Modeling potential as fiber entropy and pressureprotect unhbox voidb@x penalty @M {}as entropy
TOMASZ DOWNAROWICZ,GUO HUA ZHANG
2014-01-01
Abstract:We first prove that topological fiber entropy potential in a relatively symbolic extension (topological joining with a subshift) of a topological dynamical system (which is a non-negative, non-decreasing, upper semicontinuous and subadditive potential H) yields topological pressure equal to the topological entropy of the extended system. The terms occurring on the other side of the variational principle for pressure are equal to the extension entropies of the invariant measures. Thus the variational principle for pressure reduces to the usual variational principle (for entropy) applied to the extended system. Next we prove our main theorem saying that every non-negative, upper semicontinuous and subadditive potential F (we drop the monotonicity assumption) is ‘nearly equal’ to the fiber entropy potential H in some relatively symbolic extension of the system, in the sense that all terms occurring in the variational principle for pressure are the same for both potentials. This gives a new interpretation of all such potentials F as a kind of additional information function enhancing the natural information arising from the dynamical system, and provides a new proof of the variational principle for pressure. At the end of the paper we provide examples showing that both assumptions, continuity and additivity, under which so-called lower pressure (defined with the help of spanning sets) equals the pressure, are essential, already in the class of non-negative, upper semicontinuous, subadditive potentials. 1. Heuristics In the theory of dynamical systems, there are two major notions of entropy: measuretheoretic (Kolmogorov–Sinai) and topological (Bowen, Adler–Konheim–McAndrew). Their information-theoretic interpretation is, roughly speaking, the following: every state Ergod. Th. & Dynam. Sys. (2015), 35, 1165–1186 c © Cambridge University Press, 2014 1166 T. Downarowicz and G. H. Zhang of the system (point in the phase space) carries some natural information. This can be imagined to be the information about the point’s ‘identity’, and initially it is fully ‘concealed’ from an observer. By making a measure-theoretic measurement (applying a measurable partition P ) or a topological measurement (distinguishing points that are more than apart) at time zero, the observer can ‘acquire’ some part of this information, by learning to which element of the partition (or to which -ball) the initial state belongs. Repeating the same measurement in every following unit of time, the observer acquires analogous information about the points in the forward orbit of the initial state; hence, by ‘backtracking’, he/she acquires more and more information about the initial state. In the measure-theoretic setup, the information acquired after n steps, if the initial state is x , equals In(x)=−log μ(Ax ), where Ax denotes the unique element of the partition P n = ∨n−1 i=0 T −i (P) containing x , andμ is the invariant measure considered. In this manner, we have represented the acquired information in the form of an increasing sequence (In)n≥1 of non-negative functions on X . Note that these functions depend on the measure and on the partition. The entropy hμ(T, P) is defined as the average flow of information, with averaging applied first over the space and then over time: hμ(T, P)= lim n 1 n ∫ In dμ. Finally, the Kolmogorov–Sinai entropy of μ is the supremum of the above terms over all finite partitions P , i.e. the global (maximal) flow of information in the system. In the topological setup the information function acquired in n steps is constant over the space and equals the logarithm of the number of ‘classes of -indistinguishability’ among the n-orbits. We prefer to skip further discussion of indistinguishability and refer to [4, §6.1]. Such a constant information function is analogous to the measurable situation in which all classes have the same measure (a measurable partition with atoms of equal measures yields a constant information function equal to the logarithm of the number of atoms; this, however, is just a rough analogy—the topological ‘classes of indistinguishability’ are usually not disjoint). Since the information function is constant, the averaging over the space is not necessary, and the topological -entropy is defined as the flow (average over the time) of this information. Finally, we take supremum over to get the topological entropy as the global flow of the topological information in the system. We have included the above discussion in order to emphasize one important fact: the acquired information has the form of an increasing sequence of non-negative functions defined on the phase space. The measure-theoretic information depends on the measure and partition, while the topological information depends on a topological parameter (alternatively, on an open cover). Moreover, the topological information need not coincide with the measure-theoretic