Beruflich Dokumente
Kultur Dokumente
STEFANO GALATOLO
Abstract. These are lecture notes for a simple minicourse approaching the
satistical properties of a dynamical system by the study of the associated
transfer operator (considered on a suitable function space).
The following questions will be addressed:
existence of a regular invariant measure;
Lasota Yorke inequalities and spectral gap;
decay of correlations and some limit theorem;
stability under perturbations of the system
linear response
hyperbolic systems
The point of view taken is to present the general construction and ideas
needed to obtain these results in the simplest way. For this, some theorem is
proved in a form which is weaker than usually known, but with an elementary
and simple proof.
Contents
1. Introduction 2
2. Physical measures 3
3. Transfer operator 4
4. Expanding maps: regularizing action of the transfer operator and
existence of a regular invariant measure 6
4.1. Lasota-Yorke inequalities 7
4.2. Existence of a regular invariant measure 9
5. Convergence to equilibrium and mixing 10
5.1. Speed of convergence to equilibrium 11
6. Spectral gap and consequences 12
6.1. Central limit 14
7. Stability and response to perturbation 15
7.1. Stability of …xed points, a general statement 16
7.2. Application to expanding maps 18
7.3. Uniform family of operators and uniform Vs contraction 20
7.4. Lipschitz continuity. 21
8. Some general Linear Response statements 21
8.1. The derivative operator and linear response for circle expanding maps 24
8.2. Rigorous numerical methods for the computation of invariant measures 27
1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
2 STEFANO GALATOLO
1. Introduction
The term statistical properties of a dynamical system refers to the long time
behavior of a trajectory x1 ; :::; xn , or a set of trajectories of the system: their dis-
tribution in the phase space, the average of a given observable along the trajectory
(consider a function f with values on R or C and the time average f (x1 )+:::+f n
(xn )
),
the speed of convergence to those averages, the frequency of a deviation from the
average behavior and so on. As we will see in the following, this relates to the prop-
erties of the evolution of large sets of trajectories, measures or even distributions
by the action of the dynamics.
In chaotic systems the statistical properties of the dynamics are often a better
object to be studied than the pointwise behavior of trajectories. In fact, due to the
initial condition sensitivity, the future behavior of initial data can be unpredictable,
but statistical properties are often regular and their description simpler. This is a
classical approach to dynamics that has been implemented in the so called Ergodic
Theory ( the reader may …nd in any library or even in the references of this notes
very good books on this general theory and about its many applications).
In Ergodic Theory often it is supposed that the dynamics preserve a given mea-
sure, and some other properties (ergodicity, mixing) from which other deep conse-
quences are deduced.
These notes focuses around some questions between dynamics and ergodic theory.
We consider dynamical systems having certain geometrical properties and show that
they have "good" invariant measures. Sometime we can prove that they satisfy
other …ner properties (mixing, fast decay of correlations, spectral gap), so that we
can apply many results from Ergodic Theory and Probability. We also consider
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 3
the important problem of the stability of these invariant measures and mixing
properties, this allows to have information about whether the statistical properties
are stable under small changes in the system or not, or even about the direction of
change of these properties when the system change. (this has of course important
applications for the understanding of the behavior of many systems)
As said before, the properties we mean to investigate are related to the evolution
of measures by the action of the dynamics. We will see that given a dynamical
system is possible to associate to the system a transfer operator, describing the
action of the dynamics on suitable functional spaces of measures (or distributions
sometime). Many important results can be obtained studying the properties of this
transfer operator. This is the main subject of the following sections. We will start
de…ning the transfer operator and its basic properties. We will see how it is possible
to deduce the existence of a regular invariant measure and the speed of convergence
to this measure by the iteration of the dynamics. We will see that under additional
assumptions we may have a precise description of the action of the transfer operator
on suitable spaces of measures (spectral gap, Section 6), and some of its statistical
consequences. We will then consider the problem of stability of all these concepts
under perturbation (Section 7).
The general theory and tools shown in the notes are applied to some of the
simplest kind of uniformly expanding or hyperbolic dynamics. We enter in details
showing how all these concepts can be applied to expanding maps, but we also give
information on the piecewise expanding (Section 9) and hyperbolic case (Section
10).
2. Physical measures
In these notes we consider discrete time dynamical systems. Let X be a metric
space, T : X 7! X a Borel measurable map. We say that a Borel probability
measure is T -invariant if for each measurable set A it holds (A) = (T 1 (A))
. Invariant measures represent equilibrium states, in the sense that probabilities of
events do not change in time. A given a map T , may have many of these invariant
measures, but some of them is particularly important to describe the statistical
properties of the dynamics associated to T . In this section we will de…ne the notion
of Physical measure, which is a particularly important kind of invariant measure.
In the following, we will see that under suitable assumptions some physical measure
measure is an attractor of many other regular measures by the dynamics, and the
speed of convergence to this equilibrium state has important consequences for the
statistical properties of the dynamics.
A set A is called T -invariant if T 1 (A) = A (mod 0). The system (X; T; ) is
said to be ergodic if each T -invariant set has total or null measure. An ergodic
system is then a system which is indecomposable, from the measure theoretical
point of view.
The celebrated Birkho¤ pointwise ergodic theorem says that in this case, time
averages computed along -typical orbits coincides with space average with respect
to : More precisely, in ergodic systems, for any f 2 L1 (X; ) it holds
Z
Snf (x)
(1) lim = fd ;
n!1 n
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
4 STEFANO GALATOLO
3. Transfer operator
Let us consider the space SM (X) of Borel measures with sign on X (equivalently
complex valued measures can be considered). A function T between metric spaces
naturally induces a function L : SM (X) ! SM (X) which is linear and is called
transfer operator (associated to T ). Let us de…ne L: if 2 SM (X) then L[ ] 2
SM (X) is such that
L[ ](A) = (T 1 (A)):
Remark that if the measure we consider is absolutely continuous: d = f dm
(here we are considering the Lebesgue measure m as a reference measure, note
that other measures can be considered) and if T is nonsingular1 , the operator
induces another operator L ~ : L1 (m) ! L1 (m) acting on the measure densities
~ = d(L(f m))
(Lf dm ). By a small abuse of notation we will still indicate by L this
operator.
It is straightforward to see that In this case L : L1 ! L1 is a positive operator2
and preserves the integral
Z Z
Lf dm = f dm;
1 A map is nonsingular (with respect to the Lebesgue measure) when m(T 1 (A)) = 0 ()
m(A) = 0
2f 0 =) Lf 0
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 5
Then
Z Z
g T f dm = [g g^ + g^] T f dm
Z Z
= [g g^] T f dm + g^ T f dm:
Moreover
Z Z
g L(f ) dm = [g g^ + g^] L(f ) dm
Z Z
= [g g^] L(f ) dm + g^ L(f ) dm
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
6 STEFANO GALATOLO
Measures which are invariant for T are …xed points of L. Since physical mea-
sures usually have some "as good as possible" regularity property we will …nd such
invariant measures in some space of "regular" measures. A …rst example which
will be explained in more details below is the one of expanding maps, where we
are going to …nd physical measures in the space of invariant measures having an
absolutely continuous density.
In this section we illustrate one approach which allows to prove the existence of
regular invariant measures. The approach is quite general, but we will show it on
a class of one dimensional maps, where the construction is technically simple. An
important step is to …nd a suitable function space on which the transfer operator
has good properties.
Let us consider a map T which is expanding on the circle. i.e.
T : S1 ! S1;
T 2 C 2;
jT 0 (x)j > 1 8x.
