SAMPLING AND DATA RECONSTRUCTION

Lalan Kumar
Assistant Professor
Multichannel Signal Processing Lab
Department of Electrical Engineering
Indian Institute of Technology Delhi

March 10, 2022

 Introduction

• We discussed about the DTFT, ZT- utlises DT signal. In practice signals

are continuous in time, then how to get DT signal?
• Sampling : Process of converting a continuous-time signal into a
discrete-time sequence. Why? → it enables the processing of signals by
digital computers.
• Objective : sampling, DT processing and the subsequent reconstruction
back into CT signal.
• Q. Given x(t), do we need to have x(t) at all times in order to represent
x(t) at all times? Ans → No.
• Q. Some part of x(t) might be redundant, so we can store only a portion
of x(t) without losing any information? Ans → Yes.
• Q. This might be useful when using computers? Ans → Yes for saving
storage.

SAMPLING AND DATA RECONSTRUCTION 2 / 38

 Periodic Sampling

• Continuous-to-discrete-time (C/D) converter using periodic sampling: x[n]

is obtained by extracting xc (t) every Ts .

• Mathematically described as x[n] = xc (nT ), ∞ < n < ∞.

• xc (t) is the input continuous time signal and x[n] is the output discrete
time signal. T is the sampling period.
• fs = T1 is the sampling frequency in samples/sec. Ωs = 2π T
is the sampling
frequency in radians/sec.
• Possible to recover xc (t) from x[n]? It depends. If xc (t) has certain
characteristics and T is small then yes, else, no.

SAMPLING AND DATA RECONSTRUCTION 3 / 38

 Continuous-time to Discrete-time (CD) converter

• The impulse train is

n=∞
X
s(t) = δ(t − nT )
n=−∞

n=∞
X
xs (t) = xc (t) δ(t − nT ) (1)
n=−∞
n=∞
X
= xc (nT )δ(t − nT ) (2)
n=−∞

SAMPLING AND DATA RECONSTRUCTION 4 / 38

 Continuous-time to Discrete-time (CD) converter

• Here x[n] = xc (nT ). The system that implements this is ideal CD

converter.
• In a practice, the operation of sampling is implemented by an analog-to
digital (AD) converter.
• CD converter requires only the sampling rate
• ADC additinally requires quantisation, linearity of quantisation steps and,
sample and hold ckt. i.e. CD+quantisation = ADC.

SAMPLING AND DATA RECONSTRUCTION 5 / 38

 Continuous-time to Discrete-time (CD) converter

Figure: Sampling illustration

• Sampling process mathematically is in the two stages, impulse train

modulator followed by conversion of the impulse train to a sequence.
• xs (t) = xc (t)s(t) consists of a sequence of CT impulse functions (dirac
delta funtion) uniformly spaced at intervals of T seconds and weighted by
the values of the signal xc (t) at the sampling instants.
• The area of such dirac delta function is the values of xc (t) at the sampling
instants.

SAMPLING AND DATA RECONSTRUCTION 6 / 38

 Continuous-time to Discrete-time (CD) converter

• Bcz area of dirac delta function is unity

⇒ Value of xc (t) at sample instant = area under xs (t)
• Hence, xs (t) is not defined at the sampling instants (∞), but
x[n] = xc (nT ) is well defined which is area under the curve of xs (t) around
nT .
• DT domain : x[n] the sequence defined by taking the values of xc (t) at
the sampling instants, i.e. x[n] = xc (t)|t=nT .
• xc (t) and x[n] are different but are equivalent models of sampling process
in CT and DT domain respectively.
• Their equivalence is established by the fact that they have equal spectra in
the Fourier domain, and that the underlying CT signal from which xs (t)
and x[n] are derived can be recovered from the sampling representation,
provided a sufficiently large sampling rate is used in the sampling
operation.

