Solutions Manual

for
Statistical and Adaptive Signal Processing
This page is intentionally left blank.

ii
 Preface

This solutions manual provides solutions to the problems contained in the first
edition of our book STATISTICAL AND ADAPTIVE SIGNAL PROCESSING.
The solutions have been prepared and written by David Marden and ourselves. We
have attempted to provide very detailed solutions to the problems with notation
consistent with that used in the book. Where applicable, we have also given a
printout of Matlab code and figures for the problem solution.

Despite our efforts, however, you may find that some of the solutions may be less
detailed and refined than others. Inevitably through the use of this solutions man-
ual, omissions and errors will be discovered. Our goal is to continually improve
upon this manual using your comments. Periodically, we will issue changes to the
solutions manual and new problems to you upon request. In this respect, we would
appreciate any feedback including improved solutions, suggestions for new prob-
lems, corrections which can be sent to Prof. Vinay K. Ingle at [email protected]
or at

Prof. Vinay K. Ingle

Department of Electrical and Computer Engineering
Northeastern University
360 Huntington Avenue
Boston, MA 02115

Thank you.

Dimitris G. Manolakis
Vinay K. Ingle
Stephen K. Kogon

iii
This page is intentionally left blank.

iv
 Chapter 2
Discrete-Time Signals and Systems

2.1 Sampling frequency Fs = 100 sam/sec

(a) Continuous-time signal xc (t) = 2 cos(40π t + π /3) has frequency of 20 Hz. Hence


40π n
x(n) = xc (t)|t =n/Fs = 2 cos + π /3
100

which implies that ω0 = 40π

100
= 2π
5
.
(b) Steady-state response yc,ss (t): Given that h(n) = 0.8n u(n), the frequency response function is

1
H (e j ω ) =
1 − 0.8e− j ω
Since ω0 = 2π /5, the system response at ω0 is
1
H (e j ω ) = = 0.9343 e− j 0.2517π
1 − 0.8e− j 2π/5
Hence yss (n) = 2(0.9343) cos(2π n/5 + π /3 − 0.2517π ), or

yc,ss(t) = 1. 868 6 cos(40π t + 0.585π )

(c) Any xc (t) that has the same digital frequency ω0 after sampling and the same phase shift as above will
have the same steady state response. Since Fs = 100 sam/sec, the two other frequencies are 120 and 220
Hz.

2.2 The discrete-time signal is

x(n) = A cos(ω0 n) wR (n)

where wR (n) is an N -point rectangular window.

(a) The DTFT of x(n) is determined as

X (e j ω ) = F [ A cos(ω0 n) wR (n)]
   
= ( A/2) F e j ω0 wR (n) + (A/2) F e− j ω0 wR (n) (1)

Using the DTFT of wR (n) as


N−1
sin(ωN/2)
F [wR (n)] = e− j ωn = e− j ω(N−1)/2 (2)
n=0
sin(ω/2)

and the fact that complex exponential causes a translation in the frequency domain (1) can be written
after a fair amount of algebra and trigonometry as

X (e j ω ) = X R (e j ω ) + j X I (e j ω )

1
2 Statistical and Adaptive Signal Processing - Solution Manual

32−point DTFT (Real) 32−point DTFT (Imaginary)

20 15

15 10

5
10
0
5
−5

0
−10

−5 −15
−4 −2 0 2 4 −4 −2 0 2 4
Figure 2.2bc: Real and Imaginary DTFT and DFT Plots(ω0 = π/4)

where
A sin[(ω − ω0 )N/2]
X R (e j ω ) = cos[(ω − ω0 )(N − 1)/2]
2 sin[(ω − ω0 )/2]
A sin{[ω − (2π − ω0 )]N/2}
+ cos[(ω + ω0 )(N − 1)/2] (3)
2 sin{[ω − (2π − ω0 )]/2}
and
A sin[(ω − ω0 )N/2]
X R (e j ω ) = − sin[(ω − ω0 )(N − 1)/2]
2 sin[(ω − ω0 )/2]
A sin{[ω − (2π − ω0 )]N/2}
− sin[(ω + ω0 )(N − 1)/2] (4)
2 sin{[ω − (2π − ω0 )]/2}

(b) N = 32 and ω0 = π /4. The DTFT plots are shown in Figure 2.2bc.
(c) The DFT samples are shown in Figure 2.2bc.
(d) N = 32 and ω0 = 1.1π /4. The plots are shown in Figure 2.2d.
The added spectrum for the second case above (ω0 = 1.1π /4) is a result of the periodic extension of the
DFT. For a 32-point sequence, the end of each extension does not line up with the beginning of the next
extension. This results in sharp edges in the periodic extension, and added frequencies in the spectrum.

2.3 The sequence is x(n) = cos(πn/4), 0 ≤ n ≤ 15.

(a) The 16-point DFT is shown in the top-left plot of Figure 2.3.
(b) The 32-point DFT is shown in the top-right plot of Figure 2.3.
(c) The 64-point DFT is shown in the bottom plot of Figure 2.3.
(d) The zero padding results in a lower frequency sampling interval. Hence there are more terms in the DFT
representation. The shape of the DTFT continues to fill in as N increases from 16 to 64.

2.4 x(n) = {1, 2, 3, 4, 3, 2, 1}; h(n) = {−1, 0, 1}

Statistical and Adaptive Signal Processing - Solution Manual 3

32−point DTFT (Real) 32−point DTFT (Imaginary)

20 15

15 10

5
10
0
5
−5

0
−10

−5 −15
−4 −2 0 2 4 −4 −2 0 2 4
Figure 2.2d: Real and Imaginary DTFT Plots (ω0 = 1.1π /4)

16−point DFT 32−point DFT

7 7

6 6

5 5

4 4

3 3

2 2

1 1

0 0
0 π/4 π/2 3π/4 π 5π/4 3π/2 7π/4 0 π/4 π/2 3π/4 π 5π/4 3π/2 7π/4
64−point DFT
8

0
0 π/4 π/2 3π/4 π 5π/4 3π/2 7π/4

Figure 2.3: The 16, 32, and 64-point DFTs of x(n) = cos(π n/4)
4 Statistical and Adaptive Signal Processing - Solution Manual

(a) Convolution using matrix-vector multiplication approach.

y = Xh
   
−1 1 0 0
 −2   2 1 0 
   
 −2   3 2 1 
    
 −2   4 3 2  −1
   
 0  =  3 4 3  0 
   
 2   2 3 4  1
   
   1 
 2   2 3 
 2   0 1 2 
1 0 0 1
(b) The Matlab function convtoep
function [y]=convtoep(x,h)
% Convolution using Toeplitz Arrays
% y = convtoep(x,h)

nx = length(x); x = reshape(x,nx,1);
nh = length(h); h = reshape(h,nh,1);
X = toeplitz([x; zeros(nh-1,1)],[x(1) zeros(1,nh-1)]);
y = X*h;
(c) Verification:

2.5 x(n) = 0.9n u(n)

(a) Analytical evaluation of x(n) ∗ x(n) :

∞

y(n) = x(n) ∗ x(n) = x(k)x(n − k)
k=−∞
∞

= (0.9)k u(k)(0.9)n−k u(n − k)
k=−∞
∞
= (0.9)k (0.9)n−k u(n − k)
k=0
y(n) = (n + 1)(0.9)n
This sequence is shown in the leftmost plot in Figure 2.5.
(b) Convolution using the conv function: The sequence x(n) is truncated to 51 samples. This convolution is
done using
n = 0:50; x =(0.9).^n; y = conv(x,x);
This sequence is in the center plot in in Figure 2.5.
(c) Convolution using the filter function: To use this function, we have to represent one of the x(n) by
coefficients in an equivalent difference equation. This difference equation is given by
x(n) = δ(n) + 0.9 x(n − 1)
which means that the filter coefficients are b = 1, a = [1,-0.9]. Thus this convolution is done using
y = filter(1,[1,-0.9],x);
This sequence is in the rightmost plot in Figure 2.5.
Statistical and Adaptive Signal Processing - Solution Manual 5

(d) The three plots appear to be identical. However, the conv function gives the largest error since both
sequences are truncated. The filter function would be best suited for infinite length sequences.

y(n)=(n+1)(0.9)n Using Conv Using filter

4 4 4

3.5 3.5 3.5

3 3 3

2.5 2.5 2.5

2 2 2

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100

Figure 2.5: x(n) = (0.9) u(n) convolved with itself

n

2.6 Hap (n) is a causal and stable all-pass system with input x(n) and output y(n). Then we have
∞
 ∞

|y(n)|2 = |x(n)|2
n=0 n=0

Consider

x(n) = x0 (n) + x1 (n) and yi (n) = h(n) ∗ xi (n)

Then
∞
 ∞

|y0 (n) + y1 (n)| 2
= |x0 (n) + x1 (n)|2
n=0 n=0
∞
 ∞
 ∞
 ∞
 ∞

|y0 (n)| +
2
|y1 (n)| +
2
|y0 (n)y1(n)| = |x0 (n)| + 2
|x1 (n)|2
n=0 n=0 n=0 n=0 n=0
∞
 ∞
 ∞

|y0 (n)|2 + |y0 (n)y1(n)| = |x0 (n)|2
n=0 n=0 n=0

Hence
∞
 ∞

|y0 (n)| ≤ 2
|x0 (n)|2
n=0 n=0

Define x0 (n) as 0 above n 0 , then


n0 
n0
|y(n)|2 ≤ |x(n)|2
n=0 n=0

2.7 Monotone phase-response property of a causal and stable PZ-AP system:

(a) A real first-order system: Consider

p − z −1
H (z) = , | p| < 1
1 − p z −1
6 Statistical and Adaptive Signal Processing - Solution Manual

Then



sin ω sin ω
] H (e ) = arctan
jω
− arctan
cos ω − 1/ p cos ω − p

Clearly, ] H (e j ω ) decreases from ] H (e j 0) = π to ] H (e j 2π ) = −π . To show that it decreases mono-

tonically, consider




d  d sin ω sin ω
] H (e j ω ) = arctan − arctan
dω dω cos ω − 1/ p cos ω − p
p −1
2
=
1 + p − 2 (cos ω) p
2

which is negative for | p| < 1 . This proves that ] H (e j ω ) decreases monotonically from ] H (e j 0 ) = π
to ] H (e j 2π ) = −π.
(b) A real second-order (complex-conjugate pole-pair) system: The system function is
  
(r ]θ) − z −1 (r ]θ )∗ − z −1
H (z) = , 0<r <1
1 − (r ]θ)∗ z −1 1 − (r ]θ ) z −1
Consider the term
(r ]θ) − z −1
H1 (z) ,
1 − (r ]θ) z −1
Then angle calculations are similar to those for the real first-order case if we rotate the coordinate system
by the angle θ, that is,

  
 
sin ω sin ω
] H1(e ) = arctan
jω
+ θ − arctan +θ
cos ω − 1/r cos ω − r



sin ω sin ω
= arctan − arctan
cos ω − 1/r cos ω − r

Thus following the arguments in (a), we conclude that ] H1 (e j ω ) decreases monotonically from ] H1 (e j 0 )
to ] H1 (e j 0 ) − π . Similarly, consider

(r ]θ) − z −1
H2 (z) ,
1 − (r ]θ ) z −1

Then ] H2 (e j ω ) decreases monotonically from ] H2 (e j 0 ) to ] H2 (e j 0 ) − π . Finally, since

] H1(e j ω ) = ] H1(e j ω ) + ] H2 (e j ω )
we conclude that ] H (e j ω ) decreases monotonically from ] H (e j 0) to ] H (e j 0) − 2π .
(c) Generalizing the above result for a real, causal, and stable PZ-AP system with P pole-zero pairs, we can
show that ] H (e j ω ) decreases monotonically from ] H (e j 0) to ] H (e j 0 ) − 2π P.

2.8 Minimum group delay property: Consider the systems

 
Hmin (z) = 1 − 0.25z −1 1 + 0.5z −1
 
Hmax (z) = 0.25 − z −1 0.5 + z −1
 
Hmix (z) = 1 − 0.25z −1 0.5 + z −1
Statistical and Adaptive Signal Processing - Solution Manual 7

Group delay (in samples)

2 H
max

1
Hmix

0 Hmin

−1
0 0.2 0.4 0.6 0.8 1
Normalized frequency (Nyquist == 1)

Figure 2.8: Group delay plots

(a) Group delay response plots are shown in Figure 2.8.

(b) Proof of minimum group delay property

2.9 Minimum- and maximum-phase components

(a)
1 − 2.5z −1 + z −2
R y (z) =
1 − 2.05z −1 + z −2
(1 − 12 z −1 )(1 − 2z −1 )
=
(1 − 45 z −1 )(1 − 54 z −1 )
(1− 21 z −1 ) (1−2z −1 )
Hence the minimum phase component is (1− 54 z −1 )
and the maximum phase component is (1− 54 z −1 )
.
(b)
3z 2 − 10 + 3z −2
R y (z) =
3z 2 + 10 + 3z −2
(1 − 13 z −2 )(1 − 3z −2 )
=
(1 + 13 z −2 )(1 + 3z −2 )
(1− 31 z −2 ) (1−3z −2 )
Hence the minimum phase component is (1+ 31 z −2 )
and the maximum phase component is (1+3z −2 )
.

2.10 Consider the all-pass system function

1 − αz −1
Hap (z) = , |α| < 1
z −1 − α ∗
(a)
∗
|Hap (z)|2 = Hap (z)Hap (z)
1 − αz −1 1 − α ∗ z −1∗
= ( )( −1∗ )
z −1 − α ∗ z −α
∗ z
1 − αz −1 1 − α |z|2 z
= ( −1 ∗
)( ) where z −1∗ = 2
z −α z
|z|2
−α |z|
|z|2 − αz −1 |z|2 − α ∗ z − |α|2
=
1 − α ∗ z − αz −1 |z|2 + |α|2 |z|2
8 Statistical and Adaptive Signal Processing - Solution Manual

−3
x 10
1

Magnitude Response (dB)

0.5

−0.5

−1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency (Nyquist == 1)

400

300

Phase (degrees) 200

100

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalized frequency (Nyquist == 1)

jω
Figure 2.11: Frequency Response of H (e )

(b)

D|H |2 (z) = |z|2 − αz −1 |z|2 − α ∗ z − |α|2

A|H |2 (z) = 1 − α ∗ z − αz −1 |z|2 + |α|2 |z|2
D|H |2 (z) − A|H |2 (z) = |z|2 − |α|2 − 1 − |α|2 |z|2
= (|z|2 − 1)(1 − |α|2 )

(c) Using the result from part 2 above (for |α| < 1),

|Hap (z)| > 1 when (D|H |2 (z) − A|H |2 (z)) > 0 ⇒ |z| > 1
|Hap (z)| = 1 when (D|H |2 (z) − A|H |2 (z)) > 0 ⇒ |z| = 1
|Hap (z)| < 1 when (D|H |2 (z) − A|H |2 (z)) > 0 ⇒ |z| < 1

2.11 Consider the system function

a + b z −1 + c z −2
H (z) =
c + b z −1 + a z −2
(a) Magnitude of the frequency response function:

|H (e j ω )|2 = H (e j ω )H (e− j ω )
a + b e− j ω + c e−2 j ω a + b e j ω + c e2 j ω
= ( )( )
c + b e− j ω + a e−2 j ω c + b e j ω + a e2 j ω
Simplifying

|H (e j ω )|2 = 1, for all ω ⇒ |H (e j ω )| = 1, for all ω

(b) Magnitude and phase response plots are shown in Figure 2.11.

2.12 Consider the third-order FIR system

H (z) = 12 + 28z −1 − 29z −2 − 60z −3

= 12(1 + 2.5z −1 )(1 + 1.33z −1 )(1 − 1.5z −1 )
Statistical and Adaptive Signal Processing - Solution Manual 9

Energy Delay of the Eight Systems in part 1

6000

Emax
Emin
5000

4000

Energy
3000

2000

1000

0
0 1 2 3 4
n

Figure 2.12: Energy Delay of Eight Systems

(a) Therefore, an FIR system with three zeros results in eight (23 ) FIR systems with identical magnitude
responses. They are obtained by reflecting each zero about the unit circle and applying the appropriate
Zero@ z1 z2 z3 Gain System
Hmax (z) -2.5 -1.33 1.5 12 12 + 28z − 29z −2 − 60z −3
−1

Hmin (z) -0.4 -0.75 0.67 60 60 + 29z −1 − 28z −2 − 12z −3

Hmix1(z) -2.5 -1.33 0.67 18 18 + 57z −1 + 14z −2 − 40z −3
gain factor. Hmix2 (z) -2.5 -0.75 1.5 16 16 + 28z −1 − 48z −2 − 45z −3
Hmix3(z) -2.5 -0.75 0.67 24 24 + 62z −1 − 7z −2 − 30z −3
Hmix4(z) -0.4 -1.33 1.5 30 30 + 7z −1 − 62z −2 − 24z −3
Hmix5(z) -0.4 -1.33 0.67 45 45 + 48z −1 − 28z −2 − 16z −3
Hmix6(z) -0.4 -0.75 1.5 40 40 − 14z −1 − 57z −2 − 18z −3
(b) Clearly from the above table, Hmax (z) and Hmin are the maximum and minimum phase systems respec-
tively.
(c) Energy delay plots are shown in Figure 2.12.

2.13 Consider the system function

1 + z −1 − 6z −2
H (z) =
1 + 14 z −1 − 18 z −2

(a) After factorization, we obtain

(1 + 3z −1 )(1 − 2z −1 )
H (z) =
(1 + 12 z −1 )(1 − 14 z −1 )

Zeros outside unit circle ⇒ Not minimum phase.

(b) Minimum-phase system:
1 1
(1 + 3z −1 )(1 − 2z −1 ) ⇒ 6(1 + z −1 )(1 − z −1 )
3 2
= 6 − z −1 − z −2

Hence
6 − z −1 − z −2
Hmin (z) =
1 + ( 14 )z −1 − ( 18 )z −2
10 Statistical and Adaptive Signal Processing - Solution Manual

0.5 0.5
-1
x(n) z y(n)
-0.5
0.5

-1
z
1/3
-1
z
-1/3
Figure 2.14: Parallel Structure of two AP systems

(c) This system is a Maximum-phase system since all zeros are outside the unit circle.

2.14 Parallel connection of two all-pass systems:

3(1 + z −1 )(1 + z −1 )(1 + z −1 )

H (z) =
12(1 + 21 z −1 )(1 + 13 z −1 )
 
3 A( 12 + z −1 ) Bz −1 ( 13 + z −1 )
= +
12 1 + 12 z −1 1 + 13 z −1
⇒ A = 2, B = 2
( + z −1 )
1 1 1 −1 1
z ( 3 + z −1 )
H (z) = 2 2
+ 2
1 + 21 z −1 1 + 13 z −1

The block diagram is shown in Figure 2.14.

2.15 Impulse response of an all-pole system with lattice parameters:

k1 = 0.2; k2 = 0.2; k3 = 0.2; k4 = 0.2;

The direct form coefficients {ak }40 can be computed using the lat2dir function. The corresponding system
function is
1
H (z) =
1+ 0.76z −1 _0.731z −2 + 0.787z −3 + 0.7z −4
Using partial fraction expansion we obtain

0.3506 − 0.4334z −1 0.6492 + 0.3656z −1

H (z) = +
1 − 0.6654z −1 + 0.9109z −2 1 + 0.4254z −1 + 0.7685z −2
Hence upon inverse transformation we obtain
 
h(n) = 0.3506 (0.9544)n cos(0.3867π n) − 0.3540 (0.9544)n sin(0.3867π n) u(n)
 
+ 0.6494 (0.8766)n cos(0.8021π n) − 0.1903 (0.8766)n sin(0.8021π n) u(n)

The direct- and lattice-form structures are shown in Figure 2.15.

Statistical and Adaptive Signal Processing - Solution Manual 11

x(n) + y(n)

-1
z

-0.76
+

-1
z
-0.731
+

-1
z

-0.787 x(n)
+ + + +
+
-0.7 -0.5 -0.3 -0.2

0.7 0.5 0.3 0.2

-1
z
-0.7 + -1
z + -1
z + -1
z + -1
z
y(n)

Figure 2.15: Direct and Lattice Form Structures

 Chapter 3
Random Variables, Random Vectors, & Stochastic Processes

3.1 Exponential Density: f x (x) = a1 e−x/a u(x)

(a) Density plot for a = 1: f x (x) = e−x u(x) is shown in Figure 3.1

f (x)=e−xu(x)
x
1

0.8

0.6

0.4

0.2

0
0 2 4 6 8 10
Figure 3.1: Exponential density function

(b) Moments:
∞
i. Mean: µx = x f x (x)dx = 1/a = 1
−∞
∞
ii. Variance: σ 2x
= −∞ (x − µx )2 f x (x)dx = (1/a)2 = 1
iii. Skewness: The third central moment is given by
 ∞ 
(3)
 3 ∞
γx = x − µx f x (x) dx = (x − 1)3 (e−x )dx = 2
−∞ 0

Hence
1 (3)
skewness = γ = 2 (⇒ leaning towards right)
σ 3x x
iv. Kurtosis: The fourth central moment is given by
 ∞  ∞
(4)
 4
γx = x − µx f x (x) dx = (x − 1)4 (e−x )dx = 9
−∞ 0

Hence
1 (4)
kurtosis = γ −3= 9−3 =4
σ 4x x
which means a much flatter shape compared to the Gaussian shape.
(c) Characteristic function:
 ∞  ∞ 1
1 −x( 1 −s) 1
x = E{e s x(ξ)
}= f x (x)e dx =
sx
e a dx = a
=
−∞ 0 a 1
a
−s 1 − as

12
Statistical and Adaptive Signal Processing - Solution Manual 13

2
fx(x)=xe−x u(x)
0.5

0.4

0.3

0.2

0.1

0
0 2 4 6 8 10
Figure 3.2: Rayleigh density function

x −x 2 /(2σ 2 )
3.2 Rayleigh density : f x (x) = σ2
e u(x)

(a) Density function for σ = 1: f x (x) = x e−x

2 /2
u(x) is shown in Figure 3.2.
(b) Moments:
∞ ∞ √
x 2 e−x
2 /2
i. Mean: µx = −∞ x f x (x)dx = 0 dx = π /2
∞ ∞ √ π
ii. Variance: σ 2x = −∞ (x − µx )2 f x (x)dx = 0 (x − π/2)2 x e−x /2 dx = 2 −
2

2
iii. Skewness: The third central moment is given by
 ∞  ∞  3
(3)
 3 1 √ 3 3 √
x − π /2 (x e−x /2 )dx = √
2
γx = x − µx f x (x) dx = π −√ π
−∞ 0 2 2
Hence
√ 3 √
1 √1 π − √32 π
skewness = 3 γ (3) = 2

 3 = . 631 11 (⇒ leaning towards right)
σx x π
2−
2
iv. Kurtosis: The fourth central moment is given by
 ∞  ∞  4
(4)
 4 3π 2
x − π /2 (x e−x /2 )dx = 8 −
2
γx = x − µx f x (x) dx =
−∞ 0 4
: Hence
2
1 (4) 8 − 3π4
kurtosis = 4 γ x − 3 =  − 3 = 0.2451
σx π 2
2−
2
which means a flatter but almost a Gaussian shape.
(c) Characteristic function (σ = 1):
 ∞  y
xe−x es x dx
2
x (s) = E{e s x(ξ)
}= f x (x)e dx =
sx
−∞ 0


1 √ 1 s2 1 1 √ 1 s2 1
= s πe 4 + + s π e erf
4 s
4 2 4 2
14 Statistical and Adaptive Signal Processing - Solution Manual

3.3 Relation between r x(m) and γ (m)

x :

  
m
k m
γ (m)
x = E{[x(ζ ) − µx ] } = E (−1)
m
x(ζ )k µm−k
x
k=0
k
m

m k
= (−1)k µx E{x(ζ )k }
k=0
k
m

k m
γ (m)
x = (−1) µkx r x(n−k)
k=0
k

Similarly
 m
 
   m
r x(m) = E{( x(ζ ) − µx + µx ) } = E m
y(ζ ) k
µm−k
   k=0
k x
y(ζ )
m

 m
= µkx E{y(ζ )k }
k=0
k
m

m k (n−k)
r x(m) = µx γ x
k=0
k

3.4 First four cumulants: The cumulants are given by

 
dm ¯ x (s) 
κ (m)
x , 

ds m
s=0

where
  
¯ x (s) = ln 
 ¯ x (s) = ln E es x(ζ ) = ln [x (s)]

and
∞
s m (m)
x (s) = r
m=0
m! x

Thus



1 dx (s) 
κ (1) = 
x
x (s) ds
!∞ s m (m+1) 
s=0
r 
= m=0 m! x
!∞ s m (m)  = r x(1) = µx = 0 [ * zero mean x(ζ )]
rx 
m=0 m! s=0

Similarly
 !∞ 
s m (m+1) 
d m=0 m! r x 
κ (2) = !∞ s m (m) 
x
ds 
m=0 m! r x s=0

!∞ sm (m+2)
!∞ s m (m)
!∞ 
s m (m+1) 2 
m=0 m! r x m=0 m! r x − r
m=0 m! x 
= ! 2 
∞ s m (m) 
m=0 m! r x  (1)
s=0,rx =0
= r x(2) = σ 2x
Statistical and Adaptive Signal Processing - Solution Manual 15

 
!∞ s m (m+2)
!∞ s m (m)
!∞ 
s m (m+1) 2 
d  m=0 m! r x m=0 m! r x − r 
κ (3) =  ! 2
m=0 m! x

x
ds ∞ s m (m) 
m=0 m! r x  (1)
s=0,rx =0
= γ (3)
x

3.5 Random vector x(ζ ) = [x1 (ζ ) x2 (ζ )]

(a) Mean vector:

   
1 3   7
1
µ y = Aµx =  −1 2  = 3 
2
2 3 8

(b) Autocorrelation matrix:

 
1 3   
4 0.8 1 −1 2
y = Ax AH =  −1 2 
0.8 1 3 2 3
2 3
 
17.8 1.2 24.2
=  1.2 4.8 −1.2 
24.2 −1.2 34.6

(c) Crosscorrelation matrix:

 
5 2.8
Rx =  x + µx µHx =
2.8 5
 
13.4 0.6 18.4
Rxy = Rx A = H
17.8 7.2 20.6

3.6 Let x(ζ ) be a Gaussian random vector with mean vector µx and covariance matrix x , then the characteristic
function is
 
x (ξ ) = exp j ξ T µx − 21 ξ T x ξ
Let y(ζ ) = Ax where A is a non-singular matrix. Its characteristic function is given by
" T # " T #
y (ξ ) = E e j ξ y = E e j ξ Ax
" # $  T  T  %
= E e j (A ξ ) x = exp j AT ξ µx − 12 AT ξ x AT ξ
T T

   
= exp j ξ T Aµx − 12 ξ T A x AT ξ

which is a characteristic function of a Gaussian random vector with mean vector Aµx and covariance matrix
A x AT .
3.7 Sum of independent exponential random variables xk (ζ ): f xk (x) = e−x u(x).

(a) y2 (ζ ) = x1 (ζ ) + x2 (ζ ). Hence
 x 
−s s−x
f y2 (x) = f x1 (x) ∗ f x2 (x) = e e ds u(x)
0
= xe−x u(x)
The plot of this density is shown in Figure 3.7(a).
16 Statistical and Adaptive Signal Processing - Solution Manual

(b) y3 (ζ ) = y2 (ζ ) + x3 (ζ ). Hence
 x 
−s s−x
f y3 (x) = f y2 (x) ∗ f x3 (x) = se e ds u(x)
0
1 2 −x
= x e u(x)
2
The plot of this density is shown in Figure 3.7(b).
(c) y4 (ζ ) = y3 (ζ ) + x4 (ζ ). Hence
 x 
1 2 −s s−x
f y4 (x) = f y3 (x) ∗ f x4 (x) = s e e ds u(x)
2 0
1 3 −x
= x e u(x)
6
The plot of this density is shown in Figure 3.7(c).

y1 = x1 y2 = x1+x2
1 1
f1(y1)

f2(y2)

0.5 0.5

0 0
0 5 10 15 0 5 10 15
y1 y2

y3 = x1+x2+x3 y4 = x1+x2+x3+x4
1 1
f3(y3)

f4(y4)

0.5 0.5

0 0
0 5 10 15 0 5 10 15
y y
3 4
Figure 3.7: Sums of IID exponentially distributed random variables

(d) As k increases, the distribution of yk (ζ ) approaches a Gaussian distribution, with a mean equal to the sum
of the exponential distribution means.

3.8 Test of WSS: µx = constant and r x (n 1 ,"n 2 ) = r x (n 1 − n 2 ) # " # N→∞

!N !N
Test of m.s. ergodicity in the mean: E 2N+1 1
−N x(n, ζ ) = µ x and var 1
2N+1 −N x(n, ζ ) −→ 0

(a) x(n, ζ ) = A(ζ ), where random variable A(ζ ) is uniformly distributed between 0 and 1. Now
1
µx (n) = E {x(n, ζ )} = E {A(ζ )} =
2
Statistical and Adaptive Signal Processing - Solution Manual 17

and
 1
r x (n 1 , n 2 ) = E {x(n 1 , ζ )x(n 2 , ζ )} = E A2 (ζ ) =
3
Hence the process is WSS. Consider
 
1  N
1 N
1
E x(n, ζ ) = E {x(n, ζ )} = E {A(ζ )} = = µx
2N + 1 −N 2N + 1 −N 2

However, the computed mean remains at the observed value of x(n, ζ 0 ) , a0 and does not converge to
the true mean of 1/2. Hence the process is not m.s. ergodic in the mean.
(b) x(n, ζ ) = A(ζ ) cos ω0 n, where random variable A(ζ ) is a unit Gaussian random variable. Then

µx (n) = E{ A(ζ ) cos ωo n} = E{ A(ζ )} cos ωo n = 0

and

r x (n 1 , n 2 ) = E {x(n 1 , ζ )x(n 2 , ζ )} = E A2 (ζ ) cos ω0 n 1 cos ω0 n 2
= cos ω0 n 1 cos ω0 n 2

Thus the process is not WSS. Therefore, the m.s. ergodicity cannot be determined because the process
must be WSS.
(c) x(n, ζ ) is Bernoulli process with Pr [x(n, ζ ) = 1] = p and Pr [x(n, ζ ) = −1] = 1 − p. Then

µx (n) = E{x(n, ζ )} = p + (−1)(1 − p) = 2 p − 1

and
&
(2 p − 1)2 , n 1  = n 2
r x (n 1 , n 2 ) = (5)
1, n1 = n2
!N
Hence it is WSS. Consider the random variable x(n, ζ ) N = 1
2N+1 −N x(n, ζ )

 1  N
E x(n, ζ ) N = E {x(n, ζ )} = 2 p − 1 = µx (n)
2N + 1 −N

Now consider the second moment of the random variable x(n, ζ ) N

 N 
" 2 # 1  N
E x(n, ζ ) N = E x(n 1 , ζ ) x(n 2 , ζ )
(2N + 1)2 −N −N
 N N 
1 
= E x(n 1 , ζ )x(n 2 , ζ )
(2N + 1)2 −N −N

1 N N
= r x (n 1 , n 2 )
(2N + 1)2 −N −N

Let M = 2N + 1 and using (5)

" 2 # 1   2 
E x(n, ζ ) N = M + M − M (2 p − 1) 2
M2
18 Statistical and Adaptive Signal Processing - Solution Manual

Thus
$ 2 % "   2 # 2
var x(n, ζ ) N = E x(n, ζ ) N
− E x(n, ζ ) N
1   2 
= 2
M + M − M (2 p − 1) 2
− (2 p − 1)2
M
1  2 
≈ 2
M (2 p − 1)2 − (2 p − 1)2 for large M
M
$ 2 %
Thus var x(n, ζ ) N → 0 as N → ∞. Hence the process is m.s. ergodic in the mean.
!N
3.9 Harmonic process: x(n) = k=1 Ak cos(ωk n + φ k ), φ k uniformly distributed over [0, 2π ]

(a) Mean:


M 
M
E{x(n)} = E{ Ak cos(ωk n + φ k )} = Ak E{cos(ωk n + φ k )}
k=1 k=1
 2π
1 
M
= Ak cos(ωk n + φ k ) dφ k = 0
2π k=1 0

(b) Autocorrelation:


M
x(n) = Ak cos(ωk n + φ k )
k=1

1
M
F
⇔ X (ω) = Ak (δ(ω − ωk )e− j φk + δ(ω + ωk )e j φk )
2 k=1

1
M
|·|2
⇔ Rx (e j ω ) = |Ak |2 (δ(ω − ωk ) + δ(ω + ωk ))
4 k=1

1
M
F −1
⇔ r x (l) = |Ak |2 cos ωk l
2 k=1

3.10 Harmonic process with φ k distributed as f φk (φ k ) = 1 + cos φ k /2π , −π ≤ φ k ≤ π . Consider the mean
sequence


M 
M
E{x(n)} = E{ Ak cos(ωk n + φ k )} = Ak E{cos(ωk n + φ k )}
k=1 k=1
 π
1  
M
= Ak cos(ωk n + φ k ) 1 + cos φ k dφ k = (cos ωk n) π
2π k=1 −π

Thus the process is not stationary.

& & '
4 − |n| ; |n| ≤ 3
3.11 µx = 4; γ x (n) = = 1, 2, 3, 4, 3, 2, 1
0; otherwise ↑
h(n) = u(n) − u(n − 4) = {1, 1, 1, 1}
!
∞
(a) Mean: µ y (n) = µx h(k) = 4 × 4 = 16
−∞
Statistical and Adaptive Signal Processing - Solution Manual 19

(b) Crosscorrelation: γ x y (n,1 n 2 ) = γ x (n 1 − n 2 ) ∗ h ∗ (n 2 − n 1 ) Hence

γ x y (k) = γ x (k) ∗ h ∗ (−k) = {1, 2, 3, 4, 3, 2, 1} ∗ {1, 1, 1, 1}
↑ ↑
= {1, 3, 6, 10, 12, 12, 10 , 6, 3, 1}
↑

(c) Autocorrelation: γ y (n,1 n 2 ) = h(n,1 n 2 ) ∗ γ x y (n,1 n 2 ) Hence

γ y (k) = h(k) ∗ γ x y (k)
= {1, 1, 1, 1} ∗ {1, 3, 6, 10, 12, 12, 10 , 6, 3, 1}
↑ ↑
= {1, 4, 10, 20, 31, 40, 44 , 40, 31, 20, 10, 4, 1}
↑

1+ 13 z −1
3.12 LTI system: y(n) = 12 y(n − 1) + x(n) + 13 x(n − 1) ⇒ H (z) = 1− 12 z −1
, |z| < 1
2
3
Input process: zero-mean, WSS with r x (l) = 0.5|l| ⇒ Rx (z) = 5 1
4
−1
, 1
< |z| < 2
4 − 2 (z+z )
2

(a) PSD and autocorrelation of the output: Consider

R y (z) = H (z)H ∗(1/z ∗ )Rx (z) = H (z)H (z −1)Rx (z)
( )( )( )
1 + 13 z −1 1 + 13 z 3
= 4
1 − 12 z −1 1 − 12 z 5
4
− 12 (z + z −1 )
5
+ 14 (z + z −1 ) 1
= 6
5 1  , < |z| < 2
− (z + z −1 ) 2 2
4 2
Hence the PSD is
jω
5
+ 14 (e j ω + e− j ω ) 5
+ 12 cos ω
R y (e ) = 6
5 1 2 =
6
5 2
4
− 2 (e j ω + e− j ω ) 4
− cos ω
5
+ 21 cos ω
= 6
33
16
− 52 cos ω + 12 cos 2ω
To determine the autocorrelation, consider the PFE of
5
+ 14 (z + z −1 ) 1
R y (z) = 6
5 1  , < |z| < 2
− (z + z −1 ) 2 2
4 2
4 3z + 10z + 3z
3 2
1
= , < |z| < 2
 − 20z + 33z − 20z + 4 2
3 4z 4 3 2

4 35 253 70 43
= + + −
3 18 (2z − 1)2 54 (2z − 1) 9 (z − 2)2 27 (z − 2)
1 −1
35 −1 z 253 −1 1
= z  2 + z 
27 1 − 1 z −1
2 81 1 − 12 z −1
2
140 −1 2z −1 172 −1 1
+ z  2 − z 
27 1 − 2z −1 81 1 − 2z −1
Hence upon inverse z-transformation, we obtain

l−1

35 1 253 1 l−1
r y (l) = (l − 1) u(l − 1) + u(l − 1)
27 2 81 2
140 172 l−1
− (l − 1) 2l−1 u(−l) + 2 u(−l)
27 81
20 Statistical and Adaptive Signal Processing - Solution Manual

(b) Cross-correlation and cross-PSD between the input and output: Consider
 
Rx y (z) = H ∗ 1/z ∗ Rx (z) = H z −1 Rx (z)
( )( )
1 + 13 z 3
1
= 4
, < |z| < 2
1 − 2z
1 5
4
1 −1
− 2 (z + z ) 2

Thus the cross-PSD is

6e j ω + 2e2 j ω
Rx y (e j ω ) =
−4 + 12e j ω − 9e j 2ω + 2e j 3ω
To determine cross-correlation, consider the PFE of
( )( )
1 + 13 z 3
1
Rx y (z) = 4
, < |z| < 2
1 − 2z
1 5
4
− 2 (z + z )
1 −1 2
6z + 2z 2 1
= , < |z| < 2
12z − 9z − 4 + 2z 2
2 3

14 20 2
= + +
9 (2z − 1) 3 (z − 2) 2 9 (z − 2)
7 −1 1 10 −1 2z −1 2 −1 1
= z  + z  2 + z 
9 1 − 2z1 −1 3 1 − 2z −1 9 1 − 2z −1

Hence upon inverse z-transformation, we obtain


7 1 l−1 10 2
r x y (l) = u(l − 1) − (l − 1) 2l−1 u(−l) − 2l−1 u(−l)
9 2 3 9

1+1.25z −1
3.13 LTI system: y(n) = 0.6y(n − 1) + x(n) + 1.25x(n − 1) ⇒ H (z) = 1−0.6z −1
, |z| < 0.6
Input process: WSS with PSD Rx (e j ω ) = 1.64+1.6
1
cos ω

(a) The PSD of the output is given by

R y (e j ω ) = |H (e j ω )|2 Rx (e j ω ) = H (e j ω )H ∗ (e j ω )Rx (e j ω )
= H (e j ω )H (e− j ω )Rx (e j ω )



1 + 1.25e− j ω 1 + 1.25e j ω
= Rx (e j ω )
1 − 0.6e− j ω 1 − 0.6e j ω



2.5625 + 2.5 cos ω 1
=
1.36 − 1.2 cos ω 1.64 + 1.6 cos ω
19. 531
=
17 − 15 cos ω

(b) The cross-PSD between input and the output is given by

Rx y (e j ω ) = H ∗ (e j ω )Rx (e j ω ) = H (e− j ω )Rx (e j ω )



1 + 1.25e j ω 1
=
1 − 0.6e j ω 1.64 + 1.6 cos ω
4 + 5e j ω
= 31. 25
205 + 200 cos ω − 123e j ω − 120e j ω cos ω
Statistical and Adaptive Signal Processing - Solution Manual 21

3.14 Valid autocorrelation matrices

(a) Valid
(b) Valid
(c) Not Valid since det R3 < 0
(d) Not Valid since non-Hermitian: R4  = R4T

3.15 x(ζ ): a normal random vector with ρ i, j = 0, i  = j

(a) Consider the (i, j )th component of x which is given by

&
ρi σ i 2 i = j
γ i j = ρi j σ i σ j =
0 otherwise
Hence x has elements only along the main diagonal
(b) Since the components of x(ζ ) are mutually Gaussian with γ i j = 0, i  = j then from the property of
Gaussian random vectors, the uncorrelatedness also impiles independence.

3.16 Eigenvalues of Rk : We have to show that Rk q = λk q. Consider

R · R · R · . . . R · qi = R · R · R
 · . . . R· λi qi
  
k k−1
= λi R · R · R
 · . . . R· qi (since λi is a scalar)
k−1
= λi R · R · R
 · . . . R· λi qi
k−2

= λ2i R · R · R
 · . . . R· qi
k−2
..
.
R k q = λk q

3.17 Trace of R:

tr(R) = tr(QQH )
= tr(QQH )
= tr(I)

= tr() = λi

3.18 Determinant of R:

det R = |R|
= |QQH |
= |Q| || |QH |
*
= || = λi (since Q is orthonormal)
i
22 Statistical and Adaptive Signal Processing - Solution Manual

3.19 Relation between the determinants of R and  :

R =  + µµT = LLT + µµT

L−1 RL−T = I + L−1 µµT L−T
|L−1 RL−T | = |I + L−1 µµT L−T | = 1 + µT L−T L−1 µ = 1 + µT  −1 µ
|L−1 ||R||L−T | = 1 + µT  −1 µ
|R| = ||(1 + µT  −1 µ)

3.20 x(n): zero-mean WSS process

x = [x(0) x(2) x(3)]; Rx : correlation matrix of x. Hence
 
r x (0) r x (2) r x (3)
Rx =  r x∗ (2) r x (0) r x (2) 
r x∗ (3) r x∗ (2) r x (0)

(a) From Rx above, the matrix is Hermitian, that is, ri j = r ∗j i , Toeplitz since ri, j = ri− j , and nonnegative
definite since it is correlation matrix of a random vector.
(b) To determine the correlation matrix of x̄ = [x(0) x(1) x(2) x(3)], we need r x (1) which is not available.

3.21 Proof of r x (0) ≥ |r x (l)| : Consider

E{|x(n + l) ± x(n)|2 } = E{[x(n + l) ± x(n)][x(n + l) ± x(n)]∗ } ≥ 0

= r x (0) ± r x (l) ± r x (−l) + r x (0) ≥ 0
= 2 [r x (0) ± r x (l)] ≥ 0

which implies r x (0) ≥ |r x (l)| ∀l.

3.22 Nonnegative definiteness of r x (l): Using the nonnegativeness of E{|α 1 x(1) + · · · + α M x(M)|2 }, we have

E{|α 1 x(1) + · · · + α M x(M)|2 } = E{[α 1 x(1) + · · · + α M x(M)][α ∗1 x ∗ (1) + · · · + α ∗M x ∗ (M)]}

M  M
= α l α ∗k E{x(l)x ∗ (k)}
l=1 k=1

M 
M 
M 
M
= α l α ∗k r x (k − l) = αl r x (k − l)α ∗k ≥ 0
l=1 k=1 l=1 k=1

3.23 AP(1) system: x(n) = αx(n − 1) + w(n); n ≥ 0, |α| < 1

w(n): IID(0, σ 2w ) or rw (n 1 , n 2 ) = σ 2w δ(n 1 − n 2 )

(a) Autocorrelation r x (n 1 , n 2 ): First consider the crosscorrelation between x(n 1 ) and w(n 2 ), n 1 , n 2 ≥ 0
 
E x(n 1 )w ∗ (n 2 ) = E [αx(n 1 − 1) + w(n 1 )] w ∗ (n 2 )
= αr xw (n 1 − 1, n 2 ) + rw (n 1 , n 2 ), n 1 , n 2 ≥ 0

which results in a partial difference equation in n 1 with n 2 as a parameter

r xw (n 1 , n 2 ) = αr xw (n 1 − 1, n 2 ) + σ 2w δ(n 1 − n 2 ), n 1 , n 2 ≥ 0 (6)
Statistical and Adaptive Signal Processing - Solution Manual 23

subject to r xw (−1, n 2 ) = E {x(−1)w ∗ (n 2 )} = 0. Taking unilateral z-transform of (6) with respect to n 1 ,

we have

Rxw (z 1 , n 2 ) = αz 1−1 Rxw (z 1 , n 2 ) + σ 2w z 1−n2 , n 2 ≥ 0

or
σ 2w z 1−n2
Rxw (z 1 , n 2 ) = ⇒ r xw (n 1 , n 2 ) = σ 2w α n1 −n2 u(n 1 − n 2 ), n 1 , n 2 ≥ 0 (7)
1 − αz 1−1

which as (n 1 , n 2 ) → ∞ tends to

r xw (n 1 − n 2 ) = r xw (l) = σ 2w α l u(l), l = (n 1 − n 2 )

Now consider the autocorrelation between x(n 1 ) and x(n 2 ), n 1 , n 2 ≥ 0

   
E x(n 1 )x ∗ (n 2 ) = E x(n 1 ) αx ∗ (n 2 − 1) + w ∗ (n 2 )
= αr x (n 1 , n 2 − 1) + r xw (n 1 , n 2 ), n 1 , n 2 ≥ 0

which results in a partial difference equation in n 2 with n 1 as a parameter

r x (n 1 , n 2 ) = αr x (n 1 , n 2 − 1) + r xw (n 1 , n 2 ), n 1 , n 2 ≥ 0 (8)

subject to r x (n 1 , −1) = E {x(n 1 )x ∗ (−1)} = 0. From (7), since r xw (n 1 , n 2 ) = 0 for n 2 > n 1 , first consider
the interval 0 ≤ n 2 ≤ n 1 . Taking unilateral z-transform of (8) with respect to n 2 , we have
 
Rx (n 1 , z 2 ) = αz 2−1 Rx (n 1 , z 2 ) + Z σ 2w α n1 −n2 u(n 1 − n 2 )

or
 
 α n1 1 z 2−1 z 2−n1
1− αz 2−1 Rx (n 1 , z 2 ) = σ 2w −
1 − α1 z 2−1 α 1 − α1 z 2−1

or
σ2 −n −1
σ 2w α n1 w
α 2
z 1
Rx (n 1 , z 2 ) =   −   (9)
1 − αz 2−1 1 − α1 z 2−1 1 − αz 2−1 1 − α1 z 2−1

which after inverse z-transformation yields [note that the second term in (9) does not contribute in the
interval 0 ≤ n 2 ≤ n 1 ]

σ 2w  n1 −n2
r x (n 1 , n 2 ) = α − α 2 α n1 +n2 , 0 ≤ n 2 ≤ n 1 (10)
1 − α2
Now consider the interval 0 ≤ n 1 < n 2 < ∞. In this interval, r xw (n 1 , n 2 ) = 0 for n 1 < n 2 , the partial
difference equation (8) has no forcing function part, that is, the partial difference equation is

r x (n 1 , n 2 ) = αr x (n 1 , n 2 − 1), n 2 > n 1 ≥ 0 (11)

subject to the boundary condition at n 2 = n 1 as

σ 2w 
r x (n 1 , n 1 ) = 1 − α 2 α 2n1
1 − α2
24 Statistical and Adaptive Signal Processing - Solution Manual

Thus (11) can be easily solved to obtain

σ 2w 
r x (n 1 , n 2 ) = 1 − α 2 α 2n1 α n2 −n1 u(n 2 − n 1 ), n 1 ≥ 0 (12)
1 − α2
Finally, combining (10) and (12), we obtain
 2 
σw
α n1 −n2 − α 2 α n1 +n2 , 0 ≤ n2 ≤ n1
r x (n 1 , n 2 ) = σ 2w 
1−α 2 (13)
1−α 2
1 − α 2 α 2n1 α n2 −n1 , n2 > n1 ≥ 0

Clearly, r x (n 1 , n 2 ) is not stationary.

(b) To probe the asymptotic behavior of r x (n 1 , n 2 ), let n 1 , n 2 → ∞ keeping n 1 − n 2 , l constant. Then
from (13), we obtain
 2
σw
α n1 −n2 , 0 ≤ n 2 ≤ n 1 ⇒ l ≥ 0
r x (n 1 , n 2 ) = 1−α 2
σw
2

1−α 2
α n2 −n1 , n 2 > n 1 ≥ 0 ⇒ l < 0
or

σ 2w
αl ,l ≥ 0 σ 2w
r x (n 1 , n 1 − l) = 1−α 2
σw 2 = α |l| (14)
α −l ,l < 0 1 − α2
1−α 2

Thus r x (n 1 , n 2 ) is asymptotically stationary.

3.24 x is a random vector with mean µx and autocorrelation Rx .

(a) Consider y = QH x where Rx Q = Q and Q is orthonormal. Then

Ry = E{yyH } = E{QH xxH Q}

= QH E{xxH }Q
= QH R x Q = 

Thus y is an orthogonal component vector.

(b) Geometric interpretation
 
1 0.5
3.25 A Gausian random vector x with µx = [1 2]T and x =
0.5 1
Contour plot of the density function is shown in Figure 3.25.

3.26 Random process x(n) = ax(n − 1) + w(n), w(n) ∼ WN(0, σ 2w )

(a) Autocorrelation matrix Rx : Consider

r x (0) = E{x(n)x(n)} = E{(ax(n − 1) + w(n − 1))2 }

σ 2w
= a 2 r x (0) + σ 2w =
1 − a2
Also

r x (1) = E{x(n)x(n + 1)} = E{x(n)(ax(n) + w(n))}

= ar x (0)
Statistical and Adaptive Signal Processing - Solution Manual 25

Sigma Concentration Ellipses

−2 −1 0 1 2 3 4
Figure 3.25: Sigma Concentration Ellipses

Similarly

r x (M) = a M r x (0)

which gives the matrix

 
1 a ··· a M−1
σ 2w  a 1 ··· a M−2 
 
Rx =  .. .. .. .. 
1 − a2  . . . . 
a M−1 a M−2 · · · 1
that is symmetric Toeplitz.
(b) We will show that Rx · Rx−1 = I. Consider
 
  1 −a 0 ··· 0
1 a ··· a M−1  −a 1 + a 2 −a ··· 0 
 
σ 2w  a 1 ··· a M−2  0 −a 1 + a 2
· · · 0 
  
Rx · Rx−1 =  .. .. .. ..  .. .. .. .. 
1 − a2  . . . .  . . . . 
 
a M−1 a M−2 · · · 1  0 0 0 1 + a 2 −a 
0 0 ··· −a 1
 
1 − a2 0 ··· 0
1   0 1 − a2 · · · 0 

=  .. .. .. .. =I
1−a 
2
. . . . 
0 0 ··· 1 − a2

(c) Let
 
1 0 ··· 0 0
 −a 1 · · · 0  0
 
 .. .. . . ..  ..
Lx =  . . . .  .
 
 0 0 ··· 1 0 
0 0 ··· −a 1
26 Statistical and Adaptive Signal Processing - Solution Manual

Then
  
1 −a · · · 0 0 1 a ··· a M−1
 0 1 ···  
 0 0  a 1 ··· a M−2 

σ 2w .. .. .. .. ..  .. .. .. .. 
LTx Rx Lx =  . . . . .  . . . . 
1 − a2 
 ..

 
 −a   a a M−3 · · · 
M−2
0 0 . 1 a
0 0 ··· 0 1 a M−1
a M−2 · · · 1
 
1 0 ··· 0 0
 −a 1 · · · 0  0
 
 .. .. . . ..  ..
× . . . . .
 
 0 0 ··· 1 0 
0 0 ··· −a 1
= σ 2w I

(d) Plots of the DKLT and the DFT is shown in Figure 3.26d. The PSD points approach the eigenvalues as

2
M=8 (Diamond=eigenvalues) a=0.95 1
M=8 (Diamond=eigenvalues) a=0.5
10 10

1
10

0 0
10 10

−1
10

−2 −1
10 10
0 2 4 6 8 0 2 4 6 8

3
M=16 (Diamond=eigenvalues) a=0.95 1
M=16 (Diamond=eigenvalues) a=0.5
10 10

2
10

1
10
0
10
0
10

−1
10

−2 −1
10 10
0 5 10 15 0 5 10 15

3
M=32 (Diamond=eigenvalues) a=0.95 1
M=32 (Diamond=eigenvalues) a=0.5
10 10

2
10

1 0
10 10

0
10

−1 −1
10 10
0 10 20 30 40 0 10 20 30 40

Figure 3.26d: Eigenvalue and PSD

the size of the correlation matrix increases.

(e) Plots of the eigenvectors are shown in Figure 3.26e.
(f) Plots of the eigenvectors are shown in Figure 3.26f.
Statistical and Adaptive Signal Processing - Solution Manual 27

M=8 a=0.95 M=16 a=0.95 M=32 a=0.95

0.5 0.4 0.3

0.2
0.2
0.1

0
0 0
−0.1

−0.2
−0.2
−0.3

−0.5 −0.4 −0.4

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.95 M=16 a=0.95 M=32 a=0.95

0.4 0.3 0.2

0.2
0.2 0.1
0.1

0
0 0
−0.1

−0.2
−0.2 −0.1
−0.3

−0.4 −0.4 −0.2

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.95 M=16 a=0.95 M=32 a=0.95

0.5 0.4 0.3

0.2
0.2
0.1

0
0 0
−0.1

−0.2
−0.2
−0.3

−0.5 −0.4 −0.4

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.95 M=16 a=0.95 M=32 a=0.95

0.365 0.28 0.2

0.36 0.27 0.19

0.355 0.26 0.18

0.35 0.25 0.17

0.345 0.24 0.16

0.34 0.23 0.15

0.335 0.22 0.14

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

Figure 3.26e: Eigenvectors

3.27 Three different innovations representations:

Eigen-decomposition
  3  
0.7071 0.7071 0 0.7071 −0.7071
Rx = Qx H Q = 4
x x
−0.7071 0.7071 0 45 0.7071 0.7071
w ∼ N(0, x ), x = Qx w
LDU Triangularization
   
1 0 1 0 1 14
Rx = Lx D(x)
L Lx
H
= 1
4
1 0 0.9375 0 1
w ∼ N(0, D(x)
L ), x = Lx w

UDL Triangularization
 1
  
1 0.9375 0 1 0
Rx = Ux D(x)
U Ux
H
= 4
0 1 0 1 1 14
w ∼ N(0, D(x)
U ), x = Ux w

3.28 Eigenvalues and Eigenvectors of a MA(1) process.

(a) Given that x(n) = w(n) + b w(n − 1) with w(n)˜WN(0, σ 2w ), the complex PSD is
 
Rx (z) = σ 2w 1 + b z −1 (1 + b z) = σ 2w b z + (1 + b2 ) + b z −1
28 Statistical and Adaptive Signal Processing - Solution Manual

M=8 a=0.5 M=16 a=0.5 M=32 a=0.5

0.5 0.4 0.3

0.2
0.2
0.1

0
0 0
−0.1

−0.2
−0.2
−0.3

−0.5 −0.4 −0.4

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.5 M=16 a=0.5 M=32 a=0.5

0.6 0.3 0.3

0.2 0.2
0.4
0.1 0.1
0.2
0 0

0 −0.1 −0.1

−0.2 −0.2
−0.2
−0.3 −0.3

−0.4 −0.4 −0.4

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.5 M=16 a=0.5 M=32 a=0.5

0.5 0.4 0.3

0.2
0.2
0.1

0
0 0
−0.1

−0.2
−0.2
−0.3

−0.5 −0.4 −0.4

0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

M=8 a=0.5 M=16 a=0.5 M=32 a=0.5

0.45 0.35 0.25

0.4 0.3 0.2

0.35 0.25 0.15

0.3 0.2 0.1

0.25 0.15 0.05

0.2 0.1 0
0 2 4 6 8 0 5 10 15 20 0 10 20 30 40

Figure 3.26f: Eigenvectors

the real PSD is


Rx (e j ω ) = σ 2w 1 + b2 + 2b cos ω

and the autocorrelation sequence is


r x (l) = σ 2w b δ(l + 1) + (1 + b2 ) + b δ(l − 1)

Hence the M × M correlation matrix is

 
1 + b2 b ··· 0 0
 b 1 + b2 ··· 0 0 
 
 .
.. .. .. .. .. 
Rx = σ 2w  . . . . 
 
 0 0 ··· 1 + b2 b 
0 0 ··· b 1 + b2
Consider M = 1: Then

Rx = σ 2w 1 + b2

with
  π (1)
λ = σ 2w 1 + b2 = σ 2w 1 + b2 + 2b cos(π /2) = Rx (e j ω1 ), ω1 = π /2 =
1+1
π (1)
qn(1) = 1 = sin π n/2 = sin ω1 n; n = 1, ω1 = π /2 =
1+1
Statistical and Adaptive Signal Processing - Solution Manual 29

Consider M = 2: Then
  
1 + b2 b
Rx = σ 2w
b 1 + b2

The eigenvalues are

  π (1)
λ1 = σ 2w 1 + b2 + b = σ 2w 1 + b2 + 2b cos(π /3) = Rx (e j ω1 ), ω1 = π/3 =
2+1
  π (2)
λ2 = σ 2w 1 + b2 − b = σ 2w 1 + b2 + 2b cos(2π /3) = Rx (e j ω2 ), ω2 = 2π /3 =
2+1
with unnormalized eigenvectors
√
(1) 3 π (1)
qn = [1, 1] = sin π n/3 = sin ω1 n; n = 1, 2; ω1 = π /3 =
2 2+1
√
3 π (2)
qn(2) = [1, −1] = sin 2π n/3 = sin ω2 n; n = 1, 2; ω2 = 2π /3 =
2 2+1
Consider M = 3: Then
 
1 + b2 b 0
Rx = σ 2w  b 1 + b2 b 
0 b 1+b 2

The eigenvalues are

 √   π (1)
λ1 = σ 2w 1 + b2 + 2b = σ 2w 1 + b2 + 2b cos(π /4) = Rx (e j ω1 ), ω1 = π/4 =
3+1
  j ω2 π (2)
λ2 = σ w 1 + b = σ w 1 + b + 2b cos(2π /4) = Rx (e ), ω2 = 2π /4 =
2 2 2 2
3+1
 √   π (3)
λ2 = σ 2w 1 + b2 − 2b = σ 2w 1 + b2 + 2b cos(3π /4) = Rx (e j ω3 ), ω3 = 3π /4 =
3+1
with unnormalized eigenvectors:
   
√1 sin π/4
1√
qn(1) = 2  2  =  sin 2π /4  = {sin ω1 n}3n=1
2
1 sin 3π /4
   
1 sin 2π /4
qn(2) =  0  =  sin 4π /4  = {sin ω2 n}3n=1
−1 sin 6π /4
   
1 sin 3π /4
1√  √  
qn(3) = 2 − 2 = sin 6π /4  = {sin ω3 n}3n=1
2
1 sin 9π /4

Therefore, for a general M, we have

λk = Rx (e j ωk ); ωk = π k/(M + 1), k = 1, 2, . . . , M

and

qn(k) = sin ωk n; ωk = π k/(M + 1), k = 1, 2, . . . , M, n = 1, 2, . . . , M

30 Statistical and Adaptive Signal Processing - Solution Manual

PSD and Eigenvalues of a MA(1) Process

4

Rx(ejω)

0
0 0.1111 0.2222 0.3333 0.4444 0.5556 0.6667 0.7778 0.8889 1
ω/π

Figure 3.28: Plot of the PSD and eigenvalues of MA(1) process

(b) Assume M = 8, b = 0.9, and σ 2w = 1. The Matlab script is shown below and the plot is shown in
Figure 3.28.

var_w = 1; M = 8; b = 0.9;
r = var_w*[(1+b*b),b,zeros(1,M-2)];
R = toeplitz(r);
Lambda = eig(R); Lambda = flipud(sort(Lambda));
k = 1:M;
omg_k = pi*k/(M+1);
omg = (0:500)*pi/500;
PSD = var_w*((1+b*b)+2*b*cos(omg));

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0328b’);

plot(omg/pi,PSD,’g’); hold on;

stem(omg_k/pi,Lambda,’md:’);
xlabel(’\omega / \pi’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itR_x}({\ite^{j\omega}})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD and Eigenvalues of a MA(1) Process’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0,omg_k/pi,1],’ytick’,[0:1:4]); Hold off;

3.29 Random process: x(n) = w(n) + b w(n − 1).

(a) DKLT for M = 3. The autocorrelation matrix is given by

 
1 α 0
b b
Rx = σ 2w  α 1 α  , where α =
α 1 + b2
0 α 1
Statistical and Adaptive Signal Processing - Solution Manual 31

Now to determine the eigenanalysis, consider

 
1−λ α 0
det(Rx − λI) = 0 = det  α 1−λ α 
0 α 1−λ
= (1 − λ)(1 − 2α 2 − 2λ + λ2 ) = 0
which gives
√ √
1 + 2α 1 − 2α
λ1 = 1, λ2 = , λ3 =
1 − 2α 2 1 − 2α 2
(b) The variances of the DKLT coefficients are given by the eigenvalues which from above in the desending
order are
1 1
1, √ , √
1 − 2α 1 + 2α

3.30 Let x(n) be a stationary random process with mean µx and covariance γ x (l). Let µ̂x be the sample mean
estimator of µx given by

1 
N−1
µ̂x = x (n)
N n=0

(a) The mean of µ̂x is given by

1  1 
N−1 N−1
E[µ̂x ] = E[x (n)] = µ = µx
N n=0 N n=0 x

and the variance of µ̂x is given by

  2  
var µ̂x = E[µ̂x − µx  ] = E[ µ̂x − µx µ̂∗x − µ∗x ]
( N−1 ) ( N−1 )
1  1  ∗
= E x (n) − µx x (m) − µ∗x
N n=0 N m=0

1    1  1   ∗ 
N−1 N−1 N−1 N−1
∗ ∗
= 2
E x (n) x (m) − E [x (n)] µ x − E x (m) µx + µx µ∗x
N n=0 m=0 N n=0 N m=0

1    1   2 1   2  2

N−1 N−1 N−1 N−1
∗
= E x (n) x (m) − µ − µ + µx
N 2 n=0 m=0 N n=0 x N m=0 x

1   
N−1 
  2   1    2
N−1 N−1 N−1 N−1 N−1
= E x (n) x ∗
(m) − µ x  = 1 E x (n) x ∗
(m) − µ
N 2 n=0 m=0 N 2 n=0 m=0 N 2 n=0 m=0 x

1 "    2 # 
N−1 N−1 N−1 
N−1
= E x (n) x ∗
(m) − µ x  = 1 γ (n − m)
N 2 n=0 m=0 N 2 n=0 m=0 x

Thus
 1 
N−1 N−1
var µ̂x = 2 γ (n − m) (15)
N n=0 m=0 x
32 Statistical and Adaptive Signal Processing - Solution Manual

Using chamge of variables n − m = l, we can reduce double summation into a single summation
N

 1  |l|
var µ̂x = 1− γ x (l)
N l=−N N
 |l|

Since 1 − N
≤ 1,

 1  1   
N N
var µ̂x ≤ γ x (l) ≤ γ x (l)
N l=−N N l=−N

(b) From (15)

N
 N

 σ 2x  |l| γ x (l) σ 2x  |l| γ (l)
var µ̂x = 1− = 1− ρ x (l), ρ x (l) , x 2
N l=−N N σx 2 N l=−N N σx

Since for a real-valued random process ρ x (l) = ρ x (−l) and using symmetry of |l|, we obtain
 −1
 N
 
 σ 2x  −l  l
var µ̂x = 1− ρ x (−l) + 1 + 1− ρ x (l)
N l=−N N l=1
N
 N
 
σ 2x  l
= 1+2 1− ρ x (l)
N l=1
N
N

σ 2x   l
= 1 +  N (ρ x ) ,  N (ρ x ) , 2 1− ρ x (l)
N l=1
N

  σ 2x
(c) When x(n)˜WN µx , σ 2x , then ρ x (l) = δ(l) or  N (ρ x ) = 0. Hence var µ̂x = N
.

3.31 Let x(n) be a stationary random process with mean µx and covariance γ x (l). Let σ̂ 2x be the sample variance
estimator of σ 2x given by

1   2
N−1
σ̂ 2x = x (n) − µ̂x 
N n=0

(a) The mean of σ̂ 2x is given by

1  $ 2 % 1  $ 2 %
N−1 N−1
E[σ̂ 2x ] = E x (n) − µ̂x  = E x (n) − µx − µ̂x + µx 
N n=0 N n=0

1  $  2 %
N−1
= E x (n) − µx − µ̂x − µx 
N n=0

1  $ 2      2 %
N−1
= E x (n) − µx  − x (n) − µx µ̂∗x − µ∗x − x ∗ (n) − µ∗x µ̂x − µx + µ̂x − µx 
N n=0
 1      
N−1
= σ 2x + var µ̂x − E x (n) − µx µ̂∗x − µ∗x + x ∗ (n) − µ∗x µ̂x − µx (16)
N n=0
Statistical and Adaptive Signal Processing - Solution Manual 33

Consider
   $  2 %
E x (n) − µx µ̂∗x − µ∗x = E x (n) µ̂∗x − x(n)µ∗x − µx µ̂∗x + µx 
   2  2  2    2
= E x (n) µ̂∗x − µx  − µx  + µx  = E x (n) µ̂∗x − µx  (17)
Similarly,
      2
E x ∗ (n) − µ∗x µ̂x − µx = E x ∗ (n) µ̂x − µx  (18)
Consider
 
  1 
N−1
1  
N−1

E x (n) µ̂∗x = E x(n) ∗
x (m) = E x(n)x ∗ (m)
N m=0
N m=0

1 $  2 % 1   2
N−1 N−1
= γ x (n − m) + µx  = γ x (n − m) + µx  (19)
N m=0 N m=0
and
  1 $ ∗  2 % 1  ∗  2
N−1 N−1
E x ∗ (n) µ̂x = γ x (n − m) + µx  = γ x (n − m) + µx  (20)
N m=0 N m=0
Substituting (17), (18), (19), and (20) in (16), we obtain
 1  1  ∗
N−1 N−1 N−1 N−1
E[σ̂ 2x ] = σ 2x + var µ̂x − 2 γ x (n − m) − 2 γ (n − m) (21)
N n=0 m=0 N n=0 m=0 x
But from (15) in Problem 3.30
 1  1  ∗
N−1 N−1 N−1 N−1
var µ̂x = 2 γ x (n − m) = 2 γ (n − m) (22)
N n=0 m=0 N n=0 m=0 x
Substituting (22) in (21), we obatin
N

 1  |l|
E[σ̂ 2x ] = σ 2x − var µ̂x = σ x −
2
1− γ x (l)
N l=−N N
The last equality follows from Problem 3.30.
 σ 2x
(b) When x(n)˜WN µx , σ 2x , then γ x (l) = σ 2x δ(l). Hence E[σ̂ 2x ] = σ 2x − N
= (N − 1) σ 2x /N .

1 1 1 !N
3.32 Cauchy distribution with mean µ: f x (x) = . The mean estimator is µ̂(ζ ) = xk (ζ ). Then
π 1_ (x − µ)2 N k=1
the mean of µ̂(ζ ) = µ and

 2
var(µ̂) = E{ µ̂ − µ } = E{µ̂2 − 2µµ̂ + µ2 |}
1 
N
= E{µ̂2 } − 2µ E{µ̂} + µ2 = E{( xk (ξ ))2 } − µ2
N k=1

1 
N
= E{xk (ξ )2 } − µ2
N 2 k=1
E{x(ξ )2 }
= − µ2  = 0 as N → ∞
N
Therefore, µ̂(ζ ) is not a consistent estimator.
 Chapter 4

Linear Signal Models

4.1 Show that a second-order pole pi contributes the term npin u(n) and a third-order pole the terms npin u(n) +
n 2 pin u(n) to the impulse response of a causal PZ model.
The general expression for a PZ model with double poles in the z-domain is
D(z) k=1Q
(1 − z k z −1 )
H (z) = = d0 ; Q < P, |z| > max (| pk |) , max | pk | < 1
A(z) (1 − p0 z −1 )2 k=1
P−2
(1 − pk z −1 )
Using partial fraction expansion on this expression and only keeping the double term simplifies to
A1 z −1 A2 B(z)
H (z) = + + P−2
(1 − p0 z −1 )2 (1 − p0 z −1 ) k=1 (1 − pk z −1 )
where B(z) is a polynomial of order less than P − 2. After taking inverse z-transform,
 
B(z)
h(n) = A1 n p0n u(n) + A2 p0n u(n) + Z −1
k=1
P−2
(1 − pk z −1 )
Similarly, for a third-order pole, we can write
A1 z −2 A2 z −1 A3 B(z)
H (z) = + + P−3
(1 − p0 z ) (1 − p0 z )
−1 3 −1 2 (1 − p0 z ) k=1 (1 − pk z −1 )
−1

which upon inverse z-transformation yields

 
−1 B(z)
h(n) = A1 n 2
p0n u(n) + A2 n p0n u(n) + A3 p0n u(n) + Z
k=1
P−2
(1 − pk z −1 )

4.2 A zero-mean random sequence x(n) with PSD:

5 + 3 cos ω
Rx (e j ω ) =
17 + 8 cos ω
(a) The innovations representations of the process x(n) can be found by transforming x(n) into the z-domain
  
z 10 + 3(z + z −1 ) 3z 2 + 10z + 3 9 1 + 13 z 1 + 13 z −1
Rx (z) =  = 2 =  
z 34 + 8(z + z −1 ) 8z + 34z + 8 32 1 + 14 z 1 + 14 z −1
which can be written as
 
∗ 9 1 + 13 z 1 + 13 z −1
Rx (z) = σ 2w Hmin (z)Hmin (1/z ∗ ) =  
32 1 + 14 z 1 + 41 z −1
Clearly, Hmin (z) is equal to
1 + 13 z −1
Hmin (z) =
1 + 14 z −1
and the resulting innovations process is
1 1
x(n) = − x(n − 1) + w(n) + w(n − 1)
4 3
with σ 2w = 9
32
= 0.28125

34
Statistical and Adaptive Signal Processing - Solution Manual 35

(b) The autocorrelation sequence r x (l) can be found using the factorization of Rx (z):
 
9 1 + 13 z 1 + 13 z −1
Rx (z) =  
32 1 + 14 z 1 + 14 z −1
3 11 11 1
= −  +  , < |z| < 4
8 120 1 + 4 z 1 −1 120 1 + 4z −1 4

Hence


3 11 1 |l|
r x (l) = δ(l) − −
8 120 4
 1 |l|  |l|
4.3 A sequence with an autocorrelation r x (l) = 2
+ − 12 .

(a) From the autocorrelation, we can see that x(n) has no dependence with x(n−1) but does have a dependence
on x(n − 2). We can assume x(n) to have the following form

x(n) = ax(n − 2) + bw(n)

Solving for a and b

r x (0) = a 2 E{x(n − 2)2 } + b2 E{w(n)2 } + 2abE{x(n − 2)w(n)}

= a 2 r x (0) + b2
r x (2) = a E{x(n − 2)} + bE{w(n)x(n − 2)} = ar x (0)

which results in a = 1
4
and b = 15
8
Checking to see that x(n) has no dependence on x(n − 1)

r x (1) = E{x(n)x(n − 1)} = ar x (1) = 0

Therefore,
1 15
x(n) = x(n − 2) + w(n)
4 8
(b) A pdf of the realization of the process is shown in Figure 4.3b. The dotted line represents a true Gaussian
pdf.
(c) The true and estimated autocorrelation are shown in Figure 4.3c.

4.4 Computing the autocorrelations directly for the two systems, where w(n) ∼ WGN(0,1)

(a) x1 (n) = w(n) + 0.3w(n − 1) − 0.4w(n − 2)

r x (l) = E{[w(n) + 0.3w(n − 1) − 0.4w(n − 2)][w(n + l) + 0.3w(n + l − 1) − 0.4w(n + l − 2)]}

= δ(l) + 0.3δ(l − 1) − 0.4δ(l − 2)
+0.3δ(l − 1) + (0.3)2 δ(l) − (0.4)(0.3)δ(l − 1)
+(−0.4)δ(l − 2) + (−0.4)(0.3)δ(l − 1) + (−0.4)2 δ(l)
= 1.25δ(l) + 0.36δ(l − 1) − 0.8δ(l − 2)
36 Statistical and Adaptive Signal Processing - Solution Manual

True (*) and Estimated PDF

100 True (dotted) and Estimated Autocorrelation
1.2

80 1

0.8
60
0.6

40 0.4

0.2
20
0

0 −0.2
−5 0 5 0 2 4 6 8 10
Figure 4.3bc: Gaussian Process

(b) x2 (n) = w(n) − 1.2w(n − 1) − 1.6w(n − 2)

r x (l) = E{[w(n) − 1.2w(n − 1) − 1.6w(n − 2)][w(n + l) − 1.2w(n + l − 1) − 1.6w(n + l − 2)]}

= δ(l) − 1.2δ(l − 1) − 1.6δ(l − 2)
−1.2δ(l − 1) + (−1.2)2 δ(l) + (−1.2)(−1.6)δ(l − 1)
+(−1.6)δ(l − 2) + (−1.6)(−1.2)δ(l − 1) + (−1.6)2 δ(l)
= 5δ(l) + 1.44δ(l − 1) − 3.2δ(l − 2)

Comparing these two autocorrelation, they both have the same ratio of dependence on the previous two inputs.

4.5 Investigate how well the all-zero systems approximate the single pole system
The single pole system with a pole at a has a system function of
 ∞
1
H (z) = −1
= a n z −n
1 − az n=0

It can be approximated with a finite number of zeros with a system function of


N
HN (z) = a n z −n
n=0

The Figure 4.5.1 shows the pole-zero plots, magnitude plots, and impulse response for a single pole system
with the poles located at a = 0.7 and 0.95 respectively.
Figure 4.5.2 shows the pole-zero plots, magnitude plots and impulse response for a = 0.7 and N = 8 Figure
4.5.3 shows the pole-zero plots, magnitude plots and impulse response for a = 0.7 and N = 16 Figure 4.5.4
shows the pole-zero plots, magnitude plots and impulse response for a = 0.7 and N = 64 Figure 4.5.5 shows
the pole-zero plots, magnitude plots and impulse response for a = 0.95 and N = 8 Figure 4.5.6 shows the
pole-zero plots, magnitude plots and impulse response for a = 0.95 and N = 16 Figure 4.5.7 shows the
pole-zero plots, magnitude plots and impulse response for a = 0.95 and N = 64
Clearly from these figures, it can be seen that an all zero system can approximate a single pole system. The
number of zeroes necessary depends on how close to the unit circle the pole lies. As the pole moves closer to
the unit circle, more zeroes are needed to have the same magnitude response.
Statistical and Adaptive Signal Processing - Solution Manual 37

a=0.7, N=∞
a=0.7 a=0.7
1
3.5 1

3
0.5 0.8
Imaginary part
2.5
0 0.6
2
0.4
−0.5 1.5

1 0.2
−1
−1 −0.5 0 0.5 1 0.5 0
Real part 0 0.2 0.4 0.6 0.8 1 0 20 40 60 80 100
a=0.95, N=∞
a=0.95
1 20
a=0.95
1

0.5 0.8
15
Imaginary part

0 0.6
10
0.4
−0.5
5
0.2
−1
−1 −0.5 0 0.5 1 0 0
Real part 0 0.2 0.4 0.6 0.8 1 0 20 40 60 80 100

Figure 4.5.1: Gaussian Process

a=0.7, N=8
1
a=0.7, N=8 a=0.7, N=8
3.5 1

0.5 3
0.8
Imaginary part

2.5
0 0.6
2
0.4
−0.5 1.5

0.2
1
−1
−1 −0.5 0 0.5 1 0.5 0
Real part 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8

Figure 4.5.2: Gaussian Process

a=0.7, N=16
1
a=0.7, N=16 a=0.7, N=16
3.5 1

0.5 3
0.8
Imaginary part

2.5
0 0.6
2
0.4
−0.5 1.5

0.2
1
−1
−1 −0.5 0 0.5 1 0.5 0
Real part 0 0.2 0.4 0.6 0.8 1 0 5 10 15 20

Figure 4.5.3: Gaussian Process

38 Statistical and Adaptive Signal Processing - Solution Manual

a=0.7, N=64
1
a=0.7, N=64 a=0.7, N=64
3.5 1

0.5 3
0.8
Imaginary part

2.5
0 0.6
2
0.4
−0.5 1.5

0.2
1
−1
−1 −0.5 0 0.5 1 0.5 0
Real part 0 0.2 0.4 0.6 0.8 1 0 20 40 60 80

Figure 4.5.4: Gaussian Process

a=0.95, N=8
a=0.95, N=8
1 8
a=0.95, N=8
1

0.5 6 0.8
Imaginary part

0 0.6
4
0.4
−0.5
2
0.2
−1
−1 −0.5 0 0.5 1 0 0
Real part 0 0.2 0.4 0.6 0.8 1 0 2 4 6 8

Figure 4.5.5: Gaussian Process

a=0.95, N=16
a=0.95, N=16 a=0.95, N=16
1 12 1

10
0.5 0.8
Imaginary part

8
0 0.6
6
0.4
−0.5 4

2 0.2
−1
−1 −0.5 0 0.5 1 0 0
Real part 0 0.2 0.4 0.6 0.8 1 0 5 10 15 20

Figure 4.5.6: Gaussian Process

a=0.95, N=64
a=0.95, N=64 a=0.95, N=64
1 20 1

0.5 0.8
15
Imaginary part

0 0.6
10
0.4
−0.5
5
0.2
−1
−1 −0.5 0 0.5 1 0 0
Real part 0 0.2 0.4 0.6 0.8 1 0 20 40 60 80

Figure 4.5.7: Gaussian Process

Statistical and Adaptive Signal Processing - Solution Manual 39

4.6 Consider (4.2.33) for P = 2:

      
ρ1 1 ρ1 a1 a1 + ρ 1 a2
− = =
ρ2 ρ1 1 a2 ρ 1 a1 + a2

Hence

−a1 = ρ 1 + ρ 1 a2 = ρ 1 (1 + a2 )
−a2 = ρ 1 a1 + ρ 2

or
     &   '  
a1 1 + a2 0 ρ1 1 0 a2 0 ρ1
−a = − = = + = Aρ
a2 a1 1 ρ2 a1 1 0 0 ρ2

Thus the matrix A is a sum of triangular Toeplitz and triangular Hankel matrices. Consider (4.2.33) for P = 3:
      
ρ1 1 ρ1 ρ2 a1 a1 + ρ 1 a2 + ρ 2 a3
−  ρ 2  =  ρ 1 1 ρ 1   a2  =  ρ 1 a1 + a2 + ρ 1 a3 
ρ3 ρ2 ρ1 1 a3 ρ 2 a1 + ρ 1 a2 + a3

Hence

−a1 = ρ 1 + ρ 1 a2 + ρ 2 a3 = ρ 1 (1 + a2 ) + ρ 2 a3
−a2 = ρ 1 a1 + ρ 2 + ρ 1 a3 = ρ 1 (a1 + a3 ) + ρ 2
−a3 = ρ 2 a1 + ρ 1 a2 + ρ 3 = ρ 1 a2 + ρ 2 a1 + ρ 3

or
    
a1 (1 + a2 ) a3 0 ρ1
−a = −  a2  =  (a1 + a3 ) 1 0   ρ2 
a3 a2 a1 1 ρ3
     
 1 0 0 a2 a3 0  ρ1
=  a1 1 0  +  a3 0 0   ρ 2  = Aρ
 
a2 a1 1 0 0 0 ρ3

Again the matrix A is a sum of triangular Toeplitz and triangular Hankel matrices. Similarly for P = 4, we
obtain
      
ρ1 1 ρ1 ρ2 ρ3 a1 a1 + ρ 1 a2 + ρ 2 a3 + ρ 3 a4
 ρ 2   ρ 1 1 ρ 1 ρ 2   a2   ρ 1 a1 + a2 + ρ 1 a3 + ρ 2 a4 
−     
 ρ 3  =  ρ 2 ρ 1 1 ρ 1   a3  =  ρ 2 a1 + ρ 1 a2 + a3 + ρ 1 a4 


ρ4 ρ3 ρ2 ρ1 1 a4 ρ 3 a1 + ρ 2 a2 + ρ 1 a3 + a4
or
 
     
a1 
 1 0 0 0 a2 a3 a4 0  ρ
 a2    1
  a1 1 0 0   a3 a4 0 0  ρ 
−a = −  =  +   2  = Aρ
a3  
 a2 a1 1 0   a4 0 0 0 

 ρ3 
 
a4 a3 a2 a1 1 0 0 0 0 ρ4

Hence for a general case, the matrix A is given by the sum of triangular Toeplitz and triangular Hankel matrices.
40 Statistical and Adaptive Signal Processing - Solution Manual

4.7 To determine the autocorrelation and partial autocorrelation coefficients of the following two AR models, we
can use the Yule-Walker equations
Rx a = −rx

(a) x(n) = 0.5x(n − 1) + w(n):

Solving directly,
    
r x (0) r x (1) 1 |d0 |2
=
r x (1) r x (0) a 0
    
r x (0) r x (1) 1 1
=
r x (1) r x (0) −0.5 0
Rearranging terms and computing the matrix inverse results in
    
−0.5 1 r x (1) 1
=
1 −0.5 r x (0) 0
      
r x (1) 0.667 1.333 1 0.667
= =
r x (0) 1.333 0.667 0 1.333
with a variance σ 2x = r x (0) = 1.333 The corresponding partial autocorrelation is
a1(1) = −ρ(1) = r x (1)/r x (0)
= −0.667/1.333
= −0.5
Therefore the autocorrelation coefficients are r x (0) = 1.33 and r x (1) = 0.667 with a partial autocorrela-
tion coefficient equal to a1(1) = −0.5.
(b) x(n) = 1.5x(n − 1) − 0.6x(n − 2) + w(n)
Solving directly,
    
r x (0) r x (1) r x (2) 1 |d0 |2
 r x (1) r x (0) r x (1)   a1  =  0 
r x (2) r x (1) r x (0) a2 0
    
r x (0) r x (1) r x (2) 1 1
 r x (1) r x (0) r x (1)   −1.5  =  0 
r x (2) r x (1) r x (0) 0.6 0
Rearranging terms and computing the matrix inverse results in
    
1 −1.5 0.6 r x (0) 1
 −1.5 1.6 0   r x (1)  =  0 
0.6 −1.5 1 r x (2) 0
      
r x (0) 12.9 4.84 −7.74 1 12.9
 r x (1)  =  12.01 5.16 −7.26   0  =  12.01 
r x (2) 10.4 4.84 −5.24 0 10.4
with a variance σ 2x = r x (0) = 12.9. The corresponding partial autocorrelation is
a1(1) = −ρ(1) = −12.01/12.9 = −0.931
    
1 ρ(1) a1(2) ρ(1)
= −
ρ(1) 1 a2(2) ρ(2)
   −1    
a1(2) 1 0.931 0.931 1.356
= − =
a2(2) 0.931 1 0.806 0.456
Statistical and Adaptive Signal Processing - Solution Manual 41

Therefore the autocorrelation coefficients are r x (0) = 12.9, r x (1) = 12.01 and r x (3) = 10.4 with the
partial autocorrelation coefficients equal to a1(1) = −0.931 and a2(2) = 0.456.

4.8 Given x(n) = x(n − 1) − 0.5x(n − 2) + w(n):

(a) To find ρ x (1) first solve for the first two autocorrelation coefficients

r x (0) = E{[x(n − 1) − 0.5x(n − 2) + w(n)][x(n − 1) − 0.5x(n − 2) + w(n)]}

= (1.25)r x (0) − r x (1) + 1
r x (1) = E{x(n)x(n − 1)} = E{[x(n − 1) − 0.5x(n − 2) + w(n)]x(n − 1)}
= r x (0) − 0.5r x (1)

or 1.5r x (1) = r x (0). Therefore, ρ x (1) = r x (1)/r x (0) = 2/3.

(b) To compute ρ x (l) use the following difference equation

ρ x (l) = ρ x (l − 1) − 0.5ρ x (l − 2)

(c) The plot of ρ x (l) is shown in the left plot of Figure 4.8cd.
(d) The corresponding PSD of the model is shown in the right plot of Figure 4.8cd.

PSD of rho(l)
rho(l) 4
1

0.8
3
0.6

0.4 2
0.2

0 1
−0.2

−0.4 0
0 5 10 15 0 0.2 0.4 0.6 0.8 1
Figure 4.8cd: Period of ρ(l)

The period (T ) shown in the plot of ρ(l) is approximately 8. Thus, the frequency is 0.125(1/T ). This is
approximately the frequency peak in the PSD plot.
42 Statistical and Adaptive Signal Processing - Solution Manual

4.9 The ACS can be computed directly by rearranging terms and using the matrix inverse
    
r x (0) r x (1) r x (2) r x (3) 1 |d0 |2
 r x (1) r x (0) r x (1) r x (2)   a1   0 
    
 r x (2) r x (1) r x (0) r x (1)   a2  =  0 
r x (3) r x (2) r x (1) r x (0) a3 0
    
1 a1 a2 a3 r x (0) |d0 |2
 a1 1 + a2 a3 0   r x (1)   
   =  0 
 a2 a1 + a3 1 0   r x (2)   0 
a3 a2 a1 1 r x (3) 0
   −1  
r x (0) 1 a1 a2 a3 |d0 |2
 r x (1)     
  =  a1 1 + a2 a3 0   0 
 r x (2)   a2 a1 + a3 1 0   0 
r x (3) a3 a2 a1 1 0
Substituting in the values for the parameters from Example 4.2.3 results in the correct values for the ACS.
      
r x (0) 2.8840 2.0011 0.0003 0.4508 0.6937 2.0
 r x (1)      1.6 
  =  2.3074 2.5236 0.1444 0.3607   0 = 
 r x (2)   1.7309 2.1813 1.1534 0.2705   0   1.2 
r x (3) 1.4426 1.6588 1.0093 1.2255 0 1.0
Note: this problem can also be solved using Cramer’s rule. The solution for Problem 4.12 shows how to use
Cramer’s rule.

4.10 Given the AP(3) model: x(n) = 0.98x(n − 3) + w(n)

(a) In order to plot the PSD of x(n) first compute the ACS

r x (0) = E{x(n)x(n)} = E{[0.98x(n − 3) + w(n)][0.98x(n − 3) + w(n)]}

= 0.982 r x (0) + 1 = 25.25
r x (1) = E{[0.98x(n − 3) + w(n)][x(n − 1)]} = 0.98r x (2)
r x (2) = E{[0.98x(n − 3) + w(n)][x(n − 2)]} = 0.98r x (1)
r x (3) = E{[0.98x(n − 3) + w(n)][x(n − 3)]} = 0.98r x (0)
r x (6) = E{[0.98x(n − 3) + w(n)][x(n − 6)]} = 0.98r x (3) = 0.982 r x (0)

This results in an ACS of


 l
 (0.98) 3 r x (0) l = 0, 3, 6, 9, ...
r x (l) = r x (0) = 25.25


0 otherwise
The PSD can be found by taking the Fourier transform of the ACS. This is shown in the left plot of Figure
4.10ab.
(b) The PSD is computed by taking the squared-magnitude of the Fourier transform of realization of x(n)
This is shown in the right plot of Figure 4.10ab.
(c) Given y(n) = 13 [x(n − 1) + x(n) + x(n + 1)], its ACS is computed as follows

r y (l) = ( 13 )2 [3r x (l) + 4r x (l + 1) + 2r x (l + 2)]

Statistical and Adaptive Signal Processing - Solution Manual 43

PSD of x(n) PSD of realization of x(n)

1400 250

1200
200
1000
150
800

600 100
400
50
200

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 4.10ab: Power Spectral Density

PSD of y(n) Magnitude of H(z)=Y(z)/X(z)

150 1

0.8

100
0.6

0.4
50

0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 4.10cd: Power Spectral Density

(d) The PSD of y(n) is computed in a similar manner as x(n) above. It is shown in the left plot of Figure
4.10cd. The right plot of Figure 4.10cd shows the magnitude response of the system H (z) = Y (z)/x(z)
Clearly, H (z) is a notch filter, with the frequency notches corresponding to the peak frequency of the PSD
of x(n). This filter removes the strong correlation found in x(n) and generates an uncorrelated signal
found in part (c) above.

4.11 Given two AR(2) models:

(a) The normalized autocorrelation sequence, for each of the two models, can be found directly as follows
i. x(n) = 0.6x(n − 1) + 0.3x(n − 2) + w(n):

r x (0) = E{[0.6x(n − 1) + 0.3x(n − 2) + w(n)][0.6x(n − 1) + 0.3x(n − 2) + w(n)]}

= r x (0)(0.62 + 0.32 ) + 2(0.6)(0.3)r x (1) + 1

r x (1) = E{x(n)x(n − 1)} = E{[0.6x(n − 1) + 0.3x(n − 2) + w(n)][x(n − 1]}

= 0.6r x (0) + 0.3r x (1) = 0.857r x (0)
44 Statistical and Adaptive Signal Processing - Solution Manual

r x (l) = E{x(n + l)x(n)} = E{x(n + l)[0.6x(n − 1) + 0.3x(n − 2) + w(n)]}

= 0.6r x (l − 1) + 0.3r x (l − 2)

Thus r x (0) = σ 2x = 4.146, and r x (1) = 3.553. Therefore, the normalized autocorrelation sequence
is

ρ x (l) = r x (l)/r x (0) = 0.6ρ(l − 1) + 0.3ρ(l − 2)

ii. x(n) = 0.8x(n − 1) − 0.5x(n − 2) + w(n):

r x (0) = E{[0.8x(n − 1) − 0.5x(n − 2) + w(n)][0.8x(n − 1) − 0.5x(n − 2) + w(n)]}

= r x (0)(0.82 + 0.52 ) + 2(0.8)(−0.5)r x (1) + 1

r x (1) = E{x(n)x(n − 1)} = E{[0.8x(n − 1) − 0.5x(n − 2) + w(n)][x(n − 1]}

= 0.8r x (0) − 0.5r x (1) = 0.533r x (0)

r x (l) = E{x(n + l)x(n)} = E{x(n + l)[0.8x(n − 1) − 0.5x(n − 2) + w(n)]}

= 0.8r x (l − 1) − 0.5r x (l − 2)

Thus r x (0) = σ 2x = 1.863 and r x (1) = 0.994. Therefore, the normalized autocorrelation sequence
is

ρ x (l) = r x (l)/r x (0) = 0.8ρ(l − 1) − 0.5ρ(l − 2)

rho1(l) rho2(l)
1 1

0.8
0.8
0.6
0.6
0.4

0.4 0.2

0
0.2
−0.2

0 −0.4
0 5 10 15 0 5 10 15
Figure 4.11b: AR(2) System

(b) Figure 4.11b shows the normalized autocorrelation sequences of the two models respectively. From these
plots, x(n) = 0.6x(n − 1) + 0.3x(n − 2) + w(n) does not show any pseudo-periodic behavior, while
x(n) = 0.8x(n − 1) − 0.5x(n − 2) + w(n) does show pseudo-periodic behavior.
(c) Figure 4.11c shows the PSD of the two models. Clearly, the first model has no fundamental frequency.
While the second system does indeed have a large no zero frequency component.

4.12 The PACS for an AP(3) model can be derived from the Yule-Walker equations and Cramer’s rule.
Statistical and Adaptive Signal Processing - Solution Manual 45

PSD of x1(n) PSD of x2(n)

10 2

1.5
6

4
1

0 0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 4.11c: AR(2) System

(a) The Yule-Walker equations are

Rx a = −rx

Cramer’s rule is used to determine vector values without having to compute the inverse of the entire
matrix. If x = A−1 b then the j th component of x can be computed using the determinants of two
det B
matrices x j = det Aj where B j is the A matrix with the j th column replaced with the b vector
 
a1,1 · · · a1, j −1 b1 a1, j +1 · · · a1,n
 .. 
B j =  ... ..
.
..
.
..
. . 
an,1 · · · an, j −1 bn an, j +1 · · · an,n

Using Yule-Walker and Cramer’s rule, the PACS can be computed using
(M)
detRx(m)
am(M) =
detRx(M)
where
 
r x (0) r x (1) r x (2)
Rx(M) =  r x∗ (1) r x (0) r x (1) 
r x∗ (2) r x∗ (1) r x (0)
 
r x (1)
(M)  .. 
and Rx(m) = Rx(M) with the j th column replaced with r x =  . 
r x (M)
(b) The following MATLAB code can be used to compute the PACS of the AP(3) model in Example 4.2.3
MATLAB Code:
r0=2; r1=1.6; r2=1.2; r3=1;
R=det([r0 r1 r2; r1 r0 r1; r2 r1 r0])
a1 = -det([r1 r1 r2; r2 r0 r1; r3 r1 r0])/R
a2 = -det([r0 r1 r2; r1 r2 r1; r2 r3 r0])/R
a3 = -det([r0 r1 r1; r1 r0 r2; r2 r1 r3])/R
46 Statistical and Adaptive Signal Processing - Solution Manual

This code produces the following results

R = 1.0240
a1 = -0.9063
a2 = 0.2500
a3 = -0.1563
which is equivalent to the results in Example 4.2.3

4.13 To show that the spectrum of any PZ model with real coefficients has zero slope at ω = 0 and ω = π, the
first derivative of the PSD of the model is computed and found to be equal to zero at ω = 0 and ω = π. The
general expression for a PZ model in the z-domain is
!
α i z −i
H (z) = !
β i z −i

Its squared-magnitude is
! ! !
| α i z −i |2 n αm αn z
−m+n
|H (z)|2 = ! = !m
!
| β i z −i |2 k l β k βl z
−k+l

Replacing z = e− j ω gives the following general expression for the PSD of a PZ model
! !
jω 2 α m α n e− j (m−n)ω
|H (e )| = !m !n − j (k−l)ω
k l β k βl e

Taking its first derivative

d|H (e j ω )|2  
= ( β k β l e− j (k−l)ω )( α m α n (− j )(m − n)e− j (m−n)ω )
dω k l m n
 
− j (m−n)ω
−( αm αn e )( β k β l (− j )(k − l)e− j (k−l)ω )
m n k l

= α m α n β k β l (− j )((m − n) − (k − l))e− j (k−l+m−n)ω
k l m n

= −j α m α n β k β l (m − n − k + l)e− j (k−l+m−n)ω
k l m n

At ω = 0
d|H (e j ω )|2 
= −j α m α n β k β l (m − n − k + l) = 0
dω k l m n

and at ω = π
d|H (e j ω )|2 
= j α m α n β k β l (m − n − k + l) = 0
dω k l m n

Therefore, the spectrum has zero slope.

4.14 Derive the equations for the minimum-phase region of the AP(2) model, which are

|a2 | < 1; a2 − a1 > −1; a2 + a1 > −1

Statistical and Adaptive Signal Processing - Solution Manual 47

(a) | p1 | < 1 and | p2 | < 1:

Separating the magnitude into two one-sided inequalities p1 < 1 and p1 > 1 The derivation follows
[where a1 = −( p1 + p2 ) and a2 = p1 p2 ]

p1 < 1

Multiplying both sides by the negative quantity p2 − 1

p1 ( p2 − 1) > p2 − 1
p1 p2 − p1 − p2 > −1
a2 + a1 > −1

Similarly,

p1 > −1

Multiplying both sides by the positive quantity p2 + 1

p1 ( p2 + 1) > − p2 − 1
p1 p2 + p1 + p2 > −1
a2 − a1 > −1

Lastly, with the magnitude of both p1 and p2 less then 1, there product is also less then one, and |a2 | < 1.
(b) |k1 | < 1 and |k2 | < 1:
Given k2 = a2 , then |a2 | < 1. With
a1
k1 =
1 + a2
and direct substitution into |k1 | < 1 the equations can be derived as follows
 
 a1 
 
1 + a  < 1
2

Solving both side of the magnitude equation separately results in

a1 > 1 + a2 :: ⇒:: a2 − a1 > −1

and

a1 > −(1 + a2 ) :: ⇒:: a2 + a1 > −1

4.15 Spectrum of AP models

(a) Show that the spectrum of an AP(2) model with real coefficients can be obtained by the cascade connection
of two AP(1) models with real coefficients.
First, the general equation for the denominator of the spectrum of an AP(2) model with real coefficients
is computed. Starting in the z-domain

A2 (z) = (1 + p1 z −1 )(1 + p2 z −1 )

Finding its squared-magnitude

|A2 (z)|2 = A2 (z)A∗2 ( 1z ) = (1 + p1 z −1 )(1 + p2 z −1 )(1 + p1 z 1 )(1 + p2 z 1 )

48 Statistical and Adaptive Signal Processing - Solution Manual

Converting to the frequency domain by substituting z = e− j ω result in the spectrum as follows

|A2 (ω)|2 = 1 + ( p1 + p2 )2 + ( p1 p2 )2 + 2( p1 + p2 )(1 + p1 p2 ) cos ω + 2( p1 p2 ) cos 2ω

The denominator of the spectrum of an AP(1) model is

|A1 (ω)|2 = 1 + p12 + 2 p1 cos ω

Cascading two systems multiplies their corresponding spectra. In the case of AP models, it results in the
multiplication of their two denominators.

|A1 (ω)|2 |A2 (ω)|2 = (1 + p12 + 2 p1 cos ω)(1 + p22 + 2 p2 cos ω)

Using the trigonometric identity cos 2α = 2 cos2 α − 1 and completing the above multiplication, results
in the Therefore being equal.

ACS of AP(2) (p1=0.6, p2=−0.9)

Impulse Response (p1=0.6, p2=−0.9) 3
1

0.8 2
0.6
1
0.4

0.2 0

0
−1
−0.2

−0.4 −2
0 20 40 60 80 100 0 20 40 60 80 100
Spectrum of AP(2) (p1=06, p2=−0.9)
PACS of AP(2) (p1=0.6, p2=−0.9) 7
0.8
6
0.6
5

0.4 4

3
0.2
2
0
1

−0.2 0
0 1 2 3 4 5 0 0.2 0.4 0.6 0.8 1
Figure 4.15.1: AP(2) System

(b) Figure 4.15.1 and 4.15.2 show the impulse response, ACS, PACS, and spectrum of an AP(2) model.
Figure 4.15.1 has two real poles located at p1 = 0.6 and p2 = −0.9. Figure 4.15.2 has a double pole
located at p = 0.9.
Statistical and Adaptive Signal Processing - Solution Manual 49

Impulse Response (p1=0.9, p2=0.9)

4 ACS of AP(2) (p1=0.9, p2=0.9)
300

250
3
200

2 150

100
1
50

0 0
0 20 40 60 80 100 0 20 40 60 80 100

PACS of AP(2) (p1=0.9, p2=0.9) Spectrum of AP(2) (p1=0.9, p2=0.9)

0.4 100

0.2
80
0
60
−0.2

−0.4 40
−0.6
20
−0.8

−1 0
0 1 2 3 4 5 0 0.2 0.4 0.6 0.8 1
Figure 4.15.2: AP(2) System with double poles

4.16 Prove Equation 4.2.89 which states that the peak of the spectrum in an AP(2) system with complex conjugate
poles is

(1 + r 2 )
cos ωc = cos θ
2r
This can be solved directly from the spectrum of an AP(2) system with complex conjugate poles

d02
R(e j ω ) =
[1 − 2r cos(ωc − θ) + r 2 ][1 − 2r cos(ωc + θ) + r 2 ]
Taking the derivative of the denominator and setting it equal to zero results in

(1 + r 2 )(sin ωc cos θ) − r sin(2ω c ) = 0

Using the following trigonometric identity sin 2α = 2 sin α cos α and simplifying finishes the proof. Figure
4.16 shows the spectrum for various values of r and θ. The vertical line is computed using Equation 4.2.89
50 Statistical and Adaptive Signal Processing - Solution Manual

r=0.5 theta=0 r=0.9 theta=pi/4

20 60

50
15
40

10 30

20
5
10

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
r=0.5 theta=pi/4 r=0.9 theta=3pi/4
4 60

50
3
40

2 30

20
1
10

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
r=0.5 theta=3pi/4
4
r=0.5 theta=5pi/8
2.5

3 2

1.5
2
1

1
0.5

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
r=0.9 theta=7pi/8
r=0.9 theta=pi 200
10000

8000 150

6000
100
4000

50
2000

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 4.16: AP(2) Complex Pole Systems

Statistical and Adaptive Signal Processing - Solution Manual 51

4.17 Prove that if the AP(P) model A(z) is minimum-phase, then

 π
1 1
log dω = 0
2π −π |A(e j ω )|2
Start with
 π
1
= log |A(e− j ω )|2 dω
2π −π

Using the log identity

 π
1
= 2Re( log A(e− j ω )dω)
2π −π

Change to the z-domain, where the integral becomes a contour integral

2
1
= 2Re( log A(z −1 )dz)
2 j zπ 

and  is a contour that includes all the zeros of A(z). This equation then simplifies to

= 2Re(log[A(∞)]) = 2Re[log(1)]

Since the log(1) = 0, this proves the above equation.

4.18 Recreate plots in Figure 4.8.

(a) Prove Equations (4.2.101 and 4.2.102) which are respectively

4k2
0 < k2 < 1 and k12 <
(1 + k2 )2
Starting with the denominator of the system

A(z) = 1 + a1 z −1 + a2 z −2

Using the quadratic formula, this system has complex poles when

a12 < 4a2

as long as a2 is positive. Using a1 = k1 (1 + k2 ) and a2 = k2 , we have

4k2
0 < a2 < 1 ⇒ 0 < k2 < 1 and a12 < 4a2 ⇒ k12 <
(1 + k2 )2
which proves the above equations. The Matlab script file to plot Figure 4.8a is given below and the plot
is shown in Figure 4.18a.
% (a) Plot in Figure 4.8a
k2 = 0:0.01:1; k1 = 2*sqrt(k2)./(1+k2);
k2 = [fliplr(k2(2:end)),k2,1]; k1 = [-fliplr(k1(2:end)),k1,-1];

subplot(1,2,1);
fill([-1,1,1,-1,-1],[-1,-1,1,1,-1],[0.5,0.5,0.5]); hold on;
fill(k1,k2,[0.8,0.8,0.8]);
52 Statistical and Adaptive Signal Processing - Solution Manual

(a) (k1,k2) space (b) (ρ(1),ρ(2)) space

1 1

Complex poles

0.5 0.5

Real and equal poles

ρ(2)
k2

0 0

−0.5 Real poles −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
k1 ρ(1)

Figure 4.18: Minimum-phase and positive definiteness regions for the AP(2) model

plot([-1,1],[0,0],’w:’,[0,0],[-1,1],’w:’);
set(gca,’xtick’,[-1:0.5:1],’ytick’,[-1:0.5:1]);
xlabel(’k_1’,’fontsize’,label_fontsize);
ylabel(’k_2’,’fontsize’,label_fontsize);
axis([-1,1,-1,1]); axis(’square’);
text(-0.3,-0.5,’Real poles’);
text(-0.45,0.8,’Complex poles’);
text(-0.65,0.2,’Real and equal poles’);
title(’(a) (k_1,k_2) space’,’fontsize’,title_fontsize);
hold off;

(b) The Matlab script file to plot Figure 4.8b is given below and the plot is shown in Figure 4.18b.
% (b) Plot in Figure 4.8b
rho1 = -1:0.01:1;
rho2 = 2*rho1.*rho1-1;

subplot(1,2,2);
fill([rho1,-1],[rho2,1],[0.8,0.8,0.8]); hold on;
plot([-1,1],[0,0],’w:’,[0,0],[-1,1],’w:’);
set(gca,’xtick’,[-1:0.5:1],’ytick’,[-1:0.5:1]);
xlabel(’\rho(1)’,’fontsize’,label_fontsize);
ylabel(’\rho(2)’,’fontsize’,label_fontsize);
axis([-1,1,-1,1]); axis(’square’);
title(’(b) (\rho(1),\rho(2)) space’,’fontsize’,title_fontsize);
hold off;

4.19 Given an AR(2) process x(n) with d0 = 1, a1 = −1.6454, a2 = 0.9025, and w(n) ∼ WGN(0,1)

(a) Figure 4.19a shows a realization of x(n)

(b) Figure 4.19b shows the estimate of the ACS ρ̂ x (l) and theoretical values
(c) Compute the system parameters from the ACS ρ̂ x (l). This can be done using the Yule-Walker equations.

â = − R̂x−1r̂ x
Statistical and Adaptive Signal Processing - Solution Manual 53

Realization of x(n)
10

−5

−10
0 20 40 60 80 100
Figure 4.19a:

hat
rho (l) rho(l)
1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 5 10 15 20 0 5 10 15 20
Figure 4.19b:

Figure 4.19c shows the estimate of the system parameters.

(d) Figure 4.19d shows the estimated and true PSDs of the process
(e) Figure 4.19e shows the PACS

4.20 Find a minimum-phase model with autocorrelation ρ(0) = 1, ρ(±1) = 0.25, and ρ(l) = 0 for |l| ≥ 2. Using
Equation 4.3.13


 1 l=0
ρ(l) =
d1
1+d 2
l = ±1

 0 1 |l| ≥ 2

Clearly,
d1
= 0.25
1 + d12
and solving for d1 results in d1 = 0.2679. Inserting d1 into the general model equation generates a minimum-
phase model of

x(n) = w(n) + 0.2679w(n − 1)

4.21 Given the MA(2) model x(n) = w(n) − 0.1w(n − 1) + 0.2w(n − 2)

54 Statistical and Adaptive Signal Processing - Solution Manual

hat
a (l)
1

0.5

−0.5

−1

−1.5

−2
0 5 10 15 20
Figure 4.19c:

PSD of rhohat PSD of rho

10 10

8 8

6 6

4 4

2 2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Figure 4.19d:

(a) Is the process x(n) stationary?

The process x(n) is a linear combination of a stationary process w(n), therefore x(n) is stationary.
(b) Is the model minimum-phase?
Looking at the transfer function
H (z) = 1 − 0.1z −1 + 0.2z −2
This has d1 = −0.1 and d2 = 0.2. Using Equation 4.3.21, which are the minimum-phase conditions for
an AZ(2) model
d2 < |1|
d2 − d1 > −1
d2 + d1 > −1
Clearly, using the above equations, the model is stationary.
(c) Determine the autocorrelation and partial autocorrelation of the process.
The autocorrelation can be found directly using Equation 4.3.22


 1 l=0

 d1 (1+d 2)
1+d12 +d22
= −0.114 l = ±1
ρ(l) =


d2
2 = 0.191 l = ±2

 1+d1 +d2
2

0 |l| > 3
Statistical and Adaptive Signal Processing - Solution Manual 55

PACS of AP(2)
1

0.5

−0.5

−1
0 1 2 3 4 5 6
Figure 4.19e:

The partial autocorrelation sequence can be computed using the Yule-Walker equations and Cramer’s
rule. Where the PACS is
|Pm(m) |
am(m) = −
|Pm |
and Pm(m) and Pm are defined below as
 
ρ x (0) ρ x (1) ··· ρ x (m − 1) ρ x (1)
 ρ ∗ (1) ρ x (0) · ρ x (m − 2) ρ x (2) 
 x 
Pm(m) =  .. .. .. .. .. 
 . . . . . 
ρ ∗x (m) ρ ∗x (m − 1) · · · ρ x (1) ρ x (m + 1)
 
ρ x (0) ρ x (1) ··· ρ x (m)
 ρ ∗x (1) ρ x (0) · ρ x (m − 1) 
 
Pm =  .. .. .. .. 
 . . . . 
ρ ∗x (m) ρ ∗x (m − 1) · · · ρ x (0)
Using the above equation, the first five coefficients of the PACS are
am(m) = [0.1140, −0.1803, −0.0402, 0.0321, 0.0113]
and Figure 4.21c shows the first 20 coefficients.

