Professional Documents
Culture Documents
Ergodic Theorems
The Birkhoff Ergodic Theorem The von Neumann Mean Ergodic Theorem Applications of Ergodic Theorems
Related Topics
4.1. 4.2.
Flobenius Mixing
Perron operators
Bibliography
Chapter
1.1
o-Algebras
The theory of o-algebras plays a significant role in the probability theory and in the ergodic theory.
is
CHAPTER
1. ELEMENTS
OF MEASURE THEORY
It is an easy exercise to show that any algebra is closed under operations -4 n B and,4A B - (A \ B) U (B \ A)
Definition 1.L.2 An algebra D 'is called a o-algebra ,f UAn D yi,elds for any countable fami,ly {An}n g I.
k
As in the previous exercise, it is easy to show that any oalgebra is closed under countable intersections. A bit more difficult (but stil trivial) exercise is to show that an algebra, which is ciosed under countable unions of pairwise dis.joint
sequences of sets, is a o-algebra.
(X) of
As an example of algebra which is not a o-algebra, one may consider the collection A(T,R) of all intervals (o,,b), (o,bl, lo,b),, lo,b1, (-oo, b) , (o,+oo) in IR and their finite unions.
Exercise 1.1. 1 Show that the fami,Iy of all countable and all co-countable subsets of a non-empty set,is a oalgebra. What would be 'in the case i,f finite subsets ,instead of countable subsets are considered.
Given a non-empty family F of subsets of a non-empty set X, the algebra A(F) generated by F is the intersection of all algebras containing F (at least one such an algebra exists, the algebra P (X)). The definition of the o-algebra A"(F) generated by F is similar.
1.1.
r-ALGEBRAS
1
Exercise
.L.2
- A(F), where
The o-algebra A,(A(T,R)) is called the Borel o-algebra of subsets of lR. More generally, the Borel o-algebra of a topological space (X, r) is the o-algebr a Ao(r) generated by the collection r of all open subsets of X .
It is often a challenging problem to describe the o-algebra A"(F) generated by a family .F in a more constructive way
than just the intersection of all a-algebras containing F.
In the next two subsections we study several important results about o-algebras and recommend the reader to complete all details of the proofs below, or to consult textbooks Iike [1 , 4, 6, I7].
of subsets of X ,is sa,id to be o zr-system ,f i,t 'is closed underintersect,ion of ang two of its sets, and A e K.
Denote by
r(F)
F gP6)
CHAPTER
1.
collect'ion
D g P(X) is called
a,
impl'ies
Denote by
EP6)
system generated by
Theorem 1 If
be a n-system then
A"(K) -D(K).
A,(K) g D(K)
Note that any Dynkin system, which is a a"-system, is a o-algebra. So, it is enough to show that D(K) is
a n'-system.
Let A
family:
I *.
Ute n ail
- An U Bn e D(K)
1.1. r-ALGEBRAS
and then
U Bn e Dn.
k
Then the condition (3) is satisfied. To check the conditio" (2), take B e Da. By
A)),
u(x\ A):s
Exercise
Let A K. Using K g Dt, we obtain D(K) g Da D(K), and hence D(K) - Da for every -c A e K. Consider the foilowing implications:
AeK+DA:D(K)+ li,f A e K and B e D(K) - DA, then A)B e D(K)I =+ lx q Dn e D@) i,f B e D(K)) + DB : D(K) (VB D(K)) =+ A)B eD(K) (VA,B e D(K)).
The last condition means that intersection, what is required. I
8
(optional
)
CHAPTER
1. ELEMENTS
OF MEASURE THEORY
Let us consider now two more results of the same nature which are known as the zr-)-Theorem and the Monotone
Class Theorem.
Definition 1.L.5
collect'ion
L C P(X)
'is called
a ,\-system or \-class
o)XeL;
Denote by
Theorern
U
Proof: Obviously To prove )(f) ) Ao(f), it suffices to show that )(f) is a o-algebra. It is enough to show that )(.F) is closed under
countable intersections. (WhA? Show
Consider
.\(.F) g A"(F)
^(.F).
thisl
f).
F.
Hence
Consider now
^rV)-{A.\(r)
:AnBe^(F)
VBe )(-r)}
1.1. r-ALGEBRAS
Then
is (again as above) a ,\-system containi ng F ^r(F) (Hint: start w'ith x(F) instead of f and remark that
Hence
^r(f) The last equality means that ,\(.F) is ^(F). closed under countable intersections, what is required). n
A fami,lA
Fe
g ,\(4
and there-
i,)x#a;
i,i,)
M, An I A + A e M; iii) {Br}r I M, Bn I B + B e M.
{An}r g
Denote
F CP6)
by M(F) the
Theorem
If A
Proof: Obviously M(A) S A"(A). To show the converse A"(A) g M(A), it is enough to prove that M(A) is an
algebra.
For each G
g X,
define
10
CHAPTER
1.
Ag
Ms.
which is equivalent to
E e Mc (WhA? Show thi,sl. So, for any C e M(A)), M(/)) q Ms (since A g Ms and Ms is a monotone class). The condition
M(A) e Mc e M(A) (vc M(/)) means that for every B, C e M(A): B\C, BuC, C\B M(A). Together with the fact that X M(A)., this implies that M(A) is an algebra. The proof is complete. n
1.2. MEASURES
11
L.2
,f
Measures
A-
nR
- U{+*}
o) p(a) - 0; b) p(A))0forallAA;
c) t-r (?
r-) :
probabilistic (p(X) - 1), finite (p(X)_. oo), or o-finite (={&}8, q "4 such that X - [J X" and
lr(X,)
space.
Except of counting measures on countable sets, constructing a measure on a o-algebra could be a difficult task. We shall not go deeply in details and only mention key steps of construction of the Lebesgue measure on IR.
oo for each
n).
n':1
called a measure
L.z.L
Lebesgue measure on
IR
First of all, we need a large enough a-algebra (containing the Borel algebra 6(R)) otr which our measure will live.
12
CHAPTER
1.
Secondly, this measure should agree with a certain function on a certain subset of 6(R) (for example, with the length on the collection J g 6(R) of all intervals).
For this purpose, w extend the length I . J -+ R+ IR+ U {+*} to a function \ . P(R) * R1, and then take the restriction of ) on a certain o-algebra D ) B(R), otr which ,\ acts as a measure. Some of details follows.
ren): 11, e r k Asp.r} - inr {: I - {A, )(E) : \(E.A)+)(tr.A") for all E e 2(R)},
)(A)
R \ A.
algebra containing 6(R) and )(1) l(I) for every Moreover, D is complete w.r.t. (i.e. )(B) 0 =+ B and ,\lr is a measure.
is a o-
I J. e I)
Remark that the Caratheodory theorem holds true if we start with any countabiy additive nonnegative function on an algebra of sets (not necessary with the length of intervals).
Definition 1.2.3 The o-algebra D constructed aboue called a Lebesgue algebra, and the n'Leasure )1"
called o Lebesgue measure.
