Dispersing billiards with moving scatterers
DDISPERSING BILLIARDS WITH MOVING SCATTERERS
MIKKO STENLUND, LAI-SANG YOUNG, AND HONGKUN ZHANG
Abstract.
We propose a model of Sinai billiards with moving scatterers, in which the locationsand shapes of the scatterers may change by small amounts between collisions. Our main resultis the exponential loss of memory of initial data at uniform rates, and our proof consistsof a coupling argument for non-stationary compositions of maps similar to classical billiardmaps. This can be seen as a prototypical result on the statistical properties of time-dependentdynamical systems.
Acknowledgements.
Stenlund is supported by the Academy of Finland; he also wishes tothank Pertti Mattila for valuable correspondence. Young is supported by NSF Grant DMS-1101594, and Zhang is supported by NSF Grant DMS-0901448.1.
Introduction
Motivation.
The physical motivation for our paper is a setting in which a finite numberof larger and heavier particles move about slowly as they are bombarded by a large number oflightweight (gas) particles. Following the language of billiards, we refer to the heavy particles as scatterers . In classical billiards theory, scatterers are assumed to be stationary, an assumptionjustified by first letting the ratios of masses of heavy-to-light particles tend to infinity. We donot fix the scatterers here. Indeed the system may be open — gas particles can be injectedor ejected, heated up or cooled down. We consider a window of observation [0 , T ] , T ≤ ∞ ,and assume that during this time interval the total energy stays uniformly below a constantvalue E >
0. This places an upper bound proportional to √ E on the translational and rotationalspeeds of the scatterers. The constant of proportionality depends inversely on the masses andmoments of inertia of the scatterers. Suppose the scatterers are also pairwise repelling due toan interaction with a short but positive effective range, such as a weak Coulomb force, whosestrength tends to infinity with the inverse of the distance. The distance between any pair ofscatterers has then a lower bound, which in the Coulomb case is proportional to 1 /E . In brief,fixing a maximum value for the total energy E , the scatterers are guaranteed to be uniformlybounded away from each other; and assuming that the ratios of masses are sufficiently large,the scatterers will move arbitrarily slowly. Our goal is to study the dynamics of a tagged gasparticle in such a system on the time interval [0 , T ]. As a simplification we assume our taggedparticle is passive: it is massless, does not interact with the other light particles, and does notinterfere with the motion of the scatterers. It experiences an elastic collision each time it meetsa scatterer, and moves on with its own energy unchanged. This model was proposed in thepaper [16].The setting above is an example of a time-dependent dynamical system . Much of dynamicalsystems theory as it exists today is concerned with autonomous systems, i.e., systems for whichthe rules of the dynamics remain constant through time. Non-autonomous systems studiedinclude those driven by a time-periodic or random forcing (as described by SDEs), or moregenerally, systems driven by another autonomous dynamical system (as in a skew-productsetup). For time-varying systems without any assumption of periodicity or stationarity, even
Mathematics Subject Classification.
Key words and phrases.
Memory loss, dispersing billiards, time-dependent dynamical systems, non-stationarycompositions, coupling. The model here should not be confused with [8], which describes the motion of a heavy particle bombardedby a fast-moving light particle reflected off the walls of a bounded domain. a r X i v : . [ m a t h . D S ] N ov MIKKO STENLUND, LAI-SANG YOUNG, AND HONGKUN ZHANG the formulation of results poses obvious mathematical challenges, yet many real-world systemsare of this type. Thus while the moving scatterers model above is of independent interest, wehad another motive for undertaking the present project: we wanted to use this prototypicalexample to catch a glimpse of the challenges ahead, and at the same time to identify techniquesof stationary theory that carry over to time-dependent systems.1.2.
Main results and issues.
We focus in this paper on the evolution of densities. Let ρ be an initial distribution, and ρ t its time evolution. In the case of an autonomous systemwith good statistical properties, one would expect ρ t to tend to the system’s natural invariantdistribution (e.g. SRB measure) as t → ∞ . The question is: How quickly is ρ “forgotten”?Since “forgetting” the features of an initial distribution is generally associated with mixing ofthe dynamical system, one may pose the question as follows: Given two initial distributions ρ and ρ (cid:48) , how quickly does | ρ t − ρ (cid:48) t | tend to zero (in some measure of distance)? In thetime-dependent case, ρ t and ρ (cid:48) t may never settle down, as the rules of the dynamics may bechanging perpetually. Nevertheless the question continues to makes sense. We say a systemhas exponential memory loss if | ρ t − ρ (cid:48) t | decreases exponentially with time.Since memory loss is equivalent to mixing for a fixed map, a natural setting with exponentialmemory loss for time-dependent sequences is when the maps to be composed have, individually,strong mixing properties, and the rules of the dynamics, or the maps to be composed, varyslowly. (In the case of continuous time, this is equivalent to the vector field changing veryslowly.) In such a setting, we may think of ρ t above as slowly varying as well. Furthermore,in the case of exponential loss of memory, we may view these probability distributions asrepresenting, after an initial transient, quasi-stationary states .Our main result in this paper is the exponential memory loss of initial data for the collisionmaps of 2D models of the type described in Section 1.1, where the scatterers are assumed to bemoving very slowly. Precise results are formulated in Section 2. Billiard maps with fixed, convexscatterers are known to have exponential correlation decay; thus the setting in Section 1.1 isa natural illustration of the scenario in the last paragraph. (Incidentally, when the source andtarget configurations differ, the collision map does not necessarily preserve the usual invariantmeasure).If we were to iterate a single map long enough for exponential mixing to set in, then changethe map ever so slightly so as not to disturb the convergence in | ρ t − ρ (cid:48) t | already achieved, anditerate the second map for as long as needed before making an even smaller change, and soon, then exponential loss of memory for the sequence is immediate for as long as all the mapsinvolved are individually exponentially mixing. This is not the type of result we are after. Amore meaningful result — and this is what we will prove — is one in which one identifies aspace of dynamical systems and an upper bound in the speed with which the sequence is allowedto vary, and prove exponential memory loss for any sequence in this space that varies slowlyenough. This involves more than the exponential mixing property of individual maps; the classof maps in question has to satisfy a uniform mixing condition for slowly-varying compositions .This in some sense is the crux of the matter.A technical but fundamental issue has to do with stable and unstable directions, the staples ofhyperbolic dynamics. In time-dependent systems with slowly-varying parameters, approximatestable and unstable directions can be defined, but they depend on the time interval of interest,e.g., which direction is contracting depends on how long one chooses to look. Standard dy-namical tools have to be adapted to the new setting of non-stationary sequences; consequentlytechnical estimates of single billiard maps have to be re-examined as well.1.3. Relevant works.
Our work lies at the confluence of the following two sets of results:The study of statistical properties of billiard maps in the case of fixed convex scattererswas pioneered by Sinai et al [3, 4, 17]. The result for exponential correlation decay was firstproved in [20]; another proof using a coupling argument is given in [6]. Our exposition herefollows closely that in [6]. Coupling, which is the main tool of the present paper, is a standard
ISPERSING BILLIARDS WITH MOVING SCATTERERS 3 technique in probability. To our knowledge it was imported into hyperbolic dynamical systemsin [21]. The very convenient formulation in [6] was first used in [8]. (Despite appearing in 2009,the latter circulated as a preprint already in 2004.) We refer the reader to [10], which containsa detailed exposition of this and many other important technical facts related to billiards.The paper [16] proved exponential loss of memory for expanding maps and for one-dimensionalpiecewise expanding maps with slowly varying parameters. An earlier study in the same spiritis [13]. A similar result was obtained for topologically transitive Anosov diffeomorphisms in twodimensions in [18] and for piecewise expanding maps in higher dimensions in [12]. We mentionalso [2], where exponential memory loss was established for arbitrary sequences of finitely manytoral automorphisms satisfying a common-cone condition. Recent central-limit-type results inthe time-dependent setting can be found in [11, 15, 19].1.4.
About the exposition.
One of the goals of this paper is to stress the (strong) similaritiesbetween stationary dynamics and their time-dependent counterparts, and to highlight at thesame time the new issues that need to be addressed. For this reason, and also to keep the lengthof the manuscript reasonable, we have elected to omit the proofs of some technical preliminariesfor which no substantial modifications are needed from the fixed-scatterers case, referring thereader instead to [10]. We do not know to what degree we have succeeded, but we have triedvery hard to make transparent the logic of the argument, in the hope that it will be accessibleto a wider audience. The main ideas are contained in Section 5.The paper is organized as follows. In Section 2 we describe the model in detail, after which weimmediately state our main results in a form as accessible as possible, leaving generalizationsfor later. Theorems 1–3 of Section 2 are the main results of this paper, and Theorem 4 isa more technical formulation which easily implies the other two. Sections 3 and 4 contain acollection of facts about dispersing billiard maps that are easily adapted to the time-dependentcase. Section 5 gives a nearly complete outline of the proof of Theorem 4. In Section 6 wecontinue with technical preliminaries necessary for a rigorous proof of that theorem. UnlikeSections 3 and 4, more stringent conditions on the speeds at which the scatterers are allowedto move are needed for the results in Section 6. In Section 7 we prove Theorem 4 in the specialcase of initial distributions supported on countably many curves, and in Section 8 we provethe extension of Theorem 4 to more general settings. Finally, we collect in the Appendix someproofs which are deferred to the end in order not to disrupt the flow of the presentation in thebody of the text. 2.
Precise statement of main results
Setup.
We fix here a space of scatterer configurations, and make precise the definition ofbilliard maps with possibly different source and target configurations.Throughout this paper, the physical space of our system is the 2-torus T . We assume, tobegin with (this condition will be relaxed later on), that the number of scatterers as well as theirsizes and shapes are fixed, though rigid rotations and translations are permitted. Formally, let B , . . . , B s be pairwise disjoint closed convex domains in R with C boundaries of strictlypositive curvature. In the interior of each B i we fix a reference point c i and a unit vector u i at c i . A configuration K of { B , . . . , B s } in T is an embedding of ∪ s i =1 B i into T , onethat maps each B i isometrically onto a set we call B i . Thus K can be identified with a point( c i , u i ) s i =1 ∈ ( T × S ) s , c i and u i being images of c i and u i . The space of configurations K isthe subset of ( T × S ) s for which the B i are pairwise disjoint and every half-line in T meets ascatterer non-tangentially. More conditions will be imposed on K later on. The set K inheritsthe Euclidean metric from ( T × S ) s , and the ε -neighborhood of K is denoted by N ε ( K ).Given a configuration K ∈ K , let τ min K be the shortest length of a line segment in T \ ∪ s i =1 B i which originates and terminates (possibly tangentially) in the set ∪ s i =1 ∂ B i , and let τ max K be In general, τ min K (cid:54) = min ≤ i Rules of the dynamics. Scatterers in source configuration K andtarget configuration K (cid:48) are drawn in dashed and solid line, respectively. A particleshoots off the boundary of a scatterer B i at the point q with unit velocity v andexits the gray buffer zone B i,β \ B i . Before it re-enters the buffer zone of anyscatterer B j , the configuration is switched instantaneously from K to K (cid:48) at sometime τ (cid:63) during mid-flight. The particle then hits the boundary of a scatterer B (cid:48) i (cid:48) elastically at the point q (cid:48) , resulting in post-collision velocity v (cid:48) .the supremum of the lengths of all line segments in the closure of T \ ∪ s i =1 B i which originateand terminate non-tangentially in the set ∪ s i =1 ∂ B i (this segment may meet the scatterers tan-gentially between its endpoints). As a function of K , τ min K is continuous, but τ max K in general isonly upper semi-continuous. Notice that 0 < τ min K < τ max K ≤ ∞ .A basic question is: Given K , K (cid:48) ∈ K , is there always a well-defined billiard map (analogousto classical billiard maps) with source configuration K and target configuration K (cid:48) ? That is tosay, if B , . . . , B s are the scatterers in configuration K , and B (cid:48) , . . . , B (cid:48) s are the correspondingscatterers in K (cid:48) , is there a well defined mapping F K (cid:48) , K : T +1 ( ∪ s i =1 ∂ B i ) → T +1 ( ∪ s i =1 ∂ B (cid:48) i )where T +1 ( ∪ s i =1 ∂ B i ) is the set of ( q, v ) such that q ∈ ∪ s i =1 ∂ B i and v is a unit vector at q pointinginto the region T \ ∪ s i =1 B i , and similarly for T +1 ( ∪ s i =1 ∂ B (cid:48) i )? Is the map F K (cid:48) , K uniquely defined,or does it depend on when the changeover from K to K (cid:48) occurs? The answer can be verygeneral, but let us confine ourselves to the special case where K (cid:48) is very close to K and thechangeover occurs when the particle is in “mid-flight” (to avoid having scatterers land on topof the particle, or meet it at the exact moment of the changeover).To do this systematically, we introduce the idea of a buffer zone. For β > 0, we let B i,β ⊂ T denote the β -neighborhood of B i , and define τ esc β , the escape time from the β -neighborhood of ∪ i B i , to be the maximum length of a line in ∪ s i =1 ( B i,β \ B i ) connecting ∪ s i =1 ∂ B i to ∪ s i =1 ∂ ( B i,β ).We then fix a value of β > τ esc β < τ min K − β , and require that B (cid:48) i ⊂ B i,β for each i = 1 , . . . , s . Notice that β < τ esc β , so that β < τ min K / 2, implying in particular that theneighborhoods B i,β are pairwise disjoint. For a particle starting from ∪ s i =1 ∂ B i , its trajectory isguaranteed to be outside of ∪ s i =1 B i,β during the time interval ( τ esc β , τ min K − β ): reaching ∪ s i =1 B i,β before time τ min K − β would contradict the definition of τ min K . We permit the configuration changeto take place at any time τ (cid:63) ∈ ( τ esc β , τ min K − β ). Notice that τ esc β depends only on the shapesof the scatterers, not their configuration, and that the billiard trajectory starting from ∪ i ∂ B i and ending in ∪ i ∂ B (cid:48) i does not depend on the precise moment τ (cid:63) at which the configurationis updated. For the billiard map F K (cid:48) , K to be defined, every particle trajectory starting from ∪ s i =1 ∂ B i must meet a scatterer in K (cid:48) . This is guaranteed by K (cid:48) ∈ K , due to the requirementthat any half-line intersects a scatterer boundary.To summarize, we have argued that given K , K (cid:48) ∈ K , there is a canonical way to define F K (cid:48) , K if B (cid:48) i ⊂ B i,β for all i where β = β ( τ min K ) > τ min K (and the curvatures ofthe B i ), and the flight time τ K (cid:48) , K satisfies τ K (cid:48) , K ≥ τ min K − β ≥ τ min K / F K (cid:48) , K operate on a single phase space M , so that our time-dependent billiard system defined by compositions of these maps can be studied in a ISPERSING BILLIARDS WITH MOVING SCATTERERS 5 ( q , v )( q , v ) Figure 2. Action of the map F K (cid:48) , K . With the same conventions as in Figure 1,the point in M corresponding to the plane vector ( q (cid:48) , v (cid:48) ) has more than onepreimage, whereas the point corresponding to ( q (cid:48)(cid:48) , v (cid:48)(cid:48) ) has no preimage at all.way analogous to iterated classical billiard maps. As usual, we let Γ i be a fixed clockwiseparametrization by arclength of ∂B i , and let M = ∪ i M i with M i = Γ i × [ − π/ , π/ . Recall that each K ∈ K is defined by an isometric embedding of ∪ s i =1 B i into T . This embeddingextends to a neighborhood of ∪ s i =1 B i ⊂ R , inducing a diffeomorphism Φ K : M → T +1 ( ∪ s i =1 ∂ B i ).For K , K (cid:48) for which F K (cid:48) , K is defined then, we have F K (cid:48) , K := Φ − K (cid:48) ◦ F K (cid:48) , K ◦ Φ K : M → M . Furthermore, given a sequence ( K n ) Nn =0 of configurations, we let F n = F K n , K n − assuming thismapping is well defined, and write F n + m,n = F n + m ◦ · · · ◦ F n and F n = F n ◦ · · · ◦ F for all n, m with 1 ≤ n ≤ n + m ≤ N .It is easy to believe — and we will confirm mathematically — that F K (cid:48) , K has many of theproperties of the section map of the 2D periodic Lorentz gas. The following differences, however,are of note: unlike classical billiard maps, F K (cid:48) , K is in general neither one-to-one nor onto , andas a result of that it also does not preserve the usual measure on M . This is illustrated inFigure 2.2.2. Main results. First we introduce the following uniform finite-horizon condition : For t , ϕ > ϕ small, we say K ∈ K has ( t , ϕ )-horizon if every directed open line segment in T oflength t meets a scatterer B i of K at an angle > ϕ (measured from its tangent line), with thesegment approaching this point of contact from T \ B i . Other intersection points between ourline segment and ∪ j ∂ B j are permitted and no requirements are placed on the angles at whichthey meet; we require only that there be at least one intersection point meeting the conditionabove. Notice that this condition is not affected by the sudden appearance or disappearanceof nearly tangential collisions of billiard trajectories with scatterers as the positions of thescatterers are shifted.The space in which we will permit our time-dependent configurations to wander is definedas follows: We fix 0 < ¯ τ min < t < ∞ and ϕ > 0, chosen so that the set K = K (¯ τ min , ( t , ϕ )) = {K ∈ K : ¯ τ min < τ min K and K has ( t , ϕ ) − horizon } is nonempty. Clearly, K is an open set, and its closure ¯ K as a subset of ( T × S ) s consists ofthose configurations