17

Interpolation

In developing the sampling theorem, we based the reconstruction procedure

for recovering the original signal from its samples on the use of a lowpass fil-
ter. This follows naturally from the interpretation of the sampling process in
the frequency domain. Correspondingly, in the time domain the reconstruc-
tion is represented by the convolution of the impulse train of samples with the
impulse response of the lowpass filter. Convolution of an impulse response
with an impulse train can be viewed as a superposition of weighted delayed
impulse responses with amplitudes and positions corresponding to the im-
pulses in the impulse train. This superposition represents an interpolation
process between the samples. When the reconstruction filter is an ideal low-
pass filter, the interpolating function is a sinc function. This is often referred
to as bandlimited interpolationbecause it interpolates between sample
points by explicitly assuming that the original signal is bandlimited to less
than half the sampling frequency.
In addition to bandlimited interpolation, a variety of other interpolation
procedures are commonly used. One, referred to as a zero-orderhold, interpo-
lates between sample points by holding each sample value until the next sam-
pling instant. This generates a staircase-like approximation to the original sig-
nal. Linear interpolation, also commonly referred to as a first-order hold,
corresponds to connecting the sample points by straight line segments. Both
the zero-order hold and first-order hold can be alternatively viewed in much
the same way as we have discussed ideal bandlimited interpolation. Specifi-
cally, the zero-order hold corresponds to convolving the impulse train of sam-
ples with a rectangular pulse of duration exactly equal to the sampling period.
The first-order hold corresponds to an impulse response for the reconstruc-
tion filter that is a triangle of duration equal to twice the sampling period. In
the frequency domain, then, the zero-order hold corresponds to processing
the samples with an approximation to a lowpass filter corresponding to the
Fourier transform of a rectangular pulse. With the first-order hold the ap-
proximate lowpass filter has a frequency response that is the Fourier trans-
form of a triangle.
One of the important applications of the concept of sampling is its use in
converting continuous-time signals to discrete-time signals corresponding to
a sequence of sample values. This provides one basis for storing, coding, or

17-1
Signals and Systems
17-2

transmitting continuous-time signals. In addition, it offers the possibility for

discrete-time processing of continuous-time signals. In many situations such
processing is highly advantageous. For example, digital technologies for sig-
nal processing, which inherently are oriented toward discrete-time signals
and systems, are extremely flexible and often lend themselves to implement-
ing more sophisticated and flexible algorithms than a continuous-time system
might. By exploiting the sampling theorem, a continuous-time signal to be
processed can be converted to a discrete-time signal, processed by a discrete-
time system, and then converted back to a continuous-time signal. In develop-
ing insight into this process, it is important to clearly understand in both the
time and the frequency domains the process of converting from a continuous-
time signal to a sequence of samples. This continuous-to-discrete-time con-
version (abbreviated as C/D) is conveniently thought of in two stages. The
first represents sampling of the continuous-time signal with a periodic im-
pulse train to generate an impulse train of samples. This impulse train is then
converted to a discrete-time sequence essentially by relabeling; that is, a dis-
crete-time sequence is generated in which each impulse is represented by its
area. After the first stage, the impulses in the impulse train occur at multiples
of the sampling period. After the second stage, the discrete-time sequence
representing the impulse values is indexed on sample number and conse-
quently the sample spacing has been normalized to unity. For example, if a
continuous-time signal were to be sampled and stored in a computer memory,
it would first be sampled in time and the sample values converted through an
analog-to-digital converter to digital numbers. These numbers would then be
placed in memory. The resulting discrete-time sequence would be the se-
quence of numbers in successive memory locations, and the independent
variable indexing the discrete-time sequence could in fact be thought of as
memory location number.
In the frequency domain, the two-step process described above has a rel-
atively straightforward interpretation. Through the process of sampling, as-
suming that the continuous-time signal is bandlimited and the conditions of
the sampling theorem are met, the spectrum of the continuous-time signal is
periodically replicated at integer multiples of the sampling frequency. Con-
version of the impulse train to a discrete-time sequence corresponds in the
time domain to a time normalization, in effect normalizing out the sampling
period. Correspondingly, in the frequency domain, the frequency axis is nor-
malized with the sampling frequency being scaled to a discrete-time frequen-
cy of 2Tr. Thus, as we naturally expect, the Fourier transform of the discrete-
time sequence is periodic with a period of 2 7r. The periodicity can be
interpreted as being a consequence of the basic sampling process. The nor-
malization of the period in frequency to 27r is a consequence of the inherent
time normalization in converting the impulse train of samples to a discrete-
time sequence.