,is ,is
It
is worth to note that the Caratheodory theorem is rather deep result and that the ideas that had led to the Lebesgue go back to the end of 19th cenmeasure space (R, I(lR),
tury.
^)
1.2. MEASURES
13
Remark that the Lebesgue algebra D(R) is strictly bigger than the Borel algebra 6(lR). It can be shown (for example, by using of Monotone Class Theorem) that the cardinality of the Borel algebra Card(6(R)) is c - 2N. On the other hand, the ternary Cantor set C is Borel measurable, Card(C) : c, and )(C) : 0. Therefore by the completeness of D w.r.t. ), all subsets of C belong to I and thus
Card(D) ) Cu'd(P(C))
:2')
Card(B(R)).
Although Card(I)
proper subalgebra of 2(R). By using the Axiom of Choice, it is rather easy to construct an example of a subset of IR. which is not Lebesgue measurable.
is
0) and a bijection Xo I ,'rsatisfying 1) A A' 1"6, <+ 0-'@) Alxo, z) p(d-'(A)) t''(A)YA e Alro
isomorphism.
Example 1 .2.L Let F e [0, 1] be a closed set such that )(F) > 0. Defi,ne a n'Leasure p on Dlp by
p(A)
- I(,4) l^@)
(A e
tlr).
14
CHAPTER
1. ELEMENTS
OF MEASURE THEORY
Defineo,rnapQ:F---+
[0,
1] by
_ )(F n [0, r)) d@) XF) It i,s not ltard to uerify that (f, f le,lt) 'is 'isomorpltic to
(10, 1],
Illo,tl,
))
bg the 'isomorphi,sm Q.
nonatomic complete separable measure space is isomorphic to (1, Ilr, )), where 1 g R is an interval (bounded or unbounded).
More interesting fact is that any
L.2.3
be a
measure space
(X,A, p)
1) (X, A, p) 'is o -fini,te; 2) X - D I I'is a d'isjoi,nt un'ion of a countable set D and an'interual / q R such that (L,Alr, p) : (I ,t, )).
The component (D, Alo,p) is called a canonical atomic LMS. The component (I ,Alr,F) (which is nothing else than the interval -I equipped with the Lebesgue measure .\) is called a canonical nonatomic tMS.
Definition 1 .2.5 A complete measure space (X, A, p) is called a Lebesgue space if it 'is 'isomorph'ic mod 0 to a
canon'ical LMS. Remark that in the ergodic theory, Lebesgue spaces are of main interest among all measure spaces.
1.2. MEASURES
15
semirings (optionar)
Definition 1.2.6 A
disjoi,nt sets
Et,...,En R
such that A
\ B - lJ Er.
k:l
As an example of semiring, one may consider the collection J(R) of all intervals in IR. The following proposition is almost obvious.
be a sem'iring.
If A - U
n:I
An,
R, then A - i)
k:L
,r, where Cn e R.
by 81
Proof:
n
Define a sequence
only thing that is left to write each B, asairi#":i union of sets belonging to R. We show this by induction. Obviously,
m,p
{8"}"
f,.
The
81 R.
p:
B, - l)
nn
R, for
same property.
Bn+r:An+r\U Aj:An+r
j:I
\U Bj:
j:1
16
CHAPTER
1.
nnpj
"
Etk
all p1
by
lJ t E/,rn tr'*rn...
/}}
By the induction, all Bn have such representation, which completes the proof. n
Definition 1.2.7 Let (X,A, p) be a measure space. A sem'iring C g A of sets o/ finite measure is sa,id to be a suffi.cient semiring for (X,A,p,) xf for euery A e A:
,r
ci. c}
)
Let us mention the following useful approximation property which follows directly from Definition I.2.7 and Proposi-
tion
7.2.7.
p)
1.2, MEASURES
17
C. Let A e A, p(A)
erists a fin'ite disjoi,nt un'ion
such tlt at
and let
Ao:
lJ Cp of sets Cn k:I
p,(AAAo) < r.
Among others applications of the concept of the sufficient semiring, w mention the following useful approximation
lemma.
be a nleo,sure space w'ith a sufficient sem'iring C. Then, for any A wi,th p(A) oo, there erists a sequence Hn in A such that p(H") oo, H" I H A, p(H \ A) 0, and each Hn 'is a countable di,sjoi,nt union of elements of the sem'iring C.
( (
Proof:
exists
0, there
H(r)- 3t,,Cj C,such thatAq HG)and j:7 p(H(e) \ A) ( e. Write Hn: H(1) ) H(+) n . . . ) H(*)
Since C is a n-system, it is easy to verify that each Hn ts a countable union of elements of C. By the Proposition I.2.L, Hn is a disjoint countable union of elements of C. Furthermore, Hn H, which obviously has the required proper-
ties. n
18
CHAPTER
1.
1.3
Lebesgue Integration
an
Definition 1.3.1 A function f : X ---+ IR zs called measurable i,f f-'(I) e A for euery I /(R).
Note that the collection M(X, IR) of all measurable realvalued functions on X is ciosed under all point-wise algebraic
and lattice operations (like addition, multiplication, taking the positive part, etc.). Moreover, M(X, R) is closed under the point-wise convergence.
be a
f (")
-t i,:l
\tn,qu
(); e R).
19
is defi,ned
by
I f @) t'@r) '- !
\p(At).
:_1
L-T
It is also reasonable to have an integration that allows passing to uniform limits of integrands on sets of finite measure. So the second step of our construction is following
si,mple funct,ions
uni,formly. Then
ff
XX
where
tol u@,),-;,*
r.@) p,(d,r)
20
CHAPTER
1.
Now we remote the condition p(X) oo, by using an idea similar to that one in the last definition.
f e M(X,R). Then
r@) p(d*),p(A).
{x"l
where
p@r):-
sup
{l
*}
I r wl rr@*) ,: I AX
agreement:
na,(*)
f (r)
p,(d,r).
e M(X, lR).
Tlr,en
\ @) p,(d,r)
- [ r@) p(dr),
i,f at least one of integrals 'in the ri,ght-hand si,de i,s fini,te.
Remark that the last technical agreement, which allows us to deal with integrals which are not necessarily finite, cannot be used for complex-valued functions (or, more generally, for functions with values in IR" for n > 2).
1.3. LEBESGUE
INTEGRATIOAi
21
instead ot
I f @) p(d") X
L,(p)
- Lt(p)t
{r
ll
. L,(p) ,
ll
I
-
tftdr,: ,}
R defi,ned by
w.r.t.
il
r : Lt(p)
t/t il,
'- I
X
,,t du
The Lebesgue integral has some important properties that we shall often use. We state them without proofs.
J XX
If
rf dp-lim
NJI
f"dp.
CHAPTER
1. ELEMENTS
OF MEASURE THEORY
We also refer to the monotone convergence theorem as to MCT. Among other things, it allows the integration of pointwise convergent series with non-negative terms:
XX
Remark that the inequality in the Fatou lemma may be strict. For example, if (X, A, p) - ((0, 1], t, )) and fr: , . [(0,*1 then liminf f n - 0, but [- f" dp - 1 for all n. The situation will be changed drarrro{ti.ulty if we suppose a sequence to be bounded by a Lebesgue integrable function.