4.22 Given two ARMA models

(a) A general expression for the autocorrelation can be found directly

i. x(n) = 0.6x(n − 1) + w(n) − 0.9w(n − 1)
r x (0) = E{x(n)x(n)} = E{[0.6x(n − 1) + w(n) − 0.9w(n − 1)][0.6x(n − 1) + w(n) − 0.9w(n − 1)]}
1 + (0.9)2
= = 2.828
1 − (0.6)2
r x (1) = E{x(n − 1)x(n)} = E{[x(n − 1)][0.6x(n − 1) + w(n) − 0.9w(n − 1)]}
= 0.6r x (0) = 1.697
r x (2) = E{x(n − 2)x(n)} = E{[x(n − 2)][0.6x(n − 1) + w(n) − 0.9w(n − 1)]}
= 0.6r x (1) = 0.62r x (0) = 1.018
Therefore, r x (l) = (0.6)l r x (0) = 0.6l (2.828)
56 Statistical and Adaptive Signal Processing - Solution Manual

PACS of MA(2)
0.15

0.1

0.05

−0.05

−0.1

−0.15

−0.2
0 5 10 15 20
Figure 4.21c:

ii. x(n) = 1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1):

r x (0) = E{x(n)x(n)} = E{[1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1)]

×[1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1)]}
= r x (0)(1.42 + 0.62 ) − r x (1)(2)(1.4)(0.6) + 1 + 0.82

r x (1) = E{[x(n − 1)][1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1)]}

1.4r x (0)
= = 0.875r x (0)
1 + 0.6
1+0.82
Hence r x (0) = 1−1.42 −0.62 +2(1.4)(0.6)(0.875)
= 10.93 and r x (1) = (0.875)(10.93) = 9.567. Finally

r x (2) = E{[x(n − 2)][1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1)]}

= 1.4r x (1) − 0.6r x (0) = 6.836

Therefore, r x (l) = 1.4r x (l − 1) − 0.6r x (l − 2).

(b) Compute the partial autocorrelation
The PACS can be computed using Yule-Walker equations and Cramer’s rule. The first three coefficient
(k1 , k2 , k3 ) of the PACS can be found using the following three equations
 
 r x (0) r x (1) r x (1) 
  
 r x (0) r x (1)   r x (1) r x (0) r x (2) 
  
 r x (1) r x (2)   r x (2) r x (1) r x (3) 
k1 = −ρ(1) = −r x (1)/r x (0), k2 = −   , k3 = −  
 r x (0) r x (1)   r x (0) r x (1) r x (2) 
  
 r x (1) r x (0)   r x (1) r x (0) r x (1) 

 r x (2) r x (1) r x (0) 

i. The first four coefficients of the ACS for x(n) = 0.6x(n − 1) + w(n) − 0.9w(n − 1) are

r x = [2.828, 1.697, 1.018, 0.611]

Substituting these values into the above equations for the PACS, the resulting PACS coefficients are

km = [−0.6, 0, 0]
Statistical and Adaptive Signal Processing - Solution Manual 57

ii. The first four coefficients of the ACS for x(n) = 1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1)
are

r x = [10.93, 9.567, 6.836, 3.830]

Substituting these values into the above equations for the PACS, the resulting PACS coefficients are

km = [−0.8753, 0.6017, 0]

(c) Generate a realization of x(n) and compute ρ̂(l)

i. Figure 4.22c.1 shows a realization of x(n) = 0.6x(n − 1) + w(n) − 0.9w(n − 1) and ρ̂(l)

Realization of x(n) hat

rho (l)
3 1

2 0.8

1 0.6

0 0.4

−1 0.2

−2 0

−3 −0.2

−4 −0.4
0 20 40 60 80 100 0 5 10 15 20
Figure 4.22c.1:

ii. Figure 4.22c.2 shows a realization of x(n) = 1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1) and
ρ̂(l)

Realization of x(n) rhohat(l)

3 1

1
0.5
0

−1
0
−2

−3

−4 −0.5
0 20 40 60 80 100 0 5 10 15 20
Figure 4.22c.2:

(d) Figure 4.22d shows the estimate of kˆm . The left and right plot correspond to x(n) = 0.6x(n − 1) +
w(n) − 0.9w(n − 1) and x(n) = 1.4x(n − 1) − 0.6x(n − 2) + w(n) − 0.8w(n − 1) respectively.

4.23 see Example 4.4.1, page 179

58 Statistical and Adaptive Signal Processing - Solution Manual

PACS PACS
0.4 0.6

0.4
0.3
0.2
0.2 0

0.1 −0.2

−0.4
0
−0.6

−0.1 −0.8
0 5 10 15 20 0 5 10 15 20
Figure 4.22d:

4.24 Given an AZ(Q) model

(a) Show that the impulse response of an AZ(Q) model can be recovered from its response 3
h(n) to a periodic
train with period L if L > Q
Assume the x(n) has the general form

x(n) = w(n) + d1 w(n − 1) + d2 w(n − 2) + · · · + dq w(n − q)

The impulse response to a periodic input is 3 h(n). Since the model is an AZ(Q) model, the impulse
response will be zero for each point beyond Q taps. If the period of the periodic input pulses is greater
then the order of the model, then the system response to each periodic impulse is


q

L
3
h(n) = dk h(n − k) + 0
k=0 k=q+1

and the AZ(Q) model can be completely recovered from 3

h(n).
(b) Show that the ACS of an AZ(Q) model can be recovered from the ACS or spectrum of 3
h(n) if L ≥ 2Q + 1
Start with the definition of the ACS
∞

rh (l) =
3 3
h(n)3
h ∗ (n − l)
n=−∞

Since 3 rh (l) is also periodic. The ACS is similar to convolution, with a signal being
h(n) is periodic, 3
convolved with itself. As long as the period of the periodic impulses is greater then the aliasing point
of the convolution, then the signal can be completely recovered. In order for there to be no overlap, the
period must be greater then twice the length of the filter plus one. This can be seen by substituting directly
into the above equation
∞
 
q

L 
q

L
rh (l) =
3 ( dk h(n − k) + 0)( dk h(n − l − k) + 0)
n=−∞ k=0 k=q+1 k=0 k=q+1

These two signal do not overlap as long as L ≥ 2Q + 1.

Statistical and Adaptive Signal Processing - Solution Manual 59

4.25 To prove (4.3.17), note that

d1
ρ(0) = 1, ρ(1) = ρ(−1) = , and ρ(l) = 0, |l| ≥ 2 (23)
1 + d12

and that
 
1 ρ(1) 0 ··· 0 0
 ρ(1) 1 ρ(1) · ·· 0 0 
 
 . 
 0 ρ(1) 1 .. 0 0 
 
Rm =  . . .. .. .. ..  (24)
 .. .. . . . . 
 
 .. 
 0 0 0 . 1 ρ(1) 
0 0 0 ··· ρ(1) 1

is an m × m matrix and that the PACS (or the reflection coefficients) are computed using (4.2.23), that is,
 (m)   
a1 ρ(1)
 a (m)   0 
 2   
Rm  .  = −  .  (25)
 ..   .. 
km 0

We will consider explicit evaluations for m = 1, 2, and 3.

• m = 1: Then from (24) and (25)

 
d1 1 − d12 (−d1 )1 1 − d12 (−d1 )1
[1] k1 = −ρ(1) ⇒ k1 = − =  =   (26)
1 + d12 1 − d12 1 + d12 1 + d12(1+1)

using (23).
• m = 2: Then from (24) and (25)
    
1 ρ(1) a1(2) ρ(1)
=−
ρ(1) 1 k2 0
or
 
 1 −ρ(1)   2
 
 ρ(1) 0  ρ (1)
2
d1
1+d12 d12
k2 =   = =   =
 1 ρ(1)  1 − ρ 2 (1) 2
1 + d12 + d14
 1 − 1+d d1
 ρ(1) 1 
2
1
  
1 − d12 (−d1 )2 1 − d12 (−d1 )2 1 − d12 (−d1 )2
=   = = (27)
1 − d12 1 + d12 + d14 1 − d16 1 − d12(2+1)

• m = 3: Then from (24) and (25)

   (3)   
1 ρ(1) 0 a1 ρ(1)
 ρ(1) 1 ρ(1)   a2(3)  = −  0 
0 ρ(1) 1 k3 0
60 Statistical and Adaptive Signal Processing - Solution Manual

or
 
 1 ρ(1) −ρ(1) 
 
 ρ(1) 1 0   3
 
 0 ρ(1) 0  ρ (1)
3 − d1
1+d1 2 −d13
k3 =   =− =   =  
 1 ρ(1) 0  1 − 2ρ 2 (1) 2
1 + d 4
1 + d 2
  1 − 2 1+d 2d1 1 1
 ρ(1) 1 ρ(1) 
 1
 0 ρ(1) 1 
  
1 − d12 (−d1 )3 1 − d12 (−d1 )3 1 − d12 (−d1 )3
=    = = (28)
1 − d12 1 + d14 1 + d12 1 − d18 1 − d12(3+1)

Thus from (26), (27) and (28) we conclude that


1 − d12 (−d1 )m
km =
1 − d12(m+1)

For a general proof, consider (24). Then

 
ρ(1) · · ·
1 0 0
 .. 
 ρ(1) 1 . 0 0  
 
 .. 
det Rm = (1) det  ... ..
.
..
.
..
. . 
 
 .. 
 0 0 . 1 ρ(1) 
0 0 · · · ρ(1) 1
 
..
 1 . 0 0 
 . . . .. . .. .. 
 .  
−ρ(1)ρ(1) det  . 
 .. 
 0 . 1 ρ(1) 
0 · · · ρ(1) 1

where the first determinant is of (m − 1) × (m − 1) matrix Rm−1 and the second determinant is of (m − 2) ×
(m − 2) matrix Rm−2 . Hence

det Rm = det Rm−1 − ρ 2 (1) det Rm−2 (29)

Also from (25) [and from (26), (27), and (28)], we have
−ρ m (1)
km = (30)
det Rm
Dividing (29) by −ρ m (1) and using (30), we have
1 1 1 1
= − , m≥3 (31)
km −ρ (1) km−1 km−2
which is a recursive relation with initial conditions
d1 ρ 2 (1) d12
k1 = −ρ(1) = − and k 2 = = (32)
1 + d12 1 − ρ 2 (1) 1 + d12 + d14
Statistical and Adaptive Signal Processing - Solution Manual 61

Since the AZ(1) model is minimum phase, |d1 | < 1 which implies that the roots of the difference equation (31)
in k1m are real. Using p(n) = k1m , n ≥ 0 with p(−1) = k12 , p(−2) = k11 , and using the unilateral z-transform
approach along with fair amount of algebra, we obtain

1 − d12 (−d1 )m
km = , m ≥ 1.
1 − d12(m+1)

The validity for a particular d1 for lower values of m is established in (26) through (28). Follow that approach
for H (z) = 1 − 0.8 z −1 .

4.26 Prove Equation (4.3.24) that describe minimum-phase region of the AZ(2) model. Equation (4.3.24) is

ρ(2) + ρ(1) > −0.5

ρ(2) − ρ(1) > −0.5
ρ 2 (1) = 4ρ(2)[1 − 2ρ(2)]

Using Equation (4.3.22) and direct substitution into ρ(2) + ρ(1) > −0.5 yields

d2 d1 (1 + d2 )
+ > −0.5
1 + d1 + d2
2 2
1 + d12 + d22
(d1 + d2 )2 + 2(d1 + d2 ) + 1 > 0

Clearly, in order for the above equation to hold, d1 + d2 > −1, which is the minimum-phase region described
in Equation (4.3.21).
Using Equation (4.3.22) and direct substitution into ρ(2) − ρ(1) > −0.5 yields

d2 d1 (1 + d2 )
− > −0.5
1 + d1 + d2
2 2
1 + d12 + d22
(d1 − d2 )2 + 2(d1 − d2 ) + 1 > 0

Clearly, in order for the above equation to hold, d1 − d2 > −1, which is the minimum-phase region described
in Equation (4.3.21).

4.27 Given an AZ(2) model with d0 = 2 and zeros z 1,2 = 0.95e± j π /3

(a) The upper left plot of Figure 4.27ab shows a realization of x(n) = −0.95x(n −1)+0.9025x(n −2)+w(n)
(b) Figure 4.27ab shows the ACS, PACS, and the spectrum of the model
(c) Repeating parts a and b above with an AP(2) model with poles p1,2 = 0.95e± j π /3 . Figure 4.27c has the
plots for the realization, ACS, PACS, and spectrum to the AP(2) model.
(d) The duality of the AZ(2) and AP(2) is shown in the above figures. The finite ACS for the AZ(2) model
is similar to the finite PACS of the AP(2) model. The periodic nature of the PACS for the AZ(2) also
appears in the ACS of the AP(2) model. Lastly, frequency of notch filter for the AZ(2) corresponds to the
frequency peaks of the AP(2) model.

4.28 Given the autocorrelation sequence relations in (4.4.29) and (4.4.30), that is,

(d1 − a1 ) (1 − a1 d1 )
ρ(1) = (33)
1 + d12 − 2a1 d1
ρ(2) = −a1 ρ(1) (34)
62 Statistical and Adaptive Signal Processing - Solution Manual

Realization of x(n) rho(l) of Az(2)

10 1

5 0.5

0 0

−5 −0.5

−10 −1
0 20 40 60 80 100 0 5 10 15 20
PSD of AZ(2)
PACS of AZ(2) 35
1
30
0.8
25
0.6

0.4
20

0.2 15

0 10

−0.2 5

−0.4 0
0 5 10 15 20 0 0.2 0.4 0.6 0.8 1
Figure 4.27ab: AZ(2) System

and the inequalities

−1 < d1 < 1, −1 < a1 < 1 (35)

we want to determine conditions on ρ(1) and ρ(2).

First note from (34) that
ρ(2)
a1 = − (36)
ρ(1)
and using (35)

|a1 | < 1 ⇒ |ρ(2)| < |ρ(1)| (37)

Now for the remaining conditions the general approach is the following: first substitute a1 from (36) in (33)
and solve for d1 in terms of ρ(1) and ρ(2); then use (35) to determine conditions on ρ(1) and ρ(2). Consider
(33). After simplification we obtain

ρ(1) + ρ(1)d12 − 2ρ(1)a1 d1 = d1 − a1 d12 − a1 + a12 d1

Substituting a1 from (36) and simplifying, we obtain

 
ρ 3 (1) + ρ 3 (1)d12 + 2ρ 2 (1)ρ(2)d1 = ρ 2 (1)d1 + ρ(1)ρ(2)d12 + ρ(1)ρ(2) + ρ 2 (2)d1
Statistical and Adaptive Signal Processing - Solution Manual 63

Realization of x(n) rho(l) of AP(2)

10 1

5 0.5

0 0

−5 −0.5

−10 −1
0 20 40 60 80 100 0 5 10 15 20
PSD of AP(2)
PACS of AP(2) 10
1

0.5
6

4
0

−0.5 0
0 1 2 3 4 5 6 0 0.2 0.4 0.6 0.8 1
Figure 4.27c: AP(2) System

or
     
d12 ρ 3 (1) − ρ(1)ρ(2) + d1 2ρ 2 (1)ρ(2) − ρ 2 (1) − ρ 2 (2) + ρ 3 (1) − ρ(1)ρ(2) = 0 (38)

Solving (38) for d1 , we obtain

: 
1
d1 =  3 2b c − b − c + 4b c + 4b c − 4b c + b − 2b c + c − 4b
2 2 2 4 2 4 2 3 4 2 2 4 6 (39)
2 −b + bc
and

: 
1
d1 =  2b c − b − c −
2 2 2
4b4 c2 + 4b4 c − 4b2 c3 + b4 − 2b2 c2 + c4 − 4b6 (40)
2 −b3 + bc
where we have used b = ρ(1) and c = ρ(2) for compactness. Now since −1 < d < 1 we have four possible
equations: d1 = 1 and d1 = −1 in (39) and in (40). Thus

: 
1
1=  3 2b c − b − c + 4b c + 4b c − 4b c + b − 2b c + c − 4b
2 2 2 4 2 4 2 3 4 2 2 4 6
2 −b + bc
yields c = −b or ρ(2) = −ρ(1) [which is part of (37)] and

c = 2b2 − b or ρ(2) = ρ(1) [2ρ(1) − 1] (41)

64 Statistical and Adaptive Signal Processing - Solution Manual

Similarly

: 
1
−1 =  2b c − b − c + 4b c + 4b c − 4b c + b − 2b c + c − 4b
2 2 2 4 2 4 2 3 4 2 2 4 6
2 −b3 + bc

yields c = b or ρ(2) = −ρ(1) [which is part of (37)] and

c = 2b2 + b or ρ(2) = ρ(1) [2ρ(1) + 1] (42)

Likewise from (40), we have

ρ(2) = −ρ(1) and ρ(2) = ρ(1) [2ρ(1) − 1]

and

ρ(2) = ρ(1) and ρ(2) = ρ(1) [2ρ(1) + 1]

which are the same conditions as those in (41) and (42). Since |ρ(2)| < 1, we conclude that in (41) ρ(1) must
be greater than 0 and in (42) ρ(1) must be less than 0. Thus we conclude

ρ(2) > ρ(1) [2ρ(1) − 1] , ρ(1) > 0

ρ(2) > ρ(1) [2ρ(1) + 1] , ρ(1) < 0

which together with (37) is (4.4.31).

Matlab Script for drawing Figure 4.12b is shown below and the resulting figure in shown in Figure 4.28.

% Generate figure for rho1 > 0

rho1p = 0:0.01:1;
rho2p = rho1p.*(2*rho1p-1);
rho1p = [rho1p,0]; rho2p = [rho2p,0];

% Generate figure for rho1 < 0

rho1n = -1:0.01:0;
rho2n = rho1n.*(2*rho1n+1);
rho1n = [rho1n,-1]; rho2n = [rho2n,1];

% Plot
fill(rho1p,rho2p,’g’); hold on;
fill(rho1n,rho2n,’r’);
plot([-1,1],[0,0],’w:’,[0,0],[-1,1],’w:’);
set(gca,’xtick’,[-1:0.5:1],’ytick’,[-1:0.5:1]);
xlabel(’\rho(1)’,’fontsize’,label_fontsize);
ylabel(’\rho(2)’,’fontsize’,label_fontsize);
text(0.2,-0.2,’\rho(2)=\rho(1)[2\rho(1)-1]’);
text(-0.9,-0.2,’\rho(2)=\rho(1)[2\rho(1)+1]’);
text(0.2,0.6,’\rho(2)=\rho(1)’);
text(-0.55,0.6,’\rho(2)=-\rho(1)’);
axis([-1,1,-1,1]); axis(’square’);
Statistical and Adaptive Signal Processing - Solution Manual 65

ρ(2)=−ρ(1) ρ(2)=ρ(1)
0.5

ρ(2)
0

ρ(2)=ρ(1)[2ρ(1)+1] ρ(2)=ρ(1)[2ρ(1)−1]

−0.5

−1
−1 −0.5 0 0.5 1
ρ(1)

Figure 4.28: Minimum-phase and positive definite region for the PZ(1,1) model in the ρ(1) – ρ(2) plane

4.29 The spectral flatness can be computed using Equation (4.1.20)

σ 2w 1
SFMx = = π
σ 2x 1
2π
jω 2
−π |H (e )| dω

Using Parseval’s Theorem, this can be expressed as

1
SFMx = !∞
n=−∞ |h(n)|2

or in other words, one over the sum of the squared-magnitude of the impulse response.

(a) x(n) = a1 x(n − 1) + a2 x(n − 2) + w(n)

Finding the system function H (z) = X (z)/W (z) is
1
H (z) =
1 − a1 z −1 − a2 z −2
This is an AP(2) system. The general equation for the impulse response to an AP(2) system is
1
h(n) = ( pn+1 − p2n+1 )u(n)
p1 − p2 1
Therefore, the SFMx is
1
SFMx = !∞
n=0 | p1 − p2 ( p1 − p2n+1 )|2
1 n+1

where p1 and p2 are the two poles of the system.

(b) x(n) = w(n) + b1 w(n − 1) + b2 w(n − 2)
Finding the system function H (z) = X (z)/W (z) is

H (z) = 1 + b1 z −1 + b2 z −2
66 Statistical and Adaptive Signal Processing - Solution Manual

This is an AZ(2) system. The general equation for the impulse response to an AZ(2) system is

h(n) = δ(n) + b1 δ(n − 1) + b2 δ(n − 2)

Therefore, the SFMx is

1
SFMx =
1 + |b1 |2 + |b2 |2

4.30 The WSS process x(n) is zero-mean with PSD Rx (e j ω ). The matrix Rx is an M × M correlation matrix with
eigenvalues {λi }i=1
M
. Szegö’s theorem states that
 π
1   
M
1
lim g(λi ) = g Rx (e j ω ) dω
M→∞ M 2π −π
i=1

Let g(·) = ln(·). Then

 π
1   
M
1
lim ln λi = ln Rx (e j ω ) dω (43)
M→∞ M 2π −π
i=1

;
M
Using the property of determinant of Rx , namely det (Rx ) = λi , we can write
i=1


M
ln [det (Rx )] = ln (λi ) (44)
i=1

Substituting (44) in (43), we obtain


1 1 π  
lim ln [det (Rx )] = ln Rx (e j ω ) dω (45)
M→∞ M 2π −π

1
Using 1
M
ln [det (Rx )] = ln [det (Rx )] M , we can rewrite (45) as
&  π '
1 1  jω

lim [det (Rx )] = exp
M ln Rx (e ) dω
M→∞ 2π −π

which is the desired result.

4.31 Given the two linear random processes with system functions

1 − 0.81z −1 − 0.4z −2
H1 (z) =
(1 − z −1 )2
and
1 − 0.5z −1
H2 (z) =
(1 − z −1 )

(a) Find a difference equation for the two system functions:

Since H (z) = X (z)/W (z), the system function can be found directly
Statistical and Adaptive Signal Processing - Solution Manual 67

Realization of x(n) Realization of x(n)

250 40

200
20
150

100 0

50
−20
0

−50 −40
0 20 40 60 80 100 0 20 40 60 80 100
Figure 4.31b:

1−0.81z −1 −0.4z −2
i. for H1 (z) = (1−z −1 )2
The system function can be separated into

X (z)(1 − 2z −1 + z −2 ) = W (z)(1 − 0.81z −1 − 0.4z −2 )

Taking the inverse z-transform gives a difference equation of
x(n) = 2x(n − 1) − x(n − 2) + w(n) − 0.81w(n − 1) − 0.4w(n − 2)

1−0.5z −1
ii. for H2 (z) = (1−z −1 )
The system function can be separated into

X (z)(1 − z −1 ) = W (z)(1 − 0.5z −1 )

Taking the inverse z-transform gives a difference equation of
x(n) = x(n − 1) + w(n) − 0.5w(n − 1)

Second Difference of x(n) First Difference of x(n)

4 3

2
2
1

0 0

−1
−2
−2

−4 −3
0 20 40 60 80 100 0 20 40 60 80 100
Figure 4.31c:

(b) Figure 4.31b shows realization of x(n) = 2x(n − 1) − x(n − 2) + w(n) − 0.81w(n − 1) − 0.4w(n − 2)
and x(n) = x(n − 1) + w(n) − 0.5w(n − 1) respectively. Both plots display aspects of non-stationarity.
The left plot shows a random walk type system, with the mean changing with more samples. The second
plot remains zero mean, but its variance is not constant with the number of samples.
68 Statistical and Adaptive Signal Processing - Solution Manual

(c) The first and second difference for the above two systems are shown in Figure 4.31c. Clearly, these two
systems appear stationary. The second difference of x(n) = 2x(n − 1) − x(n − 2) + w(n) − 0.81w(n −
1) − 0.4w(n − 2) is

y(n) = x(n) − 2x(n − 1) + x(n − 2)

= (2x(n − 1) − x(n − 2) + w(n) − 0.81w(n − 1) − 0.4w(n − 2)) − 2x(n − 1) + x(n − 2)
= w(n) − 0.81w(n − 1) − 0.4w(n − 2)

and the first difference of x(n) = x(n − 1) + w(n) − 0.5w(n − 1) is

y(n) = x(n) − x(n − 1) = (x(n − 1) + w(n) − 0.5w(n − 1)) − x(n − 1)

= w(n) − 0.5w(n − 1)

4.32 Generate and plot 100 samples for each of the following linear processes and then estimate and examine the
values of the ACS and PACS

(a) H (z) = 1
(1−z −1 )(1−0.9z −1 )
: The plots are shown in Figure 4.32a.

rhohat(l)
Realization of x(n) 1
0

−20 0.8

−40
0.6

−60
0.4
−80
0.2
−100

−120 0
0 20 40 60 80 100 0 20 40 60 80
PACS
0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 5 10 15 20
Figure 4.32a:

1−0.5z −1
(b) H (z) = (1−z −1 )(1−0.9z −1 )
: The plots are shown in Figure 4.32b.

4.33 Given y(n) = d0 + d1 n + d2 n 2 + x(n), where x(n) is a stationary process with know autocorrelation r x (l)
Statistical and Adaptive Signal Processing - Solution Manual 69

hat
Realization of x(n) rho (l)
150 1

0.8
100 0.6

0.4
50
0.2

0
0
−0.2

−50 −0.4
0 20 40 60 80 100 0 20 40 60 80
PACS
0.4

0.2

−0.2

−0.4

−0.6

−0.8

−1
0 5 10 15 20
Figure 4.32b:

(a) Show that the process y (2) (n) obtained by passing y(n) through the filter H (z) = (1 − z −1)2 is stationary.
Since H (z) = Y (2)(z)/Y (z), the output of the system is

Y (2)(z) = Y (z)(1 − 2z −1 + z −2 )

Taking the inverse z-transform

y (2) (n) = y(n) − 2y(n − 1) + y(n − 2)

Substituting y(n) = d0 + d1 n + d2 n 2 + x(n) back into the above equation and simplifying

y (2) (n) = 2d2 + x(n) − 2x(n − 1) + x(n − 2)

Since the output is a linear combination of a stationary system, y (2) (n) is stationary.
(b) The autocorrelation of y (2) (n) can be computed directly

r y(2) (l) = E{y (2)(n + l)y (2) (n)}

= E{[2d2 + x(n + l) − 2x(n + l − 1) + x(n + l − 2)][2d2 + x(n) − 2x(n − 1) + x(n − 2)]}

Therefore, r y(2) (l) = 4d2(2) + 5r x (l) − 8r x (l) + 2r x (l − 2).

70 Statistical and Adaptive Signal Processing - Solution Manual

4.34 Proof of (4.4.7).

Consider the AP(P) model
d0 d0
H (z) = =
A(z) 1 + a1 z + a2 z −2 + · · · + a P z −P
−1

Then the complex cepstrum is given by

 
C(z) = log H (z) = log d0 − log 1 + a1 z −1 + a2 z −2 + · · · + a P z −P

and the cepstrum is given by

c(n) = Z −1 [C(z)]

• P = 1: Then A(z) = 1 + a1 z −1 and




 −1
−1 1 2 −2 1 3 −3 1 4 −4 
log 1 + a1 z = a1 z + − a1 z + a1 z + − a1 z + O z −5
2 3 4

Thus

c(n) = 0, n < 0; c(0) = log d0

c(1) = −a1 ,
1 2 1 1
c(2) = a1 = − (2 − 1) a1 (−a1 ) = − (2 − 1) a1 c(1)
2 2
 2
1 3 1 1 2 1
c(3) = − a1 = − (3 − 1)a1 a = − (3 − 1)a1 c(2)
3 3 2 1 3
.. .
. = ..
1
P=1
1
c(n) = − (n − 1)a1 c(n − 1) = − (n − k)ak c(n − k), n > P
n n k=1

• P = 2: Then A(z) = 1 + a1 z −1 + a2 z −2 and



 −1 −2
−1 1 2 −2 1 1
log 1 + a1 z + a2 z = a1 z + a2 − a1 z + − a1 a2 + −2a2 + a1 a1 z −3 +
2
2 3 3


1 1 
−2a2 + a1 a2 +
2
3a1 a2 − a1 a1 z −4 + O z −5
3
4 4

Thus

c(n) = 0, n < 0; c(0) = log d0

1
c(1) = −a1 = −a1 − (1 − 1) a0 c(0)
1
1 2 1
c(2) = −a2 + a1 = −a2 − (2 − 1) a1 c(2 − 1)
2 2
1
n−1
= −an − (n − k) ak c(n − k), n = 2
n k=1
Statistical and Adaptive Signal Processing - Solution Manual 71



 
1 1 1 1
c(3) = − − a1 a2 + −2a2 + a12 a1 = − 2a1 −a2 + a12 − a1 a2
3 3 3 2


 
1 1
= − (3 − 1) a1 −a2 + a12 + a2 (−a1 )
3 2
1
= − ((3 − 1) a1 c(3 − 1) + (3 − 2) a2 c(3 − 2))
3
1
P=2
= − (n − k) ak c(n − k), n = 3
3 k=1
.. .
. = ..
1
P=2
c(n) = − (n − k) ak c(n − k), n > P
n k=1

Thus from the above verifications, we conclude that


 !
1 n−1

 −an − (n − k) ak c(n − k), 1≤n≤ P
n k=1
c(n) =

 !
1 P=2
 − (n − k) ak c(n − k) n>P
n k=1
For a general case, consider the complex cepstrum
C(z) = log H (z)
Differentiating with respect to z, we obtain
!
! H (z) ! !
C (z) = ⇒ H (z) = C (z) H (z)
H (z)
or
! !
z H (z) = z C (z) H (z)
Taking inverse z-transform, we get
∞

n h(n) = [n c(n)] ∗ h(n) = h(k)(n − k)c(n − k)
k=−∞

or after dividing by n, we obtain

∞
1 
h(n) = (n − k) h(k)c(n − k) (46)
n k=−∞

Using the property of complex cepstrum of a minimum-phase system which states that if H (z) is minimum-
phase then the complex cepstrum c(n) is causal, we have from (46)

1
n−1
h(n) = (n − k) h(k)c(n − k), n > 0
n k=0

since h(n) is also causal and (n − k) = 0 for k = n. Thus

1
n−1
h(n) = h(0)c(n) + (n − k) h(k)c(n − k), n > 0
n k=1
72 Statistical and Adaptive Signal Processing - Solution Manual

Solving for c(n), we obtain

h(n) 1 
n−1
h(k)
c(n) = − (n − k) c(n − k), n > 0 (47)
h(0) n k=1 h(0)

Now for H (z) = A(z) = 1 + a1 z −1 + a2 z −2 + · · · + a P z −P , we have

h(n) = 0, n < 0; h(0) = 1; h(n) = an , 1 ≤ n ≤ P; h(n) = 0, n > P (48)

Thus in (47) we have for 1 ≤ n ≤ P

1
n−1
c(n) = an − (n − k) ak c(n − k), 1 ≤ n ≤ P (49)
n k=1

Note also from (48) that for n > P, h(n) = 0 and the summation ends at P. Thus for n > P

1
P
c(n) = − (n − k) ak c(n − k), n > P (50)
n k=1

Finally, note that for H (z) = 1/ A(z), c(n) → −c(n). Thus in (49) and (50), we have

 !
1 n−1

 −an − (n − k) ak c(n − k), 1 ≤ n ≤ P
n k=1
c(n) =

 1! P
 − (n − k) ak c(n − k), n>P
n k=1

which completes the proof.

!Q
4.35 Consider a minimum-phase AZ (Q) model D(z) = k=0 dk z −k with complex cepstrum c(k). Another AZ
model is created with coefficients d̃k = α k dk and complex cepstrum c̃(k).

(a) Let 0 < α < 1. Since D(z) is minimum phase, we have all zeros of D(z) inside the unit circle. Now


Q 
Q 
Q
−k −k
D̃(z) = d̃k z = α dk z
k
= dk (z/α)−k = D (z/α)
k=0 k=0 k=0

Thus if the zeros of D(z) are at {z i }i=1

Q
then the zeros of D̃(z) are at {α z i }i=1
Q
. Then clearly D̃(z) is also
minimum phase since 0 < α < 1 and furthermore, no zeros of D(z) cross the unit circle to become those
of D̃(z). This implies that the region of convergence of D̃(z) also includes the unit circle. Hence
  Q 
$ % *
−1 −1 −1
c̃(k) = Z log D̃(z) = Z log d0 1 − α zn z
n=1
& &
log |d0 | , k=0 log |d0 | , k=0
= !Q = αk !Q k
− n=1 α z n , k > 0
k k
− n=1 z n , k > 0
= α k c(k)

Note that both c(k) and c̃(k) are causal sequences due to minimum-phase models.
Statistical and Adaptive Signal Processing - Solution Manual 73

(b) To obtain nonminimum-phase model, we need to create zeros that are inside as well as outside the unit
circle. If z i,min = min |z i | and z i,max = max |z i |, then we want

z i,min α < 1 and z i,max α > 1

This gives
1 1
<α<
z i,max z i,min

(c) To obtain maximum-phase model, we need to create zeros that are outside the unit circle. If z i,min =
min |z i |, then we want

z i,min α > 1

This gives
1
α>
z i,min

4.36 Prove Equation 4.6.27, which determines the cepstral distance in the frequency and time domains.
The cepstral distance is defined, in (4.6.27), as
 π
1
CD ≡ | log R1 (e j ω ) − log R2 (e j ω )|2 dω
2π −π

Using the power series expansion, the cepstral distance can be shown as
 π ∞
 ∞
  π ∞

1 − j nω − j nω 2 1
CD = | c1 (n)e − c2 (n)e | dω = | (c1 (n) − c2 (n))e− j nω |2 dω
2π −π n=−∞ n=−∞
2π −π n=−∞
 π ∞
 ∞

1
= (c1 (n) − c2 (n))(c1∗ (m) − c2∗ (m))e− j (n−m) ω)dω
2π −π n=−∞ m=−∞
∞
 ∞  π
1
= (c1 (n) − c2 (n))(c1∗ (m) − c2∗ (m)) e− j (n−m) ω)dω
n=−∞ m=−∞
2π −π

∞ ∞
 ∞

= (c1 (n) − c2 (n))(c1∗ (m) − c2∗ (m))δ(n − m) = |c1 (n) − c2 (n)|2
n=−∞ m=−∞ n=−∞
!∞
Therefore, CD = n=−∞ (c1 (n) − c2 (n))2 , which proves (4.6.27).
 Chapter 5
Nonparametric Power Spectrum Estimation

5.1 Let xc (t), −∞ < t < ∞, be a continuous-time signal with Fourier transform X c (F), −∞ < F < ∞, and let
x(n) be obtained by sampling xc (t) every T per sampling interval with its DTFT X (e j ω ).

(a) Show that the DTFT X (e j ω ) is given by

∞
 1
X (e j ω ) = Fs X c ( f Fs − l Fs ) ω = 2π f Fs =
l=−∞
T

Start with the definition of the DTFT

∞

jω
X (e ) = x(n)e− j ωn (1)
n=−∞
∞
Substitute x(n) = −∞ xc (t)δ(t − nT ) dt into (1) results in
∞ 
 ∞
jω
X (e ) = xc (t)δ(t − nT )e− j ωn dt (2)
n=−∞ −∞

Since the delta train is zero everywhere except the point where n = t
T
, (2) can be written as
 ∞ ∞

X (e j ω ) = [xc (t)][ δ(t − nT )]e− j 2π f Fs t dt
−∞ n=−∞

This is the Fourier Transform between the product of two signals. The Fourier Transform of the impulse
train is
∞
 ∞

F
δ(t − nT ) ←→ 2π Fs δ( f Fs − l Fs )
n=−∞ l=−∞

Therefore, the DTFT X (e j ω ) is

∞

jω
X (e ) = X c ( f Fs ) ∗ Fs δ( f Fs − l Fs )
l=−∞

Bringing X c ( f Fs ) inside the summation and using the fact that convolution with a delta result in a shift
in the signal, the above equation results in
∞

X (e j ω ) = Fs X c ( f Fs − l Fs ) (3)
l=−∞

(b) Let
∞
 k Fs
X̃ p (k) = X (e j 2πk/N ) = Fs X c( − l Fs ) (4)
l=−∞
N

74
Statistical and Adaptive Signal Processing - Solution Manual 75

Start with the definition of the IDFT( X̃ p (k)) which is

1 
N−1
2π kn
xp (n) = X̃ p (k)e j N (5)
N k=0

Substituting the equation for X̃ p (k) from (4) into (5) is

N−1 ∞
Fs   k Fs 2π kn
xp (n) = X c( − l Fs )e j N
N k=0 l=−∞ N
∞
Substitute X c ( kNFs − l Fs ) = −∞ X c ( f Fs − l Fs )δ( f Fs − k Fs
N
)d f into the above equation and bringing
the sum on k inside the integral results in
∞  ∞
Fs  ∞  
N−1
k Fs j 2πn f
[X c ( f Fs − l Fs )][ δ( f Fs − )]e df = xc (nT − m N T ) (6)
N l=−∞ −∞ k=0
N m=−∞

since we have a multiplication in the frequency domain with a delta train.

5.2 MATLAB has two built in functions bartlett and triang. They each generate trinagular windows with
slightly different coefficients.

(a) Figure 5.2a shows plots of both the Bartlett Window and the Triangular Window for N = 11, 31, 51.
Bartlett Window Triangular Window
1 1
N=11 N=11
0.5 0.5

0 0
−5 0 5 −5 0 5
1 1
N=31 N=31
0.5 0.5

0 0
−15 −10 −5 0 5 10 15 −15 −10 −5 0 5 10 15
1 1
N=51 N=51
0.5 0.5

0 0
−30 −20 −10 0 10 20 30 −30 −20 −10 0 10 20 30

Figure 5.2a: Bartlett and Triangular Windows

(b) Figure 5.2b show the magnitude plots for the windows given in Figure 5.2a. The mainlobe of the Triangular
windows are smaller then the mainlobes of the corresponding Bartlett Windows. For each N value, the
higher order lobes are spaced evenly apart for both the Bartlett and Triangular Windows.
(c) A Bartlett Window with N = 100 has approximately the same mainlobe width as a rectangular window
with N = 51. This can be seen in Figure 5.2c. This stands to reason, since a Bartlett Window is triangular
in shape, and two rectangular windows convolved together results in a triangular window. Convolution in
the time domain is multiplication in the frequency domain, which means that the spectrum of a triangular
window is equivalent to the squared spectrum of an appropriate sized rectangular window. Lastly, when
two signals of length N are convolved, the resulting output has a length of 2N −1. Therefore, a rectangular
window of length N and a triangular window of length 2N − 1 will have the smae size mainlobe width.

5.3 This problem deals with the maximum sidelobe heights of the rectangular, Hanning, Hamming, Kaiser and
Dolph-Chebyshev windows. The maximum sidelobe height is an indicator of how much spectral leakage a
certain window will contribute.
76 Statistical and Adaptive Signal Processing - Solution Manual

10
Spectrum of Bartlett Window 10
Spectrum of Triangular Window
10 10
N=11 N=11
0 0
10 10

−10 −10
10 10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
10 5
10 10
N=31 N=31
0 0
10 10

−10 −5
10 10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
5 10
10 10
N=51 N=51
0 0
10 10

−5 −10
10 10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.2b: Spectrum of Bartlett and Triangular Windows

2
Rectangular vs. Bartlett
10
Rectangular Spectrum (N=51)
1
10

0
10

−1
10
0 0.02 0.04 0.06 0.08 0.1

2
10
Bartlett Spectrum (N=100)

0
10

−2
10
0 0.02 0.04 0.06 0.08 0.1

Figure 5.2c: Mainlobe Width of Bartlett and Recantangular Windows

(a) Figure 5.3a shows magnitude spectrum plots of the rectangular, Hamming and Hanning windows for
N = 11, 31, 51. Clearly the sidelobe height is independent of the window size.
(b) The maximum sidelobe height of the Kaiser window is controlled by β the shape parameter. Figure
5.3b shows three plots for three different β values. Each plot has the magnitude spectrum of the Kaiser
window for N = 11, 31, 51 shown. It is seen that the window length N does not effect the height of the
sidelobe, however, the shape parameter β does.
(c) A Kaiser window can be computed that has the same sidelobe height as the Hamming. Since the sidelobe
height of the Hamming window is fixed, and from Figure 5.3a the maximum sidelobe height for a
Hamming window is approximately −42.5d B. The Kaiser window of β = 5.8 and length N = 51 has a
maximum sidelobe height of approximately −42.5d B. This is seen in Figure 5.3c.
(d) The Dolph-Chebyshev has the same height A in decibels for all of the sidelobes. Figure 5.3d shows the
−3 dB point for three Dolph-Chebyshev windows for sidelobe height A = 40, 50, 60 dB.

Kaiser vs. Hamming Window Dolph−Chebyshev Window 3dB Point(N=31)

0
Kaiser (β =5.8, N=51) A=40
Hamming (N=51) A=50
A=60

−42.5
−3

0 0.1 0.2 0.3 0.4 0.5 0.0199 0.022 0.0238

Figure 5.3c: Kaiser and Hamming Windows Figure 5.3d: Dolph-Chebyshev Windows
Statistical and Adaptive Signal Processing - Solution Manual 77

Rectangular Windows Hamming Window

0 0
−13 N=11 N=11
−43

0 0.05 0.1 0.15 0.2 0.25 0 0.1 0.2 0.3 0.4 0.5
0 0
−13 N=31 N=31
−43

0 0.05 0.1 0.15 0.2 0.25 0 0.1 0.2 0.3 0.4 0.5
0 0
−13 N=51 N=51
−43

0 0.05 0.1 0.15 0.2 0.25 0 0.1 0.2 0.3 0.4 0.5
Hanning Window
0
−32 N=11

0 0.1 0.2 0.3 0.4 0.5

0
−32 N=31

0 0.1 0.2 0.3 0.4 0.5

0
−32 N=51

0 0.1 0.2 0.3 0.4 0.5

Figure 5.3a: Rectangular, Hamming, and Hanning Windows

Kaiser Window (β = 0)
Kaiser Window (β = 5)
0 N=11 0
N=31 N=11
N=51 N=31
N=51

−13

−37.5

0 0.05 0.1 0.15 0.2 0.25 0 0.05 0.1 0.15 0.2 0.25
Kaiser Window (β = 10)
0
N=11
N=31
N=51

−74

0 0.1 0.2 0.3 0.4 0.5

Figure 5.3b:Kaiser Windows

78 Statistical and Adaptive Signal Processing - Solution Manual

5.4 For x(n) = y(n)w(n) where y(n) is

y(n) = cos ω1 n + cos(ω2 n + φ)

and w(n) is either a rectangular, Hamming, or Blackman window, the goal is to determine the smallest window
length that will allow the two frequencies to be separable in the |X (e j ω )|2 plots.

(a) Figure 5.4a shows the plot of y(n) as well as its true spectrum on the left hand side. The right hand side
show six different rectangular windowed spectra. The top row in Figure 5.4a corresponds to a φ = 0
and the bottom row corresponds to φ = 0.8π . It is seen that a minimum rectangular window of length
N = 27 is required to differentiate between the two frequencies for φ = 0 and N = 55 is required to
separate the frequencies for φ = 0.8π .
Rectangular Window
y(n)=cos(0.25π n) + cos(0.3π n)
2 N=119 N=51
0.2 0.2
1
0.1 0.1
0
0 0
0 0.2 0.4 0 0.2 0.4
−1
N=31 N=27
−2 0.2 0.2
0 20 40 60 80 100 120
Spectrum of y(n) 0.1 0.1
4000
0 0
0 0.2 0.4 0 0.2 0.4
3000
N=26 N=25
2000 0.2 0.2
1000 0.1 0.1

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0 0.2 0.4
Rectangular Window
y(n)=cos(0.25π n) + cos(0.3π n + 0.8π n) 0.5 0.5
2 N=119 N=71

0
0 0
0 0.2 0.4 0 0.2 0.4
−1 0.5 0.5
N=59 N=57
−2
0 20 40 60 80 100 120
Spectrum of y(n)
4000
0 0
0 0.2 0.4 0 0.2 0.4
3000
0.5 0.5
N=55 N=53
2000

1000

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0 0.2 0.4

Figure 5.4a: Rectangular Window

(b) Part (a) above is repeated for the φ = 0.8π signal, and Figure 5.4bc shows the window spectra for the
Hamming and Blackman windows. For Figure 5.4bc it is seen that the Hamming window requires an
approxmate minimum length N = 65 and the Blackman window length N = 69 is required.

5.5 The goal of this problem is to prove that the autocorrelation matrix R̂x is a nonnegative definite matrix, that is,

zH R̂x z ≥ 0 ∀ z = 0

(a) Show that R̂x can be decomposed into the product XH X, where X is called a data matrix.
Statistical and Adaptive Signal Processing - Solution Manual 79

Hamming Window Blackman Window

500 500 500 500
N=119 N=71 N=119 N=75

0 0 0 0
0 0.2 0.4 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4
500 500 500 500
N=69 N=67 N=73 N=71

0 0 0 0
0 0.2 0.4 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4
500 500 500 500
N=65 N=63 N=69 N=67

0 0 0 0
0 0.2 0.4 0 0.2 0.4 0 0.2 0.4 0 0.2 0.4

Figure 5.4bc: Hamming and Blackman Windows

Start with the definition of R̂x which is

 
r̂ x (0) r̂ x∗ (1) ··· r̂ x∗ (N − 1)
 .. 
 r̂ x (1) r̂ x (0) . r̂ x∗ (N − 2) 
R̂x =  
.. .. .. .. 
 . . . . 
r̂ x (N − 1) r̂ x (N − 2) · · · r̂ x (0)
and


 
N−l−1


1
x(n + l)x ∗ (n) 0 ≤ l ≤ N − 1

 N
 n=0
r̂ x (l) = 
N−l−1

 1
x(n − l)x ∗ (n) −(N − 1) ≤ l ≤ 0

 N


 n=0
0 otherwise
Let X be
 
x(0) 0 ··· 0
 x(1) x(0) ··· 0 
 
 .. .. .. 
 . x(0). . 
 
 .. ..
.. .. 
 . .. . 
 
1  . 
X= √  x(N − 1) x(N − 2) x(N − 3) . . 
N
 .


 0 x(N − 1) x(N − 2) . . 
 .. . 
 . x(N − 1) . . 
 0 
 .. .. .. 
 . . . 
 0 
.. .. .. ..
. . . .

then XH X is
1 
N−1
1
X X(k, l) =
H
x(n − k)x ∗ (n − l) = r̂ x (k − l)
N n=0 N

Clearly, from the above equation XH X = R̂x

80 Statistical and Adaptive Signal Processing - Solution Manual

(b) Show that zH R̂x z ≥ 0, for every z  =0.

Since R̂x = XH X then zH R̂x z = zH XH Xz = ||Xz||2 . Where || · ||2 is the 2-norm.
Clearly ||Xz||2 ≥ 0 therefore zH R̂x z ≥ 0.

5.6 An alternate autocorrelation estimate ř x (l) is given as

 1 ! N−l−1
 N−l n=0 x(n + l)x ∗ (n) 0 ≤ l ≤ L < N
ř x (l) = ř ∗ (−l) −N < −L ≤ l < 0
 x
0 elsewhere
This is an unbiased estimator, but it is not used in spectral estimation due to its negative definiteness.

(a) Show that the mean of ř x (l) is equal to r x (l) and determine an approximate expression for the variance
of ř x (l).
The mean can be shown in a straight forward manner as
 
  1 
N−l−1
1  
N−l−1

∗
E ř x (l) = E x(n + l)x (n) = E x(n + l)x ∗ (n)
N − l n=0 N − l n=0

1 
N−l−1
= r x (l) = r x (l)
N − l n=0
 
The variance of ř x (l) is var ř x (l) = E{ř x (l)ř x∗ (l)} − |r x (l)|2 . Consider
  
  1 
N−l−1
1 
N−l−1
∗ ∗ ∗
E ř x (l)ř x (l) = E x(n + l)x (n) x (m + l)x(m)
N − l n=0 N − l m=0

1  N−l−1
N−l−1   
= E x(n + l)x ∗ (n)x ∗ (m + l)x(m)
(N − l) n=0 m=0
2

Thus
  1  N−l−1
N−l−1   
var ř x (l) = E x(n + l)x ∗ (n)x ∗ (m + l)x(m) − |r x (l)|2
(N − l) n=0 m=0
2

which under mild conditions can be shown to be

∞
  1  
var ř x (l) & |r x (k)|2 + r x (k + l)r x∗ (k − l)
N − l k=−∞

(b) Show that the mean of the corresponding periodogram is given by

 ∞
jω 1
E{ Řx (e )} = Rx (e j ϕ )WR (e j ω−ϕ )dϕ
2π −∞
where WR (e j ω ) is the DTFT of the rectangular window and is something called the Dirichlet kernal.
Start with
∞
 ∞
 
N−1
E{ Řx (e j ω )} = E{ ř x (l)e− j ωl } = E{ ř x (l)e− j ωl } = E{ř x (l)}e− j ωl
l=−∞ l=−∞ l=−N+1


N−1
= r x (l)e− j ωl
l=−N+1
Statistical and Adaptive Signal Processing - Solution Manual 81

Then if we define wR (l) as

&
1 −N + 1 ≤ l ≤ N − 1
wR (l) =
0 elsewhere
!∞
Then the mean is equivalent to l=−∞ wR (l)r x (l)e− j ωl which is the DTFT of the product of two signals.
Multiplication in the time domain is convolution in the frequency domain, therefore
∞  ∞
jω − j ωl 1
E{ Řx (e )} = wR (l)r x (l)e = Rx (e j ϕ )WR (e j ω−ϕ )dϕ
l=−∞
2π −∞

5.7 Let ř x (l) be the unbiased autocorrelation estimator (5.2.13) of a zero-mean white Gaussian process with variance
σ 2x . For simplicity, we will assume a real-valued Gaussian process.
   
(a) The variance of ř x (l) is var ř x (l) = E ř x (l)ř x∗ (l) − r x2 (l) which from Problem 5.6 is

  1  N−l−1
N−l−1 
var ř x (l) = E [x(n + l)x(n)x(m + l)x(m)] − r x2 (l) (1)
(N − l) m=0 n=0
2

Using (3.2.53), we have

E [x(n + l)x(n)x(m + l)x(m)] = E [x(n + l)x(n)] E [x(m + l)x(m)]

+E [x(n + l)x(m + l)] E [x(n)x(m)]
+E [x(n + l)x(m)] E [x(n)x(m + l)]
= r x2 (l) + r x2 (m − n) + r x (n − m + l)r x (n − m − l) (2)

Substituting (2) in (1),

  1  N−l−1
N−l−1   
var ř x (l) = r x2 (m − n) + r x (m − n − l)r x (m − n + l)
(N − l) m=0 n=0
2

Assuming stationarity, we can reduce the double summation to one, that is,

  1  N−l−1
N−l−1   
var ř x (l) = r x2 (m − n) + r x (m − n − l)r x (m − n + l)
(N − l) m=0 n=0
2



1
N−l
|k|  2 
= 1− r x (k) + r x (k − l)r x (k + l)
N − l k=−(N−l) N −l
 |k|
Now for large values of l and N , 1 − N−l
→ 1 for small enough values of k. Hence

∞
  1  2 
var ř x (l) & r x (k) + r x (k − l)r x (k + l)
N − l k=−∞

which does not become small as N, l → ∞.

(b) For the biased estimator r̂ x (l), under the same conditions as those in part (a), we have
∞
  1   2 
var r̂ x (l) & r x (k) + r x (k − l)r x (k + l)
N k=−∞

which becomes zero as N → ∞.

82 Statistical and Adaptive Signal Processing - Solution Manual

5.8 To be completed.

5.9 The periodogram R̂x (e j ω ) can also be expressed as a DTFT of the autocorreltaion estimate r̂ x (l) given in (5.2.1).

(a) Let v(n) = x(n)wR (n), where wR (n) is a rectangular window of length N . Show that
1
r̂ x (l) = v(l) ∗ v ∗ (−l)
N
This can be shown directly as follows
∞
1  1 
N−l−1
r̂ x (l) = x(n + l)x ∗ (n) = x(n + l)wR (n + l)x ∗ (n)w ∗R (n)
N n=0 N n=−∞
∞
1 
= v(n + l)v ∗ (n)
N n=−∞

which is the convolution summation between v(l) and v ∗ (−l). Therefore, r̂ x (l) = 1
N
v(l) ∗ v ∗ (−l).
(b) Show that


N−1
R̂x (e j ω ) = r̂ x (l)e− j ωl
l=−N+1

The definition of the DTFT is

∞

jω
R̂x (e ) = r̂ x (l)e− j ωl
l=−∞

but r̂ x (l) = 0 if l ≤ −N or l ≥ N . Therefore


N−1
R̂x (e j ω ) = r̂ x (l)e− j ωl
l=−N+1

5.10 Consider the following simple windows over 0 ≤ n ≤ N − 1: rectangular, Bartlett, Hanning, and Hamming.

(a) Determine the DTFT of the following window analytically.

i. Rectangular: The shape of the rectangular window is
&
1 n = 0, 1, ..., N − 1
wR (n) =
0 otherwise

Inserting this directly into the equation for the DTFT gives the following
∞
 
N−1
jω − j ωn 1 − e− j ωN
WR (e ) = wR (n)e = e− j ωn =
n=−∞ n=0
1 − e− j ω

Using Eulers relation sin(ω) = 12 (e j ω − e− j ω ), the above equation can be simplified to

ω sin(ωN/2)
WR (e j ω ) = e− j 2 (N−1)
sin(ω/2)
ii. Bartlett:
Statistical and Adaptive Signal Processing - Solution Manual 83

iii. Hanning: The Hanning window is defined as

&
0.5 − 0.5 cos 2πn
0≤n ≤ N −1
wHn (n) = N−1
0 otherwise

Computing the DTFT directly follows

∞
 N−1

 
jω − j ωn 2π n
WHn (e ) = wHn (n)e = 0.5 − 0.5 cos e− j ωn
n=−∞ n=0
N − 1

N−1 
N−1

− j ωn 2π n
= 0.5 e − 0.5 cos e− j ωn
n=0 n=0
N − 1

The first part of the

!above equation is just 0.5 times the spectrum of the rectangular window. The
N−1
second part −0.5 n=0 cos( N−1 2πn
)e− j ωn can be computed by first replacing the cos(ω) with Eulers
relation cos ω = 12 (e j ω + e− j ω ) which is


N−1 j 2πn − j 2πn
− 14 (e N−1 − e N−1 )e− j ωn
n=0

where the above equation can be seperated as follows


N−1
− j (ω−
2π
)n 
N−1
− j (ω+
2π
n)
− 14 e N−1 + 1
4
e N−1
n=0 n=0

Each side can be computed independently as

1 − exp − j (ω ± 2π
)N
± 14 N−1
1 − exp − j (ω ± N−1
2π
)

and using Eulers relation, the above equation is simplified to

− 1)) sin( 2 (ω ± N−1 )N )

1 2π
−j
1
(ω ± 2π
(N
± 14 e 2 N−1
sin( (ω ± ))
1 2π
2 N−1

Combining all of the above pieces, the spectrum for the Hanning window WHn (e j ω ) is

sin( (ω − )N )
ω 1 2π
WHn (e j ω ) = 1 − j 2 (N
e − 1) sin(ωN/2) − 1 e− j 21 (ω − 2π
N−1
)(N − 1) 2 N−1
sin( (ω − N − 1)
4 sin(ω/2) 4 1 2π
2 )

− 1)) sin( 2 (ω + N−1 )N )

1 2π
−j
1
(ω + 2π
(N
+ 14 e 2 N−1
sin( (ω + ))
1 2π
2 N−1

iv. Hamming: The Hamming window is defined as

&
0.54 − 0.46 cos 2πn
0 ≤n ≤ N −1
wHm (n) = N−1
0 otherwise

It is easily seen that the only difference between the Hamming and Hanning windows are the multi-
plicative constants. The spectrum for the Hamming window is the same as for the Hanning window
84 Statistical and Adaptive Signal Processing - Solution Manual

except for the change in the constant values. Therefore, the spectrum for the Hamming window
W H m (e j ω ) is

− 1) sin( 2 (ω − N−1 )N )
ω 1 2π
− j (ω − )(N
WHm (e j ω ) = (0.27)e− j 2 (N − 1) sin(ωN/2)
1 2π
− (0.23)e 2 N−1
sin( (ω − ))
sin(ω/2) 1 2π
2 N−1

− 1) sin( 2 (ω + N−1 )N )
1 2π
−j
1
(ω + 2π
)(N
+(0.23)e 2 N−1
sin( (ω + ))
1 2π
2 N−1

(b) Figure 5.10bc shows the DTFT plots calculated by Matlab for the rectangular, Bartlett, Hanning and
Hamming windows of length N = 31. On the same plots for each window are the DTFT plots determined
analytically above.

0
−13 Rectangular Window (N=31)

−50
0 0.05 0.1 0.15 0.2 0.25
0
−27 Bartlett Window (N=31)

−80
0 0.05 0.1 0.15 0.2 0.25
0
Hanning Window (N=31)
−32

−80
0 0.05 0.1 0.15 0.2 0.25
0
Hamming Window (N=31)
−43
−80
0 0.05 0.1 0.15 0.2 0.25

Figure 5.10bc: DTFT of simple windows

5.11 The Parzen window is given by

  l 2  l 3

 1−6 L +6 L 0 ≤ |l| ≤ L
 3
wP (l) = 2 1− Ll L
≤ |l| ≤ L

 0 2
elsewhere

(a) Show that its DTFT is given by

 4
sin(ωL/4)
jω
WP (e ) & ≥0
sin(ω/4)
(b) The left hand side of Figure 5.11bc shows the Parzen window at length L = 5, 10, 20. The right hand
side of Figure 5.11bc has the magnitude spectrum of the Parzen windows with the −3 dB points noted.

5.12 The variance reduction ratio of a correlation window ωa (l) is defined as

var{ R̂x( P S) (e j ω )} Eω
& 0<ω<π
var{ R̂x (e j ω )} N
where
 π 
L−1
1
Eω = Wa2 (e j ω )dω = ω2a (l)
2π −π l=−(L−1)
Statistical and Adaptive Signal Processing - Solution Manual 85

Parzen Window Spectrum of Parzen Window

1
0
L=5 −3 L=5
0.5 −10

−20
0
0 5 10 15 20 0 0.15 0.25 0.5
1
0
L=10 −3 L=10
0.5 −10

−20
0
0 5 10 15 20 0 0.08 0.25 0.5
1
0
L=20 −3 L=20
0.5 −10

−20
0
0 5 10 15 20 0 0.04 0.25 0.5

Figure 5.11bc: Parzen Window

Variance Reduction Factor

40
Rectangular
35 Bartlett
Hanning
30 Hamming
Parzen
25

20

15

10

0
0 50 100 150

Figure 5.12a: Variance Reduction Ratio

(a) Figure 5.12a shows the plot of E ω as a function of L for the rectangular, Bartlett, Hanning, Hamming,
and Parzen windows.
(b) From Figure 5.12a, the slope of E ω for each window for L ' 1 can be computed. The slopes for each
window approach the values in the following table:
Window Name Variance Reduction Ratio
Rectangular 2L/N
Bartlett 0.667L/N
Hanning 0.75L/N
Hamming 0.7948L/N
Parzen 0.539L/N

5.13 Matlab function to compute autocorrelations using FFT.

function rx = autocfft(x,L);
% Autocorrelation sequence computation using FFT
% rx = autocfft(x,L)
% Inputs:
86 Statistical and Adaptive Signal Processing - Solution Manual

% x : data vector; L : maximum correlation distance

% Output:
% rx : estimate of autocorrelation over 0 <= l <= L-1
%
%-----------------------------------------------------------
% Copyright 2000, by Dimitris G. Manolakis, Vinay K. Ingle,
% and Stephen M. Kogon. For use with the book
%
% "Statistical and Adaptive Signal Processing"
%
% McGraw-Hill Higher Education.
%-----------------------------------------------------------

N = length(x); x = x - mean(x);
if L > (N-1); error(’*** L must be <= N-1 ***’); end;
x = reshape(x,N,1); x = [x; zeros(N,1)];
Nfft = 2^ceil(log10(2*N-1)/log10(2));
Rx = (1/N)*(abs(fft(x,Nfft)).^2);
rx = ifft(Rx);
%rx = [flipud(rx(end:-1:end-L+1));rx(1:1:L+1)];
rx = rx(1:1:L);
if isreal(x); rx = real(rx); end;

For comparison between and autocfft.m execution time consider the following Matlab script.

x = rand(1000,1); flops(0); rx = autocfft(x,200); flops

ans =
160532
x = rand(1000,1); flops(0); rx = autoc(x,200); flops
ans =
482603
x = rand(1000,1); flops(0); rx = autocfft(x,500); flops
ans =
160532
x = rand(1000,1); flops(0); rx = autoc(x,500); flops
ans =
1503503

Clearly the autocfft.m function is more efficient for larger values of lag l.

5.14 The Welch-Bartlett estimate R̂x( P A) (k) is given by

K −1
1 
R̂x( P A) (k) = |X i (k)|2
K L i=0

Let x(n) be real valued. Then two real-valued sequences can be combined into one complex-valued sequence
K
gr (n) = x2r (n) + j x2r+1 (n) n = 0, 1, ..., L − 1, r = 0, 1, ..., −1
2
Statistical and Adaptive Signal Processing - Solution Manual 87

and the L-point DFT of gr (n) is

K
G̃ r (k) = X̃ 2r (k) + j X̃ 2r+1 (k) k = 0, 1, ..., L − 1 r = 0, 1, ..., −1
2
(a) Show that
K
|G̃ r (k)|2 + |G̃ r (L − k)|2 = 2[| X̃ 2r (k)|2 + | X̃ 2r+1 (k)|2 ] k = r = 0, 1, ..., −1
2
Start with

|G̃ r (k)|2 + |G̃ r (L − k)|2 = | X̃ 2r (k) + j X̃ 2r+1 (k)|2 + | X̃ 2r (L − k) + j X̃ 2r+1 (L − k)|2

The right hand side can be expanded into

∗ ∗
r.h.s. = ( X̃ 2r (k) + j X̃ 2r+1 (k))( X̃ 2r (k) + j X̃ 2r+1 (k))
∗ ∗
+( X̃ 2r (L − k) + j X̃ 2r+1 (L − k))( X̃ 2r (L − k) + j X̃ 2r+1 (L − k))

Working out this multiplication directly results in the following cross terms
∗ ∗
r.h.s. = | X̃ 2r (k)|2 + | X̃ 2r+1 (k)|2 − j X̃ 2r (k) X̃ 2r+1 (k) + j X̃ 2r+1 (k) X̃ 2r (k)
+| X̃ 2r (L − k)|2 + | X̃ 2r+1 (L − k)|2
∗ ∗
− j X̃ 2r (L − k) X̃ 2r+1 (L − k) + j X̃ 2r+1 (L − k) X̃ 2r (L − k)
∗
This can be simplified using the fact that for real-valued x2r , X̃ 2r (L − k) = X̃ 2r (k). Most of the cross
terms cancel, leaving

= | X̃ 2r (k)|2 + | X̃ 2r+1 (k)|2 + | X̃ 2r (L − k)|2 + | X̃ 2r+1 (L − k)|2

This can be further simplified since for real-valued x2r the following is true
∗ ∗
| X̃ 2r (L − k)|2 = ( X̃ 2r (L − k))( X̃ 2r (L − k)) = ( X̃ 2r (k))( X̃ 2r (k)) = | X̃ 2r (k)|2

Similarly | X̃ 2r+1 (L − k)|2 = | X̃ 2r+1 (k)|2 . The final result is shown as

|G̃ r (k)|2 + |G̃ r (L − k)|2 = 2[| X̃ 2r (k)|2 + | X̃ 2r+1 (k)|2 ] k = r = 0, 1, ..., K

2
−1

(b) The expression for R̂x( P A) (k) in terms of G̃(k) is shown below

K −1
K
−1
1  1 2
R̂x( P A) (k) = |X i (k)| =
2
(| X̃ 2i (k)|2 + | X̃ 2i+1 (k)|2 )
K L i=0 K L i=0
K
−1
1 2
= (|G̃ r (k)|2 + |G̃ r (L − k)|2 )
K L i=0

(c) If k is odd, then there in one DFT which is unpaired.

(K −1)
 2 −1
1 1
R̂x( P A) (k) = (| X̃ 2i (k)|2 + | X̃ 2i+1 (k)|2 ) + | X̃ K −1 (k)|2
KL i=0
KL
88 Statistical and Adaptive Signal Processing - Solution Manual

5.15 The autocorrelation estimate from a Welch’s method PSD estimate is

 π
(PA) 1
r̂ x (l) = R̂ (PA) (e j ω )e j ωl dω
2π −π x
(PA)
Let R̂˜ x (k) be samples of R̂x(PA) (k) according to
(PA)
R̂˜ x (k) = R̂x(PA) (e j 2πk/NFFT ) 0 ≤ k ≤ NFFT − 1
(PA)
(a) Show that the IDFT r̂˜ x (l) of R̂˜ x (k) is an aliased version of the autocorralation estimate r̂ x(PA) (l)
(PA)

˜ (PA)
The IDFT of R̂ x (k) is
FFT −1
N FFT −1
N
(PA) 1 ˜ (PA) 1
r̂˜ x (l) = R̂ x (k)e j 2πkl/NFFT = R̂x(PA) (e j 2πk/NFFT )e j 2πkl/NFFT
NFFT k=0
N FFT k=0
 2π $  N −1

1 %  FFT
2π k
= R̂x(PA) (e j ω )e j ωl δ ω− dω
NFFT 0 k=0
NFFT

This is the convolution between r̂ x(PA) (l) and an impulse train. Using the identity that convolution with an
(PA)
impulse train results in aliased version of the signal, therefore r̂˜ (l) is x

FFT −1
N
(PA)
r̂˜ x (l) = r̂ x(PA) (l − k NFFT )
k=0

(b) Since R̂x(PA) (e j ω ) is obtained by averaging periodograms, each of which, according to (5.3.6), is a DTFT
(PA)
of length (2L − 1) autocorrelation estimates r̂ x,i (l). Clearly then R̂x(PA) (e j ω ) is also a DTFT of length
(2L − 1) autocorrelation estimates r̂ x(PA) (l). Hence to avoid aliasing, NFFT ≥ (2L − 1).

5.16 The coherence function Gx2y is

|Rx y (e j ω )|2
Gx2y =
Rx (e j ω )R y (e j ω )
Show that it is invariant under linear transformation x1 = h 1 (n) ∗ x(n) and y1 = h 2 (n) ∗ y(n).
Start with the definition of the coherence function
|Rx1 y2 (e j ω )|2
Gx21 y1 =
Rx1 (e j ω )R y2 (e j ω )
Substitute into the above equation the following appropriate filtered autocorrelation relations

Rx1 (e j ω ) = H1 (e j ω )H1∗ (e j ω )Rx (e j ω ), R y2 (e j ω ) = H2 (e j ω )H2∗ (e j ω )R y (e j ω )

|Rx1 y2 (e j ω )|2 = Rx1 y2 (e j ω )Rx∗1 y2 (e j ω ) = [H1∗ (e j ω )H2 (e j ω )Rx y (e j ω )][H1 (e j ω )H2∗ (e j ω )Rx∗y (e j ω )]

Therefore
|Rx1 y2 (e j ω )|2 [H1∗ (e j ω )H2 (e j ω )Rx y (e j ω )][H1 (e j ω )H2∗ (e j ω )Rx∗y (e j ω )]
Gx21 y1 = =
Rx1 (e j ω )R y2 (e j ω ) H1 (e j ω )H1∗ (e j ω )Rx (e j ω )H2 (e j ω )H2∗ (e j ω )R y (e j ω )
|Rx y (e j ω )|2
=
Rx (e j ω )R y (e j ω )
Statistical and Adaptive Signal Processing - Solution Manual 89

5.17 Bartlett’s Method as a special case of Welch’s method:

(a) Consider the ith periodogram:

 T
(b) Let u(e j ω ) = 1 e j ω · · · e j (L−1)ω and
 
r̂ x,i (0) r̂ x,i (−1) · · · r̂ x,i (1 − L)
 r̂ x,i (1) r̂ x,i (0) ··· r̂ x,i (2 − L) 
 
R̂x,i =  .. .. .. .. 
 . . . . 
r̂ x,i (L − 1) r̂ x,i (L − 2) · · · r̂ x,i (0)
We will first show (P.6) using explicit calculations. First note that


L − |l|
wB (l) = , |l| ≤ L (1)
L

Consider L = 1: Then
1 H jω  
u (e )R̂x,i u(e j ω ) = [1] r̂ x,i (0) [1] = (0) r̂ x,i (−1)e j ω + (1) r̂ x,i (0) + (0) r̂ x,i (1)e− j ω
1
1 1

− j ωl 1 − |l|
= r̂ x,i (l) (1 − |l|) e = r̂ x,i (l) e− j ωl
l=−1 l=−1
1

1
= r̂ x,i (l) wB (l) e− j ωl
l=−1

Consider L = 2:
  
1 H jω 1  r̂ x,i (0) r̂ x,i (−1) 1
u (e )R̂x,i u(e j ω ) = 1 e− j ω
2 2 r̂ x,i (1) r̂ x,i (0) e jω
1  
= r̂ x,i (0) + e− j ωr̂ x,i (1) + r̂ x,i (−1) + e− j ωr̂ x,i (0) e j ω
2
1
= (0) r̂ x,i (−2)e2 j ω + (1) r̂ x,i (−1)e j ω + (2) r̂ x,i (0)
2
+ (1) r̂ x,i (1)e− j ω + (0) r̂ x,i (2)e−3 j ω
 2
 2
2 − |l| − j ωl
= r̂ x,i (l) e = r̂ x,i (l) wB (l) e− j ωl
l=−2
2 l=−2

Consider L = 3:
  
r̂ x,i (0) r̂ x,i (−1) r̂ x,i (−2) 1
1 H jω 1 
u (e )R̂x,i u(e j ω ) = 1 e− j ω e− j 2ω  r̂ x,i (1) r̂ x,i (0) r̂ x,i (−1)   e j ω 
3 3
r̂ x,i (2) r̂ x,i (1) r̂ x,i (0) e j 2ω
1    
= r̂ x,i (0) + e− j ω r̂ x,i (1) + e−2 j ωr̂ x,i (2) + r̂ x,i (−1) + r̂ x,i (0)e− j ω + r̂ x,i (1)e− j 2ω e
3  
+ r̂ x,i (−2) + e− j ωr̂ x,i (−1) + e−2 j ωr̂ x,i (0) e2 j ω
1
= (0) r̂ x,i (−3)e3 j ω + r̂ x,i (−2)e2 j ω + 2r̂ x,i (−1)e j ω + 3r̂ x,i (0) + 2r̂ x,i (1)e− j ω
3
+r̂ x,i (2)e−2 j ω + (0) r̂ x,i (2)e−3 j ω
 3
 3
3 − |l| − j ωl
= r̂ x,i (l) e = r̂ x,i (l) wB (l) e− j ωl
l=−3
3 l=−3
90 Statistical and Adaptive Signal Processing - Solution Manual

Now consider the general expression


L 
L

jω − j ωl L − |l|
R̂x,i (e ) = r̂ x,i (l) wB (l) e = r̂ x,i (l) e− j ωl
l=−L l=−L
L

1 
L
= r̂ x,i (l) (L − |l|) e− j ωl
L l=−L

Using the identity


L 
L−1 
L−1
(L − |l|) f (l) = f (m − n)
l=−L m=0 n=0

we have

1  1 
L L−1 L−1
R̂x,i (e j ω ) = r̂ x,i (l) (L − |l|) e− j ωl = r̂ x,i (m − n)e− j ω(m−n)
L l=−L L m=0 n=0
1 H jω
= u (e )R̂x,i u(e j ω )
L
(c) Finally, using (5.3.55)

1  1  H jω
K K
R̂x(B) (e j ω ) = R̂x,i (e j ω ) = u (e )R̂x,i u(e j ω )
K i=1 K L i=1

5.18 Combined data and correlation-lag windowing (Carter-Nuttall Approach):

Let R̂x(PA) (e j ω ) be the Welch-Bartlett PSD estimator of Rx (e j ω ). Then the Carter-Nuttall PSD estimator
R̂x(CN) (e j ω ) is given by
   
R̂x(CN) (e j ω ) = F r̂ x(CN) (l) , F r̂ x(PA) (l)wa (l) , R̂x(PA) (e j ω ) ~ Wa (e j ω )
 π
1
= R̂ (PA) (e j θ )Wa (e j (ω−θ) ) dθ (1)
2π −π x

where ~ denotes the periodic convolution and Wa (e j ω ) is the spectrum of the correlation lag window wa (l),
−L ≤ l ≤ L.

(a) Bias of R̂x(CN) (e j ω ): Note that from (5.3.57) the mean of R̂x(PA) (e j ω ) is given by
$ % 1
E R̂x(PA) (e j ω ) = R̂x(PA) (e j ω ) ~ Rw (e j ω )
L
where Rw (e j ω ) is the spectrum of the data window w(n), 0 ≤ n ≤ L − 1. Thus from (1)
$ %  π $ %
(CN) j ω 1
E R̂x (e ) = E R̂x(PA) (e j θ ) Wa (e j (ω−θ ) ) dθ
2π −π
$ %
= E R̂x(PA) (e j ω ) ~ Wa (e j ω ) (2)
1 (PA) j ω
= R̂ (e ) ~ Rw (e j ω ) ~ Wa (e j ω ) (3)
L x
Thus R̂x(CN) (e j ω ) is a biased estimator.
Statistical and Adaptive Signal Processing - Solution Manual 91

(b) From (1), the Carter-Nuttall PSD estimator, R̂x(CN) (e j ω ), is obtained by performing additional windowing
on the Welch-Bartlett PSD estimator, R̂x(PA) (e j ω ). Thus the variance of R̂x(PA) (e j ω ) is further reduced while
broadening frequency peaks (or decreasing resolution). The larger the maximum lag-length L, the lesser
are these effects.
(c) Matlab function CN_psd:
function [Rx,Freq] = CN_psd(x,L,Lagwindow,Fs)
% Spectrum estimation using Carter and Nuttall approach that
% combines data and lag windows.
% Rx = CN_psd(x,L,Window,Fs)
%
% Inputs:
% x : data values
% L : maximum Lag distance
% LagWindow : Lag window function (must be previously defined)
% Fs : sampling rate
%
% Output:
% Rx : Spectral estimates over 0 -- Fs/2
% Freq : Frequency samples between 0 and Fs/2
%-----------------------------------------------------------
% Copyright 2000, by Dimitris G. Manolakis, Vinay K. Ingle,
% and Stephen M. Kogon. For use with the book
%
% "Statistical and Adaptive Signal Processing"
%
% McGraw-Hill Higher Education.
%-----------------------------------------------------------

Nfft = 2^ceil(log10(2*L-1)/log10(2)); % Nfft : power of 2 & > (2L-1)

datawin = hamming(L); % Hamming window for Welch PSD
Rx = psd(x,Nfft,Fs,datawin,L/2,’none’); % Welch PSD with 50% overlap
Rx = [Rx;flipud(Rx(2:end-1))]; % Assemble full period for IFFT
rx = real(ifft(Rx)); % IFFT -> correlation estimates
rx = [rx(end:-1:end-L+1);rx(1:L+1)]; % extract over -L to L
win = feval(Lagwindow,2*L+1); % Lag window of length 2*L+1
rx = rx.*win; % Window correlation estimates
rx = fftshift(rx(1:end-1)); % assemble for FFT for
rx = [rx(1:L+1);zeros(Nfft-2*L,1);rx(L+2:end)]; % Nfft length DFT
Rx = real(fft(rx)); % Carter Nuttall estimates
Rx = Rx(1:Nfft/2+1); % extract over 0 to Fs/2
Freq = [0:Nfft/2]*Fs/Nfft; % Frequencies between 0 and Fs/2
! L−1
5.19 The scaling factor n=0 w 2 (n) = L normalizes the data window. Hence using the Parseval’s theorem
 π
1 2  
L−1
1 W (e j ω )2 dω = 1
w (n) =
L n=0 2π L −π

This makes the estimate R̂x(PA) (e j ω ) asymptotically unbiased.

92 Statistical and Adaptive Signal Processing - Solution Manual

5.20 Consider the basic periodogram estimator

 N−1 
 
1  
R̂(e j ω ) =  x(n)e− j ωn 
N  
n=0

(a) At ω = 0, we have
 N−1 2  N−1 2
   
1  − j 0n  1  
R̂(e ) =
j0
 x(n)e  =  x(n) 
N   N  
n=0 n=0

(b) Assume that x(n) ia a real-valued zero-mean white Gaussian process with variance σ 2x . Then r x (l) = σ 2x .
The mean of R̂(e j 0 ) is given by
  2   N−1   N−1 
$ % N−1   
1   1
E R̂(e ) = E  
j0
x(n)  = E x(n) x(m)
N  n=0  N n=0 m=0
 N−1 N−1 
1  1 
N−1 N−1
1 
N−1 N−1
= E x(n)x(m) = E [x(n)x(m)] = r x (n − m)
N n=0 m=0
N n=0 m=0 N n=0 m=0

1 2
N−1 N−1
= σ δ(n − m) = σ 2x
N x n=0 m=0

The variance of R̂(e j 0 ) is given by

$ % $ % $ % $ %
var R̂(e j 0 ) = E R̂ 2 (e j 0 ) − E2 R̂(e j 0 ) = E R̂ 2 (e j 0 ) − σ 4x

Consider
   N−1   N−1  
$ % 1  
N−1    
N−1 
E R̂ 2 (e j 0 ) = E x(m) x(n) x( p) x(q) 
N2   
m=0 n=0 p=0 q=0
 
1 
N−1 
N−1 
N−1 
N−1
= E x(q)x( p)x(n)x(m)
N2 m=0 n=0 p=0 q=0

1 
N−1 N−1 N−1 N−1
= E [x(q)x( p)x(n)x(m)]
N 2 m=0 n=0 p=0 q=0

Now using the property (3.2.53) of real-valued Gaussian random variables, that is,

E [x1 x2 x3 x4 ] = E [x1 x2 ] E [x3 x4 ] + E [x1 x3 ] E [x2 x4 ] + E [x1 x4 ] E [x2 x3 ]

we obtain

E [x(q)x( p)x(n)x(m)] = E [x(q)x( p)] E [x(n)x(m)] + E [x(q)x(n)] E [x( p)x(m)]

+E [x(q)x(m)] E [x( p)x(n)]
= σ 2x δ(q − p)σ 2x δ(n − m) + σ 2x δ(q − n)σ 2x δ( p − m) + σ 2x δ(q − m)σ 2x δ( p − n)
= σ 4x [δ(q − p)δ(n − m) + δ(q − n)δ( p − m) + δ(q − m)δ( p − n)]
Statistical and Adaptive Signal Processing - Solution Manual 93

Hence
$ % σ 4x    
N−1 N−1 N−1 N−1
E R̂ 2 (e j 0 ) = [δ(q − p)δ(n − m) + δ(q − n)δ( p − m) + δ(q − m)δ( p − n)]
N 2 m=0 n=0 p=0 q=0
 N−1 N−1   N−1 N−1 
σ 4    
= 3 x2 δ(n − m)  δ(q − p) = 3σ 4x
N m=0 n=0 p=0 q=0

Therefore,
$ %
var R̂(e j 0 ) = 3σ 4x − σ 4x = 2σ 4x
$ %
(c) Clearly the estimator R̂ 2 (e j 0 ) is not consistent since var R̂(e j 0 is independent of N .

5.21 Consider Bartlett’s method for estimating R̂x (e j 0 ) using L = 1. The periodogram of one sample x(n) is simply
|x(n)|2 . Thus

1  1 
N−1 N−1
R̂x(B) (e j 0 ) = R̂x,n (e j 0 ) = |x(n)|2
N n=0 N n=0
 
Assume that x(n) is a real-valued white Gaussian process with variance σ 2x . Then r x (l) = E x 2 (n) = σ 2x .

(a) The mean of R̂x(B) (e j 0 ) is given by

 
$ 1 
N−1 % 1   2 
N−1
(B) j 0
E R̂x (e ) = E |x(n)| =
2
E x (n) = σ 2x
N n=0 N n=0

The variance of R̂x(B) (e j 0 ) is given by

$ % " #2  $ % " #2 
(B) j 0 (B) j 0 (B) j 0 (B) j 0
var R̂x (e ) = E R̂x (e ) − E R̂x (e ) = E R̂x (e )
2
− σ 4x

Consider
"   
#2  1  2
N−1
1  2
N−1
E R̂x(B) (e j 0 ) = E x (m) x (n)
N m=0 N n=0

1   2 
N−1 N−1
= E x (n)x 2
(m)
N 2 m=0 n=0

Using the property (3.2.53) of real-valued Gaussian random variables, that is,

E [x1 x2 x3 x4 ] = E [x1 x2 ] E [x3 x4 ] + E [x1 x3 ] E [x2 x4 ] + E [x1 x4 ] E [x2 x3 ]

we obtain
     
E x 2 (n)x 2 (m) = E x 2 (n) E x 2 (m) + E [x(n)x(m)] E [x(n)x(m)] + E [x(n)x(m)] E [x(n)x(m)]
= σ 4x + 2σ 4x δ(n − m)
94 Statistical and Adaptive Signal Processing - Solution Manual

Hence
" #2  1   2  1  4 
N−1 N−1 N−1 N−1
E R̂x(B) (e j 0 ) = 2
E x (n)x 2
(m) = 2
σ x + 2σ 4x δ(n − m)
N m=0 n=0 N m=0 n=0
1 2σ 4x
= σ 4x + 2σ 4
N = σ 4
+
N2 x x
N
Finally
$ % " #2  2σ 4x
var R̂x(B) (e j 0 ) =E R̂x(B) (e j 0 ) − σ 4x =
N

(b) Comparison with the results of Problem 5.20: The means and variances of R̂x(B) (e j 0 ) are

Problem Mean Variance

5.20 σ 2x 2σ 4x
5.21 σ 2x 2σ 4x /N

Clearly, the variance estimator of Problem 5.21 is consistent.

5.22 Design of a lag window using an arbitrary data window.

(a) Let v(n), 0 ≤ n ≤ L − 1 be a data window. Let the lag window w(n) de defined by
   
w(n) , F −1 |F {v(n)}|2 = F −1 V (e j ω )V ∗ (e j ω )

Then using the convolution property of the DTFT, the lag window is

w(n) = v(n) ∗ v ∗ (n), 0 ≤ n ≤ 2L − 2

Thus w(n) has the desired property that its DTFT is nonnegative. We also the length of the window equal
to 2L + 1.
(b) Matlab script to design a lag window with The Hanning window as the prototype is given below and the
plots of the window and its DTFT are shown in Figure P5.22.
% (b) Design of a lag window of length 31 using the
% Hanning window prototype
N = 31; L = (N+1)/2; n = 0:N-1;
v = hanning(L);
w = conv(v,v);
% DTFT of w(n);
W = fft(w,1024); magW = abs(W(1:513));
omg = [0:512]/512;

% Plots
subplot(’position’,[0.1,0.65,0.85,0.3])
stem(n,w,’g’); axis([-1,N,0,8]);
set(gca,’xtick’,[0,15,30],’ytick’,[0:2:8]);
xlabel(’sample index n’,’fontsize’,label_fontsize);
ylabel(’x(n)’,’fontsize’,label_fontsize);
title(’Lag Window’,’fontsize’,title_fontsize);
Statistical and Adaptive Signal Processing - Solution Manual 95

Lag Window
8

x(n)
4

0
0 15 30
sample index n

Spectrum of the Lag Window

80

60
W( ejω )

40

20

0
0 0.25 0.5 0.75 1
normalized frequency ω/π

Figure 5.22: Plots of the lag window in the time and frequency domains.

subplot(’position’,[0.1,0.15,0.85,0.3])
plot(omg,magW,’m’); axis([0,1,0,80]);
set(gca,’xtick’,[0:0.25:1],’ytick’,[0:20:80]);
xlabel(’normalized frequency \omega/\pi’,’fontsize’,label_fontsize);
ylabel(’W( e^{j\omega} )’,’fontsize’,label_fontsize);
title(’Spectrum of the Lag Window’,’fontsize’,title_fontsize);

5.23 Consider the following random process


4
x(n) = Ak sin(ωk n + φ k ) + ν(n)
k=1

A1 = 1 A2 = 0.5 A3 = 0.5 A4 = 0.25

where and the phases {φ i }4i=1 are IID random variables
ω1 = 0.1π ω2 = 0.6π ω3 = 0.65π ω4 = 0.8π
uniformly distributed over [−π , π]. The Blackman-Tukey estimates will be computed from fifty realizations
of x(n) for 0 ≤ n ≤ 256

(a) The top two row of Figure 5.23ab shows the Blackman-Tukey estimate using a Bartlett window for
L = 32, 64, 128 in dB. The left sided plots are the Blackman-Tukey estimates for each realization, while
the right sided plots are the average of the Blackman-Tukey estimates.
(b) The bottom row of Figure 5.23ab shows the Blackman-Tukey estimate using a Parzen window for L =
32, 64, 128 in dB. The left sided plots are the B-T estimates for each realization, while the right sided
plots are the average of the B-T estimates.
(c) It can be seen in Figure 5.23ab that neither window does a good job of estimating the true spectrum with a
window length less then L = 128. At this length, all four frequency components are visable. For lengths
less then L = 128, the Parzen window is more tightly bound around the true spectrum, with the Bartlett
window estimate having a larger envelope. This stand to reason, the Bartlett window used for part (a) has
higher sidelobes then the Parzen window used in part (b). The higher the sidelobe the greater the spectral
leakage from the true spectra.
96 Statistical and Adaptive Signal Processing - Solution Manual

B−T estimates w/ Bartlett Window (overlay) B−T estimates w/ Bartlett Window (average)
20 N=32 20 N=32
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 N=64 20 N=64
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 N=128 20 N=128
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
B−T estimates w/ Parzen Window (overlay) B−T estimates w/ Parzen Window (average)
20 N=32 20 N=32
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 N=64 20 N=64
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 N=128 20 N=128
0 0
−20 −20
−40 −40
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.23ab:
Statistical and Adaptive Signal Processing - Solution Manual 97

5.24 Consider the random process given in Problem 5.23

Bartlett Estimate (overlay) Bartlett Estimate (average)
10 10
L=16 L=16
5 5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 10
L=32 L=32
10 5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
40 20
L=64 L=64
20 10

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Welch Estimate (overlay) Welch Estimate (average)
5 4
L=16 L=16
2

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
10 10
L=32 L=32
5 5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L=64 L=64
10 10

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.24ab:

(a) The top row of Figure 5.24ab shows the Bartlett estimate in dB, which uses no overlap, a rectangular
window, and L = 32, 64, 128. The left sided plots are the Bartlett estimates for each realization, while
the right sided plots are the average of the Bartlett estimates.
(b) The bottom row of Figure 5.24ab shows the Welch estimate in dB using 50% overlap, a Hamming window,
and L = 32, 64, 128. The left sided plots are the Welch estimates for each realization, while the right
sided plots are the average of the Bartlett estimates.
(c) The Bartlett estimates have better resolution but slightly higher variance.

5.25 Consider the random process given in Problem 5.23.

(a) Figure 5.25a shows the multitaper spectrum estimate, using K = 3, 5, 7 Slepian tapers. The left sided
plot is the estimates for each realizations while the right sided plot is the average of the estimates.
(b) As expected the multitaper spectral estimates have wider (and almost rectangular) peak and hence less
resolution.

5.26 This problem deals with spectral estimation using three different techniques. A 1000 sample AR(1) process is
generated with a = −0.9.

(a) The top row of Figure 5.26 shows a periodogram estimate for the AR(1) process. It is not very accurate
over the entire spectrum and has a high variance.
(b) The bottom left side plot of Figure 5.26 shows a Blackman-Tukey estimate for L = 10, 20, 50, 100.
Clearly, as L increases the estimated spectrum more closely matches the true spectrum.
(c) The bottom right side plot of Figure 5.26 shows a Welch estimate,with 50% overlap, using a Hamming
window, for L = 10, 20, 50, 100. Clearly, as L increases the estimated spectrum more closely matches
the true spectrum. This appears to be the best estimator of the three.
98 Statistical and Adaptive Signal Processing - Solution Manual

Multitaper Estimate (overlay) Multitaper Estimate (average)

20 20
K=3 K=3
10 10

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 10
K=5 K=5
10 5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
10 10
K=7 K=7
5 5

0 0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.25a:

Periodogram estimate vs. True Spectrum (a=−0.9)

25

20

15

10

−5

−10

−15
0 0.1 0.2 0.3 0.4 0.5
B−T estimate vs. True Spectrum (a=−0.9) Welch estimate vs. True Spectrum (a=−0.9)
20 20
L= 10 L= 10
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 20 L= 20
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 50 L= 50
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 100 L= 100
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.26:
Statistical and Adaptive Signal Processing - Solution Manual 99

5.27 This problem deals with spectral estimation using three different techniques. A 1000 sample AR(1) process is
generated with a = 0.9.

(a) The top row of Figure 5.27 shows a periodogram estimate for the AR(1) process. It is not very accurate
over the entire spectrum and has a high variance.
(b) The bottom left side plot of Figure 5.27 shows a Blackman-Tukey estimate for L = 10, 20, 50, 100.
Clearly, as L increases the estimated spectrum more closely matches the true spectrum.
(c) The bottom right side plot of Figure 5.27 shows a Welch estimate, with 50% overlap, using a Hamming
window, for L = 10, 20, 50, 100. Clearly, as L increases the estimated spectrum more closely matches
the true spectrum. This appears to be the best estimator of the three.

Periodogram estimate vs. True Spectrum (a=0.9)

20

15

10

−5

−10

−15
0 0.1 0.2 0.3 0.4 0.5
B−T estimate vs. True Spectrum (a=0.9) Welch estimate vs. True Spectrum (a=0.9)
20 20
L= 10 L= 10
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 20 L= 20
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 50 L= 50
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
20 20
L= 100 L= 100
10 10
0 0
−10 −10
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.27:

5.28 Multitaper estimation technique requires a properly designed orthonormal set of tapers for proper performance.
One set discussed in the chapter is that of harmonically related sinusoids given by

2 π (k + 1)(n + 1)
ωk (n) = sin n = 0, 1, ..., N − 1
N −1 N +1
(a) Listed below is a MATLAB functions that will generate K < N sinusoidal tapers of length N .

function w = sine_tapers(N, K)

k = [1:K]’; n = [0:N-1]; w = sqrt(2/(N-1))* sin((pi*(k+1)*(n+1))/(N+1));

100 Statistical and Adaptive Signal Processing - Solution Manual

Spectrum of Sine Tapers (length=51)

20
K=1
K=2
10
K=3
K=4
0 K=5

−10

−20

−30

−40
0 0.05 0.1 0.15

Figure 5.28b:

(b) Figure 5.28b plots the first five tapers of length 51.

5.29 The following MATLAB function determines the multitaper estimate using the sine tapers generated above

function Pxx = psd_sinetaper(x, K)

N = length(x); w = sine_tapers(N,K);
for k = 1:K
x_w(k,:) = x.*w(1,:);
S_x(:,k) = psd(x_w(k,:),N , 1, ones(N/10,1),0);
end
Pxx = mean(S_x’);

(a) The left side of Figure 5.29 shows the estimate of the AR(1) process given in Problem 5.26
(b) The right side of Figure 5.29 shows the estimate of the AR(1) process given in Problem 5.27

Sine Taper estimate vs. True Spectrum (a=0.9) Sine Taper estimate vs. True Spectrum (a=−0.9)
20 20
10 10
0 0
−10 −10

−20 −20

−30 −30

−40 −40

−50 −50

−60 −60

−70 −70

−80 −80
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5

Figure 5.29:
 Chapter 6
Optimum Linear Filters

6.1 Show that the linear MMSE estimate ŷ of a random variable y using the data vector x is given by ŷ = yo + c H x,
where yo = E{y} − c H E{x} , c = R−1 d , R = E{xx H } , and d = E{xy ∗ }.
Let

ε 2 = | ŷ − y|2 = ŷ ŷ ∗ − ŷ y ∗ − y ŷ ∗ − yy ∗

Now, ŷ is a linear estimate of y, therefore

ŷ = yo + c H x

Substituting the expression for ŷ into the above expression for ε2 , computing its expected value and taking the
derivative with respect to yo results in

d Eε 2
= 2yo + 2µ xH c − 2µ y
dyo

Setting the above expression equal to zero yields yo = µ y − c H µx

6.2 Consider an optimum FIR filter specified by the input correlation matrix R = Toeplitz {1, 14 } and cross-
correlation vector d = [1 12 ]T .

(a) Determine the optimum impulse response co and the MMSE Po

Using (6.2.12), the optimum impulse response co is

Rco = d ⇒ co = R −1 d = [0.933 0.267]T

and, using (6.2.17), the MMSE Po is

 
0.933
Po = Py − d co = Py − [1
H 1
2
] = Py − 1.067
0.267

(b) Express co and the MMSE Po in terms of the eigenvalues and eigenvectors of R
The eigenvalues and eigenvectors of R are
   
0.7071 0.7071 0.75 0 0.7071 −0.7071
R = QQ H =
−0.7071 0.7071 0 1.25 0.7071 0.7071

Using (6.2.35) and (6.2.32), co and Po are found using the eigenvalues and eigenvectors as is shown below


2
qHd
co = QQ d = H i
qi
i=1
λi
     
0.7071 1 1 0.7071
= 1
[0.7071
− 0.7071] 1 + 1
[0.7071 0.7071]
0.75 −0.7071 1.25 1
0.7071
    2  2
0.333 0.6 0.933
= + =
−0.333 0.6 0.267

101
102 Statistical and Adaptive Signal Processing - Solution Manual


2
|d ! |2 (0.3536)2 (1.0607)2
Po = Py − i
= Py − − = Py − 1.067
i=1
λ 0.75 1.25

where d ! = Q H d = [0.3536 1.0607]T

6.3 Repeat Problem 6.2 for a third-order optimum FIR filter, with R = Toeplitz {1, 14 , 0} and cross-correlation
vector d = [1 12 0]T .

(a) Solving the same way as Problem 6.2, using (6.2.12) and (6.2.17) results in
 
0.9286
co = R −1 d =  0.2857 
−0.0714

and

Po = Py − d H co = Py − 1.071

(b) The eigenvalues and eigenvectors of R are

 
0.5 −0.7071 −0.5
Q =  0.7071 0 0.7071 
0.5 0.7071 −0.5

and
 
1.3536 0 0
= 0 1 0 
0 0 0.6464

Using (6.2.35) and (6.2.32), co and Po are

       
0.315 0.5 0.113 0.9286
co =  0.446  +  0  +  −0.160  =  0.2857 
0.315 −0.5 0.113 −0.0714

and
(0.8536)2 (0.7071)2 (0.1464)2
Po = Py − − − = Py − 1.071
1.3536 1 0.6464

6.4 A process y(n) with the autocorrelation r y (l) = a |l| , −1 < a < 1, is corrupted by additive, uncorrelated
white noise ν(n) with variance σ 2ν . To reduce the noise in the observed process x(n) = y(n) + ν(n), we use a
first-order Wiener filter.

(a) Express the coefficients co,1 and co,2 and the MMSE Po in terms of the parameters a and σ 2ν
The autocorrelation matrix for x(n) is
 
1 + σ 2ν a
Rx = E{x(n)x (n)} = R y + σ ν I =
H 2
a 1 + σ 2ν
Statistical and Adaptive Signal Processing - Solution Manual 103

and the cross-correlation matrix is

   
∗ y(n)y ∗ (n) + ν(n)y ∗ (n) 1
d = E{x(n)y (n)} = E =
y(n + 1)y ∗ (n) + ν(n + 1)y ∗ (n) a

Solving for the filter coefficients using co = R −1 d yields the following

 
   1+σ 2ν −a 2
1 + σν
2
−a 1
co = ((1 + σ 2ν )2 − a 2 ) = ((1+σ 2ν )2 −a 2 ) 
−a 1 + σ 2ν a σ 2ν a
((1+σ 2ν )2 −a 2 )

The MMSE Po is found using Po = Py − d H co where Py = E|y|2 = 1. After simplifying, the MMSE
Po is
 
2 2 1+σ ν −a
Po = 1 − [1 a]  ((1+σ 2ν )2 −a 2 )
σ 2ν a
 = σ 2ν (1 + a 2 + σ 2ν )
((1+σ 2ν )2 −a 2 )

(b) Compute and plot the PSD of x(n) and the magnitude response |Co (e j ω )| of the filter when σ 2v = 2, for
both a = 0.8 and a = −0.8, and compare the results.
% (b) Plot of the PSD of x(n) and the Mag Res of the optimal filter
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0604b’);

% Given parameters
var_v = 2;
% 1. a = 0.8
a = 0.8;
% Computed Parameters
c_o1 = (1+var_v-a*a)/((1+var_v)*(1+var_v)+a*a);
c_o2 = (var_v*a)/((1+var_v)*(1+var_v)+a*a);
% PSD of x(n)
omg = [0:500]*pi/500;
Rx = (2-(1+a*a))./((1+a*a)-2*a*cos(omg)) + var_v;
% Magnitude plot of the optimal filter
C_o = freqz([c_o1,c_o2],1,omg);
magC_o = abs(C_o);

subplot(2,2,1);
plot(omg/pi,Rx,’g’); axis([0,1,0,12]);
%xlabel(’\omega / \pi’,’fontsize’,label_fontsize);
ylabel(’Magnitudes’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:0.2:1],’ytick’,[0:2:12]);
title(’PSD of x(n) for a = 0.8’,’fontsize’,title_fontsize);

subplot(2,2,2);
plot(omg/pi,magC_o,’g’); axis([0,1,0,0.5]);
%xlabel(’\omega / \pi’,’fontsize’,label_fontsize);
ylabel(’Magnitudes’,’fontsize’,label_fontsize);
104 Statistical and Adaptive Signal Processing - Solution Manual

PSD of x(n) for a = 0.8 Magnitude Response for a = 0.8

12 0.5

10 0.4

Magnitudes

Magnitudes
8
0.3
6
0.2
4

2 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

PSD of x(n) for a = −0.8 Magnitude Response for a = −0.8

12 0.5

10 0.4
Magnitudes

Magnitudes
8
0.3
6
0.2
4

2 0.1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
ω/π ω/π

Figure 6.4b: Spectral plots

set(gca,’xtick’,[0:0.2:1],’ytick’,[0:.1:0.5]);
title(’Magnitude Response for a = 0.8’,’fontsize’,title_fontsize);

% 2. a = -0.8
a = -0.8;
% Computed Parameters
c_o1 = (1+var_v-a*a)/((1+var_v)*(1+var_v)+a*a);
c_o2 = (var_v*a)/((1+var_v)*(1+var_v)+a*a);
% PSD of x(n)
omg = [0:500]*pi/500;
Rx = (2-(1+a*a))./((1+a*a)-2*a*cos(omg)) + var_v;
% Magnitude plot of the optimal filter
C_o = freqz([c_o1,c_o2],1,omg);
magC_o = abs(C_o);

subplot(2,2,3);
plot(omg/pi,Rx,’g’); axis([0,1,0,12]);
xlabel(’\omega / \pi’,’fontsize’,label_fontsize);
ylabel(’Magnitudes’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:0.2:1],’ytick’,[0:2:12]);
title(’PSD of x(n) for a = -0.8’,’fontsize’,title_fontsize);

subplot(2,2,4);
plot(omg/pi,magC_o,’g’); axis([0,1,0,0.5]);
xlabel(’\omega / \pi’,’fontsize’,label_fontsize);
ylabel(’Magnitudes’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:0.2:1],’ytick’,[0:.1:0.5]);
title(’Magnitude Response for a = -0.8’,’fontsize’,title_fontsize);
The plot is shown in Figure 6.4b.
(c) Compute and plot the processing gain of the filter for a = −0.9, −0.8, −0.7, . . . , 0.9 as a function of a
Statistical and Adaptive Signal Processing - Solution Manual 105

Processing gain vs a
0.35

Gain(a)
0.3

0.25
−1 −0.9 −0.8 −0.7 −0.6 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
parameter a

Figure 6.4c: Plot of processing gain

and comment on the results.

% (c) Processing Gain
a = -0.9:0.1:0.9;
PG = 1./(1+var_v+a.*a);
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0604c’);
plot(a,PG,’g’); axis([-1,1,0.25,0.35]);
xlabel(’parameter a’,’fontsize’,label_fontsize);
ylabel(’Gain(a)’,’fontsize’,label_fontsize);
set(gca,’xtick’,[-1:0.1:1],’ytick’,[0.25:.05:0.35]);
title(’Processing gain vs a’,’fontsize’,title_fontsize);
The plot is shown in Figure 6.4c.

6.5 Consider the harmonic process y(n) and its noise observation x(n) given in Example 6.4.1
From Example 6.4.1, y(n) = A cos(ωo n+φ) where φ is uniformly distributed on [0, 2π ]. The noise observation
is x(n) = y(n) + ν(n) with ν(n) ∼ N (0, σ 2ν )

(a) Show that r y (l) = 12 A2 cos ωol

Solving directly, it is shown as follows

r y (l) = E{y(n)y ∗ (n + l)}

= E{ A2 cos(ωo n + φ) cos(ωo (n + l) + φ)
= A2 E{ 12 cos(ωol) + 12 cos(2ωo (n + l) + φ)}
= 1
2
A2 [cos(ωo l) + cos(2ωo (n + l))E[cos φ] − sin(2ω o (n + l))E[sin φ]
= 1
2
A2 cos ωo l

(b) The following MATLAB function will compute an Mth-order optimum FIR filter impulse response h(n)

function [h] = opt_fir(A,f0,var_v,M);

%
% Designs the Mth order optimum filter h(n)
% where y(n) = A cos(2 pi f0 n + phi)
106 Statistical and Adaptive Signal Processing - Solution Manual

Impulse Response h(n)

0.1

0.05

−0.05

−0.1
0 5 10 15 20

Magnitude Response of FIR filter

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5

Figure 6.5: 20th-order optimum FIR filter

% x(n) = y(n) + nu(n)

% nu(n) = Normal(0, var_v)

l=[0:M-1]’;

ry=((A^2)/2)*cos(2*pi*f0.*l); rv=var_v.*[1 ; zeros(M-1,1)];

rx=ry+rv;

Rxx=toeplitz(rx);

h=inv(Rxx)*ry;

(c) Figure 6.5 shows the impulse response of the 20th-order optimum FIR filter for A = 0.5, f o = 0.05, and
σ 2ν = 0.5.
(d) Figure 6.5 shows the impulse and magnitude response. Note that this matches Example 6.4.1

6.6 Consider a “desired” signal s(n) generated by the process s(n) = −0.8w(n − 1) + w(n), where w(n) ∼
WN(0, σ 2w ). This signal is passed through the causal system H (z) = 1 − 0.9z −1 whose output y(n) is
corrupted by additive white noise v(n) ∼ WN(0, σ 2v ). The processes w(n) and v(n) are uncorrelated with
σ 2w = 0.3 and σ 2v = 0.1.