SAMPLING AND DATA RECONSTRUCTION 7 / 38

 Continuous-time to Discrete-time (CD) converter

• A fundamental question:

• Note that different analog signals mapping to same sequence

• Can x[n] be used to uniquely represent xc (t)?
• Answer : yes when:
(1) xc (t) is bandlimited such that its Fourier transform X (jΩ) = 0 for
|Ω| ≥ Ωb where Ωb is called the bandwidth.
(2) Sampling period is sufficiently small.

SAMPLING AND DATA RECONSTRUCTION 8 / 38

Frequency Domain Representation of Sampling : CD Converter
• We now study the frequency domain representations of the signals and
sequence in 1 for CD converter.
• The time-frequency pair for the two domain is : ω (radian/sample) and Ω
(radian/second) is xc (t) ↔ Xc (jΩ), s(t) ↔ S(jΩ), xs (t) ↔ Xs (jΩ),
x[n] ↔ X (jω),
• From (1)
1
Xs (jΩ) = Xc (jΩ) ∗ S(jΩ)
2π
" k=∞
#
1 2π X
= Xc (jΩ) ∗ δ(j(Ω − kΩs ))
2π T
k=−∞
Pn=∞ Pk=∞
• note: n=−∞ δ(t − nT ) ↔ T
2π
k=−∞ δ(j(Ω − kΩs ))
k=∞
1 X
Xs (jΩ) = Xc (jΩ) ∗ δ(j(Ω − kΩs )) (3)
T
k=−∞

k=∞
1 X
Xs (jΩ) = Xc (jΩ − kΩs ) (4)
T
k=−∞
• Xs (jΩ) can be interpreted as superimposed version of periodically repeated
Xc (jΩ)
SAMPLING AND DATA RECONSTRUCTION 9 / 38
Frequency Domain Representation of Sampling : CD Converter

Figure: Spectrum : The sampling Process

SAMPLING AND DATA RECONSTRUCTION 10 / 38
 CD Converter : Nyquist Sampling Theorem

• Let xc (t) be a band-limited signal with Xc (jΩ) = 0|Ω| > ΩN , Then xc (t)
is uniquely determined by its samples
x[n] = xc (nT )n = · · · − 2, −1, 0, 1, 2 . . . , if Ωs = 2π
T
> 2ΩN
• Where ΩN is called the Nyquist frequency and 2ΩN is called the Nyquist
rate.
• Ωs must exceed Nyquist rate for ideal reconstruction of xc (t).
Application fN = ΩN /2π fs = Ωs /2π
Biomedical <500 Hz 1 kHz
• Telephone speech <4 kHz 8 kHz
Music <20 kHz 44.1 kHz
Ultasonic <100 kHz 250 kHz
Radar <100 MHz 200 MHz
Table: Typical Bandwidths and sampling frequencies in signal processing applications

SAMPLING AND DATA RECONSTRUCTION 11 / 38

CD Converter: Relation between DTFT (X (e jω )) and CTFT(X (jΩ))
• xc (t) → xs (t) → x[n].
• Found relation between xc (jΩ) and xs (jΩ), Now X e jω and Xc (jΩ).


• xs (jΩ) ↔ xs (t)(= xc (t)s(t))
nX o
Xs (jΩ) = F xc (nT )δ(t − nT )
∞
X
= xc (nT )F (δ(t − nT ))
n=−∞
∞
X
= xc (nT )e −jΩnT (5)
n=−∞
∞
X
= x[n]e −jΩnT
n=−∞
 
Xs (jΩ) = X e jω
ω=ΩT
• Utilizing (4)
 
X e jω = X s (jΩ)|Ω= ω
T
  
  1 X ω k2π
X e jω = Xc j −
T T T
SAMPLING AND DATA RECONSTRUCTION 12 / 38
DC Converter : Reconstruction of Band-Limited Signal from Samples
• Ideal reconstruction using discrete-to-continuous time (D/C) converter is
shown with T and x[n] as input.
• T is utilized in reconstruction filter.
• CD : xc (t) → xs (t) → x[n]
• Mathematically, DC - opposite of sampling : x[n] → xs (t) → xr (t)

Figure: DC Block

• Having Xs (jΩ), the signal can be recovered from Figure 2 using Hr (jΩ) as
Hr (jΩ) = T |Ω| < Ωc
SAMPLING AND DATA RECONSTRUCTION 13 / 38
= 0, o/w
DC Converter : Reconstruction of Band-Limited Signal from Samples

• For simplicity, we set Ωc as the average of ΩN and Ωs − ΩN .

