c dG c ()
0
G () = z () = j G (25)
dx


c0 () = j exp 2 .
G (26)

Observe that low frequencies (nearby = 0) are almost cut, due to the factor . However high frequencies
are enhanced due to the same factor. Thus, we find a high-pass filter.
To obtain its kernel in the space domain, we have just to take the derivative of (22):
 
dG x x2
(x) = exp , > 0. (27)
dx 4 3 4
 (a) =1/100 (b) =1/500

(c) =1/1000 (d) =1/5000

(e) =1/10000 (f) =1/50000

Figure 6: Convolution of step function, expression (13), and the Gaussian G , expression (22), for some specific
values of .
 Higher order derivatives of the Gaussian can be used also. For instance, by applying Property 1, we find that:

G c0 () = 2d
c00 () = j G G () , (28)

and, we have another high-pass filter. So, we get the following high-pass kernels in the space domain:
  2 
d2 G 1 x2 x
(x) = exp 1 , > 0, (29)
dx2 4 3 4 2

obtained through the derivative of expression (27) or the inverse Fourier transform of expression (28). Figures 7
shows the shape of the kernels given by expressions (27) and (29).

(a) (b)

Figure 7: (a) High-Pass filter given by expression (27) with = 1. (b) High-Pass filter given by expression (29)
with = 1.

What about the effect of high-pass filters in the space domain? The next example will be worthwhile to
understand such behavior. We take the function defined by the following expression:

0, L < x < 0,
f (x) = (30)
A, 0 x < L.

with L = 10 and take its convolution with the filter kernel given by expression (29) for = 1.0 and = 0.5. The
results are pictured on Figure (8).

(a) (b)

Figure 8: (a) Convolution of step function (equation (30), for L = 10) with high-pass filter defined by expression
(29) with = 1. (b) Convolution with the same filter but with = 0.5.
3.3 Band-Pass Filtering
As the name suggests, the frequency response in this case have relatively high gain in a band of frequencies,
allowing these frequencies to pass through the filter. Other frequencies suffer a relatively high loss or are rejected.
So, let us consider the following filter in the Fourier domain:

1, b a,
b () =
H 1, a b, (31)

0, otherwise.
This is a band-pass filter because the frequencies out of the range [b, a] [a, b] are cut off in the filtering
process. In order to obtain the filter kernel in the space domain, we just take the inverse Fourier transform of
expression (31):
Z
H (x) = Hb () exp (2jx) d, (32)

which can be obtained by computing the expression:

Z a Z b
H (x) = exp (2jx) d + exp (2jx) d. (33)
b a

A simple development generates:

sin(2bx) sin(2ax)
H (x) = . (34)
x
The Figure 9.a shows this function over the range [0.5, 0.5] for a = 1 and b = 10.

Figure 9: (a) Band-Pass filter defined by expression (34) with a = 1 and b = 10.

Also, Figure 10.a shows the band-pass filtered version of the signal defined by expression (30), for L = 1,
when using this kernel with a = 10 and b = 30. It is interesting to compare this result with the one obtained if we
truncate the Fourier series f by taking only the terms for which (n/L) [10, 30] [30, 10]. We observe that
both the results have the same behavior inside the interval [0.5, 0.5] but differs outside it due to the periodicity
of the Fourier series.

4 Discrete Holographic Fourier Transform

Optical holography technology uses interference and diffraction of light to record and reproduce 3D information
of an optical wavefront [2, 18]. This technology has been applied for 3D displays [17] and data storage devices
[14]. From an arbitrary portion of an optical hologram that encodes a scene, a low quality version of the entire
 (a) (b)

Figure 10: (a) Band-Pass filtered version of the step function defined by equation (30), for L = 1. The filter kernel
is given by expression (34) with a = 10 and b = 30. (b) Truncated Fourier series of step function defined by
equation (30), with L = 1, obtained by taking (n/L) [10, 30] [30, 10] in expression (7).

