CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS

ARI FREEDMAN

Abstract. In this expository paper, I will give an overview of the necessary

conditions for convergence in Markov chains on finite state spaces. In doing
so, I will prove the existence and uniqueness of a stationary distribution for
irreducible Markov chains, and finally the Convergence Theorem when aperi-
odicity is also satisfied.

Contents
1. Introduction and Basic Definitions 1
2. Uniqueness of Stationary Distributions 3
3. Existence of a Stationary Distributions 5
4. Convergence Theorem 8
Acknowledgments 13
References 13

1. Introduction and Basic Definitions

A Markov chain is a stochastic process, i.e., randomly determined, that moves
among a set of states over discrete time steps. Given that the chain is at a certain
state at any given time, there is a fixed probability distribution for which state the
chain will go to next (including repeating the state). If there are n states, then the
n × n transition matrix P describes the Markov chain, where the rows and columns
are indexed by the states, and P (x, y), the number in the x-th row and y-th column
gives the probability of going to state y at time t + 1, given that it is at state x at
time t. We can formalize this as follows.
Definition 1.1. A finite Markov chain with finite state space Ω and |Ω| × |Ω|
transition matrix P is a sequence of random variables X0 , X1 , . . . where
P{Xt+1 = y | Xt = x} = P (x, y),
or the probability of Xt+1 = y given Xt = x is P (x, y). Then P (x, ·), the x-th row of
P , gives the distribution of Xt+1 given Xt = x. Here P is the notation for probability
of an event, and Px notates the probability of an event given X0 = x ∈ Ω. Thus,
Px {Xt = y} = P{Xt = y | X0 = x} = P t (x, y), as multiplying a distribution by
the transition matrix P advances the distribution one step along the Markov chain,
so multiplying by P t advances it by t steps from X0 = x.
Furthermore, when the distribution of Xt+1 is conditioned on Xt , the previous
values in the chain, X0 , X1 , . . . , Xt−1 , do not affect the value of Xt+1 . This is called

Date: August 28, 2017.

1
2 ARI FREEDMAN

the Markov
Tt−1 property, and can be formalized by saying that if
H = i=0 {Xi = xi } is any event such that P(H ∩ {Xt = x}) > 0, then
P {Xt+1 = y | H ∩ {Xt = x}) = P {Xt+1 = y | Xt = x},
and then we simply define P (x, y) = P {Xt+1 = y | Xt = x}.

0.5

0.4 x y 0.3

0.2
0.6 1

Figure 1. An example Markov chain with three states x, y, and z.

We can illustrate a Markov chain with a state diagram, in which an arrow from
one state to another indicates the probability of going to the second state given we
were just in the first. For example, in this diagram, given that the Markov chain is
currently in x, we have probability .4 of staying in x, probability .6 of going to z,
and probability 0 of going to y in the next time step (Fig. 1). This Markov chain
would be represented by the transition matrix
x y z
x .4 0 .6!
P = y .5 .3 .2
z 0 1 0 .

This definition mentions distributions, so it may help to formalize what these

are.
Definition 1.2. A probability distribution, or just a distribution, is a vector
of non-negative probabilities that sums up to 1 (this is known as the law of total
probability).
For any state x ∈ Ω, it makes sense that P (x, ·), the x-th row of P , should be a
distribution, since the probability of going from state x to any state is at least 0, and
the sum of the probabilities of going from state x to state y, over all states y ∈ Ω,
should be 1, as these are disjoint events that cover all the possibilities. Distribu-
tions are generally expressed as row vectors, which can then be right-multiplied by
matrices.
The transition matrices associated with Markov chains all fall under the larger
category of what we call stochastic matrices.
 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 3

Definition 1.3. A stochastic matrix is an n × n matrix with all non-negative

values and each row summing to 1. In particular, a matrix is stochastic if and only
if it consists of n distribution row vectors in Rn .

It is fairly easy to see that if P and Q are both stochastic matrices, then P Q is
also a stochastic matrix, and if µ is a distribution, then µP is also a distribution.

Definition 1.4. A distribution π is called a stationary distribution of a Markov

chain P if πP = π.

Thus, a stationary distribution is one for which advancing it along the Markov
chain does not change the distribution: if the distribution of Xt is a stationary
distribution π, then the distribution of Xt+1 will also be π. This brings up the
question of when a Markov chain will have a stationary distribution, and if so, is this
distribution unique? Will any distribution converge to this stationary distribution
over time? It turns out that with only mild constraints, all of these are satisfied.

