Solution Manual

to accompany

Pattern
Classification
(2nd ed.)

David G. Stork

Solution Manual to accompany

Pattern Classication (2nd ed.)
by R. O. Duda, P. E. Hart and D. G. Stork
David G. Stork
Copyright 2001. All rights reserved.
This Manual is for the sole use of designated educators
and must not be distributed to students
except in short, isolated portions and in conjunction
with the use of Pattern Classication (2nd ed.)
June 18, 2003

Preface
In writing this Solution Manual I have learned a very important lesson. As a
student, I thought that the best way to master a subject was to go to a superb
university and study with an established expert. Later, I realized instead that the
best way was to teach a course on the subject. Yet later, I was convinced that the
best way was to write a detailed and extensive textbook. Now I know that all these
years I have been wrong: in fact the best way to master a subject is to write the
Solution Manual.
In solving the problems for this Manual I have been forced to confront myriad
technical details that might have tripped up the unsuspecting student. Students and
teachers can thank me for simplifying or screening out problems that required pages of
unenlightening calculations. Occasionally I had to go back to the text and delete the
word easily from problem references that read it can easily be shown (Problem ...).
Throughout, I have tried to choose data or problem conditions that are particularly
instructive. In solving these problems, I have found errors in early drafts of this text
(as well as errors in books by other authors and even in classic refereed papers), and
thus the accompanying text has been improved for the writing of this Manual.
I have tried to make the problem solutions self-contained and self-explanatory. I
have gone to great lengths to ensure that the solutions are correct and clearly presented
many have been reviewed by students in several classes. Surely there are errors and
typos in this manuscript, but rather than editing and rechecking these solutions over
months or even years, I thought it best to distribute the Manual, however awed,
as early as possible. I accept responsibility for these inevitable errors, and humbly
ask anyone nding them to contact me directly. (Please, however, do not ask me to
explain a solution or help you solve a problem!) It should be a small matter to change
the Manual for future printings, and you should contact the publisher to check that
you have the most recent version. Notice, too, that this Manual contains a list of
known typos and errata in the text which you might wish to photocopy and distribute
to students.
I have tried to be thorough in order to help students, even to the occassional fault of
verbosity. You will notice that several problems have the simple explain your answer
in words and graph your results. These were added for students to gain intuition
and a deeper understanding. Graphing per se is hardly an intellectual challenge, but
if the student graphs functions, he or she will develop intuition and remember the
problem and its results better. Furthermore, when the student later sees graphs of
data from dissertation or research work, the link to the homework problem and the
material in the text will be more readily apparent. Note that due to the vagaries
of automatic typesetting, gures may appear on pages after their reference in this
Manual; be sure to consult the full solution to any problem.
I have also included worked examples and so sample nal exams with solutions to
1

2
cover material in text. I distribute a list of important equations (without descriptions)
with the exam so students can focus understanding and using equations, rather than
memorizing them. I also include on every nal exam one problem verbatim from a
homework, taken from the book. I nd this motivates students to review carefully
their homework assignments, and allows somewhat more dicult problems to be included. These will be updated and expanded; thus if you have exam questions you
nd particularly appropriate, and would like to share them, please send a copy (with
solutions) to me.
It should be noted, too, that a set of overhead transparency masters of the gures
from the text are available to faculty adopters. I have found these to be invaluable
for lecturing, and I put a set on reserve in the library for students. The les can be
accessed through a standard web browser or an ftp client program at the Wiley STM
ftp area at:
ftp://ftp.wiley.com/public/sci tech med/pattern/
or from a link on the Wiley Electrical Engineering software supplements page at:
http://www.wiley.com/products/subject/engineering/electrical/
software supplem elec eng.html
I have taught from the text (in various stages of completion) at the University
of California at Berkeley (Extension Division) and in three Departments at Stanford University: Electrical Engineering, Statistics and Computer Science. Numerous
students and colleagues have made suggestions. Especially noteworthy in this regard are Sudeshna Adak, Jian An, Sung-Hyuk Cha, Koichi Ejiri, Rick Guadette,
John Heumann, Travis Kopp, Yaxin Liu, Yunqian Ma, Sayan Mukherjee, Hirobumi
Nishida, Erhan Oztop, Steven Rogers, Charles Roosen, Sergio Bermejo Sanchez, Godfried Toussaint, Namrata Vaswani, Mohammed Yousuf and Yu Zhong. Thanks too
go to Dick Duda who gave several excellent suggestions.
I would greatly appreciate notices of any errors in this Manual or the text itself. I
would be especially grateful for solutions to problems not yet solved. Please send any
such information to me at the below address. I will incorporate them into subsequent
releases of this Manual.
This Manual is for the use of educators and must not be distributed in bulk to
students in any form. Short excerpts may be photocopied and distributed, but only
in conjunction with the use of Pattern Classication (2nd ed.).
I wish you all the best of luck in teaching and research.

Ricoh Innovations, Inc.

2882 Sand Hill Road Suite 115
Menlo Park, CA 94025-7022 USA
[email protected]

David G. Stork

Contents
Preface
1 Introduction

2 Bayesian decision theory

Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

7
7
74

3 Maximum likelihood and Bayesian parameter estimation

77
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
4 Nonparametric techniques
131
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
5 Linear discriminant functions
177
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
6 Multilayer neural networks
219
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
7 Stochastic methods
255
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 276
8 Nonmetric methods
277
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
9 Algorithm-independent machine learning
295
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
10 Unsupervised learning and clustering
305
Problem Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Computer Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
Sample nal exams and solutions

357
3

4
Worked examples

CONTENTS
415

Errata and ammendations in the text

417
First and second printings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
Fifth printing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443

Chapter 1

Introduction

Problem Solutions
There are neither problems nor computer exercises in Chapter 1.

CHAPTER 1. INTRODUCTION

Chapter 2

Bayesian decision theory

Problem Solutions
Section 2.1
1. Equation 7 in the text states
P (error|x) = min[P (1 |x), P (2 |x)].
(a) We assume, without loss of generality, that for a given particular x we have
P (2 |x) P (1 |x), and thus P (error|x) = P (1 |x). We have, moreover, the
normalization condition P (1 |x) = 1P (2 |x). Together these imply P (2 |x) >
1/2 or 2P (2 |x) > 1 and
2P (2 |x)P (1 |x) > P (1 |x) = P (error|x).
This is true at every x, and hence the integrals obey


2P (2 |x)P (1 |x)dx P (error|x)dx.
In short, 2P (2 |x)P (1 |x) provides an upper bound for P (error|x).
(b) From part (a), we have that P (2 |x) > 1/2, but in the current conditions not
greater than 1/ for < 2. Take as an example, = 4/3 and P (1 |x) = 0.4
and hence P (2 |x) = 0.6. In this case, P (error|x) = 0.4. Moreover, we have
P (1 |x)P (2 |x) = 4/3 0.6 0.4 < P (error|x).
This does not provide an upper bound for all values of P (1 |x).
(c) Let P (error|x) = P (1 |x). In that case, for all x we have


P (2 |x)P (1 |x) <

P (2 |x)P (1 |x)dx

P (1 |x)P (error|x)

<
P (1 |x)P (error|x)dx,

and we have a lower bound.

7

CHAPTER 2. BAYESIAN DECISION THEORY

(d) The solution to part (b) also applies here.

Section 2.2
2. We are given that the density is of the form p(x|i ) = ke|xai |/bi .
(a) We seek k so that the function is normalized, as required by a true density. We
integrate this function, set it to 1.0,
a

i

k
exp[(x ai )/bi ]dx + exp[(x ai )/bi ]dx = 1,

ai

which yields 2bi k = 1 or k = 1/(2bi ). Note that the normalization is independent

of ai , which corresponds to a shift along the axis and is hence indeed irrelevant
to normalization. The distribution is therefore written
p(x|i ) =

1 |xai |/bi
e
.
2bi

(b) The likelihood ratio can be written directly:



p(x|1 )
|x a1 | |x a2 |
b2
.
+
= exp
p(x|2 )
b1
b1
b2
(c) For the case a1 = 0, a2 = 1, b1 = 1 and b2 = 2, we have the likelihood ratio is

2e(x+1)/2
p(x|2 )
=
2e(13x)/2
p(x|1 ) (x1)/2
2e

x0
0<x1
x > 1,

as shown in the gure.

p(x|1)
p(x|2)
4
3.5
3
2.5
2
1.5
1
0.5
-2

-1

Section 2.3
3. We are are to use the standard zero-one classication cost, that is 11 = 22 = 0
and 12 = 21 = 1.

PROBLEM SOLUTIONS

(a) We have the priors P (1 ) and P (2 ) = 1 P (1 ). The Bayes risk is given by

Eqs. 12 and 13 in the text:


R(P (1 )) = P (1 ) p(x|1 )dx + (1 P (1 )) p(x|2 )dx.
R2

R1

To obtain the prior with the minimum risk, we take the derivative with respect
to P (1 ) and set it to 0, that is


d
p(x|1 )dx p(x|2 )dx = 0,
R(P (1 )) =
dP (1 )
R2

R1

which gives the desired result:



p(x|1 )dx =
p(x|2 )dx.
R2

R1

(b) This solution is not always unique, as shown in this simple counterexample. Let
P (1 ) = P (2 ) = 0.5 and

1 0.5 x 0.5
p(x|1 ) =
0 otherwise

1 0x1
p(x|2 ) =
0 otherwise.
It is easy to verify that the decision regions R1 = [0.5, 0.25] and R1 = [0, 0.5]
satisfy the equations in part (a); thus the solution is not unique.
4. Consider the minimax criterion for a two-category classication problem.
(a) The total risk is the integral over the two regions Ri of the posteriors times
their costs:

R =
[11 P (1 )p(x|1 ) + 12 P (2 )p(x|2 )] dx
R1

+
We use
nd:
R

R2

[21 P (1 )p(x|1 ) + 22 P (2 )p(x|2 )] dx.

R2

p(x|2 ) dx = 1

R1

p(x|2 ) dx and P (2 ) = 1 P (1 ), regroup to


p(x|2 ) dx 22

= 22 + 12
R1

p(x|2 ) dx

R1



+ P (1 ) (11 22 ) + 11 p(x|1 ) dx 12 p(x|2 ) dx


R2

+ 21 p(x|1 ) dx + 22 p(x|2 ) dx
R2

R1

R1

10

CHAPTER 2. BAYESIAN DECISION THEORY


= 22 + (12 22 )

p(x|2 ) dx

R1

+P (1 ) (11 22 ) + (11 + 21 )
R2


+ (22 12 )

p(x|1 ) dx


p(x|2 ) dx .

R1

(b) Consider an arbitrary prior 0 < P (1 ) < 1, and assume the decision boundary
has been set so as to achieve the minimal (Bayes) error for that prior. If one holds
the same decision boundary, but changes the prior probabilities (i.e., P (1 ) in
the gure), then the error changes linearly, as given by the formula in part (a).
The true Bayes error, however, must be less than or equal to that (linearly
bounded) value, since one has the freedom to change the decision boundary at
each value of P (1 ). Moreover, we note that the Bayes error is 0 at P (1 ) = 0
and at P (1 ) = 1, since the Bayes decision rule under those conditions is to
always decide 2 or 1 , respectively, and this gives zero error. Thus the curve
of Bayes error rate is concave down for all prior probabilities.
E(P(1))

0.2

0.1

P*(1)

0.5

P(1)

(c) According to the general minimax equation in part (a), for our case (i.e., 11 =
22 = 0 and 12 = 21 = 1) the decision boundary is chosen to satisfy



p(x|1 ) dx =

R2

p(x|2 ) dx.
R1

We assume that a single decision point suces, and thus we seek to nd x such
that

x

N (1 , 12 ) dx =


N (2 , 22 ) dx,
x

where, as usual, N (i , i2 ) denotes a Gaussian. We assume for deniteness and

without loss of generality that 2 > 1 , and that the single decision point lies
between the means. Recall the denition of an error function, given by Eq. 96

PROBLEM SOLUTIONS

11

in the Appendix of the text, that is,

2
erf(x) =

x

et dt.
2

We can rewrite the above as

erf[(x 1 )/1 ] = erf[(x 2 )/2 ].
If the values of the error function are equal, then their corresponding arguments
must be equal, that is
(x 1 )/1 = (x 2 )/2
and solving for x gives the value of the decision point


2 1 + 1 2
x =
.
1 + 2
(d) Because the mimimax error rate is independent of the prior probabilities, we can
choose a particularly simple case to evaluate the error, for instance, P (1 ) = 0.
In that case our error becomes


2 1 1 2

E = 1/2 erf[(x 1 )/1 ] = 1/2 erf

.
1 (1 + 2 )
(e) We substitute the values given in the problem into the formula in part (c) and
nd
x =

2 1 + 1 2
1/2 + 0
=
= 1/3.
1 + 2
1 + 1/2

The error from part (d) is then




1/3 0
E = 1/2 erf
= 1/2 erf[1/3] = 0.1374.
1+0
(f) Note that the distributions have the same form (in particular, the same variance). Thus, by symmetry the Bayes error for P (1 ) = P (for some value
P ) must be the same as for P (2 ) = P . Because P (2 ) = 1 P (1 ), we
know that the curve, analogous to the one in part (b), is symmetric around the
point P (1 ) = 0.5. Because the curve is concave down, therefore it must peak
at P (1 ) = 0.5, that is, equal priors. The tangent to the graph of the error
versus P (1 ) is thus horizontal at P (1 ) = 0.5. For this case of equal priors,
the Bayes decision point for this problem can be stated simply: it is the point
midway between the means of the two distributions, that is, x = 5.5.
5. We seek to generalize the notion of minimax criteria to the case where two independent prior probabilities are set.

CHAPTER 2. BAYESIAN DECISION THEORY

12

p(x|i)P(i)

p(x|2)P(2)

p(x|3)P(3)

p(x|1)P(1)

1 x1*
1

x2* 3

x
3

(a) We use the triangle distributions and conventions in the gure. We solve for the
decision points as follows (being sure to keep the signs correct, and assuming
that the decision boundary consists of just two points):




1 (x1 1 )
2 (2 x1 )
P (1 )
=
P
(
,
)
2
12
22
which has solution
x1 =

P (1 )22 1 + P (1 )22 1 P (2 )12 2 + P (2 )2 12

.
P (1 )22 + P (2 )12

An analogous derivation for the other decision point gives:





2 (x2 2 )
3 (x2 3 )
P (2 )
=
P
(
,
)
3
22
32
which has solution
x2 =

P (2 )32 2 + P (2 )32 2 + P (3 )22 3 + P (3 )22 3

.
P (2 )32 + P (3 )22

(b) Note that from our normalization condition,

3

i=1

P (i ) = 1, we can express all

priors in terms of just two independent ones, which we choose to be P (1 ) and

P (2 ). We could substitute the values for xi and integrate, but it is just a bit
simpler to go directly to the calculation of the error, E, as a function of priors
P (1 ) and P (2 ) by considering the four contributions:
E

1
2
[1 + 1 x1 ]
212
1
2
+P (2 ) 2 [2 2 + x1 ]
22
1
2
+P (2 ) 2 [2 + 2 x2 ]
22
 1

2
+ 1 P (1 ) P (2 )
[3 3 + x2 ] .


 232

= P (1 )

P (3 )

PROBLEM SOLUTIONS

13

To obtain the minimax solution, we take the two partial and set them to zero.
The rst of the derivative equations,
E
=
P (1 )
yields the equation


2
1 + 1 x1
1
1 + 1 x1
1

0,


=
=

3 3 + x2
3
3 3 + x2
.
3

2
or

Likewise, the second of the derivative equations,

E
= 0,
P (2 )
yields the equation

2 
2 
2
2 2 + x1
2 + 2 x2
3 3 + x2
+
=
.
2
2
3
These simultaneous quadratic equations have solutions of the general form:

bi + ci
xi =
i = 1, 2.
ai
After a straightforward, but very tedious calculation, we nd that:
a1

= 12 22 + 32 ,

b1

= 12 2 1 22 1 2 3 22 1 + 32 1 + 12 2 1 3 2 + 1 3 3 ,

= 12 21 23 + 224 + 21 22 3 + 223 3 + 1 22 1 + 223 1

c1

+2 1 2 3 1 2 2 3 2 1 + 2 2 1 2 3 2 1 2 2 1 2 2 2 2 1 2 2 2
+2 2 2 3 2 + 2 2 3 2 2 2 2 2 1 2 + 2 1 3 1 2 + 2 3 2 1 2
1 2 2 2 + 2 2 2 2 2 2 1 3 2 2 3 2 2 2 + 2 1 2 2 3 2 2 3 3
2 1 2 3 3 2 2 2 3 3 2 1 3 1 3 + 2 1 2 2 3 2 2 2 2 3

+2 1 3 2 3 1 2 3 2 + 2 2 3 2 .
An analogous calculation gives:
a2

= 12 22 + 32 ,

b2

= 1 2 3 + 22 3 + 2 2 32 + 1 3 1 1 3 2 + 32 2 + 12 3 22 3 ,


= 12 22 + 32
 2 2
2 3 + 2 2 32 1 + 32 21 2 32 1 2 + 2 32 22 + 2 1 2 3 3

+ 2 22 3 3 + 2 1 3 1 3 2 1 3 2 3 + 12 23 22 23 .

c2

(c) For {i , i } = {0, 1}, {.5, .5}, {1, 1}, for i = 1, 2, 3, respectively, we substitute
into the above equations to nd x1 = 0.2612 and x2 = 0.7388. It is a simple
matter to conrm that indeed these two decision points suce for the classication problem, that is, that no more than two points are needed.

CHAPTER 2. BAYESIAN DECISION THEORY

14

p(x|i)P(i)

p(x|1)P(1)

p(x|2)P(2)

E1
-2

-1

x*

6. We let x denote our decision boundary and 2 > 1 , as shown in the gure.
(a) The error for classifying a pattern that is actually in 1 as if it were in 2 is:

R2

1
p(x|1 )P (1 ) dx =
2


N (1 , 12 ) dx E1 .
x

Our problem demands that this error be less than or equal to E1 . Thus the
bound on x is a function of E1 , and could be obtained by tables of cumulative
normal distributions, or simple numerical integration.
(b) Likewise, the error for categorizing a pattern that is in 2 as if it were in 1 is:

E2 =
R1

1
p(x|2 )P (2 ) dx =
2

x

N (2 , 22 ) dx.

(c) The total error is simply the sum of these two contributions:
E

= E1 + E 2


x
1
1
=
N (1 , 12 ) dx +
N (2 , 22 ) dx.
2
2
x

(d) For p(x|1 ) N (1/2, 1) and p(x|2 ) N (1/2, 1) and E1 = 0.05, we have (by
simple numerical integration) x = 0.2815, and thus
E

1
0.05 +
2

1
0.05 +
2

0.168.

0.2815


N (2 , 22 ) dx

0.2815




1
(x 0.5)2

dx
exp
2(0.5)2
20.05

(e) The decision boundary for the (minimum error) Bayes case is clearly at x = 0.
The Bayes error for this problem is:
EB


1
2
N (1 , 12 ) dx
2
0

PROBLEM SOLUTIONS

15

=

N (1, 1) dx = erf[1] = 0.159,

0

which of course is lower than the error for the Neyman-Pearson criterion case.
Note that if the Bayes error were lower than 2 0.05 = 0.1 in this problem, we
would use the Bayes decision point for the Neyman-Pearson case, since it too
would ensure that the Neyman-Pearson criteria were obeyed and would give the
lowest total error.
7. We proceed as in Problem 6, with the gure below.
p(x|i)P(i)
0.3
0.25

p(x|1)P(1)

p(x|2)P(2)

0.2
0.15
0.1
0.05

E1
x

-4

-2

x*

(a) Recall that the Cauchy density is

1
1

 .
b 1 + xai 2

p(x|i ) =

If we denote our (single) decision boundary point as x , and note that P (i ) =

1/2, then the error for misclassifying a 1 pattern as 2 is:

E1

p(x|1 )P (1 ) dx
x

1
2


x

1
1
 xa 2 dx.
1
b 1 +
b

We substitute (x a1 )/b = y, and sin = 1/

E1

=
where = sin1

1
2

1 + y 2 to get:

=0

d
=



1
b
,
sin1 
2
b2 + (x a1 )2

. Solving for the decision point gives
2

b
b2 +(x a1 )


x = a1 + b

1
1 = a1 + b/tan[2E1 ].
sin [2E1 ]
2

CHAPTER 2. BAYESIAN DECISION THEORY

16

(b) The error for the converse case is found similarly:

E2

1
b

1
2

1
2

x

1+

1
 xa 2 P (2 ) dx
b

=

d
=

sin1 

b
b2 + (x a2 )2


+



1
b
1 1

,
+ sin
2
b2 + (x a2 )2

where is dened in part (a).

(c) The total error is merely the sum of the component errors:


b
1
1 1

E = E1 + E2 = E1 + + sin
,
2
b2 + (x a2 )2
where the numerical value of the decision point is
x = a1 + b/tan[2E1 ] = 0.376.
(d) We add the errors (for b = 1) and nd


1
b
1 1

E = 0.1 + + sin
= 0.2607.
2
b2 + (x a2 )2
(e) For the Bayes case, the decision point is midway between the peaks of the two
distributions, i.e., at x = 0 (cf. Problem 6). The Bayes error is then

EB = 2
0

1+

1
 xa 2 P (2 ) dx = 0.2489.
b

This is indeed lower than for the Neyman-Pearson case, as it must be. Note
that if the Bayes error were lower than 2 0.1 = 0.2 in this problem, we would
use the Bayes decision point for the Neyman-Pearson case, since it too would
ensure that the Neyman-Pearson criteria were obeyed and would give the lowest
total error.
8. Consider the Cauchy distribution.
(a) We let k denote the integral of p(x|i ), and check the normalization condition,
that is, whether k = 1:

k=

1
p(x|i ) dx =
b

1+

1
 xa 2 dx.
i
b

PROBLEM SOLUTIONS

17

We substitute (x ai )/b = y into the above and get

1
k=

1
dy,
1 + y2

and use the trigonometric substition 1/

d/sin2 to nd
1
k=

=0


1 + y 2 = sin , and hence dy =

sin2
d = 1.
sin2

Indeed, k = 1, and the distribution is normalized.

(b) We let x denote the decision boundary (a single point) and nd its value by
setting p(x |1 )P (1 ) = p(x |2 )P (2 ). We have then
1
1
1
1
1
1
 x a 2 =
 x a 2 ,
1
2
b 1 +
2
b 1 +
2
b

or (x a1 ) = (x a2 ). For a1 = a2 , this implies that x = (a1 + a2 )/2, that

is, the decision boundary is midway between the means of the two distributions.
(c) For the values a1 = 3, a2 = 5 and b = 1, we get the graph shown in the gure.
P(1|x)
1
0.8
0.6
0.4
0.2

x
-4

-2

10

(d) We substitute the form of P (i |x) and p(x|i ) and nd


lim P (i |x)

lim  

1
2

1
2


1
1
b 1+( xai )2
b

1
1
b 1+( xa1 )2
b

1
2



1
1
b 1+( xa2 )2
b

b2 + (x ai )2
1
= ,
x b2 + (x a1 )2 + b2 + (x a2 )2
2
lim

and likewise, lim P (i |x) = 1/2, as can be conrmed in the gure.

x

9. We follow the terminology in Section 2.3 in the text.

CHAPTER 2. BAYESIAN DECISION THEORY

18

(a) Without loss of generality, we assume that a2 > a1 , note that the decision
boundary is at (a1 + a2 )/2. The probability of error is given by
(a1 
+a2 )/2

p(2 |x)dx +

P (error) =

p(1 |x)dx

(a1 +a2 )/2

(a1 
+a2 )/2

1
=
b

(a1 +a2 )/2

(a1 
a2 )/2

1
b

1/2
1
 xa2 2 dx +
b
1+

(a1 
a2 )/2

1
1
 xa 2 dx =
2

1+
b

1+

1/2
 xa1 2 dx

1
dy,
1 + y2

where for the last step we have used the trigonometric substitution y = (xa2 )/b
as in Problem 8. The integral is a standard form for tan1 y and thus our solution
is:


a a 
1
 1
2
1
P (error) =
tan1 

tan
[]


2b


1
1
 a2 a1 
=
tan1 
.
2
2b
(b) See figure.
P(error)
0.5
0.4
0.3
0.2
0.1

|a2 - a1|/(2b)
1

a1
(c) The maximum value of the probability of error is Pmax ( a22b
) = 1/2, which
a2 a1
occurs for | 2b | = 0. This occurs when either the two distributions are the
same, which can happen because a1 = a2 , or even if a1 = a2 because b = and
both distributions are at.

10. We use the fact that the conditional error is


P (1 |x) if we decide 2
P (error|x) =
P (2 |x) if we decide 1 .
(a) Thus the decision as stated leads to:
P (error)


=
P (error|x)p(x)dx.

PROBLEM SOLUTIONS

19

Thus we can write the probability of error as

P (error)

P (x < and 1 is the true state)

+P (x > and 2 is the true state)
= P (x < |1 )P (1 ) + P (x > |2 )P (2 )


= P (1 ) p(x|1 ) dx + P (2 ) p(x|2 ) dx.

(b) We take a derivative with respect to and set it to zero to nd an extremum,

that is,
dP (error)
= P (1 )p(|1 ) P (2 )p(|2 ) = 0,
d
which yields the condition
P (1 )p(|1 ) = P (2 )p(|2 ),
where we have used the fact that p(x|i ) = 0 at x .
(c) No, this condition does not uniquely dene .
1. If P (1 )p(|1 ) = P (2 )p(|2 ) over a range of , then would be unspecied throughout such a range.
2. There can easily be multiple values of x for which the condition hold, for
instance if the distributions have the appropriate multiple peaks.
(d) If p(x|1 ) N (1, 1) and p(x|2 ) N (1, 1) with P (1 ) = P (2 ) = 1/2, then
we have a maximum for the error at = 0.
11. The deterministic risk is given by Bayes Rule and Eq. 20 in the text

R = R(i (x)|x) dx.
(a) In a random decision rule, we have the probability P (i |x) of deciding to take
action i . Thus in order to compute the full probabilistic or randomized risk,
Rran , we must integrate over all the conditional risks weighted by their probabilities, i.e.,
Rran =

 !
a


R(i (x)|x)P (i |x) p(x) dx.

i=1

(b) Consider a xed point x and note that the (deterministic) Bayes minimum risk
decision at that point obeys
R(i (x)|x) R(max (x)|x).

CHAPTER 2. BAYESIAN DECISION THEORY

20

Therefore we have the risk in the randomized case


 !
a
Rran =
R(i (x)|x)P (i |x) p(x)dx
i=1

R(max |x)

a
!


P (i |x) p(x)dx

i=1

R(max |x)p(x)dx

= RB ,
the Bayes risk. Equality holds if and only if P (max (x)|x) = 1.
12. We rst note the normalization condition
c
!

P (i |x) = 1

i=1

for all x.
(a) If P (i |x) = P (j |x) for all i and j, then P (i |x) = 1/c and hence P (max |x) =
1/c. If one of the P (i |x) < 1/c, then by our normalization condition we must
have that P (max |x) > 1/c.
(b) The probability of error is simply 1.0 minus the probability of being correct,
that is,

P (error) = 1 P (max |x)p(x) dx.
(c) We simply substitute the limit from part (a) to get

P (error) = 1 P (max |x) p(x) dx
  

= 1g

=g1/c

p(x) dx = 1 g.

Therefore, we have P (error) 1 1/c = (c 1)/c.

(d) All categories have the same prior probability and each distribution has the
same form, in other words, the distributions are indistinguishable.
Section 2.4
13. If we choose the category max that has the maximum posterior probability, our
risk at a point x is:
!
s
P (j |x) = s [1 P (max |x)],
j=max

whereas if we reject, our risk is r . If we choose a non-maximal category k (where

k = max), then our risk is
!
s
P (j |x) = s [1 P (k |x)] s [1 P (max |x)].
j=k

PROBLEM SOLUTIONS

21

This last inequality shows that we should never decide on a category other than the
one that has the maximum posterior probability, as we know from our Bayes analysis.
Consequently, we should either choose max or we should reject, depending upon
which is smaller: s [1 P (max |x)] or r . We reject if r s [1 P (max |x)], that
is, if P (max |x) 1 r /s .
14. Consider the classication problem with rejection option.
(a) The minimum-risk decision rule is given by:
Choose i if P (i |x) P (j |x), for all j
r
and if P (i |x) 1 .
s
This rule is equivalent to
Choose i if p(x|i )P (i ) p(x|j )P (j ) for all j


r
and if p(x|i )P (i )
p(x),
1
s
where by Bayes formula
p(i |x) =

p(x|i )P (i )
.
p(x)

The optimal discriminant function for this problem is given by

Choose i if gi (x) gj (x) for all i = 1, . . . , c, and j = 1, . . . , c + 1.
Thus the discriminant functions are:

p(x|
i = 1, . . . , c
i )P
# (i ),
"
gi (x) =
s r
p(x),
i = c + 1,
s

i = 1, . . . , c
p(x|i )P (i ),
c
=
s r 
p(x|j )P (j ), i = c + 1.
s
j=1

(b) Consider the case p(x|1 ) N (1, 1), p(x|2 ) N (1, 1), P (1 ) = P (2 ) = 1/2
and r /s = 1/4. In this case the discriminant functions in part (a) give
g1 (x)

1 e(x1)

= p(x|1 )P (1 ) =
2
2

/2

1 e(x+1) /2

= p(x|2 )P (2 ) =
2
2


r
g3 (x) =
[p(x|1 )P (1 ) + p(x|2 )P (2 )]
1
s


  (x1)2 /2
2
1
1e
1 e(x+1) /2

=
1
+
4
2
2
2
2
%
$
2
2
3
3

=
e(x1) /2 + e(x+1) /2 = [g1 (x) + g2 (x)].
4
8 2
2

g2 (x)

as shown in the gure.

CHAPTER 2. BAYESIAN DECISION THEORY

22

gi(x)
0.2

g2(x)

g1(x)

g3(x)

0.15

0.1

0.05

x
-3

-2

-1

(c) If r /s = 0, there is no cost in rejecting as unrecognizable. Furthermore,

P (i |x) 1 r /s is never satised if r /s = 0. In that case, the decision
rule will always reject as unrecognizable. On the other hand, as r /s
1, P (i |x) 1r /s is always satised (there is a high cost of not recognizing)
and hence the decision rule is the Bayes decision rule of choosing the class i
that maximizes the posterior probability P (i |x).
(d) Consider the case p(x|1 ) N (1, 1), p(x|2 ) N (0, 1/4), P (1 ) = 1/3, P (2 ) =
2/3 and r /s = 1/2. In this case, the discriminant functions of part (a) give
2 e(x1)

g1 (x) = p(x|1 )P (1 ) =
3
2

/2

1 2e2x

g2 (x) = p(x|2 )P (2 ) =
3
2


r
g3 (x) =
[p(x|1 )P (1 ) + p(x|2 )P (2 )]
1
s


2
2
1 2 e(x1) /2
e2x

+
2 3
2
2
% 1
$
2
2
1

=
e(x1) /2 + e2x = [g1 (x) + g2 (x)].
2
3 2
2

Note from the gure that for this problem we should never reject.
gi(x)
0.25

g2(x)

g1(x)

0.2

g3(x)
0.15

0.1

0.05

x
-3

-2

-1

PROBLEM SOLUTIONS

23

Section 2.5
15. We consider the volume of a d-dimensional hypersphere of radius 1.0, and more
generally radius x, as shown in the gure.

z
Vd xd

dz

z
1

1
x

(a) We use Eq. 47 in the text for d = odd, that is, Vd = 2d (d1)/2 ( d1
2 )!/d!. When
applied to d = 1 (a line) we have V1 = 21 0 1 = 2. Indeed, a line segment
1 x +1 has generalized volume (length) of 2. More generally, a line of
radius x has volume of 2x.
(b) We use Eq. 47 in the text for d = even, that is, Vd = (d/2) /(d/2)!. When
applied to d = 2 (a disk), we have V2 = 1 /1! = . Indeed, a disk of radius
1 has generalized volume (area) of . More generally, a disk of radius x has
volume of x2 .
(c) Given the volume of a line in d = 1, we can derive the volume of a disk by
straightforward integration. As shown in the gure, we have
1 
V2 = 2
1 z 2 dz = ,
0

as we saw in part (a).

(d) As can be seen in the gure, to nd the volume of a generalized hypersphere in
d + 1 dimensions, we merely integrate along the z (new) dimension the volume
of a generalized hypersphere in the d-dimensional space, with proper factors and
limits. Thus we have:
1
Vd (1 z )

2 d/2

Vd+1 = 2
0

Vd (d/2 + 1)
,
dz =
(d/2 + 3/2)

where for integer k the gamma function obeys

(k + 1) = k! and (k + 1/2) = 22k+1 (2k 1)!/(k 1)!.

(e) Using this formula for d = 2k even, and Vd given for even dimensions, we get
that for the next higher (odd) dimension d :


2 d/2
(d/2)!

Vd = Vd+1 =
(d/2)! 2 (d/2 + 3/2)

24

CHAPTER 2. BAYESIAN DECISION THEORY

=

d/2 k! 22k+1
(2k + 1)!

(d

1)/2 d 1
( 2 )!
(d )!

2d

where we have used 2k = d for some integer k and d = d + 1. This conrms

Eq. 47 for odd dimension given in the text.
(f) We repeat the above steps, but now use Vd for d odd, in order to derive the
volume of the hypersphere in an even dimension:

( d2 + 1)
Vd+1 = Vd
2 ( d2 + 32 )

2d (d1)/2 ( d1
((k + 1) + 12 )
2 )!
=
d!
2
(k + 1)!

d /2
,
(d /2)!

where we have used that for odd dimension d = 2k + 1 for some integer k, and
d = d + 1 is the (even) dimension of the higher space. This conrms Eq. 47 for
even dimension given in the text.
16. We approach the problem analogously to problem 15, and use the same gure.

z
Vd xd

dz

z
1

1
x

(a) The volume of a line from 1 x 1 is indeed V1 = 2.

(b) Integrating once for the general case (according to the gure) gives
1
Vd (1 z 2 )d/2 dz =

Vd+1 = 2
0

Vd (d/2 + 1)
,
(d/2 + 3/2)

where for integer k the gamma function obeys

(k + 1) = k! and (k + 1/2) = 22k+1 (2k 1)!/(k 1)!.

Integrating again thus gives:

 

(d/2 + 1)
((d + 1)/2 + 1)
Vd+2 = Vd
(d/2 + 3/2)
((d + 1)/2 + 3/2)



Vd+1

PROBLEM SOLUTIONS

25

(d/2 + 1)
((d + 1)/2 + 3/2)
(d/2 + 1)

= Vd
= Vd
.
(d/2 + 1)(d/2 + 1)
d/2 + 1
= Vd

This is the central result we will use below.

(c) For d odd, we rewrite /(d/2 + 1) as 2/(d + 2). Thus we have


Vd





2
2
2
= Vd4
=
= Vd2
(d 2) + 2
(d 4) + 2 (d 2) + 2






2
2
2
2
= V1

(d (d 1)) + 2 (d (d 3)) + 2
(d 4) + 2 (d 2) + 2



(d1)/2 terms

  

 
1
1
1
(d1)/2 (d+1)/2 1
=
2

3 5
d2 d
 
(d1)/2 d
1
2

= (d1)/2 2(d+1)/2
=
.
d!!
d!

We have used the fact that V1 = 2, from part (a), and the notation d!! =
d (d 2) (d 4) , read d double factorial.
(d) Analogously to part (c), for d even we have

Vd

= Vd2







= Vd4
=
(d 2)/2 + 1
(d 4)/2 + 1 (d 2)/2 + 1

V2 =


 





=

(d d)/2 + 1 (d (d 2))/2 + 1
(d 4)/2 + 1 (d 2)/2 + 1



d/2 terms

  



1 1
1
1
= d/2

1 2
d/2 1 d/2
=

d/2
.
(d/2)!

(e) The central mathematical reason that we express the formulas separately for
even and for odd dimensions comes from the gamma function, which has dierent
expressions as factorials depending upon whether the argument is integer valued,
or half-integer valued. One could express the volume of hyperspheres in all
dimensions by using gamma functions, but that would be computationally a bit
less elegant.
17. Consider the minimal rectangular bounding box in d dimensions that contains
the hyperellipsoid, as shown in the gure. We can work in a coordinate system
in which the principal axes of the ellipsoid are parallel to the box axes, that is,
= diag[1/ 2 , 1/ 2 , . . . , 1/ 2 ]). The squared
the covariance matrix is diagonal (
1
2
d
Mahalanobis distance from the origin to the surface of the ellipsoid is given by r2 ,

CHAPTER 2. BAYESIAN DECISION THEORY

26

2
1

that is, points x that satisfy

r2

= xt x

= xt

1/12
0
..
.

0
1/22
..
.

...
...
..
.

0
0
..
.

x.

. . . 1/d2

Thus, along each of the principal axes, the distance obeys x2i = i2 r2 . Because the
distance across the rectangular volume is twice that amount, the volume of the rectangular bounding box is
Vrect = (2x1 )(2x2 ) (2xd ) = 2 r

d d

d
,

1/2 .
i = 2d rd ||

i=1

We let V be the (unknown) volume of the hyperellipsoid, Vd the volume of the unit
hypersphere in d dimension, and Vcube be the volume of the d-dimensional cube having
length 2 on each side. Then we have the following relation:
V
Vrect

Vd
.
Vcube

We note that the volume of the hypercube is Vcube = 2d , and substitute the above to
nd that
V =

Vrect Vd
1/2 Vd ,
= rd ||
Vcube

where Vd is given by Eq. 47 in the text. Recall that the determinant of a matrix is
1/2 = ||1/2 ), and thus the value can be written
unchanged by rotation of axes (||
as
V = rd ||1/2 Vd .
18. Let X1 , . . . , Xn be a random sample of size n from N (1 , 12 ) and let Y1 , . . . , Ym
be a random sample of size m from N (2 , 22 ).
(a) Let Z = (X1 + + Xn ) + (Y1 + + Ym ). Our goal is to show that Z is also
normally distributed. From the discussion in the text, if Xd1 N (d1 , dd )

PROBLEM SOLUTIONS

27

and A is a k d matrix, then At X N (At , At A). Here, we take

X(n+m)1

X1
X2
..
.

Xn
=
Y1

Y2

.
..
Ym

Then, clearly X is normally distributed in (n + m) 1 dimensions. We can write

Z as a particular matrix At operating on X:
Z = X1 + + Xn + Y1 + + Ym = 1t X,
where 1 denotes a vector of length n + m consisting solely of 1s. By the above
fact, it follows that Z has a univariate normal distribution.
(b) We let 3 be the mean of the new distribution. Then, we have
3

= E(Z)
= E[(X1 + + Xn ) + (Y1 + + Ym )]
= E(X1 ) + + E(Xn ) + E(Y1 ) + + E(Ym )
(since X1 , . . . , Xn , Y1 , . . . , Ym are independent)
= n1 + m2 .

(c) We let 32 be the variance of the new distribution. Then, we have

32

= Var(Z)
= Var(X1 ) + + Var(Xn ) + Var(Y1 ) + + Var(Ym )
(since X1 , . . . , Xn , Y1 , . . . , Ym are independent)
= n12 + m22

(d) Dene a column vector of the samples, as:

X1
..
.

Xn

X=
Y1 .

.
..
Ym
Then, clearly X is [(nd + md) 1]-dimensional random variable that is normally
distributed. Consider the linear projection operator A dened by
At = (Idd Idd Idd ).



(n+m) times

CHAPTER 2. BAYESIAN DECISION THEORY

28
Then we have

Z = A t X = X1 + + Xn + Y1 + + Ym ,
which must therefore be normally distributed. Furthermore, the mean and
variance of the distribution are
3 = E(Z)
3 = Var(Z)

= E(X1 ) + + E(Xn ) + E(Y1 ) + + E(Ym )

= n1 + m2 .
= Var(X1 ) + + Var(Xn ) + Var(Y1 ) + + Var(Ym )
= n1 + m2 .

19. The entropy is given by Eq. 37 in the text:


H(p(x)) = p(x) lnp(x)dx
with constraints


bk (x)p(x)dx = ak

for k = 1, . . . , q.

(a) We use Lagrange factors and nd


Hs

p(x)lnp(x)dx +


q 
!
k=1
q
!

p(x) lnp(x)


bk (x)p(x)dx ak


k bk (x)

k=0

From the normalization condition

all x.

q
!

ak k .

k=0

p(x)dx = 1, we know that a0 = b0 = 1 for

(b) In order to nd the maximum or minimum value for H (having constraints), we

take the derivative of Hs (having no constraints) with respect to p(x) and set it
to zero:

 
q
!
Hs
k bk (x) + 1 dx = 0.
=
lnp(x)
p(x)
k=0

The argument of the integral must vanish, and thus

lnp(x) =

q
!

k bk (x) 1.

k=0

We exponentiate both sides and nd

 q

!
p(x) = exp
k bk (x) 1 ,
k=0

where the q + 1 parameters are determined by the constraint equations.

PROBLEM SOLUTIONS

29

20. We make use of the result of Problem 19, that is, the maximum-entropy distribution p(x) having constraints of the form

bk (x)p(x)dx = ak
for k = 1, . . . , q is


p(x) = exp

q
!


k bk (x) 1 .

k=0

(a) In the current case, the normalization constraint is

xh
p(x)dx = 1,
xl

and thus


p(x) = exp

q
!


k bk (x) 1 = exp(0 1).

k=0

In order to satisfy this constraint equation, we demand

xh
p(x)dx = exp(0 1)(xu xl ) = 1,
xl

and thus the distribution is uniform,


1/|xu xl | xl x xu
p(x) =
0 otherwise.
(b) Here the constraint equations are

p(x)dx =


xp(x)dx = .
0

Thus, using the above result, the density is of the form

p(x) = exp[0 + 1 x 1]
and we need to solve for 0 and 1 . The normalization constraint gives

exp[0 + 1 x 1]dx
0

e1 x 

1 0


1
0 1

= 1.
= e
1

= e0 1

30

CHAPTER 2. BAYESIAN DECISION THEORY

Likewise, the mean constraint gives


1 0 1

e e

0 1

xdx = e

1
21


= .

Hence 1 = 1/ and 0 = 1 ln, and the density is



(1/)ex/
0

p(x) =

x0
otherwise.

(c) Here the density has three free parameters, and is of the general form
p(x) = exp[0 1 + 1 x + 2 x2 ],
and the constraint equations are

p(x)dx = 1

()

xp(x)dx =

()

x2 p(x)dx = 2 .

( )

We rst substitute the general form of p(x) into () and nd

 



1
0 121 /(42 )
1


erf
2 x
= 1.
e
2 2
2 2 
Since 2 < 0, erf() = 1 and erf() = 1, we have

exp[0 1 21 /(42 )]

= 1.
2

Likewise, next substitute the general form of p(x) into () and nd

$
% 

1 exp[0 1 21 /(42 )]

erf
2 x 2 /(2 2 ) 
= ,
42 2

which can be simplied to yield

exp[0 1 21 /(42 )] = .
22 2
Finally, we substitute the general form of p(x) into ( ) and nd

exp[0 1 21 /(42 )] = 2 .
22 2

PROBLEM SOLUTIONS

31

We combine these three results to nd the constants:

1

2
+ ln[1/( 2)]
22

= / 2

= 1/(2 2 ).

We substitute these values back into the general form of the density and nd


1
(x )2
p(x) =
,
exp
2 2
2
that is, a Gaussian.
21. A Gaussian centered at x = 0 is of the form
p(x) =

1
exp[x2 /(2 2 )].
2

The entropy for this distribution is given by Eq. 37 in the text:


H(p(x)) =

p(x)lnp(x)dx



1
1
2
2
2
2

=
exp[x /(2 )]ln
exp[x /(2 )] dx
2
2

= ln[ 2] + 1/2 = ln[ 2e].

For the uniform distribution, the entropy is
xu
H(p(x)) =
xl





1
1
1
ln
dx = ln
= ln[xu xl ].
|xu xl |
|xu xl |
|xu xl |

Since we are given that the mean of the distribution is 0, we know that xu = xl .
Further, we are told that the variance is 2 , that is
xh
x2 p(x)dx = 2
xl

which, after integration, implies

x2u + xu xl + x2l = 3 2 .

We put these results together and nd for the uniform distribution H(p(x)) = ln[2 3].
We are told that the variance of the triangle distribution centered on 0 having
half-width w is 2 , and this implies
w

2w

x
w

|x|
dx =
w2

w

x
0

x
dx +
w2

2w

0

x2

w+x
dx = w2 /6 = 2 .
w2

CHAPTER 2. BAYESIAN DECISION THEORY

32

The entropy of the triangle distribution is then

w
H(p(x))

=
w



w |x|
w |x|
dx
ln
w2
w2





0
wx
w+x
wx
w+x
=
dx
dx
ln
ln
w2
w2
w2
w2
w
0

= lnw + 1/2 = ln[ 6] + 1/2 = ln[ 6e],

where we used the result w = 6 from the variance condition.

Thus, in order of decreasing entropy, these equal-variance distributions are Gaussian, uniform then triangle, as illustrated in the gure, where each has the same
variance 2 .
w

p(x)

0.4/

0.3/

0.2/

0.1/

-3

-2

22. As usual, we denote our multidimensional Gaussian distribution by p(x)

N (, ), or


1
1
t 1
p(x) =
(x

)
exp

(x

)
.
2
(2)d/2 ||1/2
According to Eq. 37 in the text, the entropy is

H(p(x)) = p(x)lnp(x)dx


$
%
1

t 1
d/2
1/2
= p(x)
2 (x ) (x ) ln (2) ||
dx



indep. of x


d
d
1 ! !
1
(xi i )[1 ]ij (xj j ) dx + ln[(2)d ||]
2
2
i=1 j=1
1 !!
2 i=1 j=1
d

1
(xj j )(xi i ) [1 ]ij dx + ln[(2)d ||]
2
  
indep. of x

1 !!
1
[]ji [1 ]ij + ln[(2)d ||]
2 i=1 j=1
2

1!
1
[1 ]jj + ln[(2)d ||]
2 j=1
2

PROBLEM SOLUTIONS

33

1!
1
[I]jj + ln[(2)d ||]
2 j=1
2
  
d

=
=

d 1
+ ln[(2)d ||]
2 2
1
ln[(2e)d ||],
2

where we used our common notation of I for the

23. We have p(x|) N (, ), where

1
1
= 2 and = 0
2
0

d-by-d identity matrix.

0
5
2

0
2 .
5

(a) The density at a test point xo is



1
1
t 1
p(xo |) =
exp (xo ) (xo ) .
2
(2)3/2 ||1/2

For this case we have



 1 0 0 


 5 2

|| =  0 5 2  = 1 
2 5
 0 2 5 

1
1 0 0
1 = 0 5 2 =
0 2 5



 = 21,

1
0
0

0
5
2

0
1
1
0
=
2
0
5

0
5/21
2/21

0
2/21 ,
5/21

and the squared Mahalanobis distance from the mean to xo = (.5, 0, 1)t is
(xo )t 1 (xo )

t
1

.5
1
1
0
0
.5
1
= 0 2 0 5/21 2/21 0 2
1
2
0 2/21 5/21
1
2

0.5
0.5
16
1
= 8/21 2 = 0.25 +
+
= 1.06.
21 21
1/21
1
We substitute these values to nd that the density at xo is:


1
1
p(xo |) =
exp

(1.06)
= 8.16 103 .
2
(2)3/2 (21)1/2
(b) Recall from Eq. 44 in the text that Aw = 1/2 , where contains the
normalized eigenvectors of and is the diagonal matrix of eigenvalues. The
characteristic equation, | I| = 0, in this case is


 1
0
0 



 0
5
2  = (1 ) (5 )2 4

 0
2
5 
= (1 )(3 )(7 ) = 0.

CHAPTER 2. BAYESIAN DECISION THEORY

34

The three eigenvalues are then = 1, 3, 7 can be read immediately from the
factors. The (diagonal) matrix of eigenvalues is thus

1 0 0
= 0 3 0 .
0 0 7
To nd the eigenvectors, we solve x = i x for (i = 1, 2, 3):

x1
1 0 0
x1
x = 0 5 2 x = 5x2 + 2x3 = i x2
0 2 5
2x2 + 5x3
x3
The three eigenvectors are given by:

x1
1 = 1 : 5x2 + 2x3 =
2x2 + 5x3

x1
2 = 3 : 5x2 + 2x3 =
2x2 + 5x3

x1
3 = 7 : 5x2 + 2x3 =
2x2 + 5x3
Thus our nal and Aw matrices are:

1
0

0 1/ 2
=
0 1/ 2
and

1/2

Aw =

0
=
0

1
= 0
0

i = 1, 2, 3.

x1
1
x2 1 = 0 ,
0
x3

0
3x1
3x2 2 = 1/ 2 ,
3x3
1/ 2

0
7x1
7x2 3 = 1/2 .
7x3
1/ 2

0
1/2
1/ 2

0
0
1
1/ 2 1/2 0
0
1/ 2 1/ 2

0
0

1/ 6 1/14 .
1/ 6 1/ 14

0
3
0

0
0
7

We have then, Y = Atw (x ) N (0, I).

(c) The transformed point is found by applying A, that is,
xw

= Atw (xo )

1
0
= 0 1/ 6
0 1/ 6

0
0.5
0.5

1/14 2 = 1/ 6 .
1
3/ 14
1/ 14

(d) From part (a), we have that the squared Mahalanobis distance from xo to in
the original coordinates is r2 = (xo )t 1 (xo ) = 1.06. The Mahalanobis
distance from xw to 0 in the transformed coordinates is xtw xw = (0.5)2 + 1/6 +
3/14 = 1.06. The two distances are the same, as they must be under any linear
transformation.

PROBLEM SOLUTIONS

35

(e) A Gaussian distribution is written as



1
1
t 1
p(x) =
exp (x ) (x ) .
2
(2)d/2 ||1/2
Under a general linear transformation T, we have that x = Tt x. The transformed mean is
 =

n
!

xk =

k=1

n
!

Tt xk = Tt

k=1

n
!

xk = Tt .

k=1

Likewise, the transformed covariance matrix is



n
!

(xk  )(xk  )t

k=1

= Tt

n
!


(xk )(xk ) T

k=1

= Tt T.
We note that | | = |Tt T| = || for transformations such as translation and
rotation, and thus
p(xo |N (, )) = p(Tt xo |N (Tt , Tt T)).
The volume element is proportial to |T| and for transformations such as scaling,
the transformed covariance is proportional to |T|2 , so the transformed normalization contant contains 1/|T|, which exactly compensates for the change in
volume.
(f) Recall the denition of a whitening transformation given by Eq. 44 in the text:
Aw = 1/2 . In this case we have
y = Atw x N (Atw , Atw Aw ),
and this implies that
Var(y)

= Atw (x )(x )t Aw
= Atw A
=

(1/2 )t t (1/2 )

= 1/2 t t 1/2
= 1/2 1/2
= I,
the identity matrix.
24. Recall that the general multivariate normal density in d-dimensions is:


1
1
t 1
p(x) =
exp

(x

)
.
(x

)
2
(2)d/2 ||1/2

CHAPTER 2. BAYESIAN DECISION THEORY

36

(a) Thus we have if ij = 0 and ii = i2 , then

diag( 2 , . . . , d2 )
2 1

1 0

= ... . . . ... .

d2

Thus the determinant and inverse matrix are particularly simple:

|| =

d
,

i2 ,

i=1
1

= diag(1/12 , . . . , 1/d2 ).

This leads to the density being expressed as:





1
1
t
2
2
p(x) =
exp (x ) diag(1/1 , . . . , 1/d ) (x )
2
(2)d/2 ||1/2

2 
d 
1
1 ! xi i
=
.
exp
d
/
2 i=1
i
2i
i=1

(b) The contours of constant density are concentric ellipses in d dimensions whose
centers areat (1 , . . . , d )t = , and whose axes in the ith direction are of
length 2i c for the density p(x) held constant at
ec/2
.

2i

d
/

i=1

The axes of the ellipses are parallel to the coordinate axes. The plot in 2
dimensions (d = 2) is shown:

(1 , 2)

22 c1/2

x2

21 c1/2
x1

(c) The squared Mahalanobis distance from x to is:

0
1/12

.. (x )
..
(x )t 1 (x ) = (x )t ...
.
.
0
1/d2
2
d 
!
xi i
=
.
i
i=1

PROBLEM SOLUTIONS

37

Section 2.6
25. A useful discriminant function for Gaussians is given by Eq. 52 in the text,
1
gi (x) = (x i )t 1 (x i ) + ln P (i ).
2
We expand to get
gi (x)


1  t 1
x x ti 1 x xt 1 i + ti 1 i + ln P (i )
2
%
1 $ t 1
t 1
t 1
=
+ ln P (i ).

x
2

x
+

x
i
i
i
2   

indep. of i

We drop the term that is independent of i, and this yields the equivalent discriminant
function:
gi (x)

1
= ti 1 x ti 1 i + ln P (i )
2
= wit x + wio ,

where
wi
wio

= 1 i
1
= ti 1 i + ln P (i ).
2

The decision boundary for two Gaussians is given by gi (x) = gj (x) or

1
1
ti 1 x ti 1 i + ln P (i ) = tj 1 x tj 1 j + ln P (j ).
2
2
We collect terms so as to rewrite this as:
1
1
P (i )
(i j )t 1 x ti 1 i + tj 1 j + ln
=0
2
2
P (j )


)/P
(
)](

)
ln
[P
(
1
i
j
i
j
1
x (i j ) +
(i j )t
2
(i j )t 1 i j )
1
1
tj 1 i + ti 1 j = 0.
 2
 2

=0

This is the form of a linear discriminant

wt (x xo ) = 0,
where the weight and bias (oset) are
w = 1 (i j )
and
xo =

ln [P (i )/P (j )](i j )
1
( + j )
,
2 i
(i j )t 1 (i j )

CHAPTER 2. BAYESIAN DECISION THEORY

38

respectively.
26. The densities and Mahalanobis distances for our two distributions with the same
covarance obey
p(x|i ) N (i , ),
ri2 = (x i )t 1 (x i )
for i = 1, 2.
(a) Our goal is to show that ri2 = 21 (x i ). Here ri2 is the gradient of ri2 ,
that is, the (column) vector in d-dimensions given by:
r2

x1

..
.

ri2
xd

We carefully write out the components of ri2 for j = 1, . . . , d, and have:

ri2
xj

=
=



(x i )t 1 (x i )
xj
%
$ 1
x x 2ti 1 x + ti 1 i
  
xj

xj

indep. of x
d !
d
!

xk xl kl 2

k=1 l=1

2xj jj +

k=1

d
!

xk kj +

d
!

d
!

xk kj 2


d
!

l=1

i,k kj

k=1

d
!

i,k kj

(note that 1 is symmetric)

k=1

(xk i,k )kj

k=1
th

2j

xl jl 2

l=1
l=j

k=1
k=j

k=1

k=l

d
!

k=1
k=j

2 xj jj +


(where 1 = [kl ]dd )

i,k kl xl

d
d !
d
d
d
!
!
!
!

x2k kk +
xk xl kl 2
i,k
xl kl

xj
l=1

k=1 l=1

k=1

d !
d
!

component of 1 (x i ).

Since this is true for each component j, we have

ri2 = 21 (x i ).
(b) From part (a) we know that ri2 = 21 (x). We can work in the translated
frame of reference where the mean is at the origin. This does not change the
derivatives. Moreover, since we are dealing with a single Gaussian, we can
dispense with the needless subscript indicating the category. For notational

PROBLEM SOLUTIONS
a tx

x2

=b

39

x1

simplicity, we set the constant matrix 1 = A. In this translated frame, the

derivative is then r2 = 2Ax. For points along a line through the origin, we
have at x = b for some contant vector a, which species the direction. Dierent
points have dierent osets, b. Thus for all points x on the line, we can write
x=

ba
.
a2

Thus our derivative is

r2 = 2Ax = 2A

ba
.
a2

We are interested in the direction of this derivative vector, not its magnitude.
Thus we now normalize the vector to nd a unit vector:
ba
2A a
2

[2Ax]t [2Ax]

ba
A a
2
#t
ba
ba
At A a
2
a2

0"

Aa
.
at At Aa

Note especially that this normalized vector is independent of b, and thus independent of the point x on the line. Note too that in the very special case
where the Gaussian in hyperspherical (i.e., = qI, the identity matrix and q
a scalar), then Aa = 1/qIa = 1/qa, and At A = 1/q 2 It I = 1/q 2 I. In this case,
the derivative depends only upon a and the scalar q.
(c) We seek now to show that r12 and r22 point in opposite direction along the
line from 1 to 2 . As we saw above, ri2 points in the same direction along
any line. Hence it is enough to show that r12 and r22 point in the opposite
directions at the point 1 or 2 . From above we know that 1 and 2 are
parallel to each other along the line from 1 to 2 . Also, ri2 points in the
same direction along any line through 1 and 2 . Thus, r12 points in the same
direction along a line through 1 as r22 along a line through 2 . Thus r12
points in the same direction along the line from 1 to 2 as r22 along the line
from 2 to 1 . Therefore r12 and r22 point in opposite directions along the
line from 1 to 2 .
(d) From Eqs. 60 and 62 the text, we know that the separating hyperplane has
normal vector w = 1 (1 2 ). We shall show that the vector perpendicular
to surfaces of constant probabilities is in the same direction, w, at the point of
intersecton of the line connecting 1 and 2 and the discriminanting hyperplane.
(This point is denoted xo .)

CHAPTER 2. BAYESIAN DECISION THEORY

40

As we saw in part (a), the gradient obeys r12 = 21 (xo 1 ), Thus, because
of the relationship between the Mahalanobis distance and probability density
in a Gaussian, this vector is also in the direction normal to surfaces of constant
density. Equation 65 in the text states
xo =

1
ln [P (1 )/P (2 )]
( + 2 )
(1 2 ).
2 1
(1 2 )t 1 (1 2 )

Thus r12 evaluated at xo is:



r12 
= 21 (xo 1 )
x=x0


1
ln [P (1 )/P (2 )](1 2 )
= 21 (1 + 2 )
2
(1 2 )t 1 (1 2 )


2ln [P (1 )/P (2 )]
= 1 (2 1 ) 1
(1 2 )t 1 (1 2 )



=scalar constant
1

2
= w.

(2 1 )

Indeed, these two vectors are in the same direction.

(e) True. We are given that P (1 ) = P (2 ) = 1/2. As described in the text, for
these conditions the Bayes decision boundary is given by:
wt (x xo )

where
w = 1 (1 2 )
and
=0

xo

=
=




ln [P (1 )/P (2 )]
1
( + 2 )
(1 2 )
2 1
(1 2 )t 1 (1 2 )
1
( + 2 ).
2 1

Therefore, we have
(1 2 )t 1 x

=
=

1
( 2 )t 1 (1 2 )
21

1 t 1
1 1 t2 1 2 + t1 1 2 t2 1 1 .



2
=0

Consequently the Bayes decision boundary is the set of points x that satisfy the
following equation:
(1 2 )t 1 x =


1  t 1
1 t2 1 2 .
2 1

PROBLEM SOLUTIONS

41

Equal squared Mahalanobis distances imply the following equivalent equations:

= r22
(x 1 ) (x 1 ) = (x 2 )t 1 (x 2 )
xt 1 x + t1 1 1 2t1 1 x = xt 1 x + t2 1 2 2t2 1 x
t1 1 1 t2 1 2 = 2(1 2 )t 1 x
1 t 1
(1 2 )t 1 x =
( 1 t2 1 2 ).
2 1
1

r12

From these results it follows that the Bayes decision boundary consists of points
of equal Mahalanobis distance.
27. Our Gaussian distributions are p(x|i ) N (i , i ) and prior probabilities are
P (i ). The Bayes decision boundary for this problem is linear, given by Eqs. 6365
in the text wt (x xo ) = 0, where
w
xo

= 1 (1 2 )
1
ln [P (1 )/P (2 )]
=
(1 2 )
(1 2 ).
2
(1 2 )t 1 (1 2 )

The Bayes decision boundary does not pass between the two means, 1 and 2 if and
only if wt (1 xo ) and wt (2 xo ) have the same sign, that is, either
wt (1 xo ) > 0

and wt (2 xo ) > 0

wt (1 xo ) < 0

and wt (2 xo ) < 0.

or

These conditions are equivalent to





1
P (1 )
(1 2 )t 1 1 (1 + 2 ) ln
2
P (2 )


P
(
1
)
1
=
( 2 )t 1 (1 2 ) ln
2 1
P (2 )




1
P (1 )
wt (2 xo ) = (1 2 )t 1 1 (1 + 2 ) ln
2
P (2 )


1
)
P
(
1
= (1 2 )t 1 (1 2 ) ln
,
2
P (2 )

wt (1 xo )

and therefore we have:

wt (1 xo ) > 0 and wt (2 xo ) > 0.
This last equation implies
(1 2 )t 1 (1 2 ) > 2 ln
and
1

(1 2 )
t

P (1 )
P (2 )


(1 2 ) < 2 ln


P (1 )
.
P (2 )

CHAPTER 2. BAYESIAN DECISION THEORY

42

Likewise, the conditions can be written as:

wt (1 xo ) < 0 and wt (2 xo ) < 0
or


P (1 )
and
P (2 )


P (1 )
t 1
(1 2 ) (1 2 ) > 2 ln
.
P (2 )

(1 2 )t 1 (1 2 ) < 2 ln

In sum, the condition that the Bayes decision boundary does not pass between the
two means can be stated as follows:
%
$
(1 )
Case 1 : P (1 ) P (2 ). Condition: (1 2 )t 1 (1 2 ) < 2 ln P
P (2 ) and
this ensures wt (1 xo ) > 0 and wt (2 xo ) > 0.

Case 2 : P (1 ) > P (2 ). Condition: (1 2 )t 1 (1 2 ) < 2 ln

this ensures w (1 xo ) < 0 and w (2 xo ) < 0.
t

P (1 )
P (2 )

%
and

28. We use Eqs. 42 and 43 in the text for the mean and covariance.
(a) The covariance obeys:
2
ij

= E [(xi i )(xj j )]
 
p(xi , xj ) (xi i )(xj j )dxi dxj
=
  
=p(xi )p(xj )
by indep.



(xi i )p(xi )dxi (xj j )p(xj )dxj

where we have used the fact that


xi p(xi )dxi = i

0,

and

p(xi )dxi = 1.

(b) Suppose we had a two-dimensional Gaussian distribution, that is,

 
   2

x1
1
12
N
, 1 2
,
x2
2
21 2
where 12 = E[(x1 1 )(x2 2 )]. Furthermore, we have that the joint density
is Gaussian, that is,


1
1
t 1
p(x1 , x2 ) =
exp (x ) (x ) .
2
2||1/2
If 12 = 0, then || = |12 22 | and the inverse covariance matrix is diagonal, that
is,


1/12
0
1
=
.
0 1/22

PROBLEM SOLUTIONS
In this case, we can write
p(x1 , x2 )

=
=

43

2 
2 
x1 1
x2 2
+
1
2




2 
2 
1
1
1 x1 1
1 x2 2

exp
exp
2
1
2
2
21
22
1
exp
21 2



= p(x1 )p(x2 ).
Although we have derived this for the special case of two dimensions and 12 = 0,
the same method applies to the fully general case in d dimensions and two
arbitrary coordinates i and j.
(c) Consider the following discrete distribution:

+1 with probability 1/2
x1 =
1 with probability 1/2,
and a random variable x2 conditioned on x1 by

+1/2 with probability 1/2
If x1 = +1,
x2 =
1/2 with probability 1/2.
If x1 = 1,
x2 = 0 with probability 1.
It is simple to verify that 1 = E(x1 ) = 0; we use that fact in the following
calculation:
Cov(x1 , x2 )

= E [(x1 1 )(x2 2 )]
= E[x1 x2 ] 2 E[x1 ] 1 E[x2 ] E[1 2 ]
= E[x1 x2 ] 1 2


1
1
=
P (x1 = +1, x2 = +1/2) +
P (x1 = +1, x2 = 1/2)
2
2
+0 P (x1 = 1)
= 0.

Now we consider
Pr(x1 = +1, x2 = +1/2)

Pr(x1 = 1, x2 = 1/2)

Pr(x2 = +1/2|x1 = 1)Pr(x1 = 1)

1 1
1
=
= .
2 2
4

However, we also have

Pr(x2 = +1/2) = Pr(x + 2 = +1/2|x1 = 1)Pr(x1 = 1) = 1/4.
Thus we note the inequality
Pr(x1 = 1, x2 = +1/2) = 1/4 = Pr(x1 = 1)Pr(x2 = +1/2) = 1/8,
which veries that the features are independent. This inequality shows that
even if the covariance is zero, the features need not be independent.

CHAPTER 2. BAYESIAN DECISION THEORY

44

29. Figure 2.15 in the text shows a decision region that is a mere line in three dimensions. We can understand that unusual by analogy to a one-dimensional problem
that has a decision region consisting of a single point, as follows.
(a) Suppose we have two one-dimensional Gaussians, of possibly dierent means
and variances. The posteriors are also Gaussian, and the priors P (i ) can be
adjusted so that the posteriors are equal at a single point, as shown in the gure.
That is to say, the Bayes decision rule is to choose 1 except at a single point,
where we should choose 2 , as shown in the gure.
P(i|x)
1

1
0.8
0.6
0.4

0.2

-3

-2

-1

R1

R2

R1

(b) Likewise, in the three-dimensional case of Fig. 2.15, the priors can be adjusted
so that the decision region is a single line segment.
30. Recall that the decision boundary is given by Eq. 66 in the text, i.e., g1 (x) =
g2 (x).
(a) The most general case of a hyperquadric can be written as
xt W1 x + w1t x + w10 = xt W2 x + w2t x + w20
where Wi are positive denite d-by-d matrices, wi are arbitrary vectors in d
dimensions, and wi0 scalar osets. The above equation can be written in the
form
1 t 1
1
1
x (1 1
2 )x + (1 1 2 2 )x + (w10 w20 ) = 0.
2
If we have p(x|i ) N (i , i ) and P (i ), then we can always choose i for
the distribution to be i for the classier, and choose a mean i such that
i i = wi .
(b) We can still assign these values if P (1 ) = P (2 ).
Section 2.7
31. Recall that from the Bayes rule, to minimize the probability of error, the decision
boundary is given according to:
Choose 1 if P (1 |x) > P (2 |x); otherwise choose 2 .

PROBLEM SOLUTIONS

45

(a) In this case, since P (1 ) = P (2 ) = 1/2, we have P (1 |x) p(x|1 ); and since
the prior distributions p(x|i ) have the same , therefore the decision boundary
is:
Choose 1 if |x 1 | < |x 2 |, otherwise choose 2 .
Without loss of generality, we assume 1 < 2 , and thus:
P (error)

= P (|x 1 | > |x 2 | |1 )P (1 ) + P (|x 2 | > |x 1 | |2 )P (2 )


2
1
=
eu /2 du
2
a

where a = |2 1 |/(2).
(b) We note that a Gaussian decay overcomes an inverse function for suciently
large argument, that is,
lim

2
1
ea /2 = 0
2a

and
Pe

eu

/2

du

2
1

ea /2 .
2a

u u2 /2
e
du
a

With these results we can conclude that Pe goes to 0 as a = |2 1 |/ goes

to innity.
32. We note rst that the categories have equal priors, P (1 ) = P (2 ) = 1/2, and
spherical normal distributions, that is,
p(x|i ) N (i , 2 I).
(a) The Bayes decision rule for this problem is given by:
Choose 1 if p(1 |x) > p(2 |x); otherwise choose 2 .
Equal priors, P (1 ) = P (2 ) = 1/2, imply P (i |x) p(x|i ) for i = 1, 2.
So the decision rule is given by choosing 1 in any of the below functionally
equivalent conditions:
p(x|1 ) > p(x|2 )




t 2 1
exp 1/2(x 1 ) ( I) (x 1 )
exp 1/2(x 2 )t ( 2 I)1 (x 2 )
>
(2)d/2 | 2 I|d/2
(2)d/2 | 2 I|d/2
t
t
(x 1 ) (x 1 ) < (x 2 ) (x 2 )
xt x 2t1 x + t1 1

< xt x 2t2 x + t2 2
1
(2 1 )t x <
(t1 1 + t2 2 ),
2

CHAPTER 2. BAYESIAN DECISION THEORY

46

and otherwise choose 2 .

The minimum probability of error Pe can be computed directly from the Bayes
decision rule, but an easier way of computing Pe is as follows: Since the Bayes
decision rule for the d-dimensional problem is determined by the one-dimensional
random variable Y = (2 1 )t X, it is sucient to consider the two-category
one-dimensional problem:
i,
P (Y|i ) N (
2 ),
with P (1 ) = P (2 ) = 1/2. In this case, the mean and variance are

i = E(Y|i )

2 = Var(Y|i )

(2 1 )t i

= (2 1 )t 2 I(2 1 )t
= 2 2 1 2 .

As shown in Problem 31, it follows that

1
Pe =
2

eu

/2

du,

where
a =
=
=

|
2
1 |
|(2 1 )t 2 (2 1 )t 1 |
=
2

22 1 
t
(2 1 ) (2 1 )
2 1 2
=
22 1 
22 1 
2 1 
.
2

(b) We work in a translated coordinate system in which 1 = 0 and 2 = (1 , . . . , d )t .

We use the result
Pe

2
1
ea /2 0 if a ,
2a

where for this case we have

2 1 
2 0
1
a=
=
=
2
2
2
We conclude, then, that Pe 0 if

d

i=1

(c) We have Pe as d , provided

1 d
!

21/2
2i

i=1

2i as d .
d

i=1

2i . This implies that as d ,

the two points 1 , 2 in d-dimensions become well separated (unless


i=1

2i <

implies i 0 as i ). Consequently, with increasing d, 1 and 2 can be

easily distinguished and Pe 0.

PROBLEM SOLUTIONS

47

33. Consider a line in d dimensions that is projected down to a line in the lower
dimension space.
(a) The conditional error in d dimensions is

Ed =
min[P (1 )p(x|1 ), P (2 )p(x|2 )] dx
line

P (1 )p(x|1 ) dx +
1

P (2 )p(x|2 ) dx,
2

where 1 and 2 are the disjoint segments where the posterior probabilities of
each of the distributions are minimum.
In the lower dimension space we have

$
%
Ed1 =
min P (1 )p(x|1 ) dx, P (2 )p(x|2 ) dx.
line

Let us suppose for deniteness that



P (1 )p(x|1 ) dx <
P (2 )p(x|2 ) dx,
line

line

and thus


Ed1 =

P (1 )p(x|1 ) dx.
line

If we let 1 + 2 denote the full line, we have


Ed1 =
P (1 )p(x|1 ) dx
1 +2

P (1 )p(x|1 ) dx +
1

P (1 )p(x|1 ) dx
2

[P (2 )p(x|2 ) + |f (x)|] dx

P (1 )p(x|1 ) dx +
1


|f (x)| dx

= Ed +
2

Ed ,
where we have used the fact that P (1 )p(x|1 ) = P (2 )p(x|2 ) + |f (x)| in 2
for some unknown function f .
(b) In actual pattern recognition problems we rarely have the true distributions,
but have instead just estimates. The above derivation does not hold if the
distributions used are merely estimates.

CHAPTER 2. BAYESIAN DECISION THEORY

48
Section 2.8

34. Consider the Bhattacharyya and Cherno bounds for 1 = 2 = and 12 =

22 = 2 .
(a) In this case, the Cherno bound for the one-dimensional problem is given by
Eqs. 7375 in the text,
k() =
=

(1 )
1 2
2[2 + (1 )2 ]1 2 + ln 2
2
2
2(1 ).

We set the derivative to zero,

k()
= 2(1 ) 2 = 0,

to nd the value of that leads to the extreme value of this error. This optimizing value is = 0.5, which gives k( ) = 0.5. The Cherno bound is
thus

PCher (error) P (1 )P (2 )e1/2 .
In this case, the Bhattacharyya bound is the same as the Cherno bound:

PBhat (error) P (1 )P (2 )e1/2 .
(b) Recall the Bayes error given by Eqs. 5 and 7 in the text:
P (error)

= P (x R2 |1 )P (1 ) + P (x R1 |2 )P (2 ))


=
p(x|2 )P (1 )dx + p(x|1 )P (2 )dx.
R2

R1

In order to nd the decision point, we set p(x|1 )P (1 ) = p(x|2 )P (2 ) and

nd

1
1
exp[(x )2 /(2 2 )]P (2 ) =
exp[(x + )2 /(2 2 )]P (1 ),
2
2

and thus
lnP (2 )

(x )2
(x + )2
=
lnP
(
)

.
1
2 2
2 2

The Bayes value xB that satises this equation is

xB =

2 P (1 )
ln
.
2 P (2 )

We substitute xB into the equation for the probability of error and nd

xB
P (error)


p(x|2 )P (2 )dx +

xB

p(x|1 )P (1 )dx

PROBLEM SOLUTIONS
xB
=

49

1
P (2 )
exp[(x )2 /(2 2 )]dx +
2


P (1 )

xB

1
exp[(x + )2 /(2 2 )]dx
2

1 erf[(xB )/)]P (2 ) + erf[(xB + )/]P (1 )



P (1 )
= 1 + (P (1 ) P (2 ))erf
ln

.
2 P (2 )
=

(c) Problem not yet solved

(d) Problem not yet solved
35. We let Pd (P (1 ), 1 , 1 , P (2 ), 2 , 2 ), or simply Pd , denote the Bhattacharyya
error bound if we consider the distributions restricted to d dimensions.
(a) We seek to show that k(1/2) must increase as we go from d to d + 1 dimensions. Let 1 , 1 , 2 , 2 , k(1/2) be the relevant quantities in (d+1) dimensions.
Equation 75 in the text gives us:






1
 1 +2 
 2 
2
1 +
1
1

1 )t
2
1 ) + ln 3
2
k(1/2)
= (
(
.
8
2
2
1 ||
2|
|
We make the following denitions:
A

1 + 2
2
1 +
2

=

2 1

2
Add
ut1d
d1
11

ud1
c11
,

where = 2 1 .

Substituting, we then have


1


A
t
t 1 + 2
(2 1 )
(2 1 ) = ( )
t
u
2

u
c

1 


.

= (
1 +
2 )/2 is the average
Note that A = (1 + 2 )/2 is invertible and A
of the appropriate covariance matrices,
1 


xA1 + A1 uut A1 A1 u
1 = At u
A
,
=
ut A1

u c
where = c ut A1 u. Thus, we have the following:
1
21
2
1 +

2
1)
2
1)
(
(
2



xA1 + A1 uut A1 A1 u

t
= ( )
ut A1

CHAPTER 2. BAYESIAN DECISION THEORY

50

= t A1 + t A1 uut A1 ut A1 t A1 u + 2
= t A1 + ( t A1 u)2
t A1

|A|
and A are positive denite)
> 0 (and A
|A|

1
1 + 2
t
(2 1 )
(2 1 ).
2

as = c ut A1 u =
=

Consequently, we have
21
1

1
2
1 +
1 + 2

t
2
1)
2
1 ) (2 1 )
(
(
(2 1 ),
2
2
where equality holds if an only if = t AA1 u. Now we let


i ui

i =
,
i = 1, 2.
uti ci
i | = i |i |, where i = ci ut 1 ui . We have
Then |
i i

 


2   1 + 2 
 1 +


 = |A|
|A| = 
=



2
2

where = c uA1 u. Now, we also expand our denition of A:







1 +2 u1 +u2

A u
2
2
= 1 + 2 = 1 + 2 =
A
=
.
ut1 +ut2
c1 +c2
ut c
2
2
2

Thus, u = (u1 + u2 )/2 and c = (c1 + c2 )/2. We substitute these values into the
above equations and nd

1
2
2 |
1 +2 |
| 1 +
|
2
= ln 
ln 3 2
|

1
1 ||2 2 |
|1 ||2 |
1
2
2 |
| 1 +

2
= ln 
+ ln
1 2
|1 ||2 |
i = (Xi , Xi,d+1 ) N (
i ) for i = 1, 2, then the
i,
If Xi N (i , i ), and X
conditional variance of Xi,d+1 given Xi is:
i
Y

= Var(Xi,d+1 |Xi ), i = 1, 2.
X1 + X2

=
2

1 +X
X
2,
= (Y, Yd+1 ) =
2

and thus = Var(Yd+1 |Y). Assume now that X1 and X2 are independent. In
that case we have

= Var(Yd+1 |Y) = Var(1/ 2 (X1,d+1 + X2,d+1 )|X1 , X2 )

PROBLEM SOLUTIONS

51
=
=

1
[Var(X1,d+1 X1 ) + Var(X2,d+1 |X2 )]
2
1
(1 + 2 ).
2

We substitute these values and nd

 
 
 1 +2 
 1 +2 
1 +2


2

 2 
ln 3   
= ln 3    + ln 2
1 2
2
2
  
  
  
1
1
d+1
d

0 since
1 +2
1 2
2

and thus

 
 1 +2 
 2 
ln 3   
2
  
1

d+1

 
 1 +2 
 2 
ln 3    ,
2
  
1

where equality holds if and only if (1 + 2 )/2 = 1 2 , or equivalently if

1 = 2 .

We put these results together and nd that k(1/2)

k(1/2). This implies



Pd+1 = P (1 )P (2 )ek(1/2) P (1 )P (2 )ek(1/2) = Pd .

That is, Pd+1 Pd , as desired.
(b) The above result was derived adding the (d + 1)th dimension and does not
depend on which dimension is added. This is because, by a permutation of
the coordinates, any dimension that is added can be designated as the (d + 1)th
i , , i , i = 1, 2,
i,
dimension. This gives a permutation of the coordinates of
i
but does not change the error bound.
(c) The pathological case where equality holds (that is, Pd+1 = Pd ) is when:
= t A1 u
1

= 2 ,

where

= 2 1


1 + 2
=
2
1
2 
1 +
2

A
=
=
t
u
2

u
c

= 2,d+1 1,d+1
i

= ci uti 1
i ui .

1 =
2.
1 =
2 and
For exmple, the condition of equality is satised if
(d) No, the true error can never increase as we go to a higher dimension, as illustrated in Fig. 3.3 in the text.

CHAPTER 2. BAYESIAN DECISION THEORY

52

(e) For non-pathological distributions, Pd ek(1/2) goes to zero as d . This

is because k(1/2) for d .
(f) No. First note that
Pd (error) Pd




P (1 )P (2 )ek(1/2)

P (1 )P (2 )ek(1/2) .

But, there is no clear relation between Pd+1 (error) and Pd = P (1 )(P (2 )ek(1/2) .

So, even if k(1/2)

> k(1/2), it is not guaranteed that Pd+1 (error) < Pd (error).
Pd+1 (error) Pd+1

36. First note the denition of k() given by Eq. 75 in the text:
k()

(1 )
(2 1 )t (2 + (1 )1 )1 (2 1 )
2

|2 + (1 )1 |
1
+ ln
.
2
|2 | |1 |1

(a) Recall from Eq. 74 we have


ek() =
p (x|1 )p1 (x|2 ) dx
$
%
  exp t 1 (1) t 1
1
2
1
2
2
2
=
|1 |/2 |2 |/2


1
1
1
t
exp 12 {xt (1
1 + (1 )2 )x 2x (1 + (1 )2 )}
dx

(2)d/2
(b) Again from Eq. 74 we have
k()

where



t 1
exp /2 t1 1
1 1 (1 )/2 2 2 2
=
|1 |/2 |2 |(1)/2

 1 t 1

exp 2 {x A x 2xA1 }

dx
(2)d/2


1 1
A = 1
1 + (1 )2

and
1
A1 = 1
1 1 + (1 )2 2 .

Thus we conclude that the vector is

1
= A(1
1 1 + (1 )2 2 ).

(c) For the conditions given, we have




1
exp (xt A1 x 2xt A1 ) dx
2



1
exp (x )t A1 (x ) dx
2
t 1
(2)d/2 e1/2 A |A|1/2

= e2
1

A1

PROBLEM SOLUTIONS

53

since
t

g(x)

e 2 (x ) A (x )
,
(2)d/2 |A|1/2

where g(x) has the form of a d-dimensional Gaussian density. So it follows that


1
k()
1
t 1
t 1
e
= exp {A + 1 1 + (1 )2 2 2 }
2
|A|1/2
[]
|1 |/2 |2 |(1)/2
Problem not yet solved
37. We are given that P (1 ) = P (2 ) = 0.5 and
p(x|1 ) N (0, I)
p(x|2 ) N (1, I)
where 1 is a two-component vector of 1s.
(a) The inverse matrices are simple in this case:


1 0
1
1
=

=
.
1
2
0 1
We substitute these into Eqs. 5355 in the text and nd
g1 (x)

= w1t x + w10
 
t x1
+ 0 + ln(1/2)
= 0
x2
= ln(1/2)

and
g2 (x)

= w2t x + w20
 
 
1
1
x1
(1, 1)
+ ln(1/2)
= (1, 1)
x2
2
1
= x1 + x2 1 + ln(1/2).

We set g1 (x) = g2 (x) and nd the decision boundardy is x1 + x2 = 1, which

passes through the midpoint of the two means, that is, at
 
0.5
(1 + 2 )/2 =
.
0.5
This result makes sense because these two categories have the same prior and
conditional distributions except for their means.
(b) We use Eqs. 76 in the text and substitute the values given to nd
k(1/2)

1
8

   t  1
1
0
0

1
0

0
1

+
2

 t 
 
1 1
1 0 1
1 1
=
+ ln
8 1
2 1
0 1 1
= 1/4.

1

0
0 1

 1 0 1 0 
 (0 1)+(0 1) 
 1    


2
1
0
1

+ ln 3  
2
1
0
 1 0  1 0 
0 1

0 1

CHAPTER 2. BAYESIAN DECISION THEORY

54

Equation 77 in the text gives the Bhatacharyya bound as



P (error) P (1 )P (2 )ek(1/2) = 0.5 0.5e1/4 = 0.3894.

(c) Here we have P (1 ) = P (2 ) = 0.5 and
 


0
2 0.5
1 =
, 1 =
0
0.5 2
 


1
5 4
2 =
, 2 =
.
1
4 5
The inverse matrices are
1
1
1
2


8/5
2/15
2/15
8/15


5/9 4/9
=
.
4/9
5/9
=

We use Eqs. 6669 and nd

  
 t  
 t 
8/5
2/15
x1
8/5
2/15
0
x1
1 x1
g1 (x) =
+
2/15
8/15 x2
2/15
8/15 0
2 x2
x2


 
 t 

1 0
8/5
2/15
0
1  2 0.5 
1

ln
 + ln
2/15
8/15 0
2 0
2  0.5 2 
2
=
and
g2 (x)

4 2
2
4
1
x + x1 x2 x22 0.66 + ln ,
15 1 15
15
2

 t 
  
 t  
1 x1
5/9 4/9
x1
5/9 4/9
1
x1
+
2 x2
x2
4/9
5/9 x2
4/9
5/9 1

 
 t 


1 1
5/9 4/9
1
1  5 4 
1

ln
 + ln
4/9
5/9 1
2 1
2  4 5 
2

5 2
1
8
5
1
1
1
x + x1 x2 x22 + x1 + x2 1.1 + ln .
18 1 18
18
9
9
9
2

The Bayes decision boundary is the solution to g1 (x) = g2 (x) or

x21 + x22 28x1 x2 10x1 10x2 + 50 = 0,
which consists of two hyperbolas, as shown in the gure.
We use Eqs. 76 and 77 in the text and nd
k(1/2) =

   t  2 0.5 5
1
1
0
0.5 2 + 4

8
1
0
2

4
5

 2 0.5 5 4 
 (0.5 2)+(4 5) 
 1    


2
1
0

+ ln 3




1
0
 2 0.5  5 4 

 t 
1  
1 1
3.5 2.25
1
1 7.1875
=
+ ln
8 1
2 5.8095
2.25 3.5
1
= 0.1499.

0.5 2

4 5

PROBLEM SOLUTIONS

55

x2

R2

R1
-4

-2

x1

-2

R2

-4

Equation 77 in the text gives the Bhatacharyya bound as



P (error) P (1 )P (2 )ek(1/2) = 0.5 0.5e1.5439 = 0.4304.

38. We derive the Bhattacharyya error bound without rst examining the Cherno
bound as follows.

(a) We wish to show that min[a, b] ab. We supposewithout

 loss of generality
that a b, or equivalently b a + for > 0. Thus ab = a(a + ) a2 =
a = min[a, b].
(b) Using the above result and the formula for the probability of error given by
Eq. 7 in the text, we have:

P (error) =
min [P (1 )p(x|1 ), P (2 )p(x|2 )] dx
 


P (1 )P (2 )
p(x|1 )p(x|2 ) dx






1/2
=

/2,
where for the last step we have used the fact that min[P (1 ), P (2 )] 1/2,
which follows from the normalization condition P (1 ) + P (2 ) = 1.
39. We assume the underlying distributions are Gaussian.
(a) Based on the Gaussian assumption, we can calculate (x 2 )/2 from the hit
rate Phit = P (x > x |x 2 ). We can also calculate (x 1 )/1 from the false
alarm rate Pf alse = P (x > x |x 1 ). Since 1 = 2 = , the discriminability
is simply
   x
x 2 

 2
1
1
d = 

.
=

1
2
(b) Recall the error function from Eq. 96 in the Appendix of the text:
2
erf[u] =

u
0

ex dx.
2

CHAPTER 2. BAYESIAN DECISION THEORY

56

We can express the probability of a Gaussian distribution in terms of the error

function as:


x

P (x > x ) = 1/2 erf

and thus
x
= erf 1 [1/2 P (x > x )] .

We let Phit = P (x > x |x 2 ) and Pf alse = P (x > x |x 1 ). The

discriminability can be written as
d =

x 1
x 2
2 1
=

= erf 1 [1/2 Pf alse ] erf 1 [1/2 Phit ].

We substitute the values for this problem and nd

d1

d2

erf 1 [0.2] erf 1 [0.3] = 0.52 + 0.84 = 1.36

erf 1 [0.1] erf 1 [0.2] = 0.26 + 0.52 = 0.78.

(c) According to Eq. 70 in the text, we have

Case 1 :
Case 2 :

1
[0.3 + (1 0.8)] = 0.25
2
1
P (error) = [0.4 + (1 0.7)] = 0.35.
2

P (error) =

(d) Because of the symmetry property of the ROC curve, the point (Phit , Pf alse ) and
the point (1 Phit , 1 Pf alse ) will go through the same curve corresponding to
some xed d . For case B, (0.1, 0.3) is also a point on ROC curve that (0.9, 0.7)
lies. We can compare this point with case A, going through (0.8, 0.3) and the
help of Fig. 2.20 in the text, we can see that case A has a higher discriminability
d .
40. We are to assume that the two Gaussians underlying the ROC curve have dierent
variances.
(a) From the hit rate Phit = P (x > x |x 2 ) we can calculate (x 2 )/2 . From
the false alarm rate Pf alse = P (x > x |x 1 ) we can calculate (x 1 )/1 .
Let us denote the ratio of the standard deviations as 1 /2 = K. Then we can
write the discriminability in this case as
   x
x 1   x 2
x 1 
 2
 2
1
da = 

=
=
.
1 2
1 2
1 2
2 /K
K1
Because we cannot determine K from (2 x )/2 and (x 1 )/1 , we cannot
determine d uniquely with only Phit = P (x > x |x 2 ) and Pf alse = P (x >
x |x 1 ).
(b) Suppose we are given the following four experimental rates:
Phit1 = P (x > x1 |2 )
Phit2 = P (x > x2 |2 )

Pf alse1 = P (x > x1 |1 )

Pf alse2 = P (x > x2 |1 ).