scene can be reconstructed. The quality of the recovered signal depends only on the size of the hologram portion
used, but not of the place from where it was taken [7]. The holographic representation of images tries to mimic
such property through image transforms. In this section, we summarize the results presented in [4].
The Discrete Holographic Fourier Transform (DHFT) of an image I = I (x, y) is defined as [4]:
h i
H (u, v) = z1 I (x, y) e2jP (x,y) , (35)

where P (x, y) is a random phase image so that E [P (x, y) P (x, y)] = 0 for (x, y) 6= (x, y) and P (x, y) is a
random variable uniformly distributed over [1, 1] . By holographic representation, we mean that from an image
portion H c (u, v) cropped from anywhere in the complex image H (u, v) , we can, by 2D Fourier transformation,
get a version of I (x, y) so that the degradation is proportional to the size of H c (u, v). Clearly, if H (u, v) is
available we get back I (x, y) as the amplitude of its 2D Fourier transform, and the same image recovery process
will be applied over cropped parts of H (u, v). Certainly, we can take a portion of an unitary transform and
perform the same process. However the result will be strong dependent from the place from where the portion
was cropped.
In order to demonstrate that the image transform defined in expression (35) has the holographic property, let
us follow the development found in [4] and consider the 1D discrete version of the above proposed representation
method. The one-dimensional DHFT is:

M
X 1
1 2
H (u) = I (k) ej2P (k) ej M uk , (36)
k=0
M

where {P (k)} is a set of independent and identically distributed random numbers uniformly distributed over
[0, 1]. We will represent the process of cropping a portion of H (u) by multiplying it with a window function
W (u), given by:

1, u [a, a + (L 1)] ,
W (u) = (37)
0, u
/ [a, a + (L 1)] ,

where a {0, 1, ..., M L} for simplicity. Now, we shall consider the question: what can be recovered from
H c (u) = H (u) W (u) by the one-dimensional Fourier transform. To answer this question we call I W (r) the
obtained signal. Thus:
 M
X 1
1 2
Iw (r) = H (u) W (u) ej M ur =
u=0
M
M 1 M 1
!
X X
j2P (k) 1 j 2 uk 1 2
I (k) e e M W (u) ej M ur =
u=0 k=0
M M
M 1 M 1
!
1 X X 2
I (k) ej2P (k) ej M u(kr) W (u) =
M
u=0 k=0
M 1 M 1
!
1 X X j 2 u(rk)
e M W (u) I (k) ej2P (k) . (38)
M u=0
k=0

Let us define g (r, k) as follows:

M 1 a+(L1)
X X
j 2 u(rk) 2
g (r, k) = e M W (u) = ej M u(rk) . (39)
u=0 u=a

Thus, by changing variable ( = u a), we find:

L1
X
j 2 (rk)a 2
g (r, k) = e M ej M (rk) . (40)
=0
But, from the Property A2, in the Appendix:
L1
X sincLx
ej2xu = Lej(L1)x . (41)
u=0
sincx
Through expressions (41) and (40) we can rewrite equation (38) as:
M 1 L(rk)
1 X j 2 (rk)a j(L1) (rk) Lsinc M
IW (r) = e M e M
(rk)
I (k) ej2P (k) .
M sinc
k=0 M
Thus, we finally have:
M 1
X 2 2 Lsinc L(rk)
Iw (r) = ej M ra ej M ka ej(L1) M r ej(L1) M k M
I (k) ej2P (k) =
k=0 M sinc (rk)
M
M 1 L(rk)
!
X
)] Lsinc M
h i
(L1)
j 2 a+ 2 r a+ L1
I (k) ej2[ (
k
P (k)+ M
=e M 2 =
k=0 M sinc (rk)
M
M
!
Lsinc L(rk)
h i
X 1
j 2 a+ (L1) r
=e M 2
I (k) ej2Pa (k) M
(rk)
. (42)
k=0 M sinc M
Let us define:

Lsinc L(rk)
M
L (r k) := . (43)
M sinc (rk)
M
Thus, we have:
M 1
!
j 2
h
a+ (L1)
i
X
r j2 Pa (k)
IW (r) = e M 2
I (k) e L (r k) . (44)
k=0
 Firstly, we should verify the result when L = M , and consequently, a = 0. In this case, expression (42)
becomes:
M
X 1 h
k(M 1)
i
j 2 M 1
M( 2 )
r j2 P (k)+ 2M
IW (r) = e I (k) e L (r k) . (45)
k=0

Besides, from the definition of L we can see that, when L = M , we have:


