Philosophy of Science, 76 (December 2009) pp. 997–1008. 0031-8248/2009/7605-0013$10.00 Copyright 2009 by the Philosophy of Science Association. All rights reserved. 997 Typicality and the Approach to Equilibrium in Boltzmannian Statistical Mechanics Roman Frigg†‡ An important contemporary version of Boltzmannian statistical mechanics explains the approach to equilibrium in terms of typicality. The problem with this approach is that it comes in different versions, which are, however, not recognized as such and not clearly distinguished. This article identifies three different versions of typicality-based explanations of thermodynamic-like behavior and evaluates their respective successes. The conclusion is that the first two are unsuccessful because they fail to take the system’s dynamics into account. The third, however, is promising. I give a precise formulation of the proposal and present an argument in support of its central contention. 1. Introduction. Consider a gas confined to the left half of a container. Removing the dividing wall results in the gas spreading uniformly over the entire available space. It has approached equilibrium. Statistical me- chanics (SM) aims to explain the approach to equilibrium in terms of the dynamical laws governing the individual molecules of which the gas is made up. What is it about molecules and their motions that leads them to spread out when the wall is removed? And why does this happen invariably? That is, why do we never observe gases staying in the left half, even after the shutter has been removed? An important contemporary version of the Boltzmannian approach to SM, originating in the work of Joel Lebowitz (1993a, 1993b), answers these questions in terms of the notion of typicality. Intuitively, something †To contact the author, please write to: Department of Philosophy, Logic, and Scientific Method, London School of Economics, Houghton Street, London WC2A 2AE, En- gland; e-mail: r.p.frigg@lse.ac.uk. ‡Special thanks goes to Charlotte Werndl for invaluable discussions about dynamical systems. I also would like to thank Craig Callender, Hartry Field, Shelly Goldstein, Stephan Hartmann, Carl Hoefer, Wolfgang Pietsch, and Nino Zanghı̀ for valuable comments on earlier drafts and helpful discussions. Thanks to the audiences in Madrid, Oxford, Pittsburgh, and Jerusalem for stimulating discussions. 998 ROMAN FRIGG is typical if it happens in the ‘vast majority’ of cases: typical lottery tickets are blanks, typical Olympic athletes are well trained, and in a typical series of 1,000 coin tosses the ratio of the number of heads and the number of tails is approximately one. The aim of a typicality-based approach to SM is to show that approaching equilibrium is the typical behavior of systems like gases. This approach has grown increasingly popular in recent years (refer- ences are given below). The problem with it is that it comes in different versions, which are, however, not recognized as such, much less clearly distinguished. The aim of this article is to distinguish three different kinds of typicality-based explanations of the approach to equilibrium and eval- uate their respective successes. My conclusion is that the first two are unsuccessful because they fail to take the system’s dynamics into account. The third, however, is promising. I give a precise formulation of the proposal and present the outline of a proof of its central contention. 2. Classical Boltzmannian SM. Consider a system consisting of n classical particles with 3 degrees of freedom each. The state of this system is spec- ified by a point x, also referred to as the system’s microstate, in its -6n dimensional phase space G, which is endowed with the Lebesgue measure mL. The dynamics of the system is governed by Hamilton’s equations of motion, which define a measure preserving flow ft on G, meaning that for all times t is a one-to-one mapping such thatf : G r G m(R) pt for all measurable . The system’s microstate at time t0 (itsm(f (R)) R P Gt ‘initial condition’), , evolves into at time t. In a Ham-x(t ) x(t) p f (x(t ))0 t 0 iltonian system, energy is conserved, and hence the motion of the system is confined to the –dimensional energy hypersurface GE. The mea-6n � 1 sure mL can be restricted to GE, which induces a natural invariant measure m on GE. To each macrostate Mi of the system, , there correspondsi p 1, . . . , m a macroregion consisting of all for which the macroscopic var-G x � GM Ei iables assume the values characteristic for Mi. The together form aGMi partition of GE, meaning that they do not overlap and jointly cover GE up to measure zero. The Boltzmann entropy of a macrostate Mi is defined as , where kB is the Boltzmann constant. GivenS (M ) :p k log [m(G )]B i B Mi this, we define the Boltzmann entropy of a system