Theorem 6 (Dominated Convergence Theorem) Let {f"}" q M(X,IR), g e Lt(p), and lf"l a s a.e. for alln. If f"(r) -- f (") a.e. then f e Lr(p) and
n_-*
Iim
r*
I lf"- fldp- o.
1.3. LEBESGUE
INTEGRATIO.^{
23
,({ r,,)
for any f e Lt(p). n
{gff)d,,
(optional)
Let (Xr, At, p) and (X2,, Az, pz) be measure spaces. For
simplicity, we suppose measures pr , Fz to be finite. Denote by R the following set
R'.- {At x Az: At e Ar, Az e Ar) of all measurable rectangles in X1 x Xz. Note that R
is a z-system. Denote by O the algebra generated by R.
Define the function pr
and extend it to O. This is possible because each elements of Q is a disjoint union of finitely many elements of R. Now we show that the function h & Fz is countably additive on O. For this purpose it is enough to show that h E Fz ts countably additive on measurable rectangles.
24
such that
CHAPTER
1.
D0 ,6Define a functian
- l) o".
n:I
fn: At
---+
R by
Dr"@):t'r@8)
n:7
t I fndt, :
":, i?
Thus Fr I Fz ts countably additive on the algebra O. An application of the Caratheodory theorem gives the measure ,\ on the o-algebra o(Ar* Az) which extends ht 1t2. The measure .\ will be denoted by the same symbol W g pz.
IE-'")
Definition 1.3.7
and (Xr,, Az,, pz).
The n'Leasure space (X, x X2, o(A1 x Az), h g p,2) ,is called the product of the n'Leasure spaces (Xr, At, p)
1.3. LEBESGUE
INTEGRAT/OAI
25
Theorem 8 (The F\rbini Theorem) Let f (rr, rz) be a real-ualued Lebesgue ,integrable funct'ion on X1 x Xz. Then the second and the thi,rd iterated
'integrals below make sense and enjoy the formula:
I",*r,f
(*',rz)dtttu
L,Ur,f
ou'' n
measurable
Theorem 9 (The Toneli Theorem) Let f (*r, rz) be o, real-ualued non-negat,iue funct'ion on X1 x X2. Then the funct'ions
At-
I*,,r,f
lr,(lr,f
L'U"f
Example 1.3.1
(*''rz)dtt)
o"'
The non-negatiui,ty cond,ition'in tlr,e Tonel,i theorem 'is essent'ial. Indeed, consider the measure space
CHAPTER
1. ELEMEN"S
OF MEASURE THEORY
i,s
1.3.5 The Radon - Nikodym theorem and the Riesz representation theorem
In this subsection we formulate two very useful classical theorems of the theory of the Lebesgue integration. The first theorem deals with a representation of measures through functions and the second one with a representation of positive linear functionals through measures.
u be a p"-
cont'inuous n'Leasure. Then there erists a non-negatiue functi,on f e L,,(lt) such that
,(A):lfdt, (vAeA). n
A
Let X be a locaily compact metric space. For simplicity, one may think that X g lR'. BV Co(X) we denote the
linear space of all continuous functions infinity.
X -* IR vanishing on
Definition 1.3.8 A mappins d: Co(X) * IR zs called a) a linear functional i,f d@ft + 0 fz) : ad(f t) + gdjz) for all a,0 R, ft, fz e Co(X); b) a positive functional i/ f (") > 0 (Vr e x) =+ d(/) > 0 for all f e Co(X).
Theorem 11 (The Riesz Representation Theorem)
For eaery pos'it'iue l'inear funct'ional Q : Cg(X) ---+ IR, there is a unique fini,te measure LLO on the Borel algebra B(X)
such th,at
0(h)
h(")
pro@r)
(Vh e
Co(x)). n
Remark that the measwe p6 in the Riesz representation theorem is obviously finite. Moreover, it is clear that every finite measure LL on B(X), gives a rise to a positive linear func* tional d, on Co(X) by
or@)
28
CHAPTER
1.
Chapter
An abstract semi-dynamical system with discrete time is apair (X,r), where X is aset and r . X ---+ X is a mapping. We think of X as the set of states of (X, r) and of r as a law which controls the way how the system (X, r) evolves through time. In the case of an invertible r, the pair (X, r) is called an abstract dynamical system. If we consider a one-parameter semigroup (rr)r>o (oneparameter group ("r)r.o) of transformations of X, the pair (X, (rt)t) is called an abstract continuous semidynamical (dynamical) system with the state-set X. In this lecture, we study several elementary examples of transformations of measure spaces. We require them to
measurable (not necessarily invertible) transformations of a measure space (X,D, p). The study of such dynamical (semi-dynamical) systems is the subject of the ergodic
be
CHAPTER
2.
Remark that, in many cases, the state-set possesses a typological structure. This is a very reach in applications area of theory of dynamical systems, which is called topological dynamics. Topological dynamics has been closely related with ergodic theory. The both theories go back to works of Henri Poincare in the end of 19-th century.
theory.
z.L.L
0 implies B e "4), otherwise one may take the completion of o-algebr a A. A measurable mapping r : X ---+ X is called nonsingular If p,(A) - 0 implies p(r-t(A)) - 0. Nonsingular mappings play a key role in the concept of a Perro Frobenius operator, which will be considered later. We do not need this concept in the present lecture.
and p,(A)
Let (X, A,, p) be a measure space. A mapping r : X ---+ X is called measurable if r-1(A) e A for every A A. We usually suppose A to be p-complete (i.e. B g A e A
Definition 2.f.L A
called endomorphis
r: X
--+
X ,is
p(r-'(A))
- p(A)
(vA e A).
doubling transformation below). The composition of two endomorphisms is obviously an endomorphism. Our main aim in this lecture is to give several classical examples of endomorphisms.
2.1. EXAMPLES OF
"R/NSFORMATIOI{S
31
X:
[0, 1]
S:X+Xby
s(,,r)
X.
We define a transformation
{ l\:'iq,L,y+})
i2:l l,3l
A<1; y <7.
To understand the operation of this transformation better, examine the following figure and take into the consideration the way by which the traditional baker is kneading dough.
-+'lrrrniffi-The mapping ,S : X
---+
ol-
is called a
baker transforma-
tion.
Obviously, ,S is measure preserving and invertible. Sometimes, the following 3-dimensional version of the baker transformation:
r(r,A, z) - (S(r,A),
by the figure beiow.
z)
'l-lEJrr s f tttr
---T---;
r-L!-,r a
,l
32
CHAPTER
2.
2.t.3
Rotation transformation
For any a e IR, define the rotation by a to be the transformation Ro : [0, 1) -* 10, 1)
given by
R*(*)-(r+*)(mod1).