Definition 1.5. A Markov chain is irreducible if for all states x, y ∈ Ω, there

exists a t ≥ 0 such that P t (x, y) > 0.

Intuitively, this means that it is possible to get from x to y for any x, y ∈ Ω in

some finite amount of time steps, or, equivalently, there exists a sequence of states
x = x0 , x1 , . . . , xt−1 , xt = y (which we call a path) the chain can take from x to y
such that P (xi , xi+1 ) > 0 for all 0 ≤ i < t.

2. Uniqueness of Stationary Distributions

In order to come up with a nice expression to show the existence of a stationary
distribution, we must first prove that if a distribution exists, it is unique. To do so,
we need to define harmonic functions and prove a lemma about them.

Definition 2.1. A function h : Ω → R is a harmonic at x ∈ Ω if

X
h(x) = P (x, y)h(y).
y∈Ω

If h is harmonic at all states in Ω = {x1 , x2 , . . . , xn },we sayh is harmonic on Ω,

h(x1 )
 h(x2 ) 
and then P h = h, where h is the column vector h =  . .
 
 .. 
h(xn )

Lemma 2.2. If P is irreducible and h is harmonic on Ω, then h is a constant

function.

Proof. Since Ω is finite, h attains a maximum at some state x0 ∈ Ω, such that

h(x0 ) ≥ h(y) ∀y ∈ Ω. Let z ∈ Ω be any state such that P (x0 , z) > 0, and assume
4 ARI FREEDMAN

that h(z) < h(x0 ). Since h is harmonic at x0 ,

X
h(x0 ) = P (x0 , y)h(y)
y∈Ω
X
= P (x0 , z)h(z) + P (x0 , y)h(y)
y∈Ω,y6=z
X
≤ P (x0 , z)h(z) + P (x0 , y)h(x0 )
y∈Ω,y6=z
X
< P (x0 , z)h(x0 ) + P (x0 , y)h(x0 )
y∈Ω,y6=z
 
X
= P (x0 , y) h(x0 )
y∈Ω

= h(x0 ),

where the last inequality follows from P (x0 , z) > 0 and h(z) < h(x0 ). However,
this gives us h(x0 ) < h(x0 ), a contradiction, which means h(x0 ) = h(z).
Now for any y ∈ Ω, P being irreducible implies there is a path from x0 to y, let
it be x0 , x1 , . . . , xn = y such that P (xi , xi+1 ) > 0. Thus, h(x0 ) = h(x1 ), and so x1
also maximizes h, which means h(x1 ) = h(x2 ). We carry on this logic to get

h(x0 ) = h(x1 ) = · · · = h(xn ) = h(y).

So h(y) = h(x0 ) ∀y ∈ Ω, and thus h is a constant function. 

Now we are ready to show that if a stationary distribution exists (which we show
in the next section), it must be unique.

Corollary 2.3. If P is irreducible and has a stationary distribution π, then π is

the only such stationary distribution.

Proof. By Lemma 2.2, the only functions h that are harmonic are those of the form
h(x) = c ∀x ∈ Ω for some constant c. Putting this into vector form, this means the
only solutions to the equation P h = h, or equivalently (P − I)h = 0 are
 
1
1
h = c . .
 
 .. 
1

Thus, dim(ker(P − I)) = 1, so by the rank-nullity theorem, rank(P − I) = |Ω| − 1.

And rank(P − I) = rank((P − I)T ) = rank(P T − I) = |Ω| − 1, so again by the
rank-nullity theorem, dim(ker(P T − I)) = 1, so over all row vectors v ∈ R|Ω| , the
equation (P T − I)v T = 0 has only a one-dimensional space of solutions. But this
equation is equivalent to vP = v, so, given that πP = π is a solution, all solutions
must be of the form v = λπ, for some scalar λ. However, to be a distribution whose
elements sum to 1, we must have λ = 1, and thus the only stationary distribution
is v = π. 
 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 5