(a) Design a second-order optimum FIR filter that estimates s(n) from the signal x(n) = y(n) + v(n) and
determine co and Po :
% Given parameters
var_w = 0.3;
h1 = [1,-0.8]; lh1 = length(h1);
h = [ 1,-0.9];
var_v = 0.1;

% Computed Parameters
% Autocorrelation of s(n): rs(l) = var_w*h1(l)*h1(-l)
Statistical and Adaptive Signal Processing - Solution Manual 107

rh1 = conv(h1,fliplr(h1)); lrh1 = length(rh1);

rs = var_w*rh1; rs = rs(lh1:end); Ps = rs(1);
% Signal x(n) = h(n)*h1(n)*w(n) + v(n); h2 def h1*h;
h2 = conv(h,h1); lh2 = length(h2);
% Autocorrelation of x(n): rx(l) = var_w*h2(l)*h2(-l)+var_v*delta(l)
rx = var_w*conv(h2,fliplr(h2))+var_v*[zeros(1,lh2-1),1,zeros(1,lh2-1)];
rx = rx(lh2:end);
% Crosscorrelation between x(n) and s(n): rxs(l) = vaw_w*h(l)*h1(l)*h1(-l)
lrh1c = round((lrh1-1)/2)+1;
rxs = var_w*conv(h,rh1); rxs = rxs(lrh1c:end);

% (a) Second-order (i.e., length M = 2) optimal FIR filter design

M = 2;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Second-order (i.e., length M = 2) optimal FIR filter design
Optimal FIR filter coefficients: 0.3255, -0.2793
Optimal error = 0.0709
(b) Plot the error performance surface, and verify that it is quadratic and that the optimum filter points to its
minimum.
% (b) Error performance surface
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0606’);

[co0,co1] = meshgrid([co(1)-1:0.05:co(1)+1],[co(2)-1:0.01:co(2)+1]);
P = Ps - 2*dxs(1)*co0 - 2*dxs(2)*co1 + rx(1)*co0.^2 + 2*rx(2)*co0.*co1...
+ rx(1)*co1.^2;
V = [0:0.05:1]+Po;
contour(co0,co1,P,V);
xlabel(’c_o(1)’,’fontsize’,label_fontsize);
ylabel(’c_o(2)’,’fontsize’,label_fontsize);
set(gca,’xtick’,[co(1)-1:1:co(1)+1],’ytick’,[co(2)-1:1:co(2)+1]);
title(’P6.6 : Error Performance Surface’,’fontsize’,title_fontsize);
grid;
The plot is shown in Figure 6.6.
(c) Repeat part (a) for a third-order filter, and see whether there is any improvement.
% (c) Third-order (i.e., length M = 3) optimal FIR filter design
M = 3;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
108 Statistical and Adaptive Signal Processing - Solution Manual

P6.6 : Error Performance Surface

0.7207

co(2)

−0.2793

−1.2793
−0.6745 0.3255 1.3255
c (1)
o

Figure 6.6: Plot of error surface

co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Third-order (i.e., length M = 3) optimal FIR filter design
Optimal FIR filter coefficients: 0.2775, -0.3893, -0.1303
Optimal error = 0.0579

6.7 Repeat Problem 6.6, assuming that the desired signal is generated by s(n) = −0.8s(n − 1) + w(n).

(a) Design a second-order optimum FIR filter that estimates s(n) from the signal x(n) = y(n) + v(n) and
determine co and Po :
% Given parameters
var_w = 0.3;
n = 0:20;
h1 = (-0.8).^n; lh1 = length(h1);
h = [ 1,-0.9];
var_v = 0.1;

% Computed Parameters
% Autocorrelation of s(n): rs(l) = var_w*h1(l)*h1(-l)
rh1 = conv(h1,fliplr(h1)); lrh1 = length(rh1);
rs = var_w*rh1; rs = rs(lh1:end); Ps = rs(1);
% Signal x(n) = h(n)*h1(n)*w(n) + v(n); h2 def h1*h;
h2 = conv(h,h1); lh2 = length(h2);
% Autocorrelation of x(n): rx(l) = var_w*h2(l)*h2(-l)+var_v*delta(l)
rx = var_w*conv(h2,fliplr(h2))+var_v*[zeros(1,lh2-1),1,zeros(1,lh2-1)];
rx = rx(lh2:end);
% Crosscorrelation between x(n) and s(n): rxs(l) = vaw_w*h(l)*h1(l)*h1(-l)
lrh1c = round((lrh1-1)/2)+1;
rxs = var_w*conv(h,rh1); rxs = rxs(lrh1c:end);
Statistical and Adaptive Signal Processing - Solution Manual 109

% (a) Second-order (i.e., length M = 2) optimal FIR filter design

M = 2;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Second-order (i.e., length M = 2) optimal FIR filter design
Optimal FIR filter coefficients: 0.2941, -0.2493
Optimal error = 0.0587
(b) Plot the error performance surface, and verify that it is quadratic and that the optimum filter points to its
minimum.
% (b) Error performance surface
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0607’);

[co0,co1] = meshgrid([co(1)-1:0.05:co(1)+1],[co(2)-1:0.01:co(2)+1]);
P = Ps - 2*dxs(1)*co0 - 2*dxs(2)*co1 + rx(1)*co0.^2 + 2*rx(2)*co0.*co1...
+ rx(1)*co1.^2;
V = [0:0.05:1]+Po;
contour(co0,co1,P,V);
xlabel(’c_o(1)’,’fontsize’,label_fontsize);
ylabel(’c_o(2)’,’fontsize’,label_fontsize);
set(gca,’xtick’,[co(1)-1:1:co(1)+1],’ytick’,[co(2)-1:1:co(2)+1]);
title(’P6.7 : Error Performance Surface’,’fontsize’,title_fontsize);
grid;
The plot is shown in Figure 6.7.
(c) Repeat part (a) for a third-order filter, and see whether there is any improvement.
% (c) Third-order (i.e., length M = 3) optimal FIR filter design
M = 3;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Third-order (i.e., length M = 3) optimal FIR filter design
Optimal FIR filter coefficients: 0.2769, -0.3269, -0.0723
Optimal error = 0.0552

6.8 Repeat Problem 6.6, assuming that H (z) = 1.

110 Statistical and Adaptive Signal Processing - Solution Manual

P6.7 : Error Performance Surface

0.7507

co(2)

−0.2493

−1.2493
−0.7059 0.2941 1.2941
c (1)
o

Figure 6.7: Plot of error surface

(a) Design a second-order optimum FIR filter that estimates s(n) from the signal x(n) = y(n) + v(n) and
determine co and Po :
% Given parameters
var_w = 0.3;
h1 = [1,-0.8]; lh1 = length(h1);
h = [ 1];
var_v = 0.1;

% Computed Parameters
% Autocorrelation of s(n): rs(l) = var_w*h1(l)*h1(-l)
rh1 = conv(h1,fliplr(h1)); lrh1 = length(rh1);
rs = var_w*rh1; rs = rs(lh1:end); Ps = rs(1);
% Signal x(n) = h(n)*h1(n)*w(n) + v(n); h2 def h1*h;
h2 = conv(h,h1); lh2 = length(h2);
% Autocorrelation of x(n): rx(l) = var_w*h2(l)*h2(-l)+var_v*delta(l)
rx = var_w*conv(h2,fliplr(h2))+var_v*[zeros(1,lh2-1),1,zeros(1,lh2-1)];
rx = [rx(lh2:end),0,0];
% Crosscorrelation between x(n) and s(n): rxs(l) = vaw_w*h(l)*h1(l)*h1(-l)
lrh1c = round((lrh1-1)/2)+1;
rxs = var_w*conv(h,rh1); rxs = [rxs(lrh1c:end),0,0];

% (a) Second-order (i.e., length M = 2) optimal FIR filter design

M = 2;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Statistical and Adaptive Signal Processing - Solution Manual 111

P6.8 : Error Performance Surface

0.9181

co(2)
−0.0819

−1.0819
−0.2021 0.7979 1.7979
c (1)
o

Figure 6.8: Plot of error surface

Second-order (i.e., length M = 2) optimal FIR filter design

Optimal FIR filter coefficients: 0.7979, -0.0819
Optimal error = 0.0798
(b) Plot the error performance surface, and verify that it is quadratic and that the optimum filter points to its
minimum.
% (b) Error performance surface
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0608’);

[co0,co1] = meshgrid([co(1)-1:0.05:co(1)+1],[co(2)-1:0.01:co(2)+1]);
P = Ps - 2*dxs(1)*co0 - 2*dxs(2)*co1 + rx(1)*co0.^2 + 2*rx(2)*co0.*co1...
+ rx(1)*co1.^2;
V = [0:0.05:1]+Po;
contour(co0,co1,P,V);
xlabel(’c_o(1)’,’fontsize’,label_fontsize);
ylabel(’c_o(2)’,’fontsize’,label_fontsize);
set(gca,’xtick’,[co(1)-1:1:co(1)+1],’ytick’,[co(2)-1:1:co(2)+1]);
title(’P6.8 : Error Performance Surface’,’fontsize’,title_fontsize);
grid;
The plot is shown in Figure 6.8.
(c) Repeat part (a) for a third-order filter, and see whether there is any improvement.
% (c) Third-order (i.e., length M = 3) optimal FIR filter design
M = 3;
% Optimal filter
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
112 Statistical and Adaptive Signal Processing - Solution Manual

Po = Ps - dxs’*co;
% Printout
Third-order (i.e., length M = 3) optimal FIR filter design
Optimal FIR filter coefficients: 0.7897, -0.1020, -0.0414
Optimal error = 0.0790

6.9 A stationary process x(n) is generated by the difference equation x(n) = ρx(n − 1) + w(n), where w(n) ∼
WN(0, σ 2w )

(a) Show that the correlation matrix of x(n) is given by

σ 2w
Rx = Toeplitz{1, ρ, ρ 2 , . . . , ρ M−1 }
1 − ρ2
Computing the correlation directly

r x (l) = E{x(n)x ∗ (n − l)} = E{(ρx(n − 1) + w(n))x ∗ (n − l)}

= ρ E{x(n − 1)x ∗ (n − l)} = ρr x (l − 1) = ρ 2r x (l − 2) = . . .
= ρ l r x (0)

Solving for r x (0)

r x (0) = E{ρ 2 x(n − 1)x ∗ (n − 1) + ρx(n − 1)w ∗ (n) + ρ ∗ x ∗ (n)w(n − 1) + w(n − 1)w ∗ (n − 1)}
= ρ 2r x (0) + σ 2w

Simplifying, and combining with the result above yields

σ 2w
r x (l) = ρl
1−ρ 2

Therefore, the correlation matrix is

σ 2w
Rx = Toeplitz{1, ρ, ρ 2 , . . . , ρ M−1 }
1 − ρ2

(b) Show that the Mth-order FLP is given by a1(M) = −ρ, ak(M) = 0 for k > 1 and the MMSE is PM = σ 2w .
f

The FLP, using (6.5.18), is R(n − 1)ao (n) = −r(n). Expanding this relation below
    
1 ρ ··· a1 ρ
σw  ρ 1
2
  a2  σ w  ρ2 
2
  =−  
1−ρ 2 .. .. .. 1 − ρ2 ..
. . . .

It is clearly seen, that the right-hand side of the above equation is simply the first column of the autocor-
relation matrix scaled by −ρ. Therefore a = [−ρ0 · · · 0]T , or in other words, a1(M) = −ρ, ak(M) = 0 for
k > 1.
The MMSE, using (6.5.19), is

σ 2w σ 2w ρ 2
Pof (n) = Px (n) + r f H ao(M) = − = σ 2w
1 − ρ2 1 − ρ2
Statistical and Adaptive Signal Processing - Solution Manual 113

6.10 Using Parseval’s theorem, show that (6.4.18) can be written as (6.4.21) in the frequency domain
Equation (6.4.18) is


M−1 
M−1  M−1
M−1 
∗ ∗
P = E{|e(n)| } = r y (0) −
2
h(k)r yx (k) − h (k)r yx (k) − h(k)r(l − k)h ∗ (l)
k=0 k=0 k=0 l=0

Parseval’s relation (6.4.20) is

∞
  π
1
x1 (n)x2∗ (n) = X 1 (e j ω )X 2∗ (e j ω )dω
n=−∞
2π −π

Applying Parseval’s relation to (6.4.18) and using the fact that h(k) is non-zero over [0, M − 1], we have


M−1  ∞
∗ 1
h(k)r yx (k) = H (e j ω )R ∗yx (e j ω )dω
0
2π −∞


M−1  ∞
1
h ∗ (k)r yx (k) = H ∗ (e j ω )R yx (e j ω )dω
0
2π −∞

and
 M−1
M−1  
M−1 
M−1
h(k)r yx (l − k)h ∗ (k) = h(k) r yx (l − k)h ∗ (l)
k=0 l=0 k=0 l=0

M−1  ∞
1
= h(k) R yx (e j ω )e− j ωk H ∗ (e j ω )dω
k=0
2π −∞
 ∞  M−1


1 − j ωk
= h(k)e R yx (e j ω )H ∗ (e j ω )dω
2π −∞ k=0
 ∞
1
= H (e j ω )R yx (e j ω )H ∗ (e j ω )dω
2π −∞
Substituting these relations in (6.4.18), we obtain (6.4.21).

6.11 By differentiating (6.4.21) with respect to H (e j ω ), derive the frequency response function Ho (e j ω ) of the
optimum filter in terms of R yx (e j ω ) and Rx (e j ω ).
Differentiating with respect to H (e j ω ) or equivalently with respect to H ∗ (e j ω ) [see Appendix B].
∂P
=0
∂ H ∗ (e j ω )
  π 
∂ 1 jω ∗ jω ∗ jω jω jω ∗ jω ∗ jω
0= r y (0) − [H (e )R yx (e ) + H (e )R yx (e ) − H (e )H (e )R yx (e )]dω
∂ H ∗ (e j ω ) 2π −π
Which implies that
∂  
H (e j ω )R ∗yx (e j ω ) + H ∗ (e j ω )R yx (e j ω ) − H (e j ω )H ∗ (e j ω )R ∗yx (e j ω ) = 0
∂ H ∗ (e j ω )
Using (B.13), we obtain

R yx (e j ω ) − Ho (e j ω )Rx (e j ω ) = 0
114 Statistical and Adaptive Signal Processing - Solution Manual

or
R yx (e j ω )
Ho (e j ω ) =
Rx (e j ω )
6.12 A conjugate symmetric linear smoother is obtained from (6.5.12) when M = 2L and i = L. If the process
x(n) is stationary, then, using R̄ J = J R̄ ∗ , show that c̄ = J c̄∗ .
Using (6.5.12)
 
0
R̄ c̄ =  Po(i) 
0
Multiply both sides by the exchange matrix J
   
0 0
J R̄ c̄ = J  Po(i)  =  Po(i) 
0 0
With the second equality due to the symmetry of the MMSE Po(i) (n). Using R̄ J = J R̄ ∗ the above equation
can be written
 
0
J R̄ c̄ = R ∗ J c =  Po(i) 
0
Resulting in J c̄ = c̄∗
¯ be the matrices from the eigen-decomposition of R̄, that is, R̄ = Q̄ 
6.13 Let Q̄ and  ¯ Q̄ H

(a) Substitute R into (6.5.20) and (6.5.27) to prove (6.5.43) and (6.5.44).
Starting with (6.5.20), and solving directly
   f 
1 Po (n)
R̄(n) =
ao (n) 0
   f 
¯ 1 Po (n)
Q̄  Q̄ H
=
ao (n) 0
 ∗

   f  q1,1
 1
M+1
1 ¯ −1 H Po (n) ¯ −1  .
..  ∗
= Q̄  Q̄ = Po (n) Q̄  
f
 = Po (n)
f
q̄i q̄i,1
ao (n) 0 λ̄ i
q M−1,1 i=1

A similar proof for (6.5.44).

(b) Generalize the above result for a j th-order linear signal estimator c( j )(n); that is, prove that

M+1
1
( j)
c (n) = Po( j ) q̄i q̄i,∗ j
i=1
λ̄i
Starting with (6.5.12), and solving directly
 
0
R̄ c̄( j )(n) =  Po(i) 
0
   ∗

0 q1, j  1
M+1
¯ −1 
¯ −1 Q̄ H  Po(i)  = Pof (n) Q̄ 
c̄( j )(n) = Q̄  
..
.

 = Po
( j)
q̄i q̄i,∗ j
0 i=1
λ̄i
q M−1, j
Statistical and Adaptive Signal Processing - Solution Manual 115

6.14 Let R̃(n) be the inverse of the correlation matrix R̄(n) given in (6.5.11).

(a) Using (6.5.12), show that the diagonal elements of R̃(n) are given by
1
 R̃(n)i,i = 1≤i ≤ M +1
P (i) (n)

Given
 
R11 (n) r1 (n) R12 (n)
R̄(n) = E{x̄(n)x̄ H (n)} =  r1H (n) Px (n − 1) r2H (n) 
H
R12 (n) r2 (n) R22 (n)

Let R̃(n) = R̄ −1 (n). Consider (6.5.12)

 
0
R̄(n)c̄o(i) (n) =  Po(i) (n) 
0

and divide by Po(i) (n) to obtain

 
0
c̄o(i) (n)
R̄(n) (i) = 1 
Po (n) 0
or
 
 R̃(n)1,i
   .. 
0  . 
c̄o(i) (n)  
= R̃(n) 1 = 
 
  R̃(n)i,i 

Po(i) (n) 0  .. 
 . 
 R̃(n) M+1,i

Since the ith element of c̄o(i) (n) = 1, comparing the ith row of the above equation, we obtain

1
 R̃(n)i,i = 1≤i ≤ M +1
P (i) (n)

(b) Again consider (6.5.12) above

 
 R̃(n)1,i
   .. 
0  . 
 
c̄o (n) = R̃(n)  Po (n)  = Po (n) 
(i) (i) (i)
 R̃(n)i,i  = r̃i (n)
   R̃(n)
0  ..  i,i
 . 
 R̃(n) M+1,i

6.15 The first five samples of the autocorrelation sequence of a signal x(n) are r(0) = 1, r(1) = 0.8, r(2) =
0.6, r(3) = 0.4, and r(4) = 0.3. Compute the FLP, the BLP, the optimum symmetric smoother, and the
corresponding MMSE.
116 Statistical and Adaptive Signal Processing - Solution Manual

(a) Normal equations method: The matrix R̄ is

 
1 0.8 0.6 0.4 0.3
 0.8 1 0.8 0.6 0.4 
 
R̄ = 
 0.6 0.8 1 0.8 0.6 

 0.4 0.6 0.8 1 0.8 
0.3 0.4 0.6 0.8 1
Following the method outlined in Table 6.3, we obtain the following systems of equations:
FLB:
  f     f   
1 0.8 0.6 0.4 c1 0.8 c1 −0. 892 86
 0.8 1 0.8 0.6   c2f     f   
   = −  0.6  ⇒  c2  =  0 
 0.6 0.8 1 0.8   c3f   0.4   c3f   0. 25 
0.4 0.6 0.8 1 f
c4 0.3 f
c4 −0. 142 86
 
−0. 892 86
  0 
Pof = r x (0) + d(1)T cof = 1 + 0.8 0.6 0.4 0.3 

 = 0. 342 85

0. 25
−0. 142 86
BLP:
  b    b   
1 0.8 0.6 0.4 c1 0.3 c1 −0. 142 86
 0.8 1 0.8 0.6   c2b   0.4   c2b   0. 25 
   = −   = 
 0.6 0.8 1 0.8   c3b   0.6  ⇒  c3b   0 
0.4 0.6 0.8 1 c4b 0.8 c4b −0. 892 86
 
−0. 142 86
  0. 25 
Pob = r x (0) + d(5)T cob = 1 + 0.3 0.4 0.6 0.8 

 = 0. 342 85

0
−0. 892 86
SLS:
  s    s   
1 0.8 0.4 0.3 c1 0.6 c1 0
 0.8 1 0.6 0.4   cs   0.8   cs   −0. 5 
  2  = −   2 = 
 0.4 0.6 1 0.8   c3s   0.8  ⇒  c3s   −0. 5 
0.3 0.4 0.8 1 c4s 0.6 c4s 0
 
0
   −0. 5 
Pos = r x (0) + d(3)T cos = 1 + 0.6 0.8 0.8 0.6  
 −0. 5  = 0. 2
0

(b) Inverse of normal equations method: The inverse of matrix R̄ is

 −1  
1 0.8 0.6 0.4 0.3 2. 916 7 −2. 604 2 0 . 729 17 −. 416 67
 0.8 1 0.8 0.6 0.4   −2. 604 2 5. 182 3 −2. 5 −. 651 04 . 729 17 
   
R̄−1 = 
 0.6 0.8 1 0.8 0.6  =
  0 −2. 5 5.0 −2. 5 0 

 0.4 0.6 0.8 1 0.8   . 729 17 −. 651 04 −2. 5 5. 182 3 −2. 604 2 
0.3 0.4 0.6 0.8 1 −. 416 67 . 729 17 0 −2. 604 2 2. 916 7
Statistical and Adaptive Signal Processing - Solution Manual 117

Using the Property 6.5.4, we obtain the following results:

FLB: Divide the first column by 2.9167 to obtain cof and take inverse of 2.9167 to obtain Pof .
 f     
c1 −2. 604 2 −0. 892 86
 cf  1  0   0  1
 2 =     , Pf =
 c3f  2. 916 7  . 729 17  =  0. 25  o
2. 916 7
= 0. 342 85
c4f −. 416 67 −0. 142 86

BLP:Divide the last column by 2. 916 7 to obtain cob and take inverse of 2. 916 7 to obtain Pob .
 b     
c1 −. 416 67 −0. 892 86
 cb  1  0. 729 17   0  1
 2b  =  =  , Pb = = 0. 342 85
 c3  2. 916 7  0   0. 25  o
2. 916 7
c4b −2. 604 2 −0. 142 86
SLS:Divide the middle column by 5 to obtain cos and take inverse of 5 to obtain Pos .
 s     
c1 0 0
 cs  1  −2.5   −0. 5  1
 2s  =    
 c3  5  −2.5  =  −0. 5  , Po = 5 = 0. 2
b

c4s 0 0

6.16 For the symmetric, Toeplitz autocorrelation matrix R = Toeplitz{r(0), r(1), r(2)} = r(0)× Toeplitz{1, ρ 1 , ρ 2 }
with R = LDL H and D = diag{ξ 1 , ξ 2 , ξ 3 }, the following conditions are equivalent:

• R is positive definite.
• ξ i > 0 for 1 ≤ i ≤ 3.
• |ki | < 1 for 1 ≤ i ≤ 3.

Determine the values of ρ 1 and ρ 2 for which R is positive definite, and plot the corresponding area in the
(ρ 1 , ρ 2 ) plane.
6.17 Prove the first equation in (6.5.45) by rearranging the FLP normal equations in terms of the unknowns
f f
Po (n),a1 (n),. . . ,a M (n) and then solve for Po (n), using Cramer’s rule. Repeat the procedure for the sec-
ond equation

(a) Given
   f 
1 Po (n)
R̄(n) =
ao (n) 0
and using Cramer’s rule for the first element
D1
1=
det R̄(n)
where,
 f 
P0 (n) r f H (n) f
D1 = det = P0 (n) det R(n − 1)
0 R(n − 1)
Hence,

f det R̄(n)
P0 (n) =
det R(n − 1)
118 Statistical and Adaptive Signal Processing - Solution Manual

(b) Given
   
bo (n) 0
R̄(n) =
1 Pob (n)

Using Cramer’s rule for the last element

D M+1
1=
det R̄(n)
where,
 
R(n) 0
D M+1 = det = P0b (n) det R(n)
r b H (n) P0b (n)

Hence,

det R̄(n)
P0b (n) =
det R(n)

6.18 Consider the signal x(n) = y(n) + ν(n) where y(n) is a useful random signal corrupted by noise ν(n). The
processes y(n) and ν(n) are uncorrelated with PSDs
&
jω 1 0 ≤ |ω| ≤ π2
R y (e ) =
0 π2 < |ω| ≤ π

and
&
jω 1 π4 ≤ |ω| ≤ π2
Rν (e ) =
0 0 ≤ |ω| < π4 and π
2
< |ω| ≤ π

(a) Determine the optimum IIR filter and find MMSE

Using 6.6.9,
R yx (z)
Hnc (z) =
Rx (z)

Evaluating on the unit circle z = e j ω results in the following

R yx (e j ω )
Hnc (e j ω ) =
Rx (e j ω )
Using the fact that y(n) and ν(n) are uncorrelated, the above can be re-written as

R (e jω
)  1 0 ≤ |ω| ≤ π4
Hnc (e j ω ) = 1/2 π4 < |ω| ≤ π2
y
=
R y (e j ω ) + Rν (e j ω )  π
0 2
< |ω| ≤ π

This can be broken into the sum of two filters as

Hnc (e j ω ) = Hnc1 (e j ω ) + Hnc2 (e j ω )

where
&
1/2 0 ≤ |ω| ≤ π2
Hnc1 (e j ω ) = π
0 2
< |ω| ≤ π
Statistical and Adaptive Signal Processing - Solution Manual 119

and
&
jω 1/2 0 ≤ |ω| ≤ π4
Hnc2 (e ) = π
0 4
< |ω| ≤ π

Taking the inverse Fourier transform on these two filters results in the optimum IIR filter
sin πn sin πn
h nc (n) = h nc1 + h nc2 = 2
+ 4
2π n 2π n
The MMSE is evaluated as
 π
1 1 1
Po = [R y (e j ω ) − H (e j ω )R ∗yx (e j ω )]dω = (π ) − (π − π /4) = 1/8
2π −π 2π 2π

(b) Determine the third-order optimum FIR filter and MMSE

The optimum filter can be found from


M−1
h o (k)r x (n − k) = r yx (n) 0 ≤n ≤ M −1
k=0

where from part(a) above

sin πn
r yx (n) = r y (n) = 2
2π n
and
sin πn sin πn
r x (n) = r y (n) + r y (n) = 2
+ (cos 3πn
8
) 8
2π n 2π n
Solving directly
    
2 0.3649 −0.0796 1 0.4877
h o = Rx−1 r y =  0.3649 2 0.3649   0.3183  =  0.0689 
−0.0796 0.3649 2 0 0.0068

The MMSE is evaluated as

 

M−1 0.4877
Po = Py − ∗
h o (k)r yx (k) = 1 − [1 0.3183 0]  0.0689  = 0.4904
k=0 0.0068

(c) Determine the non-causal optimum FIR filter defined by

ŷ(n) = h(−1)x(n + 1) + h(0)x(n) + h(1)x(n − 1)

Using a simple substitution, the above equation can be re-written as

ŷ(m − 1) = h(0)x(m) + h(1)x(m − 1) + h(2)x(m − 2)

resulting in an optimal filter equation of


2
h o (k)r x (m − k) = r yx (m − 1) 0≤m ≤ M−1
k=0
120 Statistical and Adaptive Signal Processing - Solution Manual

Due to the real values of x(n) and y(n), the autocorrelation is symmetric and r y (l) = r y (−l). Therefore
the optimal filter equation is
 −1    
r x (0) r x (1) r x (2) r y (1) 0.0760
h o = Rx−1r y =  r x (1) r x (0) r x (1)   r y (0)  =  0.4723 
r x (2) r x (1) r x (0) r y (1) 0.0760

The MMSE is evaluated as

 

2 0.0760
Po = Py − ∗
h o (k)r yx (k − 1) = 1 − [0.3183 1 0.3183]  0.4723  = 0.4793
k=0 0.0760

6.19 Consider the ARMA(1,1) process x(n) = 0.8x(n − 1) + ω(n) + 0.5ω(n − 1), where ω(n) ∼ W G N (0, 1).

(a) Determine the coefficients and the MMSE of (1) the one-step ahead FLP x̂(n) = a1 x(n − 1) + a2 x(n − 2)
and (2) the two-step ahead FLP x̂(n + 2) = a1 x(n − 1) + a2 x(n − 2).
Using (6.5.20)
   f 
1 Po (n)
R̄(n) =
ao (n) 0

where
 
  r (0) = 8.25 [r (1) r (2)] = [7.1 5.68]
Px (n) r (n)
fH  x     x x 
R̄(n) = =  r x (1) 7.1 r x (0) = 8.25 r x (1) = 7.1 
r f (n) R(n − 1) =
r x (2) 5.68 r x (1) = 7.1 r x (0) = 8.25

Using the inverse of R̄ to find the filter coefficients and MMSE

   f 
1 Po (n)
= R̄(n)−1
ao (n) 0

Therefore, ao = [−1.0337 0.2011]T and Po = 2.0532

Repeating the above step for the two-step ahead filter, where
 
 r x (0) =8.25   [r x (2) r x (3)] = [5.68 4.54] 
R̄(n) =  r x (2) 5.68 r x (0) = 8.25 r x (1) = 7.1 
=
r x (3) 4.54 r x (1) = 7.1 r x (0) = 8.25

Resulting in ao = [−0.8269 0.1609]T and Po = 4.289

(b) Check if the obtained prediction error filters are minimum-phase, and explain your findings.
For the filters to be minimum phase, the zeros of the filter transfer function must fall within the unit circle.
For a second order filter, the magnitude of the zero vector is
 
 a2 
 <1
a 
1

Clearly, for both filters above, the zero in within the unit circle and therefore minimum phase.

6.20 Consider a random signal x(n) = s(n) + ν(n), where ν(n) ∼ WGN(0, 1) and s(n) is the AR(1) process
s(n) = 0.9s(n − 1) + w(n), where w(n) ∼ WGN(0, 0.64). The signals s(n) and ν(n) are uncorrelated.
Statistical and Adaptive Signal Processing - Solution Manual 121

(a) Determine and plot the autocorrelation rs (l) and the PSD Rs (e j ω of s(n).
Start with

w(n) = s(n) ∗ h(n)

where,

h(n) = (0.9)n u(n)

Hence

rs (l) = h(l) ∗ h(−l) ∗ rw (l) = 0.64[h(l) ∗ h(−l)]

Using,

a |l|
[al u(n)] ∗ [a −l u(n)] = |a| < 1
1 − a2
and
0.64 64
rs (l) = (0.9)|l| = (0.9)|l|
0.19 19
The PSD is
∞
∞ −∞

 64  
Rs (e j ω ) = rs (l)e− j ωl = (0.9)l e− j ωl + (0.9)−l e− j ωl
−∞
19 0 −1
& − jω '
64 1 0.9e
= +
19 1 − 0.9e − j ω 1 − 0.9e− j ω

Therefore,
0.64
Rs (e j ω ) =
1.81 − 1.8 cos ω
The plot is obtained using the Matlab script and is shown in Figure 6.20a.
% (a) Plot of the PSD R_s
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,4,2]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0620b’);

omg = [0:1:500]*pi/500;
Rs = (0.64)./(1.81-1.8*cos(omg));
plot(omg/pi,Rs,’g’); axis([0,1,0,70]);
xlabel(’\omega / \pi’,’fontsize’,label_fontsize);
ylabel(’R_s(e^{j\omega})’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:0.2:1],’ytick’,[0:10:70]);
title(’PSD of s(n)’,’fontsize’,title_fontsize);

(b) Design a second-order optimum FIR filter to estimate s(n) from x(n). What is the MMSE?
122 Statistical and Adaptive Signal Processing - Solution Manual

PSD of s(n)
70

60

50

40

Rs(ejω)
30

20

10

0
0 0.2 0.4 0.6 0.8 1
ω/π

Figure 6.20a: plot of Rs (e j ω )

% (b) Second-order (i.e., length M = 2) optimal FIR filter design

%% Given parameters
var_w = 0.64;
a = 0.9;
var_v = 1.0;

%% Computed Parameters
% Autocorrelation of s(n): rs(l) = (var_w/(1-a*a))*a^|l|
l = 0:100; rs = (var_w/(1-a*a))*a.^l;
Ps = rs(1);
% Signal x(n) = s(n) + v(n);
% Autocorrelation of x(n): rx(l) = rs(l)+var_v*delta(l)
rx = rs + var_v*[1,zeros(1,100)];
% Crosscorrelation between x(n) and s(n): rxs(l) = rs(l)
rxs = rs;

% Optimal filter
M = 2;
Rx = toeplitz(rx(1:M),rx(1:M));
dxs = rxs(1:M)’;
co = Rx\dxs;
% Optimal error
Po = Ps - dxs’*co;
% Printout
Second-order (i.e., length M = 2) optimal FIR filter design
Optimal FIR filter coefficients: 0.5584, 0.3064
Optimal error = 0.5584
(c) Design an optimum IIR filter to estimate s(n) from x(n). What is the MMSE?
Start with
Rs x (e j ω )
Ho (e j ω ) =
Rx (e j ω )
Statistical and Adaptive Signal Processing - Solution Manual 123

Now
0.64
Rx (e j ω ) = Rs (e j ω ) + Rν (e j ω ) = +1
1.81 − 1.8 cos ω
2.45 − 1.8 cos ω
=
1.81 − 1.8 cos ω
and
0.64
Rs x (e j ω ) = Rs (e j ω ) =
1.81 − 1.8 cos ω
Thus
0.64
Ho (e j ω ) =
2.45 − 1.8 cos ω
or

h o (n) = 0.3851(0.4377)|l|

a non-causal IIR filter.

The optimum error is given by
∞
 ∞

Po = Ps − h o (k)rs∗x (l) = Ps − h o (k)rs (l)
−∞ −∞
∞
0.64 0.64
= − (0.3851) [(0.9)(0.4377)]|l| = 0.5136
0.19 0.19 −∞

6.21 A useful signal s(n) with PSD Rs (z) = [(1 − 0.9z −1 )(1 − 0.9z)]−1 is corrupted by additive uncorrelated noise
ν(n) ∼ WN(0, σ 2v ). Choose σ 2v = 1.

(a) The resulting signal x(n) = s(n) + ν(n) is passed through a causal filter with system function H (z) =
(1 − 0.8z −1 )−1 . Determine (1) the SNR at the input, (2) the SNR at the output, and (3) the processing
gain, that is, the improvement in SNR.
1. the SNR at the input: The autocorrelation of the input is

1
rs (l) = Z −1 [Rs (z)] = (0.9)|l| = 5. 263 2 (0.9)|l|
0.19
Hence the input power is rs (0) = 5. 263 2. The noise power is rv (0) = σ 2v = 1. Hence

SNRI = 5. 263 2

2. the SNR at the Output: Let the signal component at the output of the filter due to s(n) be ys (n) and
that due to noise be yv (n). Then
1
R ys (z) = H (z)H (z −1)Rs (z) =  
(1 − 0.8z) 1 − 0.8z −1 (1 − 0.9z) 1 − 0.9z −1
79. 365z 169. 17z 169. 17z 79. 365z
= − + − +
z − .8 z − .9 z − 1. 111 1 z − 1. 25
or

r ys (l) = 169. 17 (0.9)|l| − 79. 365 (0.8)|l|

124 Statistical and Adaptive Signal Processing - Solution Manual

Thus the signal power at the output of the filter is 169. 17 − 79. 365 = 89. 805. Similarly,
1
R yv (z) = H (z)H (z −1)Rv (z) = 
(1 − 0.8z) 1 − 0.8z −1
2. 777 8z 2. 777 8z
= −
z − .8 z − 1. 25
or

r yv (l) = 2. 777 8 (0.8)|l|

Thus the noise power at the output of the filter is 2. 777 8. Hence SNRO = 89. 805
2. 777
= 32. 339.
3. the processing gain:

SNRO 32. 339

PG = = = 6. 144 6
SNRI 5. 263

(b) Determine the causal optimum filter and compare its performance with that of the filter in (a).
First
1
Rx (z) = Rs (z) + Rv (z) =  +1
(1 − 0.9z) 1 − 0.9z −1

2.81 − 0.9z − 0.9z −1 (1 − 0. 362 33z) 1 − 0. 362 33z −1
=  = 0. 551 98 
(1 − 0.9z) 1 − 0.9z −1 (1 − 0.9z) 1 − 0.9z −1

Hence σ 2x = 0. 551 98 and

1 − 0. 362 33z −1
Hx (z) =
1 − 0.9z −1
Also
Rs x (z) Rs (z) 1 (1 − 0.9z)
Rsw (z) = = = 
Hx (z) Hx (z )−1
(1 − 0.9z) 1 − 0.9z −1 (1 − 0. 362 33z)
1 1. 483 9 1. 483 9
=  = −1
−
1 − 0.9z −1 (1 − 0. 362 33z) 1 − . 9z 1 − 0. 362 33z

Hence
[Rsw (z)]+ 1 1. 483 9 1 − 0.9z −1
Hc (z) = =
σ 2x Hx (z) 0. 551 98 1 − . 9z −1 1 − 0. 362 33z −1
2.6883
=
1 − 0. 362 33z −1

1. the SNR at the input: This is same as before, that is, SNRI = 5. 263 2
2. the SNR at the Output: Let the signal component at the output of the filter due to s(n) be ys (n) and
that due to noise be yv (n). Then
(2.6883) (2.6883)
R ys (z) = Hc (z)Hc (z −1 )Rs (z) =  
1 − 0. 362 33z −1 (1 − 0. 362 33z) (1 − 0.9z) 1 − 0.9z −1
8. 318 9z 94. 478z 94. 478z 8. 318 9z
= − + − +
z − . 362 33 z − .9 z − 1. 111 1 z − 2. 759 9
Statistical and Adaptive Signal Processing - Solution Manual 125

or

r ys (l) = 94. 478 (0.9)|l| − 8. 318 9 (0.36233)|l|

Thus the signal power at the output of the filter is 94.478 − 8. 318 9 = 86. 159. Similarly,
(2.6883) (2.6883)
R yv (z) = Hc (z)Hc (z −1 )Rv (z) = 
1 − 0. 362 33z −1 (1 − 0. 362 33z)
8. 319 1 8. 319 1
= −
z − . 362 33 z − 2. 759 9
or

r yv (l) = 8.3191 (0. 362 33)|l|

Thus the noise power at the output of the filter is 8.3191. Hence SNRO = 86. 159
8.3191
= 10. 357.
3. the processing gain:

SNRO 10. 357

PG = = = 1. 967 9
SNRI 5. 263

6.22 A useful signal s(n) with PSD Rs (z) = 0.36[(1 − 0.8z −1 )(1 − 0.8z −1 )]−1 is corrupted by additive uncorre-
lated noise ν(n) ∼ WN(0, 1). Determine the optimum noncausal and causal IIR filters, and compare their
performance by examining the MMSE and their magnitude response.
The optimum noncausal filter is
Rs x (z) Rs (z) 0.36
Hnc (z) = = =
Rx (z) Rs (z) + Rν (z) 0.36 + (1 − 0.8z −1 )(1 − 0.8z −1 )
0.225
=
(1 − 0.5z )(1 − 0.5z −1 )
−1

Therefore the optimum noncausal filter is h nc (l) = (0.225)(0.5)|l| . The resulting MMSE is
 
Po = rs (0) − h nc (k)rs x (k) = 1 − (0.225) (0.5)|k| (0.9)|k| = 1 − (0.225)(3.333)
k k
= 0.25

The optimum causal filter is

 
1 R yx (z)
Hc (z) =
σ x Hx (z)
2 Hx (1/z) +

where
    

R yx (z) Rs x (z) 0.36 1 − 0.8z
= =
Hx (1/z) + Hx (1/z) + (1 − 0.8z −1 )(1 − 0.8z) 1 − 0.5z +
 
−0.6 0.6 0.6
= + =
1 − 2z −1 1 − 0.8z −1
+ 1 − 0.8z −1

Also



1 − 0.5z −1 1 − 0.5z
Rx (z) = Rs (z) + Rν (z) = σ 2x Hx (z)Hx (z −1 ) = (1.6)
1 − 0.8z −1 1 − 0.8z
126 Statistical and Adaptive Signal Processing - Solution Manual

Therefore,

1 − 0.5z −1
Hx (z) =
1 − 0.8z −1
and



1 − 0.5z −1 0.6 0.375
Hc (z) = =
1.6(1 − 0.8z −1 ) 1 − 0.8z −1 1 − 0.5z −1
The optimum causal filter is

h c (l) = 0.375(0.5)l u(l)

The resulting MMSE is

 ∞

Po = rs (0) − h c (k)rs x (k) = 1 − (0.375) (0.5)k (0.9)k = 1 − (0.375)(1.667)
k k=0
= 0.375

−1
6.23 Consider a process with PSD Rx (z) = σ 2 Hx (z)H
! x (z ). Determine the D-step ahead linear predictor, and
show that the MMSE is given by P (D) = σ 2 n=0 h 2x (n). Check your results by using the PSD Rx (z) =
D−1

(1 − a 2 )[(1 − az −1 )(1 − az)]−1 .

Consider the desired response

y(n) = x(n + D)

then
∞
 [D]
e (n + D) = x(n + 1) −
f
h lp (k)x(n − k)
k=0

[D]
where h lp (n) is the impulse response of the D-step ahead predictor. Thus

r yx (l) = E{y(n)x ∗ (n − l)} = E{x(n + D)x ∗ (n − l)} = r x (l + D)

or

R yx (z) = z D Rx (z)

Thus,
 D 2 
1 z σ x Hx (z)Hx (1/z) [z D Hx (z)]+
Hlp[D] (z) = 2 =
σ x Hx (z) Hx (1/z) + Hx (z)

where

[z D Hx (z)]+ = h(D) + h(D + 1)z −1 + h(D + 2)z −2 + · · ·

Consider,

z D Hx (z) = z D h(0) + h(1)z D−1 + h(2)z D−2 + · · · + h(D)z 0 + h(D + 1)z −1 + · · · h(0) = 1
Statistical and Adaptive Signal Processing - Solution Manual 127

Hence,

[z D Hx (z)]+ = z D Hx (z) − [z D h + h(1)z D−1 + h(2)z D−2 + · · · + h(D − 1)z]

= z D Hx (z) − z D [1 + h(1)z −1 + · · · + h(D − 1)z 1−D ]

or
!   ! D−1 
−k
[z D Hx (z)]+ D−1
k=0 h(k)z h(k)z −1
= zD − zD =z D
1− k=0
Hx (z) Hx (z) Hx (z)

Substituting in Po[D]
2  ! D−1 
1 h(k)z −1
Po[D] = [Rx (z) − z D
1− k=0
z −D Rx (z)]z −1 dz
2π j Hx (z)
2  ! D−1  2 
D−1
1 h(k)z −1 −1 1
= Rx (z) k=0
z dz = σ 2x Hx (1/z) h(k)z −1 z −1 dz
2π j Hx (z) 2π j k=0
2 
D−1
σ 2x
= Hx (1/z) h(k)z −1 z −1 dz
2π j k=0

The inverse z-transform is given by

2
1
h(n) = H (z)z n−1dz
2π j

If we substitute z → z −1 = 1/z then

H (z) → H (z −1) and z n−1 → z −n+1

Also p = z −1 so dp = 1z −2 dz or dz = − p−2 dp and

2 2
1 −1 −n+1 −2 1
h(n) = H(p )p p dp = H ( p−1 ) p−n−1 dp
2π j 2π j
Hence

D−1
Po[D] = h 2 (k)
k=0

To check, consider

1 − a2
Rx (z) =
(1 − az −1 )(1 − az)
with
1 1
σ 2x = (1 − a 2 ), Hx (z) = , Hx (z −1 ) = , h x (k) = a k u(k)
1 − az −1 1 − az
Hence

[z D Hx (z)]+ = a D + a D+1 z −1 + · · · = a D [1 + az −1 + · · · ] = a D Hx (z)

128 Statistical and Adaptive Signal Processing - Solution Manual

and
H0 (z) = a D ⇒ h 0 (k) = a D δ(k)
The MMSE is
 
∗ ∗
Po = r y (0) − h 0 (k)r yx (k) = r x (0) − a D δ(k)r yx (k)
k
 ∞

∗
= r x (0) − a D r yx (0) = σ 2x h x (k)2 − σ 2x h x (k)2
D

D−1
= σ 2x h x (k)2
k=0

6.24 Let x(n) = s(n) + ν(n) with Rν (z) = 1, Rsν (z) = 0, and
0.75
Rs (z) =
(1 − 0.5z −1 )(1 − 0.5z)
Determine the optimum filters for the estimation of s(n) and s(n − 2) from {x(k)}n−∞ and the corresponding
MMSEs.
Start with
   
1 Rs x (z) 1 Rs (z)
Hc (z) = = 2
σ 2x Hx (z) Hx (1/z) σ x Hx (z) Hx (1/z)
Compute the autocorrelation Rx (z)
0.75
Rx (z) = Rs (z) + Rν (z) = +1
(1 − 0.5z −1 )(1 − 0.5z)



1 − 0.2679z −1 1 − 0.2679z
= σ 2x Hx (z)Hx (z −1 ) = (1.866)
1 − 0.5z −1 1 − 0.5z
Therefore,
1 − 0.2679z −1
Hx (z) =
1 − 0.5z −1
Substituting into the above equation for Hc (z)




1 1 − 0.5z −1 0.8659
Hc (z) =
1.866 1 − 0.2679z −1 1 − 0.5z −1


1
= 0.464
1 − 0.2679z −1
Therefore, h c (l) = 0.464(0.2679)l u(l), and is shown in Figure 6.24. The resulting MMSE is
 ∞

Po = rs (0) − h c (k)rs x (k) = 1 − 0.464(0.2679)k (0.70)(0.5)k = 0.5982
k k=0

For s(n − 2),

R yx (z) 0.75z −3
R !yw (z) = z D R yw (z) = z −2 =
H (1/z) −0.2679 + 1.134z −1 − 0.5z −2
 
 !  0.062 + 0.201z −1 + 0.75z −2
R yw (z) + =
1 − 0.5z −1
Statistical and Adaptive Signal Processing - Solution Manual 129

Optimum Filter Coefficients Optimum Smoothing Filter Coefficients (D=−2)

0.5 0.45

0.45 0.4

0.4
0.35

0.35 MMSE = 0.5982 MMSE = 0.5853

0.3
0.3
0.25
0.25
0.2
0.2
0.15
0.15

0.1
0.1

0.05 0.05

0 0
−1 0 1 2 3 4 5 6 7 8 9 −1 0 1 2 3 4 5 6 7 8 9

Figure 6.24: Optimum Estimation Filters

Plugging this into the above equation for Hc (z)

 
1 − 0.2679z −1 0.062 + 0.201z −1 + 0.75z −2
Hc (z) = (1.866)
1 − 0.5z −1 1 − 0.5z −1
and the impulse response is

h c (l) = −1.5δ(l − 1) − 6δ(l) + 6.036(0.2679)l u(l)

The resulting MMSE is

 ∞

Po = rs (0) − h c (k)rs x (k) = 0.75 − h c (k)rs x (k) = 0.5853
k k=0

6.25 For the random signal with PSD



1 − 0.2z −1 1 − 0.2z
Rx (z) =
1 − 0.9z −1 1 − 0.9z
determine the optimum two-step ahead linear predictor and the corresponding MMSE.
First find the autocorrelation
1.04 − 0.2z −1 − 0.2z −3.3567 3.3567
Rx (z) = −1
= −1
+ + 0.222
1.81 − 0.9z − 0.9z 1 − 1.11z 1 − 0.9z −1
r x (l) = 3.3567(0.9)|l| + δ(l)(0.222)

The two-step ahead linear predictor can be found directly using (6.5.18)

Rx (n − 1)ao = −r xf (n) ⇒ ao = {0.6301, 0.1259, 0.0252, 0.0050, 0.0010, 0.0002, 0 · · · }

The corresponding MMSE is

Pof (n) = Px (n) + r f H (n)ao (n) = 1.4897

130 Statistical and Adaptive Signal Processing - Solution Manual

Predictor coefficients
0.7

0.6

0.5 MMSE = 1.49

0.4

0.3

0.2

0.1

0
−2 0 2 4 6 8 10

Figure 6.25: Two-Step Ahead Linear Predictor

6.26 Repeat Problem 6.25 for

1
Rx (z) =
(1 − 0.2z −1 )(1 − 0.9z −1 )(1 − 0.2z)(1 − 0.9z)
First find the autocorrelation
1
Rx (z) =
(1.04 − 0.2z −1 − 0.2z)(1.81 − 0.9z −1 − 0.9z)
z −2
=
0.18 − 1.298z −1 + 2.242z −2 − 1.298z −3 + 0.18z −4
0.0653 −1.4854 1.4854 −0.0653
= + +
1 − 5z −1 1 − 1.11z 1 − 0.9z
−1 −1 1 − 0.2z −1
|l| |l|
r x (l) = 1.4854(0.9) − 0.0653(0.2)

Rx (n − 1)ao = −r xf (n)

Figure 6.27 shows the filter coefficients

The corresponding MMSE is

Pof (n) = Px (n) + r f H (n)ao (n) = 0.7403

6.27 Let x(n) = s(n) + ν(n) with ν(n) ∼ WN(0, 1) and s(n) = 0.6s(n − 1) + w(n), where w(n) ∼ WN(0, 0.82).
The process s(n) and ν(n) are uncorrelated.
Determine the optimum filters for the estimation of s(n), s(n + 2), and s(n − 2) from {x(k)}n−∞ and the
corresponding MMSEs.
The optimum filter equation is

Co = Rx−1 d

where Rx = Rs + Rν . The autocorrelation of the signal is symmetric, and each lag is found using

rs (l) = rs (0)(0.6)|l|
Statistical and Adaptive Signal Processing - Solution Manual 131

Predictor coefficients
0.6

0.5

0.4 MMSE = 0.7403

0.3

0.2

0.1

−0.1
−2 0 2 4 6 8 10

Figure 6.26: Two-Step Ahead Linear Predictor

where
σ 2w σ 2w
rs (0) = =
1 − ρ2 1 − (0.6)2
and rν (l) = σ 2ν δ(l). The same optimum filter equation is used for the three cases: filtering s(n), estimation
s(n + 2) or smoothing s(n − 2). The only change is the value of the cross-correlation vector between the
desired response and the input data vector. The value for this vector is shown below for the three cases Filtering
 
1
 (0.6) 
 
s(n) : d(l) = E{x(n)s ∗ (n + l)} = rs (l) =  (0.6)2 
 
..
.
Estimation
 
(0.6)2
 (0.6)3 
 
s(n + 2) : d(l) = E{x(n)s ∗ (n + l)} = rs (l + 2) =  (0.6)4 
 
..
.
Smoothing
 
(0.6)2
 (0.6) 
 
 1 
 
s(n − 2) : d(l) = E{x(n)s ∗ (n + l)} = rs (l − 2) =  (0.6) 
 
 (0.6)2 
 
..
.

The MMSE is found directly using the filter coefficients and the appropriate cross-correlation vector
Po = Py − d H Co = rs (0) − d H Co
The filter coefficients are found, using MATLAB, and shown in Figure 6.27, along with the appropriate MMSE
values.
132 Statistical and Adaptive Signal Processing - Solution Manual

0.6

0.4 Optimum filter coefficients for s(n)

0.2 MMSE = 0.5

0
0 1 2 3 4 5 6 7 8 9 10
0.6

0.4 Optimum filter coefficients for s(n+2)

0.2 MMSE = 1.18

0
0 1 2 3 4 5 6 7 8 9 10
0.6

0.4 Optimum filter coefficients for s(n−2)

0.2 MMSE = 0.4509

0
0 1 2 3 4 5 6 7 8 9 10

Figure 6.27: Optimum Filter Coefficients

6.28 A random sequence s(n) with PSD Rs (z) = 1−0.5z −11 (1−0.5z) is corrupted by noise sequence v(n) with PSD
( )
Rv (z) = 5 to obtain the observed sequence x(n) = s(n) + v(n). Also Rsv (z) = 0.
$ %
Optimal causal filter to estimate s(n) from {x(n)}n−∞ : First, rs (l) = Z −1 1−0.5z −11 (1−0.5z) = 43 (0.5)|l| . Now
( )

1 1 − 0.4z −1 (1 − 0.4z)
Rx (z) =  + 5 = 6.25 
1 − 0.5z −1 (1 − 0.5z) 1 − 0.5z −1 (1 − 0.5z)
and Rs x (z) = Rs (z). Thus after spectral factorization

1 − 0.4z −1
σ 2x = 6.25 and Hx (z) =  , |z| > 0.5
1 − 0.5z −1
Hence
Rs x (z) 1 (1 − 0.5z) 1
Rsw (z) = = =
Hx (z −1 ) 1 − 0.5z −1 (1 − 0.5z) (1 − 0.4z) 1 − 0.5z −1 (1 − 0.4z)
1. 25 1. 25
= −
1 − 0. 5z −1 1 − 2. 5z −1
or [Rsw (z)]+ = 1. 25
1−0. 5z −1
. Finally

[Rsw (z)]+ 1 1. 25 1 − 0.5z −1 0.2
Hc (z) = 2 =  = , |z| > 0.4
σ x Hx (z) 6.25 1 − 0. 5z −1 1 − 0.4z −1 1 − 0.4z −1
and h c (n) = 0.2 (0.4)n u(n). The MMSE is given by
∞


∞
4 1 4
Pc = rs (0) − h c (k)rs x (k) = − (0.4)k (0.5)k = 1
k=0
3 5 3 k=0

Optimal prediction filter to estimate s(n + 2) from {x(n)}n−∞ : The whitening part of the causal filter does not
depend on s(n + 2) and is given by Hx (z). The coloring part depends on s(n + 2) and is given by z 2 Rsw (z).
Thus
[Rsw (z)]+ 0.2 (0.5)2 0.0 5
Ho[2] (z) = (0.5) 2
=  = , |z| > 0.4
σ x Hx (z)
2 1 − 0.4z −1 1 − 0.4z −1
Statistical and Adaptive Signal Processing - Solution Manual 133

o (n) = 0.05 (0.4) u(n). The MMSE is given by

n
and h [2]
∞

 ∞
4 4
Po[2] = rs (0) − o (k)rs x (k
h [2] + 2) = − (0.05) (0.5)2 (0.4)k (0.5)k
k=0
3 3 k=0
= 1.3132 > Po[0] = Pc = 1

Optimal smoothing filter to estimate s(n − 2) from {x(n)}n−∞ : Once again the whitening part of the causal
filter does not depend on s(n + 2) and is given by Hx (z). The coloring part is obtained from
 [−2] 
rsw (l) + , rsw (l − 2)u(l)

From Rsw (z) above, we have rsw (l) = 1.25 (0.5)l u(l) + 1.25 (2.5)l u(−l − 1). Hence
 [−2]  
rsw (l) + = rsw (l − 2)u(l) = 1.25 (0.5)l−2 u(l − 2) + (2.5)l−2 u(−l + 2 − 1) u(l)
= 1.25 (0.5)l−2 u(l − 2) + 1.25 {.4δ(l) + δ(l − 1)}

or
  "  # 5 z −2 1 5
[−2]
Rsw (z) +
= Z [−2]
rsw (l) + = + + z −1
 41− 0.5z
−1 2 4
−1 −2
1 1 + 2z + 1.25z
= 
2 1 − 0.5z −1
Finally
   
[−2]
Rsw (z) + 1 1 − 0.5z −1 1 1 + 2z −1 + 1.25z −2
Ho[−2] (z) = =  
σ 2x Hx (z) 6.25 1 − 0.4z −1 2 1 − 0.5z −1

1 + 2z −1 + 1.25z −2
= 0.08  , |z| > 0.4
1 − 0.4z −1
 
1 − 0.4z −1 + 2.4z −1 1 − 0.4z −1 + 2.21z −2
= 0.08  , |z| > 0.4
1 − 0.4z −1
1
= 0.08 + 0.192z −1 + 0.1768z −2  , |z| > 0.4
1 − 0.4z −1
or

h [−2]
o (n) = 0.08δ(n) + 0.192δ(n − 1) + 0.1768 (0.4)n−2 u(n − 2)

To compute the MMSE, we first express rs x (l − 2)u(l) in the form

1 2 4
rs x (l − 2)u(l) = rs (l − 2)u(l) = δ(l) + δ(l − 1) + (0.5)l−2 u(l − 2)
3 3 3
Then
∞

Po[−2] = rs (0) − h [−2]
o (k)rs x (k − 2)
k=0



∞
4 1 2 4
= − (0.08) − (0.192) − (0.1768) (0.4)k (0.5)k
3 3 3 3 k=0
= 0.884 < Po[0] = Pc = 1
134 Statistical and Adaptive Signal Processing - Solution Manual

6.29 Consider the random sequence x(n) generated in Example 6.5.2.

1
x(n) = w(n) + w(n − 1)
2
where w(n) is WN(0, 1). Generate K sample functions {wk (n)}n=0
N
, k = 1, . . . , K of w(n), in order to generate
K sample functions {xk (n)}n=0 , k = 1, . . . , K of x(n).
N

% Generate K samples functions of x(n)

K = 1000; N = 100;
x = zeros(N,K);
w = randn(N,K);
for k = 1:K; x(:,k) = filter([1,0.5],1,w(:,k)); end;

(a) Use the second-order FLP ak to obtain predictions {x̂kf (n)}n=2

N
of xk (n), for k = 1, . . . , K . Then determine
the average error

1 
N
P̂ f = |xk (n) − x̂kf (n)|2 k = 1, . . . , K
N − 1 n=2

and plot it as a function of k. Compare it with Pof .

% (a) Second-order forward linear predictor
M = 2; Rx = toeplitz([5/4,1/2,0]);
[a,Pfo]=olsigest(Rx,1);
a = [1;a];
ef = zeros(N,K);
for k = 1:K; ef(:,k) = filter(a,1,x(:,k)); end;
Pf = sum(ef(M+1:N,:) .*ef(M+1:N,:),1)/(N-M);
subplot(3,1,1);
plot(1:K,Pf,’g’,[0,K],[Pfo,Pfo],’r--’);
axis([0,K,0,2]);
%xlabel(’sample function index k’,’fontsize’,label_fontsize);
ylabel(’MMSE’,’fontsize’,label_fontsize);
title(’MMSE of a Second-order FLP’,’fontsize’,title_fontsize);
text(700,1.7,[’Optimum P^f = ’,num2str(Pfo)],’fontsize’,text_fontsize);
The plot is shown in Figure 6.29.
(b) Use the second-order BLP bk to obtain predictions {x̂kb (n)}n=0
N−2
, k = 1, . . . , K of xk (n). Then determine
the average error

1 
N−2
P̂ b = |xk (n) − x̂kb (n)|2 k = 1, . . . , K
N − 1 n=0

and plot it as a function of k. Compare it with Pob .

% (b) Second-order backward linear predictor

[b,Pbo]=olsigest(Rx,3);
b = [b;1];
eb = zeros(N,K);
for k = 1:K; eb(:,k) = filter(b,1,x(:,k)); end;
Statistical and Adaptive Signal Processing - Solution Manual 135

MMSE of a Second−order FLP

2

Optimum Pf = 1.0119
1.5

MMSE
1

0.5

0
0 100 200 300 400 500 600 700 800 900 1000
MMSE of a Second−order BLP
2

Optimum Pb = 1.0119
1.5
MMSE

0.5

0
0 100 200 300 400 500 600 700 800 900 1000
MMSE of a Second−order SLP
2

Optimum Ps = 0.85
1.5
MMSE

0.5

0
0 100 200 300 400 500 600 700 800 900 1000
sample function index k

Figure 6.29: Plots of average errors vs sample function index k

Pb = sum(eb(1:N-M,:) .*eb(1:N-M,:),1)/(N-M);
subplot(3,1,2);
plot(1:K,Pb,’g’,[0,K],[Pbo,Pbo],’r--’);
axis([0,K,0,2]);
%xlabel(’sample function index k’,’fontsize’,label_fontsize);
ylabel(’MMSE’,’fontsize’,label_fontsize);
title(’MMSE of a Second-order BLP’,’fontsize’,title_fontsize);
text(700,1.7,[’Optimum P^b = ’,num2str(Pbo)],’fontsize’,text_fontsize);

The plot is shown in Figure 6.29.

(c) Use the second-order symmetric linear smoother ck to obtain smooth estimates {x̂kc (n)}n=0
N−2
of xk (n) for
k = 1, . . . , K . Determine the average error

1 
N−1
P̂ s = |xk (n) − x̂kc (n)|2 k = 1, . . . , K
N − 1 n=1

and plot it as a function of k. Compare it with Pos .

% (c) Second-order Symmetric linear predictor

[c,Pco]=olsigest(Rx,2);
c = [c(1);1;c(2)];
ec = zeros(N,K);
for k = 1:K; ec(:,k) = filter(c,1,x(:,k)); end;
Pc = sum(ec(2:N-M+1,:) .*ec(2:N-M+1,:),1)/(N-M);
subplot(3,1,3);
plot(1:K,Pc,’g’,[0,K],[Pco,Pco],’r--’);
axis([0,K,0,2]);
136 Statistical and Adaptive Signal Processing - Solution Manual

xlabel(’sample function index k’,’fontsize’,label_fontsize);

ylabel(’MMSE’,’fontsize’,label_fontsize);
title(’MMSE of a Second-order SLP’,’fontsize’,title_fontsize);
text(700,1.7,[’Optimum P^s = ’,num2str(Pco)],’fontsize’,text_fontsize);
The plot is shown in Figure 6.29.

6.30 Let x(n) = y(n) + v(n) be a wide-sense stationary process. The linear, symmetric smoothing filter estimator
of y(n) is given by

L
ŷ(n) = h(k)x(n − k)
k=−L

(a) Determine the normal equations for the optimum MMSE filter.
(b) Show that the smoothing filter cos has linear phase.
(c) Use the Lagrange multiplier method to determine the MMSE Mth-order estimator ŷ(n) = c H x(n), where
M = 2L + 1, when the filter vector c is constrained to be conjugate symmetric, that is, c = Jc∗ . Compare
the results with those obtained in part (a).

6.31 Consider the causal prediction filter discussed in Example 6.6.1.

(a) Determine h [D]
c (n).
Using equation 6.6.42
R yx (z) 3
0.3z
R yw (z) = −1
= 5
−1
+
Hx (z ) 1 − 5z
4
1 − 12 z
The causal part is
 
R yx (z) 3
[R yw (z)]+ = = 5
Hx (z −1 ) + 1 − 45 z −1
Using 6.6.21 and the relation Hc[D] (z) = z D Hc (z), the optimal predictive filter is
 D 
[D] 1 z R yx (z) 3
Hc (z) = 2 =z D 5
σ x Hx (z) Hx (z −1 ) + 1 − 45 z −1
Therefore, taking the inverse z-transform of the above equation result in
h [D]
c (n) = 8 ( 5 ) ( 2 ) u(n)
3 4 D 1 n

(b) Using the above h [D]

c (n), show that

Pc[D] = 1 − 58 ( 45 )2D
Solving directly using the following formula
∞

Pc[D] = r y (0) − h [D]
c (k)r yx (k)
k=0
 ∞
= 1− 3
8
( 12 )k ( 45 ) D ( 45 )k ( 45 ) D
k=0
∞
= 1− (
3 4 2
8 5
) D ( 12 45 )k
k=0

Evaluating the above equation results in Pc[D] = 1 − 58 ( 45 )2D

Statistical and Adaptive Signal Processing - Solution Manual 137

6.32 Consider the causal smoothing filter discussed in Example 6.6.1.

! !
(a) Using [r yw (l)]+ = r yw (l + D)u(l),D < 0, show that [r yw (l)]+ can be put in the form
!
[r yw (l)]+ = 35 ( 45 l+D u(l + D) + 35 (2l+D )[u(l) − u(l + D)] D<0
Solving directly
!
[r yw (l)]+ = r yw (l + D)u(l) D < 0

= 35 ( 45 )l+D u(l + D) + ( 35 )2l+D u(−l − D − 1) u(l)
= ( 35 )( 45 )l+D u(l + D) + ( 35 )2l+D u(−l − D − 1)u(l)
= ( 35 )( 45 )l+D u(l + D) + ( 35 )(2l+D )[u(l) − u(l + D)] D<0

(b) Hence, show that [R !yw (l)]+ is given by

−D−l

zD
[R !yw (l)]+ = 3
4 −1
+ (2
3 D
) 2l z −1
5
1 − 5z 5
l=0

Taking the z-transform, the above result is found directly

(c) Finally, using (6.6.21), prove (6.6.54)

( ) ( ) −D−1

5 1 − 45 z −1 3 zD 
Hc[D] (z), D < 0 = + 2D 2l z −l
8 1 − 12 z −1 5 1 − 45 z −1 l=0
 
−D−1

4 −1 D l −l
 (1 − z )2 2z 
3 
D 5
 z l=0 
=  + 
8  1 − 12 z −1 1− 1 −1
z 
 2 

 !−D−1 l −l !−D−1 l −l−1 

3 zD 2 D l=0 2z 4 2 D l=0 2z
= 1 −1
+ 1 −1
−( )
8 1 − 2z 1 − 2z 5 1 − 2 z −1
1

6.33 Proof of (6.6.57)

(a) Consider
$ ! %
r yw (l) = r yw (l + D)u(l), D < 0
+

 
3 4 l+D
= u(l + D) + 2l+D u(−l − D − 1) u(l), D < 0
5 5


3 4 l+D 3
= u(l + D) + 2l+D {u(l) − u(l + D)} , D < 0
5 5 5
Hence
$ % −D−1 −D−1

! 3 zD 3 D  l −l 3 zD 3 D  2 l
R yw (z) = + 2 2z = + 2
+ 5 1 − 45 z −1 5 l=0
5 1 − 45 z −1 5 l=0
z
3 zD 3 D 1 − (z/D) D 3 zD 3 2D − z D
= + 2 = +
5 1 − 45 z −1 5 1 − 2z −1 5 1 − 45 z −1 5 1 − 2z −1
138 Statistical and Adaptive Signal Processing - Solution Manual

or
( )
$ !
% 3 zD 2D − z D
R yw (z) = + , D<0
+ 5 1 − 45 z −1 1 − 2z −1

(b) Using (6.6.21) and substituting the above result

( ) ( )
5 1 − 45 z −1 3 zD 2D − z D
Hc (z) =
[D]
+
8 1 − 12 z −1 5 1 − 45 z −1 1 − 2z −1
( )(    )
3 1 − 45 z −1 z D 1 − 2z −1 + 2 D − z D 1 − 45 z −1
=  
8 1 − 12 z −1 1 − 45 z −1 1 − 2z −1
  
3 2 D 1 − 45 z −1 + 35 z D−1
=   , D<0
8 1 − 12 z −1 1 − 2z −1

Hence as −D → ∞ we have 2 D → 0 and z D−1 → z D . Hence

 
9 zD
lim Hc (z) =
[D]
  = z D Hnc (z), D<0
−D→∞ 40 1 − 2 z1 −1 1 − 2z −1

(c) Finally, consider

∞
 ∞

Pc[D] = r y (0) − c (k)r yx (k + D) = r y (0) −
h [D] c (k)r yx (k + D)u(k)
h [D]
k=0 k=−∞

As −D → ∞ we have

c (k) → h c (k) and r yx (k + D)u(k) → r yx (k)

h [D]

Hence
∞

lim Pc[D] = r y (0) − h c (k)r yx (k) = Pnc
−D→∞
k=−∞

6.34 Consider the block diagram of a simple communication system shown in Figure 6.38 in which H1 (z) =
1/(1 + 0.95z −1 ).

(a) Determine a second-order optimum FIR filter (M = 2) that estimates the signal s(n) from the received
signal x(n) = z(n) + ν(n). What is the corresponding MMSE Po ?
From the block diagram, s(n) = −0.95s(n − 1) + w(n) and z(n) = 0.85z(n − 1) + s(n). The optimum
filter is Co = Rz−1rs x . The first two lags of the autocorrelation are

rz (0) = 3.936 σ 2w = 1.181 rz (1) = −2.045 σ 2w = −0.614

and the first two lags of the cross-correlation are

rs x (l) = E{s(n)x ∗ (n + l)} = E{s(n)z ∗ (n + l)}

= E{[z(n) − 0.85z(n − 1)]z ∗ (n + l)} = rz (l) − 0.85rz (l − 1)
rs x (0) = rz (0) − 0.85, rz (1) = 1.702
rs x (1) = rz (1) − 0.85, rz (0) = −1.614

The optimum filter coefficients are shown in the top plot of Figure 6.34.
Statistical and Adaptive Signal Processing - Solution Manual 139

Error Performance Surface

0

−0.2

−0.4

−0.6

c0 = [ 0.94,−0.812]
−0.8

−1

−1.2

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Monte Carlo Simulations (M=2) Monte Carlo Simulations (M=3)

0.185 0.175
Average Square Error = 0.163 Average Square Error = 0.151
MMSE p0 = 0.163 MMSE p0 = 0.14
0.18 0.17

0.175 0.165

0.17 0.16

0.165 0.155

0.16 0.15

0.155 0.145

0.15 0.14

0.145 0.135

0.14 0.13
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

Figure 6.34: Error Performance Surface and Monte Carlo Simulations

140 Statistical and Adaptive Signal Processing - Solution Manual

(b) Plot the error performance surface

The top plot of Figure 6.34 has the error performance surface plot. The surface is defined by

ε2 = σ 2y − 2cT d + cT Rx c

where σ 2y = 3.0769, d = [1.7023 − 1.6172]T , and r x = [1.2808 − 0.6135]T .

(c) Use a Monte Carlo simulation to verify the theoretically obtained MMSE in part(a).
The lower left plot of Figure 6.34 shows a Monte Carlo simulation, with average square error and the
theoretical MMSE.
(d) Repeat part(a) for M = 3.
Figure 6.34 shows a Monte Carlo simulation for a third order FIR filter, where r x = [1.2808 −
0.6135 1.0945]T and d = [1.7023 − 1.6172 1.6160]T

6.35 Matlab program to reproduce the results shown in Figure 6.9.1.

(a) The Matlab script is shown below and the plots are shown in Figure 6.35a.
clc; close all;
set(0,’defaultaxesfontsize’,default_fontsize);

% Given parameters
var0 = 0.25;
a = 0.6;

% (a) Plots for rho = 0.1, -0.8, and 0.8, M = 8

M = 8; n = 0:M-1;
s = a.^n; s0 = s’;
Ps = 1;

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0635a’);

subplot(2,2,1), stem(n,s,’filled’,’g’);
ylabel(’s(n)’,’fontsize’,label_fontsize);
title(’Signal s(n)’,’fontsize’,title_fontsize);
axis([-1,M,0,1]);

ro = 0.1; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
subplot(2,2,2), stem(n,c,’filled’,’g’);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = 0.1, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,0,4]);

ro = -0.8; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
Statistical and Adaptive Signal Processing - Solution Manual 141

Signal s(n) ρ = 0.1, SNR= 5.5213 dB

1 4

3.5
0.8
3

0.6 2.5

s(n)

c(n)
2
0.4 1.5

1
0.2
0.5

0 0
0 2 4 6 8 0 2 4 6 8

ρ = −0.8, SNR= 13.6804 dB ρ = 0.8, SNR= 1.6898 dB

10 3

2.5
8
2

6 1.5
c(n)

c(n)
1
4 0.5

0
2
−0.5

0 −1
0 2 4 6 8 0 2 4 6 8
Sample index (n) Sample index (n)
Figure 6.35a: Signal and impulse responses of optimal matched filters for M = 8

subplot(2,2,3), stem(n,c,’filled’,’g’);
xlabel(’Sample index (n)’,’fontsize’,label_fontsize);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = -0.8, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,0,10]);

ro = 0.8; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
subplot(2,2,4), stem(n,c,’filled’,’g’); hold on;
plot([-1,M],[0,0],’w’);
xlabel(’Sample index (n)’,’fontsize’,label_fontsize);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = 0.8, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,-1,3]);

(b) The Matlab script is shown below and the plots are shown in Figure 6.35b.
% (b) Plots for rho = 0.1, -0.8, and 0.8, M = 16
M = 16; n = 0:M-1;
s = a.^n; s0 = s’;
Ps = 1;

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
142 Statistical and Adaptive Signal Processing - Solution Manual

Signal s(n) ρ = 0.1, SNR= 5.5225 dB

1 4

3.5
0.8
3

0.6 2.5

s(n)

c(n)
2
0.4 1.5

1
0.2
0.5

0 0
0 5 10 15 0 5 10 15

ρ = −0.8, SNR= 13.69 dB ρ = 0.8, SNR= 1.69 dB

10 3

2.5
8
2

6 1.5
c(n)

c(n)
1
4 0.5

0
2
−0.5

0 −1
0 5 10 15 0 5 10 15
Sample index (n) Sample index (n)
Figure 6.35b: Signal and impulse responses of optimal matched filters for M = 16

set(Hf_2,’NumberTitle’,’off’,’Name’,’Pr0635b’);

subplot(2,2,1), stem(n,s,’filled’,’g’);
ylabel(’s(n)’,’fontsize’,label_fontsize);
title(’Signal s(n)’,’fontsize’,title_fontsize);
axis([-1,M,0,1]);

ro = 0.1; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
subplot(2,2,2), stem(n,c,’filled’,’g’);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = 0.1, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,0,4]);

ro = -0.8; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
subplot(2,2,3), stem(n,c,’filled’,’g’);
xlabel(’Sample index (n)’,’fontsize’,label_fontsize);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = -0.8, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,0,10]);
Statistical and Adaptive Signal Processing - Solution Manual 143

ro = 0.8; Pv = var0/(1-ro*ro);
rv = ro.^n; Rv = Pv*toeplitz(rv);
c = inv(Rv)*s0;
subplot(2,2,4), stem(n,c,’filled’,’g’); hold on;
plot([-1,M],[0,0],’w’);
xlabel(’Sample index (n)’,’fontsize’,label_fontsize);
ylabel(’c(n)’,’fontsize’,label_fontsize);
SNRmax = Ps*(s0’*inv(Rv)*s0)
title([’\rho = 0.8, SNR= ’,num2str(SNRmax),’ dB’],’fontsize’,title_fontsize);
axis([-1,M,-1,3]);

6.36 Matlab program to reproduce the plot shown in Figure 6.9.2. The Matlab script is shown below and the plots
are shown in Figure 6.36.

clc; close all;

set(0,’defaultaxesfontsize’,default_fontsize);

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,3,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0636a’);

ro = -0.81; % correlation coefficient

R = toeplitz([1 ro]); % Autocorrelation matrix
th = [0:1000]*2*pi/1000; % Angle samples
x = [cos(th);sin(th)]; % c’*c = 1
[Q,L] = eig(R); % Eigenanalysis
lam1 = L(1,1); % Lambda 1
lam2 = L(2,2); % Lambda 2
y = Q*sqrt(L)*x; % c’*R*c = 1
y1 = Q*sqrt(L)*(1/sqrt(lam1))*x;
y2 = Q*sqrt(L)*(1/sqrt(lam2))*x;
minx = floor(min([x(1,:),y(1,:),y1(1,:),y2(1,:)]));
maxx = ceil(max([x(1,:),y(1,:),y1(1,:),y2(1,:)]));
miny = floor(min([x(2,:),y(2,:),y1(2,:),y2(2,:)]));
maxy = ceil(max([x(2,:),y(2,:),y1(2,:),y2(2,:)]));

plot(x(1,:),x(2,:),’g’,y(1,:),y(2,:),’m’);hold on;
plot(y1(1,:),y1(2,:),’c:’,y2(1,:),y2(2,:),’r--’);
plot([minx,maxx],[0,0],’w’,[0,0],[miny,maxy],’w’);
axis equal
axis([minx,maxx,miny,maxy]);
title(’\rho = -0.81’,’fontsize’,title_fontsize);
xlabel(’c_1’,’fontsize’,label_fontsize);
ylabel(’c_2’,’fontsize’,label_fontsize);
grid; hold off;
[legh,objh] = legend(’c^Hc=1’,’c^HRc=1’,...
’c^HRc=\lambda_1’,’c^HRc=\lambda_2’,1);
pause
144 Statistical and Adaptive Signal Processing - Solution Manual

exportfig(gcf,’p0636a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,3,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0636b’);

ro = 0.81; % correlation coefficient

R = toeplitz([1 ro]); % Autocorrelation matrix
th = [0:1000]*2*pi/1000; % Angle samples
x = [cos(th);sin(th)]; % c’*c = 1
[Q,L] = eig(R); % Eigenanalysis
lam1 = L(1,1); % Lambda 1
lam2 = L(2,2); % Lambda 2
y = Q*sqrt(L)*x; % c’*R*c = 1
y1 = Q*sqrt(L)*(1/sqrt(lam1))*x;
y2 = Q*sqrt(L)*(1/sqrt(lam2))*x;
minx = floor(min([x(1,:),y(1,:),y1(1,:),y2(1,:)]));
maxx = ceil(max([x(1,:),y(1,:),y1(1,:),y2(1,:)]));
miny = floor(min([x(2,:),y(2,:),y1(2,:),y2(2,:)]));
maxy = ceil(max([x(2,:),y(2,:),y1(2,:),y2(2,:)]));

plot(x(1,:),x(2,:),’g’,y(1,:),y(2,:),’m’);hold on;
plot(y1(1,:),y1(2,:),’c:’,y2(1,:),y2(2,:),’r--’);
plot([minx,maxx],[0,0],’w’,[0,0],[miny,maxy],’w’);
title(’\rho = 0.81’,’fontsize’,title_fontsize);
xlabel(’c_1’,’fontsize’,label_fontsize);
ylabel(’c_2’,’fontsize’,label_fontsize);
axis equal
axis([minx,maxx,miny,maxy]);
grid; hold off;
[legh,objh] = legend(’c^Hc=1’,’c^HRc=1’,...
’c^HRc=\lambda_1’,’c^HRc=\lambda_2’,1);
pause

exportfig(gcf,’p0636b.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

6.37 Interference rejection filter analysis.

2
The signal of interest is random with PSD Rs = Pa I. The background is colored with correlation rv (l) = ρ l ,
1 ≤ l ≤ M.

(a) SNRs for the matched and forward linear prediction error (LPE) filters:
M = 2: For the eigenfilter design, the autocorrelation matrix is given by
 
1 ρ
Rv =
ρ 1
Statistical and Adaptive Signal Processing - Solution Manual 145

Signal s(n) ρ = 0.1, SNR= 5.5213 dB Signal s(n) ρ = 0.1, SNR= 5.5225 dB
1 4
1 4

3.5 3.5
0.8 0.8
3 3

2.5 0.6 2.5

0.6

s(n)

c(n)
s(n)

c(n)
2 2

0.4 0.4 1.5

1.5

1 1
0.2 0.2
0.5 0.5

0 0 0 0
0 2 4 6 8 0 2 4 6 8 0 5 10 15 0 5 10 15

ρ = −0.8, SNR= 13.6804 dB ρ = 0.8, SNR= 1.6898 dB ρ = −0.8, SNR= 13.69 dB ρ = 0.8, SNR= 1.69 dB
10 3 10 3

2.5 2.5
8 8
2 2

1.5 6 1.5
6

c(n)

c(n)
c(n)

c(n)
1 1
4 4 0.5
0.5

0 0
2 2
−0.5 −0.5

0 −1 0 −1
0 2 4 6 8 0 2 4 6 8 0 5 10 15 0 5 10 15
Sample index (n) Sample index (n) Sample index (n) Sample index (n)

Figure 6.36: Geometric interpretation of the optimization process for ρ = −0.8 and ρ = 0.8

whose eigenvalues are: −ρ + 1, ρ + 1. The minimum eigenvalue is λmin = (1 − ρ) and the maximum
SNR is
1 1
SNRmax = =
λmin 1−ρ
For the LPE filter the maximum SNR is given by
aoT ao
SNRmax =
aoT Rv ao
 T
where ao is the forward linear predictor given by ao = 1 −ρ . Thus
 
  1
1 −ρ
−ρ 1 + ρ2
SNRmax =   =
  1 ρ 1 1 − ρ2
1 −ρ
ρ 1 −ρ
M = 3: For the eigenfilter design, the autocorrelation matrix is given by
 
1 ρ ρ4
Rv =  ρ 1 ρ 
ρ4 ρ 1
: :
whose eigenvalues are: 1 − ρ 4 , 1 + 21 ρ 4 + 12 ρ 8 + 8ρ 2 , 1 + 12 ρ 4 − 21 ρ 8 + 8ρ 2 . The minimum
eigenvalue is
: 1  
1 1  8
λmin = 1 + ρ 4 − ρ + 8ρ 2 = 2 + ρ 4 − ρ 4 1 + 8ρ −6
2 2 2
1$   %
= 2 + ρ 4 1 − 1 + 8ρ −6
2
Hence the maximum SNR is
1 2
SNRmax = =   
λmin 2 + ρ 4 1 − 1 + 8ρ −6
146 Statistical and Adaptive Signal Processing - Solution Manual

For the LPE we have to first solve

         
1 ρ a1 ρ a1 − 1 + ρ2 ρ
=− ⇒ =
ρ 1 a2 ρ4 a2 ρ2
  T
Thus ao = 1 − 1 + ρ 2 ρ ρ 2 . Hence
 
    1
1 − 1 + ρ2 ρ ρ2  − 1 + ρ2 ρ 
ρ2
SNRmax =    
   1 ρ ρ4  1
1 − 1 + ρ2 ρ ρ2  ρ 1 ρ   − 1 + ρ2 ρ 
ρ4 ρ 1 ρ2
1 + ρ 2 + 3ρ 4 + ρ 6 1 + ρ 2 + 3ρ 4 + ρ 6
= =  
1 − ρ2 − ρ4 + ρ6 ρ2 − 1 ρ4 − 1

M = 4: Using the above approach, we obtain

2
SNR(matched)
max = : 2  2
2 − ρ + ρ9 − ρ − ρ9 + 4 ρ − ρ4

and
1 + 3ρ 4 + 3ρ 6 + 2ρ 8 + ρ 10
SNR(LPE)
max =  2 
1 − ρ2 1 − ρ6

(b) Plots of the SNRs for M = 2, 3, and 4 and ρ = 0.6, 0.8, 0.9, 0.95, 0.99, and 0.995:
% (b) Plots of SNRs
ro = [0.6,0.8,0.9,0.95,0.99,0.995];

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0637b’);

% 1. M = 2
SNR_M = 1./(1-ro); SNR_Mdb = 10*log10(SNR_M);
SNR_L = (1+ro.*ro)./(1-ro.*ro); SNR_Ldb = 10*log10(SNR_L);
subplot(3,1,1);
plot(ro,SNR_Mdb,’g’,ro,SNR_Ldb,’r:’);
%xlabel(’\rho’,’fontsize’,label_fontsize);
ylabel(’Decibels’,’fontsize’,label_fontsize);
title(’M = 2’,’fontsize’,title_fontsize);
legend(’Matched filter’,’LPE filter’,0)

% 2. M = 3
SNR_M = 2./(2+ro.^4.*(1-sqrt(1+8*ro.^(-6))));
SNR_Mdb = 10*log10(SNR_M);
SNR_L = (1+ro.^2+3*ro.^4+ro.^6)./((ro.^2-1).*(ro.^4-1));
SNR_Ldb = 10*log10(SNR_L);
Statistical and Adaptive Signal Processing - Solution Manual 147

M=2
25
Matched filter
20 LPE filter

Decibels
15

10

0
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
M=3
50
Matched filter
40 LPE filter
Decibels

30

20

10

0
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
M=4
80
Matched filter
LPE filter
60
Decibels

40

20

0
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
ρ

Figure 6.37b: Plots of the SNRs for matched and LPE filters

subplot(3,1,2);
plot(ro,SNR_Mdb,’g’,ro,SNR_Ldb,’r:’);
%xlabel(’\rho’,’fontsize’,label_fontsize);
ylabel(’Decibels’,’fontsize’,label_fontsize);
title(’M = 3’,’fontsize’,title_fontsize);
legend(’Matched filter’,’LPE filter’,0)

% 3. M = 4
SNR_M = 2./(2-ro+ro.^9-sqrt((ro-ro.^9).^2+4*(ro-ro.^4).^2));
SNR_Mdb = 10*log10(SNR_M);
SNR_L = (1+3*ro.^4+3*ro.^6+2*ro.^8+ro.^10)./((1-ro.^2).^2.*(1-ro.^6));
SNR_Ldb = 10*log10(SNR_L);
subplot(3,1,3);
plot(ro,SNR_Mdb,’g’,ro,SNR_Ldb,’r:’);
xlabel(’\rho’,’fontsize’,label_fontsize);
ylabel(’Decibels’,’fontsize’,label_fontsize);
title(’M = 4’,’fontsize’,title_fontsize);
legend(’Matched filter’,’LPE filter’,0)
The plots are shown in Figure 6.37b.
(c) Magnitude responses of matched, LPE, and binomial filters for M = 3 and ρ = 0.9:
% (c) Magnitude response plots
M = 3; ro = 0.9
gr=(1+sqrt(5))/2;

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
148 Statistical and Adaptive Signal Processing - Solution Manual

0
Matched filter
LPE filter
Binomial filter
−5

−10

−15
Magnitude Response (dB)

−20

−25

−30

−35

−40

−45

−50
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Frequency (Cycles/Sampling interval)

Figure 6.37c: Magnitude responses of matched, LPE, and binomial filters

’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’Pr0637c’);

r = [ro^0 ro^1 ro^4 ro^9];

R = toeplitz(r(1:4));
[Q,D] = eig(R);
lmin = D(1,1)
q = Q(:,1);
[Hq,om] = freqz(q,1,1000);
[a,Pf] = olsigest(R,1);
a1 = [1;a]; a1=a1/norm(a1);
b1 = [1 -1];
b = conv(b1,conv(b1,conv(b1,b1))); b=b/norm(b);
[Hb,om] = freqz(b,1,1000);
[Ha,om] = freqz(a1,1,1000);
plot(om/(2*pi),20*log10(abs(Hq)),’g-’,om/(2*pi),20*log10(abs(Ha)),’r--’,...
om/(2*pi),20*log10(abs(Hb)),’c:’);
legend(’Matched filter’,’LPE filter’, ’Binomial filter’);
ylabel(’Magnitude Response (dB)’,’fontsize’,label_fontsize)
xlabel(’Frequency (Cycles/Sampling interval)’,’fontsize’,label_fontsize);
axis([0 0.5 -50 0]);
The plots are shown in Figure 6.37c.

6.38 Determine the matched filter for the deterministic pulse s(n) = cos ω0 n for 0 ≤ n ≤ M −1 and zero elsewhere.

(a) White noise with variance σ 2ν = 1

For the matched filter case in white noise, the spectrum of the filter matches the spectrum of the signal.
Statistical and Adaptive Signal Processing - Solution Manual 149

PSD of pulse s0

0 0 SNR = 1.6483 ρ =0.9

10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
PSD of Optimum Filter with White Noise

0 SNR = 6 0 SNR = 2.7324 ρ =0.5

10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
PSD of Optimum Filter with Colored Noise

0 SNR = 1.6483 0 SNR = 11.3926 ρ =−0.5

10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

2 2
SNR = 6 σν =1 SNR = 1.6483 σν =1
0 0
10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

2 2
SNR = 1.2 σν =5 SNR = 0.32965 σν =5
0 0
10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

2 2
SNR = 12 σν =0.5 SNR = 3.2965 σν =0.5
0 0
10 10

−5 −5
10 10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

Figure 6.38:

The equation for the optimum filter is

co = κso

where κ is chosen such that coH so = 1.

(b) Colored noise with autocorrelation rν (l) = σ 2ν ρ |l| /(1 − ρ 2 ), and ρ = 0.9.
For the matched filter case in colored noise, the equation for the optimum filter is

co = κ Rν−1 so

where κ = (soH Rν−1 so )−1 . The upper left plot of Figure 6.38 shows the spectrum of the signal, the
spectrum of the matched filter in white noise, and the spectrum of the matched filter in colored noise.
The dotted line in these plots represents the spectrum of the noise. As can be easily seen from this plot,
the matched filter will emphasis the spectrum with the least amount of noise energy, and de-emphasis the
spectrum with higher noise energy.
(c) Study the effect of the SNR in part(a) by varying the value of σ 2ν
The lower left and right plots of figure 6.38 show the matched filter for different values of σ 2ν in white
noise and colored noise respectively.
(d) Study the effect of the noise correlation in part(b) by varying the value of ρ
The upper right plot shows the matched filter for the colored noise case with different values of ρ
150 Statistical and Adaptive Signal Processing - Solution Manual

1
10 jω jω
|H(e )|max/ |H(e )|min = 6.573

0
10 W = 2.9

λ /λ = 6.078
−1 max min
10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
1
10 jω jω
|H(e )|max/ |H(e )|min = 12.54

0
10 W = 3.1

λmax / λmin = 11.12

−1
10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
1
10
|H(ejω)| jω
/ |H(e )|min = 26.16
max
0
10 W = 3.3

λmax / λmin = 21.71

−1
10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
1
10
|H(ejω)| / |H(ejω)|min = 63.81
max
0
10 W = 3.5

λmax / λmin = 46.82

−1
10
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5

Figure 6.39: Magnitude Response

6.39 Using the equalization experiment in Example 6.8.1 with M = 11 and D = 7.

(a) Compute and plot the magnitude response |H (e j ω )| of the channel and |Co (e j ω )| of the optimum equalizer
for W = 2.9,3.1,3.3, and 3.5 and comment upon the results.
Using Example 6.8.1, the magnitude of |H (e j ω )| is shown for each W in Figure 6.39 where H (e j ω ) is

H (e j ω ) = e−2 j ω [1 + (1 + cos(2π /W )) cos ω]

The optimum filter coefficients are found directly from Co = Rx−1 d where Rx = E{x(n)x ∗ (n)} and
d = E{a(n)x ∗ (n)}. Using equation 6.8.41,

D=7 d(l) = h(7 − l) ⇒ d(4) = h(3) d(5) = h(2) d(6) = h(1) d(l) = 0 elsewhere

Figure 6.39 shows the magnitude response of the optimum filter for each W . Note that the magnitude
response of the filter is the inverse of the channel. In effect, removing the influence of the channel from
the signal.
(b) Compute the spectral dynamic range |H (e j ω )|max /|H (e j ω )|min of the channel and the eigenvalue spread
λmax /λmin of the M x M input correlation matrix.
Figure 6.39 shows the values for the spectral dynamic range and the eigenvalue spread for each value of
W . Note that the value of these two parameters has a direct relation to each other.

6.40 In this problem we clarify some of the properties of the MSE equalizer discussed in Example 6.8.1.

(a) Compute and plot the MMSE Po as a function of M, and recommend how to choose a “reasonable” value.
(b) Compute and plot Po as a function of the delay D for 0 ≤ D ≤ 11. What is the best value of D?
(c) Study the effect of input SNR upon Po for M = 11 and D = 7 by fixing σ 2y = 1 and varying σ 2v .

6.41 In this problem we formulate the design of optimum linear signal estimators (LSE) using a constrained opti-
mization framework. To this end we consider the estimator e(n) = c0∗ x(n) + · · · + c∗M x(n − M) , c H x(n)
and we wish to minimize the output power E{|e(n)|2 } = c H Rc. To prevent the trivial solution c = 0 we need
to impose some constraint on the filter coefficients and use Lagrange multipliers to determine the minimum.
Let u i be an M × 1 vector with one at the ith position and zeros elsewhere.

(a) Show that minimizing c H Rc under the linear constraint u iT c = 1 provides the following estimators: FLP
if i = 0, BLP if i = M, and linear smoother if i  = 0, M.
Statistical and Adaptive Signal Processing - Solution Manual 151

(b) Determine the appropriate set of constraints for the L-steps ahead linear predictor, defined by c0 = 1
and {ck = 0}1L−1 , and solve the corresponding constrained optimization problem. Verify your answer by
obtaining the normal equations using the orthogonality principle.
(c) Determine the optimum linear estimator by minimizing c H Rc under the quadratic constraints c H c = 1
and c H W c = 1 (W is a positive definite matrix) which impose a constraint on the length of the filter
vector.
 Chapter 7

Algorithms and Structures for Optimum Linear Filters

7.1 By first computing the matrix product

  
Rm rmb Im −Rm−1 rm
rmb H ρ bm 0m 1

and then the determinants of both sides, prove Equation (7.1.25). Another proof, obtained using the L D L H
decomposition, is given by Equation (7.2.4).
Computing the matrix product directly
    
Rm rmb Im −Rm−1 rmb Rm 0m
=
rmb H ρ bm 0m 1 rmb H ρ bm − rmb H Rm−1 rmb

The determinants of both sides are found to be

det Rm+1 = (ρ bm − rmb H Rm−1 rmb )det Rm

Using (7.1.23)

α bm = ρ bm − rmb H Rm−1 rmb

then (7.1.25) is proven directly

det Rm+1
= α bm
det Rm

7.2 Prove the matrix inversion lemma for lower right corner partitioned matrices, which is described by Equations
(7.1.26) and (7.1.28).
Start with a slightly different partition of Rm+1 Qm+1 in (7.1.13)
    
f fH
ρ m rm qm qmH 1 0mH
Rm+1 Qm+1 = f f =
rm Rm q m Qm 0m Im

where the matrix multiplication equations (7.1.14) to (7.1.17) change to

ρ mf qm + rmf H qm = 1 (1)
ρ mf qmH + rmf H Qm = 0mH
rmf qm + Rmf qm = 0m
rmf qmH + Rmf Qm = Im (2)

Assuming matrix Rm is invertible

 −1 f
qm = − Rmf rm qm (3)

Substitute into (1) above to obtain qm

152
Statistical and Adaptive Signal Processing - Solution Manual 153

1
qm =  −1 (4)
f fH f f
ρ m − rm Rm rm

Substitute (3) into (2)

 −1
f f
− Rm rm
qm =  −1
f fH f f
ρ m − rm Rm rm
am
= f
(5)
αm

f
where am and α m are defined per (7.1.27) and (7.1.28). Now, (4) into (5) yields
 −1
Qm = Rmf (Im − rmf qmH )

  H 
−1
f f f
 f −1  
rm Rm rm
 
= Rm Im +  −1 
 f fH f f 
ρ m − rm Rm rm

Therefore, using (3), (4), and (6)

 
−1 qm qmH
Qm+1 = Rm+1 =
q m Qm
 
1 amH
 αm
f f
 α m   
= 
 −f
−f f −f f H
Rm rm Rm rm


am
f Rm +  −1
αm f
ρ m −rm
fH f
Rm
f
rm
 
1 amH

=  
f f
αm αm
−f (a )(a H )
am
f Rm + m f m
αm αm
 H
  
1 0m 1 1
= −f + f [1 amH ]
0m Rm α m am

7.3 This problem generalizes the matrix inversion lemmas to non-symmetric matrices.

(a) Show that if R −1 exists, the inverse of an upper left corner partitioned matrix is given by
 −1  
R r 1 α R −1 + wv T w
=
r̃ T σ α vT 1
where
Rw = −r
R T v = −r̃
α = σ − r̃ T R −1r = σ + v T r = σ + r̃ T w
154 Statistical and Adaptive Signal Processing - Solution Manual

Start with the partition of RQ = I in (7.1.13)

    
R r Q q I 0
=
r̃ σ
T
q̃T q 0T 1
and computing the product results in the following four equations
RQ + rq̃T = I (1)
Rq + rq = 0 (2)
r̃ Q + σ q̃
T T
= 0 T
(3)
r̃ q + σ q = 1
T
(4)

Assuming matrix Rm is invertible

q = −R−1 rq (5)

Substitute (5) into (4) to obtain q

1 1
q = = (6)
σ− r̃ H R−1 r α
where α = σ − r̃ H R−1 r.
Substitute (6) into (5) to obtain q

−R−1 r −R−1 r
q = = (7)
σ − r̃ H R−1 r α
Solve (1) for Q and substitute into (3) to obtain q̃T

−r̃T R−1 −r̃T R−1

q̃T = = (8)
σ − r̃ H R−1 r α
Lastly, substitute (9) into (8) to obtain Q

αR−1 + (R−1 r)(R−1 r̃)T

Q = (9)
α
−1
Therefore, Rm+1 is
 −1  
−1 R r Q q
Rm+1 = =
r̃ T σ q̃T q
 
1 α R −1 + wv T w
=
α vT 1
where
Rw = −r
R T v = −r̃
α = σ − r̃T R−1 r = σ + v T r = σ + r̃T w
Statistical and Adaptive Signal Processing - Solution Manual 155

(b) Show that if R −1 exists, the inverse of an upper left corner partitioned matrix is given by
 −1  
σ r̃T 1 1 vT
=
r R α w α R −1 + wv T
where

Rw = −r
R T v = −r̃
α = σ − r̃T R−1 r = σ + v T r = σ + r̃T w

Start with the partition of RQ = I in (7.1.13)

    
σ r̃T q q̃T 1 0
=
r R q Q 0T I

and computing the product results in the following four equations

σ q + r̃T q = 1 (1)
σ q̃ + r̃ Q = 0
T T T
(2)
rq + Rq = 0 (3)
RQ + rq̃ T
= I (4)

Assuming matrix Rm is invertible

q = −R−1 rq (5)

Substitute (5) into (1) to obtain q

1 1
q = = (6)
σ− r̃ H R−1 r α
where α = σ − r̃ H R−1 r.
Substitute (6) into (5) to obtain q

−R−1 r −R−1 r
q = = (7)
σ − r̃ H R−1 r α

Solve (4) for Q and substitute into (2) to obtain q̃T

−r̃T R−1 −r̃T R−1

q̃T = = (8)
σ − r̃ H R−1 r α
Lastly, substitute (8) into (1) to obtain Q

αR−1 + (R−1 r)(R−1 r̃)T

Q = (9)
α
156 Statistical and Adaptive Signal Processing - Solution Manual

−1
Therefore, Rm+1 is
 −1  
−1 σ r̃T q q̃T
Rm+1 = =
r R q Q
 
1 1 vT
= −1
α w α R + wv T

where

Rw = −r
R T v = −r̃
α = σ − r̃T R−1 r = σ + v T r = σ + r̃T w

(c)

7.4 Develop an order-recursive algorithm to solve the linear system in Example 7.1.2, using the lower right corner
partitioning lemma (7.1.26).
f f f
Using (7.1.26), and starting with m = 1, then R1 = 1, r1 = 1/2, and ρ 1 = 1. This leads to
 −1
f f
a1 = − R1 r1 = −1/2

and
 −1
f f fH f f
α 1 = ρ 1 − r1 R1 r1 = 3/4

−f
Using these results into (7.1.26) to find R2
   
−f 1 01H 1 1
R2 = −f + f [1 a1H ]
01 R1 α a1
   1 
1 0 4 1
= + [1 − 12 ]
0 1 3 − 1
 4  2
− 23
= 3
− 23 4
3

Therefore
 −1
f f
a2 = − R2 r2
 4  
− 23 1
= − 3 2
− 23 4 1
 4  3 3
−9
=
− 19
Statistical and Adaptive Signal Processing - Solution Manual 157

and
 −1
f f fH f f
α2 = ρ 2 − r2 R2 r2
f fH
= ρ 2 − r2 a2
 4 
−9
= 1 + [ 12 31 ]
− 19
= 20
27

Using (7.1.26) again

  
−f 1 02H 1
1
R3 = −f + f [1 a2H ]
02 R2 α a 2
 2  
1 0 0 1
27
=  0 4
3
− 23  +  − 4  [1 −
9
4
9
− 19 ]
20
0 − 23 4
3
− 1
  9
1.35 −0.6 −0.15
=  −0.6 1.6 −0.6 
−0.15 −0.6 1.35

−f
The desired solution is found using R3 and d3
 
− 20
9
−f
c3 = R3 d3 =  4
20

81
20

7.5 In this problem we consider two different approaches for inversion of symmetric and positive definite matrices by
constructing an arbitrary fourth-order positive definite correlation matrix R and comparing their computational
complexities.
(a) Given that the inverse of a lower (upper) triangular matrix is itself lower (upper) triangular, develop an
algorithm for triangular matrix inversion.
The Cholesky decomposition of the autocorrelation matrix is R = L D L H where L is lower triangular
and D is a diagonal matrix. Therefore the inverse of R is

R −1 = (L D L H )−1 = L −H D −1 L −1

The computational cost of L −1 is O(M 3 ).

(b) Compute the inverse of R, using the algorithm in part(a) and Equation (7.1.58).
We can build up the lower triangular inverse in a recursive manner from (7.1.37)
 
Lm 0
L m+1 =
lmH 1
where the inverse of L m+1 is
 −1  
Lm 0 L −1 0
L −1 = = m
m+1 lmH 1 vmT 1
−1
where vmT = L −H 3
m lm . This allows the order recursive build up of L m+1 . The computational cost is O(M ).
158 Statistical and Adaptive Signal Processing - Solution Manual

(c) Build up the inverse of R, using the recursion (7.1.24).

Using (7.1.22), (7.1.23) and (7.1.24), the recursion for the computation of the inverse of R is
 −1
−1 Rm rmb
Rm+1 =
rmb H ρ bm
   
Rm−1 0m 1 bm
= = b [bmH 1]
0mH 0 αm 1

where bm = −Rm−1 rmb and α bm = ρ bm + rmb H bm

(d) Estimate the number of operations for each method as a function of order M, and check their validity for
M = 4, using MATLAB.

7.6 Using the appropriate orthogonality principles and definitions, prove Equation (7.3.32)
f∗
The partial correlation E{emb (n − 1)em (n)} can be expanded using (7.3.39) as

E{emb (n − 1)(x ∗ (n) + xm∗ (n − 1)amH (n))} = E{emb (n − 1)x ∗ (n)}E{emb (n − 1)xm∗ (n − 1)}amH (n))

Using the fact that xm is orthogonal to em , the above equation can be simplified to

E{emb (n − 1)x ∗ (n)} = E{(bmH (n − 1)xm (n − 1) + x(n − M − 1))x ∗ (n)}

again using (7.3.39). The final step uses (7.3.2) and (7.3.26) to finish the proof
&   '
∗ xm (n − 1) ∗
E{(bm (n − 1)xm (n − 1) + x(n − M − 1))x (n)} = E [bm (n − 1) 1]
H H
x (n)
x(n − M − 1)
&  '
xm (n − 1) ∗
= [bm (n − 1) 1]E
H
x (n)
x(n − M − 1)
= [bmH (n − 1) 1]E{xm+1 (n − 1)x ∗ (n)}
 
f
r m (n)
= [bmH (n − 1) 1] f
rm+1 (n)

f∗ f f
Therefore, E{emb (n − 1)em (n)} = rm+1 (n) + bmH (n − 1)rm (n)

7.7 Prove Equations (7.3.36) to (7.3.38), using Equation (7.1.45)

f fH
Pm+1 (n) = Pmf (n) + rm+1 (n)am+1 (n)
fH −1 f
= Pmf (n) − rm+1 (n)Rm+1 (n − 1)rm+1 (n)
Pmf (n) − rm+1 (n)L −1 −1 −1
fH f
= m+1 (n − 1)Dm+1 (n − 1)L m+1 (n − 1)rm+1 (n)

Pmf (n) − rm+1 (n)L −1 −1 −1 −1

fH f
= m+1 (n − 1)Dm+1 (n − 1)Dm+1 (n − 1)Dm+1 (n − 1)L m+1 (n − 1)rm+1 (n)
= Pmf (n) − km+1
b
(n)Dm+1 (n − 1)km+1
b
(n)
= Pmf (n) − Dm+1 (n − 1)|km+1
b
(n)|2
= ???

7.8 Working as in Example 6.3.1, develop an algorithm for the upper-lower decomposition of a symmetric positive
definite matrix. Then use it to factorize the matrix in Example 6.3.1, and verify your results, using the function
[U,D]=udut(R).
Statistical and Adaptive Signal Processing - Solution Manual 159

Writing the decomposition directly for M=4, we have

     
r11 r12 r13 r14 1 u 10 u 20 u 30 ξ1 0 0 0 1 0 0 0
 r21 r22 r23 r24   0 1 u 21 u 31   0 ξ 2 0 0   u∗ 0 
      10 1 0 
 r31 r32 r33 r34  =  0 0 1 u 32   0 0 ξ 3 0   u∗ u∗ 1 0 
20 21
r41 r42 r43 r44 0 0 0 1 0 0 0 ξ4 u ∗30 u ∗31 u ∗32 1

where ri j = r j i and ξ i > 0. If we perform the matrix multiplication on the right-hand side of (6.3.7) and
equate the matrix elements on the left and right sides, we obtain,

r11 = ξ 1 ⇒ ξ 1 = r11
r21
r21 = ξ 1 u 10 ⇒ u 10 =
ξ1
r22 = ξ 1 |u 10 | + ξ 2 ⇒ ξ 2 = r22 − ξ |u 10 |2
2

r31
r31 = ξ 1 u 20 ⇒ u 20 =
ξ1
r32 − ξ 1 u 20 u ∗10
r32 = ξ 1 u 20 u ∗10 + ξ 2 u 21 ⇒ u 21 =
ξ2
r33 = ξ 1 |u 20 | + ξ 2 |u 21 | + ξ 3 ⇒ ξ 3 = r33 − ξ 1 |u 20 |2 − ξ 2 |u 21 |2
2 2

r41
r41 = ξ 1 u 30 ⇒ u 30 =
ξ1
r42 − ξ 1 u 30 u ∗10
r42 = ξ 1 u 30 u ∗10 + ξ 2 u 31 ⇒ u 31 =
ξ2
r43 − ξ 1 u 30 u ∗20 − ξ 2 u 31 u ∗21
r43 = ξ 1 u 30 u ∗20 + ξ 2 u 31 u ∗21 + ξ 3 u 32 ⇒ u 32 =
ξ3
r44 = ξ 1 |u 30 | + ξ 2 |u 31 | + ξ 3 |u 32 | + ξ 4 ⇒ ξ 4 = r44 − ξ 1 |u 30 | − ξ 2 |u 31 |2 − ξ 3 |u 32 |2
2 2 2 2

This provides a row-by-row computation of the elements of the U H DU decomposition. Note that the compu-
tation of the next row does not change the already computed rows. This recursion can be continued for higher
order matrices.

7.9
H
7.9 In this problem we explore the meaning of the various quantities in the decomposition R = UD̄U of the
correlation matrix.

(a) Show that the rows of A = U−1 are the MMSE estimator of xm from xm+1 , xm+2 , . . . , x M .
H
(b) Show that the decomposition R = UD̄U can be obtained by the Gram-Schmidt orthogonalization pro-
cess, starting with the random variable x M and ending with x1 , that is, proceeding backward.

7.10 In this problem we clarify the various quantities and the form of the partitionings involved in the U DU H
decomposition, using an m = 4 correlation matrix.

(a) Prove that the components of the forward prediction error vector (7.3.65) are uncorrelated.
The forward prediction error vector is uncorrelated if it’s autocorrelation is a diagonal matrix. Using
(7.3.66)

e(n) = A(n)x(n)
160 Statistical and Adaptive Signal Processing - Solution Manual

with an autocorrelation of

Re = AE{x x H } A H = ARx A H = A(U D̄U H ) A H

where the U DU H is the Gram-Schmidt decomposition. From table 7.1, A = U −1 and Re is

Re = A(U D̄U H ) A H = U −1 (U D̄U H )U −H = D̄

which a diagonal matrix.

