Introduction to Probability for Electrical Engineering

Prapun Suksompong
School of Electrical and Computer Engineering
Cornell University, Ithaca, NY 14853
[email protected]
January 23, 2008

Contents
1 Mathematical Background 4
1.1 Set Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Enumeration / Combinatorics / Counting . . . . . . . . . . . . . . . . . . 9
1.3 Dirac Delta Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2 Classical Probability 20
2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Probability Foundations 24
3.1 Algebra and σ-algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Kolmogorov’s Axioms for Probability . . . . . . . . . . . . . . . . . . . . . 29
3.3 Properties of Probability Measure . . . . . . . . . . . . . . . . . . . . . . . 31
3.4 Countable Ω . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Random Element 36
4.1 Random Variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Distribution Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.3 Discrete random variable . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4 Continuous random variable . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Mixed/hybrid Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.6 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.7 Misc . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

5 PMF Examples 48
5.1 Random/Uniform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Bernoulli and Binary distributions . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 Binomial: B(n, p) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.4 Geometric: G(β) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
 5.5 Poisson Distribution: P(λ) . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.6 Compound Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.7 Hypergeometric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.8 Negative Binomial Distribution (Pascal / Pólya distribution) . . . . . . . . 58
5.9 Beta-binomial distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.10 Zipf or zeta random variable . . . . . . . . . . . . . . . . . . . . . . . . . . 59

6 PDF Examples 59
6.1 Uniform Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
6.2 Gaussian Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.3 Exponential Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.4 Pareto: Par(α)–heavy-tailed model/density . . . . . . . . . . . . . . . . . . 67
6.5 Laplacian: L(α) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
6.6 Rayleigh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.7 Cauchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
6.8 Weibull . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

7 Expectation 69

8 Inequalities 76

9 Random Vectors 80
9.1 Random Sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

10 Transform Methods 86
10.1 Probability Generating Function . . . . . . . . . . . . . . . . . . . . . . . . 86
10.2 Moment Generating Function . . . . . . . . . . . . . . . . . . . . . . . . . 87
10.3 One-Sided Laplace Transform . . . . . . . . . . . . . . . . . . . . . . . . . 88
10.4 Characteristic Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

11 Functions of random variables 91

11.1 SISO case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
11.2 MISO case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
11.3 MIMO case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
11.4 Order Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

12 Convergences 108
12.1 Summation of random variables . . . . . . . . . . . . . . . . . . . . . . . . 114
12.2 Summation of independent random variables . . . . . . . . . . . . . . . . . 114
12.3 Summation of i.i.d. random variable . . . . . . . . . . . . . . . . . . . . . 114
12.4 Central Limit Theorem (CLT) . . . . . . . . . . . . . . . . . . . . . . . . . 116

13 Conditional Probability and Expectation 117

13.1 Conditional Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
13.2 Conditional Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
13.3 Conditional Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

2
14 Real-valued Jointly Gaussian 120

15 Bayesian Detection and Estimation 123

A More Math 126

A.1 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
A.2 Summations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
A.3 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
A.4 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
A.5 Gamma and Beta functions . . . . . . . . . . . . . . . . . . . . . . . . . . 135

3
1 Mathematical Background
1.1 Set Theory
1.1. Basic Set Identities:

• Idempotence: (Ac )c = A

• Commutativity (symmetry):

A∪B =B∪A, A∩B =B∩A

• Associativity:

◦ A ∩ (B ∩ C) = (A ∩ B) ∩ C
◦ A ∪ (B ∪ C) = (A ∪ B) ∪ C

• Distributivity

◦ A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C)
◦ A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)

• de Morgan laws

◦ (A ∪ B)c = Ac ∩ B c
◦ (A ∩ B)c = Ac ∪ B c

1.2. Basic Terminology:

• A ∩ B is sometimes written simply as AB.

• Sets A and B are said to be disjoint (A ⊥ B) if and only if A ∩ B = ∅

• A collection of sets (Ai : i ∈ I) is said to be pair-wise disjoint or mutually exclusive

[9, p 9] if and only if Ai ∩Aj = ∅ when i = j.

• A collection Π = (Aα : α ∈ I) of subsets of Ω (in this case, indexed or labeled by α

taking values in an index or label set I) is said to be a partition of Ω if

(a) Ω = Aα∈I and
(b) For all i = j, Ai ⊥ Aj (pairwise disjoint).

In which case, the collection (B ∩ Aα : α ∈ I) is a partition of B. In other words,

any set B can be expressed as B = (B ∩ Ai ) where the union is a disjoint union.
α

• The cardinality (or size) of a collection pr set A, denoted |A|, is the number of
elements of the collection. This number may be finite or infinite.

4
 name rule

Commutative laws A∩B =B∩A A∪B =B∪A

Associative laws A ∩ (B ∩ C) = (A ∩ B) ∩ C
A ∪ (B ∪ C) = (A ∪ B) ∪ C
Distributive laws A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C)
A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C)
DeMorgan’s laws A∩B =A∪B A∪B =A∩B
Complement laws A∩A=∅ A∪A=U
Double complement law A=A
Idempotent laws A∩A=A A∪A=A
Absorption laws A ∩ (A ∪ B) = A A ∪ (A ∩ B) = A
Dominance laws A∩∅=∅ A∪U =U
Identity laws A∪∅=A A∩U =A

Figure 1: Set Identities [17]

◦ Inclusion-Exclusion Principle:
   
n    
   
 Ai  = (−1)|I|+1  Ai .
   
i=1 φ=I⊂{1,...,n} i∈I

n   
  c   
◦  Ai  = |Ω| + (−1)  Ai .
|I|

i=1 φ=I⊂{1,...,n} i∈I

n
◦ If ∀i, Ai ⊂ B ( or equivalently, i=1 Ai ⊂ B), then
n   
    
   
 (B \ Ai ) = |B| + (−1)|I|  Ai .
   
i=1 φ=I⊂{1,...,n} i∈I

• An infinite set A is said to be countable if the elements of A can be enumerated

or listed in a sequence: a1 , a2 , . . . . Empty set and finite sets are also said to be
countable. By a countably infinite set, we mean a countable set that is not finite.

• A singleton is a set with exactly one element.

• N = {1, 2, , 3, . . . }
R = (−∞, ∞).

• For a set of sets, to avoid the repeated use of the word “set”, we will call it a
collection/class/family of sets.
Definition 1.3. Monotone sequence of sets

5
 • The sequence of events (A1 , A2 , A3 , . . . ) is monotone-increasing sequence of events if
and only if
A1 ⊂ A2 ⊂ A3 ⊂ . . ..
In which case,

n
◦ A i = An
i=1

∞
◦ lim An = Ai .
n→∞ i=1

Put A = lim An . We then write An A; that is An A if and only if ∀n

n→∞

∞
An ⊂ An+1 and An = A.
n=1

• The sequence of events (B1 , B2 , B3 , . . . ) is monotone-decreasing sequence of events if

and only if
B1 ⊃ B2 ⊃ B3 ⊃ . . ..
In which case,

n
◦ Bi = Bn
i=1

∞
◦ lim Bn = Bi
n→∞ i=1

Put B = lim Bn . We then write Bn  B; that is Bn  B if and only if ∀n

n→∞

∞
Bn+1 ⊂ Bn and Bn = B.
n=1

Note that An  A ⇔ Acn Ac .

1.4. An (Event-)indicator function IA : Ω → {0, 1} is defined by

1, if ω ∈ A
IA (ω) =
0, otherwise.

• Alternative notation: 1A .

• A = {ω : IA (ω) = 1}

• A = B if and only if IA = IB

• IAc (ω) = 1 − IA (ω)

• A ⊂ B ⇔ {∀ω, IA (ω) ≤ IB (ω)} ⇔ {∀ω, IA (ω) = 1 ⇒ IB (ω) = 1}

IA∩B (ω) = min (IA (ω) , IB (ω)) = IA (ω) · IB (ω)

IA∪B (ω) = max (IA (ω) , IB (ω)) = IA (ω) + IB (ω) − IA (ω) · IB (ω)

6
 N N  ∞
1.5. Suppose i=1 Ai = Bi for all finite N ∈ N. Then, ∞
i=1 i=1 Ai = i=1 Bi .

Proof. To show “⊂”, suppose x ∈ ∞ i=1 Ai . Then, ∃N0 such that x ∈ AN0 . Now, AN0 ⊂
 N0  N0 ∞ ∞
i=1 Ai = i=1 Bi ⊂ i=1 Bi . Therefore, x ∈ i=1 Bi . To show “⊃”, use symmetry.

1.6. Let A1 , A2 , . . . be a sequence of disjoint sets. Define Bn = ∪i>n Ai = ∪i≥n+1 Ai . Then,

(1) Bn+1 ⊂ Bn and (2) ∩∞ n=1 Bn = ∅.

Proof. Bn = ∪i≥n+1 Ai = (∪i≥n Ai ) ∪ An+1 = Bn+1 ∪ An+1 So, (1) is true. For (2), consider
two cases. (2.1) For element x ∈ / ∪i Ai , we know that x ∈
/ Bn and hence x ∈ / ∩∞
n=1 Bn . (2.2)
For x ∈ ∪i Ai , we know that ∃i0 such that x ∈ Ai0 . Note that x can’t be in other Ai because
the Ai ’s are disjoint. So, x ∈ / ∩∞
/ Bi+0 and therefore x ∈ n=1 Bn .

1.7. Any countable union can be written as a union of pairwise disjoint sets: Given any
sequence of sets Fn , define a new sequence by A1 = F1 , and
⎛ ⎞
 
An = Fn ∩ Fn−1c
∩ · · · ∩ F1c = Fn ∩ Fic = Fn \ ⎝ Fi ⎠ .
i∈[n−1] i∈[n−1]

∞ ∞
for n ≥ 2. Then, i=1 Fi = i=1 Ai where the union on the RHS is a disjoint union.
Proof. Note that the An are pairwise disjoint. To see this, consider An1 , An2 where n1 = n2 .
WLOG, assume n1 < n2 . This implies that Fnc1 get intersected in the definition of An2 . So,
⎛ ⎞
n1 −1 n2 −1
  ⎜ ⎟
An1 ∩ An2 = Fn1 ∩ Fnc1 ∩ ⎝ Fic ∩ Fn2 ∩ Fic ⎠ = ∅
   i=1 i=1
∅ i=n1
 
Also, for finite N ≥ 1, we have n∈[N ] Fn = n∈[N ] An (*). To see this, note that (*) is
 
true for N = 1. Suppose (*) is true for N = m and let B = n∈[m] An = n∈[m] Fn . Now,
for N = m + 1, by definition, we have
 
An = An ∪ Am+1 = (B ∪ Fm+1 ) ∩ Ω
n∈[m+1] n∈[m]

= B ∪ Fm+1 = Fi .
i∈[m+1]

So, (*) is true for N = m + 1. By induction, (*) is true for all finite N . The extension to
∞ is done via (1.5).
For finite union, we can modify the above statement by setting Fn = ∅ for n ≥ N .
Then, An = ∅ for n ≥ N .
By construction, {An : n ∈ N} ⊂ σ ({Fn : n ∈ N}). However, in general, it is not true
that {Fn : n ∈ N} ⊂ σ ({An : n ∈ N}). For example, for finite union with N = 2, we can’t
get F2 back from set operations on A1 , A2 because we lost information about F1 ∩ F2 . To
create a disjoint union which preserve the information about the overlapping parts of the

7
Fn ’s, we can define the A’s by ∩n∈N Bn where Bn is Fn or Fnc . This is done in (1.8). However,
this leads to uncountably many Aα , which is why we used the index α above instead of n.
The uncountability problem does not occur if we start with a finite union. This is shown
in the next result.
1.8. Decomposition:
• Fix sets A1 , A2 , . . . , An , not necessarily disjoint. Let Π be a collection of all sets of
the form
B = B1 ∩ B2 ∩ · · · ∩ Bn
where each Bi is either Aj or its complement. There are 2n of these, say
n
B (1) , B (2) , . . . , B (2 ) .
Then,
(a) Π is a partition of Ω and
 
(b) Π \ ∩j∈[n] Acj is a partition of ∪j∈[n] Aj .

Moreover, any Aj can be expressed as Aj = B (i) for some Si ⊂ [2n ]. More
i∈Sj
(i)
specifically, Aj is the union of all B which is constructed by Bj = Aj
• Fix sets
 A1 , A2 , . . ., not necessarily disjoint. Let Π be a collection of all sets of the form
B = n∈N Bn where each Bn is either An or its complement. There are uncountably
of these hence we will index the B’s by α; that is we write B (α) . Let I be the set of
all α after we eliminate all repeated B (α) ; note that I can still be uncountable. Then,
 
(a) Π = B (α) : α ∈ I is a partition of Ω and
(b) Π \ {∩n∈N Acn } is a partition of ∪n∈N An .

Moreover, any Aj can be expressed as Aj = B (α) for some Sj ⊂ I. Because I is
α∈Sj
uncountable, in general, Sj can be uncountable. More specifically, Aj is the (possibly
uncountable) union of all B (α) which is constructed by Bj = Aj . The uncountability
of Sj can be problematic because it implies that we need uncountable union to get
Aj back.
1.9. Let {Aα : α ∈ I} be a collection of disjoint sets where I is a nonempty index set. For
any set S ⊂ I, define a mapping g on 2I by g(S) = ∪α∈S Aα . Then, g is a 1:1 function if
and only if none of the Aα ’s is empty.
1.10. Let  
A = (x, y) ∈ R2 : (x + a1 , x + b1 ) ∩ (y + a2 , y + b2 ) = ∅ ,
where ai < bi . Then,
 
A = (x, y) ∈ R2 : x + (a1 − b2 ) < y < x + (b1 − a2 )
 
= (x, y) ∈ R2 : a1 − b2 < y − x < b1 − a2 .

8
 ‫ݕ‬

‫ݔ=ݕ‬

ܿ + ሺܾ1 െ ܽ2 ሻ A

ܿ + ሺܽ1 െ ܾ2 ሻ
ܿ

x
ܿ

 
Figure 2: The region A = (x, y) ∈ R2 : (x + a1 , x + b1 ) ∩ (y + a2 , y + b2 ) = ∅ .

1.2 Enumeration / Combinatorics / Counting

1.11. The four kinds of counting problems are:

(a) ordered sampling of r out of n items with replacement: nr ;

(b) ordered sampling of r ≤ n out of n items without replacement: (n)r ;

 
(c) unordered sampling of r ≤ n out of n items without replacement: nr ;
 
(d) unordered sampling of r out of n items with replacement: n+r−1
r
.

1.12. Given a set of n distinct items, select a distinct ordered sequence (word) of length
r drawn from this set.

• Sampling with replacement: μn,r = nr

◦ Ordered sampling of r out of n items with replacement.

◦ μn,1 = n
◦ μ1,r = 1
◦ μn,r = μn,r−1 for r > 1
◦ Examples:
 
∗ Suppose A is a finite set, then the cardinality of its power set is 2A  = 2|A| .
∗ There are 2r binary strings/sequences of length r.

9
 • Sampling without replacement:

r−1
n!
(n)r = (n − i) =
i=0
(n − r)!
= n · (n − 1) · · · (n − (r − 1)); r≤n
  
r terms
◦ Ordered sampling of r ≤ n out of n items without replacement.
◦ For integers r, n such that r > n, we have (n)r = 0.
◦ The definition in product form

r−1
(n)r = (n − i) = n · (n − 1) · · · (n − (r − 1))
  
i=0
r terms
can be extended to any real number n and a non-negative integer r. We define
(n)0 = 1. (This makes sense because we usually take the empty product to be
1.)
◦ (n)1 = n
◦ (n)r = (n − (r − 1))(n)r−1 . For example, (7)5 = (7 − 4)(7)4 .

1, if r = 1
◦ (1)r =
0, if r > 1
• Ratio:

r−1
(n − i) r−1 
 
(n)r i=0 i
= = 1−
nr 
r−1 n
(n) i=0
i=0
r−1 
  1 
r−1
−n i r(r−1)
− ni
≈ e =e i=0 = e− 2n

i=0
r2
≈ e− 2n

1.13. Factorial and Permutation: The number of arrangements (permutations) of n

≥ 0 distinct items is (n)n = n!.
• 0! = 1! = 1
• n! = n(n − 1)!
∞
• n! = e−t tn dt
0

• Stirling’s Formula:
√ √  1
2πe e(n+ 2 ) ln( e ) .
n
n! ≈ 2πnnn e−n =

10
1.14. Binomial coefficient:
 
n (n)r n!
= =
r r! (n − r)!r!

This gives the number of unordered sets of size r drawn from an alphabet of size n without
replacement; this is unordered sampling of r ≤ n out of n items without replacement. It
is also the number of subsets of size r that can be formed from a set of n elements. Some
properties are listed below:

(a) Use nchoosek(n,r) in MATLAB.

(b) Use combin(n,r) in Mathcad. However, to do symbolic manipulation, use the facto-
rial definition directly.
   n 
(c) Reflection property: nr = n−r .
   
(d) nn = n0 = 1.
   n 
(e) n1 = n−1 = n.
 
(f) nr = 0 if n < r or r is a negative integer.
   n 
(g) max nr =  n+1 .
r 2 

    n−1
(h) Pascal’s “triangle” rule: nk = n−1 k
+ k−1 . This property divides the process of
choosing k items into two steps where the first step is to decide whether to choose
the first item or not.

Figure 3: Pascal Triangle.

 n  n
(i) 0≤k≤n k
= 0≤k≤n k
= 2n−1
k even k odd
There are many ways to show this identity.

11
 (i) Consider the number of subsets of S = {a1 , a2 , . . . , an } of n distinct elements.
First choose subsets A of the first n − 1 elements of S. There are 2n−1 distinct
S. Then, for each A, to get set with even number of element, add element an to
A if and only if |A| is odd.
(ii) Look at binomial expansion of (x + y)n with x = 1 and y = −1.
   n 
(iii) For odd n, use the fact that nr = n−r .
min(n1 ,n2 ) n1 n2  n1 +n2  n1 +n2 
(j) k=0 k k
= n1
= n2
.

• This property divides the process of choosing k items from n1 +n2 into two steps
where the first step is to choose from the first n1 items.
 
• Can replace the min(n1 , n2 ) in the first sum by n1 or n2 if we define nk = 0 for
k > n.
n n2 2n
• r=1 r = n .

(k) Parallel summation:

n  
        
m k k+1 n n+1
= + + ··· + = .
m=k
k k k k k+1

To see this, suppose we try to choose k + 1 items from n + 1 items a1 , a2 , . . . , an+1 .

First, we choose whether to choose a1 . If so, then we  need to choose the rest k items
from the n items a2 , . . . , an+1 . Hence, we have the nk term. Now, suppose we didn’t
choose a1 . Then, we still need to choose k + 1 items from n items a2 , . . . , an+1 . We
then repeat the same argument on a2 in stead of a1 .
Equivalently,
r 
  r 
     
n+k n+k n+r+1 n+r+1
= = + . (1)
k n n+1 r
k=0 k=0

To prove the middle equality in (1), use induction on r.

1.15. Binomial theorem:

n 
 
nn
(x + y) = xr y n−r
r
r=0

n  

• Let x = y = 1, then n
r
= 2n .
r=0

• Entropy function:

H(p) = −p logb (p) − (1 − p) logb (1 − p)

◦ Binary: b = 2 ⇒

H2 (p) = −p log2 (p) − (1 − p) log2 (1 − p) .

12
 In which case,  
1 nH ( nr ) n
≤ 2nH ( n ) .
r
2 ≤
n+1 r
 
n
≈ 2nH2 ( n ) .
r
Hence,
r
• By repeated differentiating with respect to x followed by multiplication by x, we have
n n r n−r
◦ r=0 r r x y = nx(x + y)n−1 and
n 2 n r n−r
◦ r=0 r r x y = nx (x(n − 1)(x + y)n−2 + (x + y)n−1 ).
For x + y = 1, we have
n n r
◦ r=0 r r x (1 − x)
n−r
= nx and
n 2 n r 
◦ r=0 r r x (1 − x)
n−r
= nx (nx + 1 − x).

All identities above can be verified easily via Mathcad.

 n

1.16. Multinomial Counting: The multinomial coefficient n1 n2 ··· nr
is defined as

r 

i−1

 n− nk
k=0
i=1 ni
       
n n − n1 n − n1 − n2 nr
= · · ···
n1 n2 n3 nr
n!
= r
n!
i=1


r
It is the number of ways that we can arrange n = ni tokens when having r types of
i=1
symbols and ni indistinguishable copies/tokens of a type i symbol.

1.17. Multinomial Theorem


n− ij 
 1
n n−i 
j<r−1
n! n− ij 
r−1
(x1 + . . . + xr ) =n
···   xr
j<r
xikk
 
i1 =0 i2 =0 ir−1 =0 n− ik ! ik ! k=1
k<n k<n

• r-ary entropy function: Consider any vector p = (p1 , p2 , . . . , pr ) such that pi ≥ 0 and
r
pi = 1. We define
i=1

r
H (p) = − pi logb pi .
i=1

13
  n n!
Factorial expansion k = k!(n−k)! , k = 0, 1, 2, . . . , n
n  n 
Symmetry k = n−k , k
= 0, 1, 2, . . . , n
n   n  n 
Monotonicity 0 < 1 < ···
< n/2 , n≥0
n n−1 n−1
Pascal’s identity k = k−1 + k , k = 0, 1, 2, . . . , n
 n  
Binomial theorem (x + y)n = k=0 nk xk y n−k , n ≥ 0
n n n
Counting all subsets k=0 k = 2 , n ≥ 0
n  
k n
Even and odd subsets k=0 (−1) k = 0, n ≥ 0
n n2 2n
Sum of squares k=0 k = n , n≥0
n n 2 2n 2n
Square of row sums k=0 k = k=0 k , n ≥ 0
n n n−1
Absorption/extraction k = k k−1 , k = 0
 n m n n−k 
Trinomial revision m k = k m−k , 0 ≤ k ≤ m ≤ n
m n+k n+m+1
Parallel summation k=0 k = m , m, n ≥ 0
n−m m+k  n+1 
Diagonal summation k=0 m = m+1 , n ≥ m ≥ 0
r m n  m+n
Vandermonde convolution k=0 k r−k = r , m, n, r ≥ 0
n/2 n−k
Diagonal sums in Pascal’s k=0 k = Fn+1 (Fibonacci numbers), n ≥ 0
triangle (§2.3.2)
n n
  n−1
Other Common Identities k=0 k k = n2 , n≥0
n 2
 n
 n−2
k=0 k k = n(n + 1)2 , n≥0
n k
 n

k=0 (−1) k k = 0, n ≥ 0
n (nk) 2n+1 −1
k=0 k+1 = n+1 , n≥0
n k ( )
n
1
k=0 (−1) k+1
k
= n+1 , n≥0
n k−1 ( k ) 1
n
1 1
k=1 (−1) k = 1 + 2 + 3 + · · · + n, n > 0
n−1 n n   2n 
k=0 k k+1 = n−1 , n > 0
m m n  m+n
k=0 k p+k = m+p , m, n, p ≥ 0, n ≥ p + m

Figure 4: Binomial coefficient identities [17]

14
 ni
As a special case, let pi = n
, then
 
n n!
=  ≈ 2nH2 (p)
n1 n2 · · · nr r
ni !
i=1

1.18. The number of solutions

k+n−1  to x1 + x2 + · · · + xn = k for the xi ’s are nonnegative
k+n−1
integers is k
= n−1 .
(a) Suppose
 we further require that the xi are strictly positive (xi ≥ 1), then there are
k−1
n−1
solutions.
(b) Extra Lower-bound Requirement: Suppose we further require that xi ≥ ai
where
k−(a1 +athe ai are some
 given nonnegative integers, then the number of solution is
2 +···+an )+n−1
n−1  . Note that here we work with equivalent problem: y1 + y2 +
· · · + yn = k − ni=1 ai where yi ≥ 0.
(c) Extra Upper-bound Requirement: Suppose we further require that 0 ≤ xi <
bi . Let Ai be the set of solutions
 such that xi ≥ bi and xj ≥ 0 for j = i. The
number of solutions is k+n−1
n−1
− |∪ i=1 Ai | where the second term can be found via the
n

inclusion/exclusion principle
   
   
    
 Ai  = (−1)|I|+1
 Ai 
  
  i∈[n] I⊂[n] i∈I
I=∅
k−( 
and the fact that for any index set I ⊂ [n], we have |∩i∈I Ai | = bi )+n−1
i∈I
n−1
.
(d) Extra Range Requirement: Suppose we further require that ai ≤ xi < bi where
0 ≤ ai < bi , then we work instead with yi = xi − ai . The number of solutions is
⎛   ⎞ ⎛     ⎞

n 
n 
⎝ k− ai +n−1 ⎠  k− ai − (bi − ai ) + n − 1 ⎠
i=1 + (−1)|I| ⎝ i=1 i∈I .
n−1 I⊂[n] n−1
I=∅

1.19. The bars and stars argument:

• Consider the distribution of r = 10 indistinguishable balls into n = n distinguishable
cells. Then, we only concern with the number of balls in each cell. Using n − 1 = 4
bars, we can divide r = 10 stars into n = 5 groups.  For example, ****|***||**|*
n+r−1
would mean (4,3,0,2,1). In general, there are r
ways of arranging the bars and
stars.
 
• There are n+r−1 r
distinct vector x = xn1 of nonnegative integers such that x1 + x2 +
· · · + xn = r. We use n − 1 bars to separate r 1’s.
• Suppose r letters are drawn with replacement from a set {a1 , a2 , . . . , an }. Given a
drawn
n+r−1
sequence, let xinbe the number of ai in the drawn sequence. Then, there are
r
possible x = x1 .

15
 objects number of objects reference

Arranging objects in a row:

n distinct objects n! = P (n, n) = n(n − 1) . . . 2 · 1 §2.3.1
k out of n distinct objects nk = P (n, k) = n(n−1) . . . (n−k+1) §2.3.1
 n
 n!
some of the n objects are identical: k1 k2 ... kj = k1 ! k2 !...kj ! §2.3.2
k1 of a first kind, k2 of a second
kind, . . . , kj of a jth kind, and
where k1 + k2 + · · · + kj = n
 1 1

none of the n objects remains in its Dn = n! 1− 1! + · · · +(−1)n n! §2.4.2
original place (derangements)

Arranging objects in a circle (where rotations, but not reflections, are equivalent):
n distinct objects (n − 1)! §2.2.1
P (n,k)
k out of n distinct objects k §2.2.1

Choosing k objects from n distinct objects:

n!
order matters, no repetitions P (n, k) = (n−k)! = nk §2.3.1

order matters, repetitions allowed P R (n, k) = nk §2.3.3

 
order does not matter, no repeti- C(n, k) = nk = k!(n−k)!
n!
§2.3.2
tions
 k+n−1 
order does not matter, repetitions C R (n, k) = k §2.3.3
allowed

Figure 5: Counting problems and corresponding sections in [17].

16
 objects number of objects reference

Subsets:
 n
of size k from a set of size n k §2.3.2
n
of all sizes from a set of size n 2 §2.3.4
of {1, . . . , n}, without consecutive Fn+2 §3.1.2
elements

Placing n objects into k cells:

distinct objects into distinct cells kn §2.2.1
n
distinct objects into distinct cells, k k! §2.5.2
no cell empty
n n n
distinct objects into identical cells 1 + 2 +· · ·+ k = Bn §2.5.2
n
distinct objects into identical cells, k §2.5.2
no cell empty
 n

distinct objects into distinct cells, k1 k2 ... kj §2.3.2
with ki in cell i (i = 1, . . . , n),
and where k1 + k2 + · · · + kj = n
n+k−1
identical objects into distinct cells n §2.3.3
n−1
identical objects into distinct cells, k−1 §2.3.3
no cell empty
identical objects into identical pk (n) §2.5.1
cells
identical objects into identical pk (n) − pk−1 (n) §2.5.1
cells, no cell empty
n
Placing n distinct objects into k k §2.5.2
nonempty cycles

Solutions to x1 + · · · + xn = k:
k+n−1 k+n−1
nonnegative integers k = n−1 §2.3.3
k−1
positive integers n−1 §2.3.3
 k−(a1 +···+an )+n−1 
integers where 0 ≤ ai ≤ xi for all i n−1 §2.3.3
integers where 0 ≤ xi ≤ ai for one inclusion/exclusion principle §2.4.2
or more i

Figure 6: Counting problems (con’t) and corresponding sections in [17].

17
 objects number of objects reference

Functions from a k-element set to an n-element set:

all functions nk §2.2.1
n!
one-to-one functions (n ≥ k) nk = (n−k)! = P (n, k) §2.2.1
onto functions (n ≤ k) inclusion/exclusion §2.4.2
k   k   k  2 k  k
partial functions 0 + 1 n+ 2 n + · · · + k n §2.3.2
= (n + 1)k
Bit strings of length n:
all strings 2n §2.2.1
n−k
with given entries in k positions 2 §2.2.1
 n
with exactly k 0s k §2.3.2
 n  n   n
with at least k 0s k + k+1 + · · · + n §2.3.2
 n 
with equal numbers of 0s and 1s n/2 §2.3.2
palindromes 2n/2 §2.2.1
with an even number of 0s 2n−1 §2.3.4
without consecutive 0s Fn+2 §3.1.2

Figure 7: Counting problems (con’t) and corresponding sections in [17].

B(n) or Bn : Bell number nk = n(n − 1) . . . (n − k + 1) = P (n, k):

falling power
n!
b(n, k): associated Stirling number of the P (n, k) = (n−k)! : k-permutation
second kind
1
2n
Cn = n+1 n : Catalan number p(n): number of partitions of n
 n n!
C(n, k) = k = k!(n−k)! : binomial coefficient pk (n): number of partitions of n into
at most k summands
d(n, k): associated Stirling number of p∗k (n): number of partitions of n into
the first kind exactly k summands
n
En : Euler number k : Stirling cycle number
n
ϕ: Euler phi-function k : Stirling subset number

E(n, k): Eulerian number Tn : tangent number

Fn : Fibonacci number

Figure 8: Notation from [17]

18
1.3 Dirac Delta Function
The (Dirac) delta function or (unit) impulse function is denoted by δ(t). It is usually
depicted as a vertical arrow at the origin. Note that δ(t) is not a true function; it is
undefined at t = 0. We define δ(t) as a generalized function which satisfies the sampling
property (or sifting property) !
φ(t)δ(t)dt = φ(0)

for any function φ(t) which is continuous at t = 0. From this definition, It follows that
!
(δ ∗ φ)(t) = (φ ∗ δ)(t) = φ(τ )δ(t − τ )dτ = φ(t)

where we assume that φ is continuous at t. Intuitively we may  visualize δ(t) as a infinitely

tall, infinitely narrow rectangular pulse of unit area: lim ε 1 |t| ≤ 2 .
1 ε
ε→0
We list some interesting properties of δ(t) here.

• δ(t) = 0 when t = 0.
δ(t − T ) = 0 for t = T .

• A δ(t)dt = 1A (0).

(a) δ(t)dt = 1.

(b) {0} δ(t)dt = 1.
x
(c) −∞ δ(t)dt = 1[0,∞) (x). Hence, we may think of δ(t) as the “derivative” of the
unit step function U (t) = 1[0,∞) (x).

• φ(t)δ(t)dt = φ(0) for φ continuous at 0.

• φ(t)δ(t − T )dt = φ(T ) for φ continuous at T . In fact, for any ε > 0,
! T +ε
φ(t)δ(t − T )dt = φ(T ).
T −ε

• δ(at) = 1
|a|
δ(t)
• δ(t − t1 ) ∗ δ(t − t2 ) = δ (t − (t1 + t2 )).
• Fourier properties:

∞
◦ Fourier series: δ(x − a) = 1
2π
+ 1
π
cos(n(x − a)) on [−π, π].
 k=1
◦ Fourier transform: δ(t) = e j2πf t
df

• For a function g whose real-values roots are ti ,


n
δ (t − ti )
δ (g (t)) = (2)
k=1
|g  (ti )|

19
 [3, p 387]. Hence, !  f (x)
f (t)δ(g(t))dt = . (3)
|g  (x)|
x:g(x)=0

Note that the (Dirac) delta function is to be distinguished from the discrete time Kro-
necker delta function.
 massat 0; that is for any set A, we have δ(A) = 1[0 ∈ A].
As a finite measure, δ is a unit
In which case, we have again gdδ = f (x)δ(dx) = g(0) for any measurable g.
For a function g : D → Rn where D ⊂ Rn ,
 δ(x − z)
δ(g(x)) = (4)
|det dg(z)|
z:g(z)=0

[3, p 387].

2 Classical Probability
Classical probability, which is based upon the ratio of the number of outcomes favorable to
the occurrence of the event of interest to the total number of possible outcomes, provided
most of the probability models used prior to the 20th century. Classical probability remains
of importance today and provides the most accessible introduction to the more general
theory of probability.
Given a finite sample space Ω, the classical probability of an event A is
A the number of cases favorable to the outcome of the event
P (A) = = .
Ω the total number of possible cases
• In this section, we are more apt to refer to equipossible cases as ones selected at
random. Probabilities can be evaluated for events whose elements are chosen at
random by enumerating the number of elements in the event.

• The bases for identifying equipossibility were often

◦ physical symmetry (e.g. a well-balanced die, made of homogeneous material in

a cubical shape) or
◦ a balance of information or knowledge concerning the various possible outcomes.

• Equipossibility is meaningful only for finite sample space, and, in this case, the eval-
uation of probability is accomplished through the definition of classical probability.
2.1. Basic properties of classical probability:

• P (A) ≥ 0

• P (Ω) = 1

• P (∅) = 0

20
 • P (Ac ) = 1 − P (A)

• P (A ∪ B) = P (A) + P (B) − P (A ∩ B) which comes directly from

|A ∪ B| = |A| + |B| − |A ∩ B|.

• A ⊥ B is equivalent to P (A ∩ B) = 0.