Ωs π
Ωc = = (6)
2 T
• To get hr (t), we take inverse Fourier transform of Hr (jΩ).
•
Z π
T sin( πt
Z ∞
1 1 T ) t
hr (t) = H(jΩ)e jΩt dΩ = Te jΩt dΩ = T
= sinc( )
2π −∞ 2π − π πt T
T
(7)

SAMPLING AND DATA RECONSTRUCTION 14 / 38

DC Converter : Reconstruction of Band-Limited Signal from Samples
• Looking at the D/C block diagram in Figure 24, the reconstructed signal

xr (t) = xs (t) ∗ hr (t)

( n=∞ )
X
= x[n]δ(t − nT ) ∗ hr (t)
n=−∞
Z ∞ ∞
X
= x[n]δ(τ − nT )hr (t − τ )dτ
−∞ n=−∞
∞
X
= x[n]hr (t − nT )
n=−∞
∞
X t − nT
= x[n]sinc( ) for all real t
n=−∞
T

∞
X t − nT
xr (t) = x[n]sinc( ) (8)
n=−∞
T
• Eq.8 is Interpolation. The sinc is interpolating function for xr (t),
interpolating from the discreter sequence x[n].
SAMPLING AND DATA RECONSTRUCTION 15 / 38
DC Converter : Reconstruction of Band-Limited Signal from Samples

• Q. xr (t) = xc (t) ?
• Using interpolation formula at t=kT:
∞
X
xr (kT ) = x[n]sinc(k − n) (9)
n=−∞

But
sinc(k − n) = δ(k − n) (10)
• Putting Eq.10 in Eq.9, we have

xr [n] = x[n] = xc [nT ]

• Where samples are required, the reconstructed value is always the same as
the original value.

SAMPLING AND DATA RECONSTRUCTION 16 / 38

DC Converter : Reconstruction of Band-Limited Signal from Samples
• At other locations along time, xr (t) is same as xc (t) only when there is no
aliasing.
• Mathematically : The reconstruction using (8)
n=∞
X
xr (t) = x[n]hr (t − nT )
n=−∞

or
n=∞
X
Xr (jΩ) = x[n]Hr (jΩ)e −jΩnT
n=−∞
n=∞
X
= Hr (jΩ) x[n]e −jΩnT
n=−∞

= Hr (jΩ)X (e jω )
ω=ΩT

Xr (jΩ) = Hr (jΩ)X (e jΩT )

Xr (jΩ) = Hr (jΩ)X (jΩ)

• NB : X (e jω ) has 1/T that is compensated by H(jΩ)

SAMPLING AND DATA RECONSTRUCTION 17 / 38
DC Converter : Reconstruction of Band-Limited Signal from Samples

• X (jΩ) is frequency scaled (ω replaced by ΩT ).

• Hr (jΩ) selects the base period of the resulting periodic fourier transform
X (e jωT ) and compensates for the T1 scaling inherent in the sampling.
• Thus, if the sequence x[n] been obtained by sampling a band-limited signal
at the Nyquist rate or higher, then no aliasing is there in frequency
spectrum, the LPF preserves the spectrum and xr (t) will be equal to
band-limited signal.
• The band of Xr (jΩ) is always limited to Ωc of Hr (jΩ) regardless of
whether there is aliasing or not.

SAMPLING AND DATA RECONSTRUCTION 18 / 38

 DT Processing of CT Signal

• Sample a continuous signal, process it in discrete time and then

reconstruct the final continuous
• Even though the process Is done in discrete time domain, the entire
System is equivalent to a CT system.
• We assume in the figure that the C/D and D/C converter have the same
sampling rate but it is not necessary.

