2 System Overview

2.1 Applications
Consumer:
CD
Computer Audio (DAT)
HDTV
Speech Recognition
Military:
Radar
Sonar
Guidance
Fundamental Technology
Industrial:
Control
Medical Imaging
Medical Signals (EEG, ECG, EMG)
Geophysical Exploration
Nondestructive Testing
Mechanical Monitoring and Diagnosis
COMMUNICATIONS
2.2 Drawbacks
Processing Bottlenecks
A/D
Processor
High Initial Cost; justied by
High performance
High complexity
Leverage
Flexibility
2.3 Advantages
Precision
Flexibility
Low incremental Cost
High capability
2.4 Digital Signal Processing System Structure
Anti-Aliasing Filter
Sample-and-Hold (or Track-and-Hold)
A-to-D Converter
Digital Processor
D/A converter
Smoothing Filter
Filters and conversion: extremely important; will cover later.
Conversion process sets a limit on the signal bandwidth possible usually < 1/2
sampling frequency.
Present focus on processing algorithms.
3 Sampled-Signal Processing Basics
3.1 Input-Output Description
Input: a sequence u = (u
n
); output y = (y
n
)
y = B(u)
Linear:
B(u +v) = B(u) +B(v)
B(u) = B(u)
for all constants .
2
Time-Invariant: If
(y
n
) = B((u
n
))
then
(y
n+k
) = B((u
n+k
))
for all k.
Fundamental fact: If the processing is linear and time-invariant then it must be a
convolution so convolution is the main task of DSP chips.
Convolution is dened by
(y
n
) = (u
n
) (h
n
) where y
n
=

k=
h
k
u
nk
and h
n
is the impulse response given by: (h
n
) = B((
n
))
Here (
n
) is the discrete-time unit (im)pulse:

n
=
_
1 n = 0
0 n = 0
Causal: h
n
= 0 for n < 0.
3.2 FIR & IIR Systems
Special convolution cases:
1. FIR ltering Finite Impulse Response, usually causal, simplest; a direct con-
volution whose impulse response is given by the lter coefcients:
y
n
= h
0
u
n
+h
1
u
n1
+h
2
u
n2
+h
M
u
nM
=
M

i=0
h
i
u
ni
2. IIR ltering Innite Impulse Response, causal, more complex than FIR; uses
past outputs:
y
n
= b
0
u
n
+b
1
u
n1
+ +b
M
u
nM
a
1
y
n1
a
2
y
n2
a
N
y
nN
3. FFT Fast Fourier Transform: usually noncausal, can be used for causal, usually
FIR
3
FIR Terminology:
y
n
= b
0
u
n
+b
1
u
n1
+ +b
M
u
nM
=
M

i=0
b
i
u
ni
Finite Discrete Convolution
Transversal Filter
Tapped Delay Line
Note: a DSP chips architecture must have fast multiply-adds, efcient shifting
(circular buffers), and zero-overhead (non-branching) looping; also usually has bit-
reversed addressing.
Other Algorithms:
FFT is also widely used for spectral analysis and spectral estimation
Adaptive processing is nonlinear/time-varying
Many coding/modulation and decoding/demodulation and all detection techniques
are nonlinear.
3.3 Processing Elements and Difference Equations
For Linear Time-Invariant processing, the basic processing elements are:
Unit Delays
Constant gains
Adders
Mathematically, interconnections of these elements give difference equations.
Schematically, we get block diagrams or signal ow graphs.
Note: z
1
is used to denote a unit delay, even in a time-domain block diagram or
signal ow graph!
Examples: FIR: moving average lter:
y
n
= 0.5u
n
+ 0.5u
n1
IIR: (digital) integrator:
y
n
= u
n
+y
n1
Block Diagrams: note that the FIR normally has no (directed) loops, but other
implementations of FIR lters can have loops.
Every loop in any discrete-time block diagram must have at least one delay.
4
Elementary sampled signals:
(Recall) unit (im)pulse:

n
=
_
1 n = 0
0 n = 0
Note: the unit impulse is an ordinary function in DSP. When simulating continuous-
time systems, be careful about its normalization!
Unit step:
U
n
=
_
1 n 0
0 n < 0
Check unit step is impulse response of (digital) integrator.
4 Frequency-Domain and Transforms
4.1 Sampled Sinusoids
Continuous-time:
x(t) = Acos(2ft +)
If signal is sampled every T seconds, the sampling instants are t = nT, and so the
sampled (discrete-time) sinusoid is:
x
n
= Acos(2fnT +)
Amplitude: A (note: not necessarily peak value!)
Phase:
Sampling period: T
Sampling frequency: f
S
= 1/T
Angular sampling frequency:
S
= 2/T
Note: the sequence (x
n
) is unchanged if the frequency f is changed to f + k/T,
for any integer k (aliasing).
Sinusoid in complex form:
g
n
= Ae
j(2f0nT+)
= Ae
j
_
e
j2f0T
_
n
where A is the amplitude, and is the phase.
For a decaying sinusoid:
Continuous-time:
g(t) = Ae
0t
e
j(2f0t+)
Sampled:
g
n
= Ae
0nT
e
j(2f0nT+)
= Ae
j
z
n
0
where
z
0
= e
(0+j0)T
= e
s0T
5
Consequence:
The relationship between the (continuous-time) s-domain and the (discrete-time)
z-domain is given by:
z = e
sT
Restrict to frequency domain:
z = e
j2fT
= e
j
So the discrete-time (normalized) angular frequency is given by
= 2fT == 2f/f
S
= T = /f
S
and the frequency information is on the unit circle.
Also, stability is equivalent to the absence of poles on or outside the unit circle.
4.2 Frequency, Amplitude, and Phase Responses
Suppose we have an FIR system with input u
n
= e
j2fnT
.
Then, with = 2f:
y
n
= b
0
e
jnT
+b
1
e
j(n1)T
+ +b
M
e
j(nM)T
= e
jnT
_
b
0
+b
1
e
jT
+ +b
M
e
MjT
_
= e
jnT
K(T)
So if u
n
is a (complex) sampled sinusoid of angular frequency ,
y
n
= K(T)e
jnT
= K(T)u
n
where the complex number K(T) = K() is given by
K(T) = b
0
+b
1
e
jT
+ +b
M
e
MjT
For an IIRsystem, if we also assume that the output y
n
is given by y
n
= Ke
j2fnT
,
where K is a complex constant:
Ke
jnT
+a
1
Ke
j(n1)T
+ +a
N
Ke
j(nN)T
= b
0
e
jnT
+b
1
e
j(n1)T
+ +b
M
e
j(nM)T
or
Ke
jnT
_
1 +a
1
e
jT
+ +a
N
e
NjT
_
= e
jnT
_
b
0
+b
1
e
jT
+ +b
M
e
MjT
_
6
so that
K(T) = K() = H(e
jT
) = H(e
j
)
=
b
0
+b
1
e
jT
+ +b
M
e
MjT
1 +a
1
e
jT
+ +a
N
e
NjT
For a general convolution, (with = 2f)
y
n
=

k=0
h
k
e
j(nk)T
= e
jnT

k=0
h
k
e
jkT
= e
jnT
H(e
jT
)
The function H(e
jT
) is called the frequency response of the system;
The function A(T) = |H(e
jT
)| is called the amplitude response;
The function (T) = H(e
jT
) is called the phase response.
Note:
These are periodic with period 2/T = 2f
S
=
S
They depend only on the ratio f/f
S
= /
S
For this reason, digital signal processing often works with a normalized frequency
f/f
S
.
4.3 DTFT and Z-transforms
Discrete-Time Fourier Transform (DTFT):
Based on the formula for the frequency response, the Discrete-Time Fourier Trans-
form of a (bounded) sequence (g
n
), is dened by:
G(e
j
) =

k=
g
k
e
jk
Note:
Boundednes means that there is some constant M such that |g
n
| M for all n.
If (g
n
) is the impulse response of a system, the frequency response of the system
is given in terms of the DTFT of (g
n
) by
G(e
j2fT
) = G(e
j
)|
=2fT
7
The inverse transform is given by:
g
n
= 1/2
_