3. Existence of a Stationary Distributions

We will now show that all irreducible Markov chains have a stationary distribu-
tion by explicitly constructing one, and then by Corollary 2.3, we will know that
this stationary distribution is unique.
Definition 3.1. For a Markov chain X0 , X1 , . . ., the hitting time for a state
x ∈ Ω is the instance of the chain “hitting” x, notated
τx = min{t ≥ 0 : Xt = x}.
When we want the hitting time to be strictly positive, we notate it
τx+ = min{t > 0 : Xt = x},
which is called the first return time when X0 = x.
We will also be using the notation E to denote the expected value of a variable,
and again, Ex means the expected value given X0 = x.
Lemma 3.2. For any x, y ∈ Ω of an irreducible Markov chain, Ex (τy+ ) is finite.
Proof. Since P is irreducible, we know for any two states z, w ∈ Ω that there exists
an s > 0 such that P s (z, w) > 0 (if z = w, we consider a path from z to a different
state and back to itself to ensure s > 0). We let r be the maximum of all such s
over z, w ∈ Ω, and let
 = min{P s (z, w) > 0 such that 0 < s ≤ r : z, w ∈ Ω}.
Then for all z, w ∈ Ω, there exists 0 < s ≤ r such that P s (z, w) ≥  > 0.
This implies that given any Xt , the probability of the chain going to a state y
between times t and t + r is at least , or conversely,
P{Xs 6= y : ∀t < s ≤ t + r} ≥ 1 − .
In general, saying τy+ > n implies Xt 6= y ∀0 < t ≤ n. So for k > 0,
Px {τy+ > kr} = Px {Xt 6= y ∀0 < t ≤ kr}
= Px {Xt 6= y ∀0 < t ≤ (k − 1)r}Px {Xt 6= y ∀(k − 1)r < t ≤ kr}
≤ Px {Xt 6= y ∀0 < t ≤ (k − 1)r}(1 − )
≤ Px {Xt 6= y ∀0 < t ≤ (k − 2)r}(1 − )2
..
.
≤ Px {Xt 6= y ∀0 < t ≤ 0}(1 − )k
= (1 − )k ,
where Px {Xt 6= y ∀0 < t ≤ 0} = 1 by vacuous truth.
Now for any random variable Y valued on the non-negative integers, we have
∞
X
E(Y ) = tP (Y = t)
t=0
= 1 · P (Y = 1) + 2 · P (Y = 2) + 3 · P (Y = 3) + · · ·
   
= P (Y = 1) + P (Y = 2) + P (Y = 3) + · · · + P (Y = 2) + P (Y = 3) + · · · + · · ·
∞
X
= P (Y > t).
t=0
6 ARI FREEDMAN

And P{τy+ > t} is a decreasing function with respect to t, since

P{τy+ > t + 1} ≤ P{τy+ > t + 1} + P{τy+ = t + 1} = P{τy+ > t},

which we use to get

∞
X
Ex (τy+ ) = Px {τy+ > t}
t=0
∞
X
≤ rPx {τy+ > kr}
k=0
X∞
≤r (1 − )k .
k=0

By definition, 0 <  ≤ 1, so 0 ≤ 1 −  < 1, which means this sum converges, and

thus Ex (τy+ ) is finite. 

Theorem 3.3. If P is irreducible, then it has a unique stationary distribution π

with π(x) > 0 ∀x ∈ Ω, given by
1
π(x) = .
Ex (τx+ )
Proof. Fix any state z ∈ Ω. Then we define

π
e(y) = Ez (number of visits to y before returning to z)
∞
X
= Pz {Xt = y, τz+ > t},
t=0

since the expected number of visits to y before returning to z is the sum of all the
probabilities of the chain hitting y at a time step less than the return time.
For any given chain with X0 = z, the number of visits to y before return to z is
≤ τz+ , since the total number of states the chain visits before returning to z is τz+ .
Thus π e(y) ≤ Ez (τz+ ), which by Lemma 3.2 is finite, and thus all π e(y) are finite.
And since P is irreducible, it is at least possible to visit y once (a path from z
to y followed by a path from y to z), which means the expected number of visits to
y before returning to z is positive, so πe(y) > 0.
Now we show π e is stationary, or that for all y, (e
π P )(y) = π
e(y). First, see that
X ∞
XX
(e
π P )(y) = π
e(x)P (x, y) = Pz {Xt = x, τz+ ≥ t + 1}P (x, y),
x∈Ω x∈Ω t=0

e(x) and replaced τz+ > t with

where we just plugged in our earlier expression for π
+
the equivalent expression τz ≥ t + 1.
Since the event {τz+ ≥ t + 1} is only determined by X0 , X1 , . . . , Xt , it is inde-
pendent of the event Xt+1 = y, when conditioned on Xt = x, which means

