1

Comments on Sinc Interpolation of Discrete Periodic Signals

Frank Candocia, Student Member, IEEE and Jose C. Principe, Senior Member, IEEE
Computational NeuroEngineering Laboratory
Department of Electrical Engineering
University of Florida, Gainesville, 32611
Abstract: Recently, the convolution of the sinc kernel with the infinite sequence of a periodic
function was expressed as a finite summation. The expression obtained, however, is not
numerically stable when evaluated at or near integer values of time. This correspondence
presents a numerically stable formulation equivalent to the results reported in the above cited
paper and which, when sampled, is also shown to be equivalent to the inverse discrete Fourier
transform (IDFT).
To appear in the IEEE Transactions on Signal Processing
2
I. INTRODUCTION
The theory of Fourier signal analysis is extensive and has proven fruitful to a variety of
applications [1,2]. One such application is the interpolation of a continuous time signal from a
given set of samples. The Shannon sampling theory establishes that a bandlimited continuous
time signal can be completely represented by a set of equidistantly spaced signal samples. In
turn, the continuous signal could be recovered from this set of samples via sinc interpolation, i.e.
a linear filtering of the sample set with a sinc kernel that is infinite in extent.
When the set of samples is periodic, the use of an alternate kernel is known to equivalently
accomplish the interpolation of the corresponding continuous signal. This was the result recently
reported in [3]. The sinc interpolation of discrete periodic signals is then equivalent to a circular
convolution of the periodic sample set with the alternate kernel - the Fourier series kernel. This
method of interpolation of discrete periodic signals has been studied in [4].
The Fourier series kernel has been reported in [5]. It contains several indeterminate points, i.e.
points at which the function takes an indeterminate form (like 0/0), for which the interpolation
cannot be numerically implemented directly. By applying the operation of Fourier series
expansion on our signal, a numerically stable form of this interpolation can be obtained.
II. A NUMERICALLY STABLE FORMULATION
The sinc interpolation formula for discrete signals as presented in eqn. (1) of [3] is
x t x nT
t n T
t n T
n
( ) ( )
sin( ( ) / )
( ) /

(1)
where x(nT) are the uniformly spaced samples of a band-limited signal x(t) and where T is the
sampling period such that 1/T is larger than twice the highest frequency contained in x(t). The
result given by eqns. (6a) and (6b) in [3] is the expressing of eqn. (1) as a finite summation when
the signal x(t) is periodic. In this correspondence, periodic signals will be indicated by the use of
a tilde (~). The results from [3] are
3
~
( )
sin( )
~
( ) cot x t
t
N
x
t n
N
N
n
n
n L
M

_
,

1
1
even (2a)
and
~
( )
sin( )
~
( ) csc x t
t
N
x
t n
N
N
n
n
n L
M

_
,

1
1
odd (2b)
where
~ ~
( )
~
( ) x x n x nT
n
, N is the period of
~
( ) x n , T = 1, and L,M are arbitrary integers satisfying
L+M=N.
The numerical instability of eqns. (2a) and (2b), due to the cot() and csc() functions
respectively, is apparent when evaluating x(t) at or near t = n + mN, m any integer. To be more
precise, for a small positive number which is a function of machine tolerance, numerical
instability exists when evaluating
~
( ) x t at t satisfying |t-(n+mN)| < . In an implementation
setting, this observation translates to a numerical computation of the approximate form 0/0 when
evaluating
~
( ) x t at or near integer values of t. This is clearly an undesirable computation that
must be circumvented for digital computer implementation.
The formulation presented here is not afflicted with this occurrence. The relation between the
Fourier series of a signal and the discrete Fourier transform (DFT) of its samples leads to an
expression reminiscent of the inverse discrete Fourier transform (IDFT) expression. The new
formulation is valid for complex and real valued signals and will be refined when considering
real-valued
~
( ) x t . In this case eqns. (2a) and (2b) are formulated as a linear combination of
sinusoids. The DFT of x(n) is shown to contain the amplitude, phase and frequency information
of the stated sinusoids.
The discrete Fourier transform pair for N point sequences is given by [6]
DFT X k x n e k N
j
nk
N
n
N
: ( ) ( ) , , ,

2
0
1
01 1

K (3)
IDFT x n
N
X k e n N
j
nk
N
k
N
: ( ) ( ) , , ,

1
01 1
2
0
1
K (4)
4
where x n
x n n N
otherwise
( )
~
( ) , , ,

'

01 1
0
K
and it is noted that any DFT coefficient can be represented in polar form in terms of its
magnitude and phase: X k X k e
j X k
( ) ( )
( )

where the phase of the DFT coefficient, X(k),
satisfies X k ( ) .
The general sinc interpolation expression for complex (as well as real) discrete periodic signals
can be formulated into a numerically stable expression as given below
~
( ) ( ) ( ) x t A k X k e
j
k
N
t
T
k
N

2
0
1
(5)
( ) A k
k P
e k
e otherwise
N
N
j t T N
N
j t T
( )
, , ,
/
/

'