PROBLEM SOLUTIONS

57

Then we can calculate the four quantities

x1 2
= erf 1 [1/2 Phit1 ]
2
x 2
a2 = 2
= erf 1 [1/2 Phit2 ]
2

x1 1
= erf 1 [1/2 Pf alse1 ] for x1
1
x 1
b2 = 2
= erf 1 [1/2 Pf alse2 ] for x2 .
1

a1 =

b1 =

Then we have the following relations:

a1 a2

b1 b2

x1 x2
2
x1 x2
1
a1 a2
1
=
.
b 1 b2
2

Thus, with K we can calculate da as

   x
x 1 
 2
 2
1
1
= 
1
=

1 2
2 /K
K1
 (a a )a
(b1 b2 )b1 

1
2 1
= 

.
b 1 b2
a1 a2

da

(c) For all those x1 and x2 that satisfy

x1 2
x 2
= 2
2
2
or
x1 1
x 1
= 2
.
1
1
That is, the two dierent thresholds do not provide any additional information
and conveys the same information as only one observation. As explained in
part (a), this kind of result would not allow us to determine da .
(d) See figure.
1

P(x<x*|x 1)

p(x|i)
0.4

0.3

0.2

0.6

0.4

0.2

0.1

-2

0.8

x*

41. We use the notation shown in the gure.

0.2

0.4

0.6

0.8

P(x<x*|x 2)

CHAPTER 2. BAYESIAN DECISION THEORY

58

p(x|i)

x*

(a) Here our triangle distribution is


T (, ) =

|x|
2

|x |
otherwise.

We assume without loss of generality that 2 1 and then dene dT = (2

1 )/. We then limit our considerations to overlapping distributions, that is,
0 dT 2. There are two cases to consider:
1 dT 2, where the decision boundary x is conned to the region of
overlap, as shown in the gure.
0 dT 1 where the decision boundary x can be in one of three regions,
as shown in the gure.


We denote the hit rate H = P (x > x |2 )dx, false alarm rate F = P (x >
x |1 )dx, and miss rate M = 1 F . The hit rate is

x
H

= 1

T (2 , )dx = 1

2
1 +

T (1 , )dx =
x

dT

= 1

(2 dT
= 1
2

2F

(x 2 + )2
2 2

(1 + x )2
2 2

2(1 H)

2F )2

which is valid for 0 F (2 dT )2 /2.

For the second case 2 x < 1 , that is, 0 dT 1, we have
H
F
dT
H

(x 1 + )2
2 2

(x 1 + )2
= 1
2 2 

=
2(1 F ) 2(1 H)

( 2(1 F ) dT )2
= 1
2
=

PROBLEM SOLUTIONS

59

which is valid for 1/2 F 1 d2

T /2.
For the third case, 2 x 1 + and 0 dT 1 we have
F
H
dT

(1 + x )2
2 2
(2 + x )2
=
2
2

=
2H 2F ,
=

valid for 0 F (1 dT )2 /2.

(b) If dT = 2.0, then the two densities do not overlap. If x is closer to the mean
of 1 than to the mean of 2 , then the modied ROC curve goes through
P (x > x |x 1 ) = 1, P (x > x |x 2 ) = 0. Alternatively, if x is closer to
the mean of 2 than to the mean of 1 , then we get the converse. In short,
there is no meaningful value of dT greater than 2.0.
1
dT'=1.0

P(x<x*|x 1)

0.8

0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1

0.6

0.4

0.2

0.0
0.2

0.4

0.6

0.8

P(x<x*|x 2)

(c) For H = 0.4 and F = 0.2, we can rule out case I because H < 0.5. We can
rule out CaseII-i because
F < 0.5. Can rule out Case II-ii because H < 0.5.

Thus dT = 2H 2F = 0.262. We assume P (1 ) = P (2 ) = 0.5, then

by symmetry, x = 1 + 0.131 = 2 0.131. The Bayes optimal is then
PE = 1/2(0.378 + 0.378) = 0.378.
(d) Here x = 1 + (dT /2) = 2 (dT /2).

(e) Here H = 0.9 and F = 0.3. Case I: dT = 2 2F 2(1 H) = 0.778.

This is not valide because dT < 1. Cannot be case II-i since F < 0.5. Cannot
be case II-iii since H > 0.5. Thus it is case II-ii and dT = 0.778. Thus x =
1 + 0.389 = 2 0.389. Also, F = M = 0.187, and the Bayes error is
PE = 0.187.

42. We consider bounds on min[p, 1 p], where p = p(x|1 ).

(a) Here the candidate lower bound is


1
1 + e
bL (p) = ln p

e
+ e(1p)
with > 0. First, note that bL (p) is symmetry with respect to the interchange
p (1 p), and thus it is symmetric with respect to the value p = 0.5. We need
consider only the range 0 p 1/2 as the limit properties we prove there will
also hold for 1/2 p 1. In the range 0 p 1/2, we have min[p, 1 p] = p.

CHAPTER 2. BAYESIAN DECISION THEORY

60

The simplest way to prove that bL (p) is a lower bound for p in the range 0
p 1/2 is to note that bL (0) = 0, and the derivative is always less than the
derivative of min[p, 1 p]. Indeed, at p = 0 we have


1
1 + e
1
bL (0) = ln
= ln[1] = 0.

1 + e

Moreover the derivative is

bL (p)
p

e e2p
e + e2p

< 1=
min[p, 1 p]
p
=

in the range 0 p 1/2 for < .

(b) To show that bL (p) is an arbitrarily tight bound, we need show only that in the
limit , the derivative, bL (p)/p approaches 1, the same as

min[p, 1 p]
p
in this range. Using the results from part (a) we nd

bL (p)
p
lim

e e2p
=1
e + e2p
lim

in the range 0 p < 1/2.

(c) Our candidate upper bound is specied by
bU (p) = bL (p) + [1 2bL (0.5)]bG (p),
where gU (p) obeys several simple conditions, restated in part (d) below. We let
bL (p) = p (p), where from part (a) we know that (p) is non-negative and in
fact is at least linear in p. By the conditions given, we can write bG (p) = p+(p),
where (p) is non-negative and (0) = (1/2) = 0. Then our candidate upper
limit obeys
bU (p)

= p (p) + [1 2(1/2 (1/2))](p + (p))

= p (p) + (1/2)(p + (p)).

We show that this is an upper bound by calculating the dierence between this
bound and the Bayes limit (which is min[p, 1 p] = p in the range 0 p 1/2).
Thus we have
bU (p) p

= (p) + p(1/2) + (1/2)(p)

> 0.

(d) We seek to conrm that bG (p) = 1/2sin[p] has the following four properties:
bG (p) min[p, 1p]: Indeed, 1/2 sin[p] p for 0 p 1/2, with equality
holding at the extremes of the interval (that is, at p = 0 and p = 1/2). By
symmetry (see below), the relation holds for the interval 1/2 p 1.

PROBLEM SOLUTIONS

61

bG (p) = bG (1 p): Indeed, the sine function is symmetric about the point
/2, that is, 1/2 sin[/2 + ] = 1/2 sin[/2 ]. Hence by a simple
substitution we see that 1/2 sin[p] = 1/2 sin[(1 p)].
bG (0) = bG (1) = 0: Indeed, 1/2 sin[ 0] = 1/2 sin[ 1] = 0 a special
case of the fact bG (p) = bG (1 p), as shown immediately above.
bG (0.5) = 0.5: Indeed, 1/2 sin[0.5] = 1/2 1 = 0.5.
(e) See figure.
min[p,1-p]

0.5

0.5
bL(p;=50)

0.4

bU(p;=1)
bU(p;=10)

0.4

min[p,1-p]

bL(p;=10)
0.3

0.3

0.2

0.2

bU(p;=50)
bL(p;=1)
0.1

0.1

0.2

0.4

0.6

0.8

0.2

0.4

0.6

0.8

Section 2.9
43. Here the components of the vector x = (x1 , . . . , xd )t are binary-valued (0 or 1),
and
pij = Pr[xi = 1|j ]

i = 1, . . . , d
j = 1, . . . , c.

(a) Thus pij is simply the probability we get a 1 in feature xi given that the category
is j . This is the kind of probability structure we nd when each category has
a set of independent binary features (or even real-valued features, thresholded
in the form yi > yi0 ?).
(b) The discriminant functions are then
gj (x) = ln p(x|j ) + ln P (j ).
The components of x are statistically independent for all x in j , then we can
write the density as a product:
p(x|j )

= p((x1 , . . . , xd )t |j )
=

d
,

p(xi |j ) =

i=1

d
,

pxiji (1 pij )1xi .

i=1

Thus, we have the discriminant function

gj (x)

d
!

[xi ln pij + (1 xi ) ln (1 pij )] + ln P (j )

i=1

d
!

!
pij
xi ln
+
ln (1 pij ) + ln P (j ).
1 pij
i=1
i=1
d

CHAPTER 2. BAYESIAN DECISION THEORY

62

44. The minimum probability of error is achieved by the following decision rule:
Choose k if gk (x) gj (x) for all j = k,
where here we will use the discriminant function
gj (x) = ln p(x|j ) + ln P (j ).
The components of x are statistically independent for all x in j , and therefore,
p(x|j ) = p((x1 , . . . , xd )t |j ) =

d
,

p(xi |j ),

i=1

where
pij

Pr[xi = 1 |j ],

qij

Pr[xi = 0 |j ],

rij

Pr[xi = 1 |j ].

As in Sect. 2.9.1 in the text, we use exponents to select the proper probability,
that is, exponents that have value 1.0 when xi has the value corresponding to the
particular probability and value 0.0 for the other values of xi . For instance, for the
pij term, we seek an exponent that has value 1.0 when xi = +1 but is 0.0 when xi = 0
and when xi = 1. The simplest such exponent is 12 xi + 12 x2i . For the qij term, the
simplest exponent is 1x2i , and so on. Thus we write the class-conditional probability
for a single component xi as:
p(xi |j )

xi + 12 x2i 1x2i 12 xi + 12 x2i

qij rij

= pij2

i = 1, . . . , d
j = 1, . . . , c

and thus for the full vector x the conditional probability is

p(x|j )

d
,

xi + 12 x2i 1x2i 12 xi + 12 x2i

qij rij
.

pij2

i=1

Thus the discriminant functions can be written as

gj (x)

ln p(x|j ) + ln P (j )



1 2
1
1 2
2
=
+ ln P (j )
xi + xi ln pij + (1 xi )ln qij + xi + xi ln rij
2
2
2
2
i=1

d
d
d
!
pij rij
1!
pij !
2
=
xi ln
+
xi ln
+
ln qij + ln P (j ),
qij
2 i=1
rij
i=1
i+1
d 
!
1

which are quadratic functions of the components xi .

45. We are given that P (1 ) = P (2 ) = 1/2 and
pi1

= p > 1/2

pi2

1p

i = 1, . . . , d,

where d is the dimension, or number of features.

PROBLEM SOLUTIONS

63

(a) The minimum error decision rule is

d


Choose 1 if

i=1

>

xi ln
d

i=1

p
1p

xi ln

d


ln (1 p) + ln

i=1
1p
p

d


i=1

1
2

ln p + ln 12 .

This rule can be expressed as

1 d
2

!
p
1p
ln
ln
> d ln p d ln (1 p)
xi
1p
p
i=1
1 d
2

!
p
p
ln
xi
2 > d ln
1p
1p
i=1
or simply
d
!

xi > d/2.

i=1

(b) We denote the minimum probability of error as Pe (d, p). Then we have:
Pe (d, p)

= P (error|1 )P (1 ) + P (error|2 )P (2 )
1 d
2
1 d
2
!
!


= P
xi d/21 1/2 + P
xi > d/21 1/2.
i=1

As d is odd, and

d


i=1

is an integer, we have

i=1

Pe (d, p)

= P

 2
 2
1 d
!
d 1 
d + 1 
xi
xi
1 1/2 + P
2 1/2
2 
2 
i=1
i=1

1 d
!

(d1)/2

= 1/2

k=0


d k
p (1 p)dk + 1/2
k

d
!

 
d
(1 p)k pk .
k

k=(d+1)/2

We substitute k  = d k in the second summation, use the fact that

 d 
dk , and nd
(d1)/2

Pe (d, p)

1/2

k=0

(d1)/2

k=0

k =0

d k
p (1 p)dk .
k

(d1)/2 

lim Pe (d, p)

k


(d1)/2 
! d


d k
dk
p (1 p)
pk (1 p)dk
+ 1/2

k
k


(c) We seek the limit for p 1/2 of Pe (d, p). Formally, we write

p1/2

d

k=0

d
k


lim pk (1 p)dk

p1/2

CHAPTER 2. BAYESIAN DECISION THEORY

64

  d  d (d1)/2
! d
1
d
1
=
=
2
2
k
k
k=0
k=0

 d  !




d
d
1
d
1
1
1
2d
=
=
= .

k
2
2
2
2
2
(d1)/2 

k=0

Indeed, in the case p 1/2, the probability of error will be 1/2.

 
(d) Note that as d , the binomial probability kd pk (1 p)dk can be approximated by the normal density with mean dp and variance dp q. Thus we have
(d1)/2 

lim Pe (d, p)

lim

k=0


d k
p (1 p)dk
k

lim P (0 X (d 1)/2) where X N (dp, dp q)

d


dp
(d 1)/2 dp

= lim P
Z
where Z N (0, 1)
d
dp q
dp q




z(1/2 p) d/2

= lim P dp q Z
.
d
dp q

As P > 1/2, lim dp q = and lim d (1/2 p) = . Thus in the

=

limit of very large dimension, we have

lim Pe (d, p) = Pr( Z ) = 0.

46. The general minimum-risk discriminant rule is given by Eq. 16 in the text:
Choose 1 if (11 21 )p(x|1 )P (1 ) < (22 12 )p(x|2 )P (2 ); otherwise choose 2 .
Under the assumption 21 > 11 and 12 > 22 , we thus have
Choose 1 if

(21 11 )p(x|1 )P (1 )
> 1,
(12 22 )p(x|2 )P (2 )

or
Choose 1 if ln

p(x|1 )
P (1 )
21 11
> 0.
+ ln
+ ln
p(x|2 )
P (2 )
12 22

Thus the discriminant function for minimum risk, derived by taking logarithms in
Eq. 17 in the text, is:
g(x) = ln

p(x|1 )
P (1 )
21 11
.
+ ln
+ ln
p(x|2 )
P (2 )
12 22

For the case of independent features given in this problem, we have

ln

d
d
!
p(x|1 ) !
1 pi
=
wi xi +
ln
,
p(x|2 )
1 qi
i=1
i=1

where
wi = ln

pi (1 qi )
,
qi (1 pi )

i = 1, . . . , d.

PROBLEM SOLUTIONS

65

Therefore, the discriminant function can be written as:

g(x)

d
!

wi xi +

i=1

d
!

ln

i=1

21 11
1 pi
P (1 )
+ ln
+ ln
1 qi
P (2 )
12 22

= wt x + w0 .
47. Recall the Poisson distribution for a discrete variable x = 0, 1, 2, . . ., is given by
P (x|) = e

x
.
x!

(a) The mean, or expectation of this distribution is dened as

E[x] =

xP (x|) =

x=0

xe

x=0

!
x
x1
= e
= e e = .
x!
(x

1)!
x=1

(b) The variance of this distribution is

Var[x] = E[(x E[x])2 ] = E[x2 ] (E[x])2 .
  
2

To evaluate this variance, we need E[x2 ]:

E[x2 ]

= E[x(x 1) + x] = E[x(x 1)] + E[x]

!
x
=
+
(x(x 1)e
x!
x=0
= 2 e
= 2 e

!
x2
+
(x 2)!
x=2


!
x
+
x !


x =0

= 2 e e +
= 2 + ,
where we made the substitution x (x 2) in the fourth step. We put the
above results together and nd
Var[x] = E[(x E[x])2 ] = E[x2 ] (E[x])2 = 2 + 2 = .
(c) First recall the notation , read oor of , indicates the greatest integer less
than . For all x <  , then, we have
x
x1

x1
=
.
>
x!
(x 1)! 
x
(x 1)!
>1

That is, for x <  , the probability increases as x increases. Conversely,

for x >  we have
x
x1
x1

>
=
.
x!
(x 1)! 
x
(x 1)!
<1

CHAPTER 2. BAYESIAN DECISION THEORY

66

That is, for x > , the probability decreases as x increases. Thus, the
probability is maximized for integers x [, ]. In short, if is not an integer,
then x =  is the mode; if is an integer, then both  and are modes.
(d) We denote the two distributions
P (x|i ) = ei

xi
x!

for i = 1, 2, and by assumption 1 > 2 . The likelihood ratio is

 x
P (x|1 )
e1 x1
1
2 1
= 2 x = e
.
P (x|2 )
e
2
2
Thus the Bayes decision rule is
Choose 2

if

or equivalently

if

x
1
e
> 1,
2
(2 1 )
x<
;
ln[1 ] ln[2 ]
2 1

otherwise choose 1 ,
as illustrated in the gure (where the actual x values are discrete).
P(x|i)

0.2

0.15

0.1

0.05

x*

10

12

14

(e) The conditional Bayes error rate is



x
x
1 1
2 2
P (error|x) = min e
,e
.
x!
x!

The Bayes error, given the decision rule in part (d) is

PB (error) =

x
!
x=0

e2

!
x2
x
+
e1 1 ,
x! x=x
x!

where x = (2 1 )/(ln[1 ] ln[2 ]).

Section 2.10
48. In two dimensions, the Gaussian distribution is


1
1
t 1
p(x|i ) =
exp

(x

)
.
(x

i
i
i
2
2|i |1/2

PROBLEM SOLUTIONS

67

(a) By direct
  calculation using the densities stated in the problem, we nd that for
x = .3
.3 that p(x|1 )P (1 ) = 0.04849, p(x|2 )P (2 ) = 0.03250 and p(x|3 )P (3 ) =
0.04437, and thus the pattern should be classied as category 1 .

(b) To classify .3
, that is, a vector whose rst component is missing and its second
component is 0.3, we need to marginalize over the unknown feature. Thus we
compute numerically
  
   

x 
P (i )p
p
i = P (i )
i dx
.3
.3

and nd that P (1 )p((, .3) |1 ) = 0.12713, P (1 )p((, .3)t |2 ) = 0.10409, and

P (1 )p((, .3)t |3 ) = 0.13035. Thus the pattern should be categorized as 3 .
t

(c) As in part (a), we calculate numerically

  
   
.3 
.3 
P (i )
p
p
i = P (i )
i dy

and nd that P (1 )p((.3, ) |1 ) = 0.12713, P (1 )p((.3, )t |2 ) = 0.10409, and

P (1 )p((.3, )t |3 ) = 0.11346. Thus the pattern should be categorized as 1 .
t

(d) We follow the procedure in part (c) above:

x = (.2, .6)t
P (1 )p(x|1 ) = 0.04344.
P (2 )p(x|2 ) = 0.03556.
P (3 )p(x|3 ) = 0.04589.
Thus x = (.2, .6)t should be categorized as 3 .
x = (, .6)t
P (1 )p(x|1 ) = 0.11108.
P (2 )p(x|2 ) = 0.12276.
P (3 )p(x|3 ) = 0.13232.
Thus x = (, .6)t should be categorized as 3 .
x = (.2, )t
P (1 )p(x|1 ) = 0.11108.
P (2 )p(x|2 ) = 0.12276.
P (3 )p(x|3 ) = 0.10247.
Thus x = (, .6)t should be categorized as 2 .
49. Equation 95 in the text states
P (i |xg , xb ) =

gi (x)p(x)p(xb |xt )dxt

.
p(x)p(xb |xt )dxt

We are given that the true features, xt , are corrupted by Gaussian noise to give us
the measured bad data, that is,


1
1
t 1
p(xb |xt ) =
exp (xt i ) (xt i ) .
2
(2)d/2 ||

CHAPTER 2. BAYESIAN DECISION THEORY

68

We drop the needless subscript i indicating category membership and substitute this
Gaussian form into the above. The constant terms, (2)d/2 ||, cancel, and we then
nd that the probability of the category given the good and the measured bad data
is

g(x)p(x)exp[ 12 (xt )t 1 (xt )]dxt
P (|xg , xb ) =
.
p(x)exp[ 12 (xt )t 1 (xt )]dxt ]dxt
After integration, this gives us the nal answer,
P (|xg , xb ) =

P ()p(xg , xb |)
,
p(xg , xb )

which is the result from standard Bayesian considerations.

Section 2.11
50. We use the values from Example 4 in the text.
(a) For this case, the probabilities are:
P (a1 )

= P (a4 ) = 0.5

P (a2 )

= P (a3 ) = 0

P (b1 )

P (b2 )

P (d1 )

P (d2 )

1.

Then using Eq. 99 in the text we have

PP (x1 ) P (x1 |a1 , b1 )P (a1 )P (b1 ) + 0 + 0 + 0 + 0 + 0 + P (x1 |a4 , b1 )P (a4 )P (b1 ) + 0
0.9 0.65
0.8 0.65
=
0.5 1 +
0.5 1
0.9 0.65 + 0.1 0.35
0.8 0.65 + 0.2 0.35
= 0.472 + 0.441
= 0.913.
A similar calculation gives
PP (x2 ) P (x2 |a1 , b1 )P (a1 )P (b1 ) + 0 + 0 + 0 + 0 + 0 + P (x2 |a4 , b1 )P (a4 )P (b1 ) + 0
0.35 0.1
0.2 0.35
=
0.5 1 +
0.5 1
0.9 0.65 + 0.1 0.35
0.8 0.65 + 0.2 0.35
= 0.87.
Since the lightness is not measured, we can consider only thickness
PC (x1 ) P (eD |x2 )
= P (eD |d1 )P (d1 |x1 ) + P (eD |d2 )P (d2 |x1 )
=
=

0 0.4 + 1 0.6
0.6.

PROBLEM SOLUTIONS

69

We normalize these and nd

P (x1 |e)

P (x2 |e)

0.913 0.6
= 0.992
0.913 0.6 + 0.087 0.05
0.087 0.05
= 0.008.
0.913 0.6 + 0.087 0.05

Thus, given all the evidence e throughout the belief net, the most probable
outcome is x1 , that is, salmon. The expected error is the probability of nding
a sea bass, that is, 0.008.
(b) Here we are given that the sh is thin and medium lightness, which implies
P (eD |d1 ) = 0,

P (eD |d2 ) = 1

P (eC |c1 ) = 0,

P (eC |c2 ) = 1, P (eC |c3 ) = 0.

and as in part (a) we have

PC (x1 ) P (eC |x1 )P (eD |x1 )
= [P (eC |c1 )P (c1 |x1 ) + P (eC |c2 )P (c2 |x1 ) + P (eC |c3 )P (c3 |x1 )]
[P (eD |d1 )P (d1 |x1 ) + P (eD |d2 )P (d2 |x1 )]
= [0 + 1 0.33 + 0][0 + 1 0.6]
=

0.198.

Likewise we have
PC (x2 ) P (eC |x2 )P (eD |x2 )
= [P (eC |c1 )P (c2 |x2 ) + P (eC |c2 )P (c2 |x2 ) + P (eC |c3 )P (c3 |x2 )]
[P (eD |d1 )P (d1 |x2 ) + P (eD |d2 )P (d2 |x2 )]
= [0 + 1 0.1 + 0][0 + 1 0.05]
=

0.005.

We normalize and nd
P (x1 |eC,D )

P (x2 |eC,D )

0.198
= 0.975
0.198 + 0.005
0.005
= 0.025.
0.198 + 0.005

Thus, from the evidence of the children nodes, we classify the sh as x1 , that
is, salmon.
Now we infer the probability P (ai |x1 ). We have P (a1 ) = P (a2 ) = P (a3 ) =
P (a4 ) = 1/4. Then we normalize and nd
P (a1 |x1 )

P (x1 |a1 )P (a1 )

P (x1 |a1 ) P (a1 ) + P (x1 |a2 ) P (a2 ) + P (x1 |a3 )P (a3 ) + P (x1 |a4 )P (a4 )
0.9 0.25
=
0.9 0.25 + 0.3 0.25 + 0.4 0.25 + 0.8 0.25
= 0.375.
=

CHAPTER 2. BAYESIAN DECISION THEORY

70
We also have

P (a2 |x1 )

0.125

P (a3 |x1 )

=
=

0.167
0.333.

P (a4 |x1 )

Thus the most probable season is a1 , winter. The probability of being correct
is
P (x1 |eC,D )P (a1 |x1 ) = 0.975 0.375 = 0.367.
(c) Fish is thin and medium lightness, so from part (b) we have
P (x1 |eC,D ) = 0.975, P (x2 |eC,D ) = 0.025,
and we classify the sh as salmon.
For the sh caught in the north Atlantic, we have P (b1 |eB ) = 1 and P (b2 |eB ) =
0.
P (a1 |x1 , b1 ) P (x1 |a1 )P (x1 |b1 )P (a1 )P (b1 )
= 0.9 0.65 0.25 = 0.146
P (a2 |x1 , b1 ) P (x1 |a2 )P (x1 |b1 )P (a2 )P (b1 )
= 0.3 0.65 0.25 = 0.049
P (a3 |x1 , b1 ) P (x1 |a3 )P (x1 |b1 )P (a3 )P (b1 )
= 0.4 0.65 0.25 = 0.065
P (a4 |x1 , b1 ) P (x1 |a4 )P (x1 |b1 )P (a4 )P (b1 )
= 0.8 0.65 0.25 = 0.13.
So the most probable season is a1 , winter. The probability of being correct is
P (x1 |eC,D )P (a1 |x1 ) = 0.975 0.375 = 0.367.
51. Problem not yet solved
Section 2.12
52. We have the priors P (1 ) = 1/2 and P (2 ) = P (3 ) = 1/4, and the Gaussian
densities p(x|1 ) N (0, 1), p(x|2 ) N (0.5, 1), and p(x|3 ) N (1, 1). We use the
general form of a Gaussian, and by straightforward calculation nd:
x
0.6
0.1
0.9
1.1

p(x|1 )
0.333225
0.396953
0.266085
0.217852

p(x|2 )
0.396953
0.368270
0.368270
0.333225

p(x|3 )
0.368270
0.266085
0.396953
0.396953

We denote X = (x1 , x2 , x3 , x4 ) and = ((1), (2), (3), (4)). Using the notation
in Section 2.12, we have n = 4 and c = 3, and thus there are cn = 34 = 81 possible

PROBLEM SOLUTIONS

71

values of , such as
(1 , 1 , 1 , 1 ), (1 , 1 , 1 , 2 ), (1 , 1 , 1 , 3 ),
(1 , 1 , 2 , 1 ), (1 , 1 , 2 , 2 ), (1 , 1 , 2 , 3 ),
(1 , 1 , 3 , 1 ), (1 , 1 , 3 , 2 ), (1 , 1 , 3 , 3 ),
..
..
..
.
.
.
(3 , 3 , 3 , 1 ), (3 , 3 , 3 , 2 ), (3 , 3 , 3 , 3 )
For each possible value of , we calculate P () and p(X|) using the following, which
assume the independences of xi and (i):
p(X|)

4
,

p(xi |(i))

i=1

P ()

4
,

P ((i)).

i=1

For example, if = (1 , 3 , 3 , 2 ) and X = (0.6, 0.1, 0.9, 1.1), then we have

p(X|)

= p((0.6, 0.1, 0.9, 1.1)|(1 , 3 , 3 , 2 ))

= p(0.6|1 )p(0.1|3 )p(0.9|3 )p(1.1|2 )
= 0.333225 0.266085 0.396953 0.333225
=

0.01173

and
P ()

= P (1 )P (3 )P (3 )P (2 )
1 1 1 1
=

2 4 4 4
1
=
= 0.0078125.
128

(a) Here we have X = (0.6, 0.1, 0.9, 1.1) and = (1 , 3 , 3 , 2 ). Thus, we have
p(X)

= p(x1 = 0.6, x2 = 0.1, x3 = 0.9, x4 = 1.1)

!
=
p(x1 = 0.6, x2 = 0.1, x3 = 0.9, x4 = 1.1|)P ()

= p((x1 = 0.6, x2 = 0.1, x3 = 0.9, x4 = 1.1)|(1 , 1 , 1 , 1 ))P (1 , 1 , 1 , 1 )

+p((x1 = 0.6, x2 = 0.1, x3 = 0.9, x4 = 1.1)|(1 , 1 , 1 , 2 ))P (1 , 1 , 1 , 2 )
..
.
+p((x1 = 0.6, x2 = 0.1, x3 = 0.9, x4 = 1.1)|(3 , 3 , 3 , 3 ))P (3 , 3 , 3 , 3 )
= p(0.6|1 )p(0.1|1 )p(0.9|1 )p(1.1|1 )P (1 )P (1 )P (1 )P (1 )
+p(0.6|1 )p(0.1|1 )p(0.9|1 )p(1.1|2 )P (1 )P (1 )P (1 )P (2 )
..
.
+p(0.6|3 )p(0.1|3 )p(0.9|3 )p(1.1|3 )P (3 )P (3 )P (3 )P (3 )
=

0.012083,

CHAPTER 2. BAYESIAN DECISION THEORY

72

where the sum of the 81 terms can be computed by a simple program, or somewhat tediously by hand. We also have
p(X|)

= p(0.6, 0.1, 0.9, 1.1|1 , 3 , 3 , 2 ) = p(0.6|1 )p(0.1|3 )p(0.9|3 )p(1.1|2 )

= 0.33325 0.266085 0.396953 0.333225
= 0.01173.

Also we have
P () = P (1 )P (3 )P (3 )P (2 )
1 1 1 1
=

2 4 4 4
1
=
= 0.0078125.
128
According to Eq. 103 in the text, we have
P (|X)

= P (1 , 3 , 3 , 2 |0.6, 0.1, 0.9, 1.1)

p(0.6, 0.1, 0.8, 1.1|1 , 3 , 3 , 2 )P (1 , 3 , 3 , 2 )
=
p(X)
0.01173 0.0078125
=
= 0.007584.
0.012083

(b) We follow the procedure in part (a), with the values X = (0.6, 0.1, 0.9, 1.1) and
= (1 , 2 , 2 , 3 ). We have
p(X|)

= p(0.6, 0.1, 0.9, 1.1|1 , 2 , 2 , 3 )

= p(0.6|1 )p(0.1|2 )p(0.9|2 )p(1.1|3 )
=

0.33225 0.368270 0.368270 0.396953 = 0.01794.

Likewise, we have
P () = P (1 )P (2 )P (2 )P (3 )
1 1 1 1
=

2 4 4 4
1
=
= 0.0078125.
128
Thus
P (|X)

= P (1 , 2 , 2 , 3 |0.6, 0.1, 0.9, 1.1)

p(0.6, 0.1, 0.9, 1.1|1 , 2 , 2 , 3 )P (1 , 2 , 2 , 3 )
=
P (X)
0.01794 0.0078125
=
= 0.01160.
0.012083

(c) Here we have X = (0.6, 0.1, 0.9, 1.1) and = ((1), (2), (3), (4)). According
to Eq. 103 in the text, the sequence that maximizes p(X|)P () has the
maximum probability, since p(X) is xed for a given observed X. With the
simplications above, we have
p(X|)P () = p(x1 |(1))p(x2 |(2))p(x3 |(3))p(x4 |(4))P ((1))P ((2))P ((3))P ((4)).

PROBLEM SOLUTIONS

73

For X = (0.6, 0.1, 0.9, 1.1), we have

max [p(0.6|1 )P (1 ), p(0.6|2 )P (2 ), p(0.6|3 )P (3 )]
= max [0.333225 0.5, 0.396953 0.25, 0.368270 0.25]
= 0.333225 0.5
=

0.1666125.

Likewise, for the second step we have

max [p(0.1|1 )P (1 ), p(0.1|2 )P (2 ), p(0.1|3 )P (3 )]
= max [0.396953 0.5, 0.368270 0.25, 0.266085 0.25]
= 0.396953 0.5
= 0.1984765.
For the third step we have
max [p(0.9|1 )P (1 ), p(0.9|2 )P (2 ), p(0.9|3 )P (3 )]
=

max [0.266085 0.5, 0.368270 0.25, 0.396953 0.25]

= 0.266085 0.5
= 0.133042.
For the nal step we have
max [p(1.1|1 )P (1 ), p(1.1|2 )P (2 ), p(1.1|3 )P (3 )]
=

max [0.217852 0.5, 0.333225 0.25, 0.396953 0.25]

= 0.217852 0.5
= 0.108926.
Thus the sequence = (1 , 1 , 1 , 1 ) has the maximum probability to observe
X = (0.6, 0.1, 0.9, 1.1). This maximum probability is
0.166625 0.1984765 0.133042 0.108926

1
= 0.03966.
0.012083

CHAPTER 2. BAYESIAN DECISION THEORY

74

Computer Exercises
Section 2.5
1. Computer exercise not yet solved
2.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38

MATLAB program
load samples.mat;
the data
[n,m] = size(samples);
for i=1:3
mu{i} = mean(samples(:, (i-1)*3+1;i*3));
sigma{i} = zeros(3);
for j=1:n
sigma{i} = sigma{i} + ...
The ... continues the line
(samples(j,(i-1)*3+1:i*3) - mu{i}) ...
* (samples(j,(i-1)*3+1;i*3) - mu{i});
end
sigma{i} = sigma{i}./n;
end
s = [1 2 1; 5 3 2; 0 0 0; 1 0 0]
for j=1:size(s,2)
for i=1;3
d = sqrt((s:,j)-mu{i})*inv(sigma{i})*(s(:,j)-mu{i}));
fprintf(Mahal. dist. for class %d and point %d: %: %f\n, i, j, d);
end
end
pw(1:) =[1/3 0.8];
pw(2:) =[1/3 0.1];
pw(3:) =[1/3 0.1];
for p=1:2
fprintf(\n\n\n\n);
for j=1:size(s,2)
class = 0; max gi = -1000000;
for i=1:3
d i = (s(:,j)-mu{i})*inv(sigma{i})*(s(:,j)-mu{i});
g i = -0.5*d i - 1.5*log(2*pi) - 0.5*log(det(sigma{i})) + log(pw(i,p));
if gi > max gi,
max gi = gi;
class = i;
end
end
fprintf(Point %d classified in category %d\n, j, class);
end
end

Output

Mahal. dist. for class 1 and point 1: 1.069873

Mahal. dist. for class 2 and point 1: 0.904465
Mahal. dist. for class 3 and point 1: 2.819441

COMPUTER EXERCISES
Mahal.
Mahal.
Mahal.
Mahal.
Mahal.
Mahal.
Mahal.
Mahal.
Mahal.

dist.
dist.
dist.
dist.
dist.
dist.
dist.
dist.
dist.

for
for
for
for
for
for
for
for
for

class
class
class
class
class
class
class
class
class

1
2
3
1
2
3
1
2
3

and
and
and
and
and
and
and
and
and

75
point
point
point
point
point
point
point
point
point

Point
Point
Point
Point

1
2
3
4

classified
classified
classified
classified

in
in
in
in

category
category
category
category

2
3
1
1

Point
Point
Point
Point

1
2
3
4

classified
classified
classified
classified

in
in
in
in

category
category
category
category

1
1
1
1

2:
2:
2:
3:
3:
3:
4:
4:
4:

3. Computer exercise not yet solved

4. Computer exercise not yet solved
5. Computer exercise not yet solved
Section 2.8
6. Computer exercise not yet solved
7. Computer exercise not yet solved
8. Computer exercise not yet solved
Section 2.11
9. Computer exercise not yet solved

1.641368
1.850650
0.682007
0.516465
0.282953
2.362750
0.513593
0.476275
1.541438

76

CHAPTER 2. BAYESIAN DECISION THEORY

Chapter 3

Maximum likelihood and

Bayesian parameter
estimation

Problem Solutions
Section 3.2
1. Our exponential function is:

p(x|) =

ex
0

x0
otherwise.

(a) See Figure. Note that p(x = 2|) is not maximized when = 2 but instead
for a value less than 1.0.
p(x| =1)

p(x=2| )
0.2

0.1

0.5

x
1

=1 (part c)

(b) The log-likelihood function is

l() =

n
!
k=1

ln p(xk |) =

n
!

[ln xk ] = nln

k=1

n
!
k=1

77

xk .

78CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

We solve l() = 0 to nd as
l()



n
!

n ln
xk

k=1

n !

xk = 0.

k=1

Thus the maximum-likelihood solution is

=

1
n

1
n

k=1

.
xk

(c) Here we approximate the mean

1!
xn
n
n

k=1

by the integral

xp(x)dx,
0

which is valid in the large n limit. Noting that



xex dx = 1,

we put these results together and see that = 1, as shown on the gure in
part (a).
2. Our (normalized) distribution function is

1/ 0 x
p(x|) =
0 otherwise.
(a) We will use the notation of an indicator function I(), whose value is equal to
1.0 if the logical value of its argument is true, and 0.0 otherwise. We can write
the likelihood function using I() as
p(D|)

n
,

p(xk |)

k=1
n
,

1
I (0 xk )

k=1
 


1
=
I max xk I min xk 0 .
k
k
n

We note that 1/n decreases monotonically as increases but also that I(

max xk ) is 0.0 if is less than the maximum value of xk . Therefore, our likelihood
k

function is maximized at = max xk .

k

PROBLEM SOLUTIONS

79

p(D| )
12
10
8
6
4
2

0.2

0.4

0.6

0.8

(b) See Figure.

3. We are given that

1 if the state of nature for the k th sample is i
zik =
0 otherwise.
(a) The samples are drawn by successive independent selection of a state of nature
i with probability P (i ). We have then
Pr[zik = 1|P (i )] = P (i )
and
Pr[zik = 0|P (i )] = 1 P (i ).
These two equations can be unied as
P (zik |P (i ))

[P (i )]zik [1 P (i )]1zik .

By the independence of the successive selections, we have

P (zi1 , , zin |P (i ))

=
=

n
,
k=1
n
,

P (zik |P (i ))
[P (i )]zik [1 P (i )]1zik .

k=1

(b) The log-likelihood as a function of P (i ) is

l(P (i ))

ln P (zi1 , , zin |P (i ))
 n

,
zik
1zik
ln
[P (i )] [1 P (i )]

n
!


zik ln P (i ) + (1 zik ) ln (1 P (i )) .

k=1

k=1

Therefore, the maximum-likelihood values for the P (i ) must satisfy

!
1 !
1
P (i ) l(P (i )) =
zik
(1 zik ) = 0.
P (i )
1 P (i )
n

k=1

k=1

80CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

We solve this equation and nd

(1 P (i ))

n
!

= P (i )

zik

k=1

n
!

(1 zik ),

k=1

which can be rewritten as

n
!

zik

= P (i )

k=1

n
!

zik + nP (i ) P (i )

k=1

n
!

zik .

k=1

The nal solution is then

1!
zik .
n
n

P (i )

k=1

That is, the estimate of the probability of category i is merely the probability
of obtaining its indicatory value in the training data, just as we would expect.
4. We have n samples {x1 , . . . , xn } from the discrete distribution
P (x|) =

d
,

ixi (1 i )1xi .

i=1

The likelihood for a particular sequence of n samples is

P (x1 , . . . , xn |) =

n ,
d
,

ixki (1 i )1xki ,

k=1 i=1

and the log-likelihood function is then

l()

n !
d
!

xki ln i + (1 xki ) ln (1 i ).

k=1 i=1

To nd the maximum of l(), we set l() = 0 and evaluate component by component (i = 1, . . . , d) and get
[ l()]i

= i l()
=

n
n
1 !
1 !

(1 xki )
i
1 i

0.

k=1

k=1

This implies that for any i

n
n
1 !
1 !
xki =
(1 xki ),
1 i k=1
i k=1

which can be rewritten as

(1 i )

n
!
k=1

1
xki

= i

n
!
k=1

2
xki

PROBLEM SOLUTIONS

81

The nal solution is then

1!
i =
xki .
n
n

k=1

Since this result is valid for all i = 1, . . . , d, we can write this last equation in vector
form as
!
= 1

xk .
n
n

k=1

Thus the maximum-likelihood value of is merely the sample mean, just as we would
expect.
5. The probability of nding feature xi to be 1.0 in category 1 is denoted p:
p(xi = 1|1 ) = 1 p(xi = 0|1 ) = pi1 = p >

1
,
2

for i = 1, . . . , d. Moreover, the normalization condition gives pi2 = p(xi |2 ) = 1 pi1 .

(a) A single observation x = (x1 , . . . , xd ) is drawn from class 1 , and thus have
p(x|1 ) =

d
,

p(xi |1 ) =

i=1

d
,

pxi (1 p)1xi ,

i=1

and the log-likelihood function for p is

l(p) = ln p(x|1 ) =

d
!

[xi ln p + (1 xi ) ln (1 p)].

i=1

Thus the derivative is

!
1!
1
xi
(1 xi ).
p i=1
(1 p) i=1
d

p l(p)

We set this derivative to zero, which gives

d
d
1!
1 !
xi =
(1 xi ),
p i=1
1 p i=1

which after simple rearrangement gives

(1 p)

d
!

xi

= p d

i=1

d
!

2
xi

i=1

Thus our nal solution is

1!
xi .
d i=1
d

p =

That is, the maximum-likelihood estimate of the probability of obtaining a 1 in

any position is simply the ratio of the number of 1s in a single sample divided
by the total number of features, given that the number of features is large.

82CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

(b) We dene T = (1/d)

d

i=1

xi to be the proportion of 1s in a single observation x.

As the number of dimensions d approaches innity, we have

1!
E(xi |1 ) = [1 p + 0 (1 p)] = p.
d i=1
d

T =

Likewise, the variance of T , given that were considering just one class, 1 , is
=

1!
Var(xi |1 )
d i=1

d
1 ! 2
[1 p + 02 (1 p) p p]
d2 i=1

p(1 p)
,
d

Var(T |1 )

which vanishes as d . Clearly, for minimum error, we choose 1 if T >

T = 1/2 for p > 1/2. Since the variance vanishes for large d, the probability
of error is zero for a single sample having a suciently large d.
(c) See figure.
P(T|i)

d = 111

d = 11

0.2

0.4

T* 0.6

T
0.8

6. The d-dimensional multivariate normal density is given by



1
1
t 1
p(x|, ) =
exp

(x

)
.
(x

)
2
(2)d/2 ||1/2
We choose x1 , x2 , . . . , xn independent observations from p(x|, ). The joint density
is


n
1
1!
p(x1 , x2 , . . . , xn |, ) =
exp
(xk )t 1 (xk ) .
2
(2)n/2 ||n/2
k=1

The log-likelihood function of and is

n
n
1!
= ln (2) ln ||
(xk )t 1 (xk )
2
2
2
k=1

 n
n
n
1 ! t
= ln (2) ln ||
xk xk 2t 1 xk + nt 1 .
2
2
2
n

l(, )

k=1

PROBLEM SOLUTIONS

83

We set the derivative of the log-likelihood to zero, that is,



n
!
l(, )
1
1
1
xk + n2 = 0,
= 2

2
k=1

and nd that
1

n
!

1 .

xk = n

k=1

This gives the maximum-likelihood solution,

1!
xk ,
n
n

k=1

we temporarily substitute
as expected. In order to simplify our calculation of ,
A = 1 , and thus have
n
n
1!
l(, ) = ln (2) + ln |A|
(xk )t A(xk ).
2
2
2
n

k=1

We use the above results and seek the solution to

n
l(, )
n 1 1 !
(xk )(xk )t = 0.
=
A
A
2
2
k=1

We now replace A by

and nd that

n
1!
k )
t,
=
(xk )(x
2
2
n

k=1

and then multiply both sides by 2/n and nd our solution:

!
= 1
k )
t.

(xk )(x
n
n

k=1

As we would expect, the maximum-likelihood estimate of the covariance matrix is

merely the covariance of the samples actually found.
7. The gure shows the model for 1 , the incorrect model for 2 , that is, p(x|2 )
N (0, 1) and the true model for 2 (dashed, and scaled by 30000 for visibility).
(a) According to Eq. 18 in the text, the maximum-likelihood estimate of the mean
(for the incorrect model) is merely the sample mean of the data from the true
model. In this case, then, we have
= 1.0.
(b) Given two equal-variance Gaussians and equal priors, the decision boundary
point is midway between the two means, here, x = 0.5.
(c) For the incorrect case, according to part (b), we have R1 is the line segment
x < 0.5 and for R2 the line segment x > 0.5. For the correct case, we must
solve numerically
1
1
exp[x2 /2] =

exp[1(x 1)2 /(2 106 )],

2 1
2 106
which gives the values x = 3.717. The Bayesian decision boundaries and
regions are shown along the bottom of the gure above.

84CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

p(x|i)
0.4
p(x|2) as incorrect model

p(x|1)
0.3

0.2
true p(x|2) model
(shown 30000 x)

0.1

-6
R2

-4
-2
-3.717

0 x*
R1

4
+3.717

R2

(d) For the poor model, which has equal variances for the two Gaussians, the only
decision boundary possible is a single point, as was illustrated in part (a). The
best we can do with the incorrect model is to adjust the mean so as to match
the rightmost of the decision boundaries given in part (c), that is, x = 3.717.
This decision point is midway between the means of the two Gaussians that
is, at 0 and at . As such, we need to choose such that (0 + )/2 = 3.717.
This leads directly to the solution, = 7.43.
(e) The maximum-likelihood solution in the incorrect model here p(x|2 )
N (, 1) does not yield the minimum classication error. A dierent value
of the parameter (here, = 7.43) approximates the Bayes decision boundary
better and gives a lower error than the maximum-likelihood solution in the
incorrect model. As such, the faulty prior information about the model, which
did not permit the true model to be expressed, could not be overcome with even
an innite amount of training data.
8. Consider a case in which the maximum-likelihood solution gives the worst possible
classier.
(a) In this case, the symmetry operation x x takes p(x|1 ) p(x|2 ), and
thus we are assured that the estimated distributions have this same symmetry
property. For that reason, we are guaranteed that these distribution have the
same value at x = 0, and thus x = 0 is a decision boundary. Since the Gaussian
estimates must have the same variance, we are assured that there will be only
a single intersection, and hence a single decision point, at x = 0.
(b) See Figure.
(c) We have the estimate of the mean as


1 = p(x|1 )dx = (1 k)1 + k(X) = 1 k(1 + X).

We are asked to switch the mean, that is, have
1 < 0 and
2 > 0. This can
be assured if X > (1 k)/k. (We get the symmetric answer for 2 .)
(d) Since the decision boundary is at x = 0, the error is 1 k, that is, the value of
the distribution spikes on the wrong side of the decision boundary.

PROBLEM SOLUTIONS

85
p(x|i)
1-k=.8

1-k=.8

k=.2

k=.2

x
-6

-4

-2

x*

(e) Thus for the error to approach 1, we must have k 0, and this, in turn, requires
X to become arbitrarily large, that is, X (1 k)/k . In this way, a tiny
spike in density suciently far away from the origin is sucient to switch
the estimated mean and give error equal 100%.
(f) There is no substantive dierence in the above discussion if we constrain the
variances to have a xed value, since they will always be equal to each other, that
is,
12 =
22 . The equality of variances preserves the property that the decision
point remains at x = 0; it consequently also ensures that for suciently small
k and large X the error will be 100%.
(g) All the maximum-likelihood methods demand that the optimal solution exists
in the model space. If the optimal solution does not lie in the solution space,
then the proofs do not hold. This is actually a very strong restriction. Note
that obtaining the error equal 100% solution is not dependent upon limited
data, or getting caught in a local minimum it arises because of the error in
assuming that the optimal solution lies in the solution set. This is model error,
as described on page 101 of the text.
9. The standard maximum-likelihood solution is
= arg max p(x|).

Now consider a mapping x (x) where () is continuous. Then we can write

p( |)d = p(x|)dx, and
p( |) =

p(x|)
.
d /dx

Then we nd the value of maximizing p( (x)|) as

arg max p( (x)|)

=
=

p(x|)
d /dx
arg max p(x|)
arg max

= ,

86CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

In short, then, the maximum-likelihood

where we have assumed d /dx = 0 at = .

value of () is indeed . In practice, however, we must check whether the value of

derived this way gives a maximum or a minimum (or possibly inection point) for
p( |).
10. Consider the novel method of estimating the mean of a set of points as taking its
rst value, which we denote M = x1 .
(a) Clearly, this unusual estimator of the mean is unbiased, that is, the expected
value of this statistic is equal to the true value. In other words, if we repeat the
selection of the rst point of a data set we have
K
1 !
M(k) = 0,
K K

bias = E[M] = lim

k=1

where M(k) is the rst point in data set k drawn from the given distribution.
(b) While the unusual method for estimating the mean may indeed be unbiased, it
will generally have large variance, and this is an undesirable property. Note that
E[(xi )2 ] = 2 , and the RMS error, , is independent of n. This undesirable
behavior is quite dierent from that of the measurement of
1!
xi ,
n i=1
n

x
=
where we see

1
22
n
!
1
E[(
x )2 ] = E
xi
n i=1
=

n

1 !
E[(xi )2 ]
2
n i=1

2
.
n

Thus the RMS error, / n, approches 0 as 1/ n. Note that there are many
superior methods for estimating the mean, for instance the sample mean. (In
Chapter 9 we shall see other techniques ones based on resampling such as
the so-called bootstrap and jackknife methods.)
11. We assume p2 (x) p(x|2 ) N (, ) but that p1 (x) p(x|1 ) is arbitrary.
The Kullback-Leibler divergence from p1 (x) to p2 (x) is




1
DKL (p1 , p2 ) = p1 (x)lnp1 (x)dx +
p1 (x) dln(2) + ln|| + (x )t 1 (x ) dx,
2
where we used the fact that p2 is a Gaussian, that is,


1
(x )t 1 (x )
p2 (x) =
exp

.
2
(2)d/2 ||1/2
We now seek and to minimize this distance. We set the derivative to zero and
nd


DKL (p1 , p2 ) = 1 (x )p1 (x)dx = 0,

PROBLEM SOLUTIONS
and this implies
1

87


p1 (x)(x )dx = 0.

We assume is non-singular, and hence this equation implies


p1 (x)(x )dx = E1 [x ] = 0,
or simply, E1 [x] = . In short, the mean of the second distribution should be the
same as that of the Gaussian.
Now we turn to the covariance of the second distribution. Here for notational
convenience we denote A = . Again, we take a derivative of the Kullback-Leibler
divergence and nd:




DKL (p1 , p2 ) = 0 = p1 (x) A1 + (x )(x )t dx,

A
and thus


E1 (x )(x )t ,
or


E1 (x )(x )t = .
In short, the covariance of the second distribution should indeed match that of the
Gaussian.
Note that above, in taking the derivative above,
|A|
= |A|A1
A
we relied on the fact that A = 1 is symmetric since is a covariance matrix. More
generally, for an arbitrary non-singular matrix we would use
|M|
= |M|(M1 )t .
M
Section 3.3
12. In the text we saw the following results:
1. The posterior density can be computed as

p(x) = p(x, |D) d.
2. p(x, |D) = p(x|, D)p(|D)).
3. p(x|, D) = p(x|), that is, the distribution of x is known completely once we
know the value of the parameter vector, regardless of the data D.

4. p(x|D) = p(x|)p(|D) d.

88CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

These are justied as follows:

1. This statement reects the conceptual dierence between the maximum-likelihood
estimator and Bayesian estimator. The Bayesian learning method considers the
parameter vector to be a random variable rather than a xed value, as in
maximum-likelihood estimation. The posterior density p(x|D) also depends
upon the probability density p() distributed over the entire space instead of
a single value. Therefore, the p(x|D) is the integration of p(x, |D) over the
entire parameter space. The maximum-likelihood estimator can be regarded as
a special case of Bayesian estimator, where p() is uniformly distributed so that
its eect disappears after the integration.
2. The p(x, |D) implies two steps in computation. One is the computation of
the probability density given the data set D, that is, p(|D). The other is
the computation of the probability density of x given , that is, p(x|, D). The
above two steps are independent of each other, and thus p(x, |D) is the product
of the results of the two steps.
3. As mentioned in the text, the selection of x and that of the training samples
D is done independently, that is, the selection of x does not depend upont D.
Therefore we have p(x|, D) = p(x|).
4. We substitute the above relations into Eq. 24 in the text and get Eq. 25.
Section 3.4
13. We seek a novel approach for nding the maximum-likelihood estimate for .
(a) We rst inspect the forms of a general vector

a1
A11
a2

..
a = . and A = .
..
An1
an

a and matrix A:
...
..
.
...

A1n
.. .
.
Ann

Consider the scalar

at Aa =

n !
n
!

aj Aij aj .

i=1 j=1

The (i, i)th element of this scalar is

n

j=1

Aij aj ai , and the trace of Aaat is the

sum of these diagonal elements, that is,

tr (Aaat ) =

n !
n
!

Aij aj ai = at Aa.

i=1 j=1

(b) We seek to show that the likelihood function can be written as


2
1
n
!
1
1
p(x1 , . . . , xn |) =
exp tr 1
(xk )(xk )t
.
2
(2)nd/2 |1 |n/2
k=1

PROBLEM SOLUTIONS

89

We note that p(x|) N (, ) where is known and x1 , . . . , xn are independent observations from p(x|). Therefore the likelihood is
p(x1 , . . . , xn |)



1
1
t 1
exp

(x

)
(x
k
k
2
(2)d/2 |1 |1/2
k=1


n
||n/2
1!
t 1
=
exp
(xk ) (xk ) .
2
(2)nd/2
=

n
,

k=1

From the results in part (a), with a = xk and |A| = |1 |, we have



n

||n/2
1 !  1
t
p(x1 , . . . , xn |) =
exp
tr (xk )(xk )
2
(2)nd/2
k=1
2

1
n
!
|1 |n/2
1
1
t
=
,
exp tr
(xk )(xk )
2
(2)nd/2
k=1

where we used

n

k=1


tr (Ak ) = tr

n

k=1


Ak

and |1 | = ||1 .

(c) Recall our denition of the sample covariance matrix:

!
= 1

(xk )(xk )t .
n
n

k=1

which easily leads to the following equalities

Here we let A = 1 ,
1

,
= A
1 | = |A||1 |
|1 | = |A
1 = |A|
= |A|||

||
=

1 2 d
,

||

where 1 , . . . , n are the eigenvalues of A. We substitute these into our result

in part (b) to get
2

1
n
!
|1 |n/2
1
1
t
p(x1 , . . . , xn |) =
exp tr
(xk )(xk )
2
(2)nd/2
k=1

#
(1 . . . n )n/2
1 "
.
=
exp tr n1
n/2
2
(2)nd/2 ||
= n[tr (A)] = n(1 + + d ), and thus we
Note, however, that tr[n1 ]
have
p(x1 , . . . , xn |) =

$ n
%
(1 . . . n )n/2
exp (1 + + d ) .
n/2
2
(2)nd/2 ||

90CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

(d) The expression for p(x1 , , xn |) in part (c) depends on only through
We can write our likelihood, then, as
1 , , d , the eigenvalues of A = 1 .
p(x1 , . . . , xn |) =

1
n/2
(2)nd/2 ||

d
,

n/2
i

i e

i=1

Maximizing p(x1 , . . . , xn |) with respect to is equivalent to maximizing

i ei with respect to i . We do this by setting the derivative to zero, that is,
[i ei ]
= ei + i (ei ) = 0,
i
which has solution i = 1. In short, p(x1 , . . . , xn |) is maximized by choosing
or
= , as expected.
1 = 2 = = n = 1. This means that A = 1 ,
14. First we note that p(x|i , , i ) N (i , ). We have also lk = i if the state of
nature for xk was i .
(a) From Bayes Rule we can write
p(x1 , . . . , xn , l1 , . . . , ln |1 , . . . , c , ) = p(x1 , . . . , xn |1 , . . . , c , l1 , . . . , ln , )p(l1 , . . . , ln ).
Because the distribution of l1 , . . . , ln does not depend on 1 , . . . , c or , we
can write
p(x1 , . . . , xn |1 , . . . , c , , l1 , . . . , ln )
n
,
=
p(xk |1 , . . . , c , , lk )
=

k=1
n
,



1
t 1
(x
exp

(x

)
.
k
k
lk
lk
2
(2)d/2 ||1/2
k=1
1

The li are independent, and thus the probability density of the ls is a product,
p(l1 , . . . , ln )

n
,

p(lk ) =

k=1

n
,

p(lk ).

k=1

We combine the above equations and get

p(x1 , . . . , xn , l1 , . . . , ln |1 , . . . , c , )
n
/


P (lk )
n
1!
k=1
=
exp
(xk lk )t 1 (xk lk ) .
2
(2)nd/2 ||n/2
k=1

(b) We sum the result of part (a) over n samples to nd

n
!
k=1

(xk lk )t 1 (xk lk ) =

l !
!
i=1 lk =1

(xk i )t 1 (xk i ).

PROBLEM SOLUTIONS

91

The likelihood function for i is

n
/

l(i )

c
!
!
1
k=1
=
exp
(xk i )t 1 (xk i )
2 j=1
(2)nd/2 ||n/2
k:lk =j


1!
t 1
exp
(xk i ) (xk i ) .
2
P (i )

lk =i

Thus, the maximum-likelihood estimate of i is a function of xk restricted to

lk = i. But these xk s are from an independent identically distributed (i.i.d.)
sample from N (i , ). Thus, by the result for samples drawn from a single
normal population, it follows that
i =

1 !
xk ,
ni
lk =i

where ni = {xk : lk = i} =

1. Therefore, our solution is

lk =i


lk =i

xk

i = 

lk =i

We also know that if we have a maximum-likelihood estimate for , then the

maximum-likelihood estimate for is
!
= 1
k )
t.

(xk )(x
n
n

k=1

corresponds to the distribution of xk

But the maximum-likelihood estimate
lk . Thus, we have the estimate
is
!
= 1
lk )(xk
lk )t .

(xk
n
n

k=1

15. Consider the problem of learning the mean of a univariate normal distribution.
(a) From Eqs. 34 and 35 in the text, we have
n =

no2
2
mn +
o ,
2
2
2
no +
no + 2

and
n2

o2 2
,
no2 + 2

where the sample mean is

1!
xk .
n
n

mn =

k=1

92CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

Here o is formed by averaging no ctitious samples xk for k = no + 1, no +

2, . . . , 0. Thus we can write
0
1 !
xk ,
no

o =

k=no +1

and
n


n


xk

2 /o2 1
+
n + 2 /o2
2 /o2 + n no
k=1

k=1

xk

n + no

no 1
n + no no

0
!

0
!

xk

k=no +1

xk .

k=1no

We can use the fact that no = 2 /o2 to write

n =

1
n + no

n
!

xk .

k=no +1

Likewise, we have
n2

=
=

2 o2
no2 + 2
2
2
=
.
n + 2 /o2
n + no

(b) The result of part (a) can be interpreted as follows: For a suitable choice of
the prior density p() N (o , o2 ), maximum-likelihood inference on the full
sample on n + no observations coincides with Bayesian inference on the second
sample of n observations. Thus, by suitable choice of prior, Bayesian learning
can be interpreted as maximum-likelihood learning and here the suitable choice
of prior in Bayesian learning is
o

0
1 !
xk ,
no

2
.
no

k=no +1

o2

Here o is the sample mean of the rst no observations and o2 is the variance
based on those no observations.
16. We assume that A and B are non-singular matrices of the same order.
(a) Consider Eq. 44 in the text. We write
A(A + B)1 B

= A[(A + B)1 (B1 )1 ] = A[B1 (A + B)]1

= A[B1 A + I]1 = (A1 )1 (B1 A + I)1
= [(B1 A + I)A1 ]1 = (B1 + A1 )1 .

PROBLEM SOLUTIONS

93

We interchange the roles of A and B in this equation to get our desired answer:
B(A + B)1 A = (A1 + B1 )1 .
(b) Recall Eqs. 41 and 42 in the text:
1
n

= n1 + 1
o
= n1 n + 1
o o .

1
n n
We have solutions




1
1
1
1
n = o o + n + o +
o ,
n
n
n
and

n = o

1
1
1
.
o +
n
n

Taking the inverse on both sides of Eq. 41 in the text gives

1

n = n1 + 1
.
o
We use the result from part (a), letting A =
n
o

1
n

and B = o to get

1

1
1

+ o
=
n
n

1
1
= o o +
,
n

which proves Eqs. 41 and 42 in the text. We also compute the mean as
n

= n (n1 mn + 1
o o )
1
= n n mn + n 1
o o

1

1
1
1
1
1
1
= o o +
o 1
n mn + o +
o o
n
n
n
n

1

1
1
1
1
= o o +
mn + o +
o .
n
n
n

Section 3.5
17. The Bernoulli distribution is written
p(x|) =

d
,

ixi (1 i )1xi .

i=1

Let D be a set of n samples x1 , . . . , xn independently drawn according to p(x|).

94CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

(a) We denote s = (s1 , , sd )t as the sum of the n samples. If we denote xk =

n

(xk1 , , xkd )t for k = 1, . . . , n, then si =
xki , i = 1, . . . , d, and the likelihood
k=1

is

= P (x1 , . . . , xn |) =

P (D|)

n
,
k=1

P (xk |)


xk are indep.

d
n ,
,

ixki (1 i )1xki

k=1 i=1

n
d
n
,
xki
i k=1 (1 i ) k=1 (1xki )

i=1

d
,

isi (1 i )nsi .

i=1

(b) We assume an (unnormalized) uniform prior for , that is, p() = 1 for 0
i 1 for i = 1, , d, and have by Bayes Theorem
p(|D) =

p(D|)p()
.
p(D)

From part (a), we know that p(D|) =

d
/
i=1

isi (1 )nsi , and therefore the

probability density of obtaining data set D is


p(D)

p(D|)p()d =

 ,
d

isi (1 i )nsi d

i=1

1

isi (1 i )nsi d1 d2 dd

0 i=1

1 ,
d

d 1
,

isi (1 i )nsi di .

i=1 0

Now si =

n

k=1

xki takes values in the set {0, 1, . . . , n} for i = 1, . . . , d, and if we

use the identity

1
m (1 )n d =
0

m!n!
,
(m + n + 1)!

and substitute into the above equation, we get

d 
,

p(D) =

i=1 0

isi (1 i )nsi di =

d
,
si !(n si )!
i=1

(n + 1)!

PROBLEM SOLUTIONS

95

We consolidate these partial results and nd

p(|D)

p(D|)p()
p(D)
d
/
isi (1 i )nsi
i=1

d
/

i=1

si !(n si )!/(n + 1)!

d
,
(n + 1)! si
=
i (1 i )nsi .
s
!(n

s
)!
i
i=1 i

(c) We have d = 1, n = 1, and thus

p(1 |D) =

2!
2
1s1 (1 1 )ns1 =
s1 (1 1 )1s1 .
s1 !(n s1 )!
s1 !(1 s1 )! 1

Note that s1 takes the discrete values 0 and 1. Thus the densities are of the
form
: p(1 |D) = 2(1 1 )

s1 = 0
s1 = 1

: p(1 |D) = 21 ,

for 0 1 1, as shown in the gure.

p(1|D)
2

1
s 1=

1.5

s1 =

0.5

0
1

0.2

0.4

0.6

0.8

18. Consider how knowledge of an invariance can guide our choice of priors.
(a) We are given that s is actually the number of times that x = 1 in the
 rst n tests.
Consider the (n + 1)st test. If again x = 1, then there are n+1
s+1 permutations
of 0s and 1s in the (n + 1) tests, in which the number of 1s is (s + 1). Given the
assumption of invariance of exchangeablility (that is, all permutations have the
same chance to appear), the probability of each permutation is
1
Pinstance = n+1 .
s+1

Therefore, the probability of x = 1 after n tests is the product of two probabilities: one is the probability of having (s + 1) number of 1s, and the other is the
probability for a particular instance with (s + 1) number of 1s, that is,
p(s + 1)
Pr[xn+1 = 1|Dn ] = Pr[x1 + + xn = s + 1] Pinstance = n+1 .
s+1

96CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

The same analysis yields

p(s)
P (xn+1 = 0|Dn ) = n+1 .
s

Therefore, the ratio can be written as



p(s + 1)/ n+1
Pr[xn+1 = 1|Dn ]
s+1
 .

=
Pr[xn+1 = 0|Dn ]
p(s)/ n+1
s
(b) Given p(s)  p(s + 1) for large s, we have


n+1
p(s + 1)/ n+1
(n + 1)! (s + 1)!(n + 1 s 1)!
s+1
s
n+1
=
 n+1
s!(n
+ 1 s)!
(n + 1)!
p(s)/ s
s+1
=
=

(s + 1)!(n + 1 s 1)!
s!(s + 1)(n + 1 s 1)!
=
s!(n + 1 s)!
s!(n + 1 s 1)!(n + 1 s)
s+1
.
n+1s

We can see that for some n, the ratio will be small if s is small, and that the ratio
will be large if s is large. This implies that with n increasing, if x is unlikely
to be 1 in the rst n tests (i.e., small s), it would be unlikely to be 1 in the
(n + 1)st test, and thus s remains small. On the contrary, if there are more 1s
than 0s in the rst n tests (i.e., large s), it is very likely that in the (n + 1)st
test the x remains 1, and thus retains s large.
(c) If p() U (0, 1) for some n, then p() is a constant, which we call c. Thus
1
p(s)

s (1 )ns p()d

=
0

  1
n
= c
s (1 )ns d
s
0

 
 1
n
s+1 (1 )ns+1 
= c



s
s+1 ns+1
0

0,

which of course does not depend upon s.

19. Consider MAP estimators, that is, ones that maximize l()p().
(a) In this problem, the parameter needed to be estimated is . Given the training
data, we have
l()p() = ln[p(D|)p()]
where for the Gaussian
ln[p(D|)] =

1
ln

n
,

k=1

2
p(xk |)

PROBLEM SOLUTIONS

97

n
!

ln[p(xk |)]

k=1

n
 !
n 
1
d
= ln (2) ||
(xk )t 1 (xk )
2
2
k=1

and
p() =



1
1
t 1
exp

m
)
.
(

m
0
0
o
2
(2)d/2 |0 |1/2

The MAP estimator for the mean is then



n
 !
n 
1
d
t 1
= arg max

ln (2) ||
(xk ) (xk )

2
2
k=1


 
1
1
t 1

exp ( m0 ) o ( m0 )
.
2
(2)d/2 |0 |1/2
(b) After the linear transform governed by the matrix A, we have
 = E[x ] = E[Ax] = AE[x] = A,
and


= E[(x  )(x  )t ]
= E[(Ax A )(Ax A )t ]
= E[A(x  )(x  )t At ]
= AE[(x  )(x  )t ]At
= AAt .

Thus we have the log-likelihood

1 n
2
,
 


ln[p(D | )] = ln
p(xk | )
k=1
n
,

1
=
=

ln
n
!

2
p(Axk |A)

k=1

ln[p(Axk |A)]

k=1

n
 !
n 
1
= ln (2)d |AAt |
(Axk A)t (AAt )1 (Axk A)
2
2
k=1

n
 !
n 
1
= ln (2)d |AAt |
((x )t At )((A1 )t 1 A1 )(A(xk ))
2
2
k=1

n
 !
n 
1
= ln (2)d |AAt |
(xk )t (At (A1 )t )1 (A1 A)(xk )
2
2


n 
= ln (2)d |AAt |
2

k=1
n
!
k=1

1
(xk )t 1 (xk ).
2

98CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

Likewise we have that the density of  is a Gaussian of the form



1
1 

t 1

p( ) =
exp ( m0 ) 0 ( m0 )
2
(2)d/2 |0 |1/2


1
1
t
t 1
=
exp (A Am0 ) (A0 A ) (A Am0 )
2
(2)d/2 |0 |1/2


1
1
t t
1 t 1 1
=
exp ( m0 ) A (A ) 0 A A( m0 )
2
(2)d/2 |0 |1/2


1
1
t 1
=
exp ( m0 ) 0 ( m0 ) .
2
(2)d/2 |0 |1/2
Thus the new MAP estimator is


n 

= arg max ln (2)d |AAt |

2


 
n
!1
1
1
t 1
t 1

exp ( m0 ) 0 ( m0 )
.
(xk ) (xk )
2
2
(2)d/2 |0 |1/2
k=1
and see that the two equations are the same, up to a constant.
We compare
Therefore the estimator gives the appropriate estimate for the transformed mean
.

20. Consider the problem of noninformative priors.

(a) Assume that
has a uniform distribution, that is, p(
) = c, where c is a
constant. We make the correspondence
= ln and thus = exp[
] = f (
),
for some f (). Since p(
) = c, we have
p()

df 1 ()
d
dln()
= p(
)
d
c
=
.

= p(f 1 ())

For the case c = 1, we have p() = 1/.

(b) The noninformative priors here are very simple,
p(0 )

p( )

1
1
=
2 0
2
1/ .

21. Note that 0 p(|D) 1. In order to converge while n , it must be true

that
lim p(|D) = lim

p(xk |)p(|Dn1 )
= lim p(|Dn1 ).
p(xk |)p(|Dn1 )d n

Note that lim p(|Dn ) p(). We assume lim p(|D) p() = 0 and lim xn
n
n
n
x . Then the above equation implies
p(x |)

lim p(xn |)

PROBLEM SOLUTIONS

99

lim
p(xn |)p()d

=
lim p(xn |)p()d
n

=
p(x |)p()d
=

= p(x ).
In summary, we have the conditions:
lim p(|Dn ) p() = 0
n

lim xn x
n

p(x |) = p(x ), that is, p(x |) is independent of .

22. Consider the Gibbs algorithm.
(a) Note that p(x|2 , ) = 0 for |x | < 1 and this implies x 1 < < x + 1 and
p() = 0 for 0 2. We have the following cases:
x 1 > 2 and thus x > 3 or x + 1 < 0 and thus x < 1.
0 x 1 2 and thus 1 x 3. In that case,
2
p(x|2 ) =

2
p(x|2 , )p()d =

x1

x1

11
3x
d =
.
22
4

0 x + 1 2 and thus 1 x 1. In that case,

x+1


p(x|2 ) =

x+1


p(x|2 , )p()d =
0

Therefore, the class-conditional density is:

(x + 1)/4
p(x|2 ) =
(3 x)/4

11
x+1
d =
.
22
4

x < 1
1 x 1
1x3
x > 3.

(b) The decision point is at

x = arg[p(x|1 )p(1 ) = p(x|2 )p(2 )].
x

(c) We let P P (1 ) and note the normalization condition P (1 ) + P (2 ) = 1.

There are two cases:
(1 + x)P (x + 1)/4 (1 P ) and thus 1/5 P 1, as shown in the
gure.
In this case, (1 x )P = (x + 1)/4 (1 P ) and this implies x =
(5P 1)/(3P + 1), and the error is

x
P (error) =

1
p(x|2 )P (2 )dx +

p(x|1 )P (1 )dx =
x

24P 3 + 8P 2
.
(3P + 1)2

100CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

p(x|i)P(i)
0.5
0.4

1
0.3

2
0.2
0.1

-1

error

0 P 1/5, as shown in the gure.

The error is
1
P (error) = p(x|1 )P (1 )dx = P.
1

(d) Again, we are considering p(x|1 )P (1 ) and p(x|2 )P (2 ), but this time, we
rst use a single value of as the true one, that is, p() = 1. There are two
cases for dierent relationships between P and .
p(x|1 )P (1 ) p(x|2 , )P (2 ) which implies P 1/2 (1 P ) or 0
P 1/3. This case can be further divided into two sub-cases:
subcase 1: 0 1
The error is


1
1
2
P (error) =
(1 (1)) P
d = P P/6 = 5P/6.
2
2
0

subcase 2: 1 < 2
The error is
2
P2 (error) =
1

(2 )2
P d = P/6
2

and the total error for this case is

P
5P
+
= P.
6
6
Other case: 1/3 P 1. This case can be further subdivided into three
subcases.
subcase 1: 0 (1 P )/(2P ) as shown in the gure.
The error is
(1P


)/(2P)
1P 9
1
P1 (error) =
2
4
2
2P
0


 
1
1P
1 3P

P
+1
d
2
2
2P
PGibbs (error) = P1 (error) + P2 (error) =

(1 P )2 (31P 7)
.
48P 2

PROBLEM SOLUTIONS

101
p(x|i)P(i)
0.5
0.4

2
0.3
0.2

0.1

error
-1

subcase 2: (1 P )/(2P ) (5P 1)/(2P ), as shown in the gure.

The error is
(5P 1)/(2P
)


P2 (error)

1P
4

9
1
2
2
2P


d

(1P )/(2P )

(5P 1)(1 P )(3P 1)

.
8P 2

subcase 3: (5P 1)/(2P ) 2, as shown in the gure.

The error is
2
P3 (error)

P
(2 )2 d
2

=
(5P 1)/(2P )

P
6

1P
2P

3
.

Thus the probability of error under Gibbs sampling is

PGibbs (error)

= P1 (error) + P2 (error) + P3 (error)

5P 2 6P + 1
=
4P

(e) It can be easily conrmed that PGibbs (error) < 2PBayes (error).
Section 3.6
23. Let s be a sucient statistic for which p(|s, D) = p(|s); we assume p(|s) = 0.
In that case, we can write Bayes law
p(D|s, ) =

p(|s, D)p(D|s)
p(|s)

as
p(D|s, )

=
=

p(D, s, )
p(s, )
p(|s, D)P (D, s)
p(|s)p(s)

102CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

p(x|i)P(i)
0.5
0.4

2
0.3
0.2

1
0.1

error
-1

p(|s, D)p(D|s)p(s)
p(|s)p(s)
p(|s, D)p(D|s)
.
p(|s)

=
=

Note that the probability density of the parameter is fully specied by the sucient
statistic; the data gives no further information, and this implies
p(|s, D) = p(|s).
Since p(|s) = 0, we can write
p(|s, D)p(D|s)
p(|s)
p(|s)p(D|s)
=
p(|s)
= p(D|s),

p(D|s, ) =

which does not involve . Thus, p(D|s, ) is indeed independent of .

24. To obtain the maximum-likelihood estimate, we must maximize the likelihood
function p(D|) = p(x1 , . . . , xn |) with respect to . However, by the Factorization
Theorem (Theorem 3.1) in the text, we have
p(D|) = g(s, )h(D),
where s is a sucient statistic for . Thus, if we maximize g(s, ) or equivalently
[g(s, )]1/n , we will have the maximum-likelihoood solution we seek.
For the Rayleigh distribution, we have from Table 3.1 in the text,
[g(s, )]1/n = es
for > 0, where
1! 2
xk .
n
n

s=

k=1

Then, we take the derivative with respect to and nd

[g(s, )]1/n = es ses .

PROBLEM SOLUTIONS

103
p(x|i)P(i)
2

0.5
0.4
0.3

1
0.2
0.1

error
-1

We set this to 0 and solve to get

es

s ,
= se

which gives the maximum-likelihood solution,

21
1 n
!
1
1
2
= =
xk
.
s
n
k=1

We next evaluate the second derivative at this value of to see if the solution represents
a maximum, a minimum, or possibly an inection point:




2 [g(s, )]1/n 
= ses ses + s2 es 

= e

(s 2s) = se

< 0.

Thus indeed gives a maximum (and not a minimum or an inection point).

25. The maximum-likelihood solution is obtained by maximizing [g(s, )]1/n . From
Table 3.1 in the text, we have for a Maxwell distribution
[g(s, )]1/n
where s =

1
n

n

k=1

= 3/2 es

x2k . The derivative is

[g(s, )]1/n =

3 1/2 s
s3/2 es .
e
2

We set this to zero to obtain

3 1/2 s
e
= s3/2 es ,
2
and thus the maximum-likelihood solution is
21
1 n
3/2
3 1! 2

=
xk
.
=
s
2 n
k=1

We next evaluate the second derivative at this value of to see if the solution represents
a maximum, a minimum, or possibly an inection point:


3 1 1/2 s 3 1/2 s 3 1/2 s


2 [g(s, )]1/n 
=
se
se
+ s2 3/2 es 
e
22
2
2
=
=

104CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

p(x|i)P(i)
0.5
0.4

0.3

2
0.2
0.1

-1

error
0

3 1/2 s

e
3s1/2 es + s2 3/2 es
4


3 9 1/2
3
3
3 +

= e3/2
= 1/2 e3/2 < 0.
4
2 4
2
=

Thus indeed gives a maximum (and not a minimum or an inection point).

26. We ne the maximum-likelihood solution by maximizing [g(s, )]1/n . In this case,
we have
[g(s, )]1/n =

d
,

isi

i=1

and
1!
xk .
n
n

s = (s1 , . . . , sd )t =

k=1

Our goal is to maximize [g(s, )]1/n with respect to over the set 0 < i < 1, i =
d

1, . . . , d subject to the constraint
i = 1. We set up the objective function
i=1

l(, |s)

=
=

1/n

ln [g(s, )]
d
!

si ln i +

i=1

1 d
!

1 d
!

i=1

l
si
=
+
i
i

for i = 1, . . . , d. We set this derivative to zero and get

si
= ,
i
which yields our intermediate solution:
si
i = .

i 1
2

i 1 .

i=1

The derivative is thus

( l)i =

PROBLEM SOLUTIONS

105
p(x|i)P(i)
0.5
0.4

0.3

2
0.2
0.1

-1

error
0

We impose the normalization condition,

d

i = 1, to nd , which leads to our nal

i=1

solution
i =

sj
d

i=1

si

for j = 1, . . . , d.
27. Consider the notion that suciency is an integral concept.
(a) Using Eq. 51 in the text, we have
p(D|)

=
=
=
=
=

n
,

p(x|)

k=1
n
,



1
1
2

exp 2 (xk )
2
2
k=1



n
n

1
1 ! 2
2 2

exp
xk 2xk +
2 2
2
k=1
1 n
2


n
n
n
!
!
!
1
1
2
2

exp 2
xk 2
xk +

2
2
k=1
k=1
k=1



n

1
1 
2

exp 2 ns2 2ns1 + n

.
2
2

Now let g(s, ) be a Gaussian of the form




n
1
1
g(s, ) =
exp 2 (ns2 2ns1 + n2 )
2
2
and h(D) = 1. Then we can write
p(D|) = g(s, )h(D).
According to Theorem 3.1 in the text, the statistic s is indeed sucient for .
(b) We apply the result from part (a) to the case of g(s1 , , 2 ), that is,



n
1
1
2
g(s1 , , 2 ) =
exp
(2ns
+
n
)
1
2 2
2

106CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

p(x|i)P(i)
0.5
0.4

0.3

2
0.2
0.1

error
-1

and


1
h(D, ) = exp
ns2 .
2 2


In the general case, the g(s1 , , 2 ) and h(D, 2 ) are dependent on 2 , that is,
there is o way to have p(D|, 2 ) = g(s1 , )h(D). Thus according to Theorem 3.1
in the text, s1 is not sucient for . However, if 2 is known, then g(s1 , , 2 ) =
g(s1 , ) and h(D, 2 ) = h(D). Then we indeed have p(D|) = g(s1 , )h(D), and
thus s1 is sucient for .
(c) As in part (b), we let

2

g(s2 , , ) =

n


1
2
exp
(ns2 2ns1 + n )
2 2

and h(D) = 1. In the general case, the g(s2 , , 2 ) is dependent upon , that
is, there is no way to have p(D|, 2 ) = g(s2 , 2 )h(D). Thus according to
Theorem 3.1 in the text, s2 is not sucient for 2 . However, if is known, then
g(s2 , , 2 ) = g(s2 , 2 ). Then we have p(D| 2 ) = g(s2 , 2 )h(D), and thus s2 is
sucient for 2 .
28. We are to suppose that s is a statistic for which p(|x, D) = p(|s), that is, that
s does not depend upon the data set D.
(a) Let s be a sucient statistic for which p(|s, D) = p(|s); we assume p(|s) = 0.
In that case, we can write Bayes law
p(D|s, ) =

p(|s, D)p(D|s)
p(|s)

as
p(D|s, )

=
=
=
=

p(D, s, )
p(s, )
p(|s, D)P (D, s)
p(|s)p(s)
p(|s, D)p(D|s)p(s)
p(|s)p(s)
p(|s, D)p(D|s)
.
p(|s)

PROBLEM SOLUTIONS

107

Note that the probability density of the parameter is fully specied by the
sucient statistic; the data gives no further information, and this implies
p(|s, D) = p(|s).
Since p(|s) = 0, we can write
p(D|s, )

p(|s, D)p(D|s)
p(|s)
p(|s)p(D|s)
=
p(|s)
= p(D|s),
=

which does not involve . Thus, p(D|s, ) is indeed independent of .

(b) Assume the variable x comes from a uniform distribution, p(x) U (, 1). Let
n

s = n1
xk be a statistic of a sample data set D of x, that is, s is the sample
k=1

mean of D. Now assume that we estimate = 0 = 0, but from a particular

sample data set D0 we have s0 = 5. Note that we estimate p(x) U (0, 1), we
have p(0 |s0 , D) = p(0 |s0 ) = 0, that is, it is impossible that given the data set
D0 whose mean is s0 = 5 that the distribution is p(x) U (0, 1). Therefore, if
given = 0 = 0 and s0 = 5, there does not exist a data set D0 whose mean
is s0 = 5 but whose elements are distributed according to p(x) U (0, 1). Said
another way, we have p(D0 |s0 , 0 ) = 0 for these particular values of s0 and 0 .
However, p(D0 |s0 ) is not necessarily zero. It is easy to set up a data set whose
mean is s0 , giving p(D0 |s0 , 0 ) = p(D0 |s0 ). Therefore, p(|s) = 0, as required
for the proof.
29. Recall the Cauchy distribution,
p(x) =

1
1

 xa 2 .
b 1 +
b

(a) We integrate p(x) over all x, and have



1
p(x)dx =
b

1+

1
 xa 2 dx.
b

We substitute y = (x a)/b and dx = b dy and nd

1
b

1
1
 xa 2 dx =

1+
b


1
1 " " ##
1
dy = tan1 [y] =

= 1.
2
1+y

2
2

We nd, then, that indeed p(x) is normalized.

(b) According to the denition of the mean, we have

=

xp(x)dx =

1
x

 dx.
b 1 + xa 2
b

108CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

As in part (a), we let y = (x a)/b and dx = b dy. Then the mean is


1
by + a
b dy
b
1 + y2

y
1

dy
+a
b
dy
b
.

1 + y2
1 + y2




0

Note that f (y) = y/(1 + y 2 ) is an odd function, and thus the rst integral
vanishes. We substitute our result from part (a) and nd
a
=

1
b dy = a.
1 + y2

This conclusion is satisfying, since when x = a, the distribution has its maximum; moreover, the distribution dies symmetrically around this peak.
According to the denition of the standard deviation,


x2 p(x)dx 2

1
b


x2

1+

1
2
 xa 2 dx .
b

We substitute y = (x a)/b and dx = bdy, and have

 2 2
b y + 2aby + a2
1
b dy 2
b
1 + y2




y
1
1
b2 dy + 2ab
dy + (a2 b2 )
dy 2

1 + y2
1 + y2


1
=
+ 2ab 0 + (a2 b2 ) 2

= .

This result is again as we would expect, since according to the above calculation,
E[(x )2 ] = .
(c) Since we know = ( )t = (a )t , for any dened s = (s1 , . . . , sn ), si < ,
it is impossible to have p(D|s, (a )t ) = p(D|s). Thus the Cauchy distribution
has no sucient statistics for the mean and standard deviation.
30. We consider the univariate case. Let and 2 denote the
mean and variance of
n
the Gaussian, Cauchy and binomial distributions and
= 1/n i=1 xi be the sample
mean of the data points xi , for i = 1, . . . , n.

PROBLEM SOLUTIONS

109

(a) According to Eq. 21 in the text, we have


n
1 !
= E
(xi
)2
n 1 i=1

 n
!
1
2
=
E
((xi ) (
))
n1
i=1

 n
!

1
=
(xi )2 2(xi )(
E
) + (
)2
n1
i=1


n2

n

1 !
)] + E[(
)2 ] .
E[(xi )2 ] 2E[(xi )(
n 1 i=1

Note that

n
!
1
E[(xi )(
)] = E (xi )
xj
n j=1

n
!

x
1
i
= E (xi )
xk
+
n
n
k=1; k=i



n
!
1
1
= E
(xi )2 + E (xi )
xk (n 1)
n
n
k=1; k=i

1 2
+ 0 = 2 /n.
n

Similarly, we have the variance,

E[(
)2 ] = 2 /n,
and furthermore
n2

=
=


n 
1 !
2
1
2 2 + 2
n 1 i=1
n
n
n1 2
= 2 .
n1

Thus indeed the estimator in Eq. 21 in the text is unbiased.

(b) The result in part (a) applies to a Cauchy distribution.
(c) The result in part (a) applies to a Binomial distribution.
(d) From Eq. 20 in the text, we have
lim

that is, asymptotically unbiased.

n1 2
= 2 ,
n

110CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

Section 3.7
31. We assume that a and b are positive constants and n a variable. Recall the
denition on page 633 in the text that
O(g(x))

= {f (x) : There exist positive constants c and x0 such that

0 f (x) cg(x) for all x x0 }.

(a) Is an+1 = O(an )? Yes. If let c = a and x0 = 1, then c an = an+1 an+1 for
x 1.
(b) Is abn = O(an )? No. No matter what constant c we choose, we can always nd
a value n for which abn > c an .
(c) Is an+b = O(an )? Yes. If choose c = ab and x0 = 1, then c an = an+b an+b
for x > 1.
(d) Clearly f (n) = O(f (n)). To prove this, we let c = 1 and x0 = 0. Then of course
f (x) f (x) for x 0. (In fact, we have f (n) = f (n).)
n1
32. We seek to evaluate f (x) = i=0 ai xi at a point x where the n coecients ai
are given.
(a) A straightforward (n2 ) algorithm is:
Algorithm 0 (Basic polynomial evaluation)
1
2
3
4
5
6
7
8
9
10
11
12

begin initialize x, ai for i = 1, . . . , n 1

f a0
i0
for i i + 1
x0 1
j0
for j j + 1
xj xj1 x
until j = i
f f + ai xi
until i = n 1
end

(b) A (n) algorithm that takes advantage of Horners rule is:

Algorithm 0 (Horner evaluation)
1
2
3
4
5
6
7

begin initialize x, ai for i = 1, . . . , n 1

f an1
in
for i i 1
f f x + ai1
until i = 1
end

33. The complexities are as follows:

PROBLEM SOLUTIONS

111

(a) O(N )
(b) O(N )
(c) O(J(I + K))
34. We assume that the uniprocessor can perform one operation per nanosecond
(109 second). There are 3600 seconds in an hour, 24 3600 = 86, 400 seconds in
a day, and 31, 557, 600 seconds in a year. Given one operation per nanosecond, the
total number of basic operations in the periods are: 109 in one second; 3.6 1012 in
one hour; 8.64 1013 in one day; 3.156 1016 in one year. Thus, for the table below,
we nd n such that f (n) is equal to these total number of operations. For instance,
for the 1 day entry in the nlog2 n row, we solve for n such that nlog2 n = 8.64 1013 .
f (n)
operations:

1 sec
109

1 hour
3.6 1012
12

log2 n

n
n
nlog2 n *
n2

23.610

101012
(3.61012 )2

1.31025
12

210

4.610301029995
9 2
18

n3

(10 ) = 10
109
3.9 107

9
10  3.16 104

3
109 = 103

2n
en
n! **

log2 109 = 29.90

ln109 = 10.72
13.1

3.6 10
9.86
1010

3.61012

1.9106

3
3.61012

1.53104
log2 (3.61012 )

41.71
ln(3.61012 )

28.91

16.1

1 day
8.64 1013
13

28.6410

101013
(8.641013 )2

7.461027
13

8.64 10
2.11
1012

8.641013

9.3106

3
8.641013

4.42104
log2 (8.641013 )

46.30
ln(8.641013 )

32.09

17.8

1 year
3.156 1016
16

23.15610

101016
(3.1561016 )2

9.961032
16

3.156 10
6.41
1014

3.1561016

1.78108

3
3.1561016

3.16105
log2 (3.1561016 )

54.81
ln(3.1561016 )

38.0

19.3

* For entries in this row, we solve numerically for n such that, for instance, nlog2 n =
3.9 107 , and so on.
** For entries in this row, we use Stirlings approximation n!  (n/e)n for large n
and solve numerically.
35. Recall rst the sample mean and sample covariance based on n samples:
1!
xk
n
n

mn

k=1

1 !
(xk mn )(xk mn )t .
n1
n

Cn

k=1

(a) The computational complexity for mn is O(dn), and for Cn is O(dn2 ).