Pz {Xt = x, Xt+1 = y, τz+ ≥ t + 1} = Pz {Xt = x, τz+ ≥ t + 1}Pz {Xt+1 = y | Xt = x}

= Pz {Xt = x, τz+ ≥ t + 1}P (x, y).
 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 7

We can then plug this in to our earlier expression for (e

π P )(y) and switch around
the order of summation, since the inner sum converges for all x ∈ Ω, to get
∞ X
X
(e
π P )(y) = Pz {Xt = x, Xt+1 = y, τz+ ≥ t + 1}
t=0 x∈Ω
∞
X
= Pz {Xt+1 = y, τz+ ≥ t + 1}
t=0
∞
X
= Pz {Xt = y, τz+ ≥ t},
t=1

using the fact that the sum of the probabilities of {Xt = x} over all x ∈ Ω will
just equal 1. We notice that this final summation is very similar to our original
expression for π
e(y); in particular,
∞
X
(e e(y) − Pz {X0 = y, τz+ > 0} +
π P )(y) = π Pz {Xt = y, τz+ = t}.
t=1

But this final term is just accounting for all the occurrences of Xτz+ = y, and so it
sums up to Pz {Xτz+ = y}, which is equal to 1 when y = z and 0 otherwise (since
the Markov chain at its return time should be back at its starting state). Similarly,
Pz {X0 = y, τz+ > 0} is equal to 1 when y = z and 0 otherwise, since z = X0 and
τz+ > 0 are always true by definition. Thus, these two terms are always equal, so
they cancel out, leaving us with (e
π P )(y) = πe(y) ∀y ∈ Ω, or
π
eP = π
e.
This proves, that π
e is stationary, so to make it a stationary distribution, we must
divide each element by the sum of the elements, which is equal to
X X
π
e(x) = Ez (number of visits to x before returning to z) = Ez (τz+ ),
x∈Ω x∈Ω

as the return time for any chain is equal to the total number of states it visits before
returning to its start.
Thus, we define
π
e(x)
π(x) = ,
Ez (τz+ )
which exists since τz+ > 0 by definition, and we will get a stationary distribution
(a stationary vector multiplied by a scalar is still stationary). So by Corollary 2.3,
this π is the only such stationary distribution. As such, for any choice of z ∈ Ω, we
will get the same stationary distribution π, so
π
e(x)
π(x) = ∀x ∈ Ω.
Ex (τx+ )
Note that choosing z = x also changes the definition of πe, so that π
e(x) is now the
expected number of visits to x before returning to x, which is exactly 1, for the one
time the chain hits x upon returning. Thus,
1
π(x) = ∀x ∈ Ω
Ex (τx+ )
is a unique stationary distribution for P . 
8 ARI FREEDMAN

4. Convergence Theorem
We have now shown that all irreducible Markov chains have a unique stationary
distribution π. However, in order to ensure that any distribution over such a chain
will converge to π, we require one more condition, called aperiodicity.
Definition 4.1. Let T (x) = {t ≥ 1 : P t (x, x) > 0} be the set of all time steps
for which a Markov chain can start and end in a state x. Then the period of x is
gcd T (x).
Lemma 4.2. If P is irreducible, then the period of all states is equal, or
gcd T (x) = gcd T (y) ∀x, y ∈ Ω.
Proof. Fix states x and y. Since P is irreducible, ∃r, l ≥ 0 such that P r (x, y) > 0
and P l (y, x) > 0 (Fig. 2).

l steps

a steps x y

r steps

Figure 2. Since P is irreducible, we can get from x to y in r steps,

from y to x in l steps, and from x to itself in a ∈ T (x) steps.

