ECE 3640

Lecture 6 – Sampling
Objective: To learn and prove the sampling theorem and understand its impli-
cations.

The sampling theorem

Due to the increased use of computers in all engineering applications, including
signal processing, it is important to spend some more time examining issues of
sampling. In this chapter we will look at sampling both in the time domain and
the frequency domain.
We have already encountered the sampling theorem and, arguing purely from
a trigonometric-identity point of view, have established the Nyquist sampling cri-
terion for sinusoidal signals. However, we have not fully addressed the sampling
of more general signals, nor provided a general proof. Nor have we indicated how
to reconstruct a signal from its samples. With the tools of Fourier transforms and
Fourier series available to us we are now ready to finish the job that was started
months ago.
To begin with, suppose we have a signal x(t) which we wish to sample. Let us
suppose further that the signal is bandlimited to B Hz. This means that its Fourier
transform is nonzero for −2πB < ω < 2πB. Plot spectrum.
We will model the sampling process as multiplication of x(t) by the “picket
fence” function X
δT (t) = δ(t − nT ).
n

We encountered this periodic function when we studied Fourier series. Recall that
by its Fourier series representation we can write
1 X jnωs t
δT (t) = e
T n

where ωs = 2π T . The frequency fs = ωs /(2π) = 1/T is the sampling frequency in

samples/sec. Suppose that the sampling frequency is chosen so that fs > 2B, or
equivalently, ωs > 4πB.
The sampled output is denoted as x(t), where

x(t) = x(t)δT (t)

Using the F.S. representation we get

1 X jnωs t
x(t) = x(t) e
T n

Now lets look at the spectrum of the transformed signal. Using the convolution
property,
1 1 X 1X
X(ω) = X(ω) ∗ 2πδ(ω − nωs ) = X(ω − nωs ).
2π T n
T n

Plot the spectrum of the sampled signal with both ω frequency and f frequency.
Observe the following:
ECE 3640: Lecture 6 – Sampling 2

• The spectrum is periodic, with period 2π, because of the multiple copies of
the spectrum.
• The spectrum is scaled down by a factor of 1/T .
• Note that in this case there is no overlap between the images of the spectrum.
Now consider the effect of reducing the sampling rate to fs < 2B. In this case,
the duplicates of the spectrum overlap each other. The overlap of the spectrum is
aliasing.
This demonstration more-or-less proves the sampling theorem for general signals.
Provided that we sample fast enough, the signal spectrum is not distorted by the
sampling process. If we don’t sample fast enough, there will be distortion. The
next question is: given a set of samples, how do we get the signal back? From the
spectrum, the answer is to filter the signal with a lowpass filter with cutoff ωc ≥ 2πB.
This cuts out the images and leaves us with the original spectrum. This is a sort of
idealized point of view, because it assumes that we are filtering a continuous-time
function x(t), which is a sequence of weighted delta functions. In practice, we have
numbers x[n] representing the value of the function x[n] = x(nT ) = x(n/fs ). How
can we recover the time function from this?

Theorem 1 (The sampling theorem)

If x(t) is bandlimited to B Hz then it can be recovered from signals taken at a
sampling rate fs > 2B. The recovery formula is
X
x(t) = x(nT )g(t − nT )
n

where
sin(πfs t)
g(t) = = sinc(πfs t).
πfs t
Show what the formula means: we are interpolating in time between samples using
the sinc function.
We will prove this theorem. Because we are actually lacking a few theoretical
tools, it will take a bit of work. What makes this interesting is we will end up using
in a very essential way most of the transform ideas we have talked about.

