38.06.04 · dynamics / entropy

Topological Pressure, the Variational Principle, and Equilibrium States

shipped3 tiersLean: none

Anchor (Master): Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 9 (pressure, the variational principle, equilibrium states, the measure of maximal entropy); Bowen 1975 *Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms* (Springer LNM 470); Ruelle 1978 *Thermodynamic Formalism* (Addison-Wesley, Encyclopedia of Mathematics 5); Keller 1998 *Equilibrium States in Ergodic Theory* (Cambridge LMS Student Texts 42)

Intuition Beginner

Picture all the long histories a system can produce, the same growing tree of distinguishable orbits that measures entropy. Entropy counted those histories as if each mattered equally. Now suppose you care about some more than others: you attach a running score to every step, so a long history earns a total reward equal to the sum of its step scores. Pressure answers a weighted version of the counting question — not just how many histories there are, but how much total reward the whole crowd carries, again as an exponential growth rate. When the score is zero everywhere, every history earns the same reward and pressure collapses back to ordinary entropy.

The name comes from physics. A physicist studying a gas does not just count microscopic configurations; she weights each by its energy and asks for the system's free energy, the quantity that balances having many configurations against those configurations being cheap. Pressure is exactly this free energy for a dynamical system, with the step score playing the role of minus the energy. High pressure means either there are many histories or the favored ones are richly rewarded, and the single number trades these two effects against each other.

There is a beautiful accounting identity hiding here. You can compute pressure two ways. One way looks at the system from the outside, counting weighted histories directly. The other way looks from the inside, by choosing a long-run statistical habit for the system — a way it spreads its time across states — and adding that habit's own entropy to its average score. The variational principle says these always agree: the outside count equals the best inside choice. The habit that achieves the best balance is called an equilibrium state, the statistical law the system would settle into if it were trying to be as rich and rewarding as possible at once.

The takeaway: pressure generalizes entropy by attaching scores to histories, it is the dynamical version of free energy, and the variational principle equates the externally counted pressure with the best achievable sum of entropy plus average score. The optimizing habit is the equilibrium state, and when all scores are zero it is the most chaotic habit of all, the measure of maximal entropy.

Visual Beginner

Picture two columns of bars competing to set a single number, with a dial in the middle that always reads the same whichever column you trust.

The left column counts weighted histories from outside and reads off a growth rate. The right column tries every statistical habit and, for each, stacks its entropy on top of its average score; the best stack reaches exactly the same height as the outside count. The bottom table shows the zero-score case, where pressure is just entropy and the optimizing habit is the most chaotic one the system allows.

Worked example Beginner

We compute the pressure of a two-state system whose allowed moves are recorded by a simple rule, and we find its best habit.

Step 1. The system. There are two states, call them $a$ and $b$ . At each step the system may stay or switch freely, so every two-letter combination $aa, ab, ba, bb$ is allowed; this is a fair two-symbol shift. With score zero everywhere, the number of allowed length- $n$ histories is $2^{n}$ , so the pressure with zero score is $lo g 2 = 0.693$ , the topological entropy. This is the unweighted baseline.

Step 2. Attach a score. Now reward being in state $a$ : assign score $+ 1$ each step the system reads $a$ , and score $0$ each step it reads $b$ . A length- $n$ history earns total reward equal to the number of $a$ 's it contains. Histories heavy in $a$ now count for more.

Step 3. Count with weights. Add up the weight of each of the $2^{n}$ histories, where a history's weight is $e$ raised to its total reward. Each step independently contributes a factor $e^{1}$ if it reads $a$ and $e^{0} = 1$ if it reads $b$ , so the per-step factor totalled over the two choices is $e^{1} + e^{0} = e + 1$ . Over $n$ steps the total weighted count is $(e + 1)^{n}$ . The growth rate is $lo g (e + 1) = lo g (3.718) = 1.313$ .

Step 4. Read off the pressure and the habit. So the pressure is $lo g (e + 1) = 1.313$ , larger than the entropy $0.693$ because the reward inflates the count. The best habit spends a fraction $e / (e + 1) = 0.731$ of its time in the rewarded state $a$ and $1/ (e + 1) = 0.269$ in $b$ — tilted toward $a$ but not all the way, because spending all its time in $a$ would throw away the entropy of having choices.

What this tells us: pressure blends counting and reward into one growth rate, here $lo g (e + 1)$ . The equilibrium habit does not greedily chase the reward; it balances earning score against keeping its options open, settling at the tilt $0.731$ versus $0.269$ that maximizes entropy plus average score.

Check your understanding Beginner

Exercise (easy, multiple choice).

The topological pressure of a system with a step score (potential) is:

A. The same as the topological entropy, regardless of the score B. The exponential growth rate of the total weighted reward summed over all length- $n$ histories C. The largest single step-score the system can ever earn D. The average return time of a typical orbit

Hint

Pressure generalizes entropy by weighting each history by its total reward, then asking for the growth rate of that weighted count.

Answer

B. The exponential growth rate of the total weighted reward over length- $n$ histories.

Feedback-correct: pressure sums $e^{(total reward)}$ over distinguishable length- $n$ histories and takes the growth rate, reducing to entropy when the score is zero. Feedback-wrong: A is false because pressure depends on the score and equals entropy only when the score is zero; C confuses a single step-reward with the aggregate growth rate; D describes recurrence and return times, a separate invariant.