SAMPLING AND DATA RECONSTRUCTION 19 / 38

 DT Processing of CT Signal

• For C/D converter:

x[n] = xc (nT )
k=∞
1 X
X (e jω ) = Xc (j(Ω − kΩs ))
T Ω= ω
T
k=−∞

= Xs (jΩ)
• For D/C converter: Having just Hr
k=∞
X t − nT
yr (t) = y [n]sinc( ) (11)
T
k=−∞

Yr (jΩ) = Hr (jΩ)Y (e jΩT ) (12)

TY (e jΩT ), for |Ω| < Tπ


Yr (jΩ) = (13)
0, else

SAMPLING AND DATA RECONSTRUCTION 20 / 38

 DT Processing of CT Signal

• Let H(e jω ) be the frequency response of an LTI discrete time system, i.e.
Y (e jω ) = H(e jω )X (e jω ) . Substituting in (12), we have

Yr (jΩ) = Hr (jΩ)H(e jΩT )X (e jΩT ) (14)

k=∞
1 X 2πk
= Hr (jΩ)H(e jΩT ) Xc (j(Ω − )) (15)
T T
k=−∞

• Then for no aliasing case.

Ωs > 2ΩN
2π
> 2ΩN
T
π
ΩN <
T
• So, Xc (jΩ) = 0 for |Ω| ≥ π
T
, a band limited.

SAMPLING AND DATA RECONSTRUCTION 21 / 38

 DT Processing of CT Signal

• Hr (jΩ) of the lowpass filter selects only the k=0 term in (15).
• (
H(e jΩT )Xc (jΩ), |Ω| < Tπ
Yr (jΩ) =
0, |Ω| ≥ Tπ
Let
(
H(e jΩT ), |Ω| < π
T
Heff (jΩ) = π
(16)
0, |Ω| ≥ T

• Therefore, Yr (jΩ) = Heff (jΩ)Xc (jΩ)

SAMPLING AND DATA RECONSTRUCTION 22 / 38

 DT Processing of CT Signal

• This means the effective continuous time system is also LTI, whose
frequency response is Heff (jΩ).
• In the derivation, we assumed the discrete time system is LTI (time
domain convolution and frequency domain multiplication) and also we
assumed no aliasing occured.
• Q. What are sufficient conditions for the effective continuous system to be
LTI?
1. The discrete time system has to be LTI.
2. Input must be bandlimited.
3. Sampling rate must be high enough above Nyquist rate so that there is no
aliasing occuring.
4. If input xc (t) is not precisely bandlimited → a low-pass filter or anti-aliasing
filter is needed to process.

SAMPLING AND DATA RECONSTRUCTION 23 / 38

 DT Processing of CT Signal
• Q. DT processing of CT signal with
(
jω 1, |ω| < ωc
H(e ) = (17)
0, ωc < |ω| < π

Plot for Heff (jΩ) and H(e jω )?

• Cut-off frequency for reconstruction filter Ωc = ΩN +Ω2s +ΩN = Ω2s = Tπ
• As Heff (jΩ) comes as product of Hr (jΩ) and H(e jω ) or H(e jΩT ), and
because H(e jω ) is non-zero only for |ω| < ωc , Heff (jΩ) will be non zero for
|ω| < ωTc
• Utilising (16), Heff (jΩ) for the given system is given by replacing ω by ΩT
in (17)

SAMPLING AND DATA RECONSTRUCTION 24 / 38

 DT Processing of CT Signal
• Let us understand the DT processing of CT signal graphically step by step
for the above given signal? Assumption : no aliasing and band limited.

SAMPLING AND DATA RECONSTRUCTION 25 / 38

 DT Processing of CT Signal

•
• now passing the signal through a low pass reconstruction filter, bandwidth
of the reconstruction filter is taken to be Tπ

SAMPLING AND DATA RECONSTRUCTION 26 / 38

 DT Processing of CT Signal

• Comparing Figure (a) and (f), it is clear that Heff acts a low pass filter.
• By varying the sampling period T (in C/D part), an equivalent continuous
time low pass filter (Heff ) with variable cut-off frequency can be
implemented.
• The cut-off frequency for such LPF Ωc = ωTc where ωc is the cut-off
frequency of the DT system.
• The system behaves as an LTI system with frequency response given by
Heff (jΩ), bcz o/p is Xc Heff .