Clearly, ,R* is invertible and measure preserving. It is also clear that B* is periodic rff a a Here we consider the Lebesgue measure on 10, 1). If a measurable transformation on (X, A) is given, then a measure p is called invariant measure for r lf r is an endomorphism of (X, A, p). Thus the Lebesgue measure ) is an invariant measure for ,Ro.
format'ion of (X, A, p). Show that ris an endomorph'ism iff ,-' is an endomorphism.
We usually identify [0, 1) with the unit circle
by
It
Theorern L2 (Kronecker) Letr 10, 1). The sequence @T,@))Po 'is dense in 10,1) iff a R \ a
for simplicity Ro by R. AII the points Eo(") : n, Rt(*), ...) R(*)) ...are obviously distinct. By the Bolzano Weierstrass theorem, there exists a convergent sequence in (R(*))Po. Thus for a given j > r > 0, there exist nonnegative integers p
Proof: It is enough to show that the set {RX(r) > 0} ' ", is dense when a is irrational. Fix a R \ Q and denote
- p-q and d: lR'(*)-rl. Then the consecutive terms in the sequenc" (Rt'(r))pn are d-apart of each other. In fact, for I ) 0, lg(t+t),@)-R,(*)l _ lRt,(R,r)-R,(r)l _ lR,*_rl_
Hence
di.
Although the statement of the last theorem is pretty simple, it will lead us to several important concepts. The first one is the notion of a minimal map r . X ---+ X on a metric space X. The minimality of T means just the density of the positive orbit {r"*}70 for any r X. The Kronecker theorem states exactly that all irrationai rotations -R* are minimal.
Among others applications of the rotation transformation Er, is the construction of a subset of 10, 1) that is not Lebesgue measurable. Fix an irrational a. Then the full orbit l, : @2")*?-* consists of distinct points. Moreover, for every frr, fr2, the orbits lrr,,lrrare either disjoint or coincide. By using the axiom of choice, take in any family of all orbits, which are equal to the same set l, a point r f. Denote
34
CHAPTER
2.
A. Then
- rL, r". [J
By the construction of A,
lo, 1) Assume
- nZ R3(^) ll
0 then
1:)([0,
If)(A) >0then
1))
-I)(A)
N?L
-0.
:oo.
A
is not
1:)flo,
1))
-I)(A)
nZ
r(r)-2r(m,od1)
Iterations of
-['*' Izr-r
if o.a"tit if *5"<i.
35
;\ffiVl/+"W,-,
An easy investigation of the graph of r on the figure above gives us that r is a non-invertible endomorphism of [0, 1). Let D consist of all reals in X : 10, 1) of the form fi, and let Xs - X \ ,. The numbers in Xs have a unique representation in binary form as
[l'
rl 17
oo
(ou
e {0, 1}).
r(r)-IT
i:7
note that the set
@exo)
This representation of r shows among their things the density in [0, 1) of the set of periodic points of r. To see this,
i - i,:I #,
36
CHAPTER
2.
[0,
i,:L
Y+
2i'
i,:k*1
Y'2k+t -I "' r ?n ./
I
2k(
n+I)k i,:nk
,nW-r
I
a'i
gives us another dense in [0, 1) set. Obviously, which justify the density of r-periodic points.
rkn -
tr,
Remark that the doubling map can be constructed on the unit circle instead of the interval 10, 1). This construction makes the rotation transformation continuous.
2.2
space
Recurrence
endomorph'ism
Definition 2.2.I An
(X,A, p) is called recurrent i,f for any A e A, almost eaerA poi,nt of A (i.". all po'ints of A\ltr, /r(lf) - 0)
return to A at some future time.
r of a measure
(1)
n(r) above as a return time to A. Remark that if the measure pl is infinite, then it may easily
happened that an endomorphism is not recurrent. For example, the shift transformation ,(*) r 1 on IR. Later
2.2.
RECURRENCE
37
on in Theorem 13 we shall see that this may not happened when p is finite.
r be a recurrent endomorplt,ism of (X, A,, p) and A e A. Show that there erists a null set lf1 such that for any r A\ m tltere'is an,increas'ing
sequence (depending on
wi,th
r*itrA
(Ulntt
(Jse
(VieN).
ULot-*(l[),
and, note that,
2.2.L
Conservative endomorphisms
The following observation is an immediate consequence of Definition 2.2.1: an endomorphism r of (X,A, p) is recurrent iff
'(^\C
rence.
'-r@)) -0
(VAeA).
(2)
Lemma 2.2.L An endomorph'ism r on (X,A, p) ,is recurrent iff r is conservative (i,.e. for anA g, p(B) > 0, there erists n ) 0 such that p(B ) r-n(B)) > 0).
CHAPTER
2.
Proof:
Suppose
p@)
0. Thus
r is conservative.
A.
Denote by
Let
B the
set A
then there is n
contradicts the definition of B, since no point of B belongs to ,-"(B) C r-n(A) Thus p(B) 0 and hence r is recurrent.
tr
, :X
---+
pt) be
'i,s
an endomor-
is con-
2.2. RECURRENCE
39
p(B n r-"(B))
k>m)0,
0 for aII n
F(r-u'-^)(F)nB) _0.
pI
/x \
\n:o
U,-"(u) I /
oo
OO,
a contradiction, as p(X)
i,f
r-I(C) g
incompressible if there is no
\ '-t(C))
cornpressi,ble set
for r.
Theorern L4 Let r
40
CHAPTER
2.
r to be recurrent and let ,-t (C) g C for some C A. Put B - C \ The sequence "-'(C).
Proof:
Suppose
(r-"(B))Po
t-/(c
>
C
In particular, ,-"(B) n B - g for all n 2.2.I, p(B) - 0. So r is compressible. Now let
1. By Lemma
Put
c\'-'(c)-A\Ur-*(A).
k:I
oo
oo
o, which means
4l
2.3
Ergodic Transformations
Ergodicity is one of most important concepts in our course. It was introduced by George Birkhoff and Paul Smith in
T928.
Definition 2.3.1
Let
of a rneasure space (X, A, p). Then ris called ergodi c i,f ,is euery r-'inuariant set A e A (i.e. p(r-'(A)LA) - 0) tri,u'ial i,n the sense that e'itlter p(A) - 0 or p(X \A) - 0.
be a nons'ingular transformat'ion
ergodic endomorphisms
on
f . X - R, f o, = f (p-u..) + f : const (t -u..). Proof: ((-;)' Let r be ergodic, and let / . X -+ IR be a measurable function satisfying f r = f (lt-".".). Assume " that / is not a constant (p-*..). Then, for some a IR, the sets A - {f < a} and B - {f
measurable funct'ion measure. These sets are obviously r-invariant. For example, p-Le'
Theorem 15 Let r be a nons'ingular transformation of a measure space (X,A, p). Tlr,en r is ergodic i,ff, for euerA
,-'(A)
- {r: rtr A} f-
{f or < a}
{f < u}:
A.
r. Hence
const (p-u.. ).
