Continuity properties of transport coefficientsin simple maps

# Continuity properties of transport coefficients in simple maps

## Abstract

We consider families of dynamics that can be described in terms of Perron-Frobenius operators with exponential mixing properties. For piecewise expanding interval maps we rigorously prove continuity properties of the drift and of the diffusion coefficient under parameter variation. Our main result is that has a modulus of continuity of order , i.e. is Lipschitz continuous up to quadratic logarithmic corrections. For a special class of piecewise linear maps we provide more precise estimates at specific parameter values. Our analytical findings are verified numerically for the latter class of maps by using exact formulas for the transport coefficients. We numerically observe strong local variations of all continuity properties.

## 1 Introduction

In simple deterministic dynamical systems physical quantities like transport coefficients can be fractal functions of control parameters. This finding was first reported for a one-dimensional piecewise linear map lifted periodically onto the whole real line, for which the diffusion coefficient was computed by using Markov partitions and topological transition matrices [26, 27, 29]. A generalization of this result was obtained for a map with both drift and diffusion by deriving exact analytical solutions for the transport coefficients [15, 9]. Further maps modeling chemical reaction-diffusion [14] and anomalous diffusion [33] yielded also fractal transport coefficients. Recent work aimed at physically more realistic models like (Hamiltonian) particle billiards, for which computer simulations yielded transport coefficients that are non-monotonic under parameter variation [31]. Ref. [32] contains a summary of this line of research.

These results asked for a more detailed characterization of the “fractality” of transport coefficients. A first attempt in this direction was reported by Klages and Klauß [30], who used standard techniques from the theory of fractal dimensions for characterizing the drift and diffusion coefficients of the map studied in [15]. They numerically computed a non-integer box counting dimension for these curves which varied with the parameter interval, leading to the notion of a “fractal fractal dimension”. These results were questioned by Koza [35], who computed the oscillation of these graphs at specific Markov partition parameter values. His work suggested a dimensionality of one by conjecturing that there exist non-trivial logarithmic corrections to the usual power law behaviour in the oscillation.

This research reveals the need to study the parameter dependence of transport coefficients in a rigorous mathematical setting, which can be formulated as follows: Given a parametrized family of chaotic dynamical systems on an interval with unique invariant physical measures together with a family of sufficiently regular observables one has, under suitable mixing assumptions on the systems , a law of large numbers and a central limit theorem for the partial sum processes , namely

and

 L(n−12Sλ,n)⇒N(0,2D(λ))

where . For suitable choices of the observables , the process is just the deterministic random walk generated by a lift of the map to the real axis, and and are the drift and diffusion coefficient of this random walk respectively.

There are a few rigorous results in the literature describing the dependence of and of quantities like for various classes of systems. Without going into the details they can be summarized as follows: If the maps and the observables depend smoothly on and if the topological conjugacy class of is not changed when is varied, then (and hence ) depends differentiably on [4, 8, 10, 19, 20, 39, 40]. If the topological class changes, quantities like may behave less regular and have a modulus of continuity not better than , even for very simple maps like symmetric tent maps [3]. On the other hand, this modulus of continuity is the rule for systems whose Perron-Frobenius operator (acting on a suitable space of “regular” densities) has a spectral gap [22, 24].

The goal of this paper is to explicitly relate these mathematical results to transport coefficients. We do so by rigorously proving continuity properties of and under parameter variation for certain classes of deterministic maps. In Section 2 we give a general estimate for families of dynamics (deterministic or not), which can be described in terms of Perron-Frobenius operators with exponential mixing properties. The applicability of these general results to piecewise expanding interval maps and in particular to the class of piecewise linear maps discussed in [26, 27, 29, 15, 32] is checked in Section 3. The main result is that has a modulus of continuity of order , i.e. is Lipschitz continuous up to quadratic logarithmic corrections. In Section 4 we summarize the general results for transport coefficients in the special case of piecewise linear maps and provide more precise estimates for special parameters. Our analytical findings are verified by numerical computations in Section 5, for which we use exact analytical formulas of the transport coefficients [15]. Particularly, we numerically analyze local variations of these properties. Our work corrects and amends previous results reported in [30, 35].

## 2 The general setting

Let be a compact interval, normalized Lebesgue measure on , the space of Lebesgue-integrable functions from to , and the space of -equivalence classes of functions of bounded variation. We use the following simplified notation for the two corresponding norms:

 |f|1:=∫|f|dm,∥f∥:=Var(f) (1)

where

 Var(f):=sup{∫fφ′dm:φ∈C1(R,R),|φ|∞≤1} (2)

is the variation of as a function from (i.e. extended by on ). If is differentiable as a function from integration by parts shows easily that . is obviously a semi-norm, and as , it is actually a norm. This and more details on functions of bounded variation can be found in [25, section 2.3]. The monograph [2] is a comprehensive reference for most of the background material needed in this section.