Let us consider the Banach space W 1;1 of absolutely continuous functions3 with
the norm
jjf jj = jjf jj1 + jjf 0 jj1 :
We will show that the transfer operator is regularizing for the jj jj norm. This
implies that iterates of a starting measure have bounded jj jj norm, allowing to
…nd a suitable invariant measure (and much more information on the statistical
behavior of the system, as it will be described in the following sections).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 7
Note that
0 1 0 T 00
(3) (Lf ) = L( f) L( f)
T0 (T 0 )2
0 1 0 T 00
jj(Lf ) jj1 jj f jj1 + jj f jj1
T0 (T 0 )2
T 00
jjf 0 jj1 + jj 0 2 jj1 jjf jj1
(T )
where = max( T10 ).
Hence
0 T 00
jj(Lf ) jj1 + jjLf jj1 jjf 0 jj1 + jjf jj1 + (jj jj1 + 1)jjf jj1
(T 0 )2
and
T 00
jj(Lf )jj jjf jj + (jj jj1 + 1)jjf jj1 :
(T 0 )2
4 In the probabilistic context, this kind of estimations are often called as Doeblin Fortet
inequalities.
5 We will see this in the next paragraphs.
6 To prove the formula consider a small neighborhood B(x; ) of x and check the amount of
measure which is sent there by L. The value of the density Lf (x) will a.e. be the value of the
1
(T (B(x; )))
limit lim !0 2
(or see [15] pag. 85 ).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
8 STEFANO GALATOLO
and then also L satis…es a Lasota Yorke inequality with these norms.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 9
Proof. Consider gn;m = Lm (gn ): Remark that, gn;m 2 W 1;1 and the norms are
uniformly bounded. By the Lasota Yorke inequality
m
(8) jjgn1 ;a+m gn2 ;b+m jjW 1;1 jjgn1 ;a gn2 ;b jjW 1;1 + Bjjgn1 ;a gn2 ;b jj1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
10 STEFANO G ALATOLO
also remark that if jjgnk ;0 hjj1 then jjgnk ;j hjj1 for all j 0. Then the
sequence gnk ;k ! h in L1 ; and by 8 is a Cauchy sequence in the W 1;1 norm, indeed,
suppose k1 k2
k1
(9) jjgnk1 ;k1 gnk2 ;k2 jjW 1;1 jjgnk1 ;0 gnk2 ;k2 k1
jjW 1;1 +Bjjgnk1 ;0 gnk2 ;k2 k1
jj1 :
Since W 1;1 is complete, this implies that it converges in W 1;1 to some limit which
is forced to be h: Hence h 2 W 1;1 : 00 T
(jj 0 jj1 +1)
(T )2
By the Lasota Yorke inequality, since h is invariant then jjhjj = jjLhjj 1 .
Remark 7. Using the procedure explained in this section, using the Lasota Yorke
inequality 7 (remark that starting with the constant density 1 we are constructing
at each step a C 1 function and we are controlling its Lipschitz norm, which is
equivalent to the C 1 norm on C 1 ). It is possible to prove that a C 2 expanding
map has a C 1 invariant density. In a similar way it is possible to prove that a C 3
expanding map of the circle has a C 2 invariant density.
Proof. First let us suppose that jjgjjl < 1. By 7 we know that all the iterates of g
have uniformly bounded l norm
jjLn gjjl M:
Let usRdenote by g + ; g the positive and negative parts of g. Remark that
jjgjj1 = 2 g + dm. There is a point x such that g + (x) 21 jjgjj1 . Around this point
1
consider a neighborhood N = B(x; 41 jjgjj1 M ): For each point x 2 N; g + (x)
1
4 jjgjj1 .
0
Now let d = min jT j; D = max jT 0 j. If n1 is the smallest integer such that
1 log(2jjgjj1 1 M )
dn1 12 jjgjj1 M > 1 ( i.e. n1 > log d ) then T n1 (N ) = S 1 and Ln1 g + has
then density at least
jjgjj1 jjgjj1 jjgjj1 1 log D
= (2jjgjj1 M ) log d
4Dn1 log(2jjgjj1 1 M ) log D
4D
4De log d
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 11
on S 1 . The same is true for g and then, after iterating n1 times this positive
constant part of the density and the corresponding negative one annihilates, and
log D
1
(2M ) log d
setting C = 4D ; it holds
log D
jjLn1 gjj1 jjgjj1 Cjjgjj1 log d +1 :
Let us denote g1 = Ln1 g. We can repeat the above construction and obtain n2 such
that
log D
jjg2 jj1 = jjLn2 g1 jj1 jjg1 jj1 Cjjg1 jj1 log d +1
and so on. Continuing, we have a sequence gn such that
log D
jjgn+1 jj1 jjgn jj1 Cjjgn jj1 log d +1
and then jjgn jj1 ! 0:
If now more generally, g 2 W 1;1 we can approximate g with a g~ such that jj~
g jjl <
1 in a way that jjg g~jj1 . Since jjLjjL1 !L1 1, limn!1 jjLn (g g~)jj1 .
And then the statement follows.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
12 STEFANO G ALATOLO
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 13
(3) (Compact inclusion) the strong zero average space Vs is compactly immersed
in the weak one Vw (more precisely, the strong unit ball in the weak topology
has a …nite net for each );
(4) (Weak boundedness) the weak norm of the operator restricted to Vs satis…es
sup jjLn jVs jjw < 1:
n
Under these assumptions there are C2 > 0; 2 < 1 such that for all g 2 Vs
(11) jjLn gjjs C2 n
2 jjgjjs :
Proof. We …rst show that assumptions 2 and 3 and 4 imply that L is uniformly
contracting from Vs to Vw : there is n1 > 0 such that 8g 2 Vs
(12) jjLn1 gjjw 2 jjgjjs
where 2 B < 1:
Indeed, by item (3), for any there is a …nite set fgi gi2(1;:::;k) in the strong unit
ball B of Vs such that for each g in B there is a gi 2 Vs such that jjg gi jjw :
Hence
sup jjLn gjjw sup jjLn (gi + v)jjw :
g2Vs ;jjgjjs 1 1 i k;v2fv2Vs s:t: jjvjjw g
Since can be chosen as small as wanted and by item (2) for each i, limn!1 jjLn (gi )jjw =
0 and we have Eq. (12) (…rst …x small enough and then choose i big enough) .
Let us apply the Lasota Yorke inequality to strengthen Eq. (12) to an estimate
for the strong norm. For each f 2 Vs
m
jjLn1 +m f jjs A n1
1 jjL f jjs + BjjLn1 f jjw
then
m
jjLn1 +m f jjs A n1
1 jjL f jjs + B 2 jjf jjs
m n1
A 1 [A 1 jjf jjs + Bjjf jjw ] +B 2 jjf jjs :
If m is big enough
jjLn1 +m f jjs 3 jjf jjs
with 3 < 1:
This easily implies the statement. Indeed set n2 = n1 + m; for each k; q 2 N;
q n2 , g 2 V s ,
k
jjLkn2 +q gjjs q
3 jjL gjjs
k q
3 ( 1 jjgjjs + Bjjgjjw ):
Implying that for each g 2 Vs there are C2 > 0; 2 < 1 such that
(13) jjLn gjjs C2 n
2 jjgjjs :
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
14 STEFANO G ALATOLO
By this theorem and the spectral radius formula, the spectral radius of L re-
stricted to Vs is strictly smaller than 1, and the spectral gap (as de…ned in De…nition
15) follows: indeed, …rst remark that by the Lasota Yorke inequality and the spec-
tral radius formula, the spectral radius of L on Bs is not greater than than 1. Since
there is an invariant measure in Bs 10 then this radius is 1. By item 2) there can be
only one …xed point of L in Bs and only one invariant probability measure which we
denote by h (if there were two, consider the di¤erence which is in Vs and iterate...).