at time t, , as theS (t)B entropy of the system’s macrostate at t: , where isS (t) :p S (M ) x(t)B B x(t) the system’s microstate at t and is the macrostate corresponding toMx(t) .x(t) Among the macrostates of a system, two are of particular importance: the equilibrium state, Meq, and the system’s state at the beginning of the process, Mp (also referred to as the ‘past state’). The latter is, by as- TYPICALITY AND STATISTICAL MECHANICS 999 sumption, a low-entropy state.1 The idea now is that the behavior of should mirror the behavior of the thermodynamic entropy STD, atS (t)B least approximately.2 So we expect the Boltzmann entropy of a system initially prepared in Mp to increase more or less monotonically, reach its maximum fairly quickly, and then remain at or near the maximum for a long time. In other words, we expect the dynamics to be such that it carries the system’s initial state into reasonably quicklyx(t ) � G G0 M Mp eq and then keeps it there for a long time. I refer to this as ‘thermodynamic- like behavior’.3 The explanandum then is this: Why does the system under investigation behave in a thermodynamic-like way? The standard Boltzmannian response to this question is to introduce probabilities and argue that the values of these probabilities come out such that the system is overwhelmingly likely to evolve in a thermody- namic-like way.4 Typicality approaches to SM eschew commitment to probabilities and offer a different kind of explanation: the system behaves in a thermodynamic-like way because it is typical for systems of this kind to behave in this way.5 Before turning to a discussion of this approach, an important technical result needs to be stated. Under certain conditions, it is the case that is vastly larger (with respect to ) than any other macroregion. I referG mMeq to this matter of fact as the ‘dominance of the equilibrium macrostate’. This dominance is then often glossed as being equivalent to the claim that for large n GE is almost entirely taken up by equilibrium microstates (Bric- mont 1996, 146; Goldstein 2001, 45; Zanghı̀ 2005, 191, 196). Some caution is needed here. In certain systems, nonequilibrium states can take up a substantial part of the phase space due to the degeneracy of below-equilibrium entropy values, and hence it is not true that GE is almost entirely filled with equilibrium states, even if the equilibrium mac- rostate is by far the largest macrostate (Lavis 2005, 255–258; 2008, Section 1. If we study laboratory systems like the above-mentioned gas, Mp has low entropy by construction. If we take the universe as a whole to be the object of study, then that Mp is of low entropy is the subject matter of the so-called Past Hypothesis (Albert 2000, 96). 2. This ‘mirroring’ need not be perfect, and occasional deviations of the Boltzmann entropy from its thermodynamic counterpart are no cause for concern (Callender 2001). 3. This definition of thermodynamic-like behavior is the one adopted by those writing on typicality; see, e.g., Goldstein 2001, 43–44. Lavis (2005, 255) gives a somewhat different definition. These differences are inconsequential for what follows. 4. A discussion of the different ways of introducing probabilities into the Boltzmannian framework can be found in Frigg 2009. 5. This explanatory strategy is reminiscent of probabilistic explanations appealing to high probabilities and, hence, is open to similar objections. For the sake of argument, I set these worries aside and accept that something being typical has explanatory power. 1000 ROMAN FRIGG 2). However, it turns out that those nonequilibrium states that occupy most of the nonequilibrium area are close to equilibrium (in the sense of having close-to-equilibrium entropy values). We can then lump the equi- librium and these close-to-equilibrium states together and get an ‘equi- librium or almost equilibrium’ region, which indeed takes up most of GE. The approach to equilibrium has then to be understood as the approach to this ‘equilibrium or almost equilibrium’ state, which is sufficient to give us thermodynamic-like behavior in the sense introduced above. 