ELEC 6218 — Signal Processing —

Discrete-time Signals (Basics)

Professor Eric Rogers

Department of Electronics and Computer Science

University of Southampton
[email protected]
Office: Building 1, Room 2041

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Background
These signals, denoted by DT from this point onwards, are defined
only at discrete time values. Usually the time values are equally
spaced, i.e. at multiples of constant period T

t = n T , n an integer

Often DT signals are obtained by sampling continuous-time

(CT) signals or waveforms, eg digitized input to a computer

x(n) = x(nT )

where here we follow the common practice of omitting T from the

argument. Note also that DT signals can be inherently discrete,
e.g. daily takings in a shop.

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Normalised Frequency

Consider the DT sinusoid

n
-2 -1 0 1 2 3 11 12 13 14 15 16

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Normalised Frequency

x(n) = A cos (ωnT + θ), ω = 2πf

= A cos (Ωn + θ), Ω = 2πF

where Ω and F are normalized frequencies with respect to the

sample rate fs = T1 , i.e.

2πf
Ω = ωT = (rad/sample)
fs
f
F = fT = (cycles/sample)
fs

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Normalised Frequency

Consider also the sinusoidal sequence x0 (n) = A cos (2πF0 n) where

the normalized frequency F = F0 < 0.5, i.e. the actual frequency
f0 < f2s . Then it is clear from the next figure that a higher
frequency CT sinusoid could be fitted to the same samples,
e.g. with actual and normalized frequencies as given next.

x1(n) x 0 (n)

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Normalised Frequency

f1 = fs − f0
F1 = 1 − F0

The samples of this higher frequency are given by

x1 (n) = A cos [2π(1 − F0 )n]

= A cos (2nπ − 2nπF0 )
= A cos (2nπF0 )
= x0 (n)

This phenomenon is known as aliasing and will be considered in

detail later.
Professor Eric Rogers ECS
ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Power and Energy Signals
DT signals can be classified as power and energy signals in the
same way as for CT signals, the only difference being in the
definitions which become
N
" #
1 X
2
P = limN→+ ∞ |x(n)|
2N + 1
n=− N

and
∞
X
E = |x(n)|2 < + ∞
n=− ∞

respectively.
A DT signal is causal if

x(n) = 0, for n < 0

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Sampling Theory

Sampling is an essential process in digital control/signal/image

processing. It imposes certain limitations on what can be
achieved in subsequent processing and a through
understanding of the principles behind it is essential.
Two forms of idealized sampling of a CT signal will be considered
here:
◮ natural
◮ instantaneous

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Natural Sampling

This can be regarded as an on/off switching operation which

can be modelled as multiplication of the CT signal by a train
of unit amplitude rectangular impulses of width τ and period
T , i.e.
xS (t) = x(t)p(t)
where the periodic pulse train p(t) is defined as
∞
X t − nT
p(t) = rect( )
n=− ∞
τ

Since p(t) is periodic, it can be represented by its Fourier series

(work the details)

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Natural Sampling

Hence

τ X sin ( nπτ
∞
T ) jnωs t
xS (t) = x(t) nπτ e
T n=− ∞ T
∞
τ X
= sinc(nfs τ )x(t)e jnωs t
T n=− ∞

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Natural Sampling

Applying the Fourier transform (and making use of Property 4 in

the notes on the Fourier transform) gives the spectrum of the
sampled signal as
∞
τ X
XS (f ) = sinc(nfs τ )X(f − nfs )
T n=− ∞

This spectrum consists of the CT signal spectrum X (f ) plus

an infinite number of images of X (f ) one centered on each
multiple of the sampling frequency nfs - see the next Figure
(first diagram).

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Natural Sampling
X(f)

f
fh 0 f h

Natural Sampling Xs(f)

X(f) sinc(f τ)

f
-2f s -1/ τ -f s 0 fs 1/τ 2fs
(a)

Instantaneous Sampling ideal LPF

Xs(f) H i(f)
(i) fh < fs /2

f
-2f s -f s -f s /2 0 fs /2 fs 2fs

(ii) fh > fs /2 Xs(f)

f
-2f s -f s -f s /2 0 fs /2 fs 2fs

(b)

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Instantaneous Sampling
In this form of idealised sampling (above Figure (last two
diagrams)) the samples are represented as impulses with
magnitudes (or areas) equal to the values of x(t) at times t = nT ,
i.e.
∞
X
xS (t) = x(nT )δ(t − nT )
n=− ∞

or (by a basic property of the unit impulse function)

∞
X
xS (t) = x(t) δ(t − nT ) = x(t)c(t)
n=− ∞

where
∞
X
c(t) = δ(t − nT )
n=− ∞

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Instantaneous Sampling
This is a periodic train of unit impulses known as a comb
function. This comb function can be described by its Fourier series
∞
1 X jnωs t
c(t) = e
T n=− ∞