1, r k = 0,
L (r k) = (46)
0, r k 6= 0,
because 0 r, k M 1. Finally, by using the result (46) in expression (45) we find:
M
X 1 h
k(M 1)
i
j 2 M 1
M( 2 )
r j2 P (k)+ 2M
IW (r) = e I (k) e L (r k) =
k=0
h i
j 2 M 1 j2 P (r)+ r(M
M( 2 )
1)
r
e I (r) e 2M
=

2 M +1 M +1
= ej M ( 2 )r ej 2
M ( 2 ) I (r) e
r 2P (r)
=

= I (r) ej2P (r) . (47)

So, as expected, the whole image is recovered as the amplitude of the above result. In order to know how good
will be the approximation of the original image generated when L < M , we can perform a statistical analysis of
the recovered complex image.
Following [4], we will assume that I (k) = I 0 constant:
M 1
!
j 2
h
a+ (L1)
i
r X L1
I (k) ej2[P (k)+ M (a+ 2 )] (r k)
k
I (r) = e M
W
2
L
k=0

M
X 1 h i
j2 P (k)+ (kr)
M (
a+ L1
2 )
= I0 e L (r k) =
k=0

M
X 1

= I0 ej2Pa (kr) L (r k) .
k=0

Henceforth:

|IW (r)|2 = IW (r) IW (r) =

M 1
! M 1
!
X X
2 j2 Pa (kr) j2 Pa (lr)
= |I0 | e L (r k) e L (r l) =
k=0 l=0

MX1 XM
X 1

= |I0 |2 2L (r k) + ej2Pa (kr) L (r k) ej2Pa (lr) L (r l) =

k=0 k6=l k,l=0

MX1 XM
X 1

= |I0 |2 2L (r k) + ej2[Pa (kr)Pa (lr)] L (r k) L (r l) .

k=0 k6=l k,l=0

To answer how good an estimate is |IW (r)| for the value I0 , it is used in [4] the following result:
 Lemma 1: Let 0 , 1 , ..., M 1 be independent and identically distributed random numbers uniformly chosen
from [0, 1], and the random variable:
M
X 1
V = ej2k (k) ,
k=0

where (k) is a sequence of real numbers. Thus, we have the following statistics for V :

 2 
MX1
E (V ) = 0, 
E V  = 2 (k) , (48)
k=0
v
u
 2 
uX M
X 1

V  =t 2 (k) 2 (l) . (49)
k6=l k,l=0

MP
1
As IW (r) is of the form ej2k (k), it follows from this Lemma that:
k=0
h i M
X 1
2 2
E |IW (r)| = |I0 | 2L (r k) ,
k=0
v
h i u
2
uX M
2t
X 1
|IW (r)| = |I0 | 2L (r k) 2L (r l).
k6=l k,l=0

If L = M , expression (46) implies that:

h i h i
E |IW (r)|2 = |I0 |2 , |IW (r)|2 = 0,
as expected from the result (47).
Supposing L < M , it follows:

h i M
X 1
E |IW (r)|2 = |I0 |2 2L (r k) =
k=0
M
L2 sinc2 L(rk)
X 1
= |I0 |2 M
=
k=0 M 2 sinc2 (rk)
M
M
sinc2 L(rk)
X 1
L2 2 M
= 2 |I0 | (rk)
. (50)
M 2
k=0 sinc M
Simplifying the summation:
M
sinc2 L(rk)
X 1
M M
= ,
k=0 sinc2 (rk)
M
L
so, returning to expression (50) we find:

h i L2
M
X 1
sinc2 L(rk) L
E |IW (r)|2 = 2 I0 2 M
(rk)
= I0 2 ,
M sinc 2 M
k=0 M
h i L
E |IW (r)|2 = I0 2 .
M
q
L
So the image recovered is an approximation of the original image multiplied by the factor M.
 In the previous demonstrations we recovered the original image from a window in the holographic domain
extended by zeros. We consider now recovering the image from the window without extending it. The Fourier
transform of the cropped sequence H(u), u [a, a + (L 1)], is:
L1
X 1 2
IT (r) = F T [H (u)] = H (k + a) ej L rk . (51)
k=0
L