SAMPLING AND DATA RECONSTRUCTION 27 / 38

 DT Processing of CT Signal

• See the above Figure (c), if T were chosen such that ΩN T < ωc then o/p
of the system of figure (7) would be yr (t) = xc (t), i.e. no distortion.
• For ΩN T > ωc , Nyquist requirement for DT processing of CT signal, From
figure (c)
(2π − ΩN T ) > ωc
Rather than the original Nyquist criteria

(2π − ΩN T ) > ΩN T
• We can tolerate some aliasing while sampling as long as the distorted part
are eliminated by the DT system H(e jω ).
• (16) is valid even if there is aliasing in CD converter, as long as H(e jω )
does not pass the aliased component-Prob 4.31, oppeinheim 3rd edition.
• (2π − ΩN T ) < ωc , aliasing while sampling, can not be filtered by the DT
system, and (16) or Yr (jΩ) = Heff (jΩ)Xc (jΩ) is no more valid, bcz the
eqn assumes only zeroth spectrum of Xc (j(Ω − kΩs ))

SAMPLING AND DATA RECONSTRUCTION 28 / 38

 Continuous Processing of Discrete-Time Signals

• Complementary of discrete processing of continuous-time signals.

• Discrete time signals to a continuous signal, process it in continuous-time,
and then sample the continuous output and get another discrete time
signal, as shown below

• Even though the process is done in continuous-time domain, the entire

system is equivalent to a discrete time system. .

SAMPLING AND DATA RECONSTRUCTION 29 / 38

 Continuous Processing of Discrete-Time Signals

• Although the system of the fig. is not typically used to implement DT

system, it provides a useful interpretation of certain DT systems that have
no simple interpretation in discrete domain
• From the defination of ideal D/C converter, Xc (jΩ) and Yc (jΩ) are zero
for |Ω| ≥ Tπ (basically band-limited assumption).
• Only C/D should samples yc (t) without aliasing.
• From the D/C converter:
n=∞
X sin[ π(t−nT
T
)
]
xc (t) = x[n] π(t−nT )
n=−∞ T
n=∞
X sin[ π(t−nT
T
)
]
yc (t) = y [n] π(t−nT )
n=−∞ T

x[n] = xc (nT ), y [n] = yc (nT )

SAMPLING AND DATA RECONSTRUCTION 30 / 38

 Continuous Processing of Discrete-Time Signals
Frequency Domain Relation
• D/C (x[n] to x(t)) has reconstruction filter with gain as T and bandwidth
as ω/T
π
Xc (jΩ) = TX (e jω ) |Ω| < (18)
T
• Following y [n]− > yc (t)− > xc (t)− > x[n] in frequency domain, and
utilizing (18), we have
 
1 ω
Y (e jω ) = Yc j (19)
T T
• Utilising
π
Yc (jΩ) = Hc (jΩ)Xc (jΩ) |Ω| < (20)
T
and (18), we have
Y (e jω ) = Hc (jω/T )X (e jω )|ω| < π (21)
• ⇒ Overall system behaves as a DT system with frequency response
 
jω ω
H(e ) = Hc j |ω| < π
T
• Applivcation : Non-integer sampling : book opeinheim.
SAMPLING AND DATA RECONSTRUCTION 31 / 38
 Up-Sampling and Down-Sampling

Why we need the sampling rate conversion?

