38.06.01 · dynamics / entropy

Topological Entropy

shipped3 tiersLean: none

Anchor (Master): Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 7; Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge University Press) §3.1 and Suppl.; Bowen 1971 *Entropy for group endomorphisms and homogeneous spaces* (Trans. Amer. Math