Module F: Sampling
Question: Can we recover a continuous-time signal x(t) from regular sam-
pling xs [k] = x(kT )?
T
2
1
t
2 2
1
2
Question: Is the recovered signal unique?
Answer: No in general!
But if we know the signal x(t) is band-limited, then with sufficient samples,
we can!
1
Sampling by Impulse Train
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥
Figure 1: Diagram of impulse train sampling system.
The above diagram (system) described impulse train sampling.
The time-domain representation of various signals:
2
x(t)
xp (t)
1
T t
2 1 1 2 3
1
2
Impulse Train Sampling: Frequency Domain Analysis
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥
In frequency domain, we have:
1
Xp (j!) = P (j!) ⇤ X(j!) (1)
2⇡
Note that p(t) is periodic with period T .
P
Therefore, p(t) = 1 k= 1 ak e
jk!s t
, where
Z
1 T /2 1
ak = p(⌧ )e jk!s ⌧ d⌧ = .
T T /2 T
2⇡
P1
As a result P (j!) = T k= 1 (! k!s ).
Replacing this in (1), we get:
1
X 1
Xp (j!) = X(j(! k!s )).
T
k= 1
3
Bandlimited Signals
Define the bandwidth of a signal x(t) to be the smallest !M > 0 such that:
X(j!) = 0 for |!| > !M .
1
!
!M !M
If !M < 1 we call this signal a bandlimited signal.
P
So if !s > 2!M , for Xp (j!) = T1 1k= 1 X(j(! k!s )) we get:
1
T
!
2!s !s !M !M !s 2!s
How to recover x(t)?
4
Optimal Recovery
If x(t) is a band-limited signal and !s > 2!M , then x(t) can be recovered
by (ideal) low-pass filtering of sampled signal xp (t).
For perfect recovery, we need to have Hr (j!) = T rect!s /2 (!).
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥ Hr (j!) = T rect!s /2 (!) xr (t)
2!M is called the Nyquist-rate (of signal x(t)).
5
Sampling Theorem (Shannon-Nyquist)
Suppose that x(t) is a band-limited signal with bandwidth !M . Then x(t)
can be recovered from the samples x(kT ) for k = 0, ±1, ±2, . . . if 2⇡
T =
!s > 2!M . This can be achived by impulse train sampling and recovery
system with Hr (j!) = T rect!s /2 (!).
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥ Hr (j!) = T rect!s /2 (!) xr (t)
6
Question
Let x(t) be a signal band-limited to the interval [-20kHz, 20kHz].
What condition should the sampling period T satisfy to avoid aliasing and allow
reconstruction of the signal?
1 3
a. T < 20 ⇥ 10 sec
1 3
b. T < 40 ⇥ 10 sec
⇡ 3
c. T < 20 ⇥ 10 sec
⇡ 3
d. T < 40 ⇥ 10 sec
7
Aliasing
Overlapping copies of X(j(! k!0 )) is called aliasing.
1
T
!
!s !s
P1
In this case, what would be Xp (j!) = k= 1 X(j(! k!0 ))?
Therefore, in order not to have aliasing we need to have: !s > 2!M .
8
Exercise
What is the bandwidth !M of cos(t)?
Determine Xp (j!), Xr (j!), and xr (t) for !s = 2!M .
9
Reconstruction after Impulse Train Sampling
A naive approach would be to do linear interpolation.
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥ Hr (j!) = T rect!s /2 (!) xr (t)
T sin( T⇡ t)
In TD: hr (t) = ⇡t
Therefore,
xr (t) = xp (t) ⇤ h(t)
" 1 #
X
= x(nT ) (t nT ) ⇤ h(t)
n= 1
1
X
= x(nT ) [ (t nT ) ⇤ h(t)]
n= 1
X1 1
X T sin( T⇡ (t nT ))
= x(nT )h(t nT ) = x(nT )
n= 1 n= 1
⇡(t nT )
10
Time Domain View of Reconstruction Process
P1
p(t) = k= 1 (t kT )
xp (t)
x(t) ⇥ Hr (j!) = T rect!s /2 (!) xr (t)
P1 T sin( T⇡ (t nT ))
xr (t) = n= 1 x(nT ) ⇡(t nT )
This is a linear combination of time-shifted sinc functions!
11
Zero-Order-Hold (ZOH) Sampling
Since creating an ideal impulse train and implementing an ideal low-pass filter
is not possible, in practice, we often do Zero-Order-Hold (ZOH) sampling.