(b) We can express the sample mean for n + 1 observations as

 n
n+1
!
1 !
1
mn+1 =
xk =
xk + xn+1
n+1
n+1
k=1

k=1

1
=
[nmn + xn+1 ]
n+1
1
1
= mn
mn +
xn+1
n+1
n+1
1
= mn +
(xn+1 mn ).
n+1

112CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

For the sample covariance, we have

Cn+1

=
=

n+1
1!
(xk mn+1 )(xk mn+1 )t
n
k=1

 n
1 !
t
t
(xk mn+1 )(xk mn+1 ) + (xn+1 mn+1 )(xn+1 mn+1 )
n
k=1

n
1$!

k=1

!
1
(xk mn )t
(xn+1 mn )
(n + 1)
n

(xk mn )(xk mn )t
1

k=1

%
1
1
(xk mn ) (xn+1 mn )t +
(xn+1 mn )(xn+1 mn )t
2
n+1
(n + 1)
k=1
k=1


t
1
1
1
+
(xn+1 mn )
(xn+1 mn )
(xn+1 mn )
(xn+1 mn )
n
n+1
n+1


1
n
=
(xn+1 mn )(xn+1 mn )t
(n 1)Cn +
n
(n + 1)2
2
1
2
1
n
t
+
(xn+1 mn )(xn+1 mn )
n
n+1


n1
1
n
=
(xn+1 mn )(xn+1 bf mn )t
Cn +
+
n
(n + 1)2
(n + 1)2
n1
1
=
Cn +
(xn+1 mn )(xn+1 mn )t .
n
n+1

n
!

n
!

(c) To compute mn from scratch, requires O(dn) operations. To update mn according to

mn+1 = mn +

1
(xn+1 mn )
n+1

requires O(d) operations, one for each component.

To update Cn according to
Cn+1 =

n1
1
Cn +
(xn+1 mn )(xn+1 mn )t
n
(n + 1)

given Cn , mn and xn requires O(d) operations for computing xn+1 mn ; and

O(d2 ) operations for computing (xn+1 mn )(xn+1 mn )t . Thus, Cn+1 is
computed in O(d2 ) operations, given Cn , mn and xn . If we must compute
Cn+1 from scratch, then we need O(dn2 ) operations.
(d) The recursive method is on-line, and the classier can be used as the data
is coming in. One possible advantage of the non-recursive method is that it
might avoid compounding roundo errors in the successive additions of new
information from samples, and hence be more accurate.
36. We seek a recursive method for computing C1
n+1 .
(a) We rst seek to show
(A + xxt )1 = A1

A1 xxt A1
.
1 + xt A1 x

PROBLEM SOLUTIONS

113

We prove this most simply by verifying the inverse as:



A1 xxt A1
t
1
(A + xx ) A
1 + xt A1 x
AA1 xxt A1
xxt A1 xxt A1
t 1
A

= AA1
+
xx
1 + xt A1 x
1 + xt A1 x
t 1
xx A
xxt A1
t 1
=I
xt A1 x
t 1 + xx A
t 1
1+x
1+x
 A
 A
 x
 x
a scalar
t

= I + x xA

a scalar

1
xt A1 x
1

t
1
1 + x A x 1 + xt A1 x

= I + xt xA1 0
= I.
Note the left-hand-side and the right-hand-side of this equation and see
(A + xt x)1

= A1

A1 xxt A1
.
1 + xt A1 x

(b) We apply the result in part (a) to the result discussed in Problem 35 and see
that
n1
1
Cn +
(xn+1 mn )(xn+1 mn )t
n
n+1


n1
n
=
Cn + 2
(xn+1 mn )(xn+1 mn )t
n
n 1
n1
=
(A + xxt ).
n
3
We set A = Cn and x = n2n1 (xn+1 mn ) and substitute to get
Cn+1

C1
n+1


=
=
=

1
n1
n
t
=
(A + xx )
(A + xxt )1
n
n1


n
A1 xxt A1
1
A
n+1
1 + xt A1 x
3
3

n
n
1
t 1
(x

m
)
C
2
n+1
n
n
n 1
n2 1 (xn+1 mn ) Cn
n 1

3
3
Cn
n
n+1
1 + n2n1 (xn+1 mn )t C1
(x

m
)
n
2
n+1
n
n 1


t 1
n
C1
n (xn+1 mn )(xn+1 mn ) Cn
1
,
Cn n2 1
t 1
n1
n + (xn+1 mn ) Cn (xn+1 mn )

where we used our result from part (a).

(c) We determine the computational complexity as follows. Consider the formula
for C1
n+1 :


t
n
uu
C1
,
C1
n n2 1
n+1 =
t
n1
n + u (xn+1 mn )

114CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

2
1
where u = C1
n (xn+1 mn ) is of O(d ) complexity, given that Cn , xn+1 and
1
1
mn are known. Hence, clearly Cn can be computed from Cn1 in O(d2 ) operations, as uut , ut (xn+1 mn ) is computed in O(d2 ) operations. The complexity
2
associated with determining C1
n is O(nd ).

37. We assume the symmetric non-negative covariance matrix is of otherwise general

form:

11 12 1n
21 22 2n

= .
..
.. .
..
..
.
.
.
n1

nn

n2

To employ shrinkage of an assumed common covariance toward the identity matrix,

then Eq. 77 requires
() = (1 ) + I = I,
and this implies (1 )ii + 1 = 1, and thus
ii =

1
=1
1

for all 0 < < 1. Therefore, we must rst normalize the data to have unit variance.
Section 3.8
38. Note that in this problem our densities need not be normal.
(a) Here we have the criterion function
J1 (w) =

(1 2 )2
.
12 + 22

We make use of the following facts for i = 1, 2:

y
i
i2

= wt x
1 !
1 ! t
=
y=
w x = wt i
ni
ni
yYi
xDi


!
!
2
t
t
=
(y i ) = w
(x i )(x i ) w
xDi

yYi

(x i )(x i )t .

xDi

We dene the within- and between-scatter matrices to be

SW

= 1 + 2

SB

(1 2 )(1 2 )t .

Then we can write

12 + 22

= wt SW w

(1 2 )2

= wt SB w.

PROBLEM SOLUTIONS

115

The criterion function can be written as

J1 (w) =

wt SB w
.
wt SW w

For the same reason Eq. 103 in the text is maximized, we have that J1 (w)
is maximized at wS1
W (1 2 ). In sum, that J1 (w) is maximized at w =
(1 + 2 )1 (1 2 ).
(b) Consider the criterion function
J2 (w) =

(1 2 )2
.
P (1 )12 + P (2 )22

Except for letting SW = P (1 )1 + P (2 )2 , we retain all the notations in

part (a). Then we write the criterion function as a Rayleigh quotient
J2 (w) =

wt SB w
.
wt SW w

For the same reason Eq. 103 is maximized, we have that J2 (w) is maximized at
1

w = (P (1 )1 + P (2 )2 )

(1 2 ).

(c) Equation 96 of the text is more closely related to the criterion function in part (a)
above. In Eq. 96 in the text, we let m
i = i , and s2i = i2 and the statistical
meanings are unchanged. Then we see the exact correspondence between J(w)
and J1 (w).
39. The expression for the criterion function
J1 =

1 ! !
(yi yj )2
n 1 n2
yi Y1 yj Y2

clearly measures the total within-group scatter.

(a) We can rewrite J1 by expanding
J1

1 ! !
[(yi m1 ) (yj m2 ) + (m1 m2 )]2
n1 n2

1 ! ! 
(yi m1 )2 + (yj m2 )2 + (m1 m2 )2
n1 n2

yi Y1 yj Y2

yi Y1 yj Y2


+2(yi m1 )(yj m2 ) + 2(yi m1 )(m1 m2 ) + 2(yj m2 )(m1 m2 )
1 ! !
1 ! !
(yi m1 )2 +
(yj m2 )2 + (m1 m2 )2
n1 n2
n1 n 2
yi Y1 yj Y2

yi Y1 yj Y2

1 ! !
1 ! !
+
2(yi m1 )(yj m2 ) +
2(yi m1 )(m1 m2 )
n1 n2
n1 n2
yi Y1 yj Y2

1 ! !
+
2(yj m2 )(m1 m2 )
n 1 n2
yi Y1 yj Y2

1 2
1 2
s +
s + (m1 m2 )2 ,
n1 1 n2 2

yi Y1 yj Y2

116CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

where
m1

yi

yi Y1

m2

yj

yj Y2

s21

(yi m1 )2

yi Y1

s22

(yj m2 )2 .

yj Y2

(b) The prior probability of class one is P (1 ) = 1/n1 , and likewise the prior probability of class two is P (2 ) = 1/n2 . Thus the total within-class scatter is
J2 = P (1 )s21 + P (2 )s22 =

1 2
1 2
s +
s .
n1 1 n2 2

(c) We write the criterion function as

J(w) =

(m1 m2 )2 +
1 2
n1 s1

1 2
1 2
n1 s1 + n2 s2
1 2
n2 s2

(m1 m2 )2
1 2
1 2 + 1,
n1 s1 + n2 s2

then J(w) = J1 /J2 . Thus optimizing J1 subject to the constraint J2 = 1 is

equivalent to optimizing J(w), except that the optimal solution of the latter is
not unique. We let y = wt x, and this implies
(m1 m2 )2
J2

= wt (m1 m2 )(m1 m2 )t w


1
1 2
1 2
1
t
=
s +
s =w
S1 +
S2 w
n 1 1 n2 2
n1
n2

where the mean and scatter matrices are

1 !
mi =
x
ni
xDi
!
Si =
(x mi )(x mi )t ,
xDi

respectively. Thus the criterion function is

J(w) =

wt (m1 m2 )(m1 m2 )t w
$
%
+ 1.
wt n11 S1 + n12 S2 w

For the same reason Eq. 103 in the text is optimized, we have that J(w) is
maximized at

1
1
1
w=
S1 +
S2
(m1 m2 ).
n1
n2
We need to guarantee J2 = 1, and this requires


1
1
t
w
S1 +
S2 w = 1,
n1
n2

PROBLEM SOLUTIONS

117

which in turn requires


(m1 m2 )
2

1
1
S1 +
S2
n1
n2

1
(m1 m2 ) = 1.

This equation implies that the parameter has the value


1/2

1
1
1
t
= (m1 m2 )
S1 +
S2
(m1 m2 )
.
n1
n2
40. In the below, W is a d-by-n matrix whose columns correspond to n distinct
eigenvectors.
(a) Suppose that the set {ei } are normalized eigenvectors, then we have
eti SB ej
eti SW ej

= i ij
= i ij ,

where ij is the Kronecker symbol. We denote the matrix consisting of eigenvectors as W = [e1 e2 . . . en ]. Then we can write the within scatter matrix
in the new representation as
t
e1
et2

W = Wt SW W =
S
.. SW (e1 e2 en )
.
etn
t

e1 SW e1 et1 SW en

..
..
..
=
= I.
.
.
.
etn SW e1

etn SW en

Likewise we can write the between scatter matrix in the new representation as
t
e1
et2

B = Wt SB W =
S
.. SB (e1 e2 en )
.

etn

et1 SB e1

..
.
etn SB e1

..
..
=
.
.

etn SB en
et1 SB en

1
0
..
.

0
2
..
.

..
.

0
0
..
.

W is an n-by-n matrix and S

B is a diagonal matrix whose elements are
Thus S
the corresponding eigenvalues.
(b) Here we have the determinant of the transformed between-scatter matrix is just
the product of the eigenvalues, that is,
B | = 1 2 n ,
|S
W | = 1. Then the criterion function is simply J = 1 2 n .
and |S

118CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

(c) We make the following denitions:

t
W
5

= QDWt
t SW W
= QDWt SW WDQt .
= W

5 
W  = |D|2 and
Then we have S
5
B = W
= QDWt SB WDQt = QDS
t SB W
B DQt ,
S
5 
B  = |D|2 1 2 n . This implies that the criterion function obeys
then S
5 
B
S
J=
,
5
W 
S
and thus J is invariant to this transformation.
41. Our two Gaussian distributions are p(x|i ) N (i , ) for i = 1, 2. We denote
i and the distributions
the samples after projection as D
i) =
p(y|
i =
and

i 

for i = 1, 2. The log-likelihood ratio is


 n
/

p(y
|

)
ln
k 1
1)

lnp(D|

= k=1
n
2)

/
lnp(D|
2)
ln
p(yk |
n


k=1
n


$
ln
$
ln

k=1

c1 +
=

1
s2 )],
exp[(y
)2 /(2
2
s

c1 +

k=1

1 exp
2
s
1 exp
2
s


yk D1

yk D1

c1 +

1
2

c1 +

1
2

c1 +

1
2

c1 +

1
2

+
+

(yk
1 )2
2
s2
(yk
2 )2
2
s2

(yk
1 )2
2
s2

(yk
2 )2
2
s2

yk D2

yk D2

1
2
s2

1
2
s2

c1 + 1 +
c1 + 1 +

%%

k=1
%% = 
n

yk D2

yk D2

(yk
1 )2
2
s2

 

2
yk D

1
yk D

1
2
2
s2 n2 (
1
1
2
s2 n1 (

$
ln
$
ln

k=1

(yk
2 )2
2
s2

n


1
2
s
1
2
s

%
+
%
+

n

k=1
n

k=1

(yk
1 )2
2
s2
(yk
2 )2
2
s2

(yk
1 )2
2
s2
(yk
2 )2
2
s2

c1 +

1
2

c1 +

1
2


yk D2

yk D1

(yk
2 )+(
2
1 ))2
2
s2
(yk
2 )+(
2
1 ))2
2
s2


(yk
2 ) + (
2
1 )2 + 2(yk
2 )(
2
1 )
2

((yk
1 )2 + (
1
2 )2 + 2(yk
1 )(
1
2 ))

1 )2
c + n2 J(w)
=
.
2
c + n1 J(w)

2 )

Thus we can write the criterion function as

rc c
J(w) =
.
n2 rn1

PROBLEM SOLUTIONS

119

This implies that the Fisher linear discriminant can be derived from the negative of
the log-likelihood ratio.
42. Consider the criterion function J(w) required for the Fisher linear discriminant.
(a) We are given Eqs. 96, 97, and 98 in the text:
J1 (w)

Si

2 |2
|m
1 m
(96)
2
s1 + s22
!
(x mi )(x mi )t

(97)

xD

SW

= S1 + S2 (98)

where y = wt x, m
i = 1/ni
y = wt mi . From these we can write Eq. 99 in
yYi

the text, that is,

s2i

(y m
i )2

yYi

(wt x wt mi )2

xD

wt (x mi )(x mi )t w

xD
t

= w Si w.
Therefore, the sum of the scatter matrixes can be written as
s21 + s22 = wt SW w

(100)

(m
1 m
2 ) = (w m1 w m2 )
= wt (m1 m2 )(m1 m2 )t w
= wt SB w,
2

(101)

where SB = (m1 m2 )(m1 m2 )t , as given by Eq. 102 in the text. Putting

these together we get Eq. 103 in the text,
J(w) =

wt SB w
.
wt SW w

(103)

(b) Part (a) gave us Eq. 103. It is easy to see that the w that optimizes Eq. 103
is not unique. Here we optimize J1 (w) = wt SB w subject to the constraint
that J2 (w) = wt SW w = 1. We use the method of Lagrange undetermined
multipliers and form the functional
g(w, ) = J1 (w) (J2 (w) 1).
We set its derivative to zero, that is,
g(w, )
wi

=
=




uti SB w + wt SB ui uti SW w + wt Sw ui

2uti (SB w SW w) = 0,

where ui = (0 0 1 0 0)t is the n-dimensional unit vector in the ith

direction. This equation implies
SB w = SW w.

120CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

(c) We assume that J(w) reaches the extremum at w and that w is a small
deviation. Then we can expand the criterion function around w as
J(w + w) =

(w + w)t SB (w + w)
.
(w + w)t SW (w + w)

From J(w) = J(w + w) = , to rst order we have

=

wt SB w
wt SB w + 2wt SB w
=
,
wt SW w
wt SW w + 2wt SW w

that is,
=

wt SB w
.
wt SW w

Therefore the following equation holds:

wt (SB w SW w) = 0.
Because w is arbitrary, we conclude that SB w = SW w.
43. Here we have the between-scatter matrix is
SB =

c
!

ni (mi m)(mi m)t ,

i=1

where the group and full means are

mi

1 !
x
ni
xDi

!
xD

1!
ni mi ,
n i=1
c

x=

and for this case of equal covariances,

SW =

c !
!

(x mi )(x mi )t =

i=1 xD

c
!

= c.

i=1

In order to maximize
J(W) =

|Wt SB W|
,
|Wt SW W|

the columns of an optimal W must be generalized eigenvectors that correspond to

the largest eigenvalues in
SB wi = i SW wi = ci wi ,
with W = [w1 w2 wc1 ] in terms of and d-dimensional mean vectors.
Section 3.9
44. Consider the convergence of the expectation-maximization algorithm, where as
usual Db and Dg are the badand the good data, as described in the text.

PROBLEM SOLUTIONS

121

(a) The expected value of as determined in the primed coordinates is


E  (; Dg ) =
l(; Dg )p(Db |Dg ;  )dDb

=
(lnp(Dg , Dg ; ) lnp(Db |Dg ; )) p(Db |Dg ;  )dDb




=
lnp(Dg , Db ; )p(Db |Dg ;  ) dDb lnp(Db |Dg ; )p(Db |Dg ;  )dDb

= EDb [lnp(Dg , Db ; )|Dg ;  ] ED
[lnp(Db : Dg ; )]
b


= Q(; ) EDb [lnp(Db : Dg ; )].

(b) If we dene (Db ) = p(Db |Dg ; )/p(Db |Dg ;  ), then


p(Db )|Dg ; )

E [(Db )] 1 =
p(Db )|Dg ; )dDb 1
p(Db |Dg ; 

=
p(Db |Dg ; )dDb 1
=

1 1 = 0.

According to Jensens inequality for this case E  [ln()] E  [] 1. Thus we

have in our particular case E  [ln(Db )] E  [(Db )] 1 = 0.
(c) The expected value of the log-likelihood at step t + 1 and at step t are
t
E[l( t+1 ; Dg )] = Q( t+1 ; t ) ED
[lnp(Db |Dg ; t+1 )]
b
t
E[l( t ; Dg )] = Q( t ; t ) ED
[lnp(Db |Dg ; t )],
b

respectively. We take the dierence and nd



E[l( t+1 ; Dg )] E[l( t ; Dg )] = Q( t+1 ; t ) Q( t ; t )
 t

t
ED
[lnp(Db |Dg ; t+1 )] ED
[lnp(Db |Dg ; t )]
b
b




p(Db |Dg ; t+1 )
t+1
t
t
t
t
= Q( ; ) Q( ; ) EDb ln
.
p(Db |Dg ; t )
From part (b), we have


p(Db |Dg ; t+1 )
t
t
ED
ln
= ED
[ln(Db )] < 0,
b
b
p(Db |Dg ; t
t
and thus ED
[ln(Db )] > 0 . Given the fact that Q( t+1 ; t ) > Q( t ; t ) and
b
t+1
that E[l( ; Dg )] E[l( t ; Dg )] > 0, we conclude

E[l( t+1 ; Dg )] > E[l( t ; Dg )],

and thus
l( t+1 ; Dg ) > l( t ; Dg )
for increasing t. In short, the log-likelihood increases through the application of
the expectation-maximization algorithm.

122CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

45. We consider an iterative algorithm in which the maximum-likelihood value of

missing values is calculated, then assumed to be correct for the purposes of reestimating and the process iterated.
(a) This is a generalized expectation-maximization algorithm, but not necessarily a
full expectation-maximization algorithm because there is no guarantee that the
optimal likelihood value is achieved on each step.
(b) We have Q(; i ) = EDb [lnp(Db ; )|Dg ; i )].
46. Recall rst a special case Jensens inequality, that is
E[lnx] < ln[E[x]].
     
Our data set is D = { 23 , 31 , 54 , 4 , 86 } sampled from a two-dimensional uniform
distribution bounded by xl1 x1 xu1 and xl2 x2 xl2 .
(a) According to the denition of Q(, 0 ) of Eq. 129 in the text, we have
Q(; 0 )

= Ex42 x51 [lnp(xg , xb ; | 0 ; Dg )]

2

  1!
3
=
lnp(xk |) + lnp(x4 |) + lnp(x5 |)
k=1

p(x42 | 0 ; x41 = 4)p(x51 | 0 ; x52 = 6)dx42 dx51


3
!
lnp(xk |) +
lnp(x4 |)p(x42 | 0 ; x41 = 4)dx42
k=1



p(x5 |)p(x51 | 0 ; x52 = 6)dx51 .



We now calculate the second term, K:


K

lnp(x4 |)p(x42 | 0 ; x41 = 4)dx42

"   #
4  0
 2
p
x42 

=
lnp
dx42
 "  #

4  0

dx
p



42
x42


#
"
 1  2 p  4  0
x42
4 
=
ln
dx42
 10
1
x42 


1010 dx42
1

1


=

10

4
x42

lnp

 2 1  2
4  0
4 
 p
 dx42 .
x42 
x42 

There are four cases we must consider when calculating K:

PROBLEM SOLUTIONS

123

1. 0 xl2 < xu2 10 and xl1 4 xu1 , which gives:

K

xu2   
1
4 
10
lnp
dx42

x42
10 10

xl2


1
10

(xu2 xl2 )ln

1
.
|xu1 xl1 | |xu2 xl2 |

2. xl2 0 < xu2 10 and xl1 4 xu1 :

K

xu2   
1
4 
10
lnp
dx42

x42
10 10

1
1
xu2 ln
.
10
|xu1 xl1 | |xu2 xl2 |

3. 0 xl2 < 10 xu2 and xl1 4 xu1 :



10
K

 
1
4 
dx42

10 10
x42

10

1
1
(10 xu1 )ln
.
10
|xu1 xl1 | |xu2 xl2 |

lnp

xu1

4. Otherwise, K = 0.
Similarly, there are four cases we must consider when calculating L:
1. xl1 < xu1 0 < 10 or 0 < 10 xl1 < xu1 or 6 xl2 or xu2 6: then L = 0.
2. 0 xl1 < xu1 10 and xl2 6 xu2 , then
L =

xu1   
x41 
1
10
lnp
dx41

6
10 10
xl1

1
1
(xu1 xl1 )ln
.
10
|xu1 xl1 | |xu2 xl2 |

3. xl1 0 < xu1 10 and xl2 6 xu2 , then

L =

10

xu1   
x41 
1
lnp
dx41

6
10 10
0

1
1
xu1 ln
.
10
|xu1 xl1 | |xu2 xl2 |

4. 0 xl1 < 10 xu1 and xl2 6 xu2 :



10
L =

10
xl1

lnp

 
1
x41 
dx41

6
10 10

1
1
(10 xl1 )ln
.
10
|xu1 xl1 | |xu2 xl2 |

124CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

Therefore Q(; 0 ) =

3

k=1

lnp(xk |) + K + L has dierent forms depending upon

the dierent values of , as shown above.

(b) Here we have = (2 1 5 6)t .
(c) See figure.
x2

x1 = 5

x1 = 2

10

x2 = 6

x2 = 1

x1

0
0

10

(d) Here we have = (2 1 5 6)t .

6     7
47. Our data set is D = 11 , 33 , 2 , where represents an unknown value for the
x2 component of the third data point.
(a) For the E step we have:


Q(; 0 ) = Ex32 lnp(xg , xb ; )| 0 , Dg

=
(lnp(x1 |) + lnp(x2 |) + lnp(x3 |)) p(x32 | 0 , x31 = 2)dx32

lnp(x1 |) + lnp(x2 |) +

lnp(x3 |) p(x32 | 0 , x31 = 2)dx32

"   #
2  0
   
p
x32 
2 

lnp(x1 |) + lnp(x2 |) +
p
dx32

   
x32
2
 0 

p
 d32
x32




1/(2e)




=

lnp(x1 |) + lnp(x2 |) + 2e

lnp(x1 |) + lnp(x2 |) + K.

lnp

    
2  0
2 
 p
 dx32
x32
x32

There are three cases for K:

1. 3 2 4:
K

1
4

2
ln
0

1 21 1
e
1
2


dx32

PROBLEM SOLUTIONS

125
=

1
2 ln
4

1 21 1
e
1
2


.

2. 2 4:
1
4

4

1 21 1
=
ln
e
1
2
0


1
1 21 1
=
e
4ln
4
1
2


1 21 1
= ln
.
e
1
2


dx32

3. Otherwise K = 0.
Thus we have
Q(; 0 )

lnp(x1 |) + lnp(x2 |) + K




1 1 1
1 31 1
= ln
+ ln
+K
e
e
1
2
1
2
= 1 ln(1 2 ) 31 ln(1 2 ) + K
= 41 2ln(1 2 ) + K,
=

according to the dierent cases of 2 . Note the normalization condition

1, or


p(x1 )dx1 =

1 1 x1
e
dx1 = 1.
1

This equation yields the solution 1 = 1.

(b) There are two cases:
1. 3 2 4:


1
Q(; ) = 4 2ln2 + 2 (2 + ln2 ) .
4
0

Note that this is a monotonic function and that arg max2 Q(; 0 ) = 3,
which leads to max Q(; 0 )  8.5212.
2. 2 4: In this case Q(; 0 ) = 6 3ln2 . Note that this is a monotomic
function and that arg max2 Q(; 0 ) = 4, which leads to max Q(; 0 ) 
10.1589.
In short, then, = (1 3)t .
(c) See figure.
48. Our data set is D =

61 3 7

1 . 3 , 2 .

126CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

t
0 = (2 4)

t
= (1 3)
x2

x2
p(x1 , x2)
0.3
0.2
0.1
0
0

p(x1 , x2)
0.3
0.2
0.1
0
0

5
4
3
2

5
4
3
2

1
4

x1

1
4

x1

(a) The E step is:

Q(; 0 )



= Ex31 lnp(xg , xb ; )| 0 , Dg

=
(lnp(x1 |) + lnp(x2 |) + lnp(x3 |)) p(x31 | 0 , x32 = 2)dx31




=

lnp(x1 |) + lnp(x2 |) +

lnp

"   #
x31  0
 
p
2 
x31 
dx31
    
2
x31  0

p
 dx31
2




1/16




=

lnp(x1 |) + lnp(x2 |) + 16

lnp(x1 |) + lnp(x2 |) + K.

lnp

    
x31 
x31  0
 p
 dx31
2
2

There are two cases for K:

1. 2 2:

K

2x31

2e


ln


=

1 1 x31 1
e
1
2


dx31

2e2x31 (1 x31 ln(1 2 ))dx31

= ln(1 2 )

2e2x31 dx31 21

1
= ln(1 2 ) 21
4
1
=
1 ln(1 2 ).
2
2. Otherwise K = 0.
Therefore we have
Q(; 0 )

lnp(x1 |) + lnp(x2 |) + K

x31 e2x31 dx31

PROBLEM SOLUTIONS

127





1 1 1
1 31 1
= ln
+ ln
+K
e
e
1
2
1
2
= 1 ln(1 2 ) 31 ln(1 2 ) + K
= 41 2ln(1 2 ) + K,
according to the cases for K, directly above. Note that

p(x1 )dx1 = 1, and

thus


1 1 x
e
dx1 = 1,
1

and thus 1 = 1. The above results can be simplied.

(b) When 2 2, we have
Q(; 0 )

= 4 2ln2 +
7
= 3ln2 .
2

1
ln2
2

Note that this is a monotonic function, and thus arg max2 Q(; 0 ) = 2. Thus
the parameter vector is = (1 2)t .
(c) See figure.

t
0 = (2 4)
x2

p(x1 , x2)
0.4
0.3
0.2
0.1
0
0

t
= (1 2)

3
1

2
3

0.4
0.3
0.2
0.1
0
0

1
4

x2

p(x1 , x2)

x1

3
1

2
3

1
4

x1

Section 3.10
49. A single revision of a
ij s and bij s involves computing (via Eqs. ?? ??)
T


a
ij

ij (t)

t=1
T 

t=1 k
??


bij

??

T


t=1

ij (t)

ik (t)

ij

ij (t)

i (t 1)aij bij i (t)

.
P (V T |M )

128CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

i (t)s and P (V T |M ) are computed by the Forward Algorithm, which requires O(c2 T )
operations. The i (t)s can be computed recursively as follows:
For t=T to 1 (by -1)
For i=1to c
i (t) = j aij bjk v(t + 1)j (t + 1)
End
This requires O(c2 T ) operations.
Similarly, ijs can be computed by O(c2 T ) operations given i (t)s, aij s, bij s,
i (t)s and P (V T |M ). So, ij (t)s are computed by
O(c2 T ) + O(c2 T ) + O(c2 T ) = O(c2 T )operations.
        
i (t)s
i (t)s
ij (t)s
Then, given ij (t)s, the a
ij s can be computed by O(c2 T ) operations and bij s by
O(c2 T ) operations. Therefore, a single revision requires O(c2 T ) operations.
50. The standard method for calculating the probability of a sequence in a given
HMM is to use the forward probabilities i (t).
(a) In the forward algorithm, for t = 0, 1, . . . , T , we have

0
t = 0 and j = initial status

1
t = 0 and j = initial status
j (t) =
c


i (t 1)aij bjk v(t) otherwise.

i=1

In the backward algorithm, we use for t = T, T 1, . . . , 0,

0
t = T and j = nal status

1
t = T and j = nal status
j (t) =
c


i (t + 1)aij bjk v(t + 1) otherwise.

i=1

Thus in the forward algorithm, if we rst reverse the observed sequence VT

(that is, set bjk v(t) = bjk (T + 1 t) and then set j (t) = j (T t), we can
obtain the backward algorithm.
(b) Consider splitting the sequence VT into two parts V1 and V2 before,
during, and after each time step T  where T  < T . We know that i (T  )
represents the probability that the HMM is in hidden state i at step T  , having
generated the rt T  elements of VT , that is V1 . Likewise, i (T  ) represents
the probability that the HMM given that it is in i at step T  generates the
remaining elements of VT , that is, V2 . Hence, for the complete sequence we
have
c
!
P (VT ) = P (V1 , V2 ) =
P (V1 , V2 , hidden state i at step T  )
=

c
!

i=1

P (V1 , hidden state i at step T  )P (V2 |hidden state i at step T  )

i=1

c
!
i=1

i (T  )i (T  ).

PROBLEM SOLUTIONS

129

(c) At T  = 0, the above reduces to P (VT ) =

c

i=1

i (0)i (0) = j (0), where j is

the known initial state. This is the same as line 5 in Algorithm 3. Likewise, at
c

T  = T , the above reduces to P (VT ) =
i (T )i (T ) = j (T ), where j is the
i=1

known nal state. This is the same as line 5 in Algorithm 2.

51. From the learning algorithm in the text, we have for a giveen HMM with model
parameters :
ij (t) =

a
ij

i (t 1)aij bjk v(t)j (t)

()
P (VT |)
T

ij (t)
t=1
.
()
c
T 

ik (t)
t=1 k=1

For a new HMM with ai j  = 0, from () we have i j  = 0 for all t. Substituting
i j  (t) into (), we have a
i j  = 0. Therefore, keeping this substitution throughout
the iterations in the learning algorithm, we see that a
i j  = 0 remains unchanged.
52. Consider the decoding algorithm (Algorithm 4).
(a) the algorithm is:
Algorithm 0 (Modied decoding)
1
2
3
4
5

begin initialize Path {}, t 0

for t t + 1
j 0; 0 0
for j j + 1
j (t) min [i (t 1) ln(aij )] ln[bjk v(t)]
1ic

until j = c
j  arg min[j (t)]

6
7

Append j  to Path
until t = T
return Path

8
9
10
11

end

(b) Taking the logarithm is an O(c2 ) computation since we only need to calculate
lnaij for all i, j = 1, 2, . . . , c, and ln[bjk v(t)] for j = 1, 2, . . . , c. Then, the whole
complexity of this algorithm is O(c2 T ).

130CHAPTER 3.

MAXIMUM LIKELIHOOD AND BAYESIAN ESTIMATION

Computer Exercises
Section 3.2
1. Computer exercise not yet solved
Section 3.3
2. Computer exercise not yet solved
Section 3.4
3. Computer exercise not yet solved
Section 3.5
4. Computer exercise not yet solved
Section 3.6
5. Computer exercise not yet solved
Section 3.7
6. Computer exercise not yet solved
7. Computer exercise not yet solved
8. Computer exercise not yet solved
Section 3.8
9. Computer exercise not yet solved
10. Computer exercise not yet solved
Section 3.9
11. Computer exercise not yet solved
12. Computer exercise not yet solved
Section 3.10
13. Computer exercise not yet solved

Chapter 4

Nonparametric techniques

Problem Solutions
Section 4.3
1. Our goal is to show that Eqs. 1922 are sucient to assure convergence of Eqs. 17
and 18. We rst need to show that lim pn (x) = p(x), and this requires us to show
n
that


1
xV
lim
= n (x V) (x V).

n Vn
hn
Without loss of generality, we set V = 0 and write
1

Vn

x
hn


=

1
d
/
i=1

xi

 
d
,
xi
x
= n (x),

h
hn
i=1 n

where we used the fact that Vn = hdn . We also see that n (x) 0 if x = 0, with
normalization

n (x)dx = 1.
Thus we have in the limit n , n (x) (x) as required.
We substitute our denition of pn (x) and get

lim pn (x) = E[
p(x)] = (x V)p(V)dV = p(x).
n

Furthermore, the variance is

n2 (x)

1
nVn

1 2

Vn

xV
hn

131


p(V)dv

1
[
pn (x)]2
n

CHAPTER 4. NONPARAMETRIC TECHNIQUES

132




1 2 xV
1
p(V)dV

nVn
Vn
hn



Sup()
xV
1

p(V)dV
nVn
Vn
hn
Sup()
pn (x)
.
nVn

We note that Sup() < and that in the limit n we have pn (x) p(x) and
nVn . We put these results together to conclude that
lim n2 (x) = 0,

for all x.
2. Our normal distribution is p(x) N (, 2 ) and our Parzen window is (x)
N (0, 1), or more explicitly,
(x)

2
1
ex /2 ,
2

and our estimate is

pn (x)



n
1 !
x xi
.

nhn i=1
hn

(a) The expected value of the probability at x, based on a window width parameter
hn , is

 
n
1 !
x xi
pn (x) = E[pn (x)] =
E
nhn i=1
hn
=

1
hn
1
hn

xv
hn

p(v)dv



2 


2 
1
1
1 xv
1 v
exp

dv
exp
2
hn
2

2
2

 







1
x
2
1 2 1
1

1 x2
dv
+ 2
exp v
+ 2 2v
+ 2
exp
2hn
2 h2n

2
h2n

h2n





2 

 
1
1 x2
1 2
2
1 v
dv,
exp
+ 2 +
exp
2hn
2 h2n

2 2
2

where we have dened

2 =

1
h2n 2
=
1/h2n + 1/ 2
h2n + 2

and


=

+ 2
h2n


.

PROBLEM SOLUTIONS

133

We perform the integration and nd





2
1 x2
1 2
2
pn (x) =
exp
+ 2 +
2hn
2 h2n

2 2

 2

1
1 x
2
2
hn

=
.
exp
+ 2 2
2 h2n

2hn h2n + 2
The argument of the exponentiation can be expressed as follows
x2
2
2
+

h2n
2
2


2
x2
2
4 x

+ 2 2
+ 2
h2n

h2n

x2 h2n
2 2
2x
+ 2
2
2
2
2
2
2
(hn + )hn
(hn + )
hn + 2
2
(x )
.
h2n + 2

=
=
=

We substitute this back to nd

pn (x)



1
1 (x )2

,
exp
2 h2n + 2
2 h2n + 2

which is the form of a Gaussian, denoted

pn (x) N (, h2n + 2 ).
(b) We calculate the variance as follows:



n
1 !
x xi
Var[pn (x)] = Var

nhn i=1
hn



n
1 !
x xi
=
Var

n2 h2n i=1
hn

 
1
xv
=
Var
2
nhn
hn
  
   
2 
1
xv
xv
2
=
E
,
E
nh2n
hn
hn
where in the rst step we used the fact that x1 , . . . , xn are independent samples
drawn according to p(x). We thus now need to calculate the expected value of
the square of the kernel function
 




xv
xv
2
2
E
=

p(v)dv
hn
hn

 
2 

2 

1 v
1
1
xv

=
exp
exp
dv
2
hn
2





2 
2 

1
1 v
1
1
1 xv
exp

=
exp
dv.
2
2

2
2
hn / 2
2

CHAPTER 4. NONPARAMETRIC TECHNIQUES

134

From part (a) by a similar argument with hn / 2 replacing hn , we have



2 


2 

1
1 v
1
1
1 xv
exp

exp
dv
2 hn / 2
2

hn / 2
2
2



1
1 (x )2
= 
.
exp 2
2 hn /2 + 2
2 h2n /2 + 2
We make the substitution and nd

 



xv
h / 2
1 (x )2
n
E 2
=
,
exp 2
hn
2 hn /2 + 2
2 h2n /2 + 2
and thus conclude


 

1
xv
1
1
1 (x )2
1
2


=
.
E

exp

nh2n
hn
nhn 2 2 h2n /2 + 2
2 h2n /2 + 2

For small hn , h2n /2 + 2  , and thus the above equation can be approximated as


 

2 
1
xv
1
1
1 x
2


E
exp
nh2n
hn
2

nhn 2 2
1
p(x).
=
()
2nhn
Similarly, we have
  
 2
1
xv
E
nh2n
hn

=
=




1
1 2
1 (x )2
h 
exp
nh2n n 2 h2n + 2
2 h2n + 2


hn
1
1 (x )2

exp
nhn 2 h2n + 2
2 h2n + 2


hn
1
1 (x )2

 0, ()
exp
nhn 2
2
2

valid for small hn . From () and () we have, (still for small hn )

Var[Pn (x)] 
(c) We write the bias as
p(x) pn (x)

=
=
=
=

p(x)
.
2nhn





1
1
1 (x )2
1 (x )2


exp
exp
2
2
2 h2n + 2
2
2 h2n + 2





1
1 (x )2
1 (x )2

1 (x )2

exp
+
1 
exp
2 h2n + 2
2 h2n + 2
2
2
2
h2n + 2




1
(x )2 8
19
1
p(x) 1 
exp

2
h2n + 2
2
1 + (hn /)2



 2
1
1 hn (x )2
p(x) 1 
.
exp
2 h2n + 2
1 + (hn /)2

PROBLEM SOLUTIONS

135

For small hn we expand to second order:

1
1
2
3
 1 (hn /)
2
hn 2
1+( )
and

exp


2
h2n (x )
h2n (x )2

1
+
.
2 2 h2n + 2
2 2 h2n + 2

We ignore terms of order higher than h2n and nd



2 2 

h2 (x )2
1 + n2 2
p(x) 1
2 hn + 2


1 h2n
h2n (x )2
p(x) 1 1 +
2 2
2 2
2 hn + 2

 2 
1 hn
(x )2
1 2
p(x)
2
hn + 2


 2
2 
1 hn
x
1
p(x).
2

p(x) pn (x) 




1
1
2

hn



3. Our (normalized) distribution is



1/a 0 x a
0
otherwise,

p(x) =
and our Parzen window is


(x) =

ex
0

x0
x < 0.

(a) The expected value of the Parzen window estimate is




 
n
1
xv
1 !
x xi
pn (x) = E
=
p(v)dv

nhn i=1
hn
hn
hn

(xv)
1
=
e hn p(v)dv
hn


xv

exp[x/hn ]
hn

1
exp[v/hn ]dv
a

xv,
0<v<a

x

ex/hn

ev/hn dv

if

if x a

ahn

if x < 0
0xa

ex/hn
ahn

a
ev/hn dv
0

CHAPTER 4. NONPARAMETRIC TECHNIQUES

136

0
=

1
x/hn
)
a (1 e
1 a/hn
x/hn
(e

1)e
a

if x < 0
if 0 x a
if x a,

where in the rst step we used the fact that x1 , . . . , xn are independent samples
drawn according to p(v).
(b) For the case a = 1, we have

0
1 ex/hn
pn (x) =
1/hn
(e
1)ex/hn

x<0
0x1
x > 1,

as shown in the gure.

p(x)
1

hn = 1/16

0.8
0.6
0.4

hn = 1
0.2

hn = 1/4
x
0.5

1.5

2.5

(c) The bias is

p(x) pn (x)

a1 (1 ex/hn )

0 a1 (ea/hn 1)ex/hn

if
0
1 x/hn
e
if
=
a 1 a/hn
a (e
1)ex/hn if
=

1
a

if x < 0
if 0 x a
if x a
x<0
0xa
x a.

Formally, a bias lower than 1% over 99% of the range 0 < x < a, means that
p(x) p(x)
0.01
p(x)
over 99% of the range 0 < x < a. This, in turn, implies
1/aex/hn
0.01 over 99% of 0 < x < a or
1/a
0.01a
hn
.
ln (100)
For the case a = 1, we have that hn 0.01/(ln100) = 0.0022, as shown in
the gure. Notice that the estimate is within 1% of p(x) = 1/a = 1.0 above
x 0.01, fullling the conditions of the problem.

PROBLEM SOLUTIONS

137

p(x)
1

hn = 0.0022

0.8
0.6
0.4
0.2

x
0

0.01

0.02

0.03

0.04

0.05

4. We have from Algorithm 2 in the text that the discriminant functions of the PNN
classier are given by

 t
ni
!
wk x 1
gi (x) =
aki
exp
i = 1, . . . , c
2
k=1

where wk  = x = 1, ni is the number of training patterns belonging to i and


1 if wk i
aki =
0 otherwise.
(a) Since wk  = x = 1, gi (x) can be written as


ni
!
x wk 2
gi (x) =
aki .
exp
2 2
i=1
Note that gi (x)/ni is a radial Gaussian based kernel estimate, pn (x|i ), of
p(x|i ). If we use ni /n as the estimate of the prior class probability P (i ),
then gi (x) can be rewritten as
gi (x) = nPn (i )p(x|i ).
Thus gi (x) properly accounts for the class priors.
(b) The optimal classication rule for unequal costs is given by
Choose k if gk = min gi ,
ic

where the ij represent the costs and

gi (x) =

c
!

ij P (j |x) =

j=1

c
!
P (j )p(x|j
j=1

p(x)

This discriminant function can be written simply as

gi (x) =

c
!

ij P (j )p(x|j ).

j=1

Consequently, the PNN classier must estimate gi (x). From part (a) we have
that gi = nPn (j )p(x|j ). Thus the new discriminant functions are simply
1!
ij gj (x),
n j=1
c

gi (x) =

CHAPTER 4. NONPARAMETRIC TECHNIQUES

138

where gi (x) are the discriminant functions of the PNN classier previously dened. An equivalent discriminant function is
gi (x) =

c
!

ij gj (x).

j=1

(c) The training algorithm as dened is unaected since it only performs a normalization of the training patterns. However the PNN classication algorithm must
be rewritten as:
Algorithm 0 (PNN with costs)

11

begin initialize x test pattern, gi 0, gi 0, ij

k0
do k k + 1
netk wkt x
if aki = 1 then gi gi + exp[(netk 1)/ 2 ]
until k = n
k0
do k k + 1
gi gi + ik gk
until k = n
return class arg max gi (x)

12

end

1
2
3
4
5
6
7
8
9
10

Section 4.4
5. Our goal is to show that pn (x) converges in probability to p(x) given the below
conditions:
lim kn

(condition 1)

lim kn /n 0

(condition 2)

n
n

As we know from probability theory, when pn (x) converges in the rth mean (r 1)
to p(x) that is, E[|pn (x) p(x)|r ] 0 as n 0 then p(x) also converges in
probability to p(x).
It is simple to show for r = 2 that
E[(pn (x) p(x))2 ] = Var[pn (x)] + bias2 [pn (x)],
where
Var[pn (x)]
bias2 [pn (x)]



= E pn (x) E[pn (x)]
= E[pn (x) p(x)].

Consequently, pn (x) converges in probability to p(x) if and only if

Var[pn (x)] 0

as n

(condition 3)

E[pn (x)] = p(x)

as n

(condition 4).

PROBLEM SOLUTIONS

139

Note that the k-nearest-neighbor estimate pn (x) = knVn/n can be rewritten as a kernel
estimate according to Eq. 11 in the text:


n
1! 1
x xi
pn (x) =
,

n i=1 Vn
hn
where hn denotes here the radius of a ball which includes the kn prototypes nearest
to x, Vn is the volume of the ball, and
 

x xi
1 if x xi  hn
=

0 otherwise.
hn
Hence we can apply the results of page 167 in the text. According to Eq. 23 in the
text,
 
1
x
E[pn (x)] = p(x) as n if and only if
= (x) as n ,

Vn
hn
where (x) is the delta function.
Given a test pattern x (see gure), we will have a very small ball (or more precisely,
have Vn 0) as n , since we can get k prototypes arbitrarily close to x due to
condition 2. If Vn 0, hn 0 and therefore
  
x
1 if x = 0

=
0 otherwise.
hn
1
Vn (x/hn )

Consequently,

= (x) and hence condition 4 is fullled.

hn

x1
xkn

x3

x
x2

x4
x5

We can compute lim Var[pn (x)] using Eq. 24 as follows:

n

lim Var[pn (x)]




1
1 2 xu

p(u)du
n nVn
Vn
hn
$
"
#%

lim V1n 2 xu
p(u)du
hn
lim

lim nVn

(x u)p(u)du
lim nVn
n

p(x)
.
lim nVn

n
kn
n nVn

Since lim

= P (x) in probability and condition 1 is met, then lim nVn = .

Thus Var[p(x)] 0 as n 0.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

140

6. We are given that P (1 ) = P (2 ) = 1/2,


1
P (1 |x) =
0

1
P (2 |x) =
0

and that
if ||x|| 1
otherwise
if ||x a|| 1
otherwise,

where we have assumed, without loss of generality, that category 1 is centered on

the origin 0, and category 2 is centered on a, a point other than the origin.
(a) We denote by Pn (e) the average probability of error based on n points.
Pn (e)

Pr[true category is 1 while 2 is most frequently labeled]

+Pr[true category is 2 while 1 is most frequently labeled]

=
=

2Pr[true category is 1 while 2 is most frequently labeled]

2P (1 )Pr[label of 1 for fewer than (k 1)/2points, and the rest labeled 2 ]

1
2

(k1)/2

(k1)/2

!
j=0

Pr[j of n chosen points are labeled 1 , the rest 2 ]

j=0


n 1
1
j 2j 2(nj)

(k1)/2 
1 ! n
.
2n j=0 j

(b) We make explicit the k dependence on the probability by writing Pn (e) =

Pn (e; k) and have
Pn (e; 1) =

1
1
< Pn (e; k) = n
2n
2

(k1)/2 


n
.
j
j=0



!

>0 f or k>1

(c) We have in this case

Pn (e)

=
=

1
2n

(k1)/2 

!
j=0

n
j


= Pr [B(n, 1/2) (k 1)/2]



k1
Pr Y1 + + Yn
,
2

where Y1 , , Yn are independent, B(, ) is a binomial distribution and Pr[Yi =

1] = Pr[Y
i = 0] = 1/2. If k is allowed to increase with n, but is restricted by
k < a/ n, then we have



a/ n 1
Pn (e) Pr Y1 + + Yn
2
= P r(Y1 + + Yn 0) for n suciently large
= 0,
and this guarantees Pn (e) 0 as n .

PROBLEM SOLUTIONS

141

Section 4.5
7. Our goal is to show that Voronoi cells induced by the nearest-neighbor algorithm
are convex, that is, given any two points in the cell, the line connecting them also
lies in the cell. We let x be the labeled sample point in the Voronoi cell, and y be

x(0)

y
x*

x()

x(1)

any other labeled sample point. A unique hyperplane separates the space into those
that are closer to x than to y, as shown in the gure. Consider any two points
x(0) and x(1) inside the Voronoi cell of x ; these points are surely on the side of the
hyperplane nearest x . Now consider the line connecting those points, parameterized
as x() = (1 )x(0) + x(1) for 0 1. Because the half-space dened by
the hyperplane is convex, all the points x() lie nearer to x than to y. This is true
for every pair of points x(0) and x(1) in the Voronoi cell. Furthermore, the result
holds for every other sample point yi . Thus x() remains closer to x than any other
labeled point. By our denition of convexity, we have, then, that the Voronoi cell is
convex.
8. It is indeed possible to have the nearest-neighbor error rate P equal to the Bayes
error rate P for non-trivial distributions.
(a) Consider uniform priors over c categories, that is, P (i ) = 1/c, and onedimensional distributions

cr
0 x c1
1
cr
1
i x i + 1 c1
p(x|i ) =

0
elsewhere.
The evidence is

1
1/c
p(x) =
p(x|i )P (i ) =

i=1
0
c
!

cr
0 x c1
i x (i + 1)
elsewhere.

Note that this automatically imposes the restriction

0

cr
1.
c1

cr
c1

CHAPTER 4. NONPARAMETRIC TECHNIQUES

142

Because the P (i )s are constant, we have P (i |x) p(x|i ) and thus

P (i |x) =

P (i )
p(x)

1/c
p(x)

0x

cr
c1

if i = j
if i = j

0
1

j xj+1

cr
c1

otherwise,

as shown in the gure for c = 3 and r = 0.1.

P(i|x)
1/c = 1/3
1

1 , 2 , 3

0.3
0.25

0.2
0.15
0.1
0.05

x = r = 0.1

x
1

The conditional Bayesian probability of error at a point x is

P (e|x)

1 P (max |x)

1/c
cr
1 p(x)
if 0 x c1
=
1 1 = 0 if i x i + 1
=

cr
c1 ,

and to calculate the full Bayes probability of error, we integrate as



P (e|x)p(x)dx

cr/(c1)




1/c
=
1
p(x)dx
p(x)
0


1
cr
=
1
= r.
c c1
(b) The nearest-neighbor error rate is
 
P

c
!


P 2 (i |x) p(x)dx

i=1
cr/(c1)



=
0

cr

j+1 c1


c
!
c( 1c )2
1 2
[1 1] p(x)dx
p(x)dx +
p (x)
j=1 j



0

PROBLEM SOLUTIONS
cr/(c1)



=
0

143
1/c
1 2
p (x)

cr/(c1)



=
0

1
c


p(x)dx


dx =

1
c

cr
= r.
c1

Thus we have demonstrated that P = P = r in this nontrivial case.

9. Our data are given in the table.
1
(10,0)
(0,-10)
(5,-2)

2
(5,10)
(0,5)
(5,5)

3
(2,8)
(-5,2)
(10,-4)

Throughout the gures below, we let dark gray represent category 1 , white represent
category 2 and light gray represent category 3 . The data points are labeled by their
category numbers, the means are shown as small black dots, and the dashed straight
lines separate the means.
a)
b)
15

15

10

10

3
2

3
1

1
3

-5

-5
-10

-10

-15

-15
-10

-5

10

-10

15

-5

c)

10

15

d)

15

15

10

10

3
5

3
5

1
1

-5

-5

-10

-10

-15

-15
-10

-5

10

15

-10

-5

10

15

10. The Voronoi diagram of n points in d-dimensional space is the same as the
convex hull of those same points projected orthogonally to a hyperboloid in (d + 1)dimensional space. So the editing algorithm can be solved either with a Voronoi
diagram algorithm in d-space or a convex hull algorithm in (d + 1)-dimensional space.
Now there are scores of algorithms available for both problems all with dierent
complexities.
A theorem in the book by Preparata and Shamos refers to the complexity of
the Voronoi diagram itself, which is of course a lower bound on the complexity of
computing it. This complexity was solved by Victor Klee, On the complexity of
d-dimensional Voronoi diagrams, Archiv. de Mathematik., vol. 34, 1980, pp. 7580. The complexity formula given in this problem is the complexity of the convex
hull algorithm of Raimund Seidel, Constructing higher dimensional convex hulls at

144

CHAPTER 4. NONPARAMETRIC TECHNIQUES

logarithmic cost per face, Proc. 18th ACM Conf. on the Theory of Computing, 1986,
pp. 404-413.
So here d is one bigger than for Voronoi diagrams. If we substitute d in the
formula in this problem with (d 1) we get the complexity of Seidels algorithm
for Voronoi diagrams, as discussed in A. Okabe, B. Boots and K. Sugihara, Spatial
Tessellations: Concepts and Applications of Voronoi Diagrams, John Wiley,
1992.
11. Consider the curse of dimensionality and its relation to the separation of points
randomly selected in a high-dimensional space.
(a) The sampling density goes as n1/d , and thus if we need n1 samples in d = 1
dimensions, an equivalent number samples in d dimensions is nd1 . Thus if we
needed 100 points in a line (i.e., n1 = 100), then for d = 20, we would need
n20 = (100)20 = 1040 points roughly the number of atoms in the universe.
(b) We assume roughly uniform distribution, and thus the typical inter-point Euclidean (i.e., L2 ) distance goes as d volume, or (volume)1/d .
(c) Consider points uniformly distributed in the unit interval 0 x 1. The length
containing fraction p of all the points is of course p. In d dimensions, the width
of a hypercube containing fraction p of points is ld (p) = p1/d . Thus we have
l5 (0.01) = (0.01)1/5
l5 (0.1) = (0.1)1/5
l20 (0.01) = (0.01)1/20
l20 (0.1) = (0.1)1/20

= 0.3910
= 0.7248
= 0.8609
= 0.8609.

(d) The L distance between two points in d-dimensional space is given by Eq. 57
in the text, with k :
;
< d
<!
1/k
L (x, y) = lim =
|xi yi |k
k

i=1

= max[|x1 y1 |, |x2 y2 |, . . . , |xd yd |]

= max |xi yi |.
i

In other words, consider each axis separately, i = 1, . . . , d. There is a separation

between two points x and y along each individual direction i, that is, |xi yi |.
One of these distances is the greatest. The L distance between two points is
merely this maximum distance.
Informally we can see that for two points randomly selected in the unit ddimensional hypercube [0, 1]d , this L distance approaches 1.0 as we can nearly
always nd an axis i for which the separation is large. In contrast, the L
distance to the closest of the faces of the hypercube approaches 0.0, because
we can nearly always nd an axis for which the distance to a face is small.
Thus, nearly every point is closer to a face than to another randomly selected
point. In short, nearly every point is on the outside (that is, on the convex
hull) of the set of points in a high-dimensional space nearly every point is
an outlier.
We now demonstrate this result formally. Of course, x is always closer to a
wall than 0.5 even for d = 1 and thus we consider distances l in the

PROBLEM SOLUTIONS

145

range 0 l < 0.5. The gure at the left shows the xi and yi coordinates
plotted in a unit square. There are d points, one corresponding to each axis in
the hypercube. The probability that a particular single coordinate i will have
|xi yi | l is equal to 1.0 minus the area in the small triangles, that is,
Pr[|xi yi | l for a particular i] = (1 l )2 .
The probability that all of the d points will have |xi yi | l is then
Pr[|xi yi | l for all i] = (1 (1 l )2 )d .
The probability that at least one coordinate has separation less than l is just
1.0 minus the above, that is,
Pr[at least one coordinate is closer than l ] = 1 (1 (1 l )2 )d .
Now consider the distance of a point x to the faces of the hypercube. The gure
L-infinity distance between x
and any hypercube face

0.
4

L-infinity distance between x and y

-y

i|

yi

d points

|x
i

|x
i

-y

i|

0.
4

0.5
l*
l*
0
0

l*
0.5

1-l*
xi

xi
1

0.5

at the right shows d points corresponding to the coordinates of x. The probability that a single particular coordinate value xi is closer to a corresponding
face (0 or 1) than a distance l is clearly 2l since the position of xi is uniformly
distributed in the unit interval. The probability that any of the coordinates is
closer to a face than l is
Pr[at least one coordinate is closer to a face than l ] = 1 (1 2l )d .
We put these two results together to nd that for a given l and d, the probability
that x is closer to a wall than l and that L (x, y) l is the product of the
two independent probabilities, that is,
[1 (1 2l )d ][1 (1 l )2d ],
which approaches 1.0 quickly as a function of d, as shown in the graph (for the
case l = 0.4). As shown in Fig. 4.19 in the text, the unit hypersphere in
the L distance always encloses the unit hypersphere in the L2 or Euclidean
metric. Therefore, our discussion above is pessimistic, that is, if x is closer to
a face than to point y according to the L metric, than it is even closer to
the wall than to a face in the L2 metric. Thus our conclusion above holds for
the Euclidean metric too. Consequently in high dimensions, we generally must
extrapolate from sample points, rather than interpolate between them.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

146

probability x is closer to a face than l* = 0.4

and x is farther from y (in L-infinity distant) than l*
1
0.8
0.6

l* = 0.4
0.4
0.2
2

12. The curse of dimensionality can be overcome if we know the form of the
target function.
(a) Our target function is linear,
f (x) =

d
!

aj xj = at x,

j=1

and y = f (x) + N (0, 2 ), that is, we have Gaussian noise. The approximation
d

is f(x) =
a
j xj , where
j=1

a
j = arg min
aj

n
!

yi

i=1

d
!

2
aj xij ,

j=1

where xij is the jth component of point i, for j = 1, . . . , d and i = 1, . . . , n.

In short, these are the best t coecients in a sum-squared-error sense. The
expected value (over the data set) of the dierence between the target and t
functions (squared) is

E[f (x) f(x)]2

2 2
d
d
n
d
!
!
!
!

yi
= E
aj xj
aj xij xj .
arg min
j=1

aj

j=1

i=1

j=1

For some set of sample points xi , we let yi be the corresponding sample values
yi = f (xi ) + N (0, 2 ). Consider the n-by-d matrix X = [xij ] where xij is the
jth component of xi , that is

X=

x1
x2
..
.

xn
where xi is a d-dimensional vector. As stated above, we have f (x) =

d

j=1

aj xj .

PROBLEM SOLUTIONS

147

We dene the weight vector

a1
a2
..
.

a=

an
then we can write

Xa =

f (x1 )
f (x2 )
..
.

f (xn )
and hence y = Xa + N (0, 2 )n where

y=

y1
y2
..
.

yn
d

We approximate f (x) with f(x) =
a
j xj where the a
j are chosen to minimize

n
!

yi

i=1

j=1

d
!

a
j xij = y X
a 2

j=1

is self-evident. Thus we choose a

such that X
where the denition of a
a is as
close to y as possible. However, X
a is in CX , the column space of X, that is,
the d-dimensional subspace of Rn spanned by the columns of X. (Of course, we
assume n > d.)
such that X
It is clear that we wish to choose a
a is the projection of y onto CX ,
which we denote ProjCX y. Now y is distributed according to an n-dimensional
Gaussian distribution with mean Xn. Projecting an n-dimensional Gaussian
onto a d-dimensional subspace yields a d-dimensional Gaussian, that is,
X
a = ProjCX y = ProjCX [Xn + N (0, 2 )n ] = Xa + N (0, 2 )d ,
where N (0, 2 )d is rotated to lie in CX . Thus we have




E X
a Xa2 = Var N (0, 2 )d = d 2 .
But X
a Xa2 =

n

(f(xi ) f (xi ))2 . Since the terms (f(xi ) f (xi ))2 are

i=1

independent for each i, we have

%
$


E X
a Xa2 = nE (f(x) f (x))2 = d 2 ,
and thus

$
% d 2
E (f (x) f(x))2 =
.
n
In short, the squared error increases linearly with the dimension d, not exponentially as we might expect from the general curse of dimension.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

148

(b) We follow the logic of part (a). Now our target function is f (x) =

d

i=1

ai Bi (x)

where each member in the basis set of M basis functions Bi (x) is a function of
the d-component vector x. The approximation function is
M
!

f(x) =

a
m Bm (x),

m=1

and, as before, the coecients are least-squares estimates

a
i = arg min
ai

n
!

q
!

yi

2
am Bm (x)

m=1

i=1

and yi = f (xi )+N (0, 2 ). Now y will be approximated by B

a, the projection of
y onto the column space of Bj , that is, the subspace spanned by the M vectors

Bi (x1 )
Bi (x2 )
..
.

Bi (xn )
As in part (a), we have
M 2
E[(f (x) f(x))2 ] =
,
n
which is independent of d, the dimensionality of the original space.
13. We assume P (1 ) = P (2 ) = 0.5 and the distributions are as given in the gure.
p(x|i)

2
2

x2

0
0

x1
x*

0.5

x
1

(a) Clearly, by the symmetry of the problem, the Bayes decision boundary is x =
0.5. The error is then the area of the dark shading in the gure, divided by the
total possible area, that is
P

1
=

min[P (1 )p(x|1 ), P (2 )p(x|2 )] dx

0

PROBLEM SOLUTIONS

149

0.5
1
= P (1 ) 2x dx + P (2 ) (2 2x) dx
0

0.5

1
1
0.5 + 0.5 = 0.25.
4
4

(b) Suppose a point is randomly selected from 1 according to p(x|1 ) and another
point from 2 according to p(x|2 ). We have that the error is
1

1
dx1 p(x1 |1 )

(x1 
x2 )/2

1
dx2 p(x2 |2 )

dx p(x|2 )

(x1 x2 )/2

dx p(x|1 ).
0

(c) From part (d), below, we have for the special case n = 2,
1
1
1
51
+
+
=
= 0.425.
3 (2 + 1)(2 + 3) 2(2 + 2)(2 + 3)
120

P2 (e) =

(d) By symmetry, we may assume that the test point belongs to category 2 . Then
the chance of error is the chance that the nearest sample point ot the test point
is in 1 . Thus the probability of error is
1
Pn (e)

P (x|2 )Pr[nearest yi to x is in 1 ]dx

0

1
=

P (x|2 )

n
!

Pr[yi 1 and yi is closer to x than yj , j = i]dx.

i=1

By symmetry the summands are the same for all i, and thus we can write
1
Pn (e)

P (x|2 )nPr[y1 1 and |y1 x| > |yi x|, i > 1]dx

=
0

1
=

1
P (1 |y1 )Pr[|yi x| > |y1 x|, i > 1]dy1 dx

P (x|2 )n
0

1

1

P (1 |y1 )Pr[|y2 x| > |y1 x|]n1 dy dx,

P (x|2 )n
0

where the last step again relies on the symmetry of the problem.
To evalute Pr[|y2 x| > |y1 x|], we divide the integral into six cases, as shown
in the gure.
We substitute these values into the above integral, and break the integral into
the six cases as
Pn (e)

 x
1/2
P (x|2 )n
P (1 |y1 )(1 + 2y1 2x)n1 dy1
0

CHAPTER 4. NONPARAMETRIC TECHNIQUES

150

denotes possible locations

of y2 with |y2-x|>|y1-x|

Case 1.1:

x  [0,1/2] 0 < y1 < x

y1 x
2x-y1

1/2

Case 2.1:

x  [0,1/2] x < y1 < 2x

2x-y1 x y1
1/2

Case 2.2:

x  [0,1/2]

x
1/2

x  [1/2,1]

2x-1 < y1 < x

x 2x-y1

1/2

Pr[|y2-x|>|y1-x|] = 1 + 2y1 - 2x

2x < y1 < 1

x  [1/2,1] x < y1 < 1

2x-y1
x y1

Case 2.3:

y1

y1

Pr[|y2-x|>|y1-x|] = 1 + 2x - 2y1

Case 1.3:

1/2
Pr[|y2-x|>|y1-x|] = y1

Case 1.2:

Pr[|y2-x|>|y1-x|] = 1 + 2y1 - 2x

0 < y1 < 2x - 1

x  [1/2,1]

y1

Pr[|y2-x|>|y1-x|] = 1 - y1

1/2

Pr[|y2-x|>|y1-x|] = 1 + 2x - 2y1

2x
P (1 |y1 )(1 + 2x 2y1 )n1 dy1

+
x

1
P (1 |y1 )(1 y1 )

n1

dy1 dx

2x

1
+

 2x1

P (x|2 )n
P (1 |y1 )y1n1 dy1
0

1/2

x
P (1 |y1 )(1 + 2y1 2x)n1 dy1

+
2x1

1
P (1 |y1 )(1 + 2x 2y1 )

n1

dy1 dx.

Our density and posterior are given as p(x|2 ) = 2(1 x) and P (1 |y) = y for
x [0, 1] and y [0, 1]. We substitute these forms into the above large integral
and nd
 x
1/2
Pn (e)

2n(1 x)

=
0

y1 (1 + 2y1 2x)n1 dy1

0

2x
y1 (1 + 2x 2y1 )n1 dy1
x

1
y1 (1 y1 )

n1

2x

dy1 dx

PROBLEM SOLUTIONS

151
1

 2x1

2n(1 x)
y1n dy1
0

1/2

x
y1 (1 + 2y1 2x)n1 dy1
2x1

1
y1 (1 + 2x 2y1 )

n1

dy1 dx.

There are two integrals we must do twice with dierent bounds. The rst is:
b
y1 (1 + 2y1 2x)n1 dy1 .
a

We dene the function u(y1 ) = 1 + 2y1 2x, and thus y1 = (u + 2x 1)/2 and
dy1 = du/2. Then the integral is
b
y(1 + 2y1 2x)

n1

dy1

1
4

u(b)


(u + 2x 1)un1 du
u(a)


u(b)
1 2x 1 n
1
n+1
u +
u
.
4
n
n+1
u(a)

The second general integral is:

b
y1 (1 + 2x 2y1 )n1 dy1 .
a

We dene the function u(y1 ) = 1 + 2x 2y1 , and thus y1 = (1 + 2x u)/2 and

dy1 = du/2. Then the integral is
b
y1 (1 + 2x 2y1 )

n1

dy1

1
=
4

u(b)


(1 + 2x + u)un1 du
u(a)

u(b)

1 2x + 1 n
1
u
un+1
.
4
n
n+1
u(a)

We use these general forms to evaluate three of the six components of our full
integral for Pn (e):
x
y1 (1 + 2y1 2x)n1 dy1

12x=u(0)

1 2x 1 n
1
u
un+1
4
n
n+1
1=u(x)




1
1 2x + 1
1
1
1
n+1
+
+ (1 2x)

4
n
n+1
4
n n+1

CHAPTER 4. NONPARAMETRIC TECHNIQUES

152


12x=u(2x)
1 2x + 1 n
1
u
un+1
4
n
n+1
1=u(x)




1
1 2x + 1
1
1
1
1
n
n+1

(1 2x) + (1 2x)
+
4
n
n+1
2n
4
n n+1
12x
12x


1 n
1
n1
n+1
u
u
(1 u)u
du =
n
n+1
0

2x
y1 (1 + 2x 2y1 )n1 dy1

=
1
y1 (1 y1 )n1 dy1

2x

1
1
(1 2x)n
(1 2x)n+1 .
n
n+1

We add these three integrals together and nd, after a straightforward calculation that the sum is


x
1
1
1
n
+
(1 2x) +

(1 2x)n+1 .
()
n 2n
2n n + 1
We next turn to the three remaining three of the six components of our full
integral for Pn (e):
2x1


y1n dy1

x

y1 (1 + 2y1 2x)

n1

dy1

2x1

=
1
y1 (1 + 2x 2y1 )n1 dy1

1
(2x 1)n+1
n+1

1=u(x)

1 2x 1 n
1
n+1
u +
u
4
n
n+1
2x1=u(2x1)




1
1 2x 1
1
1
1
n+1
+
(2x 1)
+
4
n
n+1
4
n n+1
2x1=u(1)

1 2x + 1 n
1

u
un+1
4
n
n+1
1=u(x)




1
1 2x + 1
1
1
1
1
n
n+1

(2x 1) (2x 1)

4
n
n+1
2n
4
n n+1

The sum of these latter three terms is



x
1
1
1

(2x 1)n

(2x 1)n+1 .
n 2n
2n n + 1
Thus the sum of the three Case 1 terms is
 x
1/2
2n(1 x)
y1 (1 + 2y1 2x)n1 dy1
0

2x
y1 (1 + 2x 2y1 )n1 dy1

+
x

1


y1 (1 y1 )n1 dy1 dx

+
2x

()

PROBLEM SOLUTIONS
=

153




1/2
1
1
1
x
n
n+1
+
(1 2x) + (1 2x)

dx
2n(1 x)
n 2n
2n n + 1
0

1
=

n
0

1+u
2



1u
1 n
+
u +
2n
2n

1
1

2n n + 1


u

n+1

du,

where we used the substitution u = 1 2x, and thus x = (1 u)/2, and

dx = du/2 and 1 x = (1 + u)/2.
Likewise, we have for the three Case 2 terms
1

 2x1

2n(1 x)
y1n dy1
0

1/2

x
y1 (1 + 2y1 2x)n1 dy1

+
2x1

1
y1 (1 + 2x 2y1 )

n1

+


1
2n(1 x)

dy1 dx

x
1

(2x 1)n (2x 1)n+1

n 2n

1
1

2n n + 1


dx

1/2

1
=

n
0

1u
2



1+u
1 n

u
2n
2n

1
1

2n n + 1


u

n+1

du,

where we used the substitution u = 2x1, and thus x = (1+u)/2 and dx = du/2
and 1 x = (1 u)/2.
Now we are ready to put all these results together:


1
Pn (e)

n
0

1
+

n
0

1 
= n
0

1+u
2



1u
2

1u
1 n
+
u +
2n
2n



1+u
1 n

u
2n
2n

1 n+1
1
+
(1 u2 ) +
u
2n
2n

1
1

2n n + 1


1
1

2n n + 1

1
1

2n n + 1


un+1 du



un+1 du


u

n+2

du





1
1
1
u3
1
1
1
n+2
n+3
= n
u
u

u
+
2n
3 2n(n + 2
n + 3 2n n + 1
0



1
1
1n
1 2
+
+
= n
2n 3 2n(n + 2) n + 3 2n(n + 1)
1 (n + 1)(n + 3) (n 1)(n + 2)
=
+
3
2(n + 1)(n + 2)(n + 3)

CHAPTER 4. NONPARAMETRIC TECHNIQUES

154
=
=

1
3n + 5
+
3 2(n + 1)(n + 2)(n + 3)
1
1
1
+
+
,
3 (n + 1)(n + 3) 2(n + 2)(n + 3)

which decays as 1/n2 , as shown in the gure.

Pn(e)
0.5
0.4
1/3
0.2
2

10

We may check this result for the case n = 1 where there is only one sample. Of
course, the error in that case is P1 (e) = 1/2, since the true label on the test
point may either match or mismatch that of the single sample point with equal
probability. The above formula above conrms this
P1 (e)

=
=

1
1
1
+
+
3 (1 + 1)(1 + 3) 2(1 + 2)(1 + 3)
1 1
1
1
+ +
= .
3 8 24
2

(e) The limit for innite data is simply

lim Pn (e) =

1
,
3

which is larger than the Bayes error, as indeed it must be. In fact, this solution
also illustrates the bounds of Eq. 52 in the text:
P
1
4

P P (2 2P )
1
3

.
3
8

14. We assume P (1 ) = P (2 ) = 0.5 and the distributions are as given in the gure.
p(x|i)
1

0
0

x1

x*

0.5

x2

x
1

(a) This is a somewhat unusual problem in that the Bayes decision can be any point
1/3 x 2/3. For simplicity, we can take x = 1/3. Then the Bayes error is

PROBLEM SOLUTIONS

155

then simply
P

1

min[P (1 )p(x|1 ), P (2 )p(x|2 )]dx

0

2/3
P (1 )p(x|1 )dx
1/3

= 0.5(1/3)(3/2) = 0.25.
(b) The shaded area in the gure shows the possible (and equally likely) values of
a point x1 chosen from p(x|1 ) and a point x2 chosen from p(x|2 ).
x2

2/3

2
1/3
0

x1

1/3

2/3

There are two functionally separate cases, as numbered, corresponding to the

position of the decision boundary x = (x1 + x2 )/2. (Note that we will also
have to consider which is larger, x1 or x2 . We now turn to the decision rule and
probability of error in the single nearest-neighbor classier in these two cases:
case 1 : x2 x1 and 1/3 (x1 + x2 )/2 2/3: Here the decision point x
is between 1/3 and 2/3, with R2 at large values of x. This is just the
Bayes case described in part (a) and the error rate is thus 0.25, as we saw.
The relative probability of case 2 occuring is the relative area of the gray
region, that is, 7/8.
case 2 : x1 x2 and 1/3 (x1 + x2 )/2 2/3: Here the decision boundary
is between 1/3 and 2/3 (in the Bayes region) but note especially that R1
is for large values of x, that is, the decision is the opposite of the Bayes
decision. Thus the error is 1.0 minus the Bayes error, or 0.75. The relative
probability of case 2 occuring is the relative area of the gray region, that
is, 1/8.
We calculate the average error rate in the case of one point from each category
by merely adding the probability of occurrence of each of the three cases (proportional to the area in the gure), times the expected error given that case,
that is,
P1 =

7
1
5
0.25 + 0.75 =
= 0.3125,
8
8
16

which is of course greater than the Bayes error.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

156

(c) Problem not yet solved

(d) Problem not yet solved
(e) In the limit n , every test point x in the range 0 x 1/3 will be properly
classied as 1 and every point in the range 2/3 x 1 will be properly
classied as 2 . Test points in the range 1/3 x 2/3 will be misclassied
half of the time, of course. Thus the expected error in the n case is
P

= P (1 )Pr[0 x 1/3|1 ] 0 + P (1 )Pr[1/3 x 2/3|1 ] 0.5

+P (2 )Pr[1/3 x 2/3|2 ] 0.5 + P (2 )Pr[2/3 x 1|2 ] 0
= 0.5 0.5 0.5 + 0.5 0.5 0.5 = 0.25.

Note that this is the same as the Bayes rate. This problem is closely related to
the zero information case, where the posterior probabilities of the two categories are equal over a range of x. If the problem specied that the distributions
were equal throughout the full range of x, then the Bayes error and the P errors
would equal 0.5.
15. An faster version of Algorithm 3 in the text deletes prototypes as follows:
Algorithm 0 (Faster nearest-neighbor)
1
2
3
4
5
6
7
8
9

begin initialize j 0, D, n = number of prototypes

Construct the full Voronoi diagram of D
do j j + 1 (for each prototype xj )
if xj is not marked then nd the Voronoi neighbors of xj
if any neighbor is not from thexj class then mark xj and its neighbors in other classes
until j = n
Discard all unmarked prototypes
returnVoronoi diagram of the remaining (marked) prototypes
end

If we have k Voronoi neighbors on average of any point xj , then the probability that
i out of these k neighbors are not from the same class as xj is given by the binomial
law:
 
k
P (i) =
(1 1/c)k (1/c)k1 ,
i
where we have assumed that all the classes have the same prior probability. Then the
expected number of neighbors of any point xj belonging to dierent class is
E(i) = k(1 1/c).
Since each time we nd a prototype to delete we will remove k(11/c) more prototypes
on average, we will be able to speed up the search by a factor k(1 1/c).
16. Consider Algorithm 3 in the text.
(a) In the gure, the training points (black for 1 , white for 2 ) are constrained
to the intersections of a two-dimensional grid. Note that prototype f does not
contribute to the class boundary due to the existence of points e and d. Hence
f should be removed from the set of prototypes by the editing algorithm (Algorithm 3 in the text). However, this algorithm detects that f has a prototype

PROBLEM SOLUTIONS

157

from another class (prototype c) as a neighbor, and thus f is retained according

to step 5 in the algorithm. Consequently, the editing algorithm does not only
select the points a, b, c, d and e which are are the minimum set of points, but
it also retains the useless prototype f .

R1

R2

f
l

(b) A sequential editing algorithm in which each point is considered in turn and
retained or rejected before the next point is considered is as follows:
Algorithm 0 (Sequential editing)
1
2
3
4
5
6
7

begin initialize j 0, D, n = number of prototypes

do j j + 1
Remove xj from D
if xj is not well classied by D, then restore xj to D
until J = n
return D
end
This sequential editing algorithm picks up one training sample and checks
whether it can be correctly classied with the remaining prototypes. If an error
is detected, the prototype under consideration must be kept, since it may contribute to the class boundaries. Otherwise it may be removed. This procedure
is repeated for all the training patterns.
Given a set of training points, the solution computed by the sequential editing
algorithm is not unique, since it clearly depends upon the order in which the data
are presented. This can be seen in the following simple example in the gure,
where black points are in 1 , and white points in 2 . If d is the rst point
b

presented to the editing algorithm, then it will be removed, since its nearestneighbor is f so d can be correctly classied. Then, the other points are kept,
except e, which can be removed since f is also its nearest neighbor. Suppose
now that the rst point to be considered is f . Then, f will be removed since
d is its nearest neighbor. However, point e will not be deleted once f has been

CHAPTER 4. NONPARAMETRIC TECHNIQUES

158

removed, due to c, which will be its nearest neighbor. According to the above
considerations, the algorithm will return points a, b, c, f in the rst ordering,
and will return points a, b, c, d, e for the second ordering.
17. We have that P (i ) = 1/c and p(x|i ) = p(x) for i = 1, . . . , c. Thus the
probability density of nding x is
p(x) =

c
!

p(x|i )P (i ) =

i=1

c
!

p(x|)

i=1

1
= p(x|),
c

and accordingly
p(x|)1/c
1
p(x|i )P (i )
=
= .
p(x)
p(x|)
c

P (i |x) =

()

We use () in Eq. 45 in the text to nd that the asymptotic nearest-neighbor error

rate is
P

lim Pn (e)

 
c
!
2
=
1
P (i |x) p(x)dx
n

 

i=1


c
!
1
=
1
p(x)dx
c2
i=1


1
1
=
1
p(x)dx = 1 .
c
c
On the other hand, the Bayes error is

P =
P (error|x)p(x)dx
=

c 
!
i=1R

[1 P (i |x)]p(x)dx,

()

where Ri denotes the Bayes decision regions for class i . We now substitute () into
() and nd



c  
!
1
1
1

P =
1
p(x)dx = 1
p(x)dx = 1 .
c
c
c
i=1
Ri

c
Thus we have P = P , which agrees with the upper bound P (2 c1
P ) = 1 1/c
in this no information case.
18. The probability of error of a classier, and the k-nearest-neighbor classier in
particular, is

P (e) = p(e|x)dx.

In a two-class case, where P (1 |x) + P (2 |x) = 1, we have

P (e|x) = P (e|x, 1 )P (1 |x) + P (e|x, 2 )P (2 |x).

PROBLEM SOLUTIONS

159

The probability of error given a pattern x which belongs to class 1 can be computed
as the probability that the number of nearest neighbors which belong to 1 (which
we denote by a) is less than k/2 for k odd. Thus we have
P (e|x, 1 ) = P r[a (k 1)/2].
The above probability is the sum of the probabilities of nding a = 0 to a = (k 1)/2,
that is,
!

(k1)/2

Pr[a (k 1)/2] =

Pr[a = i].

i=0

The probability of nding i nearest prototopyes which belong to i among

  k is
P (1 |x)i P (2 |x)ki multiplied by the number of possible combinations, ki ; that
is,
(k1)/2 

P (e|x, 1 ) =

!
i=0


k
P (1 |x)i P (2 |x)ki .
i

By a simple interchange 1 2 , we nd
(k1)/2 

P (e|x, 2 ) =

!
i=0


k
P (2 |x)i P (1 |x)ki .
i

We put these results together and nd

(k1)/2 

P (e|x) =

!
i=0



k 
P (1 |x)i+1 [1 P (1 |x)]ki + P (1 |x)ki [1 P (1 |x)]i+1 .
i

Recall that the conditional Bayes error rate is

P (e|x) = min[P (1 |x), P (2 |x)].
Hence we can write P (e|x) as
(k1)/2 

P (e|x)

!
i=0



k 
P (e|x)i+1 [1 P (e|x))]ki + P (e|x)ki [1 P (e|x)]i+1
i

= fk [P (e|x)],
where we have dened fk to show the dependency upon k. This is Eq. 54 in the text.
Now we denote Ck [P (e|x)] as the smallest concave function of P (e|x) greater
than fk [P (e|x)], that is,
fk [P (e|x)] Ck [P (e|x)].
Since this is true at any x, we can integrate and nd



P (e) = fk [P (e|x)]p(x)dx Ck [P (e|x)]p(x)dx.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

160

Jensens inequality here states that under very broad conditions, Ex [u(x)] u[E[x]].
We apply Jensens inequality to the expression for P (e) and nd




P (e) Ck [P (e|x)]p(x)dx Ck
P (e|x)p(x)dx = Ck [P ],
where again P is the Bayes error. In short, we see P (e) Ck [P ], and from above
fk [P ] Ck [P ].
Section 4.6
19. We must show that the use of the distance measure
;
< d
<!
D(a, b) = =
k (ak bk )2
k=1

for k > 0 generates a metric space. This happens if and only if D(a, b) is a metric.
Thus we rst check whether D obeys the properties of a metric. First, D must be
non-negative, that is, D(a, b) 0, which indeed holds because the sum of squares
is non-negative. Next we consider reexivity, that is, D(a, b) = 0 if and only if
a = b. Clearly, D = 0 if and only if each term in the sum vanishes, and thus
indeed a = b. Conversely, if a = b, then D = 0. Next we consider symmetry, that
is, D(a, b) = D(b, a) for all a and b. We can alter the order of the elements in
the squared sum without aecting its value, that is, (ak bk )2 = (bk ak )2 . Thus
symmetry holds. The last property is the triangle inequality, that is,
D(a, b) + D(b, c) D(a, c)
for all a, b, and c. Note that D can be written
D(a, b) = A(a b)2
where A = diag[1 , . . . , d ]. We let x = a c and y = c b, and then the triangle
inequality test can be written
Ax2 + Ay2 A(x + y)2 .
We denote x = Ax and y = Ay, as the feature vectors computed using the matrix
A from the input space. Then, the above inequality gives
x 2 + y 2 x + y 2 .
We can square both sides of this equation and see
(x 2 + y 2 ) = x 22 + y 22 + 2x 2 y 2 x 22 + y 22 + 2
2

d
!
i=1

which is equivalent to the simple test

x 2 y 2

d
!
i=1

xi yi .

xi yi = x + y 2 ,

PROBLEM SOLUTIONS

161

Observe that the above inequality if fullled, since the Cauchy-Schwarz inequality
states that

 d

!



 
x 2 y 2 
xi yi .


i=1

If we work with metric spaces in the nearest-neighbor method, we can ensure the
existence of best approximations in this space. In other words, there will always be a
stored prototype p of minimum distance from an input pattern x . Let
= min D(p, x),
x

where x belongs to a set S of the metric space which is compact. (A subset S is said
to be compact if every sequence of points in S has a sequence that converges to a
point in S.) Suppse we dene a sequence of points {x1 , x2 , . . . xn } with the property
D(p, xn ) as n ,
where xn x using the compactness of S. By the triangle inequality, we have that
D(p, xn ) + D(xn , x ) D(p, x ).
The right-hand side of the inequality does not depend on n, and the left side approches
as n . Thus we have D(p, x ). Nevertheless, we also have D(p, x )
because x belongs to S. Hence we have that D(p, x ) = .
The import of this property for nearest-neighbor classiers is that given enough
prototypes, we can nd a prototype p very close to an input pattern x . Then, a good
approximation of the posterior class probability of the input pattern can be achieved
with that of the stored prototype, that is, p  x implies
P (i |p)  P (i |x ).
20. We must prove for
Lk (a, b) =

1 d
!

21/k
|ai bi |

= a bk

i=1

the four properties a metric are obeyed.

The rst states that Lk must be positive denite. Since |ai bi | 0 for all i,
Lk (a, b) 0.
The second states that Lk (a, b) = 0 if and only if a = b. If a = b, then each
term of the sum |ai bi | is 0, and Lk (a, b) = 0. Conversely, if Lk (a, b) = 0,
then each term |ai bi | must be 0, and thus ai = bi for all i, that is, a = b.
The third condition is symmetry, that is Lk (a, b) = Lk (b, a). This follows
directly from the fact
|ai bi | = | (bi ai )| = | 1| |bi ai | = |bi ai | for i = 1, . . . , d.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

162

The last property is the triangle inequality, that is

Lk (a, b) + Lk (b, c) Lk (a, c)
or
a bk + b ck a ck

()

for arbitrary a, b and c. We dene a b = x and b c = y, then () can be

written
xk + yk

x + yk .

()

We exponentiate both sides and nd an equivalent condition

k

(xk + yk )

(x + yk )

( )

for k > 1. We expand the left-hand side of ( ) to nd

k

(xk + yk ) =

d
!

|xi | +
k

d
!

i=1

|yi | +
k

i=1

k1
!
j=1


k
xkj
yjk ,
k
j

and the right-hand side to nd

x + ykk =

d
!

|xi |k +

i=1

d
!

|bi |k +

k1
!

i=1

j=1

 d
k !
|xi |kj |bi |j .
j i=1

In short, then ( ) can be written as

xkj
yjk
k

d
!

|xi |kj |bi |j

j = 1, . . . , k 1.

i=1

Note that
xkj

d
!

|xi |kj

and yjk

i=1

because

d
!

|yi |j

i=1

"! # q !
a
aq

for a > 0.

Then we have for the above special case

21 d
2
1 d
!
!
kj
j
xk yk
|xi |kj
|yi |j .
i=1

Since

1
!
i

i=1

21
2
!
!
ai
bi
ai bi
i

for ai , bi > 0,

PROBLEM SOLUTIONS
we have
xkj
yjk
k

163

1 d
!

|xi |

kj

21 d
!

i=1

2
|yi |

i=1

d
!

|xi |kj |bi |j .

i=1

We replace x = a b and y = b c and conclude

Lk (a, b) + Lk (b, c) Lk (a, c),
and hence the triangle inequality is obeyed.
21. Problem not yet solved
22. Problem not yet solved
23. Problem not yet solved
24. Problem not yet solved
25. Consider the computational complexity of calculating distances under dierent
metrics.
(a) The Euclidean distance between two prototypes is given by
x x 2 =

d
!

(xi xi )2 ,

i=1

where x denotes the transformed image, x is a stored prototype, and d is

the number of pixels in the handwritten digit. As we can see, d subtractions,
d multiplications and d additions are needed. Consequently, this is an O(d)
process.
(b) Given a text sample x and a prototype x , we must compute r non-linear transforms, which depend on a set of parameters subject to optimization. If an iterative algorithm is used for optimizing these parameters, we will have to compute
the Euclidean distance of the transformed image and the stored prototype x for
each step. Since the computation of the transformed image invlves r non-linear
transformations (with ai k 2 operations per transform), the number of operations
required is
ai k 2 r + 3k 2 = (ai r + 3)k 2 .
(c) For each prototype, we will have to perform A searches, and this implies (ai r + 3)k 2 A
operations. Accordingly, the total number of operations is
(ai r + 3)k 2 An.
(d) If n = 106 , r = 6, ai  10, A  5, and basic operations on our computer require
109 second, then the classication of a single point is performed in
(10 6 + 3)k 2 5 106 operations

109 seconds
= 0.315k 2 seconds.
operation

Suppose that k = 64. Then a test sample will be classied in 21 minutes and
30 seconds.

CHAPTER 4. NONPARAMETRIC TECHNIQUES

164

26. Explore the eect of r on the accuracy of nearest-neighbor search based on partial
distance.
(a) We make some simplications, specically that points are uniformly distributed
in a d-dimensional unit hypercube. The closest point to a given test point in, on
average, enclosed in a small hypercube whose volume is 1/n times the volume of
the full space. The length of a side of such a hypercube is thus 1/n1/d . Consider
a single dimension. The probability that any point falls within the projection
of the hypercube is thus 1/n1/d . Now in r d dimensions, the probability of
falling inside the projection of the small hypercube onto r-dimensional subspace
is 1/nr/d , i.e., the product of r indpendendent events, each having probability
1/n1/d . So, for each point the probability it will not fall inside the volume in
the r-dimensional space is 1 1/nr/d . Since we have n indpendent such points,
the probability that the partial distance in r dimensions will give us the true
nearest neighbor is

P r[nearest neighbor in ddim. is found in r < d dim.] =

1
nr/d

n
,

as shown in the gure for 0 < r d = 10 and n = 5, 10 and 100.

P
1

100

n=5

0.8
0.6
0.4

n=10

0.2
2

10

(b) The Bayes decision boundary in one dimenion is clearly x1 = 0.5. In two
dimensions it is the line x2 = 1 x1 . In three dimension the boundardy occurs
when
x1 x2 x3 = (1 x1 )(1 x2 )(1 x3 )
which implies

x3 =

(x1 1) (x2 1)
,
1 x1 x2 + 2x1 x2

as shown in the gure.

