Topics

Random Variables
Integration Expectation
Limit Theorems Part I
Inequalities
Bounds
Convergence

Markov chains part 2

IE [Link] eetetion
Martingales
Poisson Processes
Gaussian Processes Port 4
Brownian Motion
PART 1

Random Variables

A function f R F S S if YA es
f A wer flu EA E F

A Random variable is a measurable function on a prob

spare
X R F IR IB R

Compositions of RVs are RVs

X X2 X X2 Xi cos X2
ex exp

CDF of X is Fx x IP X EX

i increasing
o
ioFixs
II sÉ[Link]
F x
Mix Fly
iv left limits
Fix Jim Fly IP X X
 X X it
X X ere identically distributed

IP VFA IP X'EA YA

For Ur unit Lo D Y log u find density of Y

PLY Edy Ey Phys y

IP L 10g u ey

Ey ie t I tell
Pl u z e s

Ey l e
d
e'd

Integration and Expectation

Discrete expectation E x x IPLX x

x

Lebesgue Integral of a function f

f tan
expectation of RV X on R F IP

E IX XLW dip
 Almost sure convergence

X n x a s as he is if

IPL X xn x X x 1 0

Dominated Convergence Theorem

It Xn X as and I Igl st lx ul E g and E Tx cos

then E Xn E x a s as non

Monotone convergence Theorem

If Xn x as and Xu 20 then

E Xn E x as

The Kth moment of X is E x

The Variance of X is

E X E X

Independence Two events A Bef are independent if

IP AMB PLA IP B

a collection of events IA A An are indep if V5 E th

IP IP Ai
A Ai I
 IP ARB
Conditional Probability P a 1B For PIB O
IPFB

Product of expectation holds For indep Xi X2 Xh

EL Xi IIE Xi

Covariance Cov Xi X2 E X X E x E X2

if X I Xz Cov Xi X27 0

if 2
JI Xi then

var 52J Cov Xi Xi

cov X X var x

If Xi X2 Xn independent the linearity of variance

holds

Var É Xi
É VarCXi

Symmetry is cool

Cov X y
Corry y E El I
Van X Vary
Inequalities and Limit theorems

Markov's Inequality If X 20 then

ELY la
IPL XI al e

Chebyshev's Inequality

e
El x'd an
IP IX e a

IP IX E Lx Salvat E la

PI IX Ml 2 as e Ye things are
somewhat concentrate
around men
rel to o

Chernoff Bound
can choose a

P S s p 8 n I exp na xp S

I 82
exp

Convergence
Almost sure Convergence of an event

Xu X a e if IP w Xn w X w
Is
Probability to a constant
Convergence in

nth Xn X if HE so IP I Xn X1 e o

Convergence in hp

IIb Xn X in Lp if I Xn X p o as non

Weak Law of Large Numbers

For id Xu with E Xn y Var Xn 52 0 then

I xn
y
as new

in probability and in
LI
Law of Large Numbers
Strong
 Infinitely often

An occurs is if
É And of
Then Xu X a s HE IP Ellyn x1 E io

Borel Cantelli
a For Al Az

É IP A a o Plan i o o

bl If An are independent

If PLA D IP An i o

Strong Law of Large Numbers

If Xn iid E Xn M

I Xn M as as new
PART 2

Weak Convergence and the Central Limit Theorem

Central Limit Theorem If Xi are id E X M

Var Xi 02 then

IN
I
ont as no
im
In r

Xn X
Convergence in Distribution weak Convergence

it Fx X VX er
x
I Fx

Xu X Xn s x

Normal Distribution N
1,6
202
if
x exp x p

X
For N
mp3
EEx M Vart X 02
 Triangular Arrays Suppose Xnin Xyz [Link] e
independent RVs with
are sequences of men 0

It
is EI El Xii

Ii Lindeberg's Condition HE O

EY E Xn 11 Xml e o

then
IT xn Nco i es a

Markov chains

X Xz in countable set A is a Markov chain

it far az EA

IP Xu an I Xo Go X an Xn i an a

P Xn an I Xn an Markov Property
 Transition matrix is a stochastic matrix if

FX Pxy I
I
For all n I

IPL Xn y IX x IPL Xue y I XK X

P
xy

Recurrent it

IP In I Xu x I Xo x 1

II IP Xu x IX x n

else transient

MC with transition matrix P has stationary distribution

it if
P

A Mc is irreducible if ti In

IP Xn j 1 Xo so
 ie its

Perron Frobenius If P is a stochastic matrix

it left eigenvector st
has
up
a
n p
1 1 d o ti If P is irreducible
µ Mi
unique and o
then is Mi
ye