whose τ min will be ≥ ¯ τ min , and line segments of length t with their endpoints added will meet scatterers with angles ≥ ϕ . From Section 2.1, we know that there exists¯ β = β (¯ τ min ) > F K (cid:48) , K is defined for all K , K (cid:48) ∈ K with B (cid:48) i ⊂ B i, ¯ β for all i where { B i } and { B (cid:48) i } are the scatterers in K and K (cid:48) respectively. For simplicity, we will call the pair( K , K (cid:48) ) admissible (with respect to K ) if they satisfy the condition above. Clearly, if K , K (cid:48) ∈ K MIKKO STENLUND, LAI-SANG YOUNG, AND HONGKUN ZHANG are such that d ( K , K (cid:48) ) < ε for small enough ε , then the pair is admissible. We also noted inSection 2.1 that for all admissible pairs,¯ τ min / ≤ τ K (cid:48) , K ≤ t . (1)We will denote by | f | γ the H¨older constant of a γ -H¨older continuous f : M → R .Our main result is Theorem 1. Given K = K (¯ τ min , ( t , ϕ )) , there exists ε > such that the following holds. Let µ and µ be probability measures on M , with strictly positive, -H¨older continuous densities ρ and ρ with respect to the measure cos ϕ d r d ϕ . Given γ > , there exist < θ γ < and C γ > such that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) M f ◦ F n d µ − (cid:90) M f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ C γ ( (cid:107) f (cid:107) ∞ + | f | γ ) θ nγ , n ≤ N, for all finite or infinite sequences ( K n ) Nn =0 ⊂ K ( N ∈ N ∪ {∞} ) satisfying d ( K n − , K n ) < ε for ≤ n ≤ N , and all γ -H¨older continuous f : M → R . The constant C γ = C γ ( ρ , ρ ) dependson the densities ρ i through the H¨older constants of log ρ i , while θ γ does not depend on the µ i .Both constants depend on K and ε . None of the constants in the theorem depends on N . We have included the finite N caseto stress that our results do not depend on knowledge of scatterer movements in the infinitefuture; requiring such knowledge would be unereasonable for time-dependent systems. Thenotation “( K n ) Nn =0 , N ∈ N ∪ {∞} ” is intended as shorthand for K , . . . , K N for N < ∞ , and K , K , . . . (infinite sequence) for N = ∞ .Our next result is an extension of Theorem 1 to a situation where the geometries of thescatterers are also allowed to vary with time. We use κ to denote the curvature of the scatterers,and use the convention that κ > < ¯ κ min < ¯ κ max < ∞ , 0 < ¯ τ min < t < ∞ , ϕ > < ∆ < ∞ , we let (cid:101) K = (cid:101) K (¯ κ min , ¯ κ max ; ¯ τ min , ( t , ϕ ); ∆)denote the set of configurations K = (cid:0) ( B , o ) , . . . , ( B s , o s ) (cid:1) where ( B , . . . , B s ) is an orderedset of disjoint scatterers on T , o i ∈ ∂ B i is a marked point for each i , s ∈ N is arbitrary, andthe following conditions are satisfied:(i) the scatterer boundaries ∂ B i are C with (cid:107) D ( ∂ B i ) (cid:107) C < ∆ and Lip( D ( ∂ B i )) < ∆,(ii) the curvatures of ∂ B i lie between ¯ κ min and ¯ κ max , and(iii) τ min K > ¯ τ min , and K has ( t , ϕ )-horizon.In (i), (cid:107) D ( ∂ B i ) (cid:107) C and Lip( D ( ∂ B i )) are defined to be max ≤ k ≤ (cid:107) D k γ i (cid:107) ∞ and Lip( D γ i ),respectively, where γ i is the unit speed clockwise parametrization of B i . For two configu-rations K = (( B , o ) , . . . , ( B s , o s )) and K (cid:48) = (( B (cid:48) , o (cid:48) ) , . . . , ( B (cid:48) s , o (cid:48) s )) with the same numberof scatterers, we define d ( K , K (cid:48) ) to be the maximum of max i ≤ s sup x ∈M d M (ˆ γ i ( x ) , ˆ γ (cid:48) i ( x )) andmax i ≤ s max ≤ k ≤ (cid:107) D k ˆ γ i − D k ˆ γ (cid:48) i (cid:107) ∞ where ˆ γ i : S → T denotes the constant speed clockwiseparametrization of ∂ B i with ˆ γ i (0) = o i , ˆ γ (cid:48) i is the corresponding parametrization of ∂ B (cid:48) i withˆ γ (cid:48) i (0) = o (cid:48) i , and d M is the natural distance on M . The definition of admissibility for K and K (cid:48) is as above, and the billiard map F K (cid:48) , K is defined as before for admissible pairs. Configura-tions K , K (cid:48) with different numbers of scatterers are not admissible, and the distance betweenthem is set arbitrarily to be d ( K , K (cid:48) ) = 1. Theorem 2. The statement of Theorem 1 holds verbatim with ( K , d ) replaced by ( (cid:101) K , d ) . The differentiability assumption on the scatterer boundaries can be relaxed, but the pursuit of minimaltechnical conditions is not the goal of our paper. ISPERSING BILLIARDS WITH MOVING SCATTERERS 7 Theorems 1’ and 2’: The regularity assumption on the measures µ i in Theorems 1–2 abovecan be much relaxed. It suffices to assume that the µ i have regular conditional measures onunstable curves; they can be singular in the transverse direction and can, e.g., be supportedon a single unstable curve. Convex combinations of such measures are also admissible. Preciseconditions are given in Section 4, after we have introduced the relevant technical definitions.Theorems 1’–2’, which are the extensions of Theorems 1–2 respectively to the case where theserelaxed conditions on µ i are permitted, are stated in Section 4.4.Theorems 2 and 2’ obviously apply as a special case to classical billiards, giving uniformbounds of the kind above for all F K , K , K ∈ (cid:101) K . It is also a standard fact that correlation decayresults can be deduced from the type of convergence in Theorems 1–2. To our knowledge, thefollowing result on correlation decay for classical billiards is new. (See also pp. 149–150 in [9]for related observations.) The proof can be found in Section 8.2. Theorem 3. Let µ denote the measure obtained by normalizing cos ϕ d r d ϕ to a probabilitymeasure. Let (cid:101) K be fixed, and let γ > be arbitrary. Then for any γ -H¨older continuous f andany -H¨older continuous g , there exists a constant C (cid:48) γ such that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n · g d µ − (cid:90) f d µ (cid:90) g d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ C (cid:48) γ θ nγ hold for all n ≥ and for all F = F K , K with K ∈ (cid:101) K . Here θ γ is as in the theorems above. Theconstant C (cid:48) γ depends on (cid:107) f (cid:107) ∞ , | f | γ , (cid:107) g (cid:107) ∞ and | g | . We remark that Theorem 3 can also be formulated for sequences of maps. In that case thequantity bounded is (cid:82) f ◦ F n · g d µ − (cid:82) f ◦ F n d µ (cid:82) g d µ and µ is an arbitrary measure satisfyingthe conditions in Theorems 1’ and 2’. The proof is unchanged.In addition to the broader class of measures, Theorem 2 could be extended to less regularobservables f , which would allow for a corresponding generalization of Theorem 3. In particular,the observables could be allowed to have discontinuities at the singularities of the map F ; see,e.g., [10]. In order to keep the focus on what is new, we do not pursue that here.We state one further extension of the above theorems, to include the situation where the testparticle is also under the influence of an external field. Given an admissible pair ( K , K (cid:48) ) in (cid:101) K and a vector field E = E ( q , v ), we define first a continuous time system in which the trajectoryof the test particle between collisions is determined by the equations˙ q = v and ˙ v = E , where q is the position and v the velocity of the particle, together with the initial condition.For the sake of simplicity, let us assume that the field is isokinetic — that is, v · E = 0 —which allows to normalize | v | = 1. This class of forced billiards includes “electric fields withGaussian thermostats” studied in [9, 14] and many other papers. (Instead of the speed, moregeneral integrals of the motion could be considered, allowing for other types of fields, such asgradients of weak potentials; see [5, 7].) Assuming that the field E is smooth and small, thetrajectories are almost linear, and a billiard map F E K (cid:48) , K : M → M can be defined exactly asbefore. (See Section 8.3 for more details.) Note that F K (cid:48) , K = F K (cid:48) , K .The setup for our time-dependent systems result is as follows: We consider the space (cid:101) K × E where (cid:101) K is as above and E = E ( ε E ) for some ε E > E ∈ C with (cid:107) E (cid:107) C =max ≤ k ≤ (cid:107) D k E (cid:107) ∞ < ε E . In the theorem below, it is to be understood that F n = F E n K n , K n − and F n = F n ◦ · · · ◦ F . Theorem E. Given (cid:101) K , there exist ε > and ε E > such that the statement of Theorem 2holds for all sequences (( K n , E n )) n ≤ N in (cid:101) K × E ( ε E ) satisfying d ( K , K (cid:48) ) < ε for all n ≤ N . MIKKO STENLUND, LAI-SANG YOUNG, AND HONGKUN ZHANG Like the zero-field case, Theorem E also admits a generalization of measures (and observables)and also implies an exponential correlation bound.2.3. Main technical result. To prove Theorem 1, we will, in fact, prove the following tech-nical result. All configurations below are in K . Let ( (cid:101) K q ) Qq =1 ( Q ∈ Z + arbitrary) be a sequenceof configurations, (˜ ε q ) Qq =1 a sequence of positive numbers, and ( (cid:101) N q ) Qq =1 a sequence of positiveintegers. We say the configuration sequence ( K n ) Nn =0 (arbitrary N ) is adapted to ( (cid:101) K q , ˜ ε q , (cid:101) N q ) Qq =1 if there exist numbers 0 = n < n < · · · < n Q = N such that for 1 ≤ q ≤ Q , we have n q − n q − ≥ (cid:101) N q and K n ∈ N ˜ ε q ( (cid:101) K q ) for n q − ≤ n ≤ n q . That is to say, we think of the ( (cid:101) K q ) Qq =1 asreference configurations, and view the sequence of interest, ( K n ) Nn =0 , as going from one referenceconfiguration to the next, spending a long time ( ≥ (cid:101) N q ) near (within ˜ ε q of) each (cid:101) K q . Theorem 4. For any K ∈ K , there exist (cid:101) N ( K ) ≥ and ˜ ε ( K ) > such that the followingholds for every sequence of reference configurations ( (cid:101) K q ) Qq =1 ( Q < ∞ ) with (cid:101) K q +1 ∈ N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ) for ≤ q < Q and every sequence ( K n ) Nn =0 adapted to ( (cid:101) K q , ˜ ε ( (cid:101) K q ) , (cid:101) N ( (cid:101) K q )) Qq =1 , all configurationsto be taken in K : Let µ and µ be probability measures on M , with strictly positive, -H¨oldercontinuous densities ρ and ρ with respect to the measure cos ϕ d r d ϕ . Given any γ > , thereexist < θ γ < and C γ > such that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) M f ◦ F n d µ − (cid:90) M f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ C γ ( (cid:107) f (cid:107) ∞ + | f | γ ) θ nγ , n ≤ N, (2) for all γ -H¨older continuous f : M → R . The constants C γ and θ γ depend on the collection { (cid:101) K q , ≤ q ≤ Q } (see Remark 5 below) ; additionally C γ = C γ ( ρ , ρ ) depends on the densities ρ i through the H¨older constants of log ρ i , while θ γ does not depend on the µ i . Remark 5. We clarify that the constants C γ and θ γ depend on the collection of distinct con-figurations that appear in the sequence ( (cid:101) K q ) Qq =1 , not on the order in which these configurationsare listed; in particular, each (cid:101) K q may appear multiple times. This observation is essential forthe proofs of Theorems 1–3.Proof of Theorem 1 assuming Theorem 4. Given K , consider a slightly larger K (cid:48) ⊃ ¯ K , obtainedby decreasing ¯ τ min and ϕ and increasing t . We apply Theorem 4 to K (cid:48) , obtaining ˜ ε ( K ) and (cid:101) N ( K )for K ∈ K (cid:48) . Since ¯ K is compact, there exists a finite collection of configurations ( (cid:101) K q ) q ∈Q ⊂ ¯ K such that the sets (cid:101) N q = N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ) ∩ K , q ∈ Q , form a cover of K . Let ε (cid:63) = min q ∈Q ˜ ε ( (cid:101) K q ) and N (cid:63) = max q ∈Q (cid:101) N ( (cid:101) K q ). We claim that Theorem 1 holds with ε = ε (cid:63) / (2 N (cid:63) ). Let ( K n ) Nn =0 ⊂ K with d ( K n , K n +1 ) < ε be given. Suppose K ∈ (cid:101) N q . Then K i is guaranteed to be in N ˜ ε ( (cid:101) K q ) ( (cid:101) K q )for all i < N (cid:63) . Before the sequence leaves N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ), we select another (cid:101) N q (cid:48) and repeat theprocess. Thus, the assumptions of Theorem 4 are satisfied (add more copies of K N at theend if necessary). Taking note of Remark 5, this yields a uniform rate of memory loss for allsequences. Of course the constants thus obtained for K in Theorem 1 are the constants aboveobtained for the larger K (cid:48) in Theorem 4. (cid:3) Standing Hypothesis for Sections 3–8.1: We assume K as defined by ¯ τ min , t and ϕ isfixed throughout. For definiteness we fix also ¯ β , and declare once and for all that all pairs( K , K (cid:48) ) for which we consider the billiard map F K (cid:48) , K are assumed to be admissible, as are( K n , K n +1 ) in all the sequences ( K n ) studied. These are the only billiard maps we will consider.3. Preliminaries I: Geometry of billiard maps In this section, we record some basic facts about time-dependent billiard maps related totheir hyperbolicity, discontinuities, etc. The results here are entirely analogous to the fixedscatterers case. They depend on certain geometric facts that are uniform for all the billiard ISPERSING BILLIARDS WITH MOVING SCATTERERS 9 maps considered; indeed one does not know from step to step in the proofs whether or notthe source and target configurations are different. Thus we will state the facts but not givethe proofs, referring the reader instead to sources where proofs are easily modified to give theresults here.An important point is that the estimates of this section are uniform , i.e., the constants inthe statements of the lemmas depend only on K . Notation: Throughout the paper, the length of a smooth curve W ⊂ M is denoted by | W | andthe Riemannian measure induced on W is denoted by m W . Thus, m W ( W ) = | W | . We denoteby U ε ( E ) the open ε -neighborhood of a set E in the phase space M . For x = ( r, ϕ ) ∈ M , wedenote by T x M the tangent space of M and by D x F the derivative of a map F at x . Whereno ambiguity exists, we sometimes write F instead of F K (cid:48) , K .3.1. Hyperbolicity. Given ( K , K (cid:48) ) and a point x = ( r, ϕ ) ∈ M , we let x (cid:48) = ( r (cid:48) , ϕ (cid:48) ) = F x andcompute D x F as follows: Let κ ( x ) denote the curvature of ∪ i ∂B i at the point correspondingto x , and define κ ( x (cid:48) ) analogously. The flight time between x and x (cid:48) is denoted by τ ( x ) = τ K (cid:48) , K ( x ). Then D x F is given by − ϕ (cid:48) (cid:18) τ ( x ) κ ( x ) + cos ϕ τ ( x ) τ ( x ) κ ( x ) κ ( x (cid:48) ) + κ ( x ) cos ϕ (cid:48) + κ ( x (cid:48) ) cos ϕ τ ( x ) κ ( x (cid:48) ) + cos ϕ (cid:48) (cid:19) provided x / ∈ F − ∂ M , the discontinuity set of F . This computation is identical to the casewith fixed scatterers. As in the fixed scatterers case, notice that as x approaches F − ∂ M ,cos ϕ (cid:48) → F blows up. Notice also thatdet D x F = cos ϕ/ cos ϕ (cid:48) , (3)so that F is locally invertible.The next result asserts the uniform hyperbolicity of F for orbits that do not meet F − ∂ M .Let κ min and κ max denote the minimum and maximum curvature of the boundaries of thescatterers B i . Lemma 6 (Invariant cones) . The unstable cones C ux = { (d r, d ϕ ) ∈ T x M : κ min ≤ d ϕ/ d r ≤ κ max + 2 / ¯ τ min } , x ∈ M , are D x F -invariant for all pairs ( K , K (cid:48) ) , i.e., D x F ( C ux ) ⊂ C uF x for all x / ∈ F − ∂ M , and thereexist uniform constants ˆ c > and Λ > such that for every ( K n ) Nn =0 , (cid:107) D x F n v (cid:107) ≥ ˆ c Λ n (cid:107) v (cid:107) (4) for all n ∈ { , . . . , N } , v ∈ C ux , and x / ∈ ∪ Nm =1 ( F m ) − ∂ M .Similarly, the stable cones C sx = { (d r, d ϕ ) ∈ T x M : − κ max − / ¯ τ min ≤ d ϕ/ d r ≤ − κ min } are ( D x F ) − -invariant for all ( K , K (cid:48) ) , i.e., ( D x F ) − C sF x ⊂ C sx for all x / ∈ ∂ M ∪ F − ∂ M , andfor every ( K n ) Nn =0 , (cid:107) ( D x F n ) − v (cid:107) ≥ ˆ c Λ n (cid:107) v (cid:107) for all n ∈ { , . . . , N } , v ∈ C s F n x , and x / ∈ ∂ M ∪ ∪ Nm =1 ( F m ) − ∂ M . Notice that the cones here can be chosen independently of x and of the scatterer configura-tions involved. The proof follows verbatim that of the fixed scatterers case; see [10].Following convention, we introduce for purposes of controlling distortion (see Lemma 9) thehomogeneity strips H k = { ( r, ϕ ) ∈ M : π/ − k − < ϕ ≤ π/ − ( k + 1) − } H − k = { ( r, ϕ ) ∈ M : − π/ k + 1) − ≤ ϕ < − π/ k − } for all integers k ≥ k , where k is a sufficiently large uniform constant. It follows, for example,that for each k , D x F is uniformly bounded for x ∈ F − ( H − k ∪ H k ), as C − k − ≤ cos ϕ (cid:48) ≤ C cos k − (5)for a constant C cos > 0. We will also use the notation H = { ( r, ϕ ) ∈ M : − π/ k ≤ ϕ ≤ π/ − k − } . Discontinuity sets and homogeneous components. For each ( K , K (cid:48) ), the singularityset ( F K (cid:48) , K ) − ∂ M has similar geometry as in the case K (cid:48) = K . In particular, it is the unionof finitely many C -smooth curves which are negatively sloped, and there are uniform boundsdepending only on K for the number of smooth segments (as follows from (1)) and their deriva-tives. One of the geometric facts, true for fixed scatterers as for the time-dependent case, thatwill be useful later is the following: Through every point in F − ∂ M , there is a continuous pathin F − ∂ M that goes monotonically in ϕ from one component of ∂ M to the other.In our proofs it will be necessary to know that the structure of the singularity set varies ina controlled way with changing configurations. Let us denote S K (cid:48) , K = ∂ M ∪ ( F K (cid:48) , K ) − ∂ M . If K and K (cid:48) are small perturbations of (cid:101) K , then S K (cid:48) , K is contained in a small neighborhood of S (cid:101) K , (cid:101) K (albeit the topology of S K (cid:48) , K may be slightly different from that of S (cid:101) K , (cid:101) K ). A proof of thefollowing result, which suffices for our purposes, is given in the Appendix. Lemma 7. Given a configuration (cid:101) K ∈ K and a compact subset E ⊂ M \ S (cid:101) K , (cid:101) K , there exists δ > such that the map ( x, K , K (cid:48) ) (cid:55)→ F K (cid:48) , K ( x ) is uniformly continuous on E × N δ ( (cid:101) K ) × N δ ( (cid:101) K ) . While F − ∂ M is the genuine discontinuity set for F , for purposes of distortion control oneoften treats the preimages of homogeneity lines as though they were discontinuity curves also.We introduce the following language: A set E ⊂ M is said to be homogeneous if it is completelycontained in a connected component of one of the H k , | k | ≥ k or k = 0. Let E ⊂ M be ahomogeneous set. Then F ( E ) may have more than one connected component. We furthersubdivide each connected component into maximal homogeneous subsets and call these the homogeneous components of F ( E ). For n ≥ 2, the homogeneous components of F n ( E ) aredefined inductively: Suppose E n − ,i , i ∈ I n − , are the homogeneous components of F n − ( E ),for some index set I n − which is at most countable. For each i ∈ I n − , the set E n − ,i is ahomogeneous set, and we can thus define the homogeneous components of the single-step image F n ( E n − ,i ) as above. The subsets so obtained, for all i ∈ I n − , are the homogeneous componentsof F n ( E ). Let E − n,i = E ∩ F − n ( E n,i ). We call { E − n,i } i the canonical n -step subdivision of E ,leaving the dependence on the sequence implicit when there is no ambiguity.For x, y ∈ M , we define the separation