MATH 275B notes 168

25. R ECURRENCE
Our next topic finds its motivation in Boltzmann’s work. Recall that his (and Maxwell’s)
Ergodic Hypothesis asserts that the time average of a function tends, in the limit of large
time intervals, to the average of f against a measure. In 1930s that convergence was
proved to take place reducing the problem to the characterization of ergodic measures.
However, several decades earlier, physicists P. Ehrenfest and his wife T. Afanassjewa
wondered and wrote extensively about the reasons why such a statement should be
true. Their conclusion was that this is because the trajectory of a Hamiltonian systems
visits every point in the configuration space.
Such a statement was surprising and was soon found to be mathematically impossi-
ble, but a slight modification sounded reasonable: the trajectory comes arbitrarily close
to every point. This is the so called quasi-ergodic hypothesis which is naturally linked with
the concept of recurrence that we will address in this chapter.

25.1 Poincaré, Khinchin and Kac recurrence theorems.

We start with an old result that addresses the above problem:

Theorem 25.1 (Poincaré 1899) Let (X , G , µ, j) be a m.p.s. with µ(X ) † 8. Then
⇣ (⌘
@A P G : µ x P A : (@k • 1 : jk ( x ) R A) = 0 (25.1)

In words, almost every x P A will return to A after a finite number of iterations of j.

Proof. Denote
( §
B := x P A : (@k • 1 : jk ( x ) R A) = A r j´k ( A) (25.2)
k•1

and observe that j´1 ( B) Ñ j´1 ( A) and thus B X j´1 ( B) = 0. Similarly we show that
@n • 1 : j´n ( B) X B = H. Applying j´k on this we get
@n ° k • 0 : j´n ( B) X j´k ( B) = H (25.3)
It follows∞that tj´n ( B) : n • 0u is a disjoint collection with µ( j´n ( B)) = µ( B). Since
µ(X ) • n•0 µ( j´n ( B)) we must have µ( B) = 0 whenever µ(X ) † 8. ⇤
This naturally leads to:
Definition 25.2 (Recurrent point) Let (X , G , µ, j) be a m.p.s. and A P G . A point x P X
is recurrent with respect to A if there exists k • 1 such that jk ( x ) P A.
The Poincaré Recurrence Theorem thus states that almost every point in A is recurrent
with respect to A. Another way to phrase this is by introducing the first return time to A,
(
n A ( x ) := inf n • 1 : jn ( x ) P A (25.4)
The Poincaré Recurrence Theorem then says
n A † 8 a.e. on A (25.5)

Preliminary version (subject to change anytime!) Typeset: April 21, 2025

169 MATH 275B notes

Note that, as our example of j( x ) = x + 1 on R shows, this fails in infinite measure

spaces; we will return to what may happen there later.
A natural question is how long will it take for a point to return to A. We have two
theorems to discuss in this regard, the first of which focuses on returns of a set rather
than a point itself:
Theorem 25.3 (Khinchin 1934) Let (X , G , µ, j) be a m.p.s. with µ(X ) = 1. Then for all
A P G and all e ° 0, the set
(
Ee = n P N : µ( j´n ( A) X A) • µ( A)2 ´ e (25.6)
has “bounded gaps” in the sense that
Dn • 1 @k P N : Ee X tk, k + 1, . . . , k + n ´ 1u ‰ H (25.7)

Proof. Consider the L2 -isometry defined by T f = f ˝ j. Let A P G and observe that

µ( j´n ( A) X A) = xT n 1 A , 1 A y (25.8)
The Mean Ergodic Theorem implies that for each e ° 0 there exists n • 0 such that
› 1 n´1 ›
› ÿ i › e
› T 1 A ´ P1 A › † (25.9)
n 2 1 + }1 A }
i =0

