Pattern Classification (2nd ed.

Richard O. Duda, Peter E. Hart and David G. Stork

September 3, 1997

NOT FOR GENERAL DISTRIBUTION; for use only by students of designated

faculty. This is a pre-publication print of material to appear in Duda, Hart and
Stork: Pattern Classification and Scene Analysis: Part I Pattern
Classification, to be published in 1998 by John Wiley & Sons, Inc. This is a
preliminary version and may contain errors; comments and suggestions are heartily
encouraged.

Contact: Dr. David G. Stork

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

°1997
c R. O. Duda, P. E. Hart and D. G. Stork
All rights reserved.
2
Contents

A.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
A.2 Linear algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
A.2.1 Notation and preliminaries . . . . . . . . . . . . . . . . . . . . 8
A.2.2 Outer product . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
A.2.3 Derivatives of matrices . . . . . . . . . . . . . . . . . . . . . . . 10
A.2.4 Determinant and trace . . . . . . . . . . . . . . . . . . . . . . . 11
A.2.5 Eigenvectors and eigenvalues . . . . . . . . . . . . . . . . . . . 12
A.2.6 Matrix inversion . . . . . . . . . . . . . . . . . . . . . . . . . . 12
A.3 Lagrange optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
A.4 Probability Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
A.4.1 Discrete random variables . . . . . . . . . . . . . . . . . . . . . 13
A.4.2 Expected values . . . . . . . . . . . . . . . . . . . . . . . . . . 14
A.4.3 Pairs of discrete random variables . . . . . . . . . . . . . . . . 15
A.4.4 Statistical independence . . . . . . . . . . . . . . . . . . . . . . 16
A.4.5 Expected values of functions of two variables . . . . . . . . . . 16
A.4.6 Conditional probability . . . . . . . . . . . . . . . . . . . . . . 18
A.4.7 The Law of Total Probability and Bayes’ rule . . . . . . . . . . 18
A.4.8 Vector random variables . . . . . . . . . . . . . . . . . . . . . . 19
A.4.9 Expectations, mean vectors and covariance matrices . . . . . . 20
A.4.10 Continuous random variables . . . . . . . . . . . . . . . . . . . 21
A.4.11 Distributions of sums of independent random variables . . . . . 23
A.4.12 Univariate normal density . . . . . . . . . . . . . . . . . . . . . 24
A.5 Gaussian derivatives and integrals . . . . . . . . . . . . . . . . . . . . 25
A.5.1 Multivariate normal densities . . . . . . . . . . . . . . . . . . . 27
A.5.2 Bivariate normal densities . . . . . . . . . . . . . . . . . . . . . 28
A.6 Information theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
A.6.1 Entropy and information . . . . . . . . . . . . . . . . . . . . . 31
A.6.2 Relative entropy . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A.6.3 Mutual information . . . . . . . . . . . . . . . . . . . . . . . . 32
A.7 Computational complexity . . . . . . . . . . . . . . . . . . . . . . . . . 33
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

3
4 CONTENTS
Mathematical foundations

ur goal here is to present the basic results and definitions from linear algebra,
O probability theory, information theory and computational complexity that serve
as the mathematical foundations for the pattern recognition techniques discussed
throughout this book. We will try to give intuition whenever appropriate, but we
do not attempt to prove these results; systematic expositions can be found in the
references.

A.1 Notation
Here are the terms and notation used throughout the book. In addition, there are
numerous specialized variables and functions whose definitions are usage should be
clear from the text.

variables, symbols and operations

' approximately equal to
≈ approximately equal to (in an expansion)
≡ equivalent to (or defined to be)
∝ proportional to
∞ infinity
x→a x approaches a
t←t+1 in an algorithm: assign to variable t the new value t + 1
lim f (x) the value of f (x) in the limit x approaching a
x→a
arg max f (x) the value of x that leads to the maximum value of f (x)
x
arg min f (x) the value of x that leads to the minimum value of f (x)
x
ln(x) logarithm base e, or natural logarithm of x
log(x) logarithm base 10 of x
log2 (x) logarithm base 2 of x
exp[x] or ex exponential of x
∂f (x)/∂x partial derivative
Rb
f (x)dx the integral of f (x) between a and b. If no limits are written, the
a full space is assumed
Q.E.D., quod erat demonstrandum (“which was to be proved ”) —
used to signal the end of a proof

5
6 CONTENTS

mathematical operations
E[f (x)] the expected value of function f (x)
Ey [f (x, y)] the expected value of function over several variables, f (x, y), taken
over a subset y of them
Varf [·] Ef [(x − E[x])2 ]
<x> expected value of random variable
Pn
ai the sum from i = 1 to n: a1 + a2 + ... + an
i=1
Q
n
ai the product from i = 1 to n: a1 × a2 × ... × an
i=1

vectors and matrices

Rd d-dimensional Euclidean space
x, Σ boldface for (column) vectors and matrices
I identity matrix, square matrix having 1s on the diagonal and 0
everywhere else
diag(a1 , a2 , ..., ad ) matrix whose diagonal elements are a1 , a2 , ..., ad , and off-diagonal
elements zero
xt the transpose of vector x
kxk the Euclidean norm of vector x.
Σ covariance matrix
tr[A] the transpose of A , with ij entry changed to ji
A−1 the inverse of matrix A
A† pseudoinverse of matrix A
|A| or Det[A] determinant of A
λ eigenvalue
e eigenvector
ei unit vector in the i direction in Euclidean space
A.1. NOTATION 7

probability and distributions

ω state of nature
P (·) probability
p(·) probability density
P (a, b) the joint probability , i.e., of having both a and b
p(a, b) the joint probability density, i.e., of having both a and b
p(x|θ) the conditional probability density of x given that θ
F (x; θ) function of x, with implied (nonexplicit) dependence upon θ
w weight
λ(·, ·) loss function
 d 
dx1
 d 
 dx2 
∆= ..  gradient operator in Rd
 . 
d
dxdd 
dθ1
 d 
 dθ2 
∆θ =  ..  gradient operator in θ coordinates
 . 
d
dθd
θ̂ maximum likelihood value of θ
∼ “has the distribution” e.g., p(x) ∼ N (µ, σ 2 ) means that the density
of x is normal, with mean µ and variance σ 2
N (µ, σ 2 ) normal or Gaussian distribution with mean µ and variance σ 2
N (µ, Σ) multidimensional normal or Gaussian distribution with mean µ
and covariance matrix Σ
U (xl , xu ) a one-dimensional uniform distribution between xl and xu .
U (xl , xu ) a d-dimensional uniform density, having the smallest axes-aligned
bounding box containing both xl and xu
T (µ, δ) triangle distribution, having center µ and full half-width δ
δ(x) Dirac delta function
Γ(·) Gamma function
n!
¡a¢ n factorial = n × (n − 1) × (n − 2) × ... × 1
a!
b = b!(a−b)! binomial coefficient, a choose b
O(h(x)) big oh order of h(x)
Θ(h(x)) big theta order of h(x)
x̄ mean or average value of x
lim f (x) the value of f (x) in the limit x approaches y
n→y
sup f (x) the supremum value of f (x)
x
 8 CONTENTS

sets
A, B, C, D, ... “Calligraphic” font generally denotes sets or lists, e.g., data set
D = {x1 , ..., xn }
x∈D x is an element of set D
x∈/D x is not an element of set D
A∪B union of two sets, i.e., the set containing all elements of A and B
|D| the cardinality of set D, i.e., the number of (possibly non-distinct)
elements in it
max[D] the x value in set D that is maximum
x

A.2 Linear algebra

A.2.1 Notation and preliminaries
A d-dimensional (column) vector x and its (row) transpose xt can be written as
 
x1
 x2 
 
x= .  and xt = (x1 x2 . . . xd ), (1)
 .. 
xd
where here and below, all components take on real values. We denote an n × d
(rectangular) matrix M and its d × n transpose Mt as

 
m11 m12 m13 ... m1d
 m21 m22 m23 ... m2d 
 
M =  .. .. .. .. ..  and (2)
 . . . . . 
mn1 mn2 mn3 . . . mnd
 
m11 m21 ... mn1
 m12 m22 ... mn2 
 
 mn3 
Mt =  m13 m23 ... . (3)
 .. .. .. .. 
 . . . . 
m1d m2d . . . mnd

In other words, the ij th entry of Mt is the jith entry of M.

A square (d × d) matrix is called symmetric if its entries obey mij = mji ; it is
called skew-symmetric (or anti-symmetric) if mij = −mji . An general matrix is called
non-negative matrix if mij ≥ 0 for all i and j. A particularly important matrix is the
identity identity matrix, I — a d × d (square) matrix whose diagonal entries are 1’s, and all
matrix other entries 0. The Kronecker delta function or Kronecker symbol, defined as
½
Kronecker 1 if i = j
δij = (4)
delta 0 otherwise,
can function as an identity matrix. A general diagonal matrix (i.e., one having 0 for all
off diagonal entries) is denoted diag(m11 , m22 , ..., mdd ), the entries being the successive
elements m11 , m22 , . . . , mdd . Addition of vectors and of matrices is component by
component.
A.2. LINEAR ALGEBRA 9

We can multiply a vector by a matrix, Mx = y, i.e.,

 
   y1
m11 m12 ... m1d x1  y2 
 m21    
 m22 ... m2d  x2   .. 
 ..  = . , (5)
  
.. .. .. ..
 . . . .  .  .. 
 . 
mn1 mn2 ... mnd xd
yn
where

X
d
yj = mji xi . (6)
i=1

Note that if M is not square, the dimensionality of y differs from that of x.

