Stat205A: Probability Theory (Fall 2002) Lecture: 23-24

Recurrence and Transience of Random Walks

Lecturer: James W. Pitman Scribe: Tianbing Chen [email protected]

In this lecture, let X1 , X2 , · · · be i.i.d. and Sn = X1 + X2 + · · · + Xn , Sn = 0. Sn is a random

walk.

Theorem 23.1 Let X1 , X2 , · · · be i.i.d. , Fn = σ(X1 , · · · , Xn ), τ is a stopping time. Conditional

on τ < ∞, Sτ , Sτ +1 , · · · is a R.W.1 started at Sτ . i.e. Xτ +1 , Xτ +2 , · · · are i.i.d. and independent of
Fτ .

Proof Sketch: Conditional on τ

P(τ = n, (X1 , · · · , Xτ ) ∈ A, (Xτ +1 , · · · , Xτ +m ) ∈ B)

= P(τ = n, (X1 , · · · , Xn ) ∈ A, (Xn+1 , · · · , Xn+m ∈ B)
= P(τ = n, (X1 , · · · , Xn ) ∈ A)P((X1 , · · · , Xm ) ∈ B)

Summing over n gets the desired result.

Definition 23.1 The number x ∈ R is said to be a recurrent value for the R.W. Sn if for every
 > 0, P(Sn ∈ (x ± )2 ) = 1.

Definition 23.2 We say F is Lattice with period d if F (Zd) = 1 and d is the greatest positive
number with this property. Otherwise, there is no such d, and it’s called Non-lattice.

Example 23.1 F = 12 (δe + δ1 ) is non-lattice.

Theorem 23.2 If R.W. Sn is lattice with range Zd as above, then either 1) or 2)

1) each x ∈ Zd is recurrent,
2) each x ∈ Zd is transient.

Proof Sketch: Just Markov chain theory.

Definition 23.3 y is said to be possible if for every open interval I, there exists k, s.t.
P(Sk ∈ I) > 0.

Lemma 23.3 If x is recurrent and y is possible, then x−y is recurrent.

1 random walk
2x ±  := [x − , x + ]

23-1
23-2 Recurrence and Transience of Random Walks

Proof Sketch: Take  > 0, then there exists k, s.t. P(|Sk − y| < ) > 0. From Theorem 23.1

P(|Sn − x| < 3, f.o.3 ) ≥ P(|Sk − y| < , |S( k + n) − Sk − (x − y)| < 2, f.o.)
= P(|Sk − y| < ) P(Sn − (x − y)| < 2, f.o.)

If P(Sn − (x − y)| < 2, f.o.) > 0, then P(|Sn − x| < 3, f.o.) > 0, which is a contradiction!

Theorem 23.4 If R.W. Sn is non-lattice, then similarly either 1) or 2)

1) each x ∈ R is recurrent,
2) each x ∈ R is transient.

Proof Sketch: Let G={x ∈ R: x is recurrent}. Suppose G 6= ∅, then

• It is clear that Gc is open, so G is closed4 .

• From the above lemma, if x ∈ G and y ∈ G, then x − y ∈ G. Therefore, G is a group.

Since G is a closed subgroup of R and the R.W. is non-lattice, it follows that G=R.

Note: If E(X) is defined , finite and not 0, then the R.W. is transient (i.e. {recurrent points}=∅)
by S.L.L.N5 .

P P
Definition 23.4 U: potential measure. For any interval I , U (I):= n P(Sn ∈ I) = E( n 1(Sn ∈I) ).

Lemma 23.5 P(Sn ∈ (x ± /2) for some n) U(±/2) ≤ U(x ± ) ≤ U(±2)

Proof Sketch: Let τ :=the first hit of (x ± ), then Sτ +n ∈ (x − , x + ) ⇒ (Sτ +n − Sτ ) ∈ (±2).
Therefore from Theorem 23.1

U(x ± ) = E[the number of times n that Sn ∈ (x ± )]

= E[the number of times n that(Sτ +n − Sτ ) ∈ (±)]
≤ E[the number of times n that Sn ∈ (±2)]
= U(±2)

Let τ1 :=the first hit of (x ± /2). Use the same argument

P(Sn ∈ (x ± /2) for some n) U(±/2) = E[the number of times n that(Sτ1 +n − Sτ1 ) ∈ (±/2)]
≤ E[the number of times n that Sn ∈ (x ± )]
= U(x ± )

Corollary 23.6 U(±k) ≤ (2k + 1) U(±), ∀ k ∈ N.

3 finitely often
4 topologically closed
5 Strong Law of Large Numbers
Recurrence and Transience of Random Walks 23-3

Proof Sketch: Cover (−k, k) with (2k+1) intervals of the form (x ± /2). Use the fact that U
is a measure.