Now let us remark that every g 2 Bs can be written as follows:
g = [g h g(X)] + [h g(X)]:
11
the function P : Bs ! Bs de…ned as
P (g) = h g(X)
is a projection. The function N : Bs ! Bs de…ned as
N (g) = L[g h g(X)]
is such that N (Bs ) Vs , N jVs = LjVs , and by (11) satis…es (N) < j j. It holds
L=P +N
and P N = P N = 0. Thus, under the assumption of Proposition 16, L has spectral
gap according to the De…nition 15.
We remark that in several texts the role of Theorem 16 is played by a general
result referred to Hennion, Hervé or Ionescu-Tulcea and Marinescu (see e.g. [27],
[31]).
Remark 17. Equation 13 obviously implies exponential convergence to equilibrium.
Remark 18 (spectral gap for expanding maps of the circle). By Proposition 512 ,
Proposition 8 and the Lasota Yorke inequality, the assumptions of Theorem 16 are
veri…ed on our expanding maps of the circle for the W 1;1 norm (with the L1 norm
as a weak norm). Then their transfer operator have spectral gap.
6.1. Central limit. We see an application of Theorem 16 to the estimation of the
‡uctuations of an observable, obtaining a sort of central limit theorem. A proof of
the result can be found in [31].
Theorem 19. Let (X; T; ) be a mixing probability preserving transformation hav-
ing spectral gap on some Banach space (B; jj jj) containing the constants and satis-
fying
(14) jjf gjj jjf jj jjgjj
jj jj jj jj1 :
R
Let f 2 B be bounded and f d = 0. If there is no 2 B such that f = T
a.e., then 9 > 0 s.t. for all intervals [a; b];
( n 1
) Z
1 X 1 b
t2 =2 2
x: p f T k 2 [a; b] ! p e dt:
n 2 2
a
k=0
10 To prove it apply the same construction as in section 4.2 using the compact immersion
(assumption 3 of Theorem 16) instead of Proposition 5 .
11 Where g(X) stands for the g measure of the whole space.
12 Which can be easily adapted to V , by considering an integral preserving projection
R s 2f =
f f .
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 15
Remark 20. We remark that the assumption in Equation (14) can be relaxed.
Indeed if jjf gjj Kjjf jj jjgjj, then considering the new rescaled norm jj jjK = Kjj jj
it holds
jjf gjjK = Kjjf gjj K 2 jjf jj jjgjj jjf jjK jjgjjK :
We remark, that the W 1;1 norm satis…es (14) after rescaling. Since there is
spectral gap, then Theorem 19 applies to our class of expanding maps on S 1 .
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
16 STEFANO G ALATOLO
Remark 21. We remark that UF3 and UF1 together implies that L0 eventually
contracts the zero average space Vs . Indeed let f 2 Vs , using the inequality and then
the convergence to equilibrium
n
jjLn+m
0 f jjs A m m
1 jjL0 f jjs + BjjL0 f jjw
n m m
A 1 jjL0 f jjs + BE 2 jjf jjs
n m
A 1 (B + A)jjf jjs + BE 2 jjf jjs
by which there are n; m big enough that jjLn+m
0 f jjs 1
2 jjf jjs .
k=1
and
N
X
(LN
0 LN )f = L0N k
(L0 L )Lk 1
f
k=1
N
X
= L0N k
(L0 L )f
k=1
by item b)
N
X
jj(LN
0
N
L )f jjB CN k jj(L0 L )f jjB
k=1
X
jj(L0 L )f jjB Ci
i2[0;N 1]
then X
jjf f0 jjB jjLN
0 (f0 f )jjB + jj(L0 L )f jjB Ci :
i2[0;N 1]
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 17
Now, let us apply the statement to our family of operators satisfying assumptions
UF 1,...,4, supposing Bw = B. We have the following
Proof. We remark that by the uniform Lasota Yorke inequality jjf jjs M are
uniformly bounded.
Hence
jjL f L0 f jjw CM
(see item a) of Lemma 22). Moreover by UF4, Ci M2 :
Hence
jjf f0 jjw CM M2 N + jjLN
0 (f0 f )jjw :
Now by the exponential convergence to equilibrium of L0
jjLN
0 (f f0 )jjw C2 N
2 jj(f f0 )jjs
C2 N
2 M
hence
N
jjf f0 jjB CM M2 N + C2 2 M
j k
log
choosing N = log 2
j k
log
log log 2
(18) jjf f0 jjw CM M2 + C2 2 M
log 2
1
log CM2 M + C2 M:
log 2
Remark 24. We remark that in this statement we did not really used the Lasota
Yorke inequality in its full strength. We used it only to get jjf jjs M: Moreover
the statement could be generalized to slower than exponential convergence to equilib-
rium (see [20]). In the following sections we apply these statements to some classes
of maps, we remark that the modulus of continuity log is sharp for Piecewise
Expanding map (see Section 9).
Remark 25. We remark that in UF2 the size of the perturbation is measured
in the strong-weak norm. This allows general perturbations (allowing to move
discontinuities, like when making small perturbation in the Skorokhod distance, see
Eq. 38) but it is a weak assumption. Measuring the size of the perturbation in
the strong-strong norm, will lead to stronger results, like Lipschitz or di¤ erentiable
stability (see Sections 7.4, 8).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
18 STEFANO G ALATOLO
X f (y) X f (y)
j[L f ](x) [L0 f ](x)j = j j
1
jT 0 (y)j jT00 (y)j
y2T (x) y2T0 1 (x)
X f (y) X f (y)
j j+
1
jT 0 (y)j 1
jT00 (y)j
y2T (x) y2T (x)
X f (y) X f (y)
+j j:
1
jT00 (y)j jT00 (y)j
y2T (x) y2T0 1 (x)
D1 ( )jL f (x)j
jT 0 (y)j
where D1 ( ) = sup j1 jT00 (y)j j and remark that D1 = O( ). For second summand let
1
us denote T (x) = fy1 ; :::; yn g , T0 1 (x) = fy10 ; :::; yn0 g. Let y = supx;i (jyi yi0 j)
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 19
note that y = O( )
X X n
X Xn
f (y) f (y) f (yi ) f (yi0 ) 1 1
j j j j + j f (yi0 )( 0 )j
1
jT00 (y)j jT00 (y)j i=1
jT00 (yi )j i=1
jT0 i )j
(y jT00 (yi0 )j
y2T (x) y2T0 1 (x)
n
X Xn
f (yi ) f (yi0 ) f (yi0 ) jT00 (yi0 )j
j 0 j + j ( 1)j
i=1
jT0 (yi )j i=1
jT00 (yi0 )j jT00 (yi )j
Xn Xn
f (yi ) f (yi0 ) f (yi0 )
j 0 j + D 2 ( )j j
i=1
jT0 (yi )j i=1
jT00 (yi0 )j
R yi 0
Xn f (t)dt
yi0
j j + D2 ( )jL0 f (x)j
i=1
jT00 (yi )j
jT 0 (y 0 )j
where D2 ( ) := sup j jT00 (yii )j 1j = O( ). Hence
0
R yi 0
n
X f (t)dt
yi0
jjL f L0 f jj1 D1 ( )jjL f (x)jj1 + jj jj1 + D2 ( )jjL0 f (x)jj1
jT00 (yi )j
i=1
R yi 0
Xn f (t)dt
yi0
(D1 ( ) + D2 ( ))jjf (x)jj1 + jj jj1
i=1
jT00 (yi )j
n
R yi 0
X yi y
jf (t)jdt
O( )jjf (x)jj1 + jj jj1
i=1
jT00 (yi )j
n 0
X [1[ y ;0]
jf j](yi )
O( )jjf (x)jj1 + jj jj1
i=1
jT00 (yi )j
0
O( )jjf (x)jj1 + D3 jjL [1[ y ;0]
jf j]jj1
0
O( )jjf (x)jj1 + D3 jj1[ y ;0]
jf jjj1
0
O( )jjf (x)jj1 + D3 jj1[ y ;0]
jj1 jjf jj1
O( )jjf jjW 1:1 :
0
where D3 is a constant and [1[ y ;0] jf j] stands for the convolution function be-
tween the characteristic of the interval [ y ;0] (mod 1) and jf 0 j. And the statement
is proved.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
20 STEFANO G ALATOLO
We remark that the contraction of the zero average space for the operator L0
can be obtained simply by directly 21. Before the proof we need the following
Lemma 30. Suppose that L0 satis…es a Lasota Yorke inequality
k
jjLk0 gjjs A 1 jjgjjs + Bjjgjjw
and
8g 2 Bs jj(L L0 )gjjw C jjgjjs ;
8 ; n; g 2 Bs jjLn gjjw M jjgjjw for some M 0;
then Ln approximates Ln0 in the following sense: there are constants C; D such
that 8g 2 Bs ; 8n 0
(22) jj(Ln Ln0 )gjjw (Cjjgjjs + nDjjgjjw ):
Proof.