Let m = r + l. Then m ∈ T (x), since we can get from x to y in r steps and then
from y back to x in l steps, adding up to r + l = m steps. Similarly, m ∈ T (y),
going from y to x and back to y. If a ∈ T (x), there exists a path from x to itself in
a steps (Fig. 2), so then a + m ∈ T (y) by going from y to x, from x to itself, and
from x back to y, totalling l + a + r = a + m steps. Thus, a ∈ T (y) − m ∀a ∈ T (x),
where T (y) − m = {n − m | n ∈ T (y)}, so
T (x) ⊂ T (y) − m.
Take any n ∈ T (y), so gcd T (y) | n by the definition of gcd . Thus, m ∈ T (y)
implies gcd T (y) | m as well. This means gcd T (y) | n − m ∀n ∈ T (y), or equiva-
lently,
gcd T (y) | a ∀a ∈ T (y) − m.
And we showed that T (x) ⊂ T (y) − m, so this also gives us gcd T (y) | a ∀a ∈ T (x).
So gcd T (y) is a common divisor of T (x), which implies, by the definition of gcd,
that
gcd T (y) | gcd T (x).
By a completely parallel argument, switching around x and y, we also get that
gcd T (y) | gcd T (x). Therefore, gcd T (x) = gcd T (y) ∀x, y ∈ Ω. 
This shows that an irreducible Markov chain has a period common to all of its
states, which we then call the period of the chain.
Definition 4.3. An irreducible Markov chain is called aperiodic if its period is
equal to 1, or equivalently, gcd T (x) = 1 ∀x ∈ Ω.
Before being able to prove the Convergence Theorem, we need one result con-
cerning aperiodic chains, and a number-theoretic lemma to prove it.
 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 9

Lemma 4.4. If S ⊂ Z+ ∪ {0} is closed under addition (a + b ∈ S ∀a, b ∈ S) and

gcd S = 1, then there exists M such that a ∈ S ∀a ≥ M.
Proof. We begin by showing that there exists a finite subset T ⊂ S such that
gcd T = 1. Let S0 = {a0 }, for any a0 ∈ S. Either gcd S0 = 1, in which case we let
T = S0 and we are done, or there exists a1 ∈ S for which gcd(S0 ∪ {a1 }) < gcd S0 ,
since otherwise we would have gcd S = gcd S0 6= 1. So we let S1 = S0 ∪ {a1 },
and then gcd S1 < gcd S0 . We continue this process of finding ai ∈ S such that if
Si = gcd(Si−1 ∪ {ai }), then gcd Si < gcd Si−1 , creating a sequence of finite sets Si
whose gcd decreases until eventually gcd Si = 1, at which point we let T = Si and
we are done. We know this will occur at some point, since any strictly decreasing
sequence of positive integers must hit 1 in a finite number of steps.
Since gcd T = 1, there exists a linear combination of the elements in T that
evaluates to 1, so if T = {t1 , . . . , tn }, then there are c1 , . . . , cn ∈ Z such that
c1 t1 + · · · + cn tn = 1.
Without loss of generality, we can say that all of c1 , . . . , ck ≥ 0 and ck+1 , . . . , cn < 0,
for some 1 ≤ k ≤ n, so we can move all the negative terms to the other side to get
c1 t1 + · · · + ck tk = 1 + ck+1 tk+1 + · · · + cn tn .
Since S is closed under addition, and all ti ∈ T ⊂ S, the non-negative linear
combinations on the left side and right side of this equation are also in S, so if we
let these be equal to p and q respectively, we get p, q ∈ S such that p = 1 + q. It is
possible that there are no terms other than 1 on the right side, but this just means
p = 1 ∈ S, so since S is closed under addition, we can let M = 1 and we are done.
Otherwise we have p = 1 + q for p, q ∈ S.
Let M = q(q − 1). Then for any a ≥ M, we have a = kq + r for 0 ≤ r < q and
k ≥ q − 1, by the remainder theorem. Thus r ≤ q − 1 ≤ k, so k − r ≥ 0, and we
can express a as a non-negative linear combination of p and q:
a = kq + r = (k − r)q + r(q + 1) = rp + (k − r)q.
Therefore, a ∈ S ∀a ≥ M , as desired. 
Lemma 4.5. If P is aperiodic and irreducible, then there exists r ≥ 0 such that
P r (x, y) > 0 ∀x, y ∈ Ω.
Proof. For any x ∈ Ω, we see that T (x) is a set of non-negative integers closed
under addition, since if a, b ∈ T (x), we have P a (x, x) > 0 and P b (x, x) > 0, so
P a+b (x, x) ≥ P a (x, x)P b (x, x) > 0 and thus a + b ∈ T . And since P is aperiodic,
gcd T (x) = 1, which means we can apply Lemma 4.4 to get that there is an Mx
such that P t (x, x) > 0 ∀t ≥ Mx . If we then let M = max{Mx : x ∈ Ω}, we have
P t (x, x) > 0 ∀t ≥ M ∀x ∈ Ω.
Since P is irreducible, there exists a t for any x, y ∈ Ω such that P t (x, y) > 0.
We let t0 = max{any one t satisfying P t (x, y) > 0 : x, y ∈ Ω} be the maximum of
all of these, and finally let r = t0 + M . Then for any states x, y ∈ Ω, there exists
a t ≤ t0 such that P t (x, y) > 0, and then r − t = M + (t0 − t) ≥ M, which means
P r−t (x, x) > 0, so then we get
P r (x, y) ≥ P r−t (x, x)P t (x, y) > 0.

