Lecture 2: Sampling of Continuous-Time Signals

Reading: Sec. 4.0-4.5, 4.8

Sampling
• Sampling is usually the first operation in modern signal processing systems,
converting an analog signal to a digital one. We have learned that sampling is
not a LTI operation; therefore we cannot use a transfer function to analyze it.
• To understand sampling, we define an analogous signal xs(t) with respect to
an analog signal xc(t) as:
∞
xs ( t ) = xc ( t ) ∑ δ ( t – nT ) ,
n = –∞

which is xc(t) multiplied by an impulse train at time nT.

xs(t) has several properties. First it has infinite amplitude and energy (obvi-
ously from the fact that it is composed of impulses) and second it resembles
xc(t) alot, so we expect that its Fourier transform is related to the Fourier
transform of xc(t).
xc(t)

xs(t)

time
T T T T T T T

Assume that the Fourier transform of xc(t) is Xc(jΩ) (Fourier transform of an

analog signal is denoted by big Ω), then the Fourier transform of xs(t) can be
expressed as:
∞ ∞ ∞
X s ( jΩ ) = ∫ ∑ x c ( nT )δ ( t – nT )e – jΩt dt = ∑ x c ( nT )e –jΩnT .
–∞ n = –∞ n = –∞
∞
We know that X ( e jω ) = ∑ x ( n )e – jωn , which looks very much like
n = –∞
Xs(jΩ), if ΩT is replaced by ω. We will come back to this point later.

2- 1
 Since xs(t) is derived from xc(t), we like to have a relationship between Xs(jΩ)
and Xc(jΩ) as well.

From the equation of xs(t), we can use the Fourier transform property that
multiplication in the time domain means convolution in the frequency
domain to identify the relationship between Xc(jΩ) and Xs(jΩ).
∞
1 1 2π
X s ( jΩ ) = ------X c ( jΩ ) ⊗ S ( jΩ ) = ------X c ( jΩ ) ⊗ ------
2π 2π T ∑ δ ( jΩ – jkΩ s )
k = –∞
∞ ∞
ω 2πk
X c  j ---- – j --------- .
1 1
X s ( jΩ ) = --- ∑ X c ( jΩ – jkΩ s ) , and X ( e jω ) = --- ∑
T T  T T 
k = –∞ k = –∞

Graphically, we can represent the above relationship by

Xc(jΩ) S(jΩ)

Ω Ω
−2π/Τ Ωs=2π/Τ

Xs(jΩ)

Ω
−2π/Τ Ωs=2π/Τ

• Nyquist Sampling Theorem: If Xc(jΩ)=0 for |Ω| > ΩN, then xc(t) is uniquely
defined by its samples x(n)=xc(nT) if Ωs=2π/T > 2ΩN.
• Is the band limited condition necessary?

2- 2
Example:
Xc(jΩ)
1

Ω
-2π/T -π/T ΩN π/T 2π/T

Xs(jΩ)

Ω
-2π/T -π/T π/T 2π/T

Xc(jΩ)
1

Ω
-2π/T -π/T π/T 2π/T

Xs(jΩ)

Ω
-2π/T -π/T π/T 2π/T

Aliasing is introduced at every integer multiple of Ωs. X(ejω) and Xs(jΩ) have
exactly the same spectrum except for a change of variable (ω vs. Ω). It should
also be noted that the Sampling Theorem is only a sufficient condition; that
is, if the sampling rate Ωs is smaller than 2ΩN, it may still be possible to
uniquely define xc(t) from its samples (very rarely though).

2- 3
Reconstruction of Band-Limited Signals
• After the sampling operation, we have discrete-time samples and perform
digital signal processing on these samples. Most often we need to convert dis-
crete-time samples back to continuous-time signals as most real-world sig-
nals are in the analog form.
The conversion is nothing more than an ideal analog low-pass filter, to elimi-
nate all the high-frequency aliases introduced by sampling. But let’s look at
the conversion process in both time and frequency domains, and verify that
they are equivalent.

Xs(jΩ) or X(ejω)
xs(t)
1/T
time
Ω
−2π/Τ 2π/Τ
Convolved with sinc(πt/T)
LPF with gain T

xr(t)
Xc(jΩ)
1
time

∞ ∞
xr ( t ) = xs ( t ) ⊗ hr ( t ) = ∑ x ( n )δ ( t – nT ) ⊗ h r ( t ) = ∑ x ( n )h r ( t – nT )
n = –∞ n = –∞

t ( t – nT )
sin π --- ∞ sin π -------------------
T T
Since h r ( t ) = -------------------- , x r ( t ) = ∑ x ( n ) ------------------------------------ .
t ( t – nT )
π --- n = –∞ π -------------------
T T
Note that the sinc(πt/T) has a gain of T at DC in its frequency spectrum,
which compensates for the DC gain (or loss) of 1/T resulted from sampling.