3. Typicality. Consider an element e of a set S. Typicality is a relational property of e, which e possesses with respect to S, a property P, and a measure , often referred to as a ‘typicality measure’.6 Roughly speaking,n e is typical if most members of S have property P and e is one of them. More precisely, let P be the subset of S consisting of all elements that have property P. Then the element e is typical if and only if (iff ) e � P and , where is a small real number;n (P) :p n(P)/n(S) ≥ 1 � � � ≥ 0S is referred to as the ‘measure conditional on S’, or simply ‘conditionaln (7)S measure’.7 Derivatively, one can then refer to P as the ‘typical set’ and to those elements that possess property P (i.e., the members of P) as ‘typical elements’. Conversely, an element e is atypical iff it belongs to the complement of P, , in which case we refer to Q as the ‘atypicalQ :p S/P set’ and to its members as ‘atypical elements’. As an example, consider the number , which is typical with respectp/4 to the interval , the property ‘not being specifiable by a finite number[0, 1] of digits’, and the usual Lebesgue measure on the real numbers because it is a theorem of number theory that the set of all numbers that have this property has measure one. The element of interest in SM is a micro- state x, and it is generally agreed that the relevant measure is the Lebesgue measure . However, views diverge when it comes to specifying the relevantm set S and relevant property P. I now turn to a discussion of three different typicality-based accounts of SM that emerge from the writings of Goldstein, Lebowitz, and Zanghı̀. In conversation, Goldstein and Zanghı̀ have pointed out to me that they would not subscribe to Accounts 1 and 2 and that (something like) Ac- count 3 is what they had intended. However, since the relevant papers can reasonably be read as proposing Accounts 1 and 2, it worth discussing them briefly to set the record straight (Sections 4 and 5) before turning to a detailed discussion of Account 3 (Section 6). 6. Typicality measures often are, but need not be, probability measures. 7. This definition of typicality is adapted from Dürr 1998, Section 2; Lavis 2005, 258; Zanghı̀ 2005, 185; and Volchan 2007, 805. TYPICALITY AND STATISTICAL MECHANICS 1001 4. First Account. The first account sets out to explain the approach to equilibrium in terms of the dominance of the equilibrium macrostate. Zanghı̀ explains, “reaching the equilibrium distribution in the course of the temporal evolution of a system is inevitable due to the fact that the overwhelming majority of microstates in the phase space have this dis- tribution; a fact often not understood by the critics of Boltzmann” (2005, 196; my translation).8 On this view, then, a system approaches equilibrium simply because the overwhelming majority of states in GE are equilibrium microstates. If we now associate S with GE and property P with ‘being an equilibrium state’ (and, as indicated above, regard microstates as el- ements of interest and use the Lebesgue measure m as a typicality measure), the dominance of the equilibrium macrostate implies that equilibrium microstates are typical and the view put forward in the above quote can be summarized as the claim that systems approach equilibrium because equilibrium microstates are typical and nonequilibrium microstates are atypical. This explanation is unsuccessful. If a system is in an atypical microstate, it does not evolve into an equilibrium microstate just because the latter is typical. Typical states do not automatically attract trajectories.9 In fact, there are Hamiltonians—for instance, the null Hamiltonian or a collection of uncoupled harmonic oscillators—that give rise to phase flows that do not carry nonequilibrium states into equilibrium. To explain why non- equilibrium microstates eventually wind up in equilibrium, the typicality of is not enough, and appeal has to be made to the system’s dynamics.GMeq 5. Second Account. A different line of argument can be found in Lebowitz 1993a, 1993b, 1999 and Goldstein and Lebowitz 2004. This account differs from the above in that it focuses on the internal structure of the mic- roregions rather than the entire phase space: “By ‘typicality’ we meanGMi that for any [ ] . . . the relative volume of the set of microstates [x] inGMi [ ] for which the second law is violated . . . goes to zero rapidly (ex-GMi ponentially) in the number of atoms and molecules in the system” (Gold- stein and Lebowitz 2004, 57).10 This definition contains different elements that need to be distinguished. Let be the subset of containing all x that lie on trajectories that(��)G GM Mi i come into from a macrostate of higher entropy and that leaveG GM Mi i entering into a macrostate of higher entropy; , , and are(��) (��) (��)G G GM M Mi i i 8. Goldstein 2001, 43–44, 49, can be read as making a similar claim. 9. Uffink (2007, 979–980) illustrates this with the example of a trajectory. 10. Square brackets indicate that the original notation has been replaced by the no- tation used in this article. I use this convention throughout. 