G()e
jn
d
Proof: If G(e
j
) =

k=
g
k
e
jk
, then
_

G()e
jn
d =
_

k=
g
k
e
jk
e
jn
d
=

k=
g
k
_

e
jk
e
jn
d
=

k=
g
k
2
nk
= 2g
n
Note: the previous slide uses the fundamental fact that, for integer m,
_

e
jm
d = 2
m
This follows from:
If m = 0,
_

e
j0
d =
_

d = 2
If m = 0,
_

e
jm
d =
1
jm
_
e
jm

=
1
jm
[(1)
m
(1)
m
]
= 0
Two-sided Z-transform:
The two-sided Z-transform of a bounded sequence (g
n
) is dened by
G(z) = Z{g
n
}
=

n=
g
n
z
n
= +g
1
z +g
0
+g
1
z
1
+g
2
z
2
+
for all complex numbers z for which the series converges assumed to include the unit
circle.
8
Two-Sided Z-transform properties:
Linearity
Shifting: Z{g
nk
} = z
k
G(z) for all k
Convolution: If f
n
and g
n
are two bounded sequences, then Z{(f
n
) (g
n
)} =
F(z)G(z)
The Z-transform transforms convolution into multiplication.
Examples:
1. a
n
U
n
with |a| < 1
2. a
n
U
n1
with |a| > 1
3. sinusoids
DTFT and Z-transforms (continued)
Relation to frequency response and DTFT:
Frequency response is given by
K(2fT) =

k=
h
k
e
j2kfT
so that
K(2fT) = G(e
j2fT
)
= G(e
j
)|
=2f/f
S
= G(z)|
z=e
j2f/f
S
The main use of the two-sided Z-transform is in calculating the inverse DTFT of
functions which are rational in z.
The inverse DTFT of some other types of functions of is found by the integral
inverse formula (will see later).
DTFT and Z-transforms (continued)
The one-sided Z-transform of a sequence (g
n
) is dened by
G(z) = Z{g
n
} =

n=0
g
n
z
n
= g
0
+g
1
z
1
+g
2
z
2
+
for all complex numbers z for which the series converges (region of convergence, or
ROC).
Differs from the two-sided Z-transform and DTFT:
9
assumes that (g
n
) is one-sided
does not assume that (g
n
) is bounded
therefore does not necessarily converge on unit circle
converges in a region of the form {z | |z| > R} for some R.
handles initial conditions and transients.
DTFT and Z-transforms (continued)
Relation to frequency response, for a causal, stable system :
As before, frequency response is given by
K(2fT) =

k=0
h
k
e
j2fTk
so that
K(2fT) = H(e
j2fT
)
= H(z)|
z=e
j2fT
Here, stable means BIBOstable, which is equivalent to h
n
being absolutely summable,
i.e.,

k=0
|h
k
| is nite (stronger than bounded).
DTFT and Z-transforms (continued)
If g
n
is the impulse response, h
n
, of a causal system, its one-sided Z-transform
H(z) = Z{h
n
} =

n=0
h
n
z
n
is called the transfer function of the system
Examples:
1. a
n
U
n
2. sinusoids
3. exponentially decaying and growing sinusoids
DTFT and Z-transforms (continued)
(One-sided) Z-transform properties:
Linearity
Shifting: Z{g
nk
} = z
k
G(z) for k 0
10
Non-Causal Shift: Z{g
n+1
} = zG(z) zg
0
Convolution: If f
n
= 0 and g
n
= 0 for n < 0, then Z{(f
n
)(g
n
)} = F(z)G(z)
The Z-transform transforms convolution into multiplication.
Note 1: The noncausal shifting property enables the Z-transform to handle initial
conditions and transients.
Note 2: Because the signals are one-sided, the convolution here is given by:
p
n
=
n

k=0
f
k
g
nk
DTFT and Z-transforms (continued)
Inverse one-sided Z-transform:
Theoretical formula (rarely used, actually a contour integral):
g
n
=
1
2
_