In ZOH sampling, the sampled signal xs (t) holds the value of x(t) at kth
sampling time for T seconds, i.e.,
xs (t) = x(kT ) for t 2 [kT, (k + 1)T )
2 T
x(t)
xs (t)
1
2 1 1 2 3
12
ZOH Sampling and Perfect Recovery
ZOH sampling can be viewed as series interconnection of impulse train sam-
pling and an LTI system with the impulse response h0 (t) = rectT /2 (t T2 ).
P1
p(t) = k= 1 (t kT )
xp (t) h0 (t) xs (t)
x(t) ⇥ Hr (j!) xr (t)
The main question is that if !s > 2!M , then can we find a recovery system
Hr so that xr (t) = x(t)?
For perfect recovery, the system in blue box should be the ideal low-pass
filter Hr (j!) = T rect!s /2 (!).
13
ZOH Sampling and Perfect Recovery
ZOH sampling can be viewed as series interconnection of impulse train sam-
pling and an LTI system with the impulse response h0 (t) = rectT /2 (t T2 ).
P1
p(t) = k= 1 (t kT )
xp (t) h0 (t) xs (t)
x(t) ⇥ Hr (j!) xr (t)
For perfect recovery, the system in blue box should be the ideal low-pass
filter Hr (j!) = T rect!s /2 (!).
Therefore, for perfect recovery, we need to have:
!T
Hr (j!) = ej!T /2 rect!s /2 (!).
2 sin(T !/2)
Note that the first zero of sin(T!!/2) occurs at !1 = 2⇡ T = !s . Since,
sin(T !/2)
rect!s /2 (!s ) = 0, the zeros of ! are not going to be problematic
in the above fraction.
14
First Order Hold Sampling
Corresponds to linear interpolation of the impulse train samples.
15
Under-Sampling and Aliasing
Let x(t) = cos(!0 t). We vary !s by choosing !s = 6!0 , 3!0 , 1.5!0 , 1.2!0 .
16
Under-Sampling and Aliasing Continued
When !s = 6!0 , 3!0 , the reconstructed signal xr (t) = x(t).
17
Under-Sampling and Aliasing Continued
When !s = 1.5!0 , 1.2!0 , the reconstructed signal xr (t) = cos((!s !0 )t) 6=
x(t).
As !0 increases relative to !s , the frequence of output decreases. Look up
Wagon-Wheel E↵ect.
When !s = !0 , the output is constant.
Phase reversal: If x(t) = cos(!0 t + ), then xr (t) = cos((!s !0 )t ).
18
DT Processing of CT Signals
xd [n] = x(nT ) yd [n] = y(nT )
xc (t) C\D Conversion DT System Hd (ej⌦ ) D\C Conversion yc (t)
How to process a CT signal using discrete-time processors?
1. Convert x(t) to xd [n] by C/D conversion (with sampling period T )
2. (design and implement) Filter xd [n], Xd (ej⌦ ) in discrete-time with hd [n],
Hd (ej⌦ ) to get desired yd [n], Yd (ej⌦ ).
3. Convert yd [n] to y(t) by D/C conversion (with sampling period T ).
19
DT Processing of CT Signals
xd [n] = x(nT ) yd [n] = y(nT )
xc (t) C\D Conversion DT System Hd (ej⌦ ) D\C Conversion yc (t)
The discrete-time sequence xd [n] = xc (nT ).
What is the relationship between Xc (j!) and Xd (ej⌦ )?
We use ! to denote frequency in continuous-time FT and ⌦ to denote
frequency in discrete-time FT.
20
DT Processing of CT Signals
P
CT Sampling - Perspective 1: xp (t) = xc (t)p(t) = 1n= 1 x(nT ) (t nT )
and
1
1 X
Xp (j!) = Xc (j(! k!s ))
T
k= 1
j!nT
CT Sampling - Perspective 2: On the other hand (t nT ) !e and
hence: 1
X
Xp (j!) = x(nT )e j!nT
n= 1
DT Sampling: xd [n] = x(nT ) for n and hence
1
X
j⌦ j⌦n
Xd (e ) = x(nT )e
n= 1
Conclusion:
1
j⌦ 1 X
Xd (e ) = Xp (j⌦/T ) = Xc (j(⌦/T k!s ))
T
k= 1
Converting xp (t) to xd [n] scales the time-axis by 1/T . Consequently, in the
frequency domain, Xd (ej⌦ ) is obtained by scaling Xc (j!) by factor T , i.e.,
⌦ = T !.
21
DT Processing of CT Signals
xd [n] = x(nT ) yd [n] = y(nT )
xc (t) C\D Conversion DT System Hd (ej⌦ ) D\C Conversion yc (t)
22
DT Processing of CT Signals
23