42 ((
CHAPTER
2.
z )) Assume that r is not ergodic. Take a nontrivial set A e A that is r-invariant. Then the indicator function Ila of A is not a constant (p-a.e.). However, Ila o r :n,-rpa1q-Lt'
[o.
if the
Exercise 2.3.L Show that the i,dentity transformat,ion on a canon'ical Lebesgue space X 'is ergod'ic i,ff X ,is &
s'ingleton.
less trivial but stiil rather simple situation arises case of baker endomorphisms.
in the
Exercise 2.3.2 Show that the 3-d'imens'ional baker endomorph'ism 'is not ergodic. Show that 2-d'imens'ional baker endomorpltism ,is ergod'ic.
Sometimes the question on ergodicity of a given transformation is fairly simple. Here we mention only that the irrational rotation and the doubling map are both ergodic. However, in certain important cases) the question on ergodicity may be very deep and difficult.
43
p).
are equ'iualent: (1) 'is recurrent and ergodic; (2) " euery A e A, p(A) > 0, For
loo\
en'ists
n)
such that
,-"(A)nBla;
erists n
(4)
such that
B -- U ,-" (A)n:1
oo
is recurrent,
p l r-'(A)
/oo\
\n:r/
0.
Moreover
ergodicity
r-"(A)
(2) =+ (3)
44
CHAPTER
2.
by (2),
oo
n:7
,-"(l)
p-Le'
B
(3)
n:I
U ("-"(A) n B) _ s.
(4)
t'?-"(A)
B)oo
Let p(A)
0 for all n
(3).
C-
(4) =+ (1). The recurrence of r follows, by Lemma 2.2.I, if we take B Ain (a). Assume that r is not mean ergodic. Then there exists a r-invariant (lr-u...) set A such that p,(A) 0, p,(X \ A) > 0. By (+) applied to B X \ A, there exists
n)0suchthat
0
a contradiction. Therefore
Let
r be an endomorphism
rby
S
f (")
t'-Le'
f (r("))
(f
e Lr(p))-
to
)). Obviousiy, ) :
r.
1.
Proof:
: I gdp by the linear approximation, I g "rdp for any e for all S e Lz(p). Consequently, we obtain that f Lr(p)
satisfying
S
Remark that [a
r:
A.
Hence,
l,rf
which
Iv?ordLt: Ivo,r'dt,d,t
impliesl)l
,^rrrl'
- r^ r I
^
vr'd,r,,
-tn
'is e'igenualue, then the
E())
'is
-{f
e Lz(p)
,f o,-^f
(p-u")}
Exercise 2.3.4 Let f1, fz b" ei,genfunct'ions corresponding to e'igenualues )4 + ),2. Show that f1 and f2 a,re
OfthOgOnAl. (nmt, (Jse the property
that
sor
d,pr,
se
rr|t).)
46
CHAPTER
2.
coristant (p-a.e.); (2) T 'is ergodic i,ff, for each e'igenualue ),, the subspace jt-u.".)} ,r l-d,imens,ional. Es - {f Lz(p) , f o, -
l/l :
r : \f
^f
l)l -
The ergodicity of r impiies that l/l - constant (p-u.. ) Moreover, f + 0 (p-a.e.) since / is an eigenfunction, therefore we may choose an eigenfunction u such that is ergodic and fi o r - \ft jt-u." ) From part (1), /r l0 (p,-a.e.). Let f2be another eigenvector corresponding to ,\. Then the function g : fzlf, e Lz(p).
Since
l/l "r-lf"rl
-l)/l -l/l
lrl -1(p-ae)
#,
90T-
then
(rf.
g -
fzor fto,
p-d.e.
fz
constant
that dim(81)
fr: C'f,
-
-- 9,,
1.
Suppose that dim(E1) 1 for every eigenvalue Consider the eigenvalue 1. Thus the only functions satisfying are constant. By Theorem 15, is f (r(")) f (r) }t-u
).
ergodic.
".)
By the previous lemma, we may assume that eigenfunctions of an ergodic r have absolute value 1.
47
Theorern LT Let r
o/S1
an ergodic endomorphism of a Lebesgue probability space (X, A,, p). Then the set oo?) of eigenualues of r is a countable multi,pli,cat'iue subgroup
be
-{lC,l)l -1}
Proof: If )r, ,\z or(r) then fi o r : \th, fz o , - \zfz for some nonzero h, fz Lr(p). We may assume that l/tl - lfrl:1 (p-a.e.). Then (fr- fr) " r - (/t " r) . (ft o r) - .\r . ),2. (ft. f) 0t-".".). Thus f, . f, Lz(p) is an eigenfunction corresponding to
,trr ' )2.
Also,
ft
and
r-
p-d.e.
ho,
,\r ft
e Lz(p). This shows that oo?) is a multiplicative semigroup. Notice that the eigenfunctions corresponding to different eigenval-
t.
Jx
hLfz
48
CHAPTER
2.
Chapter
Ergodic Theorems
3.1
The Birkhoff Ergodic Theorem
The aim of this section is to prove celebrated Birkhoff's individual (: point-wise) ergodic theorem (it will be also referred to as the BlBtheorem) which was obtained by George David Birkhoff in 1931. Nowadays it has enormous amount of applications in different areas of mathematics (probability theory, dynamical systems, analysis, geometry, etc.). We formulate also two famous generahzattons of the BlE-theorem. For more advanced individual ergodic theorems we refer to Krengel 15] To avoid unnecessary compiications, we consider an endomorphism r : O ---) O of a o-finite measure space (Q, A,/r) We shall use the following notations:
: D Tuf,
n-7
i,:0
i,:0
49
50
CHAPTER
3.
ERGODIC THEOREMS
Mf
, Azf - supf,, Anf , . . ., Anf), M*f where is any real-valued measurable function on O.
,
M"f
f-
Szf , . . ., Snf),
3.1.1 Birkhoff's individual ergodic theorem Theorem 18 (Birkhoff's ergodic theorem) Let r be an endomorph'ism of (Q,A, p,) and f e Lt Lt(Q,A, p). u-4",e' Then Arf t L1 fo, a" r -,inuariant t
(i.". Tf
for
euery r-'inuariant
'^l A,
Ae
p(A) <
oo.
The proof of the theorem, which will be given below, is based on the following maximal ergodic theorem, that is due to Shizuo Kakutani and Kosaku Yosida (1939) for T generated by endomorphisms, and that is due to Eberhard Hopf (1954) for arbitrary positive contraction 7 in L1.
Yosida
Hopf)
Let T
En
The proof of the Kakutani - Yosida - Hopf theorem which is given below is due to Adriano Garsia (1965).
THEOREM
51
Proof:
Yk: r,...,fr (M:f)* 2 Snf and hence f +r((M:f)*) > f +Tskf : sr,+tf
.
Therefore obtain
> Snf -f (Mf,/)*) for all k - 1,... ,ntI, because it is obvious for k - 1. Passing to suprema? we
f > mf f -r(rwfl)*)
En
rou. I
En
@rf
r _ r(w:ry))dr,-
:
:
En
(Qwf
I cl
f)
@:
f)*d'1t'
>
. lw:fl*d'pt- | ,uu:f)*)dp> o.