1002 ROMAN FRIGG defined accordingly. These four subsets form a partition of .11 Fur-GMi thermore, and are the subsets(�) (��) (��) (�) (��) (��)G :p G ∪ G G :p G ∪ GM M M M M Mi i i i i i of that have a higher- and lower-entropy future, respectively.GMi There is an interpretative question about how to understand the notion of a set of microstates in violating the Second Law. A plausible readingGMi takes these to be states that have an entropy-decreasing future. Let us call this property D. Hence, x has D iff . Entropy-decreasing states(�)x � GMi are atypical in iff , where . Furthermore,(�)G m (G ) ! � m (7) :p m(7)/m(G )M i M i Mi i i let us say that a system has the property global atypical entropy decrease, GAD, iff entropy-decreasing microstates are atypical in every . TheGMi claim made in the above quote then is tantamount to saying that a system with a reasonably large number of molecules is GAD. This claim needs to be qualified. The atypicality of x with property D in trivially implies . Due to the time reversal invariance of(��)G m(G ) ! �M Mi i the Hamiltonian dynamics, we have , and therefore(��) (��)m(G ) p m(G )M Mi i . Since , we obtain .(��) (�) (��) (��) (��)m(G ) ! � G p G ∪ G m(G ) 1 1 � 2�M M M M Mi i i i i Hence, even if x with property D is atypical in , it is not the case that,GMi as we would expect, most states in act thermodynamic-like since mostGMi states have a higher entropy past. But this is a familiar problem, and remedy can be found in conditionalizing on (Albert 2000, ChapterGMp 4).12 So the correct requirement is that, at any time t, microstates violating the Second Law must be atypical in rather than only .G ∩ f (G ) GM t M Mi p i Do relevant systems meet this requirement? Immediately after the pas- sage quoted above, Goldstein and Lebowitz offer the following answer: “Boltzmann then argued that given this disparity in sizes of different M’s, the time evolved [ ] will be such that [ ] and thus [ ] willM m(M ) S (t)x(t) x(t) B typically increase in accord with the law” (2004, 57). So the argument seems to be that the relevant condition must be true because the equilib- rium state is much larger than other macrostates. This is unconvincing. The disparity of sizes of macroregions is, of course, compatible with GAD, but the latter does not follow from the former. Whether macroregions have the above internal structure depends on the system’s phase flow ft, and every attempt to answer this question without even mentioning the system’s dynamics is doomed to failure right from the start (and this is true of both the qualified and the unqualified versions of the claim). 6. Third Account. As we have seen, the basic problem with the two ac- 11. For the sake of simplicity, I omit points lying on trajectories that stay at the same entropy level. Taking these into account would not alter the argument. 12. Notice that an attempt to define D in terms of rather than only(�) (��) (�)G ∪ G GM M Mi i i leads to a contradiction. TYPICALITY AND STATISTICAL MECHANICS 1003 counts discussed so far is that they attempt to explain the approach to equilibrium without reference to the system’s dynamics. The third account, which emerges from a passage in Goldstein 2001, rectifies this problem: “[GE] consists almost entirely of phase points in the equilibrium macrostate [ ], with ridiculously few exceptions whose totality has volume of orderGMeq relative to that of [GE]. For a non-equilibrium phase point [x] of 20�1010 energy E, the Hamiltonian dynamics governing the motion [ ] wouldx(t) have to be ridiculously special to avoid reasonably quickly carrying [ ] into [ ] and keeping it there for an extremely long time—unless,x(t) GMeq of course, [x] itself were ridiculously special” (Goldstein 2001, 43–44). This is an interesting claim but one that stands in need of clarification. A reasonable reading of this passage seems to be that an argument in- volving three different typicality claims is made: Premise 1. The system’s macrostate structure is such that equilibrium states are typical in GE in the sense introduced in Section 4. Premise 2. The system’s Hamiltonian is typical in the class of all Hamiltonians. Conclusion. Initial conditions lying on trajectories showing thermo- dynamic-like behavior are typical in with respect toG m (7) :pM pp .m(7)/m(G )Mp Let us refer to this as the ‘T-Argument’. Premise 1 is familiar from Section 4 and is taken for granted here. Premise 2 and the conclusion are restate- ments in the language of typicality of the claims that the Hamiltonian of the system and the initial condition be not ‘ridiculously special’. The T-Argument, if sound, gives us the sought-after explanation of the approach to equilibrium in terms of typicality. But before we can address the question of soundness, we need to make Premise 2 more precise. What does it mean for a system’s Hamiltonian to be typical in the class of all Hamiltonians? More specifically, what is the typicality measure, and what is the relevant property P? Let us begin with the first question. The problem is that function spaces do not come equipped with normalized measures that can plausibly be used to capture the intuitive