(b) Writing explicitly the matrix R, identify and express the quantities in Equations (7.3.62) through (7.3.67)
Using (7.3.62) the order-recursion of R for m = 4 is

R1 (n − 3) → R2 (n − 2) → R3 (n − 1) → R4 (n)

where R1 (n − 3) = Px (n − 3) and
 
fH
Px (n − 2) r1 (n − 2)
R2 (n − 2) = f
r1 (n − 2) R1 (n − 3)
 
fH
Px (n − 1) r2 (n − 1)
R3 (n − 1) = f
r2 (n − 1) R2 (n − 2)
 
fH
Px (n) r3 (n)
R4 (n) = f
r3 (n) R3 (n − 1)

Using (7.3.63), the order-recursion for the FLP coefficients for m = 4 is

a1 (n − 3) → a2 (n − 2) → a3 (n − 1) → a4 (n)

where a1 (n − 3) = a1(1) (n − 3) and

 
a1(2) (n − 2)
a2 (n − 2) =
a2(2) (n − 2)
 
a1(3) (n − 1)
 
a3 (n − 1) =  a2(3) (n − 1) 
(3)
a3 (n − 1)
 
a1(4) (n)
 
 a2(4) (n) 
a4 (n) =  
 a3(4) (n) 
a4(4) (n)

Using (7.3.64) and (7.3.65) the order-recursion for the FLP errors for m = 4 is
f f f f
e1 (n − 3) → e2 (n − 2) → e3 (n − 1) → e4 (n)
Statistical and Adaptive Signal Processing - Solution Manual 161

and the FLP error vector is

f f
e1 (n − 3) = e1 (n − 3)
 
f
f e2 (n − 2)
e2 (n − 2) = f
e1 (n − 3)
 f 
e (n − 1)
f  3f 
e3 (n − 1) =  e2 (n − 2) 
f
e3 (n − 3)
 f

e4 (n)
 f 
 e (n − 1) 
e4 (n) =  3f 
 e2 (n − 2) 
f
e1 (n − 3)

(c)

7.11 Given an all-zero lattice filter with coefficients k0 and k1 , determine the MSE P(k0 , k1 ) as a function of the
required second-order moments, assumed jointly stationary, and plot the error performance surface. Use the
statistics in Example 6.2.1
For an all-zero lattice filter, the estimate ŷ(n) is

ŷ(n) = x(n) + (k0∗ + k0 k1∗ )x(n − 1) + k1∗ x(n − 2)

where x(n) is the input to the lattice. The MSE P(k0 , k1 ) is

P(k0 , k1 ) = Py − 2d1 c1 − 2d2 c2 + r11 c12 + 2r12 c1 c2 + r22 c22

where c1 = k0∗ + k0 k1∗ and c2 = k1∗ . Using the statistics from Example 6.2.1 (Py = 0.5, r11 = r22 = 4.5,
r12 = r21 = −0.1545, d1 = −0.5, and d2 = −0.1545), the MSE P(k0 , k1 ) is

P(k0 , k1 ) = 0.5 + k0 + k0 k1 + (0.3090)k1 + 4.5(k0 + k0 k1 )2 − (0.3090)(k0 + k0 k1 )(k1 ) + 4.5(k1 )2

7.12 Given the autocorrelation r(0) = 1,r(1) = r(2) = 12 , and r(3) = 14 , determine all possible representations for
the third-order prediction error filter (see Figure 7.7)
Given the autocorrelation, we need to find the forward predictor coefficients am and the latice parameters km .
Using Table 7.2, step (2a)

P0 = r(0) = 1, β 0 = r ∗ (1) = 1/2

and step (2b)

r(1)
k0 = − , a1( 1) = k0
r(0)
162 Statistical and Adaptive Signal Processing - Solution Manual

Using step (3), for m = 1

(a) P1 = P0 + β 0 k0 = 1 + ( 12 )(− 12 ) = 1/4
(b) r1 = r(1) = 1/2
(c) β 1 = a1T J r1∗ + r ∗ (2) = (1/2)(−1/2) + (1/2) = 1/4
β −1/4
(d) k1 = − 1 = = −1/3
P1 3/4
   
a1 J a1∗
(e) a2 = + k1
0 1
     
−1/2 −1/2 −1/3
= + (−1/3) =
0 1 −1/3

Repeating step (3) for m = 2

(a) P2 = P1 + β 1 k1 = 3/4 + ( 14 )(− 13 ) = 2/3
(b) r2 = [r(1)r(2)]T = [1/2 1/2]T
 
−1/3
(c) β 2 = a2T J r2∗ ∗
+ r (3) = [−1/3 − 1/3] + (1/4) = −1/2
−1/3
β 1/2
(d) k2 = − 2 = = 3/4
P2 2/3
   
a2 J a2∗
(e) a3 = + k2
0 1
     
−1/3 −1/3 −7/12
=  −1/3  +  −1/3  (3/4) =  −7/12 
0 1 3/4

Lastly, using step (4)

P3 = P2 + β 2 k2 = 2/3 − (1/2)(3/4) = 7/24
This can be checked using (7.4.30)
P3 = r(0)3m=1 (1 − |km−1 |2 )

Therefore, the forward predictor coefficients am and the latice parameters km are
 
−7/12
am =  −7/12 
3/4
and
k0 = −1/2 k1 = −1/3 k2 = 3/4

7.13 Repeat Problem 7.12 for k0 = k1 = k2 = 1/3 and P3 = (2/3)3 .

Given the lattice coeffiecients, we need to find the autocorrelation rm and forward predictor coefficients am .
The lag zero autocorreltaion coefficient can be found using (7.4.30),
P3
r(0) = = 27/64
3m=1 (1 − |km−1 |2 )
Statistical and Adaptive Signal Processing - Solution Manual 163

Using (7.5.28) we can find Pm

Pm = Pm−1 (1 − |km−1 |2 )
P1 = P0 (1 − |k0 |2 )
= (27/64)(1 − |1/3|)2 = 3/8
P2 = P1 (1 − |k1 |2 )
= (3/8)(1 − |1/3|2 ) = 1/3

The autocorrelation coefficients can be found using (7.5.36)

r(m + 1) = −km∗ Pm − amH J rm

r(1) = −k0∗ P0
= −(1/3)(27/64) = −9/64
r(2) = −k1∗ P1 − a1H r1
= (−1/3)(3/8) − (1/3)(−9/64) = −5/64
r(3) = −k2∗ P2 − a2H J r2
= −k2∗ P2 − [a1(2)r(2) + a2(2)r(1)]
= (−1/3)(1/3) + (4/9)(5/64) + (1/3)(9/64) = 17/576

where a1(1) = k0 = 1/3, a2(2) = k1 = 1/3, and a1(2) = k0 + k0∗ k1 = 4/9.

The forward predictor coefficients am can be found using (7.5.27)
   ∗

am−1 J am−1
am = + km−1
0 1
a1 = k1 = 1/3
     
1/3 1/3 4/9
a2 = + (1/3) =
0 1 1/3
     
4/9 1/3 5/9
a3 =  1/3  +  4/9  (1/3) =  13/27 
0 1 1/3

7.14 Use Levinson’s algorithm to solve the normal equations Rc = d where R =Toeplitz{3, 2, 1} and d = [6 6 2]T .
Using the Levinson algorithm (Table 7.3), starting with step 2

(a) P0 = r(0) = 3, β 0 = r ∗ (1) = 2

(b) k0 = −β 0 /P0 = −2/3, a1(1) = k0 = −2/3
(c) β c0 = d1 = 6
(d) k0c = −β c0 /P0 = −6/3 = −2, c1(1) = k0c = −2
164 Statistical and Adaptive Signal Processing - Solution Manual

Solving step 3, with m = 1 yields the following

(a) r1 = [r(1)]
(b) β 1 = a1T Jr1∗ + r(2) = (−2/3)(2) + 1 = −1/3
(c)P1 = P0 + β 0 k0∗ = 3 + 2(−2/3) = 5/9
1/3
(d) k1 = −β 1 /P1 = = 3/5
5/9
         
a1 a1 −2/3 −2/3 −16/15
(e) a2 = + k1 = + (3/5) =
0 1 0 1 3/5
( f ) β c1 = −c1(1)r(1) + d2 = (2)(2) + 6 = 10
10
(g) k1c = β c1 /P1 = = 18
5/9
         
c1 J a1 −2 −2/3 −14
(h) c2 = + k1 =
c
+ (18) =
0 1 0 1 18

The last recursion for m = 2 repeats step 3 only for the optimum filter
 
2
( f ) β 2 = −c2 Jr(2) + d3 = −[−14 18]
c H
+ 2 = 12
1
β c2
(g) k2c = β c2 /P2 = = 33.75
P1 + β 1 k1
     
    −14 3/5 6.25
c2 J a2
(h) c3 = + k2c =  18  +  −16/15  (33.75) =  −18.00 
0 1
0 1 33.75

Therefore, the optimum filter is c3 = [6.25 − 18.0 33.75]T

7.15 Consider a random sequence with autocorrelation {r(L)}30 = {1, 0.8, 0.6, 0.4}
f
(a) Determine the FLP am and the corresponding error Pm for m = 1, 2, 3.
Given the autocorrelation coefficients, the FLP and the corresponding errors can be found using Levinson-
Durbin algorithm (Table 7.2). Using Table 7.2, step (2a)

P0 = r(0) = 1, β 0 = r ∗ (1) = 0.8

and step (2b)

r(1)
k0 = − = −0.8, a1(1) = k0 = −0.8
r(0)
Statistical and Adaptive Signal Processing - Solution Manual 165

Using step (3), for m = 1

(a) P1 = P0 + β 0 k0 = 1 − (0.8)2 = 0.36

(b) r1 = r(1) = 0.8
(c) β 1 = a1T J r1∗ + r ∗ (2) = (−0.8)(0.8) + (0.6) = −0.04
β 0.04
(d) k1 = − 1 = = 0.111
P 0.36
 1   
a1 J a1∗
(e) a2 = + k1
0 1
     
−0.8 −0.8 −0.889
= + (−0.111) =
0 1 0.111

Repeating step (3) for m = 2

(a) P2 = P1 + β 1 k1 = 0.36 − (0.04)(0.111) = 0.356

(b) r2 = [r(1)r(2)]T = [0.8 0.6]T
 
0.6
(c) β 2 = a2T J r2∗ ∗
+ r (3) = [−0.889 0.111] + (0.4) = −0.044
0.8
β 0.044
(d) k2 = − 2 = = 0.1248
P 0.356
 2   
a2 J a2∗
(e) a3 = + k2
0 1
     
−0.889 0.111 −0.8765
=  0.111  +  −0.889  (0.1248) =  0.0123 
0 1 0.1111

Lastly, using step (4)

P3 = P2 + β 2 k2 = 0.356 − (0.044)(0.1248) = 0.35

This can be checked using (7.4.30)

P3 = r(0)3m=1 (1 − |km−1 |2 )

(b) Determine and draw the flow diagram of the third-order lattice prediction error filter.
Figure 7.15 shows the third-order lattice prediction error filter where the lattice parameters are k0 = −0.8,
k1 = 0.1111, and k2 = 0.1248 which have been computed above in part (a).

7.16 Using the Levinson-Durbin algorithm, determine the third-order linear predictor a3 and the MMSE P3 for the
signal with autocorrelation r(0) = 1,r(1) = r(2) = 1/2, and r(3) = 1/4.
Given the autocorrelation coefficients, the FLP and the corresponding error can be found using Levinson-Durbin
algorithm (Table 7.2). Using Table 7.2, step (2a)

P0 = r(0) = 1, β 0 = r ∗ (1) = 1/2

and step (2b)

r(1)
k0 = − = −1/2, a1(1) = k0 = −1/2
r(0)
166 Statistical and Adaptive Signal Processing - Solution Manual

− − −

Figure 7.15: Third-order lattice prediction error filter

Using step (3), for m = 1

(a) P1 = P0 + β 0 k0 = 1 − (1/2)(1/2) = 3/4

(b) r1 = r(1) = 1/2
(c) β 1 = a1T J r1∗ + r ∗ (2) = (−1/2)(1/2) + (1/2) = 1/4
β −1/4
(d) k1 = − 1 = = −1/3
P1 3/4
   
a1 J a1∗
(e) a2 = + k1
0 1
     
−1/2 −1/2 −1/3
= + (−1/3) =
0 1 −1/3

Repeating step (3) for m = 2

(a) P2 = P1 + β 1 k1 = 3/4 − (1/4)(1/3) = 2/3

(b) r2 = [r(1) r(2)]T = [1/2 1/2]T
 
1/2
(c) β 2 = a2T J r2∗ ∗
+ r (3) = [−1/3 1/3] + (1/4) = −1/12
1/2
β 1/12
(d) k2 = − 2 = = 1/8
P2 2/3
   
a2 J a2∗
(e) a3 = + k2
0 1
     
−1/3 −1/3 −3/8
=  −1/3  +  −1/3  (1/8) =  −3/8 
0 1 1/8

Lastly, using step (4)

P3 = P2 + β 2 k2 = 2/3 − (1/12)(1/8) = 21/32

Therefore, the third-order linear predictor a3 is

 
−3/8
a3 =  −3/8 
1/8

and the MMSE is P3 = 21/32

Statistical and Adaptive Signal Processing - Solution Manual 167

7.17 Given the autocorrelation sequence r(0) = 1,r(1) = r(2) = 1/2, and r(3) = 1/4, compute the lattice and
direct-form coefficients of the prediction error filter, using the algorithm of Schür.
Using Table 7.4 to help solve for the lattice coefficients km , starting with step 2
f
(a) ξ 0 (l) = ξ b0 (l) = r(l)
f
ξ 0 (1) −1/2
(b) k0 = − = = −1/2
ξ 0 (0)
b 1
(c) P1 = r(0)(1 − |k0 |2 ) = (1)(1 − |1/4|2 ) = 3/4

Continuing the algorithm with step 3, starting with m = 1

f ∗
(a) ξ mf (l) = ξ m−1 (l) + km−1 ξ bm−1 (l − 1)
ξ 1 (2) = ξ 0 (2) + k0∗ ξ b0 (1) = 1/2 + (−1/2)(1/2) = 1/4
f f

ξ b1 (1) = ξ 0 (0) + k0∗ ξ b0 (1) = 1 + (−1/2)(1/2) = 3/4

f

ξ f (m + 1)
(b) km = − m b
ξ m (m)
f
ξ (2) −1/4
k1 = − 1b = = −1/3
ξ 1 (1) 3/4
(c) Pm+1 = Pm (1 − |km |2 )
P2 = P1 (1 − |k1 |2 ) = 2/3

Repeating step 3, with m = 2

ξ 1 (3) = ξ 0 (3) + k0∗ ξ b0 (2) = 1/4 + (−1/2)(1/2) = 0

f f
(a)
ξ b1 (2) = ξ 0 (1) + k0∗ ξ b0 (2) = 1/2 + (−1/2)(1/2) = 1/4
f

ξ 2 (3) = ξ 1 (3) + k1∗ ξ b1 (2) = 0 + (−1/3)(1/4) = −1/12

f f

ξ b2 (2) = ξ 1 (1) + k1∗ ξ b1 (2) = 3/4 + (−1/3)(1/4) = 2/3

f

f
ξ (3) 1/12
(b) k2 = − 2b = = 1/8
ξ 2 (2) 2/3
(c) P2 = P1 (1 − |k1 |2 ) = (2/3)(1 − |1/8|2 ) = 21/32

Lastly, the output of the Schür algorithm is k0 = −1/2,k1 = −1/3,k2 = 1/8, and P3 = 21/32.
The direct-form coefficients can be found using (7.5.27)
   ∗

am−1 J am−1
am = + km−1
0 1
a1 = k0 = −1/2
     
−1/2 −1/2 −1/3
a2 = + (−1/3) =
0 1 −1/3
     
−1/3 −1/3 −3/8
a3 =  −1/3  +  −1/3  (1/8) =  −3/8 
0 1 1/8
168 Statistical and Adaptive Signal Processing - Solution Manual

7.18 Determine ρ 1 and ρ 2 so that the matrix R =Toeplitz{1, ρ 1 , ρ 2 } is positive definite.

If detR < 1 then R is positive definite. The determinant can be written as

 detR < 1
 1 ρ1 ρ2 
 
 ρ 1 1 ρ 1  < 1
 
 ρ ρ 1 
2 1
1 + 2ρ 21 ρ 2 − 2ρ 21 − ρ 22 < 1
(ρ 2 − 1)(2ρ 21 − ρ 2 − 1) < 1

This last line yields two conditions

ρ2 < 1

and

2ρ 21 − 1 < ρ 2

Combining the two yields 2ρ 21 − 1 < ρ 2 < 1 which is exactly the condition shown in (4.2.105).

7.19 An AR(2) model fit for a sinusoidal random-phase signal in additive noise with autocorrelation

r(l) = P0 cos ω0l + σ 1v δ(l)

(a) Model parameters a1(2) , a2(2) , and σ 2w in terms of P0 , ω0 , and σ 2v : The model parameters are given by the
normal equation
   (2)   
r(0) r(1) a1 r(1)
=−
r(1) r(0) a2(2) r(2)

or
    
P0 + σ 2v P0 cos ω0 a1(2) P0 cos ω0
=−
P0 cos ω0 P0 + σ 2v a2(2) P0 cos 2ω 0

Hence
   
P0 +σ 2v −P0 cos 2ω0
a1(2) −P0 (cos ω0 )
= P0 +2 P0 σ 2v +σ 4v −P02 cos2 ω0 
2

a2(2) −P0
−P0 cos2 ω0 +P0 cos 2ω0 +(cos 2ω0 )σ 2v
P02 +2 P0 σ 2v +σ 4v −P02 cos2 ω0

Also
$ %  r(1) 
σ 2w = P0 + a1(2) a2(2)
r(2)
$ %  P cos ω 
P0 +σ 2v −P0 cos 2ω0 −P cos2 ω +P cos 2ω +(cos 2ω )σ 2
= P0 + −P0 (cos ω0 ) P02 +2 P0 σ 2v +σ 4v −P02 cos2 ω0
−P0 0 P 2 +2 0P σ 20+σ 4 −P0 2 cos2 ω 0 v
0 0
0P0 cos 2ω0
0 v v 0 0
 
−2P02 cos2 ω0 − P0 cos2 ω0 σ 2v + 2P02 cos2 ω0 cos 2ω0 − P02 cos2 2ω0 − P0 cos2 2ω0 σ 2v + P02 + 2P0 σ 2v +
= P0
P02 + 2P0 σ 2v + σ 4v − P02 cos2 ω0
Statistical and Adaptive Signal Processing - Solution Manual 169

(b) Lattice parameters k1 and k2 in terms of P0 , ω0 , and σ 2v : We have

−P0 cos2 ω0 + P0 cos 2ω0 + (cos 2ω0 ) σ 2v

k1 = a2(2) = −P0
P02 + 2P0 σ 2v + σ 4v − P02 cos2 ω0

and

a1(2) − a1(2)∗ k1 a1(2) a1(2)

k0 = = =
1 − k12 1 + k1 1 + a2(2)
cos ω0
= −P0
P0 + σ 2v

(c) Limiting values as σ 2v → 0:

P0 − P0 cos 2ω 0 (cos ω0 ) (1 − cos 2ω0 )

a1(1 = −P0 (cos ω0 ) =
P0 − P0 cos2 ω0
2 2
(cos ω0 − 1) (cos ω0 + 1)


−P0 cos2 ω0 + P0 cos 2ω 0 cos 2ω0 − cos2 ω0
a2(2) = −P0 = = 1 = k1
P02 − P02 cos2 ω0 (cos ω0 − 1) (cos ω0 + 1)

cos ω0
k0 = −P0 = − cos ω0
P0
and
−2P02 cos2 ω0 + 2P02 cos2 ω0 cos 2ω0 − P02 cos2 2ω0 + P02
σ 2w = P0
P02 − P02 cos2 ω0

(cos 2ω 0 − 1) − cos 2ω0 + 2 cos2 ω0 − 1
= P0 =0
− (cos ω0 − 1) (cos ω0 + 1)

7.20 Given the parameters r(0) = 1,k0 = k1 = 1/2, and k2 = 1/4, determine all other equivalent representations
of the prediction error filter (see Figure 7.7)
The forward predictor can be found directly using (7.5.27)
   ∗

am−1 J am−1
am = + km−1
0 1
a1 = k0 = 1/2
   
1/2 1/2
a2 = + (1/2)
0 1
 
3/4
=
1/2
   
3/4 1/2
a3 =  1/2  +  3/4  (1/4)
0 1
 
7/8
=  11/16 
1/4
170 Statistical and Adaptive Signal Processing - Solution Manual

The autocorrelation can be found directly using (7.5.36) and using the fact that r(0) = P0 then for m = 0
r(1) = −k0∗ P0 = −(1/2)(1) = −1/2
For m = 2
r(2) = −k1∗ P1 − a1(1)∗r(1) = −(1/2)(3/4) − (1/2)(−1/2) = −1/8
where
P1 = P0 (1 − |k0 |2 ) = (1)(1 − |1/2|2 ) = 3/4
Finally, for m = 2
r(3) = −k2∗ P2 − [a1(2)∗r(2) + k1∗r(1)]
= −(1/4)(3/4)2 − [(3/4)(−1/8) + (1/2)(−1/2)
= 5/16
where
P2 = P1 (1 − |k1 |2 ) = (3/4)(1 − |1/2|2 ) = (3/4)2
Therefore the autocorrelation is r(0) = 1, r(1) = −1/2, r(2) = −1/8, and r(3) = 5/16
7.21 Let {r(l)}0P be samples of the autocorrelation sequence of a stationary random signal x(n)

(a) Is it possible to extend r(l) for |l| > P so the the PSD
∞

R(e j ω ) = r(l)e− j ωl
l=−∞

is valid, that is, R(e j ω ) ≥ 0?

We first extend R(l) for negative lags using r(−l) = r ∗ (l). We next recall that
detRm+1 = detRm Pm+1 = detRm Pm (1 − |km |2 )
Since Rm is positive definite and Pm > 0 (because r(l) is a valid autocorrelation of a stationary process),
detRm+1 is positive definite if and only if (1 − |km |2 ) > 0, but
bm r ∗ (m + 1) + amT Jrm∗
km = =
Pm Pm
Therefore detRm+1 > 0 if and only if
|r ∗ (m + 1) + amT Jrm∗ | < Pm2
which defines an open disk in the complex plane. Any value of r(m + 1) inside the disk provides a valid
extension of r(l) which inturn leads to a valid PSD R(e j ω )
(b) Using the algorithm of Levinson-Durbin, develop a procedure to check if a given autocorrelation extension
is valid
For a given autocorrelation extension r(m + 1), the Levinson-Durbin algorithm (Table 7.2) can be used
to determine its validity. Specifically computing β m in step 3(c)
β m = amT Jrm∗ + r ∗ (m + 1)
and using this result to compute km in step 3(d)
km = −β m /Pm
If |km | < 1, then the autocorrelation is valid.
Statistical and Adaptive Signal Processing - Solution Manual 171

(c) Use the algorithm in part (b) to find the necessary and sufficient conditions so that r(0) = 1, r(1) = ρ 1 ,
and r(2) = ρ 2 are a valid autocorrelation sequence. Is the resulting extension unique?
Using the algorithm from step (b) above
β 1 = a1T Jr1∗ + r ∗ (2) = ρ 2 − ρ 21
where a1(1) = −r(1)/r(0) and with P1 = P0 + β 0 k0 = r(0) + r(1) −r(1)
r(0)

a1T Jr1∗ + r ∗ (2) ρ 21 − ρ 2

k1 = −β 1 /P1 = =
r 2 (0) − |r(1)|2 1 − ρ 22
Setting |k1 | < 1 results in 2ρ 21 − 1 < ρ2 which is the necessary and sufficient condition and this agrees
with (4.2.105). Since the disk defined by (1) is finite, the autocorrelation extension is not unique.

7.22 Justify the following statements.

(a) The whitening filter for stationary process x(n) is time-varying
Suppose that we use x(0), x(1), . . . , x(m − 1) to predict x(m). The FPEF provides an approximation
to the whitening filter of the process x(n) determined by spectral factorization (see section 6.6.4). Note
that, due to stationarity, we can use x(0), x(1), . . . , x(m) or x(n), x(n − 1), . . . , x(n − m) to determine
the mth-order FLP. The two filters are identical for an AR(m) process. Since, the coefficients of am and
am+1 are different and m represents time in x(0), x(1), . . . , x(m), the finite order whitening filter is time
varying.
(b) The filter in (a) can be implemented by using a lattice structure and switching its stages on one by one
with the arrival of each new sample.
In a lattice implementation, order recursion corresponds to adding lattice stages. Therefore, to determine
the new FPEF when x(m) arrives, we add a new stage to the lattice.
(c) If x(n) is AR(P), the whitening filter becomes time-invariant P + 1 sampling intervals after the first
sample is applied. N ote : We assume that the input is applied to the filter at n = 0. IF the input is applied
at n = −∞, the whitening filter of a stationary process is always time-invariant.
For an AR(P) process, km = 0 for m > P. Therefore, the lattice does not change for m = n > P.
Hence, it is time-invariant.
7.23 Given the parameters r(0) = 1, k0 = 1/2, k1 = 1/3, and k2 = 1/4, compute the determinant of the matrix
R4 =Toeplitz{r(0), r(1), r(2), r(3)}.
The value of the determinant can be computed without determining the autocorrelation using (7.4.19)
detR4 = 3m=0 Pm = r(0)P1 P2 P3
where r(0) = 1, and
P1 = P0 (1 − |k0 |2 ) = (1)(1 − 1/4) = 3/4
P2 = P1 (1 − |k1 |2 ) = (3/4)(1 − 1/9) = 2/3
P3 = P2 (1 − |k2 |2 ) = (2/3)(1 − 1/16) = 5/8

Therefore the detR4 = (3/4)(2/3)(5/8) = 5/16

7.24 (a) Determine the lattice second-order prediction error filter (PEF) for a sequence x(n) with autocorrelation
r(l) = ( 12 )|l| . (b) Repeat part (a) for the sequence y(n) = x(n) + v(n), where v(n) ∼ WN(0, 0.2) is
uncorrelated to x(n). (c) Explain the change in the lattice parameters using frequency domain reasoning (think
of the PEF as a whitening filter).
172 Statistical and Adaptive Signal Processing - Solution Manual

7.25 Consider a prediction error filter specified by P3 = ( 15

16
)2 , k0 = 1/4, k1 = 1/2, and k2 = 1/4.

(a) Determine the direct-form filter coefficients.

The forward predictor can be found directly using (7.5.27)
   ∗

am−1 J am−1
am = + km−1
0 1
a1 = k0 = 1/4
   
1/4 1/4
a2 = + (1/2)
0 1
 
3/8
=
1/2
   
3/8 1/2
a3 =  1/2  +  3/8  (1/4)
0 1
 
1/2
=  19/32 
1/4

(b) Determine the autocorrelation values r(1), r(2), and r(3).

The autocorrelation can be found directly using (7.5.36) and using the fact that
P3
r(0) = P0 =
2m=0 (1− |km |2 )
( 15 )2
= 16
(15/16)(3/4)(15/16)
= 4/3

then for m = 0
r(1) = −k0∗ P0 = −(1/4)(4/3) = −1/3
For m = 2
r(2) = −k1∗ P1 − a1(1)∗r(1) = −(1/2)(5/4) + (1/4)(1/3) = −13/24
where
P1 = P0 (1 − |k0 |2 ) = (4/3)(1 − |1/4|2 ) = 5/4
Finally, for m = 2
r(3) = −k2∗ P2 − [a1(2)∗r(2) + k1∗r(1)]
= −(1/4)(15/16) − [(3/8)(−13/24) + (1/2)(−1/3)
= 13/96
where
P2 = P1 (1 − |k1 |2 ) = (5/4)(1 − |1/2|2 ) = 15/16
Therefore the autocorrelation is r(0) = 4/3, r(1) = −1/3, r(2) = −13/24, and r(3) = 13/96
Statistical and Adaptive Signal Processing - Solution Manual 173

(c) Determine the value r(4) so that the MMSE P4 for the corresponding fourth-order filter is the minimum
possible.
The autocorrelation extension formula (7.5.36)
r(m + 1) = −km∗ Pm − amH J rm
can be used to compute the next autocorrelation value. Since the value of k3 is not set, an infinite number
of possible values can be used provided that |k3 | < 1, therefore insuring that R is positive definite.

7.26 Consider a prediction error filter A M (z) = 1+a1(M) z −1 +· · ·+a (M)

M z
−M
with lattice parameters k1 , k2 , . . . , , k M .
(a) Show that if we set k̂m = (−1)m km , then âm(M) = (−1)m am(M) .
For M = 2, the prediction error filter is
A2 = 1 + a1 z −1 + a2 z −2
= 1 + (k0∗ + k0 k1∗ )z −1 + k1∗ z −2

With the substitution of k̂m into the above equation

Â2 = 1 + â1 z −1 + â1 z −2
= 1 + (k̂0∗ + k̂0 k̂1∗ )z −1 + k̂1∗ z −2
= 1 + (−k0∗ − k0 k1∗ )z −1 − k1∗ z −2

then â1(2) = −a1(2) and â2(2) = a2(2) . ????

(b) What are the new filter coefficients if we set k̂m = ρ m km , where ρ is a complex number with |ρ| = 1?
What happens if |ρ| < 1 ?
If k̂0 = ρk0 and k̂1 = ρ 2 k1 where ρ = re j ω then
â1(2) = k̂0∗ + k̂0 k̂1∗
= ρ ∗ k0∗ + ρk0 ρ 2 k1∗
= re j ω k0∗ + re j ω k0r 2 e j 2ω k1∗
= re j ω (k0∗ + r 2 k0 k1∗ )

When r = 1 then â1(2) = e− j ω a1(2) . For â2(2)

â2(2) = k̂1∗ = ρ ∗2 k1∗ = r 2 e− j 2ω k1∗
so when r = 1, â2(2) = e− j 2ω a2(2)
7.27 Suppose that we are given the values {r(l)}m−1
−m+1 of an autocorrelation sequence such that the Toeplitz matrix
Rm is positive definite.

(a) Show that the values of r(m) such that Rm+1 is positive definite determine a disk in the complex plane.
Find the center α m and the radius ζ m of the disk
Figure similar to 6.3 in Haykin
If Rm is positive definite, then the necessary and sufficient condition for Rm+1 to be positive definite is
r(m + 1) = −km∗ Pm − amH Jrm
for any value of km such that |km | < 1. This establishes a circular region in the Real plane of possible
values for r(m + 1), with the circle centered at {Re[amH Jrm ],Im[amH Jrm ]} and a maximum radius of Pm .
174 Statistical and Adaptive Signal Processing - Solution Manual

∞
(b) By induction show that there are infinitely many extension of {r(l)}m−1
−m+1 that make {r(l)}−∞ a valid
autocorrelation sequence.
Any choice of |km | < 1 provides a valid r(m + 1). Hence, there are infinitely many extensions.

7.28 Consider the MA(1) sequence x(n) = w(n) + d1 w(n − 1), w(n) ∼ WN(0, σ 2w ). Then

r(0) = σ 2w 1 + |d1 |2 , r(1) = r ∗ (−1) = d1 σ 2w , and r(l) = 0, |l| > 0 (1)

and
 
r(0) r(1) 0 ··· 0 0
 r ∗ (1) r(0) r(1) · · · 0 0 
 
 . 
 0
 r ∗ (1) r(0) . . 0 0  
Rm =  . .. .. .. .. ..  (2)
 .. . . . . . 
 
 .. 
 0 0 0 . r(0) r(1) 
0 0 0 ··· r ∗ (1) r(0)

(a) Show that

det Rm = r(0) det Rm−1 − |r(1)|2 Rm−2 m≥2

From (2) for m ≥ 2

 
r(0) r(1) · · · 0 0
 ∗ 
 r (1) r(0) . . . 0 0  
 
 .. . .. . .. . .. .. 
. 
det Rm = r(0) det  . −
 
 .. 
 0 0 . r(0) r(1) 
0 0 · · · r ∗ (1) r(0)
 
..
 r(0) . 0 0 
 . . . .. . .. .. 
 . .  
r(1)r ∗ (1) det  
 .. 
 0 . r(0) r(1) 
0 · · · r ∗ (1) r(0)

where the first determinant is of (m − 1) × (m − 1) matrix Rm−1 and the second determinant is of
(m − 2) × (m − 2) matrix Rm−2 . Hence

det Rm = r(0) det Rm−1 − |r(1)|2 det Rm−2 , m ≥ 2 (3)

(b) Show that km = −r m (1)/ det Rm and that

1 r(0) 1 r ∗ (1) 1
=− −
km r(1) km−1 r(1) km−2
The reflection coefficients ki ’s are given by (4.2.23)
 (m)   
a1 r(1)
 a (m)   0 
 2   
Rm  .  = −  . 
 ..   .. 
km 0
Statistical and Adaptive Signal Processing - Solution Manual 175

Hence
 
r(0) r(1) 0 ··· 0 −r(1)
 r ∗ (1) r(0) r(1) · · · 0 0 
 
 . 
 0
 r ∗ (1) r(0) . . 0 0  
det  . .. .. .. .. .. 
 .. . . . . . 
 
 .. 
 0 0 0 . r(0) 0 
0 0 0 ··· r ∗ (1) 0 −r m (1)
km = = (4)
det Rm det Rm
Dividing both sides of (3) by −r m (1), we obtain
det Rm r(0) det Rm−1 det Rm−2
= − |r(1)|2
−r m (1) −r (1)
m −r m (1)



1 r(0) det Rm−1 r(1)r ∗ (1) det Rm−2
= − −
km −r(1) −r m−1 (1) r(1)r(1) −r m−2 (1)
∗
1 r(0) 1 r (1) 1
= − − (5)
km −r(1) km−1 r(1) km−2
(c) Determine the initial conditions and solve the recursion in (b) to show that
(1 − |d1 |2 )(−d1 )m
km =
1 − |d1 |2m+2
which tends to zero as m → ∞.
Note that (5) which is a difference equation in p(m) , 1
km+3
, m ≥ 0 with initial conditions (see Prob-
lem 4.25)
1 1 + |d1 |2 1 1 + |d1 |6
p(−2) = = and p(−1) = = − (6)
k1 (−d1 ) k2 1 − |d1 |2 (−d1 )2
Using unilateral z-transform along with (6) with fair amount of algebra, we obtain
(1 − |d1 |2 )(−d1 )m
km =
1 − |d1 |2m+2
which tends to zero as m → ∞ since |d1 | < 1.

7.29 Proof of (7.6.6).

From (7.6.3), substituting m0l for l, we have
∗ H∗ ∗
m (m − l) = r (m − l) + am r̃m (m − l − 1) = r(l − m) + am r̃m (m − l − 1)
ξ f∗ H∗
(1)

Using bm = J a∗m or amH∗ = bH

m J in (1), we have
∗
m (m − l) = r(l − m) + bm J r̃m (m − l − 1)
ξ f∗ H
(2)

From (7.6.5)
 
J r̃∗m (m − l − 1) = J r ∗ (m − l − 1) r ∗ (m − l − 2) · · · r ∗ (−l)
 
= J r(l − m + 1) r(l − m + 2) · · · r(l)
 
= r(l) r(l − 1) · · · r(l − m + 1) = r̃m (l) (3)
176 Statistical and Adaptive Signal Processing - Solution Manual

ξ − ξ − ξ
− −

− −

− −

ξ − − ξ − − ξ
− −

Figure 7.30: Flow diagram of the lattice filter with the lattice parameters

Substituing (3) in (2), we obtain

m (m − l) = r(l − m) + bm J r̃m (l) = ξ m (l)

ξ f∗ H b

which proves (7.6.6).

7.30 In this problem we show that the lattice parameters can be obtained by "feeding" the autocorrelation sequence
through the lattice filter as a signal and switching on the stages one by one after the required lattice coefficient
is computed. The value of km is computed at time n = m from the inputs to stage m.

(a) Using (7.6.10), draw the flow diagram of a third-order lattice filter that implements this algorithm.
Figure 7.30 shows the flow diagram of the lattice filter with the lattice parameter computed per

ξ mf (l)
km = − (1)
ξ bm (m)

(b) Using the autocorrelation sequence in Example 7.6.1, "feed" the sequence {r(n)}30 = {3, 2, 1, 12 } through
the filter one sample at a time, and compute the lattice parameters. H int : Use Example 7.6.1 for
guidance.
Using (7.6.10)
f
ξ 0 (l) = ξ b0 (l) = r(l)
f ∗
ξ mf (l) = ξ m−1 (l) + km−1 ξ bm−1 (l − 1)
f
ξ bm (l) = ξ bm−1 (l − 1) + km−1 ξ m−1 (l)

and the flow diagram above, the lattice parameters can be computed using the autocorrelation as the input.
For each instance in time, the lattice has a number of stages equal to the time index starting with time
f
equal zero. With l = 0, the first stage is ξ 0 (0) = ξ b0 (0) = r(0) = 3. The second stage is computed with
l = 1 and r(1) = 2 as the input, with the following output
f
ξ 0 (1) = ξ b0 (1) = r(1) = 2
f
−ξ 0 (1)
k0 = = −2/3
ξ b0 (0)
f
ξ b1 (1) = ξ b0 (0) + k0 ξ 0 (1) = 3 + (−2/3)(2) = 5/3
Statistical and Adaptive Signal Processing - Solution Manual 177

The second lattice parameter is computed with l = 2 and r(2) = 1

f
ξ 0 (2) = ξ b0 (2) = r(2) = 1
ξ 1 (2) = ξ 0 (2) + k0∗ ξ b0 (1) = 1 + (−2/3)(2) = −1/3
f f

f
ξ b1 (2) = ξ b0 (1) + k0 ξ 0 (2) = 2 + (−2/3)(1) = 4/3
f
ξ 1 (2) −1/3
k1 = − =− = 1/5
ξ 1 (1)
b 4/3
f
ξ b2 (2) = ξ b1 (1) + k1 ξ 1 (2) = 5/3 + (1/5)(−1/3) = 8/5

and the last lattice parameter is computed with l = 3 and r(3) = 1/2
f
ξ 0 (3) = ξ b0 (3) = r(3) = 1/2
ξ 1 (3) = ξ 0 (3) + k0∗ ξ b0 (2) = 1/2 + (−2/3)(1) = −1/6
f f

ξ 2 (3) = ξ 1 (3) + k1∗ ξ b1 (2) = (−1/6) + (1/5)(4/3) = 1/10

f f

f
ξ 2 (3) 1/10
k2 = − =− = −1/16
ξ 2 (2)
b 8/5

Therefore the lattice parameters are k0 = −2/3, k1 = −1/4, and k2 = −1/16

7.31 Draw the superlattice structure for M = 8, and show how it can be partitioned to distribute the computations
to three processors for parallel execution.
Figure 7.31 shows the superlattice structure, and a partition of computations for three parallel processors.

7.32 Derive the superladder structure shown in Figure 7.10.

The supperladder structure shown in Figure 7.10 are computed using (7.6.15), (7.6.16), and (7.6.17). Each
stage of the supperladder is shown below. Given the autocorrelation and the cross-correlation parameters, the
first stage, working left to right, is

ξ c0 (0) = d1
ξ c0 (1) = d2
ξ c0 (2) = d3
ξ c0 (3) = d4
ξ b0 (0) = r(0)
ξ b0 (1) = r(1)
ξ b0 (2) = r(2)
ξ b0 (3) = r(3)

and the lattice-ladder coefficient k0c is

−ξ c0 (0)
k0c =
ξ b0 (0)
178 Statistical and Adaptive Signal Processing - Solution Manual

ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ

ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ
ξ

ξ ξ

ξ
ξ

Figure 7.31: Superlattice structure

Statistical and Adaptive Signal Processing - Solution Manual 179

The second stage is

ξ c1 (1) = ξ c0 (1) + k0c ξ b0 (1)

ξ c1 (2) = ξ c0 (2) + k0c ξ b0 (2)
ξ c1 (3) = ξ c0 (3) + k0c ξ b0 (3)

and the lattice-ladder coefficient k1c is

−ξ c1 (1)
k1c =
ξ b1 (1)
The third stage is

ξ c2 (2) = ξ c1 (2) + k1c ξ b1 (2)

ξ c2 (3) = ξ c1 (3) + k1c ξ b1 (3)

and the lattice-ladder coefficient k1c is

−ξ c2 (2)
k2c =
ξ b2 (2)
Finally, the last stage is

ξ c3 (3) = ξ c2 (3) + k2c ξ b2 (3)

and
−ξ c3 (3)
k3c =
ξ b3 (3)

7.33 Extend the algorithm of Schür to compute the LDL H decomposition of a Hermitian Toeplitz matrix, and write
a Matlab function for its implementation.
The L D L H decomposition of a Hermitian Toeplitz matrix can be accomplished using (7.7.12)
 
1 0 ··· 0 0
 ξ̄ b (1) ··· 0 
 0 1 0 
 b b 
−1
L=B = 0  ξ̄ (2) ξ̄ (1) · · · 0 0 
1 
 .. .. .. .. .. 
 . . . . . 
b b b
ξ̄ 0 (M) ξ̄ 1 (M) · · · ξ̄ M−1 (M) 1
b ξ bm (l) ξ bm (l)
where ξ̄ m (l) = ξ bm (m)
= Pm
and D =diag{P0 , P1 , . . . , PM }.

7.34 Given the matrix R3 =Toeplitz{1, 1/2, 1/2}, use the appropriate order-recursive algorithms to compute the
following:

(a) The L D L H and U DU H decompositions of R

The decompositions of R into L D L H can be done using (7.1.38)

Rm+1 = L m+1 Dm+1 L m+1

H
180 Statistical and Adaptive Signal Processing - Solution Manual

where L m+1 and Dm+1 are generated from L m and Dm using (7.1.37)
   
Lm 0 Dm 0
L m+1 = D =
lmH 1 m+1
0 H ξ m+1
and ξ m+1 =detRm+1 /detRm . The vector lm can be found by solving (7.1.39)

(L m Dm )lm = rmb

Starting with m = 1, then L 1 = 1, D1 = 1, and r1b = 1/2 so

(L 1 D1 )l1 = r1b ⇒ l1 = 1/2

and ξ 2 =detR2 /detR1 = 3/4. For m = 2

   
1 0 1 0 1/2
(L 2 D2 )l2 = ⇒ r2b l =
1/2 1 0 3/4 2 1/2
     
1 0 1/2 1/2
l = ⇒ l2 =
1/2 3/4 2 1/2 1/3

and ξ 3 =detR3 /detR2 = 1/2

3/4
= 2/3. So for m = 3 the L D L H decomposition is
   
  1 0 0   1 0 0
Lm 0 Dm 0
L m+1 = =  1/2 1 0  Dm+1 = =  0 3/4 0 
lmH 1 0 H
ξ m+1
1/2 1/3 1 0 0 2/3

The U DU H decomposition is easily found using (7.7.15)

R = J R∗ J = (J L ∗ J )(J D J )(J L H J ) = U D̄U H

(b) The L D L H and U DU H decompositions of R−1

The L D L H decomposition of R−1 can be computed in an order-recursive manor using
 −1  −1   −1 
−1 Łm 0 Łm 0 Łm 0
L m+1 = = =
lmH 1 vmH 1 bmH 1

where bm is computed as shown in part (c) below. Also Dm =diag{P0b , P1b , . . . , Pm−1
b
}, where Pmb is

Pmb = ρ bm − rmb H Rm−1 rmb = ρ bm + rmb H bm = α bm

where α bm is computed in part (c) below. Then L −1

m is
 
−1 1 0
b1 = −1/2 ⇒ L 2 =
−1/2 1
 b   b   
P0 0 α1 0 1 0
D2 = = =
0 P1b 0 α b2 0 3/4
 
  1 0 0
−1/3 
b1 = −1
⇒ L 3 = −1/2 1 0 
−1/3
−1/3 −1/3 1
 b   b   
P0 0 0 α1 0 0 1 0 0
D3 =  0 P1b 0  =  0 α b2 0  =  0 3/4 0 
0 0 P2 b
0 0 α3b
0 0 2/3
Statistical and Adaptive Signal Processing - Solution Manual 181

(c) The inverse matrix R−1 .

The order-recursive algorithm for the decomposition of the inverse of R is (7.1.24)
 −1  −1   
−1 Rm rmb Rm 0m 1 bm
Rm+1 = = + b [bmH 1]
rmb H ρ bm 0mH 0 αm 1
where bm and α bm are found using (7.1.22) and (7.1.23) respectively. Starting with m = 1, R1 = 1
therefore R1−1 = 1, r1b = 1/2, and ρ b1 = 1. Then R2−1 can be computed as
b1 = −R1−1 r1b = −(1)(1/2) = −1/2
α b1 = ρ b1 + r2b H b2 = 1 + (1/2)(−1/2) = 3/4
 −1   
−1 R1 01 1 b1
R2 = + b [b1H 1]
01H 0 α1 1
   
1 0 4 −1/2
= + [−1/2 1]
0 0 3 1
 
4/3 −1/2
=
−1/2 4/3

The above algorithm is repeated for m = 2, with r2b = [1/2 1/2]T and ρ b2 = 1
    
−1 b 4/3 −1/2 1/2 −1/3
b2 = −R2 r2 = − =
−1/2 4/3 1/2 −1/3
 
−1/3
α b2 = ρ b2 + r2b H b2 = 1 + [1/2 1/2] = 2/3
−1/3
 −1   
−1 R2 02 1 b2
R3 = + b [b2H 1]
02H 0 α2 1
   
7/3 −2/3 0 −1/3
3
=  −2/3 4/3 0  +  −1/3  [−1/3 − 1/3 1]
2
0 0 0 1
 
3/2 −1/2 −1/2
=  −1/2 3/2 −1/2 
−1/2 −1/2 3/2

7.35 Consider the AR(1) process x(n) = ρx(n − 1) + w(n), where w(n) ∼ WN(0, σ 2w ) and −1 < ρ < 1. The
z-domain PSD of x(n) is given by
σ 2w 1
Rx (z) =  ; < |z| < |ρ| , −1 < ρ < 1
(1 − ρ z) 1 − ρ z −1 |ρ|
σ 2w
and the correlation sequence is given by r x (l) = ρ |l| .
1 − ρ2
(a) Determine the correlation matrix R M+1 of the process.
For an arbitrary value of M, we have
 
1 ρ ρ2 ··· ρM
 ρ 1 ρ ··· ρ M−1 
 
 ρ2 ρ ··· ρ M−2 
R M+1 = 1 
 .. .. .. .. .. 
 . . . . . 
ρM ρ M−1 ρ M−2 · · · 1
182 Statistical and Adaptive Signal Processing - Solution Manual

The following Matlab script is given for M = 3, ρ = 0.8, and σ 2w = 1.

rho = 0.8; M = 3; var_w = 1;

% (a) Correlation Matrix

R = var_w/(1-rho*rho)*toeplitz(rho.^[0:M])
R =
2.7778 2.2222 1.7778 1.4222
2.2222 2.7778 2.2222 1.7778
1.7778 2.2222 2.7778 2.2222
1.4222 1.7778 2.2222 2.7778
(b) Determine the Mth-order FLP, using the algorithm of Levinson-Durbin.
The following Matlab script is given for M = 3, ρ = 0.8, and σ 2w = 1.
% (b) Mth-order FLP using Levinson-Durbin
[a,k,Po] = durbin(R(:,1),M); a
a =
1.0000
-0.8000
-0.0000
0
(c) Determine the inverse matrix R−1
M+1 , using the triangular decomposition discussed in Section 7.7.

The following Matlab script is given for M = 3, ρ = 0.8, and σ 2w = 1.

% (c) Inverse of the R matrix
Q = invtoepl(R(:,1),M+1)
Q =
1.0000 -0.8000 -0.0000 0
-0.8000 1.6400 -0.8000 -0.0000
-0.0000 -0.8000 1.6400 -0.8000
0 -0.0000 -0.8000 1.0000
% Check
R_inv = inv(R);
err = max(max(abs(Q-R_inv)))
err =
4.4409e-016

7.36 If r(l) = cos ω0l, determine the second-order prediction error filter and check whether it is minimum-phase.
The second-order forward prediction error filter is given by

e(n) = x(n) + a1 x(n − 1) + a2 x(n − 2)

where prediction error filter coefficients are given by R a = −r, that is,
    
r(0) r(1) a1 r(1)
=−
r(1) r(0) a2 r(2)
or
    
1 cos ω0 a1 cos ω0
=−
cos ω0 1 a2 cos 2ω0
Statistical and Adaptive Signal Processing - Solution Manual 183

or
   cos ω0 ω0
  
a1 cos2 ω0 −1
− coscos
2 ω −1 cos 2ω 0 −2 cos ω0
= 0
=
− coscos2 ωω0 −1
2
a2 0
+ cos2 1ω0 −1 cos 2ω0 1

Thus the prediction error filter is

e(n) = x(n) − 2 cos ω0 x(n − 1) + x(n − 2)
with zeros at z = ±]ω0 , that is, on the unit circle. Thus the system is (critically) minimum-phase system.
7.37 Show that the MMSE linear predictor of x(n + D) in terms of x(n), x(n − 1), . . . , x(n − M + 1) for D ≥ 1
is given by
Ra(D)= −r(D)
where r(D) = [r(D) r(D + 1) · · · r(D + M − 1)]T .
The predictor is given by

M−1
x̂(n + D) = − ak x(n − k)
k=0

and the prediction error is given by


M−1
e D (n) , x(n + D) − x̂(n + D) = x(n + D) + ak x(n − k)
k=0

Using the orthogonality principal, e D (n) ⊥ x(n − l), l = 0, 1, ..., M − 1, we obtain

  

M−1
E x(n + D) + ak x(n − k) x(n − l) = 0, l = 0, 1, ..., M − 1
k=0
or

M−1 
M−1
r(D + l) + ak r(l − k) = 0 ⇒ ak r(l − k) = −r(D + l), l = 0, 1, ..., M − 1
k=0 k=0
or
    
r(0) r(−1) ··· r(1 − M) a0 r(D)
 r(1) r(0) ··· r(2 − M)  a1   r(D + 1) 
    
 .. .. .. ..  ..  = − .. 
 . . . .  .   . 
r(M − 1) r(M − 2) · · · r(0) a M−1 r(D + M − 1)
R a(D) , −r(D)

Develop a recursion that computes a(D+1) from a(D) by exploring the shift invariance of the vector r(D).
We want to determine a(D+1) (which is otherwise obtained from R a(D+1) , −r(D+1) ) directly from a(D). We
will use subscripts to denote the size of a matrix or vector. Consider the partitioning
   
r(D)  r(D) 
 r(D + 1)   r(D + 1) 
(D)    
RM aM = −  ..  =   ..   (1)
 .    .  
r(D + M − 1) r(D + M − 1)
184 Statistical and Adaptive Signal Processing - Solution Manual

and
   
r(D + 1)
    ..
    .
R M a(D+1) = −   (2)
M
 r(D + M − 1) 
r(D + M)

Since R M is Hermitian Toeplitz, we have R M = JR M J, m = 1, . . . , M. Substituting in (1), we have

 
 r(D) 
 r(D + 1) 
 
JR M J a(D) = −   .  
M
  .
.  
r(D + M − 1)

Premultiplying by J
   
r(D + M − 1)
    ..  
  .  
R M J a(D) = −   (3)
M
 r(D + 1) 
r(D)

Now consoder the system of order(M − 1) from (2)

 
r(D + 1)
 .. 
R M−1 a(D+1)
M−1 = −  . 
r(D + M − 1)
or
 
  r(D + M − 1)
 .. 
R M−1 Ja(D+1)
M−1 = − .  (4)
r(D + 1)

From (3) and (4), we have using the Levinson Algorithm

 (D+1)   (D) 
(D) Ja M−1 (D) b M−1
J aM = + kM
0 0
or
   
Ja(D+1) b(D)
M−1 = J a(D) (D)
M − kM
M−1 : (Step-down recursion) (5)
0 0

Finally, to obtain a(D+1)

M from a(D+1)
M−1 , we again use the Levinson recursion
   
a(D+1) b(D)
a(D+1)
M = M−1 + k (D+1)
M
M−1 : (Step-up recursion) (6)
0 0

Thus the overall approach requires a step-down and step-up recursions.

Statistical and Adaptive Signal Processing - Solution Manual 185

7.38 The normal equations for the optimum symmetric signal smoother (see Section 6.5.1) can be written as
 
0
 
R2m+1 c2m+1 =  P2m+1 
0

where P2m+1 is the MMSE, c2m+1 = Jc∗2m+1 , and cm(2m+1) = 1. (a) Using a “central” partitioning of R2m+3
and the persymmetry property of Toeplitz matrices, develop a recursion to determine c2m+3 from c2m+1 . (b)
Develop a complete order-recursive algorithm for the computation of {c2m+1 , P2m+1 }0M (see Kok et al. 1993).
7.39 Using the triangular decomposition of a Toeplitz correlation matrix, show that

(a) the forward prediction errors of various orders and at the same time instant, that is,
ef (n) = [e0f (n) e1f (n) · · · emf (n)]T
are correlated.
 m
Using the definition of ekf (n) k=0 , we obntain
  
1 0 0 ··· 0 x(n)
 1 a1(1)∗ 0 ··· 0  x(n − 1) 
  
 1 a (2)∗ a (2)∗ · · ·  
e (n) = 
f
1 2 0  x(n − 2)  , Ā M+1 x M+1 (n)
 . . .. .. ..  .. 
 .. .
. . . .  . 
1 a1(M)∗ a2(M)∗ · · · a (M)∗
M
x(n − M)
Hence
   
E ef (n)efH (n) = Ā M+1 E x M+1 (n)x M+1 (n) ĀHM+1 = Ā M+1 R M+1 ĀHM+1

Using triangular decomposition of R M+1 = A−1 −H

M+1 D̄ M+1 A M+1 from (7.7.8), we have
 
E ef (n)efH (n) = Ā M+1 A−1 −H H
M+1 D̄ M+1 A M+1 Ā M+1

which, in general, is not a diagonal matrix. Hence components of ef (n) are correlated.
(b) the forward prediction errors
ēf (n) = [efM (n) efM−1 (n − 1) · · · e0f (n − M)]T
are uncorrelated.
Using (7.3.66)
ēf (n) = A M+1 x M+1 (n)
Hence
   
E ēf (n)ēfH (n) = A M+1 E x M+1 (n)x M+1 (n) AHM+1 = A M+1 R M+1 AHM+1
where
 
1 a1(M)∗ a2(M)∗ ··· a (M)∗
M
 0 1 a1(M−1)∗ ··· a (M−1)∗ 
 M−1 
 ··· a (M−2)∗ 
A M+1 = 0 0 1 
 .. .. .. ..
M−2
.. 
 . . . . . 
0 0 0 ··· 1
186 Statistical and Adaptive Signal Processing - Solution Manual

Using triangular decomposition of R M+1 = A−1 −H

M+1 D̄ M+1 A M+1 from (7.7.8), we have
 
E ēf (n)ēfH (n) = A M+1 A−1 −H
M+1 D̄ M+1 A M+1 A M+1 = D̄ M+1
H

which is a diagonal matrix. Hence components of ēf (n) are uncorrelated.

7.40 Generalize the inversion algorithm described in Section 7.7.3 to handle Hermitian Toeplitz matrices.

   
R M−1 Q M−1 q
Given: R M = , Determine: Q M =
rH ρ qH q

where
1 1
q = −1
, , b = −R−1
M−1 r
ρ − r R M−1 r
H ρ + r Hb

q = − R−1
M−1 r q = bq
Q M−1 = R−1
M−1 + qbb
H

Let P = 1/q which gives q = b/P, b = q P and Q M−1 = R−1 M−1 + Pqq or
H

< =
Q M−1 i j = R−1 ∗
M−1 i j + Pqi q j (1)
 −1 ∗
Using the fact that R M−1 is Hermitian Toeplitz, that is, JR M−1 J = R∗M−1 or R−1
M−1 = J R M−1 J (Per-
Hermitian), we obtain
< −1 = < −1 = ∗
R M−1 = R M−1 M− j,M−i

Substituting in (1)
< = ∗
Q M−1 i j = R−1
M−1 M− j,M−i
+ Pqi q∗j (2)

But also from (1)

< = ∗ < = ∗  ∗
Q−1
M−1 M− j,M−i
= R−1
M−1 M− j,M−i
+ P q M− j q∗M−i (3)

Subtracting (3) from (2)

< = ∗  ∗
Q M−1 i j − Q−1
M−1 M− j,M−i
= Pqi q∗j − P q M− j q∗M−i

or
< = ∗ 
Q M−1 i j = Q−1
M−1 M− j,M−i
+ P qi q∗j − q∗M− j q M−i

The Matlab function invtoeplC is shown below.

function Q = invtoeplC(r,M)
% Computes inverse of a Hermitian Toeplitz matrix R
% Q = invtoeplC(r,M)
% Input: First column of Hermitian Toeplitz matrix R
% r(m), 0 <= m <= M-1
Statistical and Adaptive Signal Processing - Solution Manual 187

% Output: Q = Inverse of R using Section 7.7.3

%
% Programmed by: Vinay K. Ingle, 1999
%
%-----------------------------------------------------------
% Copyright 2000, by Dimitris G. Manolakis, Vinay K. Ingle,
% and Stephen M. Kogon. For use with the book
%
% "Statistical and Adaptive Signal Processing"
%
% McGraw-Hill Higher Education.
%
%-----------------------------------------------------------

% Requires the function sympersym

M = length(r);

% Determine vector b using the durbin function

[a,k,Po] = durbin(r,M-1);
b = flipud(a(2:M));
%R = toeplitz(r,conj(r)); b = -R(1:M-1,1:M-1)\flipud(conj(r(2:M)));

% Determine P and q
P = r(1) + (conj(r(M:-1:2)))’*b;
q = b/P;

% Computation of Q matrix using eq. (7.7.24)

QM = zeros(M); % Initialize QM
QM(1:M-1,M) = q; QM(M,M) = 1/P; % Last column of QM
QM = sympersymC(QM,QM(:,M)); % symmetry and persymmetry operations
Q = QM(1:M-1,1:M-1); % Extract Q

Mend = ceil((M+1)/2);
for j = M-1:-1:Mend
for i = (M-j+1):1:j
% Use eq. (7.7.24)
Q(i,j) = conj(Q(M-j,M-i)) - P*(conj(q(M-j))*q(M-i)-q(i)*conj(q(j)));
end
% Perform symmetry and persymmetry operations
QM = sympersymC(QM,Q((M-j+1):1:j,j));
Q = QM(1:M-1,1:M-1);
end
Q = QM;

7.41 Consider the estimation of a constant α from its noisy observations. The signal and observation models are
188 Statistical and Adaptive Signal Processing - Solution Manual

given by
y(n + 1) = y(n) n>0 y(0) = α
x(n) = y(n) + v(n) v(n) ∼ WGN(0, σ 2v )

(a) Develop scalar Kalman filter equations, assuming the initial condition on the a posteriori error variance
R ỹ (0|0) equal to r0 .
Signal model:

y(n) = y(n − 1), n > 0, y(0) = α ⇒ A(n − 1) , A = 1, B(n) , B = 0

Observation model:

x(n) = y(n) + v(n) ⇒ H(n) , H = 1

Signal Prediction:

ŷ(n|n − 1) = A ŷ(n − 1|n − 1) = ŷ(n − 1|n − 1), ŷ(1|0) = y(0) = α

Signal filter:

ŷ(n|n) = ŷ(n|n − 1) + K (n) x(n) − H ŷ(n|n − 1)

= ŷ(n|n − 1) + K (n) x(n) − ŷ(n|n − 1)

A-priori error covariance:

R ỹ (n|n − 1) = AR ỹ (n|n − 1)A + B Rη (n)B

= R ỹ (n − 1|n − 1), R ỹ (0|0) , r0

Kalman gain:
 −1
K (n) = R ỹ (n|n − 1)H R ỹ (n|n − 1) + Rv (n)
R ỹ (n|n − 1) R ỹ (n|n − 1)
= =
R ỹ (n|n − 1) + Rv (n) R ỹ (n|n − 1) + σ 2v
A-posteriori error covariance:

R ỹ (n|n) = [1 − K (n)H ] R ỹ (n|n − 1) = [1 − K (n)] R ỹ (n|n − 1)

(b) Show that the a posteriori error variance R ỹ (n|n) is given by

r0
R ỹ (n|n) = (1)
1 + (r0 /σ 2v )n
From part (a) we have
R ỹ (n|n − 1) R ỹ (n − 1|n − 1)
K (n) = =
R ỹ (n|n − 1) + σ 2v R ỹ (n − 1|n − 1) + σ 2v
with
 
R ỹ (n − 1|n − 1)
R ỹ (n|n) = [1 − K (n)] R ỹ (n|n − 1) = 1 − R ỹ (n − 1|n − 1)
R ỹ (n − 1|n − 1) + σ 2v
σ 2v R ỹ (n − 1|n − 1)
=
R ỹ (n − 1|n − 1) + σ 2v
Statistical and Adaptive Signal Processing - Solution Manual 189

Thus for n = 1:
r0
r0 σ 2v r0
K (1) = = , R ỹ (1|1) =
r0 + σ v
2 1 + σr02 1 + σr02
v v

n = 2:
r0
r0 σ 2v r0
K (2) = = , R ỹ (2|2) =
2r0 + σ v
2 1 + 2 σr02 1 + 2 σr02
v v

Hence
r0
r0 σ 2v r0
K (n) = = , R ỹ (n|n) =
nr0 + σ 2v 1 + n σr02 1 + n σr02
v v

(c) Show that the optimal filter for the estimation of the constant α is given by

r0 /σ 2v
ŷ(n) = ŷ(n − 1) + [x(n) − ŷ(n − 1)]
1 + (r0 /σ 2v )n
Using K (n) from the above part we obtain the desired result.

7.42 Consider a random process with PSD given by

4
Rs (e j ω ) =
2.4661 − 1.629 cos ω + 0.81 cos 2ω

(a) Using Matlab, plot the PSD Rs (e j ω ) and determine the resonant frequency ω0 .
% Given PSD
w = [0:1:500]*pi/500;
Rs = 4./(2.4661-1.629*cos(w)+0.81*cos(2*w));

% (a) Plot of PSD and Resonant frequency

[maxRs,I_max] = max(Rs); w_max = w(I_max);

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3],...
’numbertitle’,’off’,’name’,’P0742a’);
plot(w/pi,Rs,’g’,’linewidth’,1.5); hold on; axis([0,1,0,3.5]);
plot([w_max,w_max]/pi,[0,maxRs],’r:’,[w_max/pi],[maxRs],’md’);
xlabel(’Digital frequency in \pi units’,’fontsize’,8);
ylabel(’{\itR}_{\its}( e^{j\omega} )’,’fontsize’,8);
title(’Power Spetral Density’,’fontsize’,10,’fontweight’,’bold’);
text(w_max/pi+0.02,0.5,[’Resonant frequency: ’,num2str(w_max),’ radians’]);
The plot is shown in Figure 7.42a.
(b) Using spectral factorization, develop a signal model for the process of the form

y(n) = Ay(n − 1) + Bη(n)

s(n) = [1 0]y(n)

where y(n) is a 2 × 1 vector, η(n) ∼ WGN(0, 1), and A and B are matrices with appropriate dimensions.
190 Statistical and Adaptive Signal Processing - Solution Manual

Power Spetral Density

3.5

2.5

2
Rs( ejω )

1.5

0.5 Resonant frequency: 1.043 radians

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Digital frequency in π units

Figure 7.42a: Plot of the PSD Rs (e j ω )

 
e j ω +e− j ω z+z −1  e j 2ω +e− j 2ω z 2 +z −2 
Substituting cos ω = 2
= 2 z=e j ω
and cos 2ω = 2
= 2  jω in Rs (e j ω )
z=e

4
Rs (z) = −1 2 +z −2
2.4661 − 1.629 z+z2 + 0.81 z 2
40000z 2
=
4050z 4 − 8145z 3 + 24661z 2 −8145z


+ 4050 
3.1427 3.1427
=
1 − 0.3158z −1 + 0.1863z −2 z 2 − 1.6953z + 5.3674
Thus the AR model is

y(n) = 0.3158 y(n − 1) − 0.1863 y(n − 2) + 3.1427 η(n)

or the state equations are:

      
y(n) 0.3158 −0.1863 y(n − 1) 3.1427
= + η(n)
y(n − 1) 1 0 y(n − 2) 0
y(n) , A y(n) + B η(n)

 
  y(n)
s(n) = y(n) = 1 0
y(n − 1)
, H y(n)

% (b) Spetcral factorization

B = 8; Rs_den = [0.81,-1.629,4.9322,-1.629,.81];
B = B/Rs_den(1); Rs_den = Rs_den/Rs_den(1);
poles = roots(Rs_den);
a_out = poly(poles(1:2)); a_in = poly(poles(3:4)); b = sqrt(B);
Statistical and Adaptive Signal Processing - Solution Manual 191

% Verification
if 0
N = 8192; eta = randn(N,1);
sn = filter(b,a_in,eta);
Nfft = 512; L = 16; w_ham = hamming(L)’; Fs = 1;
Rs_hat = psd(sn,Nfft,Fs,w_ham,L/2,’none’);
figure;
w = [0:1:Nfft/2]*Fs/Nfft; plot(w,Rs_hat); axis([0,Fs/2,0,18]);
end
(c) Let x(n) be the observed values of s(n) given by

x(n) = s(n) + v(n) v(n) ∼ WGN(0, 1)

Assuming reasonable initial conditions, develop Kalman filter equations and implement them, using
Matlab. Study the performance of the filter by simulating a few sample functions of the signal process
s(n) and its observation x(n).
% (c) Simulation

% Generate the AR(2) process s(n)

a = a_in; b = sqrt(B);
N = 100; n = [0:N];
eta = randn(N+1,1);
sn = filter(b,a,eta);
s_std = std(sn);

% Generate vector process y(n)

yn = sn; y_std = s_std;

% Generate the observation process x(n)

v = randn(N+1,1);
yn1 = [0;yn(1:N)];
h1 = 1.0; h2 = 0.0; g1 = sqrt(1);
xn = h1*yn + h2*yn1 + g1*v;
x_std = std(xn);
in_snr = 20*log10(y_std/g1)

% Design KF parameters
A = [-a(2),-a(3);1,0]; B = [b;0];
H = [h1,h2]; G = g1;
Rv = g1*g1; R_eta = eye(1);

% Initialization
y_post = [0;0]; R_post = zeros(2); I = eye(2);
y_hat = zeros(N+1,1); gain = zeros(N+1,2); mse = y_hat;

% Tracking
for n = 0:N
R_pri = A*R_post*A’ + B*R_eta*B’;
y_pri = A*y_post;
192 Statistical and Adaptive Signal Processing - Solution Manual

x_pri = H*y_pri;
Rw = H*R_pri*H’+Rv;
K = R_pri*H’*inv(Rw);
y_post = y_pri + K*(xn(n+1) - x_pri);
R_post = (I-K*H)*R_pri;
y_hat(n+1) = y_post(2);
gain(n+1,:) = K’;
mse(n+1) = R_post(1,1);
end
y_hat = [y_hat(2:N+1);y_hat(N+1)];
error = yn - y_hat; e_std = std(error);
out_snr = 20*log10(y_std/e_std)

%Plots
Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3],...
’numbertitle’,’off’,’name’,’P0742c1’);

n = 0:N;
plot(n,yn,’g’,n,xn,’w:’,n,y_hat,’c--’,’linewidth’,1);
ylabel(’Amplitude’,’fontsize’,8,’fontname’,’Helv’);
xlabel(’n’,’fontsize’,8,’fontname’,’Helv’);
legend(’y(n)’,’x(n)’,’yhat(n)’,0);
set(gca,’xtick’,[0:20:N],’ytick’,[-10:5:10],’fontname’,’times’,’fontsize’,6);
title(’Estimation of AR(2) Process’,’fontname’,’Helv’,’Fontsize’,10);

Hf_3 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,4,3],...
’numbertitle’,’off’,’name’,’P0742c2’);

subplot(2,1,1);
plot(n,gain(:,1),’g’,n,gain(:,2),’r--’,’linewidth’,1); axis([0,N,0,.1]);
legend(’K_1(n)’,’K_2(n)’,0);
ylabel(’Gains’,’fontsize’,8,’fontname’,’Helv’);
set(gca,’xtick’,[0:20:N],’ytick’,[0,0.1],’fontname’,’symbol’,’fontsize’,6);
title(’Kalman Gain Values’,’fontname’,’Helv’,’Fontsize’,10);

subplot(2,1,2);
plot(n,mse,’g’,’linewidth’,1); axis([0,N,0,1]);
ylabel(’MSE’,’fontsize’,8,’fontname’,’Helv’);
set(gca,’xtick’,[0:20:N],’ytick’,[0,1],’fontname’,’symbol’,’fontsize’,6);
title(’Mean Squared Error’,’fontname’,’Helv’,’Fontsize’,10);
xlabel(’n’,’fontsize’,8,’fontname’,’Helv’);
The plots are shown in Figures 7.42c1 and 7.42c2.

7.43 Alternative form of the Kalman filter. A number of different identities and expressions can be obtained for the
quantities defining the Kalman filter.
Statistical and Adaptive Signal Processing - Solution Manual 193

Estimation of AR(2) Process

y(n)
x(n)
yhat(n)

5
Amplitude

−5

−10
0 20 40 60 80 100
n

Figure 7.42c1: Plots of true and estimated values

Kalman Gain Values

0.1
Κ (ν)
1
Κ (ν)
2
Gains

0
0 20 40 60 80 100

Mean Squared Error

1
MSE

0
0 20 40 60 80 100
n

Figure 7.42c2: Plots of gains and MSE

194 Statistical and Adaptive Signal Processing - Solution Manual

(a) By manipulating the last two equations in (7.8.39) show that

R ỹ (n|n) = R ỹ (n|n − 1) − R ỹ (n|n − 1)HH (n)

(1)
× [H(n)R ỹ (n|n − 1)HH (n) + Rv (n)]−1 HR ỹ (n|n − 1)
Consider

R ỹ (n|n) = [I − K(n)H(n)] R ỹ (n|n − 1)

where K(n) = R ỹ (n|n − 1)HH (n)Rw−1 (n). Substituting K(n) in the above equation
 
R ỹ (n|n) = I − R ỹ (n|n − 1)HH (n)Rw−1 (n)H(n) R ỹ (n|n − 1)
= R ỹ (n|n − 1) − R ỹ (n|n − 1)HH (n)Rw−1 (n)H(n)R ỹ (n|n − 1)

Now substituting Rw (n) = H(n)R ỹ (n|n − 1)H H (n) + Rv (n) from (7.8.29) we obtain the desired result.
(b) If the inverses of R ỹ (n|n), R ỹ (n|n − 1), and Rv exist, then show that

R−1 −1 −1
ỹ (n|n) = R ỹ (n|n − 1) + H (n)Rv (n)H(n)
H
(2)

Consider
 −1
R−1 −1
ỹ (n|n) = R ỹ (n|n − 1) − R ỹ (n|n − 1)H (n)Rw (n)H(n)R ỹ (n|n − 1)
H

Using matrix inversion result

 −1
[A − A B A]−1 = A−1 + B−1 − A

we obtain
 −1 −1
R−1 −1
ỹ (n|n) = R ỹ (n|n − 1) + H (n)Rw (n)H
−H
(n) − R ỹ (n|n − 1)

Substituting Rw (n) = H(n)R ỹ (n|n − 1)H H (n) + Rv (n), we obtain

 −1 −1
R−1 −1
ỹ (n|n) = R ỹ (n|n − 1) + H (n)Rv (n)H
−H
(n)
= R−1 −1
ỹ (n|n − 1) + H (n)Rv (n)H (n)
H 1

This shows that the update of the error covariance matrix does not require the Kalman gain matrix (but
does require matrix inverses).
(c) Finally show that the gain matrix is given by

K(n) = R ỹ (n|n)HH (n)Rv−1 (n) (3)

which is computed by using the a posteriori error covariance matrix.

Premultiplying (2) by R ỹ (n|n), we obtain

I = R ỹ (n|n)R−1 −1
ỹ (n|n − 1) + R ỹ (n|n)H (n)Rv (n)H (n)
H 1

Postmultiplying by R ỹ (n|n − 1)

R ỹ (n|n − 1) = R ỹ (n|n) + R ỹ (n|n)HH (n)Rv−1 (n)H1 (n)R ỹ (n|n − 1)

Comparing with (7.8.39)

K(n) = R ỹ (n|n)HH (n)Rv−1 (n)

Statistical and Adaptive Signal Processing - Solution Manual 195

7.44 In Example 7.8.3 we assumed that only the position measurements were available for estimation. In this
problem we will assume that we also have a noisy sensor to measure velocity measurements. Hence the
observation model is
   
xp (n) yp (n) + v1 (n)
x(n) , = (1)
xv (n) yv (n) + v2 (n)

where v1 (n) and v2 (n) are two independent zero-mean white Gaussian noise sources with variances σ 2v1 and
σ 2v2 , respectively.

(a) Using the state vector model given in Example 7.8.3 and the observation model in (1), develop Kalman
filter equations to estimate position and velocity of the object at each n.
The vector state equations are:
      2 
yp (n) 1 T yp (n − 1) T /2
= + η(n)
yv (n) 0 1 yv (n − 1) T
y(n) , A y(n − 1) + B η(n)

The output vector equation is:

     
xp (n) yp (n) v1 (n)
= + =
xv (n) yv (n) v2 (n)
x(n) , H y(n) + D v(n)

(b) Using the parameter values

T = 0.1 σ 2v1 = σ 2v2 = σ 2η = 0.25 yp (−1) = 0 yv (−1) = 1

simulate the true and observed positions and velocities of the object. Using your Kalman filter equations,
generate plots similar to the ones given in Figures 7.14 and 7.15.
% Target and sensor models
T = 0.1;
A = [1 T; 0 1]; % PHI
B = [T*T/2; T]; % G
vareta =0.25;
H = eye(2);
D = eye(2);
varv1 = 0.25;
varv2 = 0.25;
Rv = diag([varv1,varv2]);

% Generate State and Observation signals

Nt = 100;
eta = sqrt(vareta)*randn(1,1,Nt); y = zeros(2,1,Nt);
y(1:2,1,1) = A*[0;1] + B*eta(1,1,1);
for n = 2:Nt
y(1:2,1,n) = A*y(1:2,1,n-1) + B*eta(1,1,n);
end
196 Statistical and Adaptive Signal Processing - Solution Manual

v1 = sqrt(varv1)*randn(1,Nt);
v2 = sqrt(varv2)*randn(1,Nt);
v = zeros(2,1,Nt); v(1,1,:) = v1(:); v(2,1,:) = v2(:);
for n = 1:Nt
x(1:2,1,n) = H*y(1:2,1,n) + D*v(1:2,1,n);
end

% A-priori estimates
yhat_ini = [0;0];
R_y = 2*eye(2,2);

% Kalman Filter
[y_hat,GainK,Re_pre,Re_post] = skf(A,B,vareta,H,D,Rv,x,yhat_ini,R_y);

% Extract quantities for plotting

time = [0:Nt]*T;
yp = y(1,1,:); yp = [0;yp(:)]; % True position
yv = y(2,1,:); yv = [1;yv(:)]; % True velocity
xp = x(1,1,:); xp = [0;xp(:)]; % Noisy position
xv = x(2,1,:); xv = [1;xv(:)]; % Noisy velocity
yp_hat = y_hat(1,1,:); yp_hat = [0;yp_hat(:)]; % Estimated position
yv_hat = y_hat(2,1,:); yv_hat = [1;yv_hat(:)]; % Estimated velocity
Kg1 = GainK(1,1,:); Kg1 = Kg1(:); % Kalman Gain (1,1)
Kg2 = GainK(2,2,:); Kg2 = Kg2(:); % Kalman Gain (2,2)
for n = 1:Nt
Rpre(n) = trace(Re_pre(:,:,n));
Rpost(n) = trace(Re_post(:,:,n));
end

% Plotting
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);

subplot(2,1,1);
plot(time,yp,’g’,time,xp,’r:’,time,yp_hat,’m--’);
axis([time(1),time(end),-5,15]);
ylabel(’Position (meters)’,’fontsize’,8);
title(’True, Noisy, and Estimated Positions’,’fontsize’,10);
legend(’True’,’Noisy’,’Estimate’,4);
set(gca,’xtick’,[0:2:10],’fontsize’,6)
set(gca,’ytickmode’,’auto’,’fontsize’,6);

subplot(2,1,2);
plot(time,yv,’g’,time,xv,’r:’,time,yv_hat,’m--’);
axis([0,10,-3,3]);
xlabel(’t (sec)’,’fontsize’,8); ylabel(’velocity (m/sec)’,’fontsize’,8);
legend(’True’,’Noisy’,’Estimate’,4);
title(’True, Noisy, and Estimated Velocities’,’fontsize’,10);
Statistical and Adaptive Signal Processing - Solution Manual 197

set(gca,’xtick’,[0:2:10],’fontsize’,6);
set(gca,’ytickmode’,’auto’,’fontsize’,6);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);
subplot(2,1,1);
H_kg = plot(time(2:end),Kg1,’go’,time(2:end),Kg2,’rd’);
set(H_kg,’markersize’,3); axis([0,10,0,1]);
title(’Kalman gain components’,’fontsize’,10);
legend(’K_p’,’K_v’,1);xlabel(’t (sec)’,’fontsize’,8);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:0.2:1],’fontsize’,6);
set(gca,’yticklabel’,[’0.0’;’0.2’;’0.4’;’0.6’;’0.8’;’1.0’]);

subplot(2,2,3);
H_pt1 = plot(time(2:41),Rpre(1:40),’ro’,time(2:41),Rpost(1:40),’gd’);
set(H_pt1,’markersize’,3);
legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
axis([0,4.0,0,1]);xlabel(’t (sec)’,’fontsize’,8);
%subplot(2,2,3), plot(time,tPm,’.’,time,tP,’g.’)
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:0.5:1],’fontsize’,6);

subplot(2,2,4);
H_pt2 = plot(time(61:72),Rpre(60:71),’ro’,time(61:72),Rpost(60:71),’gd’);
set(H_pt2,’markersize’,3); hold on;
timenew = [1;1]*time(61:72); timenew = (timenew(:))’;
tpmtp = zeros(1,23);
tpmtp(1:2:23) = Rpre(60:71); tpmtp(2:2:24) = Rpost(60:71);
plot(timenew,tpmtp,’w:’); hold off;
legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
axis([5.95,7.05,0.04,0.05]);
xlabel(’t (sec)’,’fontsize’,8);
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtick’,[6,7],’fontsize’,6);
set(gca,’ytick’,[0.04:0.005:0.05],’fontsize’,6);
The plots are shown in Figures 7.44b1 and 7.44b2.
(c) Discuss the effects of velocity measurements on the estimates.
Clearly, the velocity estimates follow the true velocity values. The error is due to noisy observations.

7.45 In this problem, we will assume that the acceleration ya (n) is an AR(1) process rather than a white noise
process. Let ya (n) be given by

ya (n) = α ya (n − 1) + η(n) η(n) ∼ WGN(0, σ 2η ) ya (−1) = 0 (1)

(a) Augment the state vector y(n) in (7.8.48), using variable ya (n), and develop the state vector as well as
the observation model, assuming that only the position is measured.
198 Statistical and Adaptive Signal Processing - Solution Manual

True, Noisy, and Estimated Positions

15

10

Position (meters)

0
True
Noisy
Estimate
−5
0 2 4 6 8 10

True, Noisy, and Estimated Velocities

3

2
velocity (m/sec)

−1

True
−2 Noisy
Estimate
−3
0 2 4 6 8 10
t (sec)

Figure 7.44b1: Plots of position, velocity and acceleration

Kalman gain components

1.0
Kp
Kv
0.8

0.6

0.4

0.2

0.0
0 1 2 3 4 5 6 7 8 9 10
t (sec)

Trace of covariance matrix Trace of covariance matrix

1 0.05
a−priori a−priori
a−posteriori a−posteriori

0.5 0.045

0 0.04
0 1 2 3 4 6 7
t (sec) t (sec)

Figure 7.44b2: Plots of gains and covariance traces

Statistical and Adaptive Signal Processing - Solution Manual 199

The scalar state equations are:

1
yp (n) = yp (n − 1) + yv (n − 1)T + ya (n − 1)T 2
2
yv (n) = yv (n − 1) + ya (n − 1)T
ya (n) = α ya (n − 1) + η(n)

Hence the vector state equations are:

      
yp (n) 1 T T 2 /2 yp (n − 1) 0
 yv (n)  =  0 1 T   yv (n − 1)  +  0  η(n)
ya (n) 0 0 α ya (n − 1) 1
y(n) , A y(n − 1) + B η(n)

The output equation is:

 
x(n) = yp (n) + v(n) = 1 0 0 y(n) + v(n)
, H y(n) + D v(n)

(b) Using the above model and the parameter values

T = 0.1 α = 0.9 σ 2v = σ 2η = 0.25

yp (−1) = 0 yv (−1) = 1

simulate the linear motion of the object. Using Kalman filter equations, estimate the position, velocity,
and acceleration values of the object at each n. Generate performance plots similar to the ones given in
Figures 7.14 and 7.15.
%% (b) AR(1) model for acceleration and observation of position
% Target and sensor models
T = 0.1; alpha = 0.9;

A = [1,T,T*T/2; 0,1,T;0,0,alpha]; K = length(A);

B = [0;0;1];
vareta =0.25; stdeta = sqrt(vareta); L = 1;
H = [1,0,0]; [M,K] = size(H);
D = eye(1);
varv = 0.25; stdv = sqrt(varv);
%varv2 = 0.25;
Rv = varv*eye(1); P = length(Rv);

% Generate State and Observation signals

Nt = 100; y = zeros(K,1,Nt); x = zeros(M,1,Nt);
eta = stdeta*randn(1,1,Nt); y = zeros(2,1,Nt);
y(1:K,1,1) = A*[0;1;0] + B*eta(1:L,1,1);
for n = 2:Nt
y(1:K,1,n) = A*y(1:K,1,n-1) + B*eta(1:L,1,n);
end

v1 = stdv*randn(1,Nt);
200 Statistical and Adaptive Signal Processing - Solution Manual

v = zeros(1,1,Nt); v(1,1,:) = v1(:);

for n = 1:Nt
x(1:M,1,n) = H*y(1:K,1,n) + D*v(1:P,1,n);
end

% A-priori estimates
yhat_ini = zeros(K,1); %[0;0;0];
R_y = 2*eye(K);

% Kalman Filter
[y_hat,GainK,Re_pre,Re_post] = skf(A,B,vareta,H,D,Rv,x,yhat_ini,R_y);

% Extract quantities for plotting

time = [0:Nt]*T;
yp = y(1,1,:); yp = [0;yp(:)]; % True position
yv = y(2,1,:); yv = [1;yv(:)]; % True velocity
ya = y(3,1,:); ya = [0;ya(:)]; % True acceleration
xp = x(1,1,:); xp = [0;xp(:)]; % Noisy position
%xv = x(2,1,:); xv = [1;xv(:)]; % Noisy velocity
yp_hat = y_hat(1,1,:); yp_hat = [0;yp_hat(:)]; % Estimated position
yv_hat = y_hat(2,1,:); yv_hat = [1;yv_hat(:)]; % Estimated velocity
ya_hat = y_hat(3,1,:); ya_hat = [1;ya_hat(:)]; % Estimated acceleration

Kg1 = GainK(1,1,:); Kg1 = Kg1(:); % Kalman Gain 1

Kg2 = GainK(2,1,:); Kg2 = Kg2(:); % Kalman Gain 2
Kg3 = GainK(3,1,:); Kg3 = Kg3(:); % Kalman Gain 3

for n = 1:Nt
Rpre(n) = trace(Re_pre(:,:,n));
Rpost(n) = trace(Re_post(:,:,n));
end

% Plotting
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);
xaxismin = time(1); xaxismax = time(end);

subplot(3,1,1);
plot(time,yp,’g’,time,xp,’r:’,time,yp_hat,’m--’);
yaxismin = floor(min(min([yp,xp,yp_hat]))/5)*5;
yaxismax = ceil(max(max([yp,xp,yp_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
ylabel(’Position (meters)’,’fontsize’,8);
title(’True, Noisy, and Estimated Positions’,’fontsize’,10);
legend(’True’,’Noisy’,’Estimate’,4);
set(gca,’xtick’,[0:2:10],’fontsize’,6)
set(gca,’ytickmode’,’auto’,’fontsize’,6);

subplot(3,1,2);
Statistical and Adaptive Signal Processing - Solution Manual 201

plot(time,yv,’g’,time,yv_hat,’m--’);
yaxismin = floor(min(min([yv,yv_hat]))/5)*5;
yaxismax = ceil(max(max([yv,yv_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
ylabel(’velocity (m/sec)’,’fontsize’,8);
legend(’True’,’Estimate’,4);
title(’True and Estimated Velocities’,’fontsize’,10);
set(gca,’xtick’,[0:2:10],’fontsize’,6);
set(gca,’ytickmode’,’auto’,’fontsize’,6);

subplot(3,1,3);
plot(time,ya,’g’,time,ya_hat,’m--’);
yaxismin = floor(min(min([ya,ya_hat]))/5)*5;
yaxismax = ceil(max(max([ya,ya_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
xlabel(’t (sec)’,’fontsize’,8); ylabel(’velocity (m/sec)’,’fontsize’,8);
legend(’True’,’Estimate’,4);
title(’True and Estimated Accelerations’,’fontsize’,10);
set(gca,’xtick’,[0:2:10],’fontsize’,6);
set(gca,’ytickmode’,’auto’,’fontsize’,6);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);
subplot(2,1,1);
H_kg = plot(time(2:end),Kg1,’mo’,time(2:end),Kg2,’rd’,time(2:end),Kg3,’g*’);
set(H_kg,’markersize’,3); axis([0,10,0,1]);
title(’Kalman gain components’,’fontsize’,10);
legend(’K_p’,’K_v’,’K_a’,1);xlabel(’t (sec)’,’fontsize’,8);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:0.2:1],’fontsize’,6);
set(gca,’yticklabel’,[’0.0’;’0.2’;’0.4’;’0.6’;’0.8’;’1.0’]);

subplot(2,2,3);
H_pt1 = plot(time(2:41),Rpre(1:40),’mo’,time(2:41),Rpost(1:40),’gd’);
set(H_pt1,’markersize’,3);
legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
%axis([0,4.0,1,6]);
xlabel(’t (sec)’,’fontsize’,8);
%subplot(2,2,3), plot(time,tPm,’.’,time,tP,’g.’)
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:1:6],’fontsize’,6);

subplot(2,2,4);
H_pt2 = plot(time(61:72),Rpre(60:71),’mo’,time(61:72),Rpost(60:71),’gd’);
set(H_pt2,’markersize’,3); hold on;
timenew = [1;1]*time(61:72); timenew = (timenew(:))’;
tpmtp = zeros(1,23);
tpmtp(1:2:23) = Rpre(60:71); tpmtp(2:2:24) = Rpost(60:71);
202 Statistical and Adaptive Signal Processing - Solution Manual

True, Noisy, and Estimated Positions

15

Position (meters)
10

True
0 Noisy
Estimate
−5
0 2 4 6 8 10
True and Estimated Velocities
5
velocity (m/sec)

True
Estimate
−5
0 2 4 6 8 10
True and Estimated Accelerations
5
acceleration (m/sec )
2

True
Estimate
−5
0 2 4 6 8 10
t (sec)

Figure 7.45b1: Plots of position, velocity and acceleration

plot(timenew,tpmtp,’w:’); hold off;

legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
axis([5.95,7.05,1.55,1.9]);
xlabel(’t (sec)’,’fontsize’,8);
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtick’,[6,7],’fontsize’,6);
set(gca,’ytick’,[1.5:0.1:1.9],’fontsize’,6);
The plots are shown in Figures 7.45b1 and 7.45b2.
(c) Now assume that noisy measurements of yv (n) and ya (n) are also available, that is, the observation model
is
   
xp (n) yp (n) + v1 (n)
   
x(n) , xv (n) =  yv (n) + v2 (n) (2)
xa (n) ya (n) + v3 (n)

where v1 (n), v2 (n), and v3 (n) are IID zero-mean white Gaussian noise sources with variance σ 2v . Repeat
parts (a) and (b) above.
%% (c) AR(1) model for acceleration and
% observations of position, velocity, and acceleration
% Target and sensor models
T = 0.1; alpha = 0.9;

A = [1,T,T*T/2; 0,1,T;0,0,alpha]; K = length(A);

B = [0;0;1];
vareta =0.25; stdeta = sqrt(vareta); L = length(vareta);
H = eye(3); [M,K] = size(H);
D = eye(3);
Statistical and Adaptive Signal Processing - Solution Manual 203

Kalman gain components

1.0
K
p
K
0.8 v
Ka

0.6

0.4

0.2

0.0
0 1 2 3 4 5 6 7 8 9 10
t (sec)

Trace of covariance matrix Trace of covariance matrix

6 1.9
a−priori a−priori
a−posteriori a−posteriori
5
1.8

1.7
3

1.6
2

0 1 2 3 4 6 7
t (sec) t (sec)

Figure 7.45b2: Plots of gains and covariance traces

varv = 0.25; stdv = sqrt(varv);

%varv2 = 0.25;
Rv = varv*eye(3); P = length(Rv);

% Generate State and Observation signals

Nt = 100; y = zeros(K,1,Nt); x = zeros(M,1,Nt);
eta = stdeta*randn(1,1,Nt); y = zeros(2,1,Nt);
y(1:K,1,1) = A*[0;1;0] + B*eta(1:L,1,n);
for n = 2:Nt
y(1:K,1,n) = A*y(1:K,1,n-1) + B*eta(1:L,1,n);
end

v1 = stdv*randn(1,Nt);
v2 = stdv*randn(1,Nt);
v3 = stdv*randn(1,Nt);
v = zeros(1,1,Nt);
v(1,1,:) = v1(:); v(2,1,:) = v2(:); v(3,1,:) = v3(:);
for n = 1:Nt
x(1:M,1,n) = H*y(1:K,1,n) + D*v(1:P,1,n);
end

% A-priori estimates
yhat_ini = zeros(K,1); %[0;0;0];
R_y = 2*eye(K);

% Kalman Filter
[y_hat,GainK,Re_pre,Re_post] = skf(A,B,vareta,H,D,Rv,x,yhat_ini,R_y);
204 Statistical and Adaptive Signal Processing - Solution Manual

% Extract quantities for plotting

time = [0:Nt]*T;
yp = y(1,1,:); yp = [0;yp(:)]; % True position
yv = y(2,1,:); yv = [1;yv(:)]; % True velocity
ya = y(3,1,:); ya = [0;ya(:)]; % True acceleration
xp = x(1,1,:); xp = [0;xp(:)]; % Noisy position
xv = x(2,1,:); xv = [1;xv(:)]; % Noisy velocity
xa = x(3,1,:); xa = [0;xa(:)]; % Noisy Acceleration
yp_hat = y_hat(1,1,:); yp_hat = [0;yp_hat(:)]; % Estimated position
yv_hat = y_hat(2,1,:); yv_hat = [1;yv_hat(:)]; % Estimated velocity
ya_hat = y_hat(3,1,:); ya_hat = [1;ya_hat(:)]; % Estimated acceleration

Kg1 = GainK(1,1,:); Kg1 = Kg1(:); % Kalman Gain 1

Kg2 = GainK(2,2,:); Kg2 = Kg2(:); % Kalman Gain 2
Kg3 = GainK(3,3,:); Kg3 = Kg3(:); % Kalman Gain 3

for n = 1:Nt
Rpre(n) = trace(Re_pre(:,:,n));
Rpost(n) = trace(Re_post(:,:,n));
end

% Plotting
Hf_3 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);
xaxismin = time(1); xaxismax = time(end);

subplot(3,1,1);
plot(time,yp,’g’,time,xp,’r:’,time,yp_hat,’m--’);
yaxismin = floor(min(min([yp,xp,yp_hat]))/5)*5;
yaxismax = ceil(max(max([yp,xp,yp_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
ylabel(’Position (meters)’,’fontsize’,8);
title(’True, Noisy, and Estimated Positions’,’fontsize’,10);
legend(’True’,’Noisy’,’Estimate’,4);
set(gca,’xtick’,[0:2:10],’fontsize’,6)
set(gca,’ytickmode’,’auto’,’fontsize’,6);

subplot(3,1,2);
plot(time,yv,’g’,time,xv,’r:’,time,yv_hat,’m--’);
yaxismin = floor(min(min([yv,yv_hat]))/5)*5;
yaxismax = ceil(max(max([yv,yv_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
xlabel(’t (sec)’,’fontsize’,8); ylabel(’velocity (m/sec)’,’fontsize’,8);
legend(’True’,’Noisy’,’Estimate’,4);
title(’True, Noisy, and Estimated Velocities’,’fontsize’,10);
set(gca,’xtick’,[0:2:10],’fontsize’,6);
set(gca,’ytickmode’,’auto’,’fontsize’,6);
Statistical and Adaptive Signal Processing - Solution Manual 205

subplot(3,1,3);
plot(time,ya,’g’,time,xa,’r:’,time,ya_hat,’m--’);
yaxismin = floor(min(min([ya,ya_hat]))/5)*5;
yaxismax = ceil(max(max([ya,ya_hat]))/5)*5;
axis([xaxismin,xaxismax,yaxismin,yaxismax]);
xlabel(’t (sec)’,’fontsize’,8); ylabel(’velocity (m/sec)’,’fontsize’,8);
legend(’True’,’Noisy’,’Estimate’,4);
title(’True, Noisy, and Estimated Accelerations’,’fontsize’,10);
set(gca,’xtick’,[0:2:10],’fontsize’,6);
set(gca,’ytickmode’,’auto’,’fontsize’,6);

Hf_4 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,papos);
subplot(2,1,1);
H_kg = plot(time(2:end),Kg1,’mo’,time(2:end),Kg2,’rd’,time(2:end),Kg3,’g*’);
set(H_kg,’markersize’,3); axis([0,10,0,1]);
title(’Kalman gain components’,’fontsize’,10);
legend(’K_p’,’K_v’,’K_a’,1);xlabel(’t (sec)’,’fontsize’,8);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:0.2:1],’fontsize’,6);
set(gca,’yticklabel’,[’0.0’;’0.2’;’0.4’;’0.6’;’0.8’;’1.0’]);

subplot(2,2,3);
H_pt1 = plot(time(2:41),Rpre(1:40),’mo’,time(2:41),Rpost(1:40),’gd’);
set(H_pt1,’markersize’,3);
legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
axis([0,4.0,0,7]);
xlabel(’t (sec)’,’fontsize’,8);
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtickmode’,’auto’,’fontsize’,6);
set(gca,’ytick’,[0:1:7],’fontsize’,6);

subplot(2,2,4);
H_pt2 = plot(time(61:72),Rpre(60:71),’mo’,time(61:72),Rpost(60:71),’gd’);
set(H_pt2,’markersize’,3); hold on;
timenew = [1;1]*time(61:72); timenew = (timenew(:))’;
tpmtp = zeros(1,23);
tpmtp(1:2:23) = Rpre(60:71); tpmtp(2:2:24) = Rpost(60:71);
plot(timenew,tpmtp,’w:’); hold off;
legend(’a-priori’,’a-posteriori’,1);xlabel(’t (sec)’,’fontsize’,8);
axis([5.95,7.05,0.15,0.6]);
xlabel(’t (sec)’,’fontsize’,8);
title(’Trace of covariance matrix’,’fontsize’,10);
set(gca,’xtick’,[6,7],’fontsize’,6);
set(gca,’ytick’,[0.2:0.1:0.6],’fontsize’,6);
The plots are shown in Figures 7.45c1 and 7.45c2.
206 Statistical and Adaptive Signal Processing - Solution Manual

True, Noisy, and Estimated Positions

10

Position (meters)
5

0 True
Noisy
Estimate
−5
0 2 4 6 8 10
True, Noisy, and Estimated Velocities
5
velocity (m/sec)

True
Noisy
Estimate
−5
0 2 4 6 8 10
True, Noisy, and Estimated
t (sec) Accelerations
5
acceleration (m/sec2)

True
Noisy
Estimate
−5
0 2 4 6 8 10
t (sec)

Figure 7.45c1: Plots of position, velocity and acceleration

Kalman gain components

1.0
Kp
K
0.8 v
Ka

0.6

0.4

0.2

0.0
0 1 2 3 4 5 6 7 8 9 10
t (sec)

Trace of covariance matrix Trace of covariance matrix

7 0.6
a−priori a−priori
6 a−posteriori a−posteriori
0.5
5

4 0.4

3
0.3
2

1
0.2

0
0 1 2 3 4 6 7
t (sec) t (sec)

Figure 7.45c2: Plots of gains and covariance traces

 Chapter 8

Least Squares Filtering and Prediction

8.1 By differentiating (8.2.8) with respect to the vector c, show that the LSE estimator cls is given by the solution
of the normal equations (8.2.12).
From (8.2.8)

E = E y − cH d̂ − d̂H c + cH R̂c

Hence (see Appendix B)

∂ ∂  
0= E= E y − cH d̂ − d̂H c + cH R̂c = −d̂H + cH R̂
∂c ∂c
gives the desired result.

8.2 Let the weighted LSE be given by E w = eH We, where W is a Hermitian positive definite matrix.

(a) By minimizing E w with respect to the vector c, show that the weighted LSE estimator is given by (8.2.35).
Using e , y − Xc, we have
 
E w = eH We = yH − cH XH W (y − Xc) = yH − cH XH (Wy − WXc)
= yH Wy − yH WXc − cH XH Wy + cH XH WXc
, E W y − d̂HW c − cH d̂ W + cH R̂ W c

where E W y , yH Wy, d̂W , XH Wy, and R̂W , XH WX. Since W is Hermitian and positive-definite, R̂W
is also Hermitian and positive-definite. Thus the weighted LSE estimator cwls is provided by the solution
of

R̂W cwls = d̂W

or
 H −1 H
cwls = R̂−1
W d̂ W = X WX X Wy

(b) Using the LDL H decomposition W = LDLH , show that the weighted LS criterion corresponds to pre-
filtering the error or the data.
Using W = LDLH , E w = eH LDLH e. Consider LH e which due to the upper triangular structure of LH is
a prefiltering operation on the error. Similarly from
 −1 H
cwls = XH LDLH X X LDLH y

we note that LH X or LH y are prefiltering operations.

8.3 Using direct substitution of (8.4.4) into (8.4.5), show that the LS estimator cls(i) and the associated LS error E ls(i)
are determined by (8.4.5).

207
208 Statistical and Adaptive Signal Processing - Solution Manual

 
Using X̄ = X1 y X2 from (8.4.4), we have in (8.4.5)
   
0  X1H  
 E (i)  = X̄H X̄ cls(i) =  y H  X1 y X2 cls(i)
ls
0 X2H
 H   
X1 X1 X1H y X1H X2 R11 r1 R12
=  y H X1 y H y y H X2  ,  r1H E y r2H 
X2H X1 X2H y X2H X2 H
R12 r2 R22
 T
Express cls(i) as cls(i) = cls1 1 cls2 so that
    
R11 r1 R12 cls1 0
 r1H E y r2H   1  =  E ls(i) 
H
R12 r2 R22 cls2 0

Following the parallels of (6.5.4) and (6.5.6) with the above equation (or using orthogonality) we obtain
    
R11 R12 cls1 r1
H =−
R12 R22 cls2 r2

and

E ls(i) = r1H cls1 + E y + r2H cls2

which are equations for the optimal LS estimator and design.

8.4 Consider a linear system described by the difference equation y(n) = 0.9y(n − 1) + 0.1x (n −1)+v(n), where
x(n) is the input signal, y(n) is the output signal, and v(n) is an output disturbance. Suppose that we have
collected N = 1000 samples of input-output data and that we wish to estimate the system coefficients, using
the LS criterion with no windowing. Determine the coefficients of the model y(n) = ay(n − 1) + dx(n − 1)
and their estimated covariance matrix σ̂ 2e R̂−1 when

(a) x(n) ∼ WGN(0, 1) and v(n) ∼ WGN(0, 1)

% part a
sigmaX=1; sigmaV=1;

B=[0 0.1]; A=[1 0.9];

x = sqrt(sigmaX)*randn(1000,1);
y = filter(B,A,x) + sqrt(sigmaV)*randn;

[R_hat,d_hat] = lsmatvec(’nowi’,x,2,y)
R_hat =
890.9536 11.7168
11.7168 891.0335
d_hat =
-31.4221
-28.2493
cls = inv(R_hat)*d_hat
c_ls =
Statistical and Adaptive Signal Processing - Solution Manual 209

-0.0349
-0.0312

Ey = y’*y;
Els = Ey-d_hat’*cls;

N = 1000; M = 2;
sigmaE = Els/(N-M)
sigmaE =
0.5557
cov = sigmaE*inv(R_hat)
cov =
1.0e-03 *
0.6238 -0.0082
-0.0082 0.6237
(b) x(n) ∼ WGN(0, 1) and v(n) = 0.8v(n − 1) + w(n) is an AR(1) process with w(n) ∼ WGN(0, 1).
Comment upon the quality of the obtained estimates by comparing the matrices σ̂ 2e R̂−1 obtained in each
case.
% part b
sigmaW = 1;
w = sqrt(sigmaW)*randn(1000,1);
v = filter([1],[1 0.8],w);
y = filter(B,A,x) + v;

[R_hat,d_hat] = lsmatvec(’nowi’,x,2,y)
R_hat =
890.9536 11.7168
11.7168 891.0335
d_hat =
-39.7950
-40.2742
cls = inv(R_hat)*d_hat
cls =
-0.0441
-0.0446
Ey = y’*y;
Els = Ey-d_hat’*cls;

N = 1000; M=2;
sigmaE = Els/(N-M)
sigmaE =
2.6554
cov = sigmaE*inv(R_hat)
cov =
0.0030 -0.0000
-0.0000 0.0030

8.5 Use Lagrange multipliers to show that Equation (8.4.13) provides the minimum of (8.4.8) under the constraint
210 Statistical and Adaptive Signal Processing - Solution Manual

(8.4.9).

8.6 If full windowing is used in LS, then the autocorrelation matrix is Toeplitz. Using this fact, show that in the
combined FBLP predictor is given by

afb = 21 a + jb∗

8.7 Consider the noncausal “middle” sample linear signal estimator specified by (8.4.1) with M = 2L and i = L.
Then

2L
e(L) (n) = ck(L)∗ x(n − k) = c(L)H x̄(n), c(L)
L = 1
k=0

(a) Show that if we apply full windowing to the data matrix, the resulting signal estimator is conjugate
symmetric, that is, c(L) = Jc(L)∗ . This property does not hold for any other windowing method.
We have the vector equation
(L)
e(L) = X̄c

in which X̄ is a data matric with full windowing. Hence R̄ , X̄H X̄ is a Hermitian Toeplitz matrix, that is,
∗ ∗
R̄J = JR̄ or R̄ = JR̄ J

Thus in (8.4.5) we have

 
0
(L) ∗  
R̄cls = JR̄ Jc(L)
ls =  E ls(L) 
0

Multiplying both sides by J and taking complex-conjugate, we obtain

 
0
 (L)  (L)
R̄∗ Jc(L)
ls =  E ls  = R̄cls
0

since JJ = I and E ls(L) is real-valued. Thus cls(L) = Jc(L)

ls and this is possible only with a full windowing.
(b) Derive the normal equations for the signal estimator that minimizes the total squared error E (L) = )e(L) )2
under the constraint c(L) = Jc(L)∗ .
In this case we assume that R̄ = X̄H X̄ is not Toeplitz but only Hermitian. Thus we want to minimize
(L)
E (L) = e(L)H e(L) , e(L) = X̄c

subject to the constraint c(L) = Jc(L) . We will prove the real-valued case. Consider partitiong of
(L)
e(L) = X̄c as
 (L) 
c
(L)
  1 
e = X̄1 y X̄2 1 
c2(L)
Statistical and Adaptive Signal Processing - Solution Manual 211

where y = x̄(n − L) and c2(L) = Jc(L)

1 (using the constraint). Thus
 (L)
e(L) = X̄1 c1(L) + y + X̄2 c2(L) = X̄1 c1(L) + y + X̄2 Jc(L)
1 = X̄1 + X̄2 J c1 + y

, Wc(L)
1 +y

Thus normal equations are given by

WT e(L) = WT Wc(L)
1 +W y =0
T

or
 T   T  
T  
X̄1 + X̄2 J X̄1 + X̄2 J c1(L) + X̄1 + X̄2 J y = 0 = X̄1T + JX̄2 X̄1 + X̄2 J c1(L) + X̄1T + X̄2T J y

or
  
X̄1T X̄1 + X̄1T X̄2 J + JX̄2 X̄1 + JX̄2 X̄2 J c1(L) + X̄1 + X̄2 J y = 0
T T
(1)

Using another form of normal equations

 (L)   T 
c1 X̄1 X̄1 c1(L) + X̄1T X̄2 Jc(L)
1 + X̄1 y
T
   (L) 
X̄T X̄ 1  =  E ls 
(L) (L) T (L)
Jc1 X̄2T X̄1 c1 + X̄2T y + X̄2 X̄2 Jc1

and since the first and the last elements (after multiplied by J) of the right-hand side vector add to zero
due to (1), we obtain
 
0
(L) T (L)  
X̄T X̄c + JX̄ X̄c = 2E (L) 
0
or
 
$ % 0
T  
X̄T X̄ + JX̄ X̄ c(L) = 2E (L) 
0

which are the normal equations. The complex-valued proof is similar.

(c) Show that if we enforce the normal equation matrix to be centro-Hermitian, that is, we use the normal
equations
 
$ % 0
T  
X̄H X̄ + JX̄ X̄∗ J c(L) = 2E (L) 
0

then the resulting signal smoother is conjugate symmetric.

In this case W , (X̄ H X̄ + JX̄ X̄∗ J) is Hermitian Toeplitz, that is,
T

 H T
  T H
∗
JWJ = JX̄ X̄J + X̄ X̄∗ = JX̄ X̄∗ J + X̄ X̄ = W∗

Then from part (a) above, we have cls(L) = Jcls(L) .

