In this lecture we prove Birkho's Ergodic Theorem (also called the Pointwise Ergodic Theorem):
Theorem: Let µbe a probability measure andψ∈L1(µ). Then the ergodic average
ψ∗(x) := lim
n→∞
1 n
n−1
X
i=0
ψ◦Ti(x)
existsµ-a.e., andψ∗ is T -invariant, i.e.,ψ∗◦T =ψ∗ µ-a.e.
If in additionµis ergodicthenψ∗ =R
X ψdµ forµ-a.e. x Or in short:
Space Average = Time Average (for typical points).
measure preserving dynamical system. Take MN =max{Sn :0≤n≤N}. Then
Z
AN
f dµ≥0 for AN ={x ∈X :MN(x)>0}.
TheKoopman operatorUT :L1(µ)→L1(µ) is dened as
UTf =f ◦T.
Clearly UT is linear andpositive, i.e., f ≥0 implies UTf ≥0.
We write theergodic sumas
Sn=Snf =
n−1
X
k=0
f ◦Tk and S0 ≡0.
all 0≤n ≤N and by positivity of the Koopman operator, also UTMN ≥UTSn. Add f : UTMN+f ≥UTSn+f =Sn+1. For x ∈AN, this means
UTMN(x) +f(x) ≥ max
1≤n≤NSn(x)
≥x∈AN max
0≤n≤NSn(x) =MN(x).
Therefore f ≥MN −UTMN on AN, and (since MN ≥S0=0) Z
AN
f dµ ≥ Z
AN
MN dµ− Z
AN
UTMN dµ
= Z
X MN dµ− Z
AN
UTMN dµ
= Z
X MN dµ− Z
X UTMN dµ=0.
This completes the proof.
Lemma: Let(X,T,B, µ)be a probability measure preserving dynamical system, and E ⊂X a T -invariant subset. Let
Bα:={x ∈X :sup
n
1
nSng(x)> α}.
Then Z
Bα∩Eg dµ≥αµ(Bα∩E).
For a measurable set E ⊂X , we dene a probability measure µE(B) = 1
µ(E)µ(B∩E) If E is T -invariant thenµE is T -invariant too.
Proof: Ifµ(E) =0 then there is nothing to prove. So assume that µ(E)>0. Take f =g−α, so
Bα=∪NAN for AN ={x ∈X :MN(x)>0}.
Note also that AN ⊂AN+1 for all N. Therefore for each ε >0 there exists N ∈Nsuch that
Z
Bα
f dµE ≥ Z
AN
f dµE ≥ −ε.
Sinceεis arbitrary, R
Bαf dµE ≥0. Addingα again we have Z
Bα
g dµE = Z
Bα
f +α dµE ≥αµE(Bα∩E).
Multiply everything byµ(E) to get the lemma.
Proof of Birkho's Ergodic Theorem:
Recallψ∈L1(µ). Dene ψ=lim sup
n→∞
1
nSnψ and ψ=lim inf
n→∞
1 nSnψ.
Since
|n+1 n
1
n+1Sn+1ψ−1
nSnψ◦T|= 1
n|ψ(x)| →0 as n→ ∞, we have ψ◦T =ψand similarly ψ◦T =ψ. We want to show thatψ=ψ µ-a.e.
Let
Eα,β ={x ∈X :ψ(x)< β, α < ψ(x)}
Then Eα,β is T -invariant, and
{x ∈X :ψ(x)< ψ(x)}= [
α,β∈Q,β<α
Eα,β.
This is a countable union, and therefore it suces to show that µ(Eα,β) =0 for every pair of rationals β < α.
WriteBα :={x ∈X :supn1nSnψ(x)> α} as in our Lemma. Since Eα,β =Eα,β∩Bα, this Lemma gives
Z
Eα,β
ψ dµ= Z
Eα,β∩Bαψdµ≥αµ(Eα,β∩Bα) =αµ(Eα,β).
From the previous slide:
Z
Eα,β
ψ dµ≥αµ(Eα,β).
We repeat this argument replacingψ, α, β by −ψ,−α,−β. Note that−ψ=−ψ and−ψ=−ψ. This gives
Z
Eα,β
ψ dµ≤βµ(Eα,β).
Sinceβ < α, this can only be true if µ(Eα,β) =0.
Thereforeψ=ψ=ψ∗, i.e., the lim sup and lim inf are actually limitsµ-a.e.
The next step is to show thatψ∗ ∈L1(µ). From Measure Theory we have:
Fatou's LemmaIf(gn)n∈N are non-negative L1(µ)-functions, then lim inf
n gn∈L1(µ) and Z
X lim inf
n gn dµ≤lim inf
n
Z
X gn dµ.