(c) The Bayes error in one dimension is clearly
1
EB

Min[P (1 )p(x|1 ), P (2 )p(x|2 )]p(x) dx

0

PROBLEM SOLUTIONS
d=1

165
d=2

d=3
x3

x2

0.5

0.8

R2

0.8

R2
R2

0.2

0.6

0
1

R1

0.4

x2

0.6

x1

R1

R1

0.4

x1

0
0

0.2

0
0.2
0.4
0.6
0.8
1

x1

x=0.5

2 0.5

0.1875.

(1 x) dx
0

The Bayes error in two dimensions is

x
2 =1x
 1

1
EB

2 0.5

0.104167.

dx2 (1 x1 )(1 x2 )

dx1

x1 =0

The Bayes error in three dimensions is

1
EB

1 x3 (x
 1 ,x2 )
dx1 dx2
dx3 (1 x1 )(1 x2 )(1 x3 )

= 0.0551969,
where x3 (x1 , x2 ) is the position of the decision boundary, found by solving
x1 x2 x3 = (1 x1 )(1 x2 )(1 x3 )
for x3 , that is,
x3 (x1 , x2 ) =

(x1 1) (x2 1)
.
1 x1 x2 + 2x1 x2

The general decision boundary in d dimensions comes from solving

d
,
i=1

xi =

d
,

(1 xi )

i=1

or
xd

d1
,
i=1

xi = (1 xd )

d1
,
i=1

(1 xi )

CHAPTER 4. NONPARAMETRIC TECHNIQUES

166

for xd , which has solution

d1
/

xd (x1 , x2 , . . . , xd1 ) =

d1
/
j=1

i=1

(1 xi )

(1 xj ) +

d1
/
k=1

.
xk

The Bayes error is

1
EB =

1
dx2

dx1
0

,...,xd1 )
x
d (x1

1

dxd1

dxd



d
,

(1 xi ).

j=1

d integrals

We rst compute the integral(x[1] - 1)) over xd , since it has the other variables
implicit. This integral I is
x
,...,xd1 )
d (x1

(1 xd ) dxd

I=

= xd 0 d

1  2 xd
x
2 d 0

1
= xd (xd )2 .
2
Substituting xd from above we have

d1
 d1
d1
d1
/
/
/
/
(1 xi )
(1 xj ) +
xk 1/2
(1 xi )
I

i=1

j=1

k=1

d1
/
j=1

1/2
=

d1
/
i=1

(1 xi )2 +

d1
/
j=1

(1 xj ) +

d1
/
i=1

(1 xj ) +

d1
/
k=1

2

i=1

xk

xi (1 xi )

d1
/
k=1

2

xk

(d) Problem not yet solved

(e) Problem not yet solved
27. Consider the Tanimoto metric described by Eq. 58 in the text.
(a) The Tanimoto metric measures the distance between two sets A and B, according to
DT animoto (A, B) =

na + nb 2nab
na + nb nab

where na and nb are the sizes of the two sets, and nab is the number of elements
common to both sets.

PROBLEM SOLUTIONS

167

We must check whether the four properties of a metric are always fullled. The
rst property is non-negativity, that is,
DT animoto (A, B) =

na + nb 2nab
0.
na + nb nab

Since the sum of the elements in A and B is always greater than their common
elements, we can write
na + nb nab > 0.
Furthermore, the term na + nb 2nab gives account of the number of dierent
elements in sets A and B, and thus
na + nb 2nab 0.
Consequently, DT animoto (A, B) 0.
The second property, reexivity, states that
DT animoto (A, B) = 0

if and only if A = B.

From the denition of the Tanimoto measure above, we see that DT animoto (A, B)
will be 0 if and only if na + nb 2nab = 0. This numerator is the number of
dierent elements in A and B, so it will yield 0 only when A and B have no
dierent elements, that is, when A = B. Conversely, if A = B, then na + nb
2nab = 0, and hence DT animoto (A, B) = 0.
The third property is symmetry, or
DT animoto (A, B) = DT animoto (B, A)
for all A and B. Since the terms na and nb appear in the numerator and
denominator in the same way, only the term nab can aect the fullment of this
property. However, nab and nba give the same measure: the number of elements
common to both A and B. Thus the Tanimoto metric is indeed symmetric.
The nal property is the triangle inequality:
DT animoto (A, B) + DT animoto (B, C) DT animoto (A, C).
We substitute () into the denition of the Tanimoto metric and nd
na + nb 2nab
nb + nc 2nbc
na + nc 2nac
+

.
na + nb nab
nb + nc nbc
na + nc nac
After some simple manipulation, the inequality can be expressed as
3BDE 2ADE 2BCE + ACE 2BDF + ADF + BCF 0,
where
A = na + n b
B

= nab

= nb + nc

D
E

= nbc
= na + nc

= nac .

()

CHAPTER 4. NONPARAMETRIC TECHNIQUES

168
We factor to nd

E[C(A 2B) + D(3B 2A)] + F [D(A 2B) + BC] 0.

Since A 2B is the number of dierent elements in A and B and hence nonnegative, the triangle inequality is fullled if some of the below conditions are
met:
3B 2A = 3nab 2(na + nb ) 0

()
( )

D = nbc = 0
C(A 2B) + D(3B 2A) 0.
This last condition can be rewritten after some arrangment as
DT

nbc
.
nb + nc nbc

On the other hand, if we take other common factors in the above equation, we
have
B[E(3D 2C) + F (C 2D)] + A[E(C 2D) + DF ] 0.
The term C 2D is the number of dierent elements in C and D, so it is greater
than zero. Accordingly, the triangle inequality can be also fullled if some of
the below additional conditions are met:
3D 2C = 3nbc 2(nb + nc ) 0
E = nb + n a = 0
E(3D 2C) + F (C 2D) 0.
After some rearrangements, this last inequality can be written as
DT animoto (B, C)

nb + nc
.
nbc (nb + nc )

Since the denominator of the right-hand side is always negative, the inequality
is met. Thus we can conclude that the Tanimoto metric also obeys the triangle
inequality.
(b) Since the Tanimoto metric is symmetric, we consider here only 15 out of the 30
possible pairings.
A
pattern
pattern
pattern
pattern
pattern
pat
pat
pat
pat
pots
pots
pots
stop
stop
taxonomy

B
pat
pots
stop
taxonomy
elementary
pots
stop
taxonomy
elementary
stop
taxonomy
elementary
taxonomy
elementary
elementary

na
7
7
7
7
7
3
3
3
3
4
4
4
4
4
8

nb
3
4
4
8
10
4
4
8
10
4
8
10
8
10
10

nab
3
2
2
3
5
2
2
2
2
4
2
1
2
1
5

DT animoto (a, b)
0.57
0.77
0.77
0.75
0.58
0.6
0.6
0.77
0.81
0
0.8
0.92
0.8
0.92
0.61

rank
2
7
7
6
3
4
4
7
9
1
8
9
8
9
5

PROBLEM SOLUTIONS

169

(c) Yes, the Tanimoto metric fullls the triangle inequality for all possible sets given
above.
Section 4.7
28. As shown in Fig. 23 in the text, there is a big dierence between categories
of a feature and the (true) categories or classes for the patterns. For this problem
there is a fuzzy feature that we can call temperature, and a class named hot.
Given a pattern, e.g., a cup, we can observe a degree of membership of this object
onto the fuzzy feature temperature, for instance, temperature = warm. However
temperature = warm does not necessarily imply that the membership to the hot
class is less than 1.0 due to we could argue that the fuzzy feature temperature is
not very hot but warm. We could reasonably suppose that if the hot class has
something to do with temperature then it would have into account the fuzzy feature
temperature for its membership function. Accordingly, values of temperature
such as warm or hot might be active in some degree for the class hot though
the exact dependence between them is unknown for us without more information
aboute the problem.
29. Consider fuzzy classication.
(a) We rst t the designers subjective feelings into fuzzy features, as suggested by
the below table.
Feature
lightness

designers feelings
medium-light
medium-dark
dark
short
long

length

fuzzy feature
light
dark
dark
short
large

If we use Min[a, b] as the conjunction rule between fuzzy sets a and b, then the
discriminant functions can be written as:
d1
d2
d1

= Min[
c(length, 13, 5), c(lightness, 70, 30)]
= Min[
c (length, 5, 5), c (lightness, 30, 30)]
= Min[
c(length, 13, 5), c (lightness, 30, 30)]

where

c(x, , )

c (x, , )

1 x>
1 + (x )/ x

0 otherwise

0 x>+

1 + ( x)/ x +
=

1 otherwise.
=

(b) If every category membership function were rescaled by a constant , the

discriminant functions would be
d1

Min[
c(length, 13, 5),
c(lightness, 70, 30)]

CHAPTER 4. NONPARAMETRIC TECHNIQUES

170
d2
d3

=
=

Min[
c(length, 5, 5),
c (lightness, 30, 30)]
Min[
c(length, 13, 5),
c (lightness, 30, 30)]

Since Min[f1 , f2 ] = Min[f1 , f2 ], the above functions can be written

d1

= Min[
c(length, 13, 5), c(lightness, 70, 30)]

d2

= Min[
c (length, 13, 5), c (lightness, 70, 30)]
= Min[
c(length, 13, 5), c (lightness, 70, 30)].

d3

Hence, the classication borders would be unaected, since all the discriminant
functions would be rescaled by the same constant , and we would have
arg max(d1 , d2 , d3 ) = arg max(d1 , d2 , d3 ) = arg max(d1 , d2 , d3 ).
di

di

di

(c) Given the pattern x = (7.5, 60)t and the above equations, the value of the
discriminant functions can be computed as follows:
d1

Min[
c(7.5, 13, 5), c(60, 70, 30)] = Min[0, 0.66] = 0

d2

=
=

Min[
c (7.5, 5, 5), c (60, 30, 30)] = Min[0.5, 0] = 0
Min[
c(7.5, 13, 5), c (60, 30, 30)] = Min[0, 0] = 0.

d3

Since all the discriminant functions are equal to zero, the classier cannot assign
the input pattern to any of the existing classes.
(d) In this problem we deal with a handcrafted classier. The designer has selected
lightness and length as features that characterize the problem and has devined
several membership functions for them without any theoretical or empirical
justication. Moreover, the conjunction rule that fuses membership functions
of the features to dene the true discriminant functions is also imposed without
any principle. Consequently, we cannot know where the sources of error come
from.
30. Problem not yet solved
Section 4.8
31. If all the radii have been reduced to a value less than m , Algorithm 4 in the text
can be rewritten as:
Algorithm 0 (Modied RCE)
1
2
3
4
5
6
7
8
9

begin initialize j 0, n number of patterns,  small parameter

do j j + 1
wji xji for i = 1, . . . , d
arg min D(x, xj )
x
xi

j D(
x, xj )
j D(
x, xj ) 
if xj k , then ajk 1 for k = 1, . . . c
until j = n
end

PROBLEM SOLUTIONS

171

According to Algorithm 5 in the text, the decision boundaries of the RCE classier
depend on j ; an input pattern x is assigned to a class if and only if all its nearest prototypes belong to the same class, where these prototypes fulll the condition
is the nearest training sample to wj beD(x, wj ), j . Since j = D(
x, wj ), where x
longing to another class, we will have dierent class boundaries when we use dierent
training points.
Section 4.9
32. We seek a Taylor series expansion for


n
1 !
x xi
pn (x) =
.

nhn i=1
hn
(a) If the Parzen window is of the form (x) N (0, 1), or
2
1
p(x) = ex /2 ,
2

we can then write the estimate as



n
1 !
x x1
pn (x) =

nhn i=1
hn

2 

n
1 ! 1
1 x x1
exp
=
nhn i=1 2
2
hn



n
1
1!
1 x2
x2i
2xxi
=
exp
+

2 h2n
h2n
h2n
2hn n i=1
$
%
2




n
exp 12 hx2 1 !
1 x2i
xxi
n

=
exp 2 .
exp
h2n
2 hn
2hn n i=1
We express this result in terms of the normalized variables u = x/hn and ui =
xi /hn as
eu /2 1 ! u2i /2 uui
pn (x) =
e
e .
2hn n i=1
n

We expand exxi /hn = euui about the origin,

euui =

!
uj uj
i

j!

j=0

and therefore our density can be written as

=
=

eu /2 1 ! ! uj uji u2i /2

e
2hn n i=1 j=0 j!
1 n
2
2
2

eu /2 ! 1 ! uji eui /2

uj
j!
2hn j=0 n i=1
2

pn (x)

CHAPTER 4. NONPARAMETRIC TECHNIQUES

172

=
=

eu /2 !

bj u j
2hn j=0

lim pnm (x),

where we have dened

1 ! 1 j u2i /2
.
u e
n i=1 j! i
n

bj =

Thus we can express pnm (x) as the m-term approximation as

2
m1
eu /2 !
pnm (x) =
bj u j .
2hn j=0

(b) If the n samples are extremely tightly clustered about u = uo (i.e., ui  uo ) for
i = 1, . . . , n, then the two-term approximation is
eu /2
pn2 (x) =
(bo + b1 u),
2hn
2

where
2
1 ! u2i /2
e
 euo /2
n i=1

bo =
and

2
2
1!
ui eui /2  uo euo /2 .
n i=1

b1 =

The peak of pn2 (x) is given by the solution of u pn2 = 0, that is

ueu /2 (bo + b1 u) b1 eu /2

+
= 0.
2hn
2hn
2

u pn2 (x)

This equation then gives

0

= u(bo + b1 u) + b1
= u(euo /2 + uo euo /2 u) uo euo /2
2

= u2 uo + u uo = 0
u
1.
= u2 +
uo
Thus the peaks of pn2 (x) are the solutions of u2 + u/uo 1 = 0.
(c) We have from part (b) that u is the solution to u2 + u/uo 1 = 0, and thus the
solution is

1/uo 1/u2o + 4
u=
.
2

PROBLEM SOLUTIONS

173

In the case uo  1, the square root can be approximated as

and thus

1 1 + 4u2o
u =
2uo
1 (1 + 12 4u2o )

,
2uo

1/u20 + 4  1/u0 ,

and this implies that one of the peaks of the probability is at

u

1 + (1 + 2u2o )
= uo .
2uo

In the case uo  1, we have

u

1/uo

1/u2o + 4

2
1/uo 2 1/(4u2o ) + 1
2
1/uo 2(1 + 12 1/(4u2o ))
.
2

=


Thus, one of the peaks occurs at

u

1/uo + 2(1 + 1/(8u2o ))

2
 = 1.
2
2

(d) In these conditions, we nd

eu /2
pn2 (u) =
(bo + b1 u)
2hn
2

where
bo  euo /2 ,

and b1  uo euo /2 ,

as is shown in the gure. (The curves have been rescaled so that their structure
is apparent.) Indeed, as our calculation above showed, for small uo , pn2 (u)
peaks near u = 0, and for large uo , it peaks near u = 1.
~pn2(u)

uo

0.4

=1

0.3

uo
=1

uo
=

0.2

.01

0.1

u
1

CHAPTER 4. NONPARAMETRIC TECHNIQUES

174

Computer Exercises
Section 4.2
1. Computer exercise not yet solved
Section 4.3
2. Computer exercise not yet solved
Section 4.4
3. Computer exercise not yet solved
Section 4.5
4. xxx
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43

MATLAB program
x = [0.58 0.27 0.0055 0.53 0.47 0.69 0.55 0.61 0.49 0.054];
mu = mean(x);
delta = std(x);
mu arr = linspace(min(x), max(x), 100);
delta arr = linspace(2*(min(x)-mu), 8*(max(x)-mu), 100);
[M, D] = meshgrid(mu arr, delta arr);
p D theta = zeros(size(M));
for i=1:size(M,2)
for j=1:size(M,2)
x minus mu = abs(x - M(i,j));
if max(x minus mu) > D(i,j)
p D theta(i,j) = 0;
else;
a = (D(i,j) - x minus mu)/D(i,j)^2;
p D theta(i,j) = prod(a);
end
end
end
p D theta = p D theta./(sum(p D theta(:))/...
((mu arr(2) - mu arr(1))*(delta arr(2)-delta arr(1))));
p theta D = p D theta;
ind = find(p D theta>0);
x vect = linspace(min(x)-10*delta, max(x)+10*delta, 100);
post = zeros(size(x vect));
for i=1:length(x vect)
for k=1:length(ind)
if abs(x vect(i) - M(ind(k))) < D(ind(k)),
p x theta = (D(ind(k)) - abs(x vect(i) - M(ind(k))))./D(ind(k)).^2;
else
p x theta = 0;
end
post(i) = post(i) + p x theta*p theta D(ind(k));

COMPUTER EXERCISES
44
45
46
47
48
49
50
51
52
53

end
end
post = post./sum(post);
figure;
plot(x vect, post, -b);
hold on;
plot([mu mu], [min(post) max(post)], -r);
plot(x, zeros(size(x)), *m);
grid;
hold off;
[[FIGURE TO BE INSERTED]]

5. Computer exercise not yet solved

Section 4.6
6. Computer exercise not yet solved
7. Computer exercise not yet solved
8. Computer exercise not yet solved
Section 4.7
9. Computer exercise not yet solved
Section 4.8
10. Computer exercise not yet solved
Section 4.9
11. Computer exercise not yet solved

175

176

CHAPTER 4. NONPARAMETRIC TECHNIQUES

Chapter 5

Linear discriminant functions

Problem Solutions
Section 5.2
1. Consider the problem of linear discriminants with unimodal and multimodal distributions.
(a) The gure shows two bimodal distributions that obey reective symmetry about
a vertical axis. Thus their densities have the same value along this vertical
line, and as a result this is the Bayesian decision boundary, giving minimum
classication error.
R1

R2

(b) Of course, if two unimodal distributions overlap signicantly, the Bayes error will
be high. As shown in Figs 2.14 and 2.5 in Chapter 2, the (unimodal) Gaussian
case generally has quadratic (rather than linear) Bayes decision boundaries.
Moreover, the two unimodal distributions shown in the gure below would be
better classied using a curved boundary than the straight one shown.
(c) Suppose we have Gaussians of dierent variances, 12 = 22 and for deniteness
(and without loss of generality) 2 > 1 . Then at large distances from the
means, the density P (2 |x) > P (1 |x). Moreover, the position of the mean of
1 , that is, 1 , will always be categorized as 1 . No (single) straight line can
separate the position of 1 from all distant points; thus, the optimal decision
177

178

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

R1

R2

1
2

boundary cannot be a straight line. In fact, for the case shown, the optimal
boundary is a circle.
R2

R1

1
2
R2

2. Consider a linear machine for c categories with discriminant functions gi (x) =

wit x + i0 for i = 1, . . . , c. Our goal is to show that the decision regions are convex,
that is, if x(1) and x(0) are in i , then x(1) + (1 )x(0) is also in i , for all
0 1. The fact that x(1) and x(0) are in i implies
max gj (x(1)) = gi (x(1)) = wit x(1) + wi0
j

and
max gj (x(0)) = gi (x(0)) = wit x(0) + wi0 .
j

For any j and for 0 1, the above imply

wjt [x(1) + (1 )x(0)] + wj0

= [wjt x(1) + wj0 ] + (1 )[wjt x(0) + wj0 ]

[wit x(1) + wi0 ] + (1 )[wit x(0) + wi0 ].

We next determine the category of a point between x(0) and x(1) by nding the
maximum discriminant function:


max wjt [x(1) + (1 )x(0)] + wj0 = max[wj t x(1) + wj0 ] + (1 ) max[wj t x(1) + wj0 ]
j

=
=

[wit x(1)
wit [x(1)

+ wi0 ] + (1
+ (1 )x(0)] + wi0 .

)[wit x(1)

+ wi0 ]

This shows that the point x(1) + (1 )x(0) is in i . Thus any point between x(0)
and x(1) is also in category i ; and since this is true for any two points in i , the
region is convex. Furthermore, it is true for any category j all that is required in
the proof is that x(1) and x(0) be in the category region of interest, j . Thus every
decision region is convex.
3. A counterexample will serve as proof that the decision
  regions need not be convex.
In the example below, there are c = 5 classes and 2c = 10 hyperplanes separating

PROBLEM SOLUTIONS

179

the pairs of classes; here Hij means the hyperplane separating class i from class j .
To show the non-convexity of the solution retion for class 1 , we take points A and
B, as shown in the gure.
H14 H24
H15

H13

H12
H35

C
2

5
A

H25

H
H23 34

H45

The table below summarizes the voting results. Each row lists a hyperplane Hij ,
used to disginguish category i from j . Each entry under a point states the category
decision according to that hyperplane. For instance, hyperplane H12 would place A in
2 , point B in 1 , and point C in 2 . The underlining indicates the winning votes in
the full classier system. Thus point A gets three 1 votes, only one 2 vote, two 3
votes, and so forth. The bottom row shows the nal voting result of the full classier
system. Thus points A and B are classied as category 1 , while point C (on the
line between A and B) is category 2 . Thus the solution region for 1 is not convex.
hyperplane
H12
H13
H14
H15
H23
H24
H25
H34
H35
H45
Vote result:

A
2
1
1
1
3
2
5
3
5
4
1

B
1
1
1
1
2
2
2
4
3
4
1

C
2
3
1
1
2
2
2
3
5
4
2

4. Consider the hyperplane used for discriminant functions.

(a) In order to minimize x xa 2 subject to the constraint g(x) = 0, we form the
objective function
f (x, ) = x xa 2 + 2[g(x)],
where is a Lagrange undetermined multiplier; the point xa is xed while x
varies on the hyperplane. We expand f () and nd
f (x, )

= x xa 2 + 2[wt x + w0 ]
= (x xa )t (x xa ) + 2(wt x + w0 )
= xt x 2xt xa + xta xa + 2(xt w + w0 ).

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

180

We set the derivatives of f (x, ) to zero and nd

f (x, )
x
f (x, )

= x xa + w = 0
= wt x + w0 = 0,

and these equations imply

x = xa w,
and
wt x + w0

= wt (xa w) + w0
= wt xa + w0 wt w
= 0.

These can be solved for as

wt w = wt xa + w0
or (so long as w = 0)
=

wt xa + w0
.
wt w

The value of x is then

x = xa w

%
$ t
xa +w0
xa w w
w
tw
=
xa

if w = 0
if w = 0.

The minimum value of the distance is obtained by substituting the value of x

into the distance:
>
>
 t

>
>
w xa + w0
>
>
x xa  = >xa
w xa >
t
>
>
ww
>
>

> wt x + w
>
>
>
a
0
= >
w
>
t
>
>
ww
=

|g(xa )|w
|g(xa )|
=
.
w2
w

(b) From part (a) we have that the minimum distance from the hyperplane to the
point x is attained uniquely at the point x = xa w on the hyperplane. Thus
the projection of xa onto the hyperplane is
xo

= xa w
g(xa )
= xa
w,
w2

where
=

wt xa + w0
g(xa )
=
.
wt w
w2

PROBLEM SOLUTIONS

181

5. Consider the three category linear machine with discriminant functions gi (x) =
wit x wio , for i = 1, 2, 3.
(a) The decision boundaries are obtained when wit x = wjt x or equivalently (wi
wj )t x = 0. Thus, the relevant decision boundaries are the lines through the
origin that are orthogonal to the vectors w1 w2 , w1 w3 and w2 w3 , and
illustrated by the lines marked i j in the gure.

x2

w1
w2

w3
-4

-2

x1

-2

-4

2
1

x2

(b) If a constant vector c is added to the weight vectors, w1 , w2 and w3 , then the
decision boundaries are given by [(wi + c) (wj + c)]t x = (wi wj )t x = 0,
just as in part (a); thus the decision boundaries do not change. Of course, the
triangle formed by joining the heads of w1 , w2 , and w3 changes, as shown in
the gure.
w1

+
1

w2

w3
w2 + c
x1
-4

-2

w3 + c

-2

c
-4

6. We rst show that totally linearly separable samples must be linearly separable.
For samples that are totally linearly separable, for each i there exists a hyperplane
gi (x) = 0 that separates i samples from the rest, that is,
gi (x) 0

for x i

gi (x) < 0

for x  i .

The set of discriminant functions gi (x) form a linear machine that classies correctly;
the samples are linearly separable and each category can be separated from each of
the others.
We shall show that by an example that, conversely, linearly separable samples
need not be totally linearly separable. Consider three categories in two dimensions

182

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

specied by
1

= {x : x1 < 1}

= {x : 1 x1 < 1}

= {x : x1 > 1}.

Clearly, these categories are linearly separable by vertical lines, but not totally linearly
separable since there can be no linear function g2 (x) that separates the category 2
samples from the rest.
7. Consider the four categories in two dimensions corresponding to the quadrants of
the Euclidean plane:
1
2
3
4

= {x : x1 > 0, x2 > 0}
= {x : x1 > 0, x2 < 0}
= {x : x1 < 0, x2 < 0}
= {x : x1 < 0, x2 > 0}.

Clearly, these are pairwise linearly separable but not totally linearly separable.
8. We employ proof by contradiction. Suppose there were two distinct minimum
length solution vectors a1 and a2 with at1 y > b and at2 y > b. Then necessarily we
would have a1  = a2  (otherwise the longer of the two vectors would not be a
minimum length solution). Next consider the average vector ao = 12 (a1 + a2 ). We
note that
ato yi =

1
1
1
(a1 + a2 )t yi = at1 yi + at2 yi b,
2
2
2

and thus ao is indeed a solution vector. Its length is

ao  = 1/2(a1 + a2 ) = 1/2a1 + a2  1/2(a1  + a2 ) = a1  = a2 ,
where we used the triangle inequality for the Euclidean metric. Thus ao is a solution
vector such that ao  a1  = a2 . But by our hypothesis, a1 and a2 are minimum
length solution vectors. Thus we must have ao  = a1  = a2 , and thus
1
a1 + a2 
2

= a1  = a2 .

We square both sides of this equation and nd

1
a1 + a2 2 = a1 2
4
or
1
(a1 2 + a2 2 + 2at1 a2 ) = a1 2 .
4
We regroup and nd
0

= a1 2 + a2 2 2at1 a2
= a1 a2 2 ,

and thus a1 = a2 , contradicting our hypothesis. Therefore, the minimum-length

solution vector is unique.

PROBLEM SOLUTIONS

183

9. Let the two sets of vectors be S1 = {x1 , . . . , xn } and S2 = {y1 , . . . , ym }. We

assume S1 and S2 are linearly seperable, that is, there exists a linear discriminant
function g(x) = wt x + w0 such that
g(x) > 0 implies x S1

and

g(x) < 0 implies x S2 .

Consider a point x in the convex hull of S1 , or
x=

n
!

i xi ,

i=1

where the i s are non-negative and sum to 1. The discriminant function evaluated
at x is
g(x) = wt x + w0
1 n
2
!
t
= w
i xi + w0
=

n
!

i=1

i (wt xi + w0 )

i=1

> 0,
where we used the fact that wt xi + w0 > 0 for 1 i n and

n

i=1

i = 1.

Now let us assume that our point x is also in the convex hull of S2 , or
x=

m
!

j yj ,

j=1

where the j s are non-negative and sum to 1. We follow the approach immediately
above and nd
g(x)

= wt x + w0

m
!
= wt
j yj + w0
j=1

m
!

j (wt yj + w0 )



j=1
g(yj )<0

< 0,
where the last step comes from the realization that g(yj ) = wt yj + w0 < 0 for each
yi , since they are each in S2 . Thus, we have a contradiction: g(x) > 0 and g(x) < 0,
and hence clearly the intersection is empty. In short, either two sets of vectors are
either linearly separable or their convex hulls intersects.
10. Consider a piecewise linear machine.
(a) The discriminiant functions have the form
gi (x) =

max gij (x),

j=1,...,ni

184

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

where the components are linear
t
gij (x) = wij
x + wij0 ,

and the decision rule is to classify x in category i , if

max gk (x) = gi (x),
k

for j = 1, . . . , c. We can expand this discriminant function as

max gk (x) = max max gkj (x),
k

j=1,...,nk

t
where gkj (x) = wkj
x+wkjo , are linear functions. Thus our decision rule implies

max max gkj (x) = max gij (x) = gij(i) (x).

k

j=1,...,nk

1jni

Therefore, the piecewise linear machine can be viewed as a linear machine for
classifying subclasses of patterns, as follows: Classify x in category i if
max max gkj (x) = gij (x).
j=1,...,nk

(b) Consider the following two categories in one dimension:

1

= {x : |x| > 1},

= {x : |x| < 1}

which clearly are not linearly separable. However, a classier that is a piecewise
linear machine with the following discriminant functions
g11 (x)
g12 (x)
g1 (x)
g2 (x)

1x
1+x

=
=
=

max g1j (x)

j=1,2

can indeed classify the patterns, as shown in the gure.

g(x)
3

g11(x)

g12(x)
2

g2(x)

x
-2

-1

-1

11. We denote the number of non-zero components in a vector by b. We let the d

components of x be either 0 or 1 and the categories given by
1
2

= {x : b is odd}
= {x : b is even}.

PROBLEM SOLUTIONS

185

(a) Our goal is to show that 1 and 2 are not linearly separable if d > 1. We prove
this by contradiction. Suppose that there exists a linear discriminant function
g(x) = wt x + w0
such that
g(x) 0 for x 1 and
g(x) < 0 for x 2 .
Consider the unique point having b = 0, i.e., x = 0 = (0, 0, . . . , 0)t . This pattern
is in 2 and hence clearly g(0) = wt 0 + w0 = w0 < 0. Next consider patterns
in 1 having b = 1. Such patterns are of the general form
x = (0, . . . , 0, 1, 0, . . . , 0),
and thus
g(x) = wi + w0 > 0
for any i = 1, . . . , d. Next consider patterns with b = 2, for instance
x = (0, . . . , 0, 1, 0, . . . , 0, 1, 0, . . . , 0).
Because such patterns are in 2 we have
g(x) = wi + wj + w0 < 0
for i = j.
We summarize our results up to here as three equations:
w0

< 0

wi + w0

> 0

wi + wj + w0

< 0,

for i = j. The rst two of these imply

(wi + w0 ) + (wj + w0 ) + (w0 ) > 0,
        
>0

>0

>0

or

wi + wj + w0

> 0,

which contradicts the third equation. Thus our premise is false, and we have
proven that 1 , 2 are not linearly separable if d > 1.
(b) Consider the discriminant functions
g1 (x)

max g1j (x) and

g2 (x)

max g2j (x),

j
j

where
gij (x)

= ij (1, . . . , 1)t x + wijo

= ij b + wijo ,

186

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

and, as in part (a), we let b denote the number of non-zero components of x.
We can set
1j

= j + 1/2

2j

= j

wijo

= j 2 j 1/4

w2jo

= j 2 ,

for j = 0, 1, . . . , d + 1 and hence write our discriminant function as

g1 (x)
g2 (x)

= (j + 1/2)b j 2 j 1/4
= jb j 2 .

We now verify that these discriminant functions indeed solve the problem. Suppose x is in 1 ; then the number of non-zero components of x is b = 2m + 1, 0
m (d 1)/2 for m an integer. The discriminant function is
g1j (x)

= (j + 1/2) (2m + 1) j 2 j 1/4

= j(2m + 1) j 2 j + 1/2(2m + 1) 1/4
= j(2m) j 2 + 1/2(2m + 1) 1/4
= j(2m j) + 1/2(2m + 1) 1/4.

For patterns in 2 , we have b = 2m, and thus

g2j (x)

= j(2m + 1) j 2
= j(2m + 1 j).

It is a simple matter to verify that j(2m j) is maximized at j = m and that

j(2m + 1 j) is maximized at j = m + 1/2. But note that j is an integer and
the maximum occurs at j = m and j = m + 1. Thus we have
g1 (x)

max g1j (x)

max[j(2m j)] + 1/2(2m + 1) 1/4

j
j

= m(2m m) + m + 1/2 1/4

= m2 + m + 1/4
= m(m + 1) + 1/4, and
g2 (x)

max g2j (x)

max j(2m + 1 j)

j
j

= m(m + 1).
Thus if x 1 and the number of non-zero components of x is b = 2m + 1 (i.e.,
odd), we have
g1 (x) = m(m + 1) + 1/4 > m(m + 1) = g2 (x),
that is, g1 (x) > g2 (x), as desired. Conversely, if x is in 2 , then the number
of non-zero components of x is b = 2m, 0 m d/2 where m is an integer.

PROBLEM SOLUTIONS

187

Following the above logic, we have

g1j (x)

= (j + 1/2)(2m) j 2 j 1/4
= j(2m 1) j 2 + m 1/4
= j(2m j 1) + m 1/4

which is maximized at j = m. Likewise, we have

g2j (x)

= j(2m) j 2
= j(2m j)

which is maximized at j = m. Thus our discriminant functions are

g1 (x)

max g1j (x)

max j(2m 1 j) + m 1/4

j
j

= m(m 1) + m 1/4 = m2 1/4,

g2 (x)

max g2j (x) = max j(2m j) = m2 .

j

Indeed, if x is in 2 and the number of non-zero components of x is b = 2m

(i.e., even), we have
g1 (x)

= m2 1/4 < g2 (x) = m2 ,

or g1 (x) < g2 (x), and hence our piecewise linear machine solves the problem.
Section 5.3
12. Consider the quadratic discriminant function given by Eq. 4 in the text:
g1 (x) g2 (x) = g(x) = w0 +

d
!

wi xi +

i=1

d !
d
!

wij xi xj ,

i=1 i=1

which for convenience we write in vector and matrix form as

g(x) = w0 + wt x + xt Wx.
In this two-category case, the decision boundary is the set of x for which g(x) = 0.
If we did not have the wt x term, we could easily describe the boundary in the form
xW0 x = k where k is some constant. We can, in fact, eliminate the wt x terms with
a proper translation of axes. We seek a translation, described by a vector m, that
eliminates the linear term. Thus we have
g(x m)

= w0 + wt (x m) + (x m)t W(x m)

()

= (w0 w m + m Wm) + (w m W mW)x + x Wx,

t

where we used the fact that W is symmetric (i.e., Wij = Wji ) and that taking the
transpose of a scalar does not change its value. We can solve for m by setting the
coecients of x to zero, that is,
0 = wt mt W mW = wt 2mt W.

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

188

This implies mt W = wt /2 or m = W1 wt /2.

Now we substitute this translation vector into the right-hand side of () and nd
g(x m)

= w0 wt (W1 w/2) + wt W1 /2 W W1 w/2 + xt Wx

= w0 wt W1 w/4 + xt Wx.

()

The decision boundary is given by g(x m) = 0, and then () implies

xt

W
x = xt Wx = 1/4,
wt W1 w 4w0

and thus the matrix

W=

W
wt W1 w 4w0

shows the basic character of the decision boundary.

(a) If W = I, the identity matrix, then xt Ix = k/4 is the equation of the decision
d

boundary, where k is some constant. Then we have
x2i = k/4, which denes
i=1

a hypersphere of radius k/2.

(b) Since W is non-singular and symmetric, we can write W in terms of t ,
where is the matrix formed by the (linearly independent) eigenvectors of W
and is the diagonal matrix formed by the eigenvalues of W. In this way,
xt Wx = 1/4 canbe written as xt t x = 1/4. We let yt = xt ; then we get
yt y = 1/4, or i i yi2 = 1/4 where the eigenvalues i > 0 when W is positive
denite, which it will be for non-pathological data distributions. Because the
eigenvalues are positive and not necessarily all equal to each other, in this case
the decision boundary is a hyperellipsoid.
(c) With an argument following that of part (b), we can represent the characteristics
d

of the decision boundary with
i yi2 = 1/4. Since, in this case we have at
i=1

least one i < 0, the geometric shape dened is a hyperhyperboloid.

(d) Our matrix and vector in this

1 2
W= 2 5
0 1
Thus we have

W1
and

16
= 6
2

6
3
1

case are

0
1 ;
3

5
w = 2 .
3

2
1
1 ;
4
1

0.0121
W
W=
= 0.0242
82.75
0

0.0242
0.0604
0.0121

wt W1 w = 82.75

0
0.0121
0.0363

The eigenvalues are {1 , 2 , 3 } = {0.0026, 0.0716, 0.0379}. Because there is

a negative eigenvalue, the decision boundary is a hyperhyperboloid.

PROBLEM SOLUTIONS
(e) Our matrix and vector in this

1 2
W= 2 0
3 4

189
case are

3
4 ;
5

2
w = 1 .
3

Thus we have

0.3077 0.4231
0.1538
W1 = 0.4231 0.2692 0.0385 ;
0.1538
0.0385 0.0769
and

0.4407
W
W=
= 0.8814
2.2692
1.3220

0.8814
0
1.7627

wt W1 w = 2.2692

1.3220
1.7627
2.2034

The eigenvalues are {1 , 2 , 3 } = {0.6091, 2.1869, 3.3405}. Because of a negtative eigenvalue, the decision boundary is a hyperhyperboloid.
Section 5.4
13. We use a second-order Taylor series expansion of the criterion function at the
point a(k):
1
J(a)  J(a(k)) + J t (a(k))(a a(k)) + (a a(k))t H(a(k))(a a(k)),
2
where J(a(k)) is the derivative and H(a(k)) the Hessian matrix evaluated at the
current point.
The update rule is
a(k + 1) = a(k) (k)J(a(k)).
We use the expansion to write the criterion function as
1
J(a(k + 1)  J(a(k)) + J t (a(k)[(k)J(a(k)] + [(k)J t (a(k))]H(a(k))[(k)J(a(k))].
2
We minimize this with respect to (k) and nd
0 = J t (a(k))J(a(k)) + (k)J t (a(k))H(a(k))J(a(k))
which has solution
(k) =

J(a(k))2
.
J t (a(k))H(a(k))J(a(k))

We must, however, verify that this is indeed a minimum, and that (k) > 0. We
calculate the second derivative
2J
= (J)t H(J).
2

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

190

If a is in a neighborhood of a local minimum of J(a), the H will be positive denite

and hence (J)t H(J) > 0. Thus J indeed attains minima at the optimal (k) and
(k) > 0 because (J)t HJ > 0.
14. We are given samples from 1 :
 
 
 
1
2
5
,
,
5
9
3
and from 2 :


2
,
3


1
,
4

 
0
.
2

Augmenting the samples with an extra dimension, and inverting the sign of the samples from 2 dene

1
1
5
b

1
b

2
9

a1
1 5 3

a = a2 b = b .
Y=

1 2
b
3

a3

1
b
1
4
1
0 2
b
Then the sum-of-squares error criterion is
2
1 ! t
(Ya b) (Ya b)
a yi b =
.
2 i=1
2
n

Js (a) =

(a) We dierentiate the criterion function and nd

6a1
a2 +
Js (a) = Y t (Ya b) = a1 + 35a2 +
6a1 + 36a2 +
and

6
H = Yt Y = 1
6

6a3
36a3
144a3

+
0
+ 3b
16b

1
6
35 36 .
36 144

Notice that for this criterion, the Hessian is not a function of a since we assume
a quadratic form throughout.
(b) The optimal step size varies with a, that is,
=

[Js (a)]t Js (a)

.
[Js (a)]t HJs (a)

The range of optimal step sizes, however, will vary from the inverse of the largest
eigenvalue of H to the inverse of the smallest eigenvalue of H. By solving the
characteristic equation or using singular value decomposition, the eigenvalues of
H are found to be 5.417, 24.57, and 155.0, so 0.006452 0.1846.

PROBLEM SOLUTIONS

191

Section 5.5
15. We consider the Perceptron procedure (Algorithm 5.3 and Theorem 5.1 in the
text).
(a) Equation 20 in the text gives the update rule
a(1)
a(k + 1)

= arbitrary
= a(k) + yk ,

k 1.

It has been shown in the convergence proof in the text that

a(k + 1)
a2 a(k)
a2 2 + 2 ,
is a solution vector, is a scale factor and 2
where a(k) is the kth iterate, a
and are specied by Eqs. 21 and 22 in the text:
2

max yi 2

t yi > 0.
min a

Thus we have
a(k + 1)
a2 a(k)
a2 (2 2 ).
Because > 2 /, we have the following inequality:
2 2 2 2 2 = 2 > 0.
Thus a(k)
a2 decreases (strictly) by an amount (2 2 ) at each iteration,
and this implies
a(k + 1)
a2 a(1)
a2 k(2 )2 .
We denote the maximum number of correction convergence by ko , we have
0 a(k)
a2 a(1)
a2 k(2 2 ),
and thus
k

a(1)
a2
,
2 2

and thus the maximum number of corrections is

ko =

a(1)
a2
.
2 2

(b) If we start out with a zero weight vector, that is, a(1) = 0, we have
ko

=
=
=
=

a(1)
a 2
2 2

a2
2 2
a 2
2 
2 2
2
2 2

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

192

In order to nd the maximum of k0 with respect to , we set

ko ()

ko ()

to 0,

a2 2 
a2 2
(2 2 )2
(2 2 )2

= 2 (4
a2 2
a2 ) 2 2 
a 2
=
and this implies

0,



2
a2 2 2 
a2 = 0.

We note that > 2 / > 0, which minimizes ko if a(1) = 0. We assume b 0,

and therefore
= min[
at yi ] > b 0,
i

and > 0. We next use the fact that 0 < a k b < , and k 1 to nd
a(k + 1)
a2 a(k)
a2 + b2 2 + 2b b 2a .
We choose the scale factor to be
=

b2 2 + 2b b
a

and nd
a(k + 1)
a2 a(k)
a2 (b2 2 + 2b b),
where b2 2 + 2b b > 0. This means the dierence between the weight vector at
iteration k + 1 and
a obeys:
a(k + 1)
a2 a(1)
a2 k(b2 2 + 2b b).
Since a squared distance cannot become negative, it follows that the procedure
must converge after almost ko iterations where
ko =

a(1)
a2
.
b2 2 + 2b b

In the case b < 0 there is no guarantee that > 0 and thus there is no guarantee
of convergence of the algorithm.
16. Our algorithm is given by Eq. 27 in the text:
a(1)
a(k + 1)

arbitrary

= a(k) + k yk

where in this case at (k)yk b, for all k. The k s are bounded by 0 < a k
b < , for k 1. We shall modify the proof in the text to prove convergence given
the above conditions.
be a solution vector. Then a
t yi > b, for all i. For a scale factor we have
Let a
a(k + 1)
a

= a(k) + k yk
a
= (a(k)
a) + k yk ,

PROBLEM SOLUTIONS

193

and thus
a(k + 1)
a2

= (a(k)
a) + k yk 2
= a(k)
a2 + k yk 2 + 2k (a(k)
a)t yk
= a(k)
a2 + k2 yk 2 + 2k atk yk 2k
at y k

a(k)
a2 + k2 yk 2 + 2k b 2k
at y k .

As yk was misclassied, at (k)yk b, for all k. Now we let

2

max yi 2 ,

min[
at yi ] > b,

is a solution vector. Thus we substitute into the above

since a
a(k + 1)
a2 a(k)
a2 + k2 2 + 2k b 2k ,
and convergence is assured.
17. If y1 , y2 , . . . , yn are linearly separable then by denition there exists a separating
hyperplane
g(x) = wt x + w0
such that
g(yi ) > 0 if yi 1
g(yi ) < 0 if yi 2 .
We augment every yi by 1, that is,
 
1
i =
y
,
yi

i = 1, . . . , n

and augment the weight vector by appending a bias weight w0 , written as

 
w0
=
w
.
w
Then our linear descriminant function can be written as
ty
i
w
t
y
i
w

> 0

if yi 1

< 0

if yi 2 .

i > 0
such that w
ty
i by 1 if yi 2 , and thus there exists a vector w
We multiply y
for all i.
which requires
We use the logic of the Perceptron convergence to nd such an w
at most ko operations for convergence, where
ko =

2
a(1) w
,
2

is the rst iterate and

where w

= max yi 2
=

t y
i > 0.
min a
i

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

194

The remainder follows from the Perceptron convergence Theorem (Theorem 5.1) in
the text.
18. Our criterion function is
!
Jq (a) =
(at y b)2 ,
yY(a)

where Y(a) is the set of samples for which at y b.

Suppose y1 is the only sample in Y(a(k)). Then we have
Jq (a(k))

= (at (k)y1 b)2 = (at (k)y1 )2 + b2 2bat (k)y1

2

d
d
!
!
=
akj y1j + b2 2b
akj y1j .
j=1

j=1

The derivative of the criterion function with respect to the components of a is

d
!
Jq (a(k))
= 2
akj y1j y1i 2by1i
aki
j=1
= 2(at y1 b)y1i ,
for i = 1, . . . , d. Thus we have
Jq (a(k)) =
Dii

=
=
=

Jq (a(k))
= 2(at y1 b)y1
a(k)
2 Jq (a(k))
aki aki

d
!

2
akj y1j y1i 2by1i
aki
j=1
2y1i y1i

for i, i = 1, . . . , d. This implies

D = 2y1 yit .
From Eq. 34 in the text, we have that the basic gradient descent algorithm is given
by
a(1)

= arbitrary

a(k + 1)

= a(k) + k

! b at (k)y
y1 ,
y2

yY(a(k))

and thus
Y(a(k))

= {y1 },

which implies
a(k + 1)

= a(k) + k

b at (k)y
y1 .
y2

PROBLEM SOLUTIONS

195

We also have
at (k + 1)y1 b = (1 k )(at (k)y1 b).
If we choose k = 1, then a(k + 1) is moved exactly to the hyperplane at y1 = b. Thus,
if we optimally choose k = 1, we have Eq. 35 in the text:
a(k + 1) = a(k) +

b at (k)y
y1 .
y2

19. We begin by following the central equation and proof of Perceptron Convergence
given in the text:
a(k + 1)
a2 = a(k)
a2 + 2(a(k)
a)t yk + 2 yk 2 ,
where yk is the kth training presentation. (We suppress the k dependence of for
clarity.) An update occurs when yk is misclassied, that is, when at (k)yk b. We
dene 2 = max yi 2 and = min[
at yi ] > b > 0. Then we can write
i

a(k + 1)
a2 a(k)
a2 + 2(b ) + 2 2 ,
where is a positive free parameter. We can dene
2
1m1
1 !
(k) =
(k) + b

k=1

at the mth update. Then we have

a(k + 1)
a2 a(k)
a2 2(k)

m1
!

(l) + 2 2 .

l=1

Summing all such equations for the mth update gives

a(k + 1) (k)
a2 a(0) (0)
a 2 2

m
!

(k)

k=1

k
!

(l) +

l=1

m
!

2 (k) 2 ,

k=1

or equivalently
a(k + 1) (k)
a2 a(0) (0)
a2 2

m
!

(k)(l) + 2

k,l=k

On the other hand, we are given that

m


lim 

k=1
m


2 = 0
(k)

k=1

and we note

m
!

k=1

22
(k)

m
!
k=1

2 (k) + 2

m
!
(kl)

(k)(l).

m
!
k=1

2 (k).

()

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

196
We also have
lim 

m


m
!

m
!

m
!

2 (k) + 2
(k)(l) 2
(k)(l) = 0
2
k=1
(kl)
(kl)
(k)

k=1

which implies

(k)(l)

(kl)

lim 1 
=0

2
m
m


(k)
2

m


k=1

or
2

m


(k)(l)

(kl)

lim 

m


2
(k)

= 1. ()

k=1

Now we can reconsider ():

a(k + 1) (k)
a 2


m
n

2
m
2 2 (k)(l)
(k)
(kl)


2 k=1


2
a(0) (0)
a 2
(k)
2 
m
n

.

k=1
(k)

k=1

(k)

k=1

But we know from () that the rst term in the brackets goes to 1 and the coecient
of 2 goes to 0, and all the (k) terms (and their sum) are positive. Moreover, the
M

corrections will never cease, since
(k) , as long as there are incorrectly
k=1

classied samples. But the distance term on the left cannot be negative, and thus we
conclude that the corrections must cease. This implies that all the samples will be
classied correctly.
Section 5.6
20. As shown in the gure, in this case the initial weight vector is marked 0, and
after the successive updates 1, 2, . . . 12, where the updates cease. The sequence of
presentations is y1 , y2 , y3 , y1 , y2 , y3 , y1 , y3 , y1 , y3 , y2 , and y1 .
Section 5.7
21. From Eq. 54 in the text we have the weight vector
w = nS1
w (m1 m2 ),
where w satises


1
n 1 n2
t
Sw + 2 (m1 m2 )(m1 m2 ) w = m1 m2 .
n
n

PROBLEM SOLUTIONS

197
2

|y |

a2

so
r e l u ti
g i on
on

b/|

y1

||
|y 1

b/|

y3

12
9

910

11

a1

8
7
4

y2

5
1

3
2

b/||
|

y 3|

We substitute w into the above and nd



1
n1 n 2
Sw + 2 (m1 m2 )(m1 m2 )t nS1
w (m1 m2 ) = m1 m2 ,
n
n
which in turn implies
$
%
n1 n 2
m1 m2 + 2 (m1 m2 )(m1 m2 )t S1
w (m1 m2 ) = m1 m2 .
n
Thus we have
(m1 m2 ) = m1 m2
where
=1+

n1 n2
(m1 m2 )t S1
w (m1 m2 ).
n2

This equation is valid for all vectors m1 and m2 , and therefore we have = 1, or
$
%1
n1 n2
= 1 + 2 (m1 m2 )t S1
.
w (m1 m2 )
n
22. We dene the discriminant function
g0 (x) = (21 11 )P (1 |x) (12 22 )P (2 |x).
Our goal is to show that the vector a that minimizes
!
!
Js (a) =
(at y (21 11 ))2 +
(at y + (12 22 ))2
yY1

yY2

is asymptotically equivalent to the vector a that minimizes


2
= [at y go (x)]2 p(x)dx,

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

198

as given in Eq. 57 in the text. We proceed by calculating the criterion function

! 
! 
2
2
Js (a) =
at y (21 11 ) +
at y (12 22 )
yY1

yY2

!
!




n1 1
n2 1
2
2
= n
at y (21 11 ) +
at y + (12 22 ) .
n n
n n

yY1

yY2

By the law of large numbers, as n we have with probability 1

lim

1 
J (a)
n s

= J  (a)





= P (1 )E1 (at y (21 11 ))2 + P (2 )E2 (at y + (12 22 ))2 .

We expand the terms involving expectations as


 t

 t
2
2
E1 (a y (21 11 ))
=
a y (21 11 ) p(x|1 )dx,



E2 (at y + (21 11 ))2 =
(at y + (12 22 ))2 p(x|2 )dx.
We substitute these into the above equations and nd that our criterion function is

 t
2
J  (a) =
a y (21 11 ) p(x|1 )dx

 t
2
+
a y + (12 22 ) p(x|2 )dx

 t
2
=
a y (21 11 ) p(x, 1 )dx

 t
2
+
a y + (12 22 ) p(x, 2 )dx

 t 2
=
a y [p(x, 1 ) + p(x, 2 )]dx

2 at y [(21 11 )p(x, 1 ) + (12 22 )p(x, 2 )] dx


+(21 11 )2 p(x, 1 )dx + (12 22 )2 p(x, 2 )dx


t
2
=
(a y) p(x)dx + 2 at yg0 (x)p(x)dx
+(21 11 )2 P (1 ) + (12 22 )2 P (2 ),
where we have used the fact that
p(x) = p(x, 1 ) + p(x, 2 )
and
(21 11 )p(x, 1 ) + (12 22 )p(x, 2 )

= (21 11 )p(x|1 ) + (12 22 )p(x|2 )

= g0 (x)p(x).

Furthermore, the probability of nding a particular category i is simply


P (i ) = p(x, i )dx,

PROBLEM SOLUTIONS

199

and thus our criterior function can be written


 t
2
J  (a) =
a y go (x) p(x)dx

+ [(21 11 )2 P (1 ) + (12 22 )2 P (2 ) go2 (x)]p(x)dx .



independent of a

Since the second term in the above expression is independent of a, minimizing J  (a)
with respect to a is equivalent to minimizing

2
= [at y g0 (x)]2 p(x)dx.
In summary, the vector a that minimizes Js (a) provides asymptotically a minimum
mean-square error approximation to the Bayes discriminant function g0 (x).
23. Given Eqs. 66 and 67 in the text,
a(k + 1) = a(k) + (k)[(k) at (k)yk ]yk
= E[yyt ]1 E[y]
a
we need to show
2 ] = 0
lim E[a(k) a

given the two conditions

(k)

k=1

2 (k) < .

k=1

 using Eq. 67 as
We write a a
2

2
a(k + 1) a

+ (k)((k) at (k)yk )yk 2

= a a
2 + 2 (k)((k) at (k)yk )yk 2
= a a
)t (k at (k)yk )yk .
+2(k)(a(k) a

We take the expected value of both sides and nd

2 ]
E[a(k + 1) a

2 ] + 2 (k) E[(k at (k)yk )yk ]

= E[a(k) a






1

2

) ((k) a (k)yk )yk ] .

+2(k) E[(a(k) a



t

3

Note that 1 0 and 2 0. If we can show that 3 < 0 for all k > M for some nite
M , then we will have proved our needed result. This is because that the update rule


2 ] for k > R for some nite R since
must reduce E[a(k + 1) a
2 (k) is nite
whereas

k=1

 ] 0, and thus
(k) is innite. We know that E[a(k + 1) a
2

k=1

2 ]
lim E[a(k + 1) a

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

200

must vanish.
Now we seek to show that indeed 3 < 0. We nd
3

)t ((k) at (k)y)yk ]
= E[(a(k) a
t
= E[a (k)(k)yk ] E[(at (k)yk )2 ] E[(k)
at yk ] + E[
at yk at yk ]
= E[(k)at (k)yk ] + E[
at yk at yk ] E[(at (k)yk )2 ] E[(at yk )2 ]
t yk ]
at yk )2 2at (k)yk a
= E[(at yk )2 + (
t yk )2 ]
= E[(at (k)yk a
t
t yk )2 ] 0.
= E[(a (k)yk a

In the above, we used the fact that y1 , y2 , . . . , yk is determined, so we can consider at (k) as a non-random variable. This enables us to write E[(k)at (k)yk ] as
].
E[at (k)yk ykt a
Thus we have proved our claim that 3 < 0, and hence the full proof is complete.
24. Our criterion function here is
Jm (a) = E[(at y z)2 ].
(a) We expand this criterion function as
Jm (a)

= E[(at y z)2 ]
= E[{(at y go (x)) (z go (x))}2 ]
= E[(at y go (x))2 2(at y go (x))(z go (x)) + (z go (x))2 ]
= E[(at y go (x))2 ] 2E[(at y go (x))(z go (x))] + E[(z go (x))2 ].

(b) We use the fact that E[z|x] = go (x) to note that

E[(at y go (x))(z go (x))] = E[(at y(x) go (x))(z go (x))].
This leads to
E{E[(at y go (x))(z go (x))|x]} = E{(at y go (x))E[(z go (x))|x]},
since conditioned on x, at y(x) go (x) is a constant. Thus we have
E{(at y go (x))[E(z|x) go (x)]} = 0
as E[z|x] = go (x). We put these results together to get
Jm (a) = E[(at y go (x))2 ] + E[(z go (x))2 ],



independent of a

where the second term in the expression does not depend on a. Thus the vector
a that minimizes Jm also minimizes E[(at y go (x))2 ].
25. We are given that
1
k+1
= k1 + yk2 .

PROBLEM SOLUTIONS

201

(a) We solve for k1 as

k1

1
2
= k1
+ yk1
1
2
2
= k2
+ yk2
+ yk1
2
= 11 + y12 + y22 + . . . + yk1

1 + 1
=

k1

i=1

yi2
,

and thus
k

1
k1


1 + 1

i=1

yi2

for 1 > 0 and 0 < a yi2 b.

(b) With these ranges we have
a(k 1)

k1
!

yi2 b(k 1),

i=1

which implies
k

1
k1


1 + 1

i=1

yi2

1
1 + a(k 1)
1
.
1 + b(k 1)

This is true for each k, so we can sum to nd

!
!
1
k

1 + b(k 1)
k

as

!
1
= and b > 0.
k

k=1

Moreover, we have
!

k2

12
2

[1 + a(k 1)]

L1 <

as

!
1
< and a > 0.
k2

k=1

Consequently, the learning rates obey

!
!
k and
k2 L < .
k

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

202

26. We rewrite the LMS rule using the following notation:

t

y1
b1
y2t
b2

Y= .
and
b = . .
..
..
ynt

bn

Then the LMS rule before Eq. 61 in the text is

a(1)
a(k + 1)

arbitrary
n
!
= a(k) + (k)
yi (bi atk yi ).
i=1

now reads
Note that the condition we are looking for the limiting vector a
n
!

bi ) = 0,
yi (yit a

i=1

or equivalently
n
!

yi (
at yi ) =

i=1

n
!

yi bi .

i=1

changes during an update:

Now we consider how the distance between a(k) and a
2
a(k + 1) a

> n
>2
>!
>
2
(1)

>
>
t
2 +
= a(k) a
y
(b

a
(k)y
)
>
>
i
i
i
>
k 2 > i=1



Ck

n
!

2(1)
)t
yi (bi at (k)yi )
(a(k) a
k
i=1



Dk

2 +
= a(k) a

(1)
2(1)
Dk ,
Ck +
k2
k

where for convenience we have dened Ck and Dk . Clearly Ck 0 as it is the sum of

non-negative values. Consider now Dk , which we expand as
Dk =

n
!
 t

t yi at (k)yi a
t yi b i
a (k)yi at (k)yi + at (k)yi bi + a
i=1

We can substitute

i=1

written
Dk

n


n
!
i=1

yi bi with

(at (k)yi )2

n

i=1

t yi , from our denition of a

. Then Dk can be
yi a

n
!
i=1

(
at (k)yi )2 + at (k)

n
!
i=1

t
yi b i + a

n
!
i=1

n
!


t yi at (k)yi
t yi + a
=
at yi )2 + at (k)yi a
(at (k)yi )2 (
i=1

yi at (k)yi

PROBLEM SOLUTIONS

203

and thus
n
!

Dk =

t yi )2 0.
(at (k)yi a

i=1

Adding the m update equations for the LMS rule we nd

2 = a(1) a
2 + 2 (1)
a(m + 1) a

m
!
Ck
k=1

k2

+ 2(1)

m
!
Dk
k=1

Now we take the limit m for both sides:

2
lim a(m + 1) a

2 + 2 (1)
= a(1) a

!
Ck
k=1

k2

+ 2(1)

2 + 2 (1)CL + 2(1)
= a(1) a

!
Dk
k=1

!
Dk
k=1

where if Ck is bounded we have

CL =

!
Ck
k=1

k2

< .

We also know that if Dk < 0

!
Dk
k=1

for all k. But Dk < 0 cannot be true for all k (except for nite occassions), otherwise
the right-hand side of the equation will be negative while the left-hand side will be
non-negative. Thus Dk must be arbitrarily close to zero for some choice N , for all
k > N . This tells us that
t yik
lim at (k)yik = a

for i = 1, 2, . . . , N .
To see this, we start with arbitrarily smally positive . Then, we know |Dk | <
for all k > N for some integer N . This implies
|Dk | =

n
!

t yi )2 ,
(at (k)yi a

i=1

t yi |
in particular |at (k)yi a
v(k)

will also be true. Now we let

n
!

yi (bi yit a(k))

i=1

n
!

yi (bi [
at yi


i ])

i=1

n
!

t yi ) yi
yi (bi a

i=1

n
!
i=1

yi


i

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

204

t yi |.
where i = |at (k)yi a
Since was arbitrarily close to zero and |i | ||, we have
lim v(k) = 0,

proving that a(k) satises the limit condition.

Section 5.9
27. The six points are shown in the gure, labeled by their category membership.
6

-2

-4

-6
-4

-2

(a) By inspecting the plot of the points, we see that 1 and 2 are not linearly
separable, that is, there is no line that can separate all the 1 points from all
the 2 points.
(b) Equation 85 in the text shows how much the e2 value is reduced from step k
to step k + 1. For faster convergence we want the maximum possible reduction.
We can nd an expression for the learning rate by solving the maxima problem
for the right-hand side of Eq. 85,

1
e(k)2 e(k + 1)2 = (1 )e+ (k)2 + 2 e+t (k)YY e+ (k).
4

()

We take the derivative with respect to of the right-hand side, set it to zero
and solve:
e+ (k)2 2e+ (k)2 + 2e+ (k)YYt e+ (k) = 0,
which gives the optimal learning rate
opt (k) =

e2
.
2 [e+ (k)2 e+ (k)YYt e+ (k)]

The maximum eigenvalue of Yt Y is max = 66.4787. Since all the values between 0 and 2/max ensure convergence to take the largest step in each iteration
we should choose as large as possible. That is, since 0 < < 2/66.4787, we
can choose opt = 0.0301 , where  is an arbitrarily small positive constant.
However, () is not suitable for xed learning rate Ho-Kashyap algorithms. To
get a xed optimal , be proceed as discussed in pages 254255 in the text: If

PROBLEM SOLUTIONS

205

we use the algorithm expressed in Eq. 90,

b(1) > 0 but otherwise arbitrary
a(1)
arbitrary
b(k + 1) = b(k) + (e(k) + |e(k)|)
a(k + 1) = a(k) + RYt |e(k)|,
where R is arbitrary positive denite. As discussed on page 255 of the text,
if R = I, then A = 2I 2 Yt Y will be positive denite, thereby ensuring
convergence if 0 < < 2/max , where max is the largest eigenvalue of Yt Y.
For our case, we have

1
1
2
1

2 4

6
6 4
1 3 1
t

6 44 10 .
Y=
1 2 4 and Y Y =

4 10 62
1
1
5
1 5
0
The eigenvalues of Yt Y are {4.5331, 40.9882, 66.4787}.
Section 5.10
28. The linear programming problem on pages 256257 in the text involved nding
min {t : t 0, at yi + t > bi for all i}.
Our goal here is to show that the resulting weight vector a minimizes the criterion
function
Jt (a) =

max (bi at yi ).

i:at yi bi

There are two cases to consider: linearly separable and non-linearly separable.
Case 1: Suppose the samples are linearly separable. Then there exists a vector, say
ao , such that
ato yi = bi .
Then clearly ato yi + t > bi , for any t > 0 and for all i. Thus we have for all i
0

min{t : t 0, at yi + t > bi }
min{t : t 0, at yi + t > bi } = 0.

Therefore, we have
min{t : t 0, at yi + t > bi } = 0,
and the resulting weight vector is ao . The fact that Jt (a) 0 for all a and Jt (ao ) = 0
implies
arg min Jt (a) = ao .
a

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

206

This proves that the a minimizes Jt (a) is the same as the one for solving the modied
variable problem.
Case 2: If the samples are not linearly separable, then there is no vector ao such that
ato yi = bi .
This means that for all i
min{t : t 0, at yi + t > bi }
t,a

min{t : t 0, t > bi at yi }

min{t : t 0, t > max(bi at yi )}

min{t : t >

min{ max
(bi at yi )}
t

min Jt (a).

t,a
t,a

t,a
a

max (bi at yi )}

i:at yi bi

i:a yi bi

Section 5.11
29. Given n patterns xk , in d-dimensional space, we associate zk where
If xk 1 , then zk = 1
If xk 2 , then zk = 1.
We also dene a mapping : Rd Rd+1 as
(x) = (xk , zk )
where
xk = arg min(xk x).
xk

In other words, returns the nearest-neighbor prototype xk . Then clearly (x) =

(x, 0) represents a separating hyperplane with normal
a = (0, 0, . . . , 0, 1)
  
d zeros

We verify this by considering any pattern in 1 . For such a pattern

at (xk ) = (0, 0, . . . , 0, 1)t (xk , zk ) = zk = +1.
Conversely, for any pattern in 2 , we have
at (xk ) = (0, 0, . . . , 0, 1)t (xk , zk ) = zk = 1.
Intuitively speaking, this construction tells us that if we can label the samples unambiguously, then we can always nd a mapping to transform class 1 and class 2 to
points into the half spaces.
30. The points are mapped to

1 :
(1, 2, 2, 2, 1, 1)t , (1, 2, 2, 2, 1, 1)t

2 :
(1, 2, 2, 2, 1, 1)t , (1, 2, 2, 2, 1, 1)t

PROBLEM SOLUTIONS

207
z2 = x12

z2 = 2 x2

g=-1
g=1

g=-1
2

2
z1 = 2 x1

z1 = 2 x1
-3

-2 - 2 -1

1 2
-1
- 2

-3

-2 - 2-1

1 2

g=-1

-2
-3

g=-1
g=1

z2 = x22
3

z2 = x12
3

g=-1
g=1
g=-1

g=1

-3

-2 - 2 -1

1
1 2

z1 = 2 x1
2

- 2
-3

-2

2
-1

z1 = 2 x1x2
3

-1
-2
-3

as shown in the gure.

The margins are not the same, simply because the real margin is the distance of the
support vectors to the optimal hyperplane in R6 space, and their projection to lower
dimensional subspaces does not necessarily preserve the margin.
31. The Support Vector Machine algorithm can be written:
Algorithm 0 (SVM)
1
2
3
4
5
6
7
8
9
10
11
12

begin initialize a; worst1 ; worst2 ; b

i0
do i i + 1
if zi = 1 and at yi zi < worst1, then worst1 at yi zi ; kworst1 k
if zi = 1 and at yi zi < worst2, then worst2 at yi zi ; kworst2 k
until i = n
a a + ykworst2 ykworst1
a0 at (ykworst2 + ykworst1 )/2
oldb b; b at ykworst1 /  a 
until | b oldb |< 
return a0 , a
end

Note that the algorithm picks the worst classied patterns from each class and adjusts
a such that the hyperplane moves toward the center of the worst patterns and rotates
so that the angle between the hyperplane and the vector connecting the worst points
increases. Once the hyperplane separates the classes, all the updates will involve
support vectors, since the if statements can only pick the vectors with the smallest
|at yi |.
32. Consider Support Vector Machines for classication.
(a) We are given the following six points in two categories:
 
 
 
1
2
2
1 : x1 =
, x2 =
, x3 =
1
2
0

208

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

 
 
 
0
1
0
: x4 =
, x5 =
, x6 =
0
0
1

with z1 = z2 = z3 = 1 and z4 = z5 = z6 = +1.

2
3/

y 2=
y 1+

y2
2

x2

x1

1
x6

x3

x4

x51

y1

2 /4

The optimal hyperplane is y1 + y2 = 3/2, or (3/2 1 1)t (1 y1 y2 ) = 0. To

ensure zk at y 1, we have to scale (3/2 1 1)t by 2, and thus the weight
vector is (3 2 2)t . The optimal marginis the shortest distance from the
patterns to the optimal hyperplane, which is 2/4, as can be seen in the gure.
(b) Support vectors are the samples on the margin, that is, the ones with the shortest
distance to the separating hyperplane. In this case, the support vectors are
{x1 , x3 , x5 , x6 } = {(1, 1)t , (2, 0)t , (1, 0)t , (0, 1)t }.
(c) We seek to maximize the criterion given in Eq. 109 in the text,
L() =

n
!

1!
k j zk zj yjt yk
2
n

k=1

k,j

subject to the constraints

n
!

zk k = 0

k=1

for k 0. Using the constraint, we can substitute 6 = 1 + 2 + 3 4 5

in the expression for L(). Then we can get a system of linear equations by
setting the partial derivatives, L/i to zero. This yields:

1 2 2 0
1
1
2
2 5 2 1 1 2 2

2 2 5 1

3 = 2 .
0 1 1 1 1 4 0
1
1
3 1 2
0
5
Unfortunately, this is an inconsistent set of equations. Therefore, the maxima
must be achieved on the boundary (where some i vanish). We try each i = 0
and solve L/i = 0:
L(0, 2 , 3 , 4 , 5 )
=0
i

PROBLEM SOLUTIONS

209

implies = 1/5(0, 2 2 8 8 4)t , which violates the constraint i 0.

Next, both of the following vanishing derivatives,
L(1 , 0, 3 , 4 , 5 )
L(1 , 2 , 0, 4 , 5 )
=
=0
i
i
lead to inconsistent equations. Then the derivative
L(1 , 2 , 3 , 0, 5 )
=0
i
implies = 1/5(16 0 4 0 14 6)t , which does not violate the constraint i 0.
In this case the criterion function is L() = 4. Finally, we have
L(1 , 2 , 3 , 4 , 0)
=0
i
which implies = 1/5(2 2 2 0 0 6)t , and the constraint i 0 is obeyed. In
this case the criterion function is L() = 1.2.
Thus = 1/5(16 0 4 0 14 6)t is where the criterion function L reaches its
maximum within the constraints. Now we seek the weight vector a. We seek to
minimize L(a, ) of Eq. 108 in the text,
!
1
k [zk at yk 1],
a2
2
n

L(a, ) =

k=1

with respect to a. We take the derivative of the criterion function,

!
L
k zk yk = 0,
=a
a
n

k=1

which for the k found above has solution

a

= (16/5)y1 0y2 4/5y3 + 0y4 + 14/5y5 + 6/5y6

0
= 2 .
2

Note that L/a = 0 here is not sucient to allow us to nd the bias

 a0 directly
since the a2 term does not include the augmented vector a and k k zk = 0.
We determine a0 , then, by using one of the support vectors, for instance y1 =
(1 1 1)t . Since y1 is a support vector, at y1 z1 = 1 holds, and thus

0
2 (1 1 1) = a0 + 4 = 1.
2
This, then, means a0 = 3, and the full weight vector is a = (3 2 2)t .
33. Consider the Kuhn-Tucker theorem and the conversion of a constrained optimization problem for support vector machines to an unconstrained one.

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

210

(a) Here the relevant functional is given by Eq. 108 in the text, that is,
!
1
k [zk at yk 1].
a2
2
n

L(a, ) =

k=1

We seek to maximize L with respect to to guarantee that all the patterns are
correctly classied, that is zk at yk 1, and we want to minimize L with respect
to the (un-augmented) a. This will give us the optimal hyperplane. This solution
corresponds to a saddle point in a space, as shown in the gure.
saddle
point
L

(b) We write the augmented vector a as (a0 ar )t , where a0 is the augmented bias.
Then we have
!
1
k [zk atr yk + zk a0 1].
ar 2
2
n

L(ar , , a0 ) =

k=1

At the saddle point, L/a0 = 0 and L/ar = 0. The rst of these derivative
vanishing implies
n
!

k zk = 0.

k=1

(c) The second derivative vanishing implies

!
L
= ar
k zk yk
ar
n

k=1

and thus
ar =

n
!

k zk yk .

k=1

Since

n

k=1

k zk = 0, we can thus write the solution in augmented form as

a=

n
!
k=1

k zk yk .

PROBLEM SOLUTIONS

211

(d) If k (zk at yk 1) = 0 and if zk at yk = 0, then k must be zero. Respectively,

call the predicates as h, N OT p and q, then the above states h AN D N OT p q,
which is equivalent to h AN D N OT q p. In other words, given the expression
k (zk at y 1) = 0,
then is non-zero if and only if zk at yk = 1.
(e) Here we have
=
L

 1 n
2

n
n
>2 !
!
1>
>
>!
k zk yk >
k zk
l zl yl yk 1
>
2
k=1
k=1
l=1
2t 1 n
2
1 n
n
n
!
!
!
1 !
k zk yk
k zk yk
k l zk zl ykt yl +
k .
2
k=1

k=1

kl

k=1

Thus we have
=
L

n
!

k=1

1!
k l zk zl ykt yl .
2
kl

(f) See part (e).

34. We repeat Example 2 in the text but with the following four points:

y1 = (1 2 5 2 5 2 1 25)t , y2 = (1 2 2 4 2 8 2 4 16)t , z1 = z2 = 1

y3 = (1 2 3 2 6 2 4 9)t , y4 = (1 2 2 5 2 5 2 1 25)t , z3 = z4 = +1
We seek the optimal hyperplane, and thus want to maximize the functional given by
Eq. 109 in the text:
1!
l k zk zl ykt yl ,
2
4

L() = 1 + 2 + 3 + 4

kl

with constraints 1 + 2 = 3 + 4 and 1 0. We substitute 4 = 1 + 2 3

into L() and take the partial derivatives with respect to 1 , 2 and 3 and set the
derivatives to zero:
L
1
L
2
L
3

= 2 2081 2562 + 2323 = 0

= 2 2561 5922 + 4963 = 0
= 2321 + 4962 5333 = 0.

The solution to these equations 1 = 0.0154, 2 = 0.0067, 3 = 0.0126 indeed

satisfy the constraint i 0, as required.
Now we compute a using Eq. 108 in the text:
!
L
k zk yk = 0,
=a
a
4

k=1

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

212

which has solution

a =
=

0.0154(y1 ) + 0.0067(y2 ) + 0.01261y3 + 0.095y4

(0 0.0194 0.0496 0.145 0.0177 0.1413)t .

Note that this process cannot determine the bias term, a0 directly; we can use a
support vector for this in the following way: We note that at yk zk = 1 must hold for
each support vector. We pick y1 and then
(a0 0.0194 0.0496 0.145 0.0177 0.1413) y1 = 1,
which gives a0 = 3.1614, and thus the full weight vector is a = (3.1614 0.0194 0.0496
0.145 0.0177 0.1413)t .
Now we plot the discriminant function in x1 x2 space:

g(x1 , x2 ) = at (1 2x1
2x2
2x1 x2 x21 x22 )
= 0.0272x1 + 0.0699x2 0.2054x1 x2 + 0.1776x21 0.1415x22 + 3.17.
The gure shows the hyperbola corresponding to g(x1 , x2 ) = 0 as well as the margins,
g(x1 , x2 ) = 1, along with the three support vectors y2 , y3 and y4 .
x2
10

g(x1,x2)=-1

7.5

y4

y1
5

g(x1,x2)=1

y3
2.5

g(x1,x2)=0
-10

-7.5

-5

-2.5

2.5

7.5

10

x1

-2.5

y2

-5

g(x1,x2)=-1

g(x1,x2)=1
-7.5

-10

Section 5.12
35. Consider the LMS algorithm and Eq. 61 in the text.
(a) The LMS algorithm minimizes Js (a) = Ya b2 and we are given the problem
>

 
>
>2
>
a0
b1 >
> 1n Y 1
Js (a) = >

>
ar
b2 >
> 1n Y 1
We assume we have 2n data points. Moreover, we let bi = (b1i b2i . . . bni )t for
i = 1, 2 be arbitrary margin vectors of size n each, and Y1 is an n-by-2 matrix
containing the data points as the rows, and a = (a0 , a1 , a2 )t = (a0 , ar )t . Then
we have
>
>
> a 1 + Y1 ar b1 >2
Js (a) = > 0 n
>
a0 1n + Y1 ar b1

PROBLEM SOLUTIONS
=

n
!

213

(a0 + yit ar bi1 )2 +

i=1

=
=

n
!
i=1
n
!

(a0

bi1 )2

(a0 bi1 )2 +

i=1

n
!
i=1
n
!
i=1

n
!

(a0 + yit ar bi2 )2

i=1

(a0

bi2 )2

+2

(a0 + bi2 )2 2

n
!

yit ar (a0 bi1 a0 bi2 )

i=1
n
!

yit ar (bi1 + bi2 ).

i=1

Thus Js /a0 = 0 implies a0 = 0 and the minimum of J must be at (0, ar )t for

some ar . Hence we showed that a0 = 0 which tells us that the separating plane
must go through the origin.
(b) We know from part (a) that the LMS hyperplane must pass through the origin.
Thus 1 samples must lie on the same side of a separating hyperplane in order
to ensure 2 samples lie in the other half space. This is guaranteed in the shaded
(union) region. We are asked to dene the region where there is no shading.
We can state this as: if y {(x1 , x2 )t |2x1 < |x2 |}, then the LMS solution to
separate 1 and 2 will not give a separating hyperplane, where
   
1
2
1 :
,
,y
2
4
   
1
2
2 :
,
, y
2
4
as shown in the gure.
(c) Part (b) can be generalized by noting that the feasible region for y (so that LMS
will give a separating hyperplane) is the union of the half spaces Hi determined
by yi as:
H1

= {half space induced by the separating hyperplane y1 and containing y2 }

H2

= {half space induced by the separating hyperplane y2 and containing y1 }.

The LMS will give separating hyperplanes if and only if y3 H1 H2 . Thus, to

ensure that the LMS solution not to separate {y1 , y2 , y3 } from {y1 , y2 , y3 },
1 H
2.
we must have y3 H
36. The algorithm is as follows:
Algorithm 0 (Fixed increment multicategory Perceptron)
1
2
3
4
5
6
7
8
9
10

begin initialize ai k 0, n number of samples, c number of classes

do k (k + 1)modn
i class of yk
do j j + 1; (j = i)
until ati yk < atj yk j > c
if j c then ai ai + yk
a j aj y k
until no more misclassications
return a1 , a2 , . . . , ac
end

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

214

x2

x2 =
2x
1

y1

x1

-2x 1
x 2=

y2

Some of the advantages are that the algorithm is simple and admits a simple
parallel implementation. Some of the disadvantages are that there may be numerical
instability and that the search is not particularly ecient.
37. Equation 122 in the text gives the Bayes discriminant function as
g0i =

c
!

ij P (i |x).

j=1

The denition of ij from Eq. 121 in the text ensures that for a given i only one ij
will be non-zero. Thus we have g0i = P (i |x). We apply Bayes rule to nd
p(x)g0i (x) = p(x|i )P (i ).

()

On the other hand, analogous to Eq. 58 in the text, the criterion function Js1 i can be
written as
!
!
Js1 i (a) =
(ati y 1)2 +
(ati y)2
yY
/ i

yYi

!
!
1
1
n
n
1
2
= n
(ati y 1)2 +
(ati y)2 .
n n1
n n2

yY
/ i

yYi

By the law of large numbers, as n , we have that 1/nJs1 i (ai ) approaches Ji (ai )
with probability 1, where Ji (ai ) is given by
Ji (ai ) = P (i )Ei [(ati y 1)2 ] + P (N OT i )E2 [(ati y)2 ],
where


E1 [(ati y 1)2 ] =
E2 [(ati y)2 ] =

Thus we have
Ji (ai ) =

(ati y 1)2 p(x|i )dx

(ati y)2 p(x|N OT i )dx.


(ati y

1) p(x, i )dx +
2

(ati y)2 p(x, N OT i )dx.

PROBLEM SOLUTIONS

215

We expand the square to nd:





i ) = (at y)2 p(x, i )dx 2 (at y)p(x, i )dx + p(x, i )dx + (at y)2 p(x, N OT i )dx.
J(a
i
i
i
We collect the terms containing (ati y)2 and nd



t
2
t

J(ai ) = (ai y) p(x)dx 2 (ai y)p(x, i )dx + p(x, i )dx.

We use the fact that p(x)g0i (x) = p(x, i ) from (), above, we nd



t
2
t

J(ai ) = (ai y) p(x)dx 2 (a y)p(x)g0i (x)dx + p(x, i )dx,

which can be written as


J(ai ) =
[ati y g0i (x)]2 p(xdx)




2
[g0i
(x)p(x) g0i (x)p(x)]dx .




2i ,the meansquared approx. error

independent of ai

The second term in the sum is independent of weight vector ai . Hence the ai that
minimizes Js1 i also minimizes 2i . Thus the MSE solution A = Yt B (where A =
ai a2 ac and Y = [Y1 Y2 Yc ]t , and B is dened by Eq. 119 in the text) for
the multiclass case yields discriminant functions ati y that provide a minimum-meansquare error approximation to the Bayes discriminant functions g0i .
38. We are given the multi-category classication problem with sets Y1 , Y2 , . . . , Yc
to be classied as 1 , 2 , . . . , c , respectively. The aim is to nd a1 , a2 , . . . , ac such
that if y Yi , then ati y atj y for all j = i. We transform this problem into a binary
classication problem using Keslers construction. We dene G = G1 G2 Gc
where
Gi = {ij |y Yi , j = i}
and ij is as given in Eq. 115 in the text. Moreover we dene
= (at1 a2 atc )t .
Perceptron case We wish to show what the Fixed-increment single sample Perceptron algorithm given in Eq. 20 in the text does to our transformed problem. We
rewrite Eq. 20 as
(1)
(k + 1)

= arbitrary
= (k) + gk

where gk is misclassied. The condition gk being misclassied at step k implies

t (k)gk 0. Since the Gi s are disjoint, gk must belong to one and only one
of the Gi . We shall use a subscript to denote which Gi that gk belongs to; for
instance, gik is in Gi . Given this notation, the inequality t (k)gik 0 implies
ati y atj y 0 for some j = i. Thus there is an equivalence between t (k)gik 0
and ati (k)y atj (k)y.
Consider the update rule (k + 1) = (k) + gik . At once we see the equivalence:

216

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

Multi-category
(k + 1) = (k) + gik

Two-category
ai (k + 1) = ai (k) + yik
aj (k + 1) = aj (k) yjk ,

where, as dened above, is the concatenation of as.

Relaxation rule case The single sample relaxation rule becomes
(1)
(k + 1)

arbitrary

= (k) +

b t gk k
g ,
gk 2

where b is the margin. An update takes place when the sample gk is incorrectly
classied, that is, when t (k)gk < b. We use the same denition of g as in the
Perceptron case above, and can then write
t (k) ij < b (ati (k)y atj (k)y) < b.
In the update rule, can be decomposed into its sub-component as, yielding
the following equivalence:
Multi-category
(k + 1) = (k) +

t (k)gk gk
bg
k 2

Two-category

b(ati (k)atj (k))yk k

y
2yk 2
b(ati (k)atj (k))yk k
+
y
2yk 2

ai (k + 1) = ai (k) +
aj (k + 1) = aj (k)

5.0. COMPUTER EXERCISES

217

Computer Exercises
Section 5.4
1. Computer exercise not yet solved
Section 5.5
2.
3.
4.
5.
6.

Computer
Computer
Computer
Computer
Computer

exercise
exercise
exercise
exercise
exercise

not
not
not
not
not

yet
yet
yet
yet
yet

solved
solved
solved
solved
solved

Section 5.6
7. Computer exercise not yet solved
Section 5.8
8. Computer exercise not yet solved
Section 5.9
9. Computer exercise not yet solved
Section 5.10
10. Computer exercise not yet solved
Section 5.11
11. Computer exercise not yet solved
Section 5.12
12. Computer exercise not yet solved

218

CHAPTER 5. LINEAR DISCRIMINANT FUNCTIONS

Chapter 6

Multilayer neural networks

Problem Solutions
Section 6.2
1. Consider a three-layer network with linear units throughout, having input vector
x, vector at the hidden units y, and output vector z. For such a linear system we
have y = W1 x and z = W2 y for two matrices W1 and W2 . Thus we can write the
output as
z

= W2 y = W2 W1 x
= W3 x

for some matrix W3 = W2 W1 . But this equation is the same as that of a two-layer
network having connection matrix W3 . Thus a three-layer network with linear units
throughout can be implemented by a two-layer network with appropriately chosen
connections.
Clearly, a non-linearly separable problem cannot be solved by a three-layer neural network with linear hidden units. To see this, suppose a non-linearly separable
problem can be solved by a three-layer neural network with hidden units. Then, equivalently, it can be solved by a two-layer neural network. Then clearly the problem is
linearly separable. But, by assumption the problem is only non-linearly separable.
Hence there is a contradiction and the above conclusion holds true.
2. Fouriers theorem shows that a three-layer neural network with sigmoidal hidden
units can act as a universal approximator. Consider a two-dimensional input and a
single output z(x1 , x2 ) = z(x). Fouriers theorem states
!!
z(x) 
Af1 f2 cos(f1 x1 )cos(f2 x2 ).
f1

f2

(a) Fouriers theorem, as stated above, can be rewritten with the trigonometric
identity:
1
1
cos()cos() = cos( + ) + cos( ),
2
2
219

CHAPTER 6. MULTILAYER NEURAL NETWORKS

220
to give

z(x1 , x2 ) 

! ! Af
f1

f2

1 f2

z2

[cos(f1 x1 + f2 x2 ) + cos(f1 x1 f2 x2 )] .

(b) We want to show that cos(x), or indeed any continuous function, can be approximated by the following linear combination


n
!
Sgn[x xi ]
f (x)  f (x0 ) +
[f (xi+1 f (xi )]
.
2
i=0
The Fourier series of a function f (x) at a point x0 converges to f (x0 ) if f (x)
is of bounded variation in some interval (x0 h, x0 + h) centered on x0 . A
function of bounded variation is a follows, given a partition on the interval
a = x0 < x1 < x2 < . . . < xn1 < xn = b form the sum
n
!

|f (xk ) f (xk1 )|.

k=1

The least upper bound of these sums is called the total variation. For a point
f (x) in the neighborhood of f (x0 ), we can rewrite the variation as


n
!
Sgn[x xi ]
[f (xi+1 f (xi )]
,
2
i=1
which sets the interval to be (x, x + 2h). Note the function has to be either
continuous at f (x0 ) or have a discontinuity of the rst kind.
f(x)

approximation
function

x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 x11 x12 x13 x14

(c) As the eective width of the sigmoid vanishes, i.e., as 0, the sigmoids
become step functions. Then the functions cos(f1 x1 +f2 x2 ) and cos(f1 x1 f2 x2 )
can be approximated as
cos(f1 x1 + f2 x2 )  cos(f1 x10 + f2 x20 )
1 n
!
+
[cos(x1i+1 f1 + x2i+1 f2 ) cos(x1i f1 + x2i f2 )]
i=0

Sgn[x1 x1i ]Sgn[x2 x2i ]

2
and similarly for cos[f1 x1 f2 x2 ].

2
,

PROBLEM SOLUTIONS

221

(d) The construction does not necessarily guarantee that the derivative is approximated, since there might be discontinuities of the rst order, that is, noncontinuous rst derivative. Nevertheless, the one-sided limits of f (x0 + 0) and
f (x0 0) will exist.
Section 6.3
3. Consider a d nH c network trained with n patterns for me epochs.
(a) Consider the space complexity of this problem. The total number of adjustable
weights is dnH + nH c. The amount of storage for the n patterns is nd.
(b) In stochastic mode we choose pattern randomly and compute
w(t + 1) = w(t) + w(t).
until stopping criterion is met. Each iteration involves computing w(t) and
then adding it to w(t). From Eq. 17 in the text, for hidden-to-output unit
weights we have
wjk = (tk zk )f  (netk )yj ,
where netk =

n
H

j=1

wjk yj is computed by nH multiplications and nH additions.

So, wjk is computed with c(2nH + 1) operations.

From Eq. 17 in the text, for input-to-hidden unit weights we have
wji = xi f  (netk )

c
!

wkj k

k=1

where, netj is computed in 2d operations. Moreover,


k

wjk k is computed in

2c operations. Thus we have wij s are computed in [2d + 2c]nH time.

Thus, the time for one iteration is the time to compute w plus the time to
add w to w, that is,
T

= c(2nH + 10 + 2(d + c + 1)nH + (dnH + nH c)

= 3dnH + 5nH c + c + 2nH .

In summary, the time complexity is (3dnH + 5nH c + c + 2nH )me .

(c) Here, the number of iterations = nme and thus the time complexity is (3dnH +
5nH c + c + 2nH )n me .
4. Equation 20 in the text gives the sensitivity at a hidden unit
j = f  (netj )

c
!

wkj k .

k=1

For a four-layer or higher-layer neural network, the sensitivity of a hidden unit is

likewise given by


E
E Oj
j
.
=
netj
oj netj

CHAPTER 6. MULTILAYER NEURAL NETWORKS

222

By the chain rule we have

j = f  (netj )

! E netk
.
netk oj
k

Now, for a unit k in the next higher layer

!
!
netk =
wj  k oj  =
wj  k oj  + wjk oj ,
j

j

where
netk
= wjk .
oj
Thus we have
j = f  (netj )

c
$
!
E %
wjk
= f  (netj )
wjk k .
netk
k=1
k=1



c
!

5. From Eq. 21 in the text, the backpropagation rule for training input-to-hidden
weights is given by
!
wji = xi f  (netj )
wkj k = xi j .
k=1

For a xed hidden unit j, wji s are determined by xi . The larger the magnitude of
xi , the larger the weight change. If xi = 0, then there is no input from the ith unit and
hence no change in wji . On the other hand, for a xed input unit i, wji j , the
sensitivity of unit j. Change in weight wji determined how the overall error changes
with the activation of unit j. For a large magnitude of j (i.e., large sensitivity),
wji is large. Also note that the sensitivities of the subsequent layers propagate to
the layer j through weighted sum of sensitivities.
6. There is no reason to expect that the backpropagation rules should be inversely
related to f  (net). Note, the backpropagation rules are dened to provide an algorithm
that minimizes
1!
(tk zk )2 ,
2
c

J=

k=1

where zk = f (netk ). Now, the weights are chosen to minimize J. Clearly, a large
change in weight should occur only to signicantly decrease J, to ensure convergence
in the algorithm, small changes in weight should correspond to small changes in J.
But J can be signicantly decreased only by a large change in the output ok . So, large
changes in weight should occur where the output varies the most and least where the
output varies the least. Thus a change in weight should not be inversely related to
f  (net).
7. In a three-layer neural network with bias, the output can be written

1
2
!
!
zk = f
wjk f
wji xi
j

PROBLEM SOLUTIONS

223

can be equivalently expressed as

zk = f

wkj f

1
!