Formal definition Intermediate+

Throughout, $f : X \to X$ is a continuous map of a compact metric space $(X, d)$ , and $φ \in C (X, R)$ is a potential (a continuous real function). Write $S_{n} φ (x) = \sum_{k = 0}^{n - 1} φ (f^{k} x)$ for the Birkhoff sum of $φ$ along the orbit, and $d_{n} (x, y) = max_{0 \leq k < n} d (f^{k} x, f^{k} y)$ for the Bowen metric, which measures how far two orbits diverge over the first $n$ steps. Let $M (X, f)$ denote the set of $f$ -invariant Borel probability measures, a nonempty convex weak- $*$ -compact set.

Definition (separated and spanning sets). A set $E \subseteq X$ is $(n, ε)$ -separated if $d_{n} (x, y) > ε$ for distinct $x, y \in E$ ; a set $F$ is $(n, ε)$ -spanning if every $x \in X$ has $d_{n} (x, y) \leq ε$ for some $y \in F$ . These quantify how many orbit segments of length $n$ are distinguishable at resolution $ε$ .

Definition (topological pressure). The partition function at scale $ε$ is $Z_{n} (f, φ, ε) = sup {x \in E \sum e^{S_{n} φ (x)} : E is (n, ε) -separated} .$ The topological pressure of $φ$ is $P (f, φ) = ε \to 0 lim n \to \infty lim sup \frac{1}{n} lo g Z_{n} (f, φ, ε),$ the doubly-asymptotic exponential growth rate of the weighted count of distinguishable orbit segments. The spanning-set version, with $in f$ over $(n, ε)$ -spanning $F$ of $\sum_{y \in F} e^{S_{n} φ (y)}$ , yields the same value, as does the open-cover version $P (f, φ) = lim_{diam U \to 0} lim_{n} \frac{1}{n} lo g in f_{V} \sum_{V \in V} e^{s u p_{V} S_{n} φ}$ over subcovers $V$ of $⋁_{k < n} f^{- k} U$ .

Definition (topological entropy as a special pressure). Setting $φ = 0$ recovers the topological entropy $h_{top} (f) = P (f, 0) = lim_{ε} lim sup_{n} \frac{1}{n} lo g s (n, ε)$ , where $s (n, ε)$ is the maximal cardinality of an $(n, ε)$ -separated set. The general topological entropy this builds on is developed in 38.06.01; pressure is its potential-weighted refinement.

Definition (equilibrium state). Given $φ$ , a measure $μ \in M (X, f)$ is an equilibrium state for $φ$ if it attains the supremum in the variational principle below: $h_{μ} (f) + \int_{X} φ d μ = P (f, φ) .$ An equilibrium state for $φ = 0$ is a measure of maximal entropy: $h_{μ} (f) = h_{top} (f)$ .

Definition (Gibbs state). For an expansive $f$ and a potential $φ$ , an invariant measure $μ$ is a Gibbs state for $φ$ if there are constants $C \geq 1$ and a number $P$ such that for every $x$ and $n$ , $C^{- 1} \leq \frac{μ ( B _{n} ( x , ε ) )}{exp ( - n P + S _{n} φ ( x ) )} \leq C,$ where $B_{n} (x, ε) = {y : d_{n} (x, y) \leq ε}$ is the Bowen ball 38.06.03. The constant $P$ is then the pressure $P (f, φ)$ . On transitive Anosov / Axiom-A systems with Hölder $φ$ , the Gibbs state and the equilibrium state coincide and are unique.

Counterexamples to common slips Intermediate+

Pressure is not the maximum of the potential. $P (f, φ)$ is a growth rate weighted by $e^{S_{n} φ}$ , not $sup φ$ or $lim \frac{1}{n} sup S_{n} φ$ . The latter ignores the entropy contribution; pressure adds the logarithm of how many comparably-rewarded orbits exist. For $φ = 0$ , $sup φ = 0$ but $P = h_{top}$ can be positive.
The variational supremum need not be attained. On a general continuous system $h_{μ} (f)$ may fail to be upper semicontinuous in $μ$ , and the supremum $sup_{μ} (h_{μ} + \int φ)$ can be approached but unattained, so equilibrium states may not exist. Expansiveness restores upper semicontinuity and hence existence; this is not automatic.
Equilibrium states need not be unique. Uniqueness is a regularity phenomenon: it holds for Hölder potentials on transitive Anosov / Axiom-A systems (and on subshifts of finite type with the Bowen condition) but fails for merely continuous potentials, where phase transitions — several distinct equilibrium states — occur exactly as in statistical mechanics.
Topological pressure is a topological quantity; equilibrium entropy is measure-theoretic. $P (f, φ)$ depends only on $f$ , $φ$ , and the topology, not on any measure; the term $h_{μ} (f) + \int φ d μ$ is measure-dependent. The variational principle is the bridge identifying the topological supremum with the optimal measure-theoretic value, not an identity of like quantities.
The measure of maximal entropy is not always Lebesgue or "uniform". For an irreducible subshift of finite type it is the Parry measure, whose cylinder weights come from the Perron-Frobenius left/right eigenvectors of the transition matrix, generally not the uniform product measure. Only for the full shift does it reduce to the uniform Bernoulli measure.

Key theorem with proof Intermediate+

Theorem (the variational principle for pressure; Ruelle 1973, Walters 1975). Let $f : X \to X$ be a continuous map of a compact metric space and $φ \in C (X, R)$ . Then $P (f, φ) = μ \in M (X, f) sup (h_{μ} (f) + \int_{X} φ d μ) .$ In particular $h_{top} (f) = sup_{μ} h_{μ} (f)$ (the entropy variational principle, the $φ = 0$ case).