• A ⊥ B ⇒ P (A ∪ B) = P (A) + P (B)

• Suppose Ω = {ω1 , . . . , ωn } and P (ωi ) = n1 . Then P (A) = p (ω).
ω∈A

◦ The probability of an event is equal to the sum of the probabilities of its com-
ponent outcomes because outcomes are mutually exclusive

2.2. Classical Conditional Probability : The conditional classical probability P (A|B)

of event A, given that event B = ∅ occurred, is given by
|A ∩ B| P (A ∩ B)
P (A|B) = = . (5)
|B| P (B)
• It is the updated probability of the event A given that we now know that B occurred.

• Read “conditional probability of A given B”.

• P (A|B) = P (A ∩ B|B) ≥ 0

• For any A such that B ⊂ A, we have P (A|B) = 1. This implies

P (Ω|B) = P (B|B) = 1.

• If A ⊥ C, P (A ∪ C |B ) = P (A |B ) + P (C |B )

• P (A ∩ B) = P (B)P (A|B)

• P (A ∩ B) ≤ P (A|B)

• P (A ∩ B ∩ C) = P (A) × P (B|A) × P (C|A ∩ B)

• P (A ∩ B) = P (A) × P (B|A)

• P (A ∩ B ∩ C) = P (A ∩ B) × P (C|A ∩ B)

• P (A, B |C ) = P (A |C ) P (B |A, C ) = P (B |C ) P (A |B, C )

2.3. Total Probability and Bayes Theorem If {Bi , . . . , Bn } is a partition of Ω, then
for any set A,

• Total Probability Theorem: P (A) = ni=1 P (A|Bi )P (Bi ).

• Bayes Theorem: Suppose P (A) > 0, we have P (Bk |A) =  nP (A|Bk )P (Bk ) .
i=1 P (A|Bi )P (Bi )

21
2.4. Independence Events: A and B are independent (A B) if and only if

|=
P (A ∩ B) = P (A)P (B) (6)

In classical probability, this is equivalent to

|A ∩ B||Ω| = |A||B|.

• Sometimes the definition for independence above does not agree with the everyday-
language use of the word “independence”. Hence, many authors use the term “statis-
tically independence” for the definition above to distinguish it from other definitions.

2.5. Having three pairwise independent events does not imply that the three events are
jointly independent. In other words,

A B, B C, A CA B C.
|=

|=

|=

|=

|=
Example: Experiment of flipping a fair coin twice. Ω = {HH, HT, T H, T T }. Define event
A to be the event that the first flip gives a H; that is A = {HH, HT }. Event B is the
event that the second flip gives a H; that is B = {HH, T H}. C = {HH, T T }. Note also
that even though the events A and B are not disjoint, they are independent.

2.6. Consider Ω of size 2n. We are given a set A ⊂ Ω of size n. Then, P (A) = 12 . We want
to find all sets B ⊂ Ω such that A B. (Note that without the required independence,
|=

there are 22n possible B.) For independence, we need

P (A ∩ B) = P (A)P (B).

Let r = |A ∩ B|. Then, r can be any integer from 0 to n. Also, let k = |B \ A|. Then, the
condition for independence becomes
r 1r+k
=
n 2 n
which is equivalent to r = k. So, the construction of the set B is given by choosing
 r
elements from set A, then choose
   r = k elements from set Ω \ A. There are n
r
choices
for the first part and nk = nr choice for the second part. Therefore, the total number of
possible B such that A B is
|=

n  2  
n 2n
= .
r=1
r n

2.1 Examples
2.7. Chevalier de Mere’s Scandal of Arithmetic:

Which is more likely, obtaining at least one six in 4 tosses of a fair die (event
A), or obtaining at least one double six in 24 tosses of a pair of dice (event B).

22
We have  4
5
P (A) = 1 − = .518
6
and  24
35
P (B) = 1 − = .491.
36
Therefore, the first case is more probable.
2.8. A random sample of size r with replacement is taken from a population of n elements.
The probability of the event that in the sample no element appears twice (that is, no
repetition in our sample) is
(n)r
.
nr
The probability that at least one element appears twice is
r−1 
 
i r(r−1)
pu (n, r) = 1 − 1− ≈ 1 − e− 2n .
i=1
n

In fact, when r − 1 < n2 , (A.2) gives

r−1 
 
1 r(r−1) 3n+2r−1 i 1 r(r−1)
e 2 n 3n ≤ 1− ≤ e2 n .
i=1
n

• From the approximation, to have pu (n, r) = p, we need

1 1"
r≈ + 1 − 8n ln (1 − p).
2 2
• Probability of coincidence birthday: Probability that there is at least two people who
have the same birthday in your class of n students
⎧
⎪ 1, if r ≥ 365,
⎪
⎪ ⎛ ⎞
⎪
⎨
= ⎜ 365 364 365 − (r − 1) ⎟
⎪
⎪ 1 − ⎜ · · · · · · ⎟ , if 0 ≤ r ≤ 365
⎪
⎪ ⎝ 365 365 365 ⎠
⎩   
r terms
◦ Birthday Paradox : In a group of 23 randomly selected people, the probability
that at least two will share a birthday (assuming birthdays are equally likely to
occur on any given day of the year) is about 0.5. See also (3).

2.9. Monte Hall’s Game: Started with showing a contestant 3 closed doors behind of
which was a prize. The contestant selected a door but before the door was opened, Monte
Hall, who knew which door hid the prize, opened a remaining door. The contestant was
then allowed to either stay with his original guess or change to the other closed door.
Question: better to stay or to switch? Answer: Switch. Because after given that the
contestant switched, then the probability that he won the prize is 23 .

23
 pu(n,r) for n = 365
1

0.9
pu  n, r  0.6
0.8

0.5 p 0.9
0.7 r  r 1

1 e 2n p 0.7
0.6
0.4
r p 0.5
0.5
p 0.3 r 23
0.4 r 23 n 0.3
p 0.1 n 365
0.3
n 365
0.2

0.2
0.1
0.1

0 0
0 5 10 15 20 25 30 35 40 45 50 55 0 50 100 150 200 250 300 350
r n

Figure 9: pu (n, r)

2.10. False Positives on Diagnostic Tests: Let D be the event that the testee has the
disease. Let + be the event that the test returns a positive result. Denote the probability
of having the disease by pD .
Now, assume that the test always returns positive result. This is equivalent to P (+|D) =
1 and P (+c |D) = 0. Also, suppose that even when the testee does not have a disease, the
test will still return a positive result with probability p+ ; that is P (+|Dc ).
If the test returns positive result, then the probability that the testee has the disease is
P (D ∩ +) pD 1
P (D |+ ) = = = p+
P (+) pD + p+ (1 − pD ) 1+ pD
(1 − pD )
PD
≈ ; for rare disease (pD  1)
P+

D Dc
+ P (+ ∩ D) P (+ ∩ Dc ) P (+)
= P (+|D)P (D) = P (+|Dc )P (Dc ) = P (+ ∩ D) + P (+ ∩ Dc )
= 1(pD ) = pD = p+ (1 − pD ) = pD + p+ (1 − pD )
+c P (+c ∩ D) P (+c ∩ Dc ) P (+c )
= P (+c |D)P (D) = P (+c |Dc )P (Dc ) = P (+c ∩D)+P (+c ∩Dc )
= 0(pD ) = 0 = (1 − p+ )(1 − pD ) = (1 − p+ )(1 − pD )
P (D) P (Dc ) P (Ω) = 1
= P (+ ∩ D) + P (+c ∩ D) = P (+∩Dc )+P (+c ∩Dc )
= pD = 1 − pD

3 Probability Foundations
To study formal definition of probability, we start with the probability space (Ω, A, P ). Let
Ω be an arbitrary space or set of points ω. Viewed probabilistically, a subset of Ω is an
event and an element ω of Ω is a sample point. Each event is a collection of outcomes
which are elements of the sample space Ω.
The theory of probability focuses on collections of events, called event algebras and
typically denoted A (or F) that contain all the events of interest (regarding the random
experiment E) to us, and are such that we have knowledge of their likelihood of occurrence.

24
The probability P itself is defined as a number in the range [0, 1] associated with each event
in A.

3.1 Algebra and σ-algebra

The class 2Ω of all subsets can be too large1 for us to define probability measures with
consistency, across all member of the class. In this section, we present smaller classes
which have several “nice” properties.

Definition 3.1. [7, Def 1.6.1 p38] An event algebra A is a collection of subsets of the
sample space Ω such that it is

(1) nonempty (this is equivalent to Ω ∈ A);

(2) closed under complementation (if A ∈ A then Ac ∈ A);

(3) and closed under finite unions (if A, B ∈ A then A ∪ B ∈ sA).

In other words, “A class is called an algebra on Ω if it contains Ω itself and is closed under
the formation of complements and finite unions.”

3.2. Examples of algebras

• Ω = any fixed interval of R. A = {finite unions of intervals contained in Ω}

• Ω = (0, 1]. B0 = the collection of all finite unions of intervals of the form (a, b] ⊂ (0, 1]
∞ 

◦ Not a σ-field. Consider the set 1
,1
2i+1 2i
i=1

3.3. Properties of an algebra A:

(a) Nonempty: ∅ ∈ A, X ∈ A

(b) A ⊂ 2Ω

(c) An algebra is closed under finite set-theoretic operations.

• A ∈ A ⇒ Ac ∈ A
• A, B ∈ A ⇒ A ∪ B ∈ A, A ∩ B ∈ A, A\B ∈ A, AΔB = (A\B ∪ B\A) ∈ A

n 
n
• A1 , A2 , . . . , An ∈ F ⇒ Ai ∈ F and Ai ∈ F
i=1 i=1

(d) The collection of algebras in Ω is closed under arbitrary intersection. In particular,

let A1 , A2 be algebras of Ω and let A = A1 ∩ A2 be the collection of sets common to
both algebras. Then A is an algebra.

(e) The smallest A is {∅, Ω}.

1
There is no problem when Ω is countable.

25
 (f) The largest A is the set of all subsets of Ω known as the power set and denoted by
2Ω .
(g) Cardinality of Algebras: An algebra of subsets of a finite set of n elements will always
have a cardinality of the form 2k , k ≤ n. It is the intersection of all algebras which
contain C.
3.4. There is a smallest (in the sense of inclusion) algebra containing any given family C
of subsets of Ω. Let C ⊂ 2X , the algebra generated by C is

G,
G is an algebra
C⊂G

i.e., the intersection of all algebra containing C. It is the smallest algebra containing C.
Definition 3.5. A σ-algebra A is an event algebra that is also closed under countable
unions,
(∀i ∈ N)Ai ∈ A =⇒ ∪i∈N ∈ A.
Remarks:
• A σ-algebra is also an algebra.
• A finite algebra is also a σ-algebra.
3.6. Because every σ-algebra is also an algebra, it has all the properties listed in (3.3).
Extra properties of σ-algebra A are as followed:

∞ 
∞
(a) A1 , A2 , . . . ∈ A ⇒ Aj ∈ A and Aj ∈ A
j=1 j=1

(b) A σ-field is closed under countable set-theoretic operations.

(c) The collection of σ-algebra in Ω is closed under arbitrary intersection, i.e., let
Aα be σ-algebra ∀α ∈ A where A is some index set, potentially uncountable. Then,
Aα is still a σ-algebra.
α∈A

(d) An infinite σ-algebra F on X is uncountable i.e. σ-algebra is either finite or uncount-

able.
(e) If A and B are σ-algebra in X, then, it is not necessary that A ∪ B is also a σ-
algebra. For example, let E1 , E2 ⊂ X distinct, and not complement of one another.
Let Ai = {∅, Ei , Eic , X}. Then, Ei ∈ A1 ∪ A2 but E1 ∪ E2 ∈
/ A 1 ∪ A2 .
Definition 3.7. Generation of σ-algebra Let C ⊂ 2Ω , the σ-algebra generated by
C, σ (C) is 
G,
G is a σ−field
C⊂G

i.e., the intersection of all σ-field containing C. It is the smallest σ-field containing C

26
 • If the set Ω is not implicit, we will explicitly write σX (C)

• We will say that a set A can be generated by elements of C if A ∈ σ (C)

3.8. Properties of σ (C):

(a) σ (C) is a σ-field

(b) σ (C) is the smallest σ-field containing C in the sense that if H is a σ-field and
C ⊂ H, then σ (C) ⊂ H

(c) C ⊂ σ (C)

(d) σ (σ (C)) = σ (C)

(e) If H is a σ-field and C ⊂ H, then σ (C) ⊂ H

(f) σ ({∅}) = σ ({X}) = {∅, X}

(g) σ ({A}) = {∅, A, Ac , X} for A ⊂ Ω.

(h) σ ({A, B}) has at most 16 elements. They are ∅, A, B, A∩B, A∪B, A\B, B \A, AΔB,
and their corresponding complements. See also (3.11). Some of these 16 sets can be
the same and hence the use of “at-most” in the statement.

(i) σ (A) = σ (A ∪ {∅}) = σ (A ∪ {X}) = σ (A ∪ {∅, X})

(j) A ⊂ B ⇒ σ (A) ⊂ σ (B)

(k) σ (A) , σ (B) ⊂ σ (A) ∪ σ (B) ⊂ σ (A ∪ B)

(l) σ (A) = σ (A ∪ {∅}) = σ (A ∪ {Ω}) = σ (A ∪ {∅, Ω})

3.9. For the decomposition described in (1.8), let the starting collection be C1 and the
decomposed collection be C2 .

• If C1 is finite, then σ (C2 ) = σ (C1 ).

• If C1 is countable, then σ (C2 ) ⊂ σ (C2 ).

3.10. Construction of σ-algebra from countable partition: An intermediate-sized

σ-algebra (a σ-algebra smaller than 2Ω ) can be constructed by first partitioning Ω into
subsets and then forming the power set of these subsets, with an individual subset now
playing the role of an individual outcome ω.
Given a countable2 partition Π = {Ai , i ∈ I} of Ω, we can form a σ-algebra A by
including all unions of subcollections:

A = {∪α∈S Aα : S ⊂ I} (7)
2
In this case, Π is countable if and only if I is countable.

27
 
[7, Ex 1.5 p 39] where we define Ai = ∅. It turns out that A = σ(Π). Of course, a
i∈∅
σ-algebra is also an algebra. Hence, (7) is also a way to construct an algebra. Note that,
from (1.9), the necessary and sufficient condition for distinct S to produce distinct element
in (7) is that none of the Aα ’s are empty.
In particular, for countable Ω = {xi : i ∈ N}, where xi ’s are distinct. If we want a
σ-algebra which contains {xi } ∀i, then, the smallest one is 2Ω which happens to be the
biggest one. So, 2Ω is the only σ-algebra which is “reasonable” to use.
3.11. Generation of σ-algebra from finite partition: If a finite collection Π =
{Ai , i ∈ I} of non-empty sets forms a partition of Ω, then the algebra generated by Π is the
same as the σ-algebra generated by Π and it is given by (7). Moreover, |σ(Π)| is 2|I| = 2|Π| ;
that is distinct sets S in (7) produce distinct member of the (σ-)algebra.
Therefore, given a finite collection of sets C = {C1 , C2 , . . . , Cn }. To find an algebra or a
σ-algebra generated by C, the first step is to use the Ci ’s to create a partition of Ω. Using
(1.8), the partition is given by

Π = {∩ni=1 Bi : Bi = Ci or Cic } . (8)

By (3.9), we know that σ(π) = σ(C). Note that there are seemingly 2n sets in Π however,
some of them can be ∅. We can eliminate the empty set(s) from Π and it is still a partition.
So the cardinality of Π in (8) (after empty-set elimination) is k where k is at most 2n . The
partition Π is then a collection of k sets which can be renamed as A1 , . . . , Ak , all of which
are non-empty. Applying the construction in (7) (with I = [n]), we then have σ(C) whose
n
cardinality is 2k which is ≤ 22 . See also properties (3.8.7) and (3.8.8).
Definition 3.12. In general, the Borel σ-algebra or Borel algebra B is the σ-algebra
generated by the open subsets of Ω
• Call BΩ the σ-algebra of Borel subsets of Ω or σ-algebra of Borel sets on Ω

• Call set B ∈ BΩ Borel set of Ω

(a) On Ω = R, the σ-algebra generated by any of the followings are Borel σ-algebra:

(i) Open sets

(ii) Closed sets
(iii) Intervals
(iv) Open intervals
(v) Closed intervals
(vi) Intervals of the form (−∞, a], where a ∈ R
i. Can replace R by Q
ii. Can replace (−∞, a] by (−∞, a), [a, +∞), or (a, +∞)
iii. Can replace (−∞, a] by combination of those in 1(f)ii.

(b) For Ω ⊂ R, BΩ = {A ∈ BR : A ⊂ Ω} = BR ∩ Ω where BR ∩ Ω = {A ∩ Ω : A ∈ BR }

28
 (c) Borel σ-algebra on the extended real line is the extended σ-algebra

BR = {A ∪ B : A ∈ BR , B ∈ {∅, {−∞} , {∞} , {−∞, ∞}}}

It is generated by, for example

(i) A ∪ {{−∞} , {∞}} where σ (A) = BR

'( ) '( *) ' *)
(ii) â, b̂ , â, b̂ , â, b̂
(iii) {[−∞, b]} , {[−∞, b)} , {(a, +∞]} , {[a, +∞]}
'( *) '( )
(iv) −∞, b̂ , −∞, b̂ , {(â, +∞]} , {[â, +∞]}

Here, a, b ∈ R and â, b̂ ∈ R ∪ {±∞}

(d) Borel σ-algebra in Rk is generated by

(i) the class of open sets in Rk

(ii) the class of closed sets in Rk
(iii) the class of bounded semi-open rectangles (cells) I of the form
 
I = x ∈ Rk : ai < xi ≤ bi , i = 1, . . . , k .
k
Note that I = ⊗ (ai , bi ] where ⊗ denotes the Cartesian product ×
i=1

(iv) 
the class of “southwest regions”  Sx of points “southwest” to x ∈ Rk , i.e. Sx =
y ∈ Rk : yi ≤ xi , i = 1, . . . , k

3.13. The Borel σ-algebra B of subsets of the reals is the usual algebra when we deal
with real- or vector-valued quantities.
Our needs will not require us to delve into these issues beyond being assured that
events we discuss are constructed out of intervals and repeated set operations on these
intervals and these constructions will not lead us out of B. In particular, countable unions
of intervals, their complements, and much, much more are in B.

3.2 Kolmogorov’s Axioms for Probability

Definition 3.14. Kolmogorov’s Axioms for Probability [12]: A set function satisfying
K0–K4 is called a probability measure.

K0 Setup: The random experiment E is described by a probability space (Ω, A, P ) con-

sisting of an event σ-algebra A and a real-valued function P : A → R.

K1 Nonnegativity: ∀A ∈ A, P (A) ≥ 0.

K2 Unit normalization: P (Ω) = 1.

29
K3 Finite additivity: If A, B are disjoint, then P (A ∪ B) = P (A) + P (B).

K4 Monotone continuity: If (∀i > 1) Ai+1 ⊂ Ai and ∩i∈N Ai = ∅ (a nested series of sets
shrinking to the empty set), then

lim P (Ai ) = 0.
i→∞

K4 Countable or σ-additivity: If (Ai ) is a countable collection of pairwise disjoint (no

overlap) events, then +∞ ,
 ∞
P Ai = P (Ai ).
i=1 i=1

• Note that there is never a problem with the convergence of the infinite sum; all partial
sums of these non-negative summands are bounded above by 1.

• K4 is not a property of limits of sequences of relative frequencies nor meaningful in

the finite sample space context of classical probability, is offered by Kolmogorov to
ensure a degree of mathematical closure under limiting operations [7, p 111].

• K4 is an idealization that is rejected by some accounts (usually subjectivist) of prob-

ability.

• If P satisfies K0–K3, then it satisfies K4 if and only if it satisfies K4 [7, Theorem

3.5.1 p 111].

Proof. To show “⇒”, consider disjoint A1 , A2 , . . .. Define Bn = ∪i>n Ai . Then, by

(1.6), Bn+1 ⊂ Bn and ∩∞
n=1 Bn = ∅. So, by K4, lim P (Bn ) = 0. Furthermore,
n→∞
+ ,

∞ 
n
Ai = Bn ∪ Ai
i=1 i=1

where all the sets on the RHS are disjoint. Hence, by finite additivity,
+∞ ,
  n
P Ai = P (Bn ) + P (Ai ).
i=1 i=1

Taking limiting as n → ∞ gives us K4 .

To show “⇐”, see (3.17).

Equivalently, in stead of K0-K4, we can define probability measure using P0-P2 below.

Definition 3.15. A probability measure defined on a σ-algebra A of Ω is a (set) function

(P0) P : A → [0, 1]

that satisfies:

30
(P1,K2) P (Ω) = 1

(P2,K4 ) Countable additivity

 ∞ : For
 every countable sequence (An )∞
n=1 of disjoint ele-
 
∞
ments of A, one has P An = P (An )
n=1 n=1

• P (∅) = 0

• The number P (A) is called the probability of the event A

• The triple (Ω, A, P ) is called a probability measure space, or simply a probability

space

• A support of P is any A-set A for which P (A) = 1

3.3 Properties of Probability Measure

3.16. Properties of probability measures:

(a) P (∅) = 0

(b) 0 ≤ P ≤ 1: For any A ∈ A, 0 ≤ P (A) ≤ 1

(c) If P (A) = 1, A is not necessary Ω.

(d) Additivity: A, B ∈ A, A ∩ B = ∅ ⇒ P (A ∪ B) = P (A) + P (B)

(e) Monotonicity: A, B ∈ A, A ⊂ B ⇒ P (A) ≤ P (B) and P (B − A) = P (B) − P (A)

(f) P (Ac ) = 1 − P (A)

(g) P (A) + P (B) = P (A ∪ B) + P (A ∩ B) = P (A − B) + 2P (A ∩ B) + P (B − A).

P (A ∪ B) = P (A) + P (B) − P (A ∩ B).

(h) P (A ∪ B) ≥ max(P (A), P (B)) ≥ min(P (A), P (B)) ≥ P (A ∩ B)

(i) Inclusion-exclusion principle:

 n 
   
P Ak = P (Ai ) − P (Ai ∩ Aj ) + P (Ai ∩ Aj ∩ Ak )+
k=1 i i<j i<j<k
· · · + (−1)n+1 P (A1 ∩ · · · ∩ An )

In a more compact form,

+n , + ,
  
P Ai = (−1)|I|+1 P Ai .
i=1 φ=I⊂{1,...,n} i∈I

31
 For example,

P (A1 ∪ A2 ∪ A3 ) = P (A1 ) + P (A2 ) + P (A3 )

− P (A1 ∩ A2 ) − P (A1 ∩ A3 ) − P (A2 ∩ A3 )
+ P (A1 ∩ A2 ∩ A3 ) .

See also (8.2). Moreover, for any event B, we have

+ n , + ,
  
P Ack ∩ B = P (B) + (−1)|I| P Ai ∩ B . (9)
k=1 ∅=I⊂[n] i∈I


n 
n
(j) Finite additivity: If A = Aj with Aj ∈ A disjoint, then P (A) = P (Aj )
j=1 j=1

• If A and B are disjoint sets in A, then P (A ∪ B) = P (A) + P (B)

(k) Subadditivity  or Boole’s

 Inequality: If A1 , . . . , An are events, not necessarily
n 
n
disjoint, then P Ai ≤ P (Ai )
i=1 i=1

(l) σ-subadditivity: 1 , A2 , . . . is a sequence of measurable sets, not necessarily

 ∞If A
 
∞
disjoint, then P Ai ≤ P (Ai )
i=1 i=1

• This formula is known as the union bound in engineering.

• If 
A1 , A2 ,. . .  α of events, not necessarily disjoint, then ∀α ∈ (0, 1],
is a sequence

∞ 
∞
P Ai ≤ P (Ai )
i=1 i=1

∞from above. If B1 ⊃ B2 ⊃ B3 ⊃ · · · is a decreasing sequence

3.17. Conditional continuity
of measurable sets, then P ( i=1 Bi ) = lim P (Bj ). In a more compact notation, if Bi  B,
j→∞
then P (B) = lim P (Bj ).
j→∞


∞
Proof. Let B = Bi . LetAk = Bk \Bk+1 , i.e, the new part. We consider two partitions of
i=1

∞ 
n−1 
∞
B1 : (1) B1 = B ∪ Aj and (2) B1 = Bn ∪ Aj . (1) implies P (B1 ) − P (B) = P (Aj ).
j=1 j=1 j=1

n−1
(2) implies P (B1 ) − P (Bn ) = P (Aj ). We then have
j=1

(1) 
∞
(2)
lim (P (B1 ) − P (Bn )) = P (Aj ) = P (B1 ) − P (B) .
n→∞
j=1

32
 3.18. Let A be a σ-algebra. Suppose that P : A → [0, 1] satisfies (P1) and is finitely
additive. Then, the following are equivalent:
∞ 
 
∞
• (P2): If An ∈ A, disjoint, then P An = P (An )
n=1 n=1

• (K4) If An ∈ A, and An  ∅, then P (An )  0

• (Continuity from above) If An ∈ A, and An  A, then P (An )  P (A)
• If An ∈ A, and An Ω, then P (An ) P (Ω) = 1
• (Continuity from below) If An ∈ A, and An A, then P (An ) P (A)
Hence, a probability measure satisfies all five properties above. Continuity from above
and continuity from below are collectively called sequential continuity properties.
In fact, for probability measure P , let An be a sequence of events in A which converges
to A (i.e. An → A). Then A ∈ A and lim P (An ) = P (A). Of course, both An  A and
n→∞
An A imply An → A. Note also that
+∞ , +N ,
 
P An = lim P An ,
N →∞
n=1 n=1

and + , + ,

∞ 
N
P An = lim P An .
N →∞
n=1 n=1
Alternative form of the sequential continuity properties are
+∞ ,

P An = lim P (AN ) , if An ⊂ An+1
N →∞
n=1

and + ,

∞
P An = lim P (AN ) , if An+1 ⊂ An .
N →∞
n=1

3.19. Given a common event algebra A , probability measures P1 , . . ., Pm , and the numbers

m 
m
λ1 , . . ., λm , λi ≥ 0, λi = 1, a convex combination P = λi Pi of probability measures is
1 1
a probability measure
3.20. A can not contain an uncountable, disjoint collection of sets of positive probability.
Definition 3.21. Discrete probability measure P is a discrete probability measure
if ∃ finitelyor countably
 many points ωk and nonnegative masses mk such that ∀A ∈ A
P (A) = mk = mk IA (ωk )
k:ωk ∈A k
If there is just one of these points, say ω0 , with mass m0 = 1, then P is a unit mass
at ω0 . In this case, ∀A ∈ A, P (A) = IA (ω0 ).
Notation: P = δω0
• Here, Ω can be uncountable.

33
3.4 Countable Ω
A sample space Ω is countable if it is either finite or countably infinite. It is countably
infinite if it has as many elements as there are integers. In either case, the element of Ω
can be enumerated as, say, ω1 , ω2 , . . . . If the event algebra A contains each singleton set
{ωk } (from which it follows that A is the power set of Ω), then we specify probabilities
satisfying the Kolmogorov axioms through a restriction to the set S = {{ωk }} of singleton
events.

Definition 3.22. When Ω is countable, a probability mass function (pmf) is any

function p : Ω → [0, 1] such that 
p(ω) = 1.
ω∈Ω

When the elements of Ω are enumerated, then it is common to abbreviate p(ωi ) = pi .

3.23. Every pmf p defines a probability measure P and conversely. Their relationship is
given by
p(ω) = P ({ω}), (10)


P (A) = p(ω). (11)
ω∈A

The convenience of a specification by pmf becomes clear when Ω is a finite set of, say, n
elements. Specifying P requires specifying 2n values, one for each event in A, and doing so
in a manner that is consistent with the Kolmogorov axioms. However, specifying p requires
only providing n values, one for each element of Ω, satisfying the simple constraints of
nonnegativity and addition to 1. The probability measure P satisfying (11) automatically
satisfies the Kolmogorov axioms.

3.5 Independence
Definition 3.24. Independence between events and collections of events.

(a) Two events A, B are called independent if P (A ∩ B) = P (A) P (B)

(i) An event with probability 0 or 1 is independent of any event (including itself).

In particular, ∅ and Ω are independent of any events.
(ii) Two events A, B with positive probabilities are independent if and only if
P (B |A) = P (B), which is equivalent to P (A |B ) = P (A)
When A and/or B has zero probability, A and B are automatically independent.
(iii) An event A is independent of itself if and only if P (A) is 0 or 1.
(iv) If A an B are independent, then the two classes σ ({A}) = {∅, A, Ac , Ω} and
σ ({B}) = {∅, B, B c , Ω} are independent.
(v) Suppose A and B are disjoint. A and B are independent if and only if P (A) = 0
or P (B) = 0.

34
 P (B)
(vi) Suppose A ⊂ B. A and B are independent if and only if P (A) = P (B)+1
.

(b) Independence for finite collection {A1 , . . . , An } of sets:

+ ,
 
≡ P Aj = P (Aj ) ∀J ⊂ [n] and |J| ≥ 2
j∈J j∈J

◦ Note that the case when j = 1 automatically holds. The case when j = 0
can be regard as the ∅ event case, which is also trivially true.
n  
◦ There are n
j
= 2n − 1 − n constraints.
j=2
◦ Example: A1 , A2 , A3 are independent if and only if

P (A1 ∩ A2 ∩ A3 ) = P (A1 ) P (A2 ) P (A3 )

P (A1 ∩ A2 ) = P (A1 ) P (A2 )
P (A1 ∩ A3 ) = P (A1 ) P (A3 )
P (A2 ∩ A3 ) = P (A2 ) P (A3 )

Remark : The first equality alone is not enough for independence. See a
counter example below. In fact, it is possible for the first equation to hold
while the last three fail as shown in (3.26.b). It is also possible to construct
events such that the last three equations hold (pairwise independence), but
the first one does not as demonstrated in (3.26.a).
≡ P (B1 ∩ B2 ∩ · · · ∩ Bn ) = P (B1 ) P (B2 ) · · · P (Bn ) where Bi = Ai or Bi = Ω

(c) Independence for collection {Aα : α ∈ I} of sets:

 
 
≡ ∀ finite J ⊂ I, P Aα = P (A)
α∈J α∈J
≡ Each of the finite subcollection is independent.

(d) Independence for finite collection {A1 , . . . , An } of classes:

≡ the finite collection of sets A1 , . . . , An is independent where Ai ∈ Ai .

≡ P (B1 ∩ B2 ∩ · · · ∩ Bn ) = P (B1 ) P (B2 ) · · · P (Bn ) where Bi ∈ Ai or Bi = Ω
≡ P (B1 ∩ B2 ∩ · · · ∩ Bn ) = P (B1 ) P (B2 ) · · · P (Bn ) where Bi ∈ Ai ∪ {Ω}
≡ ∀i ∀Bi ⊂ Ai B1 , . . . , Bn are independent.
≡ A1 ∪ {Ω} , . . . , An ∪ {Ω} are independent.
≡ A1 ∪ {∅} , . . . , An ∪ {∅} are independent.

(e) Independence for collection {Aθ : θ ∈ Θ} of classes:

≡ Any collection {Aθ : θ ∈ Θ} of sets is independent where Aθ ∈ Aθ

≡ Any finite subcollection of classes is independent.
 
 
≡ ∀ finite Λ ⊂ Θ, P Aθ = P (Aθ )
θ∈Λ θ∈Λ

35
 • By definition, a subcollection of independent events is also independent.
• The class {∅, Ω} is independent from any class.

Definition 3.25. A collection of events {Aα } is called pairwise independent if for every
distinct events Aα1 , Aα2 , we have P (Aα1 ∩ Aα2 ) = P (Aα1 ) P (Aα2 )

• If a collection of events {A:α : α ∈ I} is independent, then it is pairwise independent.

The converse is false. See (a) in example (3.26).
   
   
• For K ⊂ J, P Aα = P (A) does not imply P Aα = P (A)
α∈J α∈J α∈K α∈K

Example 3.26.

(a) Let Ω = {1, 2, 3, 4}, A = 2Ω , P (i) = 14 , A1 = {1, 2}, A2 = {1, 3}, A3 = {2, 3}. Then
P (Ai ∩ Aj ) = P (Ai ) P (Aj ) for all i = j butP (A1 ∩ A2 ∩ A3 ) = P (A1 ) P (A2 ) P (A3 )

(b) Let Ω = {1, 2, 3, 4, 5, 6}, A = 2Ω , P (i) = 16 , A1 = {1, 2, 3, 4}, A2 = A3 = {4, 5, 6}.

Then, P (A1 ∩ A2 ∩ A3 ) = P (A1 ) P (A2 ) P (A3 ) but P (Ai ∩ Aj ) = P (Ai ) P (Aj ) for
all i = j

(c) The paradox of ”almost sure” events: Consider two random events with probabilities
of 99% and 99.99%, respectively. One could say that the two probabilities are nearly
the same, both events are almost sure to occur. Nevertheless the difference may
become significant in certain cases. Consider, for instance, independent events which
may occur on any day of the year with probability p = 99%; then the probability
P that it will occur every day of the year is less than 3%, while if p = 99.99% then
P = 97%.

4 Random Element
4.1. A function X : Ω → E is said to be a random element of E if and only if
(Ω,A) (E,BE )
X is measurable which is equivalent to each of the following statements.

≡ X −1 (BE ) ⊂ A

≡ σ (X) ⊂ A

≡ (reduced form) ∃C ⊂ BE such that σ (C) = BE and X −1 (C) ⊂ A

• When E ⊂ R, X is called a random variable.

• When E ⊂ Rd , then X is called a random vector.

Definition 4.2. X = Y almost surely (a.s.) if P [X = Y ] = 1.

• The a.s. equality is an equivalence relation.

36
4.3. Law of X or Distribution of X: P X = μX = P X −1 = L(X) : E → [0, 1]
(E,E)

 
μX (A) = P X (A) = P X −1 (A) = P ◦ X −1 (A)
= P ({ω : X (ω) ∈ A}) = P ({X ∈ A})

4.4. For X ∈ Lp , lim tp P [|X| ≥ t] → 0

t→∞

4.5. A Borel set S is called a support of X if P X (S c ) = 0 (or equivalently P X (S) = 1)

4.1 Random Variable

Definition 4.6. A real-valued function X(ω) defined for points ω in a sample space Ω is
called a random variable.

• Random variables are important because they provide a compact way of referring to
events via their numerical attributes.

• The abbreviation r.v. will be used for “real-valued random variables” [10, p 1].

• Technically, a random variable must be measurable.

4.7. At a certain point in most probability courses, the sample space is rarely mentioned
anymore and we work directly with random variables. The sample space often “disappears”
but it is really there in the background.

4.8. For B ∈ R, we use the shorthand

• [X ∈ B] = {ω ∈ Ω : X(ω) ∈ B} and

• P [X ∈ B] = P ([X ∈ B]) = P ({ω ∈ Ω : X(ω) ∈ B}) .

• In particular, P [X < x] is a shorthand for P ({ω ∈ Ω : X(ω) < x}).

4.9. If X and Y are random variables, we use the shorthand

• [X ∈ B, Y ∈ C] = {ω ∈ Ω : X(ω) ∈ B and Y (ω) ∈ C} = [X ∈ B] ∩ [Y ∈ C].

• P [X ∈ B, Y ∈ C] = P ([X ∈ B] ∩ [Y ∈ C]) .

4.10. Every random variable can be written as a sum of a discrete random variable and
a continuous random variable.

4.11. A random variable can have at most countably many point x such thatP [X = x] > 0.

4.12. Point masses probability measures / Direc measures, usually written εα , δα , is used
to denote point mass of size one at the point α. In this case,

• P X {α} = 1

• P X ({α}c ) = 0

37
 • FX (x) = 1[α,∞) (x)

4.13. There exists distributions that are neither discrete nor continuous.
Let μ (A) = 12 μ1 (A) + 12 μ2 (A) for μ1 discrete and μ2 coming from a density.

4.14. When X and Y take finitely many values, say x1 , . . . , xm and y1 , . . . , yn , respectively,
we can arrange the probabilities pX,Y (xi , yj ) in the m × n matrix
⎡ ⎤
pX,Y (x1 , y1 ) pX,Y (x1 , y2 ) . . . pX,Y (x1 , yn )
⎢ pX,Y (x2 , y1 ) pX,Y (x2 , y2 ) . . . pX,Y (x2 , yn ) ⎥
⎢ ⎥
⎢ .. .. . .. ⎥.
⎣ . . . . . ⎦
pX,Y (xm , y1 ) pX,Y (xm , y2 ) . . . pX,Y (xm , yn )

• The sum of the entries in the ith row is PX (xi ), and the sum of the entries in the jth
column is PY (yj ).
• The sum of all the entries in the matrix is one.

4.2 Distribution Function

4.15. The (cumulative) distribution function (cdf ) induced by a probability P on
R , B is the function F (x) = P (−∞, x].
=Ω
The (cumulative) distribution function (cdf ) of the random variable X is the func-
tion FX (x) = P X (−∞, x] = P [X ≤ x].

• The distribution P X can be obtained from the distribution function by setting P X (−∞, x] =
FX (x); that is FX uniquely determines P X

• 0 ≤ FX ≤ 1

C1 FX is non-decreasing

C2 FX is right continuous:
 
∀x FX x+ ≡ y→x
lim FX (y) ≡ lim FX (y) = FX (x) = P [X ≤ x]
y x
y>x

◦ The function g(x) = P [X < x] is left-continuous in x.

C3 lim FX (x) = 0 and lim FX (x) = 1.

x→−∞ x→∞

• ∀x FX (x− ) ≡ y→x
lim FX (y) ≡ lim FX (y) = P X (−∞, x) = P [X < x]
yx
y<x

• P [X = x] = P X {x} = F (x) − F (x− ) = the jump or saltus in F at x

38
 Figure 10: Right-continuous function at jump point

• ∀x<y
P ((x, y]) = F (y) − F (x)
 
P ([x, y]) = F (y) − F x−
   
P ([x, y)) = F y − − F x−
 
P ((x, y)) = F y − − F (x)
 
P ({x}) = F (x) − F x−

• A function F is the distribution function of some probability measure on (R, B) if

and only if one has (C1), (C2), and (C3).

• If F satisfies (C1), (C2), and (C3), then ∃ a unique probability measure P on (R, B)
that has P (a, b] = F (b) − F (a) ∀a, b ∈ R

• FX is continuous if and only if P [X = x] = 0

• FX is continuous if and only if P X is continuous.

• FX has at most countably many points of discontinuity.

Definition 4.16. It is traditional to write X ∼ F to indicate that “X has distribution F ”

[21, p 25].

Definition 4.17. FX,A (x) = P ([X ≤ x] ∩ A)

4.18. Left-continuous inverse: g −1 (y) = inf {x ∈ R : g (x) ≥ y}, y ∈ (0, 1)

• Trick : Just flip the graph along the line x = y, then make the graph left-continuous.

• If g is a cdf, then only consider y ∈ (0, 1). It is called the inverse CDF [7, def 8.4.1
p. 238] or quantile function.

◦ In [21, Def 2.16 p 25], the inverse CDF is defined using strict inequality “>”
rather than “≥”.

• See table 1 for examples.

39
 g 1  y  inf ^ x  \ : g  x  t y`
g  x

y6
x4
y5
y4 x3
y3
y2
y1
x2
x2 x3 x4 x y1 y3 y4 y6 y

Figure 11: Left-continuous inverse on (0,1)

Distribution F F −1
Exponential 1 − e−λx − λ1 ln (u)
x−a
Extreme value 1 − e−e b
a 3+ b ln ln4u
Geometric 1 − (1 − p)i ln u

ln(1−p)

Logistic 1− 1
x−μ μ− b ln u1 − 1
1+e b
1
Pareto 1 − x−a u− a
x b 1
Weibull 1 − e( a ) a (ln u) b

Table 1: Left-continuous inverse

40
Definition 4.19. Let X be a random variable with distribution function F . Suppose that
p ∈ (0, 1). A value of x such that F (x− ) = P [X < x] ≤ p and F (x) = P [X ≤ x] ≥ p is
called a quantile of order p for the distribution. Roughly speaking, a quantile of order p
is a value where the cumulative distribution crosses p. Note that it is not unique. Suppose
F (x) = p on an interval [a, b], then all x ∈ [a, b] are quantile of order p.
A quantile of order 12 is called a median of the distribution. When there is only one
median, it is frequently used as a measure of the center of the distribution. A quantile
of order 14 is called a first quartile and the quantile of order 34 is called a third quartile. A
median is a second quartile.
Assuming uniqueness, let q1 , q2 , and q3 denote the first, second, and third quartiles
of X. The interquartile range is defined to be q3 − q1 , and is sometimes used as a mea-
sure of the spread of the distribution with respect to the median. The five parameters
max X, q1 , q2 , q3 , min X are often referred to as the five-number summary. Graphically,
the five numbers are often displayed as a boxplot.
4.20. If F is non-decreasing, right continuous, with lim F (x) = 0 and lim F (x) = 1,
x→−∞ x→∞
then F is the CDF of some probability measure on (R, B).
In particular, let U ∼ U (0, 1) and X = F −1 (U ), then FX = F . Here, F −1 is the
left-continuous inverse of F . Note that we just explicitly define a random variable X(ω)
with distribution function F on Ω = (0, 1).

• To generate X ∼ E (λ), set X = − λ1 ln (U )

4.3 Discrete random variable

Definition 4.21. A random variable X is said to be a discrete random variable if
there exists countable distinct real numbers xk such that

P [X = xk ] = 1.
k

≡ ∃ countable set {x1 , x2 , . . .} such that μX ({x1 , x2 , . . .}) = 1

≡ X has a countable support {x1 , x2 , . . .}

⇒ X is completely determined by the values μX ({x1 }) , μX ({x2 }) , . . .

• pi = pX (xi ) = P [X = xi ]

Definition 4.22. When X is a discrete random variable taking distinct values xk , we

define its probability mass function (pmf) by

pX (xk ) = P [X = xk ].

• We can use stem plot to visualize pX .

• If Ω is countable, then there can be only countably many value of X(ω). So, any
random variable defined on countable Ω is discrete.

41
 • Sometimes, we write p(xk ) or pxk in stead of pX (xk ).

• P [X ∈ B] = xk ∈B P [X = xk ].

• FX (x) = xk pX (xk )U (x − xk ).

Definition
 4.23 (Discrete CDF). A cdf which can be written in the form Fd (x) =
k pi U (x − xk ) is called a discrete cdf [7, Def. 5.4.1 p 163]. Here, U is the unit step
function, {xk } is an arbitrary countable set of real numbers, and {pi } is a countable set of
positive numbers that sum to 1.

Definition 4.24. An integer-valued random variable is a discrete random variable whose

distinct values are xk = k.
For integer-valued random variables,

P [X ∈ B] = P [X = k].
k∈B

4.25. Properties of pmf

• p : Ω → [0, 1].

• 0 ≤ pX ≤ 1.

• k pX (xk ) = 1.

Definition 4.26. Sometimes, it is convenient to work with the “pdf” of a discrete r.v.
Given that X is a discrete random variable which is defined as in definition 4.22. Then,
the “pdf” of X is 
fX (x) = pX (xk )δ(x − xk ), x ∈ R. (12)
xk

Although the delta function is not a well-defined function3 , this technique does allow easy
manipulation of mixed distribution. The definition of quantities involving discrete random
variables and the corresponding properties can then be derived from the pdf and hence
there is no need to talk about pmf at all!

4.4 Continuous random variable

Definition 4.27. A random variable X is said to be a continuous random variable if
and only if any one of the following equivalent conditions holds.

≡ ∀x, P [X = x] = 0

≡ ∀ countable set C, P X (C) = 0

≡ FX is continuous
3
Rigorously, it is a unit measure at 0.

42
 4.28. f is (probability) density function f (with respect to Lebesgue measure) of a
random variable X (or the distribution P X )
≡ P X have density f with respect to Lebesgue measure.

≡ P X is absolutely continuous w.r.t. the Lebesgue measure (P X  λ) with f =

dP X
dλ
, the Radon-Nikodym derivative.

≡ f is a nonnegative Borel function on R such that ∀B ∈ BR PX (B) = B f (x)dx =
f dλ where λ is the Lebesgue measure. (This extends nicely to the random vector
B
case.)

≡ X is absolutely continuous

≡ X (or F X ) comes from the density f

x
≡ ∀x ∈ R FX (x) = f (t)dt
−∞

b
≡ ∀a, b FX (b) − FX (a) = f (x)dx
a

4.29. If F does differentiate to f and f is continuous, it follows by the fundamental

theorem of calculus that f is indeed a density for F . That is, if F has a continuous
derivative, this derivative can serve as the density f .
4.30. Suppose a random variable X has a density f .
• F need not differentiate to f everywhere.

◦ When X ∼ U(a, b), FX is not differentiable at a nor b.


• f (x) dx = 1
Ω

• f is determined only Lebesgue-a.e. That is, If g = f Lebesgue-a.e., then g can also

serve as a density for X and PX

• f is nonnegative a.e. [9, stated on p 138]

• X is a continuous random variable

• f at its continuity points must be the derivative of F

• P [X ∈ [a, b]] = P [X ∈ [a, b)] = P [X ∈ (a, b]] = P [X ∈ (a, b)] because the corre-
sponding integrals over an interval are not affected by whether or not the endpoints
are included or excluded. In other words, P [X = a] = P [X = b] = 0.

• P [fx (X) = 0] = 0
4.31. fX (x) = E [δ (X − x)]

43
Definition 4.32 (Absolutely Continuous CDF). An absolutely continuous cdf Fac can
be written in the form ! x
Fac (x) = f (z)dz,
−∞

where the integrand,

d
Fac (x),
f (x) =
dx
is defined a.e., and is a nonnegative, integrable function (possibly having discontinuities)
satisfying !
f (x)dx = 1.

4.33. Any nonnegative function that integrates to one is a probability density function
(pdf) [9, p 139].
4.34. Remarks: Some useful intuitions
 x+Δx
(a) Approximately, for a small Δx, P [X ∈ [x, x + Δx]] = x
fX (t)dt ≈ f − X(x)Δx.
This is why we call fX the density function.
P [x<X≤x+Δx]
(b) In fact, fX (x) = lim Δx
Δx→0

4.35. Let T be an absolutely continuous nonnegative random variable with cumulative

distribution function F and density f on the interval [0, ∞). The following terms are often
used when T denotes the lieftime of a device or system.
(a) Its survival-, survivor-, or reliability-function is:
!∞
R (t) = P [T > t] = f (x)dx = 1 − F (t) .
t

• R(0) = P [T > 0] = P [T ≥ 0] = 1.
∞
(b) The mean time of failure (MTTF) = E [T ] = 0
R(t)dt.

(c) The (age-specific) failure rate or hazard function os a device or system with lifetime
T is
P [T ≤ t + δ|T > t] R (t) f (t) d
r (t) = lim =− = = ln R(t).
δ→0 δ R(t) R (t) dt
(i) r (t) δ ≈ P [T ∈ (t, t + δ] |T > t]
t
(ii) R(t) = e− 0 r(τ )dτ
.
t
(iii) f (t) = r(t)e− 0 r(τ )dτ

• For T ∼ E(λ), r(t) = λ.

See also [9, section 5.7].

44
Definition 4.36. A random variable whose cdf is continuous but whose derivative is the
zero function is said to be singular.

• See Cantor-type distribution in [5, p 35–36].

• It has no density. (Otherwise, the cdf is the zero function.) So, ∃ continuous random
variable X with no density. Hence, ∃ random variable X with no density.
• Even when we allow the use of delta function for the density as in the case of mixed
r.v., it still has no density because there is no jump in the cdf.
• There exists singular r.v. whose cdf is strictly increasing.
d
Definition 4.37. fX,A (x) = F
dx X,A
(x). See also definition 4.17.

4.5 Mixed/hybrid Distribution

There are many occasion where we have a random variable whose distribution is a combina-
tion of a normalized linear combination of discrete and absolutely continuous distributions.
For convenience, we use the Dirac delta function to link the pmf to a pdf as in definition
4.26. Then, we only have to concern about the pdf of a mixed distribution/r.v.

4.38. By allowing density functions tocontain impulses, the cdfs of mixed random variables
can be expressed in the form F (x) = (−∞,x] f (t)dt.

4.39. Given a cdf FX of a mixed random variable X, the density fX is given by


fX (x) = f˜X (x) + P [X = xk ]δ(x − xk ),
k

where

• the xi are the distinct points at which FX has jump discontinuities, and

FX (x) , FX is differentiable at x
• f˜X (x) =
0, otherwise.
In which case, ! 
E [g(X)] = g(x)f˜X (x)dx + g(xk )P [X = xk ].
k

Note also that P [X = xk ] = F (xk ) − F (x−

k)

4.40. Suppose the cdf F can be expressed in the form F (x) = G(x)U (x − x0 ) for some
function G. Then, the density is f (x) = G (x)U (x − x0 ) + G(x0 )δ(x − x0 ). Note that
G(x0 ) = F (x0 ) = P [X = x0 ] is the jump of the cdf at x0 . When the random variable is
continuous, G(x0 ) = 0 and thus f (x) = G (x)U (x − x0 ).

45
4.6 Independence
Definition 4.41. A family of random variables {Xi : i ∈ I} is independent if ∀ finite
J ⊂ I, the family of random variables {Xi : i ∈ J} is independent. In words, “an infinite
collection of random elements is by definition independent if each finite subcollection is.”
Hence, we only need to know how to test independence for finite collection.

(a) (Ei , Ei )’s are not required to be the same.

(b) The collection of random variables {1Ai : i ∈ I} is independent iff the collection of

events (sets) {Ai : i ∈ I} is independent.

Definition 4.42. Independence among finite collection of random variables: For finite I,
the following statements are equivalent

≡ (Xi )i∈I are independent (or mutually independent [2, p 182]).


≡ P [Xi ∈ Hi ∀i ∈ I] = P [Xi ∈ Hi ]where Hi ∈ Ei
i∈I
5 6

≡ P (Xi : i ∈ I) ∈ × Hi = P [Xi ∈ Hi ] where Hi ∈ Ei
i∈I i∈I
 

≡ P (Xi :i∈I) × Hi = P Xi (Hi ) where Hi ∈ Ei
i∈I i∈I

≡ P [Xi ≤ xi ∀i ∈ I] = P [Xi ≤ xi ]
i∈I

≡ [Factorization Criterion] F(Xi :i∈I) ((xi : i ∈ I)) = FXi (xi )
i∈I

≡ Xi and X1i−1 are independent ∀i ≥ 2

 
≡ σ (Xi ) and σ X1i−1 are independent ∀i ≥ 2
≡ Discrete random variables Xi ’s with countable range E:

P [Xi = xi ∀i ∈ I] = P [Xi = xi ] ∀xi ∈ E ∀i ∈ I
i∈I


≡ Absolutely continuous Xi with density fXi : f(Xi :i∈I) ((xi : i ∈ I)) = fXi (xi )
i∈I

Definition 4.43. If the Xα , α ∈ I are independent and each has the same marginal
distribution with distribution Q, we say that the Xα ’s are iid (independent and identically
iid
distributed) and we write Xα ∼ Q

• The abbreviation can be IID [21, p 39].

Definition 4.44. A pairwise independent collection of random variables is a set of

random variables any two of which are independent.

(a) Any collection of (mutually) independent random variables is pairwise independent

46
 (b) Some pairwise independent collections are not independent. See example (4.45).
Example 4.45. Let suppose X, Y , and Z have the following joint probability distribution:
pX,Y,Z (x, y, z) = 14 for (x, y, z) ∈ {(0, 0, 0), (0, 1, , 1), (1, 0, 1), (1, 1, 0)}. This, for example,
can be constructed by starting with independent X and Y that are Bernoulli- 12 . Then set
Z = X ⊕ Y = X + Y mod 2.
(a) X, Y, Z are pairwise independent.
(b) The combination of X Z and Y Z does not imply (X, Y ) Z.

|=

|=

|=
Definition 4.46. The convolution of probability measures μ1 and μ2 on (R, B) is the
measure μ1 ∗ μ2 defined by
!
(μ1 ∗ μ2 ) (H) = μ2 (H − x)μ1 (dx) H ∈ BR

(a) μX ∗ μY = μY ∗ μX and μX ∗ (μY ∗ μZ ) = (μX ∗ μY ) ∗ μZ

(b) If FX and GX are distribution functions corresponding to μX and μY , the distribution
function corresponding to μX ∗ μY is
!
(μX ∗ μY ) (−∞, z] = FY (z − x)μX (dx)

In this case, it is notationally convenient to replace μX (dx) by dFX (x) (Stieltjes

Integral.) Then, (μX ∗ μY ) (−∞, z] = FY (z − x)dFX (x). This is denoted by
(FX ∗ FY ) (z). That is
!
FZ (z) = (FX ∗ FY ) (z) = FY (z − x)dFX (x)

(c) If densityfY exists, FX ∗ FY has density FX ∗ fY , where

!
(FX ∗ fY ) (z) = fY (z − x)dFX (x)

(i) If Y (or FY ) is absolutely continuous with density fY , then for any X (or
FX ), X + Y (or FX ∗ FY ) is absolutely continuous with density (FX ∗ fY ) (z) =
fY (z − x)dFX (x)
If, in addition, FX has density fX . Then,
! !
fY (z − x)dFX (x) = fY (z − x)fX (x) dx

This is denoted by fX ∗ fY
In other words, if densities fX , fY exist, then FX ∗ FY has density fX ∗ fY , where
!
(fX ∗ fY ) (z) = fY (z − x)fX (x) dx

47
 4.47. If random variables X and Y are independent and have distribution μX and μY ,
then X + Y has distribution μX ∗ μY

4.48. Expectation and independence

(a) Let X and Y be nonnegative independent random variables on (Ω, A, P ), then E [XY ] =
EXEY

(b) If X1 , X2 , . . . , Xn are independent and gk ’s complex-valued measurable5function. Then

6
n
gk (Xk )’s are independent. Moreover, if gk (Xk ) is integrable, then E gk (Xk ) =
k=1

n
E [gk (Xk )]
k=1


n
(c) If X1 , . . . , Xn are independent and Xk has a finite second moment, then all Xk
k=1
have finite second 6
5 n moments as well.
  n
Moreover, Var Xk = Var Xk .
k=1 k=1
5 6

k 
k
(d) If pairwise independent Xi ∈ L , then Var
2
Xi = Var [Xi ]
i=1 i=1

4.7 Misc
4.49. The mode of a discrete probability distribution is the value at which its probability
mass function takes its maximum value. The mode of a continuous probability distribution
is the value at which its probability density function attains its maximum value.

• the mode is not necessarily unique, since the probability mass function or probability
density function may achieve its maximum value at several points.

5 PMF Examples
The following pmf will be defined on its support S. For Ω larger than S, we will simply
put the pmf to be 0.

5.1 Random/Uniform
5.1. Rn , Un
When an experiment results in a finite number of “equally likely” or “totally random”
outcomes, we model it with a uniform random variable. We say that X is uniformly
distributed on [n] if
1
P [X = k] = , k ∈ [n].
n
We write X ∼ Un .

48
 X∼ Support set X pX (k) ϕX (u)
Uniform Un {1, 2, . . . , n} 1
n
1−eiun
U{0,1,...,n−1} {0, 1, . . . , n − 1} 1
n n(1−eiu )
1 − p, k = 0
Bernoulli B(1, p) {0, 1}
np,
 k k=1
n
Binomial B(n, p) {0, 1, . . . , n} k
p (1 − p)n−k (1 − p + peju )
Geometric G(p) N ∪ {0} (1 − p)pk 1−β
1−βeiu
Geometric G  (p) N (1 − p)k−1 p
eλ(e )
iu −1
e−λ λk!
k
Poisson P(λ) N ∪ {0}

Table 2: Examples of probability mass functions. Here, p, β ∈ (0, 1). λ > 0.

• pi = 1
n
for i ∈ S = {1, 2, . . . , n}.

• Examples

◦ classical game of chance / classical probability drawing at random

◦ fair gaming devices (well-balanced coins and dice, well shuffled decks of cards)
◦ experiment where
∗ there are only n possible outcomes and they are all equally probable
∗ there is a balance of information about outcomes

5.2. Uniform on a finite set: U(S)

Suppose |S| = n, then p(x) = n1 for all x ∈ S.

Example 5.3. For X uniform on [-M:1:M], we have EX = 0 and Var X = M (M3 +1) .
For X uniform on [N:1:M], we have EX = M −N
2
1
and Var X = 12 (M − N )(M − N − 2).
Example 5.4. Set S = 0, 1, 2, . . . , M , then the sum of two independent U(S) has pmf
(M + 1) − |k − M |
p(k) =
(M + 1)2
1
for k = 0, 1, . . . , 2M . Note its triangular shape with maximum value at p(M ) = M +1
. To
visualize the pmf in MATLAB, try
k = 0:2*M;
P = (1/((M+1)^2))*ones(1,M+1);
P = conv(P,P); stem(k,P)

5.2 Bernoulli and Binary distributions

5.5. Bernoulli : B(1, p) or Bernoulli(p)
• S = {0, 1}

49
 • p0 = q = 1 − p, p1 = p

• EX = E [X 2 ] = p.
Var [X] = p − p2 = p (1 − p). Note that the variance is maximized at p = 0.5.

0.25

0.2

p ˜ ( 1 p )

0.1

0
0
0 0.2 0.4 0.6 0.8 1
0 p 1

Figure 12: The plot for p(1 − p).

5.6. Binary Suppose X takes only two values a and b with b > a. P [X = b] = p. Then,
X can be expressed as X = (b − a) I + a, where I is a Bernoulli random variable with
P [I = 1] = p.

• Var X = (b − a)2 V arI = (b − a)2 p (1 − p) . Note that it is still maximized at p = 1/2.

• Suppose a = −b. Then, X = 2I + a = 2I − b. In which case, Var X = 2b2 p (1 − p).

5.3 Binomial: B(n, p)

• Binomial distribution with size n and parameter p. p ∈ [0, 1].
 
• pi = ni pi (1 − p)n−i for i ∈ S = {0, 1, 2, . . . , n}

• X is the number of success in n independent Bernoulli trials and hence the sum of n
independent, identically distributed Bernoulli r.v.
n
• ϕX (u) = (1 − p + peju )

• EX = np

• EX 2 = (np)2 + np (1 − p)

• Var [X] = np (1 − p)
n  

• Tail probability: n
r
pr (1 − p)n−r = Ip (k, n − k + 1)
r=k

50
 • Maximum probability value happens at kmax = mode X = (n + 1) p ≈ np
◦ When (n+1)p is an integer, then the maximum is achieved at kmax and kmax −1.
• If have E1 , . . ., En , n unlinked repetition of E and event A for E, the the distribution
B(n,p) describe the probability that A occurs k times in E1 , . . . ,En .
• Gaussian Approximation for Binomial Probabilities: When n is large, binomial distri-
bution becomes difficult to compute directly because of the need to calculate factorial
terms. We can use
1 (k−np)2
P [X = k] " " e− 2np(1−p) , (13)
2πnp (1 − p)
which comes from approximating X by Gaussian Y with the same mean and variance
and the relation
P [X = k] " P [X ≤ k] − P [X ≤ k − 1]
" P [Y ≤ k] − P [Y ≤ k − 1] " fY (k).

See also (12.22).

n   
n−k (np)k −np 2
• Approximation: k
p (1 − p)
k
= k!
e 1+O np2 , kn

p  0.05 n  100 O  5 p  0.05 n  800 O  40

x x
O O O O
e ˜ e ˜
* ( x 1) 0.15 * ( x 1) 0.06
1 2 1 2
( x O ) ( x O )
1 2 ˜O 1 2 ˜O
˜e ˜e
2˜ S O 2˜ S O
0.1 0.04
 x O1  x O1
e ˜x e ˜x
* ( O) * ( O)

* ( n 1) x n x 0.05 * ( n 1) x n x 0.02
˜ p ˜ ( 1 p ) ˜ p ˜ ( 1 p )
* ( n x 1) ˜ * ( x 1) * ( n x 1) ˜ * ( x 1)

0 0
0 5 10 0 20 40 60
x x

Figure 13: Gaussian approximation to Binomial, Poisson distribution, and Gamma distri-
bution.

5.4 Geometric: G(β)

A geometric distribution is defined by the fact that for some β ∈ [0, 1), pk+1 = βpk for all
k ∈ S where S can be either N or N ∪ {0}.
• When its support is N, pk = (1 − p) pk−1 . This is referred to as G1 (β) or geometric1 (β).
In MATLAB, use geopdf(k-1,1-p).

51
 • When its support is N ∪ {0}, pk = (1 − p) pk . This is referred to as G0 (β) or
geometric0 (β). In MATLAB, use geopdf(k,1-p).
5.7. Consider X ∼ G0 (β).
• pi = (1 − β) β i , for S = N ∪ {0} , 0 ≤ β < 1
• β= m
m+1
where m = average waiting time/ lifetime

• P [X = k] = P [k failures followed by a success] = (P [failures])k P [success]

P [X ≥ k] = β k = the probability of having at least k initial failure = the probability
of having to perform at least k+1 trials.
P [X > k] = β k+1 = the probability of having at least k+1 initial failure.
• Memoryless property:
◦ P [X ≥ k + c |X ≥ k ] = P [X ≥ c], k, c > 0.
◦ P [X > k + c |X ≥ k ] = P [X > c], k, c > 0.
◦ If a success has not occurred in the first k trials (already fails for k times),
then the probability of having to perform at least j more trials is the same the
probability of initially having to perform at least jtrials.
◦ Each time a failure occurs, the system “forgets” and begins anew as if it were
performing the first trial.
◦ Geometric r.v. is the only discrete r.v. that satisfies the memoryless property.
• Ex.
◦ lifetimes of components, measured in discrete time units, when the fail catas-
trophically (without degradation due to aging)
◦ waiting times
∗ for next customer in a queue
∗ between radioactive disintegrations
∗ between photon emission
◦ number of repeated, unlinked random experiments that must be performed prior
to the first occurrence of a given event A
∗ number of coin tosses prior to the first appearance of a ‘head’
number of trials required to observe the first success
• The sum of independent G0 (p) and G0 (q) has pmf
7 k+1 −pk+1
(1 − p) (1 − q) q q−p , p = q
2 k
(k + 1) (1 − p) p , p=q
for k ∈ N ∪ {0}.
5.8. Consider X ∼ G1 (β).
• P [X > k] = β k

• Suppose independent Xi ∼ G1 (βi ). min (X1 , X2 , . . . , Xn ) ∼ G1 ( ni=1 βi ).

52
5.5 Poisson Distribution: P(λ)
5.9. Characterized by
• pX (k) = P [X = k] = e−λ λk! ; or equivalently,
k

• ϕX (u) = eλ(e ),
iu −1

where λ ∈ (0, ∞) is called the parameter or intensity parameter of the distribution.

In MATLAB, use poisspdf(k,lambda).
5.10. Denoted by P (λ).
5.11. In stead of X, Poisson random variable is usually denoted by Λ.
5.12. EX = Var X = λ.
5.13. Successive probabilities are connected via the relation kpX (k) = λpX (k − 1).
5.14. mode X = λ.
• Note that when λ ∈ N, there are two maximums at λ − 1 and λ.
• When λ # 1, pX (λ) ≈ √1
2πλ
via the Stirling’s formula (1.13).

Most probable value (imax ) Associated max probability

0<λ<1 0 e−λ
λλ −λ
λ∈N λ − 1, λ λ!
e
λλ −λ
λ ≥ 1, λ ∈
/N λ λ!
e

5.15. P [X ≥ 2] = 1 − e−λ − λe−λ = O (λ2 ) .

The cumulative probabilities can be found by
8 k+1 9 !∞
(∗)  1
P [X ≤ k] = P Xi > 1 = e−t tx dt,
i=1
Γ (k + 1)
λ
8 k+1 9 !λ
(∗)  1
P [X > k] = P [X ≥ k + 1] = P Xi ≤ 1 = e−t tx dt,
i=1
Γ (k + 1)
0

where the Xi ’s are i.i.d. E(λ). The equalities given by (*) are easily obtained via
counting the number of events from rate-λ Poisson process on interval [0, 1].
Var X
5.16. Fano factor (index of dispersion): EX
=1
An important property of the Poisson and Compound Poisson laws is that their classes
are close under convolution (independent summation). In particular, we have divisibility
properties (5.21) and (5.31) which are straightforward to prove from their characteristic
functions.

53
 
5.17
 (Recursion equations). Suppose X ∼ P(λ). Let mk (λ) = E X k and μk (λ) =
E (X − EX)k .

mk+1 (λ) = λ (mk (λ) + mk (λ)) (14)

μk+1 (λ) = λ (kμk−1 (λ) + μk (λ)) (15)

[14, p 112]. Starting with m1 = λ = μ2 and μ1 = 0, the above equations lead to recursive
determination of the moments mk and μk .
 1    
−λ d
5.18. E X+1 = λ 1 − e . Because for d ∈ N, Y = X+1
1 1
n=0 an X
n
can be expressed
 
d−1
as n=0 bn X
n c
+ X+1 , the value of EY is easy to find if we know EX n .

5.19. Mixed Poisson distribution: Let X be Poisson with mean λ. Suppose, that the
mean λ is chosen in accord with a probability distribution whose characteristic function is
ϕΛ . Then,
  iuX ( * ( *   
Λ(eiu −1) i(−i(eiu −1))Λ
ϕX (u) = E E e |Λ = E e =E e = ϕΛ −i eiu − 1 .

• EX = EΛ.
• Var X = Var Λ + EΛ.
• E [X 2 ] = E [Λ2 ] + EΛ.
• Var[X|Λ] = E [X|Λ] = Λ.
• When Λ is a nonnegative z−1
 1  integer-valued random variable, we have GX (z) = GΛ (e )
and P [X = 0] = GΛ z .
• E [XΛ] = E [Λ2 ]
• Cov [X, Λ] = Var Λ

5.20. Thinned Poisson: Suppose we have X → s → Y where X ∼ P (λ). The box s

is a binomial channel with success probability s. (Each 1 in the X get through the channel
with success probability s.)

• Note that Y is in fact a random sum X i=1 Ii where i.i.d. Ii has Bernoulli distribution
with parameter s.
• Y ∼ P (sλ);
x−y
• p (x |y ) = e−λ(1−s) (λ(1−s))
(x−y)!
; x ≥ y (shifted Poisson);

[Levy and Baxter, 2002]

n
n Finite additivity: Suppose we have independent Λi ∼ P (λi ), then
5.21. i=1 Λi ∼
P ( i=1 λi ).

5.22. Raikov’s theorem: independent random variables can have their sum Poisson-
distributed only if every component of the sum is Poisson-distributed.

54
 5.23. Countable Additivity Theorem [11, p 5]: Let (Xj : j =∈ N) be independent
random variables, and assume that Xj has the distribution P(μj ) for each j. If


∞
μj (16)
j=1


∞
converges to μ, then S = Xj converges with probability 1, and S has distribution P(μ).
j=1
If on the other hand (16) diverges, then S diverges with probability 1.

2 , . . . , Xn be independent, and let Xj havedistribution P(μj ) for all

5.24. Let X1 , X
n n
j.
n Then S n = j=1 Xj has distribution P(μ), with μ = j=1 μj ; and so, whenever
j=1 rj = s,
n  r
s! μj j
P [Xj = rj ∀j|Sn = s] =
r1 !r2 ! · · · rn ! j=1 μ

which follows the multinomial distribution [11, p 6–7].

• If X and Y are independent Poisson random variables with respective parameters

 λ
and μ, then (1) Z = X +Y is P(λ+μ) and (2) conditioned on Z = z, X is B z, λ+μ .
λ

So, E [X|Z] = λ
λ+μ
Z, λμ
Var[X|Z] = Z (λ+μ) 2 , and E [Var[X|Z]] =
λμ
λ+μ
.

5.25. One of the reasons why Poisson distribution is important is because many natural
phenomenons can be modeled by Poisson processes. For example, if we consider the number
of occurrences Λ during a time interval of length τ in a rate-λ homogeneous Poisson process,
then Λ ∼ P(λτ ).

Example 5.26.

• The first use of the Poisson model is said to have been by a Prussian physician,
von Bortkiewicz, who found that the annual number of late-19th-century Prussian
soldiers kicked to death by horses followed a Poisson distribution [7, p 150].

• #photons emitted by a light source of intensity λ [photons/second] in time τ

• #atoms of radioactive material undergoing decay in time τ

• #clicks in a Geiger counter in τ seconds when the average number of click in 1 second
is λ.

• #dopant atoms deposited to make a small device such as an FET

• #customers arriving in a queue or workstations requesting service from a file server

in time τ

• Counts of demands for telephone connections

• number of occurrences of rare events in time τ

55
 • #soldiers kicked to death by horses
• Counts of defects in a semiconductor chip.
5.27. Normal Approximation to Poisson Distribution with large λ: Let X ∼ P (λ). X

n
can be though of as a sum of i.i.d.Xi ∼ P (λn ), i.e., X = Xi , where nλn = λ. Hence X
i=1
is approximately normal N (λ, λ) for λ large.
Some says that the normal approximation is good when λ > 5.
5.28. Poisson distribution can be obtained as a limit from negative binomial distributions.
Thus, the negative binomial distribution with parameters r and p can be approximated
by the Poisson distribution with parameter λ = rqp (mean-matching), provided that p is
“sufficiently” close to 1 and r is “sufficiently” large.
5.29. Convergence of sum of bernoulli random variables to the Poisson Law
Suppose that for each n ∈ N
Xn,1 , Xn,2 , . . . , Xn,rn
are independent; the probability space for the sequence may change with n. Such a collec-
tion is called a triangular array [1] or double sequence [8] which captures the nature
of the collection when it is arranged as
⎫
X1,1 , X1,2 , . . . , X1,r1 , ⎪
⎪
X2,1 , X2,2 , . . . , X2,r2 , ⎪
⎪
⎪
⎬
.. .. ..
. . ··· .
⎪
Xn,1 , Xn,2 , . . . , Xn,rn , ⎪
⎪
⎪
.. .. .. ⎪
⎭
. . ··· .
where the random variables in each row are independent. Let Sn = Xn,1 + Xn,2 + · · · + Xn,rn
be the sum of the random variables in the nth row.
Consider a triangular array of bernoulli random variables Xn,k with P [Xn,k = 1] = pn,k .
rn
If max pn,k → 0 and pn,k → λ as n → ∞, then the sums Sn converges in distribution
1≤k≤rn k=1
to the Poisson law. In other words, Poisson distribution rare events limit of the binomial
(large n, small p).
As a simple special case, consider a triangular array of bernoulli random variables Xn,k
with P [Xn,k = 1] = pn . If npn → λ as n → ∞, then the sums Sn converges in distribution
to the Poisson law. i  
To show this special case directly, we bound the first i terms of n! to get (n−i)
i!
≤ ni ≤
ni
i!
. Using the upper bound,
 
n i 1 npn n
pn (1 − pn )n−i ≤ (npn )i (1 − pn )−i (1 − ) .
i i!        n 
→λ i →1
→e−λ
 n−i i
The lower bound gives the same limit because (n − i)i = n
ni where the first term
→ 1.

56
5.6 Compound Poisson
Given an arbitrary probability measure μ and a positive real  number λ, the compound
Poisson distribution CP (λ, μ) is the distribution of the sum Λj=1 Vj where the Vj are i.i.d.
with distribution μ and Λ is a P (λ) random variable, independent of the Vj .
Sometimes, it is written as POIS (λμ). The parameter λ is called the rate of CP (λ, μ)
and μ is called the base distribution.

5.30. The mean and variance of CP (λ, L (V )) are λEV and λEV 2 respectively.

5.31. If Z ∼ CP (λ, q), then ϕZ (t) = eλ(ϕq (t)−1) .

An important property of the Poisson and Compound Poisson laws is that their classes
are close under convolution (independent summation). In particular, we have divisibility
properties (5.21) and (5.31) which are straightforward to prove from their characteristic
functions.

5.32. Divisibility
 property
 ofthe compound
 1 Poisson
n law: Suppose we
nhave inde-
n
pendent Λi ∼ CP λi , μ , then i=1 Λi ∼ CP λ, λ i=1 λi μ
(i) (i)
where λ = i=1 λi .
Proof.
+ n ,

n 
ϕ
n (t) = eλi (ϕqi (t)−1) = exp λi (ϕqi (t) − 1)
Zi
i=1 i=1
i=1
+ n , + + ,,
 1
n
= exp λi ϕqi (t) − λ = exp λ λi ϕqi (t) − 1
i=1
λ i=1

We usually focus on the case when μ is a discrete probability measure on N = {1, 2, . . .}.
q on N; q is called the base pmf. Equivalently,
In which case, we usually refer to μ by the pmf 
CP (λ, q) is also the distribution
 of the sum i∈N iΛi where (Λi : i ∈ N) are independent
with Λi ∼ P (λqi ). Note that i∈N λqi = λ. The Poisson distribution is a special case of
the compound Poisson distribution where we set q to be the point mass at 1.

5.33. The compound negative binomial [Bower, Gerber, Hickman, Jones, and Nesbitt,
1982, Ch 11] can be approximated by the compound Poisson distribution.

5.7 Hypergeometric
An urn contains N white balls and M black balls. One draws n balls without replacement,
so n ≤ N + M . One gets X white balls and n − X black balls.
7 N M
( x )(n−x)
, 0 ≤ x ≤ N and 0 ≤ n − x ≤ M
P [X = x] = (N +M
n )
0, otherwise

57
 5.34. The hypergeometric distributions “converge” to the binomial distribution: Assume
N
that n is fixed, while N and M increase to +∞ with lim N +M = p, then
N →∞
M →∞
 
n x
p (x) → p (1 − p)n−x (binomial).
x

Note that binomial is just drawing balls with replacement:

n x n−x    x  n−x
x
N M n N M
p (x) = = .
(N + M )n x N +M N +M

Intuitively, when N and M large, there is not much difference in drawing n balls with or
without replacement.

5.35. Extension: If we have m colors and Ni balls of color i. The urn contains N =
N1 + · · · + Nm balls. One draws n balls without replacement. Call Xi the number of balls
of color i drawn among n balls. (Of course, X1 + · · · + Xm = n.)

⎧ N N
⎨ ( x11 )( x22 )···(Nxmm )
, x1 + · · · + xm = n and xi ≥ 0
P [X1 = x1 , . . . , Xm = xm ] = (Nn )
⎩ 0, otherwise

5.8 Negative Binomial Distribution (Pascal / Pólya distribution)

5.36. The probability that the rth success occurs on the (x + r)th trial

    
x+r−1 x x+r−1
=p p (1 − p) =
r−1
pr (1 − p)x
r−1 r−1
 
x+r−1
= pr (1 − p)x
x
i.e. among the first (x + r − 1) trials, there are r − 1 successes and x failures.

• Fix r.

• ϕX (u) = pr (1−(1−p)e
1
iu )r

• EX = rq
p
,V ar (X) = rq
p2
,q =1−p
n n(n−1)·····(n−(x−1))
• Note that if we define x
≡ x(x−1)·····1
. Then,
   
−r r (r + 1) · · · · · (r + (x − 1))
x x r+x−1
≡ (−1) = (−1) .
x x (x − 1) · · · · · 1 x

58
  
 
• If independent Xi ∼ NegBin (ri , p), then Xi ∼ NegBin ri , p . This is easy to
i i
see from the characteristic function.

• p (x) = Γ(r+x) r
Γ(r)x!
p (1 − p)x

5.37. A negative binomial distribution can arise

 as a mixture of Poisson distributions with
p
mean distributed as a gamma distribution Γ q = r, λ = 1−p .
Let X be Poisson with mean λ. Suppose, that the mean λ is chosen in accord with a
probability distribution FΛ (λ). Then, ϕX (u) = ϕΛ (−i (eiu − 1)) [see compound Poisson
distribution]. Here, Λ ∼ Γ (q, λ0 ); hence, ϕΛ (u) =  1 u q .l
1−i λ
   q 0
−q λ0

So, ϕX (u) = 1 − e λ−1

iu

0
= 1− λ01+1 eiu , which is negative binomial with p = λ0λ+1
0
.
λ0 +1

So, a.k.a. Poisson-gamma distribution, or simply compound Poisson distribution.

5.9 Beta-binomial distribution

A variable with a beta binomial distribution is distributed as a binomial distribution with
parameter p, where p is distribution with a beta distribution with parameters αand β.
 
• P (k |p ) = nk pk (1 − p)n−k .

• f (p) = fβq1 ,q2 (p) = Γ(q1 +q2 ) q1 −1

Γ(q1 )Γ(q2 )
p (1 − p)q2 −1 1(0,1) (p)
n Γ(q1 +q2 ) Γ(k+q1 )Γ(n−k+q2 ) n β(k+q1 ,n−k+q2 )
• pmf: P (k) = k Γ(q1 )Γ(q2 ) Γ(q1 +q2 +n)
= k β(q1 ,q2 )

• EX = nq1
q1 +q2

nq1 q2 (n+q1 +q2 )

• Var X = (q1 +q2 )2 (1+q1 +q2 )

5.10 Zipf or zeta random variable

5.38. P [X = k] = 1 1
ξ(p) kp
where k ∈ N, p > 1, and ξ is the zeta function defined in (A.11).
ξ(p−n)
5.39. E [X n ] = ξ(p)
is finite for n < p − 1, and E [X n ] = ∞ for n ≥ p − 1.

6 PDF Examples
6.1 Uniform Distribution
6.1. Characterization for uniform[a, b]:

0 x < a, x > b
(a) f (x) = b−a U (x − a) U (b − x) =
1
1
b−a
a≤x≤b

59
 X∼ fX (x) ϕX (u)
b+a sin(u b−a
2 )
Uniform U(a, b) 1
1 (x)
b−a [a,b]
eiu 2
u b−a
2
Exponential E(λ) λe−λx 1[0,∞] (x) λ
λ−iu
x−s
1 − μ−s0
Shifted Exponential (μ, s0 ) μ−s0
e 01
[s0 ,∞) (x)
α −αx
Truncated Exp. e−αa −e−αb
e 1[a,b] (x)
α −α|x| α2
Laplacian L(α) 2
e α2 +u2
1 x−m 2
√1 e− 2 ( σ ) ium− 12 σ 2 u2
Normal N (m, σ 2 ) σ 2π
e
1 T −1 T m− 1 uT Λu
Normal N (m, Λ) n√ e− 2 (x−m) Λ (x−m)
1
eju 2
(2π) 2 det(Λ)
λq xq−1 e−λx 1
Gamma Γ (q, λ) 1(0,∞) (x) q
Γ(q) (1−i uλ )
Pareto Par(α) αx−(α+1) 1[1,∞] (x)
 
α c α+1
Par(α, c) = c Par(α) c x
1(c,∞) (x)
Beta β (q1 , q2 ) Γ(q1 +q2 ) q1 −1
Γ(q1 )Γ(q2 )
x (1 − x)q2 −1 1(0,1) (x)
Γ(q1 +q2 ) xq1 −1
Beta prime 1
Γ(q1 )Γ(q2 ) (x+1)(q1 +q2 ) (0,∞)
(x)
2
Rayleigh 2αxe−αx 1[0,∞] (x)
1 1
Standard Cauchy π 1+x2
1 α
Cau(α) π α2 +x2
Γ(d) 1
Cau(α, d) √ 
παΓ(d− 12 ) 1+ x 2 d

( )
α
ln x−μ 2
μ,σ 2 − 12
Log Normal eN ( ) 1
√
σx 2π
e ( σ )1
(0,∞) (x)

Table 3: Examples of probability density functions. Here, c, α, q, q1 , q2 , σ, λ are all strictly

positive and d > 12 . γ = −ψ(1) ≈ .5772 is the Euler-constant. ψ (z) = dz d
log Γ (z) =

Γ (z) Γ(q1 )Γ(q2 )
(log e) Γ(z) is the digamma function. B(q1 , q2 ) = Γ(q1 +q2 ) is the beta function.

60

0 x < a, x > b
(b) F (x) = x−a
b−a
a≤x≤b
b+a sin(u b−a
2 )
(c) ϕX (u) = eiu 2
u b−a
2

esb −esa
(d) MX (s) = s(b−a)
.

6.2. For most purpose, it does not matter whether the value of the density f at the
1
endpoints are 0 or b−a .

Example 6.3.

• Phase of oscillators ⇒ [-π, π] or [0,2π]

• Phase of received signals in incoherent communications → usual broadcast carrier

phase φ ∼ U(-ππ)

• Mobile cellular communication: multipath → path phases φc ∼ U(−π, π)

• Use with caution to represent ignorance about a parameter taking value in [a, b].
(b−a)2
6.4. EX = a+b
2
, Var X = 12
, E [X 2 ] = 13 (b2 + ab + a2 ).

6.5. The product X of two independent U[0, 1] has

fX (x) = − (ln (x)) 1[0,1] (x)

and
FX (x) = x − x ln x
1 x 1
on [0, 1]. This comes from P [X > x] = 1 − FU t
dt = 1 − xt dt.
0 x

6.2 Gaussian Distribution

6.6. Gaussian distribution:

(a) Denoted by N (m, σ 2 ) . N (0, 1) is the standard Gaussian (normal) distribution.

1 x−m 2
(b) fX (x) = √ 1 e− 2 ( σ ) .
2πσ

(c) FX (x) = normcdf(x,m,sigma).

• The standard normal cdf is sometimes denoted by Φ(x). It inherits all properties
of cdf. Moreover, note that Φ(−x) = 1 − Φ(x).
 1 2 2
(d) ϕX (v) = E ejvX = ejmv− 2 v σ .]
1 2 2
(e) MX (s) = esm+ 2 s σ

61
 fX  x
fX  x

95%
68%

P V P P V P  2V P P  2V

Figure 14: Probability density function of X ∼ N (m, σ 2 ) .

∞ 1 2 σ2
(f) Fourier transform: F {fX } = fX (x) e−jωx dt = e−jωm− 2 ω .
−∞
   x−m   x−m 
(g) P [X > x] = P [X ≥ x] = Q x−m
σ
= 1
 − Φ  σx−m = Φ − σ 
P [X < x] = P [X ≤ x] = 1 − Q x−mσ
= Q − σ
= Φ x−m
σ
.
6.7. Properties
(a) P [|X − μ| < σ] = 0.6827; P [|X − μ| > σ] = 0.3173
P [|X − μ| > 2σ] = 0.0455; P [|X − μ| < 2σ] = 0.9545

(b) Moments and central moments:

( * ( *
k k−2 0, k odd
(i) E (X − μ) = (k − 1) E (X − μ) =
1 · 3 · 5 · · · · · (k − 1) σ , k even
k
7 =
( *
k 2 · 4 · 6 · · · · · (k − 1) σ k π2 , k odd
(ii) E |X − μ| =
1 · 3 · 5 · · · · · (k − 1) σ k , k even
(iii) Var [X 2 ] = 4μ2 σ 2 + 2σ 4 .

n 0 1 2 3 4
EX n 1 μ μ + σ2
2
μ (μ + 3σ 2 )
2
μ + 6μ σ + 3σ 4
4 2 2

E [(X − μ)n ] 1 0 σ2 0 3σ 4

(c) For N (0, 1) and k ≥ 1,

  0, k odd
E X k = (k − 1) E X k−2 =
1 · 3 · 5 · · · · · (k − 1) , k even.

The first equality comes from integration by parts. Observe also that
 (2m)!
E X 2m = m .
2 m!

(d) Lévy–Cramér theorem: If the sum of two independent non-constant random variables
is normally distributed, then each of the summands is normally distributed.

62
 ∞ 2 "π
• Note that e−αx dx = α
.
−∞

6.8 (Length bound). For X ∼ N (0, 1) and any (Borel) set B,

! |B|/2  
|B|
P [X ∈ B] ≤ fX (x) = 1 − 2Q ,
−|B|/2 2

where |B| is the length (Lebesgue measure) of the set B. This is because the probability
is concentrated around 0. More generally, for X ∼ N (m, σ 2 )
 
|B|
P [X ∈ B] ≤ 1 − 2Q .
2σ

6.9 (Stein’s Lemma). Let X ∼ N (μ, σ 2 ), and let g be a differentiable function satisfying
E |g  (X)| < ∞. Then
E [g(X)(X − μ)] = σ 2 E [g  (X)] .
[2, Lemma 3.6.5 p 124]. Note that this is simply integration by parts with u = g(x) and
dv = (x − μ)fX (x)dx.
  
• E (X − μ)k = E (X − μ)k−1 (X − μ) = σ 2 (k − 1)E (X − μ)k−2 .
∞ x 2
6.10. Q-function: Q (z) = √1 e− 2 dx corresponds to P [X > z] where X ∼ N (0, 1);
2π
z
that is Q (z) is the probability of the “tail” of N (0, 1). The Q function is then a comple-
mentary cdf (ccdf).

1
&  0,1 0.9

0.8

0.7

Qz 0.6

0.5

0.4

0.3

0.2

0.1

z
0
-3 -2 -1 0 1 2 3
z
0

Figure 15: Q-function

(a) Q is a decreasing function with Q (0) = 12 .

(b) Q (−z) = 1 − Q (z) = Φ(z)

(c) Q−1 (1 − Q (z)) = −z

63
 π π
2 − x2 2 x2
(d) Craig’s formula: Q (x) = 1
π
e 2 sin2 θ dθ = 1
π
e− 2 cos2 θ dθ, x ≥ 0.
0 0
i.i.d.
To see this, consider X, Y ∼ N (0, 1). Then,
π
!! !2 !∞
Q (z) = fX,Y (x, y)dxdy = 2 fX,Y (r cos θ, r sin θ)drdθ.
0 z
(x,y)∈(z,∞)×R cos θ

where we evaluate the double integral using polar coordinates [9, Q7.22 p 322].
π
4 x2
2
(e) Q (x) = 1
π
e− 2 sin2 θ dθ
0

x2
(f) d
dx
Q (x) = − √12π e− 2
(f (x))2
(g) d
dx
Q (f (x)) = − √12π e− 2
d
dx
f (x)
 
   (f (x))2
− 2
d  x
(h) Q (f (x)) g (x)dx = Q (f (x)) g (x)dx + √1 e f (x) g (t) dt dx
2π dx
a
 
(i) P [X > x] = Q x−m
σ   x−m 
P [X < x] = 1 − Q x−mσ
= Q − σ .

(j) Approximation:
( * z2
(i) Q (z) ≈ (1−a)z+a z2 +b √12π e− 2 ;
1 √
a = π1 , b = 2π
  e− x22 x2
(ii) 1 − 1
x2
≤ Q (x) ≤ 12 e− 2
√
x 2π
  − z2
(iii) Q (z) ≈ z√12π 1 − 0.7
z 2 e 2 ;z > 2
z 2 √ 
6.11. Error function (MATLAB): erf (z) = √2
π
e−x dx = 1 − 2Q 2z
0

(a) It is an odd function of z.

 
(b) For z ≥ 0, it corresponds to P [|X| < z] where X ∼ N 0, 12 .

(c) lim erf (z) = 1

z→∞

(d) erf (−z) = −erf (z)

    
(e) Q (z) = 2 erfc 2 = 2 1 − erf √z2
1 √z 1

    
(f) Φ(x) = 1
2
1 + erf √x = 1
2
erfc − √x2
(2)

64
 √
(g) Q−1 (q) = 2 erfc−1 (2q)
√  ∞ 2
(h) The complementary error function: erfc (z) = 1−erf (z) = 2Q 2z = √2
π z
e−x dx

§ 1·
& ¨ 0, ¸
© 2¹
erf  z 

Q  2z 
0 z

Figure 16: erf-function and Q-function

6.3 Exponential Distribution

6.12. Denoted by E (λ). It is in fact Γ (1, λ). λ > 0 is a parameter of the distribution,
often called the rate parameter.

6.13. Characterized by

• fX (x) = λe−λx U (x) ;

 
• FX (x) = 1 − e−λx U (x) ;
• Survival-, survivor-, or reliability-function: P [X > x] = e−λx 1[0,∞) (x) + 1(−∞,0) (x);
• ϕX (u) = λ
λ−iu
.
• MX (s) = λ
λ−s
for Re {s} < λ.

6.14. EX = σX = λ1 , Var [X] = 1

λ2
.

6.15. median(X) = 1
λ
ln 2, mode(X) = 0, E [X n ] = n!
λn
.
σX
6.16. Coefficient of variation: CV = EX
=1

6.17. It is a continuous version of geometric distribution. In fact, X ∼ G0 (e−λ ) and

$X% ∼ G1 (e−λ )

6.18. X ∼ E(λ) is simply λ1 X1 where X1 ∼ E(1).

6.19. Suppose X1 ∼ E(1). Then E [X1n ] = n! for n ∈ N ∪ {0}.

In general, for X ∼ E(λ), we have E [X α ] = λ1α Γ(α + 1) for any α > −1. In particular, for
n ∈ N ∪ {0}, the moment E [X n ] = λn!n .

6.20. μ3 = E [(X − EX)3 ] = 2

λ3
and μ4 = 9
λ4
.

65
 f (x)
6.21. Hazard function: P [X>x]
=λ

6.22. h(X) = log λe .

6.23. Can be generated by X = − λ1 ln U where U ∼ U (0, 1).

6.24. MATLAB:

• X = exprnd(1/lambda)
• fX (x) = exppdf(x,1/lambda)
• FX (x) = expcdf(x,1/lambda)

6.25. Memoryless property : The exponential r.v. is the only continuous r.v. on [0, ∞)
that satisfies the memoryless property:

P [X > s + x |X > s] = P [X > x]

for all x > 0 and all s > 0 [13, p 157–159]. In words, the future is independent of the
past. The fact that it hasn’t happened yet, tells us nothing about how much longer it will
take before it does happen.

• In particular, suppose we define the set B + x to be {x + b : b ∈ B}. For any x > 0

and set B ⊂ [0, ∞), we have

P [X ∈ B + x|X > x] = P [X ∈ B]

because  
P [X ∈ B + x] B+x
λe−λt dt τ =t−x B
λe−λ(τ +x) dτ
= = .
P [X > x] e−λx e−λx

6.26. The difference of two independent E(λ) is L(λ).


6.27. Consider independent Xi ∼ E (λ). Let Sn = ni=1 Xi .

(a) Sn ∼ Γ(n, λ), i.e. its has n-Erlang distribution.

(b) Let N = inf {n : Sn ≥ s}. Then, N ∼ Λ + 1 where Λ ∼ P (λs).

6.28. If independent Xi ∼ E (λi ), then

 

(a) min Xi ∼ E λi .
i i
Recall order statistics. Let Y1 = min Xi
i
( * ∞ 
λ
(b) P min Xi = j = j . Note that this is fXj (t) P [Xi > t]dt.
i λi
i 0 i=j

66
 i.i.d.
6.29. If Si , Ti ∼ E (α), then
+ m , m−1 
 n  n + m − 1   1 n+m−1
P Si > Tj =
i 2
i=1 j=1 i=0

  n + i − 1   1 n+i
m−1
= .
i 2
i=0

Note that we can set up two Poisson processes. Consider the superposed process. We want
the nth arrival from the T processes to come before the mth one of the S process.

6.4 Pareto: Par(α)–heavy-tailed model/density

6.30. Characterizations: Fix α > 0.

(a) f (x) = αx−α−1 U (x − 1)

  0 x<1
(b) F (x) = 1 − 1
U (x − 1) =
xα 1− 1
xα
x≥1
Example 6.31.

• distribution of wealth

• flood heights of the Nile river

• designing dam height

• (discrete) sizes of files requested by web users

• waiting times between successive keystrokes at computer terminals

• (discrete) sizes of files stored on Unix system file servers

• running times for NP-hard problems as a function of certain parameters

6.5 Laplacian: L(α)

6.32. Characterization: α > 0

(a) Also known as Laplace or double exponential.

(b) f (x) = α2 e−α|x|

1 αx
2
e x<0
(c) F (x) = 1 −αx
1 − 2e x≥0
α2
(d) ϕX (u) = α2 +u2

67
 λ2
(e) MX (s) = λ2 −s2
, −λ < Re {s} < λ.
6.33. EX = 0, Var X = 2
α2
Example 6.34.
• amplitudes of speech signals
• amplitudes of differences of intensities between adjacent pixels in an image
• If X and Y are independent E(λ), then X − Y is L(λ). (Easy proof via ch.f.)

6.6 Rayleigh
6.35. Characterizations:
 
2
(a) F (x) = 1 − e−αx U (x)
2
(b) f (x) = 2αxe−αx u (x)
2
e−αt , t ≥ 0
(c) P [X > t] = 1 − F (t) =
1, t<0
√  
(d) Use −2σ 2 ln U to generate Rayleign σ12 from U ∼ U(0, 1).
6.36. Read “ray -lee”
Example 6.37.
• noise X at the output of AM envelope detector when no signal is present
6.38. Relationship with other distributions
(a) Let X be a Rayleigh(α) r.v., then Y = X 2 is E(α). Hence,
√
·
E(α) −
−
−
−
2
Rayleigh(α). (17)
(·)

i.i.d. √
(b) Suppose X, Y ∼ N (0, σ 2 ). R = X 2 + Y 2 has a Rayleigh distribution with density
1 1 2
fR (r) = 2r 2 e− 2σ2 r .
2σ
i.i.d.    
• Note that X 2 , Y 2 ∼ Γ 12 , 2σ1 2 . Hence, X 2 + Y 2 ∼ Γ 1, α = 2σ1 2 , exponential.
√
By (17), X 2 + Y 2 is a Rayleigh r.v. α = 2σ1 2 .
• Alternatively,
 x r  transformation from Cartesian coordinates to polar coordinates
y
→ θ
1 1 r cos θ 2 1 1 r sin θ 2
fR,Θ (r, θ) = rfX,Y (r cos θ, r sin θ) = r √ e− 2 ( σ ) √ e− 2 ( σ )
σ 2π σ 2π
  
1 1 1 2
= 2r 2 e− 2σ2 r
2π 2σ
Hence, the radius R and the angle Θ are independent, with the radius R having
a Rayleigh distribution while the angle Θ is uniformly distributed in the interval
(0, 2π).

68
6.7 Cauchy
6.39. Characterizations: Fix α > 0.
α 1
(a) fX (x) = ,
α>0
π α2 +x2
 
(b) FX (x) = π1 tan− 1 λx + 12 .

(c) ϕX (u) = e−α|u|

Note that Cau(α) is simply αX1 where X1 ∼ Cau(1). Also, because fX is even, f−X = fX
and thus −X is still ∼ Cau(α).
6.40. Odd moments are not defined; even moments are infinite.
• Because the first moment is not defined, central moments, including the variance, are
not defined.
• Mean and variance do not exist.
• Note that even though the pdf of Cauchy distribution is an even function, the mean
is not 0.

6.41. Suppose Xi are independent Cau(αi ). Then, i ai Xi ∼ Cau(|ai | αi ).
 
6.42. Suppose X ∼ Cau(α). X1 ∼ Cau α1 .

6.8 Weibull
6.43. For λ > 0 and p > 0, the Weibull(p, λ) distribution [9] is characterized by
 1
(a) X = Yλ p where Y ∼ E(1).

(b) fX (x) = λpxp−1 e−λx , x > 0

p

(c) FX (x) = 1 − e−λx , x > 0

p

Γ(1+ n
p)
(d) E [X n ] = n .
λp

7 Expectation
Consider probability space (Ω, A, P )
7.1. Let X + = max (X, 0), and X − = − min (X, 0) = max (−X, 0). Then, X = X + − X − ,
and X + , X − are nonnegative r.v.’s. Also, |X| = X + + X −
7.2. A random variable X is integrable if and only if
≡ X has a finite expectation

≡ both E [X + ] and E [X − ] are finite

69
 ≡ E |X| is finite.

≡ EX is finite ≡ EX is defined.

≡ X∈L

≡ |X| ∈ L

In which case,  
EX = E X + − E X −
! !
= X (ω) P (dω) = XdP
! !
= xdP (x) = xP X (dx)
X

and !
XdP = E [1A X] .
A

Definition 7.3. A r.v. X admits (has) an expectation if E [X + ] and E [X − ] are not

both equal to +∞. Then, the expectation of X is still given by EX = E [X + ] − E [X − ]
with the conventions +∞ + a = +∞ and −∞ + a = −∞ when a ∈ R

7.4. L 1 = L 1 (Ω, A, P ) = the set of all integrable random variables.

7.5. For 1 ≤ p < ∞, the following are equivalent:

1
(a) (E [X p ]) p = 0

(b) E [X p ] = 0

(c) X = 0 a.s.
a.s.
7.6. X = Y ⇒ EX = EY

7.7. E [1B (X)] = P (X −1 (B)) = P X (B) = P [X ∈ B]


• FX (x) = E 1(−∞,x] (X)

7.8. Expectation rule: Let X be a r.v. on (Ω, A, P ), with values in (E, E), and distri-
bution P X . Let h : (E, E) → (R, B) be measurable. If

• X ≥ 0 or
 
• h (X) ∈ L 1 (Ω, A, P ) which is equivalent to h ∈ L 1 E, E, P X

then
 
• E [h (X)] = h (X (ω)) P (dω) = h (x) P X (dx)
 
• h (X (ω)) P (dω) = h (x) P X (dx)
[X∈G] G

70
 7.9. Expectation of an absolutely continuous random variable: Suppose X has
density fX , then h is P X -integrable if and only if h · fX is integrable w.r.t. Lebesgue
measure. In which case,
! !
E [h (X)] = h (x) P (dx) = h (x) fX (x) dx
X

and ! !
X
h (x) P (dx) = h (x) fX (x) dx
G G

• Caution: Suppose h is an odd function and fX is an even function, we can not

conclude that E [h(X)] = 0. One obvious odd-function h is h(x) = x. For example,
in (6.40), when X is Cauchy, the expectation does not exist even though the pdf is
an even funciton. Of course, in general, if we also know that h(X) is integrable, then
E [h(X)] is 0.

Expectation of a discrete random variable: Suppose x is a discrete random variable.


EX = xP [X = x]
x

and 
E [g(X)] = g(x)P [X = x].
x

Similarly, 
E [g(X, Y )] = = g(x, y)P [X = x, Y = y].
x y

These are called the law/rule of the lazy statistician (LOTUS) [21, Thm 3.6 p 48],[9,
p 149].
   
7.10. P [X ≥ t] dt = P [X > t] dt and P [X ≤ t] dt = P [X < t] dt
E E E E

7.11. Expectation and cdf :

(a) For nonnegative X,

!∞ !∞ !∞
EX = P [X > y] dy = (1 − FX (y)) dy = P [X ≥ y] dy (18)
0 0 0

For p > 0,
!∞
E [X p ] = pxp−1 P [X > x]dx.
0

71
 (b) For integrable X,
!∞ !0
EX = (1 − FX (x)) dx − FX (x) dx
0 −∞
∞ ∞
(c) For nonnegative integer-valued X, EX = k=0 P [X > k] = k=1 P [X ≥ k].

Definition 7.12.
( *  
(a) Absolute moment: E |X|k = |x|k P X (dx), where we define E |X|0 = 1
( *  
(b) Moment: If E |X|k < ∞, then mk = E X k = xk P X (dx) = the k th moment of
X

(c) Variance: If E [X 2 ] < ∞, then we define

!
 
Var X = E (X − EX) = (x − EX)2 P X (dx) = E X 2 − (EX)2
2

= E [X(X − EX)]

• Notation: DX , or σ 2 (X), or σX
2
, or VX [21, p 51]
5k 6 5k 6

k   
k
• If Xi ∈ L , then
2
Xi ∈ L , Var
2
Xi exists, and E Xi = EXi
i=1 i=1 i=1 i=1

• Suppose E [X 2 ] < ∞. If Var X = 0, then X = EX a.s.

"
(d) Standard deviation: σX = Var[X]
σX
(e) Coefficient of variation: CVX = EX
Var X
(f) Fano factor (index of dispersion): EX

(g) Central moments: the nth central moment is μn = E [(X − EX)n ].

(i) μ1 = E [X − EX] = 0.
2
(ii) μ2 = σX = Var X
 n
n
(iii) μn = k
mn−k (−m1 )k
k=1
n n
(iv) mn = k
μn−k mk1
k=1
μ3
(h) Skewness coefficient: γX = 3
σX

(i) Describe the deviation of the distribution from a symmetric shape (around the
mean)
(ii) 0 for any symmetric distribution

72
 μ4
(i) Kurtosis: κX = 4 .
σX

• κX = 3 for Gaussian distribution

(j) Excess coefficient: εX = κX − 3 = μ4

4
σX
−3

1 ∂k 
(k) Cumulants or semivariants: For one variable: γk = j k ∂v k
ln (ϕX (v)) .
v=0

(i) γ1 = EX = m1 ,
γ2 = E [X − EX]2 = m2 − m21 = μ2
γ3 = E [X − EX]3 = m3 − 3m1 m2 + 2m31 = μ3
γ4 = m4 − 3m22 − 4m1 m3 + 12m21 m2 − 6m41 = μ4 − 3μ22
(ii) m1 = γ1
m2 = γ2 + γ12
m3 = γ2 + 3γ1 γ2 + γ13
m4 = γ4 + 3γ22 + 4γ1 γ3 + 6γ12 γ2 + γ14

Moment Measure of Definition Continuous variable Discrete variable Sample estimator

∞  
First Central Mean, expected value μx = −∞
x f x (x) dx μx = x p(xk )
all x  s k
x̄ = xi /n
location E( X ) = μx
∞  1

Second Dispersion Variance, Var( X ) = μ2 = σx2 σx2 = −∞
(x − μx ) 2 f x (x) dx σx2 = all x  s
(xk − μx ) 2 Px (xk ) s2 = n−1
(xi − x̄) 2
" " = 
1
Standard deviation, σx σx = Var( X ) σx = Var( X ) s= n−1
(xi − x̄) 2
Coefficient of variation, x x = σx /μx x = σx /μx Cv = s/x̄
∞  n

Third Asymmetry Skewness μ3 = −∞
(x − μx ) 3 f x (x) dx μ3 = all x  s
(xk − μx ) 3 p x (xk ) m3 = (n−1)(n−2)
(xi − x̄) 3
Skewness coefficient, γx γx = μ3 /σx3 γx = μ3 /σx3 g = m3 /s3
∞  n(n+1)

Fourth Peakedness Kurtosis, κx μ4 = −∞
(x − μx ) 4 f x (x) dx μ4 = all x  s
(xk − μx ) 4 p x (xk ) m4 = (n−1)(n−2)(n−3)
(xi − x̄) 4
Excess coefficient, εx κx = μ4 /σx4 κx = μ4 /σx4 k = m4 /s4
εx = κx − 3 εx = κx − 3

Figure 17: Product-Moments of Random Variables [20]

7.13.

• For c ∈ R, E [c] = c
• E [·] is a linear operator: E [aX + bY ] = aEX + bEY .
• In general, Var[·] is not a linear operator.

7.14. All pairs of mean and variance are possible. A random variable X with EX = m
and Var X = σ 2 can be constructed by setting P [X = m − a] = P [X = m + a] = 12 .

Definition 7.15.

• Correlation between X and Y : E [XY ].

73
 Model E [X] Var [X]
n2 −1
U{0,1,...,n−1} n−1
2 12
B(n, p) np np(1-p)
G(β) β
1−β
β
(1−β)2
P(λ) λ λ
(b−a)2
U(a, b) a+b
2 12
E(λ) 1 1
λ ⎧ λ2
α ⎨ undefined, 0 < α < 1
, α>1
Par(α) α−1 ∞, 1<α<2
∞, 0<α≤1 ⎩ α
, α>2
(α−2)(α−1)2
L(α) 0 2
α2
N (m, σ 2 ) m σ 2

N (m, Λ) m Λ = [Cov [Xi , Xj ]]

q q
Γ (q, λ) λ λ2

Table 4: Expectations and Variances

• Covariance between X and Y :

Cov [X, Y ] = E [(X − EX)(Y − EY )] = E [XY ] − EXEY
= E [X(Y − EY )] = E [Y (X − EX)] .

• X and Y are said to be uncorrelated if and only if Cov [X, Y ] = 0.

≡ E [XY ] = EXEY

• X and Y are said to be orthogonal if E [XY ] = 0.

• Correlation coefficient, autocorrelation, normalized covariance:
5  6
Cov [X, Y ] X − EX Y − EY E [XY ] − EXEY
ρXY = =E = .
σX σY σX σY σX σY

7.16. Properties
(a) Var X = Cov [X, X], ρX,X = 1

(b) Var[aX] = a2 Var X, σaX = |a| σX .

(c) If X Y , then Cov [X, Y ] = 0. The converse is not true.

|=

(d) ρXY ∈ [−1, 1].

(e) By Caychy-Schwartz inequality, (Cov [X, Y ])2 ≤ σX

2 2
σY with equality if and only if
σY (X − EX) = σX (Y − EY ) a.s.
2 2 2 2

• This implies |ρX,Y | ≤ 1.

74
 When σY , σX > 0, equality occurs if and only if the following conditions holds

≡ ∃a = 0 such that (X − EX) = a(Y − EY )

≡ ∃c = 0 and b ∈ R such that Y = cX + b
≡ ∃a = 0 and b ∈ R such that X = aY + b
≡ |ρXY | = 1

In which case, |a| = σσXY and ρXY = |a|

a
= sgn a. Hence, pXY is used to quantify linear
dependence between X and Y . The closer |ρXY | to 1, the higher degree of linear
dependence between X and Y .

(f) Linearity:

(i) Let Yi = ai Xi + bi .
i. Cov [Y1 , Y2 ] = Cov [a1 X1 + b1 , a2 X2 + b2 ] = a1 a2 Cov [X1 , X2 ].
ii. The ρ is preserved under linear transformation:

ρY1 ,Y2 = ρX1 ,X2 .

(ii) Cov [a1 X + b1 , a2 X + b2 ] = a1 a2 Var X.

(iii) ρa1 X+b1 ,a2 X+b2 = 1. In particular, if Y = aX + b, then ρX,Y = 1.

(g) ρX,Y = 0 if and only if X and Y are uncorrelated.

(h) When EX = 0 or EY = 0, orthogonality is equivalent to uncorrelatedness.

(i) For finite index set I,

8 9
  
Var ai Xi = a2i Var Xi + 2 aj aj Cov [Xi , Xj ] .
i∈I i∈I (i,j)∈I×I
i=j

In particular
Var (X + Y ) = Var X + Var Y + 2Cov [X, Y ]
and
Var (X − Y ) = Var X + Var Y − 2Cov [X, Y ] .

(j) For finite index set I and J,

8 9
  
Cov ai Xi , b j Yj = ai bj Cov [Xi , Yj ] .
i∈I j∈J i∈I j∈J

(k) Covariance Inequality: Let X be any random variable and g and h any function such
that E [g(X)], E [h(X)], and E [g(X)h(X)] exist.

75
 • If g and h are either both non-decreasing or non-increasing, then

Cov [g(X), h(X)] ≥ 0. (19)

• If g is non-decreasing and h is non-increasing, then

Cov [g(X), h(X)] ≤ 0. (20)

See also [2, p 191–192] and (8.13).

(l) Being uncorrelated does not imply independence

• Discrete: Suppose pX is an even function with pX (0) = 0. Let Y = g(X) where g

is also an even function. Then, E [XY ] = E [X] = E [X] E [Y ] = Cov [X, Y ] = 0.
Consider a point x0 such that pX (x0 ) > 0. Then, pX,Y (x0 , g(x0 )) = pX (x0 ). We
only need to show that pY (g(x0 )) = 1 to show that X and Y are not independent.
For example, let X be uniform on {±1, ±2} and Y = |X|. Consider the point
x0 = 1.
• Continuous: Let Θ be uniform on an interval of length 2π. Set X = cos Θ and
Y = sin Θ. See (11.6).

7.17. See (4.48) for relationships between expectation and independence.

Example 7.18 (Martingale betting strategy). Fix a > 0. Suppose X0 , X1 , X2 , . . .

are independent random variables with P [Xi = 1] = p and P [Xi = 0] = 1 − p. Let
N = inf {i : Xi = 1}. Also define
⎧
⎨ 0,  N =0
L(N ) = N
−1
⎩ a i
r , N ∈N
i=0


k−1
and G(N ) = arN − L(N ). To have G(N ) > 0, need ri < rk ∀k ∈ N which turns out to
i=0
require r ≥ 2. In fact, for r ≥ 2, we have G(N ) ≥ a ∀N ∈ N ∪ {0}. Hence, E [G(N )] ≥ a.
It is exactly a when r = 2.
Now, E [L (N )] = a ∞ rn −1 n
n=1 r−1 p (1 − p) = ∞ if and only if r(1 − p) ≥ 1. When
1 − p ≤ 12 , because we already have r ≥ 2, it is true that r(1 − p) ≥ 1.

8 Inequalities
8.1. Let (Ai : i ∈ I) be a finite family of events. Then
 2
 + ,
P (Ai )  
 i
≤P Ai ≤ P (Ai )
P (Ai ∩ Aj ) i i
i j

76
8.2. [19, p 14]
 n +n ,
1  n
n
− 1− ≤P Ai − P (Ai ) ≤ (n − 1) n− n−1 .
n i=1 i=1 ↓
↓
1
− 1e ≈ −0.37

See figure 18.

• |P (A1 ∩ A2 ) − P (A1 ) P (A2 )| ≤ 14 .

1
1

x
§
 ¨ 1
1· 0.5
¸
© x¹
x
x 1
( x 1) ˜ x
0
1
e

2 4 6 8 10
1 x 10
 

n 
n
Figure 18: Bound for P Ai − P (Ai ).
i=1 i=1

8.3. Markov’s Inequality : P [|X| ≥ a] ≤ a1 E |X|, a > 0.

a a1ª x t a º
¬ ¼

x
a
Figure 19: Proof of Markov’s Inequality

(a) Useless when a ≤ E |X|. Hence, good for bounding the “tails” of a distribution.

77
 (b) Remark: P [|X| > a] ≤ P [|X| ≥ a]

(c) P [|X| ≥ aE |X|] ≤ a1 , a > 0.

(d) Suppose g is a nonnegative function. Then, ∀α > 0 and p > 0, we have

(i) P [g (X) ≥ α] ≤ 1
αp
(E [(g (X))p ])
(ii) P [g (X − EX) ≥ α] ≤ 1
αp
(E [(g (X − EX))p ])

(e) Chebyshev’s Inequality : P [|X| > a] ≤ P [|X| ≥ a] ≤ 1

a2
EX 2 , a > 0.

(i) P [|X| ≥ α] ≤ 1
αp
(E [|X|p ])
2
σX
(ii) P [|X − EX| ≥ α] ≤ α2
; that is P [|X − EX| ≥ nσX ] ≤ 1
n2
• Useful only when α > σX
   
a+b 2
(iii) For a < b, P [a ≤ X ≤ b] ≥ 1 − 4
(b−a)2
2
σX + EX − 2

(f) One-sided Chebyshev inequalities: If X ∈ L2 , for a > 0,

EX 2
(i) If EX = 0, P [X ≥ a] ≤ EX 2 +a2
(ii) For general X,
2
σX
i. P [X ≥ EX + a] ≤ 2 +a2 ;
σX
that is P [X ≥ EX + nσX ] ≤ 1
1+n2
2
σX
ii. P [X ≤ EX − a] ≤ 2 +a2 ;
σX
that is P [X ≤ EX − nσX ] ≤ 1
1+n2
2
2σX
iii. P [|X − EX| ≥ a] ≤ 2 +a2 ;
σX
that is P [|X − EX| ≥ nσX ] ≤ 2
1+n2
This is a
2
σX
better bound than a2
iff σX > a

(g) Chernoff bounds:

E[e−θX ]
(i) P [X ≤ b] ≤ e−θb
∀θ > 0
E[eθX ]
(ii) P [X ≥ b] ≤ eθb
∀θ > 0

This can be optimized over θ

E|X|−a
8.4. Suppose |X| ≤ M a.s., then P [|X| ≥ a] ≥ M −a
∀a ∈ [0, M )
(EX)2
8.5. X ≥ 0 and E [X 2 ] < ∞ ⇒ P [X > 0] ≥ E[X 2 ]

Definition 8.6. If p and q are positive real numbers such that p + q = pq, or equivalently,
1
p
+ 1q = 1, then we call p and qa pair of conjugate exponents.

• 1 < p, q < ∞
• As p → 1, q → ∞. Consequently, 1 and ∞ are also regarded as a pair of conjugate
exponents.

78
8.7. Hölder’s Inequality : X ∈ Lp , Y ∈ Lq , p > 1, 1
p
+ 1
q
= 1. Then,
(a) XY ∈ L1
1 1
(b) E [|XY |] ≤ (E [|X|p ]) p (E [|Y |q ]) q with equality if and only if
E [|Y |q ] |X (ω)|p = E [|X|p ] |Y (ω)|q a.s.

8.8. Cauchy-Bunyakovskii-Schwartz Inequality: If X, Y ∈ L2 , then XY ∈ L1 and

  1   1
|E [XY ]| ≤ E [|XY |] ≤ E |X|2 2 E |Y |2 2 or equivalently,
    2 
(E [XY ])2 ≤ E X 2 E Y
with equality if and only if E [Y 2 ] X 2 = E [X 2 ] Y 2 a.s.
(a) |Cov (X, Y )| ≤ σX σY .
(b) (EX)2 ≤ EX 2
(c) (P (A ∩ B))2 ≤ P (A)P (B)
1
8.9. Minkowski’s Inequality : p ≥ 1, X, Y ∈ Lp ⇒ X + Y ∈ Lp and (E [|X + Y |p ]) p ≤
1 1
(E [|X|p ]) p + (E [|Y |p ]) p
8.10. p > q > 0
(a) E [|X|q ] ≤ 1 + E [|X|p ]
1 1
(b) Lyapounov’s inequality : (E [|X|q ]) q ≤ (E [|X|p ]) p
=    1
• E [|X|] ≤ E |X|2 ≤ E |X|3 3 ≤ · · ·

8.11. Jensen’s Inequality : For a random variable X, if 1) X ∈ L1 (and ϕ (X) ∈ L1 );

2) X ∈ (a, b) a.s.; and 3) ϕ is convex on (a, b), then ϕ (EX) ≤ E [ϕ (X)]
 
• For X > 0 (a.s.), E X1 ≥ EX 1
.
8.12.
• For p ∈ (0, 1], E [|X + Y |p ] ≤ E [|X|p ] + E [|Y |p ].
• For p ≥ 1, E [|X + Y |p ] ≤ 2p−1 (E [|X|p ] + E [|Y |p ]).
8.13 (Covariance Inequality). Let X be any random variable and g and h any function
such that E [g(X)], E [h(X)], and E [g(X)h(X)] exist.
• If g and h are either both non-decreasing or non-increasing, then
E [g(X)h(X)] ≥ E [g(X)] E [h(X)] .
In particular, for nondecreasing g, E [g(X)(X − EX)] ≥ 0.
• If g is non-decreasing and h is non-increasing, then
E [g(X)h(X)] ≤ E [g(X)] E [h(X)] .

See also (19), (20), and [2, p 191–192].

79
9 Random Vectors
In this article, a vector is a column matrix with dimension n × 1 for some n ∈ N. We use
1 to denote a vector with all element being 1. Note that 1(1T ) is a square matrix with all
element being 1. Finally, for any matrix A and constant a, we define the matrix A + a to
be the matrix A with each of the components are added by a. If A is a square matrix, then
A + a = A + a1(1T ).

Definition 9.1. Suppose I is an index set. When Xi ’s are random variables, we de-
fine a random vector XI by XI = (Xi : i ∈ I). For example, if I = [n], we have XI =
(X1 , X2 , . . . , Xn ). Note also that X[n] is usually denoted by X1n . Sometimes, we simply
write X to denote X1n .

• For disjoint A, B, XA∪B = (XA , XB ).

• For vector xI , yI , we say x ≤ y if ∀i ∈ I, xi ≤ yi [7, p 206].
• When the dimension of X is implicit, we simply write X and x to represent X1n and
xn1 , respectively.
 
• For random vector X, Y , we use (X, Y ) to represent the random vector X or
 T T T Y
equivalently X Y .

Definition 9.2. Half-open cell or bounded rectangle in Rk is set of the form Ia,b =
k
{x : ai < xi ≤ bi , ∀i ∈ [k]} = × (ai , bi ]. For a real function F on Rk , the difference of F
i=1
around the vertices of Ia,b is
  
ΔIa,b F = sgnIa,b (v) F (v) = (−1)|{i:vi =ai }| F (v) (21)
v v

where the sum extending over the 2k vertices v of Ia,b . (The ith coordinate of the vertex v
could be either ai or bi .) In particular, for k = 2, we have

F (b1 , b2 ) − F (a1 , b2 ) − F (b1 , a2 ) + F (a1 , a2 ) .

9.3 (Joint cdf ).

  ( *
FX (x) = FX1k xk1 = P [X1 ≤ x1 , . . . , Xk ≤ xk ] = P X ∈ Sxk1 = P X (Sx )

where Sx = {y : yi ≤ xi , i = 1, . . . , k} consists of the points “southwest” of x.

• ΔIa,b FX ≥ 0

• The set Sx is an orthanlike or semi-infinite corner with “northeast” vertex (vertex in

the direction of the first orthant) specified by the point x [7, p 206].

C1 FX is nondecreasing in each variable. Suppose ∀i yi ≥ xi , then FX (y) ≥ FX (x)

C2 FX is continuous from above: lim FX (x1 + h, . . . , xk + h) = FX (x)

h 0

80
 C3 xi → −∞ for some i (the other coordinates held fixed), then FX (x) → 0
If ∀i xi → ∞, then FX (x) → 1.

• lim FX (x1 − h, . . . , xk − h) = P X (Sx◦ ) where Sx◦ = {y : yi < xi , i = 1, . . . , k} is the

h 0
interior of Sx

• Given a ≤ b, P [X ∈ Ia,b ] = ΔIa,b FX .

This comes from (9) with Ai = [Xi ≤ ai ] and B = [∀ Xi ≤ bi ]. Note that
ai , i ∈ I,
P [∩i∈I Ai ∩ B] = F (v) where vi =
bi , otherwise.

• For any function F on Rk with satisfies (C1), (C2), and (C3), there is a unique
probability measure μ on BRk such that ∀a, ∀b ∈ Rk with a ≤ b, we have μ (Ia,b ) =
ΔIa,b F (and ∀x ∈ Rk μ (Sx ) = F (x)).

• TFAE:

(a) FX is continuous at x
(b) FX is continuous from below
(c) FX (x) = P X (Sx◦ )
(d) P X (Sx ) = P X (Sx◦ )
(e) P X (∂Sx ) = 0 where ∂Sx = Sx − Sx◦ = {y : yi ≤ xi ∀i, ∃j yj = xj }

• If k > 1, FX can have discontinuity points even if P X has no point masses.

• FX can be discontinuous at uncountably many points.

• The continuity points of FX are dense.

• For any j, we have lim FX (x) = FXI\{j} (xI\{j} )

xj →∞

9.4 (Joint pdf ). A function f is a multivariate or joint pdf (commonly called a density)
if and only if it satisfies the following two conditions:

(a) f ≥ 0;

(b) f (x)d(x) = 1.

• The integrability of the pdf f implies that for all i ∈ I

lim f (xI ) = 0.
xi →±∞


• P [X ∈ A] = A
fX (x)dx.
• Remarks: Roughly, we may say the following:
<Xi ≤xi +Δxi ]
P [∀i, xi P [x<X≤x+δx]

(a) fX (x) = lim = lim
∀i, Δxi →0 i Δxi Δx→0 i Δxi

81
 (b) For I = [n],
 x1  xn
◦ FX (x) = −∞ . . . −∞ fX (x)dxn . . . dx1
∂n
◦ fX (x) = ∂x1 ···∂xn FX (x) .

   (k)  (k) u, j = k
◦ ∂u FX (u, . . . , u) =
∂
fX v dx[n]\{k} where vj = .
k∈N (−∞,u]n  k
xj , j =
∂
u u
For example, ∂u FX,Y (u, u) = fX,Y (x, u)dx + fX,Y (u, y) dy.
−∞ −∞
5 6

n
(c) fX1n (xn1 ) = E δ (Xi − xi ) .
i=1

• The level sets of a density are sets where density is constant.

9.5. Consider two random vectors X : Ω → Rd1 and Y : Ω → Rd2 . Define Z = (X, Y ) :
Ω → Rd1 +d2 . Suppose that Z has density fX,Y (x, y).
 
(a) Marginal Density : fY (y) = fX,Y (x, y)dx and fX (x) = fX,Y (x, y)dy.
Rd 1 Rd 2

• In other words, to obtain the marginal densities, integrate out the unwanted
variables.

• fXI\{i} (xI\{i} ) = fXI (xI )dxi .
fX,Y (x,y)
(b) fY |X (y|x) = fY (y)
.
 y1  yd
(c) FY |X (y|x) = −∞
··· −∞
2
fY |X (t|x)dtd2 · · · dt1 .

9.6. P [(X + a1 , X + b1 ) ∩ (Y + a2 , Y + b2 ) = ∅] = A
fX,Y (x, y)dxdy where A is defined
in (1.10).
9.7. Expectation and covariance:
(a) The expectation of a random vector X is defined to be the vector of expectations of
its entries. EX is usually denoted by μX or mX .

(b) For non-random matrix A, B, C and a random vector X, E [AXB + C] = AEXB +C.

(c) The correlation matrix RX of a random vector X is defined by


RX = E XX T .

Note that it is symmetric.

(d) The covariance matrix CX of a random vector X is defined as

 
CX = ΛX = Cov [X] = E (X − EX)(X − EX)T = E XX T − (EX)(EX)T
= RX − (EX)(EX)T .

(i) The ij-entry of the Cov [X] is simply Cov [Xi , Xj ].

82
 (ii) ΛX is symmetric.
i. Properties of symmetric matrix
A. All eigenvalues are real.
B. Eigenvectors corresponding to different eigenvalues are not just linearly
independent, but mutually orthogonal.
C. Diagonalizable.
ii. Spectral theorem: The following equivalent statements hold for symmet-
ric matrix.
A. There exists a complete set of eigenvectors; that is there exists an or-
thonormal basis u(1) , . . . , u(n) of R with CX u(k) = λk u(k) .
B. CX is diagonalizable by an orthogonal matrix U (U U T = U T U = I).
C. CX can be represented as CX = U ΛU T where U is an orthogonal matrix
whose columns are eigenvectors of CX and λ = diag(λ1 , . . . , λn ) is a
diagonal matrix with the eigenvalues of CX .
(iii) Always nonnegative definite (positive
( semidefinite).* That is ∀a ∈ Rn where n is
2
the dimension of X, aT CX a = E aT (X − μX ) ≥ 0.
• det (CX ) ≥ 0.
1 √ √ √ √ √ √
(iv) We can define CX2 = CX to be CX = U ΛU T where Λ = diag( λ1 , . . . , λn ).
√ √
i. det CX = det CX .
√
ii. CX is nonnegative definite.
√ 2 √ √
iii. CX = CX CX = CX .
(v) Suppose, furthermore, that CX is positive definite.
−1
i. CX = U Λ−1 U T where Λ−1 = diag( λ11 , . . . , λ1n ).
= √  
− 12 −1 −1 T 1 1
ii. CX = CX = ( CX ) = U DU where D = √ √
, . . . , λn
λ1
√ −1
√
iii. CX CX CX = I.
1
−1 −1
iv. CX , CX2 , CX 2 are all positive definite (and hence are all symmetric).
 1 2
− −1
v. CX 2 = CX .
−1
vi. Let Y = CX 2 (X − EX). Then, EY = 0 and CY = I.
(vi) For i.i.d. Xi with each with variance σ 2 , ΛX = σ 2 I.

(e) Cov [AX + b] = ACov [X] AT

 
• Cov X T h = Cov hT X = hT Cov [X] h where h is a vector with the same
dimension as X.

(f) For Y = X + Z, ΛY = ΛX + 2ΛXZ + ΛZ .

• When X and Z are independent, ΛY = ΛX + ΛZ .

83
 • For Yi = X + Zi where X and Z are independent, ΛY = σX
2
+ ΛZ .
(g) ΛX+Y + ΛX−Y = 2ΛX + 2ΛY
(h) det (ΛX+Y ) ≤ 2n det (ΛX + ΛY ) where n is the dimension of X and Y .
2
(i) Y =⎡ (X, X, . . . ⎤
, X) where X is a random variable with variance σX , then ΛY =
1 ··· 1
2 ⎢ .. . ⎥
σX ⎣ . . . . .. ⎦. Note that Y = 1X where 1 has the same dimension as Y .
1 ··· 1

(j) Let X be a zero-mean random vector whose covariance matrix is singular. Then, one
of the Xi is a deterministic linear combination of the remaining components. In other
words, there is a nonzero vector a such that aT X = 0. In general, if ΛX is singular,
then there is a nonzero vector a such that aT X = aT EX.
(k) If X and Y are both random vectors (not necessarily of the same dimension), then
their cross-covariance matrix is

ΛXY = CXY = Cov [X, Y ] = E (X − EX)(Y − EY )T .
Note that the ij-entry of CXY is Cov [Xi , Yj ].
• CY X = (CXY )T .

(l) RXY = E XY T .
 
(m) If we stack X and Y in to a composite vector Z = XY
, then
 
CX CXY
CZ = .
CY X CY

(n) X and Y are said to be uncorrelated if CXY = 0, the zero matrix. In which case,
 
CX 0
C(X ) = ,
Y 0 CY
a block diagonal matrix.
9.8. The joint characteristic function of an n-dimensional random vector X is defined
by ( T *  
ϕX (v) = E ejv X = E ej i vi Xi .
When X has a joint density fX , ϕX is just the n-dimensional Fourier transform:
!
T
ϕX (v) = ejv x fX (x)dx,

and the joint density can be recovered using the multivariate inverse Fourier transform:
!
1
e−jv x ϕX (v)dv.
T
fX (x) = n
(2π)

84
 TX
(a) ϕX (u) = Eeiu .

e−jv x ϕX (v)dv.
1 T
(b) fX (x) = (2π)n

T  
(c) For Y = AX + b, ϕY (u) = eib u ϕX AT u .

(d) ϕX (−u) = ϕX (u).

(e) ϕX (u) = ϕX,Y (u, 0)


n

n  
υi
υi 
n
(f) Moment: υ
∂ i=1
υ υn ϕX
∂v1 1 ∂v2 2 ···∂vn
(0) = j i=1 E Xiυi
i=1

∂
(i) ϕ
∂vi X
(0) = jEXi .
∂2
(ii) ϕ
∂vi ∂vj X
(0) = j 2 E [Xi Xj ] .

(g) Central Moment:



(i) ∂
∂vi
ln (ϕX (v)) = jEXi .
v=0

∂2 
(ii) ∂vi ∂vj
ln (ϕX (v)) = −Cov [Xi , Xj ] .
v=0

∂3 
(iii) ∂vi ∂vj ∂vk
ln (ϕX (v)) = j 3 E [(Xi − EXi ) (Xj − EXj ) (Xk − EXk )] .
v=0
(iv) E [(Xi − EXi ) (Xj − EXj ) (Xk − EXk ) (X − EX )] = Ψijk + Ψij Ψk + Ψik Ψj +
4 
Ψi Ψjk where Ψijk = ∂vi ∂v∂j ∂vk ∂v ln (ϕX (v)) .
v=0

Remark : we do not require that any or all of i, j, k, and λ be distinct.

9.9 (Decorrelation and the Karhunen-Loève expansion). Let X be an n-dimensional

random vector with zero mean and covariance matrix C. X has the representation X = P Y ,
there the component of Y are uncorrelated and P is an n × n orthonormal matrix. This
representation is called the Karhunen-Loève expansion.

• Y = PTX
• P T = P −1 is called a decorrelating transformation.
• Diagonalize C = P DP T where D = diag(λ1 ). Then, Cov [Y ] = D.
• In MATLAB, use [P,D] = eig(C). To extract the diagonal elements of D as a vector,
use the command d = diag(D).
• If C is singular (equivalently, if some of the λi are zero), we only need to keep around
the Yi for which λi > 0 and can throw away the other components of Y without any
loss of information. This is because λi = EYi2 and EYi2 = 0 if and only if Yi ≡ 0 a.s.

[9, p 338–339].

85
9.1 Random Sequence
9.10. [10, p 9–10] Given a countable family X1 , X2 , . . . of random r.v.’s, their statistical
properties are regarded as defined by prescribing, for each integer n ≥ 1 and every finite
set I ⊂ N, the joint distribution function FXI of the random vector XI = (Xi : i ∈ I).
Of course, some consistency requirements must be imposed upon the infinite family FXI ,
namely, that for j ∈ I
 
(a) FXI\{j} xI\{j} = lim FXI (xI ) and that
xj →∞

(b) the distribution function obtained from FXI (xI ) by interchanging two of the indices
i1 , i2 ∈ I and the corresponding variable xi1 and xi2 should be invariant. This simply
means that the manner of labeling the random variables X1 , X2 , . . . is not relevant.
The joint distributions {FXI } are called the finite-dimensional distributions associated
with XN = (Xn )∞n=1 .

10 Transform Methods
10.1 Probability Generating Function
Definition 10.1. [9][10, p 11] Let X be a discrete random variable taking only nonnegative
integer values. The probability generating function (pgf) of X is

 X 
∞
GX (z) = E z = z k P [X = k].
k=0

• In the summation, the first term (the k = 0 term) is P [X = 0] even when z = 0.

• GX (0) = P [X = 0]
• G(z −1 ) is the z transform of the pmf.
• GX (1) = 1.
• The names derives from the fact that it can be used to compute the pmf.
• It is finite at least for any complex z with |z| ≤ 1. Hence pgf is well defined for
|z| ≤ 1.
(k) (k)
Definition 10.2. GX (1) = lim GX (z).
z1

10.3. Properties
(a) GX is infinitely differentiable at least for |z| < 1.

(b) Probability generating property:


1 d(k) 
(k)
GX (z) = P [X = k].
k! dz z=0

86
 (c) Moment generating property:
 8k−1 9
d(k)  
GX (z) =E (X − i) .
dz (k) z=1 i=0

The RHS is called the kth factorial moment of X.

(d) In particular,
EX = GX (1)
EX 2 = GX (1) + GX (1)
Var X = GX (1) + GX (1) − (GX (1))
2

(e) pgf of a sum

n of independent random variables is the product of the individual pgfs.
Let S = i=1 Xi where the Xi ’s are independent.

n
GS (z) = GXi (z).
i=1

10.2 Moment Generating Function

Definition 10.4.
 sXThe moment
 sx X generating function of a random variable X is defined
as MX (s) = E e = e P (dx) for all s for which this is finite.
10.5. Properties of moment generating funciton
(a) MX (s) is defined on some interval containing 0. It is possible that the interval consists
of 0 alone.
(i) If X ≥ 0, this interval contains (−∞, 0].
(ii) If X ≤ 0, this interval contains [0, ∞).
(b) Suppose that M (s) is defined throughout an interval (−s0 , s0 ) where s0 > 0, i.e. it
exists (is finite) in some neighborhood of 0. Then,
( *
(i) X has finite moments of all order: E |X|k < ∞∀k ≥ 0

∞
sk
 k
(ii) M (s) = k!
E X , for complex-valued s with |s| < s0 [1, eqn (21.22) p 278].
k=0
Thus M (s) has a Taylor expansion about 0 with positive radius of convergence.
∞
i. If M (s) can somehow be calculated and expanded in a series ak sk , and if
 k=0 
the coefficients ak can be identified, then ak = k!1 E X k . That is E X k =
k!ak
 
(iii) M (k) (0) = E X k = xk P X (dx)

(c) If M is defined in some neighborhood of s, then M (k) (s) = xk esx P X (dx)
(d) See also Chernoff bound.

87
10.3 One-Sided Laplace Transform
10.6. The one-sided Laplace transform a nonnegative random variable X is defined
 of−sx
for s ≥ 0 by L (s) = M (s) = E e−sX = e P X (dx)
[0,∞)

• Note that 0 is included in the range of integration.

• Always finite because e−sx ≤ 1. In fact, it is a decreasing function of s
• L (0) = 1
• L (s) ∈ [0, 1]
 
(a) Derivative: For s > 0, L(k) (s) = (−1)k xk e−sx P X (dx) = (−1)k E X k e−sX
n n
n L (s) = (−1) E [X ]
d n
(i) lim ds
s↓0
• Because the value at 0 can be ∞, it does not make sense to talk about
dn
L (0) for n > 1
dsn

(ii) ∀s ≥ 0 L (s) is differentiable and ds d
L (s) = −E Xe−sX , where at 0, this is the
d
right-derivative. ds L (s) is finite for s > 0
• d
ds
L (0) = −E [X] ∈ [0, ∞]

st
(−1)k k (k)
(b) Inversion formula: If FX is continuous at t > 0, then FX (t) = lim k!
s L (s)
s→∞ k=0

(c) FX and P X are determined by LX (s)

(i) In fact, they are determined by the values of LX (s) for s beyond any arbitrary
s0 . (That is we don’t need to know LX (s) for small s.) Also, knowing LX (s)
on N is also sufficient.

(ii) Let μ and ν be probability measures on [0, ∞). If ∃s0 ≥ 0 such that e−sx μ (dx) =
e−sx ν (dx)∀s ≥ s0 , then μ = ν
 −sx
(iii) Let f1 , f2 be real functions on [0, ∞). If ∃s0 ≥ 0 such that ∀s ≥ s0 e f1 (x)dx =
 −sx [0,∞)
e f2 (x)dx , then f1 = f2 Lebesgue-a.e.
[0,∞)

(d) Let X1 , . . . , Xn be independent nonnegative random variables, then L 

n (s) =
Xi
i=1

n
LXi (s)
i=1

(e) Suppose F is a distribution function with corresponding Laplace transform L. Then

 −sx
(i) e F (x)dx = λ1 L (s)
[0,∞)

(ii) e−sx (1 − F (x))dx = 1
λ
(1 − L (s))
[0,∞)

[16, p 183].

88
10.4 Characteristic Function
10.7. The characteristic function (abbreviated
 c.f. or ch.f.) of a probability measure
μ on the line is defined for real t by ϕ (t) = eitx μ (dx)  
A random variable X has characteristic function ϕX (t) = E eitX = eitx P X (dx)
 
(a) Always exists because |ϕ(t)| ≤ |eitx |μ (dx) = 1μ (dx) = 1 < ∞

(b) If X has a density, then ϕX (t) = eitx fX (x) dx

(c) ϕ (0) = 1

(d) ∀t ∈ R |ϕ (t)| ≤ 1

(e) ϕ is uniformly continuous.

(f) Suppose that all moments of X exists and ∀t ∈ R, EetX < ∞, then ϕX (t) =
∞
(it)k
k!
EX k
k=0
( *  
(g) If E |X|k < ∞, then ϕ(k) (t) = ik E X k eitX and ϕ(k) (0) = ik E X k .

(h) Riemann-Lebesgue theorem: If X has a density, then ϕX (t) → 0 as |t| → ∞

(i) ϕaX+b (t) = eitb ϕ (at)

(j) Conjugate Symmetry Property : ϕ−X (t) = ϕX (−t) = ϕX (t)

D
• X = −X iff ϕX is real-valued.
• |ϕX | is even.

(k) X is a.s. integer-valued if and only if ϕX (2π) = 1


n
(l) If X1 , X2 , . . . , Xn are independent, then ϕ 
n (t) = ϕXj (t)
Xj j=1
j=1

(m) Inversion

(i) The inversion formula: If the probability measure μ has characteristic func-
T e−ita −e−itb
tion ϕ and if μ {a} = μ {b} = 0, then μ (a, b] = lim 2π
1
it
ϕ (t)dt
T →∞ −T

T e−ita −e−itb
i. In fact, if a < b, then lim 1
ϕ (t)dt = μ (a, b) + 12 μ {a, b}
T →∞ 2π −T it

ii. Equivalently, if F is the distribution function, and a, b are continuity points

T e−ita −e−itb
of F , then F (b) − F (a) = lim 2π 1
it
ϕ (t)dt
T →∞ −T

(ii) Fourier inversion: Suppose that |ϕX (t)|dt < ∞, then X is absolutely con-
tinuous with

89
 
i. bounded continuous density f (x) = 1
e−itx ϕ (t)dt
1
 e−ita −e−itb 2π
ii. μ (a, b] = 2π it
ϕ (t)dt
D
(n) Continuity Theorem: Xn −
→ X if and only if ∀t ϕXn (t) → ϕX (t) (pointwise).

(o) ϕX on complex plane for X ≥ 0

(i) ϕX (z) is defined in the complex plane for Imz ≥ 0

i. |ϕX (z)| ≤ 1 for such z
(ii) In the domain Im {z} > 0, ϕX is analytic and continuous including the boundary
Im {z} = 0
(iii) ϕX determines uniquely a function LX (s) of real argument s ≥ 0 which is
equal to LX (s) = ϕX (is) = Ee−sX . Conversely, LX (s) on the half-line s ≥ 0
determines uniquely ϕX

(p) If E |X|n < ∞, then

  8 7 >9
 
n k  |t|n+1
|t|n
 (it)  2
ϕX (t) − EX k  ≤ E min |X|n+1 , |X|n (22)
 k!  (n + 1)! n!
k=0


n
(it)k
and ϕX (t) = k!
EX k + tn β (t) where lim β (t) = 0 or equivalently,
k=0 |t|→0


n
(it)k
ϕX (t) = EX k + o (tn ) (23)
k=0
k!

(i) |ϕX (t) − 1| ≤ E [min {|tX| , 2}]

( '2 )* 2
• If EX = 0, then |ϕX (t) − 1| ≤ E min t2 X 2 , 2 |t| |X| ≤ t2 EX 2
( '2 )* 2
(ii) For integrable X, |ϕX (t) − 1 − itEX| ≤ E min 2 X , 2 |t| |X| ≤ t2 EX 2
t 2

(iii) For X with finite EX 2 ,

  ' 3 )
i. ϕX (t) − 1 − itEX + 12 t2 EX 2  ≤ E min |t|6 |X|3 , t2 |X|2
ii. ϕX (t) = 1 + itEX − 12 t2 EX 2 + t2 β (t) where lim β (t) = 0
|t|→0

10.8. φX (u) = MX (ju).


10.9. If X is a continuous r.v. with density fX , then ϕX (t) = ejtx fX (x)dx (Fourier
1
transform) and fX (x) = 2π e−jtx φX (t)dt (Fourier inversion formula).
• ϕX (t) is the Fourier transform of fX evaluated at −t.
• ϕX inherit the properties of a Fourier transform.
  
(a) For nonnegative ai such that i ai = 1, if fY = i ai fXi , then ϕX = i ai ϕXi .

90
 (b) If fX is even, then ϕX is also even.

• If fX is even, ϕX = ϕ−X .

10.10. Linear combination

n of independent random n variables: Suppose X1 , . . . , Xn are
independent. Let Y = i=1 ai Xi .  Then, ϕY (t) = i=1 ϕX (ai t). Furthermore, if |ai | = 1
and all fXi are even, then ϕY (t) = ni=1 ϕX (t).

(a) ϕX+Y (t) = ϕX (t)ϕY (t).

(b) ϕX−Y (t) = ϕX (t)ϕY (−t).

(c) If fY is even, ϕX−Y = ϕX+Y = ϕX ϕY .

10.11. Characteristic function for sum of distribution: Consider nonnegative ai such that

i ai = 1. Let Pi be probability measure with corresponding ch.f. ϕi . Then, the ch.f. of
i ai Pi is i ai ϕ i .

(a) 
Discrete r.v.:
 Suppose pi is pmf with corresponding ch.f. ϕi . Then, the ch.f. of
i ai pi is i ai ϕ i .

(b) Absolutely
 continuous
 r.v.: Suppose fi is pdf with corresponding ch.f. ϕi . Then, the
ch.f. of i ai fi is i ai ϕi .

11 Functions of random variables

Definition 11.1. The preimage or inverse image of a set B is defined by g −1 (B) =
{x : g(x) = B}.

11.2. For discrete X, suppose Y = g(X). Then, pY (y) = x∈g−1 ({y}) pX (x). The joint
pmf of Y and X is given by pX,Y (x, y) = pX (x)1[y = g(x)].

• In most cases,
 we  can show that X and Y are not independent,
 (0)  pick a point x
(0)

 y = g x
(0) (1) (1) (1)
such that pX x > 0. Pick a point
 (0)y such
 (1)that and pY y > 0.
(0) (1)
Then, pX,Y x , y = 0 but pX x pY y > 0. Note that this technique does
not always work. For example, if g is a constant function which maps all values of x
to a constant c. Then, we won’t be able to find y (1) . Of course, this is to be expected
because we know that a constant is always independent of other random variables.

11.1 SISO case

There are many techniques for finding the cdf and pdf of Y = g(X).

(a) One may first find FY (y) = P [g(X) ≤ y] first and then find fY from FY (y). In which
case, the Leibniz’ rule in (38) will be useful.

(b) Formula (25) below provides a convenient way of arriving at fY from fX without
going through FY .

91
11.3 (Linear transformation). Y = aX + b where a = 0.
7  
FX y−b
a 
, a>0
FY (y) =  y−b − 
1 − FX a
, a<0

(a) Suppose X is absolutely continuous,

 
1 y−b
fY (y) = fX . (24)
|a| a
 y−b 
In fact (24) holds even for mixed r.v. if we allow delta function because 1
|a|
δ a
− xk =
δ (y − (axk + b)).
(b) Suppose X is discrete,  
y−b
pY (y) = pX .
a

If we write fX (x) = k pX (xk )δ(x − xk ), we have

fY (y) = p (xk ) δ (y − (axk + b)).
k

11.4 (Power Law Function). Y = X n , n ∈ N or n ∈ (0, ∞).

 1 1
 1
1 n −1
(a) n odd: FY (y) = FX y n and fY (y) = n y fX y n .

(b) n even: ⎧ 
  − 
⎨ 1 1
FX y n − FX −y n , y≥0
FY (y) =
⎩
0, y < 0.
and 7 1
  1  1 
y −1
1 n
n
fX y n + fX −y n , y≥0
fY (y) =
0, y < 0.
Again, the density fY in the above formula holds when X is absolutely continuous. Note
that when n < 1, fY is not defined at 0. If we allow delta  functions,
 then the density
1 1
−1
1 n
formula above are also valid for mixed r.v. because n y δ ±y n − xk = δ (y − (±xk )n ).

• Let X be an absolutely continuous random variable. The density of Y = X 2 is

0, √   √  y<0
fY (y) = 1
√ f y + 1
√ f − y , y ≥ 0.
2 y X 2 y X

11.5. In general, for Y = g(X), we solve the equation y = g(x). Denoting its real roots
by xk . Then,
 fX (xk )
fY (y) = . (25)
k
|g  (x )|
k

If g(x) = c = constant for every x in the interval (a, b), then FY (y) is discontinuous for
y = c. Hence, fY (y) contains an impulse (FX (b) − FX (a)) δ(y − c) [14, p 93–94].

92
 • To see this, consider when there is unique x such that g(x) = y. Then, For small Δx
and Δy, P [y, y < Y ≤ y+Δy] = P [x < X ≤ x+Δx] where (y+Δy] = g ((x, x + Δx])
is the image of the interval (x, x + Δx]. (Equivalently, (x, x + Δx] is the inverse image
of y + Δy].) This gives fY (y)Δy = fX (x)Δx.
• The joint density fX,Y is

fX,Y (x, y) = fX (x) δ (y − g (x)) . (26)

Let the xk be the solutions for x of g(x) = y. Then, by integrating (26) w.r.t. x, we
have (25) via the use of (3).
• When g bijective,  
 d −1   −1 
fY (y) =  g (y) fX g (y) .
dy
   
 
• For Y = a
X
, fY (y) =  ay  fX ay .
√
• Suppose X is nonnegative. For Y = X,

fY (y) = 2yfX (y 2 ).

11.6. Given Y = g(X) where X ∼ U(a, b). Then, to get fY (y0 ), plot g on (a, b). Let
A = g −1 ({y0 }) be the set of all points x such that g(x) = y0 . Suppose A can be written
as a countable disjoint union A = B ∪ ∪i Ii where B is countable and the Ii ’s are intervals.
We have + ,
1 1 
fY (y) = + (Ii ) δ(y − y0 )
b − a |g  (x)| i

at y = y0 where (I) is the length of the interval I.

• Suppose Θ is uniform on an interval of length 2π. Y1 = cos Θ and Y2 = sin Θ are

both arcsine random variables with FYi (y) = 1 − π1 cos−1 y = 21 + π1 sin−1 (y) and
fYi (y) = π1 √ 1 2 for y ∈ [−1, 1]. Note also that E [Y1 Y2 ] = EYi = Cov [Y1 , Y2 ] = 0.
1−y
Hence, Y1 and Y2 are uncorrelated. However, it is easy to see that Y1 and Y2 are not
independent by considering the joint and marginal densities at y1 = y2 = 0.

Example 11.7. Y = X 2 1[X≥0]

⎧
⎨ 0, y<0
(a) FY (y) = FX (0),
√  y = 0
⎩
FX y , y>0
?
0, √  y < 0
(b) fY (y) = 1
√ f y , y > 0 + FX (0) δ (y)
2 y X

93
 p = n1 Hypergeometric

distributions
Discrete
n3 x = 0,1⋅⋅⋅ , min(n1, n2)
X1 + ⋅⋅⋅ + XK n3 ← ∞
n1, n2, n3

Poisson v = np
x = 0,1⋅⋅⋅ n ∞ ← Binomial
v x = 0,1 ⋅⋅⋅ n X1 + ⋅⋅⋅ + XK
n, p
Bernoulli
n=1 x = 0,1
m = np
s2 = v p
s 2 = np(1 – p)
m=v
n ← ∞

Normal a = b→ ∞
Y = eX Beta
Lognormal -∞ < x < ∞ 0<x<1
y>0 log Y m, s a, b
m = ab
X- m s 2 = ab 2
X1 + ⋅⋅⋅ + XK a→ ∞ X1
X1 + ⋅⋅⋅ + XK s
1/X
m + sX X1 + X2

Cauchy Standard Gamma

a=b=1
-∞ < x < ∞

distributions
Continuous
normal x>0
a, a -∞ < x < ∞ a, b
n = 2a
X1/X2 2 a = 1/2
X1 + ⋅⋅⋅ + XK2
a=0 a=n
a + aX X1 + ⋅⋅⋅ + XK
a=1

a =1 Erlang
Standard Chi-square x>0
cauchy n→∞ x>0 b, n
-∞ < x < ∞ n

n2 → ∞
1
n=2 n=1
X n1 X X1/n1 b =2 X1 + ⋅⋅⋅ + XK
n=1 X2 /n2

F Exponential Standard
x>0 x>0 -blog X
uniform
n1, n2
b 0 < x <1
min(X1,◊◊◊,XK)
2
X
1
a + (b − a)X a=0
X X a a=1 X1 - X2
b=1
t X2
-∞ < x < ∞ Rayleigh Weibull Uniform
Triangular
n x>0 x>0 a<x<b
–1 < x < 1 a, b
b a, b

Figure 20: Relationships among univariate distributions [20]

94
 Uniform: -  a, b 
§ ·
X i ~ *  qi , O  Ÿ ¦X ~ * ¨ ¦ qi , O ¸ . 1 §O·
i
© i ¹ fX  x 1  x X ~  O  Ÿ D X ~  ¨ ¸ .
i
b  a > a , b@ ©D ¹
§ O·
X ~ *  q, O  Ÿ Y D X ~ * ¨ q, ¸ . § ·
© D¹ 1 Xi ~ (Oi) Ÿ min  X i  ~  ¨ ¦ Oi ¸ .
X ~ -  0,1 Ÿ  ln  X  ~   O 
O © i ¹
Gamma : *  q, O  Exponential: (O)
O q x q 1e O x q 1 fX  x O e O x1>0, f   x 
fX  x 1 0, f   x 
*q

n 1
q ,O §1 1 · i .i .d .
§ 1 ·
2 2V 2 X ~ &  0,V 2  Ÿ X 2 ~ * ¨ , 2 ¸ X 2  Y 2 with X , Y ~ & ¨ 0, ¸
© 2 2V ¹ © 2O ¹

Normal/Gaussian: &  m,V 2 

§n 1 ·
Chi-squared F 2 : * ¨ , 2 ¸ 1 § xP ·
2

© 2 2V ¹ 1  ¨
V ¸¹
n i .i .d . fX  x e 2©
Usually, V 1 . ¦X i
2
with X i ~ &  0,V 2
 V 2S
i 1

X1 X2
with X i ~ *  qi , O 
X1  X 2 O D i .i .d .
§ 1 ·
X 2  Y 2 with X , Y ~ & ¨ 0, ¸
© 2D ¹
Beta: 2

*  q1  q2  q2 1
Rayleigh: f  x  2D xe D x 1>0, f   x 
f E q ,q  x  x q1 1 1  x  1 0,1  x 
1 2
*  q1  *  q2 
Din
November 8, 2004

Figure 21: Another diagram demonstrating relationship among univariate distributions

95
11.2 MISO case
11.8. If X and Y are jointly continuous random variables with joint density fX,Y . The
following two methods give the density of Z = g(X, Y ).
• Condition on one of the variable, say Y = y. Then, begin conditioned, Z is simply a
function of one variable g(X, y); hence, we can use the one-variable technique to find
fZ|Y (z|y). Finally, fZ (z) = fZ|Y (z|y)fY (y)dy.
   
• Directly find the joint density of the random vector YZ = g(X,Y )
. Observe that the
 ∂g ∂g  Y
 ∂g 
Jacobian is ∂x ∂y . Hence, the magnitude of the determinant is  ∂x .
0 1.
Of course, the standard way of finding the pdf of Z is by finding the derivative of the
cdf FZ (z) = (x,y):x2 +y2 ≤z fX,Y (x, y)d(x, y). This is still good for solving specific examples.
It is also a good starting point for those who haven’t learned conditional probability nor
Jacobian.
Let the x(k) be the solutions of g(x, y) = z for fixed z and y. The first method gives

 fX|Y (x|y) 
∂ 
fZ|Y (z|y) =  g(x, y)  .
k ∂x (k) x=x

Hence, 
 fX,Y (x, y) 
∂ 
fZ,Y (z, y) =  g(x, y)  ,
k ∂x x=x(k)
which comes out of the second method directly. Both methods then gives
!  
fX,Y (x, y) 
∂ 
fZ (z) =  g(x, y)  dy.
∂xk (k) x=x

The integration for a given z is only on the value of y such that there is at least a solution
for x in z = g(x, y). If there is no such solution, fZ (z) = 0. The same technique works for
a function of more than one random variables Z = g(X1 , . . . , Xn ). For any j ∈ [n], let the
(k)
xj be the solutions for xj in z = g(x1 , . . . , xn ). Then,

!  
fX1 ,...,Xn (x1 , . . . , xn ) 
fZ (z) =    dx[n]\{j} .
 ∂ 
k  ∂xj g(x1 , . . . , xn )  (k)
xj =xj

For the second method, we consider the random vector (hr (X1 , . . . , Xn ), r ∈ [n]) where
hr (X1 , . . . , Xn ) = Xr for r = j and hj = g. The Jacobian is of the form
⎛ ⎞
1 0 0 0
⎜ 0 1 0 0 ⎟
⎜ ∂g ∂g ⎟
⎝ ∂x ∂x ∂x ∂x ⎠ .
∂g ∂g
1 2 j n
0 0 0 1

96
By swapping the row with all the partial derivatives to the first row, the magnitude of
the determinant is unchanged and we also end up with upper triangular matrix whose
determinant is simply the product of the diagonal elements.
(a) For Z = aX + bY ,
!   !  
1 z − by 1 z − ax
fZ (z) = fX,Y , y dy = fX,Y x, dx.
|a| a |b| b
 ax+by    a b 
• Note that Jacobian y
y
, xy = .
0 1

(i) When a = 1, b = −1,

! !
fX−Y (z) = fX,Y (z + y, y) dy = fX,Y (x, x − z) dx.

(ii) Note that when X and Y are independent and a = b = 1, we have the convolu-
tion formula
! !
fZ (z) = fX (z − y) fY (y)dy = fX (x)fY (z − x) dx.

(b) For Z = XY ,
!  !  
z 1 z 1
fZ (z) = fX,Y x, dx = fX,Y ,y dy
x |x| y |y|
     y x 
[9, Ex 7.2, 7.11, 7.15]. Note that Jacobian xyy
, xy = .
0 1
(c) For Z = X 2 + Y 2 ,
√ "    "  
! z fX|Y  
z−  + fX|Y − z − y  y
y2 y 2
fZ (z) = " fY (y)dy
√ 2 z − y2
− z

[9, Ex 7.16]. Alternatively,

!
1 2π √ √
fZ (z) = fX,Y ( z cos θ, z sin θ)dθ, z>0 (27)
2 0

[15, eq (9.14) p 318)].

i.i.d. 1
• This can be used to show that when X, Y ∼ N (0, 1), Z = X 2 + Y 2 ∼ E 2
.
√
(d) For R = X 2 + Y 2 ,
! 2π
fR (r) = r fX,Y (r cos θ, r sin θ)dθ, r > 0
0

[15, eq (9.13) p 318)].

97
 Y
(e) For Z = , !   !
X
y 
y 
fZ (z) =   fX,Y , y dy = |x| fX,Y (x, xz)dx.
z z
X
Similarly, when Z = Y
, !
fZ (z) = |y| fX,Y (yz, y)dy.

min(X,Y )
(f) For Z = max(X,Y )
where X and Y are strictly positive,
! ∞ ! ∞
FZ (z) = FY |X (zx|x)fX (x)dx + FX|Y (zy|y)fY (y)dy,
!
0
∞ !
0
∞
fZ (z) = xfY |X (zx|x)fX (x)dx + yfX|Y (zy|y)fY (y)dy, 0 < z < 1.
0 0

[9, Ex 7.17].
N
11.9 (Random sum). Let S = i=1 Vi where Vi ’s are i.i.d. ∼V independent of N .

(a) ϕS (u) = ϕN (−i ln (ϕV (u))).

• ϕS (u) = GN (ϕV (u))

• For non-negative integer-valued summands, we have GS (z) = GN (GV (z))

(b) ES = EN EV .

(c) Var [S] = EN (Var V ) + (EV )2 (Var N ).

Remark : If N ∼ P (λ), then ϕS (u) = exp (λ (ϕV (u) − 1)), the compound Poisson
distribution CP (λ, L (V )). Hence, the mean and variance of CP (λ, L (V )) are λEV and
λEV 2 respectively.

11.3 MIMO case

Definition 11.10 (Jacobian). In vector calculus, the Jacobian is shorthand for either
the Jacobian matrix or its determinant, the Jacobian determinant. Let g be a
function from a subset D of Rn to Rm . If g is differentiable at z ∈ D, then all partial
derivatives exists at z and the Jacobian matrix of g at a point z ∈ D is
⎛ ∂g1 ⎞
∂x1
(z) · · · ∂x
∂g1
(z)  
⎜ .. ⎟
n
. . ∂g ∂g
dg (z) = ⎝ . .. .. ⎠ = (z) , . . . , (z) .
∂x1 ∂xn
∂gm
∂x1
(z) · · · ∂xn (z)
∂gm

∂(g1 ,...,gn )
Alternative notations for the Jacobian matrix are J, ∂(x 1 ,...,xn )
[7, p 242], Jg (x) where the it
is assumed that the Jacobian matrix is evaluated at z = x = (x1 , . . . , xn ).

98
 • Let A be an n-dimensional
 “box” defined by the corners x and x+Δx. The “volume”
of the image g(A) is ( i Δxi ) |det dg(x)|. Hence, the magnitude of the Jacobian
determinant gives the ratios (scaling factor) of n-dimensional volumes (contents). In
other words,  
 ∂(y1 , . . . , yn ) 

dy1 · · · dyn =   dx1 · · · dxn .
∂(x1 , . . . , xn ) 

• d(g −1 (y)) is the Jacobian of the inverse transformation.

• In MATLAB, use jacobian.

See also (A.14).

11.11 (Jacobian formulas). Suppose g is a vector-valued function of x ∈ Rn , and X

is an Rn -valued random vector. Define Y = g(X). (Then, Y is also an Rn -valued random
vector.) If X has joint density fX , and g is a suitable invertible mapping (such that the
inversion mapping theorem is applicable), then
1  −1        
fY (y) = fX g (y) = det d g −1 (y)  fX g −1 (y) .
|det (dg (g (y)))|
−1

• Note that for any matrix A, det(A) = det(AT ). Hence, the formula above could
tolerate the incorrect “Jacobian”.

In general, let X = (X1 , X2 , . . . , Xn ) be a random vector with pdf fX (x). Let S =

{x : fX (x) > 0}. Consider a new random vector Y = (Y1 , Y2 , . . . , Yn ), defined by Yi =
gi (X). Suppose that A0 , A1 , . . . , Ar form a partition of S with these properties. The set
A0 , which may be empty, satisfies P [X ∈ A0 ] = 0. The transformation Y = g(X) =
(g1 (X), . . . , gn (X)) is a one-to-ont transformation from Ai onto some common set B for
each i ∈ [k]. Then, for each i, the inverse functions from B to Ai can be found. Denote
(k)
the kth inverse x = h(k) (u) by xj = hj (y). This kth inverse gives, for y ∈ B, the unique
x ∈ Ak such that y = g(x). Assuming that the Jacobians det(dh(k) (y)) do not vanish
identically on B, we have

r
 
fY (y) = fX (h(k) (y)) det(dh(k) (y)) , y∈B
k=1

[2, p 185].

• Suppose for some k, Yk is some functions of other Yi . In particular, suppose Yk = h(yI )

for some index set I and some deterministic function h. Then, the kth row of the
Jacobian matrix is a linear combination of other rows. In particular,
 
∂yk  ∂ ∂yi
= h (yI ) .
∂xj i∈I
∂y i ∂x j

Hence, the Jacobian determinant is 0.

99
 11.12. Suppose Y = g(X) where both X and Y have the same dimension, then the joint
density of X and Y is
fX,Y (x, y) = fX (x)δ(y − g(x)).

• In most cases,
 we  can show that X and Y are not independent,   pick a point
 x(0)
such that fX x(0) > 0. Pick a point
 y(1) such
 that
 y (1) = g x(0) and fY y (1) > 0.
Then, fX,Y x(0) , y (1) = 0 but fX x(0) fY y (1) > 0. Note that this technique does
not always work. For example, if g is a constant function which maps all values of x
to a constant c. Then, we won’t be able to find y (1) . Of course, this is to be expected
because we know that a constant is always independent of other random variables.

Example 11.13.

(a) For Y = AX + b, where A is a square, invertible matrix,

1  
fY (y) = fX A−1 (y − b) . (28)
|det A|

(b) Transformation between Cartesian coordinates (x, y) and polar coordinates (r, θ)
"  
• x = r cos θ, y = r sin θ, r = x2 + y 2 , θ = tan−1 xy .
 ∂x ∂x   
  cos θ −r sin θ
•  ∂y ∂y  = 
∂r ∂θ  = r. (Recall that dxdy = rdrdθ).
∂r ∂θ
sin θ r cos θ 
We have

fR,Θ (r, θ) = rfX,Y (r cos θ, r sin θ) , r ≥ 0 and θ ∈ (−π, π),

and "  y 
1
fX,Y (x, y) = " fR,Θ x2 + y 2 , tan−1 .
x2 + y 2 x
If, furthermore, Θ is uniform on (0, 2π) and independent of R. Then,
1 1 " 
fX,Y (x, y) = " fR x2 + y 2 .
2π x2 + y 2
√ Y 
(c) A related transformation is given by Z = X 2 + Y 2 and Θ = tan−1 . In this
√ √ X
case, X = Z cos Θ, Y = Z sin Θ, and
1 √ √ 
fZ,Θ (z, θ) = fX,Y z cos θ, z sin θ
2
which gives (27).
√
11.14. Suppose X, Y are i.i.d. N (0, σ 2 ). Then, R = X 2 + Y 2 and Θ = arctan X Y
are
 1 r 2 
independent with R being Rayleigh fR (r) = σr2 e− 2 ( σ ) U (r) and Θ being uniform on
[0, 2π].

100
 11.15 (Generation of a random sample of a normally distributed random vari-
able). Let U1 , U2 be i.i.d. U(0, 1). Then, the random variables
"
X1 = −2 ln U1 cos(2πU2 )
"
X2 = −2 ln U1 sin(2πU2 )

are i.i.d. N (0, 1). Moreover,

"
Z1 = −2σ 2 ln U1 cos(2πU2 )
"
Z2 = −2σ 2 ln U1 sin(2πU2 )

are i.i.d. N (0, σ 2 ).

• The idea is to generate R and Θ according to (11.14) first.

x2 2
1 +x2
• det(dx(u)) = − 2π
u1
, u1 = e− 2 .

 ≤ dim(X). To find the joint pdf of Y , we

11.16. In (11.11), suppose dim(Y ) = dim(g)
introduce “arbitrary” Z = h(X) so that dim YZ = dim(X).

11.4 Order Statistics

Given a sample of n random variables X1 , . . . , Xn , reorder them so that Y1 ≤ Y2 ≤ · · · ≤ Yn .
Then, Yi is called the ith order statistic, sometimes also denoted Xi:n , X i , X(i) , Xn:i , Xi,n ,
or X(i)n .
In particular

• Y1 = X1:n = Xmin is the first order statistic denoting the smallest of the Xi ’s,
• Y2 = X2:n is the second order statistic denoting the second smallest of the Xi ’s . . .,
and
• Yn = Xn:n = Xmax is the nth order statistic denoting the largest of the Xi ’s.

In words, the order statistics of a random sample are the sample values placed in ascending
order [2, p 226]. Many results in this section can be found in [4].

11.17. Events properties:

 
[Xmin ≥ y] = i [Xi ≥ y] [Xmax ≥ y] = i [Xi ≥ y]
[Xmin > y] = i [Xi > y] [Xmax > y] = i [Xi > y]
[Xmin ≤ y] = i [Xi ≤ y] [Xmax ≤ y] = i [Xi ≤ y]
[Xmin < y] = i [Xi < y] [Xmax < y] = i [Xi < y]

Let Ay = [Xmax ≤ y], By = [Xmin > y]. Then, Ay = [∀i Xi ≤ y] and By = [∀i Xi > y].

101
 11.18 (Densities). Suppose the Xi are absolutely continuous with joint density fX . Let
Sy be the set of all n! vector which comes from permuting the coordinates of y.

fY (y) = fX (x), y1 ≤ y2 ≤ · · · ≤ yn . (29)
x∈Sy
 
To see this, note that fY (y) j Δy j is the probability that Yj is in the small interval of
length Δyj around yj . This probability can be calculated from finding the probability that
all Xk fall into the above small regions.
From the joint density, we can find the joint pdf/cdf of YI for any I ⊂ [n]. However, in
many cases, we can directly reapply the above technique to find the joint pdf of YI . This
is especially useful when the Xi are independent or i.i.d.
(a) The marginal density fYk can be found by approximating fYk (y) Δy with

n 
P [Xj ∈ [y, y + Δy) and ∀i ∈ I, Xi ≤ y and ∀r ∈ (I ∪ {k})c , Xr > y],
j=1 I∈([n]\{j})
k−1

 
where for any set A and integer  ∈ |A|, we define A to be the set of all k-element
subsets of A. Note also that we assume (I ∪ {k})c = [n] \ (I ∪ {k}).
To see this, we first choose the Xj that will be Yk with value around y. Then, we
must have k − 1 of the Xi below y and have the rest of the Xi > y.
(b) For integers r < s, the joint density fYr ,Ys (yr , ys )Δyr Δys can be approximated by the
probability that two of the Xi are inside small regions around yr and ys . To make
them Yr and Ys , for the other Xi , r − 1 of them before yr , s − r − 1 of them between
yr and ys , and n − s of them beyond ys .
• fXmax ,Xmin (u, v)ΔuΔv can be approximated by by

P [Xj ∈ [u, u+Δu), Xj ∈ [v, v+Δv), and ∀i ∈ [n]\{j, k} , v < Xi ≤ u], v ≤ u,
(j,k)∈S

where S is the set of all n(n − 1) pairs (j, k) from [n] × [n] with j = k. This is
simply choosing the j, k so that Xj will be the maximum with value around u,
and Xk will be the minimum with value around v. Of course, the rest of the Xi
have to be between the min and max.
◦ When n = 2, we can use (29) to get

fXmax ,Xmin (u, v) = fX1 ,X2 (u, v) + fX1 ,X2 (v, u), v ≤ u.

Note that the joint density at point yI is0 if the the elements in yI are not arranged in the
“right” order.
11.19 (Distribution functions). We note again the the cdf may be obtained by inte-
gration of the densities in (11.18) as well as by direct arguments valid also in the discrete
case.

102
(a) The marginal cdf is

n 
FYk (y) = P [∀i ∈ I, Xi ≤ y and ∀r ∈ [n] \I, Xr > y].
j=k I∈([n])
j

This is because the event [Yk ≤ y]

is the same as event that at least k of the Xi are
n
≤ y. In other words, let N (a) = i=1 1[Xi ≤ a] be the number of Xi which are ≤ a.
8 9 8 9
  
[Yk ≤ y] = N (y) ≥ k = N (y) = j , (30)
i j≥k i

where the union is a disjoint union. Hence, we sum the probability that exactly j
of the Xi are ≤ y for j ≥ k. Alternatively, note that the event [Yk ≤ y] can also be
expressed as a disjoint union

[Xi ≤ k and exactly k − 1 of the X1 , . . . , Xj−1 are ≤ y] .
j≥k

This gives

n 
FYk (y) = P [Xj ≤ y, ∀i ∈ I, Xi ≤ y, and ∀r ∈ [j − 1] \ I, Xr > y] .
j=k I∈( [j−1]
k−1)

(b) For r < s, Because Yr ≤ Ys , we have

[Yr ≤ yr ] ∩ [Ys ≤ ys ] = [Ys ≤ ys ] , ys ≤ y r .

By (30), for yr < ys ,

+ , + ,

n 
n
[Yr ≤ yr ] ∩ [Ys ≤ ys ] = [N (yr ) = j] ∩ [N (ys ) = m]
j=r m=s

n 
m
= [N (yr ) = j and N (ys ) = m],
m=s j=r

where the upper limit of the second union is changed from n to m because we must
have N (yr ) ≤ N (ys ). Now, to have N (yr ) = j and N (ys ) = m for m > j is to put j
of the Xi in (−∞, yr ], m − j of the Xi in (yr , ys ], and n − m of the Xi in (ys , ∞).

(c) Alternatively, for Xmax , Xmin , we have

FXmax ,Xmin (u, v) = P (Au ∩ Bvc ) = P (Au ) − P (Au ∩ Bv )

= P [∀i Xi ≤ u] − P [∀i v < Xi ≤ u]

where the second term is 0 when u < v. So,

FXmax ,Xmin (u, v) = FX1 ,...,Xn (u, . . . , u)

103
 when u < v. When v ≥ u, the second term can be found by (21) which gives

FXmax ,Xmin (u, v) = FX1 ,...,Xn (u, . . . , u) − (−1)|i:wi =v| FX1 ,...,Xn (w)
w∈S

|i:wi =v|+1
= (−1) FX1 ,...,Xn (w).
w∈S\{(u,...,u)}

where S = {u, v}n is the set of all 2n vertices w of the “box” × (ai , bi ]. The joint
i∈[n]
density is 0 for u < v.

11.20. For independent Xi ’s,

 
n
(a) fY (y) = fXi (xi )
x∈Sy i=1

(b) Two forms of marginal cdf:

+ ,⎛ ⎞

n   
FYk (y) = FXi (y) ⎝ (1 − FXi (y))⎠
j=k I∈([n]) i∈I r∈[n]\I
,⎛ ⎞
j
+

n   
= FXj (y) FXi (y) ⎝ (1 − FXr (y))⎠
j=k I∈( [j−1] i∈I r∈[j−1]\I
k−1 )

7 >
 0, u≤v
(c) FXmax ,Xmin (u, v) = FXk (u) − (FXk (u) − FXk (v)), v < u
k∈[n] k∈[n]

(d) The marginal cdf is

+ ,⎛ ⎞

n   
FYk (y) = FXi (y) ⎝ (1 − FXr (y))⎠ .
j=k I∈([n]) i∈I r∈[n]\I
j


(e) FXmin (v) = 1 − (1 − FXi (v)).
i

(f) FXmax (u) = FXi (u).
i

i.i.d.
11.21. Suppose Xi ∼ X with common density f and distribution function F .

(a) The joint density is given by

fY (y) = n!f (y1 )f (y2 ) . . . f (yn ), y1 ≤ y 2 ≤ · · · ≤ y n .

104
 If we define y0 = −∞, yk+1=∞ , n0 = 0, nk+1 = n + 1, then for k ∈ [n] and 1 ≤ n1 <
· · · < nk ≤ n, the joint density fYn1 ,Yn2 ,...,Ynk (yn1 , yn2 , . . . , ynk ) is given by
+ k , k
  (F (ynj+1 ) − F (ynj ))nj+1 −nj −1
n! f (ynj ) .
j=1 j=1
(n j+1 − n j − 1)!

In particular, for r < s, the joint density fYr ,Ys (yr , ys ) is given by
n!
f (yr )f (ys )F r−1 (yr )(F (ys ) − F (yr ))s−r−1 (1 − F (ys ))n−s
(r − 1)!(s − r − 1)!(n − s)!
[2, Theorem 5.4.6 p 230].
(b) The joint cdf FYr ,Ys (yr , ys ) is given by
⎧
⎨ FYs (ys ), ys ≤ yr ,

n  m
⎩
n!
j!(m−j)!(n−m)!
(F (yr ))j (F (ys ) − F (yr ))m−j (1 − F (ys ))n−m , yr < ys .
m=s j=r
?
n 0, u≤v
(c) FXmax ,Xmin (u, v) = (F (u)) − n .
(F (u) − F (v)) , v < u

0, u≤v
(d) fXmax ,Xmin (u, v) = n−2
n (n − 1) fX (u) fX (v) (F (u) − F (v)) , v<u
(e) Marginal cdf:
n  
 n
FYk (y) = (F (y))j (1 − F (y))n−j
j
j=k
n 
  
n−k  
j−1 k+m−1
= (F (y))k (1 − F (y))j−k = (F (y))k (1 − F (y))m
k−1 k−1
j=k m=0
! F (y)
n!
= tk−1 (1 − t)n−k dt.
(k − 1)! (n − k)! 0

Note that N (y) ∼ B (n, F (y)). The last equality comes from integrating the marginal
density fYk in (31) with change of variable t = F (y).
(i) FXmax (y) = (F (y))n and fXmax (y) = n (F (y))n−1 fX (y).
(ii) FXmin (y) = 1 − (1 − F (y))n and fXmin (y) = n (1 − F (y))n−1 fX (y).
(f) Marginal density:
n!
fYk (y) = (F (y))k−1 (1 − F (y))n−k fX (y) (31)
(k − 1)! (n − k)!
[2, Theorem 5.4.4 p 229]
Consider small neighborhood Δy around y. To have Yk ∈ Δy , we must have exactly
n−1k− 1 of them
one of the Xi ’s in Δy , exactly
n!
less than y, and exactly n − k of them
1
greater than y. There are n k−1 = (k−1)!(n−k)! = B(k,n−k+1) possible setups.

105
 (g) The range R is defined as R = Xmax − Xmin .

(i) For x > 0, fR (x) = n(n − 1) (F (u) − F (u − x))n−2 f (u − x)f (u)du.

(ii) For x ≥ 0, FR (x) = n (F (u) − F (u − x))n−1 f (u)du.

Both pdf and cdf above are derived by first finding the distribution of the range
conditioned on the value of the Xmax = u.

See also [4, Sec. 2.2] and [2, Sec. 5.4].

11.22. Let X1 , X2 , . . . , Xn be a random sample from a discrete distribution with pmf

pX (xi ) = pi , where
 x1 < x2 < . . . are the possible values of X in ascending order. Define
P0 = 0 and Pi = ik=1 pk , then
n  
 n
P [Yj ≤ xi ] = Pik (1 − Pi )n−k
k=j
k

and n  
 n  
P [Yj = xi ] = Pik (1 − Pi )n−k − Pi−1
k
(1 − Pi−1 )n−k .
k=j
k

Example 11.23. If U1 , U2 , . . . , Uk are independently uniformly distributed on the interval

0 to t0 , then they have joint pdf
1
 k tk0
, 0 ≤ ui ≤ t0
fU1k u1 =
0, otherwise

The order statistics τ1 , τ2 , . . . , τk corresponding to U1 , U2 , . . . , Uk have joint pdf

7
 k k!
tk0
, 0 ≤ t1 ≤ t2 ≤ · · · ≤ tk ≤ t0
fτ1k t1 =
0, otherwise

Example 11.24 (n = 2). Suppose U = max(X, Y ) and V = min(X, Y ) where X and Y

have joint cdf FXY .

FX,Y (u, u), u ≤ v,
FU,V (u, v) = ,
FX,Y (v, u) + FX,Y (u, v) − FX,Y (v, v), u > v
FU (u) = FX,Y (u, u),
FV (v) = FX (v) − FY (v) − FX,Y (v, v).

[9, Ex 7.5, 7.6]. The joint density is

fU,V (u, v) = fX,Y (u, v) + fX,Y (v, u), v < u.

106
The marginal densities is given by
 
∂  ∂ 
fU (u) = FX,Y (x, y) + FX,Y (x, y)
∂x x=u, ∂y x=u,
y=u y=u
!u !u
= fX,Y (x, u)dx + fX,Y (u, y) dy,
−∞ −∞

fV (v) = fX (v) + fY (v) − fU (v)

+   ,
∂  ∂ 
= fX (v) + fY (v) − FX,Y (x, y) + FX,Y (x, y)
∂x x=v, ∂y x=v,
y=v y=v
⎛ v ⎞
! ! v

= fX (v) + fY (v) − ⎝ fX,Y (x, u)dx + fX,Y (u, y) dy ⎠ .

−∞ −∞
! ∞ ! ∞
= fX,Y (v, y)dy + fX,Y (x, v)dx
v v

If, furthermore, X and Y are independent, then

FX (u) FY (u) , u≤v
FU,V (u, v) =
FX (v) FY (u) + FX (u) FY (v) − FX (v) FY (v) , u > v,
FU (u) = FX (u) FY (u) ,
FV (v) = FX (v) + FY (v) − FX (v) FY (v) ,
fU (u) = fX (u) FY (u) + FX (u) fY (u) ,
fV (v) = fX (v) + fY (v) − fX (v) FY (v) − FX (v) fY (v) .
If, furthermore, X and Y are i.i.d., then
FU (u) = F 2 (u) ,
FV (v) = 2F (v) − F 2 (v) ,
fU (u) = 2f (u) F (u) ,
fV (v) = 2f (v) (1 − F (v)) .
11.25. Let the Xi be i.i.d. with density f and cdf F . The range R is defined as R =
Xmax − Xmin .

0, x<0
FR (x) =
n (F (v) − F (v − x))n−1 f (v) dv, x ≥ 0

0,  x<0
fR (x) = n−2
n (n − 1) (F (v) − F (v − x)) f (v − x) f (v) dv, x > 0.
i.i.d.
For example, when Xi ∼ U(0, 1),
fR (x) = n (n − 1) xn−2 (1 − x) , 0≤x≤1
[15, Ex 9F p 322–323].

107
12 Convergences
Definition 12.1. A sequence of random variables (Xn ) converges pointwise to X if
∀ω ∈ Ω lim Xn (ω) → X (ω)
n→∞

Definition 12.2 (Strong Convergence). The following statements are all equivalent
conditions/notations for a sequence of random variables (Xn ) to converge almost surely to
a random variable X
a.s.
(a) Xn −−→ X

(i) Xn → X a.s.
(ii) Xn → X with probability 1.
(iii) Xn → X w.p. 1
(iv) lim Xn = Xa.s.
n→∞

a.s.
(b) (Xn − X) −−→ 0

(c) P [Xn → X] = 1
(' )*
(i) P ω : lim Xn (ω) = X (ω) = 1
n→∞
(' )c *
(ii) P ω : lim Xn (ω) = X (ω) =0
n→∞

(iii) P [Xn  X] = 0
(' )*
(iv) P ω : lim |Xn (ω) − X (ω)| = 0 = 1
n→∞
(' )*
(d) ∀ε > 0 P ω : lim |Xn (ω) − X (ω)| < ε =1
n→∞

12.3. Properties of convergence a.s.

a.s. a.s.
(a) Uniqueness: if Xn −−→ X and Xn −−→ Y , then X = Y a.s.
a.s. a.s.
(b) If Xn −−→ X and Yn −−→ Y , then
a.s.
(i) Xn + Yn −−→ X + Y
a.s.
(ii) Xn Yn −−→ XY
a.s. a.s.
(c) g continuous, Xn −−→ X ⇒ g (Xn ) −−→ g (X)

∞
a.s.
(d) Suppose that ∀ε > 0, P [|Xn − X| > ε] < ∞, then Xn −−→ X
n=1

a.s. 
(e) Let A1 , A2 , . . . be independent. Then, 1An −−→ 0 if and only if P (An ) < ∞
n

108
Definition 12.4 (Convergence in probability). The following statements are all equiv-
alent conditions/notations for a sequence of random variables (Xn ) to converges in proba-
bility to a random variable X
P
(a) Xn −
→X

(i) Xn →P X
(ii) p lim Xn = X
n→∞

P
(b) (Xn − X) −
→0

(c) ∀ε > 0 lim P [|Xn − X| < ε] = 1

n→∞

(i) ∀ε > 0 lim P ({ω : |Xn (ω) − X (ω)| > ε}) = 0

n→∞
(ii) ∀ε > 0 ∀δ > 0 ∃Nδ ∈ N such that ∀n ≥ Nδ P [|Xn − X| > ε] < δ
(iii) ∀ε > 0 lim P [|Xn − X| > ε] = 0
n→∞
(iv) The strict inequality between |Xn − X| and ε can be replaced by the correspond-
ing “non-strict” version.

12.5. Properties of convergence in probability

P P
(a) Uniqueness: If Xn −
→ X and Xn −
→ Y , then X = Y a.s.
P P
(b) Suppose Xn −
→ X, Yn −
→ Y , and an → a, then
P
(i) Xn + Yn −
→X +Y
P
(ii) an Xn −
→ aX
P
(iii) Xn Yn −
→ XY

(c) Suppose (Xn ) i.i.d. with distribution U [0, θ]. Let Zn = max {Xi : 1 ≤ i ≤ n}. Then,
P
Zn −
→θ
P P
(d) g continuous, Xn −
→ X ⇒ g (Xn ) −
→ g (X)
P
(i) Suppose that g : Rd → R is continuous. Then, ∀i Xi,n −
→ Xi implies
n→∞

P
g (X1,n , . . . , Xd,n ) −
→ g (X1 , . . . , Xd )
n→∞

P P
(e) Let g be a continuous function at c. Then, Xn −
→ c ⇒ g (Xn ) −
→ g (c)
P
(f) Fatou’s lemma: 0 ≤ Xn −
→ X ⇒ lim inf EXn ≥ EX
n→∞

P
(g) Suppose Xn −
→ X and |Xn | ≤ Y with EY < ∞, then EXn → EX

109
 P
(h) Let A1 , A2 , . . . be independent. Then, 1An −
→ 0 iff P (An ) → 0
P
(i) Xn −
→ 0 iff ∃δ > 0 such that ∀t ∈ [−δ, δ] we have ϕXn (t) → 1

Definition 12.6 (Weak convergence for probability measures). Let Pn and P be

 (d ≥ 1). The sequence Pn converges weakly to P if the se-
probability measure on R d

quence of real numbers gdPn → gdP for any g which is real-valued, continuous, and
bounded on Rd

12.7. Let (Xn ), X be Rd -valued random variables with distribution functions (Fn ), F ,
distributions (μn ) , μ, and ch.f. (ϕn ) , ϕ respectively.
The following are equivalent conditions for a sequence of random variables (Xn ) to
converge in distribution to a random variable X

(a) (Xn ) converges in distribution (or in law ) to X

(i) Xn ⇒ X
L
i. Xn −
→X
D
ii. Xn −
→X
(ii) Fn ⇒ F
i. FXn ⇒ FX
ii. Fn converges weakly to F
iii. lim P Xn (A) = P X (A) for every A of the form A = (−∞, x] for which
n→∞
P {x} = 0
X

(iii) μn ⇒ μ
i. P Xn converges weakly to P X

(b) Skorohod’s theorem: ∃ random variables Yn and Y on a common probability space

D D
(Ω, F, P ) such that Yn = Xn , Y = X, and Yn → Y on (the whole) Ω

(c) lim Fn = F for all continuity points of F

n→∞

(i) FXn (x) → FX (x) ∀x such that P [X = x] = 0

(d) ∃ a (countable) set D dense in R such that Fn (x) → F (x) ∀x ∈ D

(e) Continuous Mapping theorem: lim Eg (Xn ) = Eg (X)for all g which is real-
n→∞
valued, continuous, and bounded on Rd

(i) lim Eg (Xn ) = Eg (X) for all bounded real-valued function g such that P [X ∈ Dg ] =
n→∞
0 where Dg is the set of points of discontinuity of g
(ii) lim Eg (Xn ) = Eg (X) for all bounded Lipschitz continuous functions g
n→∞

(iii) lim Eg (Xn ) = Eg (X) for all bounded uniformly continuous functions g
n→∞

110
 (iv) lim Eg (Xn ) = Eg (X) for all complex-valued functions g whose real and imag-
n→∞
inary parts are bounded and continuous

(f) Continuity Theorem: ϕXn → ϕX

(i) For nonnegative random variables: ∀s ≥ 0 LXn (s) → LX (s) where LX (s) =
Ee−sX

Note that there is no requirement that (Xn ) and X be defined on the same probability
space (Ω, A, P )

12.8. Continuity Theorem: Suppose lim ϕn (t) exists ∀t; call this limit ϕ∞ Further-
n→∞
more, suppose ϕ∞ is continuous at 0. Then there exists ∃ a probability distribution μ∞
such that μn ⇒ μ∞ and ϕ∞ is the characteristic function of μ∞

12.9. Properties of convergence in distribution

(a) If Fn ⇒ F and Fn ⇒ G, then F = G

(b) Suppose Xn ⇒ X

(i) If P [X is a discontinuity point of g] = 0, then g (Xn ) ⇒ g (X)

(ii) Eg (Xn ) → Eg (X) for every bounded real-valued function g such that P [X ∈ Dg ] =
0 where Dg is the set of points of discontinuity of g
i. g (Xn ) ⇒ g (X) for g continuous.
P
(iii) If Yn −
→ 0, then Xn + Yn ⇒ X
(iv) If Xn − Yn ⇒ 0, then Yn ⇒ X

(c) If Xn ⇒ a and g is continuous at a, then g (Xn ) ⇒ g (a)

(d) Suppose (μn ) is a sequence of probability measures on R that are all point masses
with μn ({αn }) = 1. Then, μn converges weakly to a limit μ iff αn → α; and in this
case μ is a point mass at α

(e) Scheffé’s theorem:

(i) Suppose P Xn and P X have densities δn and δ w.r.t. the same measure μ. Then,
δn → δ μ-a.e. implies
i. ∀B ∈ BR P Xn (E) → P X (E)
• FXn → FX
• FXn ⇒ FX
 
ii. Suppose g is bounded. Then, g (x) P Xn (dx) → g (x) P X (dx). In Equiv-
alently, E [g (Xn )] → E [g (X)] where the E is defined with respect to appro-
priate P .
(ii) Remarks:

111
 i. For absolutely continuous random variables, μ is the Lebesgue measure, δ
is the probability density function.
ii. For discrete random variables, μ is the counting measure, δ is the probability
mass function.

(f) Normal r.v.

(i) Let Xn ∼ N (μn , σn2 ). Suppose μn → μ ∈ R and σn2 → σ 2 ≥ 0. Then, Xn ⇒

N (μ, σ 2 )
(ii) Suppose that Xn are normal random variables and let Xn ⇒ X. Then, (1) the
mean and variance of Xn converge to some limit m and σ 2 . (2) X is normal
with mean m and variance σ 2

(g) Xn ⇒ 0 if and only if ∃δ > 0 such that ∀t ∈ [−δ, δ] we have ϕXn (t) → 1
12.10. Relationship between convergences

p
X 
nk
a.s.
L
X n  oX X n  oX
p
Xn d Y  L
f ( ˜ ) continuous

p p
 Xn o X
X n d Y  Lp X 
nk

P
X n  oX

f ( ˜ ) continuous c P  X c 1 f X n 
o fX


X n  oX
f ( ˜ ) continuous

^
x  x : F  x   F  x ` x  D dense in 
dense in 
lim Fn  x  F  x
n of

Figure 22: Relationship between convergences

a.s. P
(a) For discrete probability space, Xn −−→ X if and only if Xn −
→X
P Lp
(b) Suppose Xn −
→ X, ∀n |Xn | ≤ Y , Y ∈ Lp . Then, X ∈ Lp and Xn −→ X
a.s. Lp
(c) Suppose Xn −−→ X, ∀n |Xn | ≤ Y , Y ∈ Lp . Then, X ∈ Lp and Xn −→ X
P D
(d) Xn −
→ X ⇒ Xn −
→X

112
 D P
(e) If Xn −
→ X and if ∃a ∈ RX = a a.s., then Xn −
→X
P D
(i) Hence, when X = a a.s., Xn −
→ X and Xn −
→ X are equivalent.
See also Figure 22.
Example 12.11.

0, ω ∈ 0, 1 − n12
(a) Ω = [0, 1] , P is uniform on [0,1]. Xn (ω) =
en , ω ∈ 1 − n12 , 1
Lp
(i) Xn  0
a.s.
(ii) Xn −−→ 0
P
(iii) Xn −
→0
(b) Let Ω = [0, 1] with uniform probability distribution. Define Xn (ω) = ω +1[ j−1 , j ] (ω) ,
@ √ A k k

where k = 12 1 + 8n − 1 and j = n − 12 k (k − 1). The sequence of intervals

 j−1 j
k
, k under the indicator function is shown in figure 23.

>0,1@

ª 1º ª1 º
« 0, 2 » o « 2 ,1»
¬ ¼ ¬ ¼

ª 1 º ª1 2 º ª 2 º
« 0, 3 » o « 3 , 3 » o « 3 ,1»
¬ ¼ ¬ ¼ ¬ ¼

ª 1º ª1 1º ª1 3º ª3 º
« 0, 4 » o « 4 , 2 » o « 2 , 4 » o « 4 ,1»
¬ ¼ ¬ ¼ ¬ ¼ ¬ ¼

Figure 23: Diagram for example (12.11)

Let X (ω) = ω.
(i) The sequence of real numbers (Xn (ω)) does not converge for any ω
Lp
(ii) Xn −→ X
P
(iii) Xn −
→X
1
n
, w.p. 1 − 1
n
(c) Xn = .
n, w.p. n1
P
(i) Xn −
→0
Lp
(ii) Xn  0

113
12.1 Summation of random variables

n
Set Sn = Xi .
i=1

12.12 (Markov’s theorem; Chebyshev’s inequality). For finite variance (Xi ), if

P
1
n2
Var Sn → 0, then n1 Sn − n1 ESn −
→0

n
• If (Xi ) are pairwise independent, then 1
n2
Var Sn = 1
n2
Var Xi See also (12.16).
i=1

12.2 Summation of independent random variables


n
12.13. For independent Xn , the probability that Xi converges is either 0 or 1.
i=1

12.14 (Kolmogorov’s
 Var Xn SLLN). Consider a sequence (Xn ) of independent random vari-
a.s.
ables. If n2
< ∞, then n1 Sn − n1 ESn −−→ 0
n


n
a.s.
• In particular, for independent Xn , if 1
n
EXn → a or EXn → a, then n1 Sn −−→ a
i=1

12.15. Suppose that X1 , X2 , . . . are independent and that the series X = X1 + X2 + · · ·

converges a.s. Then, ϕX = ϕX1 ϕX2 · · ·

n
12.16. For pairwise independent (Xi ) with finite variances, if 1
n2
Var Xi → 0, then
i=1
P
1
S
n n
− n1 ESn −
→0
(a) Chebyshev’s Theorem: For pairwise independent (Xi ) with uniformly bounded
P
variances, then n1 Sn − n1 ESn −
→0

12.3 Summation of i.i.d. random variable

Let (Xi ) be i.i.d. random variables.
5 n  6
1  

12.17 (Chebyshev’s inequality). P  n 
Xn − EX1  ≥ ε ≤ 1 Var[X1 ]
ε2 n
i=1

• The Xi ’s don’t have to be independent; they only need to be pairwise uncorrelated.

12.18 (WLLN). Weak Law of Large Numbers:
(a) L2 weak law: (finite variance) Let (Xi ) be uncorrelated (or pairwise independent)
random variables

n
(i) V arSn = Var Xi
i=1

n
P
(ii) If 1
n2
Var Xi → 0, then n1 Sn − n1 ESn −
→0
i=1

114
 
n
L2
(iii) If EXi = μ and V arXi ≤ C < ∞. Then, 1
n
Xi −→ μ
i=1

(b) Let (Xi ) be i.i.d. random variables with EXi = μ and V arXi = σ 2 < ∞. Then,
5 n  6
1  

(i) P  n 
Xn − EX1  ≥ ε ≤ ε12 Var[X1]
n
i=1

n
L2
(ii) 1
n
Xi −→ μ
i=1

(The fact that σ 2 < ∞ implies μ < ∞).


n
L2 
n
P 
n
D
(c) 1
n
Xi −→ μ implies 1
n
Xi −
→ μ which in turn imply 1
n
Xi −
→μ
i=1 i=1 i=1

(n) P
(d) If Xn are i.i.d. random variables such that lim tP [|X1 | > t] → 0, then n1 Sn −EX1 −
→
t→∞
0

(e) Khintchine’s theorem: If Xn are i.i.d. random variables and E |X1 | < ∞, then
(n) P
EX1 → EX1 and n1 Sn −
→ EX1 .

• No assumption about the finiteness of variance.

12.19 (SLLN).

(a) Kolmogorov’s
 Var Xn SLLN : Consider a sequence (Xn ) of independent random variables.
a.s.
If n2
< ∞, then n1 Sn − n1 ESn −−→ 0
n


n
a.s.
• In particular, for independent Xn , if 1
n
EXn → a or EXn → a, then n1 Sn −−→ a
i=1


n
a.s.
(b) Khintchin’s SLLN : If Xi ’s are i.i.d. with finite mean μ, then 1
n
Xi −−→ μ
i=1

(c) Consider the sequence (Xn ) of i.i.d. random variables. Suppose EX1− < ∞ and
a.s.
EX1+ = ∞, then n1 Sn −−→ ∞
a.s.
• Suppose that Xn ≥ 0 are i.i.d. random variables and EXn = ∞. Then, n1 Sn −−→
∞

12.20 (Relationship between LLN and the convergence of relative  frequency to

the probability). Consider i.i.d. Zi ∼ Z. Let Xi = 1A (Zi ). Then, n Sn = n ni=1 1A (Zi ) =
1 1

rn (A), the relative

 frequency of an event A. Via LLN and appropriate conditions, rn (A)
converges to E n1 ni=1 1A (Zi ) = P [Z ∈ A].

115
12.4 Central Limit Theorem (CLT)
Suppose that (Xk )k≥1 isa sequence of i.i.d. random variables with mean m and variance
0 < σ 2 < ∞. Let Sn = nk=1 Xk .

12.21 (Lindeberg-Lévy theorem).

Sn −mc

n
Xk −m
(a) √
σ n
= √1
n σ
⇒ N (0, 1).
k=1

−mc

n
(b) Sn√
n
= √1
n
(Xk − m) ⇒ N (0, σ).
k=1

Xk −m iid 
To ∼ Z and Yn = nk=1 Zk . Then, EZ = 0, Var Z = 1, and ϕYn (t) =
 see this, n let Zk = σ
ϕZ ( √tn ) . By approximating ex ≈ 1 + x + 12 x2 . We have ϕX (t) ≈ 1 + jtEX − 12 t2 E [X 2 ]
(see also (22)) and  n
1 t2 t2
ϕYn (t) = 1 − → e− 2 .
2n
• The case of Bernoulli(1/2) was derived by Abraham de Moivre around 1733. The
case of Bernoulli(p) for 0 < p < 1 was considered by Pierre-Simon Laplace [9, p 208].

12.22 (Approximation of densities and pmfs using the CLT). Approximate the
distribution of Sn by N (nm, nσ 2 ).
 
• FSn (s) ≈ Φ s−nm
√
σ n
 2
−1 x−nm
√
• fSn (s) ≈ √ 1√ e 2
2πσ n
σ n

• If the Xi are integer-valued, then

5 6  2
1 1 1 − 12 k−nm
√
P [Sn = k] = P k − < Sn ≤ k + ≈√ √ e σ n
2 2 2πσ n

[9, eq (5.14) p 213].

The approximation is best for s, k near nm [9, p 211].

√ 1
• The approximation n! ≈ 2πnn+ 2 e−n can be derived from approximating the density
n−1 e−s
of Sn when Xi ∼ E(1). We know that fSn (s) = s(n−1)! . Approximate the density at
√ 1
s = n, gives (n − 1)! ≈ 2πnn− 2 e−n . Multiply through by n. [9, Ex 5.18 p 212]
• See also normal approximation for the binomial in (13).

116
13 Conditional Probability and Expectation
13.1 Conditional Probability
Definition 13.1. Suppose conditioned on X = x, Y has distribution Q. Then, we write
Y |X = x ∼ Q [21, p 40]. It might be clearer to write P Y |X=x = Q.
13.2. Discrete random variables
pX,Y (x,y)
(a) pX|Y (x|y) = pY (y)

(b) pX,Y (x, y) = pX|Y (x|y)pY (y) = pY |X (y|x)pX (x)


(c) The law of total probability:
 P [Y ∈ C] = x P [Y ∈ C|X = x]P [X = x]. In
particular, pY (y) = x pY |X (y|x)pX (x).
pY |X (y|x)pX (x)
(d) pX|Y (x|y) =   
x pY |X (y|x )pX (x )

(e) If Y = X + Z where X and Z are independent, then

• pY |X (y|x) = pZ (y − x)

• pY (y) = x pZ (y − x)pX (x)
• pX|Y (x|y) =  pZ (y−x)pX (x)  .
x pZ (y−x )pX (x )

(f) The substitution law of conditional probability:

P [g(X, Y ) = z|X = x] = P [g(x, Y ) = z|X = x].
• When X and Y are independent, we can “drop the conditioning”.
13.3. Absolutely continuous random variables
fX,Y (x,y)
(a) fY |X (y|x) = fY (y)
.
y
(b) FY |X (y|x) = −∞
fY |X (t|x)dt

13.4. P [(X, Y ) ∈ A] = E [P [(X, Y ) ∈ A|X]] = A
fX,Y (x, y)d(x, y)

g(z,x)
13.5. ∂
F
∂z Y |X
(g (z, x) |x ) = ∂
∂z
fY |X (y |x )dy = fY |X (g (z, x) |x ) ∂z
∂
g (z, x)
−∞
P [X≤x,y−Δ<Y ≤y]

13.6. FX|Y (x |y ) = P [X ≤ x |Y = y ] = lim P [y−Δ<Y ≤y]
= E 1(−∞,x] (X) |Y = y .
Δ→0
fX,A (x)
13.7. Define fX|A (x) to be P (A)
. Then,
x
fX,A (x ) dx
x−Δ fX,A (x) P (A) fX|A (x)
P [A |X = x ] = lim x = = .
Δ→0 fX (x) fX (x)
fX (x ) dx
x−Δ

13.8. For independent Xi , P [∀i Xi ∈ Bi |∀i Xi ∈ Ci ] = P [Xi ∈ Bi |Xi ∈ Ci ].
i

117
13.2 Conditional Expectation

Definition 13.9. E [g(Y )|X = x] = y g(y)pY |X (y|x).

• In particular, E [Y |X = x] = y ypY |X (y|x).
• Note that E [Y |X = x] is a function of x.
13.10. Properties of conditional expectation.
(a) Substitution law for conditional expectation: E [g(X, Y )|X = x] = E [g(x, Y )|X = x]
(b) E [h(X)g(Y )|X = x] = h(x)E [g(Y )|X = x].
(c) (The Rule of Iterated Expectations) EY = E [E [Y |X]].
(d) Law of total probability for expectation:
E [g(X, Y )] = E [E [g(X, Y )|X]] .
(i) E [g(Y )] = E [E [g(Y )|X]].

(ii) FY (y) = E FY |X (y|X) . (Take g(z) = 1(−∞,y] (z)).
(e) E [g (X) h(Y ) |X = x ] = g (x) E [h(Y ) |X = x ]
(f) E [g (X) h(Y )] = E [g (X) E [h(Y ) |X ]]
(g) E [X + Y |Z = z ] = E [X |Z = z ] + E [Y |Z = z ]
(h) E [AX|z] = AE [X|z].
(i) E [(X − E [X |Y ])| Y ] = 0 and E [X − E [X |Y ]] = 0
(j) ming(x) E [(Y − g(X))2 ] = E [(Y − E [Y |X])2 ] where g ranges over all functions. Hence,
E [Y |X] is sometimes called the regression of Y on X, the “best” predictor of Y con-
ditional on X.
Definition 13.11. The conditional variance is defined as
!
Var[Y |X = x] = (y − m(x))2 f (y|x)dy

where m(x) = E [Y |X = x].

13.12. Properties of conditional variance
(a) Var Y = E [Var[Y |X]] + Var[E [Y |X]].
In other words, suppose given X = x, the mean and variance of Y is m(x), v(x). Then,
the variance of Y is Var Y = E [v(X)] + Var[m(X)]. Recall that for any function g,
we have E [g(Y )] = E [E [g(Y )|X]]. Because EY is just a constant, say μ, we can
define g(y) = (y − μ)2 which then implies Var Y = E [E [g(Y )|X]]. Note, however,
that E [g(Y )|X] and Var[Y |X] are not the same. Suppose conditioned
 on X, Y has
distribution Q with mean m(x)  and variance v(x). Then, v(x) = (y − m(x))2 Q(dy).
However, E [g(Y )|X = x] = (y − μ) Q(dy); note the use of μ in stead of m(x).
2

Therefore, in general, Var Y = E [Var[Y |X]].

118
 • All three terms in the expression are nonnegative. Var Y is an upper bound for
each of the terms on the RHS.

(b) Suppose N (X, Z), then Var[X + N |Z] = Var[X|Z] + Var N .

|=
(c) Var[AX|z] = A Var[X|z]AH .

13.13. Suppose E [Y |X] = X. Then, Cov [X, Y ] = Var X. See also (5.19). This is also
true for Y = X + N with X N and N is zero-mean noise.

|=
Definition 13.14. μn [Y |X ] = E [(Y − E [Y |X ])n |X ]

13.15. Properties

(a) μ3 [Y ] = E [μ3 [Y |X ]] + μ3 [E [Y |X ]]
μ4 [Y ] = E [μ4 [Y |X ]] + 6E [Var [Y |X ]] Var [E [Y |X ]] + μ4 [E [Y |X ]]

13.3 Conditional Independence

13.16. The following statements are equivalent:conditions for X1 , X2 , . . . , Xn to be mu-
tually independent conditioning on Y (a.s.).

n
(a) p (xn1 |y ) = p (xi |y ).
i=1
  
(b) ∀i ∈ [n] \ {1} p xi xi−1
1 , y = p (xi |y ).
(c) ∀i ∈ [n] Xi and the vector (Xj )[n]\{i} are independent conditioning on Y .

Example 13.17. Suppose X and Y are independent. Conditioned on another random

variable Z, it is not true in general that X and Y are still independent. See example
(4.45). Recall that Z = X ⊕ Y which can be rewritten as Y = X ⊕ Z. Hence, when Z = 0,
we must have Y = X.

13.18. Suppose we know that fX|Y,Z (x|y, z) = g(x, y); that is fX|Y,Z does not depend on
z. Then, conditioned on Y , X and Z are independent. In which case,

fX|Y,Z (x |y, z ) = fX|Y (x |y ) = g (x, y) .

13.19. Suppose we know that fZ|V,U1 ,U2 (z |v, u1 , u2 ) = fZ|V (z |v ) for all z, v, u1 , u2 , then
conditioned on V , we can conclude that Z and (U1 , U2 ) are independent. This further
implies Z and Ui are independent. Moreover,

fZ|V,U1 ,U2 (z |v, u1 , u2 ) = fZ|V,U1 (z |v, u1 ) = fZ|V,U1 (z |v, u2 ) = fZ|V (z |v ) .

119
14 Real-valued Jointly Gaussian
Definition 14.1. Random vector Rd is jointly Gaussian or jointly normal if and
only if ∀v ∈ Rd , the random variable v T X is Gaussian.
• In order for this definition to make sense when v = 0 or when X has a singular
covariance matrix, we agree that any constant random variable is considered to be
Gaussian.
• Of course, the mean and variance are v T EX and v T ΛX v, respectively.
If X is a Gaussian random vector with mean vector m and covariance matrix Λ, we write
X ∼ N (m, Λ).
14.2. Properties of jointly Gaussian random vector X ∼ N (m, Λ)

(a) m = EX, Λ = Cov [X] = E (X − EX)(X − EX)T .
1
√
T
Λ−1 (x−m)
(b) fX (x) = n
1
e− 2 (x−m) .
(2π) 2 det(Λ)

• To remember the form of the above formulas, both exponents have to be scalar.
So, we better have (x − m)T Λ−1 (x − m) instead of having the transpose on the
last term. To make this 
more
 clear, set Λ = I, then we must have a dot product.
T
Note also that v Av = vk Ak v .
k 
• The above formula can be derived by starting form a random vector Z whose
1
components are i.i.d. N (0, 1). Let X = CX2 Z + m. Use (28) and (32).
i.i.d.
• For Xi ∼ N (mi , σi2 ),
  x−mi  2
1 − 12
fX (x) = n  e .
i σi

(2π) 2 i σi
i.i.d.
In particular, if Xi ∼ N (0, 1),
1 − 12 xT x
fX (x) = n e (32)
(2π) 2
n " "
• (2π) 2 det (Λ) = det (2πΛ)
• The Gaussian density is constant on the “ellipsoids” centered at m,
 −1

x ∈ Rn : (x − m)T CX (x − m) = constant .
   
 
T m− 1 v T Λv j vi EXi − 12 vk v Cov[Xk ,X ]
(c) ϕX (v) = ejv 2 =e i k  .
• This can be derived from definition (14.1) by noting that
⎡ ⎤

Y
( T *
⎢ T ⎥
ϕX (v) = E ejv X = E ⎣ej1 v X ⎦

is simply ϕY (1) where Y = v T X which by definition is normal.

120
 (d) Random vector Rd is Jointly Gaussian if and only if ∀v ∈ Rd , the random variable
v T X is Gaussian.

• Independent Gaussian random variables are jointly Gaussian.

(e) Joint normality

 is preserved
 under linear transformation: suppose Y = AX + b, then
Y ∼ N Am + b, AΛA . T

(f) If (X, Y ) jointly Gaussian, then X and Y are independent if and only if Cov [X, Y ] =
0. Hence, uncorrelated jointly Gaussian random variables are independent.

(g) Note that the joint density does not exists when the covariance matrix Λ is singular.

(h) For i.i.d. N (μ, σ 2 ):

?
1 1 2
fX1n (xn1 ) = n exp − 2 x − μ
(2π) 2 σ n 2σ
7 >
1  2 μ 
n n
1 μ2
= n exp − 2 x + xi − 2
(2πσ 2 ) 2 2σ i=1 i σ 2 i=1 2σ

(i) Third order joint Gaussian moments are 0:

E [(Xi − EXi ) (Xj − EXj ) (Xk − EXk )] = 0 ∀ i, j, k not necessarily distinct.


In particular, E (X − EX)3 = 0.
(j) Isserlis’s Theorem: Any forth-order central moment of jointly Gaussian r.v. is
expressible as the a sum of all possible products of pairs of their covariances:

E [(Xi − EXi ) (Xj − EXj ) (Xk − EXk ) (X − EX )]

= Cov [Xi , Xj ] Cov [Xk , X ] + Cov [Xi , Xk ] Cov [Xj , X ] + Cov [Xi , X ] Cov [Xj , Xk ] .

1 4
Note that 2 2
= 3.

• In particular, E (X − EX)4 = 3σ 4 .

(k) To generate N (m, Λ). First, by spectral theorem, Λ = V DV T where V is orthogo-

nal matrix whose columns are eigenvectors of Λ and D is diagonal matrix with the
eigenvalues of Λ. The random variable we want is V X + m where X ∼ N (0, D).

14.3. For bivariate normal, fX,Y (x, y) is

⎧  2     2 ⎫
⎪
⎨ x−EX
− 2ρ σX
x−EX y−EY y−EY
+ σY ⎪
⎬
1 σX σY
" exp − ,
2πσX σY 1 − ρ2 ⎪
⎩ 2 (1 − ρ2 ) ⎪
⎭

Cov(X,Y )
where ρ = σX σ Y
∈ [−1, 1]. Here, x, y ∈ R.

121
  
• fX,Y (x, y) = 1
σ X σY
ψρ x−mX y−mY
σX
, σY
 2  2
• fX,Y is constant on ellipses of the form x
σX
+ y
σY
= r2 .
 2
  2

σX Cov [X, Y ] σX ρσX σY
• Λ= = .
Cov [X, Y ] σY2 ρσX σY σY2
• The following are equivalent:

(a) ρ = 0
(b) Cov [X, Y ] = 0
(c) X and Y are independent.

• |ρ| = 1 if and only if (X − EX) = k (Y − EY ). In which case

◦ ρ= k
|k|
= sign (k)
◦ |k| = σσXY
"
• Suppose fX,Y (x, y) only depends on x2 + y 2 and X and Y are independent, then
X and Y are normal with zero mean and equal variance.
 
• X|Y ∼ N ρ σσXY (Y − mY ) + mX , σX2
(1 − ρ2 )
 
Y |X ∼ N ρ σσXY (X − mX ) + mY , σY2 (1 − ρ2 )

• The standard bivariate density is defined as

1 − 1
(u2 −2ρuv+v 2 )
ψρ (u, v) = " e 2(1−p)2
2π 1 − ρ2
+ , + ,
1 v − ρu 1 u − ρv
= ψ(u) " ψ " =" ψ " ψ(v)
 1 − ρ2 1 − ρ2 1 − ρ2 1 − ρ2 
fU (u)       fV (v)
fV |U (v|u) fU |V (u|v)
V |U ∼N (ρU,1−ρ2 ) U |V ∼N (ρV,1−ρ2 )

[9, eq. (7.22) p 309, eq. (7.23) p 311, eq. (7.26) p 313]. This is the joint density of
U, V where U, V ∼ N (0, 1) with Cov [U, V ] = E [U V ] = ρ.

◦ The general bivariate Gaussian pair is obtained from the transformation

        
X σX U + mX σX 0 U mX
= = + .
Y σY V + mY 0 σY V mY

◦ fU |V (u|v) is N (ρv, 1 − ρ2 ). In other words, U |V ∼ N (ρV, 1 − ρ2 ).

14.4 (Conditional Gaussian).

122
   
X   μX 
ΛX ΛXY
(a) Suppose (X, Y ) are jointly Gaussian; that is Y ∼ N μY
, . Then,
  ΛY X ΛY
fX|Y (x |y ) is N E [X |y ] , ΛX|y where E [X |y ] = μX +ΛXY Λ−1
Y (y − μY ) and ΛX|y =
−1
ΛX − ΛXY ΛY ΛY X .

• Note the direction of the formula for ΛX|y .

⎛ ⎞
ΛX → ΛXY
⎝ ↓ ⎠
ΛY X ← ΛY

(b) Suppose (X,

 Y, W ) are jointly Gaussian with W (X, Y ) . Set V = BX + W . Then,

|=

V |y ∼ N E [V |y ] , ΛV |y where E [V |y ] = BE [X |y ] + EW and ΛV |y = BΛX|y B T +
ΛW .

15 Bayesian Detection and Estimation

Consider a pair of random vectors Θ and Y , where Θ is not observed, but Y is observed.
We know the joint distribution of the pair (Θ, Y ) which is usually given in the form of the
prior distribution pΘ (θ) and the conditional distribution pY |Θ (y|θ). By an estimator of Θ
based on Y , we mean a function g such that Θ̂(Y ) = g(Y ) is our estimate or “guess” of
the value of Θ.
15.1 (Orthogonality Principle). Let D be a collection of random vectors with the same
dimension as Θ. For a random vector Z, suppose that

∀X ∈ D, Z − X ∈ D. (33)

If 
E X T (Θ − Z) = 0, ∀X ∈ D, (34)
then   
E |Θ − X|2 = E |Θ − Z|2 + E |Z − X|2 + 2 E[(Z
 −
 X) (Θ − Z)],
T

∈D
  
=0

which implies  
E |Θ − Z|2 ≤ E |Θ − X|2 , ∀X ∈ D.
• If D is a subspace and Z ∈ D, then (33) is automatically satisfied.
• (34) says that the vector Θ − Z is orthogonal to all vectors in D.
Example 15.2. Suppose Θ and N are independent Poisson random variables with respec-
tive parameters λ and μ. Let Z = Θ + N .
• Y is P(λ + μ)
 
• Conditioned on Y = y, Θ is B y, λ
λ+μ
.

123
 • Θ̂MMSE (Y ) = E [Θ|Y ] = λ
λ+μ
Y .
• Var[Θ|Y ] = Y λμ
(λ+μ)2
, and MSE = E [Var[Θ|Y ]] = λμ
λ+μ
< Var Y = λ + μ.

See also [7, Q 15.17].

 T  
15.3 (Weighted Error). Suppose we define the error by E = Θ − Θ̂ (Y ) W Θ − Θ̂ (Y )
for some positive definite matrix W . (Note that the usual MSE use W = I.) The MSE
EE is uniquely minimized by the MMSE estimator Θ̂(Y ) = E [Θ|Y ]. The resulting MSE
is E (Θ − E [Θ|Y ])T W (Θ − E [Θ|Y ]) . (  *

In fact, for any function, g(Y ), the conditional weight error E (Θ − g (Y ))T W (Θ − g (Y )) Y
is given by
(  *

E (Θ − E [Θ |Y ]) W (Θ − E [Θ |Y ]) Y + (E [Θ |Y ] − g (Y ))T W (E [Θ |Y ] − g (Y )) .
T

Hence, for each Y , it is minimized by having g(Y ) = E [Θ|Y ].

15.4 (Linear minimum mean-squared-error  estimator). A linear MMSE estimator

Θ̂LMMSE = gLMMSE (Y ) minimizes the MSE E |Θ − g(Y )|2 among all affine estimators of
the form g(y) = Ay + b.

(a) It is sometimes called Wiener filters.

(b) The scalar linear (affine) MMSE estimator is given by

Cov [Y, Θ]
Θ̂LMMSE (Y ) = EΘ + (Y − EY ) .
Var Y
• To see this in Hilbert space, note that we want the orthogonal projection of Θ
onto the subspace spanned by two elements: Y and 1. The orthogonal basis of
the subspace is {1, Y − EY }. Hence, the orthogonal projection is

(Θ, 1) (Θ, Y − EY )
+ (Y − EY ) .
(1, 1) (Y − EY, Y − EY )

• The above discussion suggest alternative ways of arriving at the LMMSE by

finding a, b in Θ̂(Y ) = aY + b such that the error E = Θ − Θ̂(Y ) is orthogonal
to both 1 and Y − EY . The condition (E, 1) = 0 requires Θ̂(Y ) to be unbiased.
The condition (E, Y − EY ) = 0 gives a = Cov[Y,Θ]
Var Y
.

(c) The vector linear (affine) MMSE estimator is given by

Θ̂LMMSE (Y ) = EΘ + ΣΘY Σ−1

Y (Y − EY )

and ( *
MMSE = Cov Θ − Θ̂LMMSE (Y ) = ΣΘ − ΣΘY Σ−1
Y ΣY Θ .

124
In fact, the optimal choice of A is any solution of

AΣY = ΣΘY .

In which case,
( *
Cov Θ − Θ̂LMMSE (Y ) = ΣΘ − AΣY Θ − ΣΘY AT + AΣY AT
= ΣΘ − ΣΘY AT = ΣΘ − AΣY Θ
= ΣΘ − AΣY AT .

When ΣY is invertible, A = ΣΘY Σ−1

Y . When ΣY is singular, see [9, Q8.38 p 359].

• The MSE can be rewrite as


E ((Θ − EΘ) − A(Y − EY ))2 + |EΘ − AEY − b|2 ,

which show that the optimal choice of b is b = EΘ − AEY . This is the b which
makes the estimator unbiased.
• Fix A. Let Θ̃ = Θ − EΘ and Ỹ = Y − EY . Suppose for any matrix B, we have
5 T   6
E B Ỹ Θ̃ − AỸ = 0 .

if and only if, for all matrix B,

5 2 6 5 2 6
   
E Θ̃ − AỸ  ≤ E Θ̃ − B Ỹ  .

In which case,
5 2 6 5 2 6 5 2 6
     
E Θ̃ − B Ỹ  = E Θ̃ − AỸ  + E (A − B)Ỹ  .

• Additive Noise in 1-D: Y = Θ + N where Cov [Θ, N ] = 0.

Cov [Θ, Y ]
Θ̂LMMSE (Y ) = EΘ + (Y − EY )
Var Y
Var Θ
= EΘ + (Y − (EΘ + EN ))
Var Θ + Var N
SNR
= EΘ + (Y − (EΘ + EN )) .
1 + SNR
and
Cov2 [Θ, Y ] Var Θ Var N
MMSE = Var Θ − = .
Var Y Var Θ + Var N

125
A More Math
A.1 Inequalities
A.1. By definition,
∞  N
• n=1 an = n∈N an = lim n=1 an and N →∞
∞  N
• n=1 an = n∈N an = lim n=1 an .
N →∞

A.2. For |x| ≤ 0.5, we have

2
ex−x ≤ 1 + x ≤ ex .
This is because
x − x2 ≤ ln (1 + x) ≤ x,
which is semi-proved by the plot in figure 24. .
-0.6838
1

‫ݔ‬
0.5

0
lnሺ1 + ‫ݔ‬ሻ

-0.5
‫ ݔ‬െ ‫ݔ‬2

-1 ‫ݔ‬
‫ݔ‬+1
-1.5
-0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8
‫ݔ‬

Figure 24: Bounds for ln(1 + x) when x is small.

A.3. Let αi and βi be complex numbers with |αi | ≤ 1 and |βi | ≤ 1. Then,
 
m m   m
 
 αi − βi  ≤ |αi − βi |.
 
i=1 i=1 i=1

In particular, |αm − β m | ≤ m |α − β|.


rn
A.4. Consider a triangular array of real numbers (xn,k ). Suppose (i) xn,k → x and (ii)
k=1

rn
x2n,k → 0. Then,
k=1

rn
(1 + xn,k ) → ex .
k=1

126
Proof. Use (A.2).
n
Suppose the sum rk=1 |xn,k | converges as n → ∞. Conditions (i) and (ii) is equivalent
to conditions (i) and (iii) where (iii) is the requirement that max |xk,n | → 0 as n → ∞.
k∈[rn ]
rn
Proof. Suppose k=1 |xn,k | → x0 . To show that (iii) implies (ii) under (i), let an =
max |xk,n |. Then,
k∈[rn ]

rn 
rn
0≤ x2n,k ≤ an |xn,k | → 0 × x0 = 0.
k=1 k=1

On the other hand, suppose we have (i) and (ii). Given any ε > 0, by (ii), ∃n0 such that

rn 
rn
∀n ≥ n0 , x2n,k ≤ ε2 . Hence, for any k, x2n,k ≤ x2n,k ≤ ε2 and hence |xn,k | ≤ ε which
k=1 k=1
implies an ≤ ε.

rnNote that when the xk,n are non-negative, condition (i) already implies that the sum
k=1 |xn,k | converges as n → ∞.
 n
A.5. Suppose lim an = a. Then lim 1 − ann = e−a [9, p 584].
n→∞ n→∞
 
Proof. Use (A.4) with rn = n, xn,k = − ann . Then, nk=1 xn,k = −an → −a and nk=1 x2n,k =
a2n n1 → a · 0 = 0.
 n
Alternatively, from L’Hôpital’s rule, lim 1 − na = e−a . (See also [18, Theorem 3.31,
n→∞
p 64]) This gives
 a direct proof for the case when a > 0. For n large enough, note that
both  1− an 
and 1 − a
are ≤ 1 where we need a > 0 here. Applying (A.3), we get
  n   n
 1 − an − 1 − a  ≤ |an − a| → 0.
n n
n n   n  n −1
b −1
For a < 0, we use the fact that, for bn → b > 0, (1) 1 + n = 1 + nb → e−b
 −1   −1 
 
and (2) for n large enough, both  1 + nb  and  1 + bnn  are ≤ 1 and hence
+ −1 ,n + −1 ,n 

 bn b   |bn− b| 
 1+ − 1+ ≤ → 0.
 n n  1 + bnn 1 + nb

A.2 Summations
A.6. Basic formulas:

n
n(n+1)
(a) k= 2
k=0


n
n(n+1)(2n+1)
(b) k2 = 6
= 16 (2n3 + 3n2 + n)
k=0
 2

n 
n
(c) k3 = k = 14 n2 (n + 1)2 = 14 (n4 + 2n3 + n2 )
k=0 k=0

127
A nicer formula is given by

n
1
k (k + 1) · · · (k + d) = n (n + 1) · · · (n + d + 1) (35)
k=1
d+2


n
A.7. Let g (n) = h (k) where h is a polynomial of degree d. Then, g is a polynomial
k=0

d+1
of degree d + 1; that is g (n) = am x m .
m=1

• To find the coefficients am , evaluate g (n) for n = 1, 2, . . . , d + 1. Note that the case
when n = 0 gives a0 = 0 and hence the sum starts with m = 1.

• Alternative, first express h (k) in terms of summation of polynomials:

+ d−1 ,

h (k) = bi k (k + 1) · · · (k + i) + c.
i=0

To do this, substitute k = 0, −1, −2, . . . , − (d − 1).

• k 3 = k (k + 1) (k + 2) − 3k (k + 1) + k

Then, to get g (n), use (35).

A.8. Geometric Sums:
∞
(a) 1
ρi = 1−ρ for |ρ| < 1
i=0


∞
ρk
(b) ρi = 1−ρ
i=k


b
ρa −ρb+1
(c) ρi = 1−ρ
i=a


∞
ρ
(d) iρi = (1−ρ)2
i=0


b
ρb+1 (bρ−b−1)−ρa (aρ−a−ρ)
(e) iρi = (1−ρ)2
i=a


∞
kρk ρk+1
(f) iρi = 1−ρ
+ (1−ρ)2
i=k


∞
ρ+ρ2
(g) i2 ρ i = (1−ρ)3
i=0

A.9. Double Sums:

 n 2
 
n n
(a) ai = ai aj
i=1 i=1 j=1

128
 
∞ 
∞ 
∞ 
i 
(b) f (i, j) = f (i, j) = 1 [i ≥ j]f (i, j)
j=1 i=j i=1 j=1 (i,j)

A.10. Exponential Sums:


∞
λk λ2 λ3
• eλ = k!
=1+λ+ 2!
+ 3!
+ ...
k=0

2 3 
∞ k−1 
∞ k−1
• λeλ + eλ = 1 + 2λ + 3 λ2! + 4 λ3! + . . . = λ
k (k−1)! = λ
k (k−1)!
k=1 k=0

A.11. Zeta function ξ (s) is defined for any complex number s with Re {s} > 1 by the

∞
1
Dirichlet series: ξ (s) = ns
.
n=1

• For real-valued nonnegative x

(a) ξ (x) converges for x > 1

(b) ξ (x) diverges for 0 < x ≤ 1

[9, Q2.48 p 105].

• ξ (1) = ∞ corresponds to harmonic series.

A.12. Abel’s theorem: Let a = (ai : i ∈ N) be any sequence of real or complex numbers
and let
 ∞
Ga (z) = ai z i ,
i=0
∞
be the power series with coefficients a. Suppose that the series i=0 ai converges. Then,


∞
lim− Ga (z) = ai . (36)
z→1
i=0

In the special case where all the coefficientsai are nonnegative real numbers, then the
above formula (36) holds also when the series ∞ i=0 ai does not converge. I.e. in that case
both sides of the formula equal +∞.

A.3 Derivatives
A.13. Basic Formulas
d u
(a) dx
a = au ln a du
dx

loga e du
(b) d
dx
loga u = u dx
, a = 0, 1

129
 (c) Derivatives of the products: Suppose f (x) = g (x) h (x), then
 n  
(n) n (n−k)
f (x) = g (x) h(k) (x).
k=0
k

In fact,
dn    dni
r r
n!
fi (t) = fi (t).
dtn i=1 n +···+n
n !n ! · · · nr ! i=1 dtni
=n 1 2
1 r

Definition A.14 (Jacobian). In vector calculus, the Jacobian is shorthand for either the
Jacobian matrix or its determinant, the Jacobian determinant. Let g be a function
from a subset D of Rn to Rm . If g is differentiable at z ∈ D, then all partial derivatives
exists at z and the Jacobian matrix of g at a point z ∈ D is
⎛ ∂g1 ⎞
∂x1
(z) · · · ∂g1
∂xn
(z)  
⎜ .. . . ⎟ ∂g ∂g
dg (z) = ⎝ . .. .. ⎠ = (z) , . . . , (z) .
∂x1 ∂xn
∂gm
∂x1
(z) · · · ∂xn (z)
∂gm

∂(g1 ,...,gn )
Alternative notations for the Jacobian matrix are J, ∂(x 1 ,...,xn )
[7, p 242], Jg (x) where the it
is assumed that the Jacobian matrix is evaluated at z = x = (x1 , . . . , xn ).

• Let g : D → Rm with open D ⊂ Rn , y ∈ D. If ∀k ∀j partial derivatives ∂x ∂gk

j
exists
in a neighborhood of z and continuous at z, then g is differentiable at z. And dg is
continuous at z.
• Let A be an n-dimensional
 “box” defined by the corners x and x+Δx. The “volume”
of the image g(A) is ( i Δxi ) |det dg(x)|. Hence, the magnitude of the Jacobian
determinant gives the ratios (scaling factor) of n-dimensional volumes (contents). In
other words,  
 ∂(y1 , . . . , yn ) 

dy1 · · · dyn =   dx1 · · · dxn .
∂(x1 , . . . , xn ) 
• d(g −1 (y)) is the Jacobian of the inverse transformation.
• In MATLAB, use jacobian.
• Change of variable: Let g be a continuous differentiable map of the open set U onto
V . Suppose that g is one-to-one and that det(dg(x)) = 0 for all x.
! !
h(g(x)) |det(dg(x))| dx = h(y)dy.
U V

Definition A.15. The gradient (or gradient vector field) of a scalar function f (x) with
respect to a vector variable x = (x1 , . . . , xn ) is
⎛ ∂f ⎞
∂θ1
⎜ ⎟
∇θ f (θ) = ⎝ ... ⎠ .
∂f
∂θn

130
If the argument is row vector, then,
⎛ ∂f1 ∂f2 ∂fm
⎞
  ⎜ ∂θ1 ∂θ1 ∂θ1
⎟
∇θ f T (θ) = ⎝ ... ..
.
... ..
. ⎠.
∂f1 ∂f2 ∂fm
∂θn ∂θn ∂θn

Definition A.16. Given a scalar-valued function f , the Hessian matrix is the square
matrix of second partial derivatives
⎡ 2 2f
⎤
∂ f
∂θ12
· · · ∂θ∂1 ∂θ
⎢ . n
⎥
T ⎢
∇θ f (θ) = ∇θ (∇θ f (θ)) = ⎣ .
2 . . . .
.
. . ⎥.
⎦
∂2f ∂2
∂θn ∂θ1
· · · ∂θ 2
n

It is symmetric for nice function.

 
A.17. ∇x f T (x) = (df (x))T .

A.18. Let f, g : Ω → Rm . Ω ⊂ Rn . h (x) = (f (x) , g (x)) : Ω → R. Then

dh (x) = (f (x))T dg (x) + (g (x))T df (x) .
• For an n × n matrix A, let f (x) = (Ax, x). Then df (x) = (Ax)T I + xT A =
xT AT + xT A.
◦ If A is symmetric, then df (x) = 2xT A. So, ∂
∂xj
(Ax, x) = 2 (Ax)j .

A.19. Chain rule: If f is differentiable at y and g is differentiable at z = f (y), then

g ◦ f is differentiable at y and
d (g ◦ f ) (y) = dg (z) df (y)
p×n p×m m×n

(matrix multiplication).
• In particular,
     
d ∂ d ∂ d
g (x (t) , y (t) , z (t)) = g (x, y, z) x (t) + g (x, y, z) y (t)
dt ∂x dt ∂y dt
  
∂ d 
+ g (x, y, z) z (t)  .
∂z dt (x,y,z)=(x(t),y(t),z(t))

A.20. Let f : D → Rm where D ⊂ Rn is open and connected (so arcwise-connected).

Then, df (x) = 0 ∀x ∈ D ⇒ f is constant.
A.21. If f is differentiable at y, then all partial and directional derivative exists at y
⎛ ⎞ ⎛ ∂f1 ⎞
∇f1 (y) ∂x1
(y) · · · ∂x
∂f1
(y)  
⎜ ⎟ ⎜ ⎟
n
. . ... .. ∂f ∂f
df (y) = ⎝ .. ⎠=⎝ .. . ⎠= (y) , . . . , (y) .
∂x1 ∂xn
∇fm (y) ∂fm
∂x1
(y) · · · ∂xn (y)
∂fm

131
 A.22. Inversion (Mapping) Theorem: Let open Ω ⊂ Rn , f : Ω → Rn is C 1 , c ∈ Ω.
df (c) is bijective, then ∃U open neighborhood of c such that
n×n

(a) V = f (U ) is an open neighborhood of f (c).

(b) f |U : U → V is bijective.
(c) g = (f |U )−1 : V → U is C 1 .
(d) ∀y ∈ V dg (y) = [df (g (y))]−1 .
 
d (x) = I ∇x xT = I
  2
d x2 = 2xT ∇x x
 = 2x 
d (Ax
 T +
 b) =T A ∇x (Ax + b)T = AT
d a x =a  
∇x aT x = a
 
  ∇x f T (x) g (x)
d f T (x) g (x)
= (dg (x))T f (x) + (df (x))T g (x)
T T
= f (x) dg (x) + g (x) df (x)    
= ∇x g T (x) f (x) + ∇x f T (x) g (x)

For symmetric Q, For symmetric Q,

   
d f T (x) Qf (x) ∇x f T (x) Qf (x)
= f T (x) Qdf (x) + f T (x) QT df (x) = 2 (df (x))T Qf (x)
 
= 2f T (x) Qdf (x) = 2∇x f T (x) Qf (x)
   
d xT Qx = 2xT Q ∇x xT Qx = 2Qx.
     
d f (x)2 = 2f T (x) df (x) ∇x f (x)2 = 2∇x f T (x) f (x)

A.4 Integration
A.23. Basic Formulas
 u
(a) au du = lna a , a > 0, a = 1.
1 1
1 α , α > −1 ∞ α , α < −1
(b) t dt = α+1 and t dt = α+1 So, the integration of the
0 ∞, α ≤ −1 1 ∞, α ≥ −1
1 ∞
function 1t is the test case. In fact, 1t dt = 1t dt = ∞.
0 1
 
 xm+1
ln x − 1
, m = −1
(c) xm ln xdx = m+1 m+1
1
2
ln2 x, m = −1
A.24. Integration by Parts:
! !
udv = uv − vdu

132
 4
4

2
t
1.5 3
t

t
0.5
t
2
1
 0.5
t
1
t 1

0
0
0 1 2 3 4 5
0 t 5

Figure 25: Plots of tα


(a) Basic idea: Start with an integral of the form
 f (x) g (x)dx.
Match this with an integral of the form udv by choosing dv to be part of the
integrand including dx and possibly f (x) or g (x).
(b) In particular, repeated application of integration by parts gives
! 
n−1 !
f (x) g (x)dx = f (x) G1 (x) + (−1)i f (i) (x) Gi+1 (x) + (−1)n f (n) (x) Gn (x) dx
i=1
  (37)
(i) di
where f (x) = dxi
f
(x), G1 (x) = g (x)dx, and Gi+1 (x) = Gi (x)dx. Figure 26
can be used to derived (37).

f  x g  x
+
1
f  x G1  x 
Differentiate  2 Integrate
f  x + G2  x 
f    x  
n 1
1
n1
Gn 1  x 
f    x  
n
Gn  x 
n
1

Figure 26: Integration by Parts

To see this, note that

! !
f (x) g (x)dx = f (x) G1 (x) − f  (x) G1 (x) dx,

and ! !
f (n)
(x) Gn (x) dx = f (n)
(x) Gn+1 (x) − f (n+1) (x) Gn+1 (x) dx.

133
 x2 e3 x 2 3x §1 2 2 2 · 3x sin x + ex
+ x ³x e dx ¨ x  x  ¸e
1 3x ©3 9 27 ¹ cos x - ex
2x e
³  sin x  e dx
x
- 3 x  sin x ex
+
1 3x
2 e  sin x  cos x  e x  ³  sin x  e x dx
+ 9
1 3x 1
0 - e  sin x  cos x  e x
27 2

Figure 27: Examples of Integration by Parts using figure 26.

 xn eax

(c) xn eax dx = a
− n
a
xn−1 eax

A.25. If n is a positive integer,

!
eax  (−1)k n! n−k
n
xn eax dx = x .
a k=0 ak (n − k)!

eax
 
(a) n = 1 : a
x− 1
a
eax
 
(b) n = 2 : a
x2 − a2 x + 2
a2

t eat

n
(−1)k n! n−k (−1)n n! eat

n
(−1)k n! n−k
(c) xn eax dx = a ak (n−k)!
t − an+1
= a ak (n−k)!
t + n!
(−a)n+1
0 k=0 k=0

∞ e−at

n
e−at

n
(d) xn e−ax dx = a
n!
ak (n−k)!
tn−k = a
n!
an−k j!
tj
t k=0 j=0

∞ n!
(e) xn eax dx = (−a)n+1
, a < 0. (See also Gamma function)
0

∞
• n! = e−t tn dt.
0
• In MATLAB, consider using gamma(n+1) in stead of factorial(n). Note also
that gamma() allows vector input.
1
(f) xβ e−x dx is finite if and only if β > −1.
0
1 1 1
Note that 1
e
xβ dx ≤ xβ e−x dx ≤ xβ dx.
0 0 0

∞
(g) ∀β ∈ R, xβ e−x dx < ∞.
1
∞ ∞ ∞
For β ≤ 0, xβ e−x dx ≤ e−x dx < e−x dx = 1.
1 1 0
∞ β −x
∞ β −x
∞
For β > 0, x e dx ≤ x e dx ≤ xβ e−x dx = $β%!
1 1 0

134
 ∞
(h) xβ e−x dx is finite if and only if β > −1.
0

A.26 (Differential of integral). Leibniz’s Rule: Let g : R2 → R, a : R → R, and


b(x)
b : R → R be C 1 . Then f (x) = g (x, y)dy is C 1 and
a(x)

!b(x)
   ∂g
f (x) = b (x) g (x, b (x)) − a (x) g (x, a (x)) + (x, y)dy. (38)
∂x
a(x)

In particular, we have
!x
d
f (t)dt = f (x) , (39)
dx
a
!v(x) !v(x)
d dv d
f (t)dt = f (t)dt = f (v (x)) v  (x) , (40)
dx dx dv
a a
⎛ v(x) ⎞
!v(x) ! !
u(x)
d d ⎝
f (t)dt = f (t)dt − f (t)dt⎠ = f (v (x)) v  (x) − f (u (x)) u (x) . (41)
dx dx
u(x) a a

Note that (38) can be derived from (A.19) by considering f (x) = h(a(x), b(x), x) where
b
h (a, b, c) = g (c, y)dy. [9, p 318–319].
a

A.5 Gamma and Beta functions

A.27. Gamma function:
∞
(a) Γ (q) = xq−1 e−x dx. ; q > 0.
0

(b) Γ (0) = ∞

(c) Γ (n) = (n − 1)! for n ∈ N.

Γ (n + 1) = n! if n ∈ N ∪ {0}.

(d) 0! = 1.
  √
(e) Γ 12 = π.

(f) Γ (x + 1) = xΓ (x) (Integration by parts).

• This relationship is used to define the gamma function for negative numbers.

135
 ∞
(g) Γ(q)
αq
= xq−1 e−αx dx, α > 0.
0

A.28. The incomplete beta function is defined as

! x
B(x; a, b) = ta−1 (1 − t)b−1 dt.
0

For x = 1, the incomplete beta function coincides with the (complete) beta function.
The regularized incomplete beta function (or regularized beta function for short) is de-
fined in terms of the incomplete beta function and the (complete) beta function:

B(x; a, b)
Ix (a, b) = .
B(a, b)

• For integers m, k,


m+k−1
(m + k − 1)!
Ix (m, k) = xj (1 − x)m+k−1−j .
j=m
j!(m + k − 1 − j)!

• I0 (a, b) = 0, I1 (a, b) = 1
• Ix (a, b) = 1 − I1−x (b, a)

136
References
[1] Patrick Billingsley. Probability and Measure. John Wiley & Sons, New York, 1995.
5.29, 2b

[2] George Casella and Roger L. Berger. Statistical Inference. Duxbury Press, 2001. 4.42,
6.9, 11, 8.13, 11.11, 11.4, 1, 6, 11.21

[3] Donald G. Childers. Probability And Random Processes Using MATLAB. McGraw-
Hill, 1997. 1.3, 1.3

[4] Herbert A. David and H. N. Nagaraja. Order Statistics. Wiley-Interscience, 2003.

11.4, 11.21

[5] W. Feller. An Introduction to Probability Theory and Its Applications, volume 2. John
Wiley & Sons, 1971. 4.36

[6] William Feller. An Introduction to Probability Theory and Its Applications, Volume 1.
Wiley, 3 edition, 1968.

[7] Terrence L. Fine. Probability and Probabilistic Reasoning for Electrical Engineering.
Prentice Hall, 2005. 3.1, 3.10, 3.14, 4.18, 4.23, 5.26, 9.1, 9.3, 11.10, 15.2, A.14

[8] Boris Vladimirovich Gnedenko. Theory of probability. Chelsea Pub. Co., New York, 4
edition, 1967. Translated from the Russian by B.D. Seckler. 5.29

[9] John A. Gubner. Probability and Random Processes for Electrical and Computer En-
gineers. Cambridge University Press, 2006. 1.2, 4.30, 4.33, 4.35, 4, 6.43, 7.9, 9.9, 10.1,
2, 3, 6, 11.24, 12.21, 12.22, 14.3, 3, A.5, A.11, A.26

[10] Samuel Karlin and Howard E. Taylor. A First Course in Stochastic Processes. Aca-
demic Press, 1975. 4.6, 9.10, 10.1

[11] J. F. C. Kingman. Poisson Processes. Oxford University Press, 1993. ISBN:

0198536933. 5.23, 5.24

[12] A.N. Kolmogorov. The Foundations of Probability. 1933. 3.14

[13] Nabendu Pal, Chun Jin, and Wooi K. Lim. Handbook of Exponential and Related
Distributions for Engineers and Scientists. Chapman & Hall/CRC, 2005. 6.25

[14] Athanasios Papoulis. Probability, Random Variables and Stochastic Processes.

McGraw-Hill Companies, 1991. 5.17, 11.5

[15] E. Parzen. Stochastic Processes. Holden Day, 1962. 3, 4, 11.25

[16] Sidney I. Resnick. Adventures in Stochastic Processes. Birkhuser Boston, 1992. 5

[17] Kenneth H. Rosen, editor. Handbook of Discrete and Combinatorial Mathematics.

CRC, 1999. 1, 4, 5, 6, 7, 8

137
[18] Walter Rudin. Principles of Mathematical Analysis. McGraw-Hill, 1976. A.5

[19] Gbor J. Szkely. Paradoxes in Probability Theory and Mathematical Statistics. 1986.
8.2

[20] Tung. Fundamental of Probability and Statistics for Reliability Analysis. 2005. 17, 20

[21] Larry Wasserman. All of Statistics: A Concise Course in Statistical Inference.

Springer, 2004. 4.16, 4.18, 4.43, 7.9, 3, 13.1

138
Index
Bayes Theorem, 21
Binomial theorem, 12
Birthday Paradox, 23

Chevalier de Mere’s Scandal of Arithmetic,

22

Delta function, 19
Dirac delta function, see Delta function

event algebra, 25

False Positives on Diagnostic Tests, 24

gradient, 130
gradient vector field, 130

Hessian matrix, 131

Integration by Parts, 132

Jacobian, 98, 130

Jacobian formulas, 99

Leibniz’s Rule, 135

Monte Hall’s Game, 23

multinomial coefficient, 13
Multinomial Counting, 13
Multinomial Theorem, 13

Order Statistics, 101

Probability of coincidence birthday, 23

Total Probability Theorem, 21

uncorrelated
not independent, 76, 93

Zeta function, 129

Zipf or zeta random variable, 59

139