time s ( x, y ) to be the smallest n ≥ F n x and F n y belong in different strips H k or in different connected components of M . Observe thatthis definition is ( K n )-dependent.3.3. Unstable curves. A connected C -smooth curve W ⊂ M is called an unstable curve if T x W ⊂ C ux for every x ∈ W . It follows from the invariant cones condition that the image of anunstable curve under F n is a union of unstable curves. Our unstable curves will be parametrizedby r : for a curve W , we write ϕ = ϕ W ( r ).For an unstable curve W , define ˆ κ W = sup W | d ϕ W / d r | . Lemma 8. There exist uniform constants C c > and ϑ c ∈ (0 , such that the following holds.Let W and F n W be unstable curves. Then ˆ κ F n W ≤ C c ϑ n c ˆ κ W ) . ISPERSING BILLIARDS WITH MOVING SCATTERERS 11 We call an unstable curve W regular if it is homogeneous and satisfies the curvature boundˆ κ W ≤ C c . Thus for any unstable curve W , all homogeneous components of F n ( W ) are regularfor large enough n .Given a smooth curve W ⊂ M , define J W F n ( x ) = (cid:107) D x F n v (cid:107) / (cid:107) v (cid:107) for any nonzero vector v ∈ T x W . In other words, J W F n is the Jacobian of the restriction F n | W . Lemma 9 (Distortion bound) . There exist uniform constants C (cid:48) d > and C d > such thatthe following holds. Given ( K n ) Nn =0 , if F n W is a homogeneous unstable curve for ≤ n ≤ N ,then C − ≤ e − C (cid:48) d |F n W | / ≤ J W F n ( x ) J W F n ( y ) ≤ e C (cid:48) d |F n W | / ≤ C d (6) for every pair x, y ∈ W and ≤ n ≤ N . Finally, we state a result which asserts that very short homogeneous curves cannot acquirelengths of order one arbitrarily fast, in spite of the fact that the local expansion factor isunbounded. Lemma 10. There exists a uniform constant C e ≥ such that |F n W | ≤ C e | W | / n , if W is an unstable curve and F m W is homogeneous for ≤ m < n . The proofs of these results also follow closely those for the fixed scatterers case. For Lemma 8,see [8]. For Lemmas 9 and 10, see [10]. (Lemma 10 follows readily by iterating the correspondingone-step bound.)3.4. Local stable manifolds. Given ( K n ) n ≥ , a connected smooth curve W is called a ho-mogeneous local stable manifold , or simply local stable manifold , if the following hold for every n ≥ F n W is connected and homogeneous, and(ii) T x ( F n W ) ⊂ C sx for every x ∈ F n W .It follows from Lemma 6 that local stable manifolds are exponentially contracted under F n .We stress that unlike unstable curves, the definition of local stable manifolds depends stronglyon the infinite sequence of billiard maps defined by ( K n ) n ≥ .For x ∈ M , let W s ( x ) denote the maximal local stable manifold through x if one exists. Animportant result is the absolute continuity of local stable manifolds . Let two unstable curves W and W be given. Denote W i(cid:63) = { x ∈ W i : W s ( x ) ∩ W − i (cid:54) = 0 } for i = 1 , 2. The map h : W (cid:63) → W (cid:63) such that { h ( x ) } = W s ( x ) ∩ W for every x ∈ W (cid:63) is called the holonomymap. The Jacobian J h of the holonomy is the Radon–Nikodym derivative of the pullback h − ( m W | W ∗ ) with respect to m W . The following result gives a uniform bound on the Jacobianalmost everywhere on W (cid:63) . Lemma 11. Let W and W be regular unstable curves. Suppose h : W (cid:63) → W (cid:63) is defined ona positive m W -measure set W (cid:63) ⊂ W . Then for m W -almost every point x ∈ W (cid:63) , J h ( x ) = lim n →∞ J W F n ( x ) J W F n ( h ( x )) , (7) where the limit exists and is positive with uniform bounds. In fact, there exist uniform constants A h > and C h > such that the following holds: If α ( x ) denotes the difference between theslope of the tangent vector of W at x and that of W at h ( x ) , and if δ ( x ) is the distancebetween x and h ( x ) , then A − α − δ / h ≤ J h ≤ A α + δ / h (8) almost everywhere on W (cid:63) . Moreover, with θ = Λ − / ∈ (0 , , |J h ( x ) − J h ( y ) | ≤ C h θ s ( x,y ) (9) holds for all pairs ( x, y ) in W (cid:63) , where s ( x, y ) is the separation time defined in Section 3.2. The proof of Lemma 11 follows closely its counterpart for fixed configurations. The identityin (7) is standard for uniformly hyperbolic systems (see [1, 17]), as is (9), except for the use ofseparation time as a measure of distance in discontinuous systems; see [10, 20].4. Preliminaries II: Evolution of measured unstable curves Growth of unstable curves. Given a sequence ( K m ), an unstable curve W , a point x ∈ W , and an integer n ≥ 0, we denote by r W,n ( x ) the distance between F n x and theboundary of the homogeneous component of F n W containing F n x .The following result, known as the Growth lemma, is key in the analysis of billiard dy-namics. It expresses the fact that the expansion of unstable curves dominates the cutting by ∂ M ∪ ∪ | k |≥ k ∂ H k , in a uniform fashion for all sequences. The reason behind this fact is thatunstable curves expand at a uniform exponential rate, whereas the cuts accumulate at tan-gential collisions. A short unstable curve can meet no more than t / ¯ τ min tangencies in a singletime step (see (1)), so the number of encountered tangencies grows polynomially with timeuntil a characteristic length has been reached. The proof follows verbatim that in the fixedconfiguration case; see [10]. Lemma 12 (Growth lemma) . There exist uniform constants C gr > and ϑ ∈ (0 , such that,for all (finite or infinite) sequences ( K n ) Nn =0 , unstable curves W and ≤ n ≤ N : m W { r W,n < ε } ≤ C gr ( ϑ n + | W | ) ε . This lemma has the following interpretation: It gives no information for small n when | W | issmall. For n large enough, such as n ≥ | log | W || / | log ϑ | , one has m W { r W,n < ε } ≤ C gr ε | W | .In other words, after a sufficiently long time n (depending on the initial curve W ), the majority of points in W have their images in homogeneous components of F n W that are longer than1 / (2 C gr ), and the family of points belonging to shorter ones has a linearly decreasing tail.4.2. Measured unstable curves. A measured unstable curve is a pair ( W, ν ) where W is anunstable curve and ν is a finite Borel measure supported on it. Given a sequence ( K n ) ∞ n =0 anda measured unstable curve ( W, ν ) with density ρ = d ν/ d m W , we are interested in the followingdynamical H¨older condition of log ρ : For n ≥ 1, let { W − n,i } i be the canonical n -step subdivisionof W as defined in Section 3.2. Lemma 13. There exists a constant C (cid:48) r > for which the following holds: Suppose ρ is adensity on an unstable curve W satisfying | log ρ ( x ) − log ρ ( y ) | ≤ Cθ s ( x,y ) for all x, y ∈ W .Then, for any homogeneous component W n,i , the density ρ n,i of the push-forward of ν | W n,i bythe (invertible) map F n | W − n,i satisfies | log ρ n,i ( x ) − log ρ n,i ( y ) | ≤ (cid:18) C (cid:48) r (cid:16) C − C (cid:48) r (cid:17) θ n (cid:19) θ s ( x,y ) (10) for all x, y ∈ W n,i . Here θ is as in Lemma 11. We fix C r ≥ max { C (cid:48) r , C h , } , where C h is also introduced inLemma 11, and say a measure ν supported on an unstable curve W is regular if it is absolutelycontinuous with respect to m W and its density ρ satisfies | log ρ ( x ) − log ρ ( y ) | ≤ C r θ s ( x,y ) (11)for all x, y ∈ W . As with s ( · , · ), the regularity of ν is ( K n )-dependent. Notice that under thisdefinition, if a measure on W is regular, then so are its forward images. More precisely, in thenotation of Lemma 13, if ρ is regular, then so is each ρ n,i . We also say the pair ( W, ν ) is regular if both the unstable curve W and the measure ν are regular. ISPERSING BILLIARDS WITH MOVING SCATTERERS 13 Remark 14. The separation time s ( x, y ) is connected to the Euclidean distance d M ( x, y ) inthe following way. If x and y are connected by an unstable curve W , then |F n W | ≥ ˆ c Λ n | W | ≥ ˆ c Λ n d M ( x, y ) for ≤ n < s ( x, y ) . Since F s ( x,y ) − W is a homogeneous unstable curve, its lengthis uniformly bounded above. Thus, d M ( x, y ) ≤ C s Λ − s ( x,y ) = C s θ s ( x,y ) (12) for a uniform constant C s > . In particular, if ρ is a nonnegative density on an unstablecurve W such that log ρ is H¨older continuous with exponent / and constant C r C − / , then ρ is regular with respect to any configuration sequence .Proof of Lemma 13. Consider n = 1 and take two points x, y on one of the homogeneouscomponents W ,i . Let the corresponding preimages be x − , y − . Since s ( x − , y − ) = s ( x, y ) + 1,the bound (6) yields | log ρ ,i ( x ) − log ρ ,i ( y ) | ≤ (cid:12)(cid:12)(cid:12)(cid:12) log ρ ( x − ) J W F ( x − ) − log ρ ( y − ) J W F ( y − ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ | log ρ ( x − ) − log ρ ( y − ) | + | log J W F ( x − ) − log J W F ( y − ) |≤ Cθ s ( x,y )+1 + C (cid:48) d | W ,i ( x, y ) | / , where | W ,i ( x, y ) | is the length of the segment of the unstable curve W ,i connecting x and y .Because of the unstable cones, the latter is uniformly proportional to the distance d M ( x, y )of x and y . Recalling (12), we thus get C (cid:48) d | W ,i ( x, y ) | / ≤ C (cid:48)(cid:48) d θ s ( x,y ) ≤ C (cid:48)(cid:48) d θ s ( x,y ) for anotheruniform constant C (cid:48)(cid:48) d > 0. Let us now pick any C r such that C r ≥ C (cid:48)(cid:48) d / (1 − θ ). For then C r θ + C (cid:48)(cid:48) d ≤ θ C r , and we have | log ρ ,i ( x ) − log ρ ,i ( y ) | ≤ C (cid:48) θ s ( x,y ) with C (cid:48) ≤ Cθ + C (cid:48)(cid:48) d = ( C − C r ) θ + ( C r θ + C (cid:48)(cid:48) d ) ≤ ( C − C r ) θ + 1 + θ C r = Cθ + 1 − θ C r . (13)We may iterate (13) inductively, observing that at each step the constant C obtained at theprevious step is contracted by a factor of θ towards − θ C r . It is now a simple task to obtain (10).The constant C r was chosen so that the image of a regular density is regular and the image ofa non-regular density will become regular in finitely many steps. (cid:3) The following extension property of (11) will be necessary. We give a proof in the Appendix. Lemma 15. Suppose W (cid:63) is a closed subset of an unstable curve W , and that W (cid:63) includes theendpoints of W . Assume that the function ρ is defined on W (cid:63) and that there exists a constant C > such that | log ρ ( x ) − log ρ ( y ) | ≤ Cθ s ( x,y ) for every pair ( x, y ) in W (cid:63) . Then, ρ can beextended to all of W in such a way that the inequality involving log ρ above holds on W , theextension is piecewise constant, min W ρ = min W (cid:63) ρ , and max W ρ = max W (cid:63) ρ . Families of measured unstable curves. Here we extend the idea of measured unstablecurves to measured families of unstable stacks. We begin with the following definitions:(i) We call ∪ α ∈ E W α ⊂ M a stack of unstable curves , or simply an unstable stack , if E ⊂ R isan open interval, each W α is an unstable curve, and there is a map ψ : [0 , × E → M whichis a homeomorphism onto its image so that, for each α ∈ E , ψ (cid:0) [0 , × { α } (cid:1) = W α .(ii) The unstable stack ∪ α ∈ E W α is said to be regular if each W α is regular as an unstable curve.(iii) We call ( ∪ α ∈ E W α , µ ) a measured unstable stack if U = ∪ α ∈ E W α is an unstable stack and µ is a finite Borel measure on U .(iv) we say ( ∪ α ∈ E W α , µ ) is regular if (a) as a stack ∪ α ∈ E W α is regular and (b) the conditionalprobability measures µ α of µ on W α are regular. More precisely, { W α , α ∈ E } is a measurablepartition of ∪ α ∈ E W α , and { µ α } is a version of the disintegration of µ with respect to thispartition, that is to say, for any Borel set B ⊂ M , we have µ ( B ) = (cid:90) E µ α ( W α ∩ B ) d P ( α ) where P is a finite Borel measure on I . The conditional measures { µ α } are unique up to aset of P -measure 0, and (b) requires that ( W α , µ α ) be regular in the sense of Section 4.2 for P -a.e. α .Consider next a sequence ( K n ) ∞ n =0 and a fixed n ≥ 1. Denote by D n,i , i ≥ 1, the countablymany connected components of the set M \ ∪ ≤ m ≤ n ( F m ) − ( ∂ M ∪ ∪ | k |≥ k ∂ H k ). In analogy withunstable curves, we define the canonical n -step subdivision of a regular unstable stack ∪ α W α :Let ( n, i ) be such that ∪ α W α ∩ D n,i (cid:54) = ∅ , and let E n,i = { α ∈ E : W α ∩ D n,i (cid:54) = ∅} . Pick one ofthe (finitely our countably many) components E n,i,j of E n,i .We claim ∪ α ∈ E n,i,j ( W α ∩ D n,i ) is an unstable stack, and define ψ n,i,j : [0 , × E n,i,j → ∪ α ∈ E n,i,j W α ∩ D n,i as follows: for α ∈ E n,i , ψ n,i | [0 , ×{ α } is equal to ψ | [0 , ×{ α } followed by a linear contractionfrom W α to W α ∩ D n,i . For this construction to work, it is imperative that W α ∩ D n,i beconnected, and that is true, for by definition, W α ∩ D n,i is an element of the canonical n -stepsubdivision of W α . It is also clear that ∪ α ∈ E n,i,j F n ( W α ∩ D n,i ) is an unstable stack, with thedefining homeomorphism given by F n ◦ ψ n,i,j .What we have argued in the last paragraph is that the F n -image of an unstable stack ∪ α W α is the union of at most countably many unstable stacks. Similarly, the F n -image of a measuredunstable stack is the union of measured unstable stacks, and by Lemmas 8 and 13, regularmeasured unstable stacks are mapped to unions of regular measured unstable stacks. ψE (0 , Figure 3. A schematic illustration of an unstable stack and its dynamics. Theregular unstable curves on the right are the images of the horizontal lines underthe homeomorphism ψ . The curves with negative slopes represent the countablymany branches of the n -step singularity set ∪ ≤ m ≤ n ( F m ) − ( ∂ M ∪ ∪ | k |≥ k ∂ H k ).The canonical n -step subdivision of the unstable curves yields countably manyunstable stacks.The discussion above motivates the definition of measured unstable families , defined to beconvex combinations of measured unstable stacks. That is to say, we have a countable collectionof unstable stacks ∪ α ∈ E j W ( j ) α parametrized by j , and a measure µ = (cid:80) j a ( j ) µ ( j ) with theproperty that for each j , ( ∪ α W ( j ) α , µ ( j ) ) is a measured unstable stack and (cid:80) j a ( j ) = 1. Wepermit the stacks to overlap, i.e., for j (cid:54) = j (cid:48) , we permit ∪ α W ( j ) α and ∪ α W ( j (cid:48) ) α to meet. This isnatural because in the case of moving scatterers, the maps F n are not one-to-one; even if twostacks have disjoint supports, this property is not retained by the forward images. Regularityfor measured unstable families is defined similarly. The idea of canonical n -step subdivisionpasses easily to measured unstable families, and we can sum up the discussion by saying thatgiven ( K n ), push-forwards of measured unstable families are again measured unstable families,and regularity is preserved.So far, we have not discussed the lengths of the unstable curves in an unstable stack orfamily. Following [10], we introduce, for a measured unstable family defined by ( ∪ α W ( j ) α , µ ( j ) ) ISPERSING BILLIARDS WITH MOVING SCATTERERS 15 and µ = (cid:80) j a ( j ) µ ( j ) , the quantity Z = (cid:88) j a ( j ) (cid:90) R | W ( j ) α | d P ( j ) ( α ) . (14)Informally, the smaller the value of Z /µ ( M ) the smaller the fraction of µ supported on shortunstable curves.For µ as above, let Z n denote the quantity corresponding to Z for the push-forward (cid:80) k a n,k µ n,k of the canonical n -step subdivision of µ discussed earlier. We have the following control on Z n : Lemma 16. There exist uniform constants C p > and ϑ p ∈ (0 , such that Z n µ ( M ) ≤ C p (cid:18) ϑ n p Z µ ( M ) (cid:19) (15) holds true for any regular measured unstable family. This result can be interpreted as saying that given an initial measure µ which has a highfraction of its mass supported on short unstable curves — yielding a large value of Z /µ ( M ) —the mass gets quickly redistributed by the dynamics to the longer homogeneous componentsof the image measures, so that Z n /µ ( M ) decreases exponentially, until a level safely below C p is reached. As regular densities remain regular (Lemma 13), and as the supremum and theinfimum of a regular density are uniformly proportional, Lemma 16 is a direct consequence ofthe Growth lemma; see above. See [10] for the fixed configuration case; the time-dependentcase is analogous. Definition 17. A regular measured unstable family is called proper if Z < C p µ ( M ) . Notice that Z < C p µ ( M ) implies, by Markov’s inequality and (cid:80) j a ( j ) P ( j ) ( R ) = µ ( M ), that (cid:88) j a ( j ) P ( j ) (cid:8) α ∈ R : | W ( j ) α | ≥ (2 C p ) − (cid:9) ≥ µ ( M ) . (16)In other words, if µ is proper, then at least half of it is supported on unstable curves oflength ≥ (2 C p ) − . Notice also that for any measured unstable family with Z < ∞ , Lemma 16shows that the push-forward of such a family will eventually become proper. Starting froma proper family, it is possible that Z n ≥ C p µ ( M ) for a finite number of steps; however, (15)implies that there exists a uniform constant n p such that Z n < C p µ ( M ) for all n ≥ n p . Remark 18. The results of Sections 4.1–4.3 can be summarized as follows: (i) The F n -image of an unstable stack is the union of at most countably many such stacks.(ii) Regular measured unstable stacks are mapped to unions of the same, and(iii) the F n -image of a proper measured unstable family is proper for n ≥ n p . Statements of Theorems 1’–2’ and 4’. We are finally ready to give a precise statementof Theorem 1’, which permits more general initial distributions than Theorem 1 as stated inSection 2.2. Theorem 1’. There exists ε > such that the following holds. Let ( ∪ α W iα , µ i ) , i = 1 , , bemeasured unstable stacks. Assume Z i < ∞ and that the conditional densities satisfy | log ρ iα ( x ) − log ρ iα ( y ) | ≤ C i θ s ( x,y ) for all x, y ∈ W iα . Given γ > , there exist < θ γ < and C γ > suchthat (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) M f ◦ F n d µ − (cid:90) M f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ C γ ( (cid:107) f (cid:107) ∞ + | f | γ ) θ nγ , n ≤ N, for all sequences ( K n ) Nn =0 ⊂ K ( N ∈ N ∪ {∞} ) satisfying d ( K n − , K n ) < ε for ≤ n ≤ N ,and all γ -H¨older continuous f : M → R . The constant C γ depends on max( C , C ) and max( Z , Z ) , while θ γ does not. Let us say a finite Borel measure µ is regular on unstable curves if it admits a representationas the measure in a regular measured unstable family with Z < ∞ , proper if additionally itadmits a representation that is proper. It follows from Lemmas 13 and 16 that in Theorem 1’,after a finite number of steps depending on C i and Z i , the pushforward of µ i will becomeregular on unstable curves and proper.For completeness, we provide a proof of the following in the Appendix. Lemma 19. Theorem 1’ generalizes Theorem 1. Theorem 2’. This is obtained from Theorems 2 in exactly the same way as Theorem 1’ isobtained from Theorem 1, namely by relaxing the condition on µ i as stated. As a matter of fact, instead of just Theorem 4, the following generalization is proved inSection 8.1. Theorem 4’. This is a similar extension of Theorem 4, i.e., of the local result, to initialmeasures as stated in Theorem 1’. We finish by remarking on our use of measured unstable stacks and families: The primaryreason for considering these objects is that in the proof we really work with measured unstable curves and their images under F n . “Thin enough” stacks of unstable curves behave in a wayvery similar to unstable curves, and are treated similarly. Other generalizations are made sowe can include a larger class of initial distributions; moreover, to the extent that is possible, itis always convenient to work with a class of objects closed under the operations of interest. SeeRemark 18. We note also that our formulation here has deviated from [10] because of (fixable)measurability issues with their formulation.In view of the fact that our proof really focuses on curves, we will, for pedagogical reasonsconsider separately the following two cases:(1) The countable case , in which we assume that each initial distribution µ i is supportedon a countable family of unstable curves, i.e., the stacks above consist of single curves.