Proof. We prove the two inequalities separately.

Lower bound $P (f, φ) \geq h_{μ} (f) + \int φ d μ$ for every $μ \in M (X, f)$ . Fix $μ$ and a finite Borel partition $P = {P_{1}, \dots, P_{r}}$ with $μ (\partial P_{i}) = 0$ and $diam P_{i} < ε$ . For the join $P^{(n)} = ⋁_{k = 0}^{n - 1} f^{- k} P$ , the chain rule and concavity of $t \mapsto - t lo g t$ give, for any function $ψ$ , $H_{μ} (P^{(n)}) + \int S_{n} φ d μ = A \in P^{(n)} \sum μ (A) (- lo g μ (A) + \frac{1}{μ ( A )} \int_{A} S_{n} φ d μ) \leq lo g A \in P^{(n)} \sum e^{s u p_{A} S_{n} φ},$ the last step by the inequality $\sum a_{i} (- lo g a_{i} + b_{i}) \leq lo g \sum e^{b_{i}}$ for a probability vector $(a_{i})$ (Gibbs' inequality / Jensen applied to $- t lo g t$ ), with $a_{i} = μ (A)$ and $b_{i} = \frac{1}{μ ( A )} \int_{A} S_{n} φ d μ \leq sup_{A} S_{n} φ$ . Picking one point $x_{A} \in A$ per atom gives an $(n, ε)$ -separated set (atoms of diameter $< ε$ in the Bowen metric are $d_{n}$ -separated up to a controlled correction), so $\sum_{A} e^{s u p_{A} S_{n} φ} \leq e^{nγ (ε)} Z_{n} (f, φ, ε)$ where $γ (ε) \to 0$ as $ε \to 0$ by uniform continuity of $φ$ . Dividing by $n$ , using $\int S_{n} φ d μ = n \int φ d μ$ and $h_{μ} (f, P) = lim_{n} \frac{1}{n} H_{μ} (P^{(n)})$ 38.06.02, and letting $n \to \infty$ then $ε \to 0$ : $h_{μ} (f, P) + \int φ d μ \leq P (f, φ) .$ Taking the supremum over partitions $P$ gives $h_{μ} (f) + \int φ d μ \leq P (f, φ)$ .

Upper bound $P (f, φ) \leq sup_{μ} (h_{μ} (f) + \int φ d μ)$ . Fix $ε > 0$ and for each $n$ choose an $(n, ε)$ -separated set $E_{n}$ with $\sum_{x \in E_{n}} e^{S_{n} φ (x)} \geq \frac{1}{2} Z_{n} (f, φ, ε)$ . Form the atomic measures $σ_{n} = \frac{1}{\sum _{x \in E_{n}} e ^{S_{n} φ (x)}} x \in E_{n} \sum e^{S_{n} φ (x)} δ_{x}, μ_{n} = \frac{1}{n} k = 0 \sum n - 1 f_{*}^{k} σ_{n},$ the second being the empirical average that makes the weak- $*$ limit invariant. By weak- $*$ compactness pass to a subsequence with $μ_{n_{j}} \to μ \in M (X, f)$ . A partition argument identical to the entropy case (Misiurewicz's proof) shows that for a partition $P$ of diameter $< ε$ with $μ (\partial P_{i}) = 0$ , $j lim sup \frac{1}{n _{j}} lo g x \in E_{n_{j}} \sum e^{S_{n_{j}} φ (x)} \leq h_{μ} (f, P) + \int φ d μ \leq h_{μ} (f) + \int φ d μ .$ The left side is, up to the $\frac{1}{2}$ and the $n^{- 1} lo g$ vanishing, $lim sup_{j} \frac{1}{n _{j}} lo g Z_{n_{j}} (f, φ, ε)$ . Letting $ε \to 0$ yields $P (f, φ) \leq h_{μ} (f) + \int φ d μ \leq sup_{μ} (h_{μ} (f) + \int φ d μ)$ . Combining the two bounds proves the principle. $□$

Bridge. The variational principle builds toward every existence-and-uniqueness statement in thermodynamic formalism and appears again in the equilibrium-state and Parry-measure results of the Advanced section, where the abstract supremum is shown to be attained and, for good potentials, uniquely. The foundational reason the two sides agree is that the same separated-set count $Z_{n}$ controls both the topological growth rate from above (the lower-bound partition argument bounds $H_{μ}$ by $lo g Z_{n}$ ) and the measure-theoretic optimum from below (the empirical measures built from near-maximal separated sets converge to an optimizer) — this is exactly the duality that makes pressure a Legendre-type transform of entropy. Putting these together with 38.06.02, the entropy variational principle $h_{top} = sup_{μ} h_{μ}$ is the $φ = 0$ shadow, and it is dual to the generator theorem: where the generator theorem collapses the partition supremum inside a single measure, the variational principle collapses the measure supremum into a single topological count. The central insight is that pressure is free energy and the equilibrium state is its Gibbs measure, the bridge from ergodic theory to statistical mechanics that this unit makes precise.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Prove that $φ \mapsto P (f, φ)$ is convex on $C (X, R)$ and $1$ -Lipschitz in the supremum norm: $∣ P (f, φ) - P (f, ψ) ∣ \leq ∥ φ - ψ ∥_{\infty}$ .

Hint

Convexity follows from Hölder's inequality applied to the partition function; the Lipschitz bound from $S_{n} ψ \leq S_{n} φ + n ∥ φ - ψ ∥_{\infty}$ .