2
wji xi

by increasing the number of input units by 1 and the number of hidden inputs by 1,
xo = 1, oj = j , j = o, io = 0,
and fh is adjusted such that obias = fh (o) = 1 and wok = k .
8. We denote the number of input units as d, the number of hidden units as nH and
the number of category (output) units as c. There is a single bias unit.
(a) The number of weights is dnH + (nH + 1)c, where the rst term is the number
of input-to-hidden weights and the second term the number of hidden-to-output
weights, including the bias unit.
(b) The output at output node k is given by Eq. 7 in the text:
2
2
1
1 d
nH
!
!
zk =
f wkj f
wji xi + wj0 + wk0 .
j=1

i=1

We assume that the activation function f () is odd or antisymmetric, giving

d
d
!
!
f
wji xi f
wji xi ,
j=1

j=1

but since the weighting on this summation also ips, we have

d
d
!
!
(wkj ) f
wji xi = (wkj ) f
wji xi ,
j=1

j=1

and the original output is unchanged.

(c) The hidden units can be exchanged along with corresponding weights and this
can leave the network unaected. The number of subsets that can be constructed
over the set nH is of course 2nH . Now, because the corresponding weights for
each subset nH dierent weight orderings can be constructed, so the total hidden
unit symmetry is nH !2nH . For the case nH = 10, this is a factor of 3,715,891,200.
9. The on-line version of the backpropagation algorithm is as follows:
Algorithm 0 (On-line backpropagation)
1
2
3
4
5
6
7

begin initialize nH , w,
do
x next input pattern
wji wji + j xi ; wkj wkj + k yi
until no more patterns available
return w
end

CHAPTER 6. MULTILAYER NEURAL NETWORKS

224

10. We express the derivative of a sigmoid in terms of the sigmoid itself for positive
constants a and b for the following cases.
(a) For f (net) = 1/(1 + ea

net

), the derivative is
2

df (net)
1
ea net
= a
d net
1 + ea net
= af (net)(1 f (net)).

(b) For f (net) = atanh(b net), the derivative is

df (net)
d net

= 2b2 atanh(b net)(1 tanh2 (b net)).

11. We use the following notation: The activations at the rst (input), second, third,
and fourth (output) layers are xi , yj , vl , and zk , respectively, and the indexes are
clear from usage. The number of units in the rst hidden layer is nH1 and the number
in the second hidden layer is nH2 .

output

v1

x1

z1

z2

v2

vl

y1

y2

x2

xi

zq

zc

vlmax

yj

yjmax

xd

input
Algorithm 0 (Four-layer backpropagation)
1
2
3
4
5

begin initialize xxx

xxx
xxx
return xxx
end

Section 6.4
12. Suppose the input to hidden weights are set equal to the same value, say wo ,
then wij = wo . Then we have
netj = f (netj ) =

d
!
i=1

wji xi = wo

!
i

xi = wo x.

PROBLEM SOLUTIONS

225

This means that oj = f (netj ) is constant, say yo . Clearly, whatever the topology of
the original network, setting the wji to be a constant is equivalent to changing the
topology so that there is only a single input unit, whose input to the next layer is xo .
As, a result of this loss of one-layer and number of input units in the next layer, the
network will not train well.
13. If the labels on the two hidden units are exchanged, the shape of the error surface
is unaected. Consider a d nH c three-layer network. From Problem 8 part (c),
we know that there are nH !2nH equivalent relabelings of the hidden units that do not
aect the network output. One can also ip weights for each of these congurations,
Thus there should be nH !2nH +1 possible relabelings and weight changes that leave
the error surface unaected.
14. Consider a simple 2 1 network with bias. Suppose the training data come from
two Gaussians, p(x|1 ) N (0.5, 1) and p(x|2 ) N (0.5, 1). The teaching values
are 1.
(a) The error is a sum over the n patterns as a function of the transfer function f ()
as
J(w) =

n
!

(tk f (wt xk + w0 ))2 ,

k=1

where tk is the teaching signal and w0 the bias weight.

(b) We dierentiate twice with respect to the weights to compute the Hessian matrix, which has components
Hij =

2 J(w)
.
wi wj

We use the outer product approximation of the Hessian and nd

Hij =

n
!

f  (wt xk + w0 )xki f  (wt xk + w0 )xkj

k=1

where xkj is the jth component of sample xk .

(c) Consider two data sets drawn from p(x|1 ) N (ui , I). The Hessian matrix
then has components
Hij = n2 (1i + 2i )(1j + 2j ),
where 1i is the ith element of the mean fector of class 1, and analogously for
class 2.
(d) Problem not yet solved
(e) Problem not yet solved
(f) Problem not yet solved
15. We assume that the error function can be well described by a Hessian matrix H
having d eigenvalues and where max and min are the maximum and minimum of
these eigenvalues.

CHAPTER 6. MULTILAYER NEURAL NETWORKS

226

(a) The optimimum rate is < 2/max .

(b) The convergence criterion is |1 i | < 1 for all eigenvalues i = 1, . . . , d.
(c) The time for the system to meet the convergence criterion is
= (1 i )T
where T is the total number of steps. This factor is dominated by the smallest
eigenvalue, so we seek the value T such that
= (1 min )T .
This, in turn, implies that
T =

ln
.
ln(1 )

16. Assume the criterion function J(w) is well described to second order by a Hessian
matrix H.
(a) Stepping along the gradient gives at step T
iT = (1 i )T i0 .
To get a local minimum, we need |1 i | < 1, and this implies < 2/max .
(b) Consider (1 (2/max )i ). More steps will be required along the direction corresponding to the smallest eigenvalue i , and thus the learning rate is governed
by
1

2min
.
max

Standardization helps reduce the learning time by making the ratio max /min =
1.
(c) Standardization is, in essence, a whitening transform.
Section 6.6
17. From Eq. 25 in the text, we have

!
!
n
1
1
n

n
k
k
J(w) = n
[gk (x, w) 1]2 +
gk (x, w)2 .
n nk x
n n nk
xk

As n , the proportion of all samples that are in k approaches P (k ). By the

law of large numbers, then,
1 !
[gk (x, w) 1]2
nk x
k

approaches


E([gk (x, w) 1] |x k ) =
2

[gk (x, w) 1]2 p(x|k )dx.

PROBLEM SOLUTIONS

227

Likewise, we have
!
1
[Fk (x, )]2 ,
n nk
xk

which approaches

E([gk (x, w)] |x i=k ) =

2

[gk (x, w)]2 p(x|i=k )dx.

Thus we see, in the limit of innite data



J(w) = P (k ) [gk (x, w) 1]2 p(x|k )dx + P (i=k ) [gk (x, w)]2 p(x|i=k )dx


2
=
[gk (x, w) 1] p(x, k )dx + [gk (x, w)]2 p(x|i=k )dx


=
[gk2 (x, w) + 1 2gk (x, w)]p(x, w)]p(x, k )dx + [gk (x, w)]2 p(x|i=k )dx


2
=
[gk (x, w) P (k |x)] p(x)dx + P (k |x)[1 P (k |x)]p(x)dx


=
[gk (x, w) P (k |x)]2 p(x)dx + P (k |x)P (i=k |x)p(x)dx.
18. Consider how one of the solutions to the minimum-squared-error condition indeed
yields outputs that are estimates of posterior probabilities.
(a) From Eq. 25 in the text, we have



2
J(w) =
gk (x, w)p(x)dx 2 gk (x, w)p(x, k )dx + p(x, k )


J(w)
gk (x, w)
gk (x, w)
= 2 gk (x, w)
p(x)dx 2
p(x, k )dx.
w
w
w
We set

J(w)
w

= 0 and nd

Fk (x, )

Fk (x, )
p(x)dx =

Fk (x, )
p(x, k )dx

Clearly, w w (p(x)), the solution of the above equation, depends on the

choice of p(x). But, for all p(x), we have


gk (x, w) 
gk (x, w) 

gk (x, w (p(x))
p(x)dx
=
p(x, k )dx.


w
w
w=w (p(x))
w=w (p(x))
Thus we have
gk (x, w (p(x)))p(x) = p(x, k )
with probability 1 on the set of x : p(x, k ) > 0. This implies
gk (x, w ) =

p(x, k )
= P (k |x).
p(x)

228

CHAPTER 6. MULTILAYER NEURAL NETWORKS

(b) Already shown above.

19. The assumption that the network can represent the underlying true distribution is
not used before Eq. 28 in the text. For Eq. 29, however, we invoke gk (x; w)  p(k |x),
which is used for
c 
!
[gk (x; w) P (k |x)] p(x)dx = 0.
k=1

This is true only when the above assumption is met. If the assumption is not met, the
gradient descent procedure yields the closest projection to the posterior probability
in the class spanned by the network.
20. Recall the equation
t

k )+B(y,)+w
y
p(y|k ) = eA(w
.

(a) Given p(y|k ), we use Bayes Theorem to write the posterior as

p(k |y) =

p(y|k )P (k )
.
p(y)

k and as follows:
(b) We interpret A(), w
k)
= eA(w
kt y] P (k )
k ) + B(y, ) + w
exp [A(w
p(k |y) = 
c
t y] P ( )
m ) + B(y, ) + w
m
exp [A(w
m

P (k )

m=1

netk
,
c

enetm
m=1

t y. Thus, B(y, ) is the bias, w

is the weight vector
where netk = b(y, ) + w
k)
describing the separating plane, and eA(w
is P (k ).
21. Backpropagation with softmax is done in the usual manner; all that must be
evaluated dierently are the sensitivities at the output and hidden layer units, that
is,
eneth
zh =  net
e h

zh
= zh (1 zh ).
neth

and

(a) We are given the following terms:

netj
netk

=
=

d
!
i=1
nH
!

wji xi
wkj yj

j=1

yj
zk

= f (netj )
enetk
= 
,
c
enetm
m=1

PROBLEM SOLUTIONS

229

and the error function

1!
(tk zk )2 .
2
c

J=

k=1

To derive the learning rule we have to compute J/wkj and J/wji . We start
with the former:
J
J netk
=
.
wkj
netk wkj
We next compute
J
netk

c
!
J zs
.
zs netk
s=1

We also have

enets (1)enetk


2 = zs zk

net
m
e

zs
m=1
=
nets
nets nets
 ce
 + (1)  e e 2 = zk zk2

netk

 net


m
e
enetm

m=1

if s = k
if s = k

m=1

and nally
J
= (1)(ts zs ).
zs
Putting these together we get
!
J
=
(1)(ts zs )(zs zk ) + (1)(tk zk )(zk zk2 ).
netk
c

s=k

We use netk /wkj = yj and obtain

!
J
= yj
(ts zs )(zs zk ) yj (tk zk )(zk zk2 ).
wkj
c

s=k

Now we have to nd input-to-hidden weight contribution to J. By the chain

rule we have
J
J yj netj
=
.
wji
yj netj wji
We can at once nd out the last two partial derivatives.
yj
netj
= f  (netj ) and
= xi .
netj
wji
Now we also have
J
yj

c
!
J zs
z
s yj
s=1

230

CHAPTER 6. MULTILAYER NEURAL NETWORKS

c
!

zs
y
j
s=1


c
c
!
!
zs netr
=
(ts zs )
netr yj
s=1
r=1

(ts zs )

c
z net
!
zs netr

s
s
(ts zs )
+

net
y
net
y

s
j
r
j
s=1
      r=s      

c
!

zs zs2

c
!

zs zr

wsj

(ts zs )(zs zs2 )wsj +

s=1

c !
c
!

wrj

zs zr wrj (ts zs ).

s=1 r=s

We put all this together and nd

J
wji

= xi f  (netj )

c
!

(ts zs )

s=1

xi f  (netj )

c
!

c
!

wrj zs zr

r=s

(ts zs )wsj (zs zs2 ).

s=1

Of course, the learning rule is then

J
wji
J
=
,
wkj
=

wji
wkj

where the derivatives are as given above.

(b) We are given the cross-entropy criterion function
JCE =

c
!

tk ln

k=1

tk
.
zk

The learning rule derivation is the same as in part (a) except that we need to
replace J/z with Jce /z. Note that for the cross-entropy criterion we have
Jce /zk = tk /zk . Then, following the steps in part (a), we have
Jce
wkj

= yj

c
!
tk
tk
zs zk yj (zk zk2 )
zk
zk
s=k

= yj

c
!

tk zk yj tk (1 zk ).

s=k

Similarly, we have
Jce
wji

= xi f  (netj )

c
c
!
ts !
wrj zs zr
z
s=1 s
r=x

c
!
ts
xi f  (netj )
wsj (zs zs2 ).
z
s
s=1

PROBLEM SOLUTIONS

231

Thus we nd
Jce
wji

= xi f  (netj )

c
!

ts

s=1

xi f  (netj )

c
!

c
!

wrj zr

r=s

ts wsj (1 zs ).

s=1

Of course, the learning rule is then

wji
wkj

Jce
wji
Jce
=
,
wkj
=

where the derivatives are as given above.

22. In the two-category case, if g1  P (1 |x), then 1 g1  P (2 |x), since we
can assume the categories are mutually exclusive and exhaustive. But 1 g1 can
be computed by a network with input-to-hidden weights identical to those used to
compute g1 . From Eq. 27 in the text, we know
!
!
2
[gk1 (x, w) P (k1 |x)] dx +
[gk2 (x, w) P (k2 |x)] dx
k1

k2

is a minimum. this implies that every term in the above equation is minimized.
Section 6.7
23. Consider the weight update rules given by Eqs. 12 and 23 in the text.
(a) The weight updates are have the factor f  (net). If we take the sigmoid fb (h) =
tanh(bh), we have
fb (h) = 2b

ebh
1
1 = 2b bh
1 = 2bD = 1.
(1 + ebh )2
e + ebh + 2



D

Clearly, 0 < D < 0.25 for all b and h. If we assume D is constant, then clearly

the product /fb
(h) will be equal to fb (h), which tells us the increment
in the weight values will be the same, preserving the convergence time. The
assumption will be approximately true as long as |bh| is very small or kept
constant in spite of the change in b.
(b) If the input data is scaled by 1/, then the increment of weights at each step
will be exactly the same. That is,

f (h/) = f  (h).
b
Therefore the convergence time will be kept the same. (However, if the network
is a multi-layer one, the input scaling should be applied to the hidden units
outputs as well.)

CHAPTER 6. MULTILAYER NEURAL NETWORKS

232

24. A general additive model is given by Eq. 32 in the text:

d
!
zk = f
fjk (xj ) + wok .
j=1

The functions fk act on single components, namely the xi . This actually restricts
the additive models. To have the full power of three-layer neural networks we must
assume that fi is multivariate function of the inputs. In this case the model will
become:

d
!
zk = f
fjk (x1 , x2 , . . . , xd ) + w0 .
j=1

With this modication it is trivial to implement any three-layer neural network. We

let

d
!
fjk = wkj g
wji xi + wj0
j=1

and w0k = wk0 , where g, wkj , wji , wj0 , wk0 are dened as in Eq. 32 in the text. We
substitute fjk into the above equation and at once arrive at the three-layer neural
network function shown in Eq. 7 in the text.
25. Let px (x) and py (y) be probability density functions of x and y, respectively.
(a) From the denition of entropy given by Eq. 37 in Chapter 1 of the text (or
Eq. 118 in Section A.7.1 in the Appendix), and the one-dimensional Gaussian,


1
x2
px (x) = exp 2 .
2
2
we have the entropy

H(px (x)) =

px (x)logpx (x)dx =

1
+ log 2  1.447782 + log bits.
2

(b) From Sect. 6.8.2 in the text, if a = 2/(3), then we have f  (x)  1 for <
x < .
(c) Suppose y = f (x). If there exists a unique inverse mapping x = f 1 (y), then
py (y)dy = f  (x)dx. Thus we have

H(py (y))

py (y)logpy (y)dy


px (x)logpx (x)dx +

= H(px (x)) + log2ab 2

1.418939 + log bits.

px (x)log|f  (x)|dx



1
x2
exp 2 log[1 + exp(bx)]dx
2
2

PROBLEM SOLUTIONS

233

(d) We dene the function

y = f (x) =

a if x = 0
0 otherwise.

then we have py (y) = (y) because

1
0
Pr[Y = y] =

if y = 0
if y = a
otherwise.

his gives our entropy to be H(y) = .

(e) Information possessed by the original data is transmitted with litle loss through
sigmoid functions, whereas it is completely lost through through a Diract delta
function.
26. The transfer function is given by Eq. 34 in the text:
f (net) =

2a
1 + eb

net

a.

(a) We calculate the derivative as

f  (net)

(b) Note that f  (net) =

2a
beb net
(1 + eb net )2


2a
beb net
=
1 + eb net 1 + eb net

1
2a
1
= b
b
net
1+e
1 + eb
b 2
=
[a (f (net))2 ].
2a
=

b

a f (net)f (net),

f ()
f  ()
f  ()


net

where b > 0. At net = , we have

2a
a = 2a a = a
1 + eb net
b 2
b 2
[a (f (net))2 ] =
(a a2 ) = 0
=
2a
2a
b
= f (net)f  (net) = 0.
a
=

At net = 0, we have
f (0) =
f  (0) =
f  (0) =

2a
2a
a=
a=0
b
net
1+e
a
b 2
b 2
ab
(a (f (net))2 ) =
a =
2a
2a
2
0.

At net = , we have
f ()

f  ()

f  ()

2a
a = 0 a = a
1 + eb net
b 2
b 2
(a (f (net))2 ) =
(a a2 ) = 0
2a
2a
0.

CHAPTER 6. MULTILAYER NEURAL NETWORKS

234

27. We consider the computational burden for standardizing data, as described in

the text.
(a) The data should be shifted and scaled to ensure zero mean and unit variance
(standardization). To compute the mean, the variance we need 2nd steps, and
thus the complexity is thus O(nd).
(b) A full forward pass of the network activations requires nH d + nH c steps. A
backpropagation of the error requires nH d + nH dc steps. The rst term is for
the output-to-hidden and the second term is for hidden-to-input weights. The
extra c factor comes from the backpropagation formula given in Eq. 21 in the
text. Thus nd epochs of training require
(nd)n[nh d + nH c + nH c + nH dc]
steps. Assuming a single output netowrk we can set c = 1 and use the approximation given in Sec. 6.8.7 to get n/10 = nH d + nH . We use this result in the
above and nd
n2 d[n/10 + n/10] =

n3 d
,
5

and thus is O(n3 d) complexity.

(c) We use the results of parts (a) and (b) above and see that the ratio of steps can
be approximated as
nd
5
= 2.
n3 d/5
n
This tells us that the burden of standardizing data gets negligible as n geets
larger.
28. The Minkowski error per pattern is dened by
J=

c
!

|zk tk |R =

k=1

c
!

|tk zk |R .

k=1

We proceed as in page 290 of the text for the usual sum-square-error function:
J
J netk
=
wkj
netk wkj
1
where zk = f

n
H


j=1

2
wkj yj . We also have

J
= R|tk zk |R1 f  (netk )Sgn(tk zk ),
netk
where the signum function can be written Sgn(x) = x/|x|. We also have
netk
= yj .
wkj

PROBLEM SOLUTIONS

235

Thus we have
J
= R|tk zk |R1 f  (netk )yj Sgn(tk zk ),
wkj
and so the update rule for wkj is
wkj = |tk zk |R1 f  (netk )yj Sgn(tk zk ).
Now we compute J/wji by the chain rule:
J
J yj netj
=
.
wji
yj netj wji
The rst term on the right-hand side is
J
yj

=
=

c
!
J zk
zk yj

k=1
c
!

R|tk zk |R1 f  (netk )wkj Sgn(tk zk ).

k=1

The second term is simply yj /netj = f  (netj ). The third term is netj /wji = xi .
We put all this together and nd

 c
!
J
R1

=
R|tk zk |
Sgn(tk zk )f (netk )wkj f  (netj )xi .
wji
k=1

and thus the weight update rule is


 c
!
wji =
wkj f  (netk )Sgn(tk zk )R|tk zk |R1 f  (netj )xi .
k=1

For the Euclidean distance measure, we let R = 2 and obtain Eqs. 17 and 21 in the
text, if we note the identity
Sgn(tk zk ) |tk zk | = tk zk .
29. Consider a d nH c three-layer neural network whose input units are liner and
output units are sigmoidal but each hidden unit implements a particular polynomial
function, trained on a sum-square error criterion. Here the output of hidden unit j is
given by
oj = wji xi + wjm xm + qj xi xm
for two prespecied inputs, i and m = i.
(a) We write the network function as

zk = f

nH
!
j=1

wkj oj (x)

236

CHAPTER 6. MULTILAYER NEURAL NETWORKS

where oj (x) is
oj (x) = wjij xij + wj mj xmj + qj xij wmj .
Since each hidden unit has prespecied inputs i and m, we use a subscript on i
and m to denote their relation to the hidden unit. The criterion function is
J=

c
!

(tk zk )2 .

k=1

Now we calculate the derivatives:

J
J netk
=
= (tk zk )f  (netk )oj (x).
wkj
netk wkj
Likewise we have
J
wji


 c
! J zk oj
J oj
=
oj wji
zk oj wji
k=1





=

J/oj
c
!

(tk zk )f  (netk )wkj

k=1

Now we turn to the second term:

xi + qj xmj
oj
xi + qj xij
=

wji
0

oj
.
wji

if i = ij
if i = mj
otherwise.

Now we can write J/wji for the weights connecting inputs to the hidden layer.
Remember that each hidden unit j has three weights or parameters for the two
specied input units ij and mj , namely wjij , wjmj , and qj . Thus we have
J
wjij
J
wjmj
J
wjr


=

=

c
!
k=1
c
!


(tk zk )f  (netk )wkj (xij + qj xmj )



(tk zk )f (netk )wkj (xim + qj xim )

k=1

0 for r = ij and r = mj .

We must also compute

J
J oj
=
,
qj
oj qj
where J/qj = xij xmj and
!
J
=
(tk zk )f  (netk )wkj .
oj
c

k=1

PROBLEM SOLUTIONS
Thus we have

237


 c
!
J

=
(tk zk )f (netk )wkj xij xmj .
qj
k=1

Thus the gradient descent rule for the input-to-hidden weights is


wjij
wjmj
qj

=
=
=

c
!

k=1
 c
!
k=1
 c
!




(tk zk )f (netk )wkj (xij + qj xmj )




(tk zk )f (netk )wkj (xim + qj xij )




(tk zk )f (netk )wkj xij xmj .

k=1

(b) We observe that yi = oi , from part (a) we see that the hidden-to-output weight
update rule is the same as the standard backpropagation rule.
(c) The most obvious weakness is that the receptive elds of the network are
a mere two units. Another weakness is that the hidden layer outputs are not
bounded anymore and hence create problems in convergence and numerical stability. A key fundamental weakness is that the network is no more a universal
approximator/classier, because the function space F that the hidden units
span is merely a subset of all polynomials of degree two and less. Consider a
classication task. In essence the network does a linear classication followed
by a sigmoidal non-linearity at the output. In order for the network to be able
to perform the task, the hidden unit outputs must be linearly separable in the
F space. However, this cannot be guaranteed; in general we need a much larger
function space to ensure the linear separability.
The primary advantage of this scheme is due to the task domain. If the task
domain is known to be solvable with such a network, the convergence will be
faster and the computational load will be much less than a standard backpropagation learning network, since the number of input-to-hidden unit weights is
3nH compared to dnH of standard backpropagation.
30. The solution to Problem 28 gives the backpropagation learning rule for the
Minkowski error. In this problem we will derive the learning rule for the Manhattan
metric directly. The error per pattern is given by
J=

c
!

|tk zk |.

k=1

We proceed as in page 290 of the text for the sum-squared error criterion function:
J
wkj
J
netk

J netk
netk wkj

= f  (netk )Sgn(tk zk ),

CHAPTER 6. MULTILAYER NEURAL NETWORKS

238

where Sgn(x) = x/|x|. Finally, we have netk /wkj = yj . Thus we have

J
= f  (netk )yj Sgn(tk zk ),
wkj
and so the update rule for wkj is:
wkj = f  (netk )yj Sgn(tk zk ).
Now we turn to J/wji . By the chain rule we have
J
J yj netj
=
.
wji
yj netj wji
We compute each of the three factors on the right-hand-side of the equation:
J
yj

=
=

c
!
J zk
zk yj

k=1
c
!

f  (netk )wkj Sgn(tk zk )

k=1

yj
netj
netj
wji

= f  (netj )
= xi .

Thus we put this together and nd


 c
!
J

=
Sgn(tk zk )f (netk )wkj f  (netj )xi .
wji
k=1

Thus the weight update rule for wji is:

 c

!

wji =
wkj f (netk )Sgn(tk zk ) f  (netj )xi .
k=1

The above learning rules dene the backpropagation learning rule with the Manhattan
metric. Note that the Manhattan metric uses the direction (signum) of tk zk instead
of the actual tk zk as in the sum-squared-error criterion.
Section 6.9
31. We are given the criterion function
J=

n
1 !
(tm zm )2 .
2 m=1

The Hessian matrix is then

!
n
n
!
2J
1
2 zm
zm zm
.

=
+
(zm tm )
wji wlk
2 m=1 wji wlk m=1
wji wlk





PROBLEM SOLUTIONS

239

We drop the last term in the outer-product approximation, as given in Eq. 45 in the
text. Let Wj be any hidden-to-output weight in a single output network. Then we
have z/Wj = f  (net)yj . Now we let wji be any input-to-hidden weight. Then since

!
!

z=f
Wj f
wji xi ,
j

the derivative is
z
= f  (net)Wj f  (netj )xi .
wji
Thus the derivatives with respect to the hidden-to-output weights are:
Xtv = (f  (net)y1 , . . . , f  (net)ynH ),
as in Eq. 47 in the text. Further, the derivatives with respect to the intput-to-hidden
weights are
Xtu

= (f  (net)f  (netj )W1 x1 , . . . , f  (net)f  (netnH )WnH x1 ,

f  (net)f  (net1 )W1 xd , . . . , f  (net)f  (netnH )WnH xd ).

32. We are given that JCE =

c

k=1

tk ln(tk /zk ). For the calculation of the Hessian

matrix, we need to nd the expressions for the second derivatives:

%
%
$
$
tk z
JCE

2
wji
zk wji
JCE
=
=
wji wlk
wlk
wlk


2 zk
tk zk zk
.
=
zk
zk2 wji wlk
wji wlk
We arrive at the outer product approximation by dropping the second-order terms.
We compare the above formula to Eq. 44 in the text and see that the Hessian matrix,
HCE diers from the Hessian matrix in the traditional sum-squared error case in
Eq. 45 by a scale factor of tk /zk2 . Thus, for a d nH 1 network, the Hessian matrix
can be approximated by
HCE =

n
1 ! t1 [m]t [m]
X
X ,
n m=1 z12

where X[m] is dened by Eqs. 46, 47 and 48 in the text, as well as in Problem 31.
33. We assume that the current weight value is w(n 1) and we are doing a line
search.
(a) The next w according to the line search will be found via
w(n) = w(n 1) + J(w(n 1))
by minimizing J(w) using only . Since we are given that the Hessian matrix is
proportional to the identity, H I, the Taylor expansion of J(w) up to second

240

CHAPTER 6. MULTILAYER NEURAL NETWORKS

degree will give an exact representation of J. We expand around w(n 1) and
nd
J(w)

= J(w(n 1)) + (w w(n 1))t J(w(n 1))

+1/2(w w(n 1))t H(w w(n 1)).

We know that H = kI for some scalar k. Further, we can write w w(n 1) =

J(w(n 1)), which can be plugged into the expression for J(w) to give
J(w)

= J(w(n 1)) + J t (w(n 1))J(w(n 1))

+(k/2)2 J t (w(n 1))J(w(n 1))


k
= J(w(n 1)) + + 2 ) J(w(n 1))2 .
2

Now we see the result of the line search in J by solving J/ = 0 for , that
is,
J
= J(w(n 1))2 (1 + k) = 0,

which has solution = 1/k. Thus the next w after the line search will be
w(n) = w(n 1) 1/kJ(w(n 1)).
We can also rearrange terms and nd
1
w(n) w(n 1) = w(n 1) = J(w(n 1)).
k
Using the Taylor expansion for J(w(n) around w(n 1) we can write
J(w(n)) = J(w(n 1)) + k(w(n) w(n 1)),
then substitute w(n) w(n 1) from above into the right-hand-side and nd
J(w(n)) = J(w(n 1)) + k(1/kJ(w(n 1))) = 0.
Indeed, Eqs. 56 and 57 in the text give n = 0, given that J(wn ) = 0,
as shown. Equation 56 in the text is trivially satised since the numerator is
J(w(n))2 , and thus the Fletcher-Reeves equation gives n = 0. Equation 57
also vanishes because the numerator is the inner product of two vectors, one of
which is J(w(n)) = 0, and thus the Polak-Ribiere equation gives n = 0.
(b) The above proves that application of a line search reult with w(n 1) takes us
to the minimum of J, and hence no further weight update is necessary.
34. Problem not yet solved
35. The Quickprop algorithm assumes the weights are independent and the error
surface is quadratic. Whith these assumptions, J/w is linear. The graph shows
J/w versus w and a possible error minimization step. For nding a minimum in
the criterion function, we search for J/w = 0. That is the value w where the line
crosses the w axis.

PROBLEM SOLUTIONS

241

err
o

rm
in im
iz a

tio n

J/w

wn wn-1

w*

We can easilty compute the intercept in this one-dimensional illustration. The equation of the line passing through (x0 , y0 and x1 , y1 ) is (xx1 )/(x1 x0 ) = (yy1 )/(y1
y0 ). The x axis crossing is found by setting y = 0 and solving for x, that is
x x1 = (x1 x0 )

y1
y1
= (x1 x0 )
.
y1 y0
y0 y1

Now we can plug in our point shown on the graph, with x0 = wn1 , y0 = J/w|n1 ,
x1 = wn , y1 = J/w|n , and by noting w(m) = (x1 x0 ) and w(n+1) = (xx1 ).
The nal result is
J/w|n
w(n + 1) =
w(n).
J/w|n1 J/w|n
Section 6.10
36. Consider the matched lter described by Eq. 66 in the text.
(a) The trial weight function w(t) = w (t) + h(t) constrained of Eq. 63 must be
obeyed:


w2 (t)dt,

w (t)dt =

and we expand to nd


w dt +


2

h (t)dt + 2

w (t)h(t)dt =

w2 (t)dt

and this implies


2

h2 (t)dt 0.

h(t)w (t)dt =

(b) We have



x(t)w(t)dt =

x(t)w (t)dt +



output z

x(t)h(t)dt.

CHAPTER 6. MULTILAYER NEURAL NETWORKS

242

From Eq. 66 in the text we know that w (t) = /2x(t), and thus x(t) =
2w (t)/(). We plug this form of x(t) into the right-hand side of the above
equation and nd


2
x(t)[w (t) + h(t)]dt = z +

x(t)w(t)dt =

w (t)h(t)dt.

From part (a) we can substitute



w (t)h(t)dt =

h2 (t)dt,

and thus the output z of the matched lter is given by:


z=

x(t)w(t)dt = z


h2 (t)dt.

(c) Plugging the above into Eq. 66 in the text,



z =

w (t)x(t)dt,

we get


x(t)x(t)dt.
2

z =

Now we check on the value of z :

2
z =
2

x2 (t)dt < 0.

Thus z and have opposite signs.

(d) Part (b) ensures z = z if and only if
1


h2 (t)dt = 0,

which occurs if and only if h(t) = 0 for all t. That is, w (t) ensures the maximum
output. Assume w1 (t) also assures maximum output z as well, that is


x(t)w (t)dt

= z

x(t)w1 (t)dt

= z.

PROBLEM SOLUTIONS

243

We subtract both sides and nd



x(t)[w1 (t) w (t)]dt = 0.

Now we can let h(t) = w1 (t) w (t). We have just seen that given the trial
weight function w1 (t) = w (t) + h(t), that h(t) = 0 for all t. This tells us that
w1 (t) w (t) = 0 for all t, or equivalently w1 (t) = w (t) for all t.
37. Consider OBS and OBD algorithms.
(a) We have from Eq. 49 in the text

J

=

=

J
w

t

1
2J
w + wt
w + O(w3 )
  
2
w2



1 t
w Hw.
2

So, it is required to minimize wt Hw, subject to the constraint of deleting one

weight, say weight q. Let uq be the unit vector along the qth direction in weight
space. Then, pre- and post-operating picks out the qq entry of a matrix, in
particular,
utq H1 uq = [H1 ]qq .
Now we turn to the problem of minimizing 12 wt Hw subject to wt uq = wq .
By the method of Lagrangian multipliers, this is equivalent to minimizing
f (w) =

1 t
w Hw (wt uq + wq ) .
2



0

This in turn implies

f
= Hw (uq ) = 0.
w
Now since the derivative of matrix products obey
t
x Ax = 2Ax
x
and
t
x b = b,
x
we have
w = +H1 uq .
To solve for we compute
f
= 0,

CHAPTER 6. MULTILAYER NEURAL NETWORKS

244

and this implies

wt uq + wq = 0.
This implies
utq H1 uq = q ,
which in turn gives the value of the Lagrange undetermined multiplier,
=

wq
t
uq H1 uq

wq
.
[H1 ]qq

So, we have
w = +H1 uq =

wq
H1 uq .
[H1 ]qq

Moreover, we have the saliency of weight q is

Lq

1
wq
wq
H1 uq ]t H[ 1 H1 uq ]
[
2 [H1 ]qq
[H ]qq

wq2
1
ut H1 HH1 uq
2 ([H1 ]qq )2 q

1 wq2
1
wqq
ut H1 uq =
[H1 ]qq
2 (H1 )2qq q
2 ([H1 ]qq )2

1 wq2
.
2 [H1 ]qq

(b) If H is diagonal, then

[H1 ]qq =

1
.
Hqq

So, from part (a), the saliency of weight q for OBD is

Lq =

1 wq2
1
= wq2 Hqq .
2 [H1 ]qq
2

38. The three-layer radial basis function network is characterized by Eq. 61 in the
text:

nH
!
zk = f
wkj j (x) ,
j=0

where we can let yj = (x). The quadratic error function we use is

1!
(tk zk )2 .
2
c

J=

k=1

We begin by nding the contribution of the hidden-to-output weights to the error:

J
J netk
=
,
wkj
netk wkj

PROBLEM SOLUTIONS

245

where
J
J zk
=
= (tk zk )f  (netk ).
netk
zk netk
We put these together and nd
J
= (tk zk )f  (netk )j (x).
wkj
Now we turn to the update rule for j and bj . Here we take
j (x) = ebj xj  .
2

Note that j is a vector which can be written 1j , 2j , . . . , dj . We now compute

J/ij :
J
J yj
=
.
ij
yj ij
We look at the rst term and nd
J
yj

c
!

(tk zk )

k=1

c
!

zk netk
netk yj

(tk zk )f  (netk )wkj .

k=1

The second term is

yj
ji

= ebxj  (1)(1)2(xi ij )
2

2(x).

We put all this together to nd

!
J
= 2(x)(xi ij )
(tk zk )f  (netk )wkj .
ji
c

k=1

Finally, we nd J/bj such that the size of the spherical Gaussians be adaptive:
J
J yj
=
.
bj
yj bj
While the rst term is given above, the second can be calculated directly:
yj
bj

2
(x)
= ebxj  x j 2 (1)
bj

= (x)x j 2 .
Thus the nal derivative is
!
J
= (x)x j 2
(tk zk )f  (netk )wkj .
bj
c

k=1

CHAPTER 6. MULTILAYER NEURAL NETWORKS

246

Now it is a simple matter to write the learning rules:

wkj
ij
bj

J
wkj
J
=
ij
J
=
,
bj
=

with the derivatives as given above.

Section 6.11
39. Consider a general d-by-d matrix K and variable d-dimensional vector x.
(a) We write in component form:
t

x Kx

d !
d
!

xi Kij xj

j=1 i=1

d
!

x2i Kii +

d
!

i=1

xi xj Kij .

i=j

The derivatives are, then,

d t
x Kx
dxr

d
d
!
d ! 2
x Kii +
xi xj Kij
dxr i=1 i
i=j

d $! 2
xi Kii + x2r Krr
dxr
i=r

d !
d
!
i

xi xj Kij + xr

xj Krj + xr

j=r

0 + 2xr Krr + 0 +

d
!

xj Krj +

j=r

d
!

xj Krj +

j=1

d
!

i=r
d
!

xi Kir

i=r

xi Kir

i=1

(Kx)rth element + (Kt x)tth element

This implies
d t
x Kx = Kx + Kt x = (K + Kt )x.
dx
(b) So, when K is symmetric, we have
d t
x Kx = (K + Kt )x = 2Kx.
dx

%
xi Kir

PROBLEM SOLUTIONS

247

40. Consider weight decay and regularization.

(a) We note that the error can be written
E = Epat +

2
wij
.

ij

Then, gradient descent in the error function is

new
old
wij
= wij

E
wij

where is the learning rate. We also have

!
E

2
Epat +
=
wij
wij
wij
ij
= 2wij .
This implies
new
wij

old
old
= wij
2wij
old
= wij
(1 2).

new
old
(b) In this case we have wij
= wij
(1 ) where is a weight decay constant.
Thus
new
old
wij
= wij
(1 2)

from part (a). This implies 2 = or

=

.
2

(c) Here we have the error function

E

E
wij

= Epat +

2
wij
2
1 + wij



1
= Epat + 1
2
1 + wij


1
2wij
Epat

1
=0+
=
+
2
2 )2
wij
wij
1 + wij
(1 + wij
2wij
=
.
2 )2
(1 + wij

The gradient descent procedure gives

new
wij

old
2wij
E
old
= wij

2 )2
wij
(1 + wij


2
old
1
= wij
2 )2
(1 + wij
old
= wij

old
= wij
(1 ij )

CHAPTER 6. MULTILAYER NEURAL NETWORKS

248
where

ij =

2.

2)
(1 + wij

This implies
2

2
ij (1 + wij
)
.
2

(d) Now consider a network with a wide range of magnitudes for weights. Then, the
new
old
constant weight decay method wij
= wij
(1) is equivalent to regularization
 2
via a penalty on
ij . Thus, large magnitude weights are penalized and will
ij

be pruned to a greater degree than weights of small magnitude. On the other

new
old
hand, the variable weight decay method wij
= wij
(1 ij ) penalized on
%
$
1
1 1+w2 , which is less susceptible to large weights. Therefore the pruning
ij

ij

will not be as severe on large magnitude weights as in the rst case.

41. Suppose that the error Epat is the negative log-likelihood up to a constant, that
is,
Epat = lnp(x|w) + constant.
Then we have
p(x|w) eEpat .
Consider the following prior distribution over weights
 2

p() e

wij

i,j

where >
0 is some parameter. Clearly this prior favors small weights, that is, p(w)
2
is large if
wij
is small. The joint density of x, w is
ij

p(x, w) = p(x|w)p(w) exp Epat

2
.
wij

ij

The posterior density is

!
p(x, w)
2
p(w|x) =
.
exp Epat
wij
p(x)
ij
So, maximizing the posterior density is equivalnet to minimizing
!
2
E = Epat +
wij
,
ij

which is the E of Eq. 41 in the text.

PROBLEM SOLUTIONS

249

42. Given the denition of the criterion function

Jef = J(w) +

 t
w w,
2

we can compute the derivative Jef /wi , and thus determine the learning rule. This
derivative is
Jef
J

=
+ wi
wi
wi

for any wi . Thus the learning rule is


J 
wi (k + 1) = wi (k)
wi (k)

wi 
wi (k)

J 
=
+ (1 )wi (k).

wi 
wi (k)

This can be written in two steps as

= wiold (k)(1 )

J 
old
wi (k + 1) =
+ winew (k).

wi  old
winew (k)

wi

(k)

Continuing with the next update we have

winew (k + 1) = wiold (k + 1)(1 )

J 
old
wi (k + 2) =
+ winew (k + 1).

wi  old
wi

(k+1)

The above corresponds to the description of weight decay, that is, the weights are
updated by the standard learning rule and then shrunk by a factor, as according to
Eq. 38 in the text.
43. Equation 70 in the text computes the inverse of (H + I) when H is initialized
to I. When the value of is chosen small, this serves as a good approximation
[H + I]1  H1 .
We write the Taylor expansion of the error function around w up to second order.
The approximation will be exact if the error function is quadratic:

J(w)  J(w ) +

 t
J 
(w w ) + (w w )t H(w w ).
w w

We plug in H + I for H and get



J (w)  J(w ) +


 t
J 
(w w ) + (w w )t H(w w ) + + (w w )t I(w w ).
w w


J(w)

250

CHAPTER 6. MULTILAYER NEURAL NETWORKS

Thus we have
t
J  (w) = J(w) + wt w w
 w .
constant

Thus our calculation of H modies the error function to be

J  (w) = J(w) + w2 constant.
This constant is irrelevant for minimization, and thus our error function is equivalent
to
J  (w) + w2 ,
which is equivalent to a criterion leading to weight decay.
44. The predicted functional increase in the error for a weight change w is given by
Eq. 68 in the text:
1 t
w Hw.
2

J 

We want to nd which component of w to set to zero that will lead to the smallest
increase in the training error. If we choose the qth component, then w must satisfy
utq w + wq = 0, where uq is the unit vector parallel to the qth axis. Since we want to
minimize J with the constraint above, we write the Lagrangian
1 t
w Hw + (utq w + wq ),
2




L(w, ) =

and then solve the system of equations

L
= 0,

L
= 0.
w

These become
L

L
wj
L
wj

= utq w + wq = 0
=

Hkj wk

for j = i

Hij wk +

for j = i.

These latter equations imply Hw = uq . Thus w = H1 ()uq . We substitute

this result into the above and solve for the undetermined multiplier :
utq H1 uq + wq = 0
which yields
=

wq
.
[H1 ]qq

PROBLEM SOLUTIONS

251

Putting these results together gives us Eq. 69 in the text:

w =

wq
H1 uq .
[H1 ]q q

For the second part of Eq. 69, we just need to substitute the expression for w we
just found for the expression for J given in Eq. 68 in the text:
t 


wq
1
wq
1
1
Lq =
1 H uq H 1 H uq
2
[H ]qq
[H ]qq

1 wq
wq
=
ut HH1
H1 uq
2 [H1 ]qq q    [H1 ]qq
I

1 wq2
.
2 [H1 ]qq

45. Consider a simple 2-1 network with bias. First let w = (w1 , w2 )t be the weights
and x = (x1 , x2 )t be the inputs, and w0 the bias.
(a) Here the error is
E=

n
!
p=1

Ep =

n
!

[tp (x) zp (x)]2

p=1

where p indexes the patterns, and tp is the teaching or desired output and zp
the actual output. We can rewrite this error as
!
!
E=
[f (w1 x1 + w2 x2 + w0 ) 1]2 +
[f (w1 x1 + w2 x2 + w0 ) + 1]2 .
xD1

xD2

(b) To compute the Hessian matrix H =

E
wi

2E
w2 ,

we proceed as follows:

! $
2 f (w1 x1 + w2 x2 + w0 ) 1]f  (w1 x1 + w2 x2 + w0 )xi
x1

2[f (w1 x1 + w2 x2 + w0 ) + 1]f  (w1 x1 + w2 x2 + w0 )xi , i = 1, 2

x2

E
wj wi


=

{f  (w1 x1 + w2 x2 + w0 )}2 xj xi

x1

[f (w1 x1 + w2 x2 + w0 ) 1]f  (w1 x1 + w2 x2 + w0 )xj xi

x1

{f  (w1 x1 + w2 x2 + w0 )}2 xj xi

x2




[f (w1 x1 + w2 x2 + w0 ) + 1]f (w1 x1 + w2 x2 + w0 )xj xi .

x2

We let the net activation at the output unit be denoted

1
tx
= w t x + w0 .
net = w1 x1 + w2 x2 + w0 = (w0 w1 w2 ) x1 = w
x2

CHAPTER 6. MULTILAYER NEURAL NETWORKS

252

Then we have
2E
wj wi

[(f  (net))2 + (f (net)1)f  (net)]xj xi

x1

+2
=

[(f  (net))2 + (f (net) + 1)f  (net)]xj xi

x2

u(x)xi xj

bf x

where


u(x) =

So, we have

(f  (net))2 + (f (net) 1)f  (net) if x 1

(f  (net))2 + (f (net) + 1)f  (net) if x 2

u(x)x21

H = x
u(x)x1 x2
x

u(x)x1 x2
x
.
u(x)x22

(c) Consider two data sets N (x|i ) N (i , 1), i = 1, 2. Then, H in terms of 1 , 2

is H = ((Hij ))22 where,
Hij

2[(f  (net1 )2 ) + (f (net1 ) 1)f  (net1 )]1i 1j

+2[(f  (net2 )2 ) + (f (net2 ) + 1)f  (net2 )]2i 2j

for i, j = 1, 2, where net1 = wt 1 + w0 and net2 = wt 2 + w0 . Clearly, the

Hessian matrix is symmetric, that is, H12 = H21 .
(d) We have the Hessian matrix

H=

H11
H12

H12
H22


.

Then, the two eigenvalues of H are the roots of |H I| = 0, that is,


 h11

h12


h12 
 = 0.
h22

Now we have
2
(H11 )(H22 ) H12
=0

and
2
2 (H11 + H22 ) + H11 H22 H12
= 0,

and this implies the eigenvalue is


2 )
H11 + H22 (H11 + H22 )2 4(H11 H22 H12
=
2

2
H11 + H22 (H11 H22 )2 + 4H12
=
.
2

PROBLEM SOLUTIONS

253

Thus the minimum and maximum eigenvalues are


2
H11 + H22 (H11 H22 )2 + 4H12
min =
 2
2
H11 + H22 + (H11 H22 )2 + 4H12
max =
.
2
(e) Suppose 1 = (1, 0) and 2 = (0, 1). Then from part (c) we have
H12

= 0

H11

= 2[(f  (w1 + w0 ))2 + (f (w1 + w0 ) 1)f  (w1 + w0 )]

H22

= 2[(f  (w2 + w0 ))2 + (f (w2 + w0 ) + 1)f  (w2 + w0 )].

From part (d) we have

min

max

H11 + H22 |H11 H22 |

2
H11 + H22 + |H11 H22 |
2

and therefore in this case the minimum and maximum eigenvalues are
min

min(H11 , H22 )

max

max(H11 , H22 ),

max
min

and thus their ratio is

max(H11 , H22 )
min(H11 , H22 )
H11
H22
or
.
H22
H11

254

CHAPTER 6. MULTILAYER NEURAL NETWORKS

Computer Exercises
Section 6.2
1. Computer exercise not yet solved
Section 6.3
2.
3.
4.
5.
6.
7.

Computer
Computer
Computer
Computer
Computer
Computer

exercise
exercise
exercise
exercise
exercise
exercise

not
not
not
not
not
not

yet
yet
yet
yet
yet
yet

solved
solved
solved
solved
solved
solved

Section 6.4
8. Computer exercise not yet solved
Section 6.5
9. Computer exercise not yet solved
Section 6.6
10. Computer exercise not yet solved
Section 6.7
11. Computer exercise not yet solved
Section 6.8
12. Computer exercise not yet solved

Chapter 7

Stochastic methods

Problem Solutions
Section 7.1
1. First, the total number of typed characters in the play is approximately
m = 50 pages 80

lines
characters
40
= 160000 characters.
page
line

We assume that each character has an equal chance of being typed. Thus, the probability of typing a specic charactor is r = 1/30.
We assume that the typing of each character is an independent event, and thus
the probability of typing any particular string of length m is rm . Therefore, a rough
estimate of the length of the total string that must be typed before one copy of
Hamlet appears is r1m = 30160000  2.52 10236339 . One year is
365.25 days 24

hours
minutes
seconds
60
60
= 31557600 seconds.
day
hour
minute

We are given that the monkey types two characters per second. Hence the expected
time needed to type Hamlet under these circumstances is
2.52 10236339 characters

1 second
1
year

 4 10236331 years,
2 character 31557600 seconds

which is much larger than 1010 years, the age of the universe.
Section 7.2
2. Assume we have an optimization problem with a non-symmetric connection matrix,
W, where wij = wji .

N
N
1
1 !
1 1 !
E =
wij si sj
wji si sj
(2E) =

2
2
2 i,j=1
2 i,j=1
255

CHAPTER 7. STOCHASTIC METHODS

256
=

N
1 ! 1
(wij + wji )si sj
2 i,j=1 2

We dene a new weight matrix

w
ij =

wij + wji
,
2

which has the property w

ij = w
ji . So the original optimization problem is equivalent

to an optimization problem with a symmetric connection matrix, W.

3. Consider Fig. 7.2 in the text.
(a) In the discrete optimization problem, the variables only take the values 1,
yielding isolated points in the space. Therefore, the energy is only dened on
these points, and there is no continuous landscape as shown in Fig. 7.2 in the
text.
(b) In the discrete space, all variables take 1 values. In other words, all of the
feasible solutions are at the corners of a hypercube, and are of equal distance
to
the middle (the center) of the cube. Their distance from the center is N .
(c) Along any cut in the feature space parallel to one of the axes the energy will
be monotonic; in fact, it will be linear. This is because all other features are held
constant, and the energy depends monotonically on a single feature. However,
for cuts not parallel to a single axis (that is, when two or more features are
varied), the energy need not be monotonic, as shown in Fig. 7.6.
4. First, we recall each variable can take two values, 1. The total number of
congurations is therefore 2N . An exhaustive search must calculate energy for all
congurations. For N = 100, the time for exhaustive search is
2100 conf igurations

108 second
 1.27 1022 seconds.
conf iguration

We can express this time as

1.27 1022 seconds

1 minute
1 hour
1 day
1 year

 4.0 1014 years,

60 seconds 60 minutes 24 hours 365 days

which is ten thousand times larger than 1010 years, the age of the universe. For
N = 1000, a simliar argument gives the time as 3.4 10285 years.
5. We are to suppose that it takes 1010 seconds to perform a single multiplyaccumulate.
(a) Suppose the network is fully connected. Sincethe connection matrix is symmetric and wii = 0, we can compute E as
wij si sj and thus the number
i>j

of multiply-accumulate operations needed for calculating E for a single conguration is N (N 1)/2. We also know that the total number of congurations
is 2N for a network of N binary units. Therefore, the total time t(N ) for an
exhaustive search on a network of size N (in seconds) can be expressed as
t(N )

second
N (N 1)
operations

2N conf igurations
operation
2
conf iguration
= 1010 N (N 1)2N 1 seconds.

1010

PROBLEM SOLUTIONS

257

log[t]
700
600
500
400
300
200
100

log[N]

(b) See figure.

(c) We need to nd the largest possible N such that the total time is less than or
equal to the specied time. Unfortunately, we cannot solve the inverse of t(N )
in a closed form. However, since t(N ) is monotonically increasing with respect
to N , a simple search can be used to solve the problem; we merely increase
N until the corresponding time exceeds the specied time. A straightforward
calculation for a reasonable N would overow on most computers if no special
software is used. So instead we use the logarithm of the expression, specically
t(N ) = 10log10

10

N (N 1)2(N 1)

= 10(N 1)log2+logN +log(N 1)10 .

A day is 86400 seconds, within which an exhaustive search can solve a network
of size 20 (because 5.69 104  t(20) < 86400 < t(21)  6.22 105 ). A year is
31557600  3.16 107 seconds, within which an exhaustive search can solve a
network of size 22 (because 7.00106  t(22) < 3.16107 < t(23)  8.10107 ).
A century is 3155760000  3.16109 seconds, within which an exhaustive search
can solve a network of size 24 (because 9.65108  t(24) < 3.16109 < t(25) 
1.18 1010 ).
6. Notice that for any conguration , E does not change with T . It is sucient to
show that for any pair of congurations and  , the ratio of the probabilities P ()
and P (  ) goes to 1 as T goes to innity:
P ()
eE /T
= lim E  /T = lim e(E  E )/T


T P ( )
T e
T
lim

Since the energy of a conguration does not change with T , E and E  are constants,
so we have:
(E  E )
T
T
lim e(E  E )/T
lim

= 0
= 1,

which means the probabilities of being in dierent congurations are the same if T is
suciently high.
7. Let kuN denote the number pointing up in the subsystem of N magnets and kdN
denote the number down.

CHAPTER 7. STOCHASTIC METHODS

258

(a) The number of congurations is given by the number of ways of selecting kuN
magnets out of N and letting them point up, that is,
 
N
K(N, EN ) =
kuN
We immediately have the following set of equations for kuN and kdN :
kuN + kdN
kuN kdN

= N
= EN

The solution is given by kuN = 12 (N + EN ) and kdN = 12 (N EN ). Therefore we

have

  
N
N
K(N, EN ) =
=
.
1
kuN
2 (N + EN )
(b) We follow the approach in part (a), and nd

K(N, EM ) =

M
kuM


=


M
.
1
2 (M + EM )

(c) Problem not yet solved

(d) Problem not yet solved
(e) Problem not yet solved
8. Consider a single magnet in an applied eld.
(a) The two states  = 0 and  = 1 correspond to the two states of the binary
magnet, s = +1 and s = 1, respectively. For  = 0 or s = +1, E  = E0 ; for
 = 1 or s = 1, E  = E0 . Therefore, according to Eq. 3 in the text, the
partition function is
Z(T ) = eE0 /T + e(E0 )/T = eE0 /T + eE0 /T .
(b) According to Eq. 2 in the text, the probability of the magnet pointing up and
pointing down are
P (s = +1) =

eE0 /T
,
Z(T )

P (s = 1) =

eE0 /T
Z(T )

respectively. We use the partition function obtained in part (a) to calculate the
expected value of the state of the magnet, that is,
E[s] = P (s = +1) P (s = 1) =

eE0 /T eE0 /T
= tanh(E0 /T ),
eE0 /T + eE0 /T

which is indeed in the form of Eq. 5 in the text.

PROBLEM SOLUTIONS

259

(c) We want to calculate the expected value of the state, E[s]. We assume the other
N 1 magnets produce an average magnetic eld. In another word, we consider
the magnet in question, i, as if it is in an external magnetic eld of the same
strength as the average eld. Therefore, we can calculate the probabilities in a
similar way to that in part (b). First, we need to calculate the energy associated
with the states of i. Consider the states of the other N 1 magnets are xed. Let
i+ and i denote the conguration in which si = +1 and si = 1, respectively,
while the other N 1 magnets take the xed state values. Therefore, the energy
for conguration i+ is

N
N
N
!
!
1 !
1!
E + =
wjk sj sk =
wkj sj sk +
wki sk +
wij sj
i
2
2
j=1
k,j=i

j,k=1

k=1

since si = +1. Recall that the weight matrix is symmetric, that is, wki = wik ,
and thus we can write
N
!

wki sk =

k=1

N
!

wik sk =

N
!

wij sj ,

j=1

k=1

where the last equality is by renaming dummy summation indices. Therefore

the energy is
E + =
i

N
!

wij sj

j=1

1 !
wjk sj sk = li + C,
2
j,k=i

where, as given by Eq. 5 in the text, li =


j

wij sj , and C = 12


j,k=i

wjk sj sk is

a constant independent of i. Similarly, we can obtain

E =
i

N
!

wij sj

j=1

1 !
wjk sj sk = li + C
2
j,k=i

Next, we can apply the results from parts (a) and (b). The partition function
is thus
E + /T

Z(T ) = e

E /T

+e

= e(li C)/T + e(li C)/T = (eli /T + eli /T )/E C/T .

Therefore, according to Eq. 2 in the text, the expected value of the state of
magnet i in the average eld is
E[si ]

= P r(si = +1)(+1) + P r(si = 1)(1) =

=

E + /T

E /T

e
Z(T )

e(li C)/T e(li C)/T

eli /T eli /T
= l /T
= tanh(li /T ).
l
/T
l
/T
C/T
i
i
(e
+e
)/E
e i + eli /T

9. The two input nodes are indexed as 1 and 2, as shown in the gure. Since the input
units are supposed to have xed state values, we can ignore the connection between
the two input nodes. For this assignment, we use a version of Boltzmann networks
with biases, in another word, there is a node with constant value +1 in the network.

CHAPTER 7. STOCHASTIC METHODS

260

Suppose the constant node is numbered as 0. The representation power of Boltzmann

networks with biases is the same as those without biases. This can be shown as follows.
First, each network without the constant unit can be converted to a network with the
constant unit by including the constant node 0 and assigning the connection weights
w0j a zero value. Second, each network with the constant unit can also be converted
to a network without the constant unit by replacing the constant unit with a pair
of units, numbered 1 and 2. Node 1 assumes the same connection weights to
other units as the constant unit 0, while node 2 is only connected to node 1. The
connection weights between nodes 1 and 2 are a vary large positive number M . In
this way, the nodes 1 and 2 are forced to have the same value in the congurations
with the lowest energy, one of which requires the value be +1.
3
3

4
5

Next, we are going to determine the set of connection weights between pairs of
units. In the follows, we use a vector of state values of all units to indicate the
corresponding conguration.
(a) For the exclusive-OR problem, the input-output relationship is as follows:
Input 1
+1
+1
1
1

Input 2
+1
1
+1
1

Output (3)
1
+1
+1
1

Since the Boltzmann network always converges to a minimum energy conguration, it solves the exclusive-OR problem if and only if the following inequalities
hold:
E{+1,+1,1}

<

E{+1,+1,+1}

E{+1,1,+1}
E{1,+1,+1}

<
<

E{+1,1,1}
E{1,+1,1}

E{1,1,1}

<

E{1,1,+1} .

Note that the energy of the three-unit system can be written as

E{s1 ,s2 ,s3 } = w13 s1 s3 w23 s2 s3 w03 s3
The above set of inequalities reduce to
+w13 + w23 + w03

< w13 w23 w03

w13 + w23 w03

+w13 w23 w03

< +w13 w23 + w03

< w13 + w23 + w03

w13 w23 + w03

< +w13 + w23 w03

PROBLEM SOLUTIONS

261

or equivalently,
w13 + w23 + w03
w23
w13
w03

< 0
< w13 + w03
< w23 + w03
< w13 + w23

(1)
(2)
(3)
(4)

From (1) + (4), we have w03 < 0, and from (2) + (3), we have 0 < w03 , which
is a contradiction. Therefore, a network of the specied form cannot solve the
exclusive-OR problem.
(b) For the two-output network, the input-output relationship is as follows:
Input 1
+1
+1
1
1

Input 2
+1
1
+1
1

Output 3
1
+1
+1
1

Output 4
+1
1
1
+1

This network solves the exclusive-OR problem if and only if

E{+1,+1,1,+1}

<

E{+1,+1,1,1}

E{+1,+1,1,+1}

<

E{+1,+1,+1,1}

E{+1,+1,1,+1}

<

E{+1,+1,+1,+1}

E{+1,1,+1,1}
E{+1,1,+1,1}

<
<

E{+1,1,1,1}
E{+1,1,1,+1}

E{+1,1,+1,1}

<

E{+1,1,+1,+1}

E{1,+1,+1,1}

<

E{1,+1,1,1}

E{1,+1,+1,1}
E{1,+1,+1,1}

<
<

E{1,+1,1,+1}
E{1,+1,+1,+1}

E{1,1,1,+1}

<

E{1,1,1,1}

E{1,1,1,+1}

<

E{1,1,+1,1}

E{1,1,1,+1}

<

E{1,1,+1,+1} .

Recall the energy for the four-unit network:

E{s1 ,s2 ,s3 ,s4 } = w13 s1 s3 w14 s1 s4 w23 s2 s3 w24 s2 s4 w34 s3 s4 w03 s3 w04 s4 .
The above set of inequalities reduce to:
+w13 w14 + w23 w24 + w34 + w03 w04
+w13 w14 + w23 w24 + w34 + w03 w04

< +w13 + w14 + w23 + w24 w34 + w03 + w04

< w13 + w14 w23 + w24 + w34 w03 + w04

+w13 w14 + w23 w24 + w34 + w03 w04

w13 + w14 + w23 w24 + w34 w03 + w04

< w13 w14 w23 w24 w34 w03 w04

< +w13 + w14 w23 w24 w34 + w03 + w04

w13 + w14 + w23 w24 + w34 w03 + w04

< +w13 w14 w23 + w24 + w34 + w03 w04

w13 + w14 + w23 w24 + w34 w03 + w04

+w13 w14 w23 + w24 + w34 w03 + w04

< w13 w14 + w23 + w24 w34 w03 w04

< w13 w14 + w23 + w24 w34 + w03 + w04

+w13 w14 w23 + w24 + w34 w03 + w04

< w13 + w14 + w23 w24 + w34 + w03 w04

CHAPTER 7. STOCHASTIC METHODS

262

+w13 w14 w23 + w24 + w34 w03 + w04

w13 + w14 w23 + w24 + w34 + w03 w04

< +w13 + w14 w23 w24 w34 w03 w04

< w13 w14 w23 w24 w34 + w03 + w04

w13 + w14 w23 + w24 + w34 + w03 w04

< +w13 w14 + w23 w24 + w34 w03 + w04

w13 + w14 w23 + w24 + w34 + w03 w04

< +w13 + w14 + w23 + w24 w34 w03 w04

or equivalently
w34
w13 + w23 + w03
w13 + w23 + w34 + w03
w23 + w34
w14 + w23 + w04
w14 + w34 + w04
w13 + w34
w13 + w24 + w04
w24 + w34 + w04
w14 + w24 + w34
w14 + w24 + w03
w34 + w03

<
<
<
<
<
<
<
<
<
<
<
<

w14 + w24 + w04

w14 + w24 + w04
0
w13 + w03
w13 + w24 + w03
w24
w23 + w03
w14 + w23 + w03
w14
w04
w13 + w23 + w04
w13 + w23 .

(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)

From (2) + (11), we have the constraint w03 < w04 , and from (5) + (8), we
have w04 < w03 . These two inequalities form a contradiction and therefore the
network cannot solve the exclusive-OR problem.
(c) We consider the assignments as follows:
Input 1
+1
+1
1
1

Input 2
+1
1
+1
1

Output 3
1
+1
+1
1

Output 4
+1
1
1
+1

Hidden 5
+1
1
1
1

For the network to solve the exclusive-OR problem, it must be the case that
E{+1,+1,1,+1,+1}

E{+1,+1,1,+1,1}
< E{+1,+1,1,1,1}

E{+1,+1,1,+1,+1}
E{+1,+1,1,+1,+1}

<
<

E{+1,1,+1,1,1}

E{+1,1,+1,1,+1}

E{+1,1,+1,1,1}

<

E{+1,1,1,1,1}

E{+1,1,+1,1,1}
E{+1,1,+1,1,1}

<
<

E{+1,1,1,+1,1}
E{+1,1,+1,+1,1}

E{1,+1,+1,1,1}
E{1,+1,+1,1,1}

E{1,+1,+1,1,+1}
< E{1,+1,1,1,1}

E{1,+1,+1,1,1}

<

E{1,+1,+1,1,1}
E{1,1,1,+1,1}

< E{1,+1,+1,+1,1}
E{1,1,1,+1,+1}

E{1,1,1,+1,1}

<

E{+1,+1,1,+1,+1}

E{+1,+1,+1,1,1}
E{+1,+1,+1,+1,1}

E{1,+1,1,+1,1}

E{1,1,1,1,1}

PROBLEM SOLUTIONS

263

E{1,1,1,+1,1}
E{1,1,1,+1,1}

<
<

E{1,1,+1,1,1}
E{1,1,+1,+1,1}

The set of inequalities can be simplied as follows

w35
w34
w34 + w35
w03 + w13 + w23 + w35
w03 + w13 + w23
w03 + w13 + w23 + w34 + w35
w03 + w13 + w23 + w34
w05 + w15 + w35
w05 + w15 + w23 + w34
w23 + w34 + w35
w04 + w05 + w14 + w15 + w23
w04 + w14 + w23 + w35
w04 + w05 + w14 + w15 + w34 + w35
w04 + w14 + w34
w05 + w25 + w35
w05 + w13 + w25 + w34
w13 + w34 + w35
w04 + w05 + w13 + w24 + w25
w04 + w13 + w24 + w35
w04 + w05 + w24 + w25 + w34 + w35
w04 + w24 + w34
w05 + w45
w05 + w14 + w24 + w34
w14 + w24 + w34 + w45
w03 + w05 + w14 + w24
w03 + w14 + w24 + w45
w03 + w05 + w34 + w45
w03 + w34

<
<
<
<
<
<

<
<
<
<
<
<

<
<
<
<
<
<

<
<
<
<
<
<

w05 + w15 + w25 + w45

w04 + w14 + w24 + w45
w04 + w05 + w14 + w15 + w24 + w25
w04 + w14 + w24 + w45
w04 + w05 + w14 + w15 + w24 + w25
0
w05 + w15 + w25 + w45
w25 + w45
w03 + w13 + w25 + w45
w03 + w13
w03 + w13 + w24 + w25
w03 + w13 + w24 + w45
w24 + w25
w24 + w45
w15 + w45
w03 + w15 + w23 + w45
w03 + w23
w03 + w14 + w15 + w23
w03 + w14 + w23 + w45
w14 + w15
w14 + w45
w15 + w25 + w35
w04 + w15 + w25 + w35
w04
w04 + w13 + w15 + w23 + w25
w04 + w13 + w23 + w35
w13 + w15 + w23 + w25
w13 + w23 + w35

Consider the set of weights w05 = w35 = 1, w15 = w25 = 1, w03 = 1/2, w13 =
w23 = w45 = 1/2, w14 = w24 = 1/4, w04 = 1/4, w34 = 0. The above set of
inequalities reduce to

1
1/2
1/2
1/2
1
1/2
1/2
1/2
2
1/2

<
<
<
<
<
<

<
<

3/2
1/4
3/2
0
3/4
3/2
1/4
1
13/4
0

(1)
(4)
(7)
(10)
(13)
(16)
(19)
(22)
(25)
(28)

0
1/2
1
1/2
0
1/2
2
3/2
1/2

<
<

<
<
<
<
<
<

1/4
3/4
3/2
3/4
1/4
0
3/4
5/4
1/4

(2)
(5)
(8)
(11)
(14)
(17)
(20)
(23)
(26)

1
1/2
1/2
1/2
1
1/2
0
0
1

<
<
<
<

<
<
<
<

3/4
0
3/2
1/4
3/2
3/4
1/4
1/4
3

(3)
(6)
(9)
(12)
(15)
(18)
(21)
(24)
(27)

(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
(25)
(26)
(27)
(28)

CHAPTER 7. STOCHASTIC METHODS

264

Since all these inequalities hold, the network can indeed implement the exclusiveOR problem.
10. As with the solution to Problem 9, we consider networks with biases (that is, a
constant unit). The two input nodes are indexed as 1 and 2, the output is 4, and the
hidden unit is 3, as shown in the gure.
4

3
1

We consider the assignments of all nodes as follows:

Input 1
+1
+1
1
1

Input 2
+1
1
+1
1

Hidden (3)
+1
1
1
1

Output (4)
1
+1
+1
1

For the network to solve the exclusive-OR problem, it must be the case that
E{+1,+1,+1,1}

E{+1,+1,1,1}

E{+1,+1,+1,1}

<

E{+1,+1,+1,+1}

E{+1,+1,+1,1}

<

E{+1,+1,1,+1}

E{+1,1,1,+1}

E{+1,1,+1,+1}

E{+1,1,1,+1}

<

E{+1,1,1,+1}
E{1,+1,1,+1}

< E{+1,1,1,1}
E{1,+1,+1,+1}

E{1,+1,1,+1}
E{1,+1,1,+1}

<
<

E{1,1,1,1}
E{1,1,1,1}

E{1,1,+1,1}
< E{1,1,+1,+1}

E{1,1,1,1}

<

E{+1,1,+1,1}

E{1,+1,+1,1}
E{1,+1,1,1}

E{1,1,1,+1}

Consider the set of weights w03 = w34 = 1, w13 = w23 = 1, w04 = 1/2, and
w14 = w24 = 1/2. The energies involved in the above set of inequalities are
3
2
5
+
2
1

2
1

2
3

E{+1,+1,+1,1}

E{+1,+1,1,1}

E{+1,+1,+1,+1}

E{+1,+1,1,+1}

E{+1,1,1,+1}

PROBLEM SOLUTIONS

265
5
2
1

2
1

2
3

2
5
+
2
1

2
1

2
7

2
1
+
2
11
+
2
5

E{+1,1,+1,+1}

= +

E{+1,1,+1,1}

E{+1,1,1,1}

E{1,+1,1,+1}

E{1,+1,+1,+1}

E{1,+1,+1,1}

E{1,+1,1,1}

E{1,1,1,1}

E{1,1,+1,1}

E{1,1,+1,+1}

E{1,1,1,+1}

So, the set of inequalities hold, and the network can indeed implement the solution
to the exclusive-OR problem.
Section 7.3
11. We use the denition of Kullback-Leibler divergence from Eq. 12 in the text as
follows:
DKL (Q(o |i ), P (o |i )) =

Q(i )

Q(o |i ) log

Q(o |i )
.
P (o |i )

Recall that from Bayes rule, we have

Q(o |i ) =

Q(o , i )
Q(i )

and

P (o |i ) =

P (o , i )
.
P (i )

Thus the Kullback-Leibler divergence in this case can be written

DKL (Q(o |i ), P (o |i ))
!
!
Q(o |i )
=
Q(i )
Q(o |i ) log
P (o |i )
i
o

Q(i )

Q( )

!!
i

Q(o |i ) log

!
o

Q(o , i )P (i )
P (o , i )Q(i )

Q(o , i )
Q(i )
Q( | ) log

log
P (o , i )
P (i )
o

Q(i )Q(o |i ) log

Q(o , i ) !
Q(i ) !
i
Q(
)
log
Q(o |i )

P (o , i )
P (i ) o
i

CHAPTER 7. STOCHASTIC METHODS

266
=

Q(o , i ) log

i ,o

Q(o , i ) !
Q(i )
i

.
Q(
)
log
P (o , i )
P (i )
i

Using the denition of DKL in Eq. 7 in the text, we have

DKL (Q(o |i ), P (o |i )) = DKL (Q(o , i ), P (o , i )) DKL (Q(i ), P (i ))
According to Eq. 13 in the text, the learning rule is given by
wij

DKL (Q(o |i ), P (o |i ))
wij


DKL (Q(o , i ), P (o , i )) DKL (Q(i ), P (i ))
=

wij
wij
=

Then we can use Eqs. 8, 9 and 10 in the text to get the learning rule as follows:
wij

=
=


(EQ [si sj ]i o clamped E[si sj ]f ree ) (EQ [si sj ]i


EQ [si sj ]i o clamped EQ [si sj ]i clamped

clamped


E[si sj ]f ree )

which is indeed in the form of Eq. 14 in the text.

12. Let the constant factor in Eq. 15 in the text be 1. We have that, if i = j,
wij =

K
1 !
si (xk )sj (xk ).
K
k=1

Here, we have K = 3. The weight matrix can be

0 1
3 1
1
0
1
1

1
3
1
0
1
W=
1
1
1
0
3

1 1 1 1
3
1 3
1

calculated as

1 3
1
1

1 3
.
1
1

0
1
1
0

(a) First, we determine the energy change due to the perturbation. Let xkl denote
the pattern obtained by perturbing unit l in the pattern xk . According to Eq. 1,
the energy change is
E(xk , l) = Exk Exk
l
1!
1!
=
wij si (xkl )sj (xkl ) +
wij si (xk )sj (xk )
2 ij
2 ij
=

1!
wij (si (xkl )sj (xkl ) si (xk )sj (xk ))
2 ij

Since each unit only takes 2 possible values 1, we have sl (xkl ) = sl (xk ). We
also have si (xkl ) = si (xk ) if i = l since we only perturb one unit at a time. So,
we can cancel most of the terms in the above summation.

PROBLEM SOLUTIONS

267

Also noting wij = wji , we have

E(xk , l)

wil (si (xkl )sl (xkl ) si (xk )sl (xk ))

i=l

wil si (xk )(sl (xk ) sl (xk ))

i=l

2sl (xk )

wil si (xk )

i=l

Therefore, the change in energy obeys

E(xk , l) l = 1 l = 2 l = 3 l = 4 l = 5 l = 6
x1

14
3

x2

14
3

x3

14
3

2
3

14
3

14
3

14
3

2
3

14
3

2
3

14
3
14
3

Therefore, none of the three patterns give local minima in energy.

denote the pattern in which si (
(b) Let x
x) = si (x). We have
Ex =

1!
1!
1!
wij si (
x)sj (
x) =
wij (si (x))(sj (x)) =
wij si (x)sj (x) = Ex
2 ij
2 ij
2 ij

also gives a local minimum

Therefore, if x gives a local minimum in energy, x
in energy.
13. As in Problem 12, the symmetric

0 1
3
1
0
1

3 1
0

1
1
3
1
W=
3
1 1 1
3
1 3

1 3
1
1
3 1

connection matrix is
1
3
1
0
1
1
3
3

1
1
1
1
0
1
1
1

3
1
1 3
3
1
1
3
1
1
0 1
1
0
1 3

1
3
1
3
1
1
3
0

We thus need merely to check whether the patterns give local minima in energy. We
have
E(xk , l) l = 1 l = 2 l = 3 l = 4 l = 5 l = 6 l = 7 l = 8
x1

14
3

14
3

14
3

14
3

14
3

x2

26
3

26
3

2
3

26
3

26
3

x3

26
3

22
3

26
3

22
3

26
3

22
3

22
3

2
3

CHAPTER 7. STOCHASTIC METHODS

268

Therefore, x1 and x2 give local minima in energy, but x3 does not.

14. For simplicity, we consider there is only one pattern with a single missing feature.
Suppose the decient input pattern is d , the missing feature is represented by sk ,
and a pattern of the hidden units is . If after annealing with clamped input pattern
d , the values of sk and are sk and , respectively. We need to show that P (sk =
sk |d ) > P (sk = sk |d ). We consider the ratio of the two probabilities


P (sk = sk , |d )
P (d , sk = sk , |d )
d
P (sk = sk | )

=
=
.
P (sk = sk |d )
P (sk = sk , |d )
P (d , sk = sk , |d )

Since the pattern {d , sk = sk , } gives the minimal energy due to annealing, supposing there is only one global minimal corresponding to clamped d , we have for any
, that
E{d ,sk =sk , } < E{d ,sk =sk ,= }

and E{d ,sk =sk , } < E{d ,sk =sk ,} .

Since for any conguration including d as a subpattern, it holds that

P (|d ) = 
sk ,

eE /T
E{d ,s

k ,}

/T

Therefore, the following three limits hold:

lim P (d , sk = sk , |d )

lim P (d , sk = sk , = |d )

lim P (d , sk = sk , |d )

0.

T 0
T 0

T 0

So, for any 0 <  < 1/2, there exists a temperature T0 , such that if T < T0 ,
P (d , sk = sk , |d ) > 1 .
Therefore, under the same conditions we have the two inequalities:
!
P (d , sk = sk , |d ) < 
=

P (d , sk = sk , |d ) < .

Therefore the ratio of the probability of sk = sk to that for sk = sk , all given the
decient information d , is

P (d , sk = sk , |d )
d
P (sk = sk | )

= 
P (sk = sk |d )
P (d , sk = sk , |d )


P (d , sk = sk , |d ) +
P (d , sk = sk , |d )
=

P (d , sk = sk , |d )

>

1
> 1.


PROBLEM SOLUTIONS

269

So sk is the most probable value given d .

15. Assume the correct category is represented by sk , the input pattern is i , and
the known incorrect catagory is represented by sj , sj = 1, j = k. We consider the
ratio between probabilities P (sk = +1|i , sj = 1) and P (sk = +1|i ):
P (sk = +1|i , sj = 1)
P (sk = +1, i , sj = 1)
P (i )
=
i
i
i
P (sk = +1| )
P (sk = +1, )
P ( , sj = 1)
For a successfully trained network for classication, sk = +1 implies sj = 1. Therefore, P (sk = +1, i , sj = 1) = P (sk = +1, i ), and
P (sk = +1|i , sj = 1)
P (i )
1
=
=
1,
i
i
P (sk = +1| )
P ( , sj = 1)
P (sj = 1|i )
and hence the claim is veried.
16. Problem not yet solved
17. Here we have the variance in response to the force is

N
N
!
!
Var[li ] = Var
wij sj =
Var[wij sj ] = N Var[wij sj ].
j=1

j=1

Since Var[x] = E[x2 ] (E[x])2 , we can write

2 2
2
Var[wij sj ] = E[wij
sj ] (E[wij sj ])2 = E[wij
]E[s2j ] (E[wij ]E[sj ])2 .

Since P (sj = +1) = P (sj = 1) = 0.5, the expected value of sj is E[sj ] = 0, and
s2j = 1. Therefore, we have
2
Var[wij sj ] = E[wij
],

and thus
2
Var[li ] = N E[wij
].

We seek to have the variance of the response to be 1, that is Var[li ] = 1. which implies
2
E[wij
] = 1/N . If we randomly initialize wij according to a uniform distribution from
[h, h], we have
h
 h
1 2
1 t3 
h2
2
E[wij ] =
=
t dt =
.

2h 3 h
3
h 2h
Under these conditions, then,
1
h2
=
,
N
3

h=

3/N



So, the weights are initialized uniformly in the range 3/N < wij < + 3/N .
18. We consider the problem of setting the learning rate for a Boltzmann network
having N units.
(a) In the solution to Problem 11, we showed that the Kullback-Leibler divergence
obeyed
DKL (Q(o |i ), P (o |i )) = DKL (Q(i , o ), P (i , o )) DKL (Q(i ), P (i )).

CHAPTER 7. STOCHASTIC METHODS

270

As such, we need merely derive the result for DKL . From Eq. 7 in the text, we
obtain
DKL (Q(), P ())

Q()log

Q()
P ()

Q()logQ()

Q()logP ().

Therefore, we have
! Q() P ()
DKL (Q(), P ())
=
,
wuv
P () wuv

which yields the Hessian matrix as

H =
=

2 DKL (Q(), P ())

wij wuv
! Q() P () P ()

P ()2 wij wuv

! Q() 2 P ()
.
P () wij wuv

If the network has sucient parameters, as the number of training examples goes
to innity, the weights converge to the optimal one, w , where P () = Q().
At this limit, the second term in the above expression can be omitted, because
! Q() 2 P ()
P () wij wuv

! 2 P ()
wij wuv

1
2
!
2
=
P ()
wij wuv

21
= 0.
wij wuv

Therefore, the Hessian is simply

H

! Q() P () P ()
P ()2 wij wuv

From Eq. 9, we have

1 P ()
1 !
=
si ()sj ()P (|) E[si sj ] ,
P () wij
T

and thus we can write the Hessian as

! Q() P ()
P ()2 wij

!
!
1 !
=
Q()
si ()sj ()P (|) E[si sj ]
su (  )sv (  )P (  |) E[su sv ]
T2


H 

PROBLEM SOLUTIONS
=

271


!
!
1 !
Q()
si ()sj ()P (|)
su (  )sv (  )P (  |)
2
T



EQ [si sj ] E[su sv ] E[si sj ]EQ [su sv ] + E[si sj ]E[su sv ]


!
1 !
Q()
si ()sj ()su (  )sv (  )P (,  |)
2
T


,

EQ [si sj ] E[su sv ] E[si sj ]EQ [su sv ] + E[si sj ]E[su sv ]


1 
EQ [si sj su sv ] EQ [si sj ] E[su sv ] E[si sj ]EQ [su sv ] + E[si sj ]E[su sv ] ,
2
T

where EQ [] indicates the expected value with clamped averaged according

to Q. We can further simplify the above expression by noticing that when the
weights converge to w , P () = Q(). Thus, the Hessian is
H 




1 
EQ [si sj su sv ] EQ [si sj ] E[su sv ] E[si sj ]EQ [su sv ] + E[si sj ]E[su sv ]
T2

1 
E[si sj su sv ] E[si sj ]E[su sv ] E[si sj ]E[su sv ] + E[si sj ]E[su sv ]
T2

1 
E[si sj su sv ] E[si sj ]E[su sv ] .
T2

Recall that under general conditions

E[xy] E[x]E[y] = E[(x E[x])(y E[y])].
In this case, then, the Hessian matrix is
H=

1
,
T2

where is the covariance matrix of the random vector s = {si sj }.

(b) We calculate the curvature as
wt Hw

=
=
=
=
=
=
=

1 t
1
w w = 2 wt E[(s E[s ])(s E[s ])t ]w
T2
T
1
E[wt (s E[s ])(s E[s ])t w]
T2
1
1
E[(w (s E[s ]))2 ] = 2 E[(w s w E[s ])2 ]
T2
T
1
E[(w s )2 2(w s )(w E[s ]) + (w E[s ])2 ]
T2

1 
E[(w s )2 ] 2E[(w s )(w E[s ])] + E[(w E[s ])2 ]
T2

1 
E[(w s )2 ] 2(w E[s ])2 + (w E[s ])2
T2

1 
E[(w s )2 ] (w E[s ])2
T2

CHAPTER 7. STOCHASTIC METHODS

272

2
1
1 !

E[(w s )2 ] = 2 E
wij si sj
2
T
T
ij

2
2
!
!
1
1

E
|wij | = 2
|wij |
T2
T
ij
ij

(c) According to the inequality



we can write

a1 + a2 + + an
n

ij

|wij |

2

2

N (N 1)

2
!

|wij |

a21 + a22 + + a2n

,
n

ij

|wij |2

N (N 1)

1
N (N 1)

N (N 1).

ij

Therefore, the curvature is bounded by

wt Hw

N (N 1)
.
T2

(d) The optimal learning rate, , is inverse to the curvature, and hence
=

1
T2

.
wt Hw
N (N 1)

Section 7.4
19. It is sucient to show that Aij and Bij do not have a property which can be
translated into the constraint that the probabilities sum to 1. For the transition
probalities aij and bjk in HMM, we have
!
!
aij = 1,
bjk = 1
j

According to Eq. 23, we have

aij = eAij /T ,

bjk = eBjk /T

Therefore, if there exists a HMM equivalent for any Boltzmann chain, Aij and Bjk
must satisfy the constraints
!
!
eAij /T = 1,
eBjk /T = 1,
j

which are not required for a general Boltzmann chain.

PROBLEM SOLUTIONS

273

20. Since at each time step, there are exactly one hidden unit and one visible unit
which are on, the number of congurations for each time step is ck, and total number
of legal paths is (ck)Tf .
21. With a known initial (t = 1) hidden unit, the Boltzmann chain is annealed with
all visible units and the known initial hidden units clamped to +1, and other initial
hidden units clamped to 0. If the initial hidden unit is unknown, we can hypothetically
add an always-on hidden unit at t = 0, and connect it to all hidden units at t = 1,
and the connection weights are Ci . Since in a Boltzmann chain exactly one hidden
unit and one visible unit can be on, it follows that exactly one initial (t = 1) hidden
unit will be on, which is consistent to the semantics of Hidden Markov Models. The
generalized energy then is
Tf 1

E V = E[

Tf

, V ] = Ci
Tf

!
t=1

Aij

Tf
!

Bjk ,

t=1

where Ci indicates the connection weight between the constant unit and the initial
hidden unit which is on.
22. In a Boltzmann zipper, the hidden units are clustered into cells. For example,
in the one shown in Fig. 7.12 in the text, the fast hidden units at the rst two
time steps for the fast chain and the slow hidden units at the rst time step for
the slow chain are interconnected through E, forming a cell consisting of three
groups of hidden units. Since we require that exactly one hidden unit is on in each
chain. In a cell, three units are on after annealing, one in each group, and they are
all interconnected. Therefore, the connection matrix E cannot be simply related to
transition probabilities for two reasons. First, since each chain represents behavior
at dierent time scales, there is no time ordering between the on-unit in the slow
chain and the two on-units in the fast chain in the same cell (otherwise the slow
chain can be squeezed into the fast chain and forming a single chain). However,
it is necessary to have explicit time ordering for a state transition interpretation to
make sense. Second, even though we can exert a time ordering for the on-units, the
rst on-unit in the ordering is connected to two other on-units, which cannot have a
state transition interpretation where only a single successor state is allowed.
As an extreme case, an all-zero E means the two chains are not correlated. This
is completely meaningful, and E is not normalized. The weights in E can also be
negative to indicate a negative or suppressive correlation.
Section 7.5
23. We address the problem of a population of size L of N -bit chromosomes.
(a) The total number of dierent N -bit chromosomes is 2N . So the number of
dierent populations of size L is the number of dierent ways of selecting L
chromosomes out of 2N possible choices, allowing repetition. Since the number
of
combinations
of selection n items out of m types (n < m) with repetition is
m+n1

, we have the number of dierent populations of size L is
n


2N + L 1
L


=


L + 2N 1
.
2N 1

Next, we show that the number of combinations

of selection n items out of m

types (n < m) with repetition is m+n1
.
Denote
the number as A(m, n).
n

CHAPTER 7. STOCHASTIC METHODS

274

Suppose there are k distinctive

in the result, 1 k n. The number of

mtypes
,
and
the number of ways of choosing n items
ways of choosing the types
is
k

of exactly k types is m
B(n,
k),
where
B(n,
k) is the number of ways of parition
k
n into k positive integers. Thus the total number of ways of choosing n items
out of m types (n < m) is the sum of the above number over k
n  
!
m
B(n, k).
k
k=1

Here B(n, k) can be calculated recursively as follows:

B(n, 1)
B(n, n)

=
=

B(n, k)

1
1
nk+1
!

B(n i, k 1).

i=1

The rst two equations are obvious. The third one is due to the fact that the
last element can take values ranging from 1 to n (k 1) and thus
B(n, k)

nk+1
!

B(n i, k 1)

i=1

= B(n 1, k 1) +

nk+1
!

B(n i, k 1)

i=2

= B(n 1, k 1) +

nk
!

B(n 1 j, k 1)

j=1

= B(n 1, k 1) + B(n 1, k).

Recall the following relations:
 

 

n
n1
n1
=
+
k
k1
k
 
 
n
n
=
= 1.
0
n


Using these relations we can now prove B(n, k) = n1
If k = 1
k1 by induction.


or k = n, the result holds trivially. Suppose B(n 1, k 1) = n2
and
k2
n2
B(n 1, k) = k1 . Then we have

 
 

n2
n2
n1
B(n, k) = B(n 1, k 1) + B(n 1, k) =
+
=
.
k2
k1
k1
Therefore we have
n  
!
m
k=1

B(n, k)

=
=


n  
!
m n1
k=1
n 
!
k=1

as desired, and the proof is complete.

k
m
k



k1
n1
nk


=


m+n1
,
n

PROBLEM SOLUTIONS

275

(b) Assume each chromosome has a distinctive tness score. Since chromosomes
with the highest Ls tness scores are selected as parents, the number of all
possible sets of parents is equivalent to the number of ways of selection Ls
chromosomes from all possible chromosomes excluding those with the lowest
L Ls tness scores. Therefore, the number is

  N

Ls + (2N (L Ls )) 1
2 L + 2Ls 1
=
.
(2N (L Ls )) 1
2N L + Ls 1
(c) If Ls = L, the above expression is simply
 N

2 +L1
,
2N 1
the same as the result in part (a).
(d) If Ls = 1, the expression is
 N

  N
2 L+1
2 L+1
= 2N L + 1.
=
2N L
1
Section 7.6
24. See figure.
a)

b)

X4

SQRT

SQRT

X8

SIN

X4

d)

X0

c)

X8

X8

X0

X0 TAN

3.4

SQRT

X4

X5

X8

SQRT

X4

(a) (* (+ X4 X8) (SQRT 5))

(b) (SQRT (* X0 (+ X4 X8)))
(c) (* (- (SIN X0) (* (TAN 3.4) (SQRT X4))) X8).
(d) (* (/ X0 (+ X4 X8)) (- X5 (SQRT 5))).
(e) Constants and variables {X3, X0, 5.5, X5, -4.5, 2.7}, binary operators {+,
NOR, *, /, -, OR}, and unary operators {+, SQRT, SIN, -, NOT, TAN}. Note
+ and - can act as both binary and unary operators.