• Digital CD quality, SR = 44.1 kHz
FM Radio broadcasting, SR = 32 kHz, BW =15kHZ
Digital audio tap, SR = 48 kHz
SDTV (Standard definition TV), SR = 13.5 MHz, which is greater than
the Nyquist rate by 12.3%, the maximum video frequency is 6 MHz.
HDTV, SR = 74.25 MHz
• Q. How to put the above signals in a single platform ?
Ans: Upsamling/downsampling

SAMPLING AND DATA RECONSTRUCTION 32 / 38

 Up-Sampling
• Upsampling : introducing sample(s) between two existing sample
• The additional samples could be zero/previous value/average
• Unwanted, not adding any extra/new information
• Some Observation

• Let us introduce 2 zeros

SAMPLING AND DATA RECONSTRUCTION 33 / 38

 Up-Sampling

• We are introducing zeros between two samples, This is called expander or

upsampler.
• As T↓, fs↑.
• Pictorially : expander or upsampler is represented by

upsamples by L by introducing L-1 zero for simplicity.

• Mathematically

n
xe (n) = x , n = KL
L
= 0, else

SAMPLING AND DATA RECONSTRUCTION 34 / 38

 Up-Sampling

• By introducing L-1 zeros, sampling period reduces by L, i.e T ∗ = T /L or

fs∗ = Lfs .
• for L=1, no zero is introduced, same as original samples

X (Z ) = x0 + x1 Z −1 + x2 Z −2 + . . .
• For L=2, one zero is introduced

Xe (Z ) = x0 + (0 × Z −1 ) + x1 Z −2 + (0 × Z −3 ) + x2 Z −4 . . .
⇒ Xe (Z ) = X (Z 2 )

• In general, Xe (Z ) = X (Z L ).
• Spectrum : Xe (e jω ) = X (e jωL )
• Upsampling in time domain by L leads to shrinking in frequency domain by
L

SAMPLING AND DATA RECONSTRUCTION 35 / 38

 Up-Sampling
• How does the upsampling affect the spectrum?

•
SAMPLING AND DATA RECONSTRUCTION 36 / 38
 Up-Sampling

• within [−π, π] there was only one spectrum as in ”a”,but because of

upsampling false spectrum/contour is introduced
• How to remove?

SAMPLING AND DATA RECONSTRUCTION 37 / 38

 Removing the False contours:
Problem with up-samplers : Introduction of false contours.

 
x[n/L], n = kL
xe (n) =
0, otherwise

Xe (e jω ) = X (e jωL )

• Because of compression in spectrum ,false contours set introduced in 0 to

π.
• For removing the false contours, we need to pass the upsampler output
with a LPF, called interpolation filter given by

|ω| ≤ πL
 
L,
Hi (e jω ) =
SAMPLING AND DATA RECONSTRUCTION 0, πL < |ω| < π 38 / 38
 Removing the False contours:
Why the amplitude L ?

SAMPLING AND DATA RECONSTRUCTION 39 / 38

 Removing the False contours:
Why the amplitude L ?
• Amplitude in X (e jω ) is T1 .
• Suppose xi [n] ↔ Xi (e jω ) then xi [n] must have been received by sampling
the continuous time (CT) signal with T ′ where

1 L
= orT ′ = T /L
T′ T
• Time period is reduced and sampling frequency is increased.
• Upsampling is always followed by an interpolation filter.

SAMPLING AND DATA RECONSTRUCTION 40 / 38

 Removing the False contours:

Why the amplitude L ?

• Suppose xic (t) ←→ xi [n] ←→ Xi (e jω ) the magnitude Xi (e jω ) is limited by
the new sampling period T ′ as T1′ = TL
• The new amplitude of L
T
is possible only when gain of Hi (e jw ) = L.
• Hence, Xe (e ) is followed by LPF (interpolation filter)
jω

• LPF of Xe (e jω ) is done to remove the frequency scaled images of Xc (jΩ)

except at the integer multiples of 2π.
• After upsampling we should still be able to reconstruct Xc (jΩ)
∴ Interpolation Filter

SAMPLING AND DATA RECONSTRUCTION 41 / 38

 Why called Interpolation Filter?
• Upsampling introduces zero
• Increased frequency of variation; high to zero, but the contour follow the
signal closely ⇒ interpolation

SAMPLING AND DATA RECONSTRUCTION 42 / 38

 Down-Sampling

• Remove certain samples ⇒ loss of information.