Answer

For the Lipschitz bound, $S_{n} ψ (x) \leq S_{n} φ (x) + n ∥ φ - ψ ∥_{\infty}$ pointwise, so $Z_{n} (f, ψ, ε) \leq e^{n ∥ φ - ψ ∥_{\infty}} Z_{n} (f, φ, ε)$ ; taking $\frac{1}{n} lo g$ and limits gives $P (f, ψ) \leq P (f, φ) + ∥ φ - ψ ∥_{\infty}$ , and symmetry gives the two-sided bound. For convexity, fix $t \in [0, 1]$ and $φ, ψ$ . Hölder's inequality with exponents $1/ t, 1/ (1 - t)$ gives $x \in E \sum e^{S_{n} (tφ + (1 - t) ψ) (x)} = x \sum (e^{S_{n} φ})^{t} (e^{S_{n} ψ})^{1 - t} \leq (x \sum e^{S_{n} φ})^{t} (x \sum e^{S_{n} ψ})^{1 - t} .$ Taking the supremum over $(n, ε)$ -separated $E$ , then $\frac{1}{n} lo g$ and the limits, gives $P (f, tφ + (1 - t) ψ) \leq tP (f, φ) + (1 - t) P (f, ψ)$ . Convexity is the analytic shadow of pressure being a free energy.

Exercise 4 (medium, symbolic).

Show that the partition function $a_{n} (ε) = lo g Z_{n} (f, φ, ε)$ for fixed $ε$ is subadditive up to a bounded error, so the limit defining the pressure exists for each $ε$ . State the error term.

Hint

A separated set for $m + n$ projects to separated sets for $m$ and for $n$ on the two windows; the Birkhoff sum splits as $S_{m + n} φ = S_{m} φ + (S_{n} φ) \circ f^{m}$ .

Answer

Let $E$ be $(m + n, ε)$ -separated. Two points sharing the same first $m$ -window position and the same last $n$ -window position (to resolution $ε$ ) would be $d_{m + n}$ -close, so the map $x \mapsto (first- m class, last- n class)$ is injective into a product of an $(m, ε)$ -separated set and an $(n, ε)$ -spanning-then-separated set. Using $S_{m + n} φ (x) = S_{m} φ (x) + S_{n} φ (f^{m} x)$ , $x \in E \sum e^{S_{m + n} φ (x)} \leq (\sum e^{S_{m} φ}) (\sum e^{S_{n} φ}) e^{n var (φ, ε)},$ where $var (φ, ε) = sup {∣ φ (x) - φ (y) ∣ : d (x, y) \leq ε}$ accounts for evaluating the second factor at slightly displaced points. Hence $a_{m + n} (ε) \leq a_{m} (ε) + a_{n} (ε) + n var (φ, ε)$ , almost subadditive; the corrected sequence $a_{n} (ε) - n var (φ, ε)$ is genuinely subadditive, so Fekete's lemma 37.02.03 gives the limit, and $var (φ, ε) \to 0$ as $ε \to 0$ removes the error.

Exercise 5 (medium, numeric).

A subshift of finite type has transition matrix $A = (1110)$ (the golden-mean shift: $11$ forbidden). Compute its topological entropy (the pressure of $φ = 0$ ) in nats to three decimals, and name the measure of maximal entropy.

Hint

The entropy of an irreducible SFT is $lo g λ$ where $λ$ is the Perron-Frobenius (largest) eigenvalue of $A$ ; the measure of maximal entropy is the Parry measure.

Answer

The eigenvalues of $A$ solve $λ^{2} - λ - 1 = 0$ , so the Perron-Frobenius eigenvalue is the golden mean $λ = (1 + 5) /2 = 1.618$ . The topological entropy is $h_{top} = lo g λ = lo g 1.618 = 0.481$ nats. The measure of maximal entropy is the Parry measure: the Markov measure whose transition probabilities are $p_{ij} = A_{ij} v_{j} / (λ v_{i})$ with $v$ the right Perron eigenvector and stationary distribution from the left eigenvector. It is the unique $φ = 0$ equilibrium state because the SFT is mixing.

Exercise 6 (medium, symbolic).

Show that for $φ = 0$ the variational principle reduces to the entropy variational principle $h_{top} (f) = sup_{μ} h_{μ} (f)$ , and deduce that a measure of maximal entropy exists whenever entropy is upper semicontinuous on $M (X, f)$ .

Hint

Set $φ = 0$ in $P (f, φ) = sup_{μ} (h_{μ} + \int φ)$ ; for existence use that a u.s.c. function on a compact set attains its supremum.

Answer

With $φ = 0$ , $\int φ d μ = 0$ for all $μ$ and $P (f, 0) = h_{top} (f)$ by definition, so the variational principle becomes $h_{top} (f) = sup_{μ \in M (X, f)} h_{μ} (f)$ . The set $M (X, f)$ is weak- $*$ compact. If $μ \mapsto h_{μ} (f)$ is upper semicontinuous on it, then a u.s.c. function on a compact space attains its supremum, so there is $μ_{0}$ with $h_{μ_{0}} (f) = h_{top} (f)$ , a measure of maximal entropy. Upper semicontinuity holds for expansive $f$ (Bowen): expansiveness bounds entropy by a finite-partition computation that is u.s.c. in $μ$ , so every expansive system has a measure of maximal entropy.

Exercise 7 (hard, symbolic).