276

CHAPTER 7. STOCHASTIC METHODS

Computer Exercises
Section 7.2
1. Computer exercise not yet solved
2. Computer exercise not yet solved
3. Computer exercise not yet solved
Section 7.3
4. Computer exercise not yet solved
5. Computer exercise not yet solved
6. Computer exercise not yet solved
Section 7.4
7. Computer exercise not yet solved
Section 7.5
8. Computer exercise not yet solved
Section 7.6
9. Computer exercise not yet solved

Chapter 8

Nonmetric methods

Problem Solutions
Section 8.2
1. Assume a particular query occurs more than once along some path through a given
tree. The branch taken through the uppermost (closest to the root) occurence of this
query must also be taken in all subsequent occurences along this path. As a result,
branch(es) not taken in the uppermost instance of the query may be pruned from
lower occurences of that query along the path without altering the classication of
any pattern. After such pruning, the lower occurences of the query are left with only
a single child node. At this point, these queries serve no purpose and may be deleted
by directly connecting their parent node to their child.
By repeatedly applying the above algorithm to every possible path through a tree
and to every query which is duplicated along a path, we can convert any given tree
to an equivalent tree in which each path consists of distinct queries.
Section 8.3
2. Consider a non-binary tree, where the number of branches at nodes can vary
throughout the tree.
(a) Suppose our tree has root node R, with (local) branching ratio B. Each of the
B children is itself a root of a subtree having arbitrary depth. If we can prove
that we can replace the root node with binary nodes, then by induction we can
do it for its children the roots of the subtrees. In this way, an arbitrary tree
can be expressed as a binary tree.
We represent a tree with a root node R and B children nodes as {R, (a1 , a2 , . . . aB )}.
B = 2 This is a binary tree.
B = 3 The tree {R, (a1 , a2 , a3 )} can be represented as a binary tree {R, (a1 , R2 )}
where the subtree is {R2 , (a2 , a3 )}.
277

CHAPTER 8. NONMETRIC METHODS

278

B = k 3 We can keep replacing each root of the subtree as follows: the rst
{R, (a1 , a2 , . . . , ak )} becomes {R, (a1 , R2 )}, with subtree {R2 , (a2 , R3 )},
. . ., {Rk1 , (ak1 , ak )}. By induction, then, any tree can be replaced by
an equivalent binary tree.
While the above shows that any node with B 2 can be replaced by a node
with binary decisions, we can apply this to all nodes in the expanded tree, and
thereby make the entire tree binary.
(b) The number of levels depends on the number of classes. If the number of classes
is 2, then the functionally equivalent tree is of course 2 levels in depth. If the
number of categories is c, we can create a tree with c levels though this is not
needed. Instead we can split the root node sending c/2 categories to the left,
and c/2 categories to the right. Likewise, each of these can send c/4 to the left
and c/4 to the right. Thus a tight upper bound is logc.
(c) The lower bound is 3 and the upper bound is 2B 1.
3. Problem not yet solved
4. Problem not yet solved
5. We use the entropy impurity given in Eq. 1,
!
i(N ) =
P (i ) log2 (P (i )) = H(i ).
i

(a) After splitting on a binary feature F {R, L}, the weighted impurity at the
two child nodes is
!
P (L)i(L) + P (R)i(R) = P (L)
P (i |L) log2 (P (i |L))
i

P (R)

P (i |R) log2 (P (i |R))


P (i , L)
=
P (i , L) log2
P (L)
i


!
P (i , R)

P (i , R) log2
P (R)
i
!
=
P (i , F ) log2 (P (i , F ))
!

i,F

+P (L) log2 (P (L)) + P (R) log2 (P (R))

= H(, F ) H(F ).
Therefore, the drop in impurity is
i(N ) = H() H(, F ) + H(F ).
But H() H(, F ) H() + H(F ), and therefore we have
0 i(N ) H(F ) 1 bit.
(b) At each node, the weighted impurity at the child nodes will be less than that at
the parent, even though individual descendandant may have a greater impurity

PROBLEM SOLUTIONS

279

than their parent. For example at the (x2 < 0.61) node in the upper tree of
example 1, the impurity is 0.65, while that at the left child is 0.0, and that at
the right is 1.0. The left branch is taken 23 of the time, however, so the weighted
impurity at the children is 23 0 + 13 1 = 0.33. Similarly, at the (x1 < 0.6)
node of the lower tree, the right child has higher impurity (0.92) than the parent
(0.76), but the weighted average at the children is 23 0 + 13 0.92 = 0.304. In
each case, the reduction in impurity is between 0 and 1 bit, as required.
(c) For B-way branches, we have 0 i(N ) log2 (B) bits.
6. Problem not yet solved
7. Problem not yet solved
8. There are four attributes, {a1 , a2 , a3 , a4 } to be used in our decision tree.
(a) To select the query at the root node, we investigate queries on each of the four
attributes. The following shows the number of patterns sent to the left and
right for each value, and the entropy at the resulting children nodes:
query
a1
a2
a3
a4

sent left
21 , 22
21 , 22
01 , 22
21 , 32

left entropy
1
1
0
0.9710

sent right
21 , 22
21 , 22
41 , 22
21 , 12

right entropy
1
1
0.9183
0.9183

Because query a3 leads to the greatest weighted reduction in impurity, a3 should

be the query at the root node. We continue and grow the tree shown in the
gure.
a3
0

a1

2
0

a2

1
0

a4

a4
0

2 2

(b) We can expand the tree into rules as

1

(a3 AN D N OT a1 ) OR (a3 AN D a1 AN D N OT a2 AN D N OT a4 )
OR (a3 AN D a1 AN D a2 AN D a4 )
= a3 AN D (N OT a1 OR a1 AN D (N OT a2 AN D N OT a4 ) OR (a2 AN D a4 )).
= N OT a3 OR (a3 AN D a1 AN D N OT a2 AN D a4 )
OR (a3 AN D a1 AN D a2 AN D N OT a4 )
= N OT a3 OR(a3 AN D a1 ) AN D ((N OT a2 AN D a4 ) OR (a2 AN D N OT a4 )).

9. Problem not yet solved

10. Problem not yet solved
11. Problem not yet solved

CHAPTER 8. NONMETRIC METHODS

280
12.
13.
14.
15.
16.

Problem not yet solved

Problem not yet solved
Problem not yet solved
Problem not yet solved
The four attributes are denoted {a1 , a2 , a3 , a4 }.

(a) To select the query at the root node, we investigate queries on each of the four
attributes. The following shows the number of patterns sent to the left and
right for each value, and the entropy at the resulting children nodes:
query
a1
a2
a3
a4

sent left
21 , 12
31 , 02
21 , 12
31 , 22

left entropy
0.9183
0
0.9183
0.9710

sent right
21 , 32
11 , 42
21 , 32
11 , 22

right entropy
0.9710
0.7219
0.9710
0.9183

Because query a2 leads to the greatest weighted reduction in impurity, a2 should

be the query at the root node. We continue and grow the tree shown in the
gure, where the ? denotes the leaf having equal number of 1 and 2 patterns.
a2
0

a1

1
0

a4

2
0

a3

(b) To select the query at the root node, we investigate queries on each of the four
attributes. The following shows the number of patterns sent to the left and
right for each value, and the entropy at the resulting children nodes:
query
a1
a2
a3
a4

sent left
41 , 12
61 , 02
41 , 12
61 , 22

left entropy
0.7219
0
0.7219
0.8113

sent right
41 , 32
21 , 42
41 , 32
21 , 22

right entropy
0.9852
0.9183
0.9852
1

Because query a2 leads to the greatest weighted reduction in impurity, a2 should

be the query at the root node. We continue and grow the tree shown in the
gure.
Section 8.4
17. Problem not yet solved
Section 8.5

PROBLEM SOLUTIONS

281

a2
0

a3

1
0

18. Problem not yet solved

19. Here our strings are composed of letters in the alphabet A = {a, b, c}.
(a) Consider the following string (and shift positions)
a
1

c
2

a
3

c
4

c
5

a
6

c
7

b
8

a c
9 10

The last-occurence function gives F(a) = 9, F(b) = 8, F(c) = 10, and 0 otherwise. Likewise, the good-sux function gives G(c) = 7, G(ac) = 6, G(bac) = 0,
and 0 otherwise.
(b) Consider the following string (and shift positions)
a
1

b
2

a
3

b
4

a
5

b
6

c
7

b
8

c
9

b
10

a
11

a
12

a
13

b
14

c
15

b
16

a a
17 18

The last-occurence function gives F(a) = 18, F(b) = 16, F(c) = 15 and 0
otherwise. Likewise, the good-sux function gives G(a) = 17, G(aa) = 12,
G(baa) = 10, G(cbaa) = 9, G(bcbaa) = 8, G(abcbaa) = 0, and 0 otherwise.
(c) Consider the following string (and shift positions)
c
1

c
2

c
3

a
4

a
5

a
6

b
7

a
8

b a
9 10

c
11

c
12

c
13

The last-occurence function gives F(a) = 10, F(b) = 9, F(c) = 13, and 0
otherwise. Likewise, G(c) = 12, G(cc) = 11, G(ccc) = 1, G(accc) = 0, and 0
otherwise.
(d) Consider the following string (and shift positions)
a
1

b
2

b
3

a
4

b
5

b
6

a
7

b
8

b c
9 10

b
11

b
12

a
13

b
14

b
15

c
16

b
17

b a
18 19

The last-occurence function gives F(a) = 19, F(b) = 18, F(c) = 16, and 0
otherwise. Likewise, G(a) = 13, G(ba) = 12, G(bba) = 11, G(cbba) = 10,
G(bcbba) = 9, G(bbcbba) = 8, G(abbcbba) = 7, G(babbcbba) = 6, G(bbabbcbba) =
5, and 0 otherwise.
20. We use the information from Fig. 8.8 in the text.
(a) The string and the number of comparisons at each shift are:
p r
1 1

o
1

b
1

a
1

b
1

i
1

l
1

i
1

t
1

i
1

e
3

s
1

f
1

o
1

r
1

e
1

The sum is the total number of character comparisons: 28.

s
1

t
1

i
1

m
1

a
1

t
1

e
1

s
9

CHAPTER 8. NONMETRIC METHODS

282

(b) The test string and index positions are

e
1

s
2

t
3

i
4

m
5

a
6

t
7

e
8

s
9

Here the last-occurence function gives F(s) = 9, F(e) = 8, F(t) = 7, F(a) = 6,

F(m) = 5, F(i) = 4, and 0 otherwise. Likewise, the good-sux function gives
G(s) = 2, G(es) = 1, and 0 otherwise.
(c) The four relevant shifts are shown below, the number of letter comparison is
shown in bold at the right.
p r o b a b i l i t i e s
e s t i m a t e s
e s t i m a t e

i m a t e s
e s t i m a t

The rst shift is 9 F(i) = 5. The second shift is 9 F( ) = 9, and the last
shift is 9F(m) = 4. The total number of character comparisons is 12, as shown
at the lower right.
21. Here the test string is abcca.
(a) Naive string-matching progresses left-to-right. Each letter comparison is indicated by an underline, and the total number of such comparisons at each shift
is shown at the right. The naive string-matching algorithm requires 20 letter
comparisons in this case.
d

a
c a
c c
b c
a b
a

a
c
c
b
a

a
c a
c c a
b c c a
a b c c a
a b c c a
a b c c

5
1
1
1
1
1
3
1
2
1
3
20

The Boyer-Moore string-matching algorithm requires 9 letter comparisons, in

this case.
a
a

b
b

s
e

a b c c c
a b c c a
a b c c
a b c
a b
a

c c c d a b a c a b b c a
c c a
a b c c a
a b c c a
a b c c a

1
2
3
3
9

1
1
1
9
12

PROBLEM SOLUTIONS

283

(b) By an analysis similar to that in part (a), we nd that the naive string-matching
algorithm requires 16 letter comparisons in this case.
The Boyer-Moore string-matching algorithm requires 5 letter comparisons in
this case.
d
a

a
b

d
c

a
c

d
a

c a
a

d
1
2
2
5

c a

(c) By an analysis similar to that in part (a), we nd that the naive string-matching
algorithm requires 19 letter comparisons in this case.
The Boyer-Moore string-matching algorithm requires 7 letter comparisons in
this case.
a
a

b c b c a b c a b c a b
b c c a
a b c c a
a b c c a
a b c c a
a b c c a

c
1
3
1
1
1
7

(d) By an analysis similar to that in part (a), we nd that the naive string-matching
algorithm requires 18 letter comparisons in this case.
The Boyer-Moore string-matching algorithm requires 4 letter comparisons in
this case.
a
a

c
b

c
c

a b c a b a b a
c a
a b c c a
a b c c a

a
1
1
2
4

(e) By an analysis similar to that in part (a), we nd that the naive string-matching
algorithm requires 14 letter comparisons in this case.
The Boyer-Moore string-matching algorithm requires 18 letter comparisons in
this case. Note that in this unusual case, the naive string-matching algorithm
is more ecient than Boyer-Moore algorithm.
b b c c a c b c c a b b c c a
a b c c a
a b c c a
a b c c a
a b c c a
a b c c a
a b c c a
a b c c a

5
1
1
5
1
1
5
18

CHAPTER 8. NONMETRIC METHODS

284

22. Here is pseudocode for the last-occurrence function.

Algorithm 0 (Last-occurrence)
1
2
3
4
5
6

begin initialize F(x)

im+1
do i i 1
if [F[x(i)] = 0] then F[x(i)] i
until i = 1
end

(a) The time complexity in this serial implementation is based on the sum of the d
computations for initializing F, and the m calculations of F, that is, O(d + m).
(b) The space complexity is O(m) since we need to store the nal m values.
(c) For bonbon the number of comparisons is 26 + 6 = 32. For marmalade the
number of comparisons is 26 + 9 = 35. For abcdabdabcaabcda the number
of comparisons is 26 + 16 = 42.
23. Here the class is determined by the minimum edit distance to any pattern in a
category.
(a) By straightforward edit distance calculation, we nd that the edit distance d
from the test pattern x = abacc to each of the patterns in the categories are
as shown in the table.
1
aabbc
ababcc
babbcc

d
3
1
2

2
bccba
bbbca
cbbaaaa

d
4
3
5

3
caaaa
cbcaab
baaca

d
4
4
3

The minimum distance, 1, occurs for the second pattern in 1 , and thus we the
test pattern should be assigned to 1 .
(b) As in part (a), the table shows the edit distance d from the test x = abca to
each of the nine patterns.
1
aabbc
ababcc
babbcc

d
3
3
3

2
bccba
bbbca
cbbaaaa

d
3
2
5

3
caaaa
cbcaab
baaca

d
3
3
2

The minimum distance, 2, occurs for patterns in two categories. Hence, here
there is a tie between cateogories 2 and 3 .
(c) As in part (a), the table shows the edit distance d from the test x = ccbba
to each of the nine patterns.
1
aabbc
ababcc
babbcc

d
3
5
4

2
bccba
bbbca
cbbaaaa

d
2
3
4

3
caaaa
cbcaab
baaca

d
3
4
4

PROBLEM SOLUTIONS

285

Here the minimum distance, 2, occurs for the rst pattern in 2 , and thus the
test pattern should be assigned to cateogory 2 .
(d) As in part (a), the table shows the edit distance d from the test x = bbaaac
to each of the nine patterns.
1
aabbc
ababcc
babbcc

d
4
3
4

2
bccba
bbbca
cbbaaaa

d
4
3
2

3
caaaa
cbcaab
baaca

d
3
3
3

Here the minimum distance, 2, occurs for the third pattern in 2 , and thus the
test pattern should be assigned to category 2 .
24. Here the class is determined by the minimum edit distance to any pattern in a
category.
(a) By straightforward edit distance calculation, we nd that the edit distance d
from the test pattern x = ccab to each of the patterns in the categories are
as shown in the table.
1
aabbc
ababcc
babbcc

d
4
4
5

2
bccba
bbbca
cbbaaaa

d
3
4
5

3
caaaa
cbcaab
baaca

d
3
2
4

Here the minimum distance, 2, occurs for the second pattern in 3 , and thus
the test pattern should be assigned to 3 .
(b) As in part (a), the table shows the edit distance d from the test x = abdca
to each of the nine patterns.
1
aabbc
ababcc
babbcc

d
3
3
3

2
bccba
bbbca
cbbaaaa

d
3
2
5

3
caaaa
cbcaab
baaca

d
4
4
3

Here the minimum distance, 2, occurs for the second pattern in 2 , and thus we
assign the test pattern to 2 .
(c) As in part (a), the table shows the edit distance d from the test x = abc to
each of the nine patterns.
1
aabbc
ababcc
babbcc

d
2
3
3

2
bccba
bbbca
cbbaaaa

d
4
3
6

3
caaaa
cbcaab
baaca

d
4
4
3

Here the minimum distance, 2, occurs for the rst pattern in 1 , and thus we
assign the test pattern to 1 .

CHAPTER 8. NONMETRIC METHODS

286

(d) As in part (a), the table shows the edit distance d from the test x = bacaca
to each of the nine patterns.
1
aabbc
ababcc
babbcc

d
4
4
3

2
bccba
bbbca
cbbaaaa

d
3
3
4

3
caaaa
cbcaab
baaca

d
3
4
1

Here the minimum distance, 1, occurs for the third pattern in 3 , and thus we
assign the test pattern to 3 .
25. Here the class is determined by the minimum cost edit distance dc to any pattern
in a category, where we assume interchange is twice as costly as an insertion or a
deletion.
(a) By straightforward cost-based edit distance calculation, we nd that the edit
distance dc from the test pattern x = abacc to each of the patterns in the
categories are as shown in the table.
1
aabbc
ababcc
babbcc

dc
4
1
3

2
bccba
bbbca
cbbaaaa

dc
4
6
8

3
caaaa
cbcaab
baaca

dc
6
7
4

The nal categorization corresponds to the minimum of these cost-based distances, i.e, the distance = 1 to the second pattern in 1 . Thus we assign the
test pattern to 1 .
(b) As in part (a), the table shows the cost-based edit distance dc from the test x
= abca to each of the nine patterns.
1
aabbc
ababcc
babbcc

dc
3
4
4

2
bccba
bbbca
cbbaaaa

dc
3
3
7

3
caaaa
cbcaab
baaca

dc
5
4
3

The minimum cost distance is 3, which occurs for four patterns. There is a
three-way tie between the cateogories 1 , 2 and 3 . (In the event of a tie, in
practice we often consider the next close pattern; in this case this would lead us
to classify the test pattern as 2 .)
(c) As in part (a), the table shows the cost-based edit distance dc from the test x
= ccbba to each of the nine patterns.
1
aabbc
ababcc
babbcc

dc
6
7
7

2
bccba
bbbca
cbbaaaa

Hence x should be classied as 2 .

dc
2
4
4

3
caaaa
cbcaab
baaca

dc
6
5
6

PROBLEM SOLUTIONS

287

(d) As in part (a), the table shows the cost-based edit distance dc from the test x
= bbaaac to each of the nine patterns.
1
aabbc
ababcc
babbcc

dc
4
3
4

2
bccba
bbbca
cbbaaaa

dc
4
3
2

3
caaaa
cbcaab
baaca

dc
3
3
3

Hence x should be classied as 2 .

26. We assume the costs are all positive.
(a) Non-negativity is always guaranteed because all entries for computing edit distance are non-negative. Reexivity and symmetry are also guaranteed. The
triangle inequality is not guaranteed, however.
(b) Example: Suppose the costs for insertion and deletion are 1 and the cost for
substitution is 5. Consider the three patterns x1 = aba, x2 = aab, and x3 =
aac. In this case, the cost-based distances dc are as shown in the matrix
x1
x1
0
x2 5
x3
2

x2
2
0
2

x3

2
5 .
0

Note that dc (x2 , x3 ) > dc (x1 , x2 ) + dc (x1 , x3 ), violating the triangle inequality.
27. Problem not yet solved
28. Problem not yet solved
29. Problem not yet solved
Section 8.6
30. Recall that Lisp expressions are of the general form (operation operand1
operand2 ). (More generally, for some operations, such as multiplication, *, and addition, +, there can be an arbitrary number of operands.)
(a) Here the alphabet is A = {0, 1, 2, . . . , 9, plus, minus, quotient, (, )}, the set of
intermediate symbols is I = {A, B}, the starting symbol is S = S, and the
productions are

p1 : S (ABB)

p2 : A plus|difference|times|quotient
P=
p

3 : B 0|1| . . . |9

p4 : B (ABB)
(b) An advanced Lisp parser could eliminate excess parentheses, but for the grammar above, three of the ve expressions are ungrammatical:
(times (plus (difference 5 9)(times 3 8))(quotient 2 6)) can be
expressed in this grammar (see gure).
(7 difference 2) cannot be expressed in this grammar.

CHAPTER 8. NONMETRIC METHODS

288

S
A

times (

)(ABB)

plus (ABB)(ABB)quotient 2 6

difference 5 9 times 3 8

(quotient (7 plus 2)(plus 6 3)) cannot be expressed in this grammar.

((plus)(6 2)) cannot be expressed in this grammar.
(difference (plus 5 9)(difference 6 8)) can be expressed in this grammar (see gure).
S
A

difference ( A B B )( A B B )

plus 5

9 difference 6 8

31. Here the language is L(G) = {an b|n 1}.

(a) The alphabet is A = {a, b}, the intermediate symbol is I = {A}, the starting
symbol is S = S, and the productions are

p1 S Ab
p2 A Aa
P=

p3 A a
S

(b) See gure.

A

S
A

A
A
A
a

PROBLEM SOLUTIONS

289

32. Here the language G has alphabet A = {a, b, c}, intermediate symbols I =
{A, B}, starting symbol S = S, and the productions are

p1 S cAb

p2 A aBa
P=
p3 B aBa

p4 B cb
(a) We consider the rules in turn:
S cAb: type 3 (of the form z)
A aBa: type 3 (of the form z)
B aBa: type 3 (of the form z)
B cb: type 2 (of the form I x) and type 3 (of the form z)
Thus G is type 3 or regular grammar.
(b) We now prove by induction that the language generated by G is L(G) =
{can cban b|n 1}. Note that for n = 1, cacbab can be generated by G:
S cAb caBab cacbab.
Now suppose that can cban b is generated by G. Now can+1 cban+1 b = caan cban ab.
where A = an cban can be generated
Given the above, we know that S cAb,
n
n
by G. When we parse caa cba ab using G, we nd S cAb caBab, and
Thus can+1 cban+1 b is generated by G.
thus B = A.
(c) See gure.
S

cb

cb

33. A palindrome is a string that reads the same forward as backward.

(a) Here the grammar G has the English alphabet and the null character, A =
{a, b, . . . , z, }, the single internal symbol I = A, and starting symbol S = S.
The null symbol is needed for generating palindromes having an even number
of letters. The productions are of the form:

p1 S A

p2 A a|b| . . . |z|
P=

p3 A aAa|bAb| . . . |zAz
See gure.

CHAPTER 8. NONMETRIC METHODS

290

(b) The grammar is of type 3 because every rewrite rule is of the form z or
z.
(c) Here the grammar G has the English alphabet and the null character, A =
{a, b, . . . , z, }, the single internal symbol I = A and starting symbol S = S.
The null symbol is needed for generating palindromes having an even number
of letters. The productions are of the form:

p1 S aA|bA| . . . zA

p2 A a|b| . . . |z|
P=

p3 A aAa|bAb| . . . |zAz
See gure.

S
p

34. Consider the numbers 1, 2, . . . , 999.

(a) Here the productions

p1

p2

p3
P=
p4

p5

p6

are of the form:

digits6 digits3
digits3 digit1 hundred digits2
digits3 digit1 hundred
digits3 digits2
digits2 teens|tys|tys digit1|digit1
digit1 1|2| . . . |9

Thus there are 9 possible derivations for digit1, 10 possible derivations for teens,
8 possible derivations for tys. Thus there are (10 + 8 + 8 9 + 9) = 99 possible
derivations for digits2. Thus there are (9 99 + 9 + 99) possible derivations for
digits3. So there are 999 possible derivations for digits6 under these restrictive
conditions.

PROBLEM SOLUTIONS

291

(b) For the numbers 1, 2, . . . , 999, 999 we have

p1 digits6 digits3 thousand digits3

p2 digits6 digits3 thousand
P=

p3 digits6 digits3
We know from part (a) that |digits3| = 999. Thus |digits6| = 999 999 + 999 +
999 = 999, 999 indeed, one for each number.
(c) The grammar does not allow any pronunciation in more than one way because
there are only 999,999 numbers and there are only 999,999 possible pronunciations and these are in a one-to-one correspondence. In English, however, there
is usually more than one way to pronounce a single number. For instance, 2000
can be pronounced two thousand or twenty hundred, however the grammar
above does not allow this second pronunciation.
35. Problem not yet solved
36. Consider grammars and the Chomsky normal form.
(a) Here, most of the rewrite rules in G are

S bA

S
aB

A bAA
P=
A aS

Aa

..

not in Chomsky normal form:

not CNF

not CNF

not CNF
not CNF

CNF

..

We need only one rule to violate the conditions of being in Chomsky normal
form for the entire grammar to be not in CNF. Thus this grammar is not in
CNF.
(b) Here the rewrite rules of G are

rewrite rule

S
Cb A

S
Ca B

A
Ca S

Cb D1

Aa
P=
B Cb S

B
Ca D2

B
b

D
AA

D
BB

a
C

Cb b

CNF?
yes A BC
yes A BC
yes A BC
yes A BC
yes A z
yes A BC
yes A BC
yes A z
yes A BC
yes A BC
yes A z
yes A z

Thus, indeed all the rewrite rules are of the form A BC or A z, and the
grammar G is in CNF.
(c) Theorem: Any context-free language without the null symbol  can be generated
by a grammar G1 in which the rewrite rules are of the form A BC or A z.

CHAPTER 8. NONMETRIC METHODS

292

Proof: First, if a production ends in a string containing just a single terminal

symbol, then that production is in acceptable CNF form (i.e., A z). Now
consider a production of the form A Z1 Z2 Zn , where n 2. If Zi is a
terminal symbol a, we introduce a new internal variable Ca and the production
Ca a, which is in the permissible form. Now we expand our intermediate
variables to include this new Ca , and the rewrite rules accordingly to form G2 .
Thus, L(G1 ) L(G2 ).
Next we use induction on the number of steps in a derivation that if G2 give
a derivation A z, then G1 does as well. If so that if G2 generates a nal
string, then G1 does as well. This result is obvious for nal strings of length
1. Suppose now that it is true for derivations requiring k steps. Let D1 be a
derivation in G2 requiring k + 1 steps. The rst must be of the general form
A B1 B2 Bm for m 2. Then we can write Z = Z1 Z2 Zm where in
grammar G2 we have that Bi derives Zi .
If Bi is the same as Cai for some terminal ai , then Bi must be Zi . By our
construction, there exists a production A Z1 Z2 Zn where Zi = Bi . For
those Bi we know that the derivation in G1 requires no more than k steps, so
by induction hypothesis Zi leads to Bi .
Thus we see that any context-free language can be generated by a grammar
where every production is of the form A z or A Z1 Z2 Zn for n 2.
(d) The string aabab cannot be derived in either G or G .
37. [[state that this problem requires the pumping Lemma in Chapter 6 of Introduction to Automata Theory by Hopcroft and Ullman.]]
38. Problem not yet solved
39. The rst three numbers can be derived, but the last two cannot.
(a) See gure.
(b) See gure.
(c) See gure.
digits6
digits3
digit1
three

hundred

digits6

thousand
digits2
tys

digit1

forty

two
342,619

digits3

digits3

digit1
six

hundred

digits6
digits3

digits2

digits2

digits1

teens

teens

nine

fourteen

thirteen
13

(d) two thousand six hundred thousand five cannot be derived.

(e) one hundred sixty eleven cannot be derived.

thousand
hundred

900,000

PROBLEM SOLUTIONS

293

Section 8.7
40. Here D1 = {ab, abb, abbb} and D2 = {ba, aba, babb} are positive examples from
grammars G1 and G2 , respectively.
(a) Some candidate rewrite rules for G1 and G2 , are

p1

p1 S A
p2
p2 A ab
P1 =
P2 =
p3

p3 A Ab

p4

SA
A ba
A aA
A Ab

(b) Here we infer G1 , being sure not to include rules that would yield strings in D2 .
x+
i

i
1

ab

abb

abbb

S
S
S
A
A
S
A
A

P
A
ab
A
ab
Ab
A
ab
Ab

P produces D2 ?
No
No

No

(c) Here we infer G2 , being sure not to include rules that would yield strings in D1 .
i

x+
i

ba

aba

babb

P
SA
A ba
SA
A ba
A aA
SA
A ba
A aA
A Ab

P produces D1 ?
No
No

No

(d) Two of the strings are ambiguous, and one each is in L(G1 ) and L(G2 ):
bba is ambiguous
abab L(G2 ): S A aA aAb abab
bbb is ambiguous
abbbb L(G1 ): S A Ab Abb Abbb abbbb.
Section 8.8
41. Problem not yet solved

294

CHAPTER 8. NONMETRIC METHODS

Computer Exercises
Section 8.3
1. Computer exercise not yet solved
2. Computer exercise not yet solved
Section 8.4
3. Computer exercise not yet solved
4. Computer exercise not yet solved
5. Computer exercise not yet solved
Section 8.5
6. Computer exercise not yet solved
7. Computer exercise not yet solved
8. Computer exercise not yet solved
Section 8.6
9. Computer exercise not yet solved
10. Computer exercise not yet solved
Section 8.7
11. Computer exercise not yet solved

Chapter 9

Algorithm-independent
machine learning

Problem Solutions
Section 9.2
1.
2.
3.
4.
5.
6.
7.
8.
9.

Problem not yet solved

Problem not yet solved
Problem not yet solved
Problem not yet solved
Problem not yet solved
Problem not yet solved
Problem not yet solved
Problem not yet solved
We seek to prove the relation
n

2 =

n  
!
n

r=0

two dierent ways.

(a) Recall that the polynomial expression is
n  
!
n nr r
n
(x + y) =
y .
x
r
r=0
We substitute x = y = 1 and nd the desired result directly, as shown in Eq. 5
in the text:

(1 + 1) = 2 =

n  
!
n
r=0

295

296CHAPTER 9.

ALGORITHM-INDEPENDENT MACHINE LEARNING

(b) We dene
K(n) =

n  
!
n
r=0

and proceed by induction. Clearly, by simple substitution, K(1) = 1 + 1 = 21 =

2n . We seek to show that in general K(n + 1) = 2K(n) = 2 2n = 2n+1 . We
write this explicitly
n+1
! n + 1
K(n + 1) =
r
r=0
  !
  
n 
n
n+1
n
=
+
+
0
r
n
r=1
  !
 
n
n
(n + 1)!
n
=
+
+
0
r!(n
+
1

r)!
n
r=1
  !
 
n
n
n!(n r + 1 + r)!
n
=
+
+
0
r!(n + 1 r)!
n
r=1
  !
  
n 
n!(n r + 1)
n
n!r
n
=
+
+
+
0
r!(n
+
1

r)!
r!(n
+
1

r)!
n
r=1
  !
  
n 
n
n!
n
n!
=
+
+
+
0
r!(n

r)!
(r

1)!(n
+
1

r)!
n
r=1
  !






n
n
n
n
n
=
+
+
+
0
r
r1
n
r=1




n
!
n
n
=
+
r
r
r=0


n
! n
= 2
r
r=0
=

2K(n) = 2n+1 .

10. Problem not yet solved

11. Consider the Ugly Duckling Theorem (Theorem 9.2).
(a) If a classication problem has constraints placed on the features, the number of
patterns will be less than that of an unconstrained problem. However, the total
number o predicates depends only on the number of patterns, not the number
of features, and so is still
d  
!
d
r=0

= 2d ,

where d is the number of distinct possible patterns. The total number of predicates shared by two patterns remains

d 
!
d2
r=2

r2

= 2d2 .

PROBLEM SOLUTIONS

297

In short, there are no changes to the derivation of the Ugly Duckling Theorem.
(b) Since there are only three dierent types of cars seen, we have three dierent
patterns, x1 , x2 and x3 , with features:
x1

: f1 AND f2 AND f4 AND NOT f5 AND NOT f6

x2

: f1 AND f2 AND f3 AND NOT f4 AND f5 AND NOT f6

x3

NOT[f1 OR f2 OR f3 OR f4 OR f5 ] AND f6 ,

where f1 respresents a car model similar to car A, f2 represents an engine similar

to car A, f3 represents a four-door car, f4 represents a red car, f5 represents a
green car, f6 represents a blue car.
To determine similarity, we count the number of predicates shared by two cars.
With car A and car B, the predicates are x1 OR x2 , and x1 OR x2 OR x3 .
With car B and car C, the predicates are x2 OR x3 , and x1 OR x2 OR x3 .
Since the number of predicates is the same the patterns are equally similar,
according to the denition given in the text.
In eect, we could have described the three patterns with three features, with
f1 representing a four-door car model/engine similar to car C, f2 representing
a red car, f3 representing a green car. In this framework, the patters would be:
x1

NOT f1 AND f2 AND NOT f3

x2

NOT f1 AND NOT f2 AND f3

x3

: f1 AND NOT f2 AND NOT f3 .

The result is that the three patterns being dierent from one another by only
one feature, and hence cars A and B are equally similar as cars B and C.
12. Problem not yet solved
13. Consider the Kolmogorov complexity of dierent sequences.
(a) 010110111011110 . . .. This sequence is made up of m sequences of 0 and l 1s,
where l = 1 . . . m. Therefore the complexity is O(log2 m).
(b) 000 . . . 100 . . . 000, a sequence made up of m 0s, a single 1, and (n m 1) 0s.
The complexity is just that needed to specify m, i.e., O(log2 m).
(c) e = 10.10110111111000010 . . .2 , The complexity of this constant is O(1).
(d) 2e = 101.0110111111000010 . . .2 , The complexity of this constant is O(1).
(e) The binary digits of , but where every 100th digit is changed to the numeral
1. The constant has a complexity of O(1); it is a simple matter to wrap
a program around this to change each 100th digit. This does not change the
complexity, and the answer is O(1).
(f) As in part (e), but now we must specify n, with has complexity O(log2 n).
14. Problem not yet solved
15. For two binary strings x1 and x2 , the Kolmogorov complexity of the pair can
be at worst K(x1 , x2 ) = K(x1 ) + K(x2 ), as we can easily write concatenate two
programs, one for computing x1 and one for x2 . If the two strings share information,

298CHAPTER 9.

ALGORITHM-INDEPENDENT MACHINE LEARNING

the Kolmogorov complexity will be less than K(x1 ) + K(x2 ), since some information
from one of the strings can be used in the generatation of the other string.
16. Problem not yet solved
17. The denition the least number that cannot be dened in less than twenty words
is already a denition of less than twenty words. The denition of the Kolmogorov
complexity is the length of the shortest program to describe a string. From the above
paradoxical statement, we can see that it is possible that we are not clever enough
to determine the shortest program length for a string, and thus we will not be able
to determine easily the complexity.
Section 9.3
18. The mean-square error is
ED [(g(x; D) F (x)2 ] = ED [g 2 (x; D) 2g(x; D)F (x) + F 2 (x)]
= ED [g 2 (x; D] ED [2g(x; DF (x)] + ED [F 2 (x)]
= ED [g 2 (x; D)] 2F (x)ED [g(x; D)] + F 2 (x).
Note, however, that
ED [(g(x; D) ED [g(x; D)])2 ] = ED [g 2 (x; D) 2g(x; D)ED [g(x; D)] + [ED [g(x; D)]]2 ]
= ED [g 2 (x; D)] ED [2g(x; D)ED [g(x; D)]] + ED [(ED [g(x; D)])2 ]
= ED [g 2 (x; D)] 2ED [g(x; D)]ED [g(x; D)] + (ED [g(x; D)])2
= ED [g 2 (x; D)] (ED [g(x; D)])2 .
We put these two results together and nd
ED [g 2 (x; D)] = ED [(g(x; D) ED [g(x; D)])2 ] + (ED [g(x; D)])2 .
We now apply this result to the function g(x) F (x) and obtain
ED [(g(x; D) F (x))2 ]

= ED [g 2 (x; D)] 2F (x)ED [g(x; D)] + F 2 (x)

= ED [(g(x; D ED [g(x; D)]))2 ] + (ED [g(x; D)])2
2F (x)ED [g(x; D)] + F 2 (x)
= ED [(g(x; D) ED [g(x; D)])2 ] + (ED [g(x; D) F (x)])2
= (ED [g(x; D) F (x)])2 + ED [(g(x; D) ED [g(x; D)])2 ] .


 


bias2

variance

Since the estimate can be more or less than the function F (x), the bias can be negative.
The variance cannot be negative, as it is the expected value of a squared number.
19. For a given data set D, if g(x; D) agrees with the Bayes classier, the expected
error rate will be Min[F (x, 1 F (x)]; otherwise it will be Max[F (x, 1 F (x)]. Thus
we have
Pr[g(x; D) = y]

Min[F (x), 1 F (x)]Pr[g(x; D) = yB ]

+Max[F (x), 1 F (x)]Pr[g(x; D) = yB ].

However, under these conditions we can write

Max[F (x), 1 F (x)] = Min[F (x), 1 F (x)] + Max[F (x), 1 F (x)] Min[F (x), 1 F (x)] .



|2F (x)1

PROBLEM SOLUTIONS

299

Thus we conclude
Pr[g(x; D) = y

= |2F (x) 1|Pr[g(x; D) = yB ]

+Min[F (x), 1 F (x)](Pr[g(x; D) = yB ] + Pr[g(x; D) = yB ])
= |2F (x) 1|Pr[g(x; D) = yB ] + Min[F (x), 1 F (x)].

20. If we make the convenient assumption that p(g(x; D)) is a Gaussian, that is,
p(g(x; D)) =

1
exp[(g )2 /(2 2 )]
2

where = ED [g(x; D)] and 2 = Var[g(x; D)]. From Eq. 19 in the text, then, for
F (x) < 1/2 we have

Pr[g(x; D) = yB ]

p(g(x; D))dg

=
1/2

=
1/2

1
exp[(g )2 /(2 2 )]dg
2


exp[u2 /2]du,

(1/2)/

where u = (g )/ and du = dg/. For the other case, that is, F (x) 1/2, we have
1/2
Pr[g(x; D) = yB ]

p(g(x; D))dg

1/2

1
exp[(g )2 /(2 2 )]dg
2


exp[u2 /2]du,

(1/2)/

where u = (g )/ and du = dg/. Therefore, we have

exp[u2 /2]du if F (x) < 1/2

12
(1/2)/
Pr[g(x; D)] =

1
exp[u2 /2]du if F (x) 1/2

2
(1/2)/

where


t=


exp[u2 /2]du =
t

1/2

1/2

%
1$
1 erf[t/ 2] = (t),
2

if F (x) < 1/2

if F (x) 1/2.

300CHAPTER 9.

ALGORITHM-INDEPENDENT MACHINE LEARNING

Thus, we can write

Pr[g(x; D)]

1/2

ED [g(x; D) 1/2]
= sgn[F (x) 1/2] 
Var[g(x; D)]
= sgn[F (x) 1/2]

= sgn[F (x) 1/2][ED [g(x; D)] 1/2] Var[g(x; D)]1/2 .






variance

boundary bias

21. Problem not yet solved

22. Problem not yet solved
Section 9.4
23. The jackknife estimate of the mean is given by Eq. 25 in the text:
1!
(i)
n i=1

n
1 ! 1 !
xj
n i=1 n 1
j=i

n
n
!
!
1

xj xi
n(n 1) i=1 j=1
n

()

!
1
[n
xi ]
n(n 1) i=1
n

!
n
1
xi

n1
n(n 1) i=1
n

n
1

n1
n1
=
.

24. Problem not yet solved

25. Problem not yet solved
26. We must verify that Eq. 26 in the text for the jackknife estimate of the variance
of the mean is formally equivalent to the variance estimate given by Eq. 23 in the
text. From Eq. 26 we have
n1!
((i) () )2
n i=1
2

n 
n
x xi
n1!
()
n i=1
n1

2
n
x xi
n 1 ! n
n1

n i=1
n1
n1

2
n
x xi (n 1)
n 1 ! n
x
n i=1
n1
n

Var[
] =
=
=
=

PROBLEM SOLUTIONS

301
n1!
n i=1
n

x
xi
n1

2

!
1
(xi x
)2 ,
n(n 1) i=1
n

which is Eq. 23 in the text.

27. Consider the computational complexity of dierent statistics based on resampling.
(a) The jackknife estimate of the mean is
1!
(i) ,
n i=1
n

() =

which requires n summations, and thus has a complexity O(n).

(b) The jackknife estimate of the median has complexity just that required for the
sorting operation, which is O(nlogn).
(c) The jackknife estimate of the standard deviation is
;
<
3
n
!
<

2
= =n 1
(i) ()
Var[]
n i=1
;

<
22
1 n
n
<
!
!
n

1
1
<
2

= =
(i)
(i) ,
n
n i=1
i=1
which requires 2n summations, and thus has a complexity O(n).
(d) The bootstrap estimate of the mean is
B
1 ! (b)
()

=
,
B
b=1

which requires B summations, and thus has a complexity O(B).

(e) The bootstrap estimate of the medium has complexity the same as that as the
sorting operation, which is O(BlogB).
(f) The bootstrap estimate of the standard deviation is
;
<
n "
#2

<1 !
VarBoot [] = =
(b) ()
B i=1
;
<
22
1
B
B
<
1 ! (b )
< 1 ! (b) 2
,

= =
( )
B
B 
b=1

b =1

which requires 2B summations, and thus has a complexity O(B).

302CHAPTER 9.

ALGORITHM-INDEPENDENT MACHINE LEARNING

28. Problem not yet solved

Section 9.5
29. Problem not yet solved
30. In boosting, if none of the patterns in D2 are correctly classied by C1 , then
the classier to be trained, C2 , will not have good accuracy on the patterns correctly
classied by C1 . This means that C1 and C2 will disagree on most of the patterns,
and this means that C3 will dominate the full ensemble classier. Paradoxically, if
C2 classies none of the patterns in D1 correctly, then C2 actually shares a great deal
of information with C1 . In a c-category problem, (1 1/c) of the patterns should be
misclassied in a most informative set as this will complement C1 on its slightly
bettern than chance, 1/c of a correct classication.
31. There are a number of algorithms that are acceptable here. One approach is to
perform a simple binary search over the line between x1 and x2 , as
Algorithm 0 (Basic binary line search)
1
2
3
4
5
6

begin initialize x1 , x2 , 
while (true)
x (x1 + x2 )/2
if g1 (x) g2 (x) <  then return x
if g1 (x) < g2 (x) then x2 x
end

Another version uses the discriminant functions g1 (x) and g2 (x) to heuristically guide
the search for the boundary. As the two points under consideration, x1 and x2 become
closer together, the true discriminant functions can be approximated linearly. In the
following, x3 is the point along the line from x1 to x2 where g1 and g2 cross in the
linear approaximation.
Algorithm 0 (Binary search)
1
2
3
4
5
6
7
8

begin initialize x1 , x2 , 
while (true)
g(x1 ) g2 (x1 ) g2 (x2 )
g(x2 ) g1 (x2 ) g2 (x2 )
x3 x1 + (x2 x1 )/[g(x1 ) g(x2 )]
if g(x3 ) 0 then x1 x3 else x2 x3
if g(x3 ) <  then return x3
end

32. Problem not yet solved

Section 9.6
33. Problem not yet solved
34. Our data come from either a uniform distribution or a Gaussian.
(a) Clearly, the maximum-likelhood values for the limits of the uniform distribution
are the lower and upper limits within the data set: xl = 0.2 and xu = 0.9.
(b) The maximum-likelihood values of the mean and standard deviation are just
those statistics of the data set, which turn out to be = 0.5143 and = 0.2231,
respectively.

PROBLEM SOLUTIONS

303

(c) It is necessary to integrate over a region x about a given data point to convert
from probability density to probability. The probability of the data given a
particular model is then the product of the probability of each data point given
the model (uniform U or Gaussian G ), that is,
P (D|U )

12.1427(x)7

P (D|G )

1.7626(x)7 ,

and thus a uniform distribution is a better model for this data.

35.
36.
37.
38.
39.
40.
41.
42.
43.

Problem
Problem
Problem
Problem
Problem
Problem
Problem
Problem
Problem

not
not
not
not
not
not
not
not
not

yet
yet
yet
yet
yet
yet
yet
yet
yet

solved
solved
solved
solved
solved
solved
solved
solved
solved

Section 9.7
44. Problem not yet solved
45. Problem not yet solved

304CHAPTER 9.

ALGORITHM-INDEPENDENT MACHINE LEARNING

Computer Exercises
Section 9.2
1. Computer exercise not yet solved
Section 9.3
2. Computer exercise not yet solved
Section 9.4
3. Computer exercise not yet solved
Section 9.5
4. Computer exercise not yet solved
5. Computer exercise not yet solved
6. Computer exercise not yet solved
Section 9.6
7. Computer exercise not yet solved
8. Computer exercise not yet solved
9. Computer exercise not yet solved
Section 9.7
10. Computer exercise not yet solved

Chapter 10

Unsupervised learning and

clustering

Problem Solutions
Section 10.2
1. We are given that x can assume values 0, 1, . . . , m and that the priors P (j ) are
known.
(a) The likelihood is given by
P (x|) =

c  
!
m
j=1

with normalization constraint

x
m


jx (1 j )mx P (j ),

P (x|) = 1, for all . Thus P (x|) repre-

x=0

sents m independent equations in the c unknowns 1 , . . . , c ; there are multiple

solutions if c > m and we do not have identiability.
(b) On page 519 in the text, the case of m = 1, c = 2 was shown to be completely
unidentiable but that the sum 1 + 2 could be identied. In the present case,
too, an m-dimensional subspace of (1 , . . . , c ) is identiable. Whether it is
completely identiable or not depends on the actual values for P (x|) that are
observed. For example, when m = 1, c = 2 if P (x = 1|) = 1, then 1 + 2 = 2
and thus 1 = 2 = 1, and we have complete identiability. Thus, in general,
nothing can be said about the complete identiability of (1 , . . . , c ), though an
m-dimensional subspace will be completely identiable.
(c) Suppose now that the priors P (j ) are unknown. Then, there are a total of
c+(c1) = 2c1 unknown parameters c of them the unknowns 1 , . . . , c and
305

306CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

c 1 them the unknown P (j ) reduced by the single constraint

Thus, the problem is not identiable if 2c 1 > m.

c

j=1

P (j ) = 1.

2. Problem not yet solved

3. We are given the mixture density
P (x|) = P (1 )

2
2
2
2
1
1
ex /(21 ) + (1 P (1 ))
ex /(22 ) .
21
22

(a) When 1 = 2 , then P (1 ) can take any value in the range [0, 1], leaving the
same mixture density. Thus the density is completely unidentiable.
(b) If P (1 ) is xed (and known) but not P (1 ) = 0, 0.5, or 1.0, then the model
is identiable. For those three values of P (1 ), we cannot recover parameters
for the rst distribution. If P (1 ) = 1, we cannot recover parameters for the
second distribution. If P (1 ) = 0.5, the parameters of the two distributions are
interchangeable.
(c) If 1 = 2 , then P (1 ) cannot be identied because P (1 ) and P (2 ) are
interchangeable. If 1 = 2 , then P (1 ) can be determined uniquely.
Section 10.3
4. We are given that x is a binary vector and that P (x|) is a mixture of c multivariate
Bernoulli distributions:
c
!
P (x|) =
P (x|i , )P (i ),
i=1

where
d
,

P (x|i , i ) =

ijij (1 ij )1xij .

j=1

(a) We consider the log-likelihood

ln P (x|i , i )

d
!

[xij ln ij + (1 xij )ln (1 ij )] ,

j=1

and take the derivative

ln P (x|i , i )
ij

=
=
=
=

xij
1 xij

ij
1 ij
xij (1 ij ) ij (1 xij )
ij (1 ij )
xij xij ij ij + ij xij
ij (1 ij )
xij ij
.
ij (1 ij )

We set this to zero, which can be expressed in a more compact form as

n
!

i ) xk i = 0.
P (i |xk ,
i (1
i)

k=1

PROBLEM SOLUTIONS

307

i must
(b) Equation 7 in the text shows that the maximum-likelihood estimate
satisfy
n
!
k=1

i ) ln P (xk |i ,
i ) = 0.
P (i |xk ,
i

We can write the equation from part (a) in component form as

i) =
i ln P (xk |i ,

i
xk
,
i)
i (1

and therefore we have

n
!

i ) xk i = 0.
P (i |xk ,
i)
i (1

k=1
i (0, 1), and thus we have
We assume
n
!

i )(xk
i ) = 0,
P (i |xk ,

k=1

which gives the solution

n

i )xk
P (i |xk ,

i =

k=1
n


k=1

i)
P (i |xk ,

i , the maximum-likelihood estimate of i , is a weighted average of the

(c) Thus
xk s, with the weights being the posteriori probabilities of the mixing weights
i ) for k = 1, . . . , n.
P (i |xk ,
5. We have a c-component mixture of Gaussians with each component of the form
p(x|i , i ) N (i , i2 I),
or more explicitly,
p(x|i , i )

1
exp
(2)d/2 id



1
t
2 (x i ) (x i ) .
2i

We take the logarithm and nd

ln p(x|i , i )

d
d
1
= ln (2) ln i2 2 (x i )t (x i ),
2
2
2i

and the derivative with respect to the variance is

ln p(x|i , i )
i2

d
1
+ 4 (x i )t (x i )
2
2i
2i
1
(di2 + x i 2 ).
2i4

=
=

308CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

i must satisfy Eq. 12 in the text, that is,

The maximum-likelihood estimate
n
!
k=1

i ) ln p(xk |i ,
i ) = 0.
P (i |xk ,
i

We set the derivative with respect to i2 to zero, that is,

n
!

n
!
k=1

i ) ln p(xk |i , i )
P (i |xk ,
2
i

i ) 1 (d
i 2 )
i2 + xk
P (i |xk ,
2
i4

k=1

0,

rearrange, and nd
d
i2

n
!

i) =
P (i |xk ,

k=1

n
!

i )xk
i 2 .
P (i |xk ,

k=1

The solution is

i2 =

1
d

n

i )xk
i 2
P (i |xk ,

k=1

n

i)
P (i |xk ,

k=1

i ), the maximum-likelihood estimates of and P (i |xk , i ),

i and P (i |xk ,
where
i
are given by Eqs. 1113 in the text.
6. Our c-component normal mixture is
p(x|) =

c
!

p(x|j , )P (j ),

j=1

and the sample log-likelihood function is

l=

n
!

ln p(xk |).

k=1

We take the derivative with respect to and nd

l

=
=
=
=

n
n
!
ln p(xk |) !
1
p(xk , )
=

p(xk , )

k=1
n
!

k=1

c
!
1
p(xk |j , )P (j )
p(xk , )

k=1
n !
c
!
k=1 j=1
c
n !
!
k=1 j=1

l=1

p(xk |j , )P (j )
ln p(xk |j , )
p(xk , )

P (j |xk , )

ln p(xk |j , )
,

PROBLEM SOLUTIONS

309

where by Bayes Theorem we used

P (j |xk , ) =

p(xk |j , )P (j )
.
p(xk |)

7. Our c-component normal mixture is

p(x|) =

c
!

p(x|j , )P (j ),

j=1

and the sample log-likelihood function is

l=

n
!

ln p(xk |).

k=1

We take the derivative with respect to and nd

l

=
=
=
=

n
n
!
p(xk , )
ln p(xk |) !
1
=

p(xk , )

k=1
n
!

k=1

c
1
!
p(xk |j , )P (j )
p(xk , )

k=1
n !
c
!
k=1 j=1
n !
c
!

l=1

p(xk |j , )P (j )
ln p(xk |j , )
p(xk , )

P (j |xk , )

k=1 j=1

ln p(xk |j , )
,

where by Bayes Theorem we used

P (j |xk , ) =

p(xk |j , )P (j )
.
p(xk |)

8. We are given that 1 and 2 are statistically independent, that is, p(1 , 2 ) =
p(1 )p(2 ).
(a) We use this assumption to derive
p(1 , 2 |x1 )

=
=

p(1 , 2 , x1 )
p(x1 |1 , 2 )p(1 , 2 )
=
p(x1 )
p(x1 )
p(x1 |1 , 2 )p(1 )p(2 )
p(x1 )

p(1 )p(2 )
p(x1 )
p(1 )p(2 )
= [p(x1 |1 , 1 )P (1 ) + p(x1 |2 , 2 )P (2 )]
.
p(x1 )

= [p(x1 |1 , 1 , 2 )P (1 ) + p(x1 |2 , 1 , 2 )P (2 )]

Therefore, p(1 , 2 |x1 ) can be factored as p(1 |x1 )p(2 |x1 ) if and only if
p(x1 |1 , 1 )P (1 ) + p(x1 |2 , 2 )P (2 )

310CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

can be written as a product of two terms, one involving only 1 and the other
involving 2 . Now, for any choice of P (1 ), we have P (2 ) = 1 P (1 ). This
then implies that
p(x1 |1 , 1 )P (1 ) + p(x1 |2 , 2 )P (2 )
can be factored in terms of 1 and 2 if and only if p(x1 |i , i ) is constant in i .
This, in turn, implies that p(1 , 2 |x1 ) cannot be factored if
ln p(x1 |i , i )
= 0
i
for i = 1, 2.
(b) Posteriors of parameters need not be independent even if the priors are assumed
independent (unless of course the mixture component densities are parameter
free). To see this, we rst note 0 < P (j ) < 1 and nj /n  P (j ) as n ,
and in that case we have
max

1 ,...,c
c
!


=

1
ln p(x1 , . . . , xn |1 , . . . , c )
n

c
1
1 !
P (j )ln P (j ) ln (2 2 ) 2
P (j ) 2
2
2
j=1
j=1

c
!

1
1
P (j )ln P (j ) ln (2 2 )
2
2
j=1

c
!

1
P (j )ln P (j ) ln (2 2 e).
2
j=1

9. Problem not yet solved

10. We consider the log-likelihood function for data set D,
l

=
=

ln p(D|)
n
,
ln
p(xk |)
k=1

n
!

ln p(xk |),

k=1

where
p(xk |)

c
!

p(xk |j , j )P (j ).

j=1

We seek to maximize l with respect to and P (i ), subject to the constraints that

P (i ) 0

and

c
!
j=1

P (j ) = 1.

PROBLEM SOLUTIONS

311

We use the method of Lagrange undetermined multipliers, and dene the objective
function (to be maximized) as
1 c
2
!
f (, P (1 ), . . . , P (c )) = l + o
P (i ) 1 .
i=1

This form of objective function guarantees that the normalization constraint is obeyed.

We thus demand that at =

f (, P (1 ), . . . , P (c ))
=0
i
for i = 1, . . . , c. Furthermore, the derivative obeys
f
l
=
= i l
i
i

n
!

ln p(xk |)
i

k=1
n
!

p(xk |)
p(xk |) i
k=1

n
c
!
!
1
=
p(xk |j , j )P (j )
p(xk |) i j=1

k=1

=
=
=

n
!
k=1
n
!
k=1
n
!



1

p(xk |i , i ) P (i )
p(xk , ) i

p(xk |i , i )P (i )
1

p(xk |i , i )
p(xk , )
p(xk |i , i ) i
P (i |xk , )

k=1

ln p(xk |i , i ).

the derivative obeys f = 0, or equivalently

Thus at =
i
n
!
k=1

P (i |xk , )
i ln p(xk |i , i ) = 0,

where
=
P (i |xk , )
=

i )P (i )
p(xk |i ,

p(xk |)
p(xk |i , i )P (i )
,
c

j )P (j )
p(xk |j ,

j=1

as given by Eq. 13 in the text.

We continue with the derivative with respect to the priors
2
1 c
 n
!
!
f

ln p(xk |) + o
P (i ) 1
=
P (i )
P (i )
i=1
k=1

312CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

=
=

n
!
k=1
n
!
k=1

!
1

p(xk |j , j )P (j ) + o
p(xk |) P (i ) j=1
c

1
p(xk |i , i ) + o .
p(xk |)

and P (i ) = P (i ) and nd
We evaluate this at =
f
P (i )

n
!
i)
p(xk |i ,
k=1
n


k=1

p(xk |)

+ o

P (i |xk , )
P (i )

+ o = 0.

This implies
n
1 !

P (i ) =
P (i |xk , ),
o
k=1

and since

c

P (i ) = 1 we have

i=1

n
c
1 !!
= 1,
P (i |xk , )
o
i=1
k=1

or o = n where
1!

P (i ) =
P (i |xk , ).
n
n

k=1

Section 10.4
11. Our Gaussian densities are of the form p(x|i , i ) N (i , ), and we use the
following terminology:
pq

pq

= pqth element of
= pqth element of 1

xp (k) = pth element of xk

p (i) = pth element of i .
(a) We write the class-conditional density as a Gaussian,


1
1
t 1
p(xk |i , i ) =
exp

(x

)
.
(x
k
k
i
i
2
(2)d/2 ||1/2
The log-likelihood is
ln p(xk |i , i )

d
1
1
= ln (2) ln || (xk i )t 1 (xk i ),
2
2
2

PROBLEM SOLUTIONS

313

as given by Eq. 21 in the text. We take the derivative and nd


ln p(xk |i , i )

= pq d2 ln (2) + 12 ln |1 |
pq

12 (xk i )t 1 (xk i ) . ()
Now we also have that the squared Mahalanobis distance is
(xk i )t 1 (xk i )

d !
d
!

 

(xp (k) p (i)) p q (xq (k) q (i)) ,

p =1 q  =1

and its derivative is

(xp (k) p (i))(xq (k) q (i))

+(xq (k) q (i))(xp (k) p (i)) if p = q

=

(xp (k) p (i))2

if p = q

2(xp (k) p (i))(xq (k) q (i)) if p = q
=
(xp (k) (i))2
if p = q.



(xk i )t 1 (xk i )
pq

In component form, the derivative of the determinant of the inverse covariance

is
ln |1 |
pq

=
=

1 |1 |
|1 | pq

1  1
)pq | + |(1 )qp | ,
1 |(
| |

where we have used the fact that pq = qp (i.e., both and 1 are symmetric), and the notation (1 )ij = (i, j)th minors of 1 with ith row and jth
column deleted. We can use the above results and
d
 1  !


  =
pq (1 )pq 
q  =1

to write


ln 1 
pq

 1    1  
 



 1 pq +  1 qp






=
 1  

 1 pq

 

2pq if p = q
=
pq
if p = q.

if

p = q

if

p=q

( )

We substitute () and ( ) into () and nd


ln p(xk |i , i )
if p = q
pq (xp (k) p (i))(xq (k) q (i))
=
1
[

(x
(k)

(i))(x
(k)

(i))]
if p = q
pq
p
p
q
q
2 pq

()

314CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING


pq
[pq (xp (k) p (i))(xq (k) q (i))] ,
=
1
2

1 if p = q
=
0 if p =
 q


where pq

is the Kronecker symbol.

(b) The derivatives of the log-likelihood function are
!!
l
ln p(xk |j , )
P (j |xk , )
=

j=1
n

k=1

and
l
pq

n !
c
!

ln p(xk |j , j , )
pq
k=1 j=1


n !
c
!
pq
=
P (j |xk , j ) 1
[pq (xp (k) p (j))(xq (k) q (j))] .
2
j=1
=

P (j |xk , j , )

k=1

j of j , we have l/ pq = 0, and thereAt the maximum-likelihood estimate

fore
#
"
c
n 

j )(xp (k)
P (j |xk ,
p (j))(xq (k)
q (j))
1 pq


2
pq
k=1 j=1
1
,

pq =
c
n 

2
j )
P (j |xk ,
k=1 j=1

which yields the maximum-likelihood estimate of the pq-entry to the covariance

matrix, that is,
n 
c

j )(xp (k)
p (j))(xq (k)
q (j))
P (j |xk ,

pq =

k=1 j=1

c
n 

j )
P (j |xk ,

k=1 j=1

In vector and matrix form, this result can be written just a bit more compactly
as
c
n 

j )(xk
j )(xk
j )t
P (j |xk ,

k=1 j=1

n 
c

j )
P (j |xk ,

k=1 j=1

where we used Eq. 13 in the text,

j ) = P (xk |j , j )P (j ) ,
P (j |xk ,
c

i )P (i )
p(xk |i ,
i=1

and the normalization constraint

c
!
j=1

j ) = 1.
P (j |xk ,

PROBLEM SOLUTIONS

315

We can also write

c
!

j )(xk
j )(xk
j )t
P (j |xk ,

j=1
c
!

= xk xtk

j ) xkj
P (j |xk ,

j=1

c
!

j )
tj
P (j |xk ,

j=1

c
c
!
!
t

j ) t

j xk +

P (j |xk , j )
P (j |xk ,
j j
j=1

j=1

= xk xtk xk
+

c
!

c
!

c
!
j )t
j ) xt
P (j |xk ,
P (j |xk ,
j
j
k

j=1

j=1

j ) t
P (j |xk ,
j j

j=1

n
!

xk xtk

k=1

=
=

c
!

j=1
n
!

 n
c
!
!
j=1

n
!

k=1

j )xk t
P (j |xk ,
j

k=1

j )xt
P (j |xk ,
k

k=1
c
!

xk xtk

k=1
n
!

nP (j )j tj

j=1
c
!

xk xtk n

 n
c
!
!

j=1
c
!


j )xk t
P (j |xk ,
j j

k=1

nP (j )j tj +

j=1

c
!

P (j )j tj

j=1

P (j )j tj .

j=1

We rearrange this result, substitute it above and nd that the maximumlikelihood estimate of the covariance matrix is
n
c
!
!
= 1

P (j )j tj .
xk xtk
n
j=1
k=1

12. We are told that the distributions are normal with parameters given by p(x|1 )
N (0, 1) and p(x|2 ) N (0, 1/2). Thus the evidence is
P (1 ) x2 /2 1 P (1 ) x2

p(x) =
+
.
e
e

2
(a) If one sample x1 is observed, the likelihood function is then
l = p(x1 )

P (1 ) x21 /2 1 P (1 ) x21

+
e
e

2


2
2
2
ex1 /2
ex1
ex1

P (1 ) + .
=

2
=

Thus l is indeed linear in P (1 ), and hence must be extremal at one of the

limiting values of P (1 ), that is at either P (1 ) = 0 or P (1 ) = 1. (We ignore

316CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

the case where l is independent of P (1 ).) Thus l is maximized at P (1 ) = 0

if the slope of l is negative, i.e., if
2
2
1
1
ex1 /2 ex1 < 0.

This is equivalent to the inequality

2
2
1
1
ex1 /2 < ex1 ,

or simply x21 < ln 2. Thus if x21 < ln 2 the maximum-likelihood estimate of

P (1 ) is P (1 ) = 0.
(b) If the slope of l is positive, then l is maximized at P (1 ) = 1. A positive slope
implies
ex1 /2
ex1

> 0,

2
2

or simply x21 > ln 2. Thus if x21 > ln 2, the maximum-likelihood estimate of

P (1 ) is P (1 ) = 1.
(c) If there is a single sample, we are forced to infer that the maximum-likelihood
estimate for the prior of one of the categories is 1.0, and the other is 0.0, depending upon which density is greater at the sampled point.
1
13. We assume that ther
can be divided into c C
non-overlapping intervals
Bc real line R
1
1 , . . . , c such that j=1 j = R (the real line) and j j  = for j = j  . We
are given that at any point x, only one of the Gaussian functions diers signicantly
from zero, and thus


P (j )
1
2
p(x|1 , . . . , c ) 
exp 2 (x j ) .
2
2

We thus have for some particular j that depends upon x

1
ln p(x1 , . . . , xn |1 , . . . , c ) =
n
=


1!
ln p(xk |1 , . . . , c )
n
n

1
n

k=1
c
!

ln p(xk |1 , . . . , c )

j=1 k:xk j



c
1! !
P (j ) (xk j )2 /(22 )
.
ln
e
n j=1
2
k:x
k

We take logarithms on the right-hand side, rearrange, and nd

1
ln p(x1 , . . . , xn |1 , . . . , c ) 
n
=



c
1! !
1
1
2
2
ln P (j ) ln (2 ) 2 (xk j )
n j=1
2
2
k:xk j

c
c
!
!
!
1!
11
ln P (j )
1
1
ln (2 2 )
n j=1
n2
j=1
k:xk j

k:xk j

PROBLEM SOLUTIONS

317

c
1! ! 1
(xk j )2
n j=1
2 2
k:xk j

1
n

=
where nj =

c
!

c
1
1! !
P (j )nj ln (2 2 )
(xk j )2 ,
2
n
j=1
j=1
k:xk j

1 is the number of points in the interval j . The result above implies

k:xk j

1
ln p(x1 , . . . , xn |1 , . . . , c )
1 ,...,c n
c
c
!
1!
1
1!

nj ln P (j ) ln (2 2 ) +
max
[(xk j )2 ].
n j=1
2
n j=1 j
max

k:xk j

However, we note the fact that

max
j

[(xk j )2 ]

k:xk j

occurs at

k:xk j

xk

k:xk j

k:xk j

xk

nj

= x
j ,
for some interval, j say, and thus we have
1
p(x1 , . . . , xn |1 , . . . , c )
n
n
c
1!
1
1 1! !
nj ln P (j ) ln (2 2 ) 2
(xk x
j )2
n j=1
2
2 n j=1

max

1 ,...,c

k:xk j

c
c
1
1!
1
1 1!
2
nj ln P (j ) ln (2 ) 2
nj
n j=1
2
2 n j=1 nj

(xk x
j )2 .

k :xk j

Thus if n (i.e., the number of independently drawn samples is very large), we

have nj /n = the proportion of total samples which fall in j , and this implies (by
the law of large numbers) that we obtain P (j ).
14. We let the mean value be denoted
1!
xk .
n
n

=
x

k=1

Then we have
1!
(xk x)t 1 (xk x)
n
n

k=1

1!
+x
x)t 1 (xk x
+x
x)
(xk x
n
n

k=1

318CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

1$!
)t 1 (xk x
)
(xk x
n
n

k=1

+2(
x x)
t

n
!

%
) + n(
(xk x
x x)t 1 )(
x x)

k=1

1!
)t 1 (xk x
) + (
(xk x
x x)
x x)t 1 (
n
n

1
n

k=1
n
!

)t 1 (xk x
),
(xk x

k=1

where we used
n
!

) =
(xk x

k=1

n
!

xk n
x = n
x n
x = 0.

k=1

Since is positive denite, we have

(
x x)t 1 (
x x) 0,
. Thus
with strict inequality holding if and only if x = x
1!
(xk x)t 1 (xk x)
n
n

k=1

, that is, at
is minimized at x = x
1!
xk .
n
n

=
x=x

k=1

15. Problem not yet solved

16. The basic operation of the algorithm is the computation of the distance between
a sample and the center of a cluster which takes O(d) time since each dimension needs
to be compared seperately. During each iteration of the algorithm, we have to classify
each sample with respect to each cluster center, which amounts to a total number of
O(nc) distance computations for a total complexity O(ncd). Each cluster center than
needs to be updated, which takes O(cd) time for each cluster, therefore the update
step takes O(cd) time. Since we have T iterations of the classication and update
step, the total time complexity of the algorithm is O(T ncd).
17. We derive the equations as follows.
(a) From Eq. 14 in the text, we have
ln p(xk |i , i ) = ln

1/2
|1
1
i |
(xk i )t 1
i (xk i ).
d/2
2
(2)

It was shown in Problem 11 that



ln p(xk |i , i )
pq
= 1
[pq (i) (xp (k) p (i))(xq (k) q (i))].
pq (i)
2

PROBLEM SOLUTIONS

319

We write the squared Mahalanobis distance as

(xk i )t 1
i (xk i )

d
!

 

(xp (k) p (i))2 p p (i)

p=1

d !
d
!
p =1

 

(xp (k) p (i)) p q (i)(xq (k) q (i)).

q  =1
p=q

The derivative of the likelihood with respect to the pth coordinate of the ith
mean is
ln p(xk |i , i )
p (i)

d
 
$ 1 !

(xp (k) p (i))2 p p (i)

p (i)
2 
p =1

1
2

d
!

d
!

p =1

q  =1
p=q

%
 
(xp (k) p (i)) p q (i)(xq q (i))

1
= 2(xp (k) p (i))(1) pp (i)
2
d

1 !

(xq (k) q (i)) pq (i)(1)

2 q =1
q  =p

d

1 !
(xp (k) p (i)) pq (i)(1)
2 p =1
p =p

d
!

(xq (k) q (i)) pq (i).

q=1

i must satisfy
From Eq. xxx in the text, we know that P (i ) and
1!
P (i ) =
P (i |xk , i )
n
n

k=1

and
n
!
k=1

i) = 0
P (i |xk , i ) i ln p(xk |i ,

or
n
!
k=1

ln p(xk |i , i ) 
P (i |xk , i )
= 0.

p (i)
i = i

We evaluate the derivative, as given above, and nd

n
!
k=1

P (i |xk , i )

d
!
q=1

(xq (k)
q (i))
pq (i) = 0,

320CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

which implies
n
!

i )1 (xk i ) = 0.
P (i |xk ,
i

k=1

i and nd
We solve for

n

k=1 P (i |xk , i )xk

i = 

.
n

k=1 P (i |xk , i )

Furthermore, we have
n
!
k=1


i ) ln p(xk |i , i ) 
P (i |xk ,
= 0,
pq
i = i

and this gives

n
!
k=1



pq

P (i |xk , i ) 1
[
pq (i) (xp (k)
p (i))(xq (k)
q (i))] = 0.
2

We solve for
pq (i) and nd
n
i )[(xp (k)
p (i))(xq (k)
q (i))]
P (i |xk ,

pq (i) = k=1
n

k=1 P (i |xk , i )
and thus
i =

n
k=1

i )(xk
i )(xk
i )t
P (i |xk ,
.
n
i)
P (i |xk ,
k=1

(b) For the conditions of this part, we have i = i2 I and thus

ln p(xk |i , i ) = ln

1
d
(2i2 ) 2

1
(xk i )t (xk i ),
2i2

and thus it is easy to verify that

i ln p(xk |i , i ) =

1
(xk i ).
i2

and hence
n
!
k=1

Thus the mean

i ) 1 (xk
i ) = 0.
P (i |xk ,
i2

n

k=1 P (i |xk , i )xk

i = 

k=1 P (i |xk , i )

does not change. This implies

ln p(xk |i , i )
d
1
= 2 + 4 ||xk i ||2 .
i2
2i
2i

PROBLEM SOLUTIONS

321

We have, moreover,
n
!
k=1



ln p(xk |i , i ) 

=
P (i |xk , i )

i2
i = i

and thus

n
!

1
i) d +
P (i |xk ,

2
4
2i
2i ||xk i ||2
k=1


= 0

and this gives

i2

1
d

n

i )||xk
i ||2
P (i |xk ,
.
n

k=1 P (i |xk , i )

k=1

(c) In this case, the covariances are equal: 1 = 2 = = c = . Clearly the

maximum-likelihood estimate of i , which does not depend on the choice of i ,
will remain the same. Therefore, we have
n

k=1 P (i |xk , i )xk

i = 

.
n

k=1 P (i |xk , i )
The log-likelihood function is
l

n
!
k=1

n
!

ln p(xk |)

c
!
ln
p(xk |j , j )P (j )
j=1

k=1



n
c
1 1/2
!
!
|
|
1
=
ln
P (j )
exp (xk j )t 1 (xk j )
d/2
2
(2)
j=1
k=1
and hence the derivative is
l
pq

=
=
=
=

n
!
ln p(xk |)
pq

k=1
n
!
k=1
n
!
k=1
n
!
k=1

p(xk |)
p(xk |) pq
!
1

P (j ) pq p(xk |j , j )
p(xk |) j=1

!
1

P (j )p(xk |j , j ) pq ln p(xk |j , j )
p(xk |) j=1

However, we have from Problem 17,



ln p(xk |j , j )
pq
= 1
[pq (xp (k) p (j))(xq (k) q (j))] .
pq
2

322CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

We take derivatives and set

l 
=0

pq =

and consequently


pq

P (j )p(xk |j , ) 1
[pq (xp (k) p (j))(xq (k) q (j))] = 0.

2
p(xk |)
j=1
k=1
n
!

c
!

This gives the equation

pq
=

n
!

c
!

p(xk |)
k=1

c
!

k=1

p(xk |)

j=1

j )
P (j )p(xk |j ,

[(xp (k) p (i))(xq (k) q (i))] .

P (j )p(xk |j , )

j=1

We have, therefore,
pq =
c
!

n
!

P (j )p(xk |j , )

n
k=1

1, since

= p(xk , )

j=1
c
n !
!

j )[(xp (k)
P (j |xk ,
q (i))(xq (k)
q (i))]

k=1 j=1

and this implies that the estimated covariance matrix is

!!
j )(xk
= 1
j )(xk
j )t .

P (j |xk ,
n
j=1
n

k=1

Section 10.5
18. We shall make use of the following gure, where the largest circle represents
all possible assignments of n points to c clusters, and the smaller circle labeled C1
represents assignments of the n points that do not have points in cluster C1 , and
likewise for C2 up through Cc (the case c = 4 is shown). The gray region represents
those assigments that are valid, that is, assignments in which none of the Ci are
empty. Our goal is then to compute the number of such valid assignments, that is the
cardinality of the gray region.
(a) First we nd the cardinality of invalid assigments, that is the number of assignments in the white region:
c
!
i=1

|Ci |

c
!
i=j

|Ci Cj | +

c
!
i=j=k

|Ci Cj Ck |

 
 
 
 
c
c
c
c
=
N1
N2 +
N3
Nc
1
2
3
c
c  
!
c
(1)ci+1 Ni ,
=
i
i=1

!
i=j=k...

|Ci Cj Ck Cq |



c terms

PROBLEM SOLUTIONS

323
S

assignments that
leave no points in C1
C1

C2

C3

C4

valid assignments:
ones that have points
in every Ci

where we dene Ni to be the number of distinct assignments of n points that

leave exactly i (labeled) clusters empty. Since each of the n points can be
assigned to one of i1 clusters, we have Ni = (i1)n . We put these intermediate
results together and nd that the number of invalid assignments (where the
clusters are labeled) is
c  
!
c
(1)ci+1 (i 1)n .
i
i=1
But the above assumed that the clusters were labeled, that is, that we could
distinguish between C1 and C2 , for instance. In clustering, no such labeling is
given. Thus the above overcounts by the number of ways we can assign c labels
to c clusters, that is, by c!. Therefore, the cardinality of the white area in the
gure (the number of invalid assignments) is the above result divided by c!, that
is
c  
1! c
(1)ci+1 (i 1)n .
c! i=1 i
The number of valid assignments (the gray region in the gure) is thus the
total number of assignments (indicated by S in the gure) minus
  the number of
invalid assignments (the white region). This total number is cc in . To nd the
total number of valid assignments, we subtract:
 
c  
c n ! c
i
(1)ci+1 (i 1)n .
c
i
i=1
We perform a substitution i i 1 and regroup to obtain our nal answer:
1!
(1)ci in .
c! i=1
c

(b) For the case c = 5 and n = 100 the number of clusterings is

5  
1 ! 5
(1)5i i100
5! i=1 i


1 
5 1100 + 10 2100 + 10 3100 + 5 4100 + 1 5100
5!
= 65738408701461898606895733752711432902699495364788241645840659777500
 6.57 1067 .
=

324CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

(c) As given in part (a), the number of distinct clusterings of 1000 points into 10
clusters is
10  
1 ! 10
(1)10i i1000 .
10! i=1 i

Clearly, the term that dominates this is the case i = 10, since the 101000 is
larger
than any other in the sum. We use Stirlings aproximation, x!  xx ex 2x,
which is valid for large x. Thus we approximate
1
101000
10!

1010 990
10
10!
e10

10990
210
 2778 10990
=

 2.778 10993 .
Incidentally, this result approximates quite well the exact number of clusterings,
275573192239858906525573192239858906525573191758192446064084102276
241562179050768703410910649475589239352323968422047450227142883275
301297357171703073190953229159974914411337998286379384317700201827
699064518600319141893664588014724019101677001433584368575734051087
113663733326712187988280471413114695165527579301182484298326671957
439022946882788124670680940407335115534788347130729348996760498628
758235529529402633330738877542418150010768061269819155260960497017
554445992771571146217248438955456445157253665923558069493011112420
464236085383593253820076193214110122361976285829108747907328896777
952841254396283166368113062488965007972641703840376938569647263827
074806164568508170940144906053094712298457248498509382914917294439
593494891086897486875200023401927187173021322921939534128573207161
833932493327072510378264000407671043730179941023044697448254863059
191921920705755612206303581672239643950713829187681748502794419033
667715959241433673878199302318614319865690114473540589262344989397
5880
 2.756 10993 .
Section 10.6
19. We show that the ranking of distances between samples is invariant to any
monotonic transformation of the dissimilarity values.
(a) We dene the value vk at level k to be min (Di , Dj ), where (Di , Dj ) is the
dissimilarity between pairs of clusters Di and Dj . Recall too that by denition
in hierarchical clustering if two clusters are joined at some level (say k), they

PROBLEM SOLUTIONS

325

remain joined for all subsequent levels. (Below, we assume that only two clusters
are merged at any level.)
Suppose at level k that two clusters, D and D are merged into D at level k + 1.
Then we have
vk+1

min [(Di , Dj )]

Di ,Dj
distinct at k

min [(Di , Dj )],

min

min

Di ,Dj =D
distinct at k

min [(Di , Dj )]

Di =D
is cluster at k

[(Di , Dj )],

min

Di ,Dj =D,D 
are distinct at k1

min

[(Di , D D )] .

Di =D,D 
is cluster at k1

If = min , then we have

min

[(Di , D D )]

Di =D,D 
is cluster at k1

min (x, x )

min

xDi
Di =D,D 
is cluster at k1 x DD 

min

min

min[(Di , D), (Di , D )]

Di =D,D 
is cluster at k1

min

min (x, x ), min (x, x )

xDi
x D

min

xDi
x D 

[(Di , Dj ] .

Di ,Dj
are distance at k1

Therefore we can conclude that

vk+1

min

Di =D,D 
is cluster at k1




min

Di ,Dj =D,D 
are distinct at k1

[(Di , Dj ), (Di , D), (Di , D )]

vk
and thus vk+1 vk for all k. A similar argument goes through for = max .
(b) From part (a) we have vk+1 vk for all k, and thus
0 = v1 v2 v3 vn ,
so the similarity values retain their ordering.
Section 10.7
20. It is sucient to derive the equation for one single cluster, which allows us to
drop the cluster index. We have
!
xD

x m2

!>
1 ! >
>
>2
x>
>x
n 
xD
x D
>2
!>
!
>1
>
=
x x >
>
n 
=

xD

x D

326CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

=
=
=

>2
1 !>
> !
>
x x >
>
2
n
xD x D
1 ! ! !
(x x )t (x x )
n2
xD x D x D
1 1$ ! ! !
(x x )t (x x )
n2 2


xD x D x D
%
! ! !
+
(x x )t [(x x ) + (x x )]
xD x D x D

1 1$ ! ! !
(x x )t (x x )
n2 2


xD x D x D
%
! ! !
+
x x 2 + (x x )t (x x )
xD x D x D

1 1! ! !
=
x x 2
n2 2
xD x D x D


! ! !
1 1 ! ! !
 t


t 

+ 2
(x x ) (x x )
(x x) (x x )
n 2
xD x D x D
xD x D x D



=

1 1 ! !
x x 2
n
2 n2


1
n
s.
2

=0

xD x D

21. We employ proof by contradiction. From Problem 14 with = I, we know that

for a non-empty set of samples Di that
!
x i 2
xDi

is minimized at x = mi , the mean of the points in Di . Now our criterion function is

! !
x mi 2 .
Je =
Di = xDi

Suppose that the partition minimizing Je has an empty subset. Then n c and there
must exist at least one subset in the partition containing two or more samples; we
call that subset Dj . We write this as
D
Dj = Dj1 Dj2 ,
where Dj1 and Dj2 are disjoint and non-empty. Then we conclude that
! !
Je =
x mi 2
Di = xDi

can be written as
Je = A +

!
xDj

x mj 2 ,

PROBLEM SOLUTIONS

327

where A is some constant. However, we also have

!
!
!
x mj 2 =
x mj 2 +
x mj 2
xDj

xDj1

>

xDj2

x mj1 2 +

xDj1

x mj2 2 .

xDj2

Thus, replacing Dj by Dj1 and Dj2 and removing the empty subset yields a partition
into c disjoint subsets that reduces Je . But by assumption the partition chosen minimized Je , and hence we have a contradiction. Thus there can be no empty subsets in
a partition that minimizes Je (if n c). In short, we should partition into as many
subsets are allowed, since fewer subsets means larger dispersion within subsets.
22. The gure shows our data and terminology.
D11

Partition 1

D12

k
1
-2

Partition 2

-1

D21

D22

(a) We have n = 2k + 1 points to be placed in c = 2 clusters and of course the

clusters should be non-empty. Thus the only two cases to consider are those
shown in the gure above.
In Partition 1, we have
m11
m12

2k + 0 k
= 1,
=
2k
= a,

and thus the value of our cluster criterion function is

!
!
Je1 =
(x m11 )2 +
(x m12 )2
xD11

k
!

xD12

(2 + 1)2 +

i=1

k
!

(0 + 1)2 + (a a)2

i=1

= k + k + 0 = 2k.
In Partition 2, we have
m21
m22

= 2,
a
k0+a
=
,
=
k+1
k+1

and thus the value of our cluster criterion function is

!
!
Je2 =
(x m21 )2 +
(x m22 )2
xD21

xD22

328CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

=

(2 + 2) +

i=1

xD21

=
=

k 
!

a
0
k+1

2

a
+ a
k+1

2

a2
a2 k 2
0+
k
+
(k + 1)2
(k + 1)2
2
2
a k(k + 1)
a k
=
.
2
(k + 1)
k+1

Thus if Je2 < Je1 , that is, if a2 /(k + 1) < 2k or equivalently a2 < 2(k + 1), then
the partition that minimizes Je is Partition 2, which groups the k samples at
x = 0 with the one sample at x = a.
(b) If Je1 < Je2 , i.e., 2k < a2 /(k+1) or equivalently 2(k+1) > a2 , then the partition
that minimzes Je is Partition 1, which groups the k-samples at x = 2 with the
k samples at x = 0.
23. Our sum-of-square (scatter) criterion is tr[SW ]. We thus need to calculate SW ,
that is,
SW

=
=

c !
!
i=1 xDi
c !
!

(x mi )(x mi )t
[xxt mi xt xmti + mi mt ]

i=1 xDi

c !
!

xx
t

i=1 xDi

4
!

c
!

m1 ni mti

i=1

xk xtk

c
!

c
!

ni mi mti

i=1

c
!

ni mi mti

i=1

ni mi mti ,

i=1

k=1

where ni is the number of samples in Di and

mi =

1 !
x.
ni
xDi

For the data given in the problem we have the following:

4
!
k=1

xk xtk

 
 
 
 
4 (4 5)
1 (1 4)
0 (0 1)
5 (5
=
+
+
+
5
4
1
0

 
 
 

16 20
1 4
0 0
25 0
=
+
+
+
20 25
4 16
0 1
0 0


42 24
=
.
24 42

Partition 1: Our means are

m1

m2

    

4
1
5/2
1
+
=
,
2
5
4
9/2
    

1
0
5
5/2
+
=
,
2
1
0
1/2

0)

PROBLEM SOLUTIONS

329

and thus the matrix products are





25/4 45/4
25/4 5/4
m1 mt1 =
and m2 mt2 =
.
45/4 81/4
5/4 1/4
Our scatter matrix is therefore



42 24
25/4
SW =
2
24 42
45/4



 

45/4
25/4 5/4
17 1
2
=
,
81/4
5/4 1/4
1 1

and thus our criterion values are the trace



17 1
tr[SW ] = tr
= 17 + 1 = 18,
1 1
and the determinant
|SW | = 17 1 (1) (1) = 16.
Partition 2: Our means are
m1

m2

    

1
4
5
9/2
+
=
,
2
5
0
5/2
    

1
1
0
1/2
+
=
,
2
4
1
5/2

and thus our matrix products are





81/4 45/4
1/4 5/4
t
t
m1 m 1 =
and m2 m2 =
.
45/4 25/4
5/4 25/4
Our scatter matrix is therefore



42 24
81/4
SW =
2
24 42
45/4



 

45/4
1/4 5/4
1 1
2
=
,
25/4
5/4 25/4
1 17

and thus our criterion values are the trace



17 1
tr[SW ] = tr
= 1 + 17 = 18,
1 1
and the determinant
|SW | = 1 17 (1) (1) = 16.
Partition 3: Our means are
m1
m2

      

4
1
0
5/3
1
+
+
=
,
3
5
4
1
3
 
5
=
.
0
=

and thus our matrix products are





25/9 5
25 0
t
t
m1 m1 =
and m2 m2 =
.
5 9
0 0

330CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Our scatter matrix is therefore






 
42 24
25/9 5
25 0
26/3
SW =
3
1
=
24 42
5 9
0 0
22/3


22/3
,
26/3

and thus our criterion values are



17 1
tr SW = tr
= 26/3 + 26/3 = 17.33,
1 1
and
|SW | = 26/3 26/3 22/3 22/3 = 21.33.
We summarize our results as
Partition
1
2
3

tr[SW ]
18
18
17.33

|SW |
16
16
21.33

Thus for the tr[SW ] criterion Partition 3 is favored; for the |SW | criterion Partitions
1 and 2 are equal, and are to be favored over Partition 3.
24. Problem not yet solved
25. Consider a non-singular transformation of the feature space: y = Ax where A is
a d-by-d non-singular matrix.
i = {Ax : x Di } denote the data set transformed to the new space,
(a) If we let D
then the scatter matrix in the transformed domain can be written as
SyW

c !
!

(y myi )(y myi )t

i=1 yD
i

c !
!

(Ax Ami )(Ax Ami )t

i=1 yD
i
c !
!

= A

(x mi )(x mi )t

i=1 yD
i

where At = ASW At . We also have the between-scatter matrix

SyB

c
!

ni (myi my )(mi my )t

i=1

c
!

ni (Ami Am)(Ami Am)t

i=1

= A

c
!


ni (mi m)(mi m)

i=1

= ASB At .

At

PROBLEM SOLUTIONS

331

The product of the inverse matrices is

[SyW ]1 [SyB ]1

= (ASW At )1 (ASB At )
1
= (At )1 S1
ASB At
W A
t
= (At )1 S1
W SB A .

We let i for i = 1, . . . , d denote the eigenvalues of S1

W SB . There exist vectors
zi , . . . , zd such that
S1
W SB zi = i zi ,
for i = 1, . . . , d, and this in turn implies
t t 1
(At )1 S1
zi
W SB A A )

= i (At )1 zi ,

or
y
Sy1
W SB ui

= i ui ,

where ui = (At )1 zi . This implies that 1 , . . . , d are the eigenvalues of

y
Sy1
W SB , and nally that 1 , . . . , d are invariant to non-singular linear transformation of the data.
(b) Our total scatter matrix is ST = SB + SW , and thus
S1
T SW

(SB + SW )1 SW

1
[S1
W (SB + SW )]

1
[I + S1
.
W SB ]

If 1 , . . . , d are the eigenvalues of S1

W SB and the u1 , . . . , ud are the corresponding eigenvectors, then S1
S
u
=
i ui for i = 1, . . . , d and hence
B
i
W
ui + S1
W SB ui = ui + i ui .
This equation implies
[I + S1
W SB ]ui = (1 + i )ui .
1
We multiply both sides of the equation by (1 + i )1 [I + S1
and nd
W SB ]
1
(1 + i )1 ui = [I + S1
ui
W SB ]

and this implies i = 1/(1 + i ) are eigenvalues of I + S1

W SB .
(c) We use our result from part (a) and nd
Jd =

d
d
,
,
1
|SW |
i =
,
= |S1
T SW | =
|ST |
1
+
i
i=1
i=1

which is invariant to non-singular linear transformations described in part (a).

332CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

26. Consider a non-singular transformation of the feature space: y = Ax, where A

is a d-by-d non-singular matrix. We let Di = {Ax : x Dx }. We have
SyW

=
=
=

c !
!

(y my )(y my )t

i=1 yDi
c !
!

(Ax Amxi )(Ax Amxi )t

i=1 xDi
ASxW At .

In a similar way, we have

SyB

Syt
(Syt )1 SyW

= ASxB At
= ASxt At
= (At )1 (Sxt )1 A1 ASxW At
=

(SyW )1 SyB

=
=

(At )1 (Sxt )1 SxW At

(At )1 (SxW )1 A1 ASxB At

(At )1 (SxW )1 SxB At .

(a) From problem 25 (b), we know that

tr[(Sxt )1 SxW ]

d
!

i=1

d
!
i=1

1
1 + i

as well as tr[B SB], because they have the same eigenvalues so long as B
is non-singular. This is because if Sx = i x, then SBB1 x = i x, then also
B1 SBB1 x = B1 i x = i B1 x. Thus we have
B1 SB(B1 x) = i (B1 x).
We put this together and nd
tr[(Syt )1 SyW ]

= tr[(At )1 (Sxt )1 SxW At ]

= tr[(Sxt )1 SxW ]
=

d
!
i=1

1
.
1 + i

(b) See Solution to Problem 25 part (c).

(c) Here we have the determinant
|(SyW )1 SyB | = |(At )1 (SxW )1 SxB At |
,
=
eigenvalues of [(At )1 (SxW )1 SxB At ]
,
=
eigenvalues of [(SxW )1 SxB ]
=

d
,
i=1

i .

PROBLEM SOLUTIONS

333

(d) The typical value of the criterion is zero or close to zero. This is because SB is
often singular, even if samples are not from a subspace. Even when SB is not
singular, some i is likely to be very small, and this makes the product small.
Hence the criterion is not always useful.

 c
 

27. Equation 68 in the text denes the criterion Jd = |SW | =  Si , where
i=1

Si =

(x mi )(x mi )t

xDi

is the scatter matrix for category i , dened in Eq. 61 in the text. We let T be a
non-singular matrix and consider the change of variables x = Tx.
(a) From the conditions stated, we have
mi =

1 ! 
x
ni 
x Di

where ni is the number of points in category i . Thus we have the mean of the
transformed data is
1 !
mi =
Tx = Tmi .
ni
xDi

Furthermore, we have the transformed scatter matrix is

!
Si =
(x mi )(x mi )t
x Di

(Tx Tmi )(Tx Tmi )t

xDi

= T


(x mi )(x mi )t Tt = TSi Tt .

xDi

(b) From the conditions stated by the problem, the criterion function of the transformed data must obey
  c
  1 c
 c
2 
 

!  !
!






Jd = |SW | = 
Si  = 
TSi Tt  = T
Si Tt 
 
 


i=1
i=1
i=1

 c
! 


= |T||Tt | 
Si 


i=1

= |T| Jd .
2

Therefore, Jd diers from Jd only by an overall non-negative scale factor |T|2 .
(c) Since Jd diers from Jd only by a scale factor of |T|2 (which does not depend on
the partitioning into clusters) Jd and Jd will rank partitions in the same order.
Hence the optimal clustering based on Jd is always the optimal clustering based
on Jd . Optimal clustering is invariant to non-singular linear transformations of
the data .

334CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

28. Consider a non-singular transformation of the feature space y = Ax, where A is

a d-by-d non-singular matrix. We let Di = {Ax : x Dix } be the transformed data
set. We then have the within scatter matrix as
c !
!
SyW =
(y myi )(y myi )t
i=1 yDi

c !
!

(Ax Amxi )(y Amxi )t

i=1 yDi

= ASxW At .
In a similar way, we have
SyB

Syt

(SyW )1 SyB

= ASxB At
= ASxt At
=
=

(At )1 (SxW )1 A1 ASxB At

(At )1 (SxW )1 SxB At .

If is an eigenvalue of (SxW )1 SxB with corresponding eigenvector x, that is, (SxW )1 SxB x =
x, then we have
(SxW )1 SxB (At (At )1 ) x = x,



I

which is equivalent to
(SxW )1 SxB At ((At )1 x) = x.
We multiply both sides on the left by (At )1 and nd
(At )1 (SxW )1 SxB At ((At )1 x) = (At )1 x,
which yields
(SyW )1 SyB ((At )1 x) = ((At )1 x).
Thus we see that is an eigenvalue of (SyW )1 SyB with corresponding eigenvector
(At )1 x.
29. We consider the problems that might arise when using the determinant criterion
for clustering.
(a) Here the within-cluster matrix for category i is
!
SiW =
(x mi )(x mi )t ,
xDi

and is the sum of ni matrices, each of rank at most 1. Thus rank[SiW ] ni .

These matrices are not independent; they satisfy the constraint
!
(x mi ) = 0.
xDi

This implies that

rank[SW ]

c
!
i=1

(ni 1) = n c.

PROBLEM SOLUTIONS

335

(b) Here we have the between scatter matrix

SB =

c
!

ni (mi m)(mi m)t

i=1

is the sum of c matrices of rank at most 1, and thus rank[SB ] c. The matrices
are constrained by
c
!

ni (mi m)(mi m)t = 0,

i=1

and thus rank[SB ] c 1. The between-cluster scatter matrix, SB , is always

singular if c d, the dimension of the space. The determinant criterion for clustering is not useful under such conditions because at least one of the eigenvalues
is 0, since as a result the determinant is also 0.
Section 10.8
30. Our generalization of the basic minimum-squared-error criterion function of
Eq. 54 in the text is:
JT =

c !
!

(x mi )t S1
T (x mi ).

i=1 xDi

(a) We consider a non-singular transformation of the feature space of the form

i = {Ax : x Di }
y = Ax, where A is a d-by-d non-singular matrix. We let D
denote the transformed data set. Then, we have the criterion in the transformed
space is
JTy

c !
!
i=1 yD
i
c
! !

y
(y myi )t Sy1
T (y mi )

(Ax Ami )t Sy1

T (Ax Ami ).

i=1 xDi

As mentioned in the solution to Problem 25, the within- and between-scatter

matrices transform according to:
SyW

SyB

= ASW At and
= ASB At ,

respectively. Thus the scatter matrix in the transformed coordinates is

SyT

= SyW + SyB = A(SW + SB )At = AST At ,

and this implies

[SyT ]1

1
= (AST At )1 = (At )1 S1
.
T A

336CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Therefore, we have
JTy

c ! !
!

1
(A(x mi ))t (At )1 S1
(A(x mi ))
T A

i=1 xDi

c
!

(x mi )t (At )(At )1 S1
T (x mi )

i=1

c
!

(x mi )S1
T (x mi ) = JT .

i=1

In short, then, JT is invariant to non-singular linear transformation of the data.

being transferred from Di to Dj . Recall that the total
(b) We consider sample 
x
scatter matrix ST = x (x m)(x m)t , given by Eq. 64 in the text, does not
change as a result of changing the partition. Therefore the criterion is
JT =

c !
!
k=1

where

xDk

(x mk )t S1
T (x mk ),

if k = i, j
Dk
Di {
x} if k = i
Dk =

x} if k = j.
Dj + {

We note the following values of the means after transfer of the point:
mk
mi

= mk if k = i, j,


x

xx
xDi
xDi

=
=
1
ni 1
xDi

n i mi x
x mi )
(ni 1)mi (
=
ni 1
ni 1
mi
x
= mi
,
ni 1


x+x
xHi
=
ni + 1

n j mj + x
x mj )
(nj + 1)mj + (
=
=
nj + 1
nj + 1
mj
x
= mj +
.
nj + 1
=

mj

Thus our criterion function is

JT

c

k=1,k=i,j

(x mk )t S1
T (x mk ) +
+


xDj


xDi

(x mi )t S1
T (x mi )

(x mj )t S1
T (x mj ).

()

PROBLEM SOLUTIONS

337

We expand the sum:

!
!

(x mi )t S1
(x mj )t S1
T (x mi ) +
T (x mj )
xDi

xDi

t
xt S1
T x ni m i +

!
xDj

xDj

t 1
xt S1
T x nj mj ST mj



t

mi
mi
x
x
1
1

m
x

(n
xt S1
x

x
S

1)
m

i
i
i
T
T
T
ni 1
ni 1
xDi


t

!
mj
mj
x
x
1
1

+
m
xt S1
x
+
x
S
+
1)
m
+
S
+
x

(n
j
j
j
T
T
T
nj + 1
nj + 1
xDj
!
t 1
+ mi S1
2mti S1
t S1
=
(x mi )t S1
T (x mi ) x
T x
T mi + 2mi ST x
T mi
!

xDi

1
x mi )
(
x mi )t S1
T (
ni 1
!
t 1
mj S1
+ 2mtj S1
t S1
+
(x mj )t S1
T (x mj ) + x
T x
T mj + 2mj ST x
T mj

xDj

1
(
x mj )t S1
x mj )
T (
nj + 1
!
ni
(x mi )t S1
x mi )
(
x mi )t S1
T (x mi )
T (
ni + 1

xDi

nj
x mj ).
(
x mj )t S1
T (
nj + 1

We substitute this result in () and nd

JT

c !
!

(x mk )t S1
T (x mk )

k=1 xDk


nj
ni
t 1
(x

m
)

)
S
(
x

m
)
(x mj )t S1
(
x

m
j
i
i
T
T
nj + 1
ni 1


nj
ni
t 1
= JT +
(
x

m
)

)
S
(
x

m
)
.
(
x mj )t S1
(
x

m
j
i
i
T
T
nj + 1
ni 1
+

(c) If we let D denote the data set and n the number of points, the algorithm is:
Algorithm 0 (Minimize JT )
1
2
3
4
5
6
7
8
9

begin initialize D, c
Compute c means m1 , . . . , mc
Compute JT

do Randomly select a sample; call it x

; call it mj
Determine closest mean to x
if ni = 1 then go to line 10
nj
if j = i then j nj +1
(
x mj )t ST1 (
x mj )
ni
t 1
if j = 1 then j ni 1 (
x mi ) ST (
x mi )
to Dk
if k j for all j then transfer x

338CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Update JT , mi , and mk
until JT has not changed in n tries

10
11

end

12

31. The total scatter matrix, ST =

n

i=1

(xi m)(xi m)t , given by Eq. 64 in the text,

does not change. Thus our criterion function is

Je = tr[SW ] = tr[ST SB ] = tr[ST ] tr[SB ].
from
We let Je be the criterion function which results from transferring a sample x
Di to Dj . Thus we have
Je

= tr[ST ] tr[SB ]
!
= tr[ST ]
nk mk m2

()

= tr[ST ]

nk mk m2

k=i,j

= tr[ST ]

nk mk m2

k=i,j

nk mk m ni mi m2 nj mj m2 .
2

k=i,j

Therefore we have
ni mi m2 + nj mj m2

>
>2
mi
x
>
>
(ni 1)>mi
m>
ni 1
>
>2
mj
x
+(nj + 1)>mj +
m> ,
nj + 1

as shown in Problem 30. We thus nd that the means change by

mi
mj

mi
x
,
ni 1
mj
x
= mj +
.
nj + 1
= mi

We substitute these into () above and nd through a straightforward but tedious

calculation:


1
2

2
2
t
Je = (ni 1) mi m +
(
x mi ) (mi m)

x mi 
(ni 1)2
ni 1


1
2
2
2
t
+(nj 1) mj m +

x mj 
(
x mj ) (mj m)
(nj + 1)2
nj + 1
1
= ni mi m2 mi m2 2(
x mi )t (mi m) +

x m i 2
ni 1
1
mj 2
+nj mj m2 mj m2 + 2(
x mj )t (mj m) +
x
nj + 1
= ni mi m2 + nj mj m2 + 
x mi 2 
x mj 2
1
1
||
x mi ||2 +
||
x mj ||2
+
ni 1
nj + 1
ni
nj
= ni mi m2 + nj mj m2 +

x m i 2

x mj 2 .
ni 1
nj + 1

PROBLEM SOLUTIONS

339

Therefore our criterion function is

Je

tr[ST ] tr[SB ]
!
= tr[ST ]
nk mk m2

ni
nj

x mi 2 +

x mj 2
ni 1
nj + 1

nj
ni

x m j 2

x mi 2 .
= Je +
nj + 1
ni 1
Section 10.9
32. Our similarity measure is given by Eq. 50 in the text:
s(x, x ) =

xt x
.
xx 

(a) We have that x and x are d-dimensional vectors with xi = 1 if x possesses the
ith feature and xi = 1 otherwise. The Euclidean length of the vectors obeys
;
;
< d
< d
<!
<!


=
2
x = x  =
xi = =
1 = d,
i=1

i=1

and thus we can write

s(x, x )

1!
xt x
=
xi xi
d i=1
d d
d

=
=
=
=

1
[number of common features number of features not common]
d
1
[number of common features (d number of common features) ]
d
2
(number of common features) 1.
d

(b) The length of the dierence vector is

x x 2

= (x x )t (x x )
= xt x + xt x 2xt x
= x2 + x 2 2s(x, x )xx 

= d + d 2s(x, x ) d d

2d[1 s(x, x )],

where, from part (a), we used x = x  = d.

=

33. Consider the following candidates for metrics or pseudometrics.

(a) Squared Euclidean distance:
s(x, x ) = x x 2 =

d
!
i=1

(xi xi )2 .

340CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Clearly we have
s(x, x ) 0
(non-negativity)


s(x, x ) = 0 x = x (uniqueness)
(symmetry).
s(x, x ) = s(x , x)
Now consider the triangle inequality for the particular case d = 1, and x =
0, x = 1/2 and x = 1.
s(x, x )

(0 1)2 = 1

s(x, x )

(0 1/2)2 = 1/4

s(x , x )

(1/2 1)2 = 1/4.

Thus we have
s(x, x ) + s(x , x ) = 1/2 < s(x, x ),
and thus s(x, x ) is not less than or equal to s(x, x ) + s(x , x ). In short, the
squared Euclidean distance does not obey the triangle inequality and is not a
metric. Hence it cannot be an ultrametric either.
(b) Euclidean distance:
;
< d
<!


s(x, x ) = x x  = = (xi xi )2 .
i=1

Clearly symmetry, non-negativity and uniqueness hold, as in part (a); now we

turn to the triangle inequality. First consider two vectors x and x . We will
need to show that x + x  x + x ; we do this as follows:
x + x 2

= (x + x )t (x + x )
= xt x + xt x + 2xt x
= x2 + x 2 + 2xt x

x2 + x 2 + 2xx .

We use the Cauchy-Schwarz Inequality (a special case of the Holder inequality,

see part (c) below), i.e., |xt x | xx , and thus nd
x + x 2 (x + x )2

and thus
x + x  x + x .
We put this together to nd
s(x, x ) = x x 

= (x x ) + (x x )
x x  + x x 
= s(x, x ) + s(x , x ),

PROBLEM SOLUTIONS

341

and thus the Euclidean metric is indeed a metric. We use the same sample
points as in the example of part (a) to test whether the Euclidean metric is an
ultrametric:
s(x, x )




s(x, x ) = s(x, x )

1/4

so
max[s(x, x ), s(x , x )] = 1/4
< 1 = s(x, x ).
Thus the Euclidean metric is not an ultrametric.
(c) Minkowski metric:


s(x, x ) =

1 d
!

21/q
|xi

xi |q

i=1

It is a simple matter to show that the properties of non-negativity, uniqueness

and symmetry hold for the Minkowski metric. In order to prove that the triangle
inequality also holds, we will rst need to prove H
olders inequality, that is, for
p and q positive numbers such that 1/p + 1/q = 1
d
!

|xi xi |

i=1

1 d
!

21/q 1 d
21/p
!
 p
|xi |

|xi |
,
q

i=1

i=1

with equality holding if and only if

1/q

|xj |


1/p


d
xi
i=1

1/p

|xj |

= 


1/q


d
xi
i=1

for all j. The limiting case of p = 1 and q = can be easily veried directly. We
thus turn to the case 1 < p, q, < . Consider two real, non-negative numbers a
and b, and 0 1; we have
a b(1) a + (1 )b,
with equality if and only if a = b. To show this, we consider the function
f (t) = t t + 1
for t 0. Thus we have f  (t) = (t(1) 1) 0. Furthermore, f (t) f (1) = 0
with equality only for t = 1. Thus we have
t t + 1 .
We let t = a/b, substitute above and our intermediate inequality is thus proven.

342CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

We apply this inequality to the particular case

|xj |

a = 
,
1/p


d
|xi |p

|xj |

b = 
1/q


d
|xi |q

i=1

i=1

with = 1/p and 1 = 1/q. Thus for each component j we have

p

|xj xj |
|xj |
|xj |
1
1


+ 
.





1/p
1/q
1/p
1/q
p
q

d
d
d
d




|xi |p
|xi |q
|xi |p
|xi |q

i=1

i=1

i=1

i=1

We sum this inequality over all j = 1, . . . , d and nd

d


d

i=1

|xj xj |

j=1
1/p

|xi |p

d

i=1

|xi |q

1 1
1/q p + q = 1,

and the H
older inequality is thereby proven.
We now use the Holder inequality to prove that the triangle inequality holds for
the Minkowski metric, a result known as Minkowskis inequality. We follow the
logic above and have
d
!

|xi + xi |p

i=1

d
!

|xi + xi |p1 |xi | +

d
!

i=1

|xi + xi |p1 |xi |.

i=1

We apply H
olders inequality to each summation on the right hand side and nd
21/q 1 d
21/p 1 d
21/p
1 d
d
!
!
!
!

|xi + xi |p
|xi + xi |(p1)q
|xi |p
+
|xi |p
j=1

i=1

1 d
!

i=1

i=1

i=1


We divide both side by
1 d
!
i=1

d 


xi + x p

21/p
|xi + xi |p

1 d
!
i=1

i=1

1/q

i=1

thereby obtain

i=1

21/q 1 d
21/p 1 d
21/p
!
!

.
|xi + xi |p
|xi |p
+
|xi |p

, recall that 1 1/q = 1/p, and

21/p

|xi |p

1 d
!

21/p
|xi |p

i=1

Thus, using the notation of the Minkowski metric above, we have

s(x + x , 0) s(x, 0) + s(x , 0),

PROBLEM SOLUTIONS

343

or with simple substitution and rearrangement

s(x, x ) s(x, x ) + s(x , x ),
for arbitrary x, x and x . Thus our triangle inequality is thereby proven and
the Minkowski measure is a true metric.
Ultrametric [[more hereProblem not yet solved]]
(d) Cosine:
s(x, x ) =

xt x
.
xx 

We investigate

 1the
 condition of uniqueness in the particular case d = 2 and
x = 11 , x = 1
. For these points we have
xt x = 1 1 = 0,
but note that x = x here. Therefore s(x, x ) does not possess the uniqueness
property, and thus is not a metric and cannot be an ultrametric either.
(e) Dot product:
s(x, x ) = xt x .
We use the same counterexample as in part (c) to see that the dot product is
neither a metric nor an ultrametric.
(f) One-sided tangent distance:
s(x, x )

min x + T(x) x 2 .

There is no reason why the symmetry property will be obeyed, in general, that
is, s(x, x ) = s(x , x), or
=

min x + 1 T1 (x) x 2 = min x + 2 T2 (x ) x2 ,

1
2
and thus the one-sided tangent distance is not not a metric and not an ultrametric.
34. Consider merging two clusters Di and Dj and whether various values of the
parameters in the function
dhk = dhi + j dhj + dij + |dhi dhj |
can be used for a range of distance or similarity measures.
(a) We consider dmin dened by
dmin (Di , Dj )

dhk

= min x x 
xDi
x Dj

min minx x , minx x 

xDi
x Dj

xDj
x Dh

min (dhi , dhj )

1
1
1
=
dhi + dhj |dhi dhj |
2
2
2
= i dhi + j dhj + dij + |dhi dhj |
for i = j = 1/2, = 0 and = 1/2.

344CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

(b) We consider dmax dened by

dmax (Di , Dj )
dhk

= max x x 
xDi
x Dj

maxx x  =

xDk
x Dh


max
C x x 
Dj

xDi

x Dh

max maxx x , maxx x 

xDi
x Dj

xDj
x Dh

max (dhi , dhj )

1
1
1
=
dhi + dhj + |dhi dhj |
2
2
2
= i dhi + j dhj + dij + |dhi dhj |
for i = dj = 1/2, = 0 and = 1/2.
(c) We consider dave dened by
dave (Di , Dj )

1 !
x x 
ni nj xD

i
x Dj

dhk

!
1
x x 
nh nk xD

k
x Dh

1
nh (ni + nj )

x x 

xDi Dj
x Dh

1
nh (ni + nj )

!
!

x x  +
x x 

xDi
x Dh

xDj
xDh

1
[nh ni dhi + nh nj dhj ]
nh (ni + nj )
ni
nj
=
dhi +
dhj
ni + n j
n i + nj
= i dhi + j dhj + dij + |dhi dhj |
=

for i = ni /(ni + nj ), j = nj /(ni + nj ) and = = 0.

(d) We consider d2mean dened by
d2mean (Di , Dj )
dhk
mk

= mi mj 2
= mh mk 2


x
x
xDi Dj
xDk

=
=
1
n i + nj
xDk

x+

xDi


xDj

ni + nj

x
=

ni m i + n j m j
,
ni + nj

PROBLEM SOLUTIONS
dhk

345

>2
ni
nj
mi
mj >
ni + nj
ni + nj
>2
> ni
ni
nj
nj
>
=
mh
mi +
mh
mj >
ni + nj
ni + n j
ni + nj
ni + nj
> ni
>2 > nj
>2
= >
(mh mi )> + >
(mh mj )>
ni + nj
ni + n j
nj
ni
t
+2
(mh mi )
(mh mj )
ni + nj
n i + nj
>
= >mh

n2j + ni nj
n2i + ni nj
2
m

m

+
mh mj 2
h
i
(ni + nj )2
(ni + nj )2
n i nj
+
[(mh mi )t (mi mj ) (mh mj )t (mi mj )]
(ni + nj )2
ni
nj
ni nj
=
mh mi 2 +
mh mj 2
mh mj 2
n i + nj
ni + nj
(ni + nj )2
= i dhi + j dhj + dij + |dhi dhj |,
=

where
ni
ni + nj
nj
=
ni + nj
ni n j
=
= i j
(ni + nj )2
= 0.

35. The sum-of-squared-error criterion is given by Eq. 72 in the text:

Je

c
!
!
i =1

c
!
i =1

x mi 2

xDi

xt x ni mti mi

xDi

xt x

c
!

ni mti mi .

i =1

We merge Di and Dj into Dk and nd an increase in the criterion function Je of

Je Je =

xt x

c
!

ni mti mi nk mtk mk

i =1
i=k

c
!
!

xt x
ni mti mi ni mti mi nj mtj mj
x

i =1
i =i<j

= ni mti mi + nj mtj mj nk mtk mk ,

where
nk

= ni + n j

346CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING



mk

xDk

xDk

=
nnk mtk mk

=
=

x
1

x+

xDi


xDj

n i + nj

x
=

n i m i + nj m j
ni + nj

ni
nj
mi +
mj
ni + nj
n i + nj
t 


ni
ni
nj
nj
(ni + nj )
mi +
mj
mi +
mj
ni + n j
ni + nj
ni + n j
ni + nj
n2j
n2i
2ni nj
mti mi +
mtj mj +
mt mj .
ni + nj
n i + nj
ni + nj i

thus the dierence in criterion function, = Je Je , is





n2i
n2i
2ni nj
t
=
ni
m i mi + nj
mtj mj
m t mj
ni + nj
ni + nj
ni + n j i
ni nj
n i nj
2ni nj
=
mt mi +
mt mj
mt mj
ni + nj j
n i + nj j
ni + nj j
ni nj
=
(mti mi + mtj mj 2mti mj )
ni + nj
ni n j
=
mi mj 2 .
ni + nj
The smallest increase in Je corresponds to the smallest value of , and this arises
from the smallest value of
ni nj
mi mj 2 .
ni + nj
36. Problem not yet solved
37. Problem not yet solved
38. Problem not yet solved
39. Given a set of points, we can dene a fully connected graph by connecting each
pair of points with an edge that has the distance betyween the end points as its weight.
Assume (without loss of generality) that all the weights are mutually dierent, that
is w(e) = w(e ) if e = e . The nearest-neighbor cluster algorithm merges at each
iteration the clusters of minimal distance and this corresponds to adding the edge of
least weight that joins two dierent clusters to the spanning tree. In other words, the
algorithm examines all the edges in the graph in order of increasing weight and adds
them to the spanning tree if the edge joins two disjoint clusters.
To see the optimality of this algorithm with respect to the sum of the edge lengths,
we use reductio ad absurdum: we will assume that there exists a minimum spanning
tree T  which has not been constructed by this algorithm and then arrive at a contradiction, thus showing that the algorithm indeed constructs a minimum spanning
tree. We call the tree constructed by the algorithm T and assume that there exists a
tree T  = T with a sum of weights less than that of T . Say the choice of e1 , e2 , e3 , . . .
with w(e1 ) < w(e2 ) < w(e3 ) < . . . leads to the tree T . Sort the edges of T  such
that w(e1 ) < w(e2 ) < w(e3 ) < . . .. Now, there must exist a rst index j where these
sequences of edges dier, that is ej = ej and ei = ei for all 1 i < j. Add ej to
T  . This leads to a circle in the graph T  . In this circle, there exists an edge e with
w(ej ) < w(e) because j was the rst index with a dierence in the sequences and ej
is in T (and T has been constructed by the nearest-neighbor clustering algorithm).

PROBLEM SOLUTIONS

347

Delete e from T  . The result is a tree with lower sum of weights. Therefore T  was
not a minimal spanning tree. Thus, by contradiction, we have proven that T was
optimal.
Section 10.10
40. Problem not yet solved
41. As given in Problem 35, the change in Je due to the transfer of one point is
Je (2) = Je (1)

n 1 n2
m1 m2 2 .
n1 + n2

We calculate the expected value of Je (1) as:

2
1
!
2
E(Je (1)) = E
x m
xD

d
!
i=1

d
!

2
(xi mi )

xD

(n 1) 2 ,

i=1

where we have used the standard results

!
S2 =
(xi x)2 2 2n1
i

and
E(S 2 )

= 2 (n 1)
= (n 1)d 2
 nd 2 for n large.

We take the expectation of both sides and nd



n 1 n2
E(Je (2)) = E(Je (1)) E
m1 m2 2
n 1 + n2


n1 n2
 nd 2 E
m1 m2 2 .
n1 + n 2
We now consider only sub-optimal hyperplanes through the sample mean. For n large,
this restriction is equivalent to considering only hyperplanes through the population
mean and this, in turn, implies n1 /n 1/2 and n2 /n 1/2, by the weak Law of
Large Numbers.
Consider points x distributed symmetrically in a hypersphere about . For such
a conguration, the distribution of Je (2) will not depend on the choice of hyperplane;
in particular, we can consider the hyperplane that divides D into
D1

= {x : x1 > 1 }

D2

= {x : x1 < 1 }.

348CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Thus we have

E

n1 n 2
m1 m2 2
n1 + n 2



d
!
n 1 n2
E
(m1i m2i )2
n 1 + n2
i=1


n1 n2
 E
(m11 m21 )2 ,
n 1 + n2

as x comes from a symmetric normal distribution of the form N (, 2 I). Thus we

have m1i m2i  0 for i = 1 independent of m11 , m21 by the Strong Law of Large
Numbers. Now we have in the limit of large n
m11

E(x1 |x1 > 1 )

m21

E(x1 |x1 < 1 ).

Without loss of generality we let 1 = 0. Then we have p(x1 ) N (0, 2 ) and thus
0
2
E(x1 |x1 > 0) =

0
2
E(x1 |x1 < 0) =
.

3
Therefore, we have m11 m21 2 2 for n large. We also have
(n/2)(n/2)
E(Je (2))  nd 2
(n/2) + (n/2)

1 0
2

22

n
2
= nd 2 4 2
4

2
2 .
= n d

Thus Je (2) is approximately independent of

Je (1) Je (2)

n1 n 2
m1 m2 2 .
n1 + n 2

We write
Je (1) = Je (2) + [Je (1) Je (2)]
and take the variance of both sides and nd
Var[Je (1)]

= Var[Je (2)] + Var[],

where
=

n1 n 2
m1 m2 2 .
n1 + n2

Problem not yet solved

Section 10.11
42. Consider a simple greedy algorithm for creating a spanning tree based on the
Euclidean distance between points.

PROBLEM SOLUTIONS

349

(a) The following is known as Kruskals minimal spanning tree algorithm.

Algorithm 0 (Kruskals minimal spanning tree)
1
2
3
4
5
6
7
8
9
10
11
12
13

begin initialize i 0
T {}
do i i + 1
ci xi
until i = n
eij xi xj ; sort in non decreasing order
E ordered set of eij
do for each eij in E in non decreasing order
if xi and xj belong to disjoint clusters
then Append eij to T ; Append cj to ci
until all eij considered
return T
end

(b) The space complexity is n(n 1) if we pre-compute all point-to-point distances.

If we compute all distances on demand, we only need to store the spanning treee
T and the clusters, which are both at most of length n.
(c) The time complexity is dominated by the initial sorting step, which takes
O(n2 logn) time. We have to examine each edge in the for loop and that means
we have O(n2 ) iterations where each iteration can be done in O(logn) time assuming we store the clusters as disjoint-set-forests where searches and unions can
be done in logarithmic time. Thus the total time complexity of the algorithm
is O(n2 logn).
Section 10.12
43. Consider the XOR problem and whether it can be implemented by an ART
network as illustrated in the text.
(a) Learning the XOR problem means that the ART network is supposed to learn
the following input-output relation. We augment all the input vectors and
normalize them to unit weight wo that they lie on the unit sphere in R3 . Then
we have
0
1/ 2
0
1/ 3

input
0
0
1/2
1/ 3

1
1/2
1/2
1/ 3

output
1
0
0
1

()
()
( )
( )

Associate the weight vector w0 with cluster 0, and w1 with cluster 1, and the
subscripts with the dierent dimension. Then we get the following conditions:
w10 + x03
w20 + w30

>
>

w11
w21

x03 < w31

1
0
1
+ w3 or (w1 w1 ) > (w31 w30 )
+ w31 or (w20 w21 ) > (w31 w30 )
w10 + w20 + w30 < w11 + w21 + w31 .

()
()
( )
( )

350CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

From ( ) we see
(w10 w11 ) + (w20 w21 ) < (w31 w30 ).
We combine this with (), () and ( ) and nd
0 < 2(w31 w30 ) < (w10 w11 ) + (w20 w21 ) < (w31 w30 ).
This is a contradition. Thus there is no set of weights that the ART network
could converge to that would implement that XOR solution.
(b) Given a weight vector w for cluster C1 , the vigilance parameter and two inputs
x1 and x2 such that wt x1 > and wt x2 < , but also (w+x)t x2 > w+x1 .
Since the top-down feedback tries to push the input vector to be closer to w,
we can use the angle between w and x1 in this case to make our point, since
we cannot forecast the value of y. If we present x1 before x2 , then x1 will
be classied as belonging to cluster C1 . The weight vector w is now slightly
adjusted and now also x2 is close enough to w to be classied as belonging to
cluster C1 , thus we will not create another cluster. In constrast, if we present x2
before x1 , the angle between w and x2 is smaller than , thus we will introduce
a new cluster C2 containing x2 . Due to the xed value of but weights that are
changing in dependence on the input, the number of clusters depends on the
order the samples are presented.
(c) In a stationary environment the ART network is able to classify inputs robustly
because the feedback connections will drive the network into a stable state even
when the input is corrupted by noise. In a non-stationary environment the
feedback mechanism will delay the adaptation to the changing input vectors
because the feedback connections will interpret the changing inputs as noisy
versions of the original input and will try to force the input to stay the same.
Without the feedback the adaptation of the weights would faster account for
changing sample congurations.
Section 10.13
44. Let e be a vector of unit length and x the input vector.
(a) We can write the variance as
2

= Ex [a2 ]
= Ex [(xt e)t (xt e)]
= Ex [et xxt e]
= et Ex [xxt ]e
= et e,

since e is xed and thus can be taken out of the expectation operator.
(b) We use the Taylor expansion:
2 (e + e)

= (e + e)t (e + e)
= et e + 2et e + O(e3 ).

PROBLEM SOLUTIONS

351

If we disregard the cubic term, we see that

(e + e)t (e + e) = et e
implies (e)t e = 0.
(c) Since e is perpendicular to e, we have (e)t e = 0. We can now combine this
with the condition from part (b) and get the following equation: (e)t e
(e)t e = 0. This can be rewritten (e)t (e e) = 0. For this equation to
hold it is sucient that e = e for some . It is also a necessary condition since
(e)t e must vanish for all e perpendicular to e, thus e is itself perpendicular
to e or in other words parallel to e. In other words, we have e = e.
(d) We denote the d eigenvalues of the covariance matrix by i and the corresponding eigenvectors by ei . Since is symmetric and positive semi-denite,
we know that all eigenvalues are real and greater than or equal to zero. Since
the eigenvectors ei can be chosen orthonormal, they span a basis of the ddimensional space and we can write each vector x as
d
!

(eti x)ei .

i=1

We dene the error Ek (x) as the sum-squared error between the vector x and
its projection onto a k-dimensional subspace, that is

t 

k
k
!
!
t
t
Ek (x) =
x
x
(ei x)e
(ei x)e
i=1

d
!

t 

(eti x)ei

i=k+1

d
!

i=1
d
!

(eti x)ei

i=k+1

(eti x)2 ,

k=k+1

because et e = ij , where ij is the Kronecker delta which has value 1 if i = j

and 0 otherwise. If we now take the expected value of the error over all x, we
get using the denitions in part (a):
Ex [Ek (x)] =

d
!

Ex [(xt ei )t (xt ei )]

i=k+1

d
!

eti i ei

i=k+1

since all ei are eigenvectors of the covariance matrix , we nd

d
!
i=k+1

i eti ei =

d
!

i .

k=k+1

This then gives us the expected error. Thus to minimize the squared-error
criterion, the k-dimensional subspace should be spanned by the k largest eigenvectors, wince the error is the sum of the remaining d k eigenvalues of the
covariance matrix.

352CHAPTER
45.
46.
47.
48.
49.

Problem
Problem
Problem
Problem
Problem

10. UNSUPERVISED LEARNING AND CLUSTERING

not
not
not
not
not

yet
yet
yet
yet
yet

solved
solved
solved
solved
solved

Section 10.14
50. Consider the three points x1 =
distances

1

12 = ||x1 x2 || =
13 = ||x1 x3 || =
23 = ||x2 x3 || =

, x2 =




 0
0

, and x3 =

 0
1

. Then we have the

(1 0)2 + (0 0)2 = 1

(1 0)2 + (0 1)2 = 2
(0 0)2 + (0 1)2 = 1.

We assume for deniteness, and without loss of generality, that the transformed points
obey 0 = y1 < y2 < y3 . We dene
y2 = 21 > 0
y3 =

21

22

> 21 = y2 ,

and therefore we have

d13

= y2 y1 = 21 0 = 21
= y3 y1 = 21 + 22 0 = 21 + 22

d23

= y3 y2 = 21 + 22 21 = 22 .

d12

(a) From the denition in Eq. 107 in the text we have


(dij ij )2
i<j
 2
Jee =
ij
i<j

(21

1)2 + (21 + 22 2)2 + (22 1)2

.
1+1+2

Therefore, minimizing Jee is equivalent to minimizing

f

(21 1)2 + (22 1)2 + (21 + 22

2)2 .

In order to minimize f , we must calculate its derivatives:

f
1
f
2

2(21 1)21 + 2(21 + 22

2(2 1)22 + 2(21 + 22

We set these derivatives to zero and nd

$
%
(21 1) + (21 + 22 2) 41 = 0
$
%
(22 1) + (21 + 22 2) 42 = 0

2)21
2)22 .

PROBLEM SOLUTIONS

353

Because 1 , 2 = 0, we have
21 1 = (21 + 22

and thus 21 = 22 . We have, moreover,

21 1 + (21 + 22

and thus 321 = 1 + 2, which implies

2) = 22 1,

2) = 0

1+ 2
=
=
3
1+ 2
2
y 2 = 1 =
3
2
y3 = 1 + 22 = 221 = 2y2 .
21

22

(b) From denition in Eq. 108 in the text we have

!  dij ij 2
Jf f =
ij
i<j
22  2
 2
2 1 2
2
1 1
1 + 22 2
2 1

=
+
+
1
1
2

1
= (21 1)2 + (21 + 22 2)2 + (22 1)2 ,
2
and thus the derivatives
Jf f
1
Jf f
2

1
2(21 1)21 + 2(21 + 22 2)21
2

1
2(22 1)22 + 2(21 + 22 2)22 .
2

=
=

We set Jf f /i = 0 and obtain (for 1 , 2 = 0)

21 1

1
= (21 + 22 2)
2
= 22 1

and thus 21 = 22 . This result implies

1
21 1 + (21 + 22 2) = 0,
2
and thus
1
1
21 1 + 221
2
2
1
221 = 1 +
2
We solve for 21 and nd
21

=
=

2+1
2+ 2

.
=
2
2

2+ 2
4

354CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Using the above relations, we nd y2 and y3 to be

2+ 2
2
y 2 = 1 =
4
2
2
2
y3 = 1 + 2 = 21 = 2y2 .

10.0. COMPUTER EXERCISES

355

Computer Exercises
Section 10.4
1.
2.
3.
4.
5.

Computer
Computer
Computer
Computer
Computer

exercise
exercise
exercise
exercise
exercise

not
not
not
not
not

yet
yet
yet
yet
yet

solved
solved
solved
solved
solved

Section 10.5
6. Computer exercise not yet solved
7. Computer exercise not yet solved
Section 10.6
8. Computer exercise not yet solved
Section 10.7
9. Computer exercise not yet solved
Section 10.8
10. Computer exercise not yet solved
Section 10.9
11. Computer exercise not yet solved
12. Computer exercise not yet solved
Section 10.11
13. Computer exercise not yet solved
Section 10.12
14. Computer exercise not yet solved
Section 10.13
15. Computer exercise not yet solved
16. Computer exercise not yet solved
Section 10.14
17. Computer exercise not yet solved

356CHAPTER

10. UNSUPERVISED LEARNING AND CLUSTERING

Sample nal exams and

solutions

EXAM 1, three hours, 100 points

1. (10 points total) Consider two categories, each of which is described by a ddimensional Gaussian having the same (but arbitrary) covariance, 1 = 2 =
, arbitrary means, 1 and 2 , and arbitrary priors, P (1 ) and P (2 ). Show
that the minimum-error decision boundary is a hyperplane, described by wt (x
x0 ) = 0. Express w and x0 in terms of the variables given.
2. (20 points total) Consider a one-dimensional two-category classication problem with equal priors, P (1 ) = P (2 ) = 1/2, where the densities have the
form

" 0
# x<0
2
x
p(x|i ) =
1

0 x wi
wi
wi

0
otherwise,
where the wi for i = 1, 2, are positive but unknown parameters.
(a) Conrm that the distributions are normalized.
(b) The following data were collected: D1 = {2, 5} and D2 = {3, 9} for 1 and
1 and w
2 .
2 , respectively. Find the maximum-likelihood values w
(c) Given your answer to part (b), determine the decision boundary x for
minimum classication error. Be sure to state which category is to right
(higher) values than x , and which to the left (lower) values than x .
(d) What is the expected error of your classier in part (b)?
3. (10 points total) Consider a two-dimensional, three-category pattern classication problem, with equal priors P (1 ) = P (2 ) = P (3 ) = 1/3. We dene
the disk distribution D(, r) to be uniform inside a circular disc centered on
having radius r and elsewhere 0. Suppose we model each distribution as a
disk D(i , ri ) and after training our classier nd the following parameters:
 
3
1 :
1 =
, r1 = 2
2
357

SAMPLE FINAL EXAMS AND SOLUTIONS

358

 
4
2 =
,
1
 
5
3 =
,
4

2 :
3 :

r2 = 1
r3 = 3


(a) (1 pt) Use this information to classify the point x = 62 with minimum
probability of error.

(b) (1 pt) Use this information to classify the point x = 33 .

(c) (8 pts) Use this information to classify the point x = 0.5
, where denotes
a missing feature.
4. (10 points total) It is easy to see that the nearest-neighbor error rate P can
equal the Bayes rate P if P = 0 (the best possibility) or if P = (c 1)/c (the
worst possibility). One might ask whether or not there are problems for which
P = P where P is between these extremes.
(a) Show that the Bayes rate for the one-dimensional case where P (i ) = 1/c
and

cr
1 0 x c1
cr
1 i x i + 1 c1
P (x|i ) =

0 elsewhere
is P = r.
(b) Show that for this case the nearest-neighbor rate is P = P .
5. (10 points total) Consider a d-nH -c three-layer backpropagation network, trained
on the traditional sum-squared error criterion, where the inputs have been standardized to have mean zero and unit variance in each feature. All non-linear
units are sigmoidal, with transfer function
f (net) = atanh(b net) =

2a
1 + e2b

net

where a = 1.716 and b = 2/3, and net is the appropriate net activation at a
unit, as described in the text.
(a) State why when initializing weights in the network we never set them all
to have value zero.
(b) Instead, to speed learning we initialize weights in a range w
w +w.

For the input-to-hidden weights, what should be this range? That is, what
is a good value of w?

(c) Explain what your answer to part (b) achieves. That is, explain as specifically as possible the motivation that leads to your answer in part (b).
6. (15 points total) Consider
 training a tree-based classier with the following
eight points of the form xx12 in two categories
1

0 1 6 7

4 , 2 , 0 , 3

3 4 5 8

5 , 7 , 6 , 1

EXAM 1

359

using an entropy or information impurity and queries of the form Is xi ?

(or Is xi ?).
(a) What is the information impurity at the root node (that is, before any
splitting)?
(b) What is the optimal query at the root node?
(c) What is the information impurity at each of the immediate descendent
nodes (that is, the two nodes at the next level)?
(d) Combine the information impurities of these nodes to determine the impurity at this level. How much is the information impurity reduced by your
decision in part (b)?
(e) Continue splitting to create the full tree with pure leaf nodes. Show
your nal tree, being sure to indicate the queries and the labels on the leaf
nodes.
7. (10 points total) We dene the 20% trimmed mean of a sample to be the
mean of the data with the top 20% of the data and the bottom 20% of the data
removed. Consider the following six points, D = {0, 4, 5, 9, 14, 15}.
(a) Calculate the jackknife estimate of the 20% trimmed mean of D.
(b) State the formula for the variance of the jackknife estimate.
(c) Calculate the variance of the jackknife estimate of the 20% trimmed mean
of D.
8. (8 points total) In multi-dimensional scaling, we take points x1 , . . . , xn , with
inter-point distances ij in a high-dimensional space and map them to points
y1 . . . , yn in a low-dimensional space, having inter-point distances dij . One
measure or criterion of quality of such a mapping is
n


Jee =

i<j

(dij ij )2
n

i<j

.
2
ij

(a) Suppose we had a non-optimal mapping (conguration) and wanted to

adjust the position of one of the points yk so as to reduce the Jee criterion.
Take the derivative yk Jee to show which direction yk should be moved.
(b) Write pseudocode for an iterative procedure for full multi-dimensional scaling, using your result from part (a).
9. (7 points total) Short answer.
(a) In self-organizing feature maps (Kohonen maps, topologically correct maps),
why is it important to employ a window function (|y y |)? What
does y represent in this context?
(b) In backpropagation using sigmoids described in Problem 5 (above), why
do we train with teaching values 1 rather than 1.716, the limits of the
output units?

360

SAMPLE FINAL EXAMS AND SOLUTIONS

(c) Of all classiers that can be described by a parameter vector , must the
have the smallest error? Explain or
classier with maximum-likelihood
give a simple example.
(d) State in just a few sentences the Occams razor principle, and informally
what it implies or counsels in pattern recognition.
(e) When creating a three-component classier system for a c-category problem
through standard boosting, we train the rst component classier C1 on
a subset of the data. We then select another subset data for training the
second component classier C2 . How do we select this next set of points
for training C2 ? Why this way, and not for instance randomly?
(f) Summarize briey the No Free Lunch Theorem, referring specically to the
use of o training set data.
(g) State how cross validation is used in the training of a general classier.

EXAM 1

361

Important formulas
p(x|i )P (i )
P (i |x) = 
c
P (k )p(x|k )
k=1



1
t 1
p(x) =
exp (x ) (x )
2
(2)d/2 ||1/2
1

= E[(x )(x )t ]
n
1!
=
(xi )(xi )t
n i=1

=
(x )(x )t p(x)dx

P (i |xg ) =

P (i |xg , xb )p(xg , xb )dxb

p(xg )

l() = lnp(D|)

= arg max l()


p(x|D) =

p(x|)p(|D) d



Q(; i ) = EDb lnp(Dg , Db ; )|Dg ; i

lim Pn (e|x) = 1

c
!

P 2 (i |x)

i=1

P P P

Jp (a)

c
2
P
c1

(at y)

yY

a(k + 1) = a(k) + (k)

!
yYk

SAMPLE FINAL EXAMS AND SOLUTIONS

362

wkj
wji

= k yj = (tk zk )f  (netk )yj


 c
!
= j xi =
wkj k f  (netj )xi
k=1

f (net) = a tanh[b net] =

i(N ) =

2a
1 + e2b

net

P (j )log2 P (j )





ED (g(x; D) F (x))2 = (ED [g(x; D) F (x)])2 + ED g(x; D) ED [g(x; D)])2

(i)
()

1 , x2 , , xi1 , xi+1 , , xn )
= (x
n
1 !
=
(i)
n i=1

EXAM 1 SOLUTION

363

SOLUTION EXAM 1, three hours

1. (10 points total) Consider two categories, each of which is described by a ddimensional Gaussian having the same (but arbitrary) covariance, 1 = 2 =
, arbitrary means, 1 and 2 , and arbitrary priors, P (1 ) and P (2 ). Show
that the minimum-error decision boundary is a hyperplane, described by wt (x
x0 ) = 0. Express w and x0 in terms of the variables given.
Solution Under the stated conditions, the Gaussian densities are written


1
1
t 1
p(x|i ) =
exp (x i ) (x i ) ,
2
(2)d/2 ||1/2
and the appropriate discriminant function here is gi (x) = lnp(x|i ) + lnP (i ).
In this case, then, we have
1
d
1
gi (x) = ln(2) ln|| (x i )t 1 (x i ) + lnP (i ).
2
 2
 2

same f or both categories

We expand and cancel the quadratic term x1 x, which is the same for both
categories (and hence can be eliminated), regroup, and nd
gi (x) = wit x + wi0
where
wi = 1 i
and
1
wi0 = ti 1 i + lnP (i ),
2
where we used the symmetry of 1 to rewrite terms such as xt 1 by
t 1 x and [1 ]t by 1 .
We now seek the decision boundary, that is, where g1 (x) g2 (x) = 0, and this
implies the x-dependent term is merely the dierence between the two weights
wi , that is, our equation is wt (x x0 ) = 0 where
w = 1 (1 2 ).
Now we seek x0 . The x-independent term is
1
1
wt x0 = t1 1 1 + lnP (1 ) + t2 1 2 lnP (2 ).
2
2
Thus we have
1
P (1 )
1 (1 0 )x0 = (1 2 )t 1 (1 2 ) + ln
.
2
P (2 )
We left-multiply both sides by (1/2)(1 2 )t , then divide both sides by the
scalar (1 2 )t 1 (1 2 ) and nd
x0 =

1
lnP (1 )/P (2 )
( 2 )
(1 2 ).
2 1
(1 2 )t 1 (1 2 )

(Note, this is eectively the derivation of Eqs. 5965 on pages 3940 in the text.)

SAMPLE FINAL EXAMS AND SOLUTIONS

364

2. (20 points total) Consider a one-dimensional two-category classication problem with equal priors, P (1 ) = P (2 ) = 1/2, where the densities have the
form

" 0
# x<0
2
x
p(x|i ) =
1

0 x wi
wi
wi

0
otherwise,
where the wi for i = 1, 2, are positive but unknown parameters.
(a) Conrm that the distributions are normalized.
(b) The following data were collected: D1 = {2, 5} and D2 = {3, 9} for 1 and
2 , respectively. Find the maximum-likelihood values w
1 and w
2 .
(c) Given your answer to part (b), determine the decision boundary x for
minimum classication error. Be sure to state which category is to right
(higher) values than x , and which to the left (lower) values than x .
(d) What is the expected error of your classier in part (b)?
Solution The gure, below, should be consulted.
p(x|i)

0.3
1
0.2

0.1
2

part (d)
0
0

1 10
w

215
w

x*

(a) We can ignore subscripts as we check normalization

w
0


w
2 "
x#
2x
x2
1
dx =
2
= [(2 1) (0 0)] = 1.
w
w
w
w 0

(b) The likelihood is

p(D|w)

2 "
x1 # 2 "
x2 #
= p(x1 |w)p(x2 |w) =
1
1
w
w

w w
x1 + x2
4
x1 x2
1
.
=
+
w2
w
w2

To nd the maximum-likelihood solution, we compute

p(D|w)
w



= 4 2w3 + 3(x1 + x2 )w4 4x1 x2 w5


= 4w5 2w2 3(x1 + x2 )w + 4x1 x2
= 0.

EXAM 1 SOLUTION

365

Clearly, we are not interested in the solutions w = . Thus we use the

quadratic equation to nd the solution:

3(x1 + x2 ) 9(x1 + x2 )2 4(2)(4x1 x2 )
w
=
.
4
For category 1 , we substitute x1 = 2 and x2 = 5 and nd

21 9 49 320
w
1 =
4
21 121
=
4
= 8 or 2.5.
Clearly, the solution 2.5 is invalid, since it is smaller than one of the points
in D1 ; thus w
1 = 8.
Likewise, for 2 , we substitute x1 = 3 and x2 = 9 and nd

36 9 122 864
w
2 =
4
36 432
=
4
= 14.2 or 3.1.
Clearly, the solution 3.1 is invalid, since it is smaller than one of the points
in D2 ; thus w
1 = 14.2.
(c) We seek the value of x such that the posteriors are equal, that is, where




x
2
x
2
1
=
1
w
1
w
1
w
2
w
2
"
#
"
2
x
2
x #
1
=
1
8
8
14.2
14.2
which has solution
6.2
 = 5.1,
x =  14.2
8
8 14.2
with R1 being points less than x and R2 being points higher than x .
(d) The probability of error in this optimal case is

P (e) =
min[P (1 )p(x|1 ), P (2 )p(x|2 )]dx
x=5.1

w
1 =8
1 2 "
12"
x #
x#
1
dx +
1
dx
2 14.2
14.2
28
8
0
x =5.1
5.1
8
5.1
8
1 
x2 
1 
x2 
x
 + x

14.2 
2(14.2)2 
8 
2(8)2 
0

5.1

5.1

0.360.

A more geometrically inspired approach is to use the area of a triangle,

A = (baseheight)/2 and the gure. If you followed this approach, though,
you had to remember to divide your shaded area by the total area under
the two straight lines, and account for priors.

SAMPLE FINAL EXAMS AND SOLUTIONS

366

3. (10 points total) Consider a two-dimensional, three-category pattern classication problem, with equal priors P (1 ) = P (2 ) = P (3 ) = 1/3. We dene the
disk distribution D(, r) to be uniform inside a circular disc centered on
having radius r and elsewhere 0. Suppose we model each distribution as such a
disk D(i , ri ) and after training our classier nd the following parameters:
 
3
1 :
1 =
, r1 = 2
2
 
4
2 :
2 =
, r2 = 1
1
 
4
3 :
3 =
, r3 = 3
4

(a) (1 pt) Use this information to classify the point x = 62 with minimum
probability of error.

(b) (1 pt) Use this information to classify the point x = 33 .

(c) (8 pts) Use this information to classify the point x = 0.5
, where denotes
a missing feature.
Solution

P(1) = P(2) = P(3) = 1/3.

p(xg|1)

p(xg|2)

p(xg|3)

x2

1
x2

0
0

(a) As is clear from the gure below,

x1
5

6
2

3

should be classied as 3 .

(b) As can be seen from the gure, 3 is in the non-zero range of 1 and

3 . Because the density p(x|1 ) > p(x|3 ) at that position, however, 33
should be classied as 1 .
(c) Clearly, the decient point cannot be in 3 . We must marginalize over the
bad feature, x1 , to nd the class-conditional densities given the good feature, x2 . But notice that the dashed line goes through a greater percentage
of the diameter of the 2 disk than the 1 disk. In short, if we marginalize
the p(x|2 ) distribution over x1 with x2 = 0.5, we get a larger value than if
we perform the same marginalization
for p(x|1 ), as graphed in the gure.

as 2 .
Thus we should classify 0.5
4. (10 points total) It is easy to see that the nearest-neighbor error rate P can
equal the Bayes rate P if P = 0 (the best possibility) or if P = (c 1)/c (the
worst possibility). One might ask whether or not there are problems for which
P = P where P is between these extremes.

EXAM 1 SOLUTION

367

(a) Show that the Bayes rate for the one-dimensional case where P (i ) = 1/c
and

cr
1 0 x c1
cr
1 i x i + 1 c1
P (x|i ) =

0 elsewhere
is P = r.
(b) Show that for this case the nearest-neighbor rate is P = P .
Solution It is indeed possible to have the nearest-neighbor error rate P equal
to the Bayes error rate P for non-trivial distributions.
(a) Consider uniform priors over c categories, that is, P (i ) = 1/c, and onedimensional distributions given in the problem statement. The evidence
is

cr
0 x c1
c
1
!
cr
1/c
i x (i + 1) c1
p(x) =
p(x|i )P (i ) =

i=1
0
elsewhere.
Note that this automatically imposes the restriction
cr
0
1.
c1
Because the P (i )s are constant, we have P (i |x) p(x|i ) and thus
P (i )
1/c
cr
0 x c1

p(x) = p(x)

0 if i = j
cr
P (i |x) =
j x j + 1 c1

1
if
i
=
j

0
otherwise,
as shown in the gure for c = 3 and r = 0.1. The conditional Bayesian
P(i|x)
1/c = 1/3

0.25

1 , 2 , 3

0.3

0.2
0.15
0.1

x = r = 0.1

0.05

x
1

probability of error at a point x is

P (e|x)

1 P (max |x)

1/c
cr
1 p(x)
if 0 x c1
=
1 1 = 0 if i x i + 1
=

cr
c1 ,

SAMPLE FINAL EXAMS AND SOLUTIONS

368

and to calculate the full Bayes probability of error, we integrate as



P
=
P (e|x)p(x)dx
cr/(c1)



=

=

1
1
c


1/c
p(x)dx
p(x)

cr
= r.
c1

(b) The nearest-neighbor error rate is

 
P

c
!


P (i |x) p(x)dx
2

i=1
cr

cr/(c1)



=
0

j+1 c1


c
!
c( 1c )2
1 2
[1 1] p(x)dx
p(x)dx +
p (x)
j=1 j



0

cr/(c1)



1/c
p2 (x)

1
c

0
cr/(c1)



=
0


p(x)dx


dx =

1
c

cr
= r.
c1

Thus we have demonstrated that P = P = r in this nontrivial case. (Note:

this is Problem 8 from Chapter 4 on your homework.)
5. (10 points total) Consider a d-nH -c three-layer backpropagation network, trained
on the traditional sum-squared error criterion, where the inputs have been standardized to have mean zero and unit variance in each feature. All non-linear
units are sigmoidal, with transfer function
f (net) = atanh(b net) =

2a
1 + e2b

net

where a = 1.716 and b = 2/3, and net is the appropriate net activation at a
unit, as described in the text.
(a) State why when initializing weights in the network we never set them all
to have value zero.
(b) Instead, to speed learning we initialize weights in a range w
w +w.

For the input-to-hidden weights, what should be this range? That is, what
is a good value of w?

(c) Explain what your answer to part (b) achieves. That is, explain as specifically as possible the motivation that leads to your answer in part (b).

EXAM 1 SOLUTION

369

Solution
(a) If all the weights are zero, the net activation netj in all hidden units is zero,
and thus so is their output yj . The weight update for the input-to-hidden
weights would also be zero,
wkj = (tk zk )f  (netk ) yj .

=0

Likewise, the weight update for the input-to-hidden weights would be zero

 c
!
wji =
wkj k f  (netj )xi .


k=1



=0

In short, if all the weights were zero, learning could not progress. Moreover,
as described in the text, if all the weights are the same for a given pair of
layers (even if these weights are not zero), then learning cannot progress.
(b) Suppose all the weights had value 1.0. In calculating its net, each hidden
unit is summing d random variables with variance 1.0. The variance of the
value of net would be, then, d. We want, however, for the net activation
to be in roughly the linear range of f (net),
that is, 1 < net
< +1. Thus
we want our weights to be in the range 1/ d < w < +1/ d.
(c) We want to use the linear range of the transfer function, f (net), so as
to implement the simple linear model rst. If the problem is complex,
the training will change the weights and thus express the nonlinearities, if
necessary.
6. (15 points total) Consider
 training a tree-based classier with the following
eight points of the form xx12 in two categories
1

0 1 6 7

4 , 2 , 0 , 3

 3  4  5  8
5 , 7 , 6 , 1

using an entropy or information impurity and queries of the form Is xi ?

(or Is xi ?).
(a) What is the information impurity at the root node (that is, before any
splitting)?
(b) What is the optimal query at the root node?
(c) What is the information impurity at each of the immediate descendent
nodes (that is, the two nodes at the next level)?
(d) Combine the information impurities of these nodes to determine the impurity at this level. How much is the information impurity reduced by your
decision in part (b)?
(e) Continue splitting to create the full tree with pure leaf nodes. Show
your nal tree, being sure to indicate the queries and the labels on the leaf
nodes.

SAMPLE FINAL EXAMS AND SOLUTIONS

370
x2

= 1

= 2

Is x2 < 4.5?

ye

no

x1

0
0

no

Is x1 < 7.5?

yes

Solution
(a) The entropy impurity at the root node is
i(root)

2
!

P (j )log2 P (j )

j=1

1
1 1
1
= log2 log2
2
2 2
2
= 1 bit.
(b) As is clear from the gure, the query Is x2 < 4.5 the optimal split at the
root.
(c) The left node contains four 1 points and one 2 point. Its impurity is
thus
4
4 1
1
log2 log2 = 0.258 + 0.464 = 0.722 bits.
5
5 5
5
The right node is pure, containing three 2 points. Its impurity is
i(R) = 1log2 1 0log2 0 = 0 bits.
(d) The impurity at the level beneath the root is just the sum of the impurities
of the two nodes computed in part (b) weighted by the probability any
pattern at the root goes to that node. Thus the impurity at the level is
5
3
0.722 + 0 = 0.451 bits.
8
8
The split at the root thus reduced the information impurity by 1.00.451 =
0.549 bits.
(e) As can be seen in the gure, the query at the impure node should be Is
x1 < 7.5.
7. (10 points total) We dene the 20% trimmed mean of a sample to be the
mean of the data with the top 20% of the data and the bottom 20% of the data
removed. Consider the following six points, D = {0, 4, 5, 9, 14, 15}.
(a) Calculate the jackknife estimate of the 20% trimmed mean of D.
(b) State the formula for the variance of the jackknife estimate.
(c) Calculate the variance of the jackknife estimate of the 20% trimmed mean
of D.

EXAM 1 SOLUTION

371

Solution The jackknife estimate of a statistic is merely the average of the leaveone-out estimates. In this case, there are ve points, and thus
(a) We delete one point at a time, take the remaining data set, remove the top
20% and bottom 20% (that is, the highest point and the lowest point), and
calculate the mean. We then average these trimmed means. As shown in
the table, the jackknife estimate of the 20% trimmed mean is 8.33.
deleted point
0
4
5
9
14
15

trimmed set
{5, 9, 14}
{5, 9, 14}
{4, 9, 14}
{4, 5, 14}
{4, 5, 9}
{4, 5, 9}

mean of trimmed set

28/3 (9.33)
28/3 (9.33)
27/3 (9.00)
23/3 (7.66)
18/3 (6.00)
18/3 (6.00)
142/3
= 7.89
6

(b) The variance of the jackknife estimate is

n1!
((i) () )2 ,
n i=1
n

Varjack [] =

though other forms are occassionally used here too.

(c) We perform the standard variance calculation, here consisting of the sum
of six terms, each the square of the dierence between the leave-one-out
mean and the jackknife mean:
Varjack []

5
(9.33 7.89)2 + (9.33 7.89)2 + (9.00 7.89)2
6

+(7.66 7.89)2 + (6.00 7.89)2 + (6.00 7.89)2
0.833 [2.07 + 2.07 + 1.23 + 0.53 + 3.57 + 3.57]

10.48.

8. (8 points total) In multi-dimensional scaling, we take points x1 , . . . , xn , with

inter-point distances ij in a high-dimensional space and map them to points
y1 . . . , yn in a low-dimensional space, having inter-point distances dij . One
measure or criterion of quality of such a mapping is
n


Jee =

i<j

(dij ij )2
n

i<j

.
2
ij

(a) Suppose we had a non-optimal mapping (conguration) and wanted to

adjust the position of one of the points yk so as to reduce the Jee criterion.
Take the derivative yk Jee to show which direction yk should be moved.
(b) Write pseudocode for an iterative procedure for full multi-dimensional scaling, using your result from part (a).

SAMPLE FINAL EXAMS AND SOLUTIONS

372
Solution

(a) We can break Jee into terms that depend upon yk , and those that do not:

Jee

1
= 
n
2
ij
i<j

n
n
!

2
2
(dij ij ) +
(dik ik ) .

i=k
i=k,j=k



 


does not depend on yk

depends on yk

Thus when we take the derivative, we need only consider the second term.
Thus we have

n
d !
1
yk Jee = 
(dik ik )2
n
dyk
2
i=k
ij
i<j

n
1 !
d
2
(dik ik )
dik .
n

dyk
2
i=k
ij

i<j

We note that
dik =

(yk yi )t (yk yi )

and thus
d
1 2(yk yi )
yk yi
dik =
=
.
dyk
2
dik
dik
We put all these results together and nd
yk Jee

n
2 !
yk yi
= 
(dki ki )
.
n
dki
2 i=k
ij
i<j

(b) The multi-dimensional scaling algorithm can be written:

1 Initialize learning rate, yi , i = 1, . . . , n
2
Do Randomly select a single yk
3
Compute yk Jee
4
yk yk yk Jee
5
Until No change in any yk
6 Return yi , i = 1, . . . n
7 End
9. (7 points total) Short answer.
(a) In self-organizing feature maps (Kohonen maps, topologically correct maps),
why is it important to employ a window function (|y y |)? What
does y represent in this context?

EXAM 1 SOLUTION

(b)

(c)

(d)

(e)

373

Solution In this context, y is the output unit (in the low-dimensional

space) currently most active. The learning rule updates its weight by
adding to it the current input vector. In this way, subsequent presentations
of the input pattern will lead to an even larger activation in y . The
window function (|y y |) is large for |y y | small, and decreases
as |y y | increases. Thus, units near y will be updated by adding to
them the current input vector, though with a slightly smaller magnitude.
In this way, after such learning neighborhoods in the input space map to
neighborhoods in the output space we have a topologically correct
mapping.
In backpropagation using sigmoids described in Problem 5 (above), why
do we train with teaching values 1 rather than 1.716, the limits of the
output units?
Solution If we used teaching values 1.716, learning would be unacceptably long, as weight magnitudes increase and increase so as to ensure net
activations are . Using teaching values of 1 keeps weights in bounds
and speeds learning, since no weights are driven to very large magnitude
to obtain an output of 1.716.
Of all classiers that can be described by a parameter vector , must the
have the smallest error? Explain or
classier with maximum-likelihood
give a simple example.
Solution No. If the candidate model space does not include the true
model (e.g., distribution), then even the classier trained by maximumlikelihood methods need not be the best in this candidate model space. For
instance, suppose in a one-dimensional, two-category classication problem
the true distributions each consists of two spikes, one suciently small but
a distant outlier. Suppose we t these distributions with Gaussians. It
is then possible to switch the positions of the estimated means and yield
a classier with error = 100%, even though a classier with error = 0% is
in the model space. (This was illustrated in Problem 8 in Chapter 3.)
State in just a few sentences the Occams razor principle, and informally
what it implies or counsels in pattern recognition.
Solution William of Occam (12851349) stated that Entities should not
be multiplied beyond necessity. (Actually, since he was writing in Latin,
he stated Entia non sunt multiplicanda praeter necessitatem.) In pattern classication, this has come to be interpreted as counselling the use of
simpler models rather than complex ones, fewer parameters rather than
more, and smoother generalizers rather than those that are less smooth.
The mathematical descendents of this philosophical principle of parsimony
appear in minimum-description-length principles, having numerous manifestations in learning, for instance regularization, pruning, and overtting
avoidance.
When creating a three-component classier system for a c-category problem
through standard boosting, we train the rst component classier C1 on
a subset of the data. We then select another subset data for training the
second component classier C2 . How do we select this next set of points
for training C2 ? Why this way, and not for instance randomly?
Solution We seek a data set that contains information not already learned
by C1 , that is, is not well represented by C1 . A data set which C1 classies

374

SAMPLE FINAL EXAMS AND SOLUTIONS

with 0% accuracy is, paradoxically, not independent of C1 . Likewise, a data
set that is classied with 50% accuracy (in the general c-category case) is
also not independent of C1 . Thus we seek a data set that C1 correctly
classies 1/c of the patterns. A simple algorithm is to test all candidate
patterns and place them in two sets: ones correctly classied by C1 (call it
D+ ) and those incorrectly classied by C1 (call it D ). Then, if our data
set is to have n2 patterns, choose n2 /c patterns randomly from D+ and
n2 (1 1/c) from D .
(f) Summarize briey the No Free Lunch Theorem, referring specically to the
use of o training set data.
Solution In the absense of any information about the classication problem (or target function to be learned), on average no classier method
or learning algorithm is better than any other method, including random
guessing on points in the o training set data, that is, the points not
used for training the classier.
(g) State how cross-validation is used in the training of a general classier.
Solution Given a training set D, we randomly remove some portion of the
set, for instance 10%, and keep this as a validation set. We train the
classier on the remaining 90%, and monitor the error on the validation
set as training proceeds. This validation set acts as a representative of
future test patterns that the classier will classify. We stop training at the
rst minimum of the validation error a heuristic to avoid overtting and
improve generalization.

EXAM 2

375

EXAM 2, three hours, 100 points

1. (10 points total) Let the components of the vector x = (x1 , . . . , xd )t be ternary
valued (1, 0, or -1) with
pij

Pr[xi = 1|j ]

qij

Pr[xi = 0|j ]

rij

Pr[xi = 1|j ].

and with the components of xi being statistically independent for all x in j .

Show that a minimum probability of error decision rule can be derived that
involves discriminant functions gj (x) that are quadratic functions of the components xi .
2. (15 points total) Consider a one-dimensional two-category classication problem with equal priors, P (1 ) = P (2 ) = 1/2, where the densities have the
form

p(x|i ) =

0
i ei x

x<0
x 0,

where the i for i = 1, 2, are positive but unknown parameters.

(a) (1 pt) Conrm that the distributions are normalized.
(b) (8 pts) The following data were collected: D1 = {1, 5} and D2 = {3, 9} for
1 and 2 , respectively. Find the maximum-likelihood values 1 and 2 .
(c) (3 pts) Given your answer to part (b), determine the decision boundary
x for minimum classication error. Be sure to state which category is to
right (higher) values than x , and which to the left (lower) values than x .
(d) (3 pts) What is the expected error of your classier in part (c)?
3. (10 points total) Consider the application of the k-means clustering algorithm
to the one-dimensional data set D = {0, 1, 5, 8, 14, 16} for c = 3 clusters.
(a) (3 pt) Start with the three cluster means: m1 (0) = 2, m2 (0) = 6 and
m3 (0) = 9. What are the values of the means at the next iteration?
(b) (5 pt) What are the nal cluster means, after convergence of the algorithm?
(c) (2 pt) For your nal clusterer, to which cluster does the point x = 3
belong? To which cluster does x = 11 belong?

SAMPLE FINAL EXAMS AND SOLUTIONS

376

4. (10 points total) The task is to use Bayesian methods to estimate a one-dimensional
probability density. The fundamental density function is a normalized triangle
distribution T (, 1) with center at with half-width equal 1, dened by

p(x|) T (, 1) =

1 |x | |x | 1
0
otherwise,

as shown on the left gure. The prior information on the parameter is that
it is equally likely to come from any of the three discrete values = 1, 0 or 1.
Stated mathematically, the prior consists of three delta functions, i.e.,
p() =

1
[(x 1) + (x) + (x + 1)],
3

as shown on the gure at the right. (Recall that the delta function has negligible
width and unit integral.)

p(x|)
1

p()

x
-1

+1

-1

(a) (2 pt) Plot the estimated density before any data are collected (which
we denote by D0 = {}). That is, plot p(x|D0 ). Here and below, be sure to
label and mark your axes and ensure normalization of your nal estimated
density.
(b) (4 pts) The single point x = 0.25 was sampled, and thus D1 = {0.25}.
Plot the estimated density p(x|D1 ).
(c) (4 pts) Next the point x = 0.75 was sampled, and thus the data set is
D2 = {0.25, 0.75}. Plot the estimated density p(x|D2 ).
5. (5 points total) Construct a cluster dendrogram for the one-dimensional data
D = {2, 3, 5, 10, 13} using the distance measure Dmax (Di , Dj ).
6. (10 points total) Consider the use of traditional boosting for building a classier
for a two-category problem with n training points.
(a) (8 pts) Write pseudocode for traditional boosting, leading to three component classiers.
(b) (2 pts) How are the resulting three component classiers used to classify
a text pattern?

EXAM 2

377

7. (5 points total) Consider a standard three-layer neural net as shown. Suppose

the network is to be trained using the novel criterion function
1!
(tk zk )4 .
4
c

J=

k=1

Derive the learning rule wkj for the hidden-to-output weights.

t1

target t
output z

...

t2

z1

z2

...

tk

zk

...

tc

...

zc

output

wkj

y1

y2

yj

...

...

yn

hidden

wji

x1

input x

x1

x2

x2

xi

...
...

xi

...
...

xd

input
xd

8. (10 points total) It is World Cup Soccer season and a researcher has developed
a Bayes belief net that expresses the dependencies among the age of a person
(A), his or her nationality (B), whether he or she likes soccer (C), and how
much he or she watches sports TV during the World Cup season (D). Use the
conditional probability tables to answer the following.
P(b)
P(a)
a1 = 0-30 years
b
=
American
A
P(b
) P(b2)
B
1
P(a1) P(a2) P(a3) a = 31-40 years
1
0.3

0.6

P(c|a,b)

0.1

a1, b 1
a1 , b 2
a2 , b 1
a2 , b 2
a3 , b 1
a3 , b 2

a3 = 41-100 years

P(c1|ai, bj)
0.5
0.7
0.6
0.8
0.4
0.1

age

P(c2|ai, bj)
0.5
0.3
0.4
0.2
0.6
0.9

nation- b2 = non-American
ality

sports

0.2

0.8

c1 = likes soccer
c2 = doesn't like soccer

P(d|c)
D

watch
TV

d1 = a lot
d2 = some
d3 = none

c1
c2

P(d1|ck)
0.7
0.5

P(d2|ck) P(d3|ck)
0.2
0.1
0.3
0.2

(a) (2 pts) What is the probability we nd a non-American who is younger

than 30 who likes soccer and watches a lot of sports TV?
(b) (4 pts) Suppose we nd someone who is an American between 31-40 years
of age who watches some sports TV. What is the probability that this
person likes soccer?
(c) (4 pts) Suppose we nd someone over 40 years of age who never watches
sports TV. What is the probability that this person likes soccer?

378

SAMPLE FINAL EXAMS AND SOLUTIONS

9. (15 points total) This problem conserns the construction of a binary decision
tree for three categories from the following two-dimensional data:
1

 1  2  3
1 , 8 , 5

 4  6 7 
7 , 2 , 6

 5  7  7 
10 , 4 , 9

(a) (3 pts) What is the information impurity at the root node, i.e., before any
splitting?
(b) (4 pts) The query at the root node is: Is x1 > 3.5? What is the
information impurity at the next level?
(c) (6 pts) Continue to grow your tree fully. Show the nal tree and all queries.
(d) (2 pts) Use your tree to classify the point ( 72 ).
10. (10 points total) Short answer (1 pt each).
(a) Explain using a diagram and a few sentences the technique of learning
with hints and why it can improve a neural net pattern classier.
(b) Use the Boltzmann factor to explain why at a suciently high temperature T , all congurations in a Boltzmann network are equally probable.
(c) Explain with a simple gure the crossover operation in genetic algorithms.
(d) What is a surrogate split and when is one used?
(e) If the cost for any fundamental string operation is 1.0, state the edit distance between bookkeeper and beekeepers.
(f) Suppose the Bayes error rate for a c = 3 category classication problem is
5%. What are the bounds on the error rate of a nearest-neighbor classier
trained with an innitely large training set?
(g) What do we mean by the language induced by grammar G?
(h) In hidden Markov models, what does the term aij refer to? What does bjk
refer to?

EXAM 2

379

Important formulas
p(x|i )P (i )
P (i |x) = 
c
P (k )p(x|k )
k=1

p(x) =



1
t 1
exp

(x

)
(x

)
2
(2)d/2 ||1/2
1

= E[(x )(x )t ]
n
1!
=
(xi )(xi )t
n i=1

=
(x )(x )t p(x)dx

P (i |xg ) =

P (i |xg , xb )p(xg , xb )dxb

p(xg )

l() = lnp(D|)

= arg max l()


p(x|D) =

p(x|)p(|D) d



Q(; i ) = EDb lnp(Dg , Db ; )|Dg ; i

d =

|2 1 |

lim Pn (e|x) = 1

P P P

c
!

P 2 (i |x)

i=1

c
2
P
c1

SAMPLE FINAL EXAMS AND SOLUTIONS

380

Jp (a)

(at y)

yY

a(k + 1)

= a(k) + (k)

yYk

= k yj = (tk zk )f  (netk )yj


 c
!
= j xi =
wkj k f  (netj )xi

wkj
wji

k=1


f (net) = a tanh[b net] = a

P () =

Z=

e+b
e+b

eb
net + eb
net

net

net

eE /T
Z
eE  /T

wij =

EQ [si sj ]i o
T 


E[si sj ]i clamped
 



clamped

learning

i(N ) =

c
!

unlearning

P (j )log2 P (j )

j=1





ED (g(x; D) F (x))2 = (ED [g(x; D) F (x)])2 + ED g(x; D) ED [g(x; D)])2

(i)
()

1 , x2 , , xi1 , xi+1 , , xn )
= (x
n
1 !
(i)
=
n i=1
Si =

(x mi )(x mi )t

xDi

SW =

c
!
i=1

Si

EXAM 2

381

SB =

c
!

ni (mi m)(mi m)t

i=1

ST =

(x m)(x m)t

xD

Je = tr[SW ]

Jd = SW 

min x x 

dmin (Di , Dj )

dmax (Di , Dj )

max x x 

davg (Di , Dj )

1 ! !
x x 
ni nj


dmin (Di , Dj )

= mi mj 

xD
xD 

xD
xD 

xD xD

(dij ij )2
i<j
 2
ij

Jee

Jf f

!  dij ij 2
=
ij
i<j

Jef

i<j

1 ! (dij ij )2

ij i<j
ij

i<j

wki (t + 1) = wki (t) + (t)(|y y |)i

SAMPLE FINAL EXAMS AND SOLUTIONS

382

EXAM 2 Solutions
1. (10 points total) Let the components of the vector x = (x1 , . . . , xd )t be ternary
valued (1, 0, or -1) with
pij

Pr[xi = 1|j ]

qij

Pr[xi = 0|j ]

rij

Pr[xi = 1|j ].

and with the components of xi being statistically independent for all x in j .

Show that a minimum probability of error decision rule can be derived that
involves discriminant functions gj (x) that are quadratic functions of the components xi .
Solution (Note: This is problem 44 from Chapter 2.)
The minimum probability of error is achieved by the following decision rule:
Choose k if gk (x) gj (x) for all j = k,
where here we will use the discriminant function
gj (x) = ln p(x|j ) + ln P (j ).
The components of x are statistically independent for all x in j , and therefore,

p(x|j ) = p((x1 , . . . , xd )t |j ) =

d
,

p(xi |j ),

i=1

where
pij
qij
rij

=
=

Pr[xi = 1|j ]
Pr[xi = 0|j ]

Pr[xi = 1|j ].

As in Sect. 2.9.1 in the text, we use exponents to select the proper probability,
that is, exponents that have value 1.0 when xi has the value corresponding to
the particular probability and value 0.0 for the other values of xi . For instance,
for the pij term, we seek an exponent that has value 1.0 when xi = +1 but is
0.0 when xi = 0 and when xi = 1. The simplest such exponent is 12 xi + 12 x2i .
For the qij term, the simplest exponent is 1 x2i , and so on. Thus we write the
class-conditional probability for a single component xi as:

EXAM 2 SOLUTION
p(xi |j )

383
1

i = 1, . . . , d
j = 1, . . . , c

xi + 12 x2i 1x2i 12 xi + 12 x2i

qij rij

= pij2

and thus for the full vector x the conditional probability is

p(x|j )

d
,

xi + 12 x2i 1x2i 12 xi + 12 x2i

qij rij
.

pij2

i=1

Thus the discriminant functions can be written as

gj (x)

ln p(x|j ) + ln P (j )



d 
!
1
1
1
1
+ ln P (j )
=
xi + x2i ln pij + (1 x2i )ln qij + xi + x2i ln rij
2
2
2
2
i=1

d
d
d
!
pij rij
1!
pij !
=
x2i ln
+
xi ln
+
ln qij + ln P (j ),
qij
2 i=1
rij
i=1
i+1

which are quadratic functions of the components xi .

2. (15 points total) Consider a one-dimensional two-category classication problem with equal priors, P (1 ) = P (2 ) = 1/2, where the densities have the
form

0
x<0
p(x|i ) =
i ei x x 0,
where the i for i = 1, 2, are positive but unknown parameters.
(a) (1 pt) Conrm that the distributions are normalized.
Solution: We can drop the subscripts and perform the integral




ex dx = ex 0 = 0 (1) = 1.

(b) (8 pts) The following data were collected: D1 = {1, 5} and D2 = {3, 9} for
1 and 2 , respectively. Find the maximum-likelihood values 1 and 2 .
Solution: We temporarily drop the subscript on , denote the two training
points as x1 and x2 , and write the likelihood p(D|) for any category as
p(x1 |)p(x2 |)

= ex1 ex2
= 2 e(x1 +x2 ) .

We take the derivative of this likelihood and set it to zero:

d $ 2 (x1 +x2 ) %
= 2e(x1 +x2 ) + 2 ((x1 + x2 ))e(x1 +x2 )
e
d
(x1 +x2 )
= e
(x1 + x2 )] = 0,


 [2



=0

=0

and thus = 2/(x1 +x2 ). For our data we have, then, 1 = 2/(1+5) = 1/3,
and 2 = 2/(3 + 9) = 1/6.

SAMPLE FINAL EXAMS AND SOLUTIONS

384

(c) (3 pts) Given your answer to part (b), determine the decision boundary
x for minimum classication error. Be sure to state which category is to
right (higher) values than x , and which to the left (lower) values than x .
Solution: The decision boundary is at the point x where the two posteriors are equal, that is, where

P (1 )1 e1 x = P (2 )2 e2 x ,

or 1/3e1/3x = 1/6e1/6x . We multiply each side by 6 and take the

natural logarithm to nd ln(2) x /3 = x /6, or x = 6ln(2)  4.159
with R1 corresponding to points less than x , and R2 points greater than
x .
p(x| i)
0.25

0.2
0.15
0.1

0.05
2

R1

x*

10

R2

(d) (3 pts) What is the expected error of your classier in part (c)?
Solution: The probability of error (in this case, the Bayes error rate) is
P



1
1
Min P (1 ) ex/3 , P (2 ) ex/6 dx
3
6
0
x


 1
1 x/3
= 0.5
dx
ex/6 dx +
e
6

3
x
0

x
 :
1
1


= 0.5 6ex/6  + 3ex/3 
6
3
x
0
8
9

= 0.5 ex /6 + 1 0 + ex /3 = 0.375.


EXAM 2 SOLUTION

385

3. (10 points total) Consider the application of the k-means clustering algorithm
to the one-dimensional data set D = {0, 1, 5, 8, 14, 16} for c = 3 clusters.
(a) (3 pt) Start with the three cluster means: m1 (0) = 2, m2 (0) = 6 and
m3 (0) = 9. What are the values of the means at the next iteration?
Solution: The top of the gure shows the initial state (i.e., at iteration
t = 0). The dashed lines indicate the midpoints between adjacent means,
and thus the cluster boundaries. The two points x = 0 and x = 1 are in
the rst cluster, and thus the mean for the cluster 1 in the next iteration
is m1 (1) = (0 + 1)/2 = 0.5, as shown. Likewise, initially cluster 2 contains
the single point x = 5, and thus the mean for cluster 2 on the next iteration
is m2 (1) = 5/1 = 5. In the same manner, the mean for cluster 3 on the
next iteration is m3 (1) = (8 + 14 + 16)/3 = 12.67.
m1

m2

m3
x

t=0
0

m1

10

m2

15

20

15

20

m3
x

t=1
0

m1

10

m2

m3
x

t=2
0

x=3
(cluster 1)

10

15

20

x = 11
(cluster 3)

(b) (5 pt) What are the nal cluster means, after convergence of the algorithm?
Solution: After one more step of the algorithm, as above, we nd m1 (2) =
0.5, m2 (2) = 6.5 and m3 (2) = 15.
(c) (2 pt) For your nal clusterer, to which cluster does the point x = 3
belong? To which cluster does x = 11 belong?
Solution: As shown in the bottom gure, x = 3 is in cluster 1, and x = 11
is in cluster 3.
4. (10 points total) The task is to use Bayesian methods to estimate a one-dimensional
probability density. The fundamental density function is a normalized triangle
distribution T (, 1) with center at with half-width equal 1, dened by

1 |x | |x | 1
p(x|) T (, 1) =
0
otherwise,
as shown on the left gure. The prior information on the parameter is that
it is equally likely to come from any of the three discrete values = 1, 0 or 1.
Stated mathematically, the prior consists of three delta functions, i.e.,
p() =

1
[(x 1) + (x) + (x + 1)],
3

as shown on the gure at the right. (Recall that the delta function has negligible
width and unit integral.)

SAMPLE FINAL EXAMS AND SOLUTIONS

386

p(x|)
1

p()

x
-1

+1

-1

(a) (2 pt) Plot the estimated density before any data are collected (which
we denote by D0 = {}). That is, plot p(x|D0 ). Here and below, be sure to
label and mark your axes and ensure normalization of your nal estimated
density.
Solution: In the absense of data, the estimate density is merely the sum
of three triangle densities (of amiplitude 1/3 to ensure normalization), as
shown in the top gure.
p(x|D0)
1

a)

x
-2

-1

p(x|D1)
b)

x
-2

-1

p(x|D2)
c)

x
-2

-1

(b) (4 pts) The single point x = 0.25 was sampled, and thus D1 = {0.25}.
Plot the estimated density p(x|D1 ).
Solution: Bayesian density estimation

p(x|D) =
p(x|)p(|D) d


p(x|)p(D|)p() d
where here we do not worry about constants of proportionality as we shall
normalize densities at the end. Because the prior consists of three delta
functions, our nal estimated density consists of the sum of three triangle
distributions, centered on x = 1, 0 and 1 the only challenge is to

EXAM 2 SOLUTION

387

determine the relative weighting of these triangle distributions. For the

single point D = {0.25}, clearly the weighting of the lefthand triangle (i.e.,
the one centered on x = 1) is zero because p(D1 | = 1) = 0. The
relative weighting of the middle triangle (i.e., the one centered on x = 0) is
0.75 because p(D1 | = 0) = 0.75. Likewise, the weighting of the righthand
triangle (i.e., centered on x = 1) is 0.25 because p(D1 | = 1) = 0.25. The
priors are the same for all triangles, and thus our nal density is:
p(x|D1 ) 0.75T (0, 1) + 0.25T (1, 1),
as shown in the gure.
(c) (4 pts) Next the point x = 0.75 was sampled, and thus the data set is
D2 = {0.25, 0.75}. Plot the estimated density p(x|D2 ).
Solution: As in part (b), there will be no contribution from the lefthand
triangle because p(D2 | = 1) = 0. Because the points are placed symmetrically between the two other triangle centers, the contributions of these
triangles must be equal, that is,
p(x|D2 ) 0.5T (0, 1) + 0.5T (1, 1),
as shown in the gure.
5. (5 points total) Construct a cluster dendrogram for the one-dimensional data
D = {2, 3, 5, 10, 13} using the distance measure Dmax (Di , Dj ).
Solution: The closest two points are 2 and 3, with distance Dmax = 2, and thus
they are merged. At Dmax = 3, the point 5 is merged, since the distance is 3.
Likewise, the points 10 and 13 are merged at this level. The two clusters {2, 3, 5}
and {10, 13} are merged at Dmax = 11, the Euclidean separation between 2 and
13.
0

10
Dmax(Di|Dj)

10

15

SAMPLE FINAL EXAMS AND SOLUTIONS

388

6. (10 points total) Consider the use of traditional boosting for building a classier
for a two-category problem with n training points.
(a) (8 pts) Write pseudocode for traditional boosting, leading to three component classiers.
Solution:
Algorithm[Boosting]
Begin INITIALIZE D = {x1 . . . xn }
Train classier C1 on D1 , i.e., n/3 patterns chosen from D
Select D2 , i.e., roughly n/3 patterns that are most informative
Train classier C2 on D2
Train classier C3 on D3 , i.e., all remaining patterns
End
(b) (2 pts) How are the resulting three component classiers used to classify
a text pattern?
Solution: Majority vote.
7. (5 points total) Consider a standard three-layer neural net as shown. Suppose
the network is to be trained using the novel criterion function
1!
(tk zk )4 .
4
c

J=

k=1

Derive the learning rule wkj for the hidden-to-output weights.

t1

target t

output z

...

t2

z1

z2

...

tk

zk

...

tc

...

zc

output

wkj

y1

y2

yj

...

...

yn H

hidden

wji

x1

input x

x1

x2

x2

xi

...
...

xi

...
...

xd

input
xd

Solution: The derivation is nearly the same as the one in the text. The chain
rule for dierentiation gives us:
J
J netk
=
,
wkj
netk wkj
and because
1!
(tk zk )4 ,
4
c

J=

k=1

we have
J
netk

J zk
zk netk
= (tk zk )3 f  (netk ).
=

EXAM 2 SOLUTION

389

Recall, too, that

netk
= yj .
wkj
We put these together, and denote the learning rate by to nd the learning
rule
wkj = (tk zk )3 f  (netk )yj .
8. (10 points total) It is World Cup Soccer season and a researcher has developed
a Bayes belief net that expresses the dependencies among the age of a person
(A), his or her nationality (B), whether he or she likes soccer (C), and how
much he or she watches sports TV during the World Cup season (D). Use the
conditional probability tables to answer the following.
P(b)
P(a)
a1 = 0-30 years
b
=
American
A
P(b
) P(b2)
B
1
P(a1) P(a2) P(a3) a = 31-40 years
1
0.3

0.6

P(c|a,b)

0.1

a1, b 1
a1 , b 2
a2 , b 1
a2 , b 2
a3 , b 1
a3 , b 2

a3 = 41-100 years

P(c1|ai, bj)
0.5
0.7
0.6
0.8
0.4
0.1

age

P(c2|ai, bj)
0.5
0.3
0.4
0.2
0.6
0.9

nation- b2 = non-American
ality

sports

0.2

0.8

c1 = likes soccer
c2 = doesn't like soccer

P(d|c)
D

watch
TV

d1 = a lot
d2 = some
d3 = none

c1
c2

P(d1|ck)
0.7
0.5

P(d2|ck) P(d3|ck)
0.2
0.1
0.3
0.2

(a) (2 pts) What is the probability we nd a non-American who is younger

than 30 who likes soccer and watches a lot of sports TV?
Solution: We use the conditional probabilities and compute
= P (a1 )P (b2 )P (c1 |a1 , b2 )P (d1 |c1 )
= 0.3 0.8 0.7 0.7 = 0.1176.