Since TP = P, applying Tk inside the norm gives

› 1 n+ÿ
k´1 › e
› ›
› T i 1 A ´ P1 A › † (25.10)
n 2 1 + }1 A }
i=k
Invoking the Cauchy-Schwarz inequality, this implies
ˇA n+k´1 Eˇˇ
ˇ 1 ÿ i e}1 A }
ˇ T 1 A ´ P1 A , 1 A ˇˇ † §e (25.11)
ˇ n 1 + }1 A }
i=k
In order to rewrite the inner product, we note
xP f , 1y2 x f , 1y2
xP f , f y = xP f , P f y • = (25.12)
x1, 1y P1=1 x1, 1y
implying that
x1 A , 1y2
xP1 A , 1 A y • = µ ( A )2 (25.13)
x1, 1y µ(X )=1
Using this and (25.11) we now get
n+k´1 n+k´1
1 ÿ 1 ÿ
µ( j´k ( A) X A) = xT k 1 A , 1 A y
n n (25.14)
i=k i=k
• xP1 A , 1 A y ´ e = µ( A)2 ´ e
which proves the claim. ⇤
Theorem 25.3 seems very strong; indeed, it says that a “neighborhood” of a trajectory
of iterates of a m.p.t. will keep intersecting itself nearly regularly. However, one should

Preliminary version (subject to change anytime!) Typeset: April 21, 2025

MATH 275B notes 170

not exaggerate the interpretation because the “neighborhood” can look very different
every time it “comes back” — after all, no statement is made about how its points get
scrambled around. This should be no surprise as all that goes into the proof of Theo-
rem 25.3 is the measure-preserving property of j.
Moving to quantifying the time it takes for a point to return to a set, we get:
Theorem 25.4 (Kac 1947) Let (X , G , µ, j) be a m.p.s. with µ(X ) † 8. Assume that j
admits a measurable inverse (which is then also a m.p.t.). Then
ª
@A P G : n A dµ § µ(X ) (25.15)
A

If µ is ergodic and µ( A) ° 0, then equality holds.

Proof. Consider the sets An := A X tn A = nu and note that n A ( x ) = n implies jk ( x ) R A

for k = 1, . . . , n ´ 1. Hence jk ( An ) X A = H whenever k = 1, . . . , n ´ 1 and so
@n, m • 1 @k = 1, . . . , n ´ 1 : jk ( An ) X Am = H (25.16)
We thus get that tjk ( An ) : 0 § k † nu is a disjoint family. These sets are nicely visualized
in the Kakutani skyscraper depicted in Fig. 1.
Noting that µ( jk ( An )) = µ( An ), from (25.5) we obtain
ª ÿ
n A dµ = nµ( An )
n•1
✓ § n´1 ◆ (25.17)
ÿ n´1
ÿ §
k k
= µ j ( An ) = µ j ( An ) § µ (X )
n•1 k =0 n•1 k =0

proving the first part of the claim.

î î
For the second part observe that, since jn ( An ) Ñ A and A Ñ n•1 n´1 k
k =0 j ( An ) a.e.
(meaning that inclusion holds up to a null set), we have
✓ § n´1
§ ◆ § n´1
§
j jk ( An ) Ñ jk ( An ) a.e. (25.18)
n•1 k =0 n•1 k =0

But the two sets have the same measure ≥ and so equality holds up to null sets. Denoting
the set as S A , we have S A P I 1 and A n A dµ = µ(S A ). Assuming µ( A) ° 0, the integral
is strictly positive and so µ(S A ) ° 0. If µ is also ergodic, we must have µ(S A ) = µ(X ),
proving that equality holds. ⇤
The theorem is particularly useful in the ergodic case in which case we can rewrite
(assuming µ( A) ° 0) the result as
ª ⇣ µ( A) ⌘´1
1
n A dµ = (25.19)
µ( A) A µ (X )
This tells us that, conditional on starting from A, the expected time for the iterates of j
to return to A is equal to the inverse of the (proportional) mass of µ in A.

Preliminary version (subject to change anytime!) Typeset: April 21, 2025

171 MATH 275B notes

nA = 1 nA = 2 nA = 3 nA = 4 nA = 5 nA = 6 ...

