A transformation may have many invariant measures, but some are more improtant than others.
Example 1: To nd the digits in the standardcontinued fraction expansion of a real numberx:
x =a0+ 1 a1+ 1
a2+ 1
a3+ 1
...
= [ao :a1,a2,a3, . . .]
we need theGauÿ map G : [0,1)→(0,1]:
G(x) = 1x − b1xc.
a0 =bxc x1 =x−a0.
ai =b1/xic xi+1 =G(xi) (stop ifxi+1=0).
For example:
π= [3;7,15,1,292,1,1,1, . . .] and
e = [2;1,2,1,1,4,1,1,6,1,1,8,1,1,10, . . .] Lebesgue measure isnot G-invariant.
Denition: A measure µis called absolutely continuousw.r.t. the measureν (notation: µν) if ν(A) =0 impliesµ(A) =0. If both µν andν µ, thenµandν are calledequivalent.
Theorem of Radon-Nikodym: If µis a probability measure and µν then there is a functionh ∈L1(ν)(calledRadon-Nikodym derivativeordensity) such that µ(A) =R
Ah(x) dν(x) for every measurable setA.
Notation: h(x) = dµ(x)dν(x).
Example 1: Lebesgue measure is notG-invariant, but there is a probability measureµG that is absolutely continuous w.r.t.
Lebesgue.
µG(A) = Z
A
h(x)dx for h(x) = 1 log2
1 1+x.
Proposition 1. Suppose that µν are bothT-invariant
probability measures, with a commonσ-algebraBof measurable sets. If ν is ergodic, then µ=ν.
Proof: First we show thatµis ergodic. Indeed, otherwise there is a T-invariant setAsuch that µ(A)>0 and µ(Ac)>0. By
ergodicity ofν at least one ofAor Ac must haveν-measure 0, but this would contradict thatµν.
Now letA∈ B and letY ⊂X be the set of ν-typical points. Then ν(Yc) =0 and hence µ(Yc) =0. Applying Birkho's Ergodic Theorem toµ andν separately for ψ=1A and some µ-typical y∈Y, we get
µ(A) = lim
n
1 n
n−1
X
i=0
ψ◦T(y) =ν(A).
ButA∈ B was arbitrary, soµ=ν.
Exercise 4.1: Show that the condition of ergodicity is essential for Proposition 1?
Ifµν andµis ergodic. Does it follow that µ=ν?
Proposition 2: LetT :U ⊂Rn→U be (piecewise) dierentiable, andµ is absolutely continuous w.r.t. Lebesgue. Thenµis
T-invariant if and only if its densityh = dµdx satises h(x) = X
T(y)=x
h(y)
|detDT(y)| (1) for everyx.
Proof of Proposition 2: TheT-invariance means that dµ(x) =dµ(T−1(x)), but we need to be aware thatT−1 is multivalued. So it is more careful to split the spaceU into pieces Un such that the restrictionsTn:=T|Un are dieomorphic (onto their images) and writeyn=Tn−1(x) =T−1(x)∩Un. Then we obtain (using the change of coordinates)
h(x) dx = dµ(x) =dµ(T−1(x)) =X
n
dµ◦Tn−1(x)
= X
n
h(yn)|det(DTn−1)(x)|dyn=X
n
h(yn)
det|DT(yn)|dyn. Conversely, if (1) holds, then the above computation gives
dµ(x) =dµ◦T−1(x).
Example 1 continued: TheGauÿ map has invariant density h(x) = log121+x1 . Here log12 is just the normalising factor (so that R1
0 h(x)dx =1).
LetIn= (n+11,1n]for n=1,2,3, . . . be the domains of the branches ofG, and forx ∈(0,1), and yn:=G−1(x)∩In= x+n1 . AlsoG0(yn) =−y12
n. Therefore X
n≥1
h(yn)
|G0(yn)| = 1 log2
X
n≥1
yn2 1+yn
= 1 log2
X
n≥1
(x+n)1 2
1+x+n1
= 1
log2 X
n≥1
1
x+n · 1 x+n+1
= 1
log2 X
n≥1
1
x+n − 1
x+n+1 telescoping series
= 1
log2 1
x+1 =h(x).
Exercise 4.2: Compute the average frequency of the digit 1 for points that are normal w.r.t. the standard continued fraction.
Exercise 4.3: Show that for each integern≥2, the interval map given by
Tn(x) =
(nx if 0≤x ≤ 1n,
x1 − bx1c if 1n <x ≤1, has invariant densityh(x) = log121+x1 .
Example 2: The mapT :R\ {0} →R, T(x) =x−1x is called the Boole transformation. It is 2-to-1; the two preimages of x∈Rare y±= 12(x±√
x2+4). Clearly T0(x) =1+x12. It can be shown
that 1
|T0(y−)|+ 1
|T0(y+)| =1. Indeed,
|T0(y±)|=1+ 2 x2+2±x√
x2+4
and 1
|T0(y±)| = x2+2±x√ x2+4 x2+4±x√
x2+4, and
1
|T0(y−)|+ 1
|T0(y+)|
= x2+2−x√ x2+4 x2+4−x√
x2+4+x2+2+x√ x2+4 x2+4+x√
x2+4
= (x2+2−x√
x2+4)(x2+4+x√ x2+4) (x2+4)2−x2(x2+4)
+ (x2+2+x√
x2+4)(x2+4−x√ x2+4) (x2+4)2−x2(x2+4)
= (x2+2)2−x2(x2+4) +2(x2+2)−2x√ x2+4
4(x2+4) +
(x2+2)2−x2(x2+4) +2(x2+2) +2x√ x2+4 4(x2+4)
= 4(x2+2) +8 4(x2+4) =1.