• Down sample by N ⇒ remove N-1 sampples.
• Mathematically

xd [n] = x[nN]
X
Xd (Z ) = x[nN]Z −n (22)

• x[n] : x0 x1 x2 . . . → X (Z ) = x0 + x1 Z −1 + x2 Z −2 + . . .
• For N=2 xp [n] : x0 0 x2 . . . → Xp (Z ) = x0 + x2 Z −2 + . . .
• xd [n] : x0 x2 . . . → Xd (Z ) = x0 + x2 Z −1 + x4 Z −2 . . .
• Xd (Z ) = Xp (Z 1/2 ), or in general, Xd (Z ) = Xp (Z 1/N )
• we we dont have relation between Xd (Z ) and X (Z )

SAMPLING AND DATA RECONSTRUCTION 43 / 38

 Down-Sampling
How to generate xp [n] given x[n]
• Downsampling ⇒ Sampling the DT signal (taking some and leaving some)
• In analogy with CT sampling, the DT sampling is performed by
multiplying with impulse train p[n] = ∞
P
P k=−∞ δ[n − kN] similar to
δ(t − nT ) for CT case. For N=3, p[n] :

• xp [n] = x[n]p[n], the intermediate one.

SAMPLING AND DATA RECONSTRUCTION 44 / 38

 Down-Sampling
Relation between Xp (Z ) and X (Z )
• We have relation between Xd (Z ) and Xp (Z )
•
X∞
p[n] = δ[n − kN] (23)
k=−∞
N−1
1 X −kn
• Our Claim : p[n] given in (23) and p[n] = WN both are same,
N
k=0
j2π
with WN = e − N .
• For N=3 p[0] = 31 [1 + 1 + 1] = 1
j2π j4π
p[1] = 13 [1 + e − 3 + e− 3 ]=The Three vector sum as in the following
figure =0

j4π j8π
• p[2] = 13 [1 + e − 3 + e − 3 ] = 0
• p[3] = 1, Same, so use the alternative, rather than (23)
SAMPLING AND DATA RECONSTRUCTION 45 / 38
 Down-Sampling

Relation between Xp (Z ) and X (Z )

•
∞
X
Xp (Z ) = x[n]p[n]Z −n
n=−∞
∞ N−1
X 1 X −kn
= WN x[n]Z −n
n=−∞
N
k=0
N−1 ∞
1 X X
= x[n](ZWNk )−n
N
k=0 n=−∞
N−1
1 X
= X (ZWNk )
N
k=0

N−1
1 X
• ∴ Xp (e jω ) = X (e j(ω−kωs ) ) with ωs = 2π
N
, repeated n shifted.
N
k=0

SAMPLING AND DATA RECONSTRUCTION 46 / 38

 Down-Sampling
1
• Utilizing the relation Xd (Z ) = Xp (Z N ), we have
N−1
1 X ω 2kπ
Xd (e jω ) = X (e j( N − N ) )
N
k=0
N−1
1 X ω−2kπ
= X (e j( N
)
) (24)
N
k=0

Points To Be remembered
ω
• X (e j( N ) ) implies Spread in Frequency domain by N (X (e jω/N )) in
downsampling.
• The downsampling ≈ Sampling the DT signal ≈ multiplied by
1
PN−1 −kN
N k=0 WN .
N−1
1 X ω−2kπ
Xd (e jω ) = X (e j( N ) ) (25)
N
k=0
jω
• Periodicity of X (e N ) = 2Nπ, that of Xd (e jω ) remains 2π
• For Xd (e jω ), compute first Xp (e jω ) and then Xd (e jω )= Xp (e jω/N ) , or use
eq.(25) directly
SAMPLING AND DATA RECONSTRUCTION 47 / 38
 Down-Sampling
PN−1 2kπ )
• For N=2, downsample by 2 :Xp (e jω ) = 1
N k=0 X (e j(ω− N ), using the
first approach

• Certainly Xp (e jω ) is periodic with period 2π, but here π also an period. ∴

the fundamental period is 2π
SAMPLING AND DATA RECONSTRUCTION N
=π 48 / 38
 Down-Sampling