(2) The continuous case , where we allow the µ i to be as in Theorem 1’.For clarity of exposition, we first focus on the countable case, presenting a synopsis of the prooffollowed by a complete proof; this is carried out in Sections 5–7. Extensions to the continuouscase is discussed in Section 8.5. Theorem 4: synopsis of proof This important section contains a sketch of the proof, from beginning to end, of the “count-able case” of Theorem 4; it will serve as a guide to the supporting technical estimates in thesections to follow. We have divided the discussion into four parts: Paragraph A contains anoverview of the coupling argument on which the proof is based. The coupling procedure itselffollows closely [10]; it is reviewed in Paragraphs B and C. Having an outline of the proof in handpermits us to isolate the issues directly related to the time-dependent setting; this is discussedin Paragraph D. As mentioned in the Introduction, one of the goals of this paper is to stressthe (strong) similarities between stationary dynamics and their time-dependent counterparts,and to highlight at the same time the new issues that need to be addressed.For simplicity of notation, we will limit the discussion here to the “countable case” of The-orem 4. That is to say, we assume throughout that the initial distributions µ i , i = 1 , 2, areproper measures supported on a countable number of unstable curves; see Section 4.3. A. Overview of coupling argument The following scheme is used to produce the exponential bound in Theorem 4. Let ( K n ) , n ≤ N ≤ ∞ be an admissible (finite or infinite) sequence of configurations with associated composedmaps F n = F n ◦· · ·◦ F . Given initial probability distributions µ and µ on M , we will producetwo sequences of nonnegative measures ¯ µ in , n ≤ N , with properties (i)–(iii) below: ISPERSING BILLIARDS WITH MOVING SCATTERERS 17 (i) for i = 1 , µ i = (cid:80) j ≤ n ¯ µ ij + µ in with ¯ µ n ( M ) = ¯ µ n ( M ) for each n ;(ii) µ n ( M ) = µ n ( M ) ≤ Ce − an ;(iii) | (cid:82) f ◦ F n + m d¯ µ n − (cid:82) f ◦ F n + m d¯ µ n | ≤ C f e − am ¯ µ in ( M ), for any test function f .Here ¯ µ in , i = 1 , 2, are the components of µ i coupled at time n ; their relationship from time n on is given by (iii). By (ii), the yet-to-be-coupled part decays exponentially. In practice, acoupling occurs at a sequence of times 0 < t < t < · · · < t K < N . In particular, ¯ µ ij = 0, when j (cid:54) = t k for all 1 ≤ k ≤ K , which means that µ in remains unchanged between successive couplingtimes.It follows from (i)–(iii) above that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d µ − (cid:90) f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:107) f (cid:107) ∞ · µ in/ ( M ) + (cid:88) j ≤ n/ C f e − a ( n − j ) ¯ µ ij ( M ) ≤ (cid:107) f (cid:107) ∞ Ce − an/ + C f e − an/ . (17)We indicate briefly below how, at time n where n = t k is a coupling time, we extract ¯ µ in from µ in − and couple ¯ µ n to ¯ µ n . Recall that in the hypotheses of Theorem 4, ( K m ) Nm =0 is adaptedto ( (cid:101) K q , ˜ ε ( (cid:101) K q ) , (cid:101) N ( (cid:101) K q )) Qq =1 . We assume K n ∈ N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ) for some q . In fact, coupling times arechosen so that K m is in the same neighborhood for a large number of m ≤ n leading up to n .For simplicity, we write (cid:101) K = (cid:101) K q , and (cid:101) F = F (cid:101) K , (cid:101) K .For the coupling at time n , we construct a coupling set S n ⊂ M analogous to the “magnets”in [10] — except that it is a time-dependent object. Specifically, S n = ∪ x ∈ (cid:102) W n W sn ( x )where (cid:102) W is a piece of unstable manifold of (cid:101) F (here we mean unstable manifold of a fixed mapin the usual sense and not just an “unstable curve” as defined in Section 3.3), (cid:102) W n ⊂ (cid:102) W is aCantor subset with m (cid:102) W ( (cid:102) W n ) / | (cid:102) W | ≥ , and W sn ( x ) is the stable manifold of length ≈ | (cid:102) W | centered at x for the sequence F n +1 , F n +2 , . . . (if N < ∞ , let K m = K N for all m > N ).It will be shown that at time n , the F n -image of each of the measures µ in − , i = 1 , 2, is againthe union of countably many regular measures supported on unstable curves. Temporarily letus denote by (cid:101) ν in the part of ( F n ) ∗ µ in − that is supported on unstable curves each one of whichcrosses S n in a suitable fashion, meeting every W sn ( x ) in particular. We then show that thereis a lower bound (independent of n ) on the fraction of ( F n ) ∗ µ in − that (cid:101) ν in comprises, and couplea fraction of (cid:101) ν n to (cid:101) ν n by matching points that lie on the same local stable manifold.We comment on our construction of S n : Given that F m is close to (cid:101) F for many m before n , F n -images of unstable curves will be roughly aligned with unstable manifolds of (cid:101) F , henceour choice of (cid:102) W . In order to achieve the type of relation in (iii) above, we need to have |F n + m ( x ) − F n + m ( y ) | → m for two points x and y “matched” in our couplingat time n , hence our choice of W sn . Observe that in our setting, the “magnets” S n are necessarilytime-dependent.To further pinpoint what needs to be done, it is necessary to better acquaint ourselves withthe coupling procedure. For simplicity, we assume in Paragraphs B and C below that all theconfigurations in question lie in a small neighborhood N ˜ ε ( (cid:101) K ) of a single reference configura-tion (cid:101) K . As noted earlier, details of this procedure follow [10]. We review it to set the stageboth for the discussion in Paragraph D and for the technical estimates in the sections to follow. B. Building block of procedure: coupling of two measured unstable curves We assume in this paragraph that µ i , i = 1 , 2, is supported on a homogeneous unstablecurve W i , and that the following hold at some time n ≥ 0: (a) the image F m W i is a homogeneousunstable curve for 1 ≤ m ≤ n ; (b) the push-forward measure ( F n ) ∗ µ i = ν in has a regular density ρ in on F n W i = W in ; and (c) W in crosses the magnet S n “properly”, which means roughly that (i) it meets each stable manifold W sn ( x ), x ∈ (cid:102) W n , (ii) the excess pieces sticking “outside” ofthe magnet S n are sufficiently long, and (iii) part of W in is very close to and nearly perfectlyaligned with (cid:102) W (for a precise definition of a proper crossing, see Definition 22).Due to their regularity, the probability densities ρ in are strictly positive. Moreover, theholonomy map h , n : W n ∩ S n → W n ∩ S n has bounds on its Jacobian (Section 3.4). Thus,we may extract a fraction ¯ ν in from each measure ν in | ( W in ∩ S n ) with ( h , n ) ∗ ¯ ν n = ¯ ν n and ¯ ν n ( M ) =¯ ν n ( M ) = ζ for some ζ > 0. Because each x ∈ W n ∩ S n lies on the same stable manifold as h , n ( x ) ∈ W n ∩ S n , (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n + m,n +1 d¯ ν n − (cid:90) f ◦ F n + m,n +1 d¯ ν n (cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n + m,n +1 d¯ ν n − (cid:90) f ◦ F n + m,n +1 ◦ h , n d¯ ν n (cid:12)(cid:12)(cid:12)(cid:12) ≤ | f | γ (ˆ c − Λ − m ) γ ζ (18)by (4), for all γ -H¨older functions f and m ≥ 0. (We have assumed that the local stablemanifolds associated with the holonomy map have lengths ≤ µ i = µ in + ¯ µ in , where ( F n ) ∗ ¯ µ in = ¯ ν in corresponds to the part coupled at time n and( F n ) ∗ µ in = ν in − ¯ ν in to the part that remains uncoupled. For 0 ≤ m < n , we set ¯ µ im = 0 and µ im = µ i .In more detail, it is in fact convenient to couple each measure to a reference measure (cid:101) m n supported on (cid:102) W : Once two measures are coupled to the same reference measure, they are alsocoupled to each other. Define the uniform probability measure (cid:101) m n ( · ) = m (cid:102) W ( · ∩ S n ) / m (cid:102) W ( (cid:102) W ∩ S n ) (19)on (cid:102) W ∩ S n and write h in for the holonomy map (cid:102) W ∩ S n → W in ∩ S n . Then ( h in ) ∗ (cid:101) m n is aprobability measure on W in ∩ S n . We assume that h = h in satisfies | ( J h ◦ h − ) − − | ≤ . (20)By the regularity of the probability densities ρ in , there exists a number ζ > ν in ( W in ∩ S n ) ≥ ζe C r ( i = 1 , . (21)Setting ¯ ν in = ζ ( h in ) ∗ (cid:101) m n , we have ¯ ν n ( M ) = ¯ ν n ( M ) = ζ. Let ¯ ρ in be the density of ¯ ν in (so that it issupported on W in ∩ S n ). By (20) and (21), one checks that sup ¯ ρ in ≤ · inf W in ρ in , so that whatwe couple is strictly a fraction of ν in .In preparation for future couplings, we look at ν in − ¯ ν in , the images of the uncoupled partsof the measures. First, W in can be expressed as the union of W in ∩ S n and W in \ S n , thelatter consisting of countably many gaps “inside” the magnet S n and two excess pieces sticking“outside” of it. Moreover, there is a one-to-one correspondence between the gaps V i ⊂ W in \ S n and the gaps (cid:101) V ⊂ (cid:102) W \ (cid:102) W n . Notice that ν in − ¯ ν in has a positive density bounded away fromzero on the curve W in , but that density is not regular as ¯ ν in is only supported on the Cantorset W in ∩ S n . We decompose ν in − ¯ ν in as follows: First we separate the part that lies on theexcess pieces of W in “outside” the magnet. Let ( W in ) (cid:48) denote the curve that remains. Viewedas a density on W in ∩ S n , ¯ ρ in is regular, since C h ≤ C r . It can be continued to a regular densityon all of ( W in ) (cid:48) without affecting its bounds (Lemma 15). Letting ˇ ρ in denote this extension, wehave ( ρ in − ¯ ρ in ) | ( W in ) (cid:48) = ( ρ in − ˇ ρ in ) | ( W in ) (cid:48) + (cid:88) V i ⊂ ( W in ) (cid:48) \ S n V i ˇ ρ in , where the sum runs over the gaps V i in ( W in ) (cid:48) . Notice that each of the densities ˇ ρ in | V i on thegaps is regular. While ( ρ in − ¯ ρ in ) | ( W in ) (cid:48) is in general not regular, it is not far from regular becauseboth ρ in and ˇ ρ in are regular and ( ρ in − ˇ ρ in ) > ρ in . C. The general procedure ISPERSING BILLIARDS WITH MOVING SCATTERERS 19 Still assuming that all K n lie in a small neighborhood N ˜ ε ( (cid:101) K ) of a single reference config-uration (cid:101) K , we now consider a proper initial probability measure µ = (cid:80) α ∈A ν α , consisting ofcountably many regular measures ν α , each supported on a regular unstable curve W α . Asexplained in Paragraph B, the problem is reduced to coupling a single initial distribution toreference measures on (cid:102) W . Leaving the determination of suitable coupling times t < t < · · · for later, we first discuss what happens at the first coupling: The first coupling at time n = t . Denote by W α,n,i the components of F n W α resulting from itscanonical subdivision, where i runs over an at-most-countable index set. Similarly, the push-forward measure is ( F n ) ∗ ν α = (cid:80) i ν α,n,i , where each ν α,n,i is supported on W α,n,i . As before,each ν α,n,i has a regular density ρ α,n,i on W α,n,i .For each α ∈ A , let I α,n be the set of indices i for which W α,n,i crosses S n properly, asdiscussed earlier. This set is finite, as |F n W α | < ∞ and | W α,n,i | is uniformly bounded frombelow (by the width of the magnet) for i ∈ I α,n .Let ζ ∈ (0 , 1) be such that (cid:88) α ∈A (cid:88) i ∈I α,n ν α,n,i ( W α,n,i ∩ S n ) ≥ ζ e C r . (22)As in (19), let (cid:101) m n denote the uniform probability measure on (cid:102) W ∩ S n and write h α,n,i for theholonomy map (cid:102) W ∩ S n → W α,n,i ∩ S n , for each i ∈ I α,n . Then ( h α,n,i ) ∗ (cid:101) m n is a probabilitymeasure on W α,n,i ∩ S n which is regular and nearly uniform. We set¯ ν α,n,i = λ α,n,i ( h α,n,i ) ∗ (cid:101) m n , (23)for each i ∈ I α,n , where λ α,n,i = ζ · ν α,n,i ( W α,n,i ∩ S n ) (cid:80) β ∈A (cid:80) j ∈I β,n ν β,n,j ( W β,n,j ∩ S n ) . Then (cid:88) α ∈A (cid:88) i ∈I α,n ¯ ν α,n,i ( M ) = (cid:88) α ∈A (cid:88) i ∈I α,n λ α,n,i = ζ . Moreover, the density ¯ ρ α,n,i of ¯ ν α,n,i on W α,n,i ∩ S n is regular (and in fact nearly constant). Asin Paragraph B, the density can be extended in a regularity preserving way (Lemma 15) tothe curve ( W α,n,i ) (cid:48) obtained from W α,n,i by cropping the excess pieces outside the magnet. Wedenote the extension by ˇ ρ α,n,i . As before, ( ρ α,n,i − ˇ ρ α,n,i ) | ( W α,n,i ) (cid:48) is generally not regular, andto control it, we record the following bounds: Lemma 20. For each α ∈ A and i ∈ I α,n , ζ e − C r · sup W α,n,i ρ α,n,i ≤ inf ( W α,n,i ) (cid:48) ˇ ρ α,n,i ≤ sup ( W α,n,i ) (cid:48) ˇ ρ α,n,i ≤ · inf W α,n,i ρ α,n,i . (24) Proof. We begin by observing that the density of ( h α,n,i ) ∗ (cid:101) m n on W α,n,i ∩ S n has the expression( m (cid:102) W ( (cid:102) W ∩ S n ) J h α,n,i ◦ ( h α,n,i ) − ) − and that the supremum of a regular density is boundedby e C r times its infimum. By (20) and (22), the third inequality in (24) follows easily. Comingto the first inequality in (24), it is certainly the case that (cid:88) β ∈A (cid:88) j ∈I β,n ν β,n,j ( W β,n,j ∩ S n ) ≤ µ ( M ) ≤ . As the density ρ α,n,i is regular, λ α,n,i ≥ ζ · ν α,n,i ( W α,n,i ∩ S n ) ≥ ζe − C r · sup W α,n,i ρ α,n,i · m W α,n,i ( W α,n,i ∩ S n ) . Again by (20),inf ( W α,n,i ) (cid:48) ˇ ρ α,n,i = λ α,n,i · inf W α,n,i ∩ S n ( m (cid:102) W ( (cid:102) W ∩ S n ) J h α,n,i ◦ ( h α,n,i ) − ) − ≥ λ α,n,i m (cid:102) W ( (cid:102) W ∩ S n ) ≥ ζe − C r · m W α,n,i ( W α,n,i ∩ S n ) m (cid:102) W ( (cid:102) W ∩ S n ) · sup W α,n,i ρ α,n,i ≥ (cid:0) (cid:1) ζe − C r · sup W α,n,i ρ α,n,i . This finishes the proof. (cid:3) To recapitulate, in the language of Paragraph A, we have split µ into ¯ µ n + µ n with( F n ) ∗ ¯ µ n = (cid:88) α ∈A (cid:88) i ∈I α,n ¯ ν α,n,i . (25)The measures ( F n ) ∗ ¯ µ n and ζ (cid:101) m n are coupled. Going forward. To proceed inductively, we need to discuss the uncoupled part µ n (for n = t ),which has the form ( F n ) ∗ µ n = (cid:88) α ∈A (cid:88) i/ ∈I α,n ν α,n,i + (cid:88) α ∈A (cid:88) i ∈I α,n ( ν α,n,i − ¯ ν α,n,i ) . The measures ν α,n,i in the first term above are regular, so we leave them alone. The measures ν α,n,i − ¯ ν α,n,i in the second term are further subdivided as in Paragraph B, into the regulardensities on the excess pieces, ˇ ρ α,n,i V on the gaps V ⊂ ( W α,n,i ) (cid:48) \ S n of the Cantor sets S n ∩ W α,n,i , and ( ρ α,n,i − ˇ ρ α,n,i ) | ( W α,n,i ) (cid:48) , which are in general not quite regular. Because of thearbitrarily small gaps in ( W α,n,i ) (cid:48) \ S n , the resulting family is not proper.We allow for a recovery period of r > ρ α,n,i − ˇ ρ α,n,i ) | ( W α,n,i ) (cid:48) to be restored, and short curves to become longeron average (as a result of the hyperbolicity). Because of the arbitrarily short gaps, a fraction ofthe measure will not recover sufficiently to become proper no matter how long we wait, but thisfraction decreases exponentially with time. Specifically, for all sufficiently large m , the m -steppush-forward ( F t + m ) ∗ µ t of the uncoupled measure ( F t ) ∗ µ t can be split into the sum of twomeasures µ P t ,m and µ G t ,m , both consisting of countably many regular measured unstable curves,such that µ P t ,m is a proper measure and µ G t ,m ( M ) = C λ m for some C ≥ λ ∈ (0 , r large enough, µ G t ,r ( M ) is thus as small as we wish.At time t + r we are left with a proper measure µ P t ,r having total mass 1 − ζ − C λ r , andanother measure µ G t ,r supported on a countable union of short curves. We consider µ P t ,r , andassume that after s > t = t + r + s , we perform another coupling in thesame fashion as the one performed at time t , this time coupling a ζ -fraction of ( F t ,t ) ∗ µ P t ,r to the measure ζ (1 − ζ − C λ r ) (cid:101) m t .The cycle is repeated: Following a recovery period of length r , i.e., at time t + r , themeasure of mass (1 − ζ )(1 − ζ − C λ r ) left from the second coupling can be split into a properpart µ P t ,r and a non-proper part µ G t ,r , the latter having mass C λ r (1 − ζ − C λ r ). At the sametime, most of µ G t ,r has now become proper: the fraction of µ G t ,r that still has not recoveredat time t + r has mass C λ r +( t − t )1 . We wait another s steps, until time t = t + r + s ,for a sufficiently large fraction of the push-forward measure to cross the magnet properly. Attime t , we couple a ζ -fraction of ( F t ,t ) ∗ µ P t ,r plus the F t ,t -image of the part of µ G t ,r thathas recovered, to a measure on (cid:102) W , and so on. Our main challenge is to prove that the estimates above are uniform , i.e., there exist C ≥ ζ, λ ∈ (0 , 1) and r, s ∈ Z + , independently of the sequence ( K n ) provided each K n ∈ N ˜ ε ( (cid:101) K ), sothat the scheme above can be carried out with C i = C , ζ i = ζ, r i = r, λ i = λ and s i = s for all i . ISPERSING BILLIARDS WITH MOVING SCATTERERS 21 Assuming these uniform estimates, the situation for K n ∈ N ˜ ε ( (cid:101) K ), all n , can be summarized asfollows: Summary. We push forward the initial distribution, performing couplings with the aid of atime-dependent “magnet” at times t < t < . . . , and performing canonical subdivisions (forconnectedness and distortion control) in between. The t k ’s are r + s steps apart, with t depending additionally on the initial distribution µ . At each coupling time t k , a ζ -fraction ofthe uncoupled measure that is proper is coupled. At the same time, a small fraction of the stilluncoupled measure becomes non-proper due to the small gaps in the magnet. This non-properpart regains “properness” thereby returning to circulation exponentially fast, the exceptionalset constituting a fraction Cλ m after m steps. Simple arithmetic shows that by such a scheme,the yet-to-be coupled part of µ has exponentially small mass. This implies exponential memoryloss. D. What makes the proof work in the time-dependent case We now return to the full setting of Theorem 4, where we are handed a sequence ( K n ) Nn =0 adapted to ( (cid:101) K q , ˜ ε ( (cid:101) K q ) , (cid:101) N ( (cid:101) K q )) Qq =1 . Exponential memory loss of this sequence must necessarilycome from the corresponding property for (cid:101) F q = F (cid:101) K q , (cid:101) K q . The question is: how does the expo-nential mixing property of a system pass to compositions of nearby systems? Such a resultcannot be taken for granted, for in general mixing involves sets of all sizes, and smaller setsnaturally take longer to mix, while two systems that are a positive distance apart will havetrajectories that follow each other up to finite precision for finite times . That is to say, once theneighborhood is fixed, perturbative arguments are not effective for treating arbitrarily smallscales. These comments apply to iterations of fixed maps as well as time-dependent sequences.What is special about our situation is that there is a characteristic length (cid:96) to which imagesof all unstable curves tend to grow exponentially fast under F n , for all F = F K , K , K ∈ K , beforethey get cut — with the exception of exponentially small sets (see Lemma 12). The presenceof such a characteristic length suggests that to prove exponential mixing, it may suffice toconsider rectangles aligned with stable and unstable directions that are ≥ (cid:96) in size, and to treatseparately growth properties starting from arbitrarily small length scales. These ideas havebeen used successfully to prove exponential correlation decay for classical billiards, and will beused here as well. To carry out the program outlined in Paragraphs A–C, we need to prove that for each (cid:101) K q ,the following holds, with uniform bounds, for all ( K n ) in a sufficiently small neighborhood of (cid:101) K q :(1) Uniform mixing on finite scales. We will show that there is a uniform lower bound on thespeeds of mixing for rectangles of sides ≥ (cid:96) for the time-dependent maps defined by ( K n ). For (cid:101) F q = F (cid:101) K q , (cid:101) K q , this is proved in [3, 4], and what we prove here is effectively a perturbative versionfor time-dependent sequences in a small enough neighborhood of (cid:101) F q . Such a result is feasiblebecause it involves only finite-size objects for finite times. Caution must be exercised still, asthe maps involved are discontinuous. This result gives the s = s ( (cid:101) K q ) asserted in Paragraph C.