For a transitive subshift of finite type with a Hölder potential $φ$ , the Ruelle-Perron-Frobenius transfer operator $(L_{φ} g) (x) = \sum_{f (y) = x} e^{φ (y)} g (y)$ has a simple leading eigenvalue $λ = e^{P (f, φ)}$ . Sketch how this yields a unique equilibrium state with the Gibbs property.

Hint

The leading eigenfunction $h > 0$ ( $L_{φ} h = λh$ ) and the leading eigenmeasure $ν$ of the dual ( $L_{φ}^{*} ν = λ ν$ ) combine into the invariant measure $d μ = h d ν$ ; control of the spectral gap gives uniqueness.

Answer

By the Ruelle-Perron-Frobenius theorem on the space of Hölder functions, $L_{φ}$ has a positive simple eigenvalue $λ$ equal to its spectral radius, with strictly positive eigenfunction $h$ ( $L_{φ} h = λh$ ) and a Borel probability measure $ν$ with $L_{φ}^{*} ν = λ ν$ ; the rest of the spectrum lies in a disc of radius $< λ$ (spectral gap). Normalize so $\int h d ν = 1$ and set $d μ = h d ν$ . The eigenmeasure relation gives the Gibbs bounds $C^{- 1} \leq μ (C_{n} (x)) / e^{- n P + S_{n} φ (x)} \leq C$ with $λ = e^{P}$ , and $μ$ is $f$ -invariant. The conjugated operator $L g = (λh)^{- 1} L_{φ} (h g)$ is a Markov operator with the same gap, whose unique invariant density is constant, forcing $μ$ to be the unique measure with the Gibbs property; a separate computation shows $μ$ attains $h_{μ} + \int φ d μ = P$ , so it is the unique equilibrium state. The spectral gap also yields exponential decay of correlations for $μ$ .

Exercise 8 (hard, symbolic).

Let $f$ be a transitive $C^{2}$ Anosov diffeomorphism and $φ^{u} (x) = - lo g ∣ det D f ∣_{E^{u} (x)} ∣$ the geometric potential. Argue that the equilibrium state of $φ^{u}$ is the SRB (Sinai-Ruelle-Bowen) measure, and relate $P (f, φ^{u})$ to entropy.

Hint

Use the variational principle together with the Pesin/Ruelle inequality $h_{μ} (f) \leq \int lo g ∣ det D f ∣_{E^{u}} ∣ d μ$ , and identify when equality holds.

Answer

For any invariant $μ$ , the Ruelle inequality 38.07.02 gives $h_{μ} (f) \leq \int lo g ∣ det D f ∣_{E^{u}} ∣ d μ = - \int φ^{u} d μ$ , i.e. $h_{μ} (f) + \int φ^{u} d μ \leq 0$ . Hence $P (f, φ^{u}) = sup_{μ} (h_{μ} + \int φ^{u}) \leq 0$ . A measure attains the supremum (value $0$ ) precisely when Pesin's entropy formula holds with equality, $h_{μ} (f) = \int lo g ∣ det D f ∣_{E^{u}} ∣ d μ$ , which characterizes measures whose conditionals on unstable manifolds are absolutely continuous — the SRB measure 38.07.02. For a transitive Anosov system this measure exists and is unique (Sinai-Ruelle-Bowen), so it is the unique equilibrium state of $φ^{u}$ and $P (f, φ^{u}) = 0$ . This identifies the physically observed measure as a thermodynamic equilibrium state, the central dictionary entry of the theory.

Advanced results Master

Theorem 1 (variational principle; Ruelle 1973, Walters 1975). For continuous $f$ on a compact metric space and $φ \in C (X)$ , $P (f, φ) = sup_{μ \in M (X, f)} (h_{μ} (f) + \int φ d μ)$ . The pressure is a convex, $1$ -Lipschitz, monotone functional of $φ$ , translation-covariant under additive constants, and $P (f, φ \circ f - φ + ψ) = P (f, ψ)$ (cohomology-invariance): pressure depends on $φ$ only through its cohomology class plus the entropy term ^{[Walters 1975]}.

Theorem 2 (existence of equilibrium states under upper semicontinuity). If $μ \mapsto h_{μ} (f)$ is upper semicontinuous on the weak- $*$ -compact convex set $M (X, f)$ , then every $φ \in C (X)$ has at least one equilibrium state, and the equilibrium states form a nonempty compact convex face of $M (X, f)$ whose extreme points are ergodic. Expansive maps — in particular subshifts of finite type, Anosov diffeomorphisms, and Axiom-A systems on their basic sets — have upper semicontinuous entropy, hence equilibrium states for every continuous potential ^{[Bowen 1975]}.

Theorem 3 (measure of maximal entropy; Parry measure). For an irreducible subshift of finite type with transition matrix $A$ , $h_{top} (σ) = lo g λ_{A}$ with $λ_{A}$ the Perron-Frobenius eigenvalue, and the unique measure of maximal entropy is the Parry measure, the Markov measure with transition probabilities $p_{ij} = A_{ij} v_{j} / (λ_{A} v_{i})$ and stationary vector $u_{i} v_{i}$ , where $u, v$ are the left/right Perron eigenvectors normalized by $\sum u_{i} v_{i} = 1$ . Uniqueness holds because the SFT is expansive and mixing; for a general expansive system with specification the measure of maximal entropy is likewise unique ^{[Bowen 1974]}.