G(Re
j
)(Re
j
)
n
d
where R is such that all poles of G(z) are inside the circle of radius R.
The normal method is to use partial fractions, but the calculations are simpler if
G(z)/z, instead of G(z), is written in terms of z (not z
1
) and is then expanded in
partial fractions.
Examples:
1. 1/(1 az
1
) with |a| < 1
2. 1/(1 az
1
) with |a| > 1
3. real poles stable
4. real poles unstable
5. complex poles stable and unstable
6. comples poles on unit circle second-order oscillator; (see p.394)
4.4 Poles and Zeros
Terminology (for rational functions with no common factors between numerator
and denominator):
1. Pole: a value of z where H(z) is innite, i.e., where the denominator is zero.
2. Simple Pole: a is a simple pole if there is a factor (z a) in the denominator,
and no factor of (z a)
k
with k > 1.
11
3. Multiple Pole: a is a multiple pole if there is a factor (za)
k
in the denominator,
with k > 1.
4. Zero: a value of z where H(z) is zero
5. Residue: the coefcient Ain the termA/(za) in the partial fraction expansion.
Poles and Zeros (continued)
1. All rational functions can be expanded as product of rst-order factors with com-
plex coefcients.
2. All rational functions can be expanded as product of rst- and second-order fac-
tors with real coefcients.
3. To have a partial fraction expansion, we must have H(z) 0 as z
4. Partial fraction expansions get more complicated with multiple poles; the basic
formulas can be found by differentiating the equation
1
1 az
1
=

n=0
a
n
z
n
Poles and Zeros (continued)
1. Knowing the poles gives us a good idea of what the impulse response will look
like.
2. Knowing the poles and using the initial- and nal-value theorems gives a good
idea of the step response.
3. The poles and zeros (product expansion) can be used to give a good idea of the
frequency response.
5 Signal Sampling and Reconstruction
5.1 Sampling and Reconstruction Theory
Theoretically, sampling is a multiplication operation:
Give a sampling signal s(t) which is periodic with period T, and a continuous-time
signal g(t), dene the sampled signal g
S
(t) by
g
S
(t) = s(t)g(t)
T is called the sampling period and f
S
= 1/T is the sampling frequency.
Normally, s(t) will approximate a periodic sequence of -functions, and a further
integration step is used to obtain a specic value.
Note that the sampled function is still a continuous-time function.
12
Sampling and Reconstruction Theory (continued)
For ideal sampling, we assume that s(t) is a periodic sequence of -functions:
s(t) =

n=
(t nT)
and the sampling operation then has two steps:
Multiply by the sequence of -functions:
g
S
(t) =

n=
(t nT)g(t) =

n=
g(nT)(t nT)
Pick off numbers representing samples
g
n
= g(nT)
Sampling and Reconstruction Theory (continued)
Effect of sampling in frequency domain:
G
S
(2jf) =

k=
c
k
G(2j(f kf
S
))
where c
k
are the complex Fourier coefcients of the s(t):
c
k
=
1
T
_
T/2
T/2
s(t)e
j2kt/T
dt
Here 2f
S
= 2/T =
S
is called the sampling angular frequency.
Also, f
S
/2 is sometimes called the folding frequency.
For ideal sampling, we get
G
S
(2jf) = 1/T