212 Statistical and Adaptive Signal Processing - Solution Manual

(d) Illustrate parts (a) to (c), using the data matrix

 
3 2 1
 4 2 
 3 
 . 
 . 
 . 4 3 
 
 .. 
1001 . 4 
X=  
 ... .. 
 1001 . 
 .. 
 3 . 1001 
 
 .. 
 2 3 . 
1 2 3
and check which smoother provides the smallest total squared error. Try to justify the obtained answer.
L = 2; M = 2*L;
N = 1000; x = [1:N,N+1,N:-1:1]’; N = length(x); %x = x + 0.5*(rand(N,1)-0.5);

% Symmetric Case (Full Windowing)

disp(’Symmetric Case (Full Windowing)’);
Xbar = toeplitz([conj(x);zeros(M,1)],[conj(x(1)),zeros(1,M)]);
y = Xbar(:,L+1); X = [Xbar(:,1:L),Xbar(:,L+2:M+1)];

R = X’*X;
r = X’*y;
c = -R\r;
c = [c(1:L);1;c(L+1:end)]
Els = Xbar’*Xbar*c, Els = real(Els(L+1));
e = Xbar*c; Els1 = e’*e;

% Nonsymmetric case (No windowing)

disp(’Nonsymmetric case (No windowing)’);
Xbar = toeplitz(x(M+1:end),flipud(x(1:M+1)));
y = Xbar(:,L+1); X = [Xbar(:,1:L),Xbar(:,L+2:M+1)];

R = X’*X;
r = X’*y;
c = -R\r;
c = [c(1:L);1;c(L+1:end)]
Els = Xbar’*Xbar*c, Els = real(Els(L+1));
e = Xbar*c; Els2 = e’*e;

% No windowing but forced coefficient conjugate symmetry

disp(’No windowing but forced coefficient conjugate symmetry’);
y = Xbar(:,L+1); X = [Xbar(:,1:L)+(fliplr(Xbar(:,L+2:M+1)))];

R = X’*X;
r = X’*y;
c = -R\r;
c = [c(1:L);1;(c(L:-1:1))] %c = [c(1:L);1;c(L+1:end)]
Statistical and Adaptive Signal Processing - Solution Manual 213

Els = (Xbar’*Xbar + flipud(Xbar’*Xbar))*c/2, Els = real(Els(L+1));

e = Xbar*c; Els3 = e’*e;
%% Complex-valued approach
y = Xbar(:,L+1);
X = [Xbar(:,1:L)+(fliplr(Xbar(:,L+2:M+1))), ...
j*(Xbar(:,1:L)-(fliplr(Xbar(:,L+2:M+1))))];
R = X’*X;
r = X’*y;
c = -R\r; c = c(1:L) + j*c(L+1:end);
c = [c(1:L);1;conj(c(L:-1:1))] %c = [c(1:L);1;c(L+1:end)]
e = Xbar*c; Els3 = e’*e;

% No windowing but forced centro-Hermitian normal equation matrix

disp(’No windowing but forced centro-Hermitian normal equation matrix’);
Xbar1 = fliplr(conj(Xbar));
R = Xbar’*Xbar + Xbar1’*Xbar1;
R1 = [R(1:L,1:L),R(1:L,L+2:end);R(L+2:end,1:L),R(L+2:end,L+2:end)];
r1 = [R(1:L,L+1);R(L+2:end,L+1)];
c =-R1\r1;
c = [c(1:L);1;c(L+1:end)]
Els = R*c/2, Els = real(Els(L+1));
e = Xbar*c; Els4 = e’*e;

[Els,I] = sort([Els1,Els2,Els3,Els4])
Els =
0.6667 0.6667 0.6667 1.0000
I =
3 2 4 1
Thus the full-windowing gives the worst error.

8.8 A useful impulse response for some geophysical signal processing applications is the Mexican hat wavelet
2
g(t) = √ π −1/4 (1 − t 2 )e−t /2
2

3
which is the second derivative of a Gaussian pulse.

(a) Plot the wavelet g(t) and the magnitude and phase of its Fourier transform.
% Mexican-Hat function:
tmax = 5; t = -tmax:0.05:tmax; t2 = t.*t;
gt = (2/sqrt(3))*pi^(-1/4)*(1-t2).*exp(-t2/2);

% Fourier Transform of g(t): G(F)

F = -1:0.01:1; w = 2*pi*F; w2 = w.*w;
Gf = sqrt(8/3)*(pi^(1/4))*w2.*exp(-w2/2);
I1 = max(find(Gf(1:51) <= 0.001)); F1 = F(I1);
I2 = min(find(Gf(151:201) <= 0.001)); F2 = F(I2+150);
214 Statistical and Adaptive Signal Processing - Solution Manual

Mexican−Hat Wavelet Fourier Transform

1

1.5

0.5

G(F )
g(t )

0
0.5

−0.5 0
−5 0 5 −1 −0.74 0 0.74 1
t F
Figure 8.8a: Plots of g(t) and G( f )

% (a) Plot of g(t) and its Fourier Transform G(F)

Hf_1 = figure(’units’,’inches’,’position’,[1,1,6,3],...
’paperunits’,’inches’,’paperposition’,[0,0,6,3],...
’numbertitle’,’off’,’name’,’P0808a’);

subplot(’position’,[0.08,0.15,0.4,0.75]);
plot(t,gt,’g’,[-tmax,tmax],[0,0],’w:’,[0,0],[-.5,1],’w:’);
axis([-tmax,tmax,-0.5,1]);
title(’Mexican-Hat Wavelet’,’fontsize’,12);
xlabel(’\itt’,’fontsize’,10); ylabel(’{\itg}({\itt} )’,’fontsize’,10);

subplot(’position’,[0.58,0.15,0.4,0.75]);
plot(F,Gf,’g’,[-1,1],[0,0],’w:’,[0,0],[0,1.8],’w:’); axis([-1,1,0,1.8]);
title(’Fourier Transform’,’fontsize’,12);
xlabel(’\itF’,’fontsize’,10); ylabel(’{\itG}({\itF} )’,’fontsize’,10);
set(gca,’xtick’,[-1,F1,0,F2,1],’ytick’,[0:0.5:2]);
The plot is shown in Figure 8.8a.
(b) By examining the spectrum of the wavelet, determine a reasonable sampling frequency Fs .
%% (b) A reasonable sampling frequency (for G(F) <= 0.001)
Fs = 2*F2, Ts = 1/Fs;
Fs =
1.4800
(c) Design an optimum LS inverse FIR filter for the discrete-time wavelet g(nT ), where T = 1/Fs . Determine
a reasonable value for M by plotting the LSE E M as a function of order M. Investigate whether we can
improve the inverse filter by introducing some delay n 0 . Determine the best value of n 0 and plot the
impulse response of the resulting filter and the combined impulse response g(n) ∗ h(n − n 0 ), which
should resemble an impulse.
%%(c) Optimum LS Inverse filter
M = 14; % order of the filter
Statistical and Adaptive Signal Processing - Solution Manual 215

% Generate sample of g(t) as data

t = 0:Ts:5; t = [-fliplr(t(2:end)),t]; t2 = t.*t;
g = (2/sqrt(3))*pi^(-1/4)*(1-t2).*exp(-t2/2); g = g’;
n = 0:length(g)-1; tmax = n(end)*Ts;
G = toeplitz([g;zeros(M,1)],[g(1),zeros(1,M)]);
[L,N] = size(G); n0 = floor(L/2);
y = [zeros(n0,1);1;zeros(n0,1)];
R = G’*G; r = G’*y;
c_ls = R\r;
e = y - G*c_ls; format long; Els = e’*e, format short;
Els =
0.09340142175215
y_ls = conv(g,c_ls); m = 0:length(y_ls)-1; mmax = m(end)*Ts;

Hf_2 = figure(’units’,’inches’,’position’,[1,1,6,6],...
’paperunits’,’inches’,’paperposition’,[0,0,6,6],...
’numbertitle’,’off’,’name’,’P0808c’);

subplot(’position’,[0.08,0.6,0.4,0.3]);
stem(n*Ts,g,’g’); hold on;
plot([0,tmax],[0,0],’w:’,[0,0],[-.5,1],’w:’); axis([0,tmax,-0.5,1]);
title(’sampled Wavelet’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10); ylabel(’{\itg}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*Ts,’xticklabel’,...
[’ 0’;’ 1’;’ 2’;’ 3’;’ 4’;’ 5’;’ 6’;’ 7’;’ 8’;’ 9’;’10’;’11’;...
’12’;’13’;’14’]);

subplot(’position’,[0.58,0.6,0.4,0.3]);
stem(n*Ts,c_ls,’g’); hold on;
plot([0,tmax],[0,0],’w:’); axis([0,tmax,-20,20]);
title(’LS Inverse Filter’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10); ylabel(’{\itc}_{ls}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*Ts,’xticklabel’,...
[’ 0’;’ 1’;’ 2’;’ 3’;’ 4’;’ 5’;’ 6’;’ 7’;’ 8’;’ 9’;’10’;’11’;’12’;...
’13’;’14’]);

subplot(’position’,[0.08,0.1,0.9,0.3]);
stem(m*Ts,y_ls,’g’); hold on;
plot([0,mmax],[0,0],’w:’); axis([0,mmax,-1,1]);
title(’Verification’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10);
ylabel(’{\itc}_{ls}({\itn})*{\itg}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*M*Ts,’xticklabel’,...
[’ 0’;’14’;’29’]);
The plot is shown in Figure 8.8c.
(d) Repeat part (c) by increasing the sampling frequency by a factor of 2 and comparing with the results
obtained in part (c).
%%(d) Optimum LS Inverse filter (Twice the sampling rate)
216 Statistical and Adaptive Signal Processing - Solution Manual

sampled Wavelet LS Inverse Filter

1 20

10
0.5

cls(n )
g(n )

0
0
−10

−0.5 −20
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
n n

Verification
1

0.5
cls(n)*g(n )

−0.5

−1
0 14 29
n
Figure 8.8c: Plots of sampled wavelet and its inverse filter

sampled Wavelet LS Inverse Filter

1 1000

0.5 500
cls(n )
g(n )

0 0

−0.5 −500
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28
n n

Verification
1

0.5
cls(n)*g(n )

−0.5

−1
0 14 29
n
Figure 8.8d: Plots of sampled wavelet and its inverse filter for twice the sampling rate
Statistical and Adaptive Signal Processing - Solution Manual 217

Fs = 2*Fs; Ts = 1/Fs;

M = 28; % order of the filter

% Generate sample of g(t) as data
t = 0:Ts:5; t = [-fliplr(t(2:end)),t]; t2 = t.*t;
g = (2/sqrt(3))*pi^(-1/4)*(1-t2).*exp(-t2/2); g = g’;
n = 0:length(g)-1; tmax = n(end)*Ts;
G = toeplitz([g;zeros(M,1)],[g(1),zeros(1,M)]);
[L,N] = size(G); n0 = floor(L/2);
y = [zeros(n0,1);1;zeros(n0,1)];
R = G’*G; r = G’*y;
c_ls = R\r; k = 0:length(c_ls)-1; kmax = k(end)*Ts;
e = y - G*c_ls; format long; Els = e’*e, format short;
Els =
0.45036127872114
y_ls = conv(g,c_ls); m = 0:length(y_ls)-1; mmax = m(end)*Ts;

Hf_3 = figure(’units’,’inches’,’position’,[1,1,6,6],...
’paperunits’,’inches’,’paperposition’,[0,0,6,6],...
’numbertitle’,’off’,’name’,’P0808d’);

subplot(’position’,[0.08,0.6,0.4,0.3]);
stem(n*Ts,g,’g’); hold on;
plot([0,tmax],[0,0],’w:’,[0,0],[-.5,1],’w:’); axis([0,tmax,-0.5,1]);
title(’sampled Wavelet’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10); ylabel(’{\itg}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*2*Ts,’xticklabel’,...
[’ 0’;’ 2’;’ 4’;’ 6’;’ 8’;’10’;’12’;’14’;’16’;’18’;’20’;’22’;...
’24’;’26’;’28’]);

subplot(’position’,[0.58,0.6,0.4,0.3]);
stem(k*Ts,c_ls,’g’); hold on;
plot([0,tmax],[0,0],’w:’); %axis([0,tmax,-20,20]);
title(’LS Inverse Filter’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10); ylabel(’{\itc}_{ls}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*2*Ts,’xticklabel’,...
[’ 0’;’ 2’;’ 4’;’ 6’;’ 8’;’10’;’12’;’14’;’16’;’18’;’20’;’22’;...
’24’;’26’;’28’]);

subplot(’position’,[0.08,0.1,0.9,0.3]);
stem(m*Ts,y_ls,’g’); hold on;
plot([0,mmax],[0,0],’w:’); axis([0,mmax,-1,1]);
title(’Verification’,’fontsize’,12);
xlabel(’\itn’,’fontsize’,10);
ylabel(’{\itc}_{ls}({\itn})*{\itg}({\itn} )’,’fontsize’,10);
set(gca,’xtick’,n*M*Ts,’xticklabel’,...
[’ 0’;’14’;’29’]);
The plot is shown in Figure 8.8d.
218 Statistical and Adaptive Signal Processing - Solution Manual

8.9 (a) Prove Equation (8.5.4) regarding the LDL H decomposition of the augmented matrix R̄.
Using direct substitution
   H   
L 0 D 0 L k LDLH LDk
=
kH 1 0 E ls 0H 1 kH DLH kH Dk + E ls
 
R̂ d̂
= = R̄
d̂H E y

(b) Solve the LS estimation problem in Example 8.5.1, using the LDLH decomposition of R̄ and the parti-
tionings in (8.5.4).
% part b
X=[1 1 1; 2 2 1; 3 1 3; 1 0 1];
y=[1 2 4 3]’;
R_hat=X’*X;
d_hat=X’*y;
Ey=y’*y;
R_bar=[R_hat d_hat; d_hat’ Ey];
[L_bar D_bar] = ldlt(R_bar);
k=L_bar(end,1:end-1)’;
L=L_bar(1:end-1,1:end-1);
Cls=L’\k
Cls =
3.0000
-1.5000
-1.0000

8.10 Prove the order-recursive algorithm described by the relations given in (8.5.12). Demonstrate the validity of
this approach, using the data in Example 8.5.1.

8.11 In this problem, we wish to show that the statistical interpretations of innovation and partial correlation for
wm (n) and km+1 in (8.5.12) hold in a deterministic LSE sense. To this end, suppose that the “partial correlation”
between ỹ and x̃m+1 is defined using the residual records ẽm = ỹ − Xm cm and ẽm = x̃m+1 + Xm bm , where
bm is the LSE BLP. Show that kk+1 = β m+1 /ξ m+1 , where β m+1 , ẽmH ẽm and ξ m+1 = ẽmH ẽm . Demonstrate the
validity of these formulas using the data in Example 8.5.1.

8.12 Show that the Cholesky decomposition of a Hermitian positive definite matrix R can be computed by using
the following algorithm

for j = 1 to M
! j −1
li j = (ri j − k=1 |l j k |2 )1/2

for i = j + 1 to M
! j −1
li j = (ri j − k=1 lik∗ l j k )/l j j
end i
end j

and write a Matlab function for its implementation. Test your code using the built-in Matlab function chol.
Statistical and Adaptive Signal Processing - Solution Manual 219

8.13 Compute the LDLT and Cholesky decompositions of the following matrices:
   
9 3 −6 6 4 −2
   
X1 =  3 4 1  and X2 =  4 5 3 
−6 1 9 −2 3 6

% Part 1. Matrix X1
X1 = [9 3 -6; 3 4 1; -6 1 9];

% R1 = X1’*X1;
chX1 = chol(X1)
chX1 =
3.0000 1.0000 -2.0000
0 1.7321 1.7321
0 0 1.4142
[L1 D1] = ldlt(X1)
L1 =
1.0000 0 0
0.3333 1.0000 0
-0.6667 1.0000 1.0000
D1 =
9
3
2
A1 = L1*diag(sqrt(D1))
A1 =
3.0000 0 0
1.0000 1.7321 0
-2.0000 1.7321 1.4142

% Part 2. Matrix X2
X2 = [6 4 2; 4 5 3; 2 3 6];
%R2=X2’*X2;
[L2,D2] = ldlt(X2)
L2 =
1.0000 0 0
0.6667 1.0000 0
0.3333 0.7143 1.0000
D2 =
6.0000
2.3333
4.1429
chX2 = chol(X2)
chX2 =
2.4495 1.6330 0.8165
0 1.5275 1.0911
0 0 2.0354
A2 = L2*diag(sqrt(D2))
A2 =
220 Statistical and Adaptive Signal Processing - Solution Manual

2.4495 0 0
1.6330 1.5275 0
0.8165 1.0911 2.0354

8.14 Solve the LS problem in Example 8.6.1,

(a) using the QR decomposition of the augmented data matrix X̄ = [X y]

x=[1 1 1; 2 2 1; 3 1 3; 1 0 1];
y=[1 2 4 3]’;

x_hat=[x y];
[N M]=size(x_hat);

[q,r]=qr(x_hat)
q =
-0.2582 -0.3545 0.8006 0.4082
-0.5164 -0.7089 -0.4804 0.0000
-0.7746 0.4557 0.1601 -0.4082
-0.2582 0.4051 -0.3203 0.8165
r =
-3.8730 -2.0656 -3.3566 -5.1640
0 -1.3166 0.7089 1.2659
0 0 0.4804 -0.4804
0 0 0 1.2247
Rqr=r(1:3,1:3);

Zqr=q’*y
Zqr =
-5.1640
1.2659
-0.4804
1.2247
z1qr=Zqr(1:3);
z2qr=Zqr(4);
Cls_qr=Rqr\z1qr
Cls_qr =
3.0000
-1.5000
-1.0000
Els_qr=z2qr’*z2qr
Els_qr =
1.5000
eqr=q*[zeros(1,3) z2qr’]’
eqr =
0.5000
0.0000
-0.5000
1.0000
Statistical and Adaptive Signal Processing - Solution Manual 221

(b) using the Cholesky decomposition of the matrix R̄ = X̄ H X̄.

[R,p]=chol(x_hat’*x_hat);

Rchol=R(1:3,1:3);

Cls_ch=Rchol’\R(1:3,4)
Cls_ch =
1.3333
-3.0534
-14.8224
Els_ch=R(4,4)
Els_ch =
1.2247

We have H = I − 2wwH where )w) = wH w = 1.

8.15 (a) Show that a unit vector w is an eigenvector of the matrix H = I − 2ww H . What is the corresponding
eigenvalue?
Consider

I − 2wwH w = w − 2wwH w = w − 2w = −w

Hence w is an eigenvector woth λ = −1.

(b) If a vector z is orthogonal to w, show that z is an eigenvector of H. What is the corresponding eigenvalue?
Consider

z= z
I − 2wwH z = z − 2ww H

Hence z is an eigenvector woth λ = 1.

8.16 Solve the LS problem

   
1 2 −3
1   10 
 3   
X=  y= 
1 2   3 
1 −1 6

using the Householder transformation.

X=[1 2; 1 3; 1 2; 1 -1]; Y=[-3 10 3 6]’;

x=X(:,1);
s=norm(x);
sn=sqrt(2*s*(s+abs(x(1))));
w1=(x+[1;0;0;0]*sign(x(1))*s)/sn;
H1=eye(length(w1))-2*w1*w1’;
X1=H1*X
X1 =
-2.0000 -3.0000
222 Statistical and Adaptive Signal Processing - Solution Manual

-0.0000 1.3333
-0.0000 0.3333
-0.0000 -2.6667
x=X1(2:end,2);
s=norm(x);
sn=sqrt(2*s*(s+abs(x(1))));
w22=(x+[1;0;0]*sign(x(1))*s)/sn;
w2=[0;w22];
H22=eye(length(w22))-2*w22*w22’;
H2=eye(4,4);H2(2:end,2:end)=H22;
X2=H2*X1
X2 =
-2.0000 -3.0000
-0.0000 -3.0000
-0.0000 0
-0.0000 0
R=X2
R =
-2.0000 -3.0000
-0.0000 -3.0000
-0.0000 0
-0.0000 0
Q=H1*H2
Q =
-0.5000 -0.1667 -0.4744 -0.7051
-0.5000 -0.5000 -0.2692 0.6538
-0.5000 -0.1667 0.8333 -0.1667
-0.5000 0.8333 -0.0897 0.2179
Q*R
ans =
1.0000 2.0000
1.0000 3.0000
1.0000 2.0000
1.0000 -1.0000
w1
w1 =
0.8660
0.2887
0.2887
0.2887
w2
w2 =
0
0.8498
0.0654
-0.5230

% Solve the LS problem

z=Q’*Y;
Statistical and Adaptive Signal Processing - Solution Manual 223

cls=R(1:2,1:2)’\z(1:2)
cls =
4.0000
-4.0000
Els=z(3:4)’*z(3:4)
Els =
90.0000

8.17 Solve Problem 8.16 by using the Givens transformation.

% Given’s Rotation
X = [1,2;1,3;1,2;1,-1], A = X;
X =
1 2
1 3
1 2
1 -1
y = [-3;10;3;6];

% Step-1: Zero out X(4,1) => rotate rows 3 and 4 of Col 1

G41 = eye(4);
[c,s] = rotate(X(3,1),X(4,1));
G41(:,3:4) = G41(:,3:4)*[c,s;-s,c];
X = G41*X,
X =
1.0000 2.0000
1.0000 3.0000
1.4142 0.7071
0 -2.1213

% Step-2: Zero out X(3,1) => rotate rows 2 and 3 of Col 1

G31 = eye(4);
[c,s] = rotate(X(2,1),X(3,1));
G31(:,2:3) = G31(:,2:3)*[c,s;-s,c];
X = G31*X,
X =
1.0000 2.0000
1.7321 2.3094
-0.0000 -2.0412
0 -2.1213

% Step-3: Zero out X(2,1) => rotate rows 1 and 2 of Col 1

G21 = eye(4);
[c,s] = rotate(X(1,1),X(2,1));
G21(:,1:2) = G21(:,1:2)*[c,s;-s,c];
X = G21*X,
X =
2.0000 3.0000
-0.0000 -0.5774
224 Statistical and Adaptive Signal Processing - Solution Manual

-0.0000 -2.0412
0 -2.1213

% Step-4: Zero out X(4,2) => rotate rows 3 and 4 of Col 2

G42 = eye(4);
[c,s] = rotate(X(3,2),X(4,2));
G42(:,3:4) = G42(:,3:4)*[c,s;-s,c];
X = G42*X,
X =
2.0000 3.0000
-0.0000 -0.5774
-0.0000 -2.9439
0.0000 0

% Step-5: Zero out X(3,2) => rotate rows 2 and 3 of Col 2

G32 = eye(4);
[c,s] = rotate(X(2,2),X(3,2));
G32(:,2:3) = G32(:,2:3)*[c,s;-s,c];
X = G32*X,
X =
2.0000 3.0000
-0.0000 -3.0000
0.0000 0
0.0000 0

% The G, Q and R matrices

G = G32*G42*G21*G31*G41
G =
0.5000 0.5000 0.5000 0.5000
-0.1667 -0.5000 -0.1667 0.8333
0.8498 -0.3922 -0.3269 -0.1307
0 0.5883 -0.7845 0.1961
Q = G’
Q =
0.5000 -0.1667 0.8498 0
0.5000 -0.5000 -0.3922 0.5883
0.5000 -0.1667 -0.3269 -0.7845
0.5000 0.8333 -0.1307 0.1961
R = triu(X)
R =
2.0000 3.0000
0 -3.0000
0 0
0 0

% LS solution
R1 = R(1:2,:)
R1 =
2.0000 3.0000
Statistical and Adaptive Signal Processing - Solution Manual 225

0 -3.0000
r = G*y
r =
8.0000
0
-8.2369
4.7068
r1 = r(1:2)
r1 =
8.0000
0
c = R1\r1
c =
4.0000
0
LSE = norm(r(3:4))^2
LSE =
90

% check using Matlab’s LS tools

c = A\y
c =
4.0000
0.0000

8.18 Compute the QR decomposition of the data matrix

 
4 2 1
2 
 0 1 
X= 
2 0 −1 
1 2 1

using the GS and MGS methods, and compare the obtained results.

% QR decomposition using Gram-Schmidt (GS) and Modified GS (MGS) methods

X = [4,2,1;2,0,1;2,0,-1;1,2,1]
X =
4 2 1
2 0 1
2 0 -1
1 2 1
[N,M] = size(X);

%% GS method
R = zeros(M); Q = zeros(N,M);

% vector q1
R(1,1) = norm(X(:,1));
Q(:,1) = X(:,1)/R(1,1);
226 Statistical and Adaptive Signal Processing - Solution Manual

% vector q2
R(1,2) = Q(:,1)’*X(:,2);
p2 = X(:,2) - R(1,2)*Q(:,1); R(2,2) = norm(p2);
Q(:,2) = p2/R(2,2);

% vector q3
R(1,3) = Q(:,1)’*X(:,3); R(2,3) = Q(:,2)’*X(:,3);
p3 = X(:,3) - R(1,3)*Q(:,1) - R(2,3)*Q(:,2); R(3,3) = norm(p3);
Q(:,3) = p3/R(3,3);

% QR decomposition
Q, R
Q =
0.8000 0.2000 0
0.4000 -0.4000 0.7071
0.4000 -0.4000 -0.7071
0.2000 0.8000 0
R =
5.0000 2.0000 1.0000
0 2.0000 1.0000
0 0 1.4142

% Check
er_GS = norm(X-Q*R)
er_GS =
2.2204e-016

%% MGS method
R = zeros(M); Q = zeros(N,M); X1 = X;

% vector q1
R(1,1) = norm(X(:,1));
Q(:,1) = X(:,1)/R(1,1);
R(1,2) = Q(:,1)’*X(:,2); X(:,2) = X(:,2) - R(1,2)*Q(:,1);
R(1,3) = Q(:,1)’*X(:,3); X(:,3) = X(:,3) - R(1,3)*Q(:,1);

% vector q2
R(2,2) = norm(X(:,2));
Q(:,2) = X(:,2)/R(2,2);
R(2,3) = Q(:,2)’*X(:,3); X(:,3) = X(:,3) - R(2,3)*Q(:,2);

% vector q3
R(3,3) = norm(X(:,3));
Q(:,3) = X(:,3)/R(3,3);

% QR decomposition
Q, R
Q =
Statistical and Adaptive Signal Processing - Solution Manual 227

0.8000 0.2000 -0.0000

0.4000 -0.4000 0.7071
0.4000 -0.4000 -0.7071
0.2000 0.8000 -0.0000
R =
5.0000 2.0000 1.0000
0 2.0000 1.0000
0 0 1.4142

% Check
er_MGS = norm(X1-Q*R)
er_MGS =
1.5701e-016

8.19 Solve the following LS problem

   
1 −2 −1 −1
2   
 0 1   1 
X= , y =  
2 −4 2   1 
4 0 0 −2

by computing the QR decomposition using the GS algorithm.

Consider
 
1 −2 −1
2 1 
 0   
X=  , x1 x2 x3
2 −4 2 
4 0 0

Then we have the following steps:

• Compute p1 , r11 , and q1 :

    > >  
1 1 > 1 >
> > 0.2
2  2  >2 > 
    > >  0.4 

p1 = x1 =   , r11 = )p1 ) = 5, q1 = p1 / )p1 ) =   / > > = 
2  2  >2 > 0.4 
> >
4 4 > 4 > 0.8

Compute p2 , r12 , r22 , and q2 :

        
−2 −2 0.2 − 85
     
 T  0    0   0.4   4
 =  516


p2 = x2 − q1 x2 q1 =   −  .2 .4 .4 .8   
 −4    −4  0.4   − 5 
8
0 0 0.8 5

 
−0.4
 0.2 
r12 = q1T x2 = −2, r22 = )p2 ) = 4, and q2 = p2 / )p2 ) =  
 −0.8 
0.4
228 Statistical and Adaptive Signal Processing - Solution Manual

• Compute p3 , r13 , r23 , r33 , and q3 :

 
−1.6
   0.8 
p3 = x3 − q1T x3 q1 − q2T x3 q2 = 
 0.8 


−0.4

 
−0.8
 0.4 
r13 = q1T x3 = 1, r23 = q2T x3 = −1, r33 = )p3 ) = 2, and q3 = p3 / )p3 ) = 
 0.4 


−0.2

Thus the thin QR decomposition is

 
0.2 −0.4 −0.8  
 0.4  5 −2 1
0.2 0.4
Q1 =  0.4 −0.8
, R =  0
 4 −1 
0.4
0 0 2
0.8 0.4 −0.2

Thus
 
  −1  
0.2 0.4 0.4 0.8   −1.0
 1  
Xcls = Q1 Rcls = y ⇒ Rcls = QT1 y =  −0.4 0.2 −0.8 0.4    = −1.0 
 1 
−0.8 0.4 0.4 −0.2 2.0
−2
or
   
5 −2 1 −1.0
 0 4 −1  cls =  −1.0 
0 0 2 2.0

Finally using backsubstitution, we obtain

 
−0.4
cls =  0 
1.0

8.20 Show that the computational organization of the MGS algorithm shown in Table 8.4 can be used to compute
the GS algorithm if we replace the step rmi = qmH xi by rmi = qmH qi where qi ’s are initialized as qi = xi ,
1 ≤ i ≤ M.
This follows by observing that the rmi calculations are unchanged while xi ’s play the role of pi ’s in the GS
algorithm.
 
1 1
8.21 Compute the SVD of X = 1 1 by computing the eigenvalues and eigenvectors of X H X and XX H . Check
0 0
with the results obtained using the svd function.
Consider first XH X
 
  1 1  
1 1 0  1 1 = 2 2
XH X =
1 1 0 2 2
0 0
Statistical and Adaptive Signal Processing - Solution Manual 229

Its eigenvectors are λ1 = 4, λ2 = 0, and its eigenmatrix is

   
1 1 −1
V= √ ,
2 1 1

Similarly, consider XXH

   
1 1   2 2 0
1 1 0
XXH =  1 1  = 2 2 0 
1 1 0
0 0 0 0 0

Its eigenvectors are λ1 = 4, λ2 = 0, λ3 = 0, and its eigenmatrix is

     
1 −1 0
1   
U= √ 1 , 1 , 0 
 
2 0 0 1

From Matlab, the function [U,S,V] = svd([1,1;1,1;0,0]) gives

[U,S,V] = svd([1,1;1,1;0,0]) U =
0.7071 -0.7071 0
0.7071 0.7071 0
0 0 1.0000
S =
2 0
0 0
0 0
V =
0.7071 -0.7071
0.7071 0.7071

which agrees with the above results.

8.22 Repeat Problem 8.21 for

 
6 2
(a) X =
−7 6
% part a
X=[6 2; -7 6];
XHXa=X’*X;
XXHa=X*X’;
[Ua1,Sa1,Va1] = svd(XHXa)
Ua1 =
0.8944 -0.4472
-0.4472 -0.8944
Sa1 =
100.0000 0
0 25.0000
Va1 =
0.8944 -0.4472
230 Statistical and Adaptive Signal Processing - Solution Manual

-0.4472 -0.8944
[Ua2,Sa2,Va2] = svd(XXHa)
Ua2 =
0.4472 -0.8944
-0.8944 -0.4472
Sa2 =
100.0000 0
0 25.0000
Va2 =
0.4472 -0.8944
-0.8944 -0.4472
 
0 1 1
(b) X = .
1 1 0
% part b
X = [0 1 1; 1 1 0];
XHXb=X’*X;
XXHb=X*X’;
[Ub1,Sb1,Vb1] = svd(XHXb)
Ub1 =
0.4082 -0.7071 0.5774
0.8165 0.0000 -0.5774
0.4082 0.7071 0.5774
Sb1 =
3.0000 0 0
0 1.0000 0
0 0 0
Vb1 =
0.4082 -0.7071 0.5774
0.8165 0.0000 -0.5774
0.4082 0.7071 0.5774
[Ub2,Sb2,Vb2] = svd(XXHb)
Ub2 =
0.7071 -0.7071
0.7071 0.7071
Sb2 =
3.0000 0
0 1.0000
Vb2 =
0.7071 -0.7071
0.7071 0.7071
 
8.23 Write a Matlab program to produce the plots in Figure 8.14, using the matrix X = 6
−7
2
6
. Hint: Use a
parametric description of the circle in polar coordinates.

close all;
set(0,’defaultaxesfontsize’,8);
Hf_1 = figure(’units’,’inches’,’position’,[0.05,0.3,6,6],...
Statistical and Adaptive Signal Processing - Solution Manual 231

% Transformation matrix X
X = [6,2;-7,6];
[U,S,V] = svd(X); V(:,2) = -V(:,2); U(:,2) = -U(:,2);
disp(sprintf(’\n The singular values are %2i and %2i’,S(1,1),S(2,2)));

The singular values are 10 and 5

Pv = [1,0;0,1]; % test point after transformation by V’

P = V*Pv; % test point before transformation by V’

% Unit circle: C = r*exp(j*theta) = exp(j*theta); (since r = 1)

theta = 0:1:360; % angle in degrees
x = cos(theta*pi/180); y = sin(theta*pi/180);
C = [x;y];

subplot(2,2,1);
plot(x,y,’g’,[-4,4],[0,0],’w:’,[0,0],[-4,4],’w:’,P(1,1),P(2,1),’yo’,...
P(1,2),P(2,2),’yd’);
axis([-4,4,-4,4]); axis(’equal’);
set(gca,’xtick’,[-3:2:3],’ytick’,[-3:2:3]);
title(’Unit circle with test points’);

% Transformation of Unit circle using X

Cx = X*C; Px = X*P;
subplot(2,2,2);
plot(Cx(1,:),Cx(2,:),’g’,[-11,11],[0,0],’w:’,[0,0],[-11,11],’w:’,...
Px(1,1),Px(2,1),’yo’,Px(1,2),Px(2,2),’yd’);
axis([-11,11,-11,11]); axis(’equal’);
title(’Transformation using X’);

%% Transformation using SVD components

% X = U*S*V’;

% Transformation using V’
Cv = V’*C;
subplot(2,2,3);
plot(Cv(1,:),Cv(2,:),’g’,[-4,4],[0,0],’w:’,[0,0],[-4,4],’w:’,...
Pv(1,1),Pv(2,1),’yo’,Pv(1,2),Pv(2,2),’yd’);
axis([-4,4,-4,4]); axis(’equal’);
set(gca,’xtick’,[-3:2:3],’ytick’,[-3:2:3]);
title(’Transformation using V^H’);

% Transformation using S
Cs = S*Cv; Ps = S*Pv;
subplot(2,2,4);
plot(Cs(1,:),Cs(2,:),’g’,[-11,11],[0,0],’w:’,[0,0],[-11,11],’w:’,...
Ps(1,1),Ps(2,1),’yo’,Ps(1,2),Ps(2,2),’yd’);
axis([-11,11,-11,11]); axis(’equal’);
232 Statistical and Adaptive Signal Processing - Solution Manual

Unit circle with test points Transformation using X

10
3

5
1

−1
−5

−3
−10

−3 −1 1 3 −10 −5 0 5 10

Transformation using VH Transformation using S

10
3

5
1

−1
−5

−3
−10

−3 −1 1 3 −10 −5 0 5 10

Figure 8.23: Geometric interpretation of SVD

title(’Transformation using S’);

% Transformation using U
Cu = U*Cs; Pu = U*Ps;
hold on;
plot(Cu(1,:),Cu(2,:),’r’,Pu(1,1),Pu(2,1),’yo’,Pu(1,2),Pu(2,2),’yd’);

The plots are shown in Figure 8.23.

 T
8.24 For the matrix X = 01 11 10 determine X+ and verify that X and X+ satisfy the four Moore-Penrose conditions
(8.7.22).

X=[0 1 1; 1 1 0]’;
[U,S,V]=svd(X);
Splus=zeros(size(S’));
Splus(1:2,1:2)=diag(1./diag(S));
Xplus=V*Splus*U’;
disp(’A is’)
Statistical and Adaptive Signal Processing - Solution Manual 233

A is
disp(Xplus)
-0.3333 0.3333 0.6667
0.6667 0.3333 -0.3333

% Verify four Moore-Penrose Conditions

MP1=X*Xplus*X;
disp(’XAX = X’)
XAX = X
disp(MP1)
-0.0000 1.0000
1.0000 1.0000
1.0000 0.0000

MP2=Xplus*X*Xplus;
disp(’AXA = A’)
AXA = A
disp(MP2)
-0.3333 0.3333 0.6667
0.6667 0.3333 -0.3333

MP3=(X*Xplus)’;
disp(’(XA)^{H} = XA’)
(XA)^{H} = XA
disp(MP3)
0.6667 0.3333 -0.3333
0.3333 0.6667 0.3333
-0.3333 0.3333 0.6667

MP4=(Xplus*X)’;
disp(’(AX)^{H} = AX’)
(AX)^{H} = AX
disp(MP4)
1.0000 -0.0000
0.0000 1.0000

8.25 Prove the four Moore-Penrose conditions in (8.7.22) and explain why XX+ and X+ X are orthogonal projections
onto the range space of X and X H .
Using X = U$VH , X+ = V$ + UH , and A = X+

• Consider
  
XAX = U$VH V$ + UH U$VH = U$VH V$ + UH U$VH
= U$$ + $VH = U$VH = X

since
  
$r 0 $r−1 0
$ + $ =$$ + = =I
0 0 0 0
234 Statistical and Adaptive Signal Processing - Solution Manual

• Consider
  
AXA = V$ + UH U$VH V$ + UH = V$ + UH U$VH V$ + UH
= V$ + $$ + UH = X+ = A

• Consider
 H  H
(XA)H = AH XH = V$ + UH U$VH = U$ +H VH V$ H UH
 
= U$ +H VH V$ H UH = XA

• Similarly
 H  H
(AX)H = XH AH = U$VH V$ + UH = V$ H UH U$ +H VH
 
= V$ H UH U$ +H VH = AX

Consider

XX+ = U$V H + H
 V$ U = U$$ +UH  
$r 0 $r−1 0 Ir 0
= U U =U
H
UH
0 0 0 0 0 0
= Ur UrH

which by definition are orthogonal projections onto the range space of X (Golub & Van Loan). Similarly X+ X
are orthogonal projections onto the range space of X+ .

8.26 In this problem we examine in greater detail the radio-frequency interference cancellation experiment discussed
in Section 8.4.3. We first explain the generation of the various signals and then proceed with the design and
evaluation of the LS interference canceler.

(a) The useful signal is a pointlike target defined by


d 1 dg(t)
s(t) = αt /t
,
dt e −αt /tr +e f dt
where α = 2.3, tr = 0.4, and t f = 2. Given that Fs = 2 GHz, determine s(n) by computing the
samples g(n) = g(nT ) in the interval −2 ≤ nT ≤ 6 ns and then computing the first difference s(n) =
g(n) − g(n − 1). Plot the signal s(n) and its Fourier transform (magnitude and phase), and check whether
the pointlike and wideband assumptions are justified.
(b) Generate N = 4096 samples of the narrowband interference using the formula


L
z(n) = Ai sin (ωi n + φ i )
i=1

and the following information:

Fs=2; % All frequencies are measured in GHz.

F=0.1*[0.6 1 1.8 2.1 3 4.8 5.2 5.7 6.1 6.4 6.7 7 7.8 9.3]’;
L=length(F);
om=2*pi*F/Fs;
A=[0.5 1 1 0.5 0.1 0.3 0.5 1 1 1 0.5 0.3 1.5 0.5]’;
rand(’seed’,1954);
phi=2*pi*rand(L,1);
Statistical and Adaptive Signal Processing - Solution Manual 235

(c) Compute and plot the periodogram of z(n) to check the correctness of your code.
(d) Generate N samples of white Gaussian noise v(n) ∼ WGN (0, 0.1) and create the observed signal
x(n) = 5s(n − n 0 ) + z(n) + v(n), where n 0 = 1000. Compute and plot the periodogram of x(n).
(e) Design a one-step ahead (D = 1) linear predictor with M = 100 coefficients using the FBLP method
with no windowing. Then use the obtained FBLP to clean the corrupted signal x(n) as shown in Figure
8.7. To evaluate the performance of the canceler, generate the plots shown in Figures 8.8 and 8.9.

8.27 Careful inspection of Figure 8.9 indicates that the D-step prediction error filter, that is, the system with input
x(n) and output ef (n), acts as a whitening filter. In this problem, we try to solve Problem 8.26 by designing a
practical whitening filter using a power spectral density (PSD) estimate of the corrupted signal x(n).

(a) Estimate the PSD R̂x(PA) (e j ωk ), ωk = 2π k/NFFT , of the signal x(n), using the method of averaged
periodograms. Use a segment length of L = 256 samples, 50 percent overlap, and NFFT = 512.
% Common parameters
Fs = 2; % All frequencies are in GHz
N = 4096; n = 0:N-1; % Length of sequence

%% Pointlike Target Signal s(n)

alpha = 2.3; tr = 0.4; tf = 2;
n = [-2*Fs:6*Fs]; t = n/(Fs*1e9);
g = 1./(exp(-alpha*t/tr)+exp(alpha*t/tf));
s = diff(g);

%% Given Narrowband interference parameters

% Frequencies of Interfering sinusoids
Fk = 0.1*[0.6,1,1.8,2.1,3,4.8,5.2,5.7,6.1,6.4,6.7,7,7.8,9.3]’;
K = length(Fk);
omk = 2*pi*Fk/Fs;
% Amplitudes of Interfering sinusoids
Ak = [0.5,1,1,0.5,0.1,0.3,0.5,1,1,1,0.5,0.3,1.5,0.5]’;
% Phases of Interfering sinusoids
rand(’seed’,1954); phik = 2*pi*rand(K,1);
% Narrowband Interference signal z(n)
z = sum((Ak*ones(1,N)).*sin(omk*n + phik*ones(1,N)));

%% Additive WGN
varv = 0.1; sigv = sqrt(varv);
v = sigv*randn(1,N);

%% Observed signal x(n)

x = 5*s+z+v;

% (a) PSD Rx using the method of averaged periodogram

Nfft = 512; L = 256; w_ham = hamming(L)’;
Rx = psd(x,Nfft,Fs,w_ham,L/2,’none’);
(b) Since the PSD does not provide any phase information, we shall design a whitening FIR filter with linear
236 Statistical and Adaptive Signal Processing - Solution Manual

phase by
1 2π NFFT −1
−j k
H̃ (k) = : e NFFT 2

R̂x(PA) (e j ωk )

where H̃ (k) is the DFT of the impulse response of the filter, that is,
FFT −1
N
− j N2π nk
H̃ (k) = h(n) e FFT

n=0

with 0 ≤ k ≤ NFFT − 1.
% (b) Whitening FIR filter
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
H = magH.*exp(-j*phaH);
h = real(ifft(H));
(c) Use the obtained whitening filter to clean the corrupted signal x(n), and compare its performance with
the FBLP canceler by generating plots similar to those shown in Figures 8.8 and 8.9.
% (c) Filtering using the Whitening filter
y = filter(h,1,x);
(d) Repeat part (c) with L = 128, NFFT = 512 and L = 512, NFFT = 1024 and check whether spectral
resolution has any effect upon the performance.
% (d1) L = 128, Nfft = 512
Nfft = 512; L = 128; w_ham = hamming(L)’;
Rx = psd(x,Nfft,Fs,w_ham,L/2,’none’);
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
H = magH.*exp(-j*phaH);
h = real(ifft(H));
y = filter(h,1,x);

% (d2) L = 512, Nfft = 1024

Nfft = 1024; L = 512; w_ham = hamming(L)’;
Rx = psd(x,Nfft,Fs,w_ham,L/2,’none’);
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
H = magH.*exp(-j*phaH);
h = real(ifft(H));
y = filter(h,1,x);

8.28 Repeat Problem 8.27, using the multitaper method of PSD estimation.

(a) Estimate the PSD R̂x(PA) (e j ωk ), ωk = 2π k/NFFT , of the signal x(n), using the multitaper method. Use
NFFT = 512.
Statistical and Adaptive Signal Processing - Solution Manual 237

% Common parameters
Fs = 2; % All frequencies are in GHz
N = 4096; n = 0:N-1; % Length of sequence

%% Pointlike Target Signal s(n)

alpha = 2.3; tr = 0.4; tf = 2;
n = [-2*Fs:6*Fs]; t = n/(Fs*1e9);
g = 1./(exp(-alpha*t/tr)+exp(alpha*t/tf));
s = diff(g);

%% Given Narrowband interference parameters

% Frequencies of Interfering sinusoids
Fk = 0.1*[0.6,1,1.8,2.1,3,4.8,5.2,5.7,6.1,6.4,6.7,7,7.8,9.3]’;
K = length(Fk);
omk = 2*pi*Fk/Fs;
% Amplitudes of Interfering sinusoids
Ak = [0.5,1,1,0.5,0.1,0.3,0.5,1,1,1,0.5,0.3,1.5,0.5]’;
% Phases of Interfering sinusoids
rand(’seed’,1954); phik = 2*pi*rand(K,1);
% Narrowband Interference signal z(n)
z = sum((Ak*ones(1,N)).*sin(omk*n + phik*ones(1,N)));

%% Additive WGN
varv = 0.1; sigv = sqrt(varv);
v = sigv*randn(1,N);

%% Observed signal x(n)

x = 5*s+z+v;

% (a) PSD Rx using the Multitaper method

Nfft = 512;
Rx = pmtm(x,4,Nfft,Fs);
(b) Since the PSD does not provide any phase information, we shall design a whitening FIR filter with linear
phase by
1 2π NFFT −1
−j k
H̃ (k) = : e NFFT 2

R̂x(PA) (e j ωk )

where H̃ (k) is the DFT of the impulse response of the filter, that is,
FFT −1
N
− j N2π nk
H̃ (k) = h(n) e FFT

n=0

with 0 ≤ k ≤ NFFT − 1.
% (b) Whitening FIR filter
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
238 Statistical and Adaptive Signal Processing - Solution Manual

H = magH.*exp(-j*phaH);
h = real(ifft(H));
(c) Use the obtained whitening filter to clean the corrupted signal x(n), and compare its performance with
the FBLP canceler by generating plots similar to those shown in Figures 8.8 and 8.9.
% (c) Filtering using the Whitening filter
y = filter(h,1,x);
(d) Repeat part (c) with NFFT = 512 and NFFT = 1024 and check whether spectral resolution has any effect
upon the performance.
% (d1) Nfft = 512
Nfft = 512;
Rx = pmtm(x,4,Nfft,Fs);
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
H = magH.*exp(-j*phaH);
h = real(ifft(H));
y = filter(h,1,x);

% (d2) Nfft = 1024

Nfft = 1024;
Rx = psd(x,4,Nfft,Fs);
magH = 1./(sqrt(Rx’)); magH = [magH,fliplr(magH(2:end-1))];
k = 0:Nfft/2; phaH = 2*pi*(Nfft-1)/(Nfft*2)*k;
phaH = [phaH,fliplr(phaH(2:end-1))];
H = magH.*exp(-j*phaH);
h = real(ifft(H));
y = filter(h,1,x);

8.29 In this problem we develop an RFI canceler using a symmetric linear smoother with guard samples defined by


M
e(n) = x(n) − x̂(n) , x(n) + ck [x(n − k) + x(n − k)]
k=D

where 1 ≤ D < M prevents the use of the D adjacent samples to the estimation of x(n).

(a) Following the approach used in Section 8.4.3, demonstrate whether such a canceler can be used to mitigate
RFI and under what conditions.
a=2.3; r=0.4;
f=2; Fs=2; T=1/Fs;
n=[-4:12];

for i=1:length(n)
g(i)=1/(exp(-a*n(i)*T/r)+exp(a*n(i)*T/f));
end

s(1)=g(1);
for i=2:length(n)
Statistical and Adaptive Signal Processing - Solution Manual 239

s(i)=g(i)-g(i-1);
end

S=fft(s,501);
S=fftshift(S);
(b) If there is theoretical justification for such a canceler, estimate its coefficients, using the method of LS
with no windowing for M = 50 and D = 1 for the situation described in Problem 8.26.
% part(b)
F=0.1*[0.6 1 1.8 2.1 3 4.8 5.2 5.7 6.1 6.4 6.7 7 7.8 9.3]’;
L=length(F);
om=2*pi*F/Fs;
A=[0.5 1 1 0.5 0.1 0.3 0.5 1 1 1 0.5 0.3 1.5 0.5]’;
rand(’seed’,1954);
phi=2*pi*rand(L,1);

N=4096;
for i=1:N
z(i)=A’*(sin(om.*i+phi));
end
z=z’;
(c) Use the obtained filter to clean the corrupted signal x(n), and compare its performance with the FBLP
canceler by generating plots similar to those shown in Figures 8.8 and 8.9.
% part(d)
sigmaV=0.1;
v=sqrt(sigmaV)*randn(size(z));
s_delay=zeros(size(z));
for i=1:length(s)
s_delay(i+1000)=s(i);
end

x=5*s_delay+z+v;

M=50;
D=1;
R=lsmatvec(’nowi’,x,2*M+2);

Rb=fliplr(R);

R_bar=R+Rb;

c_bar=R_bar\R_bar(:,50);
(d) Repeat part (c) for D = 2.

8.30 In Example 6.7.1 we studied the design and performance of an optimum FIR inverse system. In this problem,
we design and analyze the performance of a similar FIR LS inverse filter, using training input-output data.
240 Statistical and Adaptive Signal Processing - Solution Manual

(a) First, we generate N = 100 observations of the input signal y(n) and the noisy output signal x(n). We
assume that x(n) ∼ WGN(0, 1) and v(n) ∼ WGN(0, 0.1). To avoid transient effects, we generate 200
samples and retain the last 100 samples to generate the required data records.
% (a)
M=10; sigmaX=1;
sigmaV=0.1; sigmaY=1;

y=sqrt(sigmaY)*randn(250,100);
v=sqrt(sigmaV)*randn(250,100);

s=filter([-3/5 7/5 -2/5],[1],[zeros(1,100) ;y]);

x=s(2:end,:)+v;
(b) Design an LS inverse filter with M = 10 for 0 ≤ D < 10, using no windowing, and choose the best
value of delay D.
% (b) No Windowing
X=x(101:200,:);

for j=1:100
for i=1:10
[R dd]=lsmatvec(’nowi’,X(:,j),10,y(99+i:198+i,j));
d(:,i)=dd;
h(:,i,j)=R\d(:,i);
Py=y(99+i:198+i,j)’*y(99+i:198+i,j);
P(i,j)=Py-d(:,i)’*h(:,i,j);
end
end

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,4,2]);
plot([0:9],mean(P,2),’g’),
title(’MMSE (no windowing)’,’fontsize’,title_fontsize);
xlabel(’D’,’fontsize’,label_fontsize);
ylabel(’Error’,’fontsize’,label_fontsize);
exportfig(gcf,’p0830a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
subplot(411);
stem([0:9],mean(h(:,2,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=1’], ’units’,’normalized’);
title(’Filter Coefficients h_{D} (no windowing)’,’fontsize’,title_fontsize);
subplot(412)
stem([0:9],mean(h(:,4,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=3’], ’units’,’normalized’,’fontsize’,title_fontsize);
Statistical and Adaptive Signal Processing - Solution Manual 241

MMSE (no windowing)

26

25

24

23

22

Error
21

20

19

18

17
0 1 2 3 4 5 6 7 8 9
D

Figure 8.30a: Plot of error variance vs D for no windowing

subplot(413)
stem([0:9],mean(h(:,6,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=5’], ’units’,’normalized’,’fontsize’,title_fontsize);
subplot(414)
stem([0:9],mean(h(:,8,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=7’], ’units’,’normalized’,’fontsize’,title_fontsize);
xlabel(’n’,’fontsize’,label_fontsize);
The plots of error variance and filter responses are shown in Figures 8.30a and 8.30b.
(c) Repeat part (b) using full windowing.
% (c) full windowing
for j=1:100
for i=1:10
[R dd]=lsmatvec(’full’,X(:,j),10,y(99+i:198+i,j));
d(:,i)=dd;
h(:,i,j)=R\d(:,i);
Py=y(99+i:198+i,j)’*y(99+i:198+i,j);
P(i,j)=Py-d(:,i)’*h(:,i,j);
end
end

Hf_3 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,4,2]);
plot([0:9],mean(P,2),’g’);
title(’MMSE (full windowing)’,’fontsize’,title_fontsize);
xlabel(’D’,’fontsize’,label_fontsize);
ylabel(’Error’,’fontsize’,label_fontsize);
exportfig(gcf,’p0830c.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

Hf_4 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,4]);
subplot(411)
242 Statistical and Adaptive Signal Processing - Solution Manual

Filter Coefficients hD (no windowing)

1
D=1
0.5

0 1 2 3 4 5 6 7 8 9
1
D=3
0.5

0 1 2 3 4 5 6 7 8 9
1
D=5
0.5

0 1 2 3 4 5 6 7 8 9
1
D=7
0.5

0 1 2 3 4 5 6 7 8 9
n
Figure 8.30b: Plots of filter responses for various values of D for no windowing

stem([0:9],mean(h(:,2,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=1’], ’units’,’normalized’,’fontsize’,title_fontsize);
title(’Filter Coefficients h_{D} (full windowing)’,’fontsize’,title_fontsize);
subplot(412)
stem([0:9],mean(h(:,4,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=3’], ’units’,’normalized’,’fontsize’,title_fontsize);
subplot(413)
stem([0:9],mean(h(:,6,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=5’], ’units’,’normalized’,’fontsize’,title_fontsize);
subplot(414)
stem([0:9],mean(h(:,8,:),3),’g’);line([0:9],zeros(10,1));
axis([0 9 -.25 1]),
text(0.85, .8,[’D=7’], ’units’,’normalized’,’fontsize’,title_fontsize);
xlabel(’n’,’fontsize’,label_fontsize);
The plots of error variance and filter responses are shown in Figures 8.30c and 8.30d.
(d) Compare the LS filters obtained in parts (b) and (c) with the optimum filter designed in Example 6.7.1.
What are your conclusions?

8.31 In this problem we estimate the equalizer discussed in Example 6.8.1, using input-output training data, and we
evaluate its performance using Monte Carlo simulation.

(a) Generate N = 1000 samples of input-desired response data {x(n), a(n)}0N−1 and use them to estimate the
Statistical and Adaptive Signal Processing - Solution Manual 243

MMSE (full windowing)

35

30

25

Error 20

15

10
0 1 2 3 4 5 6 7 8 9
D

Figure 8.30a: Plot of error variance vs D for full windowing

Filter Coefficients hD (full windowing)

1
D=1
0.5

0 1 2 3 4 5 6 7 8 9
1
D=3
0.5

0 1 2 3 4 5 6 7 8 9
1
D=5
0.5

0 1 2 3 4 5 6 7 8 9
1
D=7
0.5

0 1 2 3 4 5 6 7 8 9
n
Figure 8.30b: Plots of filter responses for various values of D for full windowing
244 Statistical and Adaptive Signal Processing - Solution Manual

correlation matrix R̂x and the cross-correlation vector d̂ between x(n) and y(n − D). Use D = 7, M = 11,
and W = 2.9. Solve the normal equations to determine the LS FIR equalizer and the corresponding LSE.

% Given Parameters
N = 1000; % Length of sequences
M = 11; % equalizer length
D = 7; % delay in desired signal
varv = 0.001; % Variance of noise
K = 500; % Number of simulations

% Initializations
h = zeros(1,3);

%% W = 2.9: channel impulse response

W = 2.9;
h(1) = 0.5*(1+cos(2*pi*(1-2)/W));
h(2) = 0.5*(1+cos(2*pi*(2-2)/W));
h(3) = 0.5*(1+cos(2*pi*(3-2)/W));
h1 = [0 h(1) h(2) h(3) 0]’;

% (a) LS FIR equalizer and LSE

% Generation of Bernoulli sequence a(n)

a = rand(N,1); a = 2*round(a)-1;

% Generation of noise sequence v(n)

v = sqrt(varv)*randn(N,1);

% Generation of received sequence x(n)

x = conv2(a,h1,’same’) + v;

% Generation of desired response y(n)

y = [zeros((M-1)/2,1);a;zeros((M-1)/2,1)]; %y = y(1:N-M+1);

% Estimated autocorrelation matrix

Xbar = toeplitz([conj(x);zeros(M-1,1)],[conj(x(1)),zeros(1,M-1)]);
R_hat = Xbar’*Xbar;

% Estimated cross-correlation vector

d_hat = Xbar’*y;

% LS FIR equalizer
c = R_hat\d_hat;

% LSE
LSE = y’*y - c’*d_hat;

% Printout
disp(sprintf(’\n *** The LS FIR equalizer is:\n %7.4f %7.4f %7.4f
Statistical and Adaptive Signal Processing - Solution Manual 245

%7.4f %7.4f %7.4f %7.4f %7.4f %7.4f %7.4f %7.4f ’,c));

*** The LS FIR equalizer is:
-0.0012 0.0024 -0.0122 0.0604 -0.2563 1.1095 -0.2550
0.0578 -0.0114 -0.0006 0.0013

disp(sprintf(’ *** The LSE is: %7.4f’, LSE));

*** The LSE is: 1.5309

(b) Repeat part (a) 500 times; by changing the seed of the random number generators, compute the average
(over the realizations) coefficient vector and average LSE, and compare with the optimum MSE equalizer
obtained in Example 6.8.1. What are your conclusions?

% (b) Monte-Carlo Analysis

c = zeros(M,K); LSE = zeros(1,K);
for k = 1:K
rand(’state’,sum(100*clock)); randn(’state’,sum(100*clock));
a = rand(N,1); a = 2*round(a)-1;
v = sqrt(varv)*randn(N,1);
x = conv2(a,h1,’same’) + v;
y = [zeros((M-1)/2,1);a;zeros((M-1)/2,1)]; %y = y(1:N-M+1);
Xbar = toeplitz([conj(x);zeros(M-1,1)],[conj(x(1)),zeros(1,M-1)]);
R_hat = Xbar’*Xbar;
d_hat = Xbar’*y;
c(:,k) = R_hat\d_hat;
LSE(k) = y’*y - c(:,k)’*d_hat;
end
c1 = mean(c,2);
LSE1 = mean(LSE);

(c) Repeat parts (a) and (b) by setting W = 3.1.

%% W = 3.5: channel impulse response

W = 3.5;
h(1) = 0.5*(1+cos(2*pi*(1-2)/W));
h(2) = 0.5*(1+cos(2*pi*(2-2)/W));
h(3) = 0.5*(1+cos(2*pi*(3-2)/W));
h2 = [0 h(1) h(2) h(3) 0]’;

% (c) LS FIR equalizer and LSE

% Generation of Bernoulli sequence a(n)

a = rand(N,1); a = 2*round(a)-1;

% Generation of noise sequence v(n)

v = sqrt(varv)*randn(N,1);

% Generation of received sequence x(n)

x = conv2(a,h2,’same’) + v;

% Generation of desired response y(n)

246 Statistical and Adaptive Signal Processing - Solution Manual

Channel response for W = 2.9 Channel response for W = 3.1

2 2

1.5 1.5

1 1
h(n)

h(n)
0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
n n

Equalizer for W = 2.9 Equalizer for W = 3.1

2 2
LSE = 1.49 LSE = 5.2407
1.5 1.5

1 1
c(n)

c(n)

0.5 0.5

0 0

−0.5 −0.5

−1 −1
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
n n

Figure 8.31: Plots of filter responses for various values of D for full windowing
Statistical and Adaptive Signal Processing - Solution Manual 247

y = [zeros((M-1)/2,1);a;zeros((M-1)/2,1)]; %y = y(1:N-M+1);

% Estimated autocorrelation matrix

Xbar = toeplitz([conj(x);zeros(M-1,1)],[conj(x(1)),zeros(1,M-1)]);
R_hat = Xbar’*Xbar;

% Estimated cross-correlation vector

d_hat = Xbar’*y;

% LS FIR equalizer
c = R_hat\d_hat;

% LSE
LSE = y’*y - c’*d_hat;

% Printout
disp(sprintf(’\n *** The LS FIR equalizer is:\n %7.4f %7.4f %7.4f
%7.4f %7.4f %7.4f %7.4f %7.4f %7.4f %7.4f %7.4f ’,c));
*** The LS FIR equalizer is:
-0.0237 0.0729 -0.1662 0.3532 -0.7476 1.5752 -0.7450
0.3513 -0.1629 0.0700 -0.0236

disp(sprintf(’ *** The LSE is: %7.4f’, LSE));

*** The LSE is: 5.3158

% (d) Monte-Carlo Analysis

c = zeros(M,K); LSE = zeros(1,K);
for k = 1:K
rand(’state’,sum(100*clock)); randn(’state’,sum(100*clock));
a = rand(N,1); a = 2*round(a)-1;
v = sqrt(varv)*randn(N,1);
x = conv2(a,h2,’same’) + v;
y = [zeros((M-1)/2,1);a;zeros((M-1)/2,1)]; %y = y(1:N-M+1);
Xbar = toeplitz([conj(x);zeros(M-1,1)],[conj(x(1)),zeros(1,M-1)]);
R_hat = Xbar’*Xbar;
d_hat = Xbar’*y;
c(:,k) = R_hat\d_hat;
LSE(k) = y’*y - c(:,k)’*d_hat;
end
c2 = mean(c,2);
LSE2 = mean(LSE);

Plots of Channel responses and their corresponding equalizers are shown in Figure-8.31.
 Chapter 9

Signal Modeling and Parametric Spectral Estimation

9.1 The random process x(n) is generated by

1
H (z) =
1− 2.7607z −1 + 3.8108z −2 − 2.6535z −3 + 0.9238z −4
using a WGN(0, 1) process.

(a) The Matlab function armodcov:

function [ahat,VarP] = armodcov(x,P)
% Modofied covariance method to compute AR(P) model coefficients
% [ahat,e,Ep_fb] = armodcov(x,P)
% Inputs:
% x : data vector; P : Model order
% Outputs:
% ahat : coefficient estimates
% e : error vector
% VarP : Error variance (forward_backword)
%
%-----------------------------------------------------------
% Copyright 2000, by Dimitris G. Manolakis, Vinay K. Ingle,
% and Stephen M. Kogon. For use with the book
%
% "Statistical and Adaptive Signal Processing"
%
% McGraw-Hill Higher Education.
%-----------------------------------------------------------

Rbar = lsmatrix(x,P+1);
if isreal(Rbar)
Rfb = Rbar + flipud(fliplr(Rbar));
else
Rfb = Rbar + flipud(fliplr(conj(Rbar)));
end

[ahat,VarP] = olsigest(Rfb,1);
Matlab Verification:
% Generate x(n)
bh = 1; ba = [1,-2.7607,3.8108,-2.6535,0.9238];
N = 250; w = randn(N,1);
x = filter(bh,ba,w);

% (a) Verification of Matlab function "armodcov"

P = 4;
[ahat,VarP] = armodcov(x,P);

248
Statistical and Adaptive Signal Processing - Solution Manual 249

disp(sprintf(’ahat = [%6.4f,%6.4f,%6.4f,%6.4f]’,ahat));
ahat = [-2.8153,3.9155,-2.7413,0.9480]
disp(sprintf(’Error variance: %8.4f’,VarP));
Error variance: 400.1909
(b) The plots are generated using the following Matlab script and shown in Figure 9.1.
% (b) Plot of error variance and MSC vs P
Pmax = 15; V = zeros(Pmax,1);
FPE = V; AIC = V; MDL = V; CAT = V;

for P = 1:Pmax
[ahat,VarP] = armodcov(x,P);
V(P) = VarP;
FPE(P) = (N+P)/(N-P)*VarP;
AIC(P) = N*log10(VarP) + 2*P;
MDL(P) = N*log10(VarP) + P*log10(N);
CAT(P) = sum([N-1:-1:N-P]’./(N*V(1:P)))/N -(N-P)/(N*VarP);
end
P = 1:Pmax;
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
stem(P,V/max(V),’g’); axis([0,Pmax+1,0,1.1]);
xlabel(’Model order P’,’fontsize’,label_fontsize);
ylabel(’Normalized error variance’,’fontsize’,label_fontsize);
title(’Modified Covariance Method’,’fontsize’,title_fontsize);
set(gca,’xtick’,[1:1:Pmax],’ytick’,[0:0.2:1]);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...

subplot(2,2,1); plot(P,FPE/max(FPE),’g’);
%xlabel(’Model order P’,’fontsize’,label_fontsize);
ylabel(’Normalized FPE(P)’,’fontsize’,label_fontsize);
title(’Final Prediction Error’,’fontsize’,title_fontsize);
set(gca,’xtick’,[1:1:Pmax],’xlim’,[0,Pmax+1]);

subplot(2,2,2); plot(P,AIC/max(AIC),’g’);
%xlabel(’Model order P’,’fontsize’,label_fontsize);
ylabel(’Normalized AIC(P)’,’fontsize’,label_fontsize);
title(’Akaike Information Criteria’,’fontsize’,title_fontsize);
set(gca,’xtick’,[1:1:Pmax],’xlim’,[0,Pmax+1]);

subplot(2,2,3); plot(P,MDL/max(MDL),’g’);
xlabel(’Model order P’,’fontsize’,label_fontsize);
ylabel(’Normalized MDL(P)’,’fontsize’,label_fontsize);
title(’Minimum Description Length’,’fontsize’,title_fontsize);
set(gca,’xtick’,[1:1:Pmax],’xlim’,[0,Pmax+1]);

subplot(2,2,4); plot(P,CAT/max(-CAT),’g’);
xlabel(’Model order P’,’fontsize’,label_fontsize);
ylabel(’Normalized CAT(P)’,’fontsize’,label_fontsize);
250 Statistical and Adaptive Signal Processing - Solution Manual

Final Prediction Error Akaike Information Criteria

1 1
Modified Covariance Method

Normalized FPE(P)

Normalized AIC(P)
0.8 0.9

1 0.8
0.6
0.7
0.4
0.6
Normalized error variance

0.8
0.2 0.5

0 0.4
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
0.6
Minimum Description Length Criterion Autoregressive Transfer
1 0

Normalized MDL(P)

Normalized CAT(P)
0.4 0.9 −0.2
0.8
−0.4
0.7
0.2 −0.6
0.6

0.5 −0.8

0 0.4 −1
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Model order P Model order P Model order P

Figure 9.1: Plots of error variance and model selection criteria vs P

title(’Criterion Autoregressive Transfer’,’fontsize’,title_fontsize);

set(gca,’xtick’,[1:1:Pmax],’xlim’,[0,Pmax+1]);
(c) Comments: Both plots indicate that the correct model order is 4.
fb
9.2 The forward/backward LS error Em is given by (9.2.33)


Nf
Em = fb
{|emf (n)|2 + |emb (n)|2 }
n=Ni

(a) Using (9.2.26), we obtain

 f 
e (n)2 = ef (n)ef∗ (n)
m
mf m
∗
 f∗
= em−1 (n) + km−1 b
em−1 (n − 1) em−1 (n) + km−1 em−1
b∗
(n − 1)
 f 2  
= em−1 (n) + |km−1 |2 em−1 (n − 1)
b 2

∗
+em−1
f
(n)km−1 em−1
b∗
(n − 1) + km−1 b
em−1 (n − 1)em−1
f∗
(n)

Similarly, from (9.2.27) we obtain

 b     
e (n)2 = eb (n − 1)2 + |km−1 |2 ef (n)2
m m−1 m−1
∗
+em−1
b
(n − 1)km−1 f∗
em−1 (n) + km−1 em−1
b∗
(n − 1)em−1
f
(n)

Adding the above two equations and summing over Ni ≤ n ≤ N f

 
N N
 f
 f   f
 b 
Emfb = 1 + |km−1 |2 e (n)2 + 1 + |km−1 |2 e (n − 1)2
m−1 m−1
Ni Ni


Nf

Nf
∗
+2km−1 f
em−1 (n)em−1
b∗
(n − 1) + 2km−1 b
em−1 (n − 1)em−1
f∗
(n)
Ni Ni

! N f  f 2 b
,E ! N f  b 2
 !N f f
Using Em−1
f
= Nie m−1 (n) m−1 = Ni em−1 (n − 1) , and β m−1 =
fb∗
Ni em−1 (n)em−1 (n − 1),
b∗

we obtain
  f  ∗
Emfb = 1 + |km−1 |2 Em−1 + Em−1
b
+ 4 Re km−1 β fb
m−1
Statistical and Adaptive Signal Processing - Solution Manual 251

∂ Emfb
(b) Using ∗
∂km−1
= 0, we obtain
 f
2km−1 Em−1 + Em−1
b
m−1 = 0
+ 4β fb

which gives

β fb 2k FP k BP
B
km−1 = − 1  f m−1 b = FP m−1 m−1
2
Em−1 + Em−1 km−1 + km−1
BP

where the last step is obtained by using (9.2.29) and (9.2.32) and the superscript B is used to emphasize
Burg’s contribution.
 FP   BP 
   
 B  km−1 ≤ 1 and km−1 ≤ 1 and since km−1 is the harmonic mean of km−1 and km−1 , it is clear that
B FP BP
(c) Since
k  ≤ 1.
m−1
IS FP BP
(d) From (9.2.36), km−1 is the geometric mean of km−1 and km−1 , that is,
 IS 2  FP   BP   IS 
k  = k    
m−1 m−1 km−1 ≤ 1 or km−1 ≤ 1

9.3 An AR(2) process is generated using the system function

1
H (z) =
1− 0.9z −1+ 0.81z −2
excited by a WGN(0,1) process. The following Matlab script provides the numerical proof.

% Generate x(n)
bh = 1; ba = [1,-0.9,0.81];
N = 250; w = randn(N,1);
x = filter(bh,ba,w);

% Implement Table 9.1 (page 461) using full windowing method

P = 2;
kFP = zeros(1,P-1);
kBP = zeros(1,P-1);
kBurg = zeros(1,P-1);
kIS = zeros(1,P-1);

% 1. Input: x(n) for Ni <= n <= Nf

x = [0;x;0]; Lx = length(x); % Full windowing
ef = zeros(Lx,P-1);
eb = zeros(Lx,P-1);

% 2. Initialization
ef0 = x; eb0 = x;
betaFB0 = ef0(2:end)’*eb0(1:end-1);
Ef0 = ef0’*ef0;
Eb0 = eb0(1:end-1)’*eb0(1:end-1);
kFP0 = -betaFB0/Eb0;
kBP0 = -betaFB0/Ef0;
kBurg0 = 2*kFP0*kBP0/(kFP0+kBP0);
kIS0 = -betaFB0/sqrt(Eb0*Ef0);
252 Statistical and Adaptive Signal Processing - Solution Manual

for n = 2:length(x)
ef(n,1) = ef0(n) + kBurg0*eb0(n-1);
eb(n,1) = eb0(n-1) + kBurg0*ef0(n);
end

% 3. Loop: m = 2,...,P
for m = 2:P
betaFB(m-1) = ef(2:end,m-1)’*eb(1:end-1,m-1);
Ef(m-1) = ef(:,m-1)’*ef(:,m-1);
Eb(m-1) = eb(1:end-1,m-1)’*eb(1:end-1,m-1);
kFP(m-1) = -betaFB(m-1)/Eb(m-1);
kBP(m-1) = -betaFB(m-1)/Ef(m-1);
kBurg(m-1) = 2*kFP(m-1)*kBP(m-1)/(kFP(m-1)+kBP(m-1));
kIS(m-1) = -betaFB(m-1)/sqrt(Eb(m-1)*Ef(m-1));
for n = 2:length(x)
ef(n,m) = ef(n,m-1) + kBurg(m-1)*eb(n-1,m-1);
eb(n,m) = eb(n-1,m-1) + kBurg(m-1)*ef(n,m-1);
end
end

% 4. Output:
kFP = [kFP0,kFP]
kFP = -0.4845 0.8483
kBP = [kBP0,kBP]
kBP = -0.4845 0.8407
kBurg = [kBurg0,kBurg]
kBurg = -0.4845 0.8445
kIS = [kIS0,kIS]
kIS = -0.4845 0.8445

9.4 An AR(2) process is generated using the difference equation

x(n) = w(n) − 1.5857x(n − 1) − 0.9604x(n − 2)

where w(n) WGN(0, 1).

(a) Matlab Script for P = 1:

% Generate x(n)
bh = 1; ba = [1,1.5857,0.9604];
N = 256; n = 0:N-1; w = randn(N,1);
x = filter(bh,ba,w);

% (a) Optimum 1st-order linear predictor

P = 1;
[a,e,V,FPE]=arls(x,P);

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0904a’);
Statistical and Adaptive Signal Processing - Solution Manual 253

L = 10; l = 0:L-1;
K = N/2; f = [0:K]/K; k = 1:K+1;
x = e;
subplot(’position’,[0.1,0.65,0.35,0.3]);
plot(n(1:100),x(1:100),’g’); axis([-1,100,-4,4]);
xlabel(’\itn’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itx}({\itn})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Error samples’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:20:100],’ytick’,[-4:1:4]);

% Autocorrelation test
r = autoc(x,L+1); rho = r(1:L)/r(1);
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.65,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l,rho,’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag \itl’,’fontsize’,label_fontsize);
ylabel(’\rho_{\itx}({\itl})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);

% PSD test
Rx = psd(x,N,2,boxcar(length(x)),’none’);
Ix = cumsum(Rx(1:K+1)); Ix = Ix/Ix(K+1);
Ibl = -1.36/sqrt(K-1) + (k-1)/(K-1);
Ibu = 1.36/sqrt(K-1) + (k-1)/(K-1);
subplot(’position’,[0.1,0.15,0.35,0.3]);
plot(f,Ix,’g’,f,Ibl,’r:’,f,Ibu,’r:’); axis([0,1,-0.2,1.2]);
xlabel(’Frequency (cycles/samp)’,’fontsize’,label_fontsize);
ylabel(’I_{\itx}({\itf})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD test’,’Fontsize’,title_fontsize);
Ibl(1) = round(Ibl(1)*100)/100;
Ibu(1) = round(Ibu(1)*100)/100;
set(gca,’xtick’,[0:0.2:1],’ytick’,[Ibl(1),0,Ibu(1),1]);
set(gca,’xticklabel’,[’ 0 ’;’0.1’;’0.2’;’0.3’;’0.4’;’0.5’]);

% PARCOR test
Vx=r(1); r=r/Vx;
[a,PACS,var]=durbin(r,L); PACS = [1;PACS(1:L-1)];
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.15,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l(2:L),PACS(2:L),’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
254 Statistical and Adaptive Signal Processing - Solution Manual

Error samples Autocorrelation test

4
3 1
2
1

ρx(l)
x(n)
0.12
0 0
−0.12
−1
−2
−3 −1
−4
0 20 40 60 80 100 0 5 10
n Lag l

PSD test Partial autocorrelation test

1
1

PACS
I (f)

0.12
0
x

−0.12

0.12
0
−0.12 −1

0 0.1 0.2 0.3 0.4 0.5 0 5 10

Frequency (cycles/samp) Lag l
Figure 9.4a: Plots of error samples and its whiteness tests for P = 1

plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag {\itl}’,’fontsize’,label_fontsize);
ylabel(’PACS’,’fontsize’,label_fontsize);
title(’Partial autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);
The plots are shown in Figure 9.4a.
(b) Matlab Script for P = 2:
% (b1) Optimum 2nd-order linear predictor
P = 2;
[a,e,V,FPE]=arls(x,P);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’Pr0904b’);

L = 10; l = 0:L-1;
K = N/2; f = [0:K]/K; k = 1:K+1;
x = e;
subplot(’position’,[0.1,0.65,0.35,0.3]);
plot(n(1:100),x(1:100),’g’); axis([-1,100,-4,4]);
xlabel(’\itn’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itx}({\itn})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Error samples’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:20:100],’ytick’,[-4:1:4]);

% Autocorrelation test
r = autoc(x,L+1); rho = r(1:L)/r(1);
conf_lim = 1.96/sqrt(N);
Statistical and Adaptive Signal Processing - Solution Manual 255

subplot(’position’,[0.6,0.65,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l,rho,’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag \itl’,’fontsize’,label_fontsize);
ylabel(’\rho_{\itx}({\itl})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);

% PSD test
Rx = psd(x,N,2,boxcar(length(x)),’none’);
Ix = cumsum(Rx(1:K+1)); Ix = Ix/Ix(K+1);
Ibl = -1.36/sqrt(K-1) + (k-1)/(K-1);
Ibu = 1.36/sqrt(K-1) + (k-1)/(K-1);
subplot(’position’,[0.1,0.15,0.35,0.3]);
plot(f,Ix,’g’,f,Ibl,’r:’,f,Ibu,’r:’); axis([0,1,-0.2,1.2]);
xlabel(’Frequency (cycles/samp)’,’fontsize’,label_fontsize);
ylabel(’I_{\itx}({\itf})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD test’,’Fontsize’,title_fontsize);
Ibl(1) = round(Ibl(1)*100)/100;
Ibu(1) = round(Ibu(1)*100)/100;
set(gca,’xtick’,[0:0.2:1],’ytick’,[Ibl(1),0,Ibu(1),1]);
set(gca,’xticklabel’,[’ 0 ’;’0.1’;’0.2’;’0.3’;’0.4’;’0.5’]);

% PARCOR test
Vx=r(1); r=r/Vx;
[a,PACS,var]=durbin(r,L); PACS = [1;PACS(1:L-1)];
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.15,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l(2:L),PACS(2:L),’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag {\itl}’,’fontsize’,label_fontsize);
ylabel(’PACS’,’fontsize’,label_fontsize);
title(’Partial autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);
The plots are shown in Figure 9.4b.
Matlab Script for P = 3:
% (b2) Optimum 3rd-order linear predictor
P = 3;
[a,e,V,FPE]=arls(x,P);
256 Statistical and Adaptive Signal Processing - Solution Manual

Error samples Autocorrelation test

4
3 1
2
1

ρx(l)
x(n)
0.12
0 0
−0.12
−1
−2
−3 −1
−4
0 20 40 60 80 100 0 5 10
n Lag l

PSD test Partial autocorrelation test

1
1

PACS
I (f)

0.12
0
x

−0.12

0.12
0
−0.12 −1

0 0.1 0.2 0.3 0.4 0.5 0 5 10

Frequency (cycles/samp) Lag l
Figure 9.4b: Plots of error samples and its whiteness tests for P = 2

Hf_3 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_3,’NumberTitle’,’off’,’Name’,’Pr0904c’);

L = 10; l = 0:L-1;
K = N/2; f = [0:K]/K; k = 1:K+1;
x = e;
subplot(’position’,[0.1,0.65,0.35,0.3]);
plot(n(1:100),x(1:100),’g’); axis([-1,100,-4,4]);
xlabel(’\itn’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itx}({\itn})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Error samples’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:20:100],’ytick’,[-4:1:4]);

% Autocorrelation test
r = autoc(x,L+1); rho = r(1:L)/r(1);
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.65,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l,rho,’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag \itl’,’fontsize’,label_fontsize);
ylabel(’\rho_{\itx}({\itl})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);

% PSD test
Rx = psd(x,N,2,boxcar(length(x)),’none’);
Statistical and Adaptive Signal Processing - Solution Manual 257

Error samples Autocorrelation test

4
3 1
2
1

ρx(l)
x(n)
0.12
0 0
−0.12
−1
−2
−3 −1
−4
0 20 40 60 80 100 0 5 10
n Lag l

PSD test Partial autocorrelation test

1
1

PACS
Ix(f)

0.12
0
−0.12

0.12
0
−0.12 −1

0 0.1 0.2 0.3 0.4 0.5 0 5 10

Frequency (cycles/samp) Lag l
Figure 9.4c: Plots of error samples and its whiteness tests for P = 2

Ix = cumsum(Rx(1:K+1)); Ix = Ix/Ix(K+1);
Ibl = -1.36/sqrt(K-1) + (k-1)/(K-1);
Ibu = 1.36/sqrt(K-1) + (k-1)/(K-1);
subplot(’position’,[0.1,0.15,0.35,0.3]);
plot(f,Ix,’g’,f,Ibl,’r:’,f,Ibu,’r:’); axis([0,1,-0.2,1.2]);
xlabel(’Frequency (cycles/samp)’,’fontsize’,label_fontsize);
ylabel(’I_{\itx}({\itf})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD test’,’Fontsize’,title_fontsize);
Ibl(1) = round(Ibl(1)*100)/100;
Ibu(1) = round(Ibu(1)*100)/100;
set(gca,’xtick’,[0:0.2:1],’ytick’,[Ibl(1),0,Ibu(1),1]);
set(gca,’xticklabel’,[’ 0 ’;’0.1’;’0.2’;’0.3’;’0.4’;’0.5’]);

% PARCOR test
Vx=r(1); r=r/Vx;
[a,PACS,var]=durbin(r,L); PACS = [1;PACS(1:L-1)];
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.15,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l(2:L),PACS(2:L),’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag {\itl}’,’fontsize’,label_fontsize);
ylabel(’PACS’,’fontsize’,label_fontsize);
title(’Partial autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);
The plots are shown in Figure 9.4c.
258 Statistical and Adaptive Signal Processing - Solution Manual

9.5 A MA(1) process is generated using the difference equation

x(n) = 0.5w(n) + 0.5w(n − 1)

where w(n) WGN(0, 1).

(a) Matlab Script for the whiteness of x(n):

% Generate x(n)
b = [0.5,0.5]; a = [1];
N = 256; n = 0:N-1; w = randn(N,1);
x = filter(b,a,w);

% (a) Whiteness of x(n)

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0905a’);

L = 10; l = 0:L-1;
K = N/2; f = [0:K]/K; k = 1:K+1;

subplot(’position’,[0.1,0.65,0.35,0.3]);
plot(n(1:100),x(1:100),’g’); axis([-1,100,-4,4]);
xlabel(’\itn’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itx}({\itn})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Signal x(n) Samples’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:20:100],’ytick’,[-4:1:4]);

% Autocorrelation test
r = autoc(x,L+1); rho = r(1:L)/r(1);
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.65,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l,rho,’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag \itl’,’fontsize’,label_fontsize);
ylabel(’\rho_{\itx}({\itl})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Autocorrelation Test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);

% PSD test
Rx = psd(x,N,2,boxcar(length(x)),’none’);
Ix = cumsum(Rx(1:K+1)); Ix = Ix/Ix(K+1);
Ibl = -1.36/sqrt(K-1) + (k-1)/(K-1);
Ibu = 1.36/sqrt(K-1) + (k-1)/(K-1);
subplot(’position’,[0.1,0.15,0.35,0.3]);
plot(f,Ix,’g’,f,Ibl,’r:’,f,Ibu,’r:’); axis([0,1,-0.2,1.2]);
Statistical and Adaptive Signal Processing - Solution Manual 259

Signal x(n) Samples Autocorrelation Test

4
3 1
2
1

ρx(l)
x(n)
0.12
0 0
−0.12
−1
−2
−3 −1
−4
0 20 40 60 80 100 0 5 10
n Lag l

PSD Test Partial Autocorrelation Test

1
1

PACS
Ix(f)

0.12
0
−0.12

0.12
0
−0.12 −1

0 0.1 0.2 0.3 0.4 0.5 0 5 10

Frequency (cycles/samp) Lag l
Figure 9.5a: Plots of signal samples and its whiteness tests for x(n)

xlabel(’Frequency (cycles/samp)’,’fontsize’,label_fontsize);
ylabel(’I_{\itx}({\itf})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD Test’,’Fontsize’,title_fontsize);
Ibl(1) = round(Ibl(1)*100)/100;
Ibu(1) = round(Ibu(1)*100)/100;
set(gca,’xtick’,[0:0.2:1],’ytick’,[Ibl(1),0,Ibu(1),1]);
set(gca,’xticklabel’,[’ 0 ’;’0.1’;’0.2’;’0.3’;’0.4’;’0.5’]);

% PARCOR test
Vx=r(1); r=r/Vx;
[a,PACS,var]=durbin(r,L); PACS = [1;PACS(1:L-1)];
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.15,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l(2:L),PACS(2:L),’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag {\itl}’,’fontsize’,label_fontsize);
ylabel(’PACS’,’fontsize’,label_fontsize);
title(’Partial Autocorrelation Test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);
The plots are shown in Figure 9.5a.
(b) The signal x(n) is processed through the AR(1) filter
1
H (z) =
1 + 0.95z −1
to generate y(n). Matlab Script for the whiteness of y(n):
% (b) Whiteness of y(n)
260 Statistical and Adaptive Signal Processing - Solution Manual

bH = 1; aH = [1,0.95];
y = filter(bH,aH,x);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’Pr0905b’);

L = 10; l = 0:L-1;
K = N/2; f = [0:K]/K; k = 1:K+1;
x = y;
subplot(’position’,[0.1,0.65,0.35,0.3]);
plot(n(1:100),x(1:100),’g’); axis([-1,100,-4,4]);
xlabel(’\itn’,’fontsize’,label_fontsize,’fontname’,’times’);
ylabel(’{\itx}({\itn})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Signal y(n) samples’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:20:100],’ytick’,[-4:1:4]);

% Autocorrelation test
r = autoc(x,L+1); rho = r(1:L)/r(1);
conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.65,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l,rho,’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag \itl’,’fontsize’,label_fontsize);
ylabel(’\rho_{\itx}({\itl})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’Autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);

% PSD test
Rx = psd(x,N,2,boxcar(length(x)),’none’);
Ix = cumsum(Rx(1:K+1)); Ix = Ix/Ix(K+1);
Ibl = -1.36/sqrt(K-1) + (k-1)/(K-1);
Ibu = 1.36/sqrt(K-1) + (k-1)/(K-1);
subplot(’position’,[0.1,0.15,0.35,0.3]);
plot(f,Ix,’g’,f,Ibl,’r:’,f,Ibu,’r:’); axis([0,1,-0.2,1.2]);
xlabel(’Frequency (cycles/samp)’,’fontsize’,label_fontsize);
ylabel(’I_{\itx}({\itf})’,’fontsize’,label_fontsize,’fontname’,’times’);
title(’PSD test’,’Fontsize’,title_fontsize);
Ibl(1) = round(Ibl(1)*100)/100;
Ibu(1) = round(Ibu(1)*100)/100;
set(gca,’xtick’,[0:0.2:1],’ytick’,[Ibl(1),0,Ibu(1),1]);
set(gca,’xticklabel’,[’ 0 ’;’0.1’;’0.2’;’0.3’;’0.4’;’0.5’]);

% PARCOR test
Vx=r(1); r=r/Vx;
Statistical and Adaptive Signal Processing - Solution Manual 261

Signal y(n) samples Autocorrelation test

4
3 1
2
1

ρx(l)
x(n)
0.12
0 0
−0.12
−1
−2
−3 −1
−4
0 20 40 60 80 100 0 5 10
n Lag l

PSD test Partial autocorrelation test

1
1

PACS
Ix(f)

0.12
0
−0.12

0.12
0
−0.12 −1

0 0.1 0.2 0.3 0.4 0.5 0 5 10

Frequency (cycles/samp) Lag l
Figure 9.5b: Plots of signal samples and its whiteness tests for y(n)

[a,PACS,var]=durbin(r,L); PACS = [1;PACS(1:L-1)];

conf_lim = 1.96/sqrt(N);
subplot(’position’,[0.6,0.15,0.35,0.3]);
plot([-1,L],[0,0],’w’); hold on;
Hr = stem(l(2:L),PACS(2:L),’filled’,’g’);
set(Hr,’markersize’,3); axis([-1,L,-1.2,1.2]);
plot([-1,L],[-conf_lim,-conf_lim],’r:’);
plot([-1,L],[conf_lim,conf_lim],’r:’);
hold off; conf_lim = round(conf_lim*100)/100;
xlabel(’Lag {\itl}’,’fontsize’,label_fontsize);
ylabel(’PACS’,’fontsize’,label_fontsize);
title(’Partial autocorrelation test’,’Fontsize’,title_fontsize);
set(gca,’xtick’,[0:5:L],’ytick’,[-1,-conf_lim,0,conf_lim,1]);
The plots are shown in Figure 9.5b.

9.6 The process x(n) is given by

x(n) = Ae j (ω0 n+θ) + w(n)

where A is real positive constant, θ is unformly distributed between [0, 2π ], ω0 is a constant between [0, π ],
and w(n)˜WGN(0, 1) that is uncorrelated with θ.

(a) The autocorrelation of x(n) is given by

 
r x (l) = E x(n)x ∗ (n − l)
  
= E Ae j (ω0 n+θ ) + w(n) Ae− j (ω0 [n−l]+θ ) + w ∗ (n − l)
= A2 e j ω0 l + rw (l) = A2 e j ω0 l + σ 2w δ(l)
262 Statistical and Adaptive Signal Processing - Solution Manual

Thus the (P + 1) × (P + 1) autocorrelation matrix is given by

 
r x (0) r x (1) r x (2) ··· r x (P)
 r ∗ (1) r x (0) r x (1) ··· r x (P − 1) 
 x 
 r ∗ (2) ∗
(1) (0) ··· r x (P − 2) 
Rx =  x r x r x 
 .. .. .. .. .. 
 . . . . . 
∗ ∗ ∗
r x (P) r x (P − 1) r x (P − 2) · · · r x (0)
 
1 e j ω0 e j 2ω0 ··· e j Pω0
 e − j ω0 1 e j ω0 ··· e j ( P−1)ω0 
 
 − j 2ω − j ω
··· e j ( P−2)ω0 
= A2  e e 0 1  + σ 2w I
0

 .. .. .. .. .. 
 . . . . . 
e− j Pω0 e− j ( P−1)ω0 e− j ( P−2)ω0 ··· 1

The matrix in the last equation can be expressed as

 
1
 e − j ω0 
 
 e− j 2ω0  
  1 e j ω0 e j 2ω0 · · · e j Pω0
 ..   
 .  eH
e− j Pω0
  
e

Hence

Rx = A2 eeH + σ 2w I

(b) To be completed
(c) To be completed

9.7 An AR(2) process y(n) is observed in noise v(n) to obtain x(n), that is,

x(n) = y(n) + v(n); v(n)˜WGN(0, σ 2v )

whwre v(n) is uncorrelated with y(n) and

yn = 1.27y(n − 1) − 0.81y(n − 2) + w(n); w(n)˜WGN(0, 1)

(a) Matlab script to compute and plot the true spectrum R(e j ω ):
% (a) Plot of the true spectrum of x(n)
b = [1]; a = [1,-1.27,0.81]; % AR(2) signal y(n) parameters
Rv = 10; sigv = sqrt(Rv); % Additive noise variance
omg = [0:500]*pi/500; % 501 frequency points over 0 to pi
H = freqz(b,a,omg); % AR(2) filter response
Ry = abs(H).*abs(H); % Spectrum of y(n)
Rx = Ry + Rv; % Spetcrum of x(n)
Rx_db = 10*log10(Rx); % Spectrum in dB

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4]);
Statistical and Adaptive Signal Processing - Solution Manual 263

set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0907’);

subplot(2,2,1);
plot(omg/pi,Rx_db,’g’);
title(’True Spectrum’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega}) in dB’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx_db)/10)*10,ceil(max(Rx_db)/10)*10]);
text(0.6,15,’R_v = 10’,’fontsize’,text_fontsize);
The plot is shown in Figure 9.7.
(b) LS estimate of power spectrum using forward-backward linear predictor with P = 2 and σ 2v = 1:
% (b) Spectrum Estimation using LS (FB-predictor) Approach (Rv = 1)
N = 256; n = 0:N-1;
Rv = 1; sigv = sqrt(Rv);
P = 2;

% Loop over realizations

M = 10;
subplot(2,2,2);
for i = 1:M
% Generate y(n) and x(n)
w = randn(N,1);
y = filter(b,a,w);
v = sigv*randn(N,1);
x = y + v;
% Estimate model parameters
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);
Rxhat_db = 10*log10(Rxhat);
plot(omg/pi,Rxhat_db,’g’); hold on;
end
title(’Estimated Spectrum’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rxhat_db)/10)*10,ceil(max(Rxhat_db)/10)*10]);
text(0.6,2,[’R_v = ’,num2str(Rv)],’fontsize’,text_fontsize);
The plot is shown in Figure 9.7.
(c) LS estimate of power spectrum using forward-backward linear predictor with P = 2 and σ 2v = 10:
% (c) Spectrum Estimation using LS (FB-predictor) Approach (Rv = 10)
N = 256; n = 0:N-1;
Rv = 10; sigv = sqrt(Rv);
P = 2;

% Loop over realizations

M = 10;
subplot(2,2,3);
264 Statistical and Adaptive Signal Processing - Solution Manual

True Spectrum Estimated Spectrum

20 10

18
5
Rx(ejω) in dB
16 Rv = 1

Rx(ejω)
Rv = 10 0
14

−5
12

10 −10
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Estimated Spectrum Modified Estimated Spectrum

10 40

30
5
20
Rv = 10
Rx(ejω)

Rx(ejω)
0 10

Rv = 1
0
−5
−10

−10 −20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
normalized frequency normalized frequency
Figure 9.7: Plots of various spectra in P9.7

Rxhatdbmin = 0; Rxhatdbmax = 0;
for i = 1:M
% Generate y(n) and x(n)
w = randn(N,1);
y = filter(b,a,w);
v = sigv*randn(N,1);
x = y + v;
% Estimate model parameters
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);
Rxhat_db = 10*log10(Rxhat);
Rxhatdbmin = min([Rxhatdbmin,min(Rxhat_db)]);
Rxhatdbmax = max([Rxhatdbmax,max(Rxhat_db)]);
plot(omg/pi,Rxhat_db,’g’); hold on;
end
title(’Estimated Spectrum’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(Rxhatdbmin/10)*10,ceil(Rxhatdbmax/10)*10]);
text(0.6,2,[’R_v = ’,num2str(Rv)],’fontsize’,text_fontsize);
The plot is shown in Figure 9.7.
(d) The effect of subtracting a small amount of noise from r x (0) on the LS estimate of power spectrum using
forward-backward linear predictor with P = 2 and σ 2v = 1:
% (d) Spectrum Estimation using LS (FB-predictor) Approach (Rv = 1)
Statistical and Adaptive Signal Processing - Solution Manual 265

N = 256; n = 0:N-1;
Rv = 1; sigv = sqrt(Rv);
P = 2;

% Loop over realizations

M = 10;
subplot(2,2,4);
Rxhatdbmin = 0; Rxhatdbmax = 0;
for i = 1:M
% Generate y(n) and x(n)
w = randn(N,1);
y = filter(b,a,w);
v = sigv*randn(N,1);
x = y + v;
% Estimate model parameters
Rbar = lsmatrix(x,P+1);
Rbar = Rbar - diag(N*ones(length(Rbar),1));
Rfb = Rbar + flipud(fliplr(Rbar));
[ahat,VarP] = olsigest(Rfb,1);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);
Rxhat_db = 10*log10(Rxhat);
Rxhatdbmin = min([Rxhatdbmin,min(Rxhat_db)]);
Rxhatdbmax = max([Rxhatdbmax,max(Rxhat_db)]);
plot(omg/pi,Rxhat_db,’g’); hold on;
end
title(’Modified Estimated Spectrum’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(Rxhatdbmin/10)*10,ceil(Rxhatdbmax/10)*10]);
text(0.6,2,[’R_v = ’,num2str(Rv)],’fontsize’,text_fontsize);
The plot is shown in Figure 9.7.

9.8 The first five estimated correlation lag values of a process x(n) are

r x (0) = 1, r x (1) = 0.7, r x (2) = 0.5, r x (3) = 0.3, and r x (4) = 0

(a) The Blackman-Tukey Power spectrum estimates:

rx = [1;0.75;0.5;0.25;0]; L = length(rx);

% (a) Blackman-Tukey Estimates

Nfft = 1024;
RxBT = real(fft([rx;zeros(Nfft-2*L+1,1);flipud(rx(2:end))]));
RxBT = RxBT(1:Nfft/2+1);
omg = [0:Nfft/2]*2*pi/Nfft;

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4]);
266 Statistical and Adaptive Signal Processing - Solution Manual

Blackman−Tuckey Estimates AP(2) Spectral Estimates

4 4

Rx(ejω) 3 3

Rx(ejω)
2 2

1 1

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
normalized frequency
AP(4) Spectral Estimates
2

1.5
Rx(ejω)

0.5

0
0 0.2 0.4 0.6 0.8 1
normalized frequency
Figure 9.8: Plots of various spectra in P9.8

set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0908’);

subplot(2,2,1);
plot(omg/pi,RxBT,’g’,’linewidth’,1);
title(’Blackman-Tuckey Estimates’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,8);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(RxBT)/1.0)*1.0,ceil(max(RxBT)/1.0)*1.0]);

The plot is shown in Figure 9.8.

(b) Spectrum estimates using AP(2) model:
% (b) Spectrum Estimation using AP(2) model
P = 2;
Rbar = toeplitz(rx(1:P),rx(1:P));
[ahat,VarP] = olsigest(Rbar,1);
Hhat = freqz(VarP,[1;ahat]’,omg);
RxAP = abs(Hhat).*abs(Hhat);
subplot(2,2,2); plot(omg/pi,RxAP,’g’,’linewidth’,1); hold on;

title(’AP(2) Spectral Estimates’,’fontsize’,title_fontsize);

xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,0,ceil(max(RxAP)/1.0)*1.0]);
The plot is shown in Figure 9.8.
(c) Spectrum estimates using AP(4) model:
Statistical and Adaptive Signal Processing - Solution Manual 267

% (c) Spectrum Estimation using AP(4) model

P = 4;
Rbar = toeplitz(rx(1:P),rx(1:P));
[ahat,VarP] = olsigest(Rbar,1);
Hhat = freqz(VarP,[1;ahat]’,omg);
RxAP = abs(Hhat).*abs(Hhat);
subplot(2,2,3); plot(omg/pi,RxAP,’g’,’linewidth’,1); hold on;

title(’AP(4) Spectral Estimates’,’fontsize’,title_fontsize);

xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,0,ceil(max(RxAP)/1.0)*1.0]);
The plot is shown in Figure 9.8.

9.9 The narrowband process x(n) is generated using the AP(4) model
1
H (z) =
1+ 0.98z −1 + 1.92z −2 + 0.94z −3 + 0.92z −4
driven by WGN(0, 0.001).

(a) Matlab script to compute and plot the true spectrum R(e j ω ):
% (a) Plot of the true spectrum of x(n)
b = [1]; a = [1,0.98,1.92,0.94,0.92]; % AP(4) signal y(n) parameters
omg = [0:500]*pi/500; % 501 frequency points over 0 to pi
H = freqz(b,a,omg); % AP(4) filter response
Rx = abs(H).*abs(H); % Spectrum of x(n)
Rx = 10*log10(Rx); % Spectrum in dB

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS+[0,0,0,0.2],...
’paperunits’,PAPUN,’paperposition’,[0,0,6,6]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0909’);

subplot(3,2,1);
plot(omg/pi,Rx,’g’); title(’True Spectrum’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);
The plot is shown in Figure 9.9.
(b) LS estimate of power spectrum using forward linear predictor with P = 4:
% (b) Spectrum Estimation using LS (Forward-predictor) Approach (P = 4)
N = 256; n = 0:N-1;
P = 4;

% Loop over realizations

M = 10;
subplot(3,2,2);
for i = 1:M
268 Statistical and Adaptive Signal Processing - Solution Manual

True Spectrum Estimated Spectrum (P = 4)

30 30

20 20
Rx(ejω)

Rx(ejω)
10 10

0 0

−10 −10

−20 −20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Estimated Spectrum (P = 8) Estimated Spectrum (P = 12)

30 30

20 20
Rx(ejω)

Rx(ejω)
10 10

0 0

−10 −10

−20 −20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

True Spectrum Estimated Spectrum (FB, P = 4)

30 30

20 20
Rx(ejω)

Rx(ejω)

10 10

0 0

−10 −10

−20 −20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
normalized frequency normalized frequency
Figure 9.9: Plots of various spectra in P9.9

% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
Rbar = lsmatrix(x,P+1); [ahat,VarP] = olsigest(Rbar,1);
%[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Estimated Spectrum (P = 4)’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);
The plot is shown in Figure 9.9.
(c) LS estimate of power spectrum using forward linear predictor with P = 8 and P = 12:
Statistical and Adaptive Signal Processing - Solution Manual 269

% (c) Spectrum Estimation using LS (Forward-predictor) Approach (P = 8 and 10)

P = 8;

% Loop over realizations

M = 10;
subplot(3,2,3);
for i = 1:M
% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
Rbar = lsmatrix(x,P+1); [ahat,VarP] = olsigest(Rbar,1);
%[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Estimated Spectrum (P = 8)’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);

% Loop over realizations (P = 12)

P = 12; M = 10;
subplot(3,2,4);
for i = 1:M
% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
Rbar = lsmatrix(x,P+1); [ahat,VarP] = olsigest(Rbar,1);
%[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Estimated Spectrum (P = 12)’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);
The plots are shown in Figure 9.9.
(d) LS estimate of power spectrum using forward-backward linear predictor with P = 4:
% (d) Spectrum Estimation using LS (FB-predictor) Approach (P = 4)
P = 4;

% Loop over realizations

M = 10;
270 Statistical and Adaptive Signal Processing - Solution Manual

subplot(3,2,5);
plot(omg/pi,Rx,’g’); title(’True Spectrum’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);

subplot(3,2,6);
for i = 1:M
% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Estimated Spectrum (FB, P = 4)’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]);
The plot is shown in Figure 9.9.

9.10 An ARMA process x(n) is generated using the PZ(4,2) model

1 − z −2
H (z) =
1 + 0.41z −4
driven by WGN(0, 1).

(a) Matlab script to compute and plot the true spectrum R(e j ω ):
% (a) Plot of the true spectrum of x(n)
b = [1,0,-1]; a = [1,0,0,0,0.41]; % AP(4) signal y(n) parameters
omg = [0:500]*pi/500; % 501 frequency points over 0 to pi
H = freqz(b,a,omg); % AP(4) filter response
Rx = abs(H).*abs(H); % Spectrum of x(n)
Rx = 10*log10(Rx + eps); % Spectrum in dB

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS+[0,0,0,0],...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0910’);

subplot(2,2,1);
plot(omg/pi,Rx,’g’);
title(’True Spectrum’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,-50,20]);
The plot is shown in Figure 9.10.
Statistical and Adaptive Signal Processing - Solution Manual 271

True Spectrum Est. Spec. LS FB Method (P = 12)

20 20

10 10

0 0

Rx(ejω)

Rx(ejω)
−10 −10

−20 −20

−30 −30

−40 −40

−50 −50
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
normalized frequency
Est. Spec. NL LS−PZ Method (4,2)
20

10

0
Rx(ejω)

−10

−20

−30

−40

−50
0 0.2 0.4 0.6 0.8 1
normalized frequency
Figure 9.10: Plots of various spectra in P9.10

(b) LS estimate of power spectrum using forward-backward linear predictor with P = 12:
% (b) Spectrum Estimation using LS (Forward-predictor) Approach (P = 12)
N = 256; n = 0:N-1;
P = 12;

% Loop over realizations

M = 10;
subplot(2,2,2);
for i = 1:M
% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(b,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat+eps);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Est. Spec. LS FB Method (P = 12)’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,-50,20]);
The plot is shown in Figure 9.10.
(c) Nonlinear LS estimate of power spectrum using pole-zero modeling algorithm with P = 4 and Q = 2:
272 Statistical and Adaptive Signal Processing - Solution Manual

% (c) Spectrum Estimation using Nonlinear LS-PZ algorithm of 9.3.3

P = 4; Q = 2;

% Loop over realizations

M = 10;
subplot(2,2,3);
for i = 1:M
% Generate x(n)
w = randn(N,1);
x = filter(b,a,w);
% Estimate model parameters
[ahat,bhat,V,FPE,C] = armals(x,P,Q,100,1e-4);
Hhat = freqz(bhat,ahat,omg);
Rxhat = abs(Hhat).*abs(Hhat); Rxhat = 10*log10(Rxhat);
plot(omg/pi,Rxhat,’g’); hold on;
end
title(’Est. Spec. NL LS-PZ Method (4,2)’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,-50,20]);
The plots are shown in Figure 9.10.

9.11 A random process x(n) is given by

 πn 

2π n
x(n) = cos + θ 1 + w(n) − w(n − 2) + cos + θ2
3 3

where w(n)˜WGN(0, 1) and θ 1 and θ 2 are IID random variables uniformly distributed over [0, 2π ].

(a) Matlab script to compute and plot the true spectrum R(e j ω ):
% (a) Plot of the true spectrum of x(n)
omg1 = pi/3; omg2 = 2*pi/3;
omg = [0:300]*pi/300; % 301 frequency points over 0 to pi
Rx = 2*(1-cos(2*omg)); % Spectrum of x(n)

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS+[0,0,0,0],...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0911a’);

subplot(2,2,1);
plot(omg/pi,Rx,’g’);
title(’True Spectrum’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,0,5]); hold on;
plot([omg1/pi,omg1/pi],[0,4],’c’);
fill([1/3-0.03,1/3+0.03,1/3,1/3-0.03]’,[4,4,4+0.3,4]’,’c’);
plot([omg2/pi,omg2/pi],[0,4],’m’);
fill([2/3-0.03,2/3+0.03,2/3,2/3-0.03]’,[4,4,4+0.3,4]’,’m’);
Statistical and Adaptive Signal Processing - Solution Manual 273

set(gca,’xtick’,[0,1/3,2/3,1]);
set(gca,’xticklabel’,[’ 0 ’;’ 1/3’;’ 2/3’;’ 1 ’])
The plot is shown in Figure 9.11a.
(b) LS estimate of power spectrum using forward-backward linear predictor with P = 10, 20, and40:
% Generate x(n)
N = 256; n = [0:N-1]’;
w = randn(N,1); y = filter([1,0,-1],1,w);
x = y + cos(pi*n/3+rand(1)*2*pi) + cos(2*pi*n/3+rand(1)*2*pi);

% (b) Spectrum Estimation using LS (FB-predictor) Approach (P = 10,20,40)

% Estimate model parameters (P = 10)

P = 10;
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(1,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);

subplot(2,2,2);
plot(omg/pi,Rxhat,’g’); hold on;
title(’Est. Spec. LS FB Method (P = 10)’,’fontsize’,title_fontsize);
%xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize); %axis([0,1,-50,20]);
set(gca,’xtick’,[0,1/3,2/3,1]);
set(gca,’xticklabel’,[’ 0 ’;’ 1/3’;’ 2/3’;’ 1 ’])

% Estimate model parameters (P = 10)

P = 20;
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(1,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);

subplot(2,2,3);
plot(omg/pi,Rxhat,’g’); hold on;
title(’Est. Spec. LS FB Method (P = 20)’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize); %axis([0,1,-50,20]);
set(gca,’xtick’,[0,1/3,2/3,1]);
set(gca,’xticklabel’,[’ 0 ’;’ 1/3’;’ 2/3’;’ 1 ’])

% Estimate model parameters (P = 10)

P = 40;
[ahat,VarP] = armodcov(x,P);
Hhat = freqz(1,[1;ahat]’,omg);
Rxhat = abs(Hhat).*abs(Hhat);

subplot(2,2,4);
plot(omg/pi,Rxhat,’g’); hold on;
title(’Est. Spec. LS FB Method (P = 40)’,’fontsize’,title_fontsize);
274 Statistical and Adaptive Signal Processing - Solution Manual

True Spectrum Est. Spec. LS FB Method (P = 10)

5 15

4
10
Rx(ejω)

Rx(ejω)
3

2
5
1

0 0
0 1/3 2/3 1 0 1/3 2/3 1

Est. Spec. LS FB Method (P = 20) Est. Spec. LS FB Method (P = 40)

30 120

25 100

20 80
Rx(ejω)

Rx(ejω)
15 60

10 40

5 20

0 0
0 1/3 2/3 1 0 1/3 2/3 1
normalized frequency normalized frequency
Figure 9.11a: Plots of various spectra in P9.11a and b

xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize); %axis([0,1,-50,20]);
set(gca,’xtick’,[0,1/3,2/3,1]);
set(gca,’xticklabel’,[’ 0 ’;’ 1/3’;’ 2/3’;’ 1 ’])

exportfig(gcf,’p0911a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);
The plots are shown in Figure 9.11a.
(c) Nonlinear LS estimate of power spectrum using pole-zero modeling algorithm with P = 4 and Q = 2:
% (c) Spectrum Estimation usinf NL PZ modeling algorithm of Sec. 9.3.3
Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS+[0,0,0,0],...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’Pr0911b’);
subplot(2,2,3);
plot(omg/pi,Rx,’g’); title(’True Spectrum’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,0,5]); hold on;
plot([omg1/pi,omg1/pi],[0,4],’c’);
fill([1/3-0.03,1/3+0.03,1/3,1/3-0.03]’,[4,4,4+0.3,4]’,’c’);
plot([omg2/pi,omg2/pi],[0,4],’m’);
fill([2/3-0.03,2/3+0.03,2/3,2/3-0.03]’,[4,4,4+0.3,4]’,’m’);
set(gca,’xtick’,[0,1/3,2/3,1]);
set(gca,’xticklabel’,[’ 0 ’;’ 1/3’;’ 2/3’;’ 1 ’])
Statistical and Adaptive Signal Processing - Solution Manual 275

True Spectrum Est. Spec. NL LS−PZ Method (4,2)

5 6

4 5

Rx(ejω)

Rx(ejω)
3
3
2
2
1 1

0 0
0 1/3 2/3 1 0 0.2 0.4 0.6 0.8 1
normalized frequency normalized frequency
Figure 9.11b: Plots of various spectra in P9.11c

% Estimate model parameters

subplot(2,2,4);
P = 4; Q = 2;
[ahat,bhat,V,FPE,C] = armals(x,P,Q,1000,1e-5);
Hhat = freqz(bhat,ahat,omg);
Rxhat = abs(Hhat).*abs(Hhat);
plot(omg/pi,Rxhat,’g’);

title(’Est. Spec. NL LS-PZ Method (4,2)’,’fontsize’,title_fontsize);

xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
The plots are shown in Figure 9.11b.