Thereforeh(x)≡1 is an invariant density, so Lebesgue measure is preserved.
Example 3: If T : [0,1]→[0,1]is (countably) piecewise linear, and each branchT :In→[0,1](on whichT is ane) is onto, thenT preserves Lebesgue measure. Indeed, the intervalsIn have pairwise disjoint interiors, and their lengths add up to 1. Ifsn is the slope of T :In→[0,1], thensn=1/|In|. Therefore
X
n
1
DT(yn) =X
n
1 sn =X
n
|In|=1.
Folklore TheoremIfT :S1 →S1 is aC2 expanding circle map, then it preserves a measureµequivalent to Lebesgue, andµis ergodic.
Expandingmeans that there is λ >1 such that|T0(x)| ≥λfor all x∈S1. The above theorem can be proved in more generality, but in the stated version it conveys the ideas more clearly.
Proof: Using the Mean Value Theorem twice, we obtain log|T0(x)|
|T0(y)| = log(1+|T0(x)| − |T0(y)|
|T0(y)| )≤ |T0(x)| − |T0(y)|
|T0(y)|
= |T00(ξ)| · |x−y|
|T0(y)| = |T00(ξ)|
|T0(y)|
|Tx−Ty| T0(ζ) .
SinceT is expanding, the denominators are ≥λand sinceT isC2 on a compact space, also|T00(ξ)|is bounded. Therefore there is someK ≤sup|T00(ξ)|/λ2 such that
log|T0(x)|
|T0(y)| ≤K|T(x)−T(y)|.
The chain rule then gives:
log|DTn(x)|
|DTn(y)| =
n−1
X
i=0
log|T0(Tix)|
|T0(Tiy)| ≤K
n
X
i=1
|Ti(x)−Ti(y)|.
SinceT is a continuous expanding map of the circle, it wraps the circled times around itself, and for each n, there aredn pairwise disjoint intervalsZn such that Tn:Zn→S1 is onto, with slope at leastλn. If we take x,y above in one suchZn, then
|x−y| ≤λ−n|Tn(x)−Tn(y)|
and in fact
|Ti(x)−Ti(y)| ≤λ−(n−i)|Tn(x)−Tn(y)|.
Therefore we obtain log|DTn(x)|
|DTn(y)| = K
n
X
i=1
λ−(n−i)|Tn(x)−Tn(y)|
≤ K
λ−1|Tn(x)−Tn(y)| ≤logK0 for someK0∈(1,∞). This means that if A⊂Zn (so Tn:A→Tn(A) is a bijection), then
1 K0
m(A)
m(Zn) ≤ m(TnA)
m(TnZn) = m(TnA)
m(S1) ≤K0 m(A)
m(Zn), (2) wherem is Lebesgue measure.
Construct theT-invariant measure µ. TakeB ⊂ B arbitrary, and setµn(B) = 1nPn−1
i=0 m(T−iB). Then by (2), 1
K0m(B)≤µn(B)≤K0m(B).
We can take a weak∗ limit of theµn's; call itµ. Then 1
K0m(B)≤µ(B)≤K0m(B),
and thereforeµ andmare equivalent. The T-invariance of µ proven in the same way as in the Theorem of Krylov-Bogul'ubov.
Now for the ergodicity ofµ, we need the Lebesgue Density Theorem, which says that ifm(A)>0, then for m-a.e. x∈A, the limit
ε→lim0
m(A∩Bε(x)) m(Bε(x)) =1,
whereBε(x) is theε-balls aroundx. Pointsx with this property are called(Lebesgue) density pointsof A. (In fact, the above also holds, ifBε(x)is just a one-sided ε-neighbourhood ofx.)
Assume by contradiction thatµis not ergodic. TakeA∈ B a T-invariant set such that µ(A)>0 andµ(Ac)>0. By equivalence ofµandm, alsoδ:=m(Ac)>0. Letx be a density point of A, andZn be a neighbourhood of x such that Tn:Zn→S1 is a bijection. Asn→ ∞,Zn→ {x}, and therefore we can choosen so large (henceZn so small) that
m(A∩Zn)
m(Zn) >1−δ/K0. Therefore m(Am(Zc∩Zn)n) < δ/K0, and using (2),
m(Tn(Ac∩Zn))
m(Tn(Zn)) ≤K0m(Ac∩Zn)
m(Zn) <K0δ/K0=δ.
SinceTn:Ac∩Zn→Ac is a bijection, and
m(TnZn) =m(S1) =1, we getδ=m(Ac)< δ, a contraction.
Thereforeµis ergodic.