n
X n
X
jj(Ln Ln0 )gjjw jjLn k
(L L0 )Lk0 1
gjjw M jj(L L0 )Lk0 1
gjjw
k=1 k=1
n
X
M CjjLk0 1
gjjs
k=1
n
X
k 1
MC (A 1 jjgjjs + Bjjgjjw )
k=1
A
M C( jjgjjs + Bnjjgjjw ):
1 1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 21
If n; m are big enough suitably chosen and small enough, then we have that there
is a 3 < 1 such that for each f 2 Vs
jjLn+m f jjs 3 jjf jjs
thus there are 4 < 1; A2 2 R such that for each f 2 Vs and small enough
(23) k
jjL f jjs A2 k4 jjf jjs :
7.4. Lipschitz continuity. We see that exploiting the uniform contraction rate of
Vs and some further assumptions we can prove Lipschitz dependence of the relevant
invariant measure under system perturbations. In this paragraph (essentially taken
from [27]). Further work also lead to di¤erentiable dependence (see next section).
Proposition 31. Let us consider a uniform family L , 2 [0; 1) of operators sat-
isfying UF1,...,UF4. Suppose that each operator L has a unique invariant measure
h in Bs and
(24) jj(L L0 )h0 jjs Ch0 ;
then the dependence is Lipschitz (with respect to the strong norm)
jjh0 h jjs O( ):
Proof. Denote h=h h0 :
(I L ) h = (I L )(h h0 )
= h L h h0 + L h0
= (L L0 )h0 :
By the uniform
P1contraction (21) we have that (I L ) is invertible on Vs ; and
(I L ) 1 = 0 Li is uniformly bounded jj(I L ) 1 jjVs !Vs M2 .
Since (L L0 )h0 2 Vs , then
1
h = (I L ) (L L0 )h0 :
and
(25) jj hjjs M2 Ch0 :
hence we have the statement.
Remark 32. The result can be easily applied to a suitable family of expanding
maps satisfying UF1,UF3,UF4. and 24 obtaining Lipschitz statistical stability on
the strong norm for this family of maps. In Section 8.1 we show a set of easy to be
veri…ed conditions on the family implying 24 .
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
22 STEFANO G ALATOLO
a Linear Response result (even for higher derivatives) was also provided in [32].
Other general results can be found in [7] .
We remark that Theorem 34 applies to systems having less than exponential
convergence to equilibrium. Examples of application of this statement in this case
are outside the scope of these lectures. In next section we apply the statement to
expanding maps.
Let us consider a system having a transfer operator L0 and a family of "nearby"
system L ; 2 [0; ], having suitable properties: suppose there are two normed
vector spaces of measures with sign Bs Bw SM (X) (the strong and weak
space) with norms jj jjw jj jjs and suppose the operators L preserve the spaces:
L (Bs ) Bs and L (Bw ) Bw .
The speed of convergence to equilibrium of a system will be measured by the
speed of contraction to 0 of this space by the iterations of the transfer operator.
jjf jjs M;
(L L0 ) ^ jjs = 0
lim jj f Lf
!0
then
f f0 1^
lim jj (1 L0 ) Lf0 jjw = 0
!0
1
P1
Where (1 L0 ) := 0 Li0 is a continuous operator:Bs ! Bw .
P1 P1
Proof. Let f 2 Bs . Since 0 (n) < 1; then (1 L0 ) 1 f := 0 Li0 f converges in
1
B
Pw1and de…nes a continuous operator Bs ! Bw . It also holds jj(1 L0 ) jjs!w
0 (n).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 23
Denote f =f f0
f f f0
(I L0 ) = (I L0 )
1
= (f L0 f f0 + L0 f0 )
1
= (L L0 )f :
f L L0
(1 + L0 + ::: + Ln0 )(I L0 ) = (1 + L0 + ::: + Ln0 ) f
f f L L0
Ln+1
0 = (1 + L0 + ::: + Ln0 ) f
f f L L0
Ln+1
0 = (1 + L0 + ::: + Ln0 ) (f + f0 f0 )
f L L0
(1 Ln+1
0 ) = (1 + L0 + ::: + Ln0 ) f0 +
L L0
(1 + L0 + ::: + Ln0 ) (f f0 ):
f
then letting n ! 1 it holds that Ln+1
0 ! 0 in the weak norm. and
f 1L L0 1L L0
= (1 L0 ) f0 + (1 L0 ) (f f0 )):
As elements of Bw :Now
X
1L L0 ^ s!s jjf
jj(1 L0 ) (f f0 )jjw ( jjLi0 jjs!w ) jjLjj f0 jjs ! 0
Remark 35. Theorem 34 is quite abstract and it is stated for families of operators.
In particular it may be adapted both to stochastic or deterministic perturbations of
(deterministic of stochastic) systems. One key point is the existence of the derivative
operator (assumption 4) ). The form of this operator is strictly related to the kind
of perturbation considered.
In the following section we will compute this operator for smooth perturbations of
expanding maps, see [7] for the derivative operator in a stochastic case.
Remark 36. In the proof of Theorem 34 the operator L ^ is applied only to f0 and
f : The assumptions of the theorem could be re…ned by supposing that there is a
stronger space Bss Bs such that f 2 Bss for each 0 and supposing that
^ : Bss ! Bs . (as before lim jj (L L0 ) f Lf
L ^ jjs = 0 but with f 2 Bss ).
!0
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
24 STEFANO G ALATOLO
8.1. The derivative operator and linear response for circle expanding
maps. In this subsection we show how to apply Theorem 34 to expanding maps.
Let us consider a family of expanding maps T satisfying UF1,UF3,UF4. We remark
that by this, the assumptions 1) and 2) of Theorem 34 are satis…ed. We now consider
a set of assumptions on the family, in a way that the derivative operator required
at assumption 4) exist, this will imply also that 24 is satis…ed, and by Remark 32
also assumption 3) of Theorem 34 will be satis…ed, giving a linear response linear
response for expanding maps in the L1 norm ( Proposition 40).