1
1
2
2
2
1 2
0 1
1
K

where ] P
N

1
2
. The operation ] c rounds c downward to the nearest integer. The equivalence
of eqn. (5) at T = 1 to eqns. (2a) and (2b) which were obtained in [3] is deferred to the appendix.
The constant P has been defined such that the interpolation formulations for N even and N odd
can be written as a single expression. Note that when N is odd the k = N/2 index is not present in
A(k) since k is necessarily an integer. The numerical stability of eqn. (5) is evident from its strict
complex exponential basis. This equation closely resembles the IDFT formulation in
appearance. In fact, by letting t = nT in eqn. (5), it is easily seen that A k
N
( )
1
for
k N 01 1 , , , K and that eqn. (5) becomes the IDFT of X(k).
For real-valued
~
( ) x t eqn. (5) is refined into the following form:
~
( ) ( ) ( ) cos ( ) x t B k X k
k
N
t
T
X k
k
Q
+

_
,

0
2
(6)
5
B k
k
otherwise
N
N
N
( )
,

'

1
2
2
0
where ] Q
N

2
. The upper limit of summation Q has also been defined (as was P) such that the
interpolation formulations for N even and N odd can be written as a single expression. The strict
cosine basis of eqn. (6) leads to its obvious numerical stability. Note that the amplitude and
phase of each of the cosinusoids that comprise
~
( ) x t are directly obtained from the DFT
coefficients of x(n). Eqn. (6) is expressed in the same manner as the harmonics form of the
trigonometric Fourier series [7].
The refined sinc interpolation expression above clearly yields an alternate representation of the
IDFT formulation. Specifically, by letting t = nT in eqn. (6) it is shown in the appendix that
~
( ) ( ) ( ) cos ( )
( )
( ) , , ,
x n B k X k
kn
N
X k
N
X k e
x n n N
k
Q
j
kn
N
k
N
+

_
,

0
2
0
1
2
1
0 1 1

K
(7)
where B(k) and Q have been defined in eqn. (6).
III. DISCUSSION AND CONCLUSIONS
It has been shown that the sinc interpolation of discrete periodic signals as presented by Schanze
can be expressed in a numerically stable manner. The formulations presented here closely
resemble the Fourier series synthesis formulas, but require the computation of the DFT of the
signal samples. Practically, this is not a burden on the interpolation process and, in fact, does not
increase the order of the computational cost of the interpolation. For example, given N signal
samples and a desired interpolation factor of two, Schanzes interpolation formula requires O(N
2
)
computations. In contrast, since the DFT of the sequence needs only be computed once, the
formulation presented here would require O(N
2
+P) computations where P represents the
additional amount of computations required to perform the DFT of an N point sequence. As a
6
worst case, directly performing the DFT on the N point sequence would yield P = N
2
computations and an interpolation cost of O(N
2
+ N
2
) = O(N
2
) computations. Obviously, the
overall computational cost could be reduced when interpolating through special length
sequences for which many fast Fourier transform algorithms are known to exist. Ultimately, the
user must decide whether the additional cost incurred by the required DFT step justifies the
numerical stability associated with the formulation presented.
APPENDIX
The discussed equivalence of the relations listed in eqns. (5), (6) to eqns. (2a) and (2b) will now
be shown. The relation between eqn. (6) and the IDFT formulation of eqn. (7) will also be
shown. The derivations for the N odd cases are omitted but are very similar to those for N even.
The difference between them is that there is one additional term in the derivations for N even
(due to the k = N/2 index) that is not present in the N odd derivations.
The derivation of eqn. (5), which relates the numerically stable general sinc interpolation formula
for complex signals to eqns. (2a) and (2b), is now shown. Begin by writing the result in eqn. (5)
for T = 1 and expand terms for the N even case:
[ ]
~
( ) ( ) ( ) ( ) ( ) x t
N
X X e e X k e X k e e
N j t j t
j
k
N
t
k
j
k
N
t
j t
k
N
N
N
+ + + +

1
]
1

1
0
1
2 2
2
1
1 2
2
1
1 2
2

(A1)
Now substituting for X(k) using the Fourier relation that X k x n e x n e
j
kn
N
n
N
j
kn
N
n L
M
( ) ( )
~
( )

2
0
1 2 1
for L and M arbitrary integers that satisfy L+M=N we have
[ ]
~
( )
~
( )
( ) ( )
x t
N
x n e e e e e e
j n j t j t
j
N
t n k
k
j t
j
N
t n k
k
N
n L
M
N
N
+ + + +

1
]
1

1
1
1
2
2
1
1
2
2
1
1 1 2
2

Now, rewriting both the k-indexed summations above in closed form using c
c c
c
k
k a
b a b

1
1
and
also using e
j n n t

( ) 1 , we simplify the resulting expression to get

7
( )( )
~
( )
~
( )( )
( )
( )
x t
N
x n
e e e
e
n
j t n j t j t
j t n
n L
M
N
N