9.12 For large values of N , the correlation matrix obtained using no windowing tends to a Toeplitz Hermitian matrix
and as N → ∞, the matrix in fact becomes a Toeplitz Hermitian matrix. Thus the model coefficients can be
obtained by the Levinson-Durbin algorithm. Therefore, the modeling error variance estimates are given by
(7.4.21) and (7.4.16). Using Pm = σ̂ 2m , we obtain
∗
σ̂ 2m = σ̂ 2m−1 + β m−1 km−1
∗

= σ̂ 2m−1 − σ̂ 2m−1 km−1 km−1 = σ̂ 2m−1 1 − |km−1 |2

9.13 To be completed

9.14 An ARMA process x(n) is generated by the PZ model

x(n) = 0.3x(n − 1) + 0.4x(n − 2) + w(n) + 0.25w(n − 2)

driven by w(n)˜WGN(0, 10). The PZ model parameters are estimated by using the equation error method of
Section 9.3.1.

(a) Estimates of model parameters:

% Generate input and output signal of the PZ(2,2) model
N = 2000;
b = [1,0,0.25]; a = [1,-0.3,-0.4];
varw = 10; w = sqrt(varw)*randn(N,1);
276 Statistical and Adaptive Signal Processing - Solution Manual

x = filter(b,a,w);

% (a) Estimates of PZ(2,2) model coefficients

P = 2; Q = 2;
[ahat,bhat,varwhat,FPE] = pzls(x,w,P,Q);

Model Parameters
bhat: 1.0000 -0.0254 0.2386
ahat: 1.0000 -0.2932 -0.4198
(b) Estimate of error variance:
% (b) Input variance Estimate

True error variance: 10.0000

Estimated error variance: 10.3212

9.15 An ARMA process x(n) is generated by the PZ(4,2) model

x(n) = 1.8766x(n − 1) − 2.6192x(n − 2) + 1.6936x(n − 3) − 0.8145x(n − 4)

+w(n) + 0.05w(n − 1) − 0.855w(n − 2)

driven by w(n)˜WGN(0, 10).

(a) Parameter and error variance estimation using nonlinear LS pole-zero modeling algorithm:
% Generate input and output signal of the PZ(4,2) model
N = 300;
b = [1,0.05,-0.855]; a = [1,-1.8766,2.6192,-1.6936,0.8145];
varw = 10; w = sqrt(varw)*randn(N,1);
x = filter(b,a,w);

% (a) Estimates of PZ(4,2) model coefficients

P = 4; Q = 2;
[ahat,bhat,varwhat,FPE,C] = armals(x,P,Q,100,0.001);

Model Parameters
bhat: 1.0000 0.0245 -0.8911
ahat: 1.0000 -1.9019 2.6539 -1.7316 0.8217

True error variance: 10.0000

Estimated error variance: 10.0175

(b) Error variance estimation using AP(10) modeling algorithm:

% (b) Estimates using AP(10) model
P2 = 10;
%Rbar = lsmatrix(x,P2+1); [ahat2,VarP] = olsigest(Rbar,1);
[ahat2,e,VarP] = arls(x,P2);

Estimated error variance: 10.4861

Statistical and Adaptive Signal Processing - Solution Manual 277

Comparison of Spectra
40
True
PZ(4,2)
30 AP(10)

20

10

Rx(ejω) 0

−10

−20

−30

−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized frequency
Figure 9.15: Comparison plots of various spectra in P9.15

(c) Plots:
% (c) Plots
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0915’);

% True Spectra
omg = [0:500]*pi/500;
H = freqz(b,a,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’g’);
title(’Comparison of Spectra’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]); hold on;

% PZ(4,2) model spectra

H = freqz(bhat,ahat,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’r:’);

% AP(10) model Spectra

H = freqz(1,ahat2,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’m--’);

legend(’True’,’PZ(4,2)’,’AP(10)’)
278 Statistical and Adaptive Signal Processing - Solution Manual

The comparison plots are shown in Figure 9.15.

9.16 To be completed

9.17 An ARMA process x(n) is generated by the PZ(4,2) model

x(n) = 1.8766x(n − 1) − 2.6192x(n − 2) + 1.6936x(n − 3) − 0.8145x(n − 4)

+w(n) + 0.05w(n − 1) − 0.855w(n − 2)

driven by w(n)˜WGN(0, 10).

(a) AP(5) model and its spectrum

% Generate input and output signal of the PZ(4,2) model
N = 300;
b = [1,0.05,-0.855]; a = [1,-1.8766,2.6192,-1.6936,0.8145];
varw = 10; w = sqrt(varw)*randn(N,1);
x = filter(b,a,w);
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’Pr0917’);

% (a) Estimates using AP(5) model

P1 = 5;
[ahat1,e,VarP1] = arls(x,P1);

% AP(5) model spectra

H = freqz(1,ahat1,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’r:’,’linewidth’,1);
The plot is shown in Figure 9.17.
(b) AP(10) model and its spectrum
% (b) Estimates using AP(10) model
P2 = 10;
[ahat2,e,VarP2] = arls(x,P2);

% AP(10) model Spectra

H = freqz(1,ahat2,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’m--’,’linewidth’,1);
The plot is shown in Figure 9.17.
(c) AP(50) model and its spectrum
% (c) Estimates using AP(50) model
P3 = 50;
[ahat3,e,VarP3] = arls(x,P3);
Statistical and Adaptive Signal Processing - Solution Manual 279

Comparison of Spectra
40
True
AP(5)
30 AP(10)
AP(50)

20

10

Rx(ejω) 0

−10

−20

−30

−40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
normalized frequency
Figure 9.17: Comparison plots of various spectra in P9.17

% AP(50) model Spectra

H = freqz(1,ahat3,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’c-.’,’linewidth’,1);
The plot is shown in Figure 9.17.
(d) True spectrum
% (d) True Spectra
omg = [0:500]*pi/500;
H = freqz(b,a,omg);
Rx = abs(H).*abs(H);
Rx = 10*log10(Rx);
plot(omg/pi,Rx,’g’,’linewidth’,1);
title(’Comparison of Spectra’,’fontsize’,title_fontsize);
xlabel(’normalized frequency’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega})’,’fontsize’,label_fontsize);
axis([0,1,floor(min(Rx)/10)*10,ceil(max(Rx)/10)*10]); hold on;

legend(’True’,’AP(5)’,’AP(10)’,’AP(50)’)
The plot is shown in Figure 9.17.

9.18 Spectral estimation of a speech signal

(a) Periodogram of the speech signal:

% Read the speech file
x = load(’testspee.dat’); Fs = 10000;
x = filter([1,-1],1,x);
N = length(x);
280 Statistical and Adaptive Signal Processing - Solution Manual

Comparison of Spectra
40
Periodogram
AP(16)
30 PZ(12,6)

20

10
Rx(ejω) in dB

−10

−20

−30

−40
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
frequency in Hz
Figure 9.18: Comparison plots of various spectra in P9.18

%(a) Periodogram of the speech signal

win = boxcar(N); Nfft = 2^(ceil(log10(N)/log10(2)));
Rx = psd(x,Nfft,Fs,win,’none’);
(b) Spectral estimate based on AP(16) model:
%(b) Estimates using AP(16) model
P2 = 16;
[ahat2,e,VarP] = arls(x,P2);
(c) Spectral estimate based on PZ(12,6) model:
% (c) Estimates using PZ(12,6) model
P = 12; Q = 6;
[ahat,bhat,varwhat,FPE,C] = armals(x,P,Q,100,0.001);
(d) Plots:
% (d) Plots
% Periodogram Spectra
freq = [0:Nfft/2]*Fs/Nfft;
Rx1 = 10*log10(Rx)-60;
plot(freq,Rx1,’g’);
title(’Comparison of Spectra’,’fontsize’,title_fontsize);
xlabel(’frequency in Hz’,’fontsize’,label_fontsize);
ylabel(’R_x(e^{j\omega}) in dB’,’fontsize’,label_fontsize);
hold on;

% AP(16) model Spectra

omg = [0:500]*pi/500;
H = freqz(1,ahat2,omg);
Rx = abs(H).*abs(H);
Rx2 = 10*log10(Rx);
Statistical and Adaptive Signal Processing - Solution Manual 281

plot(omg/pi*Fs/2,Rx2,’m--’,’linewidth’,1);

% PZ(12,6) model spectra

H = freqz(bhat,ahat,omg);
Rx = abs(H).*abs(H);
Rx3 = 10*log10(Rx);
plot(omg/pi*Fs/2,Rx3,’b:’,’linewidth’,1);

Rxmin = min([min(Rx1),min(Rx2),min(Rx3)]);
Rxmax = max([max(Rx1),max(Rx2),max(Rx3)]);
axis([0,Fs/2,floor(Rxmin/10)*10,ceil(Rxmax/10)*10]); hold off;
legend(’Periodogram’,’AP(16)’,’PZ(12,6)’)
The plots are shown in Figure 9.18.

9.19 Spectral estimation using prewhitening and postcoloring method.

(a) The Matlab function psd_lpprew:

function [Rx] = PSD_lpprew(x,P,L,Nfft);

Nfft = max(Nfft,2*L);
Nfft = 2^(ceil(log10(Nfft)/log10(2)));
x = x-mean(x);
noverlap = L/2;

% Fit an AP(P) model using forward-backward linear prediction

[ahat,VarP] = armodcov(x,P);
a = [1;ahat];

% Compute residuals
e = filter(a,1,x);
e = e - mean(e); var_e = (std(e))^2;
N = length(e);

%% Determine Welch PSD estimate of the residuals

Re = psd(e,Nfft,2,hamming(L),L/2,’none’);

%% Perform postcoloring to obtain the PSD

w = [0:Nfft/2]*2*pi/Nfft;
A = freqz(1,a,w);
magAsq = (abs(A’)).^2;
Rx = Re(1:Nfft/2+1).*magAsq;

(b) Verification of the function psd_lpprew:

% Read the speech file
x = load(’testspee.dat’); Fs = 10000;
x = filter([1,-0.95],1,x);
N = length(x);
% Periodogram of the speech signal
282 Statistical and Adaptive Signal Processing - Solution Manual

Comparison of Spectra
40
Periodogram
LP Prewhiteing
30 PZ(12,6)

Rx(ejω) in dB 20

10

−10

−20

−30
0 1000 2000 3000 4000 5000
frequency in Hz
Figure 9.19: Comparison plots of various spectra in P9.19

win = boxcar(N); NfftP = 2^(ceil(log10(N)/log10(2)));

Rx1 = psd(x,NfftP,Fs,win,’none’);

% (b) Verification of the psd_lpprew function

P = 16; L = 64; Nfft = 1024;
Rx2 = psd_lpprew(x,P,L,Nfft);

(c) Comparison using plots:

% (c) Estimates using PZ(12,6) model
P = 12; Q = 6;
[ahat,bhat,varwhat,FPE,C] = armals(x,P,Q,100,0.001);

% (d) Plots
% Periodogram Spectra
freq = [0:NfftP/2]*Fs/NfftP;
Rx1 = 10*log10(Rx1)-60;
plot(freq,Rx1,’g’); title(’Comparison of Spectra’,’fontsize’,10);
xlabel(’frequency in Hz’,’fontsize’,8);
ylabel(’R_x(e^{j\omega}) in dB’,’fontsize’,8); hold on;
set(gca,’xtick’,[0:1000:Fs/2],’fontsize’,6)

% LP_PREW Spectra
freq = [0:Nfft/2]*Fs/Nfft;
Rx2 = 10*log10(Rx2)-60;
plot(freq,Rx2,’m--’,’linewidth’,1.5);

% PZ(12,6) model spectra

omg = [0:500]*pi/500;
H = freqz(bhat,ahat,omg);
Rx = abs(H).*abs(H);
Statistical and Adaptive Signal Processing - Solution Manual 283

Rx3 = 10*log10(Rx);
plot(omg/pi*Fs/2,Rx3,’b:’,’linewidth’,1.5);

Rxmin = min([min(Rx1),min(Rx2),min(Rx3)]);
Rxmax = max([max(Rx1),max(Rx2),max(Rx3)]);
axis([0,Fs/2,floor(Rxmin/10)*10,ceil(Rxmax/10)*10]); hold off;
legend(’Periodogram’,’LP Prewhiteing’,’PZ(12,6)’)
The plots are shown in Figure 9.19.

9.20 Consider a white noise process with variance σ 2w . The correlation matrix of this white noise process is given
by

Rx = σ 2w I .

The minimum variance distortionless response weight vector for the MVDR spectrum estimator is given by
√
M Rx−1 v ( f k )
ck =
v H ( f k ) Rx−1 v ( f k )
√
M v ( fk )
= independent of σ 2w
v ( fk ) v ( fk )
H

1
= √ v ( fk )
M

since v H ( f k ) v ( f k ) = M from (9.5.6). Therefore, the MVDR spectral estimate of the white noise process is
given by
 j 2π fk 
R̂ (mv)
M e = E |y (n)|2
" 2 #
= E ckH x (n)
1 H
= v ( f k ) σ 2w I v ( f k )
M
1
= σ 2w M
M
= σ 2w

9.21 The first-order all-pole model is given by

x (n) = −a1 x (n − 1) + w (n)

where w (n) is white Gaussian noise with variance σ 2w . From pg. 167, the autocorrelation function of the
AP(1) model is

σ 2w
r x (l) = (−a1 )|l|
1 − a1 2

Using the minimum-variance spectral estimate from (9.5.11)

 j 2π f M
?(mv)
R e =
M
vH ( f ) R−1
x v(f)
284 Statistical and Adaptive Signal Processing - Solution Manual

we see that we must solve for the inverse of the correlation matrix. The Mth order correlation matrix of the
AP(1) model is
 
1 −a1 a12 · · · (−a1 ) M−1
 .. 
 −a 1 −a . 
σ 2w  1 1 
. . 
Rx =  a1 2
−a1 1 . 
1 − a12  
 .. .. .. 
 . . . −a1 
(−a1 ) M−1 · · · · · · −a1 1

>From (9.5.16), the inverse of the autocorrelation matrix is

R−1 H −1
x = A D̄ A

where
 
1 −a1 0 ··· 0
 .. .. 
 0 1 −a1 . . 
 
 
A =  ... 0 1
..
. 0 
 
 .. .. 
 . . −a1 
0 ··· 0 1

and

D̄ = diag {PM , PM−1 , . . . , P1 }

& '
σ 2w
= diag σ 2w , σ 2w , . . . , ,
1 − a12

Therefore,
1 
D̄−1 = diag 1, 1, . . . , 1 − a12
σw2

and the inverse of the correlation matrix is given by

R−1 = A H D̄−1 A
x
 
1 −a1 0 ··· 0
 . .. .. 
 −a1 1 + a 2 −a1 . 
1   1 
. . . . 
=  −a1 . . 
σ 2w  0 0 
 .. .. .. 
 . . . 1 + a12 −a1 
0 ··· 0 −a1 1

>From (9.5.6), the time-window frequency vector is

 T
v ( f ) = 1 e− j 2π f · · · e− j 2π f (M−1)
Statistical and Adaptive Signal Processing - Solution Manual 285

then
 
1 −a1 0 ··· 0
 ..  ..
 −a1 1 + a 2 −a1 .  .
 1 
1 H  .. .. 
v H ( f ) R−1
x v(f) = v ( f )  −a1 . 0 
. v(f)
σ 2w  0 
 .. .. .. 
 . . . 1 + a12 −a1 
0 ··· 0 −a1 1
1   
= 2 + (M − 2) 1 + a12 − a1 (M − 1) e j 2π f − a1 (M − 1) e− j 2π f
σw 2

1   
= 2 + (M − 2) 1 + a12 − 2a1 (M − 1) cos 2π f
σw 2

Substituting back in for the minimum-variance spectrum, we have

 j 2π f Mσ 2w
?(mv)
R M e = 
2 + (M − 2) 1 + a12 − 2a1 (M − 1) cos 2π f
Note that the actual spectrum of an AR(1) process is
 1
R e j 2π f =
1+ a12 − 2a1 cos 2π f
 j 2π f 
?(mv)
so as M → ∞ then R M e → R e j 2π f .

9.22 >From (9.5.10), the minimum-variance weight vector is given by

√
MR−1 x v ( fk )
ck = H
v ( f k ) R−1
x v ( fk )

Solving the constrained optimization problem

√
min ckH Rx ck subject to ckH v ( f k ) = M

we use Lagrange multipliers from Appendix B.2. First, we define the two functions
√
f (ck ) = ckH Rx ck g (ck ) = ckH v ( f k ) − M

The Lagrangian function is then

L (ck , λ) = f (ck ) + λg (ck )

$ √ %
= ckH Rx ck + λ ckH v ( f k ) − M

where λ is the Lagrange multiplier. We then want to minimize the Lagrangian function with respect to both ck
and λ. Taking the gradient of L (ck , λ) with respect to ck and setting it to zero, we have

∇ ck L (ck , λ) = ckH Rx + λv H ( f k ) = 0

ck = −λR−1
x v ( fk ) (1)

Similarly, taking the gradient of L (ck , λ) with respect to λ and setting it to zero
√
∇ λ L (ck , λ) = ckH v ( f k ) − M = 0 (2)
286 Statistical and Adaptive Signal Processing - Solution Manual

Substituting for (1) into (2), yields

√
−λ∗ v H ( f k ) R−1
x v ( fk ) − M =0

and solving for λ

√
− M
λ= H (3)
v ( f k ) R−1
x v ( fk )

Substituting (3) back into (1), we have the weight vector for the minimum-variance spectral estimator at the
frequency f k
√
MR−1 x v ( fk )
ck = H
v ( f k ) R−1
x v ( fk )

9.23 Recall the relation between the minimum-variance and all-pole model spectrum estimators from (9.5.22)

M  H
2
1  vm ( f ) am  1 
M
1
 = = 
?M e
R (mv) j 2π f M Pm M m=1 R (ap)
?m e j 2π f
m=1

Thus, the minimum-variance spectral estimate with a filter of length M is equal to an average of all-pole model
spectral estimates from order m = 1 to m = M. Using this relationship, the minimum-variance spectral
estimate with window length M + 1 is given by

1 
M+1
1 1
 = 
?M+1 e j 2π f
R (mv) (ap)
?m e j 2π f
M + 1 m=1 R

1 
M
1 1 1
= · (ap)  + (ap) 
?M+1 e j 2π f
M +1 R ?m e j 2π f
M + 1 m=1 R

1 
M
1 1 M 1
= · (ap)  + · 
?M+1 e j 2π f
M +1 R M + 1 M m=1 R ?m(ap) e j 2π f
1 1 M 1
= · (ap)  + · (mv) 
?M+1 e
M +1 R j 2π f M +1 R ?M e j 2π f

Therefore, the (M + 1)th order minimum-variance spectrum is equal to the Mth order minimum-variance
spectrum refined by the (M + 1)th order all-pole spectral estimate weighted by the M+1
M
and M1 , respectively.
9.24 For the Pisarenko harmonic decomposition, the pseudo-spectrum is given by
 1 1
R̄ phd e j 2π f =  2 =   
v H ( f ) q M   Q M e j 2π f 2

where q M is the eigenvector of the correlation matrix Rx associated with the smallest eigenvalue, i.e., the noise
eigenvector. Recall that the order is chosen to be M = P + 1 where P is the number of complex exponentials.
The frequency estimates are then the P = M − 1 peaks in the pseudo-spectrum fˆ1 , fˆ2 , ... , fˆP . Using the
eigenvector/eigenvalue relation
R x q m = λm q m for m = 1, 2, . . . , M

where recall that the eigenvectors have unit-norm (qmH qm = 1). Multiplying both sides of the eigenvec-
tor/eigenvalue relation by qmH

qmH Rx qm = λm qmH qm = λm (1)

Statistical and Adaptive Signal Processing - Solution Manual 287

Recall from the harmonic model, the correlation matrix from (9.6.7) is given by

P
 
Rx = A p v f p v H f p + σ 2w I (2)
p=1
 2
where A p = α p  is the power of the pth complex exponential. In the case of the Pisarenko harmonic
decomposition, we have chosen M = P+1 using our a priori knowledge of the number of complex exponentials.
In addition, the smallest eigenvalue corresponding to the noise is λ M = σ 2w . Substituting (2) into (1), we have

P
  H
A p qmH v f p v f p qm + σ 2w = λm
p=1


P
  2  P
  2
A p qmH v f p  = A p  Q m e j 2π f p  = λm − σ 2w
p=1 p=1

where Q m e j 2π f p is the Fourier transform of the eigenvector qm evaluated at the frequency f p . This same
equation can be written for all of the frequency estimates fˆ1 , fˆ2 , . . . , fˆP from the Pisarenko harmonic decom-
position obtained by picking the peaks in the pseudo-spectrum. Writing this set of equations in matrix form
we have
   2   2   2 
 j 2π fˆ1   j 2π fˆ2   j 2π fˆP     
  Q 1 e  Q1 e  · · · Q1 e 
  2   2   2  A1 λ1 − λ M
  Q e j 2π fˆ1   ˆ   ˆ 
 Q 2 e j 2π f2  · · ·  Q 2 e j 2π f P      
 2   A 2   λ2 − λ M 
   .  = . 
 .. .. ..   ..   .. 
 . . . 
   2   2   2  AP λ P − λM
 ˆ   ˆ   ˆ 
Q P e j 2π f1  Q P e j 2π f2  · · · Q 1 e j 2π f P 

This equation can then be solved for the powers of the complex exponentials A1 , A2 , . . . , A P .
9.25 The minimum-norm pseudo-spectrum from (9.6.41) is given by
 1
R̄mn e j 2π f =  
v H ( f ) umn 2
where umn is the minimum-norm vector from (9.6.40)
Pw δ 1
umn =
δ 1 Pw δ 1
H

where Pw = Qw QwH is the noise subspace projection matrix, Qw is the matrix of noise eigenvectors, and
δ 1 = [1 0 · · · 0]T . This guarantees that umn is in the noise subspace. Since the complex exponentials are all
in the signal subspace which is orthogonal to the noise subspace, they are thus orthogonal to umn . Consider
the denominator of the minimum-norm pseudo-spectrum
 M 2
 2  
v ( f ) umn  =   − (k−1) 
H
umn (k) · e j 2π f

 
k=1
  2
= Umn e j 2π f 

where Umn e j 2π f is the Fourier transform of the minimum-norm vector umn . Converting to the z-transform
(z = e− j 2π f )

M−1
Umn (z) = umn (k + 1) z −k
k=0
288 Statistical and Adaptive Signal Processing - Solution Manual

The z-transform of the denominator of the pseudo-spectrum can then be written as the following polynomial


∗ 1
P̄mn (z) = |Umn (z)| = Umn (z) Umn
2
z∗
This (2M −1)th order polynomial has (M −1) pairs of roots with one inside and one outside the unit circle. The
peaks in the pseudo-spectrum correspond to the roots of this polynomial. Since we assume that the complex
exponentials are undamped, their roots should theoretically lie on the unit circle. Therefore, looking at the
M − 1 roots of this polynomial that lie inside the unit circle, we associate the P closest ones to the unit
circle with the P complex exponentials. The phases of these roots are then the root minimum-norm frequency
estimates. A Matlab example of the use of the root minimum-norm method is given below:

% Routine to show example of root minimum-norm frequency

% estimates
clear

% Number of sinusoids
P = 2; % number of complex exponentials
M = 8; % number of samples in time-window
Nsamples = 128; % number of time samples to generate

% Define signal parameters

f_s = [0.1 0.2]’; % frequencies of complex exponentials
SNR_db = [5]*ones(P,1); % signal-to-noise ratios (dB)

% Generate signal of complex exponentials in noise

x = zeros(Nsamples,1);
phase = rand(P) - 0.5;
for p = 1:P
alpha(p) = (10^(SNR_db(p)/20))*exp(j*2*pi*phase(p));
x = x + alpha(p)*exp(j*2*pi*f_s(p)*[0:(Nsamples-1)]’);
end

% Add unit variance noise

x = x + (randn(Nsamples,1) + i*randn(Nsamples,1))/sqrt(2);

% estimate correlation matrix

% Generate data matrix
N = length(x) - M + 1;
X = zeros(N,M);
for n = 1:M
X(:,n) = x((1:N)+ (M-n));
end
R = (1/N)*X’*X;

% Compute eigendecomposition and order by descending eigenvalues

[Q0,D] = eig(R);
[lambda,index] = sort(abs(diag(D)));
lambda =lambda(M:-1:1);
Statistical and Adaptive Signal Processing - Solution Manual 289

Q=Q0(:,index(M:-1:1));

% Compute noise subspace projection matrix

% and minimum-norm vector
delta1 = zeros(M,1);
delta1(1) = 1;
Pw = Q(:,(P+1):M)*Q(:,(P+1):M)’; % noise subspace projection matrix
u_mn = (Pw*delta1)/(delta1’*Pw*delta1); % minimum-norm vector

% Compute minimum-norm polynomial and root it

Pbar = conv(u_mn,conj(u_mn(M:-1:1)));
r_Pbar = roots(Pbar); % roots of polynomial

% Find the roots closest to the unit circle

[i_min] = find(abs(r_Pbar) < 1);
r_Pbar_min = r_Pbar(i_min); % roots closest to unit circle
freq = angle(r_Pbar_min)/(2*pi);
[r_Vn_order,index] = sort(abs((abs(r_Pbar_min) -1)));
fest = freq(index(1:P)); % frequency estimates

9.26 The minimum-variance spectral estimate from (9.5.11) is given by

 j 2π f M
R̂ (mv) e = (1)
M
v H ( f ) R−1
x v(f)

The correlation matrix has the following eigendecomposition


M
Rx = λm qm qmH (2)
m=1

and the inverse of the correlation matrix is given by

M
1
R−1
x = qm qmH (3)
λ
m=1 m

Substituting (3) into (1) we have

 j 2π f M
R̂ (mv) e =
M
!
M
1 H
λm
v ( f ) qm qmH v ( f )
m=1
M
=
!
M  
1 v H ( f ) qm 2
λm
m=1
M
=
!
M   
1  Q m e j 2π f 2
λm
m=1
290 Statistical and Adaptive Signal Processing - Solution Manual


where Q m e j 2π f is the Fourier transform of the mth eigenvector qm . Recall from (9.6.30) and (9.6.31) the
pseudo-spectra of the MUSIC and eigenvector methods, respectively
 1
R̄music e j 2π f =
!
M   
 Q m e j 2π f 2
m=P+1

 1
R̄ev e j 2π f = 
!
M
1  2
 λm Q m e j 2π f 
m=P+1

If we assume that the estimated noise eigenvalues are all equal to λm = σ 2w for m = P + 1, . . . M, then
 1 
R̄music e j 2π f = 4 R̄ev e j 2π f (4)
σw

The minimum-variance spectral estimate is related to the eigenvector method pseudo-spectrum as

 j 2π f M
R̂ (mv) e = P 
M
 ! 1  2
R̄ev e j 2π f +  λm m
Q e j 2π f 
m=1

The relation for the MUSIC pseudo-spectrum follows from (4).

9.27 Recall the pseudo-spectrum from the MUSIC frequency estimation method from (9.6.30)
 1 1
R̄music e j 2π f = =
!
M   !
M   
v H ( f ) qm 2  Q m e j 2π f 2
m=P+1 m=P+1

and the minimum-variance spectrum from (9.5.11)

 j 2π f M
R̂ (mv) e = (1)
M
vH ( f ) R−1
x v(f)

The MUSIC pseudo-spectrum can be rewritten as

 1
R̄music e j 2π f = (2)
vH ( f ) Pw v ( f )
where the noise subspace projection matrix Pw is given by

Pw = Qw QwH = I − Qs QsH (3)

where Qw and Qs are matrices whose columns are the noise and signal eigenvectors of the correlation matrix,
respectively. From (9.6.7), the correlation matrix of complex exponentials in noise is

Rx = VAV H + σ 2w I

where V is the matrix of frequency vectors of the P complex exponentials

V = [v ( f 1 ) v ( f 2 ) · · · v ( f P )]
Statistical and Adaptive Signal Processing - Solution Manual 291

and A is the diagonal matrix of their respective powers


A = diag |α 1 |2 , |α 2 |2 , . . . , |α P |2 (4)

Using the matrix inversion lemma from Appendix A, the inverse of the correlation matrix of the harmonic
model from (4) is
1 $  2 −1 −1 H %
R−1
x = I − V σ w A + V H
V V (5)
σ 2w

For the case of infinite signal-to-noise ratio, the matrix A−1 = 0. Therefore, the inverse of the correlation
matrix in (5) becomes
1 $  H −1 H %
R−1 = I − V V V V
x
σ 2w
1 
= I − Qs QsH
σw 2

1
= Qw QwH
σ 2w
1
= Pw
σ 2w

making use of the projection matrix relation from (9.6.16) for the subspace spanned by the signal frequency
vectors in V. Therefore, for the case of infinite signal-to-noise ratio, the inverse of the correlation matrix
becomes the noise subspace projection matrix from (3) normalized by the noise power σ 2w . Using this result
in the equations for the minimum-variance spectrum in (1)
 j 2π f M
R̂ (mv) e =
M
( f ) R−1
vH x v(f)
2
Mσ w
=
v H ( f ) Pw v ( f )

= Mσ 2w R̄music e j 2π f (6)

which is the MUSIC pseudo-spectrum from (2) weighted by Mσ 2w . Thus, the MUSIC pseudo-spectrum can
be viewed as a minimum-variance spectrum with infinite signal-to-noise ratio. Since the resolution of the
complex exponentials is related to their signal-to-noise ratios, the MUSIC pseudo-spectrum exhibits superior
resolution capability over the minimum-variance spectrum and has led to the MUSIC frequency estimation
method commonly being referred to as a superresolution technique.

9.28 If it has not been done so in your edition, this problem should be restated as:
Find the relationship between the minimum-norm pseudo-spectrum an the all-pole model spectrum in the case
of infinite signal-to-noise ratio. What are the implications of this relationship?
First, let us examine the Mth order all-pole model spectrum

(ap)  j 2π f PM
R̂ M e = H  (1)
v ( f ) a M 

where
$ %
a M = 1 a1(M) a2(M) · · · a (M)
M−1
292 Statistical and Adaptive Signal Processing - Solution Manual

is the vector of the Mth order all-pole model coefficients which is found by solving the following set of linear
equations
 
PM
 0 
 
Rx a M =  . 
 .. 
0

Therefore, we can also write a M as

a M = PM R−1
x δ1

where δ 1 = [1 0 · · · 0]T , i.e., the column vector with a value of unity in its first element and otherwise all
zeros. Since

a M (1) = δ 1H a M = δ 1H R−1
x δ 1 PM = 1

then
1
PM = (2)
δ 1H R−1
x δ1
and
R−1
x δ1
aM = (3)
δ 1H R−1
x δ1

Substituting (2) and (3) into the all-pole model spectrum in (1), we have

(ap)  j 2π f PM
R̂ M e =  
v H ( f ) a M 
δ 1H R−1
x δ1
=   (4)
v H ( f ) R−1 δ 1 2
x

>From (9.6.7), the correlation matrix of P complex exponentials in white noise is

Rx = VAV H + σ 2w I (5)

where V is the matrix of P frequency time-window vectors

V = [v ( f 1 ) v ( f 2 ) · · · v ( f P )]

and A is the matrix of signal powers


A = diag |α 1 |2 , |α 2 |2 , · · · , |α P |2

The correlation matrix can also be written in terms of the signal and noise eigenvector and eigenvalues from
(9.6.14)

Rx = Qs s QsH + σ 2w Qw QwH

where the matrices Qs and Qw are made up of the signal and noise eigenvectors respectively
   
Qs = q 1 q 2 · · · q P Qw = q P+1 q P+2 · · · q M
Statistical and Adaptive Signal Processing - Solution Manual 293

and

s = diag {λ1 , . . . , λ P }

Also recall the signal and noise subspace projection matrices from (9.6.17)

Ps = Qs QsH Pw = Qw QwH

where from (9.6.16)

 −1 H
Ps = Qs QsH = V V H V V

Note also that the noise subspace projection matrix is given by

 −1 H
Pw = I − Ps = I − V V H V V

Using the matrix inversion lemma from Appendix A on the correlation matrix from (5)
1 $  2 −1 −1 H %
R−1 = I − V σ w A + V H
V V (6)
x
σ 2w
In the case of infinite signal-to-noise ratio, A−1 = 0 and (6) becomes
1 $  H −1 H % 1
R−1 = I − V V V V = 2 Pw (7)
x
σw2 σw
Substituting (7) into the all-pole model spectrum from (4)

(ap) 
1 H
δ P
σ 2w 1 w
δ1
R̂ M e j 2π f =  
1 v H ( f ) Pw δ 1 2
σ 4w
σ 2w Pw (1, 1)
=   (8)
v H ( f ) Pw δ 1 2

The minimum-norm pseudo-spectrum from (9.6.41) is given by

 1
R̄mn e j 2π f =   (9)
v H ( f ) umn 2

where the minimum-norm vector is given by

Pw δ 1
umn = (10)
δ 1 Pw δ 1
H

Substituting (10) into (9), the minimum-norm pseudo-spectrum is

 H 2
 j 2π f δ 1 Pw δ 1 Pw (1, 1)2
R̄mn e = 2 =   (11)
v H ( f ) Pw δ 1  v H ( f ) Pw δ 1 2

Comparing (8) and (11), we see that

 Pw (1, 1) (ap)  j 2π f
R̄mn e j 2π f = R̂ M e
σ 2w
Therefore, the minimum-norm pseudo-spectrum becomes a multiple of the all-pole model spectrum with infinite
signal-to-noise ratio for all of the complex exponentials. This relation explains the superior resolution that the
minimum-norm method can achieve over the all-pole model spectrum.
294 Statistical and Adaptive Signal Processing - Solution Manual

9.29 Recall both the MUSIC and minimum-norm pseudo-spectra from (9.6.30) and (9.6.41)
 1
R̄music e j 2π f =
!
M  
v H ( f ) qm 2
m=P+1

 1
R̄mn e j 2π f =  
v H ( f ) umn 2

where the minimum-norm vector is

Pw δ 1
umn =
δ 1 Pw δ 1
H

Substituting for the minimum-norm vector into the minimum-norm pseudo-spectrum

 Pw (1, 1)2
R̄mn e j 2π f =  
v H ( f ) Pw δ 1 2

Looking at the reciprocal of the minimum-norm pseudo-spectra

1  2
 = v H ( f ) Pw δ 1 
R̄mn e j 2π f

M
 H 
= v ( f ) qm q H δ 1 2
m
m=P+1


M
 
= qm (1) v H ( f ) qm 2 (1)
m=P+1

where qm (1) is simply the first element of the mth eigenvector qm . Therefore, the minimum-norm pseudo-
spectrum has the same form as the MUSIC pseudo-spectrum where the contribution from each eigenvector
simply has an alternate weighting, namely the first element of the corresponding eigenvector.
 Chapter 10

Adaptive Filters

10.1 Consider the process x(n) generated using the AR(3) model

x(n) = −0.729x(n − 3) + w(n)

where w(n) ∼ WGN(0, 1). We want to design a linear predictor of x(n) using the SDA algorithm. Let

ŷ(n) = x̂(n) = co,1 x(n − 1) + co,2 x(n − 2) + co,3 x(n − 3)

(a) Determine the 3 × 3 autocorrelation matrix R of x(n), and compute its eigenvalues {λi }3i=1 .
The AR(3) model system function is given by
1
H (z) =
1 + 0.729 z −3
Hence the autocorrelation sequence of the process x(n) is



−1
 −1)
 −1 1 1
r x (l) = Z H (z)H (z =Z
1 + 0.729 z −3 1 + 0.729 z 3
  
1 1
The partial fraction expansion of 1+0.729 z −3 1+0.729 z 3
is given by



1 1 1
=
1 + 0.729 z −3 1 + 0.729 z 3 0.729 z −3 + 1.5314 + 0.729 z 3
0.7114 0.7114
= −
1 + 0.9 z −1
1 + 10
9
z −1
1 − 0.45z −1
+1.4228
1 − 0.9z −1 + 0.81z −2
1 − (5/9) z −1
−1.4228  2
1 − 10
9
z −1 + 10
9
z −2

Thus after inverse z-transformation


πl
r x (l) = 0.7114 (−0.9)|l| + 1.4228 (0.9)|l| cos ((1))
3

Using (1), the 3 × 3 autocorrelation matrix R is given by

   
r x (0) r x (1) r x (2) 2.1342 0 0
R =  r x (1) r x (0) r x (1)  =  0 2.1342 0 
r x (2) r x (1) r x (0) 0 0 2.1342

The eigenvalues of R are given by

λ1 = 2.1342, λ2 = 2.1342, and λ3 = 2.1342

295
296 Statistical and Adaptive Signal Processing - Solution Manual

(b) Determine the 3 × 1 cross-correlation vector d.

For an AR(3) process, the cross-correlation vector d is given by
 T  T
d = r x (1) r x (2) r x (3) = 1.2808 0.0000405 −0.51868

Clearly
 −1    
2.1342 0 0 0 0
 0 2.1342 0   0 = 0 
0 0 2.1342 −1.5558 0.729

as expected.
(c) Choose the step size µ so that the resulting response is overdamped. Now implement the SDA

ck = [ck,1 ck,2 ck,3 ]T = ck−1 + 2µ(d − Rck−1 )

and plot the trajectories of {ck.i }3i=1 as a function of k.

Since λmax = λmin = λ = 2.1342, χ (R) = 1 and the step-size µ < λmax
1
= 0.46856. We choose µ = 0.1
for the overdamped case. The Matlab script is shown below and the plot is shown in Figure 10.1c.
% AR(3) model parameters
bx = 1; ax = [1,0,0,0.729];

% (a) 3x3 Autocorrelation matrix of x(n) and its eigenvalues

brx = [0,0,0,1]; % Num of Rx(z)
arx = conv(ax,fliplr(ax)); % Den of Rx(z)
[Res,pole,K] = residuez(brx,arx); % PFE of Rx(z)

% Computation of correlation values

l = 0:3;
rx = Res(4)*pole(4).^l + Res(5)*pole(5).^l + Res(6)*pole(6).^l;
rx = real(rx);
R = toeplitz(rx(1:3));

% Computation of eigenvalues
lambda = eig(R); lambda = lambda’;

% (b) 3x1 Crosscorrelation vector d

d = rx(2:4)’;

% (c) Implementation of SDA: Overdamped case

mu = 0.1; % Step-size
N = 30; % Number of iterations
c = zeros(3,1,N);
c(:,1,1) = 2*mu*d; % coefficient vector at k = 1
for k = 2:N
c(:,1,k) = c(:,1,k-1) + 2*mu*(d-R*c(:,1,k-1));
end
k = 0:N;
c1 = squeeze(c(1,1,:)); c1 = [0;c1];
Statistical and Adaptive Signal Processing - Solution Manual 297

Trajectories of Coefficient Vector: Overdamped case

1

ck(1)
0

−1
0 5 10 15 20 25 30

ck(2)
0

−1
0 5 10 15 20 25 30
1
ck(3)

−0.729
−1
0 5 10 15 20 25 30
Iteration index k

Figure 10.1c: Plots of trajectories for the overdamped case

c2 = squeeze(c(2,1,:)); c2 = [0;c2];
c3 = squeeze(c(3,1,:)); c3 = [0;c3];

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
subplot(3,1,1); stem(k,c1,’g’,’filled’); axis([-1,N+1,-1,1]);
set(gca,’ytick’,[-1:1:1]);
ylabel(’c_k(1)’,’fontsize’,label_fontsize);
title(’Trajectories of Coefficient Vector: Overdamped case’,...
’fontsize’,title_fontsize);
subplot(3,1,2); stem(k,c2,’g’,’filled’); axis([-1,N+1,-1,1]);
set(gca,’ytick’,[-1:1:1]);
ylabel(’c_k(2)’,’fontsize’,label_fontsize);
subplot(3,1,3); stem(k,c3,’g’,’filled’); axis([-1,N+1,-1,1]);
hold on; plot([-1,N+1],[0,0],’w’);
set(gca,’ytick’,[-1,-0.729,0,1]);
xlabel(’Iteration index k’,’fontsize’,label_fontsize);
ylabel(’c_k(3)’,’fontsize’,label_fontsize);

(d) Repeat part (c) by choosing µ so that the response is underdamped.

We choose µ = 0.4 for the underdamped case. The Matlab script is shown below and the plot is shown
in Figure 10.1d.
% (d) Implementation of SDA: Overdamped case
mu = 0.4; % Step-size
N = 30; % Number of iterations
c = zeros(3,1,N);
c(:,1,1) = 2*mu*d; % coefficient vector at k = 1
for k = 2:N
c(:,1,k) = c(:,1,k-1) + 2*mu*(d-R*c(:,1,k-1));
end
298 Statistical and Adaptive Signal Processing - Solution Manual

Trajectories of Coefficient Vector: Underdamped case

1

ck(1)
0

−1
0 5 10 15 20 25 30

1
ck(2)

−1
0 5 10 15 20 25 30
2
ck(3)

0
−0.729

−2
0 5 10 15 20 25 30
Iteration index k

Figure 10.1c: Plots of trajectories for the underdamped case

k = 0:N;
c1 = squeeze(c(1,1,:)); c1 = [0;c1];
c2 = squeeze(c(2,1,:)); c2 = [0;c2];
c3 = squeeze(c(3,1,:)); c3 = [0;c3];

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
subplot(3,1,1); stem(k,c1,’g’,’filled’); axis([-1,N+1,-1,1]);
set(gca,’ytick’,[-1:1:1]);
ylabel(’c_k(1)’,’fontsize’,label_fontsize);
title(’Trajectories of Coefficient Vector: Underdamped case’,...
’fontsize’,title_fontsize);
subplot(3,1,2); stem(k,c2,’g’,’filled’); axis([-1,N+1,-1,1]);
set(gca,’ytick’,[-1:1:1]);
ylabel(’c_k(2)’,’fontsize’,label_fontsize);
subplot(3,1,3); stem(k,c3,’g’,’filled’); axis([-1,N+1,-2,2]);
hold on; plot([-1,N+1],[0,0],’w’);
set(gca,’ytick’,[-2,-0.729,0,2]);
xlabel(’Iteration index k’,’fontsize’,label_fontsize);
ylabel(’c_k(3)’,’fontsize’,label_fontsize);

10.2 In the SDA algorithm, the index k is an iteration index and not a time index. However, we can treat it as a
time index and use the instantaneous filter coefficient vector ck to filter data at n = k. This will result in an
asymptotically optimum filter whose coefficients will converge to the optimum one. Consider the process x(n)
given in Problem 10.1.

(a) Generate 500 samples of x(n) and implement the asymptotically optimum filter. Plot the signal ŷ(n).
The design and implementation is shown in the following Matlab script and the plot is shown in Fig-
ure 10.2a.
% Generate 500 samples of AR(3) process
Statistical and Adaptive Signal Processing - Solution Manual 299

Output of the Asymptotically Optimum Filter

5

yhat(n)
0

−5
0 50 100 150 200 250 300 350 400 450 500

AR(3) Process samples

5
x(n)

−5
0 50 100 150 200 250 300 350 400 450 500
sample index n

Figure 10.2a: Asymptotically optimum filter response

bx = 1; ax = [1,0,0,0.729]; var_w = 1; % Model parameters

N = 500; w = randn(N,1)*sqrt(var_w); % white noise process
x = filter(bx,ax,w); % AR(3) process

% (a) Design and processing using AO filter

% Computation of correlation values
brx = [0,0,0,1]; % Num of Rx(z)
arx = conv(ax,fliplr(ax)); % Den of Rx(z)
[Res,pole,K] = residuez(brx,arx); % PFE of Rx(z)
l = 0:3;
rx = Res(4)*pole(4).^l + Res(5)*pole(5).^l + Res(6)*pole(6).^l;
R = toeplitz(rx(1:3)); % 3x3 Autocorrelation matrix
d = rx(2:4)’; % 3x1 Crosscorrelation vector d
% Design of the AO filter sequence
mu = 0.1; % Step-size
c = zeros(3,1,N);
c(:,1,1) = 2*mu*d; % coefficient vector at k = 1
for k = 2:N
c(:,1,k) = c(:,1,k-1) + 2*mu*(d-R*c(:,1,k-1));
end
n = 1:N;
c1 = squeeze(c(1,1,:));
c2 = squeeze(c(2,1,:));
c3 = squeeze(c(3,1,:));
% Processing using the AO filter sequence
yhat = zeros(N,1);
for k = 4:N
yhat(k) = c1(k)*x(k-1) + c2(k)*x(k-2) + c3(k)*x(k-3);
end

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
300 Statistical and Adaptive Signal Processing - Solution Manual

Output of the Optimum Filter

5

yhat(n)
0

−5
0 50 100 150 200 250 300 350 400 450 500

AR(3) Process samples

5
x(n)

−5
0 50 100 150 200 250 300 350 400 450 500
sample index n

Figure 10.2b: Optimum filter response

’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
subplot(2,1,1);
plot(n,yhat,’g’); axis([0,N+1,-5,5]);
set(gca,’ytick’,[-5:5:5]);
ylabel(’yhat(n)’,’fontsize’,label_fontsize);
title(’Output of the Asymptotically Optimum Filter’,...
’fontsize’,title_fontsize);
subplot(2,1,2);
plot(n,x,’g’); axis([0,N+1,-5,5]);
set(gca,’ytick’,[-5:5:5]);
ylabel(’x(n)’,’fontsize’,label_fontsize);
xlabel(’sample index n’,’fontsize’,label_fontsize);
title(’AR(3) Process samples’,’fontsize’,title_fontsize);

(b) Implement the optimum filter co on the same sequence, and plot the resulting ŷ(n).
The design and implementation is shown in the following Matlab script and the plot is shown in Fig-
ure 10.2b.
% (b) Design and processing using Optimum filter
c_o = R\d;
yhat = filter([0;c_o],1,x);

Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,3]);
subplot(2,1,1);
plot(n,yhat,’g’); axis([0,N+1,-5,5]);
set(gca,’ytick’,[-5:5:5]);
ylabel(’yhat(n)’,’fontsize’,label_fontsize);
title(’Output of the Optimum Filter’,...
’fontsize’,title_fontsize);
subplot(2,1,2);
Statistical and Adaptive Signal Processing - Solution Manual 301

plot(n,x,’g’); axis([0,N+1,-5,5]);
set(gca,’ytick’,[-5:5:5]);
ylabel(’x(n)’,’fontsize’,label_fontsize);
xlabel(’sample index n’,’fontsize’,label_fontsize);
title(’AR(3) Process samples’,’fontsize’,title_fontsize);
(c) Comment on the above two plots.
The plots are identical after the initial transients.

10.3 Consider the AR(2) process x(n) given in Example 10.3.1. We want to implement the Newton-type algorithm
for faster convergence using

ck = ck−1 − µR−1 ∇ P(ck−1 )

(a) Using a1 = −1.5955 and a2 = 0.95, implement the above method for µ = 0.1 and c0 = 0. Plot the
locus of ck,1 versus ck,2 .
From (10.3.29) and (10.3.35) and choosing σ 2w so that σ 2x = 1, we have

2 
−a1 a1
r(0) = 1, r(1) = r(0) = 0.81821, and r(2) = − a2 r(0) = 0.35545
1 + a2 1 + a2
Hence
   
1 0.81821 0.81821
R= , d=
0.81821 1 0.35545

The Newton-type algorithm is implemented in the Matlab script below and the plot is shown in Figure 10.3.
% Initialization
c0 = zeros(2,1);

% (a) a1 = -1.5955, a2 = 0.95, and step-size mu = 0.1

a1 = -1.5955; a2 = 0.95; mu = 0.1;
r0 = 1; % var_w selected so that var_x = 1
r1 = -(a1/(1+a2))*r0;
r2 = (-a2+a1^2/(1+a2))*r0;
R = toeplitz([r0,r1]); Rinv = R\eye(2);
d = [r1;r2];
c = zeros(2,100);
delP = -d;
c(:,1) = -mu*Rinv*delP;
for k = 2:100
delP = R*c(:,k-1) - d;
c(:,k) = c(:,k-1) - mu*Rinv*delP;
end
c = [c0,c];

subplot(2,2,1);
plot(c(1,:),c(2,:),’md’,c(1,:),c(2,:),’g-’); axis([0,-2*a1,-2*a2,0]);
%xlabel(’c_{k,1}’,’fontsize’,label_fontsize);
ylabel(’c_{k,2}’,’fontsize’,label_fontsize);
302 Statistical and Adaptive Signal Processing - Solution Manual

title(’Trajectory of c_{k,1} vs c_{k,2}, \mu = 0.1’,...

’fontsize’,title_fontsize);
set(gca,’xtick’,[0,-a1],’ytick’,[-a2,0]);grid;
text(-a1+0.05,-a2+0.05,’c_o’);
text(-a1+0.3,-a2+0.7,’a_1 = -1.5955’);
text(-a1+0.3,-a2+0.5,’a_2 = 0.95’);
(b) Repeat part (a), using a1 = −0.195 and a2 = 0.95.
In this case


−a1 a12
r(0) = 1, r(1) = r(0) = 0.1, and r(2) = − a2 r(0) = −0.9305
1 + a2 1 + a2
and
   
1 0.1 0.1
R= , d=
0.1 1 −0.9305

The Newton-type algorithm is implemented in the Matlab script below and the plot is shown in Figure 10.3.
% (b) a1 = -0.195, a2 = 0.95, and step-size mu = 0.1
a1 = -0.195; a2 = 0.95; mu = 0.1;
r0 = 1; % var_w selected so that var_x = 1
r1 = -(a1/(1+a2))*r0-eps,
r2 = (-a2+a1^2/(1+a2))*r0-eps,
R = toeplitz([r0,r1]); Rinv = R\eye(2);
d = [r1;r2];
c = zeros(2,100);
delP = -d;
c(:,1) = -mu*Rinv*delP;
for k = 2:100
delP = R*c(:,k-1) - d;
c(:,k) = c(:,k-1) - mu*Rinv*delP;
end
c = [c0,c];

subplot(2,2,2);
plot(c(1,:),c(2,:),’md’,c(1,:),c(2,:),’g-’); axis([0,-2*a1,-2*a2,0]);
%xlabel(’c_{k,1}’,’fontsize’,label_fontsize);
ylabel(’c_{k,2}’,’fontsize’,label_fontsize);
title(’Trajectory of c_{k,1} vs c_{k,2}, \mu = 0.1’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,-a1],’ytick’,[-a2,0]);grid;
text(-a1+0.005,-a2+0.05,’c_o’);
text(-a1+.03,-a2+0.7,’a_1 = -0.195’);
text(-a1+.03,-a2+0.5,’a_2 = 0.95’);
(c) Repeat parts (a) and (b), using the optimum step size for µ that results in the fastest convergence.
The optimum step-size is µ = 1. The Matlab script is shown below and the plot is shown in Figure 10.3.
% (c1) a1 = -1.5955, a2 = 0.95, and step-size mu = 1
a1 = -1.5955; a2 = 0.95; mu = 1;
Statistical and Adaptive Signal Processing - Solution Manual 303

Trajectory of ck,1 vs ck,2, µ = 0.1 Trajectory of ck,1 vs ck,2, µ = 0.1

0 0

a1 = −1.5955 a1 = −0.195
a = 0.95 a = 0.95
2 2

co co
ck,2

ck,2
−0.95 −0.95

0 1.5955 0 0.195

Trajectory of ck,1 vs ck,2, µ = 1.0 Trajectory of ck,1 vs ck,2, µ = 1.0

0 0

a1 = −1.5955 a1 = −0.195
a = 0.95 a = 0.95
2 2

co co
ck,2

ck,2
−0.95 −0.95

0 1.5955 0 0.195
ck,1 ck,1

Figure 10.3: Plots of trajectories in Problem 10.3

r0 = 1; % var_w selected so that var_x = 1

r1 = -(a1/(1+a2))*r0;
r2 = (-a2+a1^2/(1+a2))*r0;
R = toeplitz([r0,r1]); Rinv = R\eye(2);
d = [r1;r2];
c = zeros(2,100);
delP = -d;
c(:,1) = -mu*Rinv*delP;
for k = 2:100
delP = R*c(:,k-1) - d;
c(:,k) = c(:,k-1) - mu*Rinv*delP;
end
c = [c0,c];

subplot(2,2,3);
plot(c(1,:),c(2,:),’md’,c(1,:),c(2,:),’g-’); axis([0,-2*a1,-2*a2,0]);
xlabel(’c_{k,1}’,’fontsize’,label_fontsize);
ylabel(’c_{k,2}’,’fontsize’,label_fontsize);
title(’Trajectory of c_{k,1} vs c_{k,2}, \mu = 1.0’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,-a1],’ytick’,[-a2,0]);grid;
text(-a1+0.05,-a2+0.05,’c_o’);
text(-a1+0.3,-a2+0.7,’a_1 = -1.5955’);
text(-a1+0.3,-a2+0.5,’a_2 = 0.95’);
304 Statistical and Adaptive Signal Processing - Solution Manual

% (c2) a1 = -0.195, a2 = 0.95, and step-size mu = 1

a1 = -0.195; a2 = 0.95; mu = 1;
r0 = 1; % var_w selected so that var_x = 1
r1 = -(a1/(1+a2))*r0;
r2 = (-a2+a1^2/(1+a2))*r0;
R = toeplitz([r0,r1]); Rinv = R\eye(2);
d = [r1;r2];
c = zeros(2,100);
delP = -d;
c(:,1) = -mu*Rinv*delP;
for k = 2:100
delP = R*c(:,k-1) - d;
c(:,k) = c(:,k-1) - mu*Rinv*delP;
end
c = [c0,c];

subplot(2,2,4);
plot(c(1,:),c(2,:),’md’,c(1,:),c(2,:),’g-’); axis([0,-2*a1,-2*a2,0]);
xlabel(’c_{k,1}’,’fontsize’,label_fontsize);
ylabel(’c_{k,2}’,’fontsize’,label_fontsize);
title(’Trajectory of c_{k,1} vs c_{k,2}, \mu = 1.0’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,-a1],’ytick’,[-a2,0]);grid;
text(-a1+0.005,-a2+0.05,’c_o’);
text(-a1+.03,-a2+0.7,’a_1 = -0.195’);
text(-a1+.03,-a2+0.5,’a_2 = 0.95’);

10.4 Consider the adaptive linear prediction of an AR(2) process x(n) using the LMS algorithm in which

x(n) = 0.95x(n − 1) − 0.9x(n − 2) + w(n)

where w(n) ∼ WGN(0, σ 2w ). The adaptive predictor is a second-order one given by a(n) = [a1 (n) a2 (n)]T .

(a) Implement the LMS algorithm given in Table 10.3 as a Matlab function [c,e] = lplms(x,y,mu,M,c0)
which computes filter coefficients in c and the corresponding error in e, given signal x, desired signal y,
step size mu, filter order M, and the initial coefficient vector c0.
function [c,e] = lplms(x,y,mu,M,c0)
% Computes filter coefficients in c and the corresponding error in
% e, given signal x, desired signal y, step size mu, filter order M, and
% the initialcoefficient vector c0.
%
% [c,e] = lplms(x,y,mu,M,c0)
%

N = length(x); % Length of x(n)

x = reshape(x,N,1); % x as a column vector
y = x(2:end); % Generate y(n)
N = length(y); % Number of iterations
c = zeros(M,N); % Initialization of c
Statistical and Adaptive Signal Processing - Solution Manual 305

X = zeros(M,N); % Data matrix

xx = [zeros(M-1,1);x];
for i = 1:M
X(i,:) = xx(M-i+1:end-i)’;
end
e = zeros(1,N);

%--LMS Algorithm for Linear Prediction - real-valued case

% Algorithm initialization
yhat = c0’*X(:,1);
e(1) = y(1) - yhat;
c(:,1) = 2*mu*X(:,1)*e(1);;
% Iterations
for n = 2:N
yhat = c(:,n-1)’*X(:,n);
e(n) = y(n) - yhat;
c(:,n) = c(:,n-1) + 2*mu*X(:,n)*e(n);
end
c = [c0,c];
e = [x(1),e];
(b) Generate 500 samples of x(n), and obtain linear predictor coefficients using the above function. Use step
size µ so that the algorithm converges in the mean. Plot predictor coefficients as a function of time along
with the true coefficients.
The Matlab script file is shown below and plots are shown in Figure 10.4 (a) and (b).
clc; close all
set(0,’defaultaxesfontsize’,default_fontsize);
Lx = [-1,3]; % Maximum abs limit for the x-axis of contour plots
Ly = [-3,1]; % Maximum abs limit for the y-axis of contour plots
N = 501; % Signal size (also number of iterations)
Niter = 1000; % Ensemble size
mua = 0.04; % Large Step-size
mum = 0.01; % Small Step-size

% Generate AR(2) process

a1 = -0.950; a2 = 0.9; varx = 1; r0 = varx;
%a1a = a1; a2a = a2;
varw = ((1-a2)*((1+a2)^2-a1^2)/(1+a2))*varx;
r1 = -a1/(1+a2)*r0; r2 = (-a2+a1^2/(1+a2))*r0;
lam1 = (1-a1/(1+a2))*varx; lam2 = (1+a1/(1+a2))*varx;
XR = lam1/lam2;d = [r1;r2]; R = toeplitz([r0,r1,r2]);
% Ensemble averaging
e2 = zeros(N,1); e = e2; c1m = e; c2m = e; c1 = e; c2 = e;
em2 = zeros(N,1); em = em2; cm1m = em; cm2m = em; cm1 = em; cm2 = em;
for i=1:Niter
w = sqrt(varw)*randn(N,1);
x = filter(1,[1 a1 a2],w);

c1(1) = 0; c1(2) = 0; c2(1) = 0; c2(2) = 0;

306 Statistical and Adaptive Signal Processing - Solution Manual

cm1(1) = 0; cm1(2) = 0; cm2(1) = 0; cm2(2) = 0;

% LMS Algorithm
for n=3:N
e(n) = x(n)-c1(n-1)*x(n-1)-c2(n-1)*x(n-2);
c1(n) = c1(n-1)+2*mua*e(n)*x(n-1);
c2(n) = c2(n-1)+2*mua*e(n)*x(n-2);
em(n) = x(n)-cm1(n-1)*x(n-1)-cm2(n-1)*x(n-2);
cm1(n) = cm1(n-1)+2*mum*em(n)*x(n-1);
cm2(n) = cm2(n-1)+2*mum*em(n)*x(n-2);
end
c1m = c1m+c1; c2m = c2m+c2; e2 = e2+e.^2;
cm1m = cm1m+cm1; cm2m = cm2m+cm2; em2 = em2+em.^2;
end
c1ma = c1m/Niter; c2ma = c2m/Niter; e2a = e2/Niter;
cm1ma = cm1m/Niter; cm2ma = cm2m/Niter; em2a = em2/Niter;

% contour plots
a1p=(Lx(1):0.1:Lx(2))’;
a2p=(Ly(1):0.1:Ly(2))’;
L1=length(a1p);
L2=length(a2p);

for i=1:L1
for j=1:L2
a=[1 -a1p(i) -a2p(j)]’;
P(i,j)=a’*R*a;
end
end

%Plots
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4.8],...
’NumberTitle’,’off’,’Name’,’Pr1004’);

subplot(’position’,[0.075,0.55,0.4,0.4]);
contour(a1p,a2p,P’,(1:2:20)/10); hold on;
plot(c1,c2,’r’,c1ma,c2ma,’g.’);
%plot(c1ma,c2ma,’r’,’linewidth’,1.5);
%plot(Lx,[0,0],’w:’,[0,0],Ly,’w:’,Lx,[-a1,-a1],’w:’,[-a2,-a2],[0,0],’w:’);
axis([Lx,Ly]);axis(’square’);
title(’(a) Averaged Trajectory’,’fontsize’,10);
xlabel(’c_{1}’,’fontsize’,8); ylabel(’c_{2}’,’fontsize’,8);
set(gca,’xtick’,[Lx(1),0,-a1,Lx(2)],’ytick’,[Ly(1),-a2,0,Ly(2)],...
’fontsize’,6); grid;
hold off;

subplot(’position’,[0.575,0.55,0.4,0.4]); n = 0:N-1;
plot([-1,N],[-a2,-a2],’w:’,[-1,N],[-a1,-a1],’w:’); hold on;
plot(n,c1,’c’,n,c2,’g’,’linewidth’,0.5);
Statistical and Adaptive Signal Processing - Solution Manual 307

(a) Averaged Trajectory (b) c(n) Learning Curve

1

0.95

0
c (n)
1

Coefficients
−0.9
c2

c (n)
2

−0.9

−3
−1 0 0.95 3 0 500
c1 Number of iterations (n)

(c) MSE P(n) Learning Curve (d) Step−size effect on MSE

1
1

µ=0.01

0.5

0.1425
0.1425
µ=0.04
0 0
0 500 0 500
Number of iterations (n) Number of iterations (n)

Figure 10.4: Plots of trajectories and learning curves in Problem 10.4

plot(n,c1ma,’w’,n,c2ma,’w’,’linewidth’,1);
axis([-1,N,-1.2,1.2]);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Coefficients’,’fontsize’,8);
title(’(b) {\bf c}(n) Learning Curve’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[-a2,0,-a1],’fontsize’,6);
text(50, 0.5,’c_1(n)’,’fontsize’,10);
text(50,-0.5,’c_2(n)’,’fontsize’,10);
hold off;

(c) Repeat the above simulation 1000 times to obtain the learning curve, which is obtained by averaging the
squared error |e(n)|2 . Plot this curve and compare its steady-state value with the theoretical MSE.
The Matlab script file is shown below and plots are shown in Figure 10.4 (c) and (d).
subplot(’position’,[0.075,0.1,0.4,0.3]);
plot([-1,N],[varw,varw],’w:’,n,e,’g’,n,e2a,’m’); axis([-1,N,0,1.1]);
xlabel(’Number of iterations (n)’,’fontsize’,8);
title(’(c) MSE P(n) Learning Curve’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[0,varw,1],’fontsize’,6);

subplot(’position’,[0.575,0.1,0.4,0.3]);
plot([-1,N],[varw,varw],’w:’,n,e2a,’g’,n,em2a,’c’); axis([-1,N,0,1]);
308 Statistical and Adaptive Signal Processing - Solution Manual

xlabel(’Number of iterations (n)’,’fontsize’,8);

title(’(d) Step-size effect on MSE’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[0,varw,0.5,1],’fontsize’,6);
text(30,0.1,’\mu=0.04’,’fontsize’,8);
text(50,0.7,’\mu=0.01’,’fontsize’,8);

10.5 Consider the adaptive echo canceler given in Figure 10.25. The FIR filter co (n) is given by

co (n) = (0.9)n 0≤n≤2

In this simulation, ignore the far-end signal u(n). The data signal x(n) is a zero-mean, unit-variance white
Gaussian process, and y(n) is its echo.
We will first need the Matlab function firlms to implement the LMS algorithm which is given below.

function [c,e,yhat] = firlms(x,y,mu,M,c0)

% Computes filter coefficients in c and the corresponding error in
% e, given signal x, desired signal y, step size mu, filter order M, and
% the initial coefficient vector c0.
%
% [c,e] = firlms(x,y,mu,M,c0)
%

N = length(x); % Length of x(n) & number of iterations

x = reshape(x,N,1); % x as a column vector
c = zeros(M,N); % Initialization of c
X = zeros(M,N); % Data matrix
xx = [zeros(M-1,1);x]; % Prepend M-1 zeros
for i = 1:M
X(i,:) = xx(M-i+1:end-i+1)’;
end
e = zeros(1,N); yhat = zeros(N,1);

%--LMS Algorithm for FIR filter - real-valued case

% Algorithm initialization
yhat(1) = c0’*X(:,1);
e(1) = y(1) - yhat(1);
c(:,1) = c0 + 2*mu*X(:,1)*e(1);;
% Iterations
for n = 2:N
yhat(n) = c(:,n-1)’*X(:,n);
e(n) = y(n) - yhat(n);
c(:,n) = c(:,n-1) + 2*mu*X(:,n)*e(n);
end

(a) Generate 1000 samples of x(n) and determine y(n). Use these signals to obtain a fourth-order LMS
echo canceler in which the step size µ is chosen to satisfy (10.4.40) and c(0) = 0. Obtain the final echo
canceler coefficients and compare them with the true ones.
The Matlab script is given below.
Statistical and Adaptive Signal Processing - Solution Manual 309

MSE Learning Curve

3
Actual MSE
Computed MSE
2.5

MSE
1.5

0.5

0
0 200 400 600 800 1000
Number of iterations (n)

Figure 10.5: Plots of the MSE learning curve in Problem 10.5

clc; close all

set(0,’defaultaxesfontsize’,default_fontsize);
co = (0.9).^[0:2]; % Echo impulse response

% (a) Echo Canceller design

N = 1000; x = randn(N,1);
y = filter(co,1,x);
M = 4; % LMS filter order
mu = 0.005; c0 = zeros(M,1);
[c,e] = firlms(x,y,mu,M,c0);

% Final Echo-canceller coefficients

cf = c(:,N);
*** Echo Canceller coefficients ***
1.0000 0.9000 0.8100 0.0000

*** Echo System coefficients ***

1.0000 0.9000 0.8100

(b) Repeat the above simulation 500 times, and obtain the learning curve. Plot this curve along with the
actual MSE and comment on the plot.
The Matlab script is given below and the learning curve is shown in Figure 10.5b.
% (b) Monte-Carlo analysis and the learning curve
Nsim = 500; N = 1000; M = 4; mu = 0.005; c0 = zeros(M,1);

eavg = zeros(1,N,Nsim);
cavg = zeros(M,N,Nsim);

for k = 1:Nsim
x = randn(N,1);y = filter(co,1,x);
[c,e] = firlms(x,y,mu,M,c0);
cavg(:,:,k) = c;
eavg(:,:,k) = e;
end
cavg = mean(cavg,3);
eavg = eavg.^2; eavg = mean(eavg,3);
310 Statistical and Adaptive Signal Processing - Solution Manual

% Theoretical analysis
n = 1:N; alpha = (1-4*mu+4*mu*mu*M);
ethe = var_y*alpha.^n;

% Plots
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2],...
’NumberTitle’,’off’,’Name’,’Pr1005’);
subplot(’position’,[0.1,0.15,0.85,0.7]);
plot(n,ethe,’g-.’,n,eavg,’r’);axis([0,N,0,3]);
xlabel(’Number of iterations (n)’,’fontsize’,label_fontsize);
title(’MSE Learning Curve’,’fontsize’,title_fontsize);
ylabel(’MSE’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:200:N],’ytick’,[0:0.5:3]);
legend(’Actual MSE’,’Computed MSE’,1);
(c) Repeat parts (a) and (b), using a third-order echo canceler.
Modify above Matlab scripts to obtain the necessary results.
(d) Repeat parts (a) and (b), using one-half the value of µ used in the first part.
Modify above Matlab scripts to obtain the necessary results.

10.6 The normalized LMS (NLMS) algorithm is given in (10.4.67), in which the effective step size is time-varying
and is given by µ̃/)x(n))2 , where 0 < µ̃ < 1.

(a) Modify the function firlms to implement the NLMS algorithm and obtain the function [c,e] =
nfirlms(x,y,mu,M,c0).

function [c,e] = nfirlms(x,y,mu,M,c0)

% Normalized LMS algorithm for FIR filter applications
% Computes filter coefficients in c and the corresponding error in
% e, given signal x, desired signal y, step size mu, filter order M, and
% the initial coefficient vector c0.
%
% [c,e] = nfirlms(x,y,mu,M,c0)
%

N = length(x); % Length of x(n) & number of iterations

x = reshape(x,N,1); % x as a column vector
c = zeros(M,N); % Initialization of c
X = zeros(M,N); % Data matrix
xx = [zeros(M-1,1);x]; % Prepend M-1 zeros
for i = 1:M
X(i,:) = xx(M-i+1:end-i+1)’;
end
e = zeros(1,N);

% Computation of EM(n)
Statistical and Adaptive Signal Processing - Solution Manual 311

(a) Averaged Trajectory (b) c(n) Learning Curve

1

0.95

0
c (n)
1

Coefficients
−0.9
c2

c (n)
2

−0.9

−3
−1 0 0.95 3 0 500
c1 Number of iterations (n)

(c) MSE P(n) Learning Curve (d) Step−size effect on MSE

1
1

µ=0.04 (LMS)

0.5

0.1425
0.1425
µ=0.1 (NLMS)
0 0
0 500 0 500
Number of iterations (n) Number of iterations (n)

Figure 10.6b: Plots of the NLMS learning curve and its comparison with LMS in Problem 10.6b

w = abs(x).*abs(x);
b = [1,zeros(1,M-1),-1]; a = [1,-1];
EM = filter(b,a,w);

%--NLMS Algorithm for FIR filter - real-valued case

% Algorithm initialization
yhat = c0’*X(:,1);
e(1) = y(1) - yhat;
c(:,1) = c0 + mu*X(:,1)*e(1)/EM(1);;
% Iterations
for n = 2:N
yhat = c(:,n-1)’*X(:,n);
e(n) = y(n) - yhat;
c(:,n) = c(:,n-1) + mu*X(:,n)*e(n)/EM(n);
end
e = e(M+1:end);
c = c(:,M+1:end);
(b) Choose µ̃ = 0.1 and repeat Problem 10.4. Compare your results in terms of convergence speed.
The Matlab script is shown below and the plots in Figure 10.6b. From the plots we observe that the
NLMS has the faster convergence.
312 Statistical and Adaptive Signal Processing - Solution Manual

% (b) NLMS for LP Applications

% AR(2) process:
% x(n) + a1*x(n-1) + a2*x(n-2) = w(n)
%
% Linear Predictor
% y(n) = x(n); yhat = c1*x(n-1) + c2*x(n-2);

clc; close all

set(0,’defaultaxesfontsize’,default_fontsize);
Lx = [-1,3]; % Maximum abs limit for the x-axis of contour plots
Ly = [-3,1]; % Maximum abs limit for the y-axis of contour plots
N = 501; % Signal size (also number of iterations)
Niter = 1000; % Ensemble size
mua = 0.01; % Step-size for LMS
mum = 0.1; % Step-size for NLMS

% Generate AR(2) process

a1 = -0.950; a2 = 0.9; varx = 1; r0 = varx;
%a1a = a1; a2a = a2;
varw = ((1-a2)*((1+a2)^2-a1^2)/(1+a2))*varx;
r1 = -a1/(1+a2)*r0; r2 = (-a2+a1^2/(1+a2))*r0;
lam1 = (1-a1/(1+a2))*varx; lam2 = (1+a1/(1+a2))*varx;
XR = lam1/lam2;d = [r1;r2]; R = toeplitz([r0,r1,r2]);
% Ensemble averaging
e2 = zeros(N,1); e = e2; c1m = e; c2m = e; c1 = e; c2 = e;
em2 = zeros(N,1); em = em2; cm1m = em; cm2m = em; cm1 = em; cm2 = em;
for i=1:Niter
w = sqrt(varw)*randn(N,1);
x = filter(1,[1 a1 a2],w);
EM = filter([1,0,-1],[1,-1],abs(x).*abs(x))+0.001;

c1(1) = 0; c1(2) = 0; c2(1) = 0; c2(2) = 0;

cm1(1) = 0; cm1(2) = 0; cm2(1) = 0; cm2(2) = 0;
% LMS Algorithm
for n=3:N
e(n) = x(n)-c1(n-1)*x(n-1)-c2(n-1)*x(n-2);
c1(n) = c1(n-1)+2*mua*e(n)*x(n-1);
c2(n) = c2(n-1)+2*mua*e(n)*x(n-2);
em(n) = x(n)-cm1(n-1)*x(n-1)-cm2(n-1)*x(n-2);
cm1(n) = cm1(n-1)+mum*em(n)*x(n-1)/EM(n-1);
cm2(n) = cm2(n-1)+mum*em(n)*x(n-2)/EM(n-1);
end
c1m = c1m+c1; c2m = c2m+c2; e2 = e2+e.^2;
cm1m = cm1m+cm1; cm2m = cm2m+cm2; em2 = em2+em.^2;
end

c1ma = c1m/Niter; c2ma = c2m/Niter; e2a = e2/Niter;

cm1ma = cm1m/Niter; cm2ma = cm2m/Niter; em2a = em2/Niter;
Statistical and Adaptive Signal Processing - Solution Manual 313

% contour plots

a1p=(Lx(1):0.1:Lx(2))’;
a2p=(Ly(1):0.1:Ly(2))’;
L1=length(a1p);
L2=length(a2p);

for i=1:L1
for j=1:L2
a=[1 -a1p(i) -a2p(j)]’;
P(i,j)=a’*R*a;
end
end

%Plots
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,6,4.8],...
’NumberTitle’,’off’,’Name’,’Pr1006a’);

subplot(’position’,[0.075,0.55,0.4,0.4]);
contour(a1p,a2p,P’,(1:2:20)/10); hold on;
plot(cm1,cm2,’r’,cm1ma,cm2ma,’g.’);
%plot(c1ma,c2ma,’r’,’linewidth’,1.5);
%plot(Lx,[0,0],’w:’,[0,0],Ly,’w:’,Lx,[-a1,-a1],’w:’,[-a2,-a2],[0,0],’w:’);
axis([Lx,Ly]);axis(’square’);
title(’(a) Averaged Trajectory’,’fontsize’,10);
xlabel(’c_{1}’,’fontsize’,8); ylabel(’c_{2}’,’fontsize’,8);
set(gca,’xtick’,[Lx(1),0,-a1,Lx(2)],’ytick’,[Ly(1),-a2,0,Ly(2)],...
’fontsize’,6); grid;
hold off;

subplot(’position’,[0.575,0.55,0.4,0.4]); n = 0:N-1;
plot([-1,N],[-a2,-a2],’w:’,[-1,N],[-a1,-a1],’w:’); hold on;
plot(n,cm1,’c’,n,cm2,’g’,’linewidth’,0.5);
plot(n,cm1ma,’w’,n,cm2ma,’w’,’linewidth’,1);
axis([-1,N,-1.2,1.2]);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Coefficients’,’fontsize’,8);
title(’(b) {\bf c}(n) Learning Curve’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[-a2,0,-a1],’fontsize’,6);
text(50, 0.5,’c_1(n)’,’fontsize’,10);
text(50,-0.5,’c_2(n)’,’fontsize’,10);
hold off;

subplot(’position’,[0.075,0.1,0.4,0.3]);
plot([-1,N],[varw,varw],’w:’,n,em,’g’,n,em2a,’m’); axis([-1,N,0,1.1]);
xlabel(’Number of iterations (n)’,’fontsize’,8);
title(’(c) MSE P(n) Learning Curve’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[0,varw,1],’fontsize’,6);
314 Statistical and Adaptive Signal Processing - Solution Manual

subplot(’position’,[0.575,0.1,0.4,0.3]);
plot([-1,N],[varw,varw],’w:’,n,e2a,’g’,n,em2a,’c’); axis([-1,N,0,1]);
xlabel(’Number of iterations (n)’,’fontsize’,8);
title(’(d) Step-size effect on MSE’,’fontsize’,10);
set(gca,’xtick’,[0,N-1],’ytick’,[0,varw,0.5,1],’fontsize’,6);
text(30,0.1,’\mu=0.1 (NLMS)’,’fontsize’,8);
text(50,0.7,’\mu=0.04 (LMS)’,’fontsize’,8);
(c) Choose µ̃ = 0.1 and repeat Problem 10.5(a) and (b). Compare your results in terms of convergence
speed.
The Matlab script is shown below and the plots in Figure 10.6c. From the plots we again observe that the
NLMS has the faster convergence.
% (c) NLMS for Echo Cancellation
% Data signal x(n): zero-mean, unit-variance white Gaussian process
% Echo System: co(n) = 0.9^n, 0 <= n <= 2
% Echo signal: y(n) = conv(x(n),co(n))

co = (0.9).^[0:2]; % Echo impulse response

var_y = sum(co.^2); % Variance of y(n)

% (ca) Echo Canceller design

M = 4; % LMS filter order
N = 400; x = randn(N,1); % Generate x(n)
y = filter(co,1,x); % Generate y(n)
mu = 0.1; c0 = zeros(M,1); % Initialization
[c,e] = nfirlms(x,y,mu,M,c0); % LMS algorithm

% Final Echo-canceller coefficients

cf = c(:,end);
*** Echo Canceller coefficients ***
1.0000 0.9000 0.8100 -0.0000

*** Echo System coefficients ***

1.0000 0.9000 0.8100

% (cb) Monte-Carlo analysis and the learning curve

Nsim = 500;

eavg = zeros(1,N,Nsim);
cavg = zeros(M,N,Nsim);

for k = 1:Nsim
x = randn(N+M,1);y = filter(co,1,x);
[c,e] = nfirlms(x,y,mu,M,c0);
cavg(:,:,k) = c;
eavg(:,:,k) = e;
end
cavg = mean(cavg,3);
Statistical and Adaptive Signal Processing - Solution Manual 315

MSE Learning Curve

3
Actual LMS MSE
Computed NLMS MSE
2.5

MSE
1.5

0.5

0
0 200 400
Number of iterations (n)

Figure 10.6c: Plots of the NLMS learning curve and its comparison with LMS in Problem 10.6c

eavg = eavg.^2; eavg = mean(eavg,3);

eavg = [var_y,eavg];

% Theoretical analysis
muo = 0.005;
n = 0:N; alpha = (1-4*muo+4*muo*muo*M);
ethe = var_y*alpha.^n;

% Plots
Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2],...
’NumberTitle’,’off’,’Name’,’Pr1006c’);
subplot(’position’,[0.1,0.15,0.85,0.7]);
plot(n,ethe,’g-.’,n,eavg,’r’); axis([0,N,0,3]);
xlabel(’Number of iterations (n)’,’fontsize’,label_fontsize);
title(’MSE Learning Curve’,’fontsize’,title_fontsize);
ylabel(’MSE’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:200:N],’ytick’,[0:0.5:3]);
legend(’Actual LMS MSE’,’Computed NLMS MSE’,1)

10.7 Another variation of the LMS algorithm is called the sign-error LMS algorithm, in which the coefficient update
equation is given by

c(n) = c(n − 1) + 2µ sgn[e(n)]x(n), where


 1 Re[e(n)] > 0

sgn [e(n)] = 0 Re[e(n)] = 0


−1 Re[e(n)] < 0

(a) Modify the function firlms to implement the NLMS algorithm and obtain the function [c,e] = se-
firlms(x,y,mu,M,c0).

function [c,e] = sefirlms(x,y,mu,M,c0)

% Sign-Error LMS Algorithm
% Computes filter coefficients in c and the corresponding error in
316 Statistical and Adaptive Signal Processing - Solution Manual

% e, given signal x, desired signal y, step size mu, filter order M, and
% the initial coefficient vector c0.
%
% [c,e] = sefirlms(x,y,mu,M,c0)
%

N = length(x); % Length of x(n) & number of iterations

x = reshape(x,N,1); % x as a column vector
c = zeros(M,N); % Initialization of c
X = zeros(M,N); % Data matrix
xx = [zeros(M-1,1);x]; % Prepend M-1 zeros
for i = 1:M
X(i,:) = xx(M-i+1:end-i+1)’;
end
e = zeros(1,N);

%--LMS Algorithm for FIR filter - real-valued case

% Algorithm initialization
yhat = c0’*X(:,1);
e(1) = y(1) - yhat;
c(:,1) = c0 + 2*mu*X(:,1)*e(1);;
% Iterations
for n = 2:N
yhat = c(:,n-1)’*X(:,n);
e(n) = y(n) - yhat;
c(:,n) = c(:,n-1) + 2*mu*X(:,n)*sign(e(n));
end
e = e(M+1:end);
c = c(:,M+1:end);
(b) Repeat Problem 10.4 and compare your results in terms of convergence speed.
The Matlab script is shown below and the plots in Figure 10.7b. From the plots we observe that the
SELMS has just a little faster convergence.
% (b) SELMS for LP Applications
% AR(2) process:
% x(n) + a1*x(n-1) + a2*x(n-2) = w(n)
%
% Linear Predictor
% y(n) = x(n); yhat = c1*x(n-1) + c2*x(n-2);

clc; close all

set(0,’defaultaxesfontsize’,default_fontsize);
Lx = [-1,3]; % Maximum abs limit for the x-axis of contour plots
Ly = [-3,1]; % Maximum abs limit for the y-axis of contour plots
N = 501; % Signal size (also number of iterations)
Niter = 1000; % Ensemble size
mua = 0.01; % Step-size for LMS
mum = 0.01; % Step-size for NLMS
Statistical and Adaptive Signal Processing - Solution Manual 317

(a) Averaged Trajectory (b) c(n) Learning Curve

1

0.95

0
c (n)
1

Coefficients
−0.9
c2

c (n)
2

−0.9

−3
−1 0 0.95 3 0 500
c1 Number of iterations (n)

(c) MSE P(n) Learning Curve (d) Step−size effect on MSE

1
1

0.5

µ=0.01 (LMS)

0.1425
0.1425
µ=0.01 (NLMS)
0 0
0 500 0 500
Number of iterations (n) Number of iterations (n)

Figure 10.7b: Plots of the SELMS learning curve and its comparison with LMS in Problem 10.7b

% Generate AR(2) process

a1 = -0.950; a2 = 0.9; varx = 1; r0 = varx;
%a1a = a1; a2a = a2;
varw = ((1-a2)*((1+a2)^2-a1^2)/(1+a2))*varx;
r1 = -a1/(1+a2)*r0; r2 = (-a2+a1^2/(1+a2))*r0;
lam1 = (1-a1/(1+a2))*varx; lam2 = (1+a1/(1+a2))*varx;
XR = lam1/lam2;d = [r1;r2]; R = toeplitz([r0,r1,r2]);
% Ensemble averaging
e2 = zeros(N,1); e = e2; c1m = e; c2m = e; c1 = e; c2 = e;
em2 = zeros(N,1); em = em2; cm1m = em; cm2m = em; cm1 = em; cm2 = em;
for i=1:Niter
w = sqrt(varw)*randn(N,1);
x = filter(1,[1 a1 a2],w);

c1(1) = 0; c1(2) = 0; c2(1) = 0; c2(2) = 0;

cm1(1) = 0; cm1(2) = 0; cm2(1) = 0; cm2(2) = 0;
% LMS Algorithm
for n=3:N
e(n) = x(n)-c1(n-1)*x(n-1)-c2(n-1)*x(n-2);
c1(n) = c1(n-1)+2*mua*(e(n))*x(n-1);
c2(n) = c2(n-1)+2*mua*(e(n))*x(n-2);
318 Statistical and Adaptive Signal Processing - Solution Manual

em(n) = x(n)-cm1(n-1)*x(n-1)-cm2(n-1)*x(n-2);
cm1(n) = cm1(n-1)+2*mum*sign(em(n))*x(n-1);
cm2(n) = cm2(n-1)+2*mum*sign(em(n))*x(n-2);
end
c1m = c1m+c1; c2m = c2m+c2; e2 = e2+e.^2;
cm1m = cm1m+cm1; cm2m = cm2m+cm2; em2 = em2+em.^2;
end

c1ma = c1m/Niter; c2ma = c2m/Niter; e2a = e2/Niter;

cm1ma = cm1m/Niter; cm2ma = cm2m/Niter; em2a = em2/Niter;

% contour plots

a1p=(Lx(1):0.1:Lx(2))’;
a2p=(Ly(1):0.1:Ly(2))’;
L1=length(a1p);
L2=length(a2p);

for i=1:L1
for j=1:L2
a=[1 -a1p(i) -a2p(j)]’;
P(i,j)=a’*R*a;
end
end

(c) Repeat Problem 10.5(a) and (b) and compare your results in terms of convergence speed.
The Matlab script is shown below and the plots in Figure 10.7c. From the plots we again observe that the
NLMS has the faster convergence but a higher steady-state error.
% (ca) Echo Canceller design
M = 4; % LMS filter order
N = 400; x = randn(N,1); % Generate x(n)
y = filter(co,1,x); % Generate y(n)
mu = 0.1; c0 = zeros(M,1); % Initialization
[c,e] = nfirlms(x,y,mu,M,c0); % LMS algorithm

% Final Echo-canceller coefficients

cf = c(:,end);
*** Echo Canceller coefficients ***
0.9999 0.9000 0.8100 0.0000

*** Echo System coefficients ***

1.0000 0.9000 0.8100

% (cb) Monte-Carlo analysis and the learning curve

Nsim = 500;

eavg = zeros(1,N,Nsim);
cavg = zeros(M,N,Nsim);
Statistical and Adaptive Signal Processing - Solution Manual 319

MSE Learning Curve

3
Actual LMS MSE
Computed SELMS MSE
2.5

MSE
1.5

0.5

0
0 200 400
Number of iterations (n)

Figure 10.7c: Plots of the SELMS learning curve and its comparison with LMS in Problem 10.7c

for k = 1:Nsim
x = randn(N+M,1);y = filter(co,1,x);
[c,e] = sefirlms(x,y,mu,M,c0);
cavg(:,:,k) = c;
eavg(:,:,k) = e;
end
cavg = mean(cavg,3);
eavg = eavg.^2; eavg = mean(eavg,3);
eavg = [var_y,eavg];

% Theoretical analysis
muo = 0.005;
n = 0:N; alpha = (1-4*muo+4*muo*muo*M);
ethe = var_y*alpha.^n;

% Plots
Hf_2 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2],...
’NumberTitle’,’off’,’Name’,’Pr1006c’);
subplot(’position’,[0.1,0.15,0.85,0.7]);
plot(n,ethe,’g-.’,n,eavg,’r’); axis([0,N,0,3]);
xlabel(’Number of iterations (n)’,’fontsize’,label_fontsize);
title(’MSE Learning Curve’,’fontsize’,title_fontsize);
ylabel(’MSE’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:200:N],’ytick’,[0:0.5:3]);
legend(’Actual LMS MSE’,’Computed SELMS MSE’,1)

10.8 Consider an AR(1) process x(n) = ax(n − 1) + w(n), where w(n) ∼ WGN(0, σ 2w ). We wish to design a
one-step first-order linear predictor using the LMS algorithm

x̂(n) = â(n − 1) x(n − 1) (1)

e(n) = x(n) − x̂(n) (2)
â(n) = â(n − 1) + 2 µe(n) x(n − 1) (3)

where µ is the adaptation step size.

320 Statistical and Adaptive Signal Processing - Solution Manual

(a) Determine the autocorrelation r x (l), the optimum first-order linear predictor, and the corresponding
MMSE.
The z-domain SDF, Rx (z), is given by
1
Rx (z) = H (z) H (z) Rw (z) where H (z) = and Rw (z) = σ 2w
1 − a z −1



2 & '
1 1 σw 1 1
= σw
2
= − ; |a| < |z| < |a|−1
1 − a z −1 1−az 1 − a2 1 − a z −1 1 − a −1 z −1
Thus after inverse z-transformation
σ 2w |l|
r x (l) = a , ∀l
1 − a2
Now the optimum first-order linear predictor is given by
r x (1)
â = =a
r x (0)
and the MMSE is given by

σ 2w σ 2w 2
Pe = r x (0) − â r x (1) = − a = σ 2w
1 − a2 1 − a2

(b) Using the independence assumption, first determine and then solve the difference equation for E{â(n)}.
From (3), (1), and (2), we have

â(n) = â(n − 1) + 2 µ x(n) − â(n − 1) x(n − 1) x(n − 1)
= â(n − 1) + 2 µ x(n)x(n − 1) − 2 µ â(n − 1) x 2 (n − 1)

Taking expectation of both sides using the independence assumption that â(n − 1) and x(n − 1) are
mutually independent (since â(n − 1) depends only on x(m), m < n − 1), we obtain

E[â(n)] = E[â(n − 1)] + 2 µ E[x(n)x(n − 1)] − 2 µ E[â(n − 1)]E[ x 2 (n − 1)]

= [1 − 2µ r x (0)] E[â(n − 1)] + 2 µ r x (1) (4)

which is first-order difference equation for E[â(n)] driven by 2 µ r x (1) for n ≥ 1 with E[â(0)] = 0. This
equation can be solved using unilateral z-transform. Substituting

α = 1 − 2µ r x (0), β = 2 µ r x (1), and y(n − 1) , E[â(n)] (5)

in (4), we have

y(n) = α y(n − 1) + β u(n), n ≥ 0; y(−1) = 0 (6)

Taking unilateral z-transform of (6), we obtain

1
Y (z) = α Y (z) + β
1 − z −1
or
& '
β β 1 α
Y (z) =   = −1
−
1 − α z −1 1−z −1 1−α 1−z 1 − α z −1
Statistical and Adaptive Signal Processing - Solution Manual 321

or
β 
y(n) = 1 − α n+1 u(n)
1−α
Using (5), we obtain the solution
r x (1) 
E[â(n)] = 1 − [1 − 2µ r x (0)]n u(n − 1) (7)
r x (0)
Note that as n → ∞ we have that E[â(n)] → rrxx (1)(0)
= a (as expected in the absense of noise) if
|1 − 2µ r x (0)| < 1.
(c) For a = ±0.95, µ = 0.025, σ 2x = 1, and 0 ≤ n < N = 500, determine the ensemble average of E{â(n)}
using 200 independent runs and compare with the theoretical curve obtained in part (b).
The Matlab script is shown below and the plots in Figure 10.8c.
% AR(1) process parameters and LMS parameters
a = 0.95; N = 500; n = 0:N; var_x = 1; var_w = var_x*(1 - a*a);
mu = 0.025; Nsim = 200;

% Monte-Carlo analysis and the learning curve

Nsim = 500;

eavg = zeros(1,N,Nsim);
cavg = zeros(M,N,Nsim);

for k = 1:Nsim
w = var_w*randn(N+1,1); x = filter(1,[1,-a],w);
y = x(2:end); x = x(1:N);
[c,e] = firlms(x,y,mu,M,c0);
cavg(:,:,k) = c;
eavg(:,:,k) = e;
end
cavg = mean(cavg,3); cavg = [0,cavg];
eavg = eavg.^2; eavg = mean(eavg,3);
eavg = [1,eavg];

% Theoretical analysis
alpha = 1 - 2*mu*var_x;
cthe = a*(1-alpha.^n);

% Plots
Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2],...
’NumberTitle’,’off’,’Name’,’Pr1008c’);
subplot(’position’,[0.1,0.15,0.85,0.7]);
plot(n,cthe,’g-.’,n,cavg,’r’); axis([0,N,0,1]);
xlabel(’Number of iterations (n)’,’fontsize’,label_fontsize);
title(’Linear-predictor Learning Curve’,’fontsize’,title_fontsize);
ylabel(’ahat(n)’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:200:N],’ytick’,[0:0.2:1]);
legend(’Theoretical’,’Averaged’,4)
322 Statistical and Adaptive Signal Processing - Solution Manual

Linear−predictor Learning Curve

1

0.8

0.6
ahat(n)

0.4

0.2
Theoretical
Averaged
0
0 200 400
Number of iterations (n)

Figure 10.8c: Plots of learning curves in Problem 10.8c

(d) Using the independence assumption, first determine and then solve the difference equation for P(n) =
E{e2 (n)}.
An analysis similar to that in (b) can be used in this part.
(e) Repeat part (c) for P(n) and comment upon the results.

10.9 Using the a-posteriori error ε(n) = y(n) − c H (n)x(n), derive the coefficient updating formulas for the a-
posteriori error LMS algorithm.
The a-posteriori form of the LMS algorithm

c(n) = c(n − 1) + 2µx(n)ε∗ (n)

requires the a-posteriori error

ε(n) = y(n) − c H (n)x(n)

 H
= y(n) − c(n − 1) + 2µx(n)e∗ (n) x(n)
 
= y(n) − c H (n − 1)x(n) − 2µe(n)x H (n)x(n)
= e(n) − 2µe(n)x H (n)x(n)

or
 
ε(n) = I − 2µx H (n)x(n) e(n)

Therefore, the a posteriori LMS algorithm is

e(n) = y(n) − c H (n − 1)x(n)

 
ε(n) = I − 2µx H (n)x(n) e(n)
c(n) = c(n − 1) + 2µx(n)ε∗ (n)

10.10 Solve the interference cancellation problem described in Example 6.4.1, using the LMS algorithm, and compare
its performance to that of the optimum canceler.
To be completed.

10.11 Repeat the convergence analysis of the LMS algorithm for the complex case, using formula (10.4.27) instead
of (10.4.28).
This a tedious repetition of the work in Section 10.4.2 and should be given only as a project.
Statistical and Adaptive Signal Processing - Solution Manual 323

10.12 Consider the total transient excess MSE, defined by

∞

Ptr(total) = Ptr (n)
n=0

in Section 10.4.3.

(a) Show that Ptr(total) can be written as Ptr(total) = λT (I − B)−1 +θ (0), where +θ(0) is the initial (i.e., at n = 0)
deviation of the filter coefficients from their optimum setting.
From (10.4.62) and (10.4.50) we have
∞
 ∞

Ptr(t ot al) = Ptr (n) = [P(n) − P(∞)]
n=0 n=0
∞ ∞

= λ [θ (n) − θ(∞)] =
T
λT Bn [θ(0) − θ(∞)]
n=0 n=0
(∞ )

= λT Bn θ (0) = λT (I − B)−1 θ(0)
n=0
!∞
where we have used (10.4.41) and n=0 Bn = (I − B)−1 , which holds when the eigenvalues of B have
magnitude less than one.
(b) Starting with the formula in step (a), show that

!
M
θ i (0)
1−2µλi
1
Ptr(total)
i=1
=
4µ !
M
µλi
1− 1−2µλi
i=1

(b) From (10.4.33) we obtain

B = (ρ) + 4µ2 λλT

and

I − B = [I − (ρ)] − 4µ2 λλT

Using the matrix inversion lemma (A.4) we obtain

4µ2
(I − B)−1 = [I − (ρ)]−1 + [I − (ρ)]−1 λλT [I − (ρ)]−1
1 − 4µ2 λT [I − (ρ)]−1 λ
Hence,

Ptrt ot al = λT (I − B)−1 θ (0)

4µ2 αβ β
=β+ =
1 − 4µ α
2 1 − 4µ2 α
where

M
λ2i
−1
α = λ [I − (ρ)]
T
λ=
i=1
1 − ρi
324 Statistical and Adaptive Signal Processing - Solution Manual


M
λi θ i (0)
−1
β = λ [I − (ρ)]
T
θ(0) =
i=1
1 − ρi

From (10.4.36) can easily see that

λi 1
=
1 − ρi 4µ(1 − 2µλi )
Thus

1 
M
λi
α=
4µ i=1 1 − 2µλi

1  θ i (0)
M
β=
4µ i=1 1 − 2µλi

and
! M θ i (0)
1 i=1 1−2µλi
Ptrt ot al = ! M µλi
4µ 1 − i=1
1−2µλi

(c) Show that if µλk , 1, then

!
M
θ i (0)
1 
M
1 i=1
Ptr(total) & & θ i (0)
4µ 1 − µ tr(R) 4µ i=1

which is formula (10.4.62), discussed in Section 10.4.3.

When µλi , 1 we have 1 − µλi & 1 and 1 − µ tr R & 1. Hence
!M
1 
M
1 i=1 θ i (0)
Ptrt ot al & & θ i (0)
4µ 1 − µ tr R 4µ i=1

! M−1
10.13 The frequency sampling structure for the implementation of an FIR filter H (z) = n=0 h(n) · z −n is specified
by the following relation

1 − z −M  H (e j 2πk/M )
M−1
H (z) = , H1 (z) H2(z)
M k=0
1 − e j 2πk/M z −1

where H1 (z) is a comb filter with M zeros equally spaced on the unit circle and H2 (z) is a filter bank of
resonators. Note that H̃ (k) , H (e j 2πk/M ), the DFT of {h(n)}0M−1, is the coefficients of the filter. Derive an
LMS-type algorithm to update these coefficients, and sketch the resulting adaptive filter structure.
If we define
1 1 − zM
Yk (z) , X (x)
1 − e j 2πk/M z −1 M
Statistical and Adaptive Signal Processing - Solution Manual 325

we can obtain the adaptive filter structure shown in Figure 10.13. An adaptive LMS filter is obtained using the
following algorithm


M−1
ŷ(n) = ck (n − 1)yk (n)
k=0
e(n) = y(n) − ŷ(n)
ck (n) = ck (n − 1) + 2µe(n) ŷk (n)

To understand the nature of this structure, consider the frequency response of the kth parallel path of the filter
(resonator)
1 − zM
Hk (z) = 
M 1 − e j 2πk/M z −1

or for z = e2π f

1 − e−2π f M e−π f M eπ f M − e−π f M
Hk (e 2π f
)=  = $ %
M 1 − e j 2πk/M e− j 2π f Me j π ( M − f ) e− j π ( M − f ) − e j π ( M − f )
k k k

which leads to
 
   sin π M f  &
 Hk (e ) = 1   1, f = k
2π f
   = M
M  sin π Mk − f  0, f = k
M

Therefore, when

M−1
2π k
y(n) = Ak cos n
k=0
M

each coefficient of the adaptive filter can be updated without any interference from the other coefficients.

10.14 There are applications in which the use of a non-MSE criterion may be more appropriate. To this end, suppose
that we wish to design and study the behavior of an “LMS-like” algorithm that minimizes the cost function
P (k) = E{e2k (n)}, k = 1, 2, 3, . . . , using the model defined in Figure 10.19.
Since the cost function is

P (k) = E e2k (n) , k = 1, 2, ...

the error e(n) must be real-valued. Hence,

e(n) = y(n) − cT (n − 1)x(n)

(a) Use the instantaneous gradient vector to derive the coefficient updating formula for this LMS-like algo-
rithm.
An LMS-type algorithm is obtained using the instantaneous gradient
∂ P (k) ∂e(n)
= 2ke2k−1 (n) = −2ke2k−1 (n)x(n)
∂c ∂c
where we have used e(n) = y(n) − cT x(n). Hence

c(n) = c(n − 1) + 2kµe2k−1 (n)x(n)

326 Statistical and Adaptive Signal Processing - Solution Manual

1 y0 (n)
j 2π 0 x
1−e M z −1
c0 (n − 1) y(n)
1 y1 (n)
x(n) x
1−z M
1−e
j 2π 1
M z −1
M e(n)
-
c1 (n − 1) + +
. ŷ(n)
.
.

1 y M−1 (n)
j 2π (M−1) x
1−e M z −1
c M−1 (n − 1)

Figure 10.13: Adaptive LMS filter using the frequency sampling structure.

(b) Using the assumptions introduced in Section 10.4.2 show that

E{c̃(n)} = [I − 2µk(2k − 1)E{eo2(k−1) (n)}R]E{c̃(n − 1)}

where R is the input correlation matrix.

To study convergence in the mean we consider the model given by (10.4.13)

y(n) = coT x(n) + eo (n)

We start by expressing the a priori error as

e(n) = y(n) − cT (n − 1)x(n)

= coT x(n) − cT (n − 1)x(n) + eo (n)
= [co − c(n − 1)] x(n) + eo (n)

or

e(n) = −c̃T (n − 1)x(n) + eo (n)

where c̃(n) , c(n) − co is the coefficient error vector. Hence

 2k−1
e2k−1 (n) = −c̃T (n − 1)x(n) + eo (n)
 2k−1
c̃T (n − 1)x(n)
= eo (n) 1 −
2k−1
eo (n)
 
c̃T (n − 1)x(n)
& eo (n) 1 − (2k − 1)
2k−1
eo (n)
= eo (n) − (2k − 1)c̃ (n − 1)x(n)eo2k−2 (n)
2k−1 T

when c̃(n) takes very small values, that is when the coefficients are changing very slowly.
Statistical and Adaptive Signal Processing - Solution Manual 327

Substituting into the coefficient updating formula, we obtain

c̃(n) & c̃(n − 1) + 2kµeo2k−1 (n)x(n) − (2k − 1)2µkx(n)xT (n)c̃(n − 1)eo2k−2 (n)

or
 
c̃(n) & I − 2µk(2k − 1)eo2k−2 (n)x(n)xT (n) c̃(n − 1) + 2µkeo2k−1 (n)x(n)

Taking the expectation of both sides and using the independence assumptions, we have
  
E {c̃(n)} & I − 2µk(2k − 1)E eo2k−2 (n) R E {c̃(n − 1)}

(c) Show that the derived algorithm converges in the mean if

1
0 < 2µ <
k(2k − 1)E{eo2(k−1) (n)}λmax
where λmax is the largest eigenvalue of R.
Using R = QQT and v(n) , QT E {c̃(n)}, the last equation becomes
  
v(n) & I − 2µk(2k − 1)E eo2k−2 (n)  v(n − 1)

or
  
vi (n) & 1 − 2µk(2k − 1)E eo2k−2 (n) λi vi (n − 1)

whose solution is
  n
vi (n) = 1 − 2µk(2k − 1)E eo2k−2 (n) λi vi (0)

This solution converges if

  
1 − 2µk(2k − 1)E e2k−2 (n) λmax  < 1
o

or equivalently
1
0 < 2µ < 
k(2k − 1)E eo2k−2 (n) λmax

(d) Show that for k = 1 the results in parts (a) to (c) reduce to those for the standard LMS algorithm.
For k = 1 we have
1
0 < 2µ <
λmax
which is identical to (10.4.19).

10.15 Consider the noise cancellation system shown in Figure 10.6. The useful signal is a sinusoid s(n) = cos(ω0 n +
φ), where ω0 = π/16 and the phase φ is a random variable uniformly distributed from 0 to 2π . The noise
signals are given by v1 (n) = 0.9 v1 (n − 1) + w(n) and v2 (n) = −0.75 v2 (n − 1) + w(n), where the sequences
w(n) are WGN(0, 1).
328 Statistical and Adaptive Signal Processing - Solution Manual

The input signal to the optimum/adaptive filter is v2 (n) while the “desired signal” is y(n) , s(n) + v1 (n).
Hence we will need the following correlations:

σ 2w 16
rv2 (l) = (−0.75)|l| = (−0.75)|l|
1 − (−0.75) 2 7
1 σ 2w 1 100
r y (l) = rs (l) + rv1 (l) = cos ω0l + (0.9)|l| = cos ω0l + (0.9)|l|
2 1 − (0.9) 2 2 19
r yv2 (l) = rv1 v2 (l) = h 1 (l) ∗ h 2 (−l) ∗ rw (l) = h 1 (l) ∗ h 2 (−l)

where h 1 (n) = (0.9)n u(n), h 2 (n) = (−0.75)n u(n), and rw (l) = δ(l).