Remark that the last inequality is true, because
llrll < 1 I
T > 0 and
Lemma 3.1.1 (The maximal ergodic inequality) Let r be an endomorph'ism of (Q, A, p), then ll/llr 2 a. p(M"f > a) ) a- p,(M"f > a)
52
CHAPTER
3.
ERGODIC THEOREMS
This lemma was already known in 1930-th and belongs to mathematical folklore related to the low of large numbers.
Proof:
/ > 0. Because of
o
An+tf
the functions
- (n+
1)
-t f +
#(Anf)
f' :
lim inf
r,
A,f
ft : f"
f"
l.r-a.e.
{f" - +*}
oo
oo
satisfies
(vr>o).
(1)
THEOREM
53
t)
-
(Vn e N,7
>
0).
ItU"
Hence
- +*)
n-+@
< p,(M*f
p-a.e.
j) :
{'ll/ll,
>
o).
f"<x
that for some a B - {ft < CI} n {B < f"). By the r invariance of B, the function f ' : (f - 0) .ila has the property: P-le' o (Vk o)' (f' lls-,\r
Suppose
"'\'
>
Hence
lr'ou> I r'd,p- | B
()
r'dl,>o
,>o}
r{*,t>o}
?)
r{*,f
l ror>p.t@).
:
(a
f I f dp I a. p,(B). J
- f) .fie gives
CHAPTER
3.
ERGODIC THEOREMS
a<
B.
Lr(Q,A, p).
To prove the last assertion of the theorem? we may assume that A - O, lr(O) ( oo, and / > 0. For any e ) 0, take
M, such that
4,.
a,r JI r,
M,)*dl.t <
uu - Mu)+dlt |
C'
JU-M,)*dp'ae. o
Thus the sequence (A"f)" is uniformly integrable and hence it converges in ll . ll1-norrrr as it converges p-a.e. to f In other words,
I f orJnn
(-)
ll
t,p
A,fll, - Lp llA,fll,
- ll/ll, : I f dp n
c)
Probably the first one is the Hopf - Dunford - Schwartz theorem which goes back to Hopf (1954) for 7,7(llo) : [o, and to Dunford - Schwarz (1956) for an arbitrary T. Instead of (always positive, contractive w.r.t. L1 - norm, and satisfying ?(llsl) - Ilcr) operator T in Lr : Lr(Q,A, p) generated by an endomorphism, we consider now an arbi* trary positive contraction in L1 which is also a contraction w.r.t. ,L--norm. Such operators are referred to as
LrL*-contractions.
7 to be an LyL--contraction
is not very restrictive, although contraction for all p,7 < p ( oo.
it forces 7 to be an Le-
Another classical generalization of the BlE-theorem is due to Rafael Chacon and Donald Ornstein (1960). It deals with pl-a.e.-convergence of the ratio S"(T)f lS"g)g for an arbitrary positive contraction 7 rn Ly Recall that we denote by n-I S"g)/ the sum D fn
i--0
56
CHAPTER
3.
ERGODIC THEOREMS
contraction 'in L1. Then for any f Lt, g LI , the ratio S"g)f lS"g)S conaerges p"-a.e. on the set
bea
to a fini,te li,mit. n Note that the BIE theorem becomes an immediate consequence of the Chacon
3.2
The mean ergodic theorem of von Neumann deals with norm convergence of averages instead of a.e.-convergence. Although the Birkhoff's paper 12] appeared a bit earlier than von Neumann's paper lB] (both papers are published in 1931-32), the von Neumann's theorem was proved first and was known to Birkhoff. Nowadays the von Neumann mean ergodic theorem is known as a rather general resuit about norm convergence of averages of operators in Banach spaces. In this lecture, we shall prove this theorem for the case of contractions in a Hilbert space, ffi John von Neumann did. Then we mention only some classical generalizations of this theorem and refer the reader to {5] for more information on this.
3.2.
5t
bitrary Hilbert
spaces.
Theorern 22 (von Neumannts mean ergodic theorem) Let T be a contract'ion 'in a Hi,lbert spo,ce "11, and P the project'ion onto the fired space Fix(?) e'11 : Tg : g}, then A"g)f conuerges ,in norm to {S Pf for all f e '11.
Lr(Q, A, p), /r(O) - 1, and T is generated by an endomorphism of (Q, A,p), the l\ME-theorem follows directly from the BIEtheorem by using the Lebesgue dominated convergence theorem and the density of L*(Q, A,/r) in Lr(Q,A, p). Our aim now is to prove the general version of the NMEtheorem. Its proof is much simpler than the proof of the BlE-theorem. In order to give this proof (which follows to [5, p.a]), *. shall use the following lemmas.
Remark
f -rf + ll/ll' - (f ,rf) + fll'- (r.f ,f) + llr. f - fll' : llr. fll' + ll/ll' - 2(r. f, f) : llT. fll' - ll/ll' ( o =+ T. f - f
ll
.
T.f-f+f-Tf
58
CHAPTER
3.
ERGODIC THEOREMS
be a family of contractions in a Hi,lbert space 71, then the orthogonal complement of the e Tt:Tf Yf e U}'is equal fired spa,ceFix(U) to the closure o/ span{ h Th : h e H,T e U}.
- {f
-f
span{h
Th : h e'11,T eU}
(T.f
to Lemma 3.2.L
Then
llA"g)(h
n-tllh
- rh)ll -
by the approximation
llA"(r)f
for all
llA"Q)g
for all 9 Fix(T) By Lemma3.2.2 applied toU {7}, we Fix(T) O cl(..nf). Hence llA"Q)u obtain Jl Pull 0 for all u e 7{, what is required. n
- ell -
59
Kakutani (1938).
T be a l'inear bounded operatorin a Banach space X supposeT to be cesd,ro bounded (i,.e. rg llA"g)ll < *). For ana r X sati,sfyi,rg hylln-If"rll - 0 and ana a e X , the follow'ing assertions are equ'iualent:
Let
.
(t) A Fix(?) and A e co {r,Tr,T'*,. . .h (ti) y - lim A"(T)r; n (iii,) y is a weak cluster po'int of (A"Q)")T:r. n
Corollary 3.2.I If T 'is a power bounded linear operator (i.".sup llT"ll < *) 'in a refl,eriue Banach space (for ern)0 ample, all spaces Lo are refl,eriue when 1 < p < oo) then T i,s mean ergodic (i,.e. the auerages A"(T)n conaerge in norm for all r X). n Corollary 3.2.2 AnA LrL*-contraction on Ly ouer
probabilisti,c sp&ce is mean ergodic. n
a
one.
60
CHAPTER
3.
ERGODIC THEOREMS
Corollary 3.2.3 AnA linear bounded operator on L1 such that the set {A"g)r}n2t is almost order bounded (i.r.
Vs
Bh
The following very useful criterion of mean ergodicity is due to Robert Sine lI2)
Theorern 24 Let T be a Cesd,ro bounded operator sati,sfyi,ng hmn-rTny - 0 for all r X. Then T ,is n'Lean n ergod'ic i,ff Fix(T) separafes Fix(T.). u
Roughly speaking, the Sine theorem says that 7 is mean ergodic ltr T has at least so many invariant vectors as 7*
has.