10 ARI FREEDMAN

To prove anything about the convergence of a distribution, we need to first define

some measure of distance between distributions.
Definition 4.6. The total variation between two distributions µ and ν is defined
as
||µ − ν||T V = max |µ(A) − ν(A)|,
A⊂Ω
P
where µ(A) = µ(x).
x∈A
1
P
Proposition 4.7. ||µ − ν||T V = 2 |µ(x) − ν(x)|.
x∈Ω

Proof. Let B = {x ∈ Ω : µ(x) ≥ ν(x)} be the set of states for which µ(x)−ν(x) ≥ 0,
so its complement is B c = Ω \ B = {x ∈ Ω : µ(x) < ν(x)}, which is the set of states
for which µ(x) − ν(x) < 0.
Let A ⊂ Ω be any set of states. Now A is the disjoint union of A ∩ B and A ∩ B c ,
and any x ∈ A ∩ B c is in B c and thus µ(x) − ν(x) < 0, so
µ(A) − ν(A) = µ(A ∩ B) − ν(A ∩ B) + µ(A ∩ B c ) − ν(A ∩ B c ) ≤ µ(A ∩ B) − ν(A ∩ B).
And B is the disjoint union of A ∩ B and B \ A, and any x ∈ B \ A is in B and
thus µ(x) − ν(x) ≥ 0, so
µ(A ∩ B) − ν(A ∩ B) ≤ µ(A ∩ B) − ν(A ∩ B) + µ(B \ A) − ν(B \ A) = µ(B) − ν(B).
Putting these together, we get
µ(A) − ν(A) ≤ µ(B) − ν(B),
and by symmetric logic with A ∩ B c as an intermediary, we get
µ(B c ) − ν(B c ) ≤ µ(A) − ν(A).
Now µ(B) + µ(B c ) = ν(B) + ν(B c ) = 1 implies µ(B c ) − ν(B c ) = −(µ(B) − ν(B)),
so
−(µ(B) − ν(B)) ≤ µ(A) − ν(A) ≤ µ(B) − ν(B),
or |µ(A) − ν(A)| ≤ µ(B) − ν(B). Thus, |µ(A) − ν(A)| is bounded by µ(B) − ν(B),
and we can let A = B to attain this bound, since |µ(B) − ν(B)| = µ(B) − ν(B)
from the fact that µ(B) − ν(B) ≥ 0 ∀x ∈ B. Thus,
||µ − ν||T V = max |µ(A) − ν(A)|
A⊂Ω
= µ(B) − ν(B)
1 h   i
= µ(B) − ν(B) + ν(B c ) − µ(B c )
2 !
1 X X
= |µ(x) − ν(x)| + |µ(x) − ν(x)|
2
x∈B x∈B c
1X
= |µ(x) − ν(x)|
2
x∈Ω

Remark 4.8. This proof also gives us ||µ − ν||T V = µ(B) − ν(B). Since µ(B) ≤ 1
and ν(B) ≥ 0, this tells us ||µ − ν||T V ≤ 1, for any distributions µ and ν.
 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 11

Now we are finally ready to prove the main result of this paper, which tells us
that an irreducible, aperiodic Markov chain will converge at an exponential rate to
a stationary distribution over time.