k=
G(2j(f kf
S
))
Sampling and Reconstruction Theory (continued)
Classical Sampling Criterion:
Must Sample at a rate higher than the Nyquist rate = 2f
MAX
WARNING!1! Nonlinear operations internal to the signal processing can increase
the bandwidth, and require much higher sampling rates.
WARNING!2! Aliased components are not just a mathematical abstraction, but
very real; you can see them on an oscilloscope, and hear them.
WARNING!3! Higher sampling rates can also be needed to reduce phase-shift, for
example, in a feedback loop.
13
Sampling and Reconstruction Theory (continued)
To reconstruct: use lowpass lter
Reconstruction is also possible in other cases:
Bandpass
Periodic
In time-domain:
g
R
(t) =

n=
g(nT)l(t nT)
where l(t) is the impulse response of the reconstruction (smoothing) lter.
For a band-pass signal in the band f
L
f f
H
, any sampling frequency, f
S
which allows ideal reconstruction satises, for some integer k 0
kf
S
2f
L
(k + 1)f
S
2f
H
Then the possible sampling frequencies are given by
2f
H
k + 1
f
S

2f
L
k
for each integer k 0 for which the interval is non-empty, where the second inequality
is omitted for k = 0 (the broad-band case).
These intervals are disjoint, since otherwise 2f
L
/k 2f
H
/((k 1) + 1) would
imply f
L
f
H
.
The integers k for which the intervals are non-empty must satisfy
2f
H
k + 1

2f
L
k
or, equivalently,
k(f
H
f
L
) f
L
and so
0 k k
0
=
_
f
L
f
H
f
L
_
The smallest such sampling frequency is given taking the largest k, and the smallest
frequency in the corresponding interval:
f
S(min)
=
2f
H
k
0
+ 1
14
Sampling and Reconstruction Theory (continued):
Ideal reconstruction: ideal low-pass lter with theoretical gain T; unrealizable
since
l(t) =
sin(t/T)
t/T
= sinc(t/T)
(Note that this vanishes at all sample points except t = 0, and so gives the correct value
at the sample points.)
Sampling and Reconstruction Theory (continued):
Actual D/A conversion: zero-order hold followed by smoothing lter. Frequency
response of zero-order hold is given by
H(f) = e
jf/f
S
sin(f/f
S
)
f/f
S
= e
jx
sin(x)
x
where x is the normalized frequency.
It follows that the zero-order hold has a group delay of T/2, and has a rough low-
pass characteristic with innite attenuation at non-zero multiples of the sampling fre-
quency, and an attenuation of 2/ or about 3.9 dB at f
S
/2.
5.2 Antialiasing (Input) and Smoothing (Output) Filters:
Antialiasing and Smoothing Filters:
Needed from sampling theorem
To reject spurious signals and noise at input
To reject aliased components at output
Problem: Need high order analog lter, especially at output
Can alleviate requirements at output:
Upconvert, lter digitally, high-rate D/A converter,
Upconvert, high-rate D/Aconverter, switched-capacitor lter, nal simple smooth-
ing lter.
15
Antialiasing and Smoothing Filters (continued)
Can also alleviate requirements at input:
Gentle analog lter, very high sampling rate low accuracy
Use high-frequency, low word-width Digital lowpass lter
Use DSP to Down-convert to lowsampling rate, high word-width noise-shaping
- converters.
5.3 Signal Conversion
Track and Hold Amplier:
Tracks signal until strobed, and holds constant value until released.
Stores charge on capacitor
Properties
Aperture
Acquisition time
Step
Droop
Jitter
Signal Conversion (continued)
A/D Converters:
Successive Approximation
Flash
Sigma-Delta
Dual-Slope, Counting, usually too slow
Signal Conversion (continued)
D/A Converters:
Numerous types; e.g., R-2R ladders, switched current source, etc.
Virtually always use zero-order hold.
16
Recall: zero-order hold has
amplitude response: sin(x)/x
1/2 sample period delay:.
Example:
What are the requirements for the smoothing lter in a CD player?
17