The inner product of two vectors having the same dimensionality will be denoted inner
here as xt y and yields a scalar: product

X
d
xt y = xi yi = yt x. (7)
i=1

It is sometimes also called the scalar product or dot product and denoted x • y. The
Euclidean norm or length of the vector is denoted Euclidean
√ norm
kxk = xt x; (8)
we call a vector “normalized” if kxk = 1. The angle between two d-dimensional
vectors obeys

xt y
cosθ = , (9)
||x|| ||y||
and thus the inner product is a measure of the colinearity of two vectors — a natural
indication of their similarity. In particular, if xt y = 0, then the vectors are orthogonal,
and if |xt y| = ||x|| ||y||, the vectors are colinear. From Eq. 9, we have immediately
the Cauchy-Schwarz inequality, which states

kxt yk ≤ ||x|| ||y||. (10)

We say a set of vectors {x1 , x2 , . . . , xn } is linearly independent if no vector in the linear
set can be written as a linear combination of any of the others. Informally, a set of d independ-
linearly independent vectors spans an d-dimensional vector space, i.e., any vector in ence
that space can be written as a linear combination of such spanning vectors.

A.2.2 Outer product

The outer product (sometimes called matrix product) of two column vectors yields a matrix
matrix product

   
x1 (y1 y2 . . . yn ) x1 y1 x1 y2 ... x1 yn
 x2   x2 y1 x2 y2 ... x2 yn 
   
M = xyt =  ..  = .. .. .. .. , (11)
 .   . . . . 
xd xd y1 xd y2 . . . xd yn
 10 CONTENTS

that is, the components of M are mij = xi yj . Of course, if d 6= n, then M is not

square. Any matrix that can be written as the product of two vectors as in Eq. 11, is
separable called separable.

A.2.3 Derivatives of matrices

Suppose f (x) is a scalar function of d variables xi which we represent as the vector x.
Then the derivative or gradient of f with respect to this parameter vector is computed
component by component, i.e.,
 ∂f (x) 
∂x1
 
 
 ∂f (x) 
 
∂f (x)  ∂x2 
=

.
 (12)
∂x  .. 
 . 
 
 
∂f (x)
∂xd
If we have an n-dimensional vector valued function f , of a d-dimensional vector x,
Jacobian we calculate the derivatives and represent themas the Jacobian matrix
matrix  ∂f (x) ∂f1 (x)

1
. . .
∂f (x)  
∂x1 ∂xd
J(x) = =
..
.
..
.
..
.
.
 (13)
∂x
∂fn (x) ∂fn (x)
∂x1 ... ∂xd

If this matrix is square, its determinant (Sect. A.2.4) is called simply the Jacobian.
If the entries of M depend upon a scalar parameter θ, we can take the derivative
of Mmponent by component, to get another matrix, as
 ∂m11 ∂m12 
∂θ ∂θ . . . ∂m
∂θ
1d

∂M 
2d 
∂m21
 ∂θ
∂m22
∂θ . . . ∂m
∂θ 
= . .. .. .. . (14)
∂θ  .. . . . 
∂mn1 ∂mn2 ∂mnd
∂θ ∂θ ... ∂θ
In Sect. A.2.6 we shall discuss matrix inversion, but for convenience we give here the
derivative of the inverse of a matrix, M−1 :
∂ ∂M −1
M−1 = −M−1 M . (15)
∂θ ∂θ
The following vector derivative identities can be verified by writing out the com-
ponents:
∂
[Mx] = M (16)
∂x
∂ t ∂ t
[y x] = [x y] = y (17)
∂x ∂x
∂ t
[x Mx] = [M + Mt ]x. (18)
∂x
In the case where M is symmetric (as for instance a covariance matrix, cf. Sect. A.4.10),
then Eq. 18 simplifies to
A.2. LINEAR ALGEBRA 11

∂ t
[x Mx] = 2Mx. (19)
∂x
We use the second derivative of a scalar function f (x) to write a Taylor series (or
Taylor expansion) about a point x0 :

" #t " #t
2
∂f 1 t ∂ f
f (x) = f (x0 ) + (x − x0 ) + (x − x0 ) 2
(x − x0 ) + O(||x||3 ), (20)
∂x
|{z} x=x0 2 ∂x
|{z} x=x0
J H
where H is the Hessian matrix, the matrix of second-order derivatives of f (·) with Hessian
respect to the parameters, here evaluated at x0 . (We shall return in Sect. A.7 to matrix
consider the O(·) notation and the order of a function used in Eq. 20 and below.)
For a vector valued function we write the first-order expansion in terms of the
Jacobian as:
· ¸t
∂f
f (x) = f (x0 ) + (x − x0 ) + O(||x||2 ). (21)
∂x x=x0

A.2.4 Determinant and trace

The determinant of a d × d (square) matrix is a scalar, denoted |M|. If M is itself a
scalar (i.e., a 1 × 1 matrix M), then |M | = M . If M is 2 × 2, then |M| = m11 m22 −
m21 m12 . The determinant of a general square matrix can be computed by a method
called expansion by minors, and this leads to a recursive definition. If M is our d × d expansion
matrix, we define Mi|j to be the (d − 1) × (d − 1) matrix obtained by deleting the ith by minors
row and the j th column of M:

 j
N 
m11 m12 ··· N ··· ··· m1d
 m21 m22 ··· ··· ··· m2d 
 
 .. .. .. N .. 
 . . . ··· ··· . 
 
 .. .. N ..  = Mi|j . (22)
 . . ·N· · N ·N
·· ·N
·· . 
 N N N 
i  
 
 .. .. N . .. 
 . . · · · N · · · .. . 
md1 md2 ··· · · · · · · mdd
Given this defininition, we can now compute the determinant of M the expansion by
minors on the first column giving

|M| = m11 |M1|1 | − m21 |M2|1 | + m31 |M3|1 | − · · · ± md1 |Md|1 |, (23)
where the signs alternate. This process can be applied recursively to the successive
(smaller) matrixes in Eq. 23.
For a 3 × 3 matrix, this determinant calculation can be represented by “sweeping”
the matrix — taking the sum of the products of matrix terms along a diagonal, where
products from upper-left to lower-right are added with a positive sign, and those from
the lower-left to upper-right with a minus sign. That is,
 12 CONTENTS

¯ ¯
¯ m11 m12 m13 ¯
¯ ¯
|M| = ¯¯ m21 m22 m23 ¯¯ (24)
¯ m31 m32 m33 ¯
= m11 m22 m33 + m13 m21 m32 + m12 m23 m31
−m13 m22 m31 − m11 m23 m32 − m12 m21 m33 .

For two square matrices M and N, we have |MN| = |M| |N|, and furthermore |M| =
|Mt |. The determinant of any matrix is a measure of the d-dimensional hypervolume
it “subtends.” For the particular case of a covariance matrix Σ (Sect. A.4.10), |Σ|
is a measure of the hypervolume of the data taht yielded Σ.
The trace of a d × d (square) matrix, denoted tr[M], is the sum of its diagonal
elements:

X
d
tr[M] = mii . (25)
i=1

Both the determinant and trace of a matrix are invariant with respect to rotations of
the coordinate system.

A.2.5 Eigenvectors and eigenvalues

Given a d × d matrix M, a very important class of linear equations is of the form

Mx = λx, (26)
which can be rewritten as

(M − λI)x = 0, (27)
where λ is a scalar, I the identity matrix, and 0 the zero vector. This equation seeks
the set of d (possibly non-distinct) solution vectors {e1 , e2 , . . . , ed } — the eigenvectors
— and their associated eigenvalues {λ1 , λ2 , . . . , λd }. Under multiplication by M the
eigenvectors are changed only in magnitude — not direction:

Mej = λj ej . (28)
One method of finding the eigenvectors and eigenvalues is to solve the character-
character- istic equation (or secular equation),
istic
equation |M − λI| = λd + a1 λd−1 + . . . + ad−1 λ + ad = 0, (29)

secular for each of its d (possibly non-distinct) roots λj . For each such root, we then solve a
equation set of linear equations to find its associated eigenvector ej .

A.2.6 Matrix inversion

The inverse of a n × d matrix M, denoted M−1 , is the d × n matrix such that

MM−1 = I. (30)
cofactor Suppose first that M is square. We call the scalar Cij = (−1)i+j |Mi|j | the i, j cofactor
A.3. LAGRANGE OPTIMIZATION 13

or equivalently the cofactor of the i, j entry of M. As defined in Eq. 22, Mi|j is the
(d − 1) × (d − 1) matrix formed by deleting the ith row and j th column of M. The
adjoint of M, written Adj[M], is the matrix whose i, j entry is the j, i cofactor of M. adjoint
Given these definitions, we can write the inverse of a matrix as

Adj[M]
M−1 = . (31)
|M|
If M−1 does not exist — because the columns of M are not linearly independent or
M is not square — one typically uses instead the pseudoinverse M† , defined as pseudo-
inverse
M† = [Mt M]−1 Mt , (32)
which insures M† M = I. Again, note especially that here M need not be square.