A station XE A is aperiodic if

GCD s I where S n 21 pix o

If a Mc is aperiodic F X EA that's aperiodic

and irreducible then it is ergodic

If Xn is ergodic IN st Pig 0 Fn IN

Xu ergodic then Xn I T
If is

If X Y are two RVs a coupling is a joint distribution

X y St XIX y I y

Total Variation Distance

Max A 1
fu
dtv r MCA
A

I I lack
r K
Markov chain MonteCarlo

Markov chain LLN If Xn is a finite state MC

with stationary distribution it Vx it Xo X

I fix
Ey fay ly

A MC is reversible if

Xo Xi Xn I Xn Xn i Xo
so X n I Xo so Xo It is a stationary distribution

satisfies the DBE Fx y

TxPxy Ty Pyx
It satisfies DBE is reversible then it is stationary
PART 3

Conditional Expectation

Conditional expectation wrt an event

E XI A

E x a PLI

Conditional Expectation wrt a RV

EIXITY y
417 E x yay IP Y y

This is then itself a Random Variable E X 4 4 y

EE t b E Ly Iz
i E fax by 2 a z

ii EL E X ly E TX Tower

in x y EIXIY EIX

Tower Property If G E Ga then

 ELE X 623 6 E XI G

Taking G 0 of

E LEIX G ELX

Branching Processes

Each node has a random and independent number of

children with distribution X

Martingales

A filtration is an increasing set of o algebras Ft tert

F E USE t
such that Ft
A martingale is a sequence of RVs a process
Xt wat Ft if Asst
 E X IFS Xs

also E L EL Xt I Fs E Xt EL Xs

For discrete time martingales it is sufficient to show

E Yn Fn Yu
because intuition

If US Ct

EL Xt I Fs E Xs supermartingale
E Xt I Fs I Xs submentingale

sub
A martingale is both a super
martingale

If Xn is a
martingale and U Xn is convex then

4 Xu is a
martingale

Martingale Convergence Theorem If Xu is a

sup EL Xi then
and so
martingale or super

Xn X a s boundedness
criteria
and E fix I is min Xt o
 for subm it is

sup E Xt

If Xn is a non negative martingale then

Xn X are

Stopping Times
We say N I 0 is a stopping time wht a filtration
In if Yn N En E Fn

you know that it has happened

It Xn is a super sub martingale and N is a stopping

time wrt Fn then Xuan is also a sup sub martingale

A sequence Xn is uniformly integrable it

into sup E IX I 11141 Mt o

If Xn in a martingale TFA E

i Xn are UI
ii Xn X a s

iii F XE L St E X Fn Xn
 Optional stopping Theorem
If Xu is a UI super sub martingale N is

a stopping time with N s a a s then

EL Xn I Fn Xnam
2 for sub

and E
E X Xo

A MC is a birth death chain it for Xh EZ

and Xn E
Xna 1,0

Azuma Hoeffding Inequalities Let Xu be a martingale

St
I Xi Xi n I E ki Then

IP Xn Xo't exp EIKE

Random colorings of graphs

Poisson Processes

A rate A Poisson Point Process on IR is a random

measure N St
i NCA n Pais Xian
Ii disjoint then N Ai
It A 172 Ak
independent

related is the location of the first point in a PP T

T v exp X

Poissonthinning If each point of a poisson process N

is assigned a type in I 2 k with prob pi indep
and the PP of of type is
Ni is points then Ni
an independent Poisson processes with rate Api

Poisson superposition If Ni Nk are independent

PP rate
rate Xi DP then N IN is a with

X I Xi Each label is indep with p

Spatial Poisson Process intensity X R o.o

M
A Xcx ax
 NCA NPois GCA
If Ai disjoint measurable sets N Ai indep

PARTI

Gaussian Processes
called a Gaussian Process
A collection I Xi Jian is

if USCA finite s then I a ies

Es ai Xi Normal
n

for finite A also called a Gaussian vector

If X is a Gaussian vector and Xi are uncorrelated

then Xi are independent

Covariance matrices are negative semi

non definite
have positive eigenvalues
they
The covariance matrix of a random vector X
is the matrix V Vi Vij Cov Xi Xi

EIR and V NSD matrix I

Vn a
a

Nd
unique Gaussian vector distribution
G V

Yn r
St for
Nag
Yi
p
Cov Yi Yj Vi
 i e Gaussian Vectors are uniquely determined by
their mean and their co variance matrix

Brownian Motion

Brownian Motion is the sealing limit of e SRW

Brownian Motion is e random function

B r c o o such that
i B t is a Gaussian Process
ii B o o E Blt O OVCBs By EAS

iii By is continuous

Properties of Brownian Motion

i Independent Normally distributed
Usct ELBE Bs IFS ELBE Bs o

ii Bct is a Martingale
iii symmetric
iv fractal
V Non differentiable everywhere

Hitting probabilities same as RW through optimal

stopping
 Bt Bs Bit 3
For Mt EE t play it
IP N o i E3
Mt I I Bt I

For hitting time Ts int f t By s

IPL INCO I 2 YE
IPL T et

For 2 t B t O 2 he no isolated zeros

almost surely

We know if Xn is recurrent that Tcu a s

from lecture Notice for yn the states

o and K are recurrent for my n

IP 7m21 Xm o IX n 07
I
 IP Fm 21 Xm KI Xu K

I since all men satisfy it

But Xu is not irreducible since stater 0 R

cannot leave Thus we can show that Xn on

is reicurent instead because

1,2 k it
of O on K
then the probability reaching

I
[Link] rrancewe
just need to cheek recurrence of 1 state
because if I state is recurrent then

the MC is as well For state 1

É IP Xn I IX
a more restrictive
I Xn i eventoe always
I IP Xn i
staying at

P Binom K t I
n

Y E I I
O K I
 I l ta an

so I is reccurent Xn on 1 2

K l is reccurent Thus T a a s