IEEE TRANSACTIONS O N CIRCUITS A N D SYSTEMS-11: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 39, NO.

8, AUGUST 1992 587

REFERENCES tended the sampling theorem to the multichannel case [4]. He

[1] I. W. Ranmuthu, K. T. M. Ranmuthu, A. V. Pohm, and C. S. has shown that one can simultaneously sample the output of N
Comstock, “Reprogrammable logic array using MR elements,” IEEE systems having the same band-limited input with a sample fre-
Trans. Magn., vol. 26, pp. 2828, Sept. 1990. quency of 1 / N times the Nyquist rate. This work can be viewed
[2] I. W. Ranmuthu, K. T. M. Ranmuthu, C. Kohl, C. S. Comstock, and
M. Hassoun, “High density CMOS process compatible magneto as a special case of undersampling; however, it introduces the
resistive memory,” accepted for the 34th Midwest Conf. Circuits notion of multichannel sampling, and the number of
and Systems. samples/unit time is still equal to the number of samples/unit
time of the original input, sampled with the Nyquist rate.
In this contribution, we will establish a Generalized Sampling
Theorem (GST), and illustrate its significance when applied to
Generalized Sampling Theorem undersampling. In the GST context, the samples of the first
M - 1 derivatives of the underlying function and the N prespec-
Yang-Ming Zhu ified linear systems are not needed and, for the undersampling
case, the number of samples is actually reduced.
Abstract-It is shown that, for an arbitrary function f ( t ) if there
exists a reversible function g ( t ) such that g ( f ( t ) )is band-limited,then 11. MAINRESULTS
f(t) can be uniquely determined in terms of its samples f ( n T , ) sampled
with the Nyquist rate of g ( f ( t ) ) . The application of this result to In this section we present the GST as follows.
undersampling is illustrated by examples. Generalized Sampling Theorem: Let f ( t) be a function of one
variable. If a one-to-one continuous mapping g ( t ) exists such
I. INTRODUCTION
that g(f(t)) is band-limited, i.e., its Fourier transform G f ( w ) = 0
The sampling theorem for band-limited functions was origi- for 101 2 wo = ?r/T,, and if f ( t) is sampled at the points t, nT,,
nally established by Shannon et al. [l]. The theorem is stated then f ( t ) can be uniquely determined in terms of f ( t n ) as
below as the Sampling Theorem [2].
Sampling Theorem: If f ( t ) is a function of one variable having
a Fourier transform F ( w ) such that F ( w ) = 0 for I w I 2 wo =
.rr/T,, and is sampled at the points t , = nT,, then f ( t ) can be
reconstructed exactly from its samples f(nT,) as follows:
f ( t > = g-’{ 5
,=-Io
g(f(t,)) sin [ % ( t - n T , ) I / [ mo(t - n T ) 1
1
(2)

c f(nT,)sin[w,(t
3c

where g-’ is the inversion of g.

f(t) = - n T , ) I / [ w o ( t - nT,)I. (1) Proofi Recalling the sampling theorem, we have
n=-x

Since the original work, the sampling theorem has been ex-
c g(f(t,)) sin [
m

tended into various versions. The review paper [l] by Jerri is an g(f(t)) = mo(t - n T , ) I / [ wo(t - nTs)1.
encyclopedic survey of known results on sampling theorem. n=-m

More recently, Davis also reviews the various extended versions (3)
of the sampling theorem briefly in [3]. However, many of the
extensions focus on nonuniform sampling, multichannel sam- Noting that g-’g = 1, we reach (2) immediately.
pling, etc. [2]-[5]. Discussion: Whether f ( t > is band-limited or not, the GST
As far as we know, there are hardly any papers emphasizing holds provided the underlying conditions are satisfied. If f(t) is
the undersampling issue. In fact, undersampling might find di- not band-limited, then the GST can be applied to undersampled
verse applications just as nonuniform sampling and multichannel function reconstruction. If f ( t) is band-limited, in some cases
sampling do. For example, in limited-view tomography, some- one can find some specific g(t> such that the Nyquist rate of
times oversampling of the unknown function significantly is g(f(t)) may be lower than that of f(t), which means that the
required because of squashing or nonlinear transformation of GST can be applied to undersampled band-limited function
the projection functions [2], [6]. If these functions can be pre- reconstruction.
cisely reconstructed from undersampled points, then the sam-
pling frequency can be reduced. As a result, the storage for 111. EXAMPLES
saving the sampled data is also conserved. To confirm the validity of the GST further, we provide two
Shannon himself has presented a notable extension of the examples of the application of the GST to undersampling recon-
sampling theorem in his 1949 paper [7], which says that if f ( t) struction. One of the functions to be reconstructed is band-
and its first M - 1 derivatives are available, then uniformly limited, while the other is not.
spaced samples of these, taken at the reduced rate of 1/M times Example 1: f(t) = 3 c o s ( 2 r t ) + c o s ( 6 ~ t ) .
the Nyquist rate, are sufficient to reconstruct f(t). Shannon’s It is easy to see that f(t) is band-limited. According to the
extension is the earliest publication on the undersampling issue. sampling theorem, one must sample f(t) with a frequency no
However, it requires that samples of the first M - 1 derivatives less than 6 Hz if one wishes to reconstruct f ( t) precisely from its
of the underlying function be available, and the underlying samples. Here, we take g ( t ) = t1I3. As a consequence, the
function must be band-limited. Papoulis has ingeniously ex- sampling frequency is reduced to 2 Hz. Needless to say, the
number of samples is also reduced. In order to evaluate the
Manuscript received January 6, 1992; revised April 23, 1992. This performance of the reconstruction, the error magnitude E ( t ) of
paper was recommended by Associate Editor R. Liu. the reconstruction is shown in Fig. 1, where E ( t ) is defined by
The author is with the Department of Biology and Biomedical Engi-
neering, Southeast University, Nanjing 210018, China.
IEEE Log Number 9203649. E(t) = 201glfjt) -f(t)l (4)

1057-7130/92$03.00 0 1992 IEEE

Authorized licensed use limited to: Universidad Nacional de Colombia. Downloaded on January 17, 2010 at 19:56 from IEEE Xplore. Restrictions apply.
 588 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS-11: ANALOG A N D DIGITAL SIGNAL PROCESSING, VOL. 39, NO. 8, AUGUST 1992

0.0

m‘
O.O ] A

v
m
U
U
v
a,
a, U
3

I
U
3
‘e -100.0
U

c
.c -100
Y

0
0,
0
0
z
3
I
0 2
L
L W
W
-200 0
-200.0 00 05 10
0.0 05 1 .o Time ( s )
Time (s)

Fig. 1. Error magnitude of the reconstruction for Example 1. Fig. 2. Error magnitude of the reconstruction for Example 2.

and where [ ( t ) denotes the reconstructed function. If the GST is function restoration. Thus, the potential application of GST is
true, then f ( t ) - f(t) is constant and equal to zero. But consid- worth probing further.
ering the computation error, f ( t ) - f(t) will not be zero, and if
the GST is valid, then the absolute value of E ( t ) will be REFERENCES
sufficiently large. A. J. Jerri, “The Shannon sampling theorem-Its various exten-
Example 2: f ( t ) = exp(cos(2~t)). sions and applications: A tutorial review,” h o c . IEEE, vol. 65, pp.
By using Taylor series one can show that f ( t ) is not band- 1565-1596, NOV.1977.
limited, so f(t) cannot be precisely reconstructed from samples J. J. Clark, M. R. Palmer, and P. D. Lawrence, “A transformation
method for the reconstruction of functions from nonuniformly
by the classical Shannon sampling theorem. We take g ( t ) = In t , spaced samples,” IEEE Trans. Acoust. Spegch, Signal Processing,
so that g(f(t>) = cos(27rt). Consequently, the required sam- vol. 33, pp. 1151-1165, Oct. 1985.
pling frequency is 2 Hz. The error magnitude of the reconstruc- A. M. Davis, “Almost periodic extension of band-limited functions
tion is plotted in Fig. 2. and its application to nonuniform sampling,” IEEE Trans. Circuits
Notice that in the above two examples the error magnitudes S’st., vol. 33, pp. 933-938, Oct. 1986.
A. Papoulis, “Generalized sampling expansion,” IEEE Trans. Cir-
are relatively small. Considering the computation error, we cuits Syst., vol. 24, pp. 652-654, Nov. 1977.
arrive at the conclusion that both reconstructions are precise. P. P. Vaidyanathan and V. C. Liu, “Classical sampling theorems in
Considering that the sampling theorem and its extended versions the context of multirate and polyphase digital filter bank structures,”
cannot achieve such results, the GST seems to be more attrac- IEEE Trans. Acoust. Speech, Signal Processing, vol. 36, pp.
tive. 1480-1495, Sept. 1988.
P. Oskoui and H. Stark, “A comparative study of three reconstruc-
IV. CONCLUSION tion methods for a limited-view computer tomography problem,”
IEEE Trans. Med. Imaging, vol. 8, pp. 43-49, Mar. 1989.
We have developed a generalized sampling theorem (GST) C. E. Shannon, “Communications in the presence of noise,” h o c .
which, under some conditions, can be applied to undersampled IRE, vol. 37, pp. 10-21, 1949.

Authorized licensed use limited to: Universidad Nacional de Colombia. Downloaded on January 17, 2010 at 19:56 from IEEE Xplore. Restrictions apply.