idea that some sets of functions are typical while others are atypical. A natural way around this difficulty is to replace the measure-theoretic notion of typicality introduced in Section 3 by a topological one based on Baire categories. Sets can be of two kinds: meager (first Baire category) or nonmeager (second Baire category). Loosely speaking, a meager set is the ‘topological counterpart’ of a set of measure zero in measure theory, and a nonmeager set is the ‘counter- part’ of a set of nonzero measure. Given this, it is natural to say that meager sets are atypical and nonmeager sets are typical. I call this notion of typicality ‘t-typicality’ (“t” for “topological”) and, to avoid confusion, 1004 ROMAN FRIGG from now on refer to the notion of typicality introduced in Section 3 as ‘m-typicality’, in order to make it explicit that it is a measure-theoretic notion. Unfortunately there is no straightforward answer to the question about the property P. But a promising line of argument emerges from Dürr’s (1998) and Maudlin’s (2007) discussion of typicality in the so-called Gal- ton Board, a triangular arrangement of nails on an infinitely long vertical board. Balls are fed into the board from the top and then move down the board. Every time a ball collides with a nail it moves either to the right (R) or to the left (L). If we follow a ball’s trajectory and take down whether it moves to the left or to the right every time it hits a nail, we obtain a string of Rs and Ls that looks as random as one that has been generated by a coin toss: the Galton Board seems to exhibit random behavior. Why is this? Dürr’s and Maudlin’s answer is that the board appears random because random-looking trajectories are typical in the sense that the set of those initial conditions that give rise to nonrandom- looking trajectories has measure zero in the set of all possible initial conditions, and this is so because the board’s dynamics is chaotic (Dürr 1998, Section 2). Translating this idea into the context of SM suggests that the relevant property P is being chaotic. This sounds prima facie plausible and, most important, would make Premise 2 true. Markus and Meyer prove the following theorem: Completely integrable Hamiltonians are meager in the space of all normalized and infinitely differentiable Hamiltonians on a compact symplectic manyfold. (1974, 13)13 This implies that nonintegrable Hamiltonians are nonmeager, which is tantamount to saying that the class of chaotic Hamiltonians is nonmeager and, hence, t-typical.14 The question now is whether the T-Argument is valid, that is, whether the conclusion follows from the premises. This turns out to be a thorny issue. There is an entire class of systems that are chaotic but whose phase space is full of invariant curves, namely, so-called KAM systems (Argyris, 13. Two Hamiltonians that differ only by a constant are considered equivalent, and an equivalence class of Hamiltonians is called a ‘normalized Hamiltonian’. This is because in practical calculations any Hamiltonian of this class can be chosen as a representative since they all yield the same flow (Markus and Meyer 1974, 11). 14. I here follow common practice and assume that nonintegrability implies chaos, at least on some region of the phase space. However, to the best of my knowledge there is no strict mathematical proof of this. Mathematicians refer to t-typical Hamiltonians as ‘generic’. TYPICALITY AND STATISTICAL MECHANICS 1005 Faust, and Haase 1994, Chapter 4). Naturally one would expect these curves to divide GE into a set of closed volumes bounded by the invariant curves, which would prevent the system from approaching equilibrium (invariant curves are ubiquitous in KAM systems, and so it would be highly unlikely that and would not be separated by one). In systemsG GM Mp eq with 2 degrees of freedom, this is exactly what happens: the two-dimen- sional invariant surfaces divide the three-dimensional energy hypersurface into disconnected parts. Fortunately, the situation is better for systems with degrees of freedom. The energy hypersurface has di-f 1 2 2f � 1 mensions, and for another surface to divide it into two disconnected parts, it must have dimensions. But the invariant KAM tori are f-2f � 2 dimensional, and since for all , invariant KAM surfaces2f � 2 1 f f 1 2 do not divide GE into separate parts; the invariant surfaces are a bit like lines in a three-dimensional Euclidean space. So the trajectories can, in principle, wander around relatively unhindered and without being ‘sand- wiched’ between invariant surfaces. This process is known as Arnold Dif- fusion. It has first been proven analytically to exist in a particular example, and there is now numerical evidence that it can also be found in other systems. In such systems, the chaotic