Hence
∞
1 X
xS (t) = x(t)e jnωs t (1)
T n=− ∞
and, on taking the Fourier transform, the spectrum of the sampled
signal is given by
∞
X
XS (f ) = fs X (f − nfs )
n=− ∞

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Instantaneous Sampling

It is clear that
◮ all images of X (f ) are of equal magnitude
◮ if X (f ) is bandlimited such that the highest frequency
fh < f2s then X (f ) can be recovered by using an ideal low
pass filter (LPF) of the form (see also below)
f
Hi (f ) = rect( )
fs
If, however, fh > f2s then the images of X (f ) overlap giving
aliasing distortion.

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Instantaneous Sampling

Consideration of these facts leads immediately to the following

statement of the Sampling Theorem (Nyquist, Shannon).
A CT signal which is bandlimited to the frequency range ±fh is
completely defined by taking samples at a uniform rate

fs ≥ 2fh

The minimum sampling rate fs = 2fh is known as the Nyquist

rate.
Note: In practical systems an anti-aliasing low-pass filter is
used to restrict the CT signal bandwidth prior to sampling.

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation
Here this term means the recovery of the original CT signal by
filling in the gaps between sampling - a similar process to
interpolation in numerical analysis (or even graph plotting). Three
types of interpolation will be considered here — the first is purely
ideal and the other two more practical.
Ideal LP Filter
We have already seen that in the frequency domain it is possible to
recover the CT signal exactly using an ideal LP filter with transfer
function
f
Hi (f ) = rect( ) = rect(fT)
fs
and impulse response
1 t
hi (t) = sinc( )
T T
Professor Eric Rogers ECS
ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

Plot of this last expression

h i(t)

-4 -3 -2 -1 0 1 2 3 4
t/T

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation
For instantaneous sampling it is straightforward to derive a time
domain expression for the recovered CT signal, starting with the
filter input sequence
∞
X
xS (t) = x(nT )δ(t − nT )
n=− ∞

The filter output, denoted by xr (t), then is simply the sum of the
weighted and shifted impulse responses, i.e.
∞
X
xr (t) = x(nT )hi (t − nT )
n=− ∞
∞
1 X t − nT
= x(nT )sinc( )
T n=− ∞ T

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

∞
X
= fs x(nT )sinc(fs t − n)
n=− ∞

At the sampling points (t = nT ) all sinc functions are zero except

the one centered at t = nT . Hence the recovered signal is exact
at these points, and provided the sampling theorem is
satisfied, it is also exact at all intermediate values of t, i.e.

xr (t) = fs x(t)

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

Illustration of the above process by showing the summation

of the weighted and shifted impulse responses.

h i(t-nT)
xr(t) = fs.x(t)

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

Illustration of the above process by showing the summation

of the weighted and shifted impulse responses.
For natural sampling, a similar analysis shows that the recovered
signal is given by
τ
xr (t) = x(t) = τ fs x(t)
T
Non-Ideal LP Filter
Realisable LPFs can only approximate the ideal LPF response
and will always have a finite response to the higher frequency
spectral images giving distortion of the recovered CT signal.
The next figure illustrates this fact.

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

non-ideal LPF
X s(f) H(f)

f
0 fs /2 fs 2fs

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

Another minor problem with practical LPFs is that if Tτ is small

then the filter output has a small amplitude and more gain is
required.
Zero Order Hold
In this approach, each sample value is held constant until the next
sample is input to the filter. This gives a staircase
approximation of x(t) as shown below. Clearly there is high
frequency distortion due to the ‘sharp corners’ on the output
waveform. There is also low frequency distortion as shown next.

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation
For the case of instantaneous sampling, we can assume that the
unit impulse response of the filter to be a unit amplitude
rectangular pulse of duration T , i.e.

t − T2
hz (t) = rect( )
T
Hence on applying the Fourier transform, the effective transfer
function of this filter is given by
ıωT
Hz (f ) = T sinc(fT)e− 2

where the exponential term here is equivalent to a delay of

value T2 .

Professor Eric Rogers ECS

ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)
Interpolation
Plots of the gain |Hz (f )| and the amplitude spectrum of the
recovered CT signal.

Hz(f)

f
-2f s -f s -f s /2 0 fs /2 fs 2fs

Effective Frequency Response of Zero Order Hold

Xr(f)

f
-2f s -f s -f s /2 0 fs /2 fs 2fs

Professor Eric Rogers

Amplitude Spectrum of Recovered Signal ECS
ELEC 6218 — Signal Processing — Discrete-time Signals (Basics)