ω
• Now, Xd (e jω ) =Xp (e j 2 )
ω
• So Xp (e jω ) has periodicity π, then Xp (e j 2 ) has periodicity 2π with
spreaded frequency spectrum

SAMPLING AND DATA RECONSTRUCTION 49 / 38

 Down-Sampling
2nd approach : using (25)

SAMPLING AND DATA RECONSTRUCTION 50 / 38

 Down-sampling

Observation : Directly writing

• Magnitude by N1 (due to sampling of DT signal==down sampling), spread
the first spectrum by N, repeat with periodicity 2π
• Xp (e jω ) has periodicity 2π
N
ω
• Xd (e jω ) =Xp (e j N ) has periodicity 2π

SAMPLING AND DATA RECONSTRUCTION 51 / 38

 Problem in downsampling:

• Spreading in frequency by N, bcz of X (e jω/N ), there may be issue of

aliasing.
• Fig (b) : Assuming ΩN T = π/2,
• Fig (c) : downsampled by factor N=3, (replace M by N is figure.)
• Amplitude : (1/T) for CT to DT sampling, (1/N) for sampling of DT to
downsample

SAMPLING AND DATA RECONSTRUCTION 52 / 38

SAMPLING AND DATA RECONSTRUCTION 53 / 38
 Problem in downsampling

• a. CTFT of xc (t).
2π pi
• b. of x[n] = xc (nT ) where = 4ΩN or ΩN T =
T 2
• c. say downsample by factor of N=3; xd [n] = x[3n]: see aliasing in
Xd (e jω ).
• d. To avoid the aliasing in downsampling by factor of N requires
ΩN TN ≤ π (ΩN T from CT to DT, N from DT to downsampled) where
π
(ΩN T = ωN ) or ωN ≤
N
• The downsampling can be done without aliasing if we are willing to reduce
the bandwidth of signal x[n] before downsampling.
π
• If x[n] is pre-filtered by LPF with ωc = then x̃[n] can be downsampled
N
by factir N without aliasing.
• We have no aliasing but loosing certain part of spectrum due to LPF
(which was the case with aliasing as well).

SAMPLING AND DATA RECONSTRUCTION 54 / 38

 Problem in downsampling

• Such system of LPF followed by downsampler/compressor is called

decimator.
• x˜d [n] = x˜c (nTd ) where Td = NT

SAMPLING AND DATA RECONSTRUCTION 55 / 38

 Changing the sampling rate by non-integer factor:
• By combining decimator (LPF followed by downsampler) and interpolator
(Upsampler followed by LPF), it is possible to change the sampling rate by
non-integer factor.(Fig- opp.4.29)

• The interpolator decreases the sampling period from T to TL .

• The decimator increases the smpling period from TL to NT
L
.

SAMPLING AND DATA RECONSTRUCTION 56 / 38

 Changing the sampling rate by non-integer factor:

• Choosing L and N, we can approach close to any desired ratio of sampling

periods.
• In the following figure, x[n] → xe [n] → xi [n] → xd [n]
• Since the interpolation and decimation filters are in cascade, they can be
combined as

SAMPLING AND DATA RECONSTRUCTION 57 / 38

 Changing the sampling rate by non-integer factor:

• If N > L, then Nπ is the dominating cut- off frequency

• And, fs′ = T1 NL = fs . NL so, there is net reduction in sampling rate. Note
that x[n] was obtained by sampling at Nyquist rate and there is further
decrease in the sampling rate.
• The sequence x˜d [n] corresponds to a low pass filtered (π) version of the
original underlying bandlimited signal to avoid aliasing.
• If N < L, then πL is the dominant cut-off frequency and there will be no
need to further limit the bandwidth of the signal to avoid aliasing during
downsampling.

SAMPLING AND DATA RECONSTRUCTION 58 / 38

 Question/Comment/Feedback

Lalan Kumar

{lkumar}@[Link]

SAMPLING AND DATA RECONSTRUCTION 59 / 38