We consider a family of nonsingular maps . Nonsingular means that the Perron-Frobenius operator is well defined, i.e.

 ∫PTf⋅gdm=∫f⋅g∘Tdm(f∈L1m,g∈L∞m). (3)

By definition, for all , and we assume

###### Hypothesis 1

.

Our main assumption is that the maps in are uniformly exponentially mixing in the following sense:

###### Hypothesis 2

Each has a unique invariant probability density (so ), and there are constants and such that, for all ,

 |PnTf|1≤C2γnVar(f)for all f∈BV with ∫fdm=0 and for all n∈N. (4)

Observe the following consequences of Hypothesis 1 and 2:

 |PnTf−hT|1≤C2γn(Var(f)+2C1)for all probability densitiesf∈BV (5)

and

 Var(hT)≤2C1(T∈T). (6)

Indeed, as by Hypothesis 2 for each probability density , and by Hypothesis 1. Hence (6) follows from the definition (2) of , and then (5) is an immediate consequence.

Since it is our goal to investigate the dependence of various dynamical quantities as functions of , we need to introduce a distance on . At this stage the following one, which was already considered in [22], is most apropriate. It measures the distance between two maps and from in terms of a suitable norm of :

 |∥PT1−PT2|∥:=sup{|PT1f−PT2f|1:f∈BV,∥f∥≤1}. (7)

This distance can be controlled in terms of a more “hands-on” distance between the graphs of the maps:

 d(T1,T2):=inf{ϵ>0:∃Iϵ⊆I and ∃ a diffeomorphism σ:I→I s.th. m(I∖Iϵ)<ϵ,T1|Iϵ=T2∘σ|Iϵ, and ∀x∈Iϵ:|σ(x)−x|<ϵ,|1/σ′(x)−1|<ϵ}. (8)

Namely (see [22, Lemma 13]),

 |∥PT1−PT2|∥≤12⋅d(T1,T2). (9)

Now, as a warm-up exercise, we can prove the following estimate: for let

 ℓk:(0,∞)→(0,∞),ℓk(u):=u⋅(1+|logu|)k. (10)
###### Lemma 1

There exist constants such that

 |hT1−hT2|1≤K′1⋅ℓ1(|∥PT1−PT2|∥)≤K1⋅ℓ1(d(T1,T2))(T1,T2∈T) (11)
###### Proof.

Let , assume without loss of generality that , and fix . For ,

 |PNT1f−PNT2f|1≤N−1∑k=0|PN−k−1T1(PT1−PT2)PkT2f|1≤N−1∑k=0|(PT1−PT2)PkT2f|1≤N−1∑k=0|∥PT1−PT2|∥∥PkT2f∥≤N−1∑k=0~ηC1∥f∥≤C1N~η∥f∥ (12)

where we used Hypothesis 1. Hence,

 |hT1−hT2|1≤|PNT11−PNT21|1+|PNT1(1−hT1)|1+|PNT2(1−hT2)|1≤2C1N~η+2⋅C2γN(2+2C1)

where we used (4) and (6). With , this is (11). ∎

###### Remark 1

Even if is a family of piecewise linear maps and if has the Markov property, this estimate can generally not be improved. Examples for this fact within the family of symmetric mixing tent maps are provided in [3, 37].

Suppose now that to each there is associated an “observable” . We make the following assumptions:

###### Hypothesis 4

There is such that for all .

Denote

 J(T):=∫IψThTdm . (13)

Then we have immediately from (6) and Lemma 1

###### Corollary 1

There is some such that, for all ,

 |J(T1)−J(T2)|≤K2⋅ℓ1(d(T1,T2)) (14)

is the “drift” of the partial sum process

 ST,n:=n−1∑k=0ψT∘Tk=nJ(T)+n−1∑k=0^ψT∘Tk

under the invariant measure , where . Observe that

 Var(^ψT)≤2C3,|ψT1−ψT2|≤2C4d(T1,T2)for all T,T1,T2∈T. (15)

In view of Hypothesis 2 we can also define the “diffusion coefficient”1 of this process:

 D(T):=limn→∞12n∫(n−1∑k=0^ψT∘Tk)2hTdm=12∫^ψ2ThTdm+∞∑n=1∫^ψT⋅^ψT∘TnhTdm=12∫^ψ2ThTdm+∞∑n=1∫PnT(^ψThT)^ψTdm (16)

Even more, we have the central limit theorem

 L(n−12(ST,n−nJ(T))⇒N(0,2D(T)) as n→∞, (17)

see e.g. [21, 18, 38]. Among physicists (16) is known as the Taylor-Green-Kubo formula for diffusion [32]. For the dependence of on we prove:

###### Proposition 1

There is some such that, for all ,

 |D(T1)−D(T2)|≤K3⋅ℓ2(d(T1,T2)) (18)
###### Proof.