(a) Design an optimum filter of length M and choose a reasonable value for Mo by plotting the MMSE as a
function of M.
The optimum filter of length M is given by

Rv2 co = r yv2

where Rv2 is an (M × M) autocorrelation matrix and r yv2 is an (M × 1) crosscorrelation vector. The

MMSE is given by

Po = r y (0) − coT r yv2

Thus the only quantity remains to be computed is r yv2 (l). Consider, for l ≥ 0
∞

r yv2 (l) = h 1 (l) ∗ h 2 (−l) = h 1 (k) h 2 (k − l)
k=−∞
∞
 ∞

−l
= (0.9) (−0.75)
k k−l
= (−0.75) (−0.675)k
k=l k=l
∞

= (−0.75)−l (−0.675)l (−0.675)k = 0.59701 (0.9)l , l ≥ 0
k=0

The rest of the problem is solved using Matlab. The script is given below and the plot is shown in
Figure 10.15a.
% System Parameters
omg0 = pi/16; a1 = 0.9; a2 = -0.75; var_w = 1;

% (a) Design of the Optimum filter and its MMSE

Mmax = 50; Po = zeros(Mmax,1);
var_v2 = var_w/(1-a2*a2);
var_v1 = var_w/(1-a1*a1);
ry0 = 0.5*cos(omg0*0)+var_v1;
for M = 1:Mmax
Rv2 = toeplitz(var_v2*a2.^[0:M-1]);
ryv2 = (1/(1-a1*a2))*a1.^[0:M-1]’;
co = Rv2\ryv2;
Po(M) = ry0 - co’*ryv2;
end
Po = [ry0;Po]; M = 0:Mmax;
Statistical and Adaptive Signal Processing - Solution Manual 329

MMSE of the Noise Canceller

6

MMSE
3

1
0.5
0
0 10 20 30 40 50
Filter Length M

Figure 10.15a: Plot of the MSE vs M

Hf_1 = figure(’units’,SCRUN,’position’,SCRPOS,...
’paperunits’,PAPUN,’paperposition’,[0,0,5,2],...
’NumberTitle’,’off’,’Name’,’Pr1015a’);
plot([0,Mmax],[0.5,0.5],’w:’,[20,20],[0,Po(21)],...
’w:’,M,Po,’g’);
axis([0,Mmax,0,6]);
xlabel(’Filter Length M’,’fontsize’,label_fontsize);
title(’MMSE of the Noise Canceller’,’fontsize’,title_fontsize);
ylabel(’MMSE’,’fontsize’,label_fontsize);
set(gca,’xtick’,[0:10:Mmax],’ytick’,[0,0.5,1:6]);

% Choose Mo = 20;
Mo = 20;
Rv2 = toeplitz(var_v2*a2.^[0:Mo-1]);
ryv2 = (1/(1-a1*a2))*a1.^[0:Mo-1]’;
co = Rv2\ryv2;
From the plot in Figure 10.15a a reasonable value for Mo is 20.
(b) Design an LMS filter with Mo coefficients and choose the step size µ to achieve a 10 percent misadjustment.

% (b) LMS for Noise canceller

% Generate Desired and Input signals
N = 1000; n = [0:N-1]’;
phi = 2*pi*rand(1,1); s = cos(omg0*n+phi);
w = randn(N,1); v1 = filter(1,[1,-a1],w);
v2 = filter(1,[1,-a2],w); y = s + v1;
mu = 0.005;
c0 = zeros(Mo,1);
[c,e,yhat] = firlms(v2,y,mu,Mo,c0);
(c) Plot the signals s(n), s(n) + v1 (n), v2 (n), the clean signal eo (n) using the optimum filter, and the clean
signal elms (n) using the LMS filter, and comment upon the obtained results.
The script is given below and various plots are shown in Figure 10.15c. From these plots we observe that
the optimum as well as adaptive filters show a reasonable performance.
330 Statistical and Adaptive Signal Processing - Solution Manual

eo = y - filter(co,1,v2); elms = y - yhat;

subplot(’position’,[0.1,0.82,0.85,0.09]);
plot(n(500:1000),s(500:1000),’g’); axis([N/2,N,-5,5]);
title(’Original Signal s(n)’,’fontsize’,title_fontsize);
set(gca,’xtick’,[500,1000]);
subplot(’position’,[0.1,0.64,0.85,0.09]);
plot(n(500:1000),y(500:1000),’m’); axis([N/2,N,-10,10]);
title(’Primary Signal s(n)+v_1(n)’,’fontsize’,title_fontsize);
set(gca,’xtick’,[500,1000]);
subplot(’position’,[0.1,0.46,0.85,0.09]);
plot(n(500:1000),v2(500:1000),’m’); axis([N/2,N,-10,10]);
title(’Secondary Signal v_2(n)’,’fontsize’,title_fontsize);
set(gca,’xtick’,[500,1000]);
subplot(’position’,[0.1,0.28,0.85,0.09]);
plot(n(500:1000),eo(500:1000),’g’); axis([N/2,N,-5,5]);
title(’Clean Signal e_o(n)’,’fontsize’,title_fontsize);
set(gca,’xtick’,[500,1000]);
subplot(’position’,[0.1,0.10,0.85,0.09]);
plot(n(500:1000),elms(500:1000),’g’); axis([N/2,N,-5,5]);
title(’Clean Signal e_{LMS}(n)’,’fontsize’,title_fontsize);
set(gca,’xtick’,[500,1000]);
xlabel(’sample index n’,’fontsize’,label_fontsize);

10.16 A modification of the LMS algorithm, known as the momentum LMS (MLMS), is defined by

c(n) = c(n − 1) + 2µe∗ (n)x(n) + α[c(n − 1) − c(n − 2)]

where |α| < 1 (Roy and Shynk 1990).

(a) Rewrite the previous equation to show that the algorithm has the structure of a low-pass (0 < α < 1) or
a high-pass (−1 < α < 0) filter.
The coefficient updating equation can be written as

[c(n) − c(n − 1)] = α [c(n − 1) − c(n − 2)] + 2µx(n)e∗ (n)

which is a low-pass filter for 0 < α < 1 and a high-pass filter for −1 < α < 0. The structure for this
equation is given in Figure 10.16a.
(b) Explain intuitively the effect of the momentum term α[c(n − 1) − c(n − 2)] on the filter’s convergence
behavior.
The term [αc(n − 1) − c(n − 2)] = αc(n−1) adds an extra amount of change to the LMS updating term
2µx(n)e∗ (n). During the transient phase )c(n − 1)) is large and accelerates convergence. However,
the price paid is a larger amount of misadjustement in steady-state. To overcome this problem we can set
α = 0, that is we switch to the LMS algorithm, when )c(n − 1)) becomes small.
(c) Repeat the computer equalization experiment in Section 10.4.4, using both the LMS and the MLMS
algorithms for the following cases, and compare their performance:
i. W = 3.1, µlms = µmlms = 0.01, α = 0.5.
ii. W = 3.1, µlms = 0.04, µmlms = 0.01, α = 0.5.
iii. W = 3.1, µlms = µmlms = 0.04, α = 0.2.
Statistical and Adaptive Signal Processing - Solution Manual 331

Original Signal s(n)

5

−5
500 1000

Primary Signal s(n)+v 1(n)

10

−10
500 1000

Secondary Signal v2(n)

10

−10
500 1000

Clean Signal eo(n)

5

−5
500 1000

Clean Signal eLMS(n)

5

−5
500 1000
sample index n

Figure 10.15c: Signal plots

2µx(n)e∗ (n) c(n) c(n − 1) c(n − 2)

z −1 z −1

-
α
Figure 10.16a: The momentum LMS algorithm.
332 Statistical and Adaptive Signal Processing - Solution Manual

iv. W = 4, µlms = µmlms = 0.03, α = 0.3.

10.17 In Section 10.4.5 we presented the leaky LMS algorithm [see (10.4.88)]

c(n) = (1 − αµ)c(n − 1) + µe∗ (n)x(n)

where 0 < α , 1 is the leakage coefficient.

(a) Show that the coefficient updating equation can be obtained by minimizing

P(n) = |e(n)|2 + α)c(n))2

We have

P(n) = |e(n)|2 + α )c)2 = e(n)e∗ (n) + αc H c

Using results from Appendix B we obtain

∇ c P(n) = −2x(n)e∗ (n) + 2ac

Hence the a priori LMS coefficient updating is

 
c(n) = c(n − 1) − µ −2x(n)e∗ (n) + 2ac(n − 1)
= (1 − α2µ)c(n − 1) + 2µx(n)e∗ (n)
, (1 − αµ)c(n − 1) + µx(n)e∗ (n)
if we replace 2µ by µ.
(b) Using the independence assumptions, show that

E{c(n)} = [I − µ(R + αI)]E{c(n − 1)} + µd

where R = E{x(n)x H (n)} and d = E{x(n)y ∗ (n)}.

Taking expectations of both sides in the last equation, we have

E {c(n)} = (1 − αµ)E {c(n − 1)} + µE x(n)e∗ (n)

Using

e(n) = y(n) − cT (n − 1)x(n)

we obtain
 
E {c(n)} = (1 − αµ)E {c(n − 1)} + µE x(n)y ∗ (n) − µE x(n)x H (n)c(n − 1)

or using the independence assumptions

E {c(n)} = (1 − αµ)E {c(n − 1)} + µd − µRE {c(n − 1)}


where d = E {x(n)y ∗ (n)} and R = E x(n)x H (n) . Thus

E {c(n)} = [I − µ(R + αI)] E {c(n − 1)} + µd

Statistical and Adaptive Signal Processing - Solution Manual 333

(c) Show that if 0 < µ < 2/(α + λmax ), where λmax is the maximum eigenvalue of R, then

lim E{c(n)} = (R + αI)−1 d

n→∞

that is, in the steady state E{c(∞)}  = co = R−1 d.

The difference equation for E {c(n)} has a finite solution as n → ∞ if the eigenvalues of [I − µ(R + αI)]
are less than one in magnitude, that is, if
2
0<µ<
α + λmax
In this case, as n → ∞, E {c(n)} & E {c(n − 1)} and the solution of the difference equation is

lim E {c(n)} = (R + αI)−1 d

n→∞

which is identical with the optimum filter only when α = 0.

If the input x(n) is replaced by z(n) = x(n) + v(n), where v(n) is uncorrelated white noise with zero
mean and variance α, we have

Rz = R + αI

Therefore, the leaky LMS algorithm is equivalent to the LMS algorithm with white noise of variance α
added to its input.

10.18 There are various communications and speech signal processing applications that require the use of filters with
linear phase (Manolakis et al. 1984). For simplicity, assume that m is even.

(a) Derive the normal equations for an optimum FIR filter that satisfies the constraints
(lp)
i. cm = Jc(lp)
m (linear phase)
(cgd)
ii. cm = −Jc(cgd)
m (constant group delay).
(lp) (cgd)
(b) Show that the obtained optimum filters can be expressed as cm = 21 (cm + Jcm ) and cm = 12 (cm − Jcm ),
where cm is the unconstrained optimum filter.
(c) Using the results in part (b) and the algorithm of Levinson, derive lattice-ladder structure for the con-
strained optimum filters.
(lp)
(d) Repeat parts (a), (b), and (c) for the linear predictor with linear phase, which is specified by am = Ja(lp)
m .
(lp)
(e) Develop an LMS algorithm for the linear-phase filter cm = Jc(lp)
m and sketch the resulting structure. Can
you draw any conclusions regarding the step size and the misadjustment of this filter compared to those
of the unconstrained LMS algorithm?

To be completed.

10.19 In this problem, we develop and analyze by simulation an LMS-type adaptive lattice predictor introduced in
Griffiths (1977). We consider the all-zero lattice filter defined in (7.5.7), which is completely specified by the
lattice parameters {km }0M−1 . The input signal is assumed wide-sense stationary.

(a) Consider the cost function

Pmf b = E{|emf (n)|2 + |emb (n)|2 }

334 Statistical and Adaptive Signal Processing - Solution Manual

which provides the total prediction error power at the output of the mth stage, and show that

∂ Pmf b
= 2E{emf∗ (n)em−1
b
(n − 1) + em−1
f∗
(n)emb (n)}
∂km
Consider the lattice recursions
f ∗
emf (n) = em−1 (n) + km−1 b
em−1 (n − 1)
f
emb (n) = km−1 em−1 (n) + em−1
b
(n − 1)

If we assume real-valued signals, we have

fb
 f

∂ Pm ∂em (n) ∂emb (n)
= E 2em (n)
f
+ 2em (n)
b
∂km−1 ∂km−1 ∂km−1
" #
f
= E 2emf (n)em−1
b
(n − 1) + 2emb (n)em−1 (n)

For complex-valued signals, we follow Appendix B to obtain

∂ Pm
fb " #
f∗ f∗
∗
= E 2e (n)e b
(n − 1) + 2e b
(n)e (n)
∂km−1 m m−1 m m−1

(b) Derive the updating formula

km (n) = km (n − 1) + 2µ(n)[emf∗ (n)em−1

b
(n − 1) + em−1
f∗
(n)emb (n)]

where the normalized step size µ(n) = µ̄/E m−1

b
(n) is computed in practice by using the formula
b
E m−1 (n) = α E m−1
b
(n − 1) + (1 − α)[|em−1
f
(n)|2 + |em−1
b
(n − 1)|2 ]

where 0 < α < 1. Explain the role and proper choice of α, and determine the proper initialization of the
algorithm.
An LMS-type algorithm is obtained by using the instantaneous gradient. Therefore
$ %
f∗
km−1 (n) = km−1 (n − 1) − 2µ(n) emf ∗ (n)em−1
b
(n − 1) + emb (n)em−1 (n)

The quantity E m−1 (n) represents the total energy of the forward and backward linear prediction errors at
the input of the mth stage from start to present time n. The updating formula for E m−1 (n) is a first-order
recursive low-pass filter with a pole at z = a. It provides a filter with exponential memory to help track
in a nonstationary SOE.
Statistical and Adaptive Signal Processing - Solution Manual 335

The algorithm is summarized as follows

Initialization
emf (−1) = emb (−1) = 0
E m−1 (−1) = δ > 0
km (−1) = 0
For n = 0, 1, 2, ...
f
e0 (n) = e0b (n) = x(n)
For m = 0, 1, ..., M − 1
f ∗
emf (n) = em−1 (n) + km−1 b
em−1 (n − 1)
f
emb (n) = km−1 em−1 (n) + em−1
b
(n − 1)
$ %
f∗
E m−1 (n) = α E m−1 (n − 1) + (1 − α) emf ∗ (n)em−1
b
(n − 1) + emb (n)em−1 (n)
 2  
µ̃  f  2
km−1 (n) = km−1 (n − 1) − 2 
em−1 (n) + em−1 (n − 1)
b 
E m−1 (n)
(c) Write a Matlab function to implement the derived algorithm, and compare its performance with that of
the LMS algorithm in the linear prediction problem discussed in Example 10.4.1.

To be completed.

10.20 Consider a signal x(n) consisting of a harmonic process plus white noise, that is,

x(n) = A cos(ω1 n + φ) + w(n)

where φ is uniformly distributed from 0 to 2π and w(n) ∼ WGN(0, σ 2w ).

(a) Determine the output power σ 2y = E{y 2 (n)} of the causal and stable filter
∞

y(n) = h(k)x(n − k)
k=0

and show that we can cancel the harmonic process using the ideal notch filter

1 ω = ω1
H (e j ω ) =
0 otherwise

Is the obtained ideal notch filter practically realizable? That is, is the system function rational? Why?
The output power of the filter is

1 π  
E{y (n)} =
2
σ 2y =  H (e j ω )2 Rx (e j ω )dω
2π −π

where
1 2 j (ω−ω1 ) 1
Rx (e j ω ) = A δ(e ) + A2 δ(e j (ω+ω1 ) ) + σ 2w
2 2
Therefore

 2 1 π  
σ 2y = A  H (e j ω1 ) + σ 2w
2  H (e j ω )2 dω
2π −π
336 Statistical and Adaptive Signal Processing - Solution Manual

and the ideal notch filter is given by

&
  0, ω = ±ω1
 H (e j ω ) =
1, otherwise
As a result of the nature of this discontinuity the ideal notch filter cannot be realized by a rational system
function. The output power of the filter is
&
σ 2w , ω = ±ω1
σy =
2
A + σ 2w , otherwise
2

(b) Consider the second-order notch filter (see Problem 2.31)

D(z) 1 + a z −1 + z −2 D(z)
H (z) = = =
A(z) 1 + aρ z −1 + ρ 2 z −2 D(z/ρ)
where −1 < ρ < 1 determines the steepness of the notch and a = 2 cos ω0 its frequency. We fix ρ, and
we wish to design an adaptive filter by adjusting a.
i. Show that for ρ & 1, σ 2y = A2 |H (e j ω1 )|2 + σ 2w , and plot σ 2y as a function of the frequency ω0 for
ω1 = π/6.
ii. Evaluate dσ 2y (a)/da and show that the minimum of σ 2y (a) occurs for a = −2 cos ω1 .
We can easily find that H (z) has zeros on the unit circle at z 1,2 = e± j ω0 and poles at p1,2 = ρe± j ω0 .
From the pole-zero plot we can see that
&
  0 ω = ±ω0
 H (e j ω ) &
1, otherwise
The steepness of the notch increases as ρ approaches one. Therefore, if ω0 = ω1 we have σ 2y =
 2  2
A2  H (e j ω1 ) + σ 2 since the area under  H (e j ω ) is close to one. This is illustrated in Figure 10.20b.
w

|H(ejω)| for various values of ρ

1.4
0.9
1.2 0.95
0.99
0.999
1

0.8
|H(ejω)|

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5
ω/2π
Figure 10.20b: Magnitude response for pole angle ω0 = 2π (0.2) and various radius.

To find the value of a that minimizes σ 2y we compute

 
∂σ 2y (a) A2 (a + 2 cos ω1 ) 1 − ρ 3 + aρ(1 − ρ) cos ω1 − ρ(1 − ρ) cos 2ω1
=  
∂a 2  A(e j ω1 )4
Statistical and Adaptive Signal Processing - Solution Manual 337

This is zero if a = −2 cos ω1 or

1 − ρ 3 + aρ(1 − ρ) cos ω1 − ρ(1 − ρ) cos 2ω1 = 0

For −2 < a < 2 we can show that

1 − ρ 3 + aρ(1 − ρ) cos ω1 − ρ(1 − ρ) cos 2ω1 ≥ (1 − ρ 2 ) > 0

which implies that a = −2 cos ω1 is the only solution.

(c) Using a direct-form II structure for the implementation of H (z) and the property

dY (z)/da = [d H (z)/da]X (z)

show that the following relations

s2 (n) = −a(n − 1)ρs2 (n − 1) − ρ 2 s2 (n − 2) + (1 − gr)s1 (n − 1)

g(n) = s2 (n) − ρs2 (n − 2)
s1 (n) = −a(n − 1)ρs1 (n − 1) − ρ 2 s1 (n − 2) + x(n)
y(n) = s1 (n) + a(n − 1)s1 (n − 1) + s1 (n − 2)
a(n) = a(n − 1) − 2µy(n)g(n)

constitute an adaptive LMS notch filter. Draw its block diagram realization.
The instantaneous gradient is

∂ y 2 (n) ∂ y(n)
= 2y(n)
∂a ∂a
which leads to the following LMS updating

a(n + 1) = a(n) − µy(n)g(n)

where
∂ y(n)
g(n) =
∂a
We shall compute g(n) in the z-domain. From Y (z) = H (z)X (z) we have

∂Y (z) ∂ H (z)
= X (z)
∂a ∂a
and
∂ H (z) ∂ 1 + az −1 + z −2 (1 − ρ)z −1 (1 − ρz −2 )
= =  2
∂a ∂a 1 + aρz −1 + ρ 2 z −2 1 + aρz −1 + ρ 2 z −2

The direct form II implementations for the filter and the gradient are
 
 −1 −2
X (z)
Y (z) = 1 + az + z
1 + aρz −1 + ρ 2 z −2
  
S1 (z)
338 Statistical and Adaptive Signal Processing - Solution Manual

and
 
−2 (1 − ρ)z −1 S1(z)
G(z) = (1 − ρz )
1 + aρz −1 + ρ 2 z −2
  
S2 (z)

respectively. The corresponding difference equations are

s2 (n) = −a(n − 1)ρs2 (n − 1) − ρ 2 s2 (n − 2) + (1 − gr)s1 (n − 1)
g(n) = s2 (n) − ρs2 (n − 2)
s1 (n) = −a(n − 1)ρs1 (n − 1) − ρ 2 s1 (n − 2) + x(n)
y(n) = s1 (n) + a(n − 1)s1 (n − 1) + s1 (n − 2)
a(n) = a(n − 1) − 2µy(n)g(n)
and provide the complete algorithm for the implementation of the second-order adaptive notch filter. The
resulting structure is shown in Figure 10.20c.

x(n) s1 (n) y(n)

+ +

z −1

+ +
−a(n)ρ a(n)
a(n + 1)
z −1 −µ
x +

−ρ 2 z −1
1−ρ

s2 (n)
+ +

g(n)
z −1

+
−a(n)ρ
z −1

−ρ 2 −ρ
Figure 10.20c: Block diagram implementation of the adaptive notch filter.

(d) Simulate the operation of the obtained adaptive filter for ρ = 0.9, ω1 = π/6, and SNR 5 and 15 dB. Plot
ω0 (n) = arccos[−a(n)/2] as a function of n, and investigate the tradeoff between convergence rate and
misadjustment by experiment with various values of µ.
The MATLAB script for simulating this experiment is given below and Figures 10.20d1 and 10.20d2 show
filter input-output signals and frequency tracking for f 0 = 0.2, SNR=5dB, µ = 0.001 and µ = 0.0025.
The trade-off between speed of convergence and misadjustement is clearly evident.
Statistical and Adaptive Signal Processing - Solution Manual 339

µ=0.001
10
Input signal
5 Output signal

−5

−10
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
n
0.3
"Uknown" frequency
Adaptive tracking
0.25
f0(n)

0.2

0.15
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
n

Figure 10.20d1: Tracking characteristics for µ = 0.001.

µ=0.0025
10
Input signal
5 Output signal

−5

−10
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
n
0.3
"Uknown" frequency
Adaptive tracking
0.25
f0(n)

0.2

0.15
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
n

Figure 10.20d2: Tracking characteristics for µ = 0.0025.

340 Statistical and Adaptive Signal Processing - Solution Manual

10.21 Consider the AR(2) process given in Problem 10.4. We will design the adaptive linear predictor using the RLS
algorithm. The adaptive predictor is a second-order one given by c(n) = [c1 (n) c2 (n)]T .

(a) Develop a Matlab function to implement the RLS algorithm given in Table 10.6:
[c,e] = rls(x,y,lambda,delta,M,c0).
which computes filter coefficients in c and the corresponding error in e given signal x, desired signal y,
forgetting factor lambda, initialization parameter delta, filter order M, and the initial coefficient vector
c0. To update P(n), compute only the upper or lower triangular part and determine the other part by
using Hermitian symmetry.
(b) Generate 500 samples of x(n) and obtain linear predictor coefficients using the above function. Use a
very small value for δ (for example, 0.001) and various values of λ = 0.99, 0.95, 0.9, and 0.8. Plot
predictor coefficients as a function of time along with the true coefficients for each λ, and discuss your
observations. Also compare your results with those in Problem 10.4.
(c) Repeat each simulation above 1000 times to get corresponding learning curves, which are obtained by
averaging respective squared errors |e(n)|2 . Plot these curves and compare their steady-state value with
the theoretical MSE.

To be completed.
10.22 Consider a system identification problem where we observe the input x(n) and the noisy output y(n) =
yo (n) + v(n), for 0 ≤ n ≤ N − 1. The unknown system is specified by the system function
0.0675 + 0.1349z −1 + 0.0675z −2
Ho (z) =
1 − 1.1430z −1 + 0.4128z −2
and x(n) ∼ WGN(0, 1), v(n) ∼ WGN(0, 0.01), and N = 300.

(a) Model the unknown system using an LS FIR filter, with M = 15 coefficients, using the no-windowing
method. Compute the total LSE E ls in the interval n 0 ≤ n ≤ N − 1 for n 0 = 20.
(b) Repeat part (a) for 0 ≤ n ≤ n 0 − 1 (do not compute E ls ). Use the vector c(n 0 ) and the matrix P(n 0) =
! N−1 2
R̂−1 (n 0 ) to initialize the CRLS algorithm. Compute the total errors E apr = n=n e (n) and E apost =
! N−1 2 0

n=n 0 ε (n) by running the CRLS for n 0 ≤ n ≤ N − 1.

(c) Order the quantities E ls , E apr , E apost by size and justify the resulting ordering.

To be completed.
10.23 Prove Equation (10.5.25) using the identity det(I1 + AB) = det(I2 + BA), where identity matrices I1 and I2
and matrices A and B have compatible dimensions.
We have

R̂(n) = λR̂(n − 1) + x(n)x H (n)

λR̂(n − 1) = R̂(n) − x(n)x H (n)
$ %
λR̂−1 (n)R̂(n − 1) = I − R̂−1 (n)x(n) x H (n)

Taking the determinant of both sides gives

det R̂(n − 1) " $ % #

λM = det I − R̂−1 (n)x(n) x H (n)
det R̂(n)
$ %
= det 1 − x H (n)R̂−1 (n)x(n) = α(n)
Statistical and Adaptive Signal Processing - Solution Manual 341

10.24 Derive the normal equations that correspond to the minimization of the cost function (10.5.36), and show that
for δ = 0 they are reduced to the standard set (10.5.2) of normal equations. For the situation described in
Problem 10.22, run the CRLS algorithm for various values of δ and determine the range of values that provides
acceptable performance.
The cost function can be written as
E(n) = δλn c H c + E y (n) − c H d̂(n) − d̂ H (n)c + c H R̂(n)c
Computing the gradient with respect to c gives
∂ E(n) $ %∗
= δλn c∗ − d̂∗ (n) + R̂(n)c = 0
∂c
or
$ %
δλn + R̂(n) c(n) = d̂(n)

For δ = 0, we obtain the standard normal equations. It can be easily shown that the matrix R̃(n) = δλn + R̂(n)
obeys the recursion
R̃(n) = λR̃(n − 1) + x(n)x H (n)
Thus, the development of the CRLS algorithm remains unchanged. Clearly, small values of δ do not significantly
affect the transient response of the filter and have no effect on the long run due to the exponential forgetting
memory.
10.25 Modify the CRLS algorithm in Table 10.6 so that its coefficients satisfy the linear-phase constraint c = Jc∗ .
For simplicity, assume that M = 2L; that is, the filter has an even number of coefficients. To be completed.
10.26 Following the approach used in Section 7.5.1 to develop the structure shown in Figure 7.1, derive a similar
structure based on the Cholesky (not the LDLH ) decomposition.
The LS filter is specified by R̂(n)c(n) = d̂(n). Using the Cholesky decomposition R̂(n) = L̃m (n)L̃mH (n) and
the definitions in Section 10.6.1, we obtain
em (n) = y(n) − ŷm (n) = y(n) − k̃mH (n − 1)w̃m (n)
or
ŷm (n) = ŷm−1 (n) + k̃m∗ (n − 1)w̃m (n)
Using the last equation and the elements of the triangular matrix B̃m (n) , L̃−1
m (n), we can easily modify Figure
7.1 to obtain the desired structure.
10.27 Show that the partitioning (10.7.3) of R̂m+1 (n) to obtain the same partitioning structure as (10.7.2) is possible
only if we apply the prewindowing condition xm (−1) = 0. What is the form of the partitioning if we abandon
the prewindowing assumption?
We have

n
R̂m+1 (n) = λn− j xm+1 ( j )xm+1
H
( j)
j =0
 

n
xm (i)  
= λ n− j
xmH ( j ) x ∗ ( j − m)
j =0
x( j − m)
 
R̂m (n) r̂mb (n)
=
r̂mbH (n) ρ̂ bm (n)
342 Statistical and Adaptive Signal Processing - Solution Manual

where

n
r̂mb (n) = λn− j xm ( j )x ∗ ( j − m)
j =0

n
ρ̂ bm (n) = λn− j |x( j − m)|2
j =0

Similarly, using the partitioning

 
x( j )
xm+1 ( j ) =
xm ( j − 1)

we have
 f 
ρ̂ m (n) r̂mf H (n)
R̂m+1 (n) =
r̂mf (n) R̂m (n − 1)

where

n
λn− j xm ( j − 1)xmH ( j − 1) = λn xm (−1)xmH (−1) + R̂m (n − 1)
  
j =0
pre-windowing ⇒0

n
r̂mf (n) = λn− j xm ( j − 1)x ∗ ( j )
j =0

n
ρ̂ fm (n) = λn− j |x( j )|2
j =0

10.28 Derive the normal equations and the LSE formulas given in Table 10.11 for the FLP and the BLP methods.
For the FLP, the error is

ε mf (n) = x(n) + amH (n)xm (n − 1)

and the cost function is


n
E mf (n) = λn− j |ε mf ( j )|2
j =0

With the goal to minimize the cost function, the normal equations can be found directly


n
E mf (n) = λn− j |x(n) + amH (n)xm (n − 1)|2
j =0

n 
n
= λn− j x(n)x ∗ (n) + λn− j amH (n)xm (n − 1)x ∗ (n)
j =0 j =0

n 
n
+ λ n− j
x(n)xmH (n − 1)am (n) + λn− j amH (n)xm (n − 1)xmH (n − 1)am (n)
j =0 j =0

= E x (n) + amH (n)r̂mf (n) + r̂mf H (n)am (n) + amH (n)R̂m (n − 1)am (n)
Statistical and Adaptive Signal Processing - Solution Manual 343

In order to minimize the above cost function, the normal equations are

R̂m (n − 1)am (n) = −r̂mf (n)

and the least square error (LSE) is

E mf (n) = E x (n) + amH (n)r̂mf (n)

Similarly for the BLP, the error is

ε bm (n) = x(n − m) + bmH (n)xm (n)

and the cost function is


n
E mb (n) = λn− j |ε bm ( j )|2
j =0


n
E mb (n) = λn− j |x(n − m) + bmH (n)xm (n)|2
j =0

n 
n
= λn− j x(n − m)x ∗ (n − m) + λn− j bmH (n)xm (n)x ∗ (n − m)
j =0 j =0

n 
n
+ λ n− j
x(n − m)xmH (n)bm (n) + λn− j bmH (n)xm (n)xmH (n)bm (n)
j =0 j =0

= E x (n − m) + bmH (n)r̂mb (n) + r̂mb H (n)bm (n) + bmH (n)R̂m (n)bm (n)

In order to minimize the above cost function, the normal equations are

R̂m (n)bm (n) = −r̂mb (n)

and the least square error (LSE) is

E mb (n) = E x (n − m) + bmH (n)r̂mb (n)

10.29 Derive the FLP and BLP a priori and a posteriori updating formulas given in Table 10.12.
To be completed.

10.30 Modify Table10.14 for the FAEST algorithm, to obtain a table for the FTF algorithm, and write a Matlab
function for its implementation. Test the obtained function, using the equalization experiment in Example
10.5.2.
To obtain the FTF algorithm, we replace ᾱ m (n) in Table 10.14 by 1/α m (n) and Equation (h) with
f
|ε m (n)|2
α m+1 (n) = α m (n − 1) − f
E m (n)
and Equation (i) with
α m+1 (n)
α m (n) = (m+1)
1 − α m+1 (n)ḡm+1 (n)emb∗ (n)
The Matlab script is shown below and the learning curve in Figure 10.30.
344 Statistical and Adaptive Signal Processing - Solution Manual

%% Problem 10.30
%%
%%
%%

close all
clear

W=2.9;
N=1000;
Nmc=5;
M=11;
lambda=0.99;
varv=0.001;
h=zeros(3,1);
er=zeros(N,Nmc);
e2a = zeros(1,N);

h(1)=0.5*(1+cos(2*pi*(1-2)/W));
h(2)=0.5*(1+cos(2*pi*(2-2)/W));
h(3)=0.5*(1+cos(2*pi*(3-2)/W));

% Learning curves

hc=[0 h(1) h(2) h(3)]’;

n0=7;

t=(1:N)’;

for i=1:Nmc

y=sign(rand(N,1)-0.5);
v=sqrt(varv)*randn(N,1);
x=filter(hc,1,y)+v;
yd=zeros(N,1);
yd(n0:N)=y(1:N-n0+1);

% Time Initialization
aold=zeros(M,1); bold=zeros(M,1); cold=zeros(M,1); xold=zeros(M,1);
alphaold=1; gbarold=zeros(M,1);
Efold=10^3; Ebold=10^3;
xv=zeros(M,1);

for n=1:N

% Order Initialization

%ef(1)=x(n);
Statistical and Adaptive Signal Processing - Solution Manual 345

%eb(1)=x(n);
%e(1)=yd(n);
%alpha(1)=1;
%alphaold=alpha(1);

% Gain and predictor update

ef(n)=x(n) + aold’*xold;
epsilonf(n) = ef(n)*alphaold;
a = aold - gbarold.*conj(epsilonf(n));
Ef(n) = lambda*Efold + epsilonf(n)*conj(ef(n));
gM1 = [0; gbarold] + (ef(n)/(lambda*Efold))*[1; aold];
eb(n) = lambda*Ebold*gM1(end);
gbar = gM1(1:end-1) - gM1(end).*bold;
alphaM1 = alphaold - abs(epsilonf(n))^2/Ef(n);
alpha = alphaM1/(1 - alphaM1*gM1(end)*conj(eb(n)));
epsilonb(n) = eb(n)*alpha;
b = bold - gbar.*conj(epsilonb(n));
Eb(n) = lambda*Ebold + epsilonb(n)*conj(eb(n));

% Filter update
xv=[x(n); xold(1:end-1)];
e(n) = y(n) - cold’*xv;
epsilon(n) = e(n)*alpha;
c = cold + gbar.*conj(epsilon(n));

% Time Update
aold=a; bold=b; cold=c; xold=xv;
alphaold=alpha; gbarold=gbar;
Efold=Ef(n); Ebold=Eb(n);

end
i
e2a=e2a+e.^2;
end

e2a=e2a./Nmc;

t=(1:N)’;
%plot(t,Ef,’r’,t,Eb,’b’);
%text(250,300,’\leftarrow\fontsize{9pt} E^f’);
%text(100,200,’E^b \rightarrow\fontsize{9pt} ’);

Jmin1=1;

subplot(’position’,[0.1,0.55,0.85,0.4]);
semilogy(t,e2a,’r’); %,...
axis([0,N,10^(-4),10^0]);
title(’(a) MSE Learning Curve’,’fontsize’,10);
346 Statistical and Adaptive Signal Processing - Solution Manual

0
(a) MSE Learning Curve
10

−1
10
Mean squared error

−2
10

−3
10

−4
10
0 100 200 300 400 500
Number of iterations (n)

Figure 10.30: Learning curve in Problem P10.30

xlabel(’Number of iterations (n)’,’fontsize’,8);

ylabel(’Mean squared error’,’fontsize’,8);
set(gca,’xtick’,[0:100:500],’fontsize’,8);

print -deps2 P1030.eps

10.31 If we wish to initialize the fast RLS algorithms (fast Kalman, FAEST, and FTF) using an exact method, we
need to collect a set of data {x(n), y(n)}n00 for any n 0 > M.

(a) Identify the quantities needed to start the FAEST algorithm at n = n 0 . Form the normal equations and
use the LDL H decomposition method to determine these quantities.
(b) Write a Matlab function faestexact.m that implements the FAEST algorithm using the exact initial-
ization procedure described in part (a).
(c) Use the functions faest.m and faestexact.m to compare the two different initialization approaches for
the FAEST algorithm in the context of the equalization experiment in Example 10.5.2. Use n 0 = 1.5M
and n 0 = 3M. Which value of δ gives results closest to the exact initialization method?

To be completed.

10.32 Using the order-recursive approach introduced in Section 7.3.1, develop an order-recursive algorithm for the
solution of the normal equations (10.5.2), and check its validity by using it to initialize the FAEST algorithm,
as in Problem 10.31. Note: In Section 7.3.1 we could not develop a closed-form algorithm because some
recursions required the quantities bm (n − 1) and E mb (n − 1). Here we can avoid this problem by using time
recursions.
Section 7.3.1 provides an incomplete order recursive algorithm for computing the optimum filter. The algorithm
can be completed by using time updatings to determine b(n − 1) and E mb (n − 1). These can be computed using
Statistical and Adaptive Signal Processing - Solution Manual 347

time recursions from Table 10.12. The complete algorithm is as follows

Time recursions
ε bm (n) = x(n − m) + bmH (n)xm (n)
bm (n) = bm (n − 1) − ḡm (n)ε b∗
m (n)
emb (n) = ε bm (n)ᾱ m (n)
E mb (n) = λE mb (n − 1) + emb (n)εb∗
m (n)
Order recursions
β m (n) = bmH (n − 1)rmf (n) + rm+1
f
(n)
β cm (n) = bmH (n)dm (n) + dm+1 (n)
β (n)
kmf (n) = − b m
E m (n − 1)
∗
β (n)
kmb (n) = − mf
E m (n)
β (n)
c
kmc (n) = mb
E (n)
m   
am (n) bm (n − 1)
am+1 (n) = + kmf (n)
0 1
   
0 1
bm+1 (n) = + kmb (n)
bm (n − 1) am (n)
   
cm (n) bm (n)
cm+1 (n) = + kmc (n)
0 1
f
E m+1 (n) = E mf (n) + β ∗m (n)kmf (n)
b
E m+1 (n) = E mb (n − 1) + β m (n)kmb (n)
   
ḡm (n) emb (n) bm (n − 1)
ḡm+1 (n) = +
0 λE mb (n − 1) 1
 b 
e (n)2
m
ᾱ m+1 (n) = ᾱ m (n) +
λE mb (n − 1)
The last two equations are obtained from (10.7.17) and (10.7.24), respectively.
10.33 In this problem we discuss several quantities that can serve to warn of ill behavior in fast RLS algorithms for
FIR filters.

(a) Show that the variable

α m+1 (n) λE mb (n − 1)
ηm (n) , = (m+1)
= 1 − gm+1 (n)emb∗ (n)
α m (n) E mb (n)
satisfies the condition 0 ≤ ηm (n) ≤ 1.
Since 0 < λ ≤ 1 and E mb (n) ≥ 0 ⇒ ηm (n) ≥ 0. From (10.5.24) we have 0 < α m (n) ≤ 1 and from Table
10.12(j) equation
 2
E mb (n) = λE mb (n − 1) + α m (n) emb (n)
which implies that E mb (n) ≥ λE mb (n − 1). Therefore we can see that 0 ≤ ηm (n) ≤ 1.
348 Statistical and Adaptive Signal Processing - Solution Manual

(b) Prove the relations

det R̂m (n − 1) det R̂m+1 (n) det R̂m+1 (n)
α m (n) = λm E mf (n) = E mb (n) =
det R̂m (n) det R̂m (n − 1) det R̂m (n)
We have
R̂(n) = λR̂(n − 1) + x(n)x H (n)
λR̂(n − 1) = R̂(n) − x(n)x H (n)
$ %
λR̂−1 (n)R̂(n − 1) = I − R̂−1 (n)x(n) x H (n)

Taking the determinant of both sides gives

det R̂(n − 1) " $ % #
λM = det I − R̂−1 (n)x(n) x H (n)
det R̂(n)
$ %
= det 1 − x H (n)R̂−1 (n)x(n) = α(n)

Using the identities

 
A11 A12 
det = det A11 det A22 − A21 A−1
11 A12
A21 A22

= det A22 det A11 − A12 A−1
22 A21

in conjunction with (10.7.2) and (10.7.3) we can easily obtain the desired results.
(c) Show that
E mb (n)
α m (n) = λm
E mf (n)
and use it to explain why the quantity ηαm (n) = E mf (n) − λm E mb (n) can be used as a warning variable.
We have
det R̂m+1 (n)
det R̂m (n − 1) f
E m (n) E b (n)
α m (n) = λm = λm = λm mf
det R̂m (n) det R̂m+1 (n) E m (n)
E m (n)
b

Hence
ηαm (n) = E mf (n) − α m (n)E mf (n) = [1 − α m (n)] E mf (n)
and since 0 ≤ α m (n) ≤ 1 we should have, in theory, that ηαm (n) ≥ 0. Therefore, we can monitor ηαm (n)
and issue a warning when it becomes negative.
(d) Explain how the quantities
eb (n)
η ḡ (n) , ḡ M+1
(M+1)
(n) − , and
λE b (n − 1)

η b (n) , eb (n) − λE b (n − 1) ḡ (M+1)

M+1 (n)

can be used as warning variables.

From (10.7.19), ηḡ (n) should theoretically be zero. From (10.7.20), ηb (n) should theoretically be zero.
Therefore, the deviation of these quantities from zero can be used as a warning indicator.
Statistical and Adaptive Signal Processing - Solution Manual 349

10.34 When the desired response is y( j ) = δ( j − k), that is, a spike at j = k, 0 ≤ k ≤ n, the LS filter cm(k) is known
as a spiking filter or as an LS inverse filter (see Section 8.3).
NOTE: The subscript m is not necessary and it is dropped for convenience.

(a) Determine the normal equations and the LSE E m(k) (n) for the LS filter cm(k) .
When y( j ) = δ( j − k), 0 ≤ k ≤ n we have


n
R̂(n) = λn− j x( j )x H ( j )
j =0

n
d̂(k) (n) = λn− j x( j )δ( j − k) = λn−k x(k)
j =0

Hence, we have

R̂(n)c(k) (n) = λn−k x(k)

E (k) (n) = λn−k − λn−k x H (k)R̂−1 (n)x(k)λn−k

for the normal equations and the LSE.

(b) Show that cm(n) = gm (n) and E m(n) (n) = α m (n) and explain their meanings.
For k = n we obtain

R̂(n)c(n) (n) = x(k) ⇒ c(n) (n) = g(n) (adaptation gain)

E (n) (n) = 1 − x H (n)R̂−1 (n)x(n) = α(n) (angle variable)

Therefore, the adaptation gain acts like a spiking filter (see Section 8.3).
The LS filter can be expressed as


n n $
 % 
n
c(n) = R̂−1 (n) λn− j x( j )y ∗ ( j ) = λn− j R̂−1 (n)x( j ) y ∗ ( j ) = c( j ) (n)y ∗ ( j )
j =0 j =0 j =0

that is, the LS filter is a linear combination of spiking filters weighted by each individual sample of the
desired response.
(c) Use the interpretation α m (n) = E m(n) (n) to show that 0 ≤ α m (n) ≤ 1.
The worst spiking filter performance corresponds to c(n) (n) = 0 which implies E (n) (n) = 1. Hence,
0 ≤ E (n) (n) = α(n) ≤ 1.
!
(d) Show that am (n) = nk=0 cm(k) (n − 1)x(k) and explain its meaning.
We have

n 
n
a(n) = −R̂−1 (n) λn− j x( j − 1)x ∗ ( j ) = − c( j )(n − 1)x ∗ ( j )
j =0 j =0

therefore the FLP is a linear combination of delayed, by one sampling interval, spiking filters weighted
by the signal.
350 Statistical and Adaptive Signal Processing - Solution Manual

10.35 Derive Equations (10.7.33) through (10.7.35) for the a posteriori LS lattice-ladder structure, shown in Fig-
ure 10.38, starting with the partitionings (10.7.1) and the matrix by inversion by partitioning relations (10.7.7)
and (10.7.8).
Starting with (10.7.1)
   
xm (n) x(n)
xm+1 (n) = =
x(n − m) xm (n − 1)
The FLP error is
f
ε m+1 (n) = x(n) + am+1
H
(n)xm+1 (n − 1)
f
Using (7.3.27), the order-recursive lattice-ladder forward error ε m+1 (n) is found directly

    H
f am (n) bm (n − 1)
ε m+1 (n) = x(n) + + kmf (n) xm+1 (n − 1)
0 1
= x(n) + amH (n)xm (n − 1) + [bmH (n − 1)xm (n − 1) + x(n − 1 − m)]kmf ∗ (n)
= ε mf (n) + kmf ∗ (n)εbm (n − 1)

Similarly for the BLP error

ε bm+1 (n) = x(n − m) + bm+1
H
(n)xm+1 (n)
Using (7.3.23), the backward error ε bm+1 (n) is

    H
0 1
ε m+1 (n) = x(n − m) +
b
+ km (n)
b
xm+1 (n)
bm (n − 1) am (n)
= x(n − m) + bmH (n − 1)xm (n − 1) + [amH (n)xm (n − 1) + x(n)]kmb∗ (n)
= εbm (n − 1) + kmb∗ (n)εmf (n)

Lastly, for the ladder error

ε m+1 (n) = y(n) + cm+1
H
(n)xm+1 (n)
Using (7.3.15), the order-recursive ladder error is

    H
cm (n) bm (n)
ε m+1 (n) = y(n) − + km (n)
c
xm+1 (n)
0 1
= y(n) − cmH (n)xm (n) + [bmH (n)xm (n) + x(n − m)]kmc∗ (n)
= ε m (n) − kmc∗ (n)εbm (n)

10.36 Prove relations (10.7.45) and (10.7.46) for the updating of the ladder partial correlation coefficient β cm (n).
Starting with (10.7.40), and using time updating formulas
β cm (n) = bmH (n)d̂m (n) + d̂m+1 (n)
= bmH (n)[λd̂m (n − 1) + x(n)y ∗ (n)] + [λd̂m+1 (n − 1) + x(n − m)y ∗ (n)]
= λbmH (n)d̂m (n − 1) + εbm (n)y ∗ (n) + λd̂m+1 (n − 1)
= λ[bmH (n − 1) − ε bm (n)ḡm (n)]d̂m (n − 1) + λd̂m+1 (n − 1) + ε bm (n)y ∗ (n)
= λ[bmH (n − 1)d̂m (n − 1) + d̂m+1 (n − 1)] + ε bm (n)[y ∗ (n) − λḡm (n)d̂m (n − 1)
= λβ cm (n − 1) + ε bm (n)[y ∗ (n) − λḡm (n)R̂m (n − 1)cm (n − 1)]
= λβ cm (n − 1) + ε bm (n)[y ∗ (n) − xmH (n)cm (n − 1)]
= λβ cm (n − 1) + ε bm (n)em∗ (n)
Statistical and Adaptive Signal Processing - Solution Manual 351

10.37 In Section 7.3.1 we derived order-recursive relations for the FLP, BLP, and FIR filtering MMSEs.

(a) Following the derivation of (7.3.36) and (7.3.37), derive similar order-recursive relations for E mf (n) and
E mb (n).
f fH
E m+1 (n) = E x (n) + r̂m+1 (n)am+1 (n)

    
f∗ am (n) bm (n − 1)
= E x (n) + [r̂mf H (n) r̂m+1 (n)] + kmf (n)
0 1
f∗
= E x (n) + r̂mf H (n)am (n) + [r̂mf H (n)bm (n − 1) + r̂m+1 (n)]kmf (n)
= E mf (n) + β ∗m (n)kmf (n)

b
E m+1 (n) = E x (n − m − 1) + r̂m+1
bH
(n)bm+1 (n)

    
0 1
= E x (n − m − 1) + [r̂m+1
b∗
(n) r̂mb H (n − 1)] + kmb (n)
bm (n − 1) am (n)
= E x (n − m − 1) + r̂mb H (n − 1)bm (n − 1) + [r̂mb H (n − 1)am (n) + r̂m+1
b∗
(n)]kmb (n)
= E mb (n − 1) + β ∗m (n)kmb (n)
(b) Show that we can obtain a complete LS lattice-ladder algorithm by replacing, in Table 10.15, the time-
recursive updatings of E mf (n) and E mb (n) with the obtained order-recursive relations.
This is the a posteriori LS lattice-ladder algorithm, which is shown in Table 10.15. Replacing the following
steps in Table 10.15 will produce an order-recursive relation
f∗
ε bm (n − 1)ε m (n)
(b) β m+1 (n) = λβ m+1 (n − 1) +
α m (n − 1)
(c) E m+1 (n) = E mf (n) + β ∗m (n)kmf (n)
f

(d) E m+1
b
(n) = E mb (n − 1) + β ∗m (n)kmb (n)
f −β m+1 (n)
(e) km+1 (n) = b
E m+1 (n − 1)
∗
−β m+1 (n)
( f ) km+1
b
(n) = f
E m+1 (n)

(c) Write a Matlab function for this algorithm, and verify it by using the equalization experiment in Example
10.5.2.

10.38 Derive the equations for the a priori RLS lattice-ladder algorithm given in Table 10.16, and write a Matlab
function for its implementation. Test the function by using the equalization experiment in Example 10.5.2.
Using (10.7.1), (7.3.27) and (e) from Table 10.12, the a priori FLP error is
f
em+1 (n) = x(n) + am+1
H
(n − 1)xm+1 (n − 1)

    H  
am (n − 1) bm (n − 1) xm (n − 1)
= x(n) + + km (n − 1)
f
0 1 x(n − m − 1)
= x(n) + amH (n − 1)xm (n − 1) + kmf ∗ (n − 1)(bm−1 (n − 1)xm (n − 1) + x(n − m − 1))
= emf (n) + kmf ∗ (n − 1)(bm−1 (n − 1)xm (n − 1) + x(n − m − 1))
= emf (n) + kmf ∗ (n − 1)emb (n − 1)
352 Statistical and Adaptive Signal Processing - Solution Manual

Similarly for the a priori BLP error

b
em+1 (n) = x(n − m − 1) + bm+1
H
(n − 1)xm+1 (n)

    H  
0 1 x(n)
= x(n − m − 1) + + km (n − 1)
b
bm (n − 2) am (n − 1) xm (n − 1)
= x(n − m − 1) + bmH (n − 2)xm (n − 1) + kmb∗ (n − 1)(am (n − 1)xm (n − 1) + x(n))
= emb (n − 1) + kmb∗ (n − 1)(x(n) + amH (n − 1)xm (n − 1))
= emb (n − 1) + kmb∗ (n − 1)emf (n)

The Matlab function is given below and the MSE learning curve in shown in Figure 10.38.

%% Problem 10.38
%% Implement table 10.16%%
%%
%%
%%
%%

close all
clear all

W=2.9;
N=1000;
Nmc=250;
M=11;
lambda=0.99;
varv=0.001;
h=zeros(3,1);
er=zeros(N,Nmc);

h(1)=0.5*(1+cos(2*pi*(1-2)/W));
h(2)=0.5*(1+cos(2*pi*(2-2)/W));
h(3)=0.5*(1+cos(2*pi*(3-2)/W));

% Learning curves

hc=[0 h(1) h(2) h(3)]’;

n0=7;

t=(1:N)’;

for i=1:Nmc

y=sign(rand(N,1)-0.5);
v=sqrt(varv)*randn(N,1);
x=filter(hc,1,y)+v;
yd=zeros(N,1);
yd(n0:N)=y(1:N-n0+1);
Statistical and Adaptive Signal Processing - Solution Manual 353

% Time Initialization

Ef=ones(M+1,1)*10^(3);
Eb=Ef;
Ebold=Ef; Efold=Ef;
ef=zeros(M+1,1); eb=zeros(M+1,1);
ebold=eb;
kf=zeros(M,1); kb=zeros(M,1);
kfold=kf; kbold=kb;
b=zeros(M+1,1); bc=zeros(M+1,1);
bold=b; bcold=bc;
e=zeros(M+1,1); kc=zeros(M,1);
kcold=kc;
alpha=zeros(M+1,1);
alphaold=alpha;
for n=1:N

% Order Initialization

ef(1)=x(n);
eb(1)=x(n);
e(1)=yd(n);
alpha(1)=1;
alphaold(1)=alpha(1);

for m=1:M
%Lattice
ef(m+1) = ef(m) + conj(kfold(m))*ebold(m);
eb(m+1) = ebold(m) + conj(kbold(m))*ef(m);
%b(m) = lambda*bold(m) + (ebold(m)*conj(ef(m)))*alphaold(m);
Ef(m) = lambda*Efold(m) + alphaold(m)*conj(ef(m))*ef(m);
Eb(m) = lambda*Ebold(m) + alpha(m)*conj(eb(m))*eb(m);
kf(m) = kfold(m) - (alphaold(m)*ebold(m)*conj(ef(m+1)))/Ebold(m);
kb(m) = kbold(m) - (alphaold(m)*ef(m)*conj(eb(m+1)))/Ef(m);
alpha(m+1) = alpha(m) - abs(alpha(m)*eb(m))*abs(alpha(m)*eb(m))/Eb(m);

% Ladder
%bc(m)=lambda*bcold(m) + (eb(m)*conj(ef(m)))*alpha(m);
e(m+1)=e(m)-conj(kcold(m))*eb(m);
kc(m)= kcold(m)+(alpha(m)*eb(m)*conj(e(m+1)))/Eb(m);
end

er(n,i)=e(M+1);erf(n,i)=ef(M);erb(n,i)=eb(M);
Erf(n,i)=Eb(1); Erb(n,i)=Eb(M); al(n,i)=alpha(M);

% Time Delay
Ebold=Eb; Efold=Ef;
ebold=eb; efold=ef;
354 Statistical and Adaptive Signal Processing - Solution Manual

bold=b; bcold=bc;
kfold=kf; kbold=kb; kcold=kc;
alphaold=alpha;
end
i
end
er2=er.^2;
er2m=mean(er2,2);

t=(1:N)’;
%plot(t,Erf,’r’,t,Erb,’b’);
%text(250,300,’\leftarrow\fontsize{9pt} E^f’);
%text(100,200,’E^b \rightarrow\fontsize{9pt} ’);

Jmin1=1;

%subplot(’position’,[0.1,0.55,0.85,0.4]);
semilogy(t,er2m,’r’);
axis([0,N,10^(-4), 10^0]);
title([’MSE Learning Curve (\lambda=.99, W=’ num2str(W) ’ )’],’fontsize’,10);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Mean squared error’,’fontsize’,8);
set(gca,’xtick’,[0:100:N],’fontsize’,8);

print -deps2 P1038.eps

10.39 Derive the equations for the a priori RLS lattice-ladder algorithm with error feedback (see Table 10.7), and
write a Matlab function for its implementation. Test the function by using the equalization experiment in
Example 10.5.2.
β cm (n) β cm (n − 1) α m emb (n)em∗ (n)
kmc (n) = = λ +
E mb (n) E mb (n) E mb (n)
β c (n − 1) E mb (n − 1) α m emb (n)em∗ (n)
= λ mb +
E m (n − 1) E mb (n) E mb (n)
β cm (n−1)
where kmc (n − 1) = Em b (n−1) . Now

λE mb (n − 1) = E mb (n) − α m emb (n)emb∗ (n)

so, the ladder parameter can be simplified to
kmc (n − 1) b α m emb (n)em∗ (n)
kmc (n) = [E (n) − α m e b
(n)e b∗
(n)] +
E mb (n) m m m
E mb (n)
α m emb (n)emb∗ (n) α m emb (n)em∗ (n)
= kmc (n − 1) − kmc (n − 1) +
E mb (n) E mb (n)
α m emb (n) ∗
= kmc (n − 1) + [e (n) − kmc (n − 1)emb∗ (n)]
E mb (n) m
∗
α m emb (n)em+1 (n)
= kmc (n − 1) +
E m (n)
b
Statistical and Adaptive Signal Processing - Solution Manual 355

MSE Learning Curve (λ=.99, W=2.9 )

0
10

−1
10
Mean squared error

−2
10

−3
10

−4
10
0 100 200 300 400 500 600 700 800 900 1000
Number of iterations (n)

Figure 10.38: MSE learning curve in P10.38

356 Statistical and Adaptive Signal Processing - Solution Manual

f
Similarly for the lattice parameters km (n) and kmb (n)

β m (n)
kmf (n) = −
E m (n − 1)
b

f∗
λβ m (n − 1) + α m (n − 1)emb (n − 1)em (n)
= −
E mb (n − 1)
f∗
β m (n − 1) E mb (n − 2) α m (n − 1)emb (n − 1)em (n)
= −λ −
E mb (n − 2) E mb (n − 1) E mb (n − 1)
f f∗
km (n − 1) b α m (n − 1)emb (n − 1)em (n)
= [E (n − 1) − α m (n − 1)e b
(n − 1)e b∗
(n)] −
E mb (n − 1) m m m
E mb (n − 1)
α m (n − 1)em (n − 1) f
b
= kmf (n − 1) − [em (n) + kmf (n − 1)emb∗ (n − 1)]
E mb (n − 1)
f∗
α m (n − 1)emb (n − 1)em+1 (n)
= kmf (n − 1) −
E mb (n − 1)

and
β ∗m (n)
kmb (n) = − f
E m (n)
f∗
λβ ∗m (n − 1) + α m (n − 1)emb (n − 1)em (n)
= − f
E m (n)
f∗
β ∗m (n − 1) E m (n − 1)
f
α m (n − 1)emb (n − 1)em (n)
= −λ f f
− f
E m (n − 1) E m (n) E m (n)
f∗
kmb (n − 1) α m (n − 1)emb (n − 1)em (n)
= f
[E mf (n) − α m (n − 1)emf (n)emf ∗ (n)] − f
E m (n) E m (n)
f
α m (n − 1)em (n)
= kmb (n − 1) − f
[emb∗ (n − 1) + kmb (n − 1)emf ∗ (n)]
E m (n)
α m (n − 1)emb (n − 1)em+1
b∗
(n)
= kmb (n − 1) − f
E m (n)
The Matlab script is given below and the MSE learning curve in shown in Figure 10.39.

%% Problem 10.38
%% Implement table 10.16%%
%%
%%
%%
%%

close all
clear all

W=2.9;
N=1000;
Nmc=250;
M=11;
Statistical and Adaptive Signal Processing - Solution Manual 357

lambda=0.99;
varv=0.001;
h=zeros(3,1);
er=zeros(N,Nmc);

h(1)=0.5*(1+cos(2*pi*(1-2)/W));
h(2)=0.5*(1+cos(2*pi*(2-2)/W));
h(3)=0.5*(1+cos(2*pi*(3-2)/W));

% Learning curves

hc=[0 h(1) h(2) h(3)]’;

n0=7;

t=(1:N)’;

for i=1:Nmc

y=sign(rand(N,1)-0.5);
v=sqrt(varv)*randn(N,1);
x=filter(hc,1,y)+v;
yd=zeros(N,1);
yd(n0:N)=y(1:N-n0+1);

% Time Initialization

Ef=ones(M+1,1)*10^(3);
Eb=Ef;
Ebold=Ef; Efold=Ef;
ef=zeros(M+1,1); eb=zeros(M+1,1);
ebold=eb;
kf=zeros(M,1); kb=zeros(M,1);
kfold=kf; kbold=kb;
b=zeros(M+1,1); bc=zeros(M+1,1);
bold=b; bcold=bc;
e=zeros(M+1,1); kc=zeros(M,1);
kcold=kc;
alpha=zeros(M+1,1);
alphaold=alpha;
for n=1:N

% Order Initialization

ef(1)=x(n);
eb(1)=x(n);
e(1)=yd(n);
alpha(1)=1;
alphaold(1)=alpha(1);
358 Statistical and Adaptive Signal Processing - Solution Manual

for m=1:M
%Lattice
ef(m+1) = ef(m) + conj(kfold(m))*ebold(m);
eb(m+1) = ebold(m) + conj(kbold(m))*ef(m);
b(m) = lambda*bold(m) + (ebold(m)*conj(ef(m)))*alphaold(m);
Ef(m) = lambda*Efold(m) + alphaold(m)*conj(ef(m))*ef(m);
Eb(m) = lambda*Ebold(m) + alpha(m)*conj(eb(m))*eb(m);
kf(m) = -b(m)/Ebold(m);
kb(m) = -conj(b(m))/Ef(m);
alpha(m+1) = alpha(m) - abs(eb(m))*abs(eb(m))/Eb(m);

% Ladder
bc(m)=lambda*bcold(m) + (eb(m)*conj(ef(m)))*alpha(m);
e(m+1)=e(m)-conj(kcold(m))*eb(m);
kc(m)=bc(m)/Eb(m);
end

er(n,i)=e(M+1);erf(n,i)=ef(M);erb(n,i)=eb(M);
Erf(n,i)=Eb(1); Erb(n,i)=Eb(M); al(n,i)=alpha(M);

% Time Delay
Ebold=Eb; Efold=Ef;
ebold=eb; efold=ef;
bold=b; bcold=bc;
kfold=kf; kbold=kb;
alphaold=alpha;
end
i
end
er2=er.^2;
er2m=mean(er2,2);

t=(1:N)’;
%plot(t,Erf,’r’,t,Erb,’b’);
%text(250,300,’\leftarrow\fontsize{9pt} E^f’);
%text(100,200,’E^b \rightarrow\fontsize{9pt} ’);

Jmin1=1;

%subplot(’position’,[0.1,0.55,0.85,0.4]);
semilogy(t,er2m,’r’);
axis([0,N,10^(-4), 10^0]);
title([’MSE Learning Curve (\lambda=.99, W=’ num2str(W) ’ )’],’fontsize’,10);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Mean squared error’,’fontsize’,8);
set(gca,’xtick’,[0:100:N],’fontsize’,8);

print -deps2 P1039.eps

Statistical and Adaptive Signal Processing - Solution Manual 359

MSE Learning Curve (λ=.99, W=2.9 )

0
10

−1
10
Mean squared error

−2
10

−3
10

−4
10
0 100 200 300 400 500 600 700 800 900 1000
Number of iterations (n)

Figure 10.39: MSE learning curve in P10.39

10.40 Derive the equations for the a posteriori RLS lattice-ladder algorithm with error feedback (Ling et al. 1986)
and write a Matlab function for its implementation. Test the function by using the equalization experiment in
Example 10.5.2.
To be completed.
10.41 The a posteriori and the a priori RLS lattice-ladder algorithms need the conversion factor α m (n) because the
updating of the quantities E mf (n), E mb (n), β m (n), and β cm (n) requires both the a priori and a posteriori errors.
Derive a double (a priori and a posteriori) lattice-ladder RLS filter that avoids the use of the conversion factor
by updating both the a priori and the a posteriori prediction and filtering errors.
Combining the a posteriori(Table 10.15) and a priori(Table 10.16) lattice-ladder algorithms, and removing
dependence on α m . The lattice parameters are
−β m (n)
kmf (n) =
E mb (n − 1)
and
−β ∗m (n)
kmb (n) = f
E m (n)
where the time updating of β m (n) is
β m (n) = λβ m (n − 1) + ε bm (n − 1)emf ∗ (n)
360 Statistical and Adaptive Signal Processing - Solution Manual

and the minimum error energies are

E mf (n) = λE mf (n − 1) + ε mf (n)emf ∗ (n)

E mf (n) = λE mb (n − 1) + ε bm (n)emb∗ (n)

The a priori and a posteriori errors become

em+1 (n) = emf (n) + kmf ∗ (n − 1)emb (n − 1)

f

b
em+1 (n) = emb (n − 1) + kmb∗ (n − 1)emf (n)
εm+1 (n) = ε mf (n) + kmf ∗ (n)εbm (n − 1)
f

εbm+1 (n) = ε bm (n − 1) + kmb∗ (n)εmf (n)

The ladder parameters are

−β cm (n)
kmc (n) =
E mb (n)
β cm (n) = λβ cm (n − 1) + ε bm (n)em∗ (n)
em+1 (n) = em (n) − kmc∗ (n − 1)emb (n)
ε m+1 (n) = ε m (n) − kmc∗ (n)εbm (n)

10.42 Program the RLS Givens lattice-ladder filter with square roots (see Table 10.18), and study its use in the
adaptive equalization experiment of Example 10.5.2.
The Matlab script is given below and the MSE learning curve in shown in Figure 10.42.

%% Problem 10.38
%% Implement table 10.16%%
%%
%%
%%
%%

close all
clear all

W=2.9;
N=1000;
Nmc=250;
M=11;
lambda=0.99;
varv=0.001;
h=zeros(3,1);
er=zeros(N,Nmc);

h(1)=0.5*(1+cos(2*pi*(1-2)/W));
h(2)=0.5*(1+cos(2*pi*(2-2)/W));
h(3)=0.5*(1+cos(2*pi*(3-2)/W));
Statistical and Adaptive Signal Processing - Solution Manual 361

% Learning curves

hc=[0 h(1) h(2) h(3)]’;

n0=7;

t=(1:N)’;

for i=1:Nmc

y=sign(rand(N,1)-0.5);
v=sqrt(varv)*randn(N,1);
x=filter(hc,1,y)+v;
yd=zeros(N,1);
yd(n0:N)=y(1:N-n0+1);

% Time Initialization

Ef=ones(M+1,1)*10^(3);
Eb=Ef;
Ebold=Ef; Efold=Ef;
ef=zeros(M+1,1); eb=zeros(M+1,1);
ebold=eb;
kf=zeros(M,1); kb=zeros(M,1);
kfold=kf; kbold=kb;
cf=zeros(M+1,1); sf=zeros(M+1,1);
cb=zeros(M+1,1); sb=zeros(M+1,1);
cfold=cf; cbold=cb; sfold=sf; sbold=sb;
e=zeros(M+1,1); kc=zeros(M,1);
kcold=kc;
alpha=zeros(M+1,1);
alphaold=alpha;

for n=1:N

% Order Initialization

ef(1)=x(n);
eb(1)=x(n);
e(1)=yd(n);
alpha(1)=1;
alphaold(1)=alpha(1);

for m=1:M
%Lattice
Ef(m) = (lambda*Efold(m)^2 + conj(ef(m))*ef(m))^(1/2);
cf(m) = lambda^(1/2)*Efold(m)/Ef(m);
sf(m) = ef(m)/Ef(m);
Eb(m) = (lambda*Ebold(m)^2 + conj(eb(m))*eb(m))^(1/2);
cb(m) = lambda^(1/2)*Ebold(m)/Eb(m);
362 Statistical and Adaptive Signal Processing - Solution Manual

sb(m) = eb(m)/Eb(m);
ef(m+1) = cbold(m)*ef(m) + lambda^(1/2)*sbold(m)*conj(kfold(m));
kf(m) = lambda^(1/2)*cbold(m)*kfold(m) - sbold(m)*conj(ef(m));
eb(m+1) = cf(m)*ebold(m) + lambda^(1/2)*sf(m)*conj(kbold(m));
kb(m) = lambda^(1/2)*cf(m)*kbold(m) - sf(m)*conj(ebold(m));
alpha(m+1) = alpha(m) - abs(alpha(m)*eb(m))*abs(alpha(m)*eb(m))/Eb(m);

% Ladder
e(m+1) = cb(m)*e(m) - lambda^(1/2)*sb(m)*kcold(m);
kc(m)= lambda^(1/2)*cb(m)*kcold(m) + sb(m)*conj(e(m));
end

er(n,i)=e(M+1);erf(n,i)=ef(M);erb(n,i)=eb(M);
Erf(n,i)=Eb(1); Erb(n,i)=Eb(M); al(n,i)=alpha(M);

% Time Delay
cbold=cb; cfold=cf;
sbold=sb; sfold=sf;
Ebold=Eb; Efold=Ef;
ebold=eb; efold=ef;
kfold=kf; kbold=kb; kcold=kc;
alphaold=alpha;
end
i
end
er2=er.^2;
er2m=mean(er2,2);

t=(1:N)’;
%plot(t,Erf,’r’,t,Erb,’b’);
%text(250,300,’\leftarrow\fontsize{9pt} E^f’);
%text(100,200,’E^b \rightarrow\fontsize{9pt} ’);

Jmin1=1;

%subplot(’position’,[0.1,0.55,0.85,0.4]);
semilogy(t,er2m,’r’);
axis([0,N,10^(-4), 10^0]);
title([’MSE Learning Curve (\lambda=.99, W=’ num2str(W) ’ )’],’fontsize’,10);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Mean squared error’,’fontsize’,8);
set(gca,’xtick’,[0:100:N],’fontsize’,8);

print -deps2 P1042.eps

10.43 Derive the formulas and program the RLS Givens lattice-ladder filter without square roots (see Table 10.18),
and study its use in the adaptive equalization experiment of Example 10.5.2.
Statistical and Adaptive Signal Processing - Solution Manual 363

MSE Learning Curve (λ=.99, W=2.9 )

0
10

−1
10
Mean squared error

−2
10

−3
10

−4
10
0 100 200 300 400 500 600 700 800 900 1000
Number of iterations (n)

Figure 10.42: MSE learning curve in P10.42

364 Statistical and Adaptive Signal Processing - Solution Manual

The square root-free version is a modification of the error feedback form of the a priori LS lattice-ladder
algorithm. If we define

λE mb (n − 1)
cmb (n) = = |c̃mb (n)|2
E mb (n)
and
α m (n)emb (n)
smb (n) =
E mb (n)
then using (10.7.50), the ladder coefficient is

kmc (n) = cmb (n)kmc (n − 1) + smb (n)em∗ (n)

Also the lattice parameters are found using the error feedback a priori updating with

kmf (n) = cmb (n − 1)kmf (n − 1) − smb (n − 1)emf ∗ (n)

and

kmb (n) = cmf (n)kmb (n − 1) − smf (n)emb∗ (n − 1)

where
f
λE m (n − 1)
cmf (n) = f
= |c̃mf (n)|2
E m (n)
and
f
α m (n − 1)em (n)
smf (n) = f
E m (n)
are the forward rotation parameters.
The Matlab script is given below and the MSE learning curve in shown in Figure 10.43.

%% Problem 10.38
%% Implement table 10.16%%
%%
%%
%%
%%

close all
clear all

W=2.9;
N=1000;
Nmc=250;
M=11;
lambda=0.99;
varv=0.001;
h=zeros(3,1);
Statistical and Adaptive Signal Processing - Solution Manual 365

er=zeros(N,Nmc);

h(1)=0.5*(1+cos(2*pi*(1-2)/W));
h(2)=0.5*(1+cos(2*pi*(2-2)/W));
h(3)=0.5*(1+cos(2*pi*(3-2)/W));

% Learning curves

hc=[0 h(1) h(2) h(3)]’;

n0=7;

t=(1:N)’;

for i=1:Nmc

y=sign(rand(N,1)-0.5);
v=sqrt(varv)*randn(N,1);
x=filter(hc,1,y)+v;
yd=zeros(N,1);
yd(n0:N)=y(1:N-n0+1);

% Time Initialization

Ef=ones(M+1,1)*10^(3);
Eb=Ef;
Ebold=Ef; Efold=Ef;
ef=zeros(M+1,1); eb=zeros(M+1,1);
ebold=eb;
kf=zeros(M,1); kb=zeros(M,1);
kfold=kf; kbold=kb;
cf=zeros(M+1,1); sf=zeros(M+1,1);
cb=zeros(M+1,1); sb=zeros(M+1,1);
cfold=cf; cbold=cb; sfold=sf; sbold=sb;
e=zeros(M+1,1); kc=zeros(M,1);
kcold=kc;
alpha=zeros(M+1,1);
alphaold=alpha;

for n=1:N

% Order Initialization

ef(1)=x(n);
eb(1)=x(n);
e(1)=yd(n);
alpha(1)=1;
alphaold(1)=alpha(1);

for m=1:M
366 Statistical and Adaptive Signal Processing - Solution Manual

%Lattice
Ef(m) = lambda*Efold(m) + alphaold(m)*conj(ef(m))*ef(m);
cf(m) = lambda*Efold(m)/Ef(m);
sf(m) = alphaold(m)*ef(m)/Ef(m);
Eb(m) = lambda*Ebold(m) + alpha(m)*conj(eb(m))*eb(m);
cb(m) = lambda*Ebold(m)/Eb(m);
sb(m) = alpha(m)*eb(m)/Eb(m);
ef(m+1) = ef(m) + conj(kfold(m))*ebold(m);
kf(m) = cbold(m)*kfold(m) - sbold(m)*conj(ef(m));
eb(m+1) = ebold(m) + conj(kbold(m))*ef(m);
kb(m) = cf(m)*kbold(m) - sf(m)*conj(ebold(m));
alpha(m+1) = alpha(m) - abs(alpha(m)*eb(m))*abs(alpha(m)*eb(m))/Eb(m);

% Ladder
e(m+1)=e(m)-conj(kcold(m))*eb(m);
kc(m)= cb(m)*kcold(m) + sb(m)*conj(e(m));
end

er(n,i)=e(M+1);erf(n,i)=ef(M);erb(n,i)=eb(M);
Erf(n,i)=Eb(1); Erb(n,i)=Eb(M); al(n,i)=alpha(M);

% Time Delay
cbold=cb; cfold=cf;
sbold=sb; sfold=sf;
Ebold=Eb; Efold=Ef;
ebold=eb; efold=ef;
kfold=kf; kbold=kb; kcold=kc;
alphaold=alpha;
end
i
end

er2=er.^2;
er2m=mean(er2,2);

t=(1:N)’;
%plot(t,Erf,’r’,t,Erb,’b’);
%text(250,300,’\leftarrow\fontsize{9pt} E^f’);
%text(100,200,’E^b \rightarrow\fontsize{9pt} ’);

Jmin1=1;

%subplot(’position’,[0.1,0.55,0.85,0.4]);
semilogy(t,er2m,’r’);
axis([0,N,10^(-4), 10^0]);
title([’MSE Learning Curve (\lambda=.99, W=’ num2str(W) ’ )’],’fontsize’,10);
xlabel(’Number of iterations (n)’,’fontsize’,8);
ylabel(’Mean squared error’,’fontsize’,8);
Statistical and Adaptive Signal Processing - Solution Manual 367

MSE Learning Curve (λ=.99, W=2.9 )

0
10

−1
10
Mean squared error

−2
10

−3
10

−4
10
0 100 200 300 400 500 600 700 800 900 1000
Number of iterations (n)

Figure 10.43: MSE learning curve in P10.43

set(gca,’xtick’,[0:100:N],’fontsize’,8);

print -deps2 P1043.eps

10.44 In this problem we discuss the derivation of the normalized lattice-ladder RLS algorithm, which uses a smaller
number of time and order updating recursions and has better numerical behavior due to the normalization of
its variables.

(a) Define the energy and angle normalized variables

ε fm (n) ε bm (n) ε m (n)

ēmf (n) = √  ēmb (n) = √  ēm (n) = √ √
α m (n) E m (n)
f α m (n) E m (n)
b α m (n) E m (n)

β m (n) β cm (n)
k̄m (n) =   k̄mc (n) = √ 
E mf (n) E mb (n − 1) E m (n) E mb (n)

and show that the normalized errors and the partial correlation coefficients k̄m (n) and k̄mc (n) have magni-
tude less than 1.
368 Statistical and Adaptive Signal Processing - Solution Manual

(b) Derive the following normalized lattice-ladder RLS algorithm:

E 0f (−1) = E 0 (−1) = δ > 0

For n = 0, 1, 2, . . .
E 0f (n) = λE 0f (n − 1) + |x(n)|2 , E 0 (n) = λE 0 (n − 1) + |y(n)|2
ē0f (n) = ē0b (n) = √x(n) , ē0 (n) = √
y(n)
E 0 (n)
E 0 (n)
f

For m = 0 to M − 1
 
k̄m (n) = 1 − |ēmf (n)|2 1 − |ēmb (n − 1)|2 k̄m (n − 1) + ēmf∗ (n)ēmb (n − 1)
  −1
f
ēm+1 (n) = 1 − |emb (n − 1)|2 1 − |k̄m (n)|2 [ēmf (n) − k̄m (n) ēmb (n − 1)]
  −1
b
ēm+1 (n) = 1 − |ēmf (n)|2 1 − |k̄m (n)|2 [ēmb (n − 1) − k̄m (n) ēmf (n)]
 
k̄mc (n) = 1 − |ēm (n)|2 1 − |ēmb (n)|2 k̄mc (n − 1) + ēm∗ (n)ēmb (n)

: −1

ēm+1 (n) = 1 − |ēm (n)| 1 − |k̄m (n)|
b 2 c 2 [ēm (n) − k̄mc (n)ēmb (n)]

(c) Write a Matlab function to implement the derived algorithm, and test its validity by using the equalization
experiment in Example 10.5.2.

To be completed.

10.45 Prove (10.4.46) by direct manipulation of (10.6.35).

From (10.6.35) follows that the last diagonal element of Q(n) is equal to α̃(n). Therefore, using the partitioning
 
Q̄(n) q(n)
Q(n) =
q H (n) α̃(n)

and (10.6.35) we obtain

√
q H (n) λR̃(n − 1) + α̃(n)x H (n) = 0 H
√
q H (n) λk̃(n − 1) + α̃(n)y ∗ (n) = ẽ∗ (n)

Solving the top equation

√
q H (n) λ = −α̃(n)x H (n)R̃−1 (n − 1)

and substituting to the bottom one, gives

  ẽ(n)
α̃(n) y ∗ (n) − x H (n)c(n − 1) = ẽ∗ (n) ⇒ α̃(n) =
e(n)

because c(n − 1) = R̃−1 (n − 1)k̃(n − 1) and e(n) = y(n) − c H (n − 1)x(n).

10.46 Derive the formulas for the QR-RLS lattice predictor (see Table 10.18), using the approach introduced in
Section 10.6.3 (Yang and Böhme 1992).
To be completed.
Statistical and Adaptive Signal Processing - Solution Manual 369

10.47 Demonstrate how the systolic array in Figure 10.55, which is an extension of the systolic array structure shown
in Figure 10.36, can be used to determine the LS error e(n) and the LS coefficient vector c(n). Determine
the functions to be assigned to the dotted-line computing elements and the inputs with which they should be
supplied.
To be completed.
10.48 (new version) The implementation of adaptive filters using multiprocessing involves the following steps:
(1) partitioning of the overall computational job into individual tasks, (2) allocation of computational and
communications tasks to the processors, and (3) synchronization and control of the processors. Figure 10.56
shows a cascade multiprocessing architecture used for adaptive filtering. To avoid latency (i.e., a delay between
the filter’s input and output that is larger than the sampling interval), each processor should complete its task
in time less than the sampling period and use results computed by the preceding processor and the scalar
computational unit at the previous sampling interval. This is accomplished by the unit delays inserted between
the processors.

(a) Draw the block diagram of a section that updates one coefficient of the FLP using equations (a)-(h), given
in Table 10.13. Then, use the obtained diagram to explain why the fast Kalman algorithm does not satisfy
the multiprocessing requirements.
The block diagram of equation (g) is given in Figure 10.48a1 while that of (h) is given in Figure 10.48a2.
-M.
g M+1 (n)

b(n − 1)
g(n)

−g (M+1)
M+1 (n) 1
(M+1)
1−g M+1 (n)eb* (n)

Figure 10.48a1: Block diagram for equation (g)

g(n)
b(n)

z −1

eb* (n)
Figure 10.48a2: Block diagram for equation (h).

Combining these diagrams with equations (a)-(f) in Table 10.13 leads to the block diagram 10.48(a). The
lack of delay elements in the paths for g-M.
M+1 (n) and g M+1 (n) violates the multiprocessing architecture.
370 Statistical and Adaptive Signal Processing - Solution Manual

a(n)

ε f (n)
x
E f (n)

+ g M+1 (n)

b(n − 1) g(n − 1)

x −g (M+1)
M+1 (n)

+ z −1
g-M.
M+1 (n)

1
X
1− g (M+1)
M+1 (n)e (n)
b∗

x −eb∗ (n)

+ g(n)

z −1 b(n)

x −ef∗ (n − 1)

a(n) +

a(n + 1)
z −1

Figure 10.48a2: Multiprocessing element for the fast Kalman algorithm using equations (a)-(h) in table 10.13.
Statistical and Adaptive Signal Processing - Solution Manual 371

a(n) b(n − 1)
f∗
−e (n)

a(n)
z −1 z −1
ε f (n − 1) a(n − 1)
g(n − 1) z −1
E f (n − 1)
g(n − 1)
z −1 z −1
g(n − 2)

z −1 b(n − 1)
−eb∗ (n − 1)
g-M.
M+1 (n)

g (M+1)
M+1 (n − 1) 1
1− g (M+1)
M+1 (n
− 1)eb∗ (n − 1)
Figure 10.48b: Section for the multiprocessing implementation of the fast Kalman algorithm using Table 10.13 and
equations (g) and (h).

(b) Rearrange the formulas in Table 10.13 as follows: (e), (g), (h), (a), (b), (c), (d), (f ). Replace n by n − 1
in (e), (k), and (l). Show that the resulting algorithm leads to the multiprocessing section shown in Figure
10.57. Identify the input-output quantities and the various multiplier factors and show that this section can
be used in the multiprocessing architecture shown in Figure 10.56. Each processor in Figure 10.56 can
be assigned to execute one or more of the designed sections. Note: You may find useful the discussions
in Lawrence and Tewksbury (1983) and in Manolakis and Patel (1992).
Working as in (a) we obtain the results shown in Figure 10.48(b).
Comments: $ %
The multiplier factors ε f (n − 1)/E f (n − 1), g (M+1)
M+1 (n), 1/ 1 − g (M+1)
M+1 (n − 1)em
b∗
(n − 1) , −eb∗ (n − 1),
and −e f ∗ (n) at each element, should be available after the processing of the sample x(n − 1) has been
completed and the sample x(n) has arrived. To start processing the sample x(n) all these quantities should
be available to each processing section.
• e f (n − 1), eb (n − 1), E f (n − 1) are available from the previous sampling interval.
• ε f (n − 1)/E f (n − 1) and g (M+1)
M+1 (n) are computed at the first section as the first and last elements
of g M+1 (n). However, since they are used$ by all processors they should
% be computed separately at
(M+1)
the beginning. Then, we compute 1/ 1 − g M+1 (n − 1)em (n − 1) . Therefore, these quantities
b∗

should be computed “outside” the sections before we start processing the sample x(n).
• e f (n) should be available before each section finishes the computation of b(n − 1) (see Figure ?).
The solution is to compute x̂(n) = a H (n − 1)x(n − 1) at the previous processing cycle and then
compute e f (n) = x(n) + x̂(n) after the arrival of x(n).
(c) Prove the formulas
-M.
b(n − 1) − g M+1 (n)eb∗ (n)
b(n) = (k)
1 − g (M+1)
M+1 (n) e (n)
b∗

g(n) = g-M. (M+1)

M+1 (n) − g M+1 (n) b(n) (l)
and show that they can be used to replace formulas (g) and (h) in Table 10.13.
372 Statistical and Adaptive Signal Processing - Solution Manual

If we eliminate gm (n) from (10.7.13) using (10.7.14), we obtain equations (k) and (l).
(d) Draw a block diagram of a single multiprocessing section for the obtained algorithm and show that this
section can be used in the architecture of Figure 10.56.
Figure 10.48(c) shows the updating section for the fast Kalman algorithm described by (a)-(f), (k), (l). The
lack of delay elements in the paths for g-M.
M+1 (n) and g M+1 (n) violates the multiprocessing requirements.
The multiprocessing section for this version of the fast Kalman algorithm is given in Figure 10.48(d).
(e) Rearrange the formulas (a)-(l),in Table 10.14, to obtain a version of the FAEST algorithm that complies
with the multiprocessing architecture of Figure 10.56 and draw a block diagram of the resulting updating
section.
Repeat (a) and (b).

10.49 (new version) Show that the LMS algorithm given in Table 10.3 satisfies the multiprocessing architecture in
Figure 10.56.
If we modify the LMS algorithm as follows

c(n) = c(n − 1) + 2µx(n − 1)e∗ (n − 1)

ŷ(n) = c H (n)x(n)
e(n) = y(n) − ŷ(n)

we can easily see that the LMS algorithm complies with the multiprocessing architecture of Figure 10.56.

10.50 Show that the a priori RLS linear prediction lattice (i.e., without the ladder part) algorithm with error feedback
complies with the multiprocessing architecture of Figure 10.56. Explain why the addition of the ladder part
violates the multiprocessing architecture. Can we rectify these violations? (See Lawrence and Tewksbury
1983.)
The lattice predictor has a modular structure

e0f (n) = e0b (n) = x(n)

f
em+1 (n) = emf (n) + kmf∗ (n − 1)emb (n − 1) 0≤m < M −1
b
em+1 (n) = emb (n − 1) + kmb∗ (n − 1)emf (n) 0≤m < M −1

Multiprocessing is impossible because there is no delay in the emf (n) path (if emf (n) could be replaced by
emf (n − 1) there would be no problem). To overcome this problem, we notice that

emf (n) = x(n) + amH (n − 1)xm (n − 1) , x(n) − x̂m (n)

f
which combined with em+1 (n) = emf (n) + kmf∗ (n − 1)emb (n − 1) gives

x̂m+1 (n) = x̂m (n) + kmf∗ (n − 1)emb (n − 1

which is initialized x̂0 (n) = 0. Since e0f (n) = x(n) only x(n) is not available at time n. Thus, instead of
f
em+1 (n) we can compute x̂m+1 (n) in advance, store it and have it ready for use at time n when x(n) becomes
available. These observations lead to the lattice section shown in Figure 10.50.

10.51 The fixed-length sliding window RLS algorithm is given in (10.8.4) through (10.8.10).

(a) Derive the above equations of this algorithm (see Manolakis et al. 1987).
Statistical and Adaptive Signal Processing - Solution Manual 373

a(n)

ε f (n)
x
E f (n)
g(n − 1)
+

g M+1 (n)

x −eb∗ (n)

-M.
g M+1 (n) b(n − 1)
+

1
X
1− g (M+1)
M+1 (n)e (n)
b∗

b(n) z −1

x −g (M+1)
M+1 (n)

g(n − 1)
g(n) −1
z

x −e f ∗ (n + 1)

a(n) +

z −1 a(n + 1)

Figure 10.48c: Section for the multiprocessing implementation of the fast Kalman algorithm using Table 10.13 and
equations (k) and (l).
374 Statistical and Adaptive Signal Processing - Solution Manual

b(n − 1) a(n)

a(n − 1)
a(n) z −1 z −1

ε f (n) −eb∗ (n − 1) −e f ∗ (n)

x x x
E f (n)
b(n − 2)
g(n − 1) z −1 g(n − 1)
z −1
+ + +

z −1
−1 a(n)
z x

-M.
g M+1 (n − 1) b(n − 1)

1 x

1 − g (M+1)
M+1 (n − 1)e (n − 1)
b∗ g (M+1)
M+1 (n − 1) g(n − 1)
+

Figure 10.48c: Section for the multiprocessing implementation of the fast Kalman algorithm using Table 10.13 and
equations (k) and (l) This is the version proposed by Lawrence and Tewksbury.

x(n)

x̂m (n) x̂m+1 (n)

+

f∗
f
em (n) km (n − 1) f
em+1 (n)
-
+ z −1 +

kmb∗ (n − 1)

f∗
km (n − 1)
emb (n) b
em+1 (n − 1)
−1 −1
z + z
b
em+1 (n)

Figure 10.50: Multiprocessing session for the a-priori RLS lattice filter.
Statistical and Adaptive Signal Processing - Solution Manual 375

(b) Develop a Matlab function to implement the algorithm

[c,e] = slwrls(x,y,L,delta,M,c0)
where L is the fixed length of the window.
(c) Generate 500 samples of the following nonstationary process

w(n) + 0.95x(n − 1) − 0.9x(n − 2) 0 ≤ n < 200
x(n) = w(n) − 0.95x(n − 1) − 0.9x(n − 2) 200 ≤ n < 300

w(n) + 0.95x(n − 1) − 0.9x(n − 2) n ≥ 300

where w(n) is a zero-mean, unit-variance white noise process. We want to obtain a second-order linear
predictor using adaptive algorithms. Use the sliding window RLS algorithm on the data and choose
L = 50 and 100. Obtain plots of the filter coefficients and mean square error.
(d) Now use the growing memory RLS algorithm by choosing λ = 1. Compare your results with the sliding-
window RLS algorithm.
(e) Finally, use the exponentially growing memory RLS by choosing λ = (L − 1)/(L + 1) that produces the
same MSE. Compare your results.

To be completed.

10.52 Consider the definition of the MSD D(n) in (10.2.29) and that of the trace of a matrix (A.2.16).

(a) Show that D(n) = tr{(n)}, where (n) is the correlation matrix of c̃(n).
(b) For the evolution of the correlation matrix in (10.8.58), show that

tr(R−1 Rψ )
D(∞) & µMσ 2ν +
4µ

To be completed.
10.53 Consider the analysis model given in Figure 10.42. Let the parameters of this model be as follows:
 
0.9
co (n) model parameters: co (0) = M =2 ρ = 0.95
−0.8
ψ(n) ∼ WGN(0, Rψ ) Rψ = (0.01)2 I

Signal x(n) parameters x(n) ∼ WGN(0, R) R=I

Noise v(n) parameters v(n) ∼ WGN(0, σ 2v ) σ v = 0.1

Simulate the system, using three values of µ that show slow, matched, and optimum adaptations of the LMS
algorithm.

(a) Obtain the tracking plots similar to Figure 10.43 for each of the above three adaptations.
(b) Obtain the learning curve plots similar to Figure10.44 for each of the above three adaptations.

To be completed.
376 Statistical and Adaptive Signal Processing - Solution Manual

10.54 Consider the analysis model given in Figure 10.42. Let the parameters of this model be as follows
 
0.9
co (n) model parameters co (0) = M =2 ρ = 0.95
−0.8
ψ(n) ∼ WGN(0, Rψ ) Rψ = (0.01)2 I

Signal x(n) parameters x(n) ∼ WGN(0, R) R=I

Noise v(n) parameters v(n) ∼ WGN(0, σ 2v ) σ v = 0.1

Simulate the system, using three values of µ that show slow, matched, and optimum adaptations of the RLS
algorithm.

(a) Obtain the tracking plots similar to Figure 10.49 for each of the above three adaptations.
(b) Obtain the learning curve plots similar to Figure 10.50 for each of the above three adaptations.
(c) Compare your results with those obtained in Problem 10.53.

To be completed.

10.55 Consider the time-varying adaptive equalizer shown in Figure 10.58 in which the time variation of the channel
impulse response is given by

h(n) = ρh(n − 1) + 1 − ρ η(n), with
√
ρ = 0.95 η(n) ∼ WGN(0, 10) h(0) = 0.5

Let the equalizer be a single-tap equalizer and v(n) ∼ WGN(0, 0.1).

(a) Simulate the system for three different adaptations; that is, choose µ for slow, matched, and fast adaptations
of the LMS algorithm.
(b) Repeat part (a), using the RLS algorithm.

To be completed.
 Chapter 11

Array Processing

11.1 The 2 element array with inter-element spacing of d has a signal impinging on it from and angle of φ = 0◦ at
a rate of propagation c.

(a) If the wavelength of the propagating signal is given by

c
λ=
Fc
where Fc is the frequency of the signal, then the difference in the arrival time of the signal at the two
sensors is
d
τ = sin φ
λ
= 0

since φ = 0◦ . The signal arrives at the two sensors at the same time.
(b) The same delay of 0 seconds between sensors is produced by any other signals for which sin φ = 0.
Therefore, signals impinging on the array from φ = 180◦ will also produce a delay of zero.
(c) With a single sensor we cannot determine the angle from which a signal arrives because we cannot
determine a difference in arrival time. In other words, signals from all directions will have the same
characteristics.

11.2 This problem is a Matlab based problem. Below is the Matlab code and beampatterns at φ = 45◦ and
φ = 60circ . Note that mechanical steering is modeled as the beampattern for electronic steering to broadside
and shifting the angle to the desired mechanical steering angle. In both cases we observe a smaller mainbeam
for the mechanical steering than the electronic steering due to the fact that spatial frequency is a function of
sin φ and not simply φ. This effect is more pronounced the greater the steering angle, so it is more noticeable
for the case of φ = 60◦ .

%
% Problem 11.2
%
clear

M = 20; % number of elements

spacing = 0.5; % lambda/2 spacing

% First part of problem has a steering angle of 45 degrees

phi_steer = 45; % steering direction

% Compute the steering vectors for electronic and mechanical steering

v_elec = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_steer*pi/180))/sqrt(M);
v_mech = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(0*pi/180))/sqrt(M);

% Angles for beampattern

377
378 Statistical and Adaptive Signal Processing - Solution Manual

phi = -90:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the beampatterns for electronic and mechanical steering

V_elec = V’*v_elec;
V_mech = V’*v_mech;

% Mechanical steering angles are simply the true angles shifted by

% the mechanical steering angle
phi_mech = phi + phi_steer; % mechanical steering angles for beampattern
index_toobig = find(phi_mech > 90);
phi_mech(index_toobig) = phi_mech(index_toobig)-180;
[phi_mech,index] = sort(phi_mech);
V_mech = V_mech(index);

% Plot out the beampatterns in decibels

figure
plot(phi,20*log10(abs(V_elec)),’b’,...
phi_mech,20*log10(abs(V_mech)),’r--’,...
’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid;
set(gca,’xtick’,-90:15:90);

% Repeat for a steering angle of 60 degrees

phi_steer = 60; % steering direction

% Compute the steering vectors for electronic and mechanical steering

v_elec = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_steer*pi/180))/sqrt(M);
v_mech = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(0*pi/180))/sqrt(M);

% Angles for beampattern

phi = -90:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the beampatterns for electronic and mechanical steering

V_elec = V’*v_elec;
V_mech = V’*v_mech;

% Mechanical steering angles are simply the true angles shifted by

% the mechanical steering angle
phi_mech = phi + phi_steer; % mechanical steering angles for beampattern
index_toobig = find(phi_mech > 90);
phi_mech(index_toobig) = phi_mech(index_toobig)-180;
[phi_mech,index] = sort(phi_mech);
V_mech = V_mech(index);
Statistical and Adaptive Signal Processing - Solution Manual 379

0 0

−5 −5

−10 −10
Power Response (dB)

Power Response (dB)

−15 −15

−20 −20

−25 −25

−30 −30

−35 −35

−40 −40

−45 −45

−50 −50
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg)

(a) φ = 45◦ . (b) φ = 60◦ .

Figure 11.2a: Beampatterns for a M = 20 λ/2-spaced ULA with electronic steering (solid) and mechanical
steering (dashed).

% Plot out the beampatterns in decibels

figure
plot(phi,20*log10(abs(V_elec)),’b’,...
phi_mech,20*log10(abs(V_mech)),’r--’,...
’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

11.3 This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a), (b), and (c) and presented sequentially along with the accompanying figures.
(a) Both signals are clearly visible in the steered responses. The expected steered response is so close to the
estimated steered response that the two overlay in Figure 11.3a.

%
% Problem 11.3
%
clear

M = 50; % number of elements

spacing = 0.5; % lambda/2 spacing
N = 1000; % number of samples

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
380 Statistical and Adaptive Signal Processing - Solution Manual

%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% The angles in degrees of the two spatially propagating signals

phi1 = -10;
phi2 = 30;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 25;
Pn = 0;

% Compute the array response vectors of the two signals

v1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Compute the true covariance matrix

% (remember to divide by 10 not 20 for powers in the covariance)
R_x = (10^(P1_db/10)) * v1 * v1’ + (10^(P2_db/10)) * v2 * v2’ + ...
(10^(Pn/10)) * eye(M);

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

for k = 1:length(phi)
R_steer_est(k) = mean(abs(C_smf(:,k)’ * x).^2,2);
R_steer_true(k) = abs(C_smf(:,k)’* R_x * C_smf(:,k));
end

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 65]);
grid
set(gca,’xtick’,-90:15:90);
Statistical and Adaptive Signal Processing - Solution Manual 381

60 60

50 50
Steered Response (dB)

Steered Response (dB)

40 40

30 30

20 20

10 10

0 0

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg)

(a) Estimated correlation matrix. (b) True correlation matrix.

Figure 11.3a: Steered responses with estimated and true correlation matrices for scenario in Problem 11.3(a) with
signals at φ = −10◦ and φ = 30◦ with powers of 20 and 25 dB, respectively.

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_true),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 65]);
grid
set(gca,’xtick’,-90:15:90);

(b) The beampattern for the beamformer steered to φ = 30◦ is shown in Figure 11.3b. The gain at φ = 30◦
is one (0 dB) so that the steered response in Figure 11.3a shows the correct power of the 25 dB signal at
φ = 30◦ . The same is true of the signal at φ = −10◦ . Elsewhere, we see the two signals leaked through the
first few principal sidelobes in the steered response near the two sources, but otherwise the response is near 0
dB, the noise level, since the two signals are not strong enough to leak through the further out sidelobes of the
beamformer.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Compute the spatial matched filter at 30 degrees

phi_smf = 30;
c_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_smf*pi/180))/sqrt(M);

% The angles at which to evaluate the beampattern

phi = -90:0.2:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);
382 Statistical and Adaptive Signal Processing - Solution Manual

−5

−10

Power Response (dB)

−15

−20

−25

−30

−35

−40

−45

−50
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.3b: The beampattern for a M = 50 λ/2-spaced ULA steered to φ = 30◦ .

% Compute the beampatterns for electronic and mechanical steering

C_bp = V’*c_smf;

% Plot out the steered response in decibels

figure
plot(phi,20*log10(abs(C_bp)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

(c) The steered response of the spatial matched filter is shown in Figure 11.3c. Clearly the 60-dB signal is
so strong that it leaks through sidelobes of all the beamformers used to compute the steered response and the
25 dB signal at φ = 30◦ is no longer visible. Using a tapered beamformer with a Chebyshev window with
−65-dB sidelobes is sufficient to suppress this 60-dB signal when not steered to φ = −10◦ , as seen in the
steered response in Figure 11.3c. Note, however, that a slight loss is incurred for both signals and that each
one spreads out slightly further in angle as opposed to the untapered spatial matched filter.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% The power levels of the signals and the noise

P1_db = 60;

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
Statistical and Adaptive Signal Processing - Solution Manual 383

w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Compute the true covariance matrix

% (remember to divide by 10 not 20 for powers in the covariance)
R_x = (10^(P1_db/10)) * v1 * v1’ + (10^(P2_db/10)) * v2 * v2’ + ...
(10^(Pn/10)) * eye(M);

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Use a tapered beamformer with 65 dB sidelobe levels

SLL_taper = 65;
t = chebwin(M,SLL_taper);
for m = 1:length(phi)
C_tbf(:,m) = diag(t)*C_smf(:,m);
C_tbf(:,m) = C_tbf(:,m)/sqrt(C_tbf(:,m)’*C_tbf(:,m));
end

% Compute the steered response

for k = 1:length(phi)
R_steer_smf(k) = mean(abs(C_smf(:,k)’ * x).^2,2);
R_steer_tbf(k) = mean(abs(C_tbf(:,k)’ * x).^2,2);
end

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_smf),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 65]);
grid
set(gca,’xtick’,-90:15:90);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_tbf),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 65]);
grid
set(gca,’xtick’,-90:15:90);
384 Statistical and Adaptive Signal Processing - Solution Manual

60 60

50 50
Steered Response (dB)

Steered Response (dB)

40 40

30 30

20 20

10 10

0 0

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg)

(b) Tapered beamformer (−65-dB

(a) Untapered beamformer. Chebyshev).

Figure 11.3c: Steered responses for scenario in Problem 11.3(c) with signals at φ = −10◦ and φ = 30◦ with
powers of 60 and 25 dB, respectively.

11.4 This problem is a Matlab based problem. Below is the Matlab code and six output figures. Note that for
this problem the signals should be at φ = 0◦ and φ = 3◦ . If your text does not have this correction, please
pass this along to the students. The code is broken up into parts (a), (b), (c), and (d) and presented sequentially
along with the accompanying figures.
(a) The two signal at φ = 0◦ and φ = 3◦ are not resolved in the steered response. The cause is the limited
aperture and the proximity of the two signals in angle.

%
% Problem 11.4
%
clear

M = 30; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% The angles in degrees of the two spatially propagating signals

phi1 = 0;
phi2 = 3;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 20;
Pn = 0;
Statistical and Adaptive Signal Processing - Solution Manual 385

25

20

Steered Response (dB)

15

10

−5
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.4a: Steered responses for scenario in Problem 11.4(a) for an array with M = 30 λ/2-spaced elements
with signals at φ = 0◦ and φ = 3◦ both with powers of 20 dB.

% Compute the array response vectors of the two signals

v1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

R_steer_est = mean(abs(C_smf’ * x).^2,2);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 25]);
grid
set(gca,’xtick’,-90:15:90);

(b) Now the two signals have been resolved in the steered response because we have doubled the aperture of
386 Statistical and Adaptive Signal Processing - Solution Manual

the array and thus improved its resolution by a factor of two.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Clear out the entire workspace and start with 60 elements
clear

M = 60; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples

% The angles in degrees of the two spatially propagating signals

phi1 = 0;
phi2 = 3;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 20;
Pn = 0;

% Compute the array response vectors of the two signals

v1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

R_steer_est = mean(abs(C_smf’ * x).^2,2);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
Statistical and Adaptive Signal Processing - Solution Manual 387

25

20

Steered Response (dB)

15

10

−5
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.4b: Steered responses for scenario in Problem 11.4(a) for an array with M = 60 λ/2-spaced elements
with signals at φ = 0◦ and φ = 3◦ both with powers of 20 dB.

axis([-90 90 -5 25]);
grid
set(gca,’xtick’,-90:15:90);

(c) Despite the fact that the array only has M = 30 elements as in part (a), we are able to resolve the two
signals. The reason is that the aperture for the M = 30 λ-spaced element array is equal to the aperture of the
M = 60 λ/2-spaced element. Note that in this case due to the increased element spacing we observe spatial
aliasing in the steered response in the form of the additional peaks at φ ≈ 75◦ and φ ≈ 90◦

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Clear out the entire workspace and start with 30 elements but with
% element spacing of lambda
clear

M = 30; % number of elements

spacing = 1; % element spacing in wavelengths
N = 1000; % number of samples

% The angles in degrees of the two spatially propagating signals

phi1 = 0;
phi2 = 3;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 20;
Pn = 0;

% Compute the array response vectors of the two signals

388 Statistical and Adaptive Signal Processing - Solution Manual

25

20

Steered Response (dB)

15

10

−5
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.4c: Steered responses for scenario in Problem 11.4(a) for an array with M = 60 λ-spaced elements with
signals at φ = 0◦ and φ = 3◦ both with powers of 20 dB.

v1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

R_steer_est = mean(abs(C_smf’ * x).^2,2);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 25]);
grid
set(gca,’xtick’,-90:15:90);

(d) The beampatterns for the spatial matched filter (beamformer) at φ = 0◦ used to compute the steered response
are plotted in Figure 11.4d. Increasing the number of λ/2-spaced elements from M = 30 to M = 60 increases
the resolution of the beamformer by a factor of two as is evidenced by a mainbeam with half the size. In the
Statistical and Adaptive Signal Processing - Solution Manual 389

case of the M = 30 λ-spaced element array, the mainbeam size is the same as for the M = 60 λ/2-spaced
element array since the aperture of the two arrays is the same. However, in the case of the λ element spacing,
we observe a grating lobe at φ = 90◦ for the beamformer steered to φ = 0◦ .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part D
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Clear out the entire workspace and compute the beampatterns for
% the spatial matched filters in parts a, b, and c steered to array broadside
clear

phi_steer = 0; % steering angle (degrees) for spatial matched filter

% Compute beamformer from part a

M = 30; % number of elements
spacing = 0.5; % element spacing in wavelengths
c_bf1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_steer*pi/180))/sqrt(M);

% The angles at which to evaluate the beampattern

phi = -90:0.2:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);
C_bp1 = V’*c_bf1; % beampattern

% Compute beamformer from part a

M = 60; % number of elements
spacing = 0.5; % element spacing in wavelengths
c_bf2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_steer*pi/180))/sqrt(M);

% The angles at which to evaluate the beampattern

phi = -90:0.2:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);
C_bp2 = V’*c_bf2; % beampattern

% Compute beamformer from part a

M = 30; % number of elements
spacing = 1; % element spacing in wavelengths
c_bf3 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_steer*pi/180))/sqrt(M);

% The angles at which to evaluate the beampattern

phi = -90:0.2:90;
V = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi*pi/180))/sqrt(M);
C_bp3 = V’*c_bf3; % beampattern

% Plot out the beampattern in decibels

figure
plot(phi,20*log10(abs(C_bp1)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
390 Statistical and Adaptive Signal Processing - Solution Manual

0 0 0

−5 −5 −5

−10 −10 −10

Power Response (dB)

Power Response (dB)

Power Response (dB)

−15 −15 −15

−20 −20 −20

−25 −25 −25

−30 −30 −30

−35 −35 −35

−40 −40 −40

−45 −45 −45

−50 −50 −50

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg) Angle (deg)

(a) M = 30 λ/2-spacing. (b) M = 60 λ/2-spacing. (c) M = 30 λ-spacing.

Figure 11.4d: Beampatterns for the spatial matched filters steered to φ = 0◦ with (a) M = 30 λ/2-spaced
elements, (b) M = 60 λ/2-spaced elements and (c) M = 30 λ-spaced elements.

ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);

set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

% Plot out the beampattern in decibels

figure
plot(phi,20*log10(abs(C_bp2)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

% Plot out the beampattern in decibels

figure
plot(phi,20*log10(abs(C_bp3)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

11.5 This problem is a Matlab based problem. Below is the Matlab code and four output figures. Note that for
this problem the signals should be at φ = 0◦ and φ = 3◦ . If your text in Problem 11.4 does not have this
correction, please pass this along to the students.