P (a1 , b2 , c1 , d1 )

(b) (4 pts) Suppose we nd someone who is an American between 31-40 years

of age who watches some sports TV. What is the probability that this
person likes soccer?
Solution: There are several ways to approach this, but perhaps the simplest is to compute
P (c1 |a2 , b1 , d2 )

P (a2 , b1 , c1 , d2 )
= P (a2 , b1 , c1 , d2 )
P (a2 , b1 , d2 )
= P (a2 )P (b1 )P (c1 |a2 , b1 )P (d2 |c1 )
= 0.6 0.2 0.6 0.2 = 0.0144.
=

Likewise we have
P (c2 |a2 , b1 , d2 ) = P (a2 , b1 , c2 , d2 )
= P (a2 )P (b1 )P (cc |a2 , b1 )P (d2 |cc )
= 0.6 0.2 0.4 0.3 = 0.0144.

SAMPLE FINAL EXAMS AND SOLUTIONS

390

Since P (c1 |a2 , b1 , d2 ) + P (c2 |a2 , b1 , d2 ) = 1, we x the normalization and

nd
P (c1 |a2 , b1 , d2 ) =

0.0144
= 0.5.
0.0144 + 0.0144

(c) (4 pts) Suppose we nd someone over 40 years of age who never watches
sports TV. What is the probability that this person likes soccer?
Solution: Again, there are several ways to calculate this probability, but
this time we proceed as:
!
P (c1 |a3 , d3 )
P (a3 , b, c1 , d3 )
b

= P (a3 )P (d3 |c1 )

P (b)P (c1 |a3 , b)

= P (a3 )P (d3 |c1 ) [P (b1 )P (c1 |a3 , b1 ) + P (b2 )P (c1 |a3 , b2 )]

= 0.1 0.1 [0.2 0.4 + 0.8 0.1] = 0.0016.
Likewise, we have
!

P (c2 |a3 , d3 )

P (a3 , b, c2 , d3 )

= P (a3 )P (d3 |c2 )

P (b)P (c2 |a3 , b)

= P (a3 )P (d3 |c2 ) [P (b1 )P (c2 |a3 , b1 ) + P (b2 )P (c2 |a3 , b2 )]

= 0.1 0.2 [0.2 0.6 + 0.8 0.9] = 0.0168.
Since P (c1 |a3 , d3 ) + P (c2 |a3 , d3 ) = 1 we normalize our results to nd
P (c1 |a3 , d3 ) =

0.0016
= 0.087.
0.0016 + 0.0168

9. (15 points total) This problem conserns the construction of a binary decision
tree for three categories from the following two-dimensional data:
1

 1  2  3
1 , 8 , 5

 4  6 7 
7 , 2 , 6

 5  7  7 
10 , 4 , 9

(a) (3 pts) What is the information impurity at the root node, i.e., before any
splitting?
Solution: The information impurity at the root (before splitting) is
i(N )

c
!

P (j ) log2 P (j )

j=1

= 3


1
1
log2
= log2 3 = 1.585 bits.
3
3

(b) (4 pts) The query at the root node is: Is x1 > 3.5? What is the
information impurity at the next level?

EXAM 2 SOLUTION

391

Solution: The impurity at the right node is 0, as all its points are in a
single category, 1 . The impurity at the left node, NL , is:


1 1
1
1
i(NL ) =
log2 + log2
= log2 2 = 1.0 bit.
2
2 2
2
The impurity at this level is the weighted sum of the impurities at the
nodes, i.e.,
PL 1.0 + (1 PL )0.0 =

6
3
1.0 + 0.0 = 0.66 bits,
9
9

where PL and PR = 1 PL are the probabilities patterns go to the left

and right, respectively. Note that the reduction in impurity 1.585 0.66 is less than 1.0 bits, as it must from the answer to a single yes-no
question.
(c) (6 pts) Continue to grow your tree fully. If two candidate splits are equally
good, prefer the one based on x1 (rather than x2 ). Show the nal tree and
all queries.
Solution: See gure.
x2
x1 > 3.5?

10
Y

N
1

x2 > 8?
Y
3

N
x1 > 6.5?
Y

N
2

x2 > 5?
0
0

10

x1

Y
2

N
3

part d)

(d) (2 pts) Use your tree to classify the point ( 72 ).

Solution: From the tree or the plot, it is clear that ( 72 ) is classied as 3 .
10. (10 points total) Short answer (1 pt each).
(a) Explain using a diagram and a few sentences the technique of learning
with hints and why it can improve a neural net pattern classier.
Solution: If a neural network is being trained on a particular classication problem, we can add to the network output units that correspond to
a subsidiary, but related problem. We train the network on both problems simultaneously. We discard the hint output units after training. Such
learning with hints can improve classication on the central task because
the hidden units develop better features, relevant to the primary classicaation task.
(b) Use the Boltzmann factor to explain why at a suciently high temperature T , all congurations in a Boltzmann network are equally probable.

SAMPLE FINAL EXAMS AND SOLUTIONS

392

categories

output

...

h1

hints

h2

hidden

input

Solution: The probability of any conguration in a Boltzmann network

is given by
P () =

eE /T
,
Z

where the numerator is the Boltzmann factor and Z is the partition function basically a normalization. If T  E , then the exponent is nearly
0 regardless of , and the Boltzmann factor is nearly 1.0 for every state.
Thus the probability of each conguration is roughly the same.
(c) Explain with a simple gure the crossover operation in genetic algorithms.
Solution: In genetic algorithms, we represent classiers by a string of bits.
The crossover operation employs a randomly chosen place along the string,
then takes the left part from sequence A and splices it to the right side of
sequence B, and vice versa, much in analogy with sexual reproduction in
biology. This operation is fundamentally dierent from random variation
and occassionally leads to particularly good or t classiers.

A
B

1010100001010011101010010100010101111101011100

parents

1110001101100001010011110101000000100100101010

1010100001010001010011110101000000100100101010

offspring

1110001101100011101010010100010101111101011100

(d) What is a surrogate split and when is one used?

Solution: In addition to the traditional or primary split in a node in
a decision tree classier, we may provide surrogate splits, to be used
whenever a test pattern is missing the feature queried by the primary split.
The surrogate split is based on a feature other than the one used by the
primary split, and is chosen to best approximate the action of the primary
split, i.e., maximize the predictive association with the primary split.
(e) If the cost for any fundamental string operation is 1.0, state the edit distance between bookkeeper and beekeepers.
Solution: The edit distance is 4, as shown in the table.

EXAM 2 SOLUTION

393
x

bookkeeper
beokkeeper
beekkeeper
beekeeper
beekeepers

source string
substitute e for o
substitute e for o
delete k
insert s

1
2
3
4

(f) Suppose the Bayes error rate for a c = 3 category classication problem is
5%. What are the bounds on the error rate of a nearest-neighbor classier
trained with an innitely large training set?
Solution: In the limit of an innitely large training set, the nearestneighbor classier has a probability of error P bounded as

P P P

c
2
P
c1

where P is the Bayes error rate. For the case P = 0.05 and c = 3, these
limits are then



3
0.05 P 0.05 2 0.05 = 0.09625.
2

(g) What do we mean by the language induced by grammar G?

Solution: The complete set of strings (possibly innite in number) that
can be derived from the source symbol S by application of the production
rules in all possible sequences.

(h) In hidden Markov models, what does the term aij refer to? What does bjk
refer to?
Solution: The aij = P (j (t + 1)|i (t)) denote the transition probability
from hidden state i to hidden state j in each time step. The bjk =
P (vk (t)|j (t)) denote the transition probability that hidden state j will
emit visible symbol vk in each time step.

SAMPLE FINAL EXAMS AND SOLUTIONS

394

EXAM 3, three hours, 100 points

1. (15 points total) A Bayes belief net consisting of six nodes and its associated
conditional probability tables are shown below.

A
C

B
D
E
P (a1 )
0.5

P (a2 )
0.3

P (a3 )
0.2

a1
a2
a3

P (b1 |ai )
0.4
0.3
0.5

P (b2 |ai )
0.6
0.7
0.5

d1
d2

P (e1 |di )
0.1
0.8

P (e2 |di )
0.9
0.2

F
P (c1 )
0.2

b1 , c1
b1 , c2
b1 , c3
b2 , c1
b2 , c2
b2 , c3

d1
d2

P (c2 )
0.4

P (c3 )
0.4

P (d1 |bi , cj )
0.3
0.5
0.9
1.0
0.4
0.7
P (f1 |di )
0.1
0.8

P (d2 |bi , cj )
0.7
0.5
0.1
0.0
0.6
0.3

P (f2 |di )
0.5
0.0

P (f3 |di )
0.4
0.2

(a) (2 pt) Compute the probability P (a3 , b2 , c3 , d1 , e2 , f1 ).

(b) (2 pt) Compute the probability P (a2 , b2 , c2 , d2 , e1 , f2 ).
(c) (5 pt) Suppose we know the net is in the following (partial) state of evidence
e: a3 , b1 , c2 . What is the probability P (f1 |e)? What is the probability
P (e2 |e)?
(d) (6 pt) Suppose we know the net is in the following (partial) state of evidence
e: f1 , e2 , a2 . What is the probability P (d1 |e)? What is the probability
P (e2 |e)?

EXAM 3

395

2. (15 points total) Consider a one-dimensional two-category classication problem with unequal priors, P (1 ) = 0.7 and P (2 ) = 0.3, where the densities have
the form form of a half Gaussian centered at 0, i.e.,

p(x|i ) =

x2 /(2i2 )

i e

x<0
x 0,

where the i for i = 1, 2, are positive but unknown parameters.

(a) (1 pt) Find the normalization constant i as a function of i .

(b) (8 pts) The following data were collected: D1 = {1, 4} and D2 = {2, 8} for
1 and 2 , respectively. Find the maximum-likelihood values
1 and
2 .
(c) (3 pts) Given your answer to part (b), determine the decision boundary
for minimum classication error. Be sure to state which category label
applies to each range in x values.
(d) (3 pts) Recall that the standard error function is dend as erf (x)
x z2 /2
2
e
dz. Write a formula for the expected error of your classier in

part (c) in terms of error functions.

3. (10 points total) Consider the application

the
k-means
  of
 2
 1 clustering
 4   6  algorithm
3
to the two-dimensional data set D = { 5
,
,
,
,
3
1
6
7
3 , 1 } for c = 3
clusters.

 

(a) (3 pt) Start with the three cluster means: m1 (0) = 7
, m2 (0) = 74 ,
4
2
and m3 (0) = 5
. What are the values of the means at the next iteration?
(b) (5 pt) What are the nal cluster means, after convergence of the algorithm?
(c) (2 pt) For your nal clusterer, to which cluster does the point x =
 
belong? To which cluster does x = 3
4 belong?

 3
3

SAMPLE FINAL EXAMS AND SOLUTIONS

396

4. (15 points total) The task is to use Bayesian methods to estimate a one-dimensional
probability density. The fundamental density function is a normalized half circle distribution HC(, 1) with center at with half-width equal 1, dened
by
 2
1 (x )2 |x | 1

p(x|) HC(, 1) =
0
otherwise,
as shown on the left gure. The prior information on the parameter is that
it is equally likely to come from either of the two discrete values = 0.5 or
+0.5. Stated mathematically, the prior consists of two delta functions, i.e.,
p() =

1
[( 0.5) + ( + 0.5)],
2

as shown on the gure at the right. (Recall that the delta function has negligible
width and unit integral.)

p(x|)
2/pi

p()

x
-1

+1

-0.5

+0.5

(a) (3 pt) Plot (sketch) the estimated density before any data are collected
(which we denote by D0 = {}). That is, plot p(x|D0 ). Here and below, be
sure to label and mark your axes and ensure normalization of your nal
estimated density.
(b) (4 pts) The single point x = 0.25 was sampled, and thus D1 = {0.25}.
Plot the density p(x|D1 ) estimated by Bayesian methods.
(c) (5 pts) Next the point x = 0.25 was sampled, and thus the data set is
D2 = {0.25, 0.25}. Plot the estimated density p(x|D2 ).
(d) (3 pts) Suppose a very large number of points were selected and they were
all 0.25, i.e., D = {0.25, 0.25, . . . , 0.25}. Plot the estimated density p(x|D).
(You dont need to do explicit calculations for this part.)
5. (5 points total) Construct a cluster dendrogram for the one-dimensional data
D = {5, 6, 9, 11, 18, 22} using the distance measure Davg (Di , Dj ).
6. (5 points total) Consider learning a grammar from sentences.
(a) (8 pts) Write pseudocode for simple grammatical inference. Dene your
terms.
(b) (2 pts) Dene D+ and D and why your algorithm needs both.

EXAM 3

397

7. (5 points total) Consider a standard three-layer neural net as shown. Suppose

the network is to be trained using the novel criterion function
1!
(tk zk )6 .
6
c

J=

k=1

Derive the learning rule wkj for the hidden-to-output weights.

t1

target t
output z

...

t2

z1

z2

...

tk

zk

...

tc

...

zc

output

wkj

y1

y2

yj

...

...

yn

hidden

wji

x1

input x

x2

x1

x2

xi

...
...

xi

...
...

xd

input
xd

8. (5 points total) Prove that the single best representative pattern x0 for a data
set D = {x1 , . . . , xn } in the sum-squared-error criterion

J0 (x0 ) =

n
!

x0 xk 2

k=1

is the sample mean m =

1
n

n

k=1

xk .

9. (15 points total) This problem concerns the construction of a binary decision
tree for two categories from the following two-dimensional data using queries of
the form Is xi > xi ? for i = 1, 2 and the information impurity.
1

 1  2  4  5 8
5 , 9 , 10 , 7 , 6

 3  6 7 9
8 , 4 , 2 , 3

(a) (2 pts) What is the information impurity at the root node, i.e., before any
splitting?
(b) (3 pts) What should be the query at the root node?
(c) (3 pts) How much has the impurity been reduced by the query at the root?
(d) (3 pts) Continue constructing your tree fully. (Whenever two candidate
queries lead to the samee reduction in impurity,
the query

  that uses
 prefer
the x1 feature.) Use your tree to classify x = 66 and x = 34 .
(e) (2 pts) Suppose your tree is to be able to classify decient patterns. What
should be the rst (and only) surrogate split at the root node?

398

SAMPLE FINAL EXAMS AND SOLUTIONS

10. (10 points total) Short answer (1 pt each).

(a) What are the four major components of a grammar G? What do we mean
by the language induced by grammar G, i.e., L(G)?
(b) Use the Boltzmann factor to explain why at a suciently high temperature T , all congurations in a Boltzmann network are equally probable.
(c) Use an equation and a few sentences to explain the minimum description
length (MDL) principle.
(d) Use an equation and a few sentences to explain what is the discriminability
in signal detection theory.
(e) If the cost for any fundamental string operation is 1.0, state the edit distance between streets and scrams.
(f) Suppose the Bayes error rate for a c = 5 category classication problem is
1%. What are the upper and lower bounds on the error rate of a nearestneighbor classier trained with an innitely large training set?
(g) Use a formula and a sentence to explain learning with momentum in backpropagation.
(h) What is the evaluation problem in hidden Markov models?

EXAM 3

399

Important formulas
p(x|i )P (i )
P (i |x) = 
c
P (k )p(x|k )
k=1

p(x) =



1
t 1
exp

(x

)
(x

)
2
(2)d/2 ||1/2
1

= E[(x )(x )t ]
n
1!
=
(xi )(xi )t
n i=1

=
(x )(x )t p(x)dx

P (i |xg ) =

P (i |xg , xb )p(xg , xb )dxb

p(xg )

l() = lnp(D|)

= arg max l()


p(x|D) =

p(x|)p(|D) d



Q(; i ) = EDb lnp(Dg , Db ; )|Dg ; i

d =

|2 1 |

lim Pn (e|x) = 1

P P P

c
!

P 2 (i |x)

i=1

c
2
P
c1

SAMPLE FINAL EXAMS AND SOLUTIONS

400

Jp (a)

(at y)

yY

a(k + 1)

= a(k) + (k)

yYk

= k yj = (tk zk )f  (netk )yj


 c
!
= j xi =
wkj k f  (netj )xi

wkj
wji

k=1


f (net) = a tanh[b net] = a

P () =

Z=

e+b
e+b

eb
net + eb
net

net

net

eE /T
Z
eE  /T

wij =

EQ [si sj ]i o
T 


E[si sj ]i clamped
 



clamped

learning

i(N ) =

c
!

unlearning

P (j )log2 P (j )

j=1





ED (g(x; D) F (x))2 = (ED [g(x; D) F (x)])2 + ED g(x; D) ED [g(x; D)])2

(i)
()

1 , x2 , , xi1 , xi+1 , , xn )
= (x
n
1 !
(i)
=
n i=1
Si =

(x mi )(x mi )t

xDi

SW =

c
!
i=1

Si

EXAM 3

401

SB =

c
!

ni (mi m)(mi m)t

i=1

ST =

(x m)(x m)t

xD

Je = tr[SW ]

Jd = SW 

min x x 

dmin (Di , Dj )

dmax (Di , Dj )

max x x 

davg (Di , Dj )

1 ! !
x x 
ni nj


dmin (Di , Dj )

= mi mj 

xD
xD 

xD
xD 

xD xD

(dij ij )2
i<j
 2
ij

Jee

Jf f

!  dij ij 2
=
ij
i<j

Jef

i<j

1 ! (dij ij )2

ij i<j
ij

i<j

wki (t + 1) = wki (t) + (t)(|y y |)i

SAMPLE FINAL EXAMS AND SOLUTIONS

402

EXAM 3 Solutions
1. (15 points total) A Bayes belief net consisting of six nodes and its associated
conditional probability tables are shown below.

A
C

B
D
E
P (a1 )
0.5

P (a2 )
0.3

P (a3 )
0.2

a1
a2
a3

P (b1 |ai )
0.4
0.3
0.5

P (b2 |ai )
0.6
0.7
0.5

d1
d2

P (e1 |di )
0.1
0.8

P (e2 |di )
0.9
0.2

F
P (c1 )
0.2
b1 , c1
b1 , c2
b1 , c3
b2 , c1
b2 , c2
b2 , c3
d1
d2

P (c2 )
0.4

P (c3 )
0.4

P (d1 |bi , cj )
0.3
0.5
0.9
1.0
0.4
0.7
P (f1 |di )
0.1
0.8

P (d2 |bi , cj )
0.7
0.5
0.1
0.0
0.6
0.3

P (f2 |di )
0.5
0.0

P (f3 |di )
0.4
0.2

(a) (2 pt) Compute the probability P (a3 , b2 , c3 , d1 , e2 , f1 ).

Solution
P (a3 , b2 , c3 , d1 , e2 , f1 )

= P (a3 )P (b2 |a3 )P (c3 )P (d1 |b2 , c3 )P (e2 |d1 )P (f1 |d1 )
= 0.2 0.5 0.4 0.7 0.9 0.1 = 0.0025.

(b) (2 pt) Compute the probability P (a2 , b2 , c2 , d2 , e1 , f2 ).

Solution
P (a2 , b2 , c2 , d2 , e1 , f2 )

= P (a2 )P (b2 |a2 )P (c2 )P (d2 |b2 , c2 )P (e1 |d2 )P (f2 |d2 )
= 0.3 0.7 0.4 0.6 0.8 0.0 = 0.0.

(c) (5 pt) Suppose we know the net is in the following (partial) state of evidence
e: a3 , b1 , c2 . What is the probability P (f1 |e)? What is the probability
P (c2 |e)?
Solution
P (f1 |e)

= P (d1 |b1 , c2 )P (f1 |d1 ) + P (d2 |b1 , c2 )P (f1 |d2 )

=

P (e2 |e)

0.5 0.1 + 0.5 0.8 = 0.45.

= P (d1 |b1 , c2 )P (e2 |d1 ) + P (d2 |b1 , c2 )P (e2 |d2 )

= 0.5 0.9 + 0.5 0.2 = 0.55.

EXAM 3 SOLUTION

403

(d) (6 pt) Suppose we know the net is in the following (partial) state of evidence
e: f1 , e2 , a2 . What is the probability P (d1 |e)? What is the probability
P (e2 |e)?
Solution

b,c
P (d1 |e) = 

b,c,d


=

P (b, c, d1 , a2 , e2 , f1 )

b,c

P (b, c, d, a2 , e2 , f1 )

P (a2 )P (b|a2 )P (c)P (d1 |b, c)P (e2 |d1 )P (f1 |d1 )

P (a2 )P (b|a2 )P (c)P (d|b, c)P (e2 |d)P (f1 |d)


P (e2 |d1 )P (f1 |d1 ) b,c P (b|a2 )P (c)P (d1 |b, c)

b,c,d P (b|a2 )P (c)P (d|b, c)P (e2 |d)P (f1 |d)


0.9 0.1[P (b1 |a2 ) c P (c)P (d1 |b1 , c)) + P (b2 |a2 ) c P (c)P (d1 |b2 , c)]


P (e2 |d1 )P (f1 |d1 ) b,c P (b|a2 )P (c)P (d1 |b, c) + P (e2 |d2 )P (f1 |d2 ) b,c P (b|a2 )P (c)P (d2 |b, c)
b,c,d

=
=

0.9 0.1[0.3(0.2 0.3 + 0.4 0.5 + 0.4 0.9) + 0.7(0.2 1.0 + 0.4 0.4 + 0.4 0.7)]


P (e2 |d1 )P (f1 |d1 ) b,c P (b|a2 )P (c)P (d1 |b, c) + P (e2 |d2 )P (f1 |d2 ) b,c P (b|a2 )P (c)P (d2 |b, c)
0.05706
=
0.05706 + 0.2 0.8(0.3(0.2 0.7 + 0.4 0.5 + 0.4 0.1) + 0.7(0.2 0.0 + 0.4 0.6 + 0.4 0.3))
0.05706
=
0.4935.
0.05706 + 0.05856

b,d P (b, d, c2 , a2 , e2 , f1 )
P (c2 |e) = 
b,c,d P (b, c, d, a2 , e2 , f1 )

b,d P (a2 )P (b|a2 )P (c2 )P (d|b, c2 )P (e2 |d)P (f1 |d)
= 
b,c,d P (a2 )P (b|a2 )P (c)P (d|b, c)P (e2 |d)P (f1 |d)


P (c2 ) b P (b|a2 ) d P (d|b, c2 )P (e2 |d)P (f1 |d)


= 
c P (c)
b P (b|a2 )
d P (d|b, c)P (e2 |d)P (f1 |d)
0.4(0.3(0.5 0.9 0.1 + 0.5 0.2 0.8) + 0.7(0.4 0.9 0.1 + 0.6 0.2 0.8))



=
c P (c)
b P (b|a2 )
d P (d|b, c)P (e2 |d)P (f1 |d)
1
!
!
= .05196/ .05196 + P (c1 )
P (b|a2 )
P (d|b, c1 )P (e2 |d)P (f1 |d)
=

+ P (c3 )

!
b

P (b|a2 )

P (d|b, c3 )P (e2 |d)P (f1 |d)

= .05196/ (.05196 + 0.2(0.3(0.3 0.9 0.1 + 0.7 0.2 0.8) + 0.7(1.0 0.9 0.1 + 0.0 0.2 0.8))
2
!
!
+ P (c3 )
P (b|a2 )
P (d|b, c3 )P (e2 |d)P (f1 |d)
b

= .05196/(.05196 + .02094
+ 0.4(0.3(0.9 0.9 0.1 + 0.1 0.2 0.8) + 0.7(0.7 0.9 0.1 + 0.3 0.2 0.8)))
.05196
=
0.4494.
.05196 + .02094 + .04272

404

SAMPLE FINAL EXAMS AND SOLUTIONS

2. (15 points total) Consider a one-dimensional two-category classication problem with unequal priors, P (1 ) = 0.7 and P (2 ) = 0.3, where the densities have
the form form of a half Gaussian centered at 0, i.e.,

0
x<0
2
2
p(x|i ) =
i ex /(2i ) x 0,
where the i for i = 1, 2, are positive but unknown parameters.
(a) (1 pt) Find the normalization constant i as a function of i .
Solution
Because the normalization on a full one-dimensional Gaussian (as given
1
on the equations on the exam) is 2
, the normalization on a half
Gaussian must be twice as large, i.e.,
0
2 1
i =
.
i
(b) (8 pts) The following data were collected: D1 = {1, 4} and D2 = {2, 8} for
1 and
2 .
1 and 2 , respectively. Find the maximum-likelihood values
Solution
We drop the subscripts, denote the two training points x1 and x2 , and
compute the likelihood:
p(D|)

= p({x1 , x2 }|) = p(x1 |)p(x2 |)

0
0
2 1 x21 /22
2 1 x22 /22
=

e
e

2 1 (x21 +x22 )/22

=
e
.
2

Clearly, because of the exponential we will want to work with the loglikelihood,
l ln(2/) 2 ln (x21 + x22 )/2 2 .
To nd the maximum-likelihood solution, we rst take the derivative
dl
2 (x21 + x22 ) 2
,
=

2
3
and set it to zero. Clearly we must ignore the solution
= . The solution
we seek is
0
x21 + x22

=
,
2
which for the training data given yields
0
0
12 + 42
17

1 =
=
2
2
0

22 + 82

2 =
= 34.
2

EXAM 3 SOLUTION

405

(c) (3 pts) Given your answer to part (b), determine the decision boundary
for minimum classication error. Be sure to state which category label
applies to each range in x values.
Solution
We set the posteriors equal, yielding
P (1 )p(x |1 ) = P (2 )p(x |2 )
which gives the following equalities:
0
0
2 1 x2 /(212 )
2 1 x2 /(222 )
0.7
e
= 0.3
e

2




x2
0.7
1
1
=
ln
0.3
2
2
22 +
12
or
;
%
$
<
< 2 ln 0.7 1
0.3
2
<
%  5.91,
x = = $
1

22 +
12

with R1 for points 0 x < x and R2 for points x > x .

(d) (3 pts) Recall that the standard error function is dend as erf(x)
x z2 /2
2
e
dz. Write a formula for the expected error of your classier

in part (c) in terms of error functions.

Solution
Consult the gure.
P(i|x)

0.15

1
0.1

0.05

10

x*

We sum the two shaded areas to nd the probability of error,


P (error) =
min[P (1 )p(x|1 ), P (2 )p(x|2 )]dx

x
=

0
P (2 )

2 1 x2 /(222
e
dx +

0
P (1 )

2 1 x2 /(212 )
e
dx

=
=

1 1 2

0.3
2
2

x/
2

z 2 /2

2 e
0

1 1 2

dz + 0.7
1
2

0.3
0.7
erf[x /
2 ] + [1 erf[x /
1 ]]  0.237.
2
2


x /
1

1 ez

/2

dz

406

SAMPLE FINAL EXAMS AND SOLUTIONS

3. (10 points total) Consider the application

the
k-means
 2
 1 clustering
 4   6  algorithm
  of
3
to the two-dimensional data set D = { 5
3 , 1 , 6 , 7 , 3 , 1 } for c = 3
clusters.

0
m1

m2

1
2
2

0
1m
3

 

(a) (3 pt) Start with the three cluster means: m1 (0) = 7
, m2 (0) = 74 ,
4
2
. What are the values of the means at the next iteration?
and m3 (0) = 5
Solution
The points nearest each mean are as follows:
  3 2
m1 : 5
3 , 1 , 6
6
m2 :
1
1  4 
m3 :
7 , 3 .
We compute the mean for each of these sets to nd the new means, i.e.,


 
 
3.33
6
1.5
m1 (1) =
,
m2 (1) =
,
m3 (1) =
.
3
1
5
(b) (5 pt) What are the nal cluster means, after convergence of the algorithm?
Solution
On the next iteration, m1 does not change, but the others do:


 
 
3.33
5
1
m1 (2) =
m2 (2) =
m3 (2) =
,
3
2
7
which is the nal state.
(c) (2 pt) For your nal clusterer, to which cluster does the point x =
 
belong? To which cluster does x = 3
4 belong?
Solution  
 
The point 33 is in cluster 2; the point 3
4 is in cluster 3.

3
3

EXAM 3 SOLUTION

407

4. (15 points total) The task is to use Bayesian methods to estimate a one-dimensional
probability density. The fundamental density function is a normalized half circle distribution HC(, 1) with center at with half-width equal 1, dened
by
 2
1 (x )2 |x | 1

p(x|) HC(, 1) =
0
otherwise,
as shown on the left gure. The prior information on the parameter is that
it is equally likely to come from either of the two discrete values = 0.5 or
+0.5. Stated mathematically, the prior consists of two delta functions, i.e.,
p() =

1
[( 0.5) + ( + 0.5)],
2

as shown on the gure at the right. (Recall that the delta function has negligible
width and unit integral.)

p(x|)
2/pi

p()

x
-1

+1

-0.5

+0.5

(a) (3 pt) Plot (sketch) the estimated density before any data are collected
(which we denote by D0 = {}). That is, plot p(x|D0 ). Here and below, be
sure to label and mark your axes and ensure normalization of your nal
estimated density.
Solution In the absense of training data, our distribution is based merely
on the prior values of the parameters. Written out in full, we have

0
p(x|D ) =
p(x|)p()d

=
HC(, 1)0.5[( 0.5) + ( + 0.5)]d
=

0.5 [HC(0.5, 1) + HC(0.5, 1)] ,

which is just the sum of the two disk distributions, as shown in the gure.
p(x|D0)
0.5
0.4
0.3
0.2
0.1
-2

-1

408

SAMPLE FINAL EXAMS AND SOLUTIONS

(b) (4 pts) The single point x = 0.25 was sampled, and thus D1 = {0.25}.
Plot the density p(x|D1 ) estimated by Bayesian methods.
Solution
We use recursive Bayesian learning, in particular p(|D1 ) p(x1 |)p(|D0 ,
where we defer the normalization.

1
p(x|) = p(x|)p(x1 |) [( 0.5) + ( + 0.5)]
2
The ratio of contributions of the two components is given by the values of
the component densities at x1 = 0.25, i.e.,


1 (0.5 0.25)2
1 1/16 

=
= 15/7 ,
2
1 9/16
1 (.5 0.25)
where we let be this ratio. The ratio of contributions is then /(1+) for
the HC(0.5, 1) component and 1/(1 + ) for the HC(0.5, 1) component,
i.e.,
p(x|D1 ) =

1
HC(0.5, 1) +
HC(0.5, 1).
1+
1+
p(x|D1)
0.5
0.4
0.3
0.2
0.1

-2

-1

(c) (5 pts) Next the point x = 0.25 was sampled, and thus the data set is
D2 = {0.25, 0.25}. Plot the estimated density p(x|D2 ).
Solution
We use recursive Bayesian learning, in particular p(|D2 ) p(x2 |)p(|D1 ,
where again we defer the normalization. As before, we now have /(1 + )
for the HC(0.5, 1) component and 1/(1 + ) for the HC(0.5, 1) component, i.e.,
p(x|D1 ) =

2
1
HC(0.5, 1) +
HC(0.5, 1).
1 + 2
1 + 2

p(x|D2)
0.5
0.4
0.3
0.2
0.1
-2

-1

(d) (3 pts) Suppose a very large number of points were selected and they were all 0.25,
i.e., D = {0.25, 0.25, . . . , 0.25}. Plot the estimated density p(x|D). (You dont need
to do explicit calculations for this part.)

EXAM 3 SOLUTION

409

Solution
Clearly, each subsequent piont x = 0.5 introduces another factor of into the ratio
of the two components. In the limit of large n, we have
n
n 1 + n
1
lim
n 1 + n
lim

1;

0,

and thus the density consists solely of the component HC(0.5, 1). Incidentally, this
would be the appropriate maximum-likelihood solution too.
p(x|D)
0.6
0.5
0.4
0.3
0.2
0.1
-2

-1

SAMPLE FINAL EXAMS AND SOLUTIONS

410

5. (5 points total) Construct a cluster dendrogram for the one-dimensional data D = {5, 6, 9, 11, 18, 22}
using the distance measure davg (Di , Dj ).
Solution
 
See gure. Throughout, we use davg (Di , Dj ) = ni1nj
x x . The
xD xD 

distances at each successive levels are: 1, 2, 4, 4.5, 12.25.

0

10

15

20
1
2
4
4.5
12.25

6. (5 points total) Consider learning a grammar from sentences.

(a) (8 pts) Write pseudocode for simple grammatical inference. Dene
your terms.
Solution
See Algorithm 5 in Chapter 8.
(b) (2 pts) Dene D+ and D and why your algorithm needs both.
Solution
D+ is a set of sentences known to be derivable in the grammar; D
is a set of sentences known to be not derivable in the grammar. One
needs both to reduce the number of grammars consistent with training
data and make learning tractable.
7. (5 points total) Consider a standard three-layer neural net as shown. Suppose the network
is to be trained using the novel criterion function
1!
(tk zk )6 .
6
c

J=

k=1

Derive the learning rule wkj for the hidden-to-output weights.

t1

target t
output z

...

t2

z1

z2

...

tk

zk

...

tc

...

zc

output

wkj

y1

y2

yj

...

...

yn H

hidden

wji

x1

input x

x1

x2

x2

xi

...
...

xi

...
...

xd

input
xd

EXAM 3 SOLUTION

411

Solution
We have J =

c


1
6

k=1

(tk zk )6 , and thus we have the derivative

J
zk

(tk zk )5 . We use the chain rule and nd

wkj =

J
J zk netk
=
= (tk zk )5 f  (netk )yj .
wkj
zk netk wkj

8. (5 points total) Prove that the single best representative pattern x0 for a data set D =
{x1 , . . . , xn } in the sum-squared-error criterion

J0 (x0 ) =

n
!

x0 xk 2

k=1

is the sample mean m =

Solution

1
n

n

k=1

xk .

Minimizing x0 under the sum-squared-error criterion

n

k=1

x0 xk 2

2
1
n
n
n
!
!
!
2
0=
x0 xk  = 2
(x0 xk ) = 2 n x0
xk
x0
k=1

k=1

k=1

and thus
1!
xk .
n
n

x0 =

k=1

9. (15 points total) This problem concerns the construction of a binary decision tree for two
categories from the following two-dimensional data using queries of the
form Is xi > xi ? for i = 1, 2 and the information impurity.
1

 1  2  4  5 8
5 , 9 , 10 , 7 , 6

 3  6 7 9
8 , 4 , 2 , 3

(a) (2 pts) What is the information impurity at the root node, i.e., before
any splitting?
Solution

i(N ) =

2
!

5
4
5
4
P (j ) log2 P (j ) = log2 + log2 0.9911
9
9
9
9
j=1

(b) (3 pts) What should be the query at the root node?

Solution
By inspection, the rst splitting criterion should be x2 = 4.5.

SAMPLE FINAL EXAMS AND SOLUTIONS

412

(c) (3 pts) How much has the impurity been reduced by the query at the
root?
Solution
Trivially, i(NR ) = 0.
1
5
1
5
i(NL ) = log2 + log2 0.65
6
6
6
6
The reduction in extropy is
6
i(N ) = i(N ) PL i(NL ) (1 PL )i(NR ) 0.99 0.65 0 0.56
9
.
(d) (3 pts) Continue constructing your tree fully. (Whenever two candidate queries lead to the same reduction in impurity, prefer the
 query
that uses the x1 feature.) Use your tree to classify x = 66 and

x = 34 .
Solution
We need to decide whether the second splitting criterion is x1 = 2.5
or x1 = 3.5. The entropy reduction for x1 = 2.5 is




1
1 5
5
1 3
3
4
1
i(N ) = log2 log2

log2 log2
0.11.
6
6 6
6
6
4
4 4
4
The entropy reduction for x1 = 3.5 is




1
1 5
5
1 2
2
3
1
i(N ) = log2 log2

log2 log2
0.19.
6
6 6
6
6
3
3 3
3
So the second split should be x1 = 3.5. The last split is then x1 = 2.5.
Thus the tree is
(e) (2 pts) Suppose your tree is to be able to classify decient patterns.
What should be the rst (and only) surrogate split at the root node?
Solution
x = (6, 6)t 1 ; x = (3, 4)t 2 . The surrogate split tries to
achieve the same partitioning of the samples as the x2 > 4.5 primary
split. By inspection, it is x1 < 5.5.

R1

R1

R2

EXAM 3 SOLUTION

413

10. (10 points total) Short answer (1 pt each).

(a) What are the four major components of a grammar G? What do we mean
by the language induced by grammar G, i.e., L(G)?
Solution
The four major components of a grammar is symbols in the alphabet
A, variables I, root symbol in set S, and productions P. The language
induced by a grammar G is the set of all strings that can be generated by
G.
(b) Use the Boltzmann factor to explain why at a suciently high temperature T , all congurations in a Boltzmann network are equally probable.
Solution
At high temperature T , the probability of all states are equal.
lim

P ()
= lim e(E E  )/T 1.
p(  ) T

(c) Use an equation and a few sentences to explain the minimum description
length (MDL) principle.
Solution
The MDL principle states that one should seek a model h that minimizes
the sum of the models algorithmic complexity and the description of the
training data with respect to that model. That is,
h = arg min K(h, D) = arg min K(h) + K(D using h).
h

(d) Use an equation and a few sentences to explain what is the discriminability in signal detection theory.
Solution
For two normally distributed classes in one dimension (with same variance), discriminability is a measure of the ease of discriminating the two
1|
classes. It is dened as d = |2
.

(e) If the cost for any fundamental string operation is 1.0, state the edit
distance between streets and scrams.
Solution
The cost is 4. streets screets scraets scramts scrams.
(f) Suppose the Bayes error rate for a c = 5 category classication problem is
1%. What are the upper and lower bounds on the error rate of a nearestneighbor classier trained with an innitely large training set?
Solution
c
Using the equation given with the exam, P P P (2 c1
P ), we
5
0.01) = 0.019875 0.02.
get 0.01 P 0.01(2 51
(g) Use a formula and a sentence to explain learning with momentum in backpropagation.
Solution
Momentum in neural net learning is a heuristic that weight changes should
tend to keep moving in the same direction. Let w(m) = w(m)w(m1)

414

SAMPLE FINAL EXAMS AND SOLUTIONS

and wbp (m) be the change in w(m) that would be called for by the backJ
propagation algorithm (i.e., w
). Then,
w(m + 1) = w(m) + (1 )wbp (m) + w(m 1).
(h) What is the evaluation problem in hidden Markov models?
Solution
The evaluation problem is determining the probability that a particular
sequence of visible states was generated by an HMM model.

Worked examples

Below are worked examples, organized by book section, that may be of use to
students.
[to be written]

415

416

WORKED EXAMPLES

Errata and additions in the

text

Below are errata and minor alterations that improve the style or clarify the book.
To see which printing of the book you have, look on the third page of the text itself,
which faces the dedication page. At the bottom you will see:
Printed in the United States of America
10 9 8 7 6 5 4 3 2
The last number at the right gives the number of the printing; thus, as illustrated
here, 2 means that this is the second printing.

First and second printings

Below, line +7 means the seventh line from the top of the text body (not
including gure or table captions, or other headlines unless otherwise indicated), and
line -5 means the fth line from the bottom of the text body. In Algorithms, the
numbering refers to the line numbers within the algorithm itself. Thus Algorithm
4, line 5 means line 5 in Algorithm 4, not the fth line from the top of the page.

Front matter
page x line +6: Change 4.8 Reduced Coulomb to *4.8 Reduced Coulomb
page xv line -13: Change A.4.7 The Law of Total Probability and Bayes Rule to
A.4.7 The Law of Total Probability and Bayes Rule
page xviii Take the last sentence under Examples, In addition, in response to
popular demand, a Solutions Manual has been prepared to help instructors
who adopt this book for courses. and move it to be the nal sentence under
Problems, lower on the same page.
page xviii lines -10 -11: Change and they are generally to and they are typically
page xix line -15: Change Ricoh Silicon Valley to Ricoh Innovations

417

418

ERRATA IN THE TEXT

Chapter 1
page 1 line +4: Change data and taking to data and making
 
 
page 4 the only equation on the page: Change x = xx12 to x = xx12
page 11 line +21: Change us for practical, rather than to us for practical rather
than
page 14 line -4 in the caption to Figure 1.8: Change of the data impact both to
of the data aect both
page 19 line +15: Change is achieved in humans to is performed by humans

Chapter 2
page 21 line -6 in the footnote: Change should be written as pX (x|) to should
be written as px (x|)
page 21 line -4 in the footnote: Change clear that pX () and pY () to clear that
px () and py ()
page 22 second line after Eq. 3: Change probability (or posterior) probability to
probability (or posterior)
page 23 second line after Eq. 7: Change By using Eq. 1, we can to By using
Eq. 1 we can
page 26 rst line after Eq. 17: Change and 2 otherwise. to and otherwise decide
2 .
page 28 Equation 23: Change (11 22 )(21 11 ) to (11 22 )+(21 11 )
page 28 second line after Eq. 24: Change decision boundary gives to decision
boundary then gives
page 32 second line after Eq. 33: Change expected values by these to expected
values by these
page 36 rst equation after Eq. 49: Change Let us examine the discriminant to
Let us examine this discriminant
page 41 Figure 2.13, caption, line +2: Change unequal variance. to unequal
variance, as shown in this case with P (1 ) = P (2 ).
page 47 Equation 73: Change for0 to for 0 (i.e., add space)
page 47 line - 10: Change substituting the results in Eq. 73 to substituting this
into Eq. 73
page 47 line -2: Change This result is the so-called to This gives the so-called
page 48 Example 2, line +3: Change 4.11, to 4.06,
page 48 Example 2, line +4: Change 0.016382. to 0.0087.

Chapter 2

419

page 50 The x-axis label on Fig. 2.20: Change P (x < x |x 2 ) to P (x <

x |x 1 )
pages 56 62 At the time of the release of the rst printing, the problem of inference
in Bayes belief nets with loops was not fully solved. Since that time, however,
such a solution has emerged and for this reason Section 2.11 has been rewritten
accordingly. This revision is posted on the Wiley site.
page 66 Problem 2, part (b), line +2: Change for arbitrary ai and bi . to for
arbitrary ai and positive bi .
page 66 Problem 3, part (a) equation: End the equation with a period (full stop).
page 67 Problem 5, part (d): Change What is the minimax risk? to What is the
minimax risk for part (c)?
page 67 Problem 6, part (2), line +2: Change Determine the decision boundary
to Determine the single-point decision boundary
page 69 Move the title Section 2.4 to the top of the page so that Problem 13 is
now under Section 2.4.
page 71 Problem 20, part (a), line +1: Change we know only that a distribution
is nonzero in to we know solely that a distribution is nonzero only in
page 71 Problem 20, part (b), line +1: Change we know only that a distribution
is nonzero for to we know solely that a distribution is nonzero only for
page 71 Problem 23, at the center of the typeset equation, change and to and
(i.e., add space)
page 72 Problem 24, line +1: Change normal density for which ij = 0 to normal
density with mean , ij = 0
page 73 Problem 34, line +67: Change assume the distributions to assume
P (1 ) = P (2 ) = 0.5 and the distributions
page 73 Problem 34, part c), line +4: Change Bayes error is 0.5. to Bayes error
is 0.25.
page 75 Problem 37, rst equation: Change andP (1 ) to and P (1 ) (i.e., add
space)
page 75 Problem 37, part (c) equation: Change andp(x|2 ) to and p(x|2 ) (i.e.,
add space)
page 75 Problem 39, line +1: Change Use the signal detection to Use signal
detection
page 75 Problem 39, part (a), line +1: Change and P (x < x |x 2 ), taken to
and P (x > x |x 1 ), taken
page 75 Problem 39, part (b): Replace the last two sentences with Estimate d if
P (x > x |x 1 ) = 0.8 and P (x > x |x 2 ) = 0.3. Repeat for P (x > x |x
1 ) = 0.7 and P (x > x |x 1 ) = 0.4.

420

ERRATA IN THE TEXT

page 75 Problem 39, part (d): Replace the two equation lines with Case A: P (x >
x |x 1 ) = 0.8, P (x > x |x 2 ) = 0.3 or
Case B: P (x > x |x 1 ), P (x > x |x 2 ) = 0.7.
page 76 Problem 41, rst line after the equation: Change (2 1 )/i to (2
1 )/
page 76 Problem 41, part (b): Change dT = 1.0. to dT = 1.0 and 2.0.
page 76 Problem 41, part (c): Change P (x > x |x 1 ) = .2. to P (x > x |x
1 ) = .7.
page 76 Problem 76, part (e): Change measure P (x > x |x 2 ) = .9 and (x >
x |x 1 ) = .3. to measure P (x > x |x 2 ) = .3 and P (x > x |x 1 ) =
.9.
page 81 Computer exercise 6, part (b), line +1: Change Consider to Consider
the normal distributions
page 81 Computer exercise 6, part (b), equation: Change andp(x|2 ) to and
p(x|2 ) (i.e., add space)
page 81 Computer exercise 6, part (b), equation: Move with P (1 ) = P (2 ) =
1/2. out of the centered equation, and into the following line of text.
page 83 rst column, entry for [21], lines +3 4: Change Silverman edition, 1963.
to Silverman, 1963.

Chapter 3
page 88 Equation 9: Change to
page 91 Ninth line after Eq. 22: Change shall consider), the samples to shall
consider) the samples
page 99 Caption to rst gure, change starts our as a at to starts out as a at
page 100 line -5: Change are equivalent to to are more similar to
page 100 line -5 -4: Change If there are much data to If there is much data
page 102 line -5: Change (Computer exercise 22) to (Problem 22)
page 103 line -2: Change choice of an prior to choice of a prior
page 104 line +6: Change if to only if
page 104 line +16: Change only if to if


page 104 Equation 62: Make the usage and style of the summation sign ( ) uniform
in this equation. Specically, intwo places put
 the arguments beneath the
summation sign, that is, change DD to

DD

page 105 rst line after Eq 63: Change to this kind of scaling. to to such scaling.

Chapter 3

421

page 111 lines +9 10: Change constants c0 and x0 such that |f (x)| c0 |h(x)|
for all to constants c and x0 such that |f (x)| c|h(x)| for all
page 111 line +14: Change proper choice of c0 and x0 . to proper choice of c and
x0 .
page 116 Equation 86: Change et e to (et e 1)
page 125 Algorithm 1, line 1: Change i = 0 to i 0
page 126 rst line of the equation at the middle of the page: Change
Q(; 0 ) = Ex41 [lnp(xg , xb ; | 0 ; Dg )]
to
Q(; 0 ) = Ex41 [lnp(xg , xb ; )| 0 ; Dg ]
page 128 line +6, (second line after the Example): Change the EM algorithm, and
they to the EM algorithm as they
page 129 second line aboove Section 3.10.3: Change while the i are unobservable
to while the j are unobservable
page 132 line +3: Change bkj , and thus to bjk , and thus
page 132 Equation 136: Replace by:

t = 0 and j =
 initial state
0
1 
t
=
0
and
j
= initial state
j (t) =


(t

1)a
v(t)
otherwise.
ij jk
i i
page 132 third line after Eq. 136: Change Consequently, i (t) represents to Consequently, j (t) represents
page 132 fourth line after Eq. 136: Change hidden state i to hidden state j
page 132 Algorithm 2, line 1: Delete (1),
page 132 Algorithm 2, line 1: Change t = 0 to t 0
page 132 Algorithm 2, line 1: Change (0) = 1 to j (0)
page 132 Algorithm 2, line 3: Replace entire line by j (t) bjk v(t)
1)aij

c
i=1

i (t

page 132 Algorithm 3: Somehow the line numbering became incorrect. Change
the line number to be sequential, 1, 2, . . . 6.
page 132 Algorithm 3, line 1: Change (t), t = T to j (T ), t T
page 132 Algorithm
3, old line 4, renumbered to be line 3: Replace entire line by
c
i (t) j=1 j (t + 1)aij bjk v(t + 1)
page 132 Algorithm 3, old line 7, renumbered line 5: Change P (V T ) to P (VT )
(i.e., make the V bold)

422

ERRATA IN THE TEXT

page 133 Figure 3.10, change the label on the horizontal arrow from a12 to a22 .
page 133 Figure 3.10, caption, line +5: Change was in state j (t = 2) to was in
state i (t = 2)
page 133 Figure 3.10, caption, line +6: Change is j (2) for j = 1, 2 to is i (t)
for i = 1, 2
c
page 133 Figure 3.10, caption, line -1: Change 2 (3) = b2k j=1 j (2)aj2 to
c
2 (3) = b2k i=1 i (2)ai2
page 133 line -6: Change V5 = {v3 , v1 , v3 , v2 , v0 } to V4 = {v1 , v3 , v2 , v0 }
page 133 line -44: Change is shown above, to is shown at the top of the gure
page 134 Figure in Example 3, caption, line +4: Change i (t) the probability
to j (t) the probability
page 134 Figure in Example 3, caption, line +6: Change and i (0) = 0 for i = 1.
to and j (0) = 0 for j = 1.
page 134 Figure in Example 3, caption, line +6: Change calculation of i (1). to
calculation of j (1).
page 134 Figure in Example 3, caption, line -6: Change calculation of i (1) to
calculation of j (1)
page 134 igure in Example 3, caption, line -4: Change contribution to i (1).to
contribution to j (1).
page 135 Algorithm 4: somehow the line numbering became incorrect. Change
the line numbering to be sequential, 1, 2, 3, . . ., 11. In the old line 4 (now
renumbered 3): Change k = 0, 0 = 0 to j 1
page 135 Algorithm 4 old line 5 (now renumbered 4): Change k k + 1 to
j j + 1
page 135 Algorithm 4 old line 7 (now renumbered 5): Change k (t) to j (t)
page 135 Algorithm 4, old line 8 (now renumbered 6): Change k = c to j = c
page 135 Algorithm 4, old line 11 (now renumbered 8): Change AppendT o Path j 
to Append j  to Path
page 135 line -5: Change The red line to The black line
page 135 line -4: Change value of i at each step to value of j at each step
page 137 Equation 138: Replace equation by

i (t) = 0 and t = T
0
1
i (t) = 0 and t = T
i (t) =

(t
+
1)a
b
v(t
+
1)
otherwise.
j
ij
jk
j
page 137 seventh 
line after Eq. 138: Change i (T 1) =
i (T 1) = j aij bjk v(T )j (T ).


j

aij bij v(T )i (T ). to

Chapter 3

423

page 137 fourth line before Eq. 139: Change probabilities aij and bij to probabilities aij and bjk
page 137 Equation 139: Change bij to bjk
page 138 line +3: Change whereas at step t it is to whereas the total expected
number of any transitions from i is
page 138 rst line after Eq. 140: Change bij to bjk
page 138 Equation 141: Replace equation by:

bjk =

T


t=1
v(t)=vk

T 

t=1 l

jl (t)

jl (t)

page 138 Algorithm 5, line 1: Change criterion to criterion , z 0

page 143 Problem 11, second and third lines after rst equation: Change p2 (x) by
a normal p1 (x) N (, ) to p1 (x) by a normal p2 (x) N (, )
page 143 Problem 11: Second equations: Change E2 to E1 in two places
page 143 Problem 11, last line: Change over the density p2 (x) to over the density
p1 (x)
page 147 Problem 22, line between the two equations: Change has a uniform to
has a uniform distribution
page 148 Problem 27, part (a), line after the equation: Change as given in Table
3.1. to as in Table 3.1.1.
page 149 Problem 31, line +1: Change suppose a and b are constants to suppose
a and b are positive constants
page 149 Problem 32, line +1: Change where the n coecients to at a point x,
where the n coecients
page 150 Problem 34, line +4: Change the number of operations n to the maximum size n
page 151 Problem 38, line +1: Change px (x|i ) to px (x|i )
page 151 Problem 38 (b) bottom equation on page: Change (1 2 )2 to (1
2 )2
page 152 line +1: Change and to is maximized by
page 153 Problem 43, line +4: Change and the d mean vectors. to and the c
mean vectors.
page 154 Problem 46, (top equation): Change 0

otherwise. to 

otherwise.

ERRATA IN THE TEXT

424

page 154 Problem 46, line +1 (after the equation): Change missing feature values.
to missing feature values and  is a very small positive constant that can be
neglected when normalizing the density within the above bounds.
page 154 Problem 46, part (b): Add You may make some simplifying assumptions.
page 154 Problem 47, rst equation: Change e1 x1 to ex1 /1
page 156 Computer exercise 2, after the equation: Change calculate a density to
calculate the density
page 156 Computer exercise 2. After x2 feature of category 2 . add Assume your
priors on the parameters are uniform throughout the range of the data.
page 157 Computer exercise 4, line -3: Change apply it to the x1 x2 components
to apply it to the x1 x2 components (i.e., eliminate spaces and note that the
dash is not subtraction sign, but an n-dash)
page 157 Computer exercise 6, part (a), line +2: Change in the Table above. to
in the table above.
page 159 Computer exercise 13 table, sample 4 under 1 : Change AD to ADB

Chapter 4
page 172 Figure 4.9, caption, line -1: Change where I is the d d idenity matrix.
to where I is the d-by-d identity matrix.
page 173 Algorithm 1, line 1: Change j = 0 to j 0
page 173 Algorithm 2, line 1: Change test pattern, to test pattern
page 178 line +3: Change Becuase to Because


page 179 Equation 41: Change x S to
(i.e., place the limits underneath
x S

the integral sign)

page 184 lines +2 3: Keep k i > i on the same line (i.e., do not put a line
break in this equation)
page 186 Algorithm 3, line 1: Change j = 0, D = data set, n = # prototypes to
j 0, D data set, n # prototypes
page 187 Figure 4.18, caption, line -4: Change by a factor 1/3 to by a factor
= 1/3
page 188 First margin note: Change minkowski metric to Minkowski metric (i.e., capitalize the M in Minkowski)
page 188 Second margin note: Change manhattan distance to Manhattan
distance (i.e., capitalize the M in Manhattan)

Chapter 4

425

page 188 Third margin note: Change tanimoto metric to Tanimoto metric (i.e., capitalize the T in Tanimoto)
page 189 line +5: Change the relative shift is a mere to the relative shift s is a
mere
page 192 Figure 4.23, caption, lines -4 -2: Eliminate the second-to-last sentence.
page 193 Equation 61: Replace the full equation with x (x) y (y)
page 194 line +19: Change beief about memberships to belief about memberships
page 195 line +9: Change in the title 4.8 REDUCED COULOMB to *4.8
REDUCED COULOMB
page 196 Algorithm 4, line 1: Change j = 0, n = # patterns,  = small parameter, m =
max radius to j 0, n # patterns,  small parameter, m max radius
page 197 Algorithm 5, line 1: Change j = 0, k = 0, x = test pattern, Dt = {}
to j 0, k 0, x test pattern, Dt {}
page 199 line -7: Change prior knowledge. to prior beliefs.
page 202 Problem 6c, rst line: Change increases to increase
page 203 Problem 11, part (d), line +1: Change close to an edge to close to a
face
page 203 Problem 11, part (d) line +4: Change closer to an edge to closer to a
face
page 203 Problem 11, part (d), line +5: Change even though it is easier to calculate
here to and happens to be easier to calculate
page 204 Problem 13 line +1: Change from the distributions to with priors
P (1 ) = P (2 ) = 0.5 and the distributions
page 205 Problem 21, part (a), line +1: Change As given in the text, take to
Follow the treatment in the text and take
page 206 Problem 23 part (d) line +2: Change space, then the b to space, then
in the b
page 206 Problem 26, line +4: Change and nd its nearest to and seek its
nearest
page 207 Problem 26 part (d), rst line: change Calculate the probability to
Estimate through simulation the probability
page 207 Problem 27, part (a) equation: Replace current equation with DT animoto (S1 , S2 ) =
n1 +n2 2n12
n1 +n2 n12 ,
page 207 Problem 29, rst equation: Change i to 1 in all three places
i to C(x;

page 207 Problem 29, second equation: Change C(x,

i

426

ERRATA IN THE TEXT

page 207 Problem 29 second equation: Change i to 2 in all three places

page 207 Problem 29 line -5: Change we have for the length i = 5 to we have
for the length 1 = 5
page 207 Problem 29 line -4: Change and for lightness j = 30 to and for the
lightness 2 = 30

Chapter 5
page 218 line +6: Change the problem to c 1 to the problem to c
page 218 Equation 2: Change wt xi to wit x
page 219 second line after the second (unnumbered) equation: Change is given by
(gi gj )/wi wj  to is given by (gi (x) gj (x))/wi wj 
page 220 sentence before Eq. 5, Change this in turn suggest to this in turn
suggests
pages 221222 (across the page break): Change mul-tilayer to multi-layer (i.e.,
hyphenate as multi-layer)
page 225 Algorithm 1, line 1: Change k = 0 to k 0
page 228 Algorithm 3, line 1: Change k = 0 to k 0
page 229 line +5: Change We shall begin our examination to We begin our
examination
page 229 line -3: Change Thus we shall denote to Thus we denote
page 230 Algorithm 4, line 1: Change k = 0 to k 0
page 230 fourth line before Theorem 5.1: Change correction is clearly moving
to correction is hence moving
page 230 line -2: Change From Eq. 20, to From Eq. 20 we have
page 232 rst line after Eq. 24: Change Because the squared distance to Because
this squared distance
page 233 Algorithm 5, line 1: Change k = 0 to k 0
page 233 Algorithm 6, line 1: Change k = 0 to k 0
page 234 line +22 (counting from the end of the Algorithm): Change that it will
have little eect at all. to that it will have little if any eect.
page 235 lines +3 4 (counting from the end of the Algorithm): Change and this
means the gap, determined by these two vectors, can never increase in size
for separable data. to and this means that for separable data the gap,
determined by these two vectors, can never increase in size.
page 235 second and third lines after the section title 5.6 RELAXATION PROCEDURES: Change in so-called relaxation procedures to include to in
so-called relaxation procedures, to include

Chapter 5

427

page 235 rst and second lines after Eq. 32: Change misclassied by a, as do Jp ,
Jq focus attention to misclassied by a. Both Jp and Jq focus attention
page 235 second line after Eq. 32: Change Its chief dierence is that its gradient
to The chief dierence is that the gradient of Jq is
page 236 Algorithm 8, line 1: Change k = 0 to k 0
page 236 Algorithm 8, line 4: Change j = 0 to j 0
page 236 Algorithm 9, line 1: Change k = 0 to k 0
page 238 line -5: Change procedures, because to procedures because
page 242 line -2: Change We begin by writing Eq. 47 to We begin by writing
Eq. 45
page 246 lines +1 2: Dont split the equation by the line break
page 246 rst (unnumbered) equation, second line: Change (Yak b). to (Ya(k)
b).
page 246 margin note: Change lms rule to LMS rule (i.e., capitalize LMS)
page 246 Equation 61, second line: Change (bk a(k)t yk ) to (b(k) at (k)yk )
page 246 Algorithm 10, line 1: Change k = 0 to k 0
page 248 second line after Eq. 66: Change obtain the MSE optimal to obtain
the MSE-optimal
page 250 Equation 79: Co-align vertically the > in the top equation with the =
in the bottom equation
page 250 Equation 79: Change a(k) to b(k)
page 251 Algorithm 11, line 5: Change a to b
page 252 top equation: Change = 0 = to = 0 = (i.e., de-bold the 0)
page 253 line -7: Change requires that e+ (k) = 0 for to requires that e+ (k) = 0
for
page 254 line after Eq. 90: Change constant, positive-denite to constant, symmetric, positive-denite
page 255 First (unnumbered) equation on page: Change the rst term in parentheses
on the right-hand side from 2 YRYt YRY to 2 YRYt YRYt
page 255 Eq. 92: Last term on the right-hand-side, change 2 RYt R to 2 RYt YR
page 257 Figure 5.18, caption, line +2: Change form Au to form Au =
page 262 fourth line after Eq. 105: Change with the largest margin to with the
largest margin (i.e., italicize largest)
page 263 fourth line after Eq. 107: Change equation in Chapter 9, to topic in
Chapter 9,

428

ERRATA IN THE TEXT

page 266 Equation 114: Change ati (k)yk aj (k)t yk . to ati (k)yk atj (k)yk .
page 270 twelfth line under BIBLIOGRAPHICAL AND HISTORICAL REMARKS: Change error-free case [7] and [11] and to error-free case [7,11]
and
page 270 line -13: Change support vector machines to Support Vector Machines
page 270 line -8: Change support vector machines to Support Vector Machines
page 271 Problem 2, line +3: Change if 0 1. to for 0 1.
page 272 Problem 8, line +2: Change if at yi 0 to if at yi b
page 274 Problem 22, second term on the right-hand side change (at y (12
22 ))2 to (at y + (12 22 ))2
page 275 Problem 27, last line: Change by Eq. 85. to by Eq. 95.
page 277 lines +2 3: Change satises zk at yk = 0 to satises zk at yk = 1
page 277 Problem 38, line -2: Change procedures Perceptron to procedures.
Generalize the Perceptron
page 278 Problem 1, part (a): change data in in order to data in order
page 278 Computer exercise 2, part (a): Change Starting with a = 0, to Starting
with a = 0, (i.e., make bold the 0)
page 278 First heading after the table: Change Section 5.4 to Section 5.5
page 278 Computer Exercise 1: Change (Algorithm 1) and Newtons algorithm
(Algorithm 2) applied to (Algorithm 1) and the Perceptron criterion (Eq. 16)
page 278 Second heading after the table: Delete Section 5.5
page 279 line +1: Change length is greater than the pocket to length is greater
than with the pocket
page 279 Computer exercise 4, part (a), line +2: Change and 1 = 0, to and
1 = 0, (i.e., make bold the 0)

Chapter 6
page 286 Equation 5: Change zk = f (netk ). to zk = f (netk ),
page 287 line +2: Change all identical. to all the same.
page 288 line -3: Change depend on the to depends on the
page 291 Two lines before Eq. 3: Change hidden-to-output weights, wij to hiddento-output weights, wkj
page 292 Equation 19, rst line (inside brackets): Change 1/2 to 12 (i.e., typeset
as a full fraction)
page 292 After Eq. 19, line +3: Change activiation to activation

Chapter 6

429

page 293 Figure 6.5: Change wij to wji

page 294 Algorithm 2, line 3: Change wkj to wkj 0
page 295 line -5: Change In addition to the use of the training set, here are to
In addition to the use of the training set, there are
page 299 line -1: Change and can be linearly separable to and are linearly separable
page 305 line 10: Change ratio of such priors. to ratio of such priors, though this
need not ensure minimal error.
page 306 sixth and seventh line after Eq. 33: Change in a sum squared error sense
to in a sum-squared-error sense
page 307 lines +2 3: Change been found useful to found to be useful
page 307 line +6: Change as continuity of f and its derivative to as continuity
of f () and its derivative
page 308 line +10: Change that is, to or is an odd function, that is,
page 308 line +19: Change values that ensure f  (0)  1 to values that ensure
f  (0)  0.5
page 315 rst line after Eq. 38: Change reducing the criterion to reducing the
error
page 316 Figure 6.19 caption, line -1: Change trained network. to trained network (red).
page 318 line +8: Change to compute is nonnegative to to compute, is nonnegative
page 318 margin note: Change minkowski error to Minkowski error (i.e.,
captialize M in Minkowski)
page 320 line between Eqs. 50 and 51: Change The optimimum change to Therefore, the optimum change
page 319 Equation 48: Change last entry from f  (net)ynH xd to f  (net)f  (netnH xd
page 322 Equation 56: Replace the current equation by
m =

J t (w(m))J(w(m))
1))J(w(m 1))

J t (w(m

(56)

page 322 Equation 57: Replace the current equation by

m =

J t (w(m))[J(w(m)) J(w(m 1))]

J t (w(m 1))J(w(m 1))

(57)

page 323 Fourth (unnumbered) equation on the page: Replace the left portion by
1 =

J t (w(1))J(w(1))
=
J t (w(0))J(w(0))

430

ERRATA IN THE TEXT

page 323 Caption to bottom gure, line +2: Change shown in the contour plot,
to shown in the density plot,
page 325 last line of the section Special Bases: Add This is very closely related
to model-dependent maximum-likelihood techniques we saw in Chapter 3.
page 326 rst line after Eq. 63: Change of the lter in analogy to of the lter,
in analogy
page 326 line -5: Add a red margin note time delay neural network
page 330 three lines above Eq. 67: Change write the error as the sum to write
the new error as the sum
page 332 Figure 6.28, caption, line +1: Change a function of weights, J(w) to a
function of weights, J(w),
page 337 Problem 8, pagt (b) line +2: Change if the sign if ipped to if the sign
is ipped
page 337 Problem 14, part (c), line +2: Change the 2 2 identity to the 2-by-2
identity
page 339 Problem 22, line +4: Add Are the discriminant functions independent?
page 341 Problem 31, lines +1 2: Change for a sum squared error criterion to
for a sum-square-error criterion
page 344 Computer exercise 2, line +2: Change backpropagation to (Algorithm
1) to backpropagation (Algorithm 1)
page 345 Computer exercise 7, line +2: Change on a random problem. to on a
two-dimensional two-category problem with 2k patterns chosen randomly from
the unit square. Estimate k such that the expected error is 25% . Discuss your
results.
page 346 Computer exercise 10, part (c), line +1: Change Use your network to
Use your trained network
page 347 Column 2, entry for [14], line +5: Change volume 3. Morgan Kaufmann
to volume 3, pages 853859. Morgan Kaufmann
page 348 column 2, entry for [43], line +4: Change 2000. to 2001.

Chapter 7
page 351 fourteenth line after Eq. 1: Change of the magnets with the most stable
conguration to of the magnets that is the most stable
page 351 footnote, line -1: Change in a range of problem domains. to in many
problem domains.
page 352 Figure 7.1, caption, line +7: Change While our convention to While
for neural nets our convention

Chapter 7

431

page 352 Figure 7.1, caption, line +8: Change Boltzmann networks is to Boltzmann networks here is
page 352 Caption to Figure 7.1, line -1: Change 0 210 to 0 < 210
page 353 Figure 7.2, caption, line +3: Change or temperature T to avoid to
or temperature T , to avoid
page 357 Figure 7.4, caption, line +3: Change values eE /T . to values eE /T .
page 360 rst line in Section 7.3: Change will use modify the to will use the
page 360 second line in Section 7.3: Change to specically identify to and specifically identify
page 360 second line in Section 7.3: Change and other units as outputs to and
others as outputs
page 361 fourth line after Eq. 6: Change possible hidden states. to possible hidden states consistent with .
page 364 at the end of the body of the text: insert One benet of such stochastic learning is that if the nal error seems to be unacceptably high, we can
merely increase the temperature and anneal we do not need to re-initialize
the weights and re-start the full anneal.
page 365 seventh line after the subsection Pattern Completion: Change components of a partial pattern to components of the partial pattern
page 367 line +1: Change Recall, at the end to As mentioned, at the end
page 373 fourteenth line in Section 7.5: Change repeated for subsequent to repeated for the subsequent
page 373 fth line above the subsection Genetic Algorithms: Change In both
cases, a key to In both cases a key
page 374 line +2: Change used in the algorithm. Below to used in the algorithm;
below
page 374 line +3: Change Pco and Pmut , respectively, but rst we present the
general algorithm: to Pco and Pmut .
page 379 third line in the subsection Representation: Change Here the syntactic
to Whereas the syntactic
page 381 third line under BIBLIOGRAPHICAL AND HISTORICAL REMARKS: Change branch-and-bound, A to branch-and-bound and A
page 382 line +28: Change been fabricated as described to been fabricated, as
described
page 383 Problem 3, part (b), line +1: Change The gure shows to That gure
shows
page 384 Problem 7, part (c), line +1: Change magnets, total to magnets, the
total

432

ERRATA IN THE TEXT

page 384 Problem 8, part (b), equation: Change = P (s = +1)(+1) + P (s =

1)(1). to = Pr[s = +1](+1) + Pr[s = 1](1).
page 385 Problem 12, line +1: Change Train a six-unit Hopeld network with
the following three patterns using the to Determine the weights in a six-unit
Hopeld network trained with the following three patterns. Use the
page 385 Problem 15 line +1: Change not be in a set of to not be in a subset
of
page 387 Problem 24: Change crossover operator to crossover operator, and the
multiplication operator, *, and the addition operator, +, can take two or more
operands.
page 393 column 1, entry for [54], line +3: Change Evoultions- to Evolutions-
page 393 column 2, entry for [62], line +2: Change neurobilogical system. to
neurobiological systems.

Chapter 8
page 403 line above Section 8.3.4 Pruning: Change splitting is stopped. to
splitting should be stopped.
page 405 Table at top, x2 entry in fth row under 1 , change .48 to .44
page 405 caption to gure, line -2: Change marked were instead slightly lower
(marked ), to marked were instead slightly lower (marked ), (i.e., change
the color of the special symbols to red)
page 414 line +16: Change "
/"
/"
/ GACTG to "
/"
/"
/ GACTG (i.e., eliminate space)
page 416 Algorithm 2, line 2: Change F(x) to F
page 416 Algorithm 2, line 3: Change G(x) to G
page 416 Algorithm 2, line 11: Change G(0) to 1
page 416 line -1: Change F(x) to F
page 417 lines -9 -4, Replace last full paragraph by Consider target string x.
Each location j (for j < m) denes a sux of x, that is, x[j + 1, . . . , m]. The
good-sux function G(j) returns the starting location of the right-most instance
of another occurrence of that sux (if one exists). In the example in Fig. 8.8,
x = estimates and thus j = 8 denes the sux s. The right-most occurrence
of another s is 2; therefore G(8) = 2. Similarly j = 7 denes the sux es. The
right-most occurrence of another es is 2; therefore G(7) = 1. No other sux
appears repeatedly within x, and thus G is undened for j < 7.
page 418 fth line before Algorithm 3: Change consider interchanges. to consider the interchange operation.
page 418 fourth line before Algorithm 3: Change be an m n matrix to be an
m-by-n matrix

Chapter 9

433

page 422 line -3: Change specify how to transform to species how to transform
page 426 line +1: Change The grammar takes digit6 and to This grammar takes
digits6 and
page 428 line -1: Change subsequent characters. to subsequent characters, or
instead starting and the last (right) character in the sentence.
page 429 caption to Figure 8.16, add after the last line: Such nite-state machines
are sometimes favored because of their clear interpretation and learning methods
based on addition of nodes and links. In Section 8.7, though, we shall see general
methods for grammatical learning that apply to a broader range of grammatical
models.
page 438 Problem 5, line +2: Change Eqs. 1 and 5. to Eqs. 1 and 5 for the case
of an arbitrary number of categories.
page 438 Problem 6 after the rst set of equations: Replace the i () equation by
i () = i(P a (1 ) + (1 )P b (1 ), ..., P a (c ) + (1 )P b (c ))
page 438 Problem 6 last line before part (a): Replace line by then we have i
ia + (1 )ib .
page 445 Problem 36, part (d), line +1: Change Give a derivation to Attempt
a derivation
page 445 Problem 40, part (d), line +2: Change either grammar as to either
grammar, or can be parsed in both grammars, as
page 446 Table, sample 12: Change D to E
page 447 Computer exercise 3, part b): Change {C, D, J, L, M} to {C, E, J, L,
M}

Chapter 9
page 455 line -9 -10: Change a zero-one loss function, or, more generally, the
cost for a general loss function L(, ) to a zero-one or other loss function.
page 460 Table for rank r = 3, third row: Change x1 OR x3 OR x3 to x1 OR x3 OR x4
page 462 line +12: Change upon a specication method L, by upon a specication method,
page 462 line +13: Change transmitted as y, denoted L(y) = x. to transmitted
as y and decoded given some xed method L, denoted L(y) = x.
page 462 line +15 16: Change denoted min L(y) = x; this minimal...[[to end of
paragraph]] to denoted

min |y|.

|y|

y:L(y)=x

page 462 line +17 18: Change by analogy to entropy, where instead of a specication method L we consider to by analogy to communication, where instead
of a xed decoding method L, we consider

ERRATA IN THE TEXT

434

page 462 line +25: Change and so on. A universal description would to and so
on. Such a description would
page 462 line +26: Change dierent binary strings. to dierent binary strings
x1 and x2 .
page 462 third and second line above Eq. 7: Change the shortest program y (where
the length to the shortest program string y (where ys length
page 462 Equation 7: Replace the entire equation by:
K(x) =

min

y:U (y)=x

|y|,

page 463 line -10: Change No Free Lunch Theorems. to No Free Lunch Theorem.
1
1
page 472 Equation 23: Change (n1)
to n(n1)

page 472 Equation 24: Change

j=i

to

n


(i.e., place the upper limit n over the

j=i

summation sign)
page 472 line -4: Change jackknife estimate of the variance to variance of the
jackknife estimate
page 479 line +1: Change in line 4 to in line 5
page 485 line -10: Change good estimates, because to good estimates because
page 488 Caption to Fig. 9.13, line +6: Change approximated as a k-dimensional
to approximated as a p-dimensional
page 488 line +10 (i.e., second line of the second paragraph of text): Change is
k-dimensional and the to is p-dimensional and the
page 496 Equation 54: Change P (r|x, 0 ) to P (r|x, 00 )
page 497 Equation 58: Change r to r in two places
page 501 line -15: Change and learning algorithm was rst described to and
learning algorithm were rst described
page 502 Problem 6, last line: Change this way to this sense
page 504 Problem 20, line -2: Change denoted p(g(x; D)) is a to denoted
p(g(x; D)), is a
page 508 Problem 45, line +1: Change mixture of experts classier to mixtureof-experts classier
page 512 line -3: Add Now what is the training error measured using A and B ?

10.0. CHAPTER 10

435

Chapter 10

%
page 524 The second to last equation: Increase the size of the nal bracket, to
match its mate
page 525 line following Eq. 23: Change results in Eq. 12 to results into Eq. 12
page 529 line +5: Change P (wj ) to P (j )
page 529 Algorithm 2: Change (Fuzzy k-Means Clustering) to (Fuzzy kMeans Clustering)
page 534 line +2: Change overlap, thus to overlap; thus
page 535 First equation: move the second, third, and fourth lines so as to co-align
vertically the corresponding terms
page 536 line +19: Change classication analog of Chapter 3 to classication
case in Chapter 3
page 537 line -20: Change distributed, these statistics to distributed these statistics
page 571 line +1 +2: Change mi-crophones to micro-phones (i.e., re-hyphenate)
page 578 Figure 10.31, caption, line +2: Change space that leads maximally to
space that maximally
page 579 Figure 10.32, caption, line -1: Change to this center region to to this
central region
page 580 line +15: Change cluster centers being used to to cluster centers being
used to (i.e., italicize cluster centers)
page 580 line +16: Change with combined features being to with combined features being (i.e., italicize combined features)
page 582 four lines above BIBLIOGRAPHICAL and HISTORICAL REMARKS:
Change between points that, too, seeks to preserve neighborhoods to between
points that preserves neighborhoods
page 583 lines +9 10: Change The classicatory foundations of biology, cladistics
(from the Greek klados, branch) provide useful to Cladistics, the classifactory
foundation of biology (from the Greek klados, branch), provides useful
page 583 line +18: Change analysis, and explained the very close to analysis,
and in reference [36] explained the very close
page 583 line +19: Change information maximization in reference [36]. to information maximization.
page 584 Problem 2, equation: Change (1|x1 |)/(2wi ) to (wi |xi |)/wi2
page 584 Problem 4 a, right-hand side: Change xi to xj
page 585 Problem 6, line +1: Change Consider a c component to Consider a
c-component

436

ERRATA IN THE TEXT

page 587 Problem 13, line +3: Change that for any observed x, all but one to
that for any observed x all but one
page 589 Problem 24 b, Change the trace criterion to the determinant criterion
page 591 Problem 41, line +1: Change null hypothesis in associated to null
hypothesis associated
page 592 Problem 44, part (c), line +3: Change (e)t (e)t e = 0 to (e)t e
(e)t e = 0
page 592 Problem 46, line +1: Change principal componet analysis to principal
component analysis
page 593 Problem 48, line +3: Change linearity is given by to linearity is the
one given by
page 596 Problem 11, lines +3 4: Change to the date in the table above using
the distance measure indicated to to the data in the table above using the
distance measures indicated
page 597 Problem 13, line +1: Change a basic competitive learning to a basic
Competitive Learning
page 597 Problem 13, lines +4 5: Change hy-persphere to hyper-sphere (i.e.,
re-hyphenate hypersphere)

Appendix
page 608 Section A.2.5 line +5: Change In this case the absolute value of the
determinant to In this case the determinant
page 609 rst line after Eq. 26: Change the i, j cofactor or to the i, j cofactor
or (i.e., eliminate the space in i, j)
page 609 second line after Eq. 26: Change is the (d 1) (d 1) matrix to is
the (d 1)-by-(d 1) matrix
page 609 fourth line after Eq. 26: Change whose i, j entry is the j, i cofactor
to whose i, j entry is the j, i cofactor (i.e., eliminate the space in i, j and
j, i)
page 609 line -2: Add The inverse of the product of two square matrices obeys
[MN]1 = N1 M1 , as can be veried by multiplying on the right or the left
by MN.
11 /n
page 615 line -12 (i.e., just before Section A.4.7): Change and (n01n+n
is ap11 )/n
proximately to and (n01 + n11 )/n is approximately

page 624 Figure A.3,

caption,
line +2: Change between 2u and 2u, that is
to between 2u and 2u; that is

page 629 line +9: Change from a distribution to from a standarized Gaussian
distribution

Index

437

page 631 line +9: Change equally likely is to equally likely, is

page 631
and H = log2 23 = 3 to outcome and H =
7line1 +12: 3Change outcome
3
i=0 23 log2 2 = log2 2 = 3
page 631 Equation 118: Change ln p(x) to lnp(x) (i.e., reduce the space between
ln and p(x))
page 631 rst line after Eq. 119: Change For this Dirac function to For this Dirac
density
page 632 Red margin note: Change Kullback-leibler distanceto KullbackLeibler distance (i.e., capitalize the L in Leibler)
page 632 line -2: Change is always larger than to is never smaller than
page 634 line +3: Change f (x) c0 g(x) for all to f (x) cg(x) for all
page 634 line +7: Change proper choice of c0 and x0 to proper choice of c and
x0

Index
page 644 column 1, line -14: Insert Gini impurity, 399, 401
page 653 column 1, line -3: Change multi-variate to multivariate

ERRATA IN THE TEXT

438

Third and fourth printings

Chapter 2
page 53 Bottom equation (for w0 ): Change = 1.2 to = -1.75
page 58 In the gure, the entry for P (c3 |x1 ) should be changed from 0.1 to 0.2.
page 60 Equation 99: Change X to x in two places
page 73 Problem 34, line +67: Change assume the distributions to assume
P (1 ) = P (2 ) = 0.5 and the distributions
page 73 Problem 34, part c), line +4: Change Bayes error is 0.5. to Bayes error
is 0.25.

Chapter 3
page 99 Caption to rst gure, change starts our as a at to starts out as a at
page 133 Figure 3.10, change the label on the horizontal arrow from a12 to a22
page 143 Problem 11, second and third lines after rst equation: Change p2 (x) by
a normal p1 (x) N (, ) to p1 (x) by a normal p2 (x) N (, )
page 143 Problem 11: Second equations: Change E2 to E1 in two places
page 143 Problem 11, last line: Change over the density p2 (x) to over the density
p1 (x)
page 149 Problem 31, line 1: Change suppose a and b are positive constants and
to suppose a and b are constants greater than 1 and
page 151 Problem 38 (b) bottom equation on page: Change (1 2 )2 to (1
2 )2
page 156 Computer exercise 2, after the equation: Change calculate a density to
calculate the density
page 156 Computer exercise 2. After x2 feature of category 2 . add Assume your
priors on the parameters are uniform throughout the range of the data.
page 159 Computer exercise 13 table, sample 4 under 1 : Change AD to ADB

Chapter 4
page 178 line +3: Change Becuase to Because
page 202 Problem 6 part (c), rst line: Change increases to increase
page 207 Problem 26 part (d), rst line: change Calculate the probability to
Estimate through simulation the probability
page 207 Problem 26 delete part (e)

10.0. CHAPTER 5

439

Chapter 5
page 220 sentence before Eq. 5, Change this in turn suggest to this in turn
suggests
page 250 Equation 79: Change a(k) to b(k)
page 251 Algorithm 11, line 5: Change a to b
page 254 line after Eq. 90: Change constant, positive-denite to constant, symmetric, positive-denite
page 255 First (unnumbered) equation on page: Change the rst term in parentheses
on the right-hand side from 2 YRYt YRY to 2 YRYt YRYt
page 255 Eq. 92: Last term on the right-hand-side, change 2 RYt R to 2 RYt YR
page 274 Problem 22, second term on the right-hand side change (at y (12
22 ))2 to (at y + (12 22 ))2
page 275 Problem 27, last line: Change by Eq. 85. to by Eq. 95.
page 278 Problem 1, part (a), line +1: change data in in order to data in order
page 278 Problem 1, part (a), line +2: change use (k) = 0.1. to use (k) =
0.01.
page 278 First heading after the table: Change Section 5.4 to Section 5.5
page 278 Computer Exercise 1: Change (Algorithm 1) and Newtons algorithm
(Algorithm 2) applied to (Algorithm 1) and the Perceptron criterion (Eq. 16)
applied
page 278 Second heading after the table: Delete Section 5.5
page 279 line +1: Change length is greater than the pocket to length is greater
than with the pocket

Chapter 6
page 291 Two lines before Eq. 3: Change hidden-to-output weights, wij to hiddento-output weights, wkj
page 292 After Eq. 19, line +3: Change activiation to activation
page 293 Figure 6.5: Change wij to wji
page 294 Algorithm 2, line 3: Change wkj to wkj 0
page 295 Second paragraph in Section 6.3.3: Change indpendently selected to
independently selected
page 302 line +34: Change weights merely leads to weights merely lead
page 305 line 10: Change ratio of such priors. to ratio of such priors, though this
need not ensure minimal error.

440

ERRATA IN THE TEXT

page 314 line +6: Change weight changes are response to weight changes are a
response
page 330 line -2 -1: Change While it is natural to It is natural
page 337 Problem 8, pagt (b) line +2: Change if the sign if ipped to if the sign
is ipped
page 346 Problem 10, part (c), line +3: Change x5 = (0, 0, 0)t to x5 = (0, 0, 1)t

Chapter 7
page 352 Caption to Figure 7.1, line -1: Change 0 210 to 0 < 210

Chapter 8
page 405 Table at top, x2 entry in fth row under 1 , change .48 to .44
page 409 line -7: Change queries involves to queries involve
page 416 Algorithm 2, line 2: Change F(x) to F
page 416 Algorithm 2, line 3: Change G(x) to G
page 416 Algorithm 2, line 11: Change G(0) to 1
page 416 line -1: Change F(x) to F
page 417 lines -9 -4, Replace last full paragraph by Consider target string x.
Each location j (for j < m) denes a sux of x, that is, x[j + 1, . . . , m]. The
good-sux function G(j) returns the starting location of the right-most instance
of another occurrence of that sux (if one exists). In the example in Fig. 8.8,
x = estimates and thus j = 8 denes the sux s. The right-most occurrence
of another s is 2; therefore G(8) = 2. Similarly j = 7 denes the sux es. The
right-most occurrence of another es is 2; therefore G(7) = 1. No other sux
appears repeatedly within x, and thus G is undened for j < 7.
page 438 Problem 5, line +2: Change Eqs. 1 and 5. to Eqs. 1 and 5 for the case
of an arbitrary number of categories.
page 438 Problem 6 after the rst set of equations: Replace the i () equation by
i () = i(P a (1 ) + (1 )P b (1 ), ..., P a (c ) + (1 )P b (c ))
page 438 Problem 6 last line before part (a): Replace line by then we have i
ia + (1 )ib .
page 446 Table, sample 12: Change D to E
page 447 Computer exercise 3, part b): Change {C, D, J, L, M} to {C, E, J, L,
M}

10.0. CHAPTER 9

441

Chapter 9
page 460 Table for rank r = 3, third row: Change x1 OR x3 OR x3 to x1 OR x3 OR x4
page 468 Five lines after Eq. 12: Change regardless the amount to regardless of
the amount
page 474 Caption to Figure: Change and jackknife estimate to and whose jackknife estimate
page 497 Equation 58: Change r to r in two places
page 505 Take the heading Section 9.4 and move it to the top of the page, i.e.,
between Problems 20 and 21.
page 508 Problem 45, line +2: Change N (, ) to N (r , r )

Chapter 10
page 526 line -1 Change In the absense to In the absence
page 541 line +9: Change this is a symmetric functions to this is a symmetric
function
page 549 Eq. 76: Change m to mi on the righthand side.
page 573 Eq.107: Put 
the lower limit underneath the summation sign, that is,
change i<j to
i<j

page 573 Eq.109: Put 

the lower limit underneath the summation sign, that is,
change i<j to
i<j

page 574 First equation:

Put

 the lower limit underneath the summation sign, that
is, change i<j to
i<j

page 574 Third 

equation: Put
 the lower limit underneath the summation sign, that
is, change i<j to
i<j

page 576 Eq.112: Put 

the lower limit underneath the summation sign, that is,
change i<j to
i<j

page 558 Last full paragraph, line +1: Change This result agrees with out statement to This result agrees with our statement
page 584 Problem 4 part (a), right-hand side: Change xi to xj
i to
in two places only on the rightpage 585 Problem 4 part (b): Change
hand side of the equation.
page 589 Problem 24 part (b), Change the trace criterion to the determinant
criterion

442

ERRATA IN THE TEXT

Appendix
page 608 Section A.2.5 line +5: Change In this case the absolute value of the
determinant to In this case the determinant

10.0. CHAPTER 2

443

Fifth printing
Chapter 2
page 73 Problem 34, line +67: Change assume the distributions to assume
P (1 ) = P (2 ) = 0.5 and the distributions
page 73 Problem 34, part c), line +4: Change Bayes error is 0.5. to Bayes error
is 0.25.

Chapter 3
page 143 Problem 11, second and third lines after rst equation: Change p2 (x) by
a normal p1 (x) N (, ) to p1 (x) by a normal p2 (x) N (, )
page 143 Problem 11: Second equations: Change E2 to E1 in two places
page 143 Problem 11, last line: Change over the density p2 (x) to over the density
p1 (x)
page 151 Problem 38 (b) bottom equation on page: Change (1 2 )2 to (1
2 )2
page 159 Computer exercise 13 table, sample 4 under 1 : Change AD to ADB

Chapter 5
page 250 Equation 79: Change a(k) to b(k)
page 251 Algorithm 11, line 5: Change a to b
page 275 Problem 27, last line: Change by Eq. 85. to by Eq. 95.

Chapter 6
page 294 Algorithm 2, line 3: Change wkj to wkj 0
page 305 line 10: Change ratio of such priors. to ratio of such priors, though this
need not ensure minimal error.
page 337 Problem 8, pagt (b) line +2: Change if the sign if ipped to if the sign
is ipped

Chapter 8
page 438 Problem 6 after the rst set of equations: Replace the i () equation by
i () = i(P a (1 ) + (1 )P b (1 ), ..., P a (c ) + (1 )P b (c ))
page 438 Problem 6 last line before part (a): Replace line by then we have i
ia + (1 )ib .

444

ERRATA IN THE TEXT

page 446 Table, sample 12: Change D to E

page 447 Computer exercise 3, part b): Change {C, D, J, L, M} to {C, E, J, L,
M}

Chapter 9
page 460 Table for rank r = 3, third row: Change x1 OR x3 OR x3 to x1 OR x3 OR x4
page 508 Problem 45, line +2: Change N (, ) to N (r , r )

Chapter 10
page 553 line +1314: Change each of which is an O(d2 ) calculation to each of
which is an O(d) calculation
page 553 line +1718: Change the complexity is O(n(n 1)(d2 + 1)) = O(n2 d2 )
to the complexity is O(n(n 1)(d + 1)) = O(n2 d)
page 553 line +21: Change complexity is thus O(cn2 d2 ) to complexity is thus
O(cn2 d)