Observe first that, for all ,

 Var(ψh)≤Var(ψ)|h|∞+Var(h)|ψ|∞≤Var(ψ)Var(h). (19)

Indeed, for differentiable we have , so for differentiable and (19) follows from the product rule of diferentiation. General and are then approximated using mollifiers. It follows that, in view of (4),

 ∣∣∣∫PnT(^ψThT)^ψTdm∣∣∣≤|PnT(^ψThT)|1⋅|^ψT|∞≤C2C3γnVar(^ψThT)≤4C1C2C23γn. (20)

Let as before, denote (so that ), and fix . For all , eq. (20) implies

 2∞∑n=N∣∣∣∫PnT1(^ψT1hT1)^ψT1dm−∫PnT2(^ψT2hT2)^ψT2dm∣∣∣≤16C1C2C231−γη. (21)

For we use a different estimate. We decompose

 ∫PnT1(^ψT1hT1)^ψT1dm−∫PnT2(^ψT2hT2)^ψT2dm=Δn1+Δn2+Δn3+Δn4 (22)

where

 |Δn1|:=∣∣∣∫PnT1(^ψT1hT1)(^ψT1−^ψT2)dm∣∣∣≤12Var(PnT1(^ψT1hT1))|^ψT1−^ψT2|1≤4C21C3C4η (23)

and

 |Δn2|:=∣∣∣∫(PnT1−PnT2)(^ψT1hT1)^ψT2dm∣∣∣≤|∥PnT1−PnT2|∥Var(^ψT1hT1)|^ψT2|∞≤4C1C23|∥PnT1−PnT2|∥≤4(C1C3)2n~η (24)

where the last inequality follows from eq. (12). Next,

 |Δn3|:=∣∣∣∫PnT2((^ψT1−^ψT2)hT1)^ψT2dm∣∣∣≤|^ψT1−^ψT2|1|hT1|∞|^ψT2|∞≤4C1C3C4η (25)

and

 |Δn4|:=∣∣∣∫PnT2(^ψT2(hT1−hT2))^ψT2dm∣∣∣≤|hT1−hT2|1|^ψT2|∞|^ψT2|∞≤C23K1ℓ1(~η) (26)

¿From (23) - (26) we see that

 |Δn1|+|Δn2|+|Δn3|+|Δn4|≤~K(nη+ℓ1(η)) (27)

for some constant . Hence, in view of (21) and the choice of ,

 |D(T1)−D(T2)|≤4C1C2C231−γη+~K(ℓ1(η)+2N−1∑n=1(nη+ℓ1(η)))≤4C1C2C231−γη+~K(N2η+(2N−1)ℓ1(η))≤K3⋅ℓ2(η) (28)

for a suitable constant . ∎

###### Remark 2

Quite often slightly stronger forms of Hypotheses 2 and 4 are satisfied, where the mixing assumption (4) is replaced by

 Extra open brace or missing close brace (29)

and the assumption on the -dependence of is strengthened to

 Var(ψT1−ψT2)≤C′4d(T1,T2)for all T1,T2∈T. (30)

An inspection of the above estimates shows that and if (29) is assumed. If additionally (30) is assumed, then can be estimated as follows: Let . Then

 |Δn3|=∣∣∣∫PnT2((^ψT1−^ψT2)hT1−αhT2)^ψT2dm∣∣∣≤C2C3γnVar((^ψT1−^ψT2)hT1−αhT2)≤C1C2C3C′4(4+2)γnη . (31)

Hence, uniformly in and . But we see no way, in general, to bound the -terms in a similar way. However, for particular families of maps (which are all topologically conjugate), we will see in subsection 4.2 that for all and that the estimate for can be made more precise.

## 3 Checking Hypothesis 1 and 2

### 3.1 General piecewise expanding maps

In this subsection we show how the general Hypothesis 1 and 2 can be verified in the more particular setting when is a parametrized family of piecewise twice continuously differentiable and expanding interval maps. So, from now on, we look at the following setting:

 Λ⊂Rd is a compact parameter space, T={Tλ:λ∈Λ}, and (T1) there is some L>0 such that d(Tλ1,Tλ2)≤L|λ1−λ2| for all λ1,λ2∈Λ. (T2)

We start with an abstract result which reduces Hypothesis 2 essentially to a uniform Lasota-Yorke type inequality.