%
% Problem 11.5
%
clear
Statistical and Adaptive Signal Processing - Solution Manual 391

M0 = 60; % number of elements for a filled array

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples

% The angles in degrees of the two spatially propagating signals

phi1 = 0;
phi2 = 3;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 20;
Pn = 0;

% Generate the 45 randomly selected elements (75% thinning)

elements_used = 0.75;
M = elements_used*M0;
r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;

% Compute the array response vectors of the two signals

% (subtract one from elements to have first element be reference (0))
v1 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

R_steer_est = mean(abs(C_smf’ * x).^2,2);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 25]);
392 Statistical and Adaptive Signal Processing - Solution Manual

grid
set(gca,’xtick’,-90:15:90);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Repeat the same process for 50% array thinning
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
clear

M0 = 60; % number of elements for a filled array

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples

% The angles in degrees of the two spatially propagating signals

phi1 = 0;
phi2 = 3;

% The power levels of the signals and the noise

P1_db = 20;
P2_db = 20;
Pn = 0;

% Generate the 30 randomly selected elements (50% thinning)

elements_used = 0.50;
M = elements_used*M0;
r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;

% Compute the array response vectors of the two signals

% (subtract one from elements to have first element be reference (0))
v1 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi1*pi/180))/sqrt(M);
v2 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi2*pi/180))/sqrt(M);

% Generate the signals 1 and 2 and the noise

s1 = (10^(P1_db/20)) * v1 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
s2 = (10^(P2_db/20)) * v2 * ((randn(1,N) + j*randn(1,N))/sqrt(2));
w = (10^(Pn/20)) * (randn(M,N) + j*randn(M,N))/sqrt(2); % Noise

% Add signals and noise to get the total array signal

x = s1 + s2 + w;

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:0.2:90;
C_smf = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);

% Compute the steered response

Statistical and Adaptive Signal Processing - Solution Manual 393

R_steer_est = mean(abs(C_smf’ * x).^2,2);

% Plot out the steered response in decibels

figure
plot(phi,10*log10(R_steer_est),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -5 25]);
grid
set(gca,’xtick’,-90:15:90);

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Generate 3 randomly thinned arrays both at 50% and 75% and compute
% the beampatterns for the spatial matched filter at 0 degrees
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
clear

M0 = 60; % number of elements for a filled array

spacing = 0.5; % element spacing in wavelengths
phi_steer = 0; % steering direction for beamformer
phi = -90:0.2:90; % angles to compute the beampattern

% Generate the 30 randomly selected elements (75% thinning)

elements_used = 0.75;
M = elements_used*M0;

% Beamformer 1 with 75% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
c1_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
V1_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C1_75_bp = V1_75’*c1_75;

% Beamformer 2 with 75% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
c2_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
V2_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C2_75_bp = V2_75’*c2_75;

% Beamformer 3 with 75% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
394 Statistical and Adaptive Signal Processing - Solution Manual

c3_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
V3_75 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C3_75_bp = V3_75’*c3_75;

% Plot out the beampatterns in decibels

figure
plot(phi,10*log10(abs(C1_75_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(C2_75_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(C3_75_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

% Generate the 30 randomly selected elements (50% thinning)

elements_used = 0.50;
M = elements_used*M0;

% Beamformer 1 with 50% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
c1_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
V1_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C1_50_bp = V1_50’*c1_50;

% Beamformer 2 with 50% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
c2_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
Statistical and Adaptive Signal Processing - Solution Manual 395

V2_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C2_50_bp = V2_50’*c2_50;

% Beamformer 3 with 50% thinning and its beampattern

r = rand(M0,1);
[dummy,index] = sort(r);
elements = sort(index(1:(elements_used*M0)))’;
c3_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi_steer*pi/180))/sqrt(M);
V3_50 = exp(-j*2*pi*spacing*[elements-1]’*sin(phi*pi/180))/sqrt(M);
C3_50_bp = V3_50’*c3_50;

% Plot out the beampatterns in decibels

figure
plot(phi,10*log10(abs(C1_50_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(C2_50_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(C3_50_bp).^2),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Power Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -50 0]);
grid
set(gca,’xtick’,-90:15:90);

The steered responses for the case of 75% (45 element) and 50% (30 element) thinned arrays are shown in
Figures 11.5a and 11.5b. In both cases we are able to resolve the two signals at φ = 0◦ and φ = 3◦ since
the aperture is equal to the M = 60 λ/2-spaced array. Due to the reduced number of elements in the random
spacing the background levels in the steered response are raised. The beampatterns for the three different
random spacings for the 75% and 50% thinning are shown in Figures 11.5c and 11.5d. Here, we see that the
increase in the background level was due to the elevated sidelobes that get worse the more thinning is done. In
addition, we see that the sidelobes exhibit random behavior caused by the random thinning of the arrays.
396 Statistical and Adaptive Signal Processing - Solution Manual

25

20

Steered Response (dB)

15

10

−5
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.5a: Steered responses for scenario in Problem 11.4(a) for an array with M = 60 λ/2-spaced element
array thinned by 75% with signals at φ = 0◦ and φ = 3◦ both with powers of 20 dB.

25

20
Steered Response (dB)

15

10

−5
−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg)

Figure 11.5b: Steered responses for scenario in Problem 11.4(a) for an array with M = 60 λ/2-spaced element
array thinned by 50% with signals at φ = 0◦ and φ = 3◦ both with powers of 20 dB.

0 0 0

−5 −5 −5

−10 −10 −10

Power Response (dB)

Power Response (dB)

Power Response (dB)

−15 −15 −15

−20 −20 −20

−25 −25 −25

−30 −30 −30

−35 −35 −35

−40 −40 −40

−45 −45 −45

−50 −50 −50

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg) Angle (deg)

(a) 1st thinned array. (b) 2nd thinned array. (c) 3rd thinned array.

Figure 11.5c: Beampatterns for three different randomly thinned M = 60 λ/2-spaced element array thinned by
75% (M = 45 elements).
Statistical and Adaptive Signal Processing - Solution Manual 397

0 0 0

−5 −5 −5

Power Response (dB) −10 −10 −10

Power Response (dB)

Power Response (dB)

−15 −15 −15

−20 −20 −20

−25 −25 −25

−30 −30 −30

−35 −35 −35

−40 −40 −40

−45 −45 −45

−50 −50 −50

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg) Angle (deg)

(a) 1st thinned array. (b) 2nd thinned array. (c) 3rd thinned array.

Figure 11.5d: Beampatterns for three different randomly thinned M = 60 λ/2-spaced element array thinned by
50% (M = 30 elements).

11.6 We have an M = 20 element array whose even-numbered elements have a gain of 1 and whose odd-numbered
elements have a gain of 3/4. For an array whose elements all have equal gain, the beamforming gain is simply
the ratio of the array and element SNR
SNRarray
G bf =
SNRelement
Since the signal and noise powers at the array output are

Ps = (Mσ s )2 = M 2 σ 2s

 2 
M 
 
Pn = E  wm (n) = Mσ 2w
 
m=1

Then the SNR at the array output is

Ps σ2
SNRarray = = M 2s
Pn σw
while the element level SNR is
σ 2s
SNRelement =
σ 2w
so that the beamforming gain is given by

G bf = M

Now for the scenario described above with unequal gain on the elements of the array with half the elements
having a gain of 1 and half having a gain of 3/4. The signal power at the output of the array is

2
M M3
Ps = σs + σs
2 2 4

2
7
= Mσ s
8
49 2
= M
64
398 Statistical and Adaptive Signal Processing - Solution Manual

The noise power is as before

Pn = Mσ 2w
Therefore, the SNR at the array output is
Ps 49 σ 2s 49
SNRarray = = M 2 = M · SNRelement
Pn 64 σ w 64
and the array gain is
49
G bf = M
64
11.7 From (11.3.15), the optimum beamformer weight vector is given by
−1

Ri+n v φs
co =  −1 
v H φ s Ri+n v φs
As was pointed out in the text the optimum beamformer can also be formulated as the constrained minimization
problem in (11.3.17)

min c H Ri+n c subject to cH v φs = 1
we use Lagrange multipliers from Appendix B.2. First, we define the two functions

f (c) = c H Ri+n c g (c) = c H v φ s − 1
The Lagrangian function is then
L (c, λ) = f (c) + λg (c)
  
= c H Ri+n c + λ c H v φ s − 1
where λ is the Lagrange multiplier. We then want to minimize the Lagrangian function with respect to both c
and λ. Taking the gradient of L (c, λ) with respect to c and setting it to zero, we have

∇ c L (c, λ) = c H Ri+n + λv H φ s = 0

−1

c = −λRi+n v φs (1)
Similarly, taking the gradient of L (c, λ) with respect to λ and setting it to zero

∇ λ L (c, λ) = c H v φ s − 1 = 0 (2)
Substituting for (1) into (2), yields
 −1 
−λ v H φ s Ri+n v φs − 1 = 0
and solving for λ
−1
λ=  −1  (3)
vH φ s Ri+n v φ s
Substituting (3) back into (1), we have the minimum variance or optimum beamformer weight vector for the
look-direction φ s
−1

Ri+n v φs
co =  −1 
v H φ s Ri+n v φs
Statistical and Adaptive Signal Processing - Solution Manual 399

11.8 This problem is a Matlab based problem. Below is the Matlab code and one output figure. Use a total of
M = 30 λ/2-spaced elements if not stated in your text. If your text does not have this correction, please pass
this along to the students.

%
% Problem 11.8
%
clear

M = 30; % number of elements

spacing = 0.5; % element spacing in wavelengths

% The angles in degrees of the two spatial interference sources

phi_i1 = 45;
phi_i2 = 20;

% The power levels of the signals and the noise

P_i1_db = 30; % power of interference source 1 (dB)
P_i2_db = 15; % power of interference source 2 (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors of the two signals

v_i1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i1*pi/180))/sqrt(M);
v_i2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i2*pi/180))/sqrt(M);

% Compute the covariance matrix of the interference-plus-noise

R_in = (10^(P_i1_db/10))*v_i1*v_i1’ + (10^(P_i2_db/10))*v_i2*v_i2’ + ...
(sigma_w^2) * eye(M);

% Generate the spatial matched filters to compute the steered

% response for angles from -90 to 90 degrees
phi = -90:90;

for n = 1:length(phi)
v0 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi(n)*pi/180))/sqrt(M);
c_unnorm = R_in\v0;
norm_mvdr = 1/(c_unnorm’*v0);
norm_amf = 1/sqrt(c_unnorm’*v0);
norm_noise = 1/sqrt(v0’*(inv(R_in).^2)*v0);
c_mvdr = norm_mvdr * c_unnorm;
c_amf = norm_amf * c_unnorm;
c_noise = norm_noise * c_unnorm;
R_steer_mvdr(n) = c_mvdr’ * R_in * c_mvdr;
R_steer_amf(n) = c_amf’ * R_in * c_amf;
R_steer_noise(n) = c_noise’ * R_in * c_noise;
end

% Plot out the steered responses in decibels

figure
400 Statistical and Adaptive Signal Processing - Solution Manual

plot(phi,10*log10(abs(R_steer_mvdr)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -32 32]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-30:10:30);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(R_steer_amf)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -32 32]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-30:10:30);
grid
set(gca,’xtick’,-90:15:90);

figure
plot(phi,10*log10(abs(R_steer_noise)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Steered Response (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -32 32]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-30:10:30);
grid
set(gca,’xtick’,-90:15:90);

In this problem, the optimum beamformers with the three different normalizations are computed for the scenario
with two interference sources at φ = 45◦ and φ = 20◦ with powers of 30 and 15 dB, respectively. The steered
response is computed for the three different normalizations and shown in Figure 11.8a. The steered response
for the MVDR-normed optimum beamformer has the correct power estimates for both of the interference
sources as well as the noise background at 0 dB. Thus, the MVDR normalization generates a spatial power
spectrum. On the other hand, the output level of the AMF-normed optimum beamformer is constant by design
(unit-gain on interference-plus-noise). This normalization is particularly useful if we want to set a detection
threshold independent of angle to generate a constant probability of detection. The steered response for the
optimum beamformer with unit-gain on noise normalization produces the reciprocal of the MVDR-normed
steered response.

11.9

11.10 Starting with the correlation matrix with the signal at φ s with amplitude σ s present
 
Rx = Ri+n + σ 2s v φ s v H φ s (1)
Statistical and Adaptive Signal Processing - Solution Manual 401

30 30 30

Steered Response (dB) 20 20 20

Steered Response (dB)

Steered Response (dB)

10 10 10

0 0 0

−10 −10 −10

−20 −20 −20

−30 −30 −30

−90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90 −90 −75 −60 −45 −30 −15 0 15 30 45 60 75 90
Angle (deg) Angle (deg) Angle (deg)

(a) MVDR normalization. (b) AMF normalization. (c) Unit noise gain normalization.

Figure 11.8: Steered responses for scenario in Problem 11.8 for optimum beamformers with (a) MVDR , (b) AMF
and (c) unit-gain on noise normalizations.

we want to show that the optimum beamformer weights do not change when the signal is present in the case
of no mismatch, i.e.,
 −1

R−1
x v φs Ri+n v φs
co = H  −1  =  −1 
v φ s Rx v φ s v H φ s Ri+n v φs

Using the matrix inversion lemma from Appendix A, the inverse correlation matrix from (1) is
2 −1
 H  −1
−1 σ R v φ s v φ s Ri+n
R−1x = Ri+n −
s i+n
 −1 
1 + σ s v φ s Ri+n
2 H v φs

Multiplying by v φ s , we have
−1
  −1 
 −1
 σ 2s Ri+n v φ s v H φ s Ri+n v φs
R−1
φ s = Ri+n v φ s −
x v  −1 
1 + σ 2s v H φ s Ri+n v φ s
  −1  
−1
 σ 2s v H φ s Ri+n v φs
= Ri+n v φ s 1 −  −1 
1 + σ 2s v H φ s Ri+n v φ s
−1
 1
= Ri+n v φs  −1  (2)
1 + σ s v φ s Ri+n v φ s
2 H

  
Similarly, we can compute v H φ s R−1
x v φ s , by multiplying (1) by v φ s
 −1 
  v H φ s Ri+n v φs
v φ s R−1
H
x v φs =  −1  (3)
1 + σ 2s v H φ s Ri+n v φ s

Substituting (1) and (2) for the numerator and denominator of the optimum beamformer, we have

R−1
x v φs
co =  
v H φ s R−1x v φs
−1
  −1 
Ri+n v φs 1 + σ 2s v H φ s Ri+n v φs
=  −1  ·  −1 
1 + σ s v φ s Ri+n v φ s
2 H v φ s Ri+n v φ s
H

−1

R v φs
=  i+n −1 
v H φ s Ri+n v φs
402 Statistical and Adaptive Signal Processing - Solution Manual

Therefore, in the case of no signal mismatch, the optimum beamformer is not affected by the presence of the
signal of interest in the correlation matrix. Note that in practice it may not be possible to perfectly match the
signal of interest which can have significant performance implications.

11.11 This problem is a Matlab based problem. Below is the Matlab code and two output figures. The code is
broken up into parts (a) and (b) and presented sequentially along with the accompanying figures.
(a) In this part of the problem, we examine the effect of mismatch when the signal is not in the correlation
matrix. The output powers for the cases of no mismatch (signal at φ = 0◦ ) and mismatch in angle (signal at
φ = −1◦ ) shown by the dashed and solid lines in Figure 11.11a. The mismatch loss is simply the difference
between these two lines in decibels (or their ratio). From the figure, we observe an approximate loss of −3
dB for all angles independent of the signal-to-noise ratio of the signal since the signal is not included in the
correlation matrix. This result can be compared to that predicted by (11.4.12).

%
% Problem 11.11
%
clear

M = 50; % number of elements

spacing = 0.5; % element spacing in wavelengths

% The angles in degrees of the three spatially interference sources

phi_i1 = 5;
phi_i2 = 20;
phi_i3 = -30;

% The power levels of the signals and the noise

P_i1_db = 25; % power of interference source 1 (dB)
P_i2_db = 30; % power of interference source 2 (dB)
P_i3_db = 50; % power of interference source 3 (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors of the three signals

v_i1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i1*pi/180))/sqrt(M);
v_i2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i2*pi/180))/sqrt(M);
v_i3 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i3*pi/180))/sqrt(M);

% Compute the covariance matrix of the interference-plus-noise

R_in = (10^(P_i1_db/10))*v_i1*v_i1’ + (10^(P_i2_db/10))*v_i2*v_i2’ + ...
(10^(P_i3_db/10))*v_i3*v_i3’ + (sigma_w^2) * eye(M);

% Specify the signal of interest

phi_s = -1; % angle of signal of interest (deg)
SNR_db = 0:30; % signal-to-noise ratios (dB)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
Statistical and Adaptive Signal Processing - Solution Manual 403

35

30

25

Output SNR (dB)

20

15

10

−5
0 5 10 15 20 25 30
Signal SNR (dB)

Figure 11.11a: Output power as a function of signal-to-noise ratio for the case of no mismatch (φ = 0◦ , dashed
line) and angle mismatch (φ = −1◦ , solid line) when the signal is not contained in the correlation matrix.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Compute the optimum beamformer to the desired angle
phi0 = 0; % angle to which we are steering (deg)
v0 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi0*pi/180))/sqrt(M);
c_unnorm = R_in\v0; % unnormalized optimum beamformer
c_opt = c_unnorm/(v0’*c_unnorm); % MVDR norm on optimum beamformer

% Loop on SNR to compute the signal mismatch losses

for n = 1:length(SNR_db)
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
x_s = (10^(SNR_db(n)/20))*v_s; % actual array signal at SNR
x0 = (10^(SNR_db(n)/20))*v0; % array signal at the steering angle
P_s(n) = abs(c_opt’*x_s)^2; % output power of actual signal
P0(n) = abs(c_opt’*x0)^2; % output power for signal at phi0
L_sm(n) = P_s(n)/P0(n); % mismatch loss
end

% Plot out the output SNR in decibels vs. the signal SNR
figure
plot(SNR_db,10*log10(abs(P_s)),’b’,...
SNR_db,10*log10(abs(P0)),’r--’,’linewidth’,2.5);
xlabel(’Signal SNR (dB)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output SNR (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 30 -5 35]);
grid

(b) In this part of the problem, we examine the effect of mismatch when the signal is in the correlation matrix.
Here, we observe that, as predicted, the loss is strongly dependent on the signal-to-noise ratio. The stronger
the signal, the larger the loss incurred is. This loss should be compared to the loss predicted by (11.4.17). Note
that the loss computed and plotted in this problem is combination of the mismatch losses L sm and L sp .
404 Statistical and Adaptive Signal Processing - Solution Manual

−10

−20

Loss (dB)
−30

−40

−50

−60

0 5 10 15 20 25 30
Signal SNR(dB)

Figure 11.11b: Mismatch loss versus signal-to-noise ratio for a mismatched steering direction (φ = −1◦ ) when the
signal is contained in the correlation matrix .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Loop on SNR to compute the signal mismatch losses

for n = 1:length(SNR_db)
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
x_s = (10^(SNR_db(n)/20))*v_s; % actual array signal at SNR
x0 = (10^(SNR_db(n)/20))*v0; % array signal at the steering angle
R_x = R_in + x_s*x_s’; % correlation matrix containing signal
c_unnorm = R_x\v0; % unnormalized optimum beamformer
c = c_unnorm/(v0’*c_unnorm); % MVDR norm on optimum beamformer
P_s(n) = abs(c’*x_s)^2; % output power of actual signal
P0(n) = abs(c’*x0)^2; % output power for signal at phi0
L_mismatch(n) = P_s(n)/P0(n); % mismatch loss (losses due to
% SINR loss, signal present in
% correlation matrix, and mismatch
end

% Plot out the mismatch loss versus the signal SNR in decibels
figure
plot(SNR_db,10*log10(L_mismatch),’b’,’linewidth’,2.5);
xlabel(’Signal SNR(dB)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Loss (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 30 -65 0]);
grid

11.12 This problem is a Matlab based problem. Below is the Matlab code and two output figures. The code is
broken up into parts (a) and (b) and presented sequentially along with the accompanying figures. The signal
Statistical and Adaptive Signal Processing - Solution Manual 405

level should be Mσ 2s = 20 dB instead of 30 dB. In both parts (a) and (b), the problem should read: ”Compute
and plot the mismatch loss for diagonal loading levels of 0 to 20 dB". If your text does not have this correction,
please pass this along to the students.
(a) In this part of the problem, we examine the effect of diagonal loading on the mismatch loss when the signal
is not in the correlation matrix. The output powers for the cases of no mismatch (signal at φ = 0◦ , dashed
line) and mismatch in angle (signal at φ = −1◦ , solid line) are plotted in Figure 11.12a. The mismatch loss
is simply the difference between these two lines in decibels (or their ratio). From the figure, we observe an
approximate loss of −3 dB for all diagonal loading levels since the signal is not included in the correlation
matrix.

%
% Problem 11.12
%
clear

M = 50; % number of elements

spacing = 0.5; % element spacing in wavelengths

% The angles in degrees of the three spatially interference sources

phi_i1 = 5;
phi_i2 = 20;
phi_i3 = -30;

% The power levels of the signals and the noise

P_i1_db = 25; % power of interference source 1 (dB)
P_i2_db = 30; % power of interference source 2 (dB)
P_i3_db = 50; % power of interference source 3 (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors of the three signals

v_i1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i1*pi/180))/sqrt(M);
v_i2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i2*pi/180))/sqrt(M);
v_i3 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i3*pi/180))/sqrt(M);

% Compute the covariance matrix of the interference-plus-noise

R_in = (10^(P_i1_db/10))*v_i1*v_i1’ + (10^(P_i2_db/10))*v_i2*v_i2’ + ...
(10^(P_i3_db/10))*v_i3*v_i3’ + (sigma_w^2) * eye(M);

% Specify the signal of interest

phi_s = -1; % angle of signal of interest (deg)
SNR_db = 20; % signal-to-noise ratios (dB)

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Compute the optimum beamformer to the desired angle
phi0 = 0; % angle to which we are steering (deg)
406 Statistical and Adaptive Signal Processing - Solution Manual

35

30

25

Output SNR (dB)

20

15

10

−5
0 5 10 15 20 25 30
Diagonal Loading Level (dB)

Figure 11.12a: Output power as a function of diagonal loading level for the case of no mismatch (φ = 0◦ , dashed
line) and angle mismatch (φ = −1◦ , solid line) when the signal is not contained in the correlation matrix.

v0 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi0*pi/180))/sqrt(M);

% Diagonal loading levels

DLL_db = 0:30; % diagonal loading levels (dB)

% Loop on SNR to compute the signal mismatch losses

for n = 1:length(DLL_db)
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
x_s = (10^(SNR_db/20))*v_s; % actual array signal at SNR
x0 = (10^(SNR_db/20))*v0; % array signal at the steering angle
c_dl_unnorm = (R_in + (10^(DLL_db(n)/10))*eye(M))\v0;
c_dl = c_dl_unnorm/(c_dl_unnorm’*v0);
P_s(n) = abs(c_dl’*x_s)^2; % output power of actual signal
P0(n) = abs(c_dl’*x0)^2; % output power for signal at phi0
L_sm(n) = P_s(n)/P0(n); % mismatch loss
end

% Plot out the output powers versus diagonal loading level in decibels
figure
plot(DLL_db,10*log10(abs(P_s)),’b’,...
DLL_db,10*log10(abs(P0)),’r--’,’linewidth’,2.5);
xlabel(’Diagonal Loading Level (dB)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output SNR (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 30 -5 35]);
grid

(b) In this part of the problem, we examine the effect of diagonal loading on mismatch loss when the signal is
in the correlation matrix. We observe that the more heavily we diagonally load the more signal we are able to
protect from mismatch loss. This protection from mismatch loss when the signal is in the correlation matrix is
achieved at the expense of interference cancellation performance in the case of diagonal loading. Note that the
loss computed and plotted in this problem is combination of the mismatch losses L sm and L sp . Therefore, the
Statistical and Adaptive Signal Processing - Solution Manual 407

mismatch loss can never exceed the -3-dB level due to the steering direction mismatch loss that was observed
in part (a).

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Loop on SNR to compute the signal mismatch losses

for n = 1:length(DLL_db)
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
x_s = (10^(SNR_db/20))*v_s; % actual array signal at SNR
x0 = (10^(SNR_db/20))*v0; % array signal at the steering angle
R_x = R_in + x_s*x_s’; % correlation matrix containing signal
c_dl_unnorm = (R_x + (10^(DLL_db(n)/10))*eye(M))\v0;
c_dl = c_dl_unnorm/(c_dl_unnorm’*v0);
P_s(n) = abs(c_dl’*x_s)^2; % output power of actual signal
P0(n) = abs(c_dl’*x0)^2; % output power for signal at phi0
L_mismatch(n) = P_s(n)/P0(n); % mismatch loss (losses due to
% SINR loss, signal present in
% correlation matrix, and mismatch
end

% Plot out the mismatch loss versus diagonal loading level in decibels
figure
plot(DLL_db,10*log10(L_mismatch),’b’,’linewidth’,2.5);
xlabel(’Diagonal Loading Level (dB)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Loss (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 30 -65 0]);
grid

11.13 The linearly constrained minimum variance (LCMV) beamformer weights clcmv are found by solving the
constrained minimization from (11.6.5)
min c H Rx c subject to CH c = δ
where C is the constraint matrix and δ is the vector of corresponding responses to the constraints. Using
Lagrange multiplers from Appendix B.2, the LCMV beamformer weights are found to be
 H −1 −1
clcmv = R−1
x C C Rx C δ
which minimizes the Lagrangian function. Using Lagrange multiplers from Appendix B.2, we define the
following two functions
f (c) = c H Rx c g (c) = C H c − δ
The Lagrangian function is then given by
L (c, λ) = f (c) + λ H g (c)
 
= c H Rx c + λ H C H c − δ
408 Statistical and Adaptive Signal Processing - Solution Manual

−10

−20

Loss (dB)
−30

−40

−50

−60

0 5 10 15 20 25 30
Diagonal Loading Level (dB)

Figure 11.12b: Mismatch loss versus diagonal loading level for a mismatched steering direction (φ = −1◦ ) when
the signal is contained in the correlation matrix.

where λ is the vector of Lagrange multipliers whose length is equal to the number of constraints (same length
as the constraint response vector δ). Taking the gradient of the Lagrangian function with respect to c, we have

∇ c L (c, λ) = Rx c + Cλ

For the method of steepest descent (Frost’s algorithm), the adaptive weights at time n+1 are found by modifying
the weights at time n by the gradient of the Lagrangian function scaled by the update factor µ

c (n + 1) = c (n) − µ∇ c L (c, λ)
= c (n) − µ [Rx c (n) + Cλ (n)] (1)

To satisfy the constraint at time n + 1, we require that

δ = C H c (n + 1) = C H c (n) − µC H Rx c (n) − µC H Cλ (n)

Solving for λ (n), we have

1  H −1 H  −1 1  H −1
λ (n) = C C C c (n) − C H C C H Rx c (n) − C C δ (2)
µ µ
Substituting the solution for λ (n) from (2) back into the weight update equation in (1), we have

c (n + 1) = c (n) − µRx c (n)

 
1  H −1 H  H −1 H 1  H −1
−µC C C C c (n) − C C C Rx c (n) − C C δ
µ µ
 −1 H
= c (n) − µRx c (n) + µC C H C C Rx c (n)
 H −1 H  H −1
−C C C C c (n) + C C C δ
$  −1 %  −1  
= c (n) − µ I − C C H C C H Rx c (n) + C C H C δ − C H c (n) (3)

Let us now define the non-adaptive portion of the weight update equation
 −1
cna = C C H C δ
Statistical and Adaptive Signal Processing - Solution Manual 409

and the projection matrix

 −1 H
P = I − C CH C C
The weight update equation from (3) can then be written as
 −1
c (n + 1) = c (n) − µPRx c (n) − C C H C C H c (n) + cna
= cna + P [c (n) − µRx c (n)] (4)
Equation (4) is a deterministic constrained gradient descent algorithm requiring knowledge of the correlation
matrix Rx . Since Rx is unknown for any given time n, we can substitute the estimate
?
Rx = x (n) x H (n)
into (4) to obtain the weight update equation for Frost’s algorithm with multiple linear constraints
 
c (n + 1) = cna + P c (n) − µx (n) x H (n) c (n)
 
= cna + P c (n) − µx (n) y ∗ (n) (5)
since y (n) = c H (n) x (n).
11.14 The linearly constrained minimum variance (LCMV) beamformer weights clcmv are found by solving the
constrained minimization from (11.6.5)
min c H Ri+n c subject to CH c = δ
where C is the constraint matrix and δ is the vector of corresponding responses to the constraints. Using
Lagrange multiplers from Appendix B.2, we define the following two functions
f (c) = c H Ri+n c g (c) = C H c − δ
The Lagrangian function is then given by
L (c, λ) = f (c) + λ H g (c)
 
= c H Ri+n c + λ H C H c − δ
where λ is the vector of Lagrange multipliers whose length is equal to the number of constraints (same length
as the constraint response vector δ). Taking the gradient of the Lagrangian function with respect to c, we have
∇ c L (c, λ) = Ri+n c + Cλ
Setting the gradient to zero in order to minimize the Lagrangian
−1
c = −Ri+n Cλ (1)
Now, we simply need to solve for the vector of Lagrange multipliers. Taking the gradient of the Lagrangian
function with respect to λ
∇ λ L (c, λ) = C H c − δ
Setting this equal to zero, we have
CH c = δ (2)
and substituting for c from (1) into (2) and solving for λ
 −1
−1
λ = − C H Ri+n C δ (3)
Substituting (3) back into (1), we have the LCMV beamformer weight vector
−1
 −1
−1
clcmv = Ri+n C C H Ri+n C δ
410 Statistical and Adaptive Signal Processing - Solution Manual

11.15 In this problem, we study the performance of the sidelobe canceler. Let us start by defining the main channel
signal as

xmc = gs s (n) + i mc (n) + wmc (n)

where i mc (n) = gi i (n) is the interference signal i (n) with a main channel gain gi . The interference signal is
assumed to be temporally uncorrelated with unit variance. The main channel thermal noise signal wmc (n) is
also temporally uncorrelated and has a variance of σ 20 . Likewise, the signal in the auxiliary channel, consisting
of an M-element λ/2-spaced ULA, is given by
 
xa (n) = s (n) v φ s + σ i i (n) v φ i + w (n)

where w (n) is the the thermal noise signal in the array with a variance of σ 2w in each channel (element).

(a) The auxiliary channel correlation matrix is given by


Ra = E xa (n) xaH (n)
= Ps v (φ s ) v H (φ s ) + σ 2i v (φ i ) v H (φ i ) + σ 2w I

where Ps = |s (n)|2 . The cross-correlation vector between the auxiliary channels and the main channel
is
 ∗
rma = E xa (n) xmc (n)
= Ps gs∗ v (φ s ) + σ i gi∗ v (φ i )

(b) The output signal from the sidelobe canceler is given by

y (n) = xmc (n) − caH xa (n)

The interference-plus-noise portion of this signal, however, is given by

yi+n (n) = i i+n (n) + wi+n (n) − σ i i (n) caH v (φ i ) − caH w (n)

The output power of the interference-plus-noise is thus


Pi+n = E |yi+n (n)|2
 2 
= gi2 + σ 20 + σ 2i caH v (φ i ) + σ 2w caH ca − 2 · Real σ i gi∗ caH v (φ i ) (1)

Recall that

Ra = Ri+n + Ps v (φ s ) v H (φ s )

Here, we make the assumption that Ps << σ 2w < σ 2i so that the correlation matrix and cross-correlation
vector from part (a) are

Ra ≈ Ri+n

rma ≈ σ i gi∗ v (φ i )

Using the matrix inversion lemma from Appendix A, we have

 
−1 1 1/σ 2w I · σ 2i v (φ i ) v H (φ i ) 1/σ 2w I
Ra = I− 
σ 2w 1 + σ 2i v H (φ i ) 1/σ 2w I · v (φ i )
 
1 σ 2i v (φ i ) v H (φ i )
= I− (2)
σ 2w σ 2w + σ 2i
Statistical and Adaptive Signal Processing - Solution Manual 411

Therefore, the adaptive weights are given by

−1
ca ≈ Ri+n σ i gi∗ v (φ i ) (3)

Substituting (2) and (3) back into (1), we find the interference-plus-noise output power is
 2
Pi+n = gi2 + σ 20 + σ 2i gi∗  v H (φ i ) Ri+n
−1
Ri+n Ri+n−1
v (φ i )
"  #
−2 · Real σ i gi∗  v H (φ i ) Ri+n −1
2
v (φ i )
 
= gi2 + σ 20 + σ 2i gi∗  v H (φ i ) Ri+n
−1
2
v (φ i )


σ 2
= σ 20 + 1 − 2 i 2 gi2 (4)
σi + σw

using some algebraic manipulation and the fact that v H (φ i ) v (φ i ) = 1.

(c) The output power of the signal from the sidelobe canceler is

ys (n) = gs s (n) − caH v (φ s ) s (n)

where

ca = Ra−1 rma (5)

The output signal power is then

 2
Ps = σ 2s · gs − caH v (φ s ) (6)

Recall the expressions for the correlation matrix Ra and the cross-correlation vector rma from part (a).
Using the matrix inversion lemma from Appendix A, we have
−1 2 −1
−1 Ri+n σ s v (φ s ) v H (φ s ) Ri+n
Ra−1 = Ri+n + −1
(7)
1 + σ 2s v H (φ s ) Ri+n v (φ s )

Also using the matrix inversion lemma on the interference-plus-noise correlation matrix as was done for
part (b), we have
 
−1 1 σ 2i v (φ i ) v H (φ i )
Ri+n = 2 I − (8)
σw σ 2w + σ 2i

Substituting (8) into (7), and (7) into (5), and finally (5) into (6), along with some algebraic manipulation,
we have the expression for the output signal power

σ 2i + σ 2w
Ps =   ·
σ2 
σ 2i + σ 2w + σ 2s 1 + σ 2i 1 − v H (φ i ) v H (φ s )
w
 
σ 2    σ2    2
|gs |2 − 2 2 i 2 Real gi gs∗ v H φ i v φ s + 2 i 2 |gi |2 v H φ i v φ s  (9)
σi + σw σi + σw

The first term in (8) represents the signal cancellation due to its inclusion in the auxiliary channels. One can
see that this term has a maximum value of unit when σ 2s = 0. The second term in the expression for the output
signal power represents the signal gain due to interference cancellation of the sidelobe canceler.
412 Statistical and Adaptive Signal Processing - Solution Manual

11.16 (a) From (11.6.7), the LCMV beamformer weight vector is

−1
 −1
−1
clcmv = Ri+n C C H Ri+n C δ

Now for the minimum variance distortionless

 response (MVDR) beamformer, the constraint matrix is
the single look-direction constraint C = v φ s and the constraint response vector is the scalar δ = 1.
Substituting these values into the LCMV beamformer weights, we have

−1
   H −1  −1
cmvdr = Ri+n v φ s v φ s Ri+n v φs
−1

Ri+n v φs
=  H −1 
v φ s Ri+n v φ s
= co

(b) The output interference-plus-noise power for any beamformer c is simply

Pi+n = c H Ri+n c

Using the LCMV beamformer, we have

Pi+n = clcmv
H
Ri+n clcmv
 H −1 −1 H −1 −1
 −1
−1
= δ C Ri+n C C Ri+n Ri+n Ri+n
H
C C H Ri+n C δ
 −1
−1
= δ H C H Ri+n C δ

11.17 As with the minimum variance distortionless response (MVDR) beamformer, the linearly constrained minimum
variance (LCMV) beamformer can also be implemented using a generalized sidelobe canceler (GSC). Recall
the constrained minimization for the LCMV beamformer from (11.6.1)

min c H Ri+n c subject to CH c = δ

where C is the M × K matrix consisting of K constraint vectors δ is the K ×1 vector of constraint responses.
  and
Recall that for the single constraint MVDR beamformer v φ s B formed a basis for the M-dimensional
space in which the beamforming weight vector lies.  B is the M × (M − 1) blocking matrix that is orthogonal
to the look-direction constraint vector, i.e., B H v φ s = 0. In the case of the multiple constraints in the LCMV
beamformer with K constraints, [C B] must span the M-dimensional space for the LCMV beamformer weight
vector and, thus, the blocking matrix B is a M × (M − K ) matrix. Again, we require that B H C = 0. As with
the GSC for the MVDR beamformer, the GSC for the LCMV beamformer consists of an upper branch that
satisfies the constraint using a quiescent or non-adaptive weight vector cq and a lower branch with the blocking
matrix B and the unconstrained adaptive weight vector c B . The overall weight vector can then be written as

c = cq − Bc B (1)

where the quiescent weight vector

cq = Cc̃

is a linear combination of the K constraints, i.e. it must lie in the constraint subspace. Requiring (1) to satisfy
the LCMV constraint C H c = δ, we find that

C H cq = C H Cc̃ = δ
Statistical and Adaptive Signal Processing - Solution Manual 413

since B is orthogonal to C. Solving for c̃

 −1
c̃ = C H C δ (2)

and the quiescent weight vector or the non-adaptive portion in the upper branch of the GSC is
 −1
cq = Cc̃ = C C H C δ

Now, looking at the unconstrained minimization that computes an adaptive weight vector c B in the lower branch
of the GSC that minimizes the overall output power Pi+n given by
 H 
Pi+n = cq − Bc B Ri+n cq − Bc B
= cqH Ri+n cq − cqH Ri+n Bc B − c BH B H Ri+n cq + c BH B H Ri+n Bc B
 −1 H
If we complete the square by adding and subtracting the term cqH Ri+n B B H Ri+n B B Ri+n cq from Pi+n , we
have
$  −1 %H $  −1 %
Pi+n = c B − B H Ri+n B B H Ri+n Bcq B H Ri+n B c B − B H Ri+n B B H Ri+n Bcq
 −1
+cqH Ri+n cq − cqH Ri+n B B H Ri+n B B H Ri+n cq (3)

If we minimize Pi+n with respect to c B by taking the derivative with respect to c B and setting it to zero, we see
that the last two terms in (3) drop out since they are independent of c B . Thus ,the adaptive weight vector in the
lower branch of the GSC is
 −1 H
c B = B H Ri+n B B Ri+n cq (4)

Note that correlation matrix in the lower branch following the blocking matrix is

R B = E x B (n) x BH (n) = B H Ri+n B

and the cross-correlation vector between the lower branch of the GSC and the upper branch is

r B = E x B (n) y0∗ (n)

= E B H x (n) x H (n) cq
= B H Ri+n cq

Using these expressions, we can put the adaptive weight vector for the GSC of the LCMV beamformer from
(4) in the familiar form from (11.3.42)

c B = R−1
B rB

11.18 This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a), (b) and (c) and presented sequentially along with the accompanying figures. Note that
the interference scenario is not stressing enough to illustrate the fact that too few degrees of freedom (number
of subarrays) will compromise interference nulling capability. See the new problem 11.22 that is not included
in your textbook but instead found in this solution manual. This problem, however, is useful for students to
learn how to beamform using subarrays.
(a) The signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The M = 20 element array has sufficient
degrees of freedom to cancel the interferer at φ = 30◦ .
414 Statistical and Adaptive Signal Processing - Solution Manual

%
% Problem 11.18
%
clear

M = 20; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples in the realization

% The angle, power, and sample number of the signal of interest

phi_s = 0; % angle (deg)
P_s_db = 20; % signal power level (dB)
N_s = 100; % sample number of signal

% The angle in degrees of the single interference source

phi_i = 30;

% The power levels of the signals and the noise

P_i_db = 50; % power of interference source (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors for the signal of interest and
% the interference
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
v_i = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i*pi/180))/sqrt(M);

% Generate one realization of the signal of interest, the interference,

% and the noise
x_s = zeros(M,N);
x_s(:,N_s) = (10^(P_s_db/20))*v_s;
x_i = (10^(P_i_db/20))*v_i*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_n = (randn(M,N) + i*randn(M,N))/sqrt(2);

% Combine to get overall signal and interference-plus-noise signal

x = x_s + x_i + x_n;
x_in = x_i + x_n;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Estimate the covariance matrix of the interference-plus-noise

R_in = (1/N)*x_in*x_in’;

% Compute SMI beamformer and normalize to MVDR

c_opt_unnorm = R_in\v_s;
c_opt = c_opt_unnorm/(c_opt_unnorm’*v_s);
Statistical and Adaptive Signal Processing - Solution Manual 415

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.18a: Output signal for the full array SMI adaptive beamformer with M = 20 elements.

% Compute the output signal

y_out = c_opt’*x;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(b) As in part (a), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The Q = 4 subarrays or degrees
of freedom are sufficient to cancel the single interferer at φ = 30◦ .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Q = 4; % number of non-overlapping subarrays
M_sub = 5; % number of elements per subarray

% Compute the subarray transformation matrix

v_sub = exp(-j*2*pi*spacing*[0:(M_sub-1)]’*sin(phi_s*pi/180))/sqrt(M_sub);
T_sub = zeros(M,Q);
for q = 1:Q
index = (1:M_sub) + (q-1)*M_sub;
T_sub(index,q) = v_sub;
end
416 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.18b: Output signal for the SMI partially adaptive beamformer with Q = 4 nonoverlapping subarrays
with M̃ = 5 elements.

% Compute the subarray output signals

x_sub = T_sub’*x; % overall signal
x_in_sub = T_sub’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_sub = (1/N)*x_in_sub*x_in_sub’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_sub = T_sub’*v_s;
c_opt_sub_unnorm = R_in_sub\v_s_sub;
c_opt_sub = c_opt_sub_unnorm/(c_opt_sub_unnorm’*v_s_sub);

% Compute the output signal

y_out_sub = c_opt_sub’*x_sub;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_sub)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(c) As in parts (a) and (b), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The Q = 2 subarrays
or degrees of freedom are sufficient to cancel the single interferer at φ = 30◦ . The interference scenario is not
stressing enough to illustrate the fact that too few degrees of freedom will limit the ability to cancel interferers.
See problem 11.22 in this solutions manual for a more stressing case.
Statistical and Adaptive Signal Processing - Solution Manual 417

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Q = 2; % number of non-overlapping subarrays
M_sub = 10; % number of elements per subarray

% Compute the subarray transformation matrix

v_sub = exp(-j*2*pi*spacing*[0:(M_sub-1)]’*sin(phi_s*pi/180))/sqrt(M_sub);
T_sub = zeros(M,Q);
for q = 1:Q
index = (1:M_sub) + (q-1)*M_sub;
T_sub(index,q) = v_sub;
end

% Compute the subarray output signals

x_sub = T_sub’*x; % overall signal
x_in_sub = T_sub’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_sub = (1/N)*x_in_sub*x_in_sub’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_sub = T_sub’*v_s;
c_opt_sub_unnorm = R_in_sub\v_s_sub;
c_opt_sub = c_opt_sub_unnorm/(c_opt_sub_unnorm’*v_s_sub);

% Compute the output signal

y_out_sub = c_opt_sub’*x_sub;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_sub)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

11.19 This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a), (b) and (c) and presented sequentially along with the accompanying figures. Note that
the interference scenario is not stressing enough to illustrate the fact that too few degrees of freedom (number
of beams) will compromise interference nulling capability. See the new problem 11.23 that is not included in
your textbook but instead found in this solution manual. This problem, however, is useful for students to learn
how to implement an adaptive beamformer in beamspace which is commonly done in practical applications.
418 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.18c: Output signal for the SMI partially adaptive beamformer with Q = 2 nonoverlapping subarrays
with M̃ = 10 elements.

(a) The signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The M = 40 element array has sufficient
degrees of freedom to cancel the interferer at φ = 20◦ .

%
% Problem 11.19
%
clear

M = 40; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples in the realization

% The angle, power, and sample number of the signal of interest

phi_s = 0; % angle (deg)
P_s_db = 20; % signal power level (dB)
N_s = 100; % sample number of signal

% The angle in degrees of the interference source

phi_i = 20;

% The power levels of the signals and the noise

P_i_db = 50; % power of interference source (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors for the signal of interest and
% the interference
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
v_i = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i*pi/180))/sqrt(M);

% Generate one realization of the signal of interest, the interference,

% and the noise
x_s = zeros(M,N);
Statistical and Adaptive Signal Processing - Solution Manual 419

x_s(:,N_s) = (10^(P_s_db/20))*v_s;
x_i = (10^(P_i_db/20))*v_i*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_n = (randn(M,N) + i*randn(M,N))/sqrt(2);

% Combine to get overall signal and interference-plus-noise signal

x = x_s + x_i + x_n;
x_in = x_i + x_n;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Estimate the covariance matrix of the interference-plus-noise

R_in = (1/N)*x_in*x_in’;

% Compute SMI beamformer and normalize to MVDR

c_opt_unnorm = R_in\v_s;
c_opt = c_opt_unnorm/(c_opt_unnorm’*v_s);

% Compute the output signal

y_out = c_opt’*x;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(b) As in part (a), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The 11 beams or degrees of
freedom are sufficient to cancel the single interferer at φ = 20◦ .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Angles of beam in beamspace
phi_bs = -5:5;
B = length(phi_bs);

% Compute the beamspace transformation matrix

T_bs = zeros(M,B);
420 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.19a: Output signal for the full array SMI adaptive beamformer with M = 40 elements.

for b = 1:B
v_bs = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_bs(b)*pi/180))/sqrt(M);
T_bs(1:M,b) = v_bs;
end

% Compute the beamspace output signals

x_bs = T_bs’*x; % overall signal
x_in_bs = T_bs’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_bs = (1/N)*x_in_bs*x_in_bs’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_bs = T_bs’*v_s;
c_opt_bs_unnorm = R_in_bs\v_s_bs;
c_opt_bs = c_opt_bs_unnorm/(c_opt_bs_unnorm’*v_s_bs);

% Compute the output signal

y_out_bs = c_opt_bs’*x_bs;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_bs)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(c) As in parts (a) and (b), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The 3 beams or degrees
of freedom are sufficient to cancel the single interferer at φ = 20◦ The interference scenario is not stressing
Statistical and Adaptive Signal Processing - Solution Manual 421

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.19b: Output signal for the SMI partially adaptive beamformer with 11 beams between −5◦ ≤ φ ≤ 5◦ at
1◦ increments.

enough to illustrate the fact that too few degrees of freedom will limit the ability to cancel interferers. See
problem 11.23 in this solutions manual for a more stressing case.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
clear T_bs

% Angles of beam in beamspace

phi_bs = -1:1;
B = length(phi_bs);

% Compute the beamspace transformation matrix

T_bs = zeros(M,B);
for b = 1:B
v_bs = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_bs(b)*pi/180))/sqrt(M);
T_bs(1:M,b) = v_bs;
end

% Compute the beamspace output signals

x_bs = T_bs’*x; % overall signal
x_in_bs = T_bs’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_bs = (1/N)*x_in_bs*x_in_bs’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_bs = T_bs’*v_s;
c_opt_bs_unnorm = R_in_bs\v_s_bs;
c_opt_bs = c_opt_bs_unnorm/(c_opt_bs_unnorm’*v_s_bs);
422 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.19c: Output signal for the SMI partially adaptive beamformer with 3 beams at φ = −1◦ , 0◦ , and1◦ .

% Compute the output signal

y_out_bs = c_opt_bs’*x_bs;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_bs)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

11.20 Starting with the SINR loss for a beamformer steered to φ s from (11.3.18)

SINRout φ s
L sinr = (1)
SNR0

where SINRout φ s is the output SINR at φ s and SNR0 is the output signal-to-noise ratio in the absence of
interference (noise-only) given by

σ 2s
SNR0 = M (2)
σ 2w

The weight vector for the partially adaptive beamformer with a rank-reducing transformation T in a Q-
dimensional subspace is given by
−1

c̃ = α R̃i+n ṽ φ s

where α is the normalization constant and

R̃i+n = T H Ri+n T
Statistical and Adaptive Signal Processing - Solution Manual 423

 
ṽ φ s = T H v φ s

are the interference-plus-noise correlation matrix and steering vector projected into the Q-dimensional sub-
space. If we include the rank-reducing transformation, the resulting M × 1 beamforming weight vector is

cpa = Tc̃

Therefore, the output SINR of the beamformer is

  2
 H
cpa σ s v φ s 
SINRout = HR
cpa i+n cpa

Mσ 2s
=
c̃ H T H Ri+n Tc̃
Mσ 2s
=
˜ c̃
c̃ H Ri+n
Mσ 2    2
−1
=  −1 s −1  · ṽ H φ s R̃i+n ṽ φ s
ṽ H φ s R̃i+n R̃i+n R̃i+n ṽ φ s
 −1 
= Mσ 2s ṽ H φ s R̃i+n ṽ φ s (3)

Substituting (2) and (3) back into the SINR loss equation in (1), we have the SINR loss for a partially adaptive
beamformer
 −1 
Mσ 2s ṽ H φ s R̃i+n ṽ φ s
L sinr =
Mσ s /σ w
2 2
 −1 
= σ w ṽ φ s R̃i+n ṽ φ s
2 H

11.21 This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a) and (c) [no code for part(b)] and presented sequentially along with the accompanying
figures.
(a) The SINR loss for the scenario with a single interferer is plotted in Figure 11.21a. Note that the SINR loss
is computed by finding the optimum beamformer at each angle −90 ≤ φ < 90. The SINR loss has a minimum
value of the interference power at the angle of the interference. Otherwise, the SINR loss is near 0 dB since
the interference has been effectively canceled at these other angles.

%
% Problem 11.21
%
clear

M = 20; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples in the realization

% The angle in degrees of the interference source

phi_i = 30;

% The power levels of the signals and the noise

424 Statistical and Adaptive Signal Processing - Solution Manual

P_i_db = 40; % power of interference source (dB)

sigma_i = 10^(P_i_db/20); % std. deviation of interference
sigma_w = 1; % standard deviation of noise

% Compute the array response vector for the interference

u_i = spacing * sin(phi_i*pi/180); % spatial frequency
v_i = exp(-j*2*pi*[0:(M-1)]’*u_i)/sqrt(M); % array response

% Compute the full array interference-plus-noise correlation matrix

R_n = (sigma_w^2) * eye(M); % noise correlation matrix
R_in = (sigma_i^2)*v_i*v_i’ + R_n;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Set the angles at which the SINR loss will be computed

phi = -90:0.5:90;

for n = 1:length(phi)
% steering vector to current angle
u_s = spacing*sin(phi(n)*pi/180); % spatial frequency
v = exp(-j*2*pi*[0:(M-1)]’*u_s)/sqrt(M); % steering vector

L_sinr1(n) = (sigma_w^2) * v’* (R_in\v); % SINR loss from (11.3.18)

c_o = R_in\v; % unnormalized optimum beamformer

% Compute the array output SINR for the optimum beamformer

SINR = (abs(c_o’*v)^2)/(c_o’*R_in*c_o);

% Compute the array output SNR in the absence of interference

% (best we could do)
SNR_0 = (abs(v’*v)^2)/(v’*R_n*v);

% Definition of SINR loss

L_sinr(n) = SINR/SNR_0;
end

% Plot out the SINR loss of the optimum beamformer in decibels

figure
plot(phi,10*log10(abs(L_sinr)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’SINR Loss (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -45 2]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-40:10:0);
Statistical and Adaptive Signal Processing - Solution Manual 425

−10

SINR Loss (dB)

−20

−30

−40

−90 −60 −30 0 30 60 90

Angle (deg)

Figure 11.21a: SINR loss for the optimum beamformer.

grid

(b) From (11.3.39), the spatial frequencies of the spatial matched filters that are orthogonal to the spatial
matched filter in the upper branch of the generalized sidelobe cacneler (GSC) with spatial frequency u s are
m
um = us +
M
for m = 1, 2, . . . , M − 1. For φ s = 0◦ the spatial frequency of its spatial matched filter is u s = 0. Therefore,
the spatial frequencies of the spatial matched filters for the blocking matrix B for a beamspace GSC are
m
um =
M
(c) In this part of the problem, we first compute the SINR loss for the full-rank GSC which is equivalent to the
optimum beamformer. This should serve as a check to make sure the the GSC has been implemented correctly.
Again, the SINR loss is found by computing a different GSC at each angle φ and computing the SINR loss of
that beamformer to the given interference scenario. Note that since the lower branch signal is
xB (n) = B H x (n)
and the upper branch signal is

y0 (n) = v H φ s x (n)
then the theoretical value of the lower branch correlation matrix is

RB = E xB (n) xBH (n) = B H Ri+n B
and the theoretical value of the cross-correlation vector between the lower branch and the upper branch is given
by
 
rB = E xB (n) y0∗ (n) = B H Ri+n v φ s
as are used in the accompanying Matlab code.
We see that the full-rank SINR loss matches the SINR loss of the optimum beamformer from Figure 11.2a for
part (a). The SINR loss of the partially adaptive GSC, however, also matches the SINR loss of the optimum
beamformer. That is, the same performance is achieved with 2 degrees of freedom as with the full-rank M = 20
array since only one interferer is present requiring only one degree of freedom to effectively cancel it.
426 Statistical and Adaptive Signal Processing - Solution Manual

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
% - first compute the SINR loss for the full-rank GSC which is
% equivalent to the optimum beamformer
% - then compute the partially adaptive GSC only using 2 beams in
% the lower branch
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

for n = 1:length(phi)
% steering vector to current angle
u_s = spacing*sin(phi(n)*pi/180); % spatial frequency
v = exp(-j*2*pi*[0:(M-1)]’*u_s)/sqrt(M); % array response

% Compute the spatial frequencies for beamspace blocking matrix

% making sure that they lie between -0.5 and 0.5
u_bs = u_s + (1:(M-1))/M;
u_bs(find(u_bs >= 0.5)) = u_bs(find(u_bs >= 0.5)) - 1;
u_bs(find(u_bs < -0.5)) = u_bs(find(u_bs < -0.5)) + 1;
u_bs = sort(u_bs);

% Compute the full-dimensional beamspace blocking matrix

B = exp(-j*2*pi*[0:(M-1)]’*u_bs)/sqrt(M);

% Compute the full-rank GSC adaptive weights

R_B = B’*R_in*B; % correlation matrix
r_B = (sigma_i^2)*(B’*v_i)*(v’*v_i)’; % cross-correlation vector
c_B = inv(R_B)*r_B; % GSC adaptive weights
c_total = v - B*c_B; % effective end-to-end weights

% Compute the array output SINR for the full-rank GSC

SINR_full = (abs(c_total’*v)^2)/(c_total’*R_in*c_total);

% Compute the array output SNR in the absence of interference

% (best we could do)
SNR_0 = (abs(v’*v)^2)/(v’*R_n*v);

% Definition of SINR loss

L_sinr_full(n) = SINR_full/SNR_0;

% Find the 2 spatial frequencies closest to the interference

[dummy,index] = sort(abs(u_bs - u_i));
u_pa = u_bs(index(1:2));

% Form the partially adaptive beamspace blocking matrix

B_pa = exp(-j*2*pi*[0:(M-1)]’*u_pa)/sqrt(M);
Statistical and Adaptive Signal Processing - Solution Manual 427

% Compute the partially adaptive GSC adaptive weights

R_B_pa = B_pa’*R_in*B_pa; % correlation matrix
r_B_pa = (sigma_i^2)*(B_pa’*v_i)*(v’*v_i)’; % cross-correlation vector
c_B_pa = inv(R_B_pa)*r_B_pa; % GSC adaptive weights
c_total_pa = v - B_pa*c_B_pa; % effective end-to-end weights

% Compute the array output SINR for the partially adaptive GSC
SINR_pa = (abs(c_total_pa’*v)^2)/(c_total_pa’*R_in*c_total_pa);

% Compute the array output SNR in the absence of interference

% (best we could do)
SNR_0 = (abs(v’*v)^2)/(v’*R_n*v);

% Definition of SINR loss

L_sinr_pa(n) = SINR_pa/SNR_0;
end

% Plot out the SINR loss in decibels for the full-rank GSC
figure
plot(phi,10*log10(abs(L_sinr_full)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’SINR Loss (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -45 2]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-40:10:0);
grid

% Plot out the SINR loss in decibels for the partially adaptive GSC
figure
plot(phi,10*log10(abs(L_sinr_pa)),’b’,’linewidth’,2.5);
xlabel(’Angle (deg)’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’SINR Loss (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([-90 90 -45 2]);
set(gca,’xtick’,-90:30:90);
set(gca,’ytick’,-40:10:0);
grid

11.22 NEW PROBLEM

PROBLEM STATEMENT
Consider the case of a M = 20 element array with d = λ/2 inter-element spacing and thermal noise power
σ 2w = 1. Two interference sources are present at φ = 3◦ and φ = −5◦ both with powers of 50 dB. Generate
one realization of 1000 samples of this interferer. In addition a signal of interest is present at φ s = 0◦ with a
power of σ s = 100 (20 dB) in the n = 100th sample only.
428 Statistical and Adaptive Signal Processing - Solution Manual

0 0

−10 −10
SINR Loss (dB)

SINR Loss (dB)

−20 −20

−30 −30

−40 −40

−90 −60 −30 0 30 60 90 −90 −60 −30 0 30 60 90

Angle (deg) Angle (deg)

(a) SINR loss for full-rank GSC. (b) SINR loss for partially adaptive GSC.

Figure 11.21b: SINR loss for full-rank and partially adaptive GSC with 2 spatial matched filters in the lower
branch.

(a) Using an SMI adaptive beamformer for the full array compute the output signal. Is the signal of interest
visible?
(b) Using a partially adaptive beamformer with Q = 4 non-overlapping sub-arrays with M̃ = 5 elements,
compute the output of an SMI adaptive beamformer. What can you say about the signal of interest now?
(c) Repeat part (b) with Q = 2 and M̃ = 10. What are your observations now?

SOLUTION
This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a), (b) and (c) and presented sequentially along with the accompanying figures.
(a) The signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The M = 20 element array has sufficient
degrees of freedom to cancel the two interferers at φ = 3◦ and φ = −5◦ .

%
% Problem 11.22 -- new problem that is like 11.18 but with
% a more challenging interference scenario
%
clear

M = 20; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples in the realization

% The angle, power, and sample number of the signal of interest

phi_s = 0; % angle (deg)
P_s_db = 20; % signal power level (dB)
N_s = 100; % sample number of signal

% The angle in degrees of the two interference sources

phi_i1 = 3;
Statistical and Adaptive Signal Processing - Solution Manual 429

phi_i2 = -5;

% The power levels of the signals and the noise

P_i1_db = 50; % power of interference source 1 (dB)
P_i2_db = 50; % power of interference source 2 (dB)
sigma_w = 1; % standard deviation of noise

% Compute the array response vectors for the signal of interest and
% the interference
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
v_i1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i1*pi/180))/sqrt(M);
v_i2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i2*pi/180))/sqrt(M);

% Generate one realization of the signal of interest, the interference,

% and the noise
x_s = zeros(M,N);
x_s(:,N_s) = (10^(P_s_db/20))*v_s;
x_i1 = (10^(P_s_db/20))*v_i1*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_i2 = (10^(P_s_db/20))*v_i2*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_n = (randn(M,N) + i*randn(M,N))/sqrt(2);

% Combine to get overall signal and interference-plus-noise signal

x = x_s + x_i1 + x_i2 + x_n;
x_in = x_i1 + x_i2 + x_n;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Estimate the covariance matrix of the interference-plus-noise

R_in = (1/N)*x_in*x_in’;

% Compute SMI beamformer and normalize to MVDR

c_opt_unnorm = R_in\v_s;
c_opt = c_opt_unnorm/(c_opt_unnorm’*v_s);

% Compute the output signal

y_out = c_opt’*x;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
430 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.22a: Output signal for the full array SMI adaptive beamformer with M = 20 elements.

set(gca,’ytick’,0:10:40);
grid

(b) As in part (a), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The Q = 4 subarrays or degrees
of freedom are sufficient to cancel the two interferers at φ = 3◦ and φ = −5◦ .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Q = 4; % number of non-overlapping subarrays
M_sub = 5; % number of elements per subarray

% Compute the subarray transformation matrix

v_sub = exp(-j*2*pi*spacing*[0:(M_sub-1)]’*sin(phi_s*pi/180))/sqrt(M_sub);
T_sub = zeros(M,Q);
for q = 1:Q
index = (1:M_sub) + (q-1)*M_sub;
T_sub(index,q) = v_sub;
end

% Compute the subarray output signals

x_sub = T_sub’*x; % overall signal
x_in_sub = T_sub’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_sub = (1/N)*x_in_sub*x_in_sub’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_sub = T_sub’*v_s;
c_opt_sub_unnorm = R_in_sub\v_s_sub;
Statistical and Adaptive Signal Processing - Solution Manual 431

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.22b: Output signal for the SMI partially adaptive beamformer with Q = 4 nonoverlapping subarrays
with M̃ = 5 elements.

c_opt_sub = c_opt_sub_unnorm/(c_opt_sub_unnorm’*v_s_sub);

% Compute the output signal

y_out_sub = c_opt_sub’*x_sub;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_sub)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(c) Unlike parts (a) and (b), the signal is no longer visible at φ = 0◦ . The Q = 2 subarrays or degrees of
freedom are not enough to cancel the two interferers at φ = 3◦ and φ = −5◦ . Note that one degree of freedom
is used by the constraint at φ = 0◦ of the optimum beamformer.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Q = 2; % number of non-overlapping subarrays
M_sub = 10; % number of elements per subarray

% Compute the subarray transformation matrix

v_sub = exp(-j*2*pi*spacing*[0:(M_sub-1)]’*sin(phi_s*pi/180))/sqrt(M_sub);
T_sub = zeros(M,Q);
432 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.22c: Output signal for the SMI partially adaptive beamformer with Q = 2 nonoverlapping subarrays
with M̃ = 10 elements.

for q = 1:Q
index = (1:M_sub) + (q-1)*M_sub;
T_sub(index,q) = v_sub;
end

% Compute the subarray output signals

x_sub = T_sub’*x; % overall signal
x_in_sub = T_sub’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_sub = (1/N)*x_in_sub*x_in_sub’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_sub = T_sub’*v_s;
c_opt_sub_unnorm = R_in_sub\v_s_sub;
c_opt_sub = c_opt_sub_unnorm/(c_opt_sub_unnorm’*v_s_sub);

% Compute the output signal

y_out_sub = c_opt_sub’*x_sub;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_sub)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

11.23 NEW PROBLEM

Statistical and Adaptive Signal Processing - Solution Manual 433

PROBLEM STATEMENT
Consider the case of a M = 40 element array with d = λ/2 inter-element spacing and thermal noise power
σ 2w = 1. Three interference sources are present at φ = −5◦ , −10◦ and 5◦ with powers of 50, 40 and 45 dB,
respectively . Generate one realization of 1000 samples of this interferer. In addition a signal of interest is
present at φ s = 0◦ with a power of σ s = 100 (20 dB) in the n = 100th sample only.
(a) Using an SMI adaptive beamformer for the full array compute the output signal. Is the signal of interest
visible?
(b) Using a beamspace partially adaptive beamformer 11 beams at the angles −5◦ ≤ φ ≤ 5◦ at 1◦ increments,
compute the output of an SMI adaptive beamformer. What can you say about the signal of interest now?
(c) Repeat part (b) with beams only at φ = −1◦ , 0circ , and 1circ . What are your observations now?

SOLUTION
This problem is a Matlab based problem. Below is the Matlab code and three output figures. The code is
broken up into parts (a), (b) and (c) and presented sequentially along with the accompanying figures.
(a) The signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The M = 40 element array has sufficient
degrees of freedom to cancel the three interferers at φ = −5◦ , −10◦ , and 5◦ .

%
% Problem 11.23 -- new problem that is like 11.19 but with
% a more challenging interference scenario
%
clear

M = 40; % number of elements

spacing = 0.5; % element spacing in wavelengths
N = 1000; % number of samples in the realization

% The angle, power, and sample number of the signal of interest

phi_s = 0; % angle (deg)
P_s_db = 20; % signal power level (dB)
N_s = 100; % sample number of signal

% The angle in degrees of the interference sources

phi_i1 = -5;
phi_i2 = -10;
phi_i3 = 5;

% The power levels of the signals and the noise

P_i1_db = 50; % power of interference source 1 (dB)
P_i2_db = 40; % power of interference source 2 (dB)
P_i3_db = 45; % power of interference source 3 (dB)
sigma_w = 1; % standard deviation of noise
434 Statistical and Adaptive Signal Processing - Solution Manual

% Compute the array response vectors for the signal of interest and
% the interference
v_s = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_s*pi/180))/sqrt(M);
v_i1 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i1*pi/180))/sqrt(M);
v_i2 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i2*pi/180))/sqrt(M);
v_i3 = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_i3*pi/180))/sqrt(M);

% Generate one realization of the signal of interest, the interference,

% and the noise
x_s = zeros(M,N);
x_s(:,N_s) = (10^(P_s_db/20))*v_s;
x_i1 = (10^(P_i1_db/20))*v_i1*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_i2 = (10^(P_i2_db/20))*v_i2*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_i3 = (10^(P_i3_db/20))*v_i3*(randn(1,N) + i*randn(1,N))/sqrt(2);
x_n = (randn(M,N) + i*randn(M,N))/sqrt(2);

% Combine to get overall signal and interference-plus-noise signal

x = x_s + x_i1 + x_i2 + x_i3 + x_n;
x_in = x_i1 + x_i2 + x_i3 + x_n;

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part A
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

% Estimate the covariance matrix of the interference-plus-noise

R_in = (1/N)*x_in*x_in’;

% Compute SMI beamformer and normalize to MVDR

c_opt_unnorm = R_in\v_s;
c_opt = c_opt_unnorm/(c_opt_unnorm’*v_s);

% Compute the output signal

y_out = c_opt’*x;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(b) As in part (a), the signal is clearly visible at φ = 0◦ with a SNR = 20 dB. The 11 beams or degrees of
Statistical and Adaptive Signal Processing - Solution Manual 435

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.23a: Output signal for the full array SMI adaptive beamformer with M = 40 elements.

freedom are sufficient to cancel the three interferers at φ = −5◦ , −10◦ , and 5◦ .

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part B
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Angles of beam in beamspace
phi_bs = -5:5;
B = length(phi_bs);

% Compute the beamspace transformation matrix

T_bs = zeros(M,B);
for b = 1:B
v_bs = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_bs(b)*pi/180))/sqrt(M);
T_bs(1:M,b) = v_bs;
end

% Compute the beamspace output signals

x_bs = T_bs’*x; % overall signal
x_in_bs = T_bs’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_bs = (1/N)*x_in_bs*x_in_bs’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_bs = T_bs’*v_s;
c_opt_bs_unnorm = R_in_bs\v_s_bs;
c_opt_bs = c_opt_bs_unnorm/(c_opt_bs_unnorm’*v_s_bs);

% Compute the output signal

y_out_bs = c_opt_bs’*x_bs;
436 Statistical and Adaptive Signal Processing - Solution Manual

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.23b: Output signal for the SMI partially adaptive beamformer with 11 beams between −5◦ ≤ φ ≤ 5◦ at
1◦ increments.

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_bs)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid

(c) Unlike parts (a) and (b), the signal is no longer visible at φ = 0◦ . The 3 beams or degrees of freedom are
not enough to cancel the three interferers at φ = −5◦ , −10◦ , and 5◦ . Note that one degree of freedom is used
by the constraint at φ = 0◦ of the optimum beamformer.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%
% Part C
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
clear T_bs

% Angles of beam in beamspace

phi_bs = -1:1;
B = length(phi_bs);

% Compute the beamspace transformation matrix

T_bs = zeros(M,B);
for b = 1:B
v_bs = exp(-j*2*pi*spacing*[0:(M-1)]’*sin(phi_bs(b)*pi/180))/sqrt(M);
T_bs(1:M,b) = v_bs;
end
Statistical and Adaptive Signal Processing - Solution Manual 437

20

Output Power (dB)

10

0 100 200 300 400 500 600 700 800 900 1000
Sample

Figure 11.23c: Output signal for the SMI partially adaptive beamformer with 3 beams at φ = −1◦ , 0◦ , and1◦ .

% Compute the beamspace output signals

x_bs = T_bs’*x; % overall signal
x_in_bs = T_bs’*x_in; % interference-plus-noise

% Compute the subarray correlation matrix

R_in_bs = (1/N)*x_in_bs*x_in_bs’;

% Compute subarray SMI beamformer and normalize to MVDR

v_s_bs = T_bs’*v_s;
c_opt_bs_unnorm = R_in_bs\v_s_bs;
c_opt_bs = c_opt_bs_unnorm/(c_opt_bs_unnorm’*v_s_bs);

% Compute the output signal

y_out_bs = c_opt_bs’*x_bs;

% Plot out the output signal in decibels

figure
plot(1:N,20*log10(abs(y_out_bs)),’b’,’linewidth’,2.5);
xlabel(’Sample’,’fontweight’,’bold’,’fontsize’,20);
ylabel(’Output Power (dB)’,’fontweight’,’bold’,’fontsize’,20);
set(gca,’fontweight’,’bold’,’fontsize’,18);
axis([0 1000 -5 25]);
set(gca,’xtick’,0:100:1000);
set(gca,’ytick’,0:10:40);
grid
 Chapter 12

Further Topics

12.1 Prove (12.1.27), which relates the output and input fourth-order cumulants of a linear, time-invariant system.
We have

r y(4) (l1 , l2 , l3 ) = E y ∗ (n)y ∗ (n + l1 )y(n + l2 ) + y(n + l3 )
 
   
= E h ∗ (k0 )x ∗ (n − k0 ) h ∗ (k1 )x ∗ (n − k1 ) h(k2 )x(n − k2 ) h(k3 )x(n − k3 )
k0 k1 k2 k3
 
∗ ∗ ∗ ∗
= h (k0 )h (k1 )h(k2 )h(k3 )E x (n − k0 )x (n − k1 )x(n − k2 )x(n − k3 )
k0 k1 k2 k3

= h ∗ (k0 )h ∗ (k1 )h(k2 )h(k3 )r x(4) (l1 − k1 + k0 , l2 − K 2 + k0 , l3 − k3 + k0 )
k0 k1 k2 k3

using the linearity property of mathematical expectation and the definitions of higher order moments.

12.2 Derive (12.1.35) and (12.1.36), using the formulas for the cumulant of the sum of IID random variables,
developed in Section 3.2.4.
From (12.1.28) and (12.1.34), we have
∞

κ (4) (4)
x (l1 , l2 , l3 ) = γ w h(n)h(n + l1 )h(n + l2 )h(n + l3 )
n=0

or for l1 = l2 = l3 = 0
∞

κ (4)
x = γ (4)
w h 4 (n)
n=0

which also follows from (3.2.75).

12.3 If x(n) is a stationary Gaussian process, show that E{x(n)x 2 (n − l)} = ρ 2x (l) and explain how it can be used
to investigate the presence of nonlinearities.
From (3.2.53) we obtain
 
E {x(n)x(n)x(n + l)x(n + l)} = E x 2 (n) E x 2 (n + l)
+E {x(n)x(n + l)}2 + E {x(n)x(n + l)}2
= r x2 (0) + 2r x2 (l)

We can check for non-Gaussianity or nonlinearities by exciting a system with a Gaussian input w(n), computing
the output x(n), and then computing and comparing the autocorrelation of x(n) with the autocorrelation of
x 2 (n).

12.4 In this problem we use an MA(2) model to explore some properties of cumulants and bispectra.

(a) Write a MATLAB function k=cuma(b) that computes the cumulant κ (3)
x (l1 , l2 ) of the MA(2) model
H (z) = b0 + b1 z −1 + b2 z −2 for −L ≤ l1 , l2 ≤ L .

438
Statistical and Adaptive Signal Processing - Solution Manual 439

(b) Use the functions k=cuma(b), X=fft(x), and X=shiftfft(X) to compute the bispectra of the three
MA(2) models in Table 12.1. Plot your results and compare with those in Figure 12.2.
(c) Compute the bispectra of the models using the formula

Rx(3) (e j ω1 , e j ω2 ) = κ (3)
w H (e
j ω1
)H (e j ω2 )H ∗ (e j (ω1 +ω2 ) )

for ω1 = ω2 = 2π k/N, 0 ≤ k ≤ N − 1. Compare with the results in part b and Figure 12.2.
(d) Show that the bispectrum can be computed in MATLAB using the following segment of code:
H=freqz(h,1,N,’whole’);
Hc=conj(H);
R3x=(H*H’).*hankel(Hc,Hc([N,1:N-1]));
R3x=shiftfft(R3x);

The Matlab script is given below and plots are shown in Figures 12.4a through 12.4d.

% Problem 12.04

clear
close all
j=sqrt(-1);
z1=0.8*exp(j*0.6*pi);
z2=0.8*exp(j*0.9*pi);
hmin=poly([z1 conj(z1) z2 conj(z2)]);
hmax=poly([1/z1 1/conj(z1) 1/z2 1/conj(z2)]);

a=0.5;b=0.9;
hmin=[1 -(a+b) a*b];
hmax=fliplr(hmin); %hmax=hmax/hmax(1);

[Hmin,om]=freqz(hmin,1,1000);
[Hmax,om]=freqz(hmax,1,1000);
f=om/(2*pi);

subplot(2,2,1),plot(f,abs(Hmin));axis([0 0.5 -Inf Inf]);

ylabel(’|H(e^{j\omega})|’);
xlabel(’\omega/2\pi’);
title(’Minimum phase system’)
subplot(2,2,3),plot(f,angle(Hmin));axis([0 0.5 -Inf Inf]);
ylabel(’\angle H(e^{j\omega})’);
xlabel(’\omega/2\pi’);
title(’Minimum phase system’)

subplot(2,2,2),plot(f,abs(Hmax));axis([0 0.5 -Inf Inf]);

ylabel(’|H(e^{j\omega})|’);
xlabel(’\omega/2\pi’);
title(’Maximum phase system’)

subplot(2,2,4),plot(f,angle(Hmax));axis([0 0.5 -Inf Inf]);

440 Statistical and Adaptive Signal Processing - Solution Manual

ylabel(’\angle H(e^{j\omega})’);
xlabel(’\omega/2\pi’);
title(’Maximum phase system’)

NFFT=128;
H=freqz(hmin,1,NFFT,’whole’);
Hc=conj(H);
for i=1:NFFT
for k=1:NFFT
ik=mod(i+k,NFFT);
if ik==0 ik=1; end
R3min(i,k)=H(i)*H(k)*Hc(ik);
end
end
R3min=fftshift(R3min);

H=freqz(hmax,1,NFFT,’whole’);
Hc=conj(H);
for i=1:NFFT
for k=1:NFFT
ik=mod(i+k,NFFT);
if ik==0 ik=1; end
R3max(i,k)=H(i)*H(k)*Hc(ik);
end
end
R3max=fftshift(R3max);

exportfig(gcf,’p1204a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure
om = [-NFFT/2:NFFT/2-1]’/NFFT;

subplot(2,2,1),contour(om,om,abs(R3min))
title(’Bispectrum magnitude: Min phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,3),contour(om,om,angle(R3min))
title(’Bispectrum phase: Min phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,2),contour(om,om,abs(R3max))
title(’Bispectrum magnitude: Max phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,4),contour(om,om,angle(R3max))
Statistical and Adaptive Signal Processing - Solution Manual 441

title(’Bispectrum phase: Max phase’)

xlabel(’f_1’);
ylabel(’f_2’);

exportfig(gcf,’p1204b.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure

[R3min,om]=bispectrum(hmin,NFFT);
[R3max,om]=bispectrum(hmax,NFFT);

exportfig(gcf,’p1204c.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure
subplot(2,2,1),contour(om,om,abs(R3min))
title(’Bispectrum magnitude: Min phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,3),contour(om,om,angle(R3min))
title(’Bispectrum phase: Min phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,2),contour(om,om,abs(R3max))
title(’Bispectrum magnitude: Max phase’)
xlabel(’f_1’);
ylabel(’f_2’);

subplot(2,2,4),contour(om,om,angle(R3max))
title(’Bispectrum phase: Max phase’)
xlabel(’f_1’);
ylabel(’f_2’);

exportfig(gcf,’p1204d.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

12.5 Using the minimum-, maximum-, and mixed-phase systems discussed in Example 12.1.1, write a MATLAB
program to reproduce the results shown in Figures 12.3 and 12.4. Use a = 0.4, b = 0.8, and N = 300
samples.
442 Statistical and Adaptive Signal Processing - Solution Manual

Minimum phase system Maximum phase system

2.5 2.5
2 2
|H(e )|

|H(ejω)|
jω

1.5 1.5
1 1
0.5 0.5

0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4

ω/2π ω/2π
Minimum phase system Maximum phase system
1.5
2
∠ H(e )

∠ H(e )
1
jω

jω
0
0.5
−2
0
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
ω/2π ω/2π
Figure 12.4a: Plots of system responses in 12.4a

Bispectrum magnitude: Min phase Bispectrum magnitude: Max phase

0.4 0.4

0.2 0.2

0 0
f2

f2

−0.2 −0.2

−0.4 −0.4
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4
f1 f1
Bispectrum phase: Min phase Bispectrum phase: Max phase

0.4 0.4

0.2 0.2

0 0
f2

f2

−0.2 −0.2

−0.4 −0.4
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4
f1 f1

Figure 12.4b: Plots of bispectra in 12.4b

Statistical and Adaptive Signal Processing - Solution Manual 443

0.4

0.3

0.2

0.1

0
2
f

−0.1

−0.2

−0.3

−0.4

−0.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4
f1

Figure 12.4c: Plot of bispectrum in 12.4c

Bispectrum magnitude: Min phase Bispectrum magnitude: Max phase

0.4 0.4

0.2 0.2

0 0
f2

f2

−0.2 −0.2

−0.4 −0.4
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4
f1 f1
Bispectrum phase: Min phase Bispectrum phase: Max phase

0.4 0.4

0.2 0.2

0 0
f2

f2

−0.2 −0.2

−0.4 −0.4
−0.4 −0.2 0 0.2 0.4 −0.4 −0.2 0 0.2 0.4
f1 f1

Figure 12.4d: Plots of bispectra in 12.4d

444 Statistical and Adaptive Signal Processing - Solution Manual

The Matlab script is given below and plots are shown in Figures 12.5a and 12.5b.

% Min, max, and mix phase deconvolution of Gaussian

% and non-Gaussian random signals.

close all
a=0.4; b=0.8;
N=300;

hmin=[1 -(a+b) a*b]’;

hmax=flipud(hmin);
hmix=[-a (1+a*b) -b]’;

w=randn(N,1);title(’Gaussian’);
x=filter(1,hmin,w);
wmin=filter(hmin,1,x);
wmax=filter(hmax,1,x);
wmix=filter(hmix,1,x);
subplot(5,1,1),plot(w); ylabel(’w(n)’);title(’Gaussian’);
subplot(5,1,2),plot(x); ylabel(’x(n)’);
subplot(5,1,3),plot(wmin); ylabel(’w_{min}(n)’);
subplot(5,1,4),plot(wmax); ylabel(’w_{max}(n)’);
subplot(5,1,5),plot(wmix); ylabel(’w_{mix}(n)’);
xlabel(’Sample index (n)’);

exportfig(gcf,’p1205a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure
w=-log(rand(N,1));
x=filter(1,hmin,w);
wmin=filter(hmin,1,x);
wmax=filter(hmax,1,x);
wmix=filter(hmix,1,x);
subplot(5,1,1),plot(w); ylabel(’w(n)’);title(’Non-Gaussian’);
subplot(5,1,2),plot(x); ylabel(’x(n)’);
subplot(5,1,3),plot(wmin); ylabel(’w_{min}(n)’);
subplot(5,1,4),plot(wmax); ylabel(’w_{max}(n)’);
subplot(5,1,5),plot(wmix); ylabel(’w_{mix}(n)’);
xlabel(’Sample index (n)’);
exportfig(gcf,’p1205b.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

12.6 Use the Levinson-Durbin algorithm, developed in Chapter 7, to derive expressions (12.5.20), direct-form
coefficients, and (12.5.21) for the lattice parameters of the fractional pole model.
We first determine the partial correlation coefficients for m = 1 and m = 2. Then we go from order m − 1 to
order m using induction.
Statistical and Adaptive Signal Processing - Solution Manual 445

Gaussian
5

w(n)
0

−5
0 50 100 150 200 250 300
x(n) 10

−10
0 50 100 150 200 250 300
5
(n)
min

0
w

−5
0 50 100 150 200 250 300
5
(n)
max

0
w

−5
0 50 100 150 200 250 300
5
(n)
mix

0
w

−5
0 50 100 150 200 250 300
Sample index (n)
Figure 12.5a: Signal plots in 12.5a

Non−Gaussian
5
w(n)

0
0 50 100 150 200 250 300
20
x(n)

10

0
0 50 100 150 200 250 300
5
wmin(n)

0
0 50 100 150 200 250 300
5
wmax(n)

−5
0 50 100 150 200 250 300
5
wmix(n)

−5
0 50 100 150 200 250 300
Sample index (n)
Figure 12.5b: Signal plots in 12.5b
446 Statistical and Adaptive Signal Processing - Solution Manual

For m = 1 we have
d
k1 = a1(1) = −ρ(1) = −
1−d
For m = 2 we use the Levinson-Durbin recursion. Indeed, we have

ρ(2) + a1(1) ρ(1) d

k2 = a2(2) =− =−
1+ a1(1) ρ(1) 2−d

and
−2d
a1(2) = a1(1) + k2 a1(1) =
2−d
which is identical to (12.5.20) for m = 2 and k = 1.
Let us assume now that formula (12.5.20) is true for m − 1. Using the Levinson-Durbin recursion, we have

a (m)
j = a (m−1)
j
(m−1)
+ km am− j


m − 1 ( j − d − 1)!(m − 1 − d − j )!
a (m−1) =
j
j (−d − 1)!(m − 1 − d)!


m ( j − d − 1)!(m − d − j )!(m − j )(m − d)
=
j (−d − 1)!(m − d)!m(m − d − j )
But


m ( j − d − 1)!(m − d − j )!(−d) j
a (m) (m−1)
j am− j =
j (−d − 1)!(m − d)!m(m − d − j )
Using the previous relations we can show that


m ( j − d − 1)!(m − d − j )!
a (m−1) + k a (m−1)
m m− j = = a (m)
j
j (−d − 1)!(m − d)! j

which completes the proof by induction.

12.7 Consider the FPZ(1, d, 0) model

1 1
Hfpz (z) =
(1 − z ) (1 + az −1 )
−1 d

where − 12 < d < 12 and −1 < a < 1. Compute and plot the impulse response, autocorrelation, and spectrum
for a = ±0.9 and d = ±0.2, ±0.4. Identify which models have long memory and which have short memory.
The Matlab script is given below and plots are shown in Figures 12.7a and 12.7b.

% Problem 12.7

close all;
clear;

ap=0.9; am=-0.9;
d=[-0.2 -0.4 0.2.0.4];
Statistical and Adaptive Signal Processing - Solution Manual 447

PSD1=zeros(1000,4);
PSD2=PSD1;
f=(1:1000)’/2000;
om=2*pi*f;

for i=1:length(d)
PSDfp=(2*sin(om/2)).^(-2*d(i));
PSDar1=(abs(1+ap*exp(j*om))).^2;
PSD1(:,i)=PSDfp./PSDar1;
end

for i=1:length(d)
PSDfp=(2*sin(om/2)).^(-2*d(i));
PSDar1=(abs(1+am*exp(j*om))).^2;
PSD2(:,i)=PSDfp./PSDar1;
end

loglog(f,PSD1);
xlabel(’f’);
ylabel(’PSD’)
title(’a=0.9’);

exportfig(gcf,’p1207a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure
loglog(f,PSD2);
xlabel(’f’);
ylabel(’PSD’)
title(’a=-0.9’);

exportfig(gcf,’p1207b.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

12.8 Compute and plot the PSD of the FGN process, using the following approaches, and compare the results.

!∞
(a) The definition Rx (e j ω ) = l=−∞ r x (l)e
− j ωl
and formula (12.6.36) for the autocorrelation.
(b) The theoretical formula (12.6.37).

The Matlab script is given below and plots are shown in Figures 12.8a and 12.8b.

% fBn auto and PSD

close all
clear
448 Statistical and Adaptive Signal Processing - Solution Manual

a=0.9
3
10

2
10

1
10
PSD

0
10

−1
10

−2
10

−3
10 −4 −3 −2 −1 0
10 10 10 10 10
f
Figure 12.7a: PSD plot in 12.7a

a=−0.9
4
10

3
10

2
10
PSD

1
10

0
10

−1
10 −4 −3 −2 −1 0
10 10 10 10 10
f
Figure 12.7b: PSD plot in 12.7b
Statistical and Adaptive Signal Processing - Solution Manual 449

H=0.3;
f=(0:0.001:0.5)’;
om=2*pi*f;
K=50;
H1=(0.1:0.1:0.9)’;
for i=1:length(H1)
H=H1(i);
al=zeros(size(om));
for k=-K:1:K
al=al+1./(abs(om+2*k*pi).^(2*H+1));
end
Rx(:,i)=(abs(1-exp(-j*om))).^2;
Rx1(:,i)=Rx(:,i).*al;
end

loglog(f,Rx1); grid;
%xlabel(’f’);
%ylabel(’R_x(e^{j2\pif})’)
xlabel(’NORMALIZED FREQUENCY (F/Fs)’)
ylabel(’POWER SPECTRUM’)

exportfig(gcf,’p1208a.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

figure
L=20;
l=(0:L)’;
H1=(0.1:0.1:0.9)’;
for i=1:length(H1)
H=H1(i);
rx(:,i)=(abs(l-1).^(2*H)-2*abs(l).^(2*H)+abs(l+1).^(2*H))/2;
end

plot(l,rx’,’.-’); grid
%xlabel(’\fontname{times}\itl’);
%ylabel(’\fontname{times}r_x(\itl)’);
xlabel(’LAG’)
ylabel(’AUTOCORRELATION’)

exportfig(gcf,’p1208b.eps’,’fontmode’,’scaled’,...
’linemode’,’scaled’,’color’,’cmyk’);

12.9 Use the algorithm of Schür to develop a more efficient implementation of the fractional pole noise generation
method described by Equations (12.5.24) to (12.5.28).
To be completed.

12.10 In this problem we study the properties of the harmonic fractional unit-pole model specified by the system
450 Statistical and Adaptive Signal Processing - Solution Manual

2
10

1
10
POWER SPECTRUM

0
10

−1
10

−2
10 −3 −2 −1 0
10 10 10 10
NORMALIZED FREQUENCY (F/Fs)
Figure 12.8a: Power spectrum plot in 12.8a

1
AUTOCORRELATION

0.5

−0.5
0 5 10 15 20
LAG
Figure 12.8b: Autocorrelation plot in 12.8b
Statistical and Adaptive Signal Processing - Solution Manual 451

function given by (12.5.32). The impulse response is given by

0n/21
 (−1)k (d + n − k)(2 cos θ)n−2k
h θ,d (n) =
k=0
k!(n − 2k)!(d)

where (·) is the gamma function.

(a) Compute and plot h θ,d (n) for various values of θ and d.
(b) Demonstrate the validity of the above formula by evaluating h θ,d (n) from Hθ,d (z) for the same values of
θ and d.
(c) Illustrate that the model is minimum-phase if | cos θ| < 1 and − 12 < d < 1
2
or cos θ = ±1 and
− 14 < d < 14 .
(d) Illustrate that the harmonic minimum-phase model, like the FPZ(0, d, 0) one, exhibits long-memory
behavior only for positive values of d.
(e) Show that for 0 < d < 14 and cos θ = 1, the autocorrelation equals that of the FPZ(0, 2d, 0) model
[multiplied by (−1)l if cos θ = −1]. When | cos θ| < 1 and 0 < d < 12 , illustrate numerically that the
autocorrelation can be approximated by ρ(l) ∼ −l 2d−1 sin(θl − πd) as l → ∞.
(f) Compute and plot the spectrum of the model for θ = π /3 and various values of d.
(g) Generate and plot realizations of Gaussian HFPZ noise for θ = π /6 and d = −0.3, 0.1, and 0.4.

To be completed.

12.11 Determine the variogram of the process x(n) obtained by exciting the system
1
H (z) = |a| < 1
(1 − z −1 )(1 − az −1 )

with white noise w(n) ∼ WGN(0, σ 2w ).

The variogram of a process x(n) is defined by
1
vx (l) = E{[x(n + l) − x(n)]2 }
2
If the process is stationary

vx (l) = r x (0) − r x (l)

Unfortunately, the system H (z) has a pole on the unit circle. Hence, the output process is nonstationary. To
determine the variagram we note that x(n) can be obtained by integrating an AR(1) process

s(n) = as(n − 1) + w(n)


n
x(n) = x(n − 1) + s(n) = s(k)
k=−∞

Therefore

x(n + l) − x(n) = s(n + 1) + . . . + s(n + l)

452 Statistical and Adaptive Signal Processing - Solution Manual

and
1 1
vx (l) = E{[x(n + l) − x(n)]2 } = E{[s(n + 1) + . . . + s(n + l)]2 }
2 2
The autocorrelation of s(n) is

E{s(n + l)s(n)} = E{s 2 (n)}al = σ 2 al

Hence
1 1
vx (1) = E{[s(n + 1)]2 } = σ 2
2 2
1 1 2 1
vx (2) = E{[s(n + 1) + s(n + 2)]2 } = σ + 2σ 2 a + σ 2 = σ 2 (2 + 2a)
2 2 2
1 1 
vx (3) = E{[s(n + 1) + s(n + 2) + s(n + 3)]2 } = σ 2 3 + 4a + 2a 2
2 2
1 2
vx (3) = σ 4 + 6a + 4a 2 + a 3
2
or in general
 
1 2 
l−1
vx (l) = σ l + 2 (l − k)a k
2 k=1

12.12 Following the steps leading to (12.6.26), show that the fractal (Haussdorff) dimension D is related to the Hurst
exponent H by

D =2− H

Consider an arbitrary trace of a fractional Brownian motion with Hurst exponent H . Let us assume that this
trace is enclosed by a box of length (time-axis) A and height (amplitude-axis) B. We will use the method of
box counting. If we extract a piece of this trace of length r A where 0 ≤ r ≤ 1 then to preserve self-similarity
of the process, the scaled subbox needed to enclose the piece will have height r H B. On the other hand, if we
did not have this self-similarity property, then the height of the unscaled subbox would be r B. Hence the total
number of such unscaled subboxes that we can fit in one scaled subbox is equal to

area of scaled subbob (r A) r H B 1
= = 1−H
area of unscaled subbox (r A) (r B) r
However the total number of subtraces (of length r A) are 1/r. Hence the total number of unscaled boxes (that
is, the identical units carved out of the original one) are



1 1 1 1
N = 1−H
= 2−H = D
r r r r
Hence

D =2− H

12.13 Develop a Matlab function to generate the ordinary Brownian motion trace according to the steps given for
the cumulative sum method in Section 12.6.3. The format of the function should be x = obm_cumsum(N).
The Matlab function x = obm_cumsum(N) is given below.
Statistical and Adaptive Signal Processing - Solution Manual 453

Sample function of an ordinary Brownian motion

200

150

x(t)
100

50

0
0 0.25 0.5 0.75 1

Figure 12.13a: Plot of a sample function of an ordinary Brownian motion

function x = obm_cumsum(N)
% Function to generate N samples of an ordinary Brownian motion trace x
% using the cumulative sum method
% x = obm_cumsum(N)

x = randn(N,1); x = cumsum(x);

(a) Generate 16,384 samples of the Brownian motion x(t) over 0 ≤ t ≤ 1.

The Matlab script is given below and the plot is shown in Figure 12.13a.
close all;
set(0,’defaultaxesfontsize’,default_fontsize);

% (a) Generate 16,384 samples

N = 16384; n = [0:N-1]/N;
x = obm_cumsum(N);

Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,1.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1213a’);

plot(n,x,’g’); axis([0,1,min(x),max(x)]);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of an ordinary Brownian motion’,’fontsize’,title_fontsize);
set(gca,’xtick’,[0:0.25:1],’ytick’,[floor(min(x)/50)*50:50:ceil(max(x)/50)*50]);

(b) Investigate the self-affine property of x(t) by reproducing a figure similar to Figure 12.23.
The Matlab script is given below and the plots are shown in Figure 12.13b.
% (b) Self-affine property
Hf_2 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’P1213b’);

subplot(3,1,1); plot(n,x,’g’); axis([0,1,min(x),max(x)]); hold on;

plot([0.375,0.375],[min(x),max(x)],’w:’,[0.625,0.625],[min(x),max(x)],’w:’);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
454 Statistical and Adaptive Signal Processing - Solution Manual

Sample function of an ordinary Brownian motion

200

150

x(t)
100

50

0
0 0.375 0.5 0.625 1
Time t

225

200

175

150

0.375 0.4688 0.5 0.5313 0.625

200

190

180

0.4688 0.5 0.5313

Figure 12.13b: Self-affine property of the ordinary Brownian motion

title(’Sample function of an ordinary Brownian motion’,’fontsize’,title_fontsize);

set(gca,’xtick’,[0,0.375,0.5,0.625,1]);
set(gca,’ytick’,[floor(min(x)/50)*50:50:ceil(max(x)/50)*50]); hold off;

% Zoom-1
subplot(3,1,2); plot(n(6145:1:10240),x(6145:1:10240),’g’); hold on;
axis([0.375,0.625,min(x(6145:1:10240)),max(x(6145:1:10240))]);
plot([15/32,15/32],[min(x(6145:1:10240)),max(x(6145:1:10240))],’w:’,...
[17/32,17/32],[min(x(6145:1:10240)),max(x(6145:1:10240))],’w:’);
set(gca,’xtick’,[3/8,15/32,1/2,17/32,5/8],’fontsize’,8);
set(gca,’ytick’,[floor(min(x(6145:1:10240))/25)*25:25:...
ceil(max(x(6145:1:10240))/25)*25]); hold off;
hold off;

% Zoom-2
subplot(3,1,3); plot(n(7681:1:8704),x(7681:1:8704),’g’);
axis([15/32,17/32,min(x(7681:1:8704)),max(x(7681:1:8704))]);
set(gca,’xtick’,[15/32,1/2,17/32],’fontsize’,8);
set(gca,’ytick’,[floor(min(x(7681:1:8704))/10)*10:10:...
ceil(max(x(7681:1:8704))/10)*10]); hold off;

12.14 Develop a Matlab function to generate the fractional Brownian motion trace according to the steps given for the
spectral synthesis method in Section 12.6.3. The format of the function should be x = fbm_spectral(H,N).
The Matlab function x = fbm_spectral(H,N) is given below.
Statistical and Adaptive Signal Processing - Solution Manual 455

function x = fbm_spectral(H,N);
% Function to generate N samples of a fraction Brownian motion trace x
% of self-similarity index H using the Spectral Synthesis method
% x = fbm_spectral(H,N)

N1 = N; N = 2^nextpow2(2*N1);
beta = 2*H+1;

x = randn(N,1);
X = fft(x,2*N);
phase = (rand(N+1,1)-0.5)*pi;
k = (1:N)’;
X_H = 10*(k.^(-beta/2)).*abs(X(2:N+1));
X_H = [0; X_H; flipud(X_H(1:N-1))];
phase = [phase;-flipud(phase(2:N))];
X_H = X_H.*exp(j*phase);
x_H = ifft(X_H); N = length(x_H);
x_H = real(x_H(N/2-N1/2+1:N/2+N1/2));
x = x_H;

(a) Generate 1024 samples of the FBM B H (t) over 0 ≤ t ≤ 1 for H = 0.3. Investigate the self-affine
property of B0.3 (t).
The Matlab script is given below and the plots are shown in Figure 12.14a.
close all; clc;
set(0,’defaultaxesfontsize’,default_fontsize);

% (a) Generate 1024 samples for H = 0.3

N = 1024; n = [0:N-1]/N;
H = 0.3; x = fbm_spectral(H,N);

% (a1) Self-affine property

Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1214a’);

subplot(2,1,1); plot(n,x,’g’); axis([0,1,min(x),max(x)]); hold on;

plot([0.375,0.375],[min(x),max(x)],’w:’,[0.625,0.625],[min(x),max(x)],’w:’);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of a fractional Brownian motion: H = 0.3’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,0.375,0.5,0.625,1]); vr = 0.1;
set(gca,’ytick’,[floor(min(x)/vr)*vr:vr:ceil(max(x)/vr)*vr]); hold off;

% Zoom-1
I1 = (3/8)*N+1; I2 = (5/8)*N+1;
subplot(2,1,2); plot(n(I1:1:I2),x(I1:1:I2),’g’); hold on;
axis([0.375,0.625,min(x(I1:1:I2)),max(x(I1:1:I2))]);
456 Statistical and Adaptive Signal Processing - Solution Manual

Sample function of a fractional Brownian motion: H = 0.3

0.4

0.3

0.2

0.1
x(t)
0

−0.1

−0.2

−0.3

0 0.375 0.5 0.625 1

Time t

0.2

0.16

0.12
x(t)

0.08

0.04

−0.04

0.375 0.5 0.625

Time t

Figure 12.14a: Plots of a fractional Brownian motion with H = 0.3

xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
set(gca,’xtick’,[3/8,1/2,5/8],’fontsize’,8); vr = 0.04;
set(gca,’ytick’,[floor(min(x(I1:1:I2))/vr)*vr:vr:...
ceil(max(x(I1:1:I2))/vr)*vr]); hold off;

(b) Generate 1024 samples of the FBM B H (t) over 0 ≤ t ≤ 1 for H = 0.7. Investigate the self-affine
property of B0.7 (t).
The Matlab script is given below and the plots are shown in Figure 12.14b.
% (b) Generate 1024 samples for H = 0.7
N = 1024; n = [0:N-1]/N;
H = 0.7; x = fbm_spectral(H,N);

% (b1) Self-affine property

Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1214b’);

subplot(2,1,1); plot(n,x,’g’); axis([0,1,min(x),max(x)]); hold on;

plot([0.375,0.375],[min(x),max(x)],’w:’,[0.625,0.625],[min(x),max(x)],’w:’);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of a fractional Brownian motion: H = 0.7’,...
’fontsize’,title_fontsize);
Statistical and Adaptive Signal Processing - Solution Manual 457

Sample function of a fractional Brownian motion: H = 0.7

−0.05

x(t)
−0.1

−0.15

−0.2
0 0.375 0.5 0.625 1
Time t

−0.04
x(t)

−0.08

−0.12

0.375 0.5 0.625

Time t

Figure 12.14b: Plots of a fractional Brownian motion with H = 0.7

set(gca,’xtick’,[0,0.375,0.5,0.625,1]); vr = 0.05;
set(gca,’ytick’,[floor(min(x)/vr)*vr:vr:ceil(max(x)/vr)*vr]); hold off;

% Zoom-1
I1 = (3/8)*N+1; I2 = (5/8)*N+1;
subplot(2,1,2); plot(n(I1:1:I2),x(I1:1:I2),’g’); hold on;
axis([0.375,0.625,min(x(I1:1:I2)),max(x(I1:1:I2))]);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
set(gca,’xtick’,[3/8,1/2,5/8],’fontsize’,8); vr = 0.04;
set(gca,’ytick’,[floor(min(x(I1:1:I2))/vr)*vr:vr:...
ceil(max(x(I1:1:I2))/vr)*vr]); hold off;

12.15 Develop a Matlab function to generate the fractional Brownian motion trace according to the steps given
for the random midpoint replacement method in Section 12.6.3. The format of the function should be x =
fbm_replace(N).
The Matlab function x = fbm_replace(H,N) is given below.

function x = fbm_replace(H,N)
% This function generates a fractional Brownian motion with a given
% self-similarity index H, using the mipoint displacement method.
% The function receives as inputs, the self-similarity index H,
% and the number of points N to be generated.
%
458 Statistical and Adaptive Signal Processing - Solution Manual

% The function has the following format:

%
% x = fbm_replace(H,N)
%
% H = self-similarity index
% N = number of points to be generated
%

sigma = 1.0; alpha = 2;

f(1) = 0; % initialize the endpoints
f(2) = 1;
M = 2;
int = round(log(N)/log(2) + 0.5);
for j = 1:int
delta = sqrt(((1 - 2^(2 * H - 2)) * 2^(-2*H*j))/2) * sigma;
stable_rv = stable(alpha,0,M-1,1);
for i = 1:M-1
new_f(2*i-1) = f(i);
new_f(2*i) = (f(i) + f(i+1))/2 + delta * stable_rv(i);
end;
new_f(2*M-1) = f(M);
M = 2*M - 1;
f = new_f;
end

f = f - ((1:M)-1)/M;
x = f(1:N);

(a) Generate 1024 samples of the FBM B H (t) over 0 ≤ t ≤ 1 for H = 0.5. Compare visually B0.5 (t) with
that obtained by using the cumulative-sum method. Comment on your observations.
The Matlab script is given below and the plots are shown in Figure 12.15a. Visually the plots appear to
be similar.
close all; clc;
set(0,’defaultaxesfontsize’,default_fontsize);

% (a) Generate 1024 samples for H = 0.5 and compare with the cumsum method
N = 1024; n = [0:N-1]/N;
H = 0.5; x1 = fbm_replace(H,N); x2 = obm_cumsum(N);

Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1215a’);

subplot(2,1,1); plot(n,x1,’g’); axis([0,1,min(x1),max(x1)]);

xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of an ordinary Brownian motion: Replacement’,...
’fontsize’,title_fontsize);
Statistical and Adaptive Signal Processing - Solution Manual 459

Sample function of an ordinary Brownian motion: Replacement

0

−0.5

−1

x(t)
−1.5

−2

0 0.375 0.5 0.625 1

Time t

Sample function of an ordinary Brownian motion: CumSum

30

20

10
x(t)

−10

−20

−30

0 0.375 0.5 0.625 1

Time t

Figure 12.15a: Comparison of the ordinary Brownian motion plots

using the replacement and the cumsum methods

set(gca,’xtick’,[0,0.375,0.5,0.625,1]); vr = 0.1;

% Zoom-1
subplot(2,1,2); plot(n,x2,’g’); axis([0,1,min(x2),max(x2)]);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of an ordinary Brownian motion: CumSum’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,0.375,0.5,0.625,1]); vr = 0.1;

(b) Generate 1024 samples of the FBM B H (t) over 0 ≤ t ≤ 1 for H = 0.99. Investigate the artifact discussed
in the chapter for H → 1.
The Matlab script is given below and the plots are shown in Figure 12.15b. Visually, the trace has self
repeating (but increasing) structures. Thus the statistical properties will be different at different times
and at different scales.
% (b) Generate 1024 samples for H = 0.99
N = 1024; n = [0:N-1]/N;
H = 0.99; x = fbm_replace(H,N);

% (b1) Self-affine property

Hf_2 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,1.5]);
set(Hf_2,’NumberTitle’,’off’,’Name’,’P1215b’);
460 Statistical and Adaptive Signal Processing - Solution Manual

Sample function of an ordinary Brownian motion: H = 0.99

0.02

0.015

0.01

0.005

x(t) 0

−0.005

−0.01

0 0.375 0.5 0.625 1

Figure 12.15b: Plots of a fractional Brownian motion with H = 0.99

plot(n,x,’g’); axis([0,1,min(x),max(x)]);
xlabel(’Time \itt’,’fontsize’,label_fontsize);
ylabel(’\it{x}(\it{t})’,’fontsize’,label_fontsize);
title(’Sample function of an ordinary Brownian motion: H = 0.99’,...
’fontsize’,title_fontsize);
set(gca,’xtick’,[0,0.375,0.5,0.625,1]);

12.16 Based on Equation (12.6.54), develop a Matlab function [H,sigmaH] = est_H_mad(x) that computes an
estimate of the self-similarity index H and the variance σ 2H of an FBM process.
The Matlab function [H,sigmaH] = est_H_mad(x) is given below.

function [H,sigma] = est_H_mad(x)

%
% function [H,sigma] = est_H_mad(x)
%
% This program calculates the fractal dimension of an input signal
% f(n) as described in the article ’Fractal-Based Description of
% Natural Scenes’ by Alex Pentland in IEEE Transactions on Pattern
% Analysis and Machine Intelligence, Nov. 1984.

N = length(x);
maxscale = min(round(0.1*N),100);

% Find the expected values of the absolute value of f(x + delta) - f(x)
% for delta = 1 to maximum scale.

for delta = 1:maxscale

scale(delta) = delta;
expdif(delta) = (1/(N-delta))*sum(abs(x(1+delta:N)-x(1:N-delta)));
end

% Create the self-similarity plot, with the initial estimate of the fit

c_init = polyfit(log(scale),log(expdif),1);
fit_init = polyval(c_init,log(scale));
Statistical and Adaptive Signal Processing - Solution Manual 461

loglog(scale,expdif,’.’,scale,exp(fit_init));
d_est = 2 - c_init(1);
sigma = - c_init(2);
H = 2-d_est;
title([’Estimated H = ’,num2str(H)]);
xlabel(’Scale \Delta’);
ylabel(’E\{|x(n+\Delta) - x(n)|\}’);
grid;

(a) Use function x = fbm_replace(N) to generate N = 1024 samples of an FBM process with H = 0.3,
and use the function [H,sigmaH] = est_H_mad(x) to estimate H and σ H .
The Matlab script is given below and the plots are shown in Figure 12.16a.
close all; clc;
set(0,’defaultaxesfontsize’,10);

% (a) Generate 1024 samples for H = 0.3 using the replacement method
% and estimate H and the process initial variance
Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1216a’);

N = 1024; n = [0:N-1]/N;
H = 0.3; x = fbm_replace(H,N);
[H,sigma] = est_H_mad(x)
H =
0.26705987484094
sigma =
2.35161153837552

(b) Repeat the previous task for H = 0.7.

The Matlab script is given below and the plots are shown in Figure 12.16b.
% (b) Generate 1024 samples for H = 0.7 using the replacement method
% and estimate H and the process initial variance
Hf_1 = figure(’Units’,SCRUN,’position’,SCRPOS,’color’,[0,0,0],...
’Paperunits’,PAPUN,’PaperPosition’,[0,0,4.9,4.5]);
set(Hf_1,’NumberTitle’,’off’,’Name’,’P1216b’);

N = 1024; n = [0:N-1]/N;
H = 0.7; x = fbm_replace(H,N);
[H,sigma] = est_H_mad(x)
H =
0.64831632433527
sigma =
5.63245919105432

(c) Perform a Monte Carlo simulation using 100 trials and compute the mean and standard deviation of the
estimates for H and σ H in (a) and (b). To be completed.
462 Statistical and Adaptive Signal Processing - Solution Manual

Estimated H = 0.26706
0
10

E{|x(n+∆) − x(n)|}

−1
10

−2
10
0 1 2
10 10 10
Scale ∆

Figure 12.16a: The log-log plot for the estimation of H given a trace using H = 0.3

Estimated H = 0.64832
−1
10
E{|x(n+∆) − x(n)|}

−2
10

−3
10
0 1 2
10 10 10
Scale ∆

Figure 12.16b: The log-log plot for the estimation of H given a trace using H = 0.7
Statistical and Adaptive Signal Processing - Solution Manual 463

12.17 Repeat Problem 12.16 by developing a function that estimates the self-similarity index H by determining the
slope of the first 10 percent values of the periodogram in a log-log plot.
The Matlab function [H,sigmaH] = est_H_psd(x) is given below.

function [H] = est_H_psd(x)

% Estimates the self-similarity index by determining the slope of the
% first 10% of the periodogram values in log-log scale.
% [H] = est_H_psd(x)

M = nextpow2(length(x)*4);
NFFT = 2^M;
[R,F] = psd(x,NFFT,1.0,’none’);

l2f = log2(F(2:end));
l2R = log2(R(2:end));

s = [2:NFFT*0.05];
cr1 = polyfit(l2f(s),l2R(s),1);
crf1 = polyval(cr1,l2f(s));
beta1 = cr1(1);
H = (-beta1-1)/2;

The Matlab script is given below.

close all; clc;

set(0,’defaultaxesfontsize’,10);

% (a) Generate 1024 samples for H = 0.3 using the replacement method
% and estimate H
N = 1024; n = [0:N-1]/N;
H = 0.3; x = fbm_replace(H,N);
[H] = est_H_psd(x)
H =
0.30246903766353

% (b) Generate 1024 samples for H = 0.7 using the replacement method
% and estimate H
N = 1024; n = [0:N-1]/N;
H = 0.7; x = fbm_replace(H,N);
[H] = est_H_psd(x)
H =
0.61908068053019