Ideal Sampling of Continuous-Time Signals

Periodic Sampling
Continous-time signals are usually sampled periodically every T seconds:

x[n] = xc (nT ) 1<n<1

1
Since T is the sampling period, its reciprocal Fs = T (samples/sec) is the sampling
frequency, or in radians per second: ⌦s = 2⇡T .
Sampling With a Periodic Impulse Train
1
X 1
X
xs (t) = xc (t) (t nT ) = xc (nT ) (t nT )
n= 1 n= 1
Frequency-Domain Perspective of Sampling
The impulse train that we used in the model of the sampling process,
1
X
s(t) = (t nT )
n= 1

has a Fourier transform

1
2⇡ X
S(j⌦) = (⌦ k⌦s )
T
k= 1
Frequency-Domain Perspective of Sampling (Continued)

The continuous-time signal, xc (t), has a Fourier-transform represented by this

arbitrary spectrum:

Notice this spectrum is bandlimited:

Xc (j⌦) = 0 for |⌦| ⌦N

Frequency-Domain Perspective of Sampling (Continued)
In the sampling process, we multiply the impulse train, s(t), the continuous-time
signal, xc (t), which is equivalent to convolving their spectra:
1
Xs (j⌦) = Xc (j⌦) ⇤ S(j⌦)
2⇡

Spectrum of sampled signal:

1
1 X
Xs (j⌦) = Xc (j (⌦ k⌦S ))
T
k= 1
Sampling Rate to Avoid Aliasing Distortion
⌦s ⌦N ⌦N , or ⌦s 2⌦N , Fourier transform of sampled signal:

⌦s < 2⌦N , causes aliasing:

Nyquist-Shannon Sampling Theorem

Let xc (t) be a bandlimited signal with

Xc (j⌦) = 0 for |⌦| ⌦N .

Then xc (t) is uniquely determined by its samples x[n] = xc (nT ), n = 0, ±1, ±2, ...,
if
2⇡
⌦s = 2⌦N .
T
⌦N is the Nyquist frequency.
2⌦N is the Nyquist rate.
Relationship Between Continuous-Time and Discrete-Time
Frequency Variables with Periodic Sampling
Discrete-Time Sinusoidal Signals

x[n] = A cos (!n + ✓) = A cos (2⇡f n + ✓) where n is an integer

I A discrete-time sinusoid is periodic only if its frequency f is a rational number:

f = Nk , where k and N are integers (common factors cancelled, N is period)
I Discrete-time sinusoids whose frequencies are separated by an integer multiples
of 2⇡ are identical: cos [(! + 2⇡) n + ✓] = cos (!n + 2⇡n + ✓) = cos (!n + ✓).
I The highest rate of oscillation in a discrete-time sinusoid is when frequency
w = ±⇡ (or f = ±1/2); fundamental range: ⇡  !  ⇡ ( 12  f  12 )
Relative or Normalized Frequency
Periodic sampling creates relationship between time variables t and n:
n
t = nT =
Fs
This leads to a relationship between the continuous-time frequency F (or ⌦) and
the discrete-time frequency f (or !), assuming the signal is sampled at Fs = 1/T
samples per second.

xc (t) = A cos (2⇡F t + ✓)

✓ ◆
2⇡nF
x[n] = xc (nT ) = A cos (2⇡F nT + ✓) = A cos +✓
Fs

F
f= or, ! = ⌦T
Fs
Example: Sampled Sinusoidal Signal
If we sample xc (t) = cos(4000⇡t) with a sampling period of T = 1/6000 we obtain:
2⇡
x[n] = xc (nT ) = cos (4000⇡T n) = cos 3 n

The sample rate is ⌦s = 2⇡/T = 12000⇡ rad/sec which is more than twice the
frequency of the original signal so there is no aliasing.

The DTFT is a function

of the normalized
frequency: ! = ⌦T
Example: Sampling Two Analog Signals

Determine the discrete-time signals from sampling analog signals at fs = 40 Hz:

x1 (t) = cos (2⇡10t)

x2 (t) = cos (2⇡50t)
n n
replace t = nT = fs = 40 :

x1 [n] = cos 2⇡ 10
40 n = cos
⇡
2n
x2 [n] = cos 2⇡ 50
40 n = cos
5⇡
2 n

5⇡ 5⇡ ⇡
However, cos 2 n = cos cos 2 n 2⇡n = cos 2n so x2 [n] = x1 [n].