Defining the extended sequence H(u), u [0, M 1] and its Fourier transform as below:

H (k + a) , k [0, L 1] ,
H(k) =
0, k [L, M L] ,
h i MX1
1 2
IW S (v) = F T H(u) = H(u) ej M vu =
u=0
M
L1
X 1 2
= H (k + a) ej M vk . (52)
k=0
M
Comparing the two expressions (51) and (52) it can be straightforward noticed that:
r  
L L L 2L (L 1) L
IW S (v) = IT v , v = 0, , , ..., . (53)
M M M M M
We must also observe that:
a+L1
X 1 2
Iw (r) = H (u) ej M ur . (54)
u=a M
Hence, by changing k = u a, we have
L1
X 1 2
Iw (r) = H (k + a) ej M (k+a)r =
k=0
M
L1
!
X 1 2 2 2
= H (k + a) ej M kr ej M ar = IW S (r) ej M ar . (55)
k=0
M
2
IW S (r) = IW (r) ej M ar . (56)
Substituting this result in (53), we have:
  r
L M 2
IT v = IW (v) ej M av . (57)
M L
L
Considering r = M v, follows:
r  
M M 2
IT (r) = IW r ej L ar . (58)
L L
Since the image is the absolute value of the holographic representation, it follows:
r   
M  M 
|IT (r)| = IW r  . (59)
L  L
q 

Which proves that, being |IT (r)| a subsampling of M  M 
L IW L r , it is a good estimate of I0 (multiplying
q
L
|IT (r)| by M ), in case of I (k) constant.
5 Continuous Holographic Fourier Transform
In this section we extend some developments of section 4 for continuous signals. Therefore, we start with a
definition for the Continuous Fourier Holographic Transform (CHFT):
Z
H () = f (x) exp (2jx) exp (2jP (x)) dx, (60)

where P (x) is a random phase. Following the development of section 4, we will also consider a window in
the Fourier domain and check the quality of the recovered signal. Before this, we shall arrange the terms of the
integral in the following form:
Z
H () = exp (2jx) [f (x) exp (2jP (x))]dx. (61)

We can think about the Fourier transform of the signal h (x) = f (x) exp (2jP (x)) , which is well defined
if f L2 . Now, we use a window function (low-pass filter) of the type of expression (18):

1, k k,
W () = (62)
0, otherwise.
The filtering of signal h by W can be represented, in the Fourier domain by:

G () = W () H () , (63)
where G () is the Fourier transform of the output signal and H is the Fourier transform of h, that is, the CHFT
just defined in expression (60). The convolution theorem applied to expression (63) renders:
Z
g (x) = h (y) w (x y) dy, (64)

which is equivalent to:

Z
g (x) = f (y) exp (2jP (y)) w (x y) dy. (65)

A simple numerical scheme to compute expression (65) offers an interesting application of Lemma 1. In fact,
if we approximate that integral by the Riemann sum:
N
X
g (x) [f (yn ) w (x yn ) y] exp (2jP (yn )) , (66)
n=0

we can , from a straightforward application of Lemma 1, estimate the f (x) through:

N
!1/2
X 2
f (x) [f (yn ) w (x yn ) y] . (67)
n=0

This expression will be used in the experimental results check the quality of the recovered signal.

5.1 CHFT Analysis

The first point to be demonstrated is the fact that we can recover the input signal f (x) if we take the inverse
Fourier transform of the CHT given b y:
Z
F 1 (H ()) = H () exp (2jx) d. (68)

If we substitute the CHT H () by expression (60) we obtain:

 Z Z 
1
F (H ()) = f (y) exp (2jy) exp (2jP (y)) dy exp (2jx) d. (69)



Using development of [8] we shall introduce the kernels exp 2 /n2 in order to deal with convergence
problems for this integral. Therefore, we get:

Z Z   
2
fn (x) = f (y) exp (2jy) exp (2jP (y)) dy exp (2jx) exp 2 d. (70)
n

Now, we can change the order of integration and re-write this expression as:
Z Z   
2
fn (x) = f (y) exp (2jP (y)) exp [2j (x y)] exp 2 d dy. (71)
n
The inner integral has analytical solution named by k n (x y) :
Z  

2
kn (x y) = exp [2j (x y)] exp 2 d, (72)
n
which gives:

Z  
2
kn (x y) = cos[2j(x y)] exp 2 d + (73)
n
Z  
2
j sin[2j(x y)] exp 2 d. (74)
n

From the fact that sine is an odd function and exp is an even one, it is easy to shown that the imaginary part
will be zero and the real part of the last expression can be computed as:
Z  
2
kn (x y) = 2 cos[2(x y)] exp 2 d. (75)
0 n
This integral has analytical solution that can be obtained by expression [9]:
Z r  2
2

1 b
cos bx exp ax dx = exp . (76)
0 2 a 4a
Henceforth, the application of this expression gives:
" #
[2 (x y)]2 n2
kn (x y) = n exp . (77)
4
Therefore, returning to equation (71) we obtain:
Z
fn (x) = f (y) exp (2jP (y)) kn (x y) dy. (78)

It can be proved that (see [8] for details):

lim fn (x) = f (x) exp (2jP (x)) . (79)

n

But, we have also:

Z  
2
fn (x) = H () exp (2jx) exp 2 d. (80)
n
 It is also possible to prove that:
Z   Z
2
lim H () exp (2jx) exp 2 d = H () exp (2jx) d. (81)
n n

From (79) and (81) it follows that:

Z
f (x) exp (2jP (x)) = H () exp (2jx) d, (82)

and so, as expected the whole signal is recovered as the amplitude of the inverse Fourier transform of the CHT.

6 Progressive Transmission
In this section we propose a progressive transmission approach based on the DHFT. This is performed by dividing
the image (in the holographic domain) in several equally sized portions. Then, we take the recovered signal of a
portion and incrementally add those ones corresponding to the other pieces.
Formally, given W1 and W2 two window functions similar to expression (37), we have:
M
X 1
1 2
Iw1 (r) = H (u) W1 (u) ej M ur , (83)
u=0
M
M
X 1
1 2
Iw2 (r) = H (u) W2 (u) ej M ur , (84)
u=0
M
where H (u) is the DHFT defined in equation (35). Thus, the linearity of the Fourier transform allows to write:
M
X 1
1 2
Iw1 (r) + Iw2 (r) = H (u) (W1 (u) + W2 (u)) ej M ur , (85)
u=0
M
and so:

Iw1 (r) + Iw2 (r) = Iw1 +w2 (r) . (86)

Observe that if W1 , W2 are disjoint windows, W1 + W2 is another window function. Consequently, there is
not changes in the spectrum of the signal inside the window.
Thus, following expression (83)-(84), each recovered portion (I w1 (r) and Iw2 (r) above) can be considered
as a packet to be transmitted, whose arrival order does not matter.

7 Experimental Results
In this section we report some experiments with the DHFT, described on section 4, and with its continuous version,
the CHFT, presented on section 5.

7.1 DHFT Results

We implemented the DHFT using a 8-bit grayscale image and performing the calculus of the fourier transform
using double-precision for each complex component. This high precision avoids quantization effects in the holo-
graphic domain. The conversion from floating-point to 8-bit grayscale was done using uniform quantization [15].
Figure 11.a is the original image while Figure 11.b-d yields the results of the DHFT for a set of window sizes.
As expected, the larger the size of the window the better the quality of the recovered signal. However, the visual
inspection of the results in Figure 11 shows that the holographic property holds.
The expression (86) can be also verified. For instance, if we take two disjoint windows of size 1/8 of the
original image, and compute expressions (83)-(84) then the sum of the obtained results gives an image like the
one in Figure 11.c, which agrees with result (86).
 (a) (b)

(c) (d)

Figure 11: (a) Original image. Recovered images of a 8-bit grayscale image using an area corresponding to: (b)
1/8, (c) 1/4, (d) 3/4 of the original image in the holographic domain.
7.2 CHFT Results
In this section we show some 1D testes in order to check the holographic property for the CHFT definition in
expression (60). We use the simplest numerical scheme of expression (89) bellow to obtain a discrete version
of the CHFT. By comparing this expression with equation (4) we observe that the Lemma 1 (section 4) can be
applied and so the signal can be recovered by expression (48). The equation (87) defines the original signal to be
used and the equation (88) the low-pass filter applied in the tests bellow, where k defines the cut-of frequency.
The Figure 12 shows the results for k = 10, 20 and 40. To allow a qualitative analysis of the result, we show
in Figure 13 the same example but now setting P = 0 in expression (65). In this case, we have just a low-pass
version of the original signal. A visual inspection shows that the quality of the results obtained with the CHFT
methodology is higher than the counterparts using P = 0.