###### Lemma 2

Assume ((T1)) and ((T2)). Then Hypothesis 1 and 2 are valid if the transformations are mixing and satisfy a uniform Lasota-Yorke type inequality: there are constants and such that

 Var(PnTf)≤C5αnVar(f)+C6|f|1for all T∈T,n∈N,f∈BV. (LY)
###### Proof.

As and , it is straightforward to check that Hypothesis 1 holds with .

We turn to Hypothesis 2. Note first that, because of ((T1)) and ((T2)), it suffices to show that for each there are , and such that (4) holds with these constants for all with . But this is guaranteed by Corollary 2 and Remark 1c in [24]. ∎

Our next task is to give sufficient conditions for ((LY)). To this end we specialize further and assume from now on that our maps are piecewise expanding (PE) maps in the following sense:

 For each λ∈Λ there is a % finite partition (I1λ,…,INλλ) of I into sub-intervals such that all Tλ|Ijλ are monotone, C2, and κλ:=inf|T′λ|>2. (PE)

Already in [36] it was proved that each individual (PE)-map (even if ) satisfies ((LY)) with constants depending on the map. For parametrized families of maps one can generally find uniform constants, but there are counterexamples where this is not possible [22, 6, 7]. Under the above assumption one can, however, give simple sufficient conditions ensuring the uniform LY-inequality. The proof in [36] (see also [25, Proposition 2.1]) shows

 Var(PTλf)≤2κλVar(f)+(Dλ+Eλ)|f|1 (32)

where

 Dλ=supx∣∣ ∣∣(1T′λ(x))′∣∣ ∣∣,Eλ=2κλminj|Ijλ|. (33)

From this ((LY)) follows with , , and provided this supremum is finite.

In some cases of interest the , , are not bounded because there are arbitrarily short monotonicity intervals. In such situations, ad hoc arguments are needed. We give an example in the next section.

### 3.2 Piecewise linear modulo 1 maps

We now look at a particular model dealt with in [12, 13, 16, 17] from a mathematical perspective and in [26, 27, 29, 15, 32] from a physics point of view. Let , for some constants , and for consider

 Tλ(x)=ax+bmod(Z−12). (34)

Hofbauer [16] showed that these maps have always a unique invariant probability density2, but although these maps received further attention also in the mathematical literature [17, 12, 13], it is not so easy to draw Hypothesis 2 from these sources. Therefore we will take up a rather direct computation made in [23] to prove, without having to rely on the compactness assumption ((T1)), the following lemma.

###### Lemma 3

Let . Then

 Var(PTλf)≤2aVar(f)+2∣∣∣∫fdm∣∣∣for all f∈BV and n∈N. (35)

(This implies immediately Hypothesis 1 with and Hypothesis 2 as well as its strengthening (29) with and .)

###### Proof.

Denote by the family of all -functions with . In [23, eq. (11)] a number is defined for each pair of and . In view of [23, eq. (13) and (14)] it suffices to show that for each there is some such that .3 4 We will show that this is the case for .

Let and . Then , , and has monotonicity intervals , , where

 A−p=−12, Ak=a−1(k−12−b)(k=−p+1,…,q) Bq=12, Bk=a−1(k+12−b)(k=−p,…,q−1)

In order to estimate in [23, eq. (11)] one has to evaluate certain terms and . In our case, , as is the inverse of the derivative of the -th monotone branch. The quantity is an abbreviation for , where denotes a limit from the right. Therefore, using the formula on the bottom of [23, p.1779], we obtain

 aUk=a⋅(ψ(Ak)−ℓkgk(Ak)−sφAk)=−φ(−a2+b+p)+(k+p)(φ(12)−φ(−12))−sφ⋅(k−12−b)=−φ(−a2+b+p)+sφ⋅(p+12+b) if k>−p ,aU−p=−φ(−a2+b+p)+sφ⋅a2 , (36)

and similarly,

 aVk=a⋅(ψ(Ak)−ℓkgk(Ak)−sφBk)=−φ(−a2+b+p)+sφ⋅(p−12+b) if k

It follows that for all . Hence, by [23, eqs. (11) and (12)],

 Γ(φ,sφ)≤a−1+12a−1|sφ|≤2a . (38)

Next we check assumption ((T2)) on the Lipschitz dependence of the maps on the parameters, so we estimate . For the proof we extend the maps to the whole real line (keeping the same names) by applying definition (34) to all .

Suppose and denote by and , , the discontinuity points of the two maps as introduced in the proof of Lemma 3. Consider the linear map , and observe that and for all . Let and for some arbitrarily small . Define by if and extend to a diffeomorphism of . Then

• for all