1. The first step is to notice that the spectrum of the sampled signal,
1X
X(ω) = X(ω − nωs )
T n

is periodic and hence has a Fourier series. The period of the function in
frequency is ωs , and the fundamental frequency is
2π 1
p0 = = = T.
ωs fs
By the F.S. we can write
X
X(ω) = cn ejnωT
n

where the cn are the F.S. coefficients

Z ∞
1 2π 1 1
Z
cn = X(ω)e−jnωT dω = X(ω)e−jnωT dω.
ωs ω s ωs 2π −∞ T
ECE 3640: Lecture 6 – Sampling 3

But the integral is just the inverse F.T., evaluated at t = −nT :


1 2π 
cn = x(t) = x(−nT ),
T ωs t=−nT

so X X
X(ω) = x(−nt)ejnωT = x(nt)e−jnωT .
n n

2. Let g(t) = sinc(πfs t). Then

 
ω
g(t) ⇔ T rect .
2πfs

3. Let X
y(t) = x(nT )g(t − nT ).
n

We will show that y(t) = x(t) by showing that Y (ω) = X(ω). We can compute
the F.T. of y(t) using linearity and the shifting property:
   X
X ω ω
Y (ω) = x(nT )T rect e −jωnT
= T rect x(nT )e−jωnT
n
2πfs 2πfs n

Observe that the summation on the right is the same as the F.S. we derived
in step 1:
ω
Y (ω) = T rect X(ω).
2πfs
Now substituting in the spectrum of the sampled signal (derived above)
  !
ω 1X
Y (ω) = T rect X(ω − nωs ) = X(ω)
2πfs T n

since x(t) is bandlimited to −πfs < ω < πfs or −fs /2 < f < fs /2.
Notice that the reconstruction filter is based upon a sinc function, whose trans-
form is a rect function: we are really just doing the filtering implied by our initial
intuition.
In practice, of course, we want to sample at a frequency higher than just twice
the bandwidth to allow room for filter rolloff.

Other issues for sampling

Antialias filtering When sampling is done, the signal is usually passed first
through an analog low-pass filter with the cutoff frequency set to ensure that no
aliasing occurs.
The discussion above assumes that the signal is sampled by a sequence of impulse
functions. In practice, the signal is sampled by multiplying by some other function.
Is it still possible to recover the signal exactly? Provided that we satisfy the Nyquist
sampling criterion, the answer is yes. The reconstruction is similar.

Some applications
Why digital?
1. Recoverable signals
ECE 3640: Lecture 6 – Sampling 4

2. Flexibility
3. Channel coding theorem. Source coding theorem.
4. Encryption

• Time division multiplexing

• Pulse code modulation

Spectral sampling
Just as we can sample a band-limited signal in the time domain and reconstruct it,
provided that we sample often enough, so can we also sample a time-limited signal
in the frequency domain and exactly reconstruct the spectrum, provided that we
take the samples close enough together in the spectrum. There is thus a dual to the
sampling theorem which applies to sampling the spectrum. We can also use this
kind of thinking to find the F.T. of a periodic signal.
Let f (t) be a causal signal time-limited to τ seconds. Its F.T. is
Z τ
F (ω) = f (t)e−jωt dt.
0

Now create a periodic extension of f (t) by repeating it every T0 seconds.

X
fT0 (t) = f (t − nT0 ).
n

Since fT0 (t) is periodic, it has a F.S. and we can write

X
fT0 (t) = Dn ejnω0 t
n

where ω0 = 2π/T0 and (when τ < T0 )

T0 τ
1 1
Z Z
Dn = f (t)e−jnω0 t = f (t)e−jnω0 t .
T0 0 T0 0

Comparing the F.S. coefficient with the F.T. above, it follows that
1
Dn = F (nω0 ).
T0
An implication of this is that we can find the F.S. coefficients by first taking the
F.T. of one period of our signal, then sampling and scaling the F.T.
In terms of reconstructing the signal spectrum from its samples, we can see that
as long as T0 > τ , the cycles of f (t) do not overlap. We can then (at least in concept)
reconstruct the entire spectrum of F (ω) from its samples. Conceptually, we time-
limit the function, then take its F.T. The sampling condition can be expressed as
follows:
T0 > τ
1
F0 <
τ
So we can reconstruct from samples of the spectrum, provided that the samples are
close enough by comparison with the time-limit of the function.