Theorems
icity.
Theorern 25 Let r
be an endomorphi,sm of a probabi,li,ty
p).
Then
A,B A,
n---+n
ris
ergod'ic i,ff
for all
hm
1
Tl,
n-7
THEOREMS
61
Proof: Suppose that r is ergodic and \et A, B e -4. Then Ila Lt(ti and, by the Birkhoff theorem,
I ,-7
Jg ;Dno(,r(")) k:0
Hence
p(A) (r,-u.
t n-7
for all n
gives
l*:o
,
e)
J*;:
k
r31
t'('-*(A)
r3:l B):J*;: f
lo-r,A)nBdt'd,p" d,trt
iim I ;:; 1 ; il,-n6y.ns dtt: /IJ1g *-k:0 Aork).ns tfo n k:0 .l .4fr,-n67'nn - ^Lt* *D(raork)'116
n<n-I^an_7
J X
B
A to be r-invariant.
L, obtain
p(A).p(A):
Jgg
I ,-I
62
1 n-7
CHAPTER
3.
ERGODIC THEOREMS
Thus p(A)
Hence
r is ergodic. n
1 or
p(A)
0.
Remark that this theorem can also be proven by using the von Neumann ergodic theorem.
be a continuoustransformation
of.
a compact
metric
space f).
e C({l) the
a,re
*fk:0 f orn,(n>
n-I
1) e C(O)
d)
sequence. But accordingly to the Arzela - Ascoli theorem (A;f )" has even uniformly convergent subsequences.
is equi-continuous. c) Implications b) d) are trivial because the uniform convergence is stronger than the weak one and the weak convergence implies the point-wise convergence. d) c): Let converge point-wise to a function /s. Remark that it is enough to prove that for any positive linear functional m: C(O) IR, m(fo A'"f) 0. Accord-
A[f
'
J51
ingly to the Riesz representation theorem (see Theorem 11), there exists a measure p on the Borel algebra B@) satisfying
m(g):
lgdtr
CI
(vge c(o)).
The sequenc (A;f ),, is obviously bounded. By the Lebesgue " dominated convergence theorem, AT,f) dp 0.
AT,f)
0, what is required.
;* d(f, -
Let O be a compact metric space. A positive operator S in the Banach space C(0) is called a stochastic operator if SII : I[, where I[ - IIe (cf. Definition 4.1.1). For example, the operator Sf - f o, in Theorem 26 is a stochastic operator. By the Riesz representation theorem , the dual of C(O) , M - C(O)*, is the space of signed measures on
64
CHAPTER
3.
ERGODIC THEOREMS
(Sf
,p)
Exercise 3.3.L Show that, for any stochast,ic operator S i,n C (Q, there erists at least one invariant probability measure p M (that is S* p, : tt). (nln ' rake aBanach lirnit tr
and a po'i:nt a.'e
Hence p,
f).
C(A)
by
p(f)
L((S"f (u))n.).
M:
lr:
lt,
lt
0, p(lt)
1.)
Definition 3.3.L A stochast'ic operator S i,nC(O) ,is called uniquely ergo dic i,f S has only one ,inuariant probabi,li,ty n'Leasure.
Theorern 27 AnA uniquely ergod'ic stochast,ic operator S 'in C(O) 'is mean ergod'ic.
Proof: Let0lu
e M, S*u
D - * lc:0 Sk f
e
C @)
.
n-l
'is
conaerges un'iformly
THEOREMS
65
Proof: Denote the norm limit of Aflf for f C(O) bV f Suppose that f - const for all / C(O). Take
,S*-invariant probabilities u1 afid u2, then
(f ,rr)
- (f ,li- (Af ).r') - (J5g Al,f ,r') : (f ,rr) - (f ,rr) : (f ,rr) (V/ e C(O)).
:
Uz.
Hence u\
Assume now that f (rr) f @r) Take the Dirac measures 5,, and 6,r. The sequences (Af).d,, and (AE).6,2 converge in c,,r*- topology on M C* (O) to ,S*-invariant probabilities (1 and (2, which are different because of
: ??--'oo ,(Af,)*d,,) : lim (,af f ,6,r) : lim (f (f ,6,,) - f(r) I f@r) - (f ,6,,): ff,Cr). tr
(/, (r)
n--+oo
66
CHAP'I:ER
3.
ERGODIC THEOREMS
Chapter 4
Related Topics
4.L
4.1.1
FYobenius
Perron Operators
First of all we remind the following concept that appeared already in subsection 3.3.3.
Let F(O) be a uector space of compleraalued funct'ions on a, set O. A I'inear operator T : F(O) -+ lr(O) 'is called o stochastic operatoruf
Definition 4.L.L
i,s
CHAPTER
4. RELATED TOPICS
In the simplest non-trivial case O - {rt,u)2t ...,an} is discrete and finite. Then any linear operator T on C(O) -.C" can be uniquely identified with an n x n matrix (ti)i-T',:j=i This matrix is called stochastic if it represents a stochastic
operator on C'.
Exercise 4.L.L Show that a matri, (t,)ijT'l=? is stochasti,c i,ff all 'its entries t,ii are nonnegat'iue reals sati,sfyi,ng
the property
n
Itur:1 ?
J-L
Exercise 4.L.3 Show that the set of all stochastic operators on F(Q), whereA + 0, is o, non-ernptA conr)er subset of the space r(F(O)) of all linear operators on F(A).
4.1. FROBENIUS
PERROAI OPERATORS
69
Definition 4.1.2 Any positi,ue I'inear operatorT in L{p) such that llT. f llr - ll/llt is called a Markov operator.
Exercise 4.t.4 Sltow that the set of all Markou operators on L1(p,) 'is non-empty, conlefr, and closed ,in the
operator norn'L of L(L1(p))).
A matri" (tu)i\?j=? is called Markov matrix if it represents a Markov operator on the finite discrete measure space
measure on n
{0,
1,...,fr-
Exercise 4.L.6 Show that a matri, (t,.)ijT'l=T ,is Markou iff all its entries t,ii are non-negat'iue reals sati,sfging the
property
n
Irnr-1
for any j :1,2,...,fr.
fr
CHAPTER
4. RELATED TOPICS
4.L.3
Let
Frobenius
Perron operators
transformation of a measure space (X, A, p). We do not impose any restriction on r and on X apart from the common assumption that p is o-finite. Letfe Lt (t). Consider the following function on A
r be a measurable nonsingular
,f(A)
-I
:
fdp
(AeA).
,_r(A)
Clearly, uy rs a signed measure on A. Moreover, uy is frnite and pr-continuous. By the Radon - Nikodim theorem (Theorem 10), there exists a unique p.f Lt(p) such that
I
r
pf
dLL
(VA e A).