In the following we show how to obtain the existence of the derivative operator
for a smooth family of expanding maps. The approach is taken from [22]. Let us
consider T : X ! X be a family of C 3 expanding orientation preserving maps of
the circle X where 2 (0; ). Let us suppose that the dependence of the family on
is di¤erentiable at 0, hence can be written
T (x) = T0 (x) + (x) + oC 3 ( ) f or x 2 X:
where 2 C (X; R), and oC 3 ( ) denotes a term whose C 3 norm tends to zero faster
3
than , as ! 0.13
Let us see more precisely how to apply Theorem 34 to our setting. Let us consider
the following choice of spaces:
R
(1) Bw = L1 (X) with the norm kf kL1 = RX jf (x)jdx; R
(2) Bs = W 1;1 with the norm kf kW 1;1 = X jf (x)jdx + X jf 0 (x)jdx;
(3) Bss = C 2 (X) with the norm kf kC 2 (X) = kf k1 + kf 0 k1 + kf 00 k1 where
kf k1 = supx2X jf (x)j.
We hence have a family of C 3 expanding maps, and each one of them has a C 2
invariant density f (see Remark 7). The following proposition present a detailed
description of the structure of the operator L^ : C 2 (X) ! W 1;1 in our case (see
remark 37).
Proposition 38. Let w 2 C 2 (X; R). For each x 2 X we can write
(27)
^ L w(x) L0 w(x) w 0 w0 T000
Lw(x) = lim = L0 0 (x) L0 (x)+L0 w (x)
!0 T0 T00 T002
and the convergence is also in the C 1 topology.
Before presenting the proof of Proposition 38 we state a simple lemma.
1
Lemma 39. If yi 2 T (x) then we can expand
(yi0 )
yi = yi0 + + oC 2 ( ):
T00 (yi0 )
Proof of Lemma 39. We denote by fyi gdi=1 := T 1 (x) and fyi0 gdi=1 := T0 1 (x) the
d preimages under T and T0 , respectively, of a point x 2 X. Let us write
yi (x) = yi0 (x) + i (x) + Fi ( ; x):
13 More precisely we say that T is a di¤erentiable family of C 3 expanding maps if there exists
2 C 3 (X; R) such that k(T T0 )= kC 3 ! 0 as ! 0, where
kf (x)kC 3 = sup jf (x)j + sup jf 0 (x)j + sup jf 00 (x)j + sup jf 000 (x)j
x2X x2X x2X x2X
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 25
We will show that Fi ( ; x) = oC 3 ( ). Substituting this into the identity T (yi (x)) =
x and and using that T (x) = T0 (x) + (x) + oC 3 ( ) we can expand
T0 (yi0 (x) + i (x) + Fi ( ; x)) = T0 (yi0 (x)) + T00 (yi0 (x))( i (x) + Fi ( ; x)) + oC 2 ( )
and use that T0 (yi0 (x)) = x to cancel terms on either side of (28) to get that
Thus we can identify the …rst order terms as T00 (yi0 (x)) i (x) + (yi0 (x)) and then
what is left is
Proof of Proposition 38. Let us again denote by fyi gdi=1 := T 1 (x) and fyi0 gdi=1 :=
T0 1 (x) the d preimages under T and T0 , respectively, of a point x 2 X. Further-
more, we assume that the indexing is chosen so that yi is a small perturbation of
yi0 , for 1 i d. We can write
d d
!
L w(x) L0 w(x) 1 X w(yi ) X w(yi0 )
=
i=1
T 0 (yi ) i=1 T00 (yi0 )
d
! d
!
1 X 1 1 1 X w(yi ) w(yi0 )
= w(yi ) +
i=1
T 0 (y i) T00 (yi ) i=1
T00 (yi )
| {z } | {z }
=:(I) =:(II)
d
!
1 X 1 1
+ w(yi0 ) :
i=1
T00 (yi ) T00 (yi0 )
| {z }
=:(III)
For the …rst term we …rst di¤erentiate the expansion T (x) = T0 (x)+ (x)+oC 3 ( )
in x to get:
T 0 (x) = T00 (x) + 0
(x) + oC 2 ( ):
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
26 STEFANO G ALATOLO
(yi0 )
(30) w(yi ) = w(yi0 + + oC 2 ( ))
T00 (yi0 )
(yi0 )
(31) = w(yi0 ) + w0 (yi0 )( + oC 2 ( ))
T00 (yi0 )
1 (yi0 )
(32) + ( + oC 2 ( ))2 w00 ( )
2 T00 (yi0 )
Since w00 is uniformly bounded, then
(yi0 )
w(yi ) = w(yi0 ) + w0 (yi0 ) + oC 2 ( ):
T00 (yi0 )
Thus
d d
1 X w(yi ) w(yi0 ) X w0 (y 0 ) i (yi0 )
(II) = = + oC 1 (1)
i=1
T00 (yi ) i=1
T00 (yi ) T00 (yi0 )
d
X (yi0 )w0 (yi0 )
= + oC 1 (1)
i=1
T00 (yi0 )T00 (yi )
1
and therefore, both pointwise and in the C topology
d
1 X w(yi ) w(yi0 ) w0
lim = L0 (x):
!0
i=1
T00 (yi ) T00
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 27
d
!
X (yi0 )T000 (yi0 )
= w(yi0 ) + oC 1 (1)
i=1
T0 (yi0 )2 T00 (yi )
0
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
28 STEFANO G ALATOLO
the L1 distance). Thus the result of the computation has a mathematical meaning
.
This can be done by approximating the transfer operator L0 of the system by a
suitable …nite rank one L which is essentially a matrix, of which we can compute
…xed points and other properties.
There are many ways to construct a suitable L depending on the system which
is considered. The most used one (for L1 approximations) is the so called Ulam
discretization. In this approximation, the system is approximated by a Markov
chain.
In the Ulam Discretization method the phase space X is discretized by a partition
I = fIi g and the system is approximated by a (…nite state) Markov Chain with
transition probabilities
1
(33) Pij = m(T (Ij ) \ Ii )=m(Ii )
(where m is the normalized Lebesgue measure on the phase space). The approx-
imated operator L can be seen in the following way: let F be the algebra
associated to the partition I , let us consider the projection on the step functions
supported on the elements of the partition given by
(f ) = E(f jF )
(34) L = L :
In a series of works it was proved that in several cases the …xed point f of
L converges to the …xed point of L. Explicit bounds on the error have been
given, rigorous methods implemented and experimented in several classes of cases
(see e.g. [6],[11],[12],[14],[28] and [18] where several computations on nontrivial
systems are also shown). Ulam methods and similar methods have been also used
to rigorously compute (up to prescribed errors) other important quantities related
to the statistical properties of dynamics, as Linear Response (see [7]), dimension
of attractors (see [19]) or di¤usion coe¢ cients (see ??). Some more details on the
Ulam method will be given in Section 9.3 in the case of Piecewise Expanding maps
of the interval.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 29
We now consider a class of maps on the interval which are expanding, but allow
discontinuities. This class is interesting and was much studied because it presents
a quite rich behavior, while being approachable with techniques similar to the ones
introduced in the previous sections.
De…nition 41. We call a nonsingular function T : ([0; 1]; m) ! ([0; 1]; m) piece-
wise expanding if
There is a …nite set of points d1 = 0; d2 ; :::; dn = 1 such that for each i,
jT 00 j
Ti := T j(di ;di+1 ) is C 2 and sup[0;1] (T 02 ) dx < 1.
0
inf x2[0;1] jT (x)j > 1 on the set where it is de…ned.
The transfer operator associated to a map of this class has general properties
similar to the ones of the expanding maps, if we apply the transfer operator to
measures having a density we obtain the following formula (see e.g. [15] chapter 4
) for the associated operator (which we continue denoting with L )
In the presence of discontinuities of the map T the transfer operator does not
necessarily preserve spaces of continuous densities. For this the introduction of a
suitable space of regular densities including discontinuous functions is important.