A.3 Lagrange optimization

Suppose we seek the position x0 of an extremum of a scalar-valued function f (x),
subject to some constraint. For the following method to work, such a constraint
must be expressible in the form g(x) = 0. To find the extremum, we first form the
Lagrangian function

L(x, λ) = f (x) + λg(x), (33)

| {z }
=0

where λ is a scalar called the Lagrange undetermined multiplier. To find the ex- undeter-
tremum, we take the derivative mined
multiplier
∂L(x, λ) ∂f (x) ∂g(x)
= + λ = 0, (34)
∂x ∂x ∂x }
| {z
6=0 in gen.

and solve the resulting equations for λ and x0 — the position of the extremum.

A.4 Probability Theory

A.4.1 Discrete random variables
Let x be a random variable that can assume only a finite number m of different values
in the set X = {v1 , v2 , . . . , vm }. We denote pi as the probability that x assumes the
value vi :

pi = Pr{x = vi }, i = 1, . . . , m. (35)
Then the probabilities pi must satisfy the following two conditions:

pi ≥ 0 and
X
m
pi = 1. (36)
i=1
 14 CONTENTS

Sometimes it is more convenient to express the set of probabilities {p1 , p2 , . . . , pm }

in terms of the probability mass function P (x). To distinguish between a random probability
variable and the values that it can assume, it is sometime convenient to use an upper- mass
case letter for the random variable and the corresponding lower-case letter for the function
value. The mass function would then be written PX (x). While this avoids the possible
confusion in Eq. 37 and elsewhere (where x denotes a value, not a random variable), it
also significantly complicates our the notation. Since it is usually clear from context
whether one is referring to a random variable or its value, we will use the simpler
notation as whenever possible.
The probability mass function must satisfy the following two conditions:

P (x) ≥ 0 and
X
P (x) = 1. (37)
x∈X

A.4.2 Expected values

mean The mean or expected value or average of x is defined by

X X
m
E[x] = µ = xP (x) = vi pi . (38)
x∈X i=1

If one thinks of the probability mass function as defining a set of point masses, with
pi being the mass concentrated at x = vi , then the expected value µ is just the center
of mass. Alternatively, we can interpret µ as the arithmetic average of the values in a
large random sample. More generally, if f (x) is any function of x, the expected value
of f is defined by

X
E[f (x)] = f (x)P (x). (39)
x∈X

Note that the process of forming an expected value is linear, in that if α1 and α2 are
arbitrary constants,

E[α1 f1 (x) + α2 f2 (x)] = α1 E[f1 (x)] + α2 E[f2 (x)]. (40)

It is sometimes convenient to think of E as an operator — the (linear) expectation

expectation operator. Two important special-case expectations are the second moment and the
operator variance:
second X
moment E[x2 ] = x2 P (x) (41)
x∈X
variance X
Var[x] = E[(x − µ)2 ] = σ 2 = (x − µ)2 P (x), (42)
x∈X

standard where σ is the standard deviation of x. Physically, if we think of x as a random

deviation signal, the second moment is its total average power and the variance is its AC power.
A.4. PROBABILITY THEORY 15

Alternatively, the variance can be viewed as the moment of inertia of the probability
mass function. The variance is never negative, and is zero if and only if all of the
probability mass is concentrated at one point.
The standard deviation is a simple but valuable measure of how far values of x
are likely to depart from the mean. Its very name suggests that it is the standard
or typical amount one should expect a randomly drawn value for x to deviate or
differ from µ. Chebyshev’s inequality provides a mathematical relation between the Chebyshev’s
standard deviation and |x − µ|: inequality

1
Pr{|x − µ| > nσ} ≤ . (43)
n2
This inequality is not a tight bound (and it is useless for n < 1); a more practical rule
of thumb, which strictly speaking is true only for the normal distribution, is that 68%
of the values will lie within one, 95% within two, and 99.7% within three standard
deviations of the mean (Fig. A.1). Nevertheless, Chebyshev’s inequality shows the
strong link between the standard deviation and the spread of the distribution. In
addition, it suggests that |x−µ|/σ is a meaningful normalized measure of the distance
from x to the mean (cf. Sect. A.4.12).
By expanding the quadratic in Eq. 42, it is easy to prove the useful formula

Var[x] = E[x2 ] − (E[x])2 . (44)

Note that, unlike the mean, the variance is not linear. In particular, if y = αx, where
α is a constant, then Var[y] = α2 Var[x]. Moreover, the variance of the sum of two
random variables is usually not the sum of their variances. However, as we shall see
below, variances do add when the variables involved are statistically independent.
In the simple but important special case in which x is binary valued (say, v1 = 0
and v2 = 1), we can obtain simple formulas for µ and σ. If we let p = Pr{x = 1},
then it is easy to show that

µ = p and
p
σ = p(1 − p). (45)

A.4.3 Pairs of discrete random variables

Let x be a random variable whose domain is X = {v1 , v2 , . . . , vm }, and let y be
a random variable whose domain is Y = {w1 , w2 , . . . , wn . We can think of (x, y)
as a vector or a point in the product space of x and y. For each possible pair of product
values (vi , wj ) we have a joint probability pij = Pr{x = vi , y = wj }. These mn joint space
probabilities pij are non-negative and sum to 1. Alternatively, we can define a joint
probability mass function P (x, y) for which

P (x, y) ≥ 0 and
XX
P (x, y) = 1. (46)
x∈X y∈Y

The joint probability mass function is a complete characterization of the pair of ran-
dom variables (x, y); that is, everything we can compute about x and y, individually
16 CONTENTS

or together, can be computed from P (x, y). In particular, we can obtain the separate
marginal distributions for x and y by summing over the unwanted variable: marginal
distribution
P
Px (x) = P (x, y)
y∈Y
P
Py (y) = P (x, y). (47)
x∈X

As mentioned above, although the notation is more precise when we use subscripts
as in Eq. 47, it is common to omit them and write simply P (x) and P (y) whenever
the context makes it clear that these are in fact two different functions — rather than
the same function merely evaluated with different variables.

A.4.4 Statistical independence

Variables x and y are said to be statistically independent if and only if

P (x, y) = Px (x)Py (y). (48)

We can understand such independence as follows. Suppose that pi = Pr{x = vi } is
the fraction of the time that x = vi , and qj = Pr{y = wj } is the fraction of the time
that y = wj . Consider those situations where x = vi . If it is still true that the fraction
of those situations in which y = wj is the same value qj , it follows that knowing the
value of x did not give us any additional knowledge about the possible values of y;
in that sense y is independent of x. Finally, if x and y are statistically independent,
it is clear that the fraction of the time that the specific pair of values (vi , wj ) occurs
must be the product of the fractions pi qj = Px (vi )Py (wj ).

A.4.5 Expected values of functions of two variables

In the natural extension of Sect. A.4.2, we define the expected value of a function
f (x, y) of two random variables x and y by
XX
E[f (x, y)] = f (x, y)P (x, y), (49)
x∈X y∈Y

and as before the expectation operator E is linear:

E[α1 f1 (x, y) + α2 f2 (x, y)] = α1 E[f1 (x, y)] + α2 E[f2 (x, y)]. (50)
The means and variances are:

XX
µx = E[x] = xP (x, y)
x∈X y∈Y
XX
µy = E[y] = yP (x, y)
x∈X y∈Y
XX
σx2 = V [x] = E[(x − µx )2 ] = (x − µx )2 P (x, y)
x∈X y∈Y
XX
σy2 = V [y] = E[(y − µy ) ] =
2
(y − µy )2 P (x, y). (51)
x∈X y∈Y
 A.4. PROBABILITY THEORY 17

covar- An important new “cross-moment” can now be defined, the covariance of x and
iance y:
XX
σxy = E[(x − µx )(y − µy )] = (x − µx )(y − µy )P (x, y). (52)
x∈X y∈Y

We can summarize Eqs. 51 & 52 using vector notation as:

X
µ = E[x] = xP (x) (53)
x∈{X Y}
Σ = E[(x − µ)(x − µ)t ], (54)

where Σ is the covariance matrix (cf., Sect. A.4.9).