Proposition 23.7 Either U(I) < ∞ for all bounded intervals I (transient case) or U(x ± ) = ∞
for all possible x and all  > 0.

Proof Sketch: Consider U(±δ):

1) If U(±δ) < ∞ for some δ > 0, then

U(±kδ) ≤ (2k + 1)U(±δ) < ∞, ∀ k ∈ N =⇒ U(I) < ∞ for all bounded intervals I.

2) If U(±δ) = ∞ for all δ > 0, then from Lemma 23.5

P(Sn ∈ (x ± δ/2) for some n) U(±δ/2) ≤ U(x ± δ) =⇒ U(x ± δ) = ∞ for all δ > 0 if x is
possible.

Theorem 23.8 Either U(±1) < ∞ and no x is recurrent or U(±1) = ∞ and every possible x is
recurrent.

Proof Sketch: If U(±1) < ∞, no x is recurrent by Borel-Cantelli lemma.

The other way:
If Sn is in an interval I only finitely often, consider τ := the last time that Sn ∈ I.
Careful: τ is not a stopping time since {τ = n} = {Sn ∈ I, Sn+1 ∈ I, Sn+2 ∈ I, · · · }
{τ = 0} = {Sn ∈ I, for all n}, {τ = ∞} = {Sn ∈ I, i.o.}.

Since U(±1) = ∞ by assumption, we know that U(±) = ∞, ∀  > 0 by estimate:

U(±k) ≤ (2k + 1)U(±) f or k ≥ 1 . P
Let τ :=last time that the R.W. is in (±), then P(Sn ∈ (±), f.o.) = n P(τ = n).

{τ = n} = {Sn ∈ (±)} ∩ {Sn+k 6∈ (±), ∀ k ≥ 1}

⊃ {Sn ∈ (±)} ∩ {Sn+k − Sn 6∈ (±2), ∀ k ≥ 1}

Therefore P(τ = n) ≥ P(Sn ∈ (±) P(|Sk | ∈ 2, ∀ k ≥ 1). Sum over n:

1 ≥ P(Sn ∈ ±, f.o.) ≥ U(±) P(|Sk | ≥ 2, ∀ k ≥ 1)

But U(±) = ∞, which forces the term P(|Sk | ≥ 2, ∀ k ≥ 1) to be 0.

Rewrite what we have proved:

U(±1) = ∞ =⇒ P(|Sn | ≥ δ, ∀ n ≥ 1) = 0 for all δ > 0

Finish the argument:

23-4 Recurrence and Transience of Random Walks

P(Sn ∈ (±), f.o.) = P(τ < ∞ and Sτ ∈ (±))

1
= lim P(τ < ∞ and Sτ ∈ (±(1 − )))
k→∞ k
X
∞
1
= lim P(Sn ∈ (±(1 − )) and Sn+j 6∈ (±), ∀ j ≥ 1)
k→∞
n=0
k
X∞
1 
≤ lim P(Sn ∈ ±(1 − )) P(|Sj | ≥ , ∀ j ≥ 1)
k→∞
n=0
k k
= 0 since P(|Sn | ≥ δ, ∀ n ≥ 1) = 0 f or all δ > 0.

Key idea here: Think about the last time in the trip.

Theorem 23.9 (Chung-Fuchs Theorem) Suppose E|X1 | < ∞.

• If EX1 6= 0, then the R.W. is transient(by S.L.L.N.)

• If EX1 = 0, then all possible points are recurrent.

Proof Sketch: We’ll show U(±1) = ∞ when EX1 = 0. We know

1
U(±1) ≥ ( )U (±k)
2k + 1
1 X∞
= ( ) P(Sn ∈ (±k))
2k + 1 n=0

Take  > 0, and choose k so that

P( |Snn | < ) ≥ 1
2 for all n ≥ k (by W.L.L.N. 6 )

For k ≤ n ≤ k , P(|Sn | < k) ≥ 12 . Hence

6 Weak Law of Large Numbers

Recurrence and Transience of Random Walks 23-5

X
∞
1 X
∞
P(Sn ∈ (±1)) ≥ ( ) P(Sn ∈ (±k))
n=0
2k + 1 n=0
1 X
≥ ( ) P(Sn ∈ (±k))
2k + 1 k
k≤n≤ 
1 X 1
≥ ( )
2k + 1 2
k≤n≤ k

1 1 k
≥ ( ) ( − k)
2 2k + 1 
1 1 k
≥ ( ) ( − k)
2 3k 
1 1
≥ ( − 1)
6 
→ ∞ as  → 0.