Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd
Upload
81 views

18.445 Introduction To Stochastic Processes

This document summarizes key concepts from a lecture on Markov chain mixing, including: - The lecturer announced midterm and final exam dates for the course. - Total variation distance and coupling are introduced as ways to characterize convergence of a Markov chain to its stationary distribution. - The convergence theorem states that for irreducible, aperiodic Markov chains, the total variation distance between the chain's distribution and the stationary distribution decreases exponentially fast. - Mixing time is defined as the number of steps needed for the total variation distance to drop below a given threshold, and methods for bounding mixing time are discussed.

Uploaded by

AdityaPandhare
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
81 views

18.445 Introduction To Stochastic Processes

This document summarizes key concepts from a lecture on Markov chain mixing, including: - The lecturer announced midterm and final exam dates for the course. - Total variation distance and coupling are introduced as ways to characterize convergence of a Markov chain to its stationary distribution. - The convergence theorem states that for irreducible, aperiodic Markov chains, the total variation distance between the chain's distribution and the stationary distribution decreases exponentially fast. - Mixing time is defined as the number of steps needed for the total variation distance to drop below a given threshold, and methods for bounding mixing time are discussed.

Uploaded by

AdityaPandhare
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

18.

445 Introduction to Stochastic Processes


Lecture 4: Introduction to Markov chain mixing

Hao Wu

MIT

23 February 2015

Hao Wu (MIT) 18.445 23 February 2015 1/9


Announcement
Midterm : April 6th.(on class)

Final : May 18th.


The tests are closed book, closed notes, no calculators.

Recall
If (Xn )n is an irreducible Markov chain with stationary distribution π,
then
n
1n
lim 1[Xj =x] = π(x), Pµ − a.s.
n→∞ n
j=0

Today’s goal We will show that Xn converges to π under some "strong


sense".
total variation distance
the convergence theorem
mixing times

Hao Wu (MIT) 18.445 23 February 2015 2/9


Three ways to characterize the total variation distance
µ and ν : probability measures on Ω.
||µ − ν||TV = max |µ(A) − ν(A)|.
A⊂Ω

Lemma

1X
||µ − ν||TV = |µ(x) − ν(x)|.
2
x∈Ω

1
||µ − ν||TV = sup{µf − νf : f satisfying max |f (x)| ≤ 1}.
2 x∈Ω

||µ − ν||TV = inf{P[X 6= Y ] : (X , Y ) is a coupling of µ, ν}.

Definition
We call (X , Y ) the optimal coupling if P[X 6= Y ] = ||µ − ν||TV .
Hao Wu (MIT) 18.445 23 February 2015 3/9
The Convergence Theorem

Suppose that (Xn )n is a Markov chain with transition matrix P. Assume


that P is irreducible and aperiodic, then
there exists r such that P r (x, y ) > 0 for all x, y ∈ Ω ;
there exists a unique stationary distribution π and π(x) > 0 for all
x ∈ Ω.

Theorem
Suppose that P is irreducible, aperiodic, with stationary distribution π.
Then there exist constants α ∈ (0, 1) and C > 0 such that

max ||P n (x, ·) − π||TV ≤ C αn ∀n ≥ 1.


x∈Ω

Hao Wu (MIT) 18.445 23 February 2015 4/9


Mixing time
Definition
d(n) = max ||P n (x, ·) − π||TV
x∈Ω

d̄(n) = max ||P n (x, ·) − P n (y , ·)||TV


x,y ∈Ω

Lemma
d(n) ≤ d̄(n) ≤ 2d(n)

Lemma
d̄(m + n) ≤ d̄(m) · d̄(n)

Corollary

d̄(mn) ≤ d̄(n)m
Hao Wu (MIT) 18.445 23 February 2015 5/9
Mixing time

Definition
tmix = min{n : d(n) ≤ 1/4}, tmix () = min{n : d(n) ≤ }

Lemma
1 tmix
tmix () ≤ log( )
 log 2

Questions : How long does it take the Markov chain to be close to the
stationary measure ?
Lecture 5 : Upper bounds on tmix ; Lecture 6 : Lower bounds on tmix ;
Lecture 7 : Interesting models.

Hao Wu (MIT) 18.445 23 February 2015 6/9


Couple two Markov chains

Definition
A coupling of two Markov chains with transition matrix P is a process
(Xn , Yn )n≥0 with the following two properties.
Both (Xn ) and (Yn ) are Markov chains with transition matrix P.
They stay together after their first meet.

Notation : If (Xn )n≥0 and (Yn )n≥0 are coupled Markov chains with
X0 = x, Y0 = y , then we denote by Px,y the law of (Xn , Yn )n≥0 .

Hao Wu (MIT) 18.445 23 February 2015 7/9


Couple two Markov chains

Theorem
Suppose that P is irreducible with stationary distribution π. Let
(Xn , Yn )n≥0 be a coupling of Markov chains with transition matrix P for
which X0 = x, Y0 = y . Define τ to be their first meet time :

τ = min{n ≥ 0 : Xn = Yn }.

Then
||P n (x, ·) − P n (y , ·)||TV ≤ Px,y [τ > n].
In particular,
d(n) ≤ max Px,y [τ > n].
x,y

Hao Wu (MIT) 18.445 23 February 2015 8/9


Random walk on N−cycle : Upper bound on tmix

Lazy walk : it remains in current position with probability 1/2, moves


left with probability 1/4, right with probability 1/4.
It is irreducible.
The stationary measure is the uniform measure.

Theorem
For the lazy walk on N−cycle, we have

tmix ≤ N 2 .

Hao Wu (MIT) 18.445 23 February 2015 9/9


MIT OpenCourseWare
http://ocw.mit.edu

18.445 Introduction to Stochastic Processes


Spring 2015

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

You might also like