Let
on a o -fini,te n'Leasure space (X , A, p) . Then the unique operator P . Lr(p) -- Lt(p) defi,ned by the equat,ion
be a nons,ingular transformat,ion
A
'is
Irrdt,- tfdp JJ
,-r(A)
(vAeA)
correspond-
- Perron operator
Sltow that for euery two nons'ingular transformations 11, 12 the followi,ng Pr2or1 Prz o Pn holds. -
P, is a Markou operator.
4.1.
FROBENIUS
- PERROAI OPERATORS
7I
- Perron operators Pr, and P,, such that *(P"r+ P,r) is not a Frobenius - Perron
Gi,ue an eranxple oJ two Froben'ius
operator.
Although Definition 4.7.3 is highly nonconstructive because of the Radon - Nikodim theorem, in some cases explicit formulas for Py are available.
For example, lf
X:
10,
1]
P,f(,)
: * r,t or: * I
o
[0, 1]
r-1([o,r])
f dt,.
is a
that is a diffeomorphism.
P,f
(*):
The formula makes sense if we restrict P on a subspace of L,,(lt) like ClO, 1] for elements f of which the expression
f (r-t(r))
is well defined.
Remark also that certain multi-dimensional analogues of this formula can be found.
CHAPTER
4. RELATED TOPICS
L,f
fiR wi,th
(t"r) for a
"(")
P"f (r)
(Hin
(rF) + r(*F)
4\tr- r
'
Show that
'-'co,"l)
: [o, tF1,
e.)
l-F,
r]
transformat'ions: a) r : IR+ -* IR1, ,@) - 12; b) : [0, 1] 11, r(r) - sinrnt " IR - 10, c) r: -+ R, r(r) - r mod(1).
Many other interesting information on explicit representations of Frobenius - Perron operators can be found in Lasota - Mackey book [7].
4.L.4 Koopmanoperators
Let r be a nonsingular transformation on (X, A, p).
4.1.
FROBENIUS _ PERROI{
OPERATORS
---+
73
L*(p)
U,f (*)
for
- f (r(*)) (p-u. )
Koopman operator
Notice that the usingof p,-a.e.-option in this definition causes minor problems and can be easily avoided if we define LI, on indicator functions by
ll . ll*-norm.
for all f
- ,U,g)
(1)
Proof:
Both operators Pr, (J, are linear contractions. So it is enough to prove the equation (1) for indicator functions. Let g - IIA and / e Lt(p). Then
(P,f ,ra)
XA
lfr,n.
(rra) dt,
- I ,,, dp -
- I
74
:lr
X
CHAPTER
4. RELATED TOPICS
.
U,(ni dp - (f ,U,na)
4.2 Mixing
Throughout this section, r is supposed to be an endon'Lorph'ism of a probab'ili,ty space X : (X, A, p).
Accordingly to Theorem 25, Lhe ergodicity of r is equivalent to the condition: , n-| l^.p(r-r(A)n p(A)p(B) (VA, B e A). (1) _li* rL u n-+oo k:0
=)
B):
dition:
rL---+6
(z)
the con-
is worth to remark that the concept of mixing generalizes asymptotically, in a certain sense, the basic concept of the probability theory, the independence of events
It
tt(A)B) -
p(A)p(B)
4.2. MIXING
75
means
of reals converges to p(A)p@) in the usual sense. The ergodicity of T means that, for all A, B e A, the sequence (p(r-r(A)nB))Po ts Cesd,ro conuergent Lo p(A) p(B). Another type of convergence of real sequences, the strong Cesd,ro conuergence) motivates the following class of endomorphisms.
(tt(r-r(A) )B))Po
A, the se-
called
weakly
lt ,,
k:0
lr?-r(A) n B) - t@)rru;)l
-o
(3)
for all A, B e A.
It can be shown that there are weakly mixing endomorphisms
which are not mixing and ergodic endomorphisms which are not weakly mixing. Thus we have the following proper implications for endomorphisms of a probability space:
ergodic
=+
recurrent.
For more information on these implications the reader is referred to bookr [3],[7], and [11]
76
CHAPTER
4. RELATED TOPICS
,(A)eAforallAeA
n---+6
Definition 4.2.3 An
Iim p,(rr (l))
such that
-1
0)
(4)
4.2.4 Ergodicity, mixing and exactness in terms of correspondent FYobenius Recall that an element
denoted bv
Perron operators
D(p)
D(p).
be the Froben'ius
Theorem 30 Let P,
corresponding to an endomorphism r. Then (1) PiU) conaerges 'in ll . ll yTlorrrl tu n y for all e D iff 'is eract; (2) Piff) weakly conuerges to ny for all e D iff ,is m'iring;
- Perron operator
Pi(fl conuerges,in ll .llt -norn'L tofiy for all k:0 e D iffr'is ergod'ic. f
(3) lim * D ??-+oo '"
We shall prove only that the strong convergence of Pi ff ) to I[; implies exactness of and that the weak convergence of Piff) to II; implies mixing. For other, more
n-I
Proof:
4.2.
MIXING
77
book l7l.
- Mackey
0.
Let ll Pif
a, by en : llPi f o -
t?"(A))
,"(A)
tnyd"1t"- tPifa-(pifa-rrx)l JJ
,"(A)
,"(A)
t J
,"(A)
t J
pifad,-en:
I
exactness of
fadp-ena
r.
fadp-en-r-en.
1 which means
J* p?"(A)) -
f to / ll r .llx , where
II
1t(r-r(l)n 7?-+oo
lim
,, rim
B)
iim - "-* t
n'---+oo
np d,trt
r-!@)
I - *'* to Piil" dp -
r is mixing. tr
78
CHAPTER
4. RELATED TOPICS
Bibliography
[1] K.B. Athreya, S.N. Lahiri, Measure theory and probabi,li,ty theory, Springer,
656 560.
2006.
[2] G.D. Birkhoff, Proof of the ergodi,ctheorem Proc. Nat. Acad. Sci. USA, 17 (1931), [3] I.P. Cornfeld, S.V. Fomin, Ya.G. Sinai, Ergodi,c theory, Springer-Verlag, [4] A.N. Kolmogorov, S.V. Fomin, Introductory real analysis. Prentice-Hall, 1970. [5] U. Krengel, Ergodic Theorems, Walter de Gruyter,
edi,t'ion, Springer, 2007.
1985. 1982.
[6] L.B. Koralov, Ya.G. Sinai, Theory of probability and random processes. Second [7] A. Lasota, M.C. Mackey, Chaos, fractals, and noi,se. Stochastic aspects of dynami,cs.
S econd
[8] J. von Neumann, Proof of the quasi,-ergodic hypothesis Proc. Nat. Acad. Sci. USA, 18 (1932), 31-38. [9] H.L. Royden, Real analysis, Macmillan,
L988.
[10] W. Rudin, Real and compler analysis. Thi,rd edi,tion, McGraw-HilI, 1987. [11] C.E. Silva, Inui,tation to ergodi,c theory, American Mathematical Society, 2008. [12] R. Sine, ,4 mean ergodic theorem, Proc. Amer. Math. Soc. 24 (1970), 438-439.
79