parts of GE are connected and form a single net, the so-called Arnold Web, which permeates the entire phase space in the sense that a trajectory moving on the web will eventually visit almost every finite region of GE. 15 This looks like what we need, but unfortunately some difficulties arise on the finishing line. These can be circumvented only at the cost of ac- cepting three conjectures, which are supported only by plausibility ar- guments. First, there is no proof for the existence of Arnold Webs in all nonintegrable systems with . The good news is that so far there aref 1 2 no known examples in which this is not the case, and so we can venture the conjecture that all nonintegrable systems with have Arnold Websf 1 2 (Conjecture 1).16 Second, there is a question about the relative measure in GE occupied by Arnold Webs. It could in principle be that these webs are of measure zero or else fill only a small part of the phase space. If this were the case, it would be unlikely that m-typical initial conditions would come to lie on trajectories that wander around randomly (and therefore wind up in ), which would undercut the conclusion in the T-Argument. However,GMeq numerical simulations on simple systems have shown that the relative measure occupied by invariant KAM curves decreases as f increases (Ear- 15. In fact, Lichtenberg and Liebermann (1992, 61) and Ott (1993, 257) say that the system visits every finite region of GE, but this seems to be too strong. 16. Or if not all nonintegrable systems have Arnold Webs, then those that do not should be so few that the class of those with Arnold Webs is still nonmeager. 1006 ROMAN FRIGG man and Rédei 1996, 70). Furthermore, Sklar (1993, 175) observes that there are good numerical reasons to think that large systems are “at least ergodic-like” on the “overwhelmingly largest part” of the accessible phase space, and Vranas (1998, 695–698) gathers a welter of numerical evidence for the conclusion that many systems of interest in SM are �-ergodic, that is, ergodic on nearly the entire energy hypersurface. This suggests that it may well be the case not only that Arnold Webs have finite measure but that they in fact fill most of GE (Conjecture 2). Third, in order to explain thermodynamic-like behavior, we need to know how much time the system spends in different parts of the phase space. Again, little is proven rigorously, but Ott (1993, 257) suggests that the system is ergodic on the Arnold Web (Conjecture 3). The numerical evidence just mentioned supports this conjecture. If these three assumptions are correct, then the T-Argument is sound. By Premise 2 the system is chaotic, and by Conjecture 1 it has an Arnold Web, which, by Conjecture 2, fills most of the energy surface and hence most of . Therefore, points on the web are m-typical in . By Con-G GM Mp p jecture 3, these points wander around ergodically on the web and hence approach fairly soon and stay there for a long time (where “fairlyGMeq soon” means that the time taken to arrive at equilibrium is much shorter than the time spent in equilibrium) because, by Premise 1, nonequilibrium states occupy a much smaller volume than equilibrium states. To put this argument on secure footing, more would have to be said about the three conjectures. This is an extremely difficult task, and so it is worth asking whether there is not a simpler way to arrive at the same conclusion. I now discuss a plausible way of doing so and show that it is a blind alley. Hence, there is no way around trying to make progress on the conjectures. The new line of argument departs from the observation that we might have chosen too liberal a notion of chaos. In fact, there is a great deal of controversy over the correct characterization of chaos (Smith 1998, Chapter 10), and so we might say that KAM systems show the wrong kind of chaotic behavior: they exhibit ‘local chaos’, meaning that the dynamics are chaotic only on parts of the phase space. What we need, so the argument goes, is that the relevant systems show ‘global chaos’, which disqualifies KAM systems. The question then becomes how to characterize global chaos. Com- monly this is done in either of two ways: a topological way and a measure- theoretic one. The latter always involves ergodicity and is therefore un- tenable: Markus and Meyer (1974, 14) also prove that in the space of all normalized and infinitely differentiable Hamiltonians on a compact sym- plectic manyfold, the class of ergodic Hamiltonians is meager, and hence strongly chaotic systems are t-atypical. Topological definitions of chaos TYPICALITY AND STATISTICAL MECHANICS 1007 (the best known of which is Devaney’s) always involve topological tran- sitivity, the condition that for any two open regions A and B in GE, there is a trajectory initiating in A