0, 1 x < 0,
f (y) = (87)
1, 0 x 1.

1, k k,
W () = (88)
0, otherwise.
N
X
g (x) [f (yn ) w (x yn ) y] exp (2jP (yn )) . (89)
n=0

(a) (b)

(c)

Figure 12: Function g(x) computed by expression (89) with: (a) k=10, (b) k=20, (c) k=40.

The next example compares the recovered signal when replacing the window function by a band-pass filter.
In this case the filter is defined by expression (31) with values for a, b shown in Figure 14. Again, we can make a
visual comparison with the results pictured on Figure 15 to realize that the holographic property holds.

8 Conclusions
In this paper we discuss some results concerning to Fourier analysis and Holographic representation of images.
 (a) (b)

(c)

Figure 13: Function g(x) computed by expression (89) with P = 0 and: (a) k=10, (b) k=20, (c)k=40.

(a) (b)

(c)

Figure 14: Recovered signal when replacing the window function by the band-pass filter (expression (31)) in
equation (65): (a) a = 1, b = 10, (b) a = 1, b = 20, (c) a = 10, b = 30.
 (a) (b)

(c)

Figure 15: Recovered signal when replacing the window function by the band-pass filter (expression (31)) in
equation (65) but setting P = 0: (a) a = 1, b = 10, (b) a = 1, b = 20, (c) a = 10, b = 30.

The approach described in section 6 must be analyzed from the viewpoint of progressive transmission ap-
proaches [13, 5]. Certainly, the need of double precision representation and the known instability of Fourier
analysis [6] may be several limitations for such method, if compared with other ones [5, 13]. Probably, the main
areas of application for the Holographic transform would be in distributed words and watermarking methods. Be-
sides a more complete theoretical analysis for CHFT must offered in order to quantify how good the estimate in
expression (65) is for the input signal f .

References
[1] On holographic transform compression of images. In ICPR 00: Proceedings of the International Conference
on Pattern Recognition (ICPR00)-Volume 3, page 3236, Washington, DC, USA, 2000. IEEE Computer
Society.

[2] Tibor Balogh, Tams Forgcs, Tibor Agocs, Olivier Balet, Eric Bouvier, Fabio Bettio, Enrico Gobbetti, and
Gianluigi Zanetti. A scalable hardware and software system for the holographic display of interactive graph-
ics applications. In EUROGRAPHICS 2005 Short Papers Proceedings, Conference Held in Dublin, Ireland,
August 2005, 2005.

[3] A. M. Bruckstein, R. J. Holt, and A. N. Npngetravali. Holographic image representations: the subsampling
method. In ICIP 97: Proceedings of the 1997 International Conference on Image Processing (ICIP 97)
3-Volume Set-Volume 1, page 177, Washington, DC, USA, 1997. IEEE Computer Society.

[4] A.M. Bruckstein, R.J. Holt, and A.N. Netravali. Holographic representation of images. IEEE Transactions
on Image Processing, 7(11):15831597, 1998.

[5] Chin-Chen Chang and Yi-Long Liu. Progressive image transmission using singular value decomposition.
EUROGRAPHICS Multimedia Workshop, 2004.
 [6] C.K. Chui. An Introduction to Wavelets. New York Academic Press, 1992.

[7] Maurice Francon. Optical Image Formation And Processing. Academic Press Inc., New York, New York,
1979.

[8] Djairo G. de Figueiredo. Analise de Fourier e Equac oes Diferenciais Parciais. Edgard Blcher, 1977.

[9] Murray R. Spiegel e John Liu. Colec ao Schaum Manual de Formulas e Tabelas Matematicas. Bookman,
2004.png

[10] L. C. Ferri. Holographic images in multimedia information systems. Technical report, Fraunhofer Institute
for Computer Graphics, www.inigraphics.net/ini-ccc/projects/holographie/paper1.pdf, 1996.