9.1. Bounded variation and the Lasota Yorke inequality for piecewise
expanding maps. Let : [0; 1] ! R a real function. Let fx1 ; :::; xk g [0; 1] be a
sequence of points. Let us de…ne the variation of with respect to fx1 ; :::; xk g as
k
X1
V arfx1 ;:::;xk g ( ) = j (xi ) (xi+1 )j
i=1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
30 STEFANO G ALATOLO
we de…ne the variation of as the supremum of V arfx1 ;:::;xk g ( ) over all the …nite
sequences fx1 ; :::; xk g
V ar( ) = sup V arfx1 ;:::;xk g ( ):
fx1 ;:::;xk g [0;1]
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 31
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
32 STEFANO G ALATOLO
2
This inequality is for maps such that inf I ((T n )0 ) < 1. The inequality can be
i
applied to the other piecewise expanding maps by previously iterating the map
2
until inf I ((T n )0 ) < 1.
i
In this case, as before, a straightforward computation lead to the general Lasota
Yorke inequality valid for any piecewise expanding map: there are A; B 0 and
2 [0; 1) such that
n
(37) jjLnT jjBV A jj jjBV + Bjj jj1 :
The …rst consequence of the inequality is the existence of a bounded variation
invariant density for each piecewise expanding map. By (37) and Theorem 42,
repeating the arguments stated in Section 4.2 we get the existence of an absolutely
continuous invariant measure with bounded variation density for this kind of maps.
By arguments very similar to the ones presented in Section 6 it is also possible to
obtain that a mixing piecewise expanding map has spectral gap on the space of
bounded variation densities14 .
This kind of maps however, have a more complicated behavior than expanding
ones, with respect to perturbations. We point out that the Lasota Yorke inequality
we have proved, works only if the expansion rate of the map is bigger than 2. In the
previous paragraph, to prove the existence of an absolutely continuous invariant
2
measure we had to take an iterate of the map such that inf I ((T n )0 ) < 1. When
i
considering the statistical stability of a family of maps, this is not only a technical
point but is substantial, because sometime it is not possible to …nd a uniform iterate
which is suitable for the whole family. While many of the stability arguments
outlined in the previous sections applies also to piecewise expanding maps with
expansion rate greater than 2, for the maps T such that 1 inf Ii (T 0 ) 2 the
stability questions are more dedicated. We present below some examples of results
illustrating the questions.
14 Using the BV and L1 as strong and weak spaces and its Lasota Yorke inequality, Theorem
16 and Theorem 42 to obtain the compact inclusion property.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 33
The maps are piecewise expanding, let ha;d;r denote the unique invariant density
for Wa;b;r . Now let us consider a sequence (an ; dn ; rn ) ! ( 21 ; 12 ; 14 ) and the related
densities han ;dn ;rn . In [24] (page 331) is shown that h1; 12 ;r = 23 1[0; 12 ] + 12 1( 12 ;1] and
h 12 ; 21 ;r = 21[0; 21 ] while if 12 < bn 1 2rn then han ;dn ;rn ! 1=2 weakly. This is due
to the fact that for 21 < b 1 2r the interval [1 b; b] is sent to itself by the map,
and "attracts" all the measure while iterating the map. Hence for a = 1=2 and
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
34 STEFANO G ALATOLO
a = 1 the limit measure does not coincide with the invariant absolutely continuous
measure of the limit map shown above (by the way the limit map has 1=2 as a non
absolutely continuous invariant measure since 1=2 is a …xed point). We remark
that for this family of maps we cannot have a uniform Lasota Yorke inequality as
in UF1, as the slopes tend to 2. On the other hand if one takes iterates of the maps
to increase the slope, the smaller and smaller invariant interval around 1=2 let the
second coe¢ cient of the inequality to converge to 1.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 35
indeed jjLf jjL1 jjf jjL1 and jjE(f jF )jjL1 jjf jjL1 , and L comes from the com-
position of such functions.
Hence we proved that UF1,...,UF4 applies and thus we can apply Proposition
2
23, concluding that for Piecewise expanding maps with inf I ((T )0 ) < 1 the Ulam
i
method approximation h converge to the real invariant density h and we have a
quantitative estimation
jjh h0 jj1 = O( log ):
In [11] it is proved that this rate is the optimal one. There are examples of map
for which the approximation rate is asymptotically proportional to log :
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
36 STEFANO G ALATOLO
15 In the following to simplify notations, when no confusion is possible we will indicate the
generic leaf or its coordinate with .
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 37
The notation we use for this norm is similar to the usual L1 norm.
Remark 52. Indeed this is formally the case if we associate to ; by disintegration,
a path G : S 1 ! SB(D2 ) de…ned by G ( ) = j . In this case, this will be the
L1 norm of the path. For more details about the disintegration and the properties
of the restriction, see the appendix of [21].
Later, similarly we will de…ne a norm jj jjW 1;1 which will work as a Sobolev norm
for these paths (see De…nition 61).
jj jjW :
Taking the supremum over jgj1 1 and Lip(g) 1 we …nish the proof of the
inequality.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
38 STEFANO G ALATOLO
Xq Z
= jj j jjW dm1 ( )
i=1 i
= jj jj"1" :
= jj jjW + (X):
And taking the supremum over jgj1 1 and Lip(g) 1 we have jj F jjW
jj jjW + (X). In particular, if (X) = 0 we get the second part.
Proposition 57. For all signed measure 2 L1 it holds
(49) jj F jj"1" jj jj"1" + ( + 1)jj x jj1 :
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 39
j = ;y :
q Z +j + 1
(Ti 1 ( ))
X FT 1
( ) Ti 1
( ) x (Ti ( )) FT 1
( )
jT
i
1
( ) x
i i
jj F jj"1" dm1 ( )
i=1 I jTi0 j Ti 1 ( ) jTi0 j Ti 1 ( )
W
q Z +j + 1 +j 1
X FT 1
( ) Ti 1
( ) x (Ti ( )) FT 1
( ) Ti 1
( ) x (Ti ( ))
i i
1 dm1 ( )
i=1 I jTi0 j Ti ( ) jTi0 j Ti 1 ( )
W
q Z +j 1 1
X FT 1
( ) Ti 1
( ) x (Ti ( )) FT 1
( )
jT
i
1
( ) x (Ti ( ))
i i
+ 1 dm1 ( )
i=1 I jTi0 j Ti ( ) jTi0 j Ti 1 ( )
W
= I1 + I2
where
q Z +j + 1 +j (Ti 1 ( ))
X FT 1
( ) Ti 1
( ) x (Ti ( )) FT 1
( ) Ti 1
( ) x
i i
I1 = dm1 ( )
i=1 I jTi0 j Ti 1 ( ) jTi0 j Ti 1 ( )
W
and
q
X Z FT +j 1 (Ti 1 ( )) FT jT 1 (Ti 1 ( ))
1
( ) Ti ( ) x 1
( ) i ( ) x
i i
I2 = dm1 ( ):
i=1 I jTi0 j Ti 1 ( ) jTi0 j Ti 1 ( )
W
q
X Z +
j x x j
I1 = FT 1
+j
Ti 1
( ) Ti 1 ( )dm1 ( )
I i ( ) W jTi0 j
Zi=1
+j +
= F j x x j( )dm1 ( )
I W
Z
+
= j x x j( )dm1 ( )
I
= jj x jj1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
40 STEFANO G ALATOLO
q
X Z
x
I2 = FT 1
+j
Ti 1
( ) jT 1
( ) Ti 1 ( )dm1 ( )
i=1 I i ( ) i W jTi0 j
Xq Z
+j j ( )dm1 ( )
F x
Ii W
i=1
Z
+j j ( )dm1 ( )
x
I W
Z
+j +j +
x ( ) x( ) dm1 ( )
I W
Z Z
+j +j + +j +
x ( ) x( ) dm1 ( ) + x( ) j x ( ) dm1 ( )
I W I W
= jj x jj1 + jj jj"1" :
Summing the above estimates we …nish the proof.