(2) Uniform structure of magnets. To ensure that a definite fraction of measure is coupledwhen a measured unstable curve crosses the magnet, a uniform lower bound on the densityof local stable manifolds in S n is essential: we require m (cid:102) W ( (cid:102) W n ) / | (cid:102) W | ≥ ; see Paragraph A.In fact, we need more than just a minimum fraction: uniformity in the distribution of smallgaps in S n is also needed. Following a coupling, they determine how far from being proper theuncoupled part of the measure is; see Paragraph C. As S n , the magnet used for coupling at The ideas alluded to here are applicable to large classes of dynamical systems with some hyperbolic propertiesincluding but not limited to billiards; they were enunciated in some generality in [20], which also provedexponential correlation decay for the periodic Lorentz gas. time n , is constructed using the local stable manifolds of F n , F n +1 , . . . , the results above musthold uniformly for all relevant sequences.(3) Uniform growth of unstable curves. This very important fact, which takes into considerationboth the expansion due to hyperbolicity of the map and the cutting by discontinuities andhomogeneity lines, is used in more ways than one: It is used to ensure that regularity ofdensities is restored and most of the uncoupled measure becomes proper at the end of the“recovery periods”. The uniform r and λ asserted in Paragraph C are obtained largely fromthe uniform structure of magnets, i.e., item (2) above, together with the growth results inSection 4 (as well as inductive control from previous steps). Growth results are also usedto produce a large enough fraction of sufficiently long unstable curves at times t k + r . Thattogether with the uniform mixing in item (1) permits us to guarantee the coupling of a fixedfraction ζ at time t k +1 .Item (1) above is purely perturbative as we have discussed; item (2) is partially perturbative:proximity to (cid:101) K q is used to ensure that S n has some of the desired properties. Item (3) isstrictly nonperturbative: we do not derive the properties in question from the proximity of thecomposed sequence F n ◦ · · · ◦ F ◦ F to (cid:101) F nq . Instead, we show that these properties hold true,with uniform bounds, for all sequences ( K n ) with K n ∈ K . In the case of genuinely movingscatterers, the constants r, C and ζ all depend on the relevant reference configuration (cid:101) K q ,through the curve (cid:102) W in whose neighborhood the couplings occur. A priori, the same is true of λ , although, as we will show, λ is in fact independent of (cid:101) K q .6. Main ingredients in the proof of Theorem 4 We continue to develop the main ideas needed in the proof of Theorem 4, focusing first on thecountable case and addressing issues that have been raised in the synopsis in the last section.As in Sections 3 and 4, all configuration pairs whose billiard maps are discussed are assumedto be admissible and in K . Further conditions on ( K n ), such as close proximity to a referenceconfiguration, will be stated explicitly. Many of the results below are parallel to known resultsfor classical billiards; see e.g. [10].6.1. Local stable manifolds. Given ( K n ) ∞ n =0 , we let W s ( x ) denote the maximal (possiblyempty, homogeneous) local stable manifold passing through the point x ∈ M for the sequenceof maps ( F n ) n ≥ . Recall that W s ( x ) has positive length if and only if the trajectory F n x doesnot approach the “bad set” ∂ M ∪ ∪ | k |≥ k ∂ H k too fast as a function of n . Based on this fact,the size of local stable manifolds may be quantified as follows: Let r s ( x ) denote the distanceof x from the nearest endpoint of W s ( x ) as measured along W s ( x ). A standard computation,which we omit, shows that for an arbitrary unstable curve W through x , r s ≥ (cid:101) C − inf n ≥ Λ n r W,n ≡ u sW , (26)where (cid:101) C > r W,n was introduced in the beginning of Section 4.1.In Paragraph D, item (2), of the Synopsis, we identified the need for certain uniform propertiesof local stable manifolds, such as the density of stable manifolds of uniform length on unstablecurves. The next lemma provides a basic result in this direction. Lemma 21. Given a > and A > , there exist s (cid:48) ∈ Z + and L > such that for any ( K n ) ∞ n =0 ,every unstable curve (cid:102) W has the property m (cid:102) W (cid:8) u sW ≥ A (cid:12)(cid:12)(cid:102) W (cid:12)(cid:12)(cid:9) ≥ (1 − a ) | (cid:102) W | (27) provided (i) (cid:102) W is located in the middle third of a homogeneous unstable curve W for which F s (cid:48) W has a single homogeneous component, and (ii) | (cid:102) W | ≤ L | W | / . ISPERSING BILLIARDS WITH MOVING SCATTERERS 23 Proof. Since F s (cid:48) W consists of a single homogeneous component, we have r W,n ( x ) ≥ ˆ c Λ n r W, ( x ) ∀ x ∈ W, ≤ n ≤ s (cid:48) ,r W,n ( x ) ≥ r (cid:102) W ,n ( x ) ∀ x ∈ (cid:102) W , ≤ n ≤ s (cid:48) ,r W, ( x ) ≥ | W | / r (cid:102) W , ( x ) ∀ x ∈ (cid:102) W . Using these facts and the Growth Lemma, we can estimate that m (cid:102) W { u sW < A | (cid:102) W |} ≤ (cid:88) n ≥ m (cid:102) W { r W,n < (cid:101) CA | (cid:102) W | Λ − n }≤ (cid:88) n ≤ s (cid:48) m (cid:102) W { r W, < ˆ c − (cid:101) CA | (cid:102) W | Λ − n } + (cid:88) n>s (cid:48) m (cid:102) W { r (cid:102) W ,n < (cid:101) CA | (cid:102) W | Λ − n }≤ (cid:88) n ≤ s (cid:48) m (cid:102) W { r (cid:102) W , < ˆ c − (cid:101) CA | (cid:102) W | Λ − n − | W | / } + (cid:88) n>s (cid:48) C gr ( ϑ n + | (cid:102) W | ) (cid:101) CA | (cid:102) W | Λ − n . In the last line, the first sum vanishes if we take L ≤ ˆ c (cid:101) C − A − and the Growth Lemmayields the bound on the second sum. The second sum is then < a | (cid:102) W | if s (cid:48) is so large that C gr (1+ LL / (cid:101) CA (Λ − s (cid:48) ≤ a , where L here is the maximum length of a homogeneous unstable curve. (cid:3) Magnets. We now define more precisely the objects S n in Paragraph A of the Synopsis.Recall that S n is constructed using stable manifolds W sn ( x ) with respect to the sequence ofmaps F n , F n +1 , F n +2 , . . . . When what happens before time n is irrelevant to the topic underdiscussion, it is simpler notationally to set n = 0 (by shifting and renaming indices in theoriginal sequence). That is what we will do here as well as in the next few subsections.We fix a reference configuration (cid:101) K ∈ K , and denote (cid:101) F = F (cid:101) K , (cid:101) K . Let s (cid:48) and L be given byLemma 21 with A = and a = 0 . 01. We pick a piece of unstable manifold (cid:102) W u + of (cid:101) F (morethan just an unstable curve) with the property that (cid:102) W u + is homogeneous and (cid:101) F s (cid:48) (cid:102) W u + has asingle homogeneous component. Let (cid:102) W u ⊂ (cid:102) W u + be the subsegment of (cid:102) W u + half as long andlocated at the center. Then there exists ε (cid:48) > F s (cid:48) (cid:102) W u has a single homogeneouscomponent for any K n ∈ N ε (cid:48) ( (cid:101) K ) , ≤ n ≤ s (cid:48) . Let (cid:102) W ⊂ (cid:102) W u be located at the center of (cid:102) W u with | (cid:102) W | = L | (cid:102) W u | / 3. Lemma 21 then tells us that for ( K n ) as above and (cid:102) W := { x ∈ (cid:102) W : u s (cid:102) W u ( x ) ≥ | (cid:102) W | / } , we are guaranteed that m (cid:102) W ( (cid:102) W ) / | (cid:102) W | ≥ . The set S = ∪ x ∈ (cid:102) W W s ( x ) isthe magnet defined by (cid:102) W and the sequence ( K n ).Additional upper bounds will be imposed on | (cid:102) W | to obtain the magnet used in the proof ofTheorem 4. The size of the neighborhood N ε (cid:48) ( (cid:101) K ) will also be shrunk a finite number of timesas we go along.Now let W be any unstable curve that crosses S completely, in the sense that it meets W s ( x ) for each x ∈ (cid:102) W with excess pieces on both sides, and let h denote the holonomy mapfrom (cid:102) W ∩ S → W ∩ S . Definition 22. We say the crossing is proper if for a uniform constant ℵ > to be determined,the following hold: (i) W is regular, (ii) the distance between any x ∈ (cid:102) W and h ( x ) as measuredalong W s ( x ) is less than ℵ| (cid:102) W | , and (iii) each of the two excess pieces “outside” the magnet ismore than | (cid:102) W | units long. We need ℵ to be small enough that (20), i.e., | ( J h ◦ h − ) − − | ≤ , is guaranteed in propercrossings. To guarantee (20), we need, by (8), both (i) the distance between x and h ( x ) and(ii) the difference between the slopes of (cid:102) W and (cid:102) W at x and h ( x ) respectively, to be small. (i)is bounded by ℵ| (cid:102) W | . Observe that (ii) is also (indirectly) controlled by ℵ : since both W and (cid:102) W are regular curves (real unstable manifolds of (cid:101) F are automatically regular), there is a fixed upper bound on their curvatures. Thus the shorter the curves, the closer they are to straightlines. Now since W meets the stable manifolds at the two ends of (cid:102) W at distances < ℵ| (cid:102) W | from (cid:102) W , taking ℵ small forces the slopes of W and (cid:102) W to be close. Further upper bounds on ℵ maybe imposed later. Note on terminology. In the discussion to follow, the setting above is assumed, and a numberof constants referred to as “uniform constants” will be introduced. This refers to constantsthat are independent of (cid:101) K for as long as (cid:101) K ∈ K , and they are independent of ( K n ), (cid:102) W u , (cid:102) W or (cid:102) W provided these objects are chosen according to the recipe above.6.3. Gap control. We discuss here the distribution of gap sizes of the magnet, issues aboutwhich were raised in the Synopsis. The setting, including S , is as in Section 6.2.Recall that a point x ∈ (cid:102) W belongs to the Cantor set (cid:102) W if and only if r (cid:102) W u ,k ( x ) ≥ (cid:101) C Λ − k | (cid:102) W | / k ≥ 0. We define the rank of a gap (cid:101) V in (cid:102) W \ (cid:102) W to be the smallest R such that r (cid:102) W u ,R ( x ) < (cid:101) C Λ − R | (cid:102) W | / x ∈ (cid:101) V . Observe that R so defined is also the smallestnumber for which F R (cid:101) V meets the “bad set” ∂ M ∪ ∪ | k |≥ k ∂ H k : Clearly, F k (cid:101) V could not havemet the bad set for k < R . On the other hand, F R (cid:101) V must meet the bad set, or the minimumof r (cid:102) W u ,R on (cid:101) V would occur on one of its end points, which cannot happen for a gap (excesspieces are not gaps). Notice that this implies that F R − (cid:101) V must cross (transversally) F − R ∂ H k for some k , and if it crosses F − R ∂M , then it automatically crosses F − R ∂ H k for infinitely many k . Consider next an unstable curve W that crosses S properly, and let W = W ∩ S . Eachgap V ⊂ W \ W corresponds canonically to a unique gap (cid:101) V ⊂ (cid:102) W \ (cid:102) W , as their correspondingend points are connected by local stable manifolds γ s and γ s in S . We define the rank of V to be that of (cid:101) V , and claim that the rank of V is also the first time F R V meets the bad set. Tosee this, consider the F n -images of the region bounded by V, (cid:101) V , γ s and γ s . Since F n ( γ si ) avoidthe bad set (which consists of horizontal lines), it follows that for each n , F n ( V ) crosses thebad set if and only if F n ( (cid:101) V ) does.Let W be as above. For b > 0, we consider the dynamically defined Cantor set W b = { x ∈ W (cid:48) : r W,k ≥ b Λ − k | (cid:102) W | ∀ k ≥ } . (28)For W = (cid:102) W u and b = (cid:101) C , W b = (cid:102) W . We observe that, for b ≤ ˆ c , with ˆ c as in (4), the definitionof the set W b does not depend on the part W \ W (cid:48) “outside” the magnet. This is because ofthe length of each of the two components of W \ W (cid:48) and the expansion in (4). Like (cid:102) W , W b isa Cantor set, and the ranks of the gaps of W \ W b have the same characterizations as the gapsof (cid:102) W u \ (cid:102) W .The proofs below are a little sketchy, as there are no new issues in the time-dependent case. Lemma 23. There exists a uniform constant ¯ b ≤ ˆ c such that the following hold for ℵ smallenough: Let W be an arbitrary unstable curve crossing S properly, and let W = W ∩ S .Then(i) W ⊂ W ¯ b , and(ii) through every point of W ¯ b there is a local stable manifold which meets (cid:102) W .Proof. (i) Let ˜ x ∈ (cid:102) W , and denote by x ∈ W the intersection of W s ( (cid:101) x ) and W . The assertionfollows since for some a ∈ (0 , 1) depending only on the cones, d M ( F n ˜ x, ∂ M ∪ ∪ | k |≥ k ∂ H k ) ≥ a (cid:101) C Λ − n | (cid:102) W | / n ≥ 0, while d M ( F n x, F n ˜ x ) ≤ ˆ c − Λ − n ℵ| (cid:102) W | for all n ≥ 0. In particular,¯ b = min(ˆ c, a (cid:101) C/ 4) suffices. (ii) At each point x ∈ W ¯ b the local stable manifold W s ( x ) extendsat least ¯ b (cid:101) C − | (cid:102) W | units on both sides of W , proving (ii) for ℵ sufficiently small. (cid:3) We record next a tail bound for gaps of dynamically defined Cantor sets. ISPERSING BILLIARDS WITH MOVING SCATTERERS 25 Lemma 24. There exists a uniform constant C (cid:48) g > such that if W crosses S properly, thenfor any b > and R ≥ , we have m W (cid:110) x ∈ W (cid:48) : inf (cid:8) k ∈ N : r W,k ( x ) < b Λ − k | (cid:102) W | (cid:9) ∈ [ R, ∞ ) (cid:111) ≤ b C (cid:48) g Λ − R | (cid:102) W | . (29) Proof. The Growth Lemma yields the following upper bound on the left side of (29): (cid:88) k ≥ R m W { r W,k < b Λ − k | (cid:102) W |} ≤ b C gr | (cid:102) W | (cid:88) k ≥ R ( ϑ k + | W | )Λ − k ≤ b C gr (cid:101) C | (cid:102) W | (1 + L )1 − Λ − Λ − R , where L is the maximum length of a (connected) unstable curve. (cid:3) The following is the result we need. Lemma 25. There exist uniform constants C g > and c g > for which the following hold:Let S be as above, and let W be an arbitrary unstable curve which crosses S properly. Then(a) for every R ≥ , m W { x ∈ W : x is in a gap of rank ≥ R } ≤ C g Λ − R | W | ; (30) (b) for any gap V ⊂ W \ W of any rank R ≥ , F R − V is a homogeneous unstable curve and |F R V | ≥ c g Λ − R | (cid:102) W | . (31) Proof of Lemma 25. (a) For W = (cid:102) W , the result follows immediately from Lemma 24. Forgeneral W , a separate argument is needed as W ∩ S is not exactly of the form in (28). Let¯ b be such that W ⊂ W ¯ b (Lemma 23(i)), and let V be a gap of W \ W of rank R . Then V is the union V ∩ W ¯ b and a collection of gaps of W \ W ¯ b . We observe that these gapshave ranks ≥ R , because of the characterization of rank (for both kinds of gaps) as the firsttime their images meet the bad set. As for the measure of V ∩ W ¯ b , by Lemma 23(ii) andthe properties of the Jacobian of the holonomy map h : (cid:102) W ∩ ( ∪ x ∈ W ¯ b W s ( x )) → W , we have m W ( V ∩ W ¯ b ) ≤ m (cid:102) W ( h − ( V ∩ W ¯ b )) ≤ m (cid:102) W ( (cid:101) V ).Summing over gaps V of rank ≥ R in W \ W and applying Lemma 24 to the gaps of W ¯ b and (cid:102) W (recalling | (cid:102) W | < | W | ), we obtain m W (cid:0) union of all gaps V ⊂ W \ W of rank ≥ R (cid:1) ≤ (¯ b + ) C (cid:48) g Λ − R | W | and (a) is proved by choosing C g large enough.To prove (b), first make the argument for gaps (cid:101) V of (cid:102) W (which is straightfoward), andthen leverage the connection between V and (cid:101) V via the stable manifolds connecting their endpoints. (cid:3) Recovery of densities. As explained in the Synopsis, the uncoupled part of the measurehas to ‘recover’ and become proper again before it is eligible for the next coupling. Postponingthe full picture to later, we focus here on the situation of the last two subsections, i.e., areference configuration (cid:101) K , a sequence ( K n ) with K n ∈ N ε ( (cid:101) K ) for 0 ≤ n ≤ s (cid:48) , and a magnet S .We assume that a coupling takes place at time 0, and consider the recovery process thereafter.6.4.1. Single measured unstable curve. We treat first the case of a single measured unstablecurve ( W, ν ) making a proper crossing of the magnet, as described in Paragraph B of theSynopsis with n = 0. We denote by ρ and ¯ ρ the densities of ν and of the coupled part of ν respectively. We also let W (cid:48) be the shortest subsegment of W containing W ∩ S , so that W \ W (cid:48) consists of the two excess pieces. As in Paragraph B, we extend ¯ ρ to a regular densitycalled ˇ ρ on W (cid:48) , and decompose the uncoupled part of ρ into densities of the following types:(a) ρ | W \ W (cid:48) ;(b) ( ρ − ˇ ρ ) | W (cid:48) (this will be referred to as the “top density”), and We use the convention that the infimum equals ∞ if it does not exist in N . (c) ˇ ρ | V as V ranges over all the gaps of W .We consider separately each of these densities (counting ˇ ρ | V for different V as different mea-sures), and discuss their recovery times, meaning the time it takes for such a measure to becomeproper (see Definition 17).(a) Since these are regular to begin with, the only reason why they may not be proper is thatthe excess pieces may be too short. Thus their recovery times may depend on | (cid:102) W | (each excesscurve has length ≥ | (cid:102) W | ), but are otherwise uniformly bounded and independent of ( K n ).