Theorem 4.9 (Convergence Theorem). If P is irreducible and aperiodic, with

stationary π, then there exist constants 0 < α < 1 and C > 0 such that

max ||P t (x, ·) − π||T V ≤ Cαt .

x∈Ω

Proof. Since P is irreducible and aperiodic, Lemma 4.5 tells us there exists an r ≥ 0
satisfying P r   > 0 ∀x, y ∈ Ω.
(x, y)
π
Let Π =  ...  be the |Ω| × |Ω| matrix all of whose rows are the stationary
 

π
distribution π, making Π a stochastic matrix.
Since P r (x, y) > 0 ∀x, y ∈ Ω and π(y) > 0 ∀y ∈ Ω, by Theorem 3.3, then

P r (x, y)
δ 0 = min >0
x,y∈Ω π(y)

satisfies P r (x, y) ≥ δ 0 π(y) ∀x, y ∈ Ω. So δ = min{δ 0 , 21 } also satisfies this property,

as well as 0 < δ < 1, so setting θ = 1 − δ, we get 0 < θ < 1.
Now define
P r − (1 − θ)Π
Q= .
θ

Every element of Q is non-negative, since P r (x, y) − (1 − θ)Π(x, y) = P r (x, y) −

δπ(y) ≥ 0, by the definition of δ, and, because P r and Π are stochastic, each row
of Q sums to 1−(1−θ)
θ = 1, making Q a stochastic matrix.
For any n ≥ 0, Qn is stochastic, from Q being stochastic, so we get Qn Π = Π,
since
X X
(Qn Π)(x, y) = Qn (x, z)Π(z, y) = π(y) Qn (x, z) = π(y) = Π(x, y).
z∈Ω z∈Ω

And since πP = π,
     
π πP π
 ..   ..   .. 
ΠP =  .  P =  .  =  .  = Π,
π πP π

so ΠP n = (ΠP )P n−1 = ΠP n−1 = · · · = ΠP = Π.

Using these identities, we will now prove inductively that

P rk = (1 − θk )Π + θk Qk ∀k ≥ 1.
12 ARI FREEDMAN

The base case k = 1 is true, since P r = (1 − θ)Π + θQ from the definition of Q.

Now assume it is true for k = n, so P rn = (1 − θn )Π + θn Qn . Then

P r(n+1) = P rn P r
= [(1 − θn )Π + θn Qn ]P r
= (1 − θn )ΠP r + θn Qn P r
= (1 − θn )Π + θn Qn [(1 − θ)Π + θQ]
= (1 − θn )Π + θn (1 − θ)Qn Π + θn+1 Qn+1
= (1 − θn )Π + (θn − θn+1 )Π + θn+1 Qn+1
= (1 − θn+1 )Π + θn+1 Qn+1 ,

proving inductively that P rk = (1 − θk )Π + θk Qk ∀k ≥ 1. Multiplying both sides

by P j for j ≥ 0,

P rk+j = (1 − θk )ΠP j + θk Qk P j = (1 − θk )Π + θk Qk P j = Π + θk (Qk P j − Π),

or P rk+j − Π = θk (Qk P j − Π). Looking at any row x of both sides of this equation,
we have (P rk+j − Π)(x, ·) = θk (Qk P j − Π)(x, ·). Thus,

||P rk+j (x, ·) − π||T V = ||P rk+j (x, ·) − Π(x, ·)||T V

1X
= |(P rk+j − Π)(x, y)|
2
y∈Ω
1X
= θk |(Qk P j − Π)(x, y)|
2
y∈Ω

= θ ||Qk P j (x, ·) − Π(x, ·)||T V

k

≤ θk ,

since Q, P, and Π being stochastic implies Qk P j (x, ·) and Π(x, ·) are distributions,
thus their total variation is bounded by 1√from Remark 4.8.
Now we can define the constants α = r θ and C = α−r . Let t ≥ 0, and define k
and j from the division theorem by r, so t = rk + j with 0 ≤ j < r. Finally, since
0 < θ < 1 and 0 < α < 1,

||P t (x, ·) − π||T V = ||P rk+j (x, ·) − π||T V

≤ θk
= αrk
≤ αrk αj−r
= α−r αrk+j
= Cαt .

This is all true for any choice of x ∈ Ω; therefore, for any t ≥ 0,

max ||P t (x, ·) − π||T V ≤ Cαt .

x∈Ω


 CONVERGENCE THEOREM FOR FINITE MARKOV CHAINS 13

Acknowledgments. I am deeply grateful to my mentor, Nat Mayer, for devoting

his time to learning this subject with me and helping me through the understanding
and writing of this paper. I must also thank László Babai for teaching a wonderful
and instructive course which gave me much insight into the tools needed for this
paper, as well as Peter May for reviewing it and organizing the University of Chicago
REU.

References
[1] David A. Levin, Yuval Peres and Elizabeth L. Wilmer. Markov Chains and Mixing Times.
http://pages.uoregon.edu/dlevin/MARKOV/markovmixing.pdf.