50 Hz is an alias of 10 Hz when the sampling frequency is 40 Hz

Reconstruction of a Signal From Its Samples
Recovery of Signal From Its Samples: Frequency Domain
Recovery of Signal From Its Samples: Time Domain

Sampled signal represented by impulse train of successive values:

1
X
xs (t) = x[n] (t nT )
n=1
Reconstructed signal is output of ideal lowpass filter, with impulse response, hr (t):
1
X
xr (t) = xs (t) ⇤ hr (t) = ... = x[n]hr (t nT )
n=1
Impulse Responses of Ideal Reconstruction Filter
Cuto↵ frequency:

⌦c = ⌦s /2 = ⇡/T

Impulse response:

sin(⇡t/T )
hr (t) =
⇡t/T
= sinc(t/T )

Using normalized sinc function:

sin(⇡x)
sinc(x) =
⇡x
Reconstruction of the Original Signal

Output of reconstruction filter:

1
X
xr (t) = x[n]hr (t nT )
n=1

where impulse response is

hr (t) = sinc(t/T )

Output:
1
X
xr (t) = x[n] sinc ((t nT )/T )
n=1
Discrete-Time Processing of Continuous-Time Signals
Summary of Mathematical Representations

x[n] = xc (nT )
1
1 X ⇥ ⇤
X(ej! ) = Xc j !
T
2⇡k
T
T
k= 1
1
X
yr (t) = y[n] sinc [(t nT )/T ]
n= 1
(
j⌦T T Y (ej⌦T ), |⌦| < ⇡/T
Yr (j⌦) = Hr (j⌦)Y (e )=
0, otherwise
Overall Frequency Response

For the LTI discrete-time system:

Y (ej! ) = H(ej! )X(ej! )

The Fourier-transform of the output of entire system is:

Yr (j⌦) = He↵ (j⌦)Xc (j⌦)

where (
H(ej⌦T ), |⌦| < ⇡/T
He↵ (j⌦) =
0, |⌦| ⇡/T
Example: Ideal Lowpass Filter
( (
j! 1, |!| < !c 1, |⌦T | < !c or |⌦| < !c /T
H(e ) = He↵ (j⌦) =
0, !c < |!|  ⇡ 0, |⌦T | !c or |⌦| !c /T
Example: Ideal Lowpass Filter (Continued)
Impulse Invariance for Implementing Continuous-Time
System in Discrete-Time
Impulse Invariance (Given a Desired Continuous-Time System)

Desired: hc (t), Hc (j⌦)

Choose:

H(ej! ) = Hc (j!/T ), |!| < ⇡

Select T such that:

Hc (j⌦) = 0, |⌦| ⇡/T

Then the impulse-invariant

version of hc (t) is:

h[n] = T hc (nT )
Example: Lowpass Filter Using Impulse Invariance
Use the impulse invariance technique to find an ideal lowpass discrete-time filter
with cuto↵ frequency !c < ⇡, given this continuous-time ideal lowpass filter:
(
1, |⌦| < ⌦c
Hc (j⌦) = where ⌦c = !c /T < ⇡/T
0, |⌦| ⌦c

The inverse Fourier transform is the impulse response of this system:

✓ ◆
sin (⌦c t) ⌦c ⌦c t
hc (t) = = sinc
⇡t ⇡ ⇡
Finding the impulse-invariant system h[n] = T hc (nT ):
(
sin (⌦c nT ) sin (!c n) 1, |!| < !c
h[n] = T hc (nT ) = T = ! H(ej! ) =
⇡nT ⇡n 0, !c  |!|  ⇡
Example: Rational System Function Using Impulse Invariance
Use the impulse invariance technique to find discrete-time system to implement:

hc (t) = Aes0 t u(t)

Applying the impulse invariance method: h[n] = T hc (nT ) = AT es0 T n u[n]

Finding z-transform:
AT
H(z) = 1 |z| > |es0 T |
1 e s0 T z
Gives frequency response, assuming Re(s0 ) < 0:

AT A
H(ej! ) = 6= Hc (j T! ) =
1 e s0 T e j! j T! s0

(original continuous-time system did not have bandlimited frequency response)