that eventually visits B. But this condition does not fit the bill: while it is at least plausible that topological chaos is a sufficient condition for the approach to equilibrium, it requires a revision of P that seems to render Premise 2 false. As Markus and Meyer (1974, 1) point out, ergodic systems are meager because generic systems have invariant surfaces preventing the trajectory from accessing the entire phase-energy hypersurface. But a system that cannot access certain regions of GE not only fails to be ergodic; it also fails to be topologically transitive. So topologically transitive systems must be meager too and, hence, also fail to be t-typical. For this reason, shifting attention to global chaos is a dead end. 7. Conclusion. I have distinguished three different accounts of how typ- icality is used to explain thermodynamic-like behavior. I have argued that while the first two fail, the third is promising, and I have sketched a proof. The proof rests on three conjectures that need to be further substantiated to put the argument on secure footing. Furthermore, an argument needs to be given that m-typicality and t-typicality have explanatory power from the point of view of physics. REFERENCES Albert, David Z. (2000), Time and Chance. Cambridge, MA: Harvard University Press. Argyris, John, Gunter Faust, and Maria Haase (1994), An Exploration of Chaos. Amsterdam: Elsevier. Bricmont, Jean (1996), “Science of Chaos or Chaos in Science?”, in Paul R. Gross, Norman Levitt, and Martin W. Lewis (eds.), The Flight from Science and Reason. Annals of the New York Academy of Sciences 775. New York: New York Academy of Sciences, 131– 175. Callender, Craig (2001), “Taking Thermodynamics Too Seriously”, Studies in History and Philosophy of Modern Physics 32: 539–553. Dürr, Detlef (1998), “Über den Zufall in der Physik”, paper given at the 1998 Leopoldina Meeting, Halle, http://www.mathematik.uni-muenchen.de/%7Eduerr/Zufall/zufall.html. Earman, John, and Miklós Rédei (1996), “Why Ergodic Theory Does Not Explain the Success of Equilibrium Statistical Mechanics”, British Journal for the Philosophy of Science 47: 63–78. Frigg, Roman (2009), “Probability in Boltzmannian Statistical Mechanics”, in Gerhard Ernst and Andreas Hüttemann (eds.), Time, Chance and Reduction: Philosophical Aspects of Statistical Mechanics. Cambridge: Cambridge University Press. Goldstein, Sheldon (2001), “Boltzmann’s Approach to Statistical Mechanics”, in Jean Bric- mont, Detlef Dürr, Maria Carla Galavotti, Gian Carlo Ghirardi, Francesco Petruccione, and Nino Zanghı̀ (eds.), Chance in Physics: Foundations and Perspectives. Berlin: Springer, 39–54. Goldstein, Sheldon, and Joel L. Lebowitz (2004), “On the (Boltzmann) Entropy of Non- equilibrium Systems”, Physica D 193: 53–66. Lavis, David (2005), “Boltzmann and Gibbs: An Attempted Reconciliation”, Studies in History and Philosophy of Modern Physics 36: 245–273. 1008 ROMAN FRIGG ——— (2008), “Boltzmann, Gibbs, and the Concept of Equilibrium”, Philosophy of Science 75: 682–696. Lebowitz, Joel L. (1993a), “Boltzmann’s Entropy and Time’s Arrow”, Physics Today (Sep- tember): 32–38. ——— (1993b), “Macroscopic Laws, Microscopic Dynamics, Time’s Arrow and Boltz- mann’s Entropy”, Physica A 194: 1–27. ——— (1999), “Statistical Mechanics: A Selective Review of Two Central Issues”, Reviews of Modern Physics 71: 346–357. Lichtenberg, Allan J., and M. A. Liebermann (1992), Regular and Chaotic Dynamics. 2nd ed. Berlin: Springer. Markus, Lawrence, and Kenneth R. Meyer (1974), “Generic Hamiltonian Dynamical Sys- tems Are neither Integrable nor Ergodic”, Memoirs of the American Mathematical Society 144: 1–52. Maudlin, Tim (2007), “What Could Be Objective about Probabilities?”, Studies in History and Philosophy of Modern Physics 38: 275–291. Ott, Edward (1993), Chaos in Dynamical Systems. Cambridge: Cambridge University Press. Sklar, Lawrence (1993), Physics and Chance: Philosophical Issues in the Foundations of Sta- tistical Mechanics. Cambridge: Cambridge University Press. Smith, Peter (1998), Explaining Chaos. Cambridge: Cambridge University Press. Uffink, Jos (2007), “Compendium of the Foundations of Classical Statistical Physics”, in Jeremy Butterfield and John Earman (eds.), Philosophy of Physics. Amsterdam: North- Holland, 923–1047. Volchan, Sérgio B. (2007), “Probability as Typicality”, Studies in History and Philosophy of Modern Physics 38: 801–814. Vranas, Peter B. M. (1998), “Epsilon-Ergodicity and the Success of Equilibrium Statistical Mechanics”, Philosophy of Science 65: 688–708. Zanghı̀, Nino (2005), “I fondamenti concettuali dell’approccio statistico in fisica”, in Valia Allori, Mauro Dorato, Federico Laudisa, and Nino Zanghı̀ (eds.), La natura delle cose: Introduzione ai fundamenti e alla filosofia della fisica. Rome: Carocci.