Suggested Reading
Section 8.1.2, Sampling with a Zero-Order Hold, pages 519-521
Section 8.2, Reconstruction of a Signal from Its Samples Using Interpolation,
pages 521-526
Section 8.4, Discrete-Time Processing of Continuous-Time Signals, pages 531
to mid-537
 Interpolation
17-3

+oo0

P(t) 6(t - nT)

n=- oo

X,(t) TRANSPARENCY
17.1
System and an
illustration of
associated spectra
for sampling and
reconstruction.
2
WS> WM

M WS

H(W)

WM < Wc < (w - M)
M

-Ac ( cC

MARKERBOARD
17.1 (a)

LOI~J~qO4- 4ec(O+k
+* 4

Cvvto
for ~ 1 iWw de LO,,7~)n
Signals and Systems
17-4

MARKERBOARD
17.1 (b)

(t

Y,)

Xrct)=

___j

p(t)

x(t) X0 ()
TRANSPARENCY
17.2
Reconstruction of a
sampled signal with a
zero-order hold.

xp(t)
x(t) - NI -'
/M

Af f
N

N.v~FiNN%.l

- N x0 (t)
/
%.~ - N
N
/
 Interpolation
17-5

p(t)

x(t) x 1(t)
TRANSPARENCY
17.3
Reconstruction of a
sampled signal with a
first-order hold.

T*] XP(t)
x(t)

x, (t)

TRANSPARENCY
H (co) 17.4
Transparencies 17.4
and 17.5 illustrate a
comparison of the
frequency responses
/- -- Ideal interpolating
of the ideal
reconstruction filter
f ilter and the reconstruction
filter corresponding to
Zero-order a zero-order hold and
hold a first-order hold.
Shown here is the
ideal reconstruction
filter corresponding to
a zero-order hold.
I ____________ I ____________

- WS cis cO

2
Signals and Systems

TRANSPARENCY
17.5 H (w)
Ideal reconstruction
filter, reconstruction T
filter corresponding -to
a first-order hold, and
reconstruction filter
corresponding to a Ideal interpolating
zero-order hold. filter

- W, 0 Cs wsw
2 2

DEMONSTRATION
17.1
Original image.
 Interpolation
17-7

DEMONSTRATION
17.2
Image of
Demonstration 17.1
spatially sampled.

DEMONSTRATION
17.3
Optical lowpass
filtering (defocusing)
of spatial sampling.
Signals and Systems
17-8

DEMONSTRATION
17.5
Defocusing of
Demonstration 17.4.
 Interpolation
17-9

DEMONSTRATION
17.7
Close-up of
Demonstration 17.4
defocused.
I

Signals and Systems

DEMONSTRATION
17.8
A first-order hold
applied to the
sampled image of
Demonstration 17.1.

DEMONSTRATION
17.9
Image of
Demonstration 17.1
spatially sampled with
twice the sample
spacing as in
Demonstration 17.2.
 Interpolation
17-11

DEMONSTRATION
17.11
A first-order hold
applied to the
sampled image of
Demonstration 17.9.
Signals and Systems
17-12

TRANSPARENCY
17.6
Discrete-time
processing of DISCRETE-TIME PROCESSING
continuous-time OF CONTINUOUS-TIME SIGNALS
signals.

Conversion to x[n] Discrete-time y[n] Conversion to y (t)

discrete-time system continuous-time

L- _ - - __ J

TRANSPARENCY
17.7
Interpretation of the
conversion of a
C/D conversion
continuous-time signal
to a discrete-time
sequence as impulse p(t)
train sampling
followed by
conversion of the Conversion of
Ix t) xP(t) impulse train x[n] xc(nT)
impulse train to a
discrete-time c--- to discrete-timex
sequence. sequence
 Interpolation
17-13

MARKERBOARD
17.2

pct)

xx 3-

440
T

+ '--*
e t c(4 5 Zg.t-a

W24
0(0~

XC). p(o

__j

xM(t) x (w)
TRANSPARENCY
17.8
Illustration in the time
domain and frequency
domain of conversion
of a signal from
x (w) continuous time to
discrete time.

T 2T

x[n] 1

012 L-N-
MIT OpenCourseWare
http://ocw.mit.edu

Resource: Signals and Systems

Professor Alan V. Oppenheim

The following may not correspond to a particular course on MIT OpenCourseWare, but has been
provided by the author as an individual learning resource.

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.