information of any invariant measure. Nonetheless, the measure-theoretic and topological entropies are strictly related: the variational principle asserts that topological entropy equals the supremum, over all invariant measures, of their Kolmogorov–Sinai entropies. In this note we will try to extend the above information-theoretic interpretation to the notion of topological pressure. So far, this notion has been mainly attributed to Modeling potential as fiber entropy 1167 so-called thermodynamical formalism, hence we believe our interpretation might shed some new light on it. We would like to interpret a non-negative potential as a kind of additional information enhancing the natural information arising from the dynamical system. Moreover, since a potential F= ( fn)n≥1 is just one sequence of functions, this additional information must be universal, i.e. the same potential should enhance the measure-theoretic information regardless of the invariant measure and partition, as well as the topological information regardless of any topological parameter. Needless to say, we would like to be able to interpret the topological pressure associated with the potential as the global flow of the enhanced topological information, and we would like to be able to interpret the global flow of the enhanced measure-theoretic information as well. In order to understand whether such an interpretation is possible, we seek standard constructions where such universal additional information occurs naturally. As we shall see, exactly this kind of additional information occurs as so-called topological fiber entropy in relatively symbolic extensions of topological dynamical systems. In a general topological extension, topological fiber entropy depends on a topological parameter (or on an open cover), so it has the form of a family of sequences of functions. But in a relatively symbolic extension, due to the existence of a relative universal generator in form of a partition into clopen sets (so it relatively generates both topologically as well as measure-theoretically for every invariant measure), one sequence suffices to determine the topological fiber entropy completely. We will call this sequence the fiber entropy potential and denote it by H= (Hn)n≥1. It is non-negative, upper semicontinuous, subadditive and non-decreasing in n. A heuristic interpretation of the fiber entropy potential as additional information associated to each state is via the fact that the state splits in the extension into many (finer) states. Topological fiber entropy at a point equals the logarithm of the number of ‘classes of indistinguishability’ into which the point splits. So, it corresponds to the additional topological information resulting from the splitting. As we shall see, although of topological origin, such information can be successfully applied to enhance the measuretheoretic information as well, by simply adding the potential to the information function (creating, for every invariant measure and partition, the new sequence (In +Hn)n≥1). For the fiber entropy potential, we can prove the following facts. (1) The topological pressure coincides with the topological entropy of the extended system (i.e. with the global flow of the enhanced information, as we desired). (2) The global flow of the enhanced measure-theoretic information is equal, for each invariant measure μ on the factor system, to the so-called extension entropy of μ. (3) As a result, the variational principle for pressure reduces to the usual variational principle (for entropy) applied to the extended system, which is an evidence for correctness of our interpretation. Since our interpretation works in the class of fiber entropy potentials, we have attempted to interpret every non-negative, non-decreasing, upper semicontinuous, subadditive potential F as being in some sense equivalent (nearly equal) to the fiber entropy potential H in some relatively symbolic extension, so that the above three facts hold for the potential F. This attempt was successful; we were even able to drop the assumption that the potential F is non-decreasing. So, we have not only proved that our interpretation of the potential 1168 T. Downarowicz and G. H. Zhang as the additional information is correct, but also provided a method to model the potential in such a way that this interpretation can be seen in a practical mode. We have shown that the variational principle for pressure, which was a subject of many independent works and has been proved gradually with increasing generality (ultimately it is known for upper semicontinuous subadditive potentials), reduces, for non-negative potentials, to the usual variational principle. In this way our paper provides an alternative proof. 2. Preliminaries We recall some of the notions in the classic theory of topological entropy and topological pressure. Let (X, T ) be a topological dynamical system, where X is a compact metric space with a metric d and T : X→ X is a continuous transformation. We denote by dn the metric dn(x, y)=max{d(T i x