The covariance is one measure of the degree of statistical dependence between x
and y. If x and y are statistically independent, then σxy = 0. If α is a constant
and y = αx, which is a case of strong statistical dependence, it is also easy to show
that σxy = ασx2 . Thus, the covariance is positive if x and y both increase or decrease
together, and is negative if y decreases when x increases. If σxy = 0, the variables x and
y are said to be uncorrelated. It does not follow that uncorrelated variables must be uncorre-
statistically independent — covariance is just one measure of independence. However, lated
it is a fact that uncorrelated variables are statistically independent if they have a
multivariate normal distribution, and in practice statisticians often treat uncorrelated
variables as if they were statistically independent.
There is an important Cauchy-Schwarz inequality for the variances σx and σy and Cauchy-
the covariance σxy . It can be derived by observing that the variance of a random Schwarz
variable is never negative, and thus the variance of λx + y must be non-negative no inequality
matter what the value of the scalar λ. This leads to the famous inequality

2
σxy ≤ σx2 σy2 , (55)

which is analogous to the vector inequality (xt y)2 ≤ kxk2 kyk2 (Eq. 9).
The correlation coefficient, defined as correla-
tion coef-
σxy ficient
ρ= , (56)
σ x σy

is a normalized covariance, and must always be between −1 and +1. If ρ = +1,

then x and y are maximally positively correlated, while if ρ = −1, they are maxi-
mally negatively correlated. If ρ = 0, the variables are uncorrelated. It is common for
statisticians to consider variables to be uncorrelated for practical purposes if the mag-
nitude of their correlation coefficient is below some threshold, such as .05, although
the threshold that makes sense does depend on the actual situation.
If x and y are statistically independent, then for any two functions f and g

E[f (x)g(y)] = E[f (x)]E[g(y)], (57)

a result which follows from the definition of statistical independence and expectation.
Note that if f (x) = x − µx and g(y) = y − µy , this theorem again shows that
σxy = E[(x − µx )(y − µy )] is zero if x and y are statistically independent.
18 CONTENTS

A.4.6 Conditional probability

When two variables are statistically dependent, knowing the value of one of them
lets us get a better estimate of the value of the other one. This is expressed by the
following definition of the conditional probability of x given y:

Pr{x = vi , y = wj }
Pr{x = vi |y = wj } = , (58)
Pr{y = wj }
or, in terms of mass functions,

P (x, y)
P (x|y) = . (59)
Py (y)
Note that if x and y are statistically independent, this gives P (x|y) = Px (x). That
is, when x and y are independent, knowing the value of y gives you no information
about x that you didn’t already know from its marginal distribution Px (x).
To gain intuition about this definition of conditional probability, consider a simple
two-variable binary case where both x and y are either 0 or 1. Suppose that a large
number n of pairs of xy-values are randomly produced. Let nij be the number of
pairs in which we find x = i and y = j, i.e., we see the (0, 0) pair n00 times, the (0, 1)
pair n01 times, and so on, where n00 + n01 + n10 + n11 = n. Suppose we pull out
those pairs where y = 1, i.e., the (0, 1) pairs and the (1, 1) pairs. Clearly, the fraction
of those cases in which x is also 1 is

n11 n11 /n
= . (60)
n01 + n11 (n01 + n11 )/n
Intuitively, this is what we would like to get for P (x|y) when y = 1 and n is large.
And, indeed, this is what we do get, because n11 /n is approximately P (x, y) and
n11 /n
(n01 +n11 )/n is approximately Py (y) for large n.

A.4.7 The Law of Total Probability and Bayes’ rule

The expression
X
Py (y) = P (x, y) (61)
x∈X

is an instance of the Law of Total Probability. This law says that if an event A can
occur in m different ways A1 , A2 , . . . , Am , and if these m subevents are mutually
exclusive — that is, cannot occur at the same time — then the probability of A
occurring is the sum of the probabilities of the subevents Ai . In particular, the random
variable y can assume the value y in m different ways — with x = v1 , with x = v2 , . . .,
and with x = vm . Because these possibilities are mutually exclusive, it follows from
the Law of Total Probability that Py (y) is the sum of the joint probability P (x, y)
over all possible values for x. But from the definition of the conditional probability
P (y|x) we have

P (x, y) = P (y|x)Px (x), (62)

and thus, we obtain

A.4. PROBABILITY THEORY 19

P (y|x)Px (x)
P (x|y) = X , (63)
P (y|x)Px (x)
x∈X

or in words,

likelihood × prior
posterior = ,
evidence
where these terms are discussed more fully in Chapt. ??.
Equation 63 is usually called Bayes’ rule. Note that the denominator, which is
just Py (y), is obtained by summing the numerator over all x values. By writing
the denominator in this form we emphasize the fact that everything on the right-
hand side of the equation is conditioned on x. If we think of x as the important
variable, then we can say that the shape of the distribution P (x|y) depends only on
the numerator P (y|x)Px (x); the denominator is just a normalizing factor, sometimes
called the evidence, needed to insure that P (x|y) sums to one. evidence
The standard interpretation of Bayes’ rule is that it “inverts” statistical connec-
tions, turning P (y|x) into P (x|y). Suppose that we think of x as a “cause” and y
as an “effect” of that cause. That is, we assume that if the cause x is present, it is
easy to determine the probability of the effect y being observed, where the conditional
probability function P (y|x) — the likelihood — specifies this probability explicitly. If likelihood
we observe the effect y, it might not be so easy to determine the cause x, because
there might be several different causes, each of which could produce the same ob-
served effect. However, Bayes’ rule makes it easy to determine P (x|y), provided that
we know both P (y|x) and the so-called prior probability Px (x), the probability of x prior
before we make any observations about y. Said slightly differently, Bayes’ rule shows
how the probability distribution for x changes from the prior distribution Px (x) before
anything is observed about y to the posterior P (x|y) once we have observed the value posterior
of y.

A.4.8 Vector random variables

To extend these results from two variables x and y to d variables x1 , x2 , . . . , xd , it
is convenient to employ Pvector notation. The joint probability mass function P (x)
satisfies P (x) ≥ 0 and P (x) = 1 (Eq. 46), where the sum extends over all possible
values for the vector x. Note that P (x) is a function of d variables, x1 , x2 , . . . , xd ,
and can be a very complicated, multi-dimensional function. However, if the random
variables xi are statistically independent, it reduces to the product

P (x) = Px1 (x1 )Px2 (x2 ) · · · Pxd (xd )

Y
d
= Pxi (xi ). (64)
i=1

Here the separate marginal distributions Pxi (xi ) can be obtained by summing the joint
distribution over the other variables. In addition to these univariate marginals, other
marginal distributions can be obtained by this use of the Law of Total Probability.
For example, suppose that we have P (x1 , x2 , x3 , x4 , x5 ) and we want P (x1 , x4 ), we
merely calculate
 20 CONTENTS

XXX
P (x1 , x4 ) = P (x1 , x2 , x3 , x4 , x5 ). (65)
x2 x3 x5

One can define many different conditional distributions, such as P (x1 , x2 |x3 ) or
P (x2 |x1 , x4 , x5 ). For example,

P (x1 , x2 , x3 )
P (x1 , x2 |x3 ) = , (66)
P (x3 )
where all of the joint distributions can be obtained from P (x) by summing out the un-
wanted variables. If instead of scalars we have vector variables, then these conditional
distributions can also be written as

P (x1 , x2 )
P (x1 |x2 ) = , (67)
P (x2 )
and likewise, in vector form, Bayes’ rule becomes

P (x2 |x1 )P (x1 )

P (x1 |x2 ) = X . (68)
P (x2 |x1 )P (x1 )
x1

A.4.9 Expectations, mean vectors and covariance matrices

The expected value of a vector is defined to be the vector whose components are
the expected values of the original components. Thus, if f (x) is an n-dimensional,
vector-valued function of the d-dimensional random vector x,
 
f1 (x)
 f2 (x) 
 
f (x) =  .. , (69)
 . 
fn (x)
then the expected value of f is defined by
 
E[f1 (x)]
 E[f2 (x)]  X
 
E[f ] =  .. = f (x)P (x). (70)
 .  x
E[fn (x)]
mean In particular, the d-dimensional mean vector µ is defined by
vector    
E[x1 ] µ1
 E[x2 ]   µ2  X
   
µ = E[x] =  .  =  . = xP (x). (71)
 ..   ..  x
E[xd ] µd
covariance Similarly, the covariance matrix Σ is defined as the (square) matrix whose ij th element
matrix σij is the covariance of xi and xj :

σij = E[(xi − µi )(xj − µj )] i, j = 1 . . . d, (72)

as we saw in the two-variable case of Eq. 52. Therefore, in expanded form we have
A.4. PROBABILITY THEORY 21

 
E[(x1 − µ1 )(x1 − µ1 )] E[(x1 − µ1 )(x2 − µ2 )] ... E[(x1 − µ1 )(xd − µd )]
 E[(x2 − µ2 )(x1 − µ1 )] E[(x2 − µ2 )(x2 − µ2 )] ... E[(x2 − µ2 )(xd − µd )] 
 
Σ =  .. .. .. .. 
 . . . . 
E[(xd − µd )(x1 − µ1 )] E[(xd − µd )(x2 − µ2 )] . . . E[(xd − µd )(xd − µd )]
   2 
σ11 σ12 . . . σ1d σ1 σ12 . . . σ1d
 σ21 σ22 . . . σ2d   2 
   σ21 σ2 . . . σ2d 
=  ..  =  .. .. 
.. .. .. .. . . . (73)
 . . . .   . . .. 
σd1 σd2 . . . σdd σd1 σd2 . . . σd2

We can use the vector product (x − µ)(x − µ)t , to write the covariance matrix as

Σ = E[(x − µ)(x − µ)t ]. (74)

Thus, the diagonal elements of Σ are just the variances of the individual elements
of x, which can never be negative; the off-diagonal elements are the covariances,
which can be positive or negative. If the variables are statistically independent, the
covariances are zero, and the covariance matrix is diagonal. The analog to the Cauchy-
Schwarz inequality comes from recognizing that if w is any d-dimensional vector, then
the variance of wt x can never be negative. This leads to the requirement that the
quadratic form wt Σw never be negative. Matrices for which this is true are said to be
positive semi-definite; thus, the covariance matrix Σ must be positive semi-definite.
It can be shown that this is equivalent to the requirement that none of the eigenvalues
of Σ can ever be negative.