Iterating (49) we get the following corollary.
Corollary 58.
jjLnF jj"1" n
jj jj"1" + jj x jj1 ;
1+
where = 1 .
Let us consider the set of zero average measures
(50) V = f 2 L1 : (X) = 0g:
R
Since x = x m1 ( x = + x x ) we have x dm1 = 0. From the last
corollary and the convergence to equilibrium for expanding maps with respect to
L1 and W 1;1 norms (see Remarks 17 and 18) it directly follows:
Proposition 59 (Exponential convergence to equilibrium). There exist D 2 R and
0 < 1 < 1 such that, for every signed measure 2 V, it holds
n
jjLnF jj"1" D2 1 (jj jj1 + jj x jjW 1;1 )
for all n 1.
We now prove the existence of an invariant measure for the solenoidal system in
the set L1 it is not di¢ cult to deduce that this should be a physical measure (points
in the same stable leaves must have the same long time average for a continuous
observable).
Proposition 60. There is a unique 2 L1 such that LF = .
Proof. The base map T is expanding and has an absolutely continuous invariant
measure. Let us call it 'x . Consider the measure = 'x m (the measure
having marginal 'x and Lebesgue measure on the sable leaves) and the sequence
n
n = LF : By Proposition 59 jj n m jj"1" D2 n1 and n is a Cauchy sequence.
By passing to a subsequence we can …nd nk such that for almost each leaf , nk j
is a Cauchy sequence for the jj jjW norm. By Remark 48 this sequence must have a
limit which is a positive measure. This de…nes a limit measure which is invariant.
The integrability of jj jjW follows by the fact that jj nk j jjW sup 'x and thus
it is a bounded sequence and jj nk j jjW converges pointwise to jj jjW :
The uniqueness follow trivially by Proposition 59.
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 41
10.3. Strong norm and Lasota Yorke inequality. We give here an example of
a strong space satisfying a kind of Lasota Yorke inequality which holds for positive
measures.
Given 2 L1 let us denote by its marginal density. Let us consider the
following space of measures
( )
1;1
2 L1 : 2 W 1;1 8 1 lim 2 ! 1 jj j 2 j 1 jjW = 0 and
"W "= j j 1
f or almost all 1 ; D( ; 1 ) := lim sup 2 ! 1
jj 2 jjW < 1
2 1
this will play the role of the strong norm in the solenoid case. Indeed the following
Lasota-Yorke-like inequality can be proved
Proposition 62. Let F be a solenoidal map, then.LF "W 1;1 " "W 1;1 "and there
are < 1; B > 0 s.t 8 2 "W 1;1 " such that 0.
0
jjLF jj"W 1;1 " ( jj jj"W 1;1 " + jj jj1 )+Bjj jj"1" :
By Equation 47 we have
q F
X jT 1
1 ( )
T i ( ) i
(51) (LF )j = 0 1 f or almost all 2 N1 :
i=1
jTi Ti ( ))j
Then
q Z
X FT jT 1 FT jT 1
1 i
1
( 1) i ( 1) i
1
( 2) i ( 2)
jjD(LF ; 1 )jj1 lim sup jj ( 0 1 0 1 )jj dm( 1)
i=1 2! 1 2 1 jTi Ti ( 1 ))j jTi Ti ( 2 ))j W
Xq Z FT jT 1 FT jT 1
1 i
1
( 1) i ( 1) i
1
( 2) i ( 2)
lim sup jj 0 1 jjW dm
i=1 2! 1 2 1 jTi Ti ( 1 ))j
Z
1 1 1
+j lim sup jj FT 1 jT 1
( 2)
( 1 )jjW dm
Ti 1 (
( 0 0
2) jTi Ti ( 1 ))j jTi 2 ))j
i i
2! 1 2 1
q Z
X FT jT 1 FT jT 1
1 i
1
( 1) i ( 1) i
1
( 2) i ( 2)
0 1 lim sup jj jjW dm
i=1
jTi Ti ( 1 ))j 2! 1 2 1
Z
1 1 1
+ jj FT 1 jT 1
( 1)
jjW lim sup ( )jdm
Ti 1 ( Ti 1 (
( 0 0
1) jTi 2 ))j jTi 1 ))j
i i
2! 1 2 1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
42 STEFANO G ALATOLO
Hence
q Z
X FT jT 1 FT jT 1
1 i
1
( 1) i ( 1) i
1
( 1) i ( 2)
jjD(LF ; 1 )jj1 0 1 lim sup jj jjW
i=1
jTi Ti ( 1 ))j 2! 1 2 1
1 FT 1
(
jT 1
( 2)
FT 1 jT 1
( 2)
i 1) i i ( 2) i
+ 0 1 lim sup jj jj dm
jTi Ti ( 1 ))j 2! 1 2 1
Z
1 1 1
+ jj FT 1 jT 1
( 1)
jjW lim sup ( 1 )jdm
Ti 1 (
( 0 0
1) jTi Ti ( 2 ))j jTi 1 ))j
i i
2! 1 2 1
= I + II + III
q Z
X FT jT 1 jT 1
1 i
1
( 1) i ( 1) i ( 2)
I= 0 1 lim sup jj jjW
i=1
jTi Ti ( 1 ))j 2! 1 2 1
8 9
q Z
X < jT 1 jT 1 jT 1 (D 2
) jT 1 (D 2
) =
1 i ( 1) i ( 2) i ( 1) i ( 2)
I lim sup jj jjW +j j d 1
Ti 1 (
0
jTi 1 ))j 2! 1
: 2 1 2 1 ;
i=1
and
q Z
X jT jT
1 Ti 1 ( 2) Ti 1 ( 1) i
1
( 1) i
1
( 2)
I 0 1 [lim sup lim sup jj 1 1 jjW
i=1
jTi Ti ( 1 ))j 2! 1 2 1 2! 1 Ti ( 2) Ti ( 1)
Ti 1 ( Ti 1 ( jT 1 (D2 ) jT 1 (D2 )
2) 1) i ( 1) i ( 2)
+ lim sup lim sup j j]d 1
2! 1 2 1 2! 1 Ti 1 ( 2) Ti 1 ( 1)
q
X Z
Ti 1 ( 2) Ti 1 ( 1) 0
sup lim sup D( ; Ti 1 ( 1 )) +j 1 1
x (Ti ( 1 ))jdTi ( 1 ):
i=1 1 2! 1 2 1 Ii
Ti 1 ( 2) Ti 1 ( 1) 0
I sup lim sup [ jjD( ; 1 )jj1 + jj jj1 ]
i; 1 2! 1 2 1
1 0
[ jjD( ; 1 )jj1 + jj jj1 ]:
inf jT 0 j
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 43
Ti 1 ( 2) Ti 1 ( 1)
sup lim sup
i; 1 2! 1 2 1
XZ 1 FT 1
(
jT 1
( 2)
FT 1 jT 1
( 2)
i 1) i i ( 2) i
0 1 lim sup jj 1 1 jjW d 1
i I jTi Ti ( 1 ))j 2! 1 Ti ( 2) Ti ( 1)
Ti 1 ( Ti 1 ( XZ FT jT 1 FT jT 1
2) 1)
1 ( 1
( 1) 2) ( 2)
( 2)
jjW dTi 1 (
i i i i
sup lim sup lim sup jj 1 1 1)
i; 1 2! 1 2 1 i Ii 2! 1 Ti ( 2) Ti ( 1)
1 @G
0
jj jj1 jj jj1 :
inf jT j @x
where in the last step we used that 0. Finally
XZ 1 1 1
III jj FT 1 ( ) jT 1 ( 1 ) jjW lim sup ( )jdm( 1)
Ti 1 ( Ti 1 (
0 0
i
i 1 i
2! 1 2 1 jTi 2 ))j jTi 1 ))j
T 00
jjT 0 jj1 jj jj1 jj jj1 :
(T 0 )2
Summarizing
(52)
1 0 @G T 00
jjLF jjW 1;1 ( jj jj"W 1;1 " + jj jj1 +jj jj1 jj jj1 )+(1+jjT 0 jj1 jj 0 2 jj1 )jj jj1 :
inf jT 0 j @x (T )
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
44 STEFANO G ALATOLO
R
Now let us consider g such that g dm 6= 0
Z Z Z
T n g dm g dm d
Z Z Z
n g
( g dm)[ T R dm T n h dm]
g dm
Z Z
g
( g dm) Tn [R h] dm
g dm
R
and since R ggdm h dm = 0
Z Z Z Z
(53) T n g dm g dm d (n)jj jj1 jjg h g dmjjW 1;1
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 45
R
Proposition 64. If for some f 0 with f dm = 1;
n
L f !h
1
weakly in L . Then T has an absolutely continuous invariant probability measure
with density h:
R w R
Proof. Let us suppose without loss of generality hdm = 1: Ln f ! h f dm if
and only if 8 2 L1
Z Z Z
h dm = lim Ln+1 f dm = lim ( T ) Ln f dm
n!1 n!1
Z Z
= ( T ) h dm = Lh dm:
References
[1] J. F. Alves, M. Sou… Statistical Stability in Chaotic Dynamics Progress and Challenges in
Dyn. Sys. Springer Proc. in Math. & Statistics V. 54, 2013, pp 7-24
[2] V. Baladi Positive transfer operators and decay of correlations, Advanced Series in Nonlinear
Dynamics, 16 World Sci. Publ., NJ, (2000).
