The transformation !
X(ejω ) = x[n]e−jωn
n
is referred to as the DTFT of x[n]. In contrast to the CTFT we note the following. X(ejω ) is periodic with
period 2π. Given the DT F T the signal x[n] can be computed via
# 2π
1
x[n] = X(ejω )ejωn dω
2π 0
Note that x[n] need not be finitely supported in time for defining the DTFT.
Also recall that for an LTI system if the input is ejωn the output is H(ejω )ejωn where H(ejω ) is the
DTFT of the impulse response h[n] of the system.
A
"necessary and sufficient condition for the DTFT to exist is that the signal is absolutely summable,
i.e. n |x[n]| < ∞. This is similar to the case for CTFT.
Please study in detail the examples 5.1, 5.2, 5.3 (rectangular pulse), and 5.4. These examples are also
taken up in the class.
We note that this is very similar to the CTFT for periodic signals. Does this definition imply periodicity
with period 2π? Yes, since ak are periodic with period N . Take examples 5.5 and 5.6 from the book.
Note: Please regard X(ejω ) as X(ω). The notation with ejω is to relate the DTFT with the Z-Transform
that we will study in chapter 10.
2
3 Properties of the DTFT
Most are similar but the time scaling properties need a careful development in the discrete case. Also the
multiplication property is different - periodic convolution.
and
ejω0 n x[n] ←→ X(ej(ω−ω0 ) )
x∗ [n] ←→ X ∗ (e−jω )
5. Time reversal: x[−n] ←→ X(e−jω ). Combined with the conjugation property, using this property
what can you say about the DTFT of real and even signals and DTFT of real and odd signals?
6. Time Expansion: Time expansion in the case of DTFT corresponds to inserting zeros between every
successive samples. Define a signal x(k) [n] as follows.
$ n
x[ k ] if n = rk
x(k) [n] = (4)
0 otherwise
One can split the summation on the RHS into two parts. One summation ranges over all n such that
n = rk, and the other summation ranges over all n such that n ∕= rk. For the second summation
we see that the contribution is zero since for those values of n ∕= rk, x(k) [n] = 0 (by definition).
Therefore, we have that,
∞
! ∞
! ∞
!
x(k) [n]e−jωn = x(k) [rk]e−jωrk = x[k]e−jωrk = X(ejkω )
n=−∞ r=−∞ r=−∞
Take example from Figure 5.14 in class. Read example 5.9 from the book.
7. Time contraction: In the discrete case, time contraction corresponds to downsampling. We will
study this in detail in Chapter 7.
dX(ejω )
nx[n] ←→ j
dω
3
9. Parseval’s relation: Show that
! #
2 1
|x[n]| = |X(ejω )|2 dω
n
2π 2π
10. Convolution Property: Convolution in the time domain implies multiplication in the frequecny
domain. Take examples 5.11, 5.12, and 5.14 in class. Ideal low-pass and high pass filters in the
discrete case.
11. Multiplication Property: Periodic convolution. Show the mechanics via the example in the book.
Show that, #
1
x1 [n]x2 [n] ←→ X1 (ejθ )X2 (ej(ω−θ) )dθ
2π 2π
where the integration is carried out over any interval of length 2π. In general we note the following.
Let X̂1 (ejω ) = X1 (ejω ) in the interval [−π, π] and zero otherwise. Then note that,
# π # ∞
1 jθ j(ω−θ) 1
X1 (e )X2 (e )dθ = X̂1 (ejθ )X2 (ej(ω−θ) )dθ = X̂1 (ejθ ) ∗ X2 (ejω ) (5)
2π −π 2π −∞
Therefore, the periodic convolution reduces to normal convolution when we restrict one of the DTFT
to be non-zero in some interval of length 2π (You may choose any interval but often it is convenient
to choose the interval [−π, π]. Take example 5.15 from the book.
We note that such a periodic convolution can also be described for periodic DT signals. This will
appear as an exercise in the HW.
4
Fall 2018
Given the CT signal x(t) that is bandlimited to !0 (in the sense of the definition above), consider
multiplying the signal with a periodic pulse train with p(t) fundamental period T ,
1
X
p(t) = (t kT ). (1)
k= 1
Let
2⇡
!s =
T
denote the sampling frequency. Using the multiplication property of the CTFT and using the definition
of the CTFT of the periodic pulse train given by,
X 1
P (j!) = 2⇡ (! k!s ),
T
we note that,
1
Xp (j!) = X(j!) ⇤ P (j!) (3)
2⇡ ✓X ◆
1 1
= X(j!) ⇤ 2⇡ (! k!s ) (4)
2⇡ T
1X
= X(j(! k!s )) (5)
T
k
Xp (j!) is then seen to be sum of shifted versions of T1 X(j!). These shifted versions overlap if !s 2!0 .
But there is no overlap if !s > 2!0 . For this case, it can be seen that the original signal x(t) can be exactly
recovered from the sampled signal xp (t) by passing the xp (t) through an ideal low-pass filter with gain T
and cut-off !2s . This result can be summarize into a Theorem called as the Nyquist Sampling Theorem.