A.4.10 Continuous random variables

When the random variable x can take values in the continuum, it no longer makes
sense to talk about the probability that x has a particular value, such as 2.5136,
because the probability of any particular exact value will almost always be zero.
Rather, we talk about the probability that x falls in some interval (a, b); instead of
having a probability mass function P (x) we have a probability mass density function mass
p(x). The mass density has the property that density

Zb
Pr{x ∈ (a, b)} = p(x) dx. (75)
a

The name density comes by analogy with material density. If we consider a small
interval (a, a + ∆x) over which p(x) is essentially constant, having value p(a), we see
that p(a) = Pr{x ∈ (a, a + ∆x)}/∆x. That is, the probability mass density at x = a
is the probability mass Pr{x ∈ (a, a + ∆x)} per unit distance. It follows that the
probability density function must satisfy

p(x) ≥ 0 and
Z∞
p(x) dx = 1. (76)
−∞
22 CONTENTS

In general, most of the definitions and formulas for discrete random variables carry
over to continuous random variables with sums replaced by integrals. In particular,
the expected value, mean and variance for a continuous random variable are defined
by
Z∞
E[f (x)] = f (x)p(x) dx
−∞
Z∞
µ = E[x] = xp(x) dx (77)
−∞
Z∞
Var[x] = σ 2 = E[(x − µ)2 ] = (x − µ)2 p(x) dx,
−∞

and, as in Eq. 44, we have σ 2 = E[x2 ] − (E[x])2 .

The multivariate situation is similarly handled with continuous random vectors x.
The probability density function p(x) must satisfy

p(x) ≥ 0 and
Z∞
p(x) dx = 1, (78)
−∞

where the integral is understood to be a d-fold, multiple integral, and where dx is the
element of d-dimensional volume dx = dx1 dx2 · · · dxd . The corresponding moments
for a general n-dimensional vector-valued function are
Z∞ Z∞ Z∞ Z∞
E[f (x)] = ··· f (x)p(x) dx1 dx2 . . . dxd = f (x)p(x) dx (79)
−∞ −∞ −∞ −∞

and for the particular d-dimensional functions as above, we have

Z∞
µ = E[x] = xp(x) dx (80)
−∞
Z∞
Σ = E[(x − µ)(x − µ)t ] = (x − µ)(x − µ)t p(x) dx.
−∞

If the components of x are statistically independent, then the joint probability density
function factors as

Y
d
p(x) = pi (xi ) (81)
i=1

and the covariance matrix is diagonal.

Conditional probability density functions are defined just as conditional mass func-
tions. Thus, for example, the density for x given y is given by
A.4. PROBABILITY THEORY 23

p(x, y)
p(x|y) = (82)
py (y)
and Bayes’ rule for density functions is

p(y|x)px (x)
p(x|y) = Z∞ , (83)
p(y|x)px (x) dx
−∞

and likewise for the vector case.

Occassionally we will need to take the expectation with respect to a subset of the
variables, and in that case we must show this as a subscript, for instance
R∞
Ex1 [f (x1 , x2 )] = f (x1 , x2 )p(x1 ) dx1 .(83)
−∞

A.4.11 Distributions of sums of independent random variables

It frequently happens that we know the distributions for two independent random
variables x and y, and we need to know the distribution for their sum z = x + y. It
is easy to obtain the mean and the variance of the sum:

µz = E[z] = E[x + y] = E[x] + E[y] = µx + µy ,

σz2 = E[(z − µz )2 ] = E[(x + y − (µx + µy ))2 ] = E[((x − µx ) + (y − µy ))2 ]
= E[(x − µx )2 ] + 2 E[(x − µx )(y − µy )] +E[(y − µy )2 ] (84)
| {z }
=0
= σx2 + σy2 ,

where we have used the fact that the cross-term factors into E[x − µx ]E[y − µy ] when
x and y are independent; in this case the product is manifestly zero, since each of
the expectations vanishes. Thus, in words, the mean of the sum of two independent
random variables is the sum of their means, and the variance of their sum is the sum
of their variances. If the variables are random yet not independent — for instance
y = -x, where x is randomly distribution — then the variance is not the sum of the
component variances.
It is only slightly more difficult to work out the exact probability density function
for z = x + y from the separate density functions for x and y. The probability that z
is between ζ and ζ + ∆z can be found by integrating the joint density p(x, y) =
px (x)py (y) over the thin strip in the xy-plane between the lines x + y = ζ and
x + y = ζ + ∆z. It follows that, for small ∆z,
( Z∞ )
Pr{ζ < z < ζ + ∆z} = px (x)py (ζ − x) dx ∆z, (85)
−∞

and hence that the probability density function for the sum is the convolution of the convolution
probability density functions for the components:
Z∞
pz (z) = px ? py = px (x)py (z − x) dx. (86)
−∞
 24 CONTENTS

As one would expect, these results generalize. It is not hard to show that:
• The mean of the sum of d independent random variables x1 , x2 , . . . , xd is the
sum of their means. (In fact the variables need not be independent for this to
hold.)
• The variance of the sum is the sum of their variances.
• The probability density function for the sum is the convolution of the separate
density functions:

pz (z) = px1 ? px2 ? . . . ? pxd . (87)

A.4.12 Univariate normal density

Central One of the most important results of probability theory is the Central Limit Theorem,
Limit which states that, under various conditions, the distribution for the sum of d inde-
Theorem pendent random variables approaches a particular limiting form known as the normal
distribution. As such, the normal or Gaussian probability density function is very
Gaussian important, both for theoretical and practical reasons. In one dimension, it is defined
by
µ ¶2
1 x−µ
1 −
p(x) = √ e 2 σ
. (88)
2πσ
The normal density is traditionally described as a “bell-shaped curve”; it is com-
pletely determined by the numerical values for two parameters, the mean µ and the
variance σ 2 . This is often emphasized by writing p(x) ∼ N (µ, σ 2 ), which is read as
“x is distributed normally with mean µ and variance σ 2 .” The distribution is sym-
metrical about the mean, the peak occurring at x = µ and the width of the “bell” is
proportional to the standard deviation σ. The normal density satisfies the following
equations:

Z∞
E[1] = p(x) dx = 1
−∞
Z∞
E[x] = x p(x) dx = µ (89)
−∞
Z∞
E[(x − µ)2 ] = (x − µ)2 p(x) dx = σ 2 .
−∞

Normally distributed data points tend to cluster about the mean. Numerically, the
probabilities obey

Pr{|x − µ| ≤ σ} ≈ 0.68
Pr{|x − µ| ≤ 2σ} ≈ 0.95 (90)
Pr{|x − µ| ≤ 3σ} ≈ 0.997,
A.5. GAUSSIAN DERIVATIVES AND INTEGRALS 25

as shown in Fig. A.1.

p(u)

u
-4 -3 -2 -1 0 1 2 3 4
68%
95%
99.7%

Figure A.1: A one-dimensional Gaussian distribution, p(u) ∼ N (0, 1), has 68% of its
probability mass in the range |u| ≤ 1, 95% in the range |u| ≤ 2, and 99.7% in the
range |u| ≤ 3.

A natural measure of the distance from x to the mean µ is the distance |x − µ|

measured in units of standard deviations:

|x − µ|
r= , (91)
σ
the Mahalanobis distance from x to µ. Thus, the probability is .95 that the Maha- Mahalanobis
lanobis distance from x to µ will be less than 2. If a random variable x is modified distance
by (a) subtracting its mean and (b) dividing by its standard deviation, it is said to
be standardized. Clearly, a standardized normal random variable u = (x − µ)/σ has standardized
zero mean and unit standard deviation, that is,
1
p(u) = √ e−u /2 ,
2
(92)
2π
which can be written as p(u) ∼ N (0, 1).

A.5 Gaussian derivatives and integrals

Because of the prevalence of Gaussian functions throughout pattern recognition, we
often have occassion to integrate and differentiate them. The first three derivatives
of a one-dimensional (normalized) Gaussian are

· ¸
∂ 1 −x2 /(2σ 2 ) −x −x2 /(2σ2 )
√ e = √ e
∂x 2πσ 2πσ 3
· ¸
∂2 1 1 ¡ 2 ¢
e−x /(2σ ) −σ + x2 e−x /(2σ )
2 2 2 2
√ = √ (93)
∂x2 2πσ 2πσ 5
· ¸
∂3 1 1 ¡ ¢
e−x /(2σ ) 3xσ 2 − x3 e−x /(2σ ) ,
2 2 2 2
√ = √
∂x3 2πσ 2πσ 7
 26 CONTENTS

and are shown in Fig. A.2.

f'''

f'
f''
x
-4 -2 2 4

Figure A.2: A one-dimensional Gaussian distribution and its first three derivatives,
shown for f (x) ∼ N (0, 1).

error An improtant finite integral of the Gaussian is the so-called error function, defined
function as

Zu
2
e−x
2
erf(u) = √ /2
dx. (94)
π
0

Note especially the pre-factor of 2 and the lower limit of integration. As can be
seen from Fig. A.1, erf(0) = 0, erf(1) = .68 and lim erf(x) = 1. There is no closed
x→∞
analytic form for the error function, and thus we typically use tables, approximations
or numerical integration for its evaluation (Fig. A.3).