Sample Rate Conversion: Downsampling
Sample Rate Reduction: Downsampling

xd [n] = x[nM ] = xc (nM T )

I Reduce sample rate by integer factor M

I Same as sampling xc (t) if Td = M T ,
I Assume Xc (j⌦) = 0 for |⌦| ⌦N
I Using mapping ⌦ = !/T : No aliasing if ⇡/Td = ⇡/(M T ) ⌦N
I Can reduce sampling rate by factor of M without aliasing if original sample
rate is at least M times Nyquist rate (or bandwidth reduced by factor of M )
Downsampling: Frequency Domain Perspective
DTFT of x[n], the original sampled signal:
1
1 X ⇥ ⇤
X(ej! ) = Xc j !
T
2⇡k
T
T
k= 1

DTFT of xd [n] = x[nM ], in terms of Xc (j⌦):

1
1 X ⇥ ⇤
Xd (ej! ) = Xc j !
MT
2⇡r
MT
M T r= 1

or DTFT of xd [n] = x[nM ], in terms of the spectrum of x[n]

1 X ⇣ j(!/M ⌘
M 1
j! 2⇡i/M )
Xd (e ) = X e
M
i=0
Example: Downsampling, M = 2, With No Aliasing

To avoid aliasing, ensure X(ej! )

is bandlimited:

X(ej! ) = 0, !N  |!|  ⇡

and 2⇡/M is above the Nyquist

rate:
2⇡/M 2!N
or
⇡/M !N
Example: Downsampling, M = 3, With Aliasing and Prefilter
Highest frequency:
!N = ⌦N T

To avoid aliasing:
!M  ⇡ or !N  ⇡/M

In this example: M !N = 3⇡/2 ⇥ ⇡

Decimation: Lowpass filtering followed

by compression
Sample Rate Conversion: Upsampling and Changing Sample
Rate by a Rational Factor
Sample Rate Reduction: Upsampling
Goal is to obtain: xi [n] = xc (nTi )

where Ti = T /L, from original sequence: x[n] = xc (nT )

( 1
x[n/L], n = 0, ±L, ±2L, .... X
xe [n] = or xe [n] = x[k] [n kL]
0, otherwise k= 1
Upsampling: A frequency Domain Perspective
After upsampling:
1
X
xe [n] = x[k] [n kL]
k= 1

Fourier transform of upsampled sequence:

1 1
!
X X
j! j!n
Xe (e ) = x[k] [n kL] e
n= 1 k= 1
X1
j!Lk
= x[k]e = X(ej!L )
k= 1

DTFT is frequency scaled version of input:

! ! !L so that !L = ⌦T and ! = ⌦(T /L) = ⌦Ti

Example: Upsampling (Interpolation)
Upsampling: Interpolation Formula
Output of expander and input to lowpass filter:
1
X
xe [n] = x[k] [n kL]
k= 1

Impulse response of lowpass filter:

sin(⇡n/L)
hi [n] = note: hi [0] = 1 and hi [n] = 0, n = 0, ±L, ±2L, ....
⇡n/L

Output of lowpass filter:

1
X sin(⇡(n kL)/L)
xi [n] = x[k]
⇡(n kL)/L
k= 1

With this ideal filter: xi [n] = x[n/L] = xc (nT /L) = xc (nTi ), n = 0, ±L, ±2L, ...
Simple and Practical Interpolation Filters
Linear Interpolation:

(
1 |n|L, |n|  L
hlin [n] =
0, otherwise

FIR Filter for interpolation, where h̃i [n] = 0 for |n| KL:

n+KL
X 1
x̃i [n] = xe [k]h̃i [n k]
k=n KL+1
Changing Sample Rate by Rational Factor

If M > L: increase in sample period (decrease sampling rate).