Here we apply this to gn=|1nSnψ|, which belong to L1(µ) because (by T -invariance)
Z
X
|1
nSnψ|dµ≤ 1 n
n−1
X
k=0
Z
X
|ψ| ◦Tkdµ= Z
|ψ|dµ <∞.
Hence in the limit: R
X|ψ|dµ≤lim infnR
X |ψ|dµ <∞.
Next, we need to show that X ψ dµ= X ψdµ(so without absolute value signs). Take
Dk,n={x ∈X : k
n ≤ψ∗(x)< k+1 n }.
Then Dk,n is T -invariant, and ∪k∈ZDk,n =X modµ. Also Bk
n−ε∩Dk,n=Dk,n forε >0. Therefore our Lemma gives Z
Dk,n
ψ dµ = Z
Bk n−ε∩Dk,n
ψ dµ
≥ (k
n −ε)µ(Bk
n−ε∩Dk,n)
= (k
n −ε)µ(Dk,n).
n Dk,n
Z
Dk,n
ψ∗ dµ≤ k+1
n µ(Dk,n)≤ 1
nµ(Dk,n) + Z
Dk,n
ψdµ.
Summing over all k ∈Z, we ndR
Xψ∗ dµ≤ 1n+R
X ψdµ. Since n∈Nis arbitrary, also
Z
X ψ∗ dµ≤ Z
X ψ dµ, By the same argument for−ψ, we ndR
X ψ∗ dµ≥R
Xψ dµ.
Hence Z
Xψ∗= Z
Xψ dµ.
Finally, ifµis ergodic, the T -invariant functionψ∗ has to be constantµ-a.e., soψ∗ =R
ψdµ. This completes the proof of Birkho's Ergodic Theorem.
The Lp Ergodic Theorem is a generalisation of Von Neumann's L2 version of the Ergodic Theorem, which predates1 Birkho's Theorem, but nowadays, it is usually proved as a corollary of the pointwise ergodic theorem.
Theorem: Let (X,T,B, µ) be a probability measure preserving dynamical system. Ifµis ergodic, andψ∈Lp(µ)for some 1≤p <∞ then there existsψ∗ ∈Lp(µ) with ψ∗◦T =ψ∗ µ-a.e.
such that
k1
nSnψ−ψ∗kp→0 as n→ ∞.
1John von Neumann was earlier in proving his L1-version, but Birkho delayed its publication until after the appearance of his own paper.
First assume thatψis bounded (and hence in Lp(µ)). By Birkho's Ergodic Theorem there isψ∗ such that n1Snψ(x)→ψ∗(x) µ-a.e., andψ∗ is bounded (and hence in Lp(µ) too). In particular,
|1
nSnψ(x)−ψ∗(x)|p→0 µ-a.e.
By the Bounded Convergence Theorem, we can swap the limit and the integral:
nlim→∞k1
nSnψ−ψ∗kp
= lim
n→∞
Z
X
|1
nSnψ(x)−ψ∗(x)|p dµ 1/p
= Z
X lim
n→∞|1
nSnψ(x)−ψ∗(x)|p dµ 1/p
=0.
Convergent sequences are Cauchy, so for everyε >0 there is N =N(ε, ψ) such that
k1
mSmψ− 1
nSnψkp< ε
2 (*)
for all m,n≥N.
Now ifφ∈Lp(µ) is unbounded, we want to show that 1nSnφis a Cauchy sequence ink kp. Letε >0 be arbitrary, and takeψ bounded such thatkφ−ψkp< ε/4. By T -invariance,
k1
nSnφ−1
nSnψkp≤ kφ−ψkp for all n≥1. (**)
k1
mSmψ−1
nSnψkp < ε
2 for all m,n ≥N(ε, ψ) (*) and
k1
nSnφ−1
nSnψkp ≤ kφ−ψkp < ε
4 for all n≥1. (**) By the triangle inequality,
k1
mSmφ−1
nSnφkp ≤ k1
mSmφ− 1
mSmψkp +k1
mSmψ− 1 nSnψkp +k1
nSnφ−1 nSnψkp
< ε 4 +ε
2 +ε 4 =ε for all m,n≥N(ε, ψ).
Hence 1nSnφis Cauchy in k kp and thus converges to some φ∗ ∈Lp(µ). We have
n+1 n
1
n+1Sn+1φ(x)− 1
nSnφ◦T(x) =|1
nφ(x)|
for all x. Taking the limit n→ ∞ givesφ∗=φ∗◦T µ-a.e.
This concludes the proof of the Lp Ergodic Theorem.