1
1-erf(u) erf(u)
0.8

0.6

0.4

1/u2
0.2

u
1 2 3 4

Figure A.3: The error function is the corresponds to the area under a standardized
Gaussian (Eq. 94) between −u and u, i.e., is the probability that a sample is chosen
from the Gaussian |x| ≤ u. Thus, the complementary probability, 1 − erf(u) is the
probability that a sample is chosen with |x| > u for the sandardized Gaussian. Cheby-
shev’s inequality states that for an arbitrary distribution having standard deviation
= 1, this latter probability is bounded by 1/u2 . As shown, this bound is quite loose
for a Gaussian.

In calculating moments of Gaussians, we need the general integral of powers of x

gamma weighted by a Gaussian. Recall first the definition of a gamma function
function
A.5. GAUSSIAN DERIVATIVES AND INTEGRALS 27

Z∞
xn e−x dx = Γ(n + 1), (95)
0

where the gamma function obeys

Γ(n) = nΓ(n − 1) (96)

√
and Γ(1/2) = π. For n an integer we have Γ(n + 1) = n × (n − 1) × (n − 2) . . . 1 = n!,
read “n factorial.”
Changing variables in Eq. 95, we find the moments of a (normalized) Gaussian
distribution as

Z∞ −x2 /(2σ 2 )
µ ¶
ne 2n/2 σ n n+1
2 x √ dx = √ Γ , (97)
2πσ π 2
0

where again we have used a pre-factor of 2 and lower integration limit of 0 in order
give non-trivial (i.e., non-vanishing) results for odd n.

A.5.1 Multivariate normal densities

Normal random variables have many desirable theoretical properties. For example, it
turns out that the convolution of two Gaussian functions is again a Gaussian function,
and thus the distribution for the sum of two independent normal random variables is
again normal. In fact, sums of dependent normal random variables also have normal
distributions. Suppose that each of the d random variables xi is normally distributed,
each with its own mean and variance: p(xi ) ∼ N (µi , σi2 ). If these variables are
independent, their joint density has the form

µ ¶2
1 xi − µi
Y
d Y
d
1 −
p(x) = pxi (xi ) = √ e 2 σi
i=1 i=1
2πσi
d µ
X ¶2
1 xi − µi
−
1 2 σi
= e i=1 . (98)
Y
d
(2π)d/2 σi
i=1

This can be written in a compact matrix form if we observe that for this case the
covariance matrix is diagonal, i.e.,
 
σ12 0 ... 0
 0 σ22 ... 0 
 
Σ= .. .. .. .. , (99)
 . . . . 
0 0 . . . σd2

and hence the inverse of the covariance matrix is easily written as

28 CONTENTS

 
1/σ12 0 ... 0
 0 1/σ22 ... 0 
 
Σ−1 =  .. .. .. .. . (100)
 . . . . 
0 0 . . . 1/σd2
Thus, the quadratic form in Eq. 98 can be written as
d µ
X ¶2
xi − µi
= (x − µ)t Σ−1 (x − µ). (101)
i=1
σi
Finally, by noting that the determinant of Σ is just the product of the variances, we
can write the joint density compactly in the form
1
1 − (x − µ)t Σ−1 (x − µ)
p(x) = e 2 . (102)
(2π)d/2 |Σ|1/2
This is the general form of a multivariate normal density function, where the covari-
ance matrix Σ is no longer required to be diagonal. With a little linear algebra, it
can be shown that if x obeys this density function, then

Z∞
µ = E[x] = x p(x) dx
−∞
Z∞
Σ = E[(x − µ)(x − µ)t ] = (x − µ)(x − µ)t p(x) dx, (103)
−∞

just as one would expect. Multivariate normal data tend to cluster about the mean
vector, µ, falling in an ellipsoidally-shaped cloud whose principal axes are the eigen-
vectors of the covariance matrix. The natural measure of the distance from x to the
mean µ is provided by the quantity

r2 = (x − µ)t Σ−1 (x − µ), (104)

which is the square of the Mahalanobis distance from x to µ. It is not as easy
to standardize a vector random variable (reduce it to zero mean and unit covariance
matrix) as it is in the univariate case. The expression analogous to u = (x−µ)/σ is u =
Σ−1/2 (x−µ), which involves the “square root” of the inverse of the covariance matrix.
The process of obtaining Σ−1/2 requires finding the eigenvalues and eigenvectors of
Σ, and is just a bit beyond the scope of this Appendix.

A.5.2 Bivariate normal densities

It is illuminating to look at the so-called bivariate normal density, that is, the case
of two Gaussian random variables x1 and x2 . In this case, it is convenient to define
σ12 = σ11 , σ22 = σ22 , and to introduce the correlation coefficient ρ defined by
σ12
ρ= . (105)
σ1 σ2
With this notation, that the covariance matrix becomes
A.5. GAUSSIAN DERIVATIVES AND INTEGRALS 29

· ¸ · ¸
σ11 σ12 σ12 ρσ1 σ2
Σ= = , (106)
σ21 σ22 ρσ1 σ2 σ22

and its determinant simplifies to

|Σ| = σ12 σ22 (1 − ρ2 ). (107)

Thus, the inverse covariance matrix is given by

· ¸
−1 1 σ22 −ρσ1 σ2
Σ =
σ12 σ22 (1 − ρ2 ) −ρσ1 σ2 σ12
" #
1
1
σ12
− σ1 σ2
ρ
= . (108)
1 − ρ2 − σ1 σ2
ρ 1
σ2 2

Next we explicitly expand the quadratic form in the normal density:

(x − µ)t Σ−1 (x − µ)
" #· ¸
1
1
σ12
− σ1ρσ2 (x1 − µ1 )
= [(x1 − µ1 ) (x2 − µ2 )]
1 − ρ2 − σ1 σ2
ρ 1
σ22
(x2 − µ2 )
1 £¡ x1 − µ1 ¢2 ¡ x1 − µ1 ¢¡ x2 − µ2 ¢ ¡ x2 − µ2 ¢2 ¤
= − 2ρ + . (109)
1−ρ 2 σ1 σ1 σ2 σ2

Thus, the general bivariate normal density has the form

1
px1 x2 (x1 , x2 ) = p × (110)
2πσ1 σ2 1 − ρ2
1 h³ x − µ ´2 ³ x − µ ´³ x − µ ´ ³ x − µ ´2 i
1 1 1 1 2 2 2 2
− − 2ρ +
e 2(1 − ρ2) σ 1 σ 1 σ 2 σ 2 .

As we can see from Fig. A.4, p(x1 , x2 ) is a hill-shaped surface over the x1 x2 plane.
The peak of the hill occurs at the point (x1 , x2 ) = (µ1 , µ2 ), i.e., at the mean vector µ.
The shape of the hump depends on the two variances σ12 and σ22 , and the correlation
coefficient ρ. If we slice the surface with horizontal planes parallel to the x1 x2 plane,
we obtain the so-called level curves, defined by the locus of points where the quadratic
form
³ x − µ ´2 ³ x − µ ´³ x − µ ´ ³ x − µ ´2
1 1 1 1 2 2 2 2
− 2ρ + (111)
σ1 σ1 σ2 σ2
is constant. It is not hard to show that |ρ| ≤ 1, and that this implies that the level
curves are ellipses. The x and y extent of these ellipses are determined by the variances
σ12 and σ22 , and their eccentricity is determined by ρ. More specifically, the principal
√ direction of the eigenvectors ei of Σ, and the different
axes of the ellipse are in the principal
widths in these directions λi . For instance, if ρ = 0, the principal axes of the ellipses axes
are parallel to the coordinate axes, and the variables are statistically independent. In
the special cases where ρ = 1 or ρ = −1, the ellipses collapse to straight lines. Indeed,
 30 p(x) CONTENTS

x2

µ2|1
µ

xˆ 1

x1

Figure A.4: A two-dimensional Gaussian having mean µ and non-diagonal covariance

Σ. If the value on one variable is known, for instance x1 = x̂1 , the distribution over
the other variable is Gaussian with mean µ2|1 .

the joint density becomes singular in this situation, because there is really only one
independent variable. We shall avoid this degeneracy by assuming that |ρ| < 1.
One of the important properties of the multivariate normal density is that all
conditional and marginal probabilities are also normal. To find such a density expli-
cilty, which we deonte px2 |x1 (x2 |x1 ), we substitute our formulas for px1 x2 (x1 , x2 ) and
px1 (x1 ) in the defining equation

px1 x2 (x1 , x2 )
px2 |x1 (x2 |x1 ) =
px1 (x1 )
" £¡ x −µ ¢2 ¡ x −µ ¢ ¡ x −µ ¢2 ¤ #
1 − 1 1 1 −2ρ 1σ 1 + 2σ 2
= p e 2(1−ρ2 ) σ1 1 2

2πσ1 σ2 1 − ρ 2
· ¡ ¢2 ¸
√ 1 x1 −µ1
× 2πσ1 e 2 σ1
£ x −µ ¤2
1 − 1 2 2 −ρ x1 −µ1
= √ p e 2(1−ρ2 ) σ2 σ1

2πσ2 1 − ρ2
à !2
1 x2 − [µ2 + ρ σ21 (x1 − µ1 )]
σ
− p
1
p
2 σ 1 − ρ2
= √ e 2
. (112)
2πσ2 1 − ρ 2

Thus, we have verified that the conditional density px1 |x2 (x1 |x2 ) is a normal distri-
conditional bution. Moreover, we have explicit formulas for the conditional mean µ2|1 and the
2
mean conditional variance σ2|1 :

σ2
µ2|1 = µ2 + ρ (x1 − µ1 ) and
σ1
2
σ2|1 = σ22 (1 − ρ2 ), (113)
A.6. INFORMATION THEORY 31

as illustrated in Fig. A.4.