(c) As discussed in the Synopsis, the density ˇ ρ | V for each V is regular to begin with. Thusrecovery time has only to do with length. For a gap of rank R , recall that the image F R − V is a regular homogeneous unstable curve. Denoting Z = ˇ ν ( V ) / | V | where ˇ ν is the measureon V with density ˇ ρ , we have Z R − / ˇ ν ( V ) = 1 / |F R − V | ≤ C / |F R V | ≤ C c − | (cid:102) W | − Λ R byLemma 10 and (31). In the next step, the curve V R = F R − V will get cut, but we may proceedwith the aid of Lemma 16 and obtain Z n /ν ( V ) ≤ C p for n ≥ R − R log Λ + | log C c − | +2 | log | (cid:102) W || ) / | log ϑ p | . In other words, we have proved Lemma 26. There exists a uniform constant c p > such that the following holds. In thesetting of Lemma 25, after c p ( R + | log | (cid:102) W || ) steps, each one of the measures on the gaps ofrank R will have become a proper measure. (b) We begin by stating a general result. See the Appendix for a proof. Lemma 27. There exists a uniform constant C top > such that the following holds. Given asequence ( K n ) n ≥ , suppose that two regular densities ϕ, ˇ ϕ on the same unstable curve W satisfy b ≤ ϕ/ ˇ ϕ ≤ B for some B > b > everywhere on W , and let ψ = ϕ − ˇ ϕ . Then | log ψ ( x ) − log ψ ( y ) | ≤ C top B + 1 b − θ s ( x,y ) (32) for all x, y ∈ W . By Lemma 20, these assumptions are satisfied for ϕ = ρ and ˇ ϕ = ˇ ρ with b = and B = ζ − e C r where ζ is the fraction of the measure coupled (see Section 5). Even after thedensities become regular, it may take additional time for the measure to become proper. Thenext lemma, proved in the Appendix, is suited for such situations. Lemma 28. There exists a uniform constant ¯ C p > such that the following holds. Givenan admissible sequence ( K n ) n ≥ , suppose ν is a measure on a regular unstable curve W whosedensity ψ satisfies | log ψ ( x ) − log ψ ( y ) | ≤ Cθ s ( x,y ) for some C > C r . Then the push-forward ( F n ) ∗ ν will be a proper measure, if n ≥ ¯ C p ( | log | W || + C ) . From these lemmas, we conclude that Lemma 29. The maximum time it takes for each of the “top” measures to become a propermeasure is ¯ c p + ¯ C p ( | log | (cid:102) W || + | log ζ | ) , where ¯ c p > is another uniform constant. More general initial measures. Let the initial measure µ = (cid:80) α ν α be a proper probabilitymeasure consisting of countably many regular measured unstable curves, and assume that afraction of µ crosses the magnet S at time 0 with (22) holding for some ζ . As explainedin Section 5, precisely ζ units of its mass will be coupled to the reference measure (cid:101) m . Theremaining measure, which we denote by µ , has mass 1 − ζ and consists of the three kinds ofmeasures described earlier. The next result summarizes some of the results above and is veryconvenient for bookkeeping. Lemma 30. There exist constants C ≥ , λ ∈ (0 , , and r > such that, for any m ≥ r , ( F m ) ∗ µ can be split into the sum of two nonnegative measures µ P m and µ G m , both consisting ofcountably many regular measured unstable curves, with the properties that µ P m is proper , and µ G m ( M ) = Cλ m . ISPERSING BILLIARDS WITH MOVING SCATTERERS 27 The constants C and r in the lemma depend on the reference configuration used in theconstruction of the magnet (through | (cid:102) W | ) and r also depends on ζ , but neither of them dependon the initial measure µ or the sequence of configurations. The constant λ is uniform.We will use the notation in Section 5 in the proof – except for omitting the subscript n . Proof of Lemma 30. By Lemma 29, we can choose r large enough that the m -step push-forwardsof the densities on the excess pieces and the “top” densities yield proper measures for all m ≥ r .From that point on the question is about the gaps. Here we turn to Lemmas 25 and 26.By regularity of the measures ν α , (30) implies ν α { x ∈ W α : x is in a gap of rank ≥ R } ≤ C (cid:48)(cid:48) g Λ − R ν α ( W α ) for another uniform constant C (cid:48)(cid:48) g > 0. Summing over α yields the estimate µ { gaps of rank ≥ R } ≤ C (cid:48)(cid:48) g Λ − R . With the aid of Lemma 26, we see that the quantity µ { gaps needing ≥ m steps to yield a proper measure } is bounded above by the expression C (cid:48)(cid:48) g Λ − ( m/c p −| log | (cid:102) W || )+1 ≤ C (cid:48)(cid:48) g Λ | log | (cid:102) W || λ m ≤ Cλ m with C = max(1 , C (cid:48)(cid:48) g Λ | log | (cid:102) W || ) and λ =Λ − /c p ∈ (0 , r large, we may assume Cλ r < − ζ = µ ( M ). We first collect allthe gap measures from ( F r ) ∗ µ into µ G r ; they have total mass ≤ Cλ r . Next, we take a suitableconstant multiple of the remaining, proper, measure from ( F r ) ∗ µ and include it into µ G r so thatfinally µ G r ( M ) = Cλ r holds exactly. (This is mostly for purposes of keeping the statementsclean.) Note that µ P r = ( F r ) ∗ µ − µ G r is proper. By our earlier results, the push-forwards of µ P r remain proper; these will always be included in the measures µ P m for m > r . The real gapmeasures included in µ G r , on the one hand, continue to recover into being proper measures atleast at the rate λ . On the other hand, the proper measures included in µ P r will continue to beproper under push-forwards. Hence, for m > r , we return to µ P m a suitable constant multipleof the proper part of ( F m,r +1 ) ∗ µ G r , as necessary, so that the statements of the lemma continueto hold. (cid:3) Uniform mixing. We discuss here the primary reason behind the asserted exponentialmemory loss for the sequence ( K m ). Since events that occur prior to couplings are involved, wecannot assume that the coupling of interest occurs at n = 0, as was done in Sections 6.2–6.4.Our goal is to address item (1) in Paragraph D of the Synopsis.Recall from Section 6.2 that given a configuration (cid:101) K ∈ K , there exist unstable manifolds (cid:102) W ⊂ (cid:102) W u of (cid:101) F = F (cid:101) K , (cid:101) K , s (cid:48) ∈ Z + and ε (cid:48) > K m ) m ≥ with K n , . . . , K n + s (cid:48) ∈N ε (cid:48) ( (cid:101) K ), a magnet S n with desirable properties (see Sections 6.2 and 6.3) can be constructedout of (cid:102) W and stable manifolds for ( F n + m ) m ≥ . We assume for each (cid:101) K that s (cid:48) , ε (cid:48) and (cid:102) W ⊂ (cid:102) W u are fixed. Proposition 31. Given (cid:101) K ∈ K , there exist ζ > , ε ∈ (0 , ε (cid:48) ) and s ∈ Z + such that the followingholds for every ( K m ) m ≥ with K , . . . , K s + s (cid:48) ∈ N ε ( (cid:101) K ) : Let S s be the magnet defined by (cid:102) W and ( F s + m ) m ≥ . Then every regular measured unstable curve ( W, ν ) with | W | ≥ (2 C p ) − has theproperty that if W s,i = F s ( W − s,i ) are the homogeneous components of F s ( W ) which cross S s properly, then (cid:88) i ( F s ) ∗ ( ν | W − s,i )( W s,i ∩ S s ) ≥ ζe C r ν ( W ) . (33)This proposition asserts that starting from an arbitrary regular measured unstable curve( W, ν ) with | W | ≥ (2 C p ) − , at least a uniform fraction of its F s -image has sufficiently many(homogeneous) components crossing the magnet S s properly provided K m remains in a suf-ficiently small neighborhood N ε ( (cid:101) K ) of the reference configuration (cid:101) K for time ≤ s + s (cid:48) . Thelast s (cid:48) steps is used to make sure that the magnet has a high density of sufficiently long localstable manifolds (see Sections 6.1 and 6.2), whereas the first s is directly related to the mixingproperty of (cid:101) F .Note the uniformity in Proposition 31: the constants depend on (cid:101) K , with s, ζ and ε dependingalso on the choice of (cid:102) W ; but ( K m ) m ≥ is arbitrary as long as it satisfies the conditions above. Proof of Proposition 31. We begin by recalling the following known result on the mixing prop-erty of (cid:101) F n (see e.g. [10]): Let (cid:101) S be the magnet defined by (cid:102) W and powers of (cid:101) F , and let σ > s > ζ (cid:48) > W, ν ) with | W | ≥ (2 C p ) − and any n ≥ s ,(i) finitely many components of (cid:101) F n W cross (cid:101) S super -properly and(ii) denoting these components W super n,i , (cid:88) i ( (cid:101) F n ) ∗ ν ( W super n,i ∩ S (cid:101) K ) ≥ ζ (cid:48) ν ( W ) . (34)Here super-proper crossing means that the crossing is proper with room to spare. Specifically,the excess pieces are twice as long (i.e., at least 2 | (cid:102) W | units), and | ϕ W super n,i − ϕ (cid:102) W | < σ , where (cid:102) W extended by | (cid:102) W | along (cid:102) W u on each side is the graph of ϕ (cid:102) W as a function of r , and W super n,i suitably restricted is the graph of ϕ W super n,i defined on the same r -interval.Let σ be such that for any unstable curve U , | ϕ U − ϕ (cid:102) W | < σ implies that condition (ii) in thedefinition of proper crossing (Definition 22) is satisfied by U independently of the maps usedto define the stable manifolds in the magnet (provided all configurations are in K ). We haveused here the fact that there are uniform stable and unstable cones and that they are boundedaway from each other. Let s be given by the above result for (cid:101) F . Our next step is to view F s asa perturbation of (cid:101) F s , and to argue that the following holds for ε sufficiently small: Let ( W, ν )be as in the proposition, and suppose W super s,i ⊂ (cid:101) F s W crosses (cid:101) S super-properly. Then for every( K m ) with K m ∈ N ε ( (cid:101) K ) for m ≤ s + s (cid:48) , there exists a subcurve V ⊂ (cid:101) V = (cid:101) F − s W super s,i ⊂ W suchthat F s V crosses S s properly.First observe the following facts about (cid:101) F m ( (cid:101) V ): (a) There exists ¯ k ∈ N independent of W such that (cid:101) F m (cid:101) V ⊂ ∪ | k |≤ ¯ k H k \ S (cid:101) K , (cid:101) K for 0 ≤ m < s . This is because for each m , (cid:101) F m (cid:101) V is containedin a homogeneity strip H k , and would be arbitrarily short if k was arbitrarily large, and that isnot possible by Lemma 10 since | (cid:101) F s ( (cid:101) V ) | (cid:38) | (cid:102) W | . (b) Let V ⊂ (cid:101) V be the subsegment with theproperty that (cid:101) F s ( V ) crosses (cid:101) S and the excess pieces have length | (cid:102) W | . By uniform distortionbounds (Lemma 9) and (a) above, there exists δ > W or (cid:101) V such that for1 ≤ m < s , (cid:101) F m ( V ) has distance > δ from ∪ | k |≤ ¯ k ∂ H k ∪ S (cid:101) K , (cid:101) K .We wish to choose ε small enough that (i) for each m = 1 , . . . , s , F m ( V ) and (cid:101) F m ( V ) differ inHausdorff distance by < δ , and (ii) each of the excess pieces of F s ( V ) are > | (cid:102) W | in length,and | ϕ F s ( V ) − ϕ (cid:102) W | < σ . The purpose of (i) is to ensure that F s ( V ) is a single homogeneouscomponent, and (ii) is intended to ensure proper crossing, the providing some room to ac-commodate the slight difference between S s and (cid:101) S (the “end points” of S s ∩ (cid:102) W and (cid:101) S ∩ (cid:102) W may differ by | (cid:102) W | ). It is straightforward to check that (i) and (ii) are assured if each of theconstituent maps F m is sufficiently close to (cid:101) F in C -distance and (cid:101) F has a uniformly boundedderivative on the relevant domain (which is bounded away from the bad set). These propertiescan be guaranteed by taking ε small.To finish the proof, it suffices to show that there exists a constant c > K m ) or on W ) such that if W super s,i and V ⊂ (cid:101) F − s W super s,i are as above, then( F s | V ) ∗ ν ( F s V ∩ S s ) ≥ c · ( (cid:101) F s ) ∗ ν ( (cid:101) F s V ∩ (cid:101) S ) . By Lemma 9, ( F s | V ) ∗ ν ( F s V ∩ S s ) ≥ ν ( V ) e − C r m F s V ( S s ) / |F s V | and ( (cid:101) F s ) ∗ ν ( (cid:101) F s V ∩ (cid:101) S ) ≤ ν ( V ) e C r m (cid:101) F s V ( (cid:101) S ) / | (cid:101) F s V | . ISPERSING BILLIARDS WITH MOVING SCATTERERS 29 Since | (cid:101) F s V | ≥ m (cid:101) F s V ( (cid:101) S ) and |F s V | is uniformly bounded from above, it remains to show that m F s V ( S s ) is uniformly bounded from below, and that is true by the absolute continuity ofstable manifolds in S s (Lemma 11) and the fact that m (cid:102) W ( S s ∩ (cid:102) W ) ≥ | (cid:102) W | . (cid:3) We remark that s and ε in Proposition 31 depend strongly on (cid:101) K but are independent of ( K m )or W . The argument is a perturbative one, and it is feasible only because it does not involvemore than a finite, namely s , number of iterates. We remark also that stronger estimates on ζ than the one above can probably be obtained by leveraging the C proximity of F m to (cid:101) F ,The next result extends Proposition 31 to more general initial measures and coupling times.See Sections 4.3 and 4.4 for definitions. Corollary 32. Let (cid:101) K be fixed, and let s, s (cid:48) , ε and ζ be as in Proposition 31. Then the followingholds for every n ∈ Z + with n ≥ s + n p and every sequence ( K m ) satisfying K m ∈ N ε ( (cid:101) K ) for n − s ≤ m ≤ n + s (cid:48) : Let µ be an initial probability measure that is regular on unstable curvesand proper (i.e., Z < C p ). Then (22) holds for ( F n ) ∗ µ with ζ = ζ .Proof. The discussion immediately following Definition 17 is relevant here. Since n − s ≥ n p ,the measure ( F n − s ) ∗ µ is again proper; thus at least half of ( F n − s ) ∗ µ can be disintegrated intomeasures supported on regular unstable curves of length ≥ (2 C p ) − . Then Proposition 31 canbe applied, giving (33) with a factor of on the right side. (cid:3) Proof of Theorem 4: the countable case The purpose of this section is to go through the proof of the countable case of Theorem 4from beginning to end, connecting the individual ingredients discussed in the last section. Forinitial distributions, we start from the most general kind permitted in this paper, namely thoseintroduced in Section 4.4 under Theorem 1’. It was observed in Section 4 that by delaying thefirst coupling, each initial distribution can be assumed to be regular on unstable curves andproper. For simplicity, we will start from that. Also, as noted before, it suffices to consider asingle initial distribution, for the two measures will be coupled to reference measures and henceto each other.7.1. Coupling times. To each (cid:101) K ∈ K , we first assign values to the constants ˜ ε ( (cid:101) K ) and (cid:101) N ( (cid:101) K )appearing in the formulation of the theorem. Namely, we set ˜ ε ( (cid:101) K ) = ε, s ( (cid:101) K ) = s and s (cid:48) ( (cid:101) K ) = s (cid:48) where ε, s and s (cid:48) are as in Proposition 31, and let r = r ( (cid:101) K ) be the maximum of the similarlynamed constant in Lemma 30 and of s (cid:48) ( (cid:101) K ). We then set (cid:101) N ( (cid:101) K ) = s ( (cid:101) K ) + r ( (cid:101) K ). For future use,let ζ = ζ ( (cid:101) K ) be as in Proposition 31, and C = C ( (cid:101) K ) and λ as in Lemma 30.Next, we fix reference configurations ( (cid:101) K q ) Qq =1 with (cid:101) K q +1 ∈ N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ) for 1 ≤ q < Q anda sequence ( K n ) Nn =0 adapted to ( (cid:101) K q , ˜ ε ( (cid:101) K q ) , (cid:101) N ( (cid:101) K q )) Qq =1 . Such a sequence is admissible. If thesequence is finite ( N < ∞ ), augment it to an infinite one by setting K n = K N for all n > N (so stable manifolds are well defined).The following are considerations in our choice of coupling times.(a) Suppose the k th coupling occurs at time t k and n q − ≤ t k ≤ n q . We require that for m ∈ [ t k − s ( (cid:101) K q ) , t k + r ( (cid:101) K q )] , K m ∈ N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ).(b) There exists ∆ = ∆(( (cid:101) K q ) Qq =1 ) such that t k +1 − t k ≤ ∆.(c) There exists ∆ = ∆ (( (cid:101) K q ) Qq =1 ) > t k +1 − t k ≥ ∆ .The reasons for (a) are explained in Sections 6.2–6.5. The purpose of (b) is to ensure the expo-nential estimate in the theorem: at most a fraction ζ of the still uncoupled measure is matched per coupling . The reason for (c) is a little more subtle: it is not necessarily advantageous tocouple as often as one can, because each coupling matches a ζ -fraction of the measure thatis available for coupling, but renders at the same time a fraction Cλ m of it improper, henceunavailable for coupling in the near future. Intuitively at least, it may be meaningful, especially if C and λ are large, to wait till a sufficiently large part of the uncoupled measure has recovered,i.e., has rejoined µ P m , before performing the next coupling. This is discussed in more detail inSection 7.3.There are many ways to choose t k . Postponing the choice of ∆ to Section 7.3 (and assumingfor now it is a preassigned number), an algorithm may go as follows: Start by fitting into eachtime interval ( n q − , n q ) , q = 1 , . . . , Q , as many disjoint subintervals of length s ( (cid:101) K q ) + r ( (cid:101) K q )as one can; by definition, at least one such interval can be fitted into each ( n q − , n q ). Labelthese intervals as J i = [ t (cid:48) i − s ( (cid:101) K q ) , t (cid:48) i + r ( (cid:101) K q )] with t (cid:48) < t (cid:48) < · · · . This is not quite our desiredsequence of coupling times yet, as it need not respect (c) above. To fix that, we let t = t (cid:48) ,and let t = t (cid:48) i where i > t − t ≥ ∆ . Continuing, we let t = t (cid:48) i where i is the smallest number such that t (cid:48) i > t and t − t ≥ ∆ , and so on. Then (a)is satisfied by definition, and we check that (b) is satisfied with∆ = 2 max ≤ q ≤ Q s ( (cid:101) K q ) + 2 max ≤ q ≤ Q r ( (cid:101) K q ) + ∆ . The coupling procedure and recovery of densities. We review the procedure briefly,setting some notation at the same time. Once the coupling times t k have been fixed, weconstruct, for each k , a magnet S t k using the reference configuration (cid:101) K q if n q − < t k < n q . Let r k = r ( (cid:101) K q ), s k = s ( (cid:101) K q ), ζ k = ζ ( (cid:101) K q ), and C k = C ( (cid:101) K q ).Suppose at time n = t k − s k we have at our disposal a proper measure ˜ µ k with total mass P k = ˜ µ k ( M ), ready to be used in the k th coupling. In particular, ˜ µ = µ and P = 1, sincethe initial probability measure µ is assumed to be proper. By Corollary 32, a ζ k -fraction of( F t k ,t k − s k +1 ) ∗ ˜ µ k is coupled to the reference measure (cid:101) m t k on S t k . In the language of Section 5,Paragraph A, ¯ µ t k is the part of µ such that ( F t k ) ∗ ¯ µ t k is equal to the part of ( F t k ,t k − s k +1 ) ∗ ˜ µ k coupled at time t k ; in particular, ¯ µ t k ( M ) = ζ k P k . (35)The uncoupled part of ( F t k ,t k − s k +1 ) ∗ ˜ µ k consist of a countable family of measured unstablecurves, including arbitrarily short gaps among others, as discussed several times earlier, andhas total mass (1 − ζ k ) P k . With the aid of Lemma 30 (with ˜ µ k /P k in the role of µ ), we identifyout of its push-forward under F t k + m,t k +1 a proper part, and call the rest the “non-proper” part,with the latter rejoining the first at a certain rate. Deviating from the notation of the lemma inorder not to overburden the notation, we denote these parts ˜ µ P k,m and ˜ µ G k,m , respectively (“P”for proper, “G” for gap). It can be arranged so that˜ µ P k,m ( M ) = (1 − ζ k − C k λ m ) P k and ˜ µ G k,m ( M ) = C k λ m P k ( m ≥ r k ) . (36)7.3. Bookkeeping and exponential bounds. Letting u k = ( t k +1 − s k +1 ) − t k ≥ r k , the totalmass P k +1 of the proper measure available for coupling at time t k +1 satisfies P k +1 = ˜ µ P k,u k ( M ) + k − (cid:88) j =1 (cid:16) ˜ µ G j,t k − + u k − − t j ( M ) − ˜ µ G j,t k + u k − t j ( M ) (cid:17) . The first term comes directly from the k th coupling, as explained above. In the second termwe take into account the fact that at the j th coupling, 1 ≤ j < k , some measure was depositedinto the “non-proper” part, and what remains of that part at a later time n is the measure˜ µ G j,n − t j . Thus this sum represents the total mass that was not available for the k th couplingbut has become available for the ( k + 1)st. Plugging in the numbers from (36), we obtain P k +1 = (1 − ζ k − C k λ u k ) P k + k − (cid:88) j =1 C j λ t k − − t j + u k − (1 − λ u k + s k ) P j . (37) ISPERSING BILLIARDS WITH MOVING SCATTERERS 31 We also have the following expression for the total mass that remains uncoupled immediatelyafter the k th coupling, i.e., µ t k ( M ) in the language of Section 5, Paragraph A: µ t k ( M ) = (1 − ζ k ) P k + k − (cid:88) j =1 C j λ t k − − t j + u k − P j . (38)Here the first term is the measure that was “eligible” for coupling at time t k but was notcoupled, and the second sum consists of terms coming from earlier couplings that at time t k − + u k − = t k − s k were still not ready to be coupled.Notice that (37) is a recursion relation for the sequence ( P k ) with the initial condition P = 1.We need to show that P k tends to zero exponentially with k . Lemma 33. Let ˜ ζ = min ≤ q ≤ Q ζ ( (cid:101) K q ) . Suppose ∆ = (cid:24) log (cid:16) ˜ ζ (1 − ˜ ζ ) (cid:14) max ≤ q ≤ Q C ( (cid:101) K q ) (cid:17) / log λ (cid:25) + max ≤ q ≤ Q s ( (cid:101) K q ) + n p . Then a coupling strategy satisfying (a) and (c) in Section 7.1 will produce a sequence of P k with P k ≤ (1 − ˜ ζ ) k ∀ k ≥ . (39)We have included n p in the definition of ∆ to allow for the transient loss of properness whena proper measure is pushed forward (Section 4.3); it plays no role in the proof below. Proof of Lemma 33. We form a majorizing sequence ( Q k ) with Q = 1 = P and Q k +1 = (1 − ˜ ζ ) Q k + k − (cid:88) j =1 C j λ t k − − t j + u k − Q j . (40)Clearly P k ≤ Q k by comparing (37) and (40). We want to show that Q k tends to zero expo-nentially with k , with Q k +1 ≤ (1 − ˜ ζ ) Q k . The bound is certainly implied if k − (cid:88) j =1 C j λ t k − − t j + u k − Q j ≤ ˜ ζQ k , k ≥ , and this is what we will prove.Observe from t k +1 − t k ≥ ∆ and the choice of ∆ above thatmax ≤ q ≤ Q C ( (cid:101) K q ) · λ t k +1 − t k − max ≤ q ≤ Q s ( (cid:101) K q ) ≤ ˜ ζ (1 − ˜ ζ ) . (41)Together with C k − ≥ 1, this gives k − (cid:88) j =1 C j λ t k − + u k − − t j Q j = C k − λ u k − (cid:32) Q k − + λ s k − C k − k − (cid:88) j =1 C j λ t k − + u k − − t j Q j (cid:33) ≤ C k − λ u k − (cid:32) Q k − + k − (cid:88) j =1 C j λ t k − + u k − − t j Q j (cid:33) ≤ ˜ ζ (1 − ˜ ζ ) (cid:32) Q k − + k − (cid:88) j =1 C j λ t k − + u k − − t j Q j (cid:33) ≤ ˜ ζQ k . The second to last inequality uses (41), and the last inequality is from (40). Hence, P k ≤ Q k ≤ (1 − ˜ ζ ) k − Q for k ≥ 2. Finally, Q = (1 − ˜ ζ ) and P = 1, yield P k ≤ (1 − ˜ ζ ) k − for all k ≥ (cid:3) Corollary 34. For any n ≥ , ¯ µ n ( M ) ≤ ˜ ζ (1 − ˜ ζ ) n/ ∆ and µ n ( M ) ≤ (1 − ˜ ζ ) n/ ∆ − . Proof. By (41) and C k − ≥ C j λ t k − t j ≤ (cid:0) ˜ ζ (1 − ˜ ζ ) (cid:1) k − j in (38). Insert-ing also the exponential bound on P k in (39) and computing the resulting sum yields easily µ t k ( M ) ≤ (1 − ˜ ζ ) k +1 . Recalling (35), also ¯ µ t k ( M ) ≤ ˜ ζ (1 − ˜ ζ ) k . Observe that t k ≤ k ∆ by t j +1 − t j ≤ ∆ and t ≤ ∆. Thus, µ t k ( M ) ≤ (1 − ˜ ζ ) t k / ∆+1 and ¯ µ t k ( M ) ≤ ˜ ζ (1 − ˜ ζ ) t k / ∆ .By definition, µ n = µ t k and ¯ µ n = 0 for t k < n < t k +1 ; see Section 5. We therefore obtain¯ µ n ( M ) ≤ ˜ ζ (1 − ˜ ζ ) n/ ∆ for all n ≥ µ n ( M ) ≤ (1 − ˜ ζ ) t k / ∆+1 for t k ≤ n < t k +1 ( k ≥ t j +1 − t j ≤ ∆ once more, we see in the latter case that t k ≥ n − ∆. In other words, µ n ( M ) ≤ (1 − ˜ ζ ) n/ ∆ for all n ≥ t , or µ n ( M ) ≤ (1 − ˜ ζ ) n/ ∆ − for all n ≥ t ≤ ∆). (cid:3) Memory-loss estimate. Finally, we prove the estimate in (2), along the lines of (17).Consider two measures µ i , i = 1 , 2. Recalling that µ i = µ in/ + (cid:80) j ≤ n/ ¯ µ ij and using Lemma 34,we see that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d µ − (cid:90) f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ (1 − ˜ ζ ) n/ − (cid:107) f (cid:107) ∞ + (cid:88) j ≤ n/ (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d¯ µ j − (cid:90) f ◦ F n d¯ µ j (cid:12)(cid:12)(cid:12)(cid:12) . Here ¯ µ j = ¯ µ j = 0 unless j = t k for some k = 1 , , . . . . At times j = t k a coupling occurs:Recall from Paragraph C of Section 5 that (for each i = 1 , 2) ( F j ) ∗ ¯ µ ij is coupled to the referencemeasure a j (cid:101) m j ( · ∩ S j ), where a j = ¯ µ ij ( M ) = (( F j ) ∗ ¯ µ ij )( S j ) and (cid:101) m j ( S j ) = 1. Moreover, ac-cording to (25) and (23), the measure ( F j ) ∗ ¯ µ ij is a sum of countably many components, namely (cid:80) α ∈A i (cid:80) m ∈I iα,j λ iα,j,m ( h iα,j,m ) ∗ (cid:101) m j , where h iα,j,m is a holonomy map associated to the stable man-ifolds of the magnet S j and (cid:80) α ∈A i (cid:80) m ∈I iα,j λ iα,j,m = ¯ µ ij ( M ). If f is γ -H¨older continuous withsome exponent γ > 0, we can estimate, similarly to (18), that (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d¯ µ j − (cid:90) f ◦ F n d¯ µ j (cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n,j +1 d(( F j ) ∗ ¯ µ j ) − (cid:90) f ◦ F n,j +1 d(( F j ) ∗ ¯ µ j ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:88) i =1 , (cid:88) α ∈A i (cid:88) m ∈I iα,j λ iα,j,m (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:90) S j f ◦ F n,j +1 d(( h iα,j,m ) ∗ (cid:101) m j ) − (cid:90) S j f ◦ F n,j +1 d (cid:101) m j (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ (cid:88) i =1 , (cid:88) α ∈A i (cid:88) m ∈I iα,j λ iα,j,m | f | γ (ˆ c − Λ − ( n − j ) ) γ = (cid:88) i =1 , ¯ µ ij ( M ) | f | γ (ˆ c − Λ − ( n − j ) ) γ . Since (cid:80) j ¯ µ ij ( M ) = 1, (cid:88) j ≤ n/ (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d¯ µ j − (cid:90) f ◦ F n d¯ µ j (cid:12)(cid:12)(cid:12)(cid:12) ≤ | f | γ (ˆ c − Λ − n/ ) γ . Combining the above estimates yields the bound in (2) with C γ = 2 max((1 − ˜ ζ ) − , ˆ c − γ ) and θ γ = max((1 − ˜ ζ ) / , Λ − γ / ζ are determined by the set of reference configurations { (cid:101) K q } Qq =1 ; their order and number of appearances are irrelevant. This completes the proof of thecountable case of Theorem 4. (cid:3) Completing the proofs In this section we complete the proofs of Theorems 1’–2’ and 4’, restricted versions of whichare stated in Sections 2.2 and 2.3, and the full versions in Section 4.4. To do that, we mustfirst treat the “continuous case” of Theorem 4, which is used to give the full versions of all theother results. ISPERSING BILLIARDS WITH MOVING SCATTERERS 33 Proof of Theorem 4: continuous case. In Section 4.3, we introduced the idea ofmeasured unstable families, defined to be convex combinations of measured unstable stacks.The “continuous case” of Theorem 4 refers to the version of Theorem 4 for which initial measuresare of this form. The proof proceeds almost exactly as in the countable case, so we focus hereonly on the differences.Three types of processes are involved in the proof of Theorem 4: (i) canonical subdivisions,(ii) coupling to reference measures, and (iii) recovery of densities following the couplings. Theprocess of pushing forward measured unstable families and canonical subdivisions was discussedin Section 4.3. We noted that this process produces objects of the same kind, i.e., canonicalsubdivisions of measured unstable families are measured unstable families. Regularity andproperness, including their recovery properties, were also discussed: there is no substantivedifference between the countable and continuous cases since these are essentially propertieson individual unstable curves; in the continuous case, one simply replaces summations in thecountable case by integrals.We provide below more detail on (ii): The coupling procedure: continuous case. Consider the situation in Section 7.2, where at time n = t k − s k we have a proper measure ˜ µ k of continuous type ready to be used in the k thcoupling. We sketch below a few issues that require additional care:In the countable case, we observed in Corollary 32 that at least half of ˜ µ k is supported onunstable curves of length ≥ (2 C p ) − ; the same is true here, as it is a general fact. But then inthe countable case, we applied Proposition 31 to one unstable curve at a time , comparing theaction of F s to that of (cid:101) F s on each curve to obtain the asserted bound on the fraction that canbe coupled. Here it is not legitimate to argue one curve at a time, so we proceed as follows:Noting that ˜ µ k is supported on a countable number of unstable stacks, we plan to subdividethese stacks in such a way that there is a a collection of countably many “thin enough” stackswith the following properties: (i) their union carries at least half of ˜ µ k , and (ii) on each thinstack all the curves have length at least (2 C p ) − (or thereabouts). We then treat these thinstacks with long curves one at a time. The conditions for “thin enough” are basically that thestack should behave as though it was a single curve in the next s k steps.More precisely, pick one of the measured unstable stacks ( ∪ α ∈ E W α , µ ) associated to ˜ µ k , andconsider its canonical s k -step subdivision (associated to the sequence F t k − s k +1 , . . . , F t k ) intostacks of the form ∪ α ∈ E sk,i,j ( W α ∩ D s k ,i ) as discussed in Section 4.3. We first specify what wemean by a “thin enough” substack of ∪ α ∈ E W α . Let ˆ α ∈ E be such that | W ˆ α | ≥ (2 C p ) − ,and assume the images of this unstable curve in the next s k steps do not pass through branchpoints of the discontinuity set. Then there is a small neighborhood E ˆ α of ˆ α in E such thatthe following holds for all β ∈ E ˆ α : For each i such that F t k ,t k − s k +1 ( W ˆ α ∩ D s k ,i ) crosses S t k properly, the same holds for F t k ,t k − s k +1 ( W β ∩ D s k ,i ) with a slightly relaxed definition of “propercrossing” that is good enough for our purposes. Moreover, if ˆ α ∈ E s k ,i,j , then E ˆ α ⊂ E s k ,i,j . Weare guaranteed that E ˆ α exists because there are only finitely many such proper crossings foreach W ˆ α . The stack ∪ α ∈ E ˆ α W α is “thin enough”.Assuming that the transverse measure P on E has no atoms (the argument is easily modifiedif it does), there is a finite number of disjoint intervals of the form E ˆ α l where | W ˆ α l | ≥ (2 C p ) − and ∪ α ∈∪ l E ˆ αl W α carries more than 99% of the part of µ supported on W α -curves of length ≥ (2 C p ) − . The procedure is to first subdivide E into { E ˆ α l } and the connected components of E \ ∪ l E ˆ α l . This corresponds to subdividing the original stack ( ∪ α ∈ E W α , µ ) before proceedingwith the canonical subdivision. At time t k , we consider one l at a time: For each i such that F t k ,t k − s k +1 ( W ˆ α l ∩ D s k ,i ) crosses S t k properly, the F t k ,t k − s k +1 -image of ∪ α ∈ E sk,i ∩ E ˆ αl ( W α ∩ D s k ,i ) isa single unstable stack every curve in which crosses S t k properly. A fraction of the conditionalprobability measures on each unstable curve is coupled to (cid:101) m t k as before. These are the onlystacks on which couplings will be performed at time t k . To obtain the desired lower bound on the fraction of ( F t k ,t k − s k +1 ) ∗ ˜ µ k coupled, we prove a slightgeneralization of Proposition 31 in which the measured stack ( ∪ α ∈ E ˆ αl W α , µ | ∪ α ∈ E ˆ αl W α ) takesthe place of ( W, ν ). The argument is virtually identical (and omitted); since the conditionalmeasures have the same uniform bounds.After a coupling, we must also show that the uncoupled part of ( F t k ,t k − s k +1 ) ∗ ˜ µ k is againsupported on at most a countable number of measured unstable stacks. Treating first the curves(without the measures), we observe that for each i and l in the next to last paragraph, afterthe coupling there are two stacks corresponding to the excess pieces of F t k ,t k − s k +1 ( W ˆ α l ∩ D s k ,i ),a third stack which is F t k ,t k − s k +1 ( ∪ α ∈ E sk,i ∩ E ˆ αl ( W α ∩ D s k ,i )) minus the first two, plus a countablenumber of stacks one for each gap. We also need to decompose the uncoupled part of themeasure in the same way as was done in Section 6.4. In particular, a slight generalization ofthe extension lemma (Lemma 15) leading to the “top conditional densities” in the third stackis needed. We leave this technical but straightforward exercise to the reader.Finally, we observe that the subdivision of a stack into thinner stacks (without cutting anyof the unstable curves in the stack) does not increase the Z -value of a family.The rest of the proof is unchanged from the countable case.This concludes the proof of Theorem 4’, that is, the extension of Theorem 4 to the largerclass of initial measures permitted in Theorem 1’. Theorem 1’ then follows, in the same wayas Theorem 1 was deduced from Theorem 4; see Section 2.8.2. Scatterers with variable geometries. To understand what additional arguments areneeded as we go from scatterers with fixed geometries to scatterers with variable geometries,recall that the proof of Theorem 1’ has two distinct parts: one is local , and the other global .The local result is contained in Theorem 4, which treats essentially time-dependent sequences( K n ) near a fixed reference configuration (cid:101) K . It also shows how the scheme can be continuedas the time-dependent sequence moves from the sphere of influence of one reference configu-ration to that of another. The rate of memory loss, however, depends on the set of referenceconfigurations visited. The global part of the proof seeks to identify a suitable space, as largeas possible, for which one can have a uniform convergence rate for the measures involved. Forscatterers with fixed geometries, this is done by showing that the entire configuration spaceof interest can be “covered” by a finite number of reference configurations { (cid:101) K , . . . , (cid:101) K Q } , i.e.,no matter how long the time-dependent sequence, it is, at any one moment in time, always“within radar range” of one of the (cid:101) K q , ≤ q ≤ Q . The argument is thus reduced to the localone. Details are given at the end of Section 2. Proof of Theorem 2’. We discuss separately the local and global parts of the argument. Local part: We claim that the local part of the proof, i.e., Theorem 4’, extends verbatim to thesetting of variable scatterer geometry, and leave the step-by-step verification to the reader. Forexample, the arguments in Sections 3 and 4 are entirely oblivious to the fact that the shapesof the scatterers change with time, in the same way that they are oblivious to their changinglocations, for as long as their curvatures and flight times lie within specified ranges. The moresensitive parts of the proof involve ( K m ) ⊂ N ε ( (cid:101) K ), where N ε ( · ) is now defined using the d -metric introduced in Section 2.2. Notice that as before, (i) for K , K (cid:48) ∈ N ε ( (cid:101) K ), the singularityset for F = F K (cid:48) , K lies in a small neighborhood of the singularity set for (cid:101) F = F (cid:101) K , (cid:101) K , and (ii) afixed distance away from these singularity sets, F and (cid:101) F can be made arbitrary close in C as ε → 0. These properties are sufficient for the arguments needed, including the uniform mixingargument in Section 6.2. Global part: The argument is along the lines of the one at the end of Section 2, but involvesdifferent spaces and different norms. In order to reduce to the local argument, we need toestablish some compactness. Decreasing ¯ κ min and ¯ τ min , increasing ¯ κ max and ¯ τ max , as well as ISPERSING BILLIARDS WITH MOVING SCATTERERS 35 increasing ∆ to some ∆ (cid:48) ≥ ∆ (to be fixed below), we let (cid:101) K (cid:48) denote the configuration spacedefined analogously to (cid:101) K but using these relaxed bounds on curvature and flight times. Wedenote the closure of (cid:101) K with respect to the metric d by c(cid:96) ( (cid:101) K ). We will show that c(cid:96) ( (cid:101) K ) is acompact subset of (cid:101) K (cid:48) .First, a constant ˆ∆ can be fixed so that for all K ∈ (cid:101) K , if ˆ γ i : S → T is the constant speedparametrization of ∂ B i in Section 2.2, then (cid:107) D k ˆ γ i (cid:107) ∞ ≤ ˆ∆ for 1 ≤ k ≤ D ˆ γ i ) ≤ ˆ∆.This is true because of property (i) in the definition of (cid:101) K and the fact that derivatives of ˆ γ i and γ i (unit speed parametrization of the same scatterer) differ only by a factor equal to thelength of ∂ B i , which is uniformly bounded above and below due to ¯ κ min < κ < ¯ κ max . Next weargue that if the number of scatterers s were fixed, it would follow that c(cid:96) ( (cid:101) K ) is a compact set:Given s sequences (ˆ γ i,n ) n ≥ of parametrizations as above, we first note that they are uniformlybounded. The same is true of the sequences ( D k ˆ γ i,n ) n ≥ , 1 ≤ k ≤ 3, as noted above. Each ofthese is also equicontinuous because it is uniformly Lipschitz. Hence, the Arzel`a–Ascoli theoremyields the existence of uniform limits ˆ γ i ≡ lim j →∞ ˆ γ i,n j and ˆ γ ( k ) i ≡ lim j →∞ D k ˆ γ i,n j , 1 ≤ k ≤ ≤ i ≤ s , along a subsequence ( n j ) j ≥ . Since Lipschitz constants are preserved in uniformlimits, it is easy to check that ˆ γ ( k ) i = D k ˆ γ i , max ≤ k ≤ (cid:107) D k ˆ γ i (cid:107) ∞ ≤ ˆ∆ and Lip( D ˆ γ i ) ≤ ˆ∆.We now replace the limit parametrizations ˆ γ i , 1 ≤ i ≤ s , by the corresponding constant speedparametrizations γ i . Owing to the above bounds, they specify a configuration in (cid:101) K (cid:48) , if we choose∆ (cid:48) large enough. While (cid:101) K (cid:48) permits in principle an arbitrarily large number of scatterers, thereis, in fact, a finite upper bound on s imposed by ¯ τ min , which is less than or equal to the minimumdistance between any pair of scatterers. We have thus proved that c(cid:96) ( (cid:101) K ) is compact.We apply the result from the local part to (cid:101) K (cid:48) , obtaining ˜ ε ( K ) and (cid:101) N ( K ) for each K ∈ (cid:101) K (cid:48) .The collection {N ˜ ε ( K ) ( K ) : K ∈ c(cid:96) ( (cid:101) K ) } is an open cover of c(cid:96) ( (cid:101) K ), open as subsets of (cid:101) K (cid:48) . Let (cid:8) (cid:101) N q = N ˜ ε ( (cid:101) K q ) ( (cid:101) K q ) , q ∈ Q (cid:9) be a finite subcover. The rest of the proof is as in Section 2: weapply the local result to the given sequence ( K n ) ⊂ (cid:101) K , noting that any ( K n ) with d ( K n , K n +1 )sufficiently small is adapted to a sequence of reference configurations chosen from { (cid:101) N q , q ∈Q} . (cid:3) Proof of Theorem 3. Without loss of generality, we assume that (cid:82) g d µ = 0. Let a = 1 − inf g .Then d µ (cid:48) = a − ( g + a ) d µ is a probability measure. Moreover, the density ρ (cid:48) = a − ( g + a )satisfies the assumptions of Theorem 2. Indeed,(1 + (cid:107) g (cid:107) ∞ ) − ≤ ρ (cid:48) ≤ (cid:107) g (cid:107) ∞ , and, like g , ρ (cid:48) is -H¨older with its logarithm satisfying the estimate | log ρ (cid:48) ( x ) − log ρ (cid:48) ( y ) | ≤ (1 + (cid:107) g (cid:107) ∞ ) | g ( x ) − g ( y ) | ≤ (1 + (cid:107) g (cid:107) ∞ ) | g | d M ( x, y ) . We thus have (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n · g d µ (cid:12)(cid:12)(cid:12)(cid:12) = a (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) f ◦ F n d µ (cid:48) − (cid:90) f ◦ F n d µ (cid:12)(cid:12)(cid:12)(cid:12) ≤ (1 + (cid:107) g (cid:107) ∞ ) C γ ( (cid:107) f (cid:107) ∞ + | f | γ ) θ nγ , after an application of Theorem 2 with µ = µ (cid:48) and µ = µ . Here C γ depends on the bound(1 + (cid:107) g (cid:107) ∞ ) | g | on the H¨older constant of log ρ (cid:48) obtained above. (cid:3) Small external fields. In this section we discuss modifications