[3] V.Baladi, On the susceptibility function of piecewise expanding maps,Comm. Math. Phys.
275, I. 3, pp 839–859 (2007).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
46 STEFANO G ALATOLO
[4] V. Baladi, Linear response, or else, ICM Seoul 2014, Proceedings, Volume III, 525–545,
[5] V. Baladi, M. Tsujii Anisotropic Holder and Sobolev spaces for hyperbolic di¤ eomorphisms
Ann. Inst. Fourier (Grenoble) 57 (2007), no. 1, 127–154.
[6] W. Bahsoun, C. Bose Invariant densities and escape rates: rigorous and computable estima-
tion in the L 1 norm. Nonlinear Analysis, 2011, vol. 74, 4481-4495.
[7] W. Bahsoun, S. Galatolo, I Nisoli, X Niu A Rigorous Computational Approach to Linear
Response arXiv:1506.08661
[8] Bahsoun, W., Galatolo, S., Nisoli, I. and Niu, X., Rigorous approximation of di¤usion
coe¢ cients for expanding maps. Journal of Statistical Physics V. 163 (2016) Issue 6, pp
1486âe“1503.
[9] G. Basso A Hitchhiker’s guide to Wasserstein distances
http://n.ethz.ch/~gbasso/download/A%20Hitchhikers%20guide%20to%20Wasserstein/A%20Hitchhikers%20guide%20to%20Wa
[10] V. Bogachev, A. V. Kolesnikov The Monge-Kantorovich problem: achievements, connections,
and perspectives Russian Mathematical Surveys, Volume 67, Number 5 (2012)
[11] Christopher J. Bose, R. Murray. The exact rate of approximation in Ulam’s method. Discrete
and Continuous Dynamical Systems, 7(1):219-235 (2001).
[12] M. Dellnitz, O. Junge Set Oriented Numerical Methods for Dynamical Systems Handbook
of dynamical systems vol 2 - Elsevier, (2002).
[13] M Demers H.-K. Zhang A functional analytic approach to perturbations of the Lorentz Gas,
Communications in Mathematical Physics 324:3 (2013), 767-830.
[14] G. Froyland Extracting dynamical behaviour via Markov models in Alistair Mees, editor,
Nonlinear Dynamics and Statistics: Proceedings, Newton Institute, Cambridge (1998): 283-
324, Birkhauser, 2001.
[15] A. Boyarsky P. Gora Laws of Chaos: Invariant Measures and Dynamical Systems in One
Dimension Probability and Its Applications Birkhäuser Boston, 1997
[16] Evans, Lawrence C. (2010). Partial Di¤ erential Equations (2nd ed.). American Mathematical
Society. ISBN 0-8218-4974-3.
[17] Z. Li, P. Gora, A. Boyarsky, H. Proppe, P. Eslami Family of piecewise expanding maps having
singular measure as a limit of ACIMs Erg. Th. Dyn. Sys. (2013)
[18] S. Galatolo I Nisoli An elementary approach to rigorous approximation of invariant measures
SIAM Journal on Applied Dynamical Systems 2014, Vol. 13, No. 2, pp. 958-985
[19] Galatolo, S., Nisoli, I., Rigorous computation of invariant measures and fractal dimension
for piecewise hyperbolic maps: 2D Lorenz like maps. Ergodic Theory and Dynamical Systems
36, I. 6 (2016), pp. 1865-1891 doi: 10.1017/etds.2014.145
[20] S. Galatolo Quantitative statistical stability and speed of convergence to equilibrium for par-
tially hyperbolic skew products arXiv:1702.05996
[21] Galatolo S., Lucena R Spectral Gap and quantitative statistical stability for systems with
contracting …bers and Lorenz like maps. arXiv:1507.08191
[22] S. Galatolo, M. Pollicott Controlling the statistical properties of expanding maps
arXiv:1603.07762
[23] S.Gouezel, C. Liverani Banach spaces adapted to Anosov systems, Ergodic Theory and Dy-
namical Systems, 26, 1, 189–217, (2006)
[24] G. Keller Stochatic stability in some chaotic dynamical systems Mh. Math. 94, 313-333 (1982)
[25] M. Mazzolena Dinamiche espansive unidimensionali: dipendenza della misura invariante da
un parametro Master thesis, Università di Roma Tor Vergata (2006)
[26] O. E. Lanford III Informal Remarks on the Orbit Structure of Discrete Approximations to
Chaotic Maps Exp. Math. (1998), 7: 317:324
[27] C. Liverani Invariant measures and their properties. A functional analytic point of view,
Dynamical Systems. Part II: Topological Geometrical and Ergodic Properties of Dynamics.
Centro di Ricerca Matematica “Ennio De Giorgi”: Proceedings. Published by the Scuola
Normale Superiore in Pisa (2004).
[28] C. Liverani, Rigorous numerical investigations of the statistical properties of piecewise ex-
panding maps–A feasibility study, Nonlinearity (2001), 14: 463-490.
[29] A. Lasota, J.Yorke On the existence of invariant measures for piecewise monotonic transfor-
mations , Trans. Amer. Math. Soc. (1973), 186: 481-488.
[30] K. Oliveira, M. Viana Fudamentos da Teoria Ergà 3 dica, Colecao Fronteiras da Matematica,
SBM, (2014).
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19
STATISTICAL PROPERTIES OF DYNAM ICS FUNCTIONAL ANALYTIC APPROACH 47
AMS Open Math Notes: Works in Progress; Reference # OMN:201707.110711; 2017-07-31 02:35:19