Theorem 4 (uniqueness for Hölder potentials; Ruelle-Bowen). On a transitive subshift of finite type, or a transitive Anosov / Axiom-A system, a Hölder-continuous potential $φ$ has a unique equilibrium state $μ_{φ}$ , which is also the unique Gibbs state: $C^{- 1} \leq μ_{φ} (C_{n} (x)) / exp (- n P (f, φ) + S_{n} φ (x)) \leq C$ . It is constructed from the leading eigendata of the Ruelle-Perron-Frobenius transfer operator $L_{φ}$ , is ergodic with exponential decay of correlations and a central limit theorem, and depends real-analytically on $φ$ within the Hölder class away from phase transitions ^{[Bowen 1975]}.

Theorem 5 (SRB measures as equilibrium states). For a transitive $C^{2}$ Anosov diffeomorphism, the SRB measure is the unique equilibrium state of the geometric potential $φ^{u} = - lo g ∣ det D f ∣_{E^{u}} ∣$ , with $P (f, φ^{u}) = 0$ , equivalently the unique invariant measure satisfying Pesin's entropy formula $h_{μ} (f) = \int lo g ∣ det D f ∣_{E^{u}} ∣ d μ$ . Its conditionals on unstable manifolds are absolutely continuous, and it is the physical measure: time averages of continuous observables converge to $\int \cdot d μ_{SRB}$ for Lebesgue-a.e. initial point ^{[Bowen 1975]}.

Synthesis. The five results are one architecture, and the foundational reason they cohere is that the variational principle is a Legendre duality between the convex pressure functional $φ \mapsto P (f, φ)$ and the concave entropy functional $μ \mapsto h_{μ} (f)$ : the equilibrium state is exactly a tangent functional to $P$ at $φ$ , so existence of equilibrium states is the existence of supporting hyperplanes and uniqueness is the differentiability of pressure. This is dual to the generator theorem and the Shannon-McMillan-Breiman theorem of 38.06.02 and 38.06.03, which fix entropy inside a single measure: here the supremum is taken across measures, and the measure of maximal entropy is the $φ = 0$ tangent — the Parry measure on an SFT, putting the Perron-Frobenius eigenvector to work. Putting these together, the thermodynamic dictionary is exact: pressure is free energy, the equilibrium state is the Gibbs measure, the transfer operator is the transfer matrix of statistical mechanics, phase transitions are non-unique equilibrium states, and the SRB measure is the physically selected equilibrium state for the geometric potential — this is exactly the central insight that the long-run statistics of a chaotic system are governed by a variational principle of the same form that governs a lattice gas, the bridge from Boltzmann-Gibbs thermodynamics to smooth dynamics that Sinai, Ruelle, and Bowen built.

Full proof set Master

Proposition 1 (pressure is well-defined and independent of the separated/spanning choice). For each $ε > 0$ the spanning-set sum and the separated-set sum have the same exponential growth rate, and $P (f, φ) = lim_{ε \to 0} lim sup_{n} \frac{1}{n} lo g Z_{n} (f, φ, ε)$ exists in $[- \infty, + \infty)$ .

Proof. A maximal $(n, ε)$ -separated set is $(n, ε)$ -spanning, so writing $Q_{n} (ε)$ for the infimum spanning sum, $Q_{n} (ε) \leq Z_{n} (f, ε)$ . Conversely an $(n, 2 ε)$ -separated set $E$ and an $(n, ε)$ -spanning set $F$ admit an injection $E ↪ F$ (each $x \in E$ has a distinct $ε$ -spanning neighbor, distinct because $E$ is $2 ε$ -separated), and $S_{n} φ$ changes by at most $n var (φ, ε)$ between matched points, so $Z_{n} (f, 2 ε) \leq e^{n var (φ, ε)} Q_{n} (ε)$ . Thus the two growth rates agree up to $var (φ, ε) \to 0$ . The double limit exists because $lim sup_{n} \frac{1}{n} lo g Z_{n} (f, ε)$ is monotone in $ε$ (smaller $ε$ allows more separated points), hence has a limit as $ε ↓ 0$ ; finiteness above follows from $Z_{n} \leq N (n, ε) e^{n ∥ φ ∥_{\infty}}$ with $N (n, ε)$ the spanning count, whose rate is $h_{top} (f) < \infty$ for compact $X$ . $□$

Proposition 2 (Gibbs' inequality, the engine of the lower bound). For a probability vector $(a_{i})_{i = 1}^{r}$ and reals $(b_{i})$ , $\sum_{i} a_{i} (- lo g a_{i} + b_{i}) \leq lo g \sum_{i} e^{b_{i}}$ , with equality iff $a_{i} = e^{b_{i}} / \sum_{j} e^{b_{j}}$ .

Proof. Let $Z = \sum_{j} e^{b_{j}}$ and $q_{i} = e^{b_{i}} / Z$ , a probability vector. The relative entropy is non-negative: $\sum_{i} a_{i} lo g (a_{i} / q_{i}) \geq 0$ by Jensen applied to the convex $t \mapsto t lo g t$ , equivalently $- \sum_{i} a_{i} lo g (q_{i} / a_{i}) \geq - lo g \sum_{i} a_{i} (q_{i} / a_{i}) = 0$ . Expanding $lo g q_{i} = b_{i} - lo g Z$ , $0 \leq i \sum a_{i} lo g \frac{a _{i}}{q _{i}} = i \sum a_{i} lo g a_{i} - i \sum a_{i} b_{i} + lo g Z,$ which rearranges to $\sum_{i} a_{i} (- lo g a_{i} + b_{i}) \leq lo g Z = lo g \sum_{i} e^{b_{i}}$ . Equality in Jensen for the strictly convex $t lo g t$ forces $a_{i} / q_{i}$ constant, i.e. $a_{i} = q_{i}$ . $□$

Proposition 3 (the supremum is attained on ergodic measures; ergodic decomposition of pressure). $P (f, φ) = sup {h_{μ} (f) + \int φ d μ : μ \in M (X, f) ergodic}$ , and if an equilibrium state exists, an ergodic one exists.