[11] R.C. Gonzalez. Digital Image Processing. Reading Addison-Wesley, 1992.

[12] Nickolas D. Gorsky. Experiments with handwriting recognition using holographic representation of line
images. Pattern Recognition Letters, 15(9):853859, 1994.

[13] See References in:. IEEE Transactions on Image Processing, 8(1), 1999.

[14] G. W. Burr H. Coufal H. Guenther J. A. Hoffnagle C. M. Jefferson B. Marcus R. M. Macfarlane R. M. Shelby

J. Ashley, M.-P. Bernal and G. T. Sincerbox. Holographic data storage. IBM Jornal of Research and Devel-
opment, 44(3):4653, 2000.

[15] Anil K. Jain. Fundamentals of Digital Image Processing. Prentice-Hall, Inc., 1989.

[16] J. M. Jolion and A. Montanvert. The adaptive pyramid: A framework for 2d image analysis. CVGIP: Image
Understanding, 55(3):339348, May 1992.

[17] L. B. Lesem, P. M. Hirsch, and Jr. J. A. Jordan. Scientific applications: Comppnguter synthesis of holograms
for 3-d display. Commun. ACM, 11(10):661674, 1968.

[18] Ting-Chung Poon. Digital Holography and Three-Dimensional Display: Principles and Applications.
Springer, 2006.

[19] Y. Zeevi and R. Coifman. Signal and Image Representation in Combined Spaces (Wavelet Analysis and Its
Applications). Academic Press, 1998. png

png

9 Appendix
In this section we review interesting properties in discrete Fourier analysis.
Property A1: The circular convolution theorem states that the DFT of the circular convolution of two se-
quences is equal to the product of their DFTs, that is, if:
N
X 1
x2 (n) = h (n k)c x1 (k) , 0 n 1, (90)
k=0

where h (n k)c is the a circular kernel (see [15]) then:

DF T [x2 (n)] = DF T [h (n)] DF T [x1 (n)] . (91)

Expression (91) can be written as:
N
X 1 N
X 1 N
X 1
j 2 vk j 2 vk 2
x2 (k) e N = h(k)e N x1 (k) ej N vk .
k=0 k=0 k=0

If we divide this expression by N we get:

N
X 1 N
X 1 N
X 1
1 2 1 2 2
x2 (k) ej N vk = h(k) ej N vk x1 (k) ej N vk .
k=0
N k=0
N k=0

Therefore, we get the same relation for the Unitary Discrete Fourier Transform:

U DF T [x2 (n)] = U DF T [h (n)] DF T [x1 (n)] . (92)

Now, if we define

png1, k = a,
x1 (k) =
0, k 6= a,
we have that, from expression (90), x 2 (n) = h (n a)c , i.e., h (n) is a circular shift of x2 (n). Hence

N
X 1
2
U DF T [x2 (n)] = U DF T [h (n)] x1 (k) ej N vk ,
k=0

which implies
2
U DF T [x2 (n)] = U DF T [h (n)] ej N va .png
Property A2. Let us consider the summation:

L1
X L1
X
u
ej2xu = ej2x . (93)
u=0 u=0

By remembering the geometric series:

L1
X
S= au = 1 + a + a2 + + aL1 ,
u=0

we know that:
1 aL
S= .
1a
If we apply this result in expression (93) we get:
L1
L

X 1 ej2x 1 ej2xL ejxL ejxL ejxL
j2xu
e = = = , (94)
1 ej2x 1 ej2x ejx (ejx ejx )
u=0

But we know that:

ejx ejx
sin (x) = ,
2j

ejxL ejxL
sin (Lx) = .
2j
Therefore, expression (94) becomes:
L1
X ejxL 2j sin (Lx) sin (Lx)
ej2xu = = ej(L1)x . (95)
u=0
ej 2j sin (x) sin (x)

Hence, if x = 0 we get:
 L1
X
ej2xu = L.
u=0
sin(x)
We can also re-write expression (95) by using the sinc function, which is defined by: sinc (x) = x .
Thus, that expression becomes:
L1
X Lx sinc (Lx) sinc (Lx)
ej2xu = ej(L1)x = Lej(L1)x .
x sinc (x) sinc (x)
u=0