of earlier proofs neededto yield Theorem E.First we claim that for each (cid:101) K ∈ (cid:101) K , there exist ˆ δ ( (cid:101) K ) and E ( (cid:101) K ) > K , K (cid:48) ∈ N ˆ δ ( (cid:101) K ) and E ∈ C with (cid:107) E (cid:107) ∞ ≤ E ( (cid:101) K ), F E K (cid:48) , K is defined, and ¯ τ min < τ E K (cid:48) , K < t .Here τ E K (cid:48) , K is the flight time between source and target scatterers following trajectories definedby E . To prove the asserted upper bound for τ E K (cid:48) , K , notice that (i) the set of straight linesegments of length t is compact, and (ii) for a C -small E , particle trajectories deviate onlyslightly from straight lines. Thus the ( t , ϕ )-horizon property of the E = 0 case guarantees that any flow-trajectory of length t will also meet a scatterer at an angle not much below ϕ (measured from the tangent).Next we claim that there exist ˆ δ ( (cid:101) K ) ≤ ˆ δ ( (cid:101) K ) and E ( (cid:101) K ) ≤ E ( (cid:101) K ) such that the basicproperties in Sections 3 and 4 hold (with relaxed constants) for all sequences ( K n , E n ) with theproperty that for each n , there is (cid:101) K such that K n , K n +1 ∈ N ˆ δ ( (cid:101) K ) and (cid:107) E n (cid:107) C ≤ E ( (cid:101) K ). Moreprecisely, we claim that the maps F n = F E n K n +1 , K n have the same properties as their analogswith E = 0, including the geometry of the singularity sets, stable and unstable cones, uniformexpansion and contraction rates, distortion and curvature bounds for unstable curves, absolutecontinuity and bounds on the Jacobians, the Growth Lemma holds, etc. For fixed scatterers,the main technical references for fields E with small enough C -norms are [5, 7]. The resultsabove are obtained following the proofs in these references, except for Lemma 7 the proof ofwhich is also straightforward and left as an exercise.Next we proceed to the analog of Theorem 4’ for small external fields, for sequences of the form( K n , E n ) Nn =0 adapted, in a sense to be defined, to a finite sequence of reference configurations( (cid:101) K q ) q ≤ Q : For each q , there exist ˜ ε ( (cid:101) K q ) , ˜ ε field ( (cid:101) K q ) > (cid:101) N ( (cid:101) K q ) ∈ Z + such that ( K n ) Nn =0 is adapted to (cid:0) (cid:101) K q , ˜ ε ( (cid:101) K q ) , (cid:101) N ( (cid:101) K q ) (cid:1) Qq =1 in the sense of Section 2.3 and, additionally, (cid:107) E n (cid:107) C ≤ ˜ ε field ( (cid:101) K q ) for the relevant q . The argument proceeds as in Sections 6 and 7. There are exactlytwo places where the argument is perturbative, and “perturbative” here means perturbing fromsystems with fixed scatterer configurations (cid:101) K and zero external field. One is the constructionof the magnet in Section 6.2, and the other is the uniform mixing argument (Proposition 31)in Section 6.5. For each (cid:101) K ∈ (cid:101) K , these two arguments impose bounds ˜ ε ( (cid:101) K ) and ˜ ε field ( (cid:101) K ) > d ( K n , (cid:101) K ) and (cid:107) E n (cid:107) C respectively. (We may assume ˜ ε ( (cid:101) K ) ≤ ˆ δ ( (cid:101) K ) and ˜ ε field ( (cid:101) K ) ≤ E ( (cid:101) K ).)Such bounds exist because in Section 6.2 we require only that the action of F n + s (cid:48) ,n on a specificpiece of unstable curve follows that of ( F (cid:101) K , (cid:101) K ) s (cid:48) closely in the sense of C for a fixed numberof iterates, namely s (cid:48) , during which this curve stays a positive distance from any discontinuitycurve or homogeneity lines. The argument in Section 6.5 requires a little more, but that tooinvolves only curves that stay away from discontinuity and homogeneity lines and also for only afixed number of iterates. For appropriate choices of ˜ ε ( (cid:101) K ) and ˜ ε field ( (cid:101) K ), the latter made possibleby our extended version of Lemma 7, these two proofs as well as others needed go throughwithout change, yielding an analog of Theorem 4’ as formulated above.Finally it remains to go from our “local” result, i.e., the analog of Theorem 4’, to the “global”one, namely Theorem E. We cover the closure of (cid:101) K with balls centered at each (cid:101) K having d -radius ˜ ε ( (cid:101) K ) in a slightly enlarged space (cid:101) K (cid:48) and choose as before a finite subcover, consisting ofballs centered at { (cid:101) K j } . The uniform bounds ε E and ε appearing in the statement of Theorem Eare given by ε E = min j ˜ ε field ( (cid:101) K j ) and ε = min j ˜ ε ( (cid:101) K j ). Appendix. Proofs Proof of Lemma 7. We first prove continuity of the map ( x, K , K (cid:48) ) (cid:55)→ F K (cid:48) , K ( x ). Consider aninitial configuration K and a target configuration K (cid:48) and some initial condition x ∈ M whichcorresponds to a non-tangential collision . Obviously, there exists an open neighborhood U of the triplet ( x , K , K (cid:48) ) in which there are no tangential collisions: each ( x, K , K (cid:48) ) ∈ U corresponds to a head-on collision from a scatterer B in configuration K to a scatterer B (cid:48) inconfiguration K (cid:48) . We can view the scatterers B and B (cid:48) as subsets of the plane and representthem by two vectors, ( c , u ) and ( c (cid:48) , u (cid:48) ) in R × S , which depend continuously on K and K (cid:48) (as long as ( x, K , K (cid:48) ) ∈ U ). The c and u components specify the location and orientation ofthe scatterer, as was explained in Section 2. Let (¯ c , ¯ u ) be the relative polar coordinates of B with respect to the frame attached to B (cid:48) whose origin is specified by ( c (cid:48) , u (cid:48) ). Then (¯ c , ¯ u )depends continuously on the pair ( K , K (cid:48) ). We write (¯ c , ¯ u ) = G ( K , K (cid:48) ) and point out thatid M × G : ( x, K , K (cid:48) ) (cid:55)→ ( x, G ( K , K (cid:48) )) is continuous on U . Recall that x ∈ M represents theinitial condition in the intrinsic (phase space) coordinates of B . Let (¯ q, ¯ v ) be its projection to ISPERSING BILLIARDS WITH MOVING SCATTERERS 37 the plane, expressed relative to the frame attached to B (cid:48) . The map ¯ π : ( x, ¯ c , ¯ u ) (cid:55)→ (¯ q, ¯ v ) isclearly continuous. Given any plane vector (¯ q, ¯ v ) expressed relative to the frame attached to B (cid:48) ,pointing towards B (cid:48) , let x (cid:48) = F (cid:48) (¯ q, ¯ v ) ∈ M denote the post-collision vector as expressed in theintrinsic (phase space) coordinates of B (cid:48) . Then F (cid:48) is continuous (except at tangential collisions,which we have ruled out). We have F K (cid:48) , K ( x ) = x (cid:48) = F (cid:48) ◦ ¯ π ◦ (id M × G )( x, K , K (cid:48) ), where thecomposition comprises continuous functions. The uniform continuity statement follows from astandard compactness argument. (cid:3) Proof of Lemma 15. Because W (cid:63) is closed in W , the set W \ W (cid:63) is a countable union of disjoint,open (i.e., endpoints not included), connected, curves V ⊂ W , which we call gaps. Considera gap V . Notice that its endpoints x and y belong to W (cid:63) whence it follows that ρ satisfies | log ρ ( x ) − log ρ ( y ) | ≤ Cθ s ( x,y ) for the fixed pair ( x, y ). Let r > the first time such that F r V intersects the set ∂ M ∪ ∪ | k |≥ k ∂ H k , in other words r = s ( x, y ), and pick an arbitrary point z ∈ V whose image F r ( z ) is in the intersection. On the curve V , placing a discontinuity at z as needed, assign ρ the constant value ρ ( x ) between the points x and z and similarly the value ρ ( y ) between z and y . With the exception of the above bound being satisfied on all of W , theclaims of the lemma are clearly true.To check the bound, let ( x (cid:48) , y (cid:48) ) be an arbitrary pair of points in W . If x (cid:48) ∈ W (cid:63) , set x = x (cid:48) .Otherwise x (cid:48) belongs to a gap V with an endpoint x ∈ W (cid:63) satisfying ρ ( x ) = ρ ( x (cid:48) ). Similarly wedefine a point y in terms of y (cid:48) . For x = y we simply have ρ ( x (cid:48) ) − ρ ( y (cid:48) ) = 0, so let us assume fromnow on that x (cid:54) = y . Since | log ρ ( x (cid:48) ) − log ρ ( y (cid:48) ) | = | log ρ ( x ) − log ρ ( y ) | ≤ Cθ s ( x,y ) , it remainsto check that s ( x, y ) ≥ s ( x (cid:48) , y (cid:48) ) in order to prove | log ρ ( x (cid:48) ) − log ρ ( y (cid:48) ) | ≤ Cθ s ( x (cid:48) ,y (cid:48) ) . Indeed,given two points a, b on W , let W ( a, b ) denote the open subcurve of W between the two points.If W ( x, y ) ⊂ W ( x (cid:48) , y (cid:48) ), then the bound s ( x, y ) ≥ s ( x (cid:48) , y (cid:48) ) is obvious. On the other hand, if W ( x, y ) ⊃ W ( x (cid:48) , y (cid:48) ), then there exist gaps W ( x, ¯ x ) and W (¯ y, y ) on W such that x (cid:48) ∈ (cid:102) W u ( x, ¯ x )and y (cid:48) ∈ (cid:102) W u (¯ y, y ), where ¯ x and y (cid:48) are on the same side of x on W , and ¯ y and x (cid:48) are on the sameside of y . By construction, s ( x, x (cid:48) ) ≥ s ( x (cid:48) , ¯ x ) and s ( y (cid:48) , y ) ≥ s (¯ y, y (cid:48) ). These inequalities implyimmediately s ( x, y ) = s ( x (cid:48) , y (cid:48) ). Showing that s ( x, y ) ≥ s ( x (cid:48) , y (cid:48) ) also when neither W ( x, y ) nor W ( x (cid:48) , y (cid:48) ) is completely contained in the other set can be done by combining ideas from theprevious cases and is left to the reader. (cid:3) Proof of Lemma 19. Assume that µ has a strictly positive, -H¨older continuous density χ withrespect to the measure dµ = N − ρ d r d ϕ , where ρ = cos ϕ and N is the normalizing factor;we then have dµ = N − ρ d r d ϕ with ρ = χρ . Such a measure can be represented as ameasured unstable family in a canonical way: Let S j , 1 ≤ j ≤ ∞ , be an enumeration of all thesets H k ∩ M i . For each j , partition S j into straight lines W α , α ∈ E ( j ) , of slope κ min and ofmaximal length so that ∪ α ∈ E ( j ) W α is a regular unstable stack. We assume here that the sets E ( j ) are disjoint subsets of R in order to avoid having to introduce additional superscripts ( j )for the line segments. Disintegrating µ and µ using these stacks, we denote the conditionaldensities on W α by ρ α and ρ α respectively. Because of the simple geometry of the partition, ρ α and ρ α , are obtained as the normalized restrictions of ρ and ρ on W α . In particular, we havethe identity ρ α = χρ α . (42)The conditional densities ρ α have uniformly -H¨older continuous logarithms. In other words,there exists a constant C > 0, independent of α , such that | log ρ α ( x ) − log ρ α ( y ) | ≤ C d M ( x, y ) / for all x, y ∈ W α , for all α ∈ A . Indeed, denoting by W α ( x, y ) ⊂ H k the segment of W α con-necting the points x, y ∈ W α , we have | W α ( x, y ) | ≤ C H k − for a constant C H > k ≥ k . Writing x = ( r x , ϕ x ) and y = ( r y , ϕ y ), the bound (5) then yields | log ρ α ( x ) − log ρ α ( y ) | = | log cos ϕ y − log cos ϕ x | ≤ ϕ y , cos ϕ x ) | cos ϕ y − cos ϕ x |≤ C cos k | ϕ y − ϕ x | ≤ C cos ( C H | W α ( x, y ) | − ) / | W α ( x, y ) | ≤ C cos C / H d M ( x, y ) / as claimed. The extension to k = 0 is immediate, observing that cos ϕ ≥ cos( π/ − k − ) on H .The logarithm of χ is also -H¨older continuous on M ; let us denote the constant | log χ | / .In particular, | log ρ α ( x ) − log ρ α ( y ) | ≤ | log χ | / d M ( x, y ) / + C d M ( x, y ) / ≤ ( | log χ | / + C L / ) d M ( x, y ) / for all x, y ∈ W α , for all α . (Here L is an upper bound on the length of ahomogeneous unstable curve.) Following Remark 14, | log ρ α ( x ) − log ρ α ( y ) | ≤ ( | log χ | / + C L / ) C / θ s ( x,y ) for any configuration sequence. Furthermore, denoting by Z and Z the quantity appearingin (14) for µ and µ , respectively, the identity in (42) yields Z ≤ sup χ · Z . Here Z < ∞ by direct inspection and sup χ ≤ e a ·| log χ | / for a uniform constant a > 0. Thus,the initial measures of Theorem 1 also satisfy the assumptions of Theorem 1’, and | log χ | / controls the constant C γ as claimed. (cid:3) Proof of Lemma 27. Observe that ψ ( x ) ψ ( y ) = 1 + (cid:20)(cid:18) ρ ( x ) ρ ( y ) − (cid:19) ρ ( y )ˇ ρ ( y ) + (cid:18) − ˇ ρ ( x )ˇ ρ ( y ) (cid:19)(cid:21) (cid:18) ρ ( y )ˇ ρ ( y ) − (cid:19) − ≤ A (exp( C r θ s ( x,y ) ) − , where A = ( B + 1)( b − − . Using the estimate log(1 + t ) ≤ t ( t ≥ (cid:12)(cid:12)(cid:12)(cid:12) log ψ ( x ) ψ ( y ) (cid:12)(cid:12)(cid:12)(cid:12) ≤ A (exp( C r θ s ( x,y ) ) − , the absolute value on the left side being justified because the preceding bound continues tohold for x and y interchanged and because s ( x, y ) = s ( y, x ). Next, fix a constant S > C r θ S ) − ≤ C r θ S . Then, if s ( x, y ) ≥ S , we see that (32) holds if C top ≥ C r . Todeal with the case s ( x, y ) < S , we use the crude bound | log ψ ( x ) − log ψ ( y ) | ≤ ( e C r − any x and y , at least for C top = max(( e C r − θ − S , C r ). (cid:3) Proof of Lemma 28. Write W n,i for the homogeneous components of F n W and ν n,i for thepush-forward of ν ( W − n,i ∩ · ) under F n . Here W − n,i denotes the element of the canonical n -step subdivision of W which maps bijectively onto W n,i under F n . It is a regular curve. Set Z n = (cid:80) i ν n,i ( W n,i ) / | W n,i | . Our task is to show that Z n ≤ C p ν ( W ) eventually. The smallnuisance is that we cannot apply Lemma 16 directly, as ν is not necessarily regular. Our trickis to compare the evolutions of ν and the uniform measure m W . Since the latter is obviouslyregular, Lemma 16 does apply: Writing Z m W n = (cid:80) i ( m W ) n,i ( W n,i ) / | W n,i | , we have Z m W n | W | ≤ C p (cid:18) ϑ n p | W | (cid:19) , as Z m W = 1. Next, fix n and the component index i . We write x − n = ( F n | W − n,i ) − ( x ) ∈ W − n,i for the preimage of any x ∈ W n,i . Denoting by (cid:96) n,i the density of the push-forward ( m W ) n,i ,we have ν n,i ( W n,i ) = (cid:82) W n,i ρ ( x − n ) (cid:96) n,i ( x ) d m W n,i ( x ) ≤ sup W − n,i ρ · ( m W ) n,i ( W n,i ). From here, usingthe bound on ρ , Z n ≤ sup W ρ · Z m W n ≤ e C inf W ρ · Z m W n ≤ e C Z m W n | W | ν ( W ) . If n ≥ n (cid:48) = max (cid:0) log( C r / ( C − C r )) / log θ, log | W | / log ϑ p (cid:1) , Lemma 13 guarantees that themeasures ν n,i are all regular, and the bounds above yield Z n ν ( W ) ≤ e C C p . ISPERSING BILLIARDS WITH MOVING SCATTERERS 39 We can therefore apply Lemma 16, which results in Z n ν ( W ) ≤ C p (cid:16) ϑ n − n (cid:48) p e C C p (cid:17) . The measure ( F n ) ∗ ν is therefore proper, provided that n ≥ n (cid:48) + ( C + log C p ) / | log ϑ p | . Becausewe are assuming C > C r > 2, there exist a uniform constant A p > n ≥ A p (cid:0) max (cid:0) log C, | log | W || (cid:1) + C (cid:1) . The condition in the lemma follows. (cid:3) References [1] Dmitry Anosov and Yakov Sinai. Some smooth ergodic systems. Russian Mathematical Surveys , 22(5):103,1967. Available from: http://stacks.iop.org/0036-0279/22/i=5/a=R05 .[2] Arvind Ayyer and Mikko Stenlund. Exponential decay of correlations for randomly chosen hyperbolic toralautomorphisms. Chaos , 17(4):043116, 7, 2007. Available from: http://dx.doi.org/10.1063/1.2785145 , doi:10.1063/1.2785145 .[3] Leonid Bunimovich, Yakov Sinai, and Nikolai Chernov. Markov partitions for two-dimensional hyperbolicbilliards. Uspekhi Mat. Nauk , 45(3(273)):97–134, 221, 1990. Available from: http://dx.doi.org/10.1070/RM1990v045n03ABEH002355 , doi:10.1070/RM1990v045n03ABEH002355 .[4] Leonid Bunimovich, Yakov Sinai, and Nikolai Chernov. Statistical properties of two-dimensional hyperbolicbilliards. Uspekhi Mat. Nauk , 46(4(280)):43–92, 192, 1991. Available from: http://dx.doi.org/10.1070/RM1991v046n04ABEH002827 , doi:10.1070/RM1991v046n04ABEH002827 .[5] Nikolai Chernov. Sinai billiards under small external forces. Ann. Henri Poincar´e , 2(2):197–236, 2001.Available from: http://dx.doi.org/10.1007/PL00001034 , doi:10.1007/PL00001034 .[6] Nikolai Chernov. Advanced statistical properties of dispersing billiards. J. Stat. Phys. , 122(6):1061–1094, 2006. Available from: http://dx.doi.org/10.1007/s10955-006-9036-8 , doi:10.1007/s10955-006-9036-8 .[7] Nikolai Chernov. Sinai billiards under small external forces. II. Ann. Henri Poincar´e , 9(1):91–107, 2008.Available from: http://dx.doi.org/10.1007/s00023-007-0351-7 , doi:10.1007/s00023-007-0351-7 .[8] Nikolai Chernov and Dmitri Dolgopyat. Brownian Brownian motion. I. Memoirs of the American Mathe-matical Society , 198(927):viii+193, 2009.[9] Nikolai Chernov, Gregory Eyink, Joel Lebowitz, and Yakov Sinai. Steady-state electrical conduction in theperiodic Lorentz gas. Comm. Math. Phys. , 154(3):569–601, 1993. Available from: http://projecteuclid.org/getRecord?id=euclid.cmp/1104253079 .[10] Nikolai Chernov and Roberto Markarian. Chaotic billiards , volume 127 of Mathematical Surveys and Mono-graphs . American Mathematical Society, Providence, RI, 2006.[11] Jean-Pierre Conze and Albert Raugi. Limit theorems for sequential expanding dynamical systems on [0 , Ergodic theory and related fields , volume 430 of Contemp. Math. , pages 89–121. Amer. Math. Soc.,Providence, RI, 2007.[12] Chinmaya Gupta, William Ott, and Andrei T¨or¨ok. Memory loss for time-dependent piecewise-expandingsystems in higher dimension. 2012. Preprint.[13] Andrzej Lasota and James Yorke. When the long-time behavior is independent of the initial density. SIAMJournal on Mathematical Analysis , 27(1):221–240, 1996. Available from: http://link.aip.org/link/?SJM/27/221/1 , doi:10.1137/0527012 .[14] Bill Moran and William Hoover. Diffusion in a periodic Lorentz gas. J. Statist. Phys. , 48(3-4):709–726,1987. Available from: http://dx.doi.org/10.1007/BF01019693 , doi:10.1007/BF01019693 .[15] P´eter N´andori, Domokos Sz´asz, and Tam´as Varj´u. A Central Limit Theorem for Time-Dependent Dynam-ical Systems. Journal of Statistical Physics , 146(6):1213–1220, 2012. Available from: http://dx.doi.org/10.1007/s10955-012-0451-8 , doi:10.1007/s10955-012-0451-8 .[16] William Ott, Mikko Stenlund, and Lai-Sang Young. Memory loss for time-dependent dynamical sys-tems. Mathematical Research Letters , 16(3):463–475, 2009. Available from: .[17] Yakov Sinai. Dynamical systems with elastic reflections. Russian Mathematical Surveys , 25(2):137, 1970.Available from: http://stacks.iop.org/0036-0279/25/i=2/a=R05 .[18] Mikko Stenlund. Non-stationary compositions of Anosov diffeomorphisms. Nonlinearity , 24:2991–3018,2011. doi:doi:10.1088/0951-7715/24/10/016 .[19] Mikko Stenlund. A vector-valued almost sure invariance principle for Sinai billiards with random scatterers.2012. Preprint arXiv:1210.0902. Available from: http://arxiv.org/abs/1210.0902 .[20] Lai-Sang Young. Statistical properties of dynamical systems with some hyperbolicity. Ann. of Math. (2) ,147(3):585–650, 1998. Available from: http://dx.doi.org/10.2307/120960 , doi:10.2307/120960 .[21] Lai-Sang Young. Recurrence times and rates of mixing. Israel J. Math. , 110:153–188, 1999. Available from: http://dx.doi.org/10.1007/BF02808180 , doi:10.1007/BF02808180 . ISPERSING BILLIARDS WITH MOVING SCATTERERS 41 (Mikko Stenlund) Department of Mathematics, University of Rome “Tor Vergata”, Via dellaRicerca Scientifica, I-00133 Roma, Italy; Department of Mathematics and Statistics, P.O. Box68, Fin-00014 University of Helsinki, Finland. E-mail address : [email protected] URL : (Lai-Sang Young) Courant Institute of Mathematical Sciences, New York, NY 10012, USA. E-mail address : [email protected] URL : (Hongkun Zhang) Department of Mathematics & Statistics, University of Massachusetts,Amherst, 01003, USA. E-mail address : [email protected] URL ::