Proof. Entropy is affine on $M (X, f)$ : $h_{t μ + (1 - t) ν} (f) = t h_{μ} (f) + (1 - t) h_{ν} (f)$ , by the affineness of $μ \mapsto h_{μ} (f, P)$ at each partition (a consequence of the concavity-and-convexity of conditional entropy in the measure) followed by the supremum over partitions. The integral $μ \mapsto \int φ d μ$ is affine, so $g (μ) = h_{μ} (f) + \int φ d μ$ is affine. The ergodic decomposition writes any $μ = \int_{erg} ν d τ (ν)$ , and affineness gives $g (μ) = \int g (ν) d τ (ν) \leq sup_{ν erg} g (ν)$ . Hence the supremum over all invariant measures equals the supremum over ergodic ones. If $μ$ is an equilibrium state, $g (μ) = P = \int g (ν) d τ$ , forcing $g (ν) = P$ for $τ$ -a.e. ergodic $ν$ , each of which is then an ergodic equilibrium state. $□$

Proposition 4 (equilibrium states are the tangent functionals to pressure). Let entropy be upper semicontinuous. A measure $μ \in M (X, f)$ is an equilibrium state for $φ$ if and only if $μ$ is a tangent functional to $P$ at $φ$ , meaning $P (f, φ + ψ) \geq P (f, φ) + \int ψ d μ$ for all $ψ \in C (X)$ .

Proof. If $μ$ is an equilibrium state for $φ$ , then for any $ψ$ , $P (f, φ + ψ) \geq h_{μ} (f) + \int (φ + ψ) d μ = (h_{μ} (f) + \int φ d μ) + \int ψ d μ = P (f, φ) + \int ψ d μ$ , so $μ$ is tangent. Conversely, suppose $μ$ is tangent. Upper semicontinuity gives, by the duality between the convex $P$ and the concave entropy (the variational principle realizes $P$ as the convex conjugate of $- h$ ), the formula $h_{μ} (f) = in f_{ψ \in C (X)} (P (f, ψ) - \int ψ d μ)$ for every invariant $μ$ . Tangency at $φ$ says $P (f, φ + ψ) - \int (φ + ψ) d μ \geq P (f, φ) - \int φ d μ$ for all $ψ$ , so the infimum defining $h_{μ} (f)$ is attained at $ψ = φ$ , giving $h_{μ} (f) = P (f, φ) - \int φ d μ$ , i.e. $μ$ is an equilibrium state. Uniqueness of the equilibrium state is therefore equivalent to differentiability (Gateaux) of $P$ at $φ$ . $□$

Connections Master

The general topological entropy 38.06.01 is the special case $φ = 0$ of topological pressure: $h_{top} (f) = P (f, 0)$ , and the $(n, ε)$ -separated and open-cover constructions of pressure are the potential-weighted refinements of the entropy constructions in that unit. The entropy variational principle $h_{top} (f) = sup_{μ} h_{μ} (f)$ is the $φ = 0$ shadow of the pressure variational principle proved here, so this unit subsumes and quantifies the topological-entropy theory.
The Kolmogorov-Sinai entropy and generator theorem 38.06.02 supplies the measure-theoretic side of the variational principle: the term $h_{μ} (f)$ is the entropy built in that unit, the chain rule and join structure power the lower-bound partition estimate, and the affineness of $μ \mapsto h_{μ} (f)$ that gives the ergodic decomposition of pressure rests on the conditional-entropy machinery there. Pressure takes the supremum of that invariant across the simplex of invariant measures.
The Shannon-McMillan-Breiman theorem 38.06.03 is the pointwise companion: for the equilibrium (Gibbs) state, SMB makes $- \frac{1}{n} lo g μ (C_{n} (x)) \to h_{μ} (f)$ a.e., and combined with the Gibbs property $μ (C_{n} (x)) ≍ e^{- n P + S_{n} φ (x)}$ it yields the a.e. statement $\frac{1}{n} S_{n} φ (x) \to \int φ d μ$ , identifying the pressure $P = h_{μ} (f) + \int φ d μ$ orbit-by-orbit. The typical-set count $e^{nh}$ becomes the weighted count $e^{n P}$ .
The hyperbolic-sets and Smale decomposition theory 38.03.01 is where equilibrium states become geometric: Markov partitions code an Axiom-A basic set as a subshift of finite type, transporting the Parry and Gibbs measures of this unit onto the manifold, and the Bowen specification property that guarantees uniqueness is a dynamical consequence of transitive hyperbolicity.
Pesin theory and the entropy formula 38.07.02 closes the dictionary: the SRB measure is the equilibrium state of the geometric potential $φ^{u} = - lo g ∣ det D f ∣_{E^{u}} ∣$ with pressure zero, the Ruelle inequality is the one-sided variational bound $h_{μ} + \int φ^{u} \leq 0$ , and Pesin's formula is the equality case that selects the physical measure as a thermodynamic equilibrium state.