F IGURE 1. The Kakutani skyscaper for (non-null) set A with j invertible. Here the
“foundation” is the set A decomposed into “intervals” An := A X tn A = nu. (We
ignore the part where n A = 8 as it is null.) Iterations of j on each An at the point
keeps lifting it, one floor at the time, until a “patio” is hit, which j maps back to A.

25.2 Some interesting consequences.

Let us present some interesting consequences of the above conclusions. The Poincaré
Recurrence Theorem naturally leads to the concept of induced transform (a.k.a. derivative
transform). This is the map j A : A Ñ A defined for any A P G with µ( A) ° 0 as
#
j n A ( x ) ( x ), if n A ( x ) † 8
j A ( x ) := (25.20)
x, if n A ( x ) = 8
where n A is as in (25.4). Under µ(X ) † 8 and with j an m.p.t., the Poincaré Recur-
rence Theorem ensures the validity of (25.5) and the second line in the definition above
becomes a null-set proviso, which is desired as the map acts trivially in this case.
A key point about the induced transform is that it inherits the properties of j:
Proposition 25.5 Let µ be a finite measure on (X , G) and let j : X Ñ X be a map. Given
A P G such that µ( A) ° 0, define G A := tA X B : B P G u and µ A ( B) := µ( A X B). Then
(1) j is G -measurable ñ j A is G A -measurable
(2) j m.p.t. on (X , G , µ) ñ j A m.p.t. on ( A, G A , µ A )
(3) j ergodic with respect to µ ñ j A ergodic with respect to µ A
We leave the proof of this proposition to a homework assignment. The concept of
the induced transform appears (implicitly or explicitly) throughout probability. To give
an example, consider a stationary sequence tXk uk•0 of t0, 1u-valued random variables
with P( X0 = 0) ° 0. Define A := tX0 = 0u and let q be the left shift. Then q A is

Preliminary version (subject to change anytime!) Typeset: April 21, 2025

MATH 275B notes 172

map representing the “shift to the nearest zero” which is well defined because, by the
Poincaré Recurrence Theorem, tXk uk•0 will contain infinitely many zero’s a.s.
Our next observation concerns the arguments relying on the Kakutani skyscraper in
the proof of Theorem 25.4. For convenience of expression, here we say that “A Ñ B a.e.”
to denote µ( A r B) = 0 and “A = B a.e.” to denote µ( A4 B) = 0.
Corollary 25.6 Let (X , G , µ, j) be a measurably-invertible m.p.s. with µ(X ) † 8 and,
given A P G , let
§ n´1
§
S A := jk A X tn A = nu (25.21)
n•1 k =0
be the set represented by the Kakutani skyscraper. Then S A P I 1 and A Ñ S A a.e. Moreover, S A
is minimal with this property in the sense
@B P I 1 : A Ñ B a.e. ñ S A Ñ B a.e. (25.22)
In particular, µ ergodic implies µ(S A ) P t0, µ(X )u.

Proof. All the properties have been already proved except (25.22). Suppose A Ñ B a.e.
for B P I 1 . Then An := A X tn A = nu Ñ B a.e. for each n • 1 and so jk ( An ) Ñ B a.e. for
each 0 § k † n. But then also S A Ñ B a.e., as desired. ⇤
The Kakutani skyscraper picture allows one to deduce that the whole space is, to
within arbitrary small error, an orbit of a given length of a particular set:
Theorem 25.7 (Kakutani 1943, Rokhlin 1948) Let (X , G , µ, j) be an measurably-invertible
ergodic m.p.s. with µ(X ) † 8. Assume that µ is non-atomic (i.e., @A P G with µ( A) ° 0
there exists B P G with 0 † µ( B) † µ( A)). Then for each e ° 0 and m • 1 there exists B P G
such that
B, j´1 ( B), . . . , j´m+1 ( B) are disjoint (25.23)
and
✓ m´1
§ ◆
´k
µ X r j ( B) † e (25.24)
k =0

The idea of the proof is to pick A of small measure and consider the Kakutani sky-
scraper associated with A. Then we take a suitable subset of its parts to be B so that the
two properties hold true. We leave the details to a homework assignment.

Preliminary version (subject to change anytime!) Typeset: April 21, 2025