+

1
]
1
1

1
1
1
1
1
2
1
2
2

Finally, we obtain eqn. (2a) by pulling out a factor e
j t n
N

( )
from the numerator and denominator
above followed by the expressing of the resulting complex exponential terms into their
equivalent cos() and sin() forms. By a simple rescaling of the t variable in eqns. (2a) and (2b) to
t/T, the result of the numerically stable sinc interpolation formula shown in eqn. (5) is obtained.
The expression given in eqn. (6), which refines the general sinc interpolation formulation for the
case of real-valued signals, is now shown. Beginning the N even derivation, eqn. (6) is written
again and expanded as follows:
~
( ) ( ) ( ) cos ( )
( ) cos( ( )) ( ) cos ( ) ( ) cos ( )
x t B k X k
k
N
t
T
X k
N
X X X
t
T
X X k
k
N
t
T
X k
k
Q
N N
k
N
+

_
,

+ +

_
,
+ +

_
,

1
]
1

0
2 2
1
1
2
1
0 0 2
2
2

Now, for T 1 and using the identity cos y
e e
jy jy

+

2
and facts e e
j X j X

( ) ( ) 0 0
,
e e
j X N j X N

( / ) ( / ) 2 2
for real x(n) we write
( )
~
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
x t
N
X e X e e e X k e e X k e e
j X N j X j t j t j X k
j
k
N
t
k
j X k
j
k
N
t
k
N
N N
+ + + +

1
]
1

1
0
0 1
2 2
2
1
1 2
1
1
2
2 2


Now, letting p = N - k in the fourth term above yields
( )
~
( )
( ) ( ) ( )
( )
( ) ( ) ( )
( )
( )
x t
N
X e X e e e X k e e
X N p e e
j X N j X j t j t j X k
j
k
N
t
k
j X N p
j
t
N
N p
p N
N
N
N

+ + + +

1
]
1
1
1
1

1
0
0 1
2 2
2
1
1
2
1
1
2
2
2

Using Fourier properties X k X k e

j X k
( ) ( )
( )

, X k X N k ( ) ( ) and X k X N k ( ) ( );
k
N
12 1
2
, , , K for real x(n), we obtain eqn. (A1). We had previously shown how to get eqns.
(2a) and (2b) from eqn. (A1).
8
The N even case of eqn. (7) is also similarly derived. Begin by expanding terms in eqn. (7)
~
( ) ( ) ( ) cos ( )
( ) cos( ( )) ( ) cos( ( )) ( ) cos ( )
x n B k X k
kn
N
X k
N
X X X n X X k
kn
N
X k
k
Q
N N
k
N
+

_
,

+ + + +

_
,

1
]
1

0
2 2
0
1
2
1
0 0 2
2
2

Now using the identity cos y
e e
jy jy

+

2
and facts e e
j n j n

, e e
j X j X

( ) ( ) 0 0
and
e e
j X N j X N

( / ) ( / ) 2 2
for real x(n) we can write
~
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
x n
N
X e X e e X k e e X k e e
j X N j n j X
j
kn
N
j X k
k
j
kn
N
j X k
k
N
N N
+ + +

1
]
1

1
0
0
2
2
1
1 2
1
1
2
2 2


Now, with a substitution of variables p = N - k into the fourth term above and using additional
Fourier properties for real x(n) that X k X N k ( ) ( ) and X k X N k ( ) ( );
k
N
12 1
2
, , , K we have
~
( ) ( ) ( ) ( ) ( )
( ) ( ) ( ) ( )
x n
N
X e X e e X k e e X p e e
j X N
j X j n j X k
k
j
kn
N
j X p
p
N
j
pn
N
N
N
N
+ + +

1
]
1
1

1
0
0
2
1
1
2
1
1 2
2
2
2


Finally, the four terms can be combined into one summation which yields the result to be shown,
that is
~
( ) ( )
( )
( ) , , , .
( )
x n
N
X k e e
N
X k e
x n n N
j X k
j
kn
N
k
N
j
kn
N
k
N

1
1
0 1 1
2
0
1
2
0
1

K
9
REFERENCES
[1] Papoulis A., The Fourier Integral and Its Applications, McGraw Hill Book Co., New York,
1962.
[2] Bracewell, R. M., The Fourier Transform and Its Applications, McGraw Hill Book Co., New
York, 1965.
[3] Schanze T., Sinc Interpolation of Discrete Periodic Signals, IEEE Trans. on Signal
Processing, Vol. 43, No. 6, pp. 1502-1503, June 1995
[4] Ahmed N. and Rao K.R., Orthogonal Transforms for Digital Signal Processing, Springer-
Verlag, New York, Chap. 2, 1975.
[5] Schaum A. Theory and Design of Local Interpolators, Comp. Vision Graph. Image Proc.:
Models and Image Proc., Vol. 55, No. 6, pp 464-481, Nov. 1993.
[6] Oppenheim A.V. and Schafer R. W., Discrete-Time Signal Processing, Prentice Hall,
Englewood Cliffs, NJ, Chap. 8, 1989.
[7] Hsu H. P., Applied Fourier Analysis, HBJ College Outline Series, New York, Chap. 1, 1984.