These formulas provide some insight into the question of how knowledge of the
value of x1 helps us to estimate x2 . Suppose that we know the value of x1 . Then
a natural estimate for x2 is the conditional mean, µ2|1 . In general, µ2|1 is a linear
function of x1 ; if the correlation coefficient ρ is positive, the larger the value of x1 ,
the larger the value of µ2|1 . If it happens that x1 is the mean value µ1 , then the best
we can do is to guess that x2 is equal to µ2 . Also, if there is no correlation between
x1 and x2 , we ignore the value of x1 , whatever it is, and we always estimate x2 by
µ2 . Note that in that case the variance of x2 , given that we know x1 , is the same
2
as the variance for the marginal distribution, i.e., σ2|1 = σ22 . If there is correlation,
knowledge of the value of x1 , whatever the value is, reduces the variance. Indeed,
with 100% correlation there is no variance left in x2 when the value of x1 is known.

A.6 Information theory

A.6.1 Entropy and information
Assume we have a discrete set of symbols {v1 v2 . . . vm } with associated probabili-
ties pi . The entropy of the discrete distribution — a measure of the randomness or
unpredictability of a sequence of symbols drawn from it — is

X
m
H=− Pi log2 Pi , (114)
i=1

where here we use the logarithm is base 2. In case any of the probabilities vanish, we
use the relation 0 log 0 = 0. (For continuous distributions, we often use logarithm
base e, denoted ln.) If we recall the expectation operator (cf. Eq. 39), we can write
H = E[log 1/P ], where we think of P as being a random variable whose possible
values are p1 , p2 , . . . , pm . Note that the entropy does not depend on the symbols, but
just on their probabilities. The entropy is non-negative and measured in bits when bit
the base of the logarithm is 2. One bit corresponds to the uncertainty that can be
resolved by the answer to a single yes/no question. For a given number of symbols
m, the uniform distribution in which each symbol is equally likely, is the maximum
entropy distribution (and H = log2 m bits) — we have the maximum uncertainty
about the identity of each symbol that will be chosen. Conversely, if all the pi are 0
except one, we have the minimum entropy distribution (H = 0 bits) — we are certain
as to the symbol that will appear.
For a continuous distribution, the entropy is

Z∞
H=− p(x)log p(x)dx, (115)
−∞

and again H = E[log 1/p]. It is worth mentioning that among all continuous density
2
functions having a given mean µ and √ variance σ , it is the Gaussian that has the
maximum entropy (H = .5 + log2 ( 2πσ) bits). We can let σ approach zero to find
that a probability density in the form of a Dirac delta function, i.e., Dirac
delta
½
0 if x 6= a
δ(x − a) = with
∞ if x = a,
 32 CONTENTS

Z∞
δ(x)dx = 1, (116)
−∞

has the minimum entropy (H = −∞ bits). For a Dirac function, we are sure that the
value a will be selected each time.
Our use of entropy in continuous functions, such as in Eq. 115, belies some sub-
tle issues which are worth pointing out. If x had units, such as meters, then the
probability density p(x) would have to have units of 1/x. There is something funda-
mentally wrong in taking the logarithm of p(x), since the argument of any nonlinear
function has to be dimensionless. What we should really be dealing with is a dimen-
sionless quantity, say p(x)/p0 (x), where p0 (x) is some reference density function (cf.,
Sect. A.6.2).
One of the key properties of the entropy of a discrete distribution is that it is
invariant to “shuffling” the event labels; no such property is evident for continuous
variables. The related question with continuous variables concerns what happens
when one makes a change of variables. In general, if we make a change of variables,
such
R as y = x3 or even y = 10x, we will get a different value for the integral of
q(y)log q(y) dy, where q is the induced density for y. If entropy is supposed to
measure the intrinsic disorganization, it doesn’t make sense that y would have a
different amount of intrinsic disorganization than x.
Fortunately, in practice these concerns do not present important stumbling blocks
since relative entropy and differences in entropy are more fundamental than H taken
by itself. Nevertheless, questions of the foundations of entropy measures for continu-
ous variables are addressed in books listed in Bibliographical Remarks.

A.6.2 Relative entropy

Suppose we have two discrete distributions over the same variable x, p(x) and q(x).
Kullback- The relative entropy or Kullback-Leibler distance is a measure of the “distance” be-
Leibler tween these distributions:
distance
X q(x)
DKL (p(x), q(x)) = q(x)ln . (117)
x
p(x)

The continuous version is

Z∞
q(x)
DKL (p(x), q(x)) = q(x)ln dx. (118)
p(x)
−∞

Although DKL (p(·), q(·)) ≥ 0 and DKL (p(·), q(·)) = 0 if and only if p(·) = q(·), the
relative entropy is not a true metric, since DKL is not necessarily symmetric in the
interchange p ↔ q and furthermore the triangle inequality need not be satisfied.

A.6.3 Mutual information

Now suppose we have two distributions over possibly different random variables, e.g.,
p(x) and q(y). The mutual information is the reduction in uncertainty about one
variable due to the knowledge of the other variable
A.7. COMPUTATIONAL COMPLEXITY 33

X r(x, y)
I(p; q) = H(p) − H(p|q) = r(x, y)log , (119)
x,y
p(x)q(y)

where r(x, y) is the probability of finding value x and y. Mutual information is simply
the relative entropy between the joint distribution r(x, y) and the product distribution
p(x)q(y) and as such it measures how much the distributions of the variables differ
from statistical independence. Mutual information does not obey all the properties of
a metric. In particular, the metric requirement that if p(x) = q(y) then I(x; y) = 0
need not hold, in general. As an example, suppose we have two binary random
variables with r(0, 0) = r(1, 1) = 1/2, so r(0, 1) = r(1, 0) = 0. According to Eq. 119,
the mutual information between p(x) and q(y) is log 2 = 1.
The relationships among the entropy, relative entropy and mutual information
are summarized in Fig. A.5. The figure shows, for instance, that the joint entropy
H(p, q) is generally larger than individual entropies H(p) and H(q); that H(p) =
H(p|q) + I(p; q), and so on.

H(p,q)

H(p|q) I(p;q) H(q|p)

H(p)
H(q)

Figure A.5: The relationship among the entropy of distributions p and q, mutual
information I(p, q), and conditional entropies H(p|q) and H(q|p). From this figure one
can quickly see relationships among the information functions, for instance I(p; p) =
H(p); that if I(p; q) = 0 then H(q|p) = H(q), and so forth.

A.7 Computational complexity

In analyzing and describing the difficulty of problems and the algorithms designed to
solve such problems, we turn now to computational complexity. For instance, calcu-
lating the standard deviation of a distribution is somehow “harder” than calculating
its mean. Furthermore, some algorithms for computing some function may be faster
or take less memory, than another algorithm. How can we specify such differences,
independent of the current computer hardware (which is always changing anyway)?
To this end we use the concept of the order of a function and asymptotic nota-
tion and “big oh,” “big omega,” and “big theta” asymptotic notations. The three
asymptotic bounds most often used are:
 34 CONTENTS

Asymptotic upper bound O(g(x)) = {f (x): there exist positive constants c and
x0 such that 0 ≤ f (x) ≤ cg(x) for all x ≥ x0 }
Asymptotic lower bound Ω(g(x)) = {f (x): there exist positive constants c and
x0 such that 0 ≤ cg(x) ≤ f (x) for all x ≥ x0 }
Asymptotically tight bound Θ(g(x)) = {f (x): there exist positive constants c1 , c2 ,
and x0 such that 0 ≤ c1 g(x) ≤ f (x) ≤ c2 g(x) for all x ≥ x0 }

f(x) = O(g(x)) f(x) = Ω(g(x)) f(x) = Θ(g(x))

f(x) c2 g(x)
c g(x)

f(x)

c g(x)

f(x) c1 g(x)

x x x
x0 x0 x0

a) b) c)

Figure A.6: Three types of order of a function describe the upper, lower and tight
asymptotic bounds. a) f (x) = O(g(x)). b) f (x) = Ω(g(x)). c) f (x) = Θ(g(x)).