Historical & philosophical context Master

Topological pressure entered dynamics as a deliberate import from statistical mechanics. David Ruelle, working on one-dimensional lattice gases, defined a pressure for expansive systems with specification in a 1973 paper in the Transactions of the American Mathematical Society ^{[Ruelle 1973]} and proved a variational principle for that class, modeling the construction on the free energy of a spin system. Peter Walters removed the specification hypothesis in 1975 in the American Journal of Mathematics ^{[Walters 1975]}, establishing the variational principle $P (f, φ) = sup_{μ} (h_{μ} (f) + \int φ d μ)$ for every continuous map of a compact metric space and every continuous potential, the form used today. The $φ = 0$ case, the entropy variational principle $h_{top} = sup_{μ} h_{μ}$ , had been proved earlier by Dinaburg, Goodman, and Goodwyn around 1969-1971; Misiurewicz later gave the short proof of the hard inequality reproduced in the modern literature.

The theory of equilibrium states was built in parallel by Yakov Sinai, Ruelle, and Rufus Bowen. Sinai's 1972 survey ^{[Sinai 1972]} introduced Gibbs measures into ergodic theory and constructed them for Anosov systems via Markov partitions; Ruelle developed the transfer-operator (Ruelle-Perron-Frobenius) method, and Bowen's 1974 paper on unique equilibrium states ^{[Bowen 1974]} and his 1975 Springer Lecture Notes synthesized the subject, proving existence and uniqueness for Hölder potentials on Axiom-A basic sets and identifying the SRB measure as the equilibrium state of the geometric potential. The resulting dictionary — pressure as free energy, equilibrium states as Gibbs states, the transfer operator as the transfer matrix, non-uniqueness as phase transition — made the analogy between hyperbolic dynamics and one-dimensional statistical mechanics into a working mathematical tool, and the measures it produced became the standard notion of a physically observable invariant measure for chaotic systems.

Bibliography Master

@article{Ruelle1973,
  author  = {Ruelle, David},
  title   = {Statistical mechanics on a compact set with $\mathbb{Z}^\nu$ action satisfying expansiveness and specification},
  journal = {Transactions of the American Mathematical Society},
  volume  = {187},
  year    = {1973},
  pages   = {237--251}
}

@article{Walters1975,
  author  = {Walters, Peter},
  title   = {A variational principle for the pressure of continuous transformations},
  journal = {American Journal of Mathematics},
  volume  = {97},
  number  = {4},
  year    = {1975},
  pages   = {937--971}
}

@article{Bowen1974,
  author  = {Bowen, Rufus},
  title   = {Some systems with unique equilibrium states},
  journal = {Mathematical Systems Theory},
  volume  = {8},
  number  = {3},
  year    = {1974},
  pages   = {193--202}
}

@article{Sinai1972,
  author  = {Sinai, Yakov G.},
  title   = {Gibbs measures in ergodic theory},
  journal = {Russian Mathematical Surveys},
  volume  = {27},
  number  = {4},
  year    = {1972},
  pages   = {21--69}
}

@book{Bowen1975,
  author    = {Bowen, Rufus},
  title     = {Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms},
  publisher = {Springer},
  series    = {Lecture Notes in Mathematics},
  volume    = {470},
  year      = {1975}
}

@book{Ruelle1978,
  author    = {Ruelle, David},
  title     = {Thermodynamic Formalism},
  publisher = {Addison-Wesley},
  series    = {Encyclopedia of Mathematics and its Applications},
  volume    = {5},
  year      = {1978}
}

@book{Walters1982,
  author    = {Walters, Peter},
  title     = {An Introduction to Ergodic Theory},
  publisher = {Springer},
  series    = {Graduate Texts in Mathematics},
  volume    = {79},
  year      = {1982}
}

@book{Keller1998,
  author    = {Keller, Gerhard},
  title     = {Equilibrium States in Ergodic Theory},
  publisher = {Cambridge University Press},
  series    = {London Mathematical Society Student Texts},
  volume    = {42},
  year      = {1998}
}

Prerequisites

38.06.02
38.06.03

Tier anchors

beginner: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 9 (informal: pressure as a weighted growth rate, free energy of a dynamical system); Brin-Stuck 2002 *Introduction to Dynamical Systems* (Cambridge) §4.3 (topological entropy and the variational principle)
intermediate: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) §9.1-9.4 (topological pressure, the variational principle, equilibrium states); Bowen 1975 *Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms* (Springer LNM 470) §1 (pressure on subshifts)
master: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 9 (pressure, the variational principle, equilibrium states, the measure of maximal entropy); Bowen 1975 *Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms* (Springer LNM 470); Ruelle 1978 *Thermodynamic Formalism* (Addison-Wesley, Encyclopedia of Mathematics 5); Keller 1998 *Equilibrium States in Ergodic Theory* (Cambridge LMS Student Texts 42)

References

Ruelle — Statistical mechanics on a compact set with Z^nu action satisfying expansiveness and specification · Transactions of the American Mathematical Society 187 (1973), 237-251
Walters — A variational principle for the pressure of continuous transformations · American Journal of Mathematics 97 (1975), 937-971
Bowen — Some systems with unique equilibrium states · Mathematical Systems Theory 8 (1974), 193-202
Sinai — Gibbs measures in ergodic theory · Russian Mathematical Surveys 27 (1972), 21-69
Walters — An Introduction to Ergodic Theory · Springer GTM 79, 1982, Ch. 9 (pressure, the variational principle, equilibrium states)
Bowen — Equilibrium States and the Ergodic Theory of Anosov Diffeomorphisms · Springer Lecture Notes in Mathematics 470, 1975

Estimated time

beginner: 18m
intermediate: 60m
master: 98m