Theorem 1.1. Nyquist: A sufficient condition for recovery of bandlimited signals from uniform sampling is
that
!s > 2!0
Is uniform sampling and recovery using low-pass filtering the only way to sample and recover (ban-
dlimited) signals? What is the meaning of this rate of sampling? For modern perspective read about
compressed sensing and sampling finite rate of innovation signals. I will post some links on Piazza.
1.1 Is strict inequality !s > 2!0 necessary (under uniform sampling and using low-pass
filtering for reconstruction)?
Let us take an example here. Suppose x(t) = cos(!0 t + ). Then, using the definition of the CTFT for
periodic signals, we note that X(j!) = ⇡ej (! !0 ) + ⇡e j (! !0 ). We consider the following cases.
• Now suppose that !s = 2!0 . In this case we that when we try to recover x(t) from xp (t) via ideal
low-pass filter we will get xr (t) = cos( ) cos(!0 t). To see this, plot the two impulses and see that
the shift to the left causes the impulse with the negative phase add to the impulse with the positive
phase and the shift to the left causes the impulse with the positive phase add to the impulse with
the negative phase. In the worst case when = ⇡2 the output is zero.
• When !s < 2!0 the output is given by xr (t) = cos((!s !0 )t ). This effect is seen in moving
car wheels on the screen. The wheel starts to first move forward with reduced frequency and then
backwards with some other frequency.
Study Figure 7.15 and associated discussion carefully to arrive at the conclusions above.
!c T sin(!c t)
hl (t) =
⇡!c t
Then
In particular we have
X !c T sin(!c (t kT ))
xr (t) = x(kT ) (9)
⇡!c (t kT )
k
In this case !s = 2!0 is a sufficient enough sampling frequency for exact recovery of the signal, i.e.
one can sample at the Nyquist rate. For strictly bandlimited signals we can show that (BONUS),
X ✓ k⇡ ◆ sin(!0 t k⇡)
x(t) = x (10)
!0 (!0 t k⇡)
k
2
We note that the signals
sin(!0 t k⇡)
k (t) =
(!0 t k⇡)
R
are orthogonal to each other, i.e. k k0 (t) = 0 for all k 6= k . Therefore, one way to interpret Nyquist
0
Sampling theorem is to note that any bandlimited signal can be represented as a superposition of ban-
dlimited signals that are orthogonal to each other!
Let xc (t) denote the CT signal that is assumed to be strictly bandlimited, ie. X(j!) = 0 for all |!| !0 .
Let xd [n] = xc (nT ) denote the discrete time signal obtained by sampling xc (t) using periodic pulse train
sampling with period T .
All through the discussion we will assume that the sampling period T is such that !s = 2⇡
T 2!0 .
Let this signal be filtered via a DT filter hd [n] and let yd [n] = xd [n] ⇤ hd [n].
P
Let yp (t) = n yd [n] (t nT ) denote the continuous time signal obtained by converting the discrete
time signal yd [n] to a CT pulse train. Let hlp (t) denote the impulse response for the ideal low pass filter
with gain T and cut-off !2s and let yc (t) = yp (t) ⇤ hlp (t) denote the CT filtered output.
We want to find the relation between the effective CT impulse response h(t) of the end-2-end CT
system
xc (t) ! h(t) ! yc (t)
and the DT filter hd [n] for the system shown in Figure 1.
3
Let ⌦ = !T . Note: The discrete frequency (which we will now denote by ⌦) has no units! Then we have,
X
Xp (j!) = xd [n]e j⌦n (13)
n
, Xd (ej⌦ ) (14)
1X
Xp (j!) = Xc (j(! n!s ))
T n
where !s = 2⇡
T is the sampling frequency. Therefore, we have
✓ ✓ ◆◆
1X ⌦ 2⇡n
Xd (ej⌦ ) = Xc j (15)
T n T
Then note that Yd (ej⌦ ) = Xd (ej⌦ )Hd (ej⌦ ) and by the same relation as above we can see that,
✓ ✓ ◆◆
j⌦ 1X ⌦ 2⇡n
Yd (e ) = Yc j (16)
T n T
1 X
= Yc (j(! n!s )) (17)
T n
where
⇢
Hd (ej!T ) |!| < !2s
Hc (j!) = (18)
0 otherwise
Note that we can only define the CT filter for the range !s < ! < !s . Let us look at this entire process in
pictures - See Figure 2 on next page.
4
Figure 2: Schematic illustrating the relations between CT and DT frequencies for the system of Figure 1.
From this figure we can deduce the relation between the CT and DT filters.
5
3 Sampling and recovery of DT bandlimited signals
In this case, it can again be seen that the same result, that is, the Nyquist Sampling criteria holds true.
Let x[n] be multiplied with a periodic pulse train (with period N ) p[n] to obtain xp [n] = x[n]p[n]. Then
using the fact that for the discrete time signals, multiplication in the time domain is equivalent to periodic
convolution in the frequency domain, we get
N 1
1 X
Xp (ej! ) = X(ej(! k!s )
)
N
k=0
This means that there are N copies of the X̂(ej! ) in any interval of length 2⇡. The Nyquist Sampling
theorem can be deduced from Figure 3 on the next page.
It can be seen that using a low pass filter with an appropriate cut-off one can recover the original signal
back. Look at Equations 7.44 - 7.46.
6
x[n] ! x(N ) [n] ! Low Pass Filter ! xu [n]
7
Figure 3: DT sampling.