big oh Consider the asymptotic upper bound. We say that f (x) is “of the big oh order
of g(x)” (written f (x) = O(g(x)) if there exist constants c0 and x0 such that f (x) ≤
c0 g(x) for all x > x0 . We shall assume that all our functions are positive and dispense
with taking absolute values. This means simply that for sufficiently large x, an upper
bound on f (x) grows no worse than g(x). For instance, if f (x) = a + bx + cx2 then
f (x) = O(x2 ) because for sufficiently large x, the constant, linear and quadratic terms
can be “overcome” by proper choice of c0 and x0 . The generalization to functions
of two or more variables is straightforward. It should be clear that by the definition
above, the (big oh) order of a function is not unique. For instance, we can describe
our particular f (x) as being O(x2 ), O(x3 ), O(x4 ), O(x2 ln x), and so forth. We
little oh write the tightest asymptotic upper bound f (x) = o(g(x)), read “little oh of g(x)”
for the minimum in the class O(g(x)). Thus for instance if f (x) = ax2 + bx + c, then
f (x) = o(x2 ). Conversely, we use big omega notation, Ω(·), for lower bounds, and
little omega, ω(·), for the tightest lower bound.
Of these, the big oh notation has proven to be most useful since we generally
want an upper bound on the resources needed to solve a problem; it is frequently too
difficult to determine the little oh complexity.
Such a rough analysis does not tell us the constants c and x0 . For a finite size
problem it is possible (though not likely) that a particular O(x3 ) algorithm is simpler
than a particular O(x2 ) algorithm, and it is occasionally necessary for us to determine
these constants to find which of several implemementations is the simplest. Never-
theless, for our purposes the big oh notation as just described is generally the best
way to describe the computational complexity of an algorithm.
Suppose we have a set of n vectors, each of which is d-dimensional and we want to
calculate the mean vector. Clearly, this requires O(nd) multiplications. Sometimes we
stress space and time complexities, which are particularly relevant when contemplat-
ing parallel hardware implementations. For instance, the d-dimensional sample mean
A.7. BIBLIOGRAPHICAL REMARKS 35

could be calculated with d separate processors, each adding n sample values. Thus
we can describe this implementation as O(d) in space (i.e., the amount of memory space
or possibly the number of processors) and O(n) in time (i.e., number of sequential complexity
steps). Of course for any particular algorithm there may be a number of time-space
tradeoffs. time
complexity

Bibliographical Remarks
There are several good books on linear system, such as [13], and matrix computations
[9]. Lagrange optimization and related techniques are covered in the definitive book
[2]. While [12] is of historic interest and significance, readers seeking clear presen-
tations of the central ideas in probability are [11, 8, 6, 18]. Another book treating
the foundations is [3]. A handy reference to terms in probability and statistics is
[17]. The definitive collection of papers on information theory is [7], and an excellent
textbook, at the level of this one, is [5]; readers seeking a more abstract and formal
treatment should consult [10]. The multi-volume [14, 15, 16] contains a description
of computational complexity, the big oh and other asymptotic notations. Somewhat
more accessible treatments can be found in [4] and [1].
36 CONTENTS
Bibliography

[1] Alfred V. Aho, John E. Hopcroft, and Jeffrey D. Ullman. The Design and Anal-
ysis of Computer Algorithms. Addison-Wesley, Reading, MA, 1974.

[2] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Meth-

ods. Athena Scientific, xxx, xx, 1996.

[3] Patrick Billingsley. Probability and Measure. John Wiley and Sons, New York,
NY, 2 edition, 1986.

[4] Thomas H. Cormen, Charles E. Leiserson, and Ronald L. Rivest. Introduction

to Algorithms. MIT Press, Cambridge, MA, 1990.

[5] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley
Interscience, New York, NY, 1991.

[6] Alvin W. Drake. Fundamentals of Applied Probability Theory. McGraw-Hill,

New York, NY, 1967.

[7] David Slepian (editor). Key Papers in The Development oformation Theory.
IEEE Press, New York, NY, 1974.

[8] William Feller. An Introduction to Probability Theory and Its Applications, vol-
ume 1. Wiley, New York, NY, 1968.

[9] Gene H. Golub and Charles F. Van Loan. Matrix Computations. Johns Hopkins
University Press, Baltimore, MD, 3 edition, 1996.

[10] Robert M. Gray. Entropy and Information Theory. Springer-Verlag, New York,
NY, 1990.

[11] Richard W. Hamming. The Art of Probability for Scientists and Engineers.
Addison-Wesley, New York, NY, 1991.

[12] Harold Jeffreys. Theory of Probability. Oxford University Press, Oxford, UK,
1961 reprint edition, 1939.

[13] Thomas Kailath. Linear Systems. Prentice-Hall, Englewood Cliffs, NJ, 1980.

[14] Donald E. Knuth. The Art of Computer Programming, Volume I, volume 1.

Addison-Wesley, Reading, MA, 1 edition, 1973.

[15] Donald E. Knuth. The Art of Computer Programming, Volume III, volume 3.
Addison-Wesley, Reading, MA, 1 edition, 1973.

37
38 BIBLIOGRAPHY

[16] Donald E. Knuth. The Art of Computer Programming, Volume II, volume 2.
Addison-Wesley, Reading, MA, 1 edition, 1981.

[17] Francis H. C. Marriott. A Dictionary of Statistical Terms. Longman Scientific

& Technical, Essex, UK, 5 edition, 1990.
[18] Yuri A. Rozanov. Probability Theory: A Concise Course. Dover, New York, NY,
1969.
Index
†
, see matrix, pseudoinverse Gaussian
ρ, see correlation, coefficient bivariate, 28
E[·], see expectation conditional mean, 30
marginal, 30
adjoint, see matrix, adjoint mean, 24
asymptotic lower bound, see lower bound, univariate, 24
asymptotic variance, 24
asymptotic notation, 33 joint
asymptotic tight bound, see tight bound, singular, 30
asymptotic distance
asymptotic upper bound, see upper bound, Euclidean, 9
asymptotic Kullback-Leibler, 32
average, see expected value Mahalanobis, 25, 28
distribution
Bayes’ rule, 18, 19, 23
Gaussian, 24
vector, 20
area, 15
big oh, 34
covariance, 28
big omega, 34
eigenvector, 28
big theta, 34
moment, 27
Cauchy-Schwarz inequality, 9, 17 multivariate, 27
vector analog, 21 principal axes, 28
Central Limit Theorem, 24 joint, 20
Chebyshev’s inequality, 15 marginal, 19
cofactor maximum entropy, 31
matrix, see matrix, cofactor prior, 19
complexity dot product, see inner product
space, 35
time, 35 eigenvalue, 12
computational complexity, 33–35 eigenvector, 12
conditional probability, see probabil- entropy, 31
ity, conditional continuous distribution, 31
convolution, 23 discrete, 32
correlation relative, 32
coefficient, 17 error function (erf(·)), 26
coefficient (ρ), 29, 31 Euclidean norm, see distance, Euclidean
covariance, 17, 21 events
matrix, see matrix, covariance mutually exclusive, 18
normalized, 17 evidence, 19
cross moment, see covariance expectation
continuous, 22
density entropy, 31

39
40 INDEX

linearity, 14, 16 marginal, 16

vector, 20 distribution, 16
expected value, 14 mass function
two variables, 16 probability, see probability, mass
function
factorial, 27 matrix
function addition, 8
Dirac delta, 31 adjoint, 13
gamma, 26 anti-symmetric, 8
Kronecker, 8 covariance, 10
vector valued, 22 determinant, 28, 29
diagonal, 21, 22, 27
gamma function, see function, gamma
eigenvalues, 21
Gaussian derivative, 26
inverse, 27, 29
Gaussian deriviative, 25
derivative, 10–11
gradient, 10
determinant, 11–12
Hessian matrix, see matrix, Hessian hypervolume, 12
Hessian, 11
identity matrix, see matrix, identity (I) identity (I), 8
independence inverse
statistical, 16 derivative, 10
independent variables inversion, 12–13
sum, 23 Jacobean, 10
information multiplication, 9
bit, see bit non-negative, 8
mutual, 32–33 positive semi-definite, 21
information theory, 31–33 product, see outer product
inner product, 9 pseudoinverse, 13
separable, 10
Jacobean matrix, see matrix, Jacobean skew-symmetric, 8
Jacobian, 10, 11 square, 8
symmetric, 8, 10
Kronecker delta, see function, Kronecker
trace, 12
Kullback-Leibler, see distance, Kullback-
maximum entropy, 31
Leibler
mean, see expected value
Lagrange optimization, see optimiza- calculation
tion, Lagrange computational complexity, 33
Lagrange undetermined multiplier, 13 two variables, 16
Law of Total Probability, 18 mean vector, see vector, mean
level curves, 29 moment
likelihood, 19 cross, see covariance
linear independence, 9 second, 14
matrix columns, 13 multiple integral, 22
little oh, 34 mutual information, see information,
little omega, 34 mutual
lower bound
asymptotic (Ω(·)), 34 normal, see distribution, Gaussian

Mahalanobis distance, see distance, Ma- optimization

halanobis Lagrange, 13
INDEX 41

outer product, 9, 21 variables

uncorrelated, 17
principal axes, see axes, principal variance, 14
prior, 19 nonlinearity, 15
prior distribution, see distribution, prior two variables, 16
probability vector, 8
conditional, 18 addition, 8
density, 21 colinearity, 9
joint, 22 linearly independent, 9
joint, 15, 18 mean, 20
mass, 18, 21 orthogonal, 9
joint, 15 space, 9
mass function, 14 span, 9
total vector product, see outer product
law, see Bayes’rule
probability theory, 13–25
product space, 15

random variable
discrete, 13
vector, 19–21

scalar product, see inner product

second moment, see moment, second
space-time tradeoff, 35
standard deviation, 14, 25
statistical
independence
expectation, 17
statistical dependence, 17
statistical independence, see indepen-
dence, statistical, 17, 21
Gaussian, 29
vector, 19

Taylor series, 11
tight bound
asymptotic (Θ(·)), 34
trace, see matrix, trace
transpose, 8

unpredictability, see entropy

upper bound
asymptotic (O(·)), 34

variable
random
continuous, 21–23
discrete, 15
standardized, 28
standardized, 25