2- 4
Example:Discrete-time processing of continuous-time signals.

Xc(jΩ)
1

Ω
Sampled at 2π/T

1/T

Ω
-2π/T 2π/T
Filtered by H(ejω)

Y(ejω)=X(ejω)H(ejω)

ω
−2π 2π
LPF with gain T

Yr(jΩ)
1

Ω
The effective analog filter response can be expressed as
H(ejΩT), |Ω|< π/Τ
H eff ( jΩ ) =
0, |Ω| > π/Τ

Example:If we want to implement a continuous-time differentiator in dis-

crete-time, what is the impulse response of this system?
d
Define y c ( t ) = x c ( t ) , then the effective analog filter response is
dt
jΩ, |Ω| < π/T
Heff(jΩ) =
0, |Ω| > π/T

H(ejω) = jω/T, |ω| < π.

2- 5
 |H(jΩ)| |H(ejω)|

π/Τ π/Τ

Ω ω
−π/Τ π/Τ −2π 2π

Phase Phase
π/2 π/2
Ω ω
−π/Τ π/Τ −2π 2π
−π/2

πn cos ( nπ ) – sin ( nπ ) 0, n=0

h ( n ) = ---------------------------------------------------- =
πn 2 T cos(πn)/nT, otherwise.
Q: How to implement this filter?

• So far we have assumed ideal filters (non-causal IIR filters can never be prac-
tically implemented). We will discuss next how to make appropriate approxi-
mations of these ideal cases and use realistic systems for both sampling and
reconstruction.

2- 6
Practical Analog/Digital Interface
• A/D conversion is usually performed by a sample-and-hold circuit which
dumps the value of the sampled date into a capacitor. The voltage level stored
in the capacitor is compared with the reference voltages in an A/D converter
and quantized to the nearest quantization level.

xc(t) Anti-aliasing Sample xo(t)

xa(t) A/D x(n)
and hold converter
analog filter ho(t)

∞ ∞
xo ( t ) = ho ( t ) ⊗ ∑ x a ( nT )δ ( t – nT ) = ∑ x a ( nT )h o ( t – nT )
k = –∞ k = –∞

The above hold circuit is called a zero-th order hold, which provides time for
the following A/D converter to work out the nearest quantization level.
1, 0 < t < T
ho(t) =
0, otherwise

Since both sampling and A/D conversion are time-variant operations (you
should verify this), we cannot use frequency domain techniques in analyzing
them. The sampling circuit is approximated by the “dump” operation men-
tioned above while the hold circuit generally doesn’t introduce much distor-
tion.

2- 7
 • D/A conversion is, interestingly enough, a LTI system and therefore can be
described in both time and frequency domains.

x(n) Hold xo(t) Post xc(t)

circuit filtering
Ho(jΩ) Hr(jΩ)

Hold operation (zero-th order hold):

X(ejΩT)
T
1/T

n Ω
-2π/Τ 2π/Τ
Xo(jΩ)=Ho(jΩ)X(ejΩT)

The hold circuit serves as a very simple low-pass filter, which limits the sig-
nal bandwidth to approximately 2π/T. But because of the residual frequency
components left, we need to use a post filter to provide an effective filter
response with a relatively flat passband and a sharp cutoff frequency at π/T.

Example:Use the zero-th order hold circuit as an example and examine its
frequency response, therefore the design of the post filter.
The frequency response of the zero-th order hold is
1 1
Delay of
+ e-jT/2
t t
0 T -T/2 T/2

ΩT
sin ------- ΩT
- –jΩ T--- H ( jΩ ) –1 -------- jΩ T---
2
H o ( jΩ ) = ---------------- e 2 , and thus H ( jΩ ) =  ------------------
o  =
2
----------------e 2 .
Ω r  
----
T sin ΩT --------
2 2

2- 8
 Example:Continued
|X(ejω)|

1/T

ω
−2π 2π
|Ho(jΩ)|
T

Ω
−4π/Τ −2π/Τ 2π/Τ 4π/Τ

|Xo(jΩ)|

−4π/Τ −2π/Τ 2π/Τ 4π/Τ

Ideal post filer: |Hr(jΩ)|

−4π/Τ −2π/Τ 2π/Τ 4π/Τ

2- 9