Multirate Signal Processing
Interchange of Lowpass Filter with Compressor

These two systems are From downsampling, output of second system:

equivalent: M 1
j! 1 X
Y (e ) = Xb (ej(!/M 2⇡i/M )
)
M
i=0
M
X1
1
= X(ej(!/M 2⇡i/M )
)H(ej(! 2⇡i)
)
M
i=0

Since H(ej(! 2⇡i) ) = H(ej! ):

M 1
1 X
Y (ej! ) = H(ej! ) X(ej(!/M 2⇡i/M )
)
M
i=0
Xb (ej! ) = H(ej!M )X(ej! )
= H(ej! )Xa (e ) j!
Interchange of Lowpass Filter with Expander

These two systems are

Spectrum at expander output (2nd system):
equivalent:
(Linear filter modified) Xb (ej! ) = X(ej!L )

Spectrum at output (1st system):

Y (ej! ) = Xa (ej!L )
= X(ej!L )H(ej!L)
= H(ej!L )Xb (ej! )

(which corresponds to output of second system)

Multistage Decimation To Reduce Computation
Two-stage decimation system with overall rate reduction M = M1 M2 :

Equivalent single-stage lowpass filter: H(z) = H1 (z)H2 (z M1 )

Single-stage required lowpass cuto↵ frequency: ⇡/(M1 M2 )
Two-stage required lowpass cuto↵ frequencies: ⇡/M1 and ⇡/M2
Multistage Interpolation To Reduce Computation
Two-stage interpolation system with overall sample rate increase L = L1 L2 :

Equivalent single-stage lowpass filter:

H(z) = H1 (z L2 )H2 (z)

Prefiltering to Avoid Aliasing
Digital Processing of Analog Signals
Anti-Aliasing Filter to Avoid Aliasing

Remember frequency scaling: ! = ⌦T

Anti-aliasing filter:
(
1, |⌦| < ⌦c  ⇡/T Since Haa (j⌦) can’t be ideally
Haa (j⌦) = bandlimited, the overall
0, |⌦| ⌦c
frequency response is:
From xa (t) to yr (t):
He↵ (j⌦) ⇡ Haa (j⌦)H(ej⌦T )
(
H(ej⌦T ), |⌦| < ⌦c
He↵ (j⌦) =
0, |⌦| ⌦c
Oversampling to Simplify Anti-Aliasing Filter

I Simple anti-aliasing filter with gradual cuto↵ at ⌦N

I Significant attenuation at M ⌦N
I Sample at rate higher than 2⌦N (sample at 2M ⌦N )
I Implement sharp antialiasing filter in discrete-time domain when
downsampling by factor of M
I Subsequent discrete-time processing completed at lower sampling rate
Spectrum of Oversampling Followed by Decimation
Analog-to-Digital Conversion
Ideal Sample-and-Hold

Analog-to-digital conversion:
Analog-to-Digital Conversion System

Quantizer maps input sample to finite set of values:

x̂[n] = Q(x[n])

Coder will label each quantization level with binary code of (B + 1)-bits.
Quantizer for A/D Conversion (8 Levels with B + 1 bits, B = 2)

Quantizer step
size:
2Xm Xm
= B+1
= B
2 2
Example: Sampling, Quantization, Coding, D/A Conversion (3-bits)
Analysis of Quantization Errors
Quantization Error

Quantization error is di↵erence between quantized sample and true sample value:

e[n] = x̂[n] x[n]

If is the step-size and the input samples stay within the full range of

( Xm /2) < x[n]  (Xm /2)

then the quantization error will be between:

/2  e[n] < /2

otherwise, then |e[n]| > /2 and samples are clipped (quantizer is overloaded)
Model of Quantizer

I e[n] is sample sequence of stationary

random process
I e[n] is uncorrelated with x[n]
I e[n] is a white noise process
I uniform distribution over the range
of quantization error
Statistical Model for Quantization Noise
Probability density function of e[n], the quantization error using (B + 1) bits:

Variance (and quantization noise power):

Z /2 2
2 1 2 2B Xm
2 Xm
e = e2 de = = since =
/2 12 12 2B

Since autocorrelation is 2
ee [m] = e [m], the power spectral density is:
2 2B X 2
m
Pee (ej! ) = 2
e = |!|  ⇡
12
Signal-to-Quantization-Noise Ratio (SNR)

SNR for (B + 1)-bit uniform quantizer is

✓ 2◆ ✓ ◆
x 12 · 22B 2
x
SN RQ = 10 log10 2
= 10 log 10
e X2
✓ ◆ m
Xm
= 6.02B + 10.8 20 log10
x

I SNR increases about 6 dB for each bit added (doubling of quantization levels)
I To prevent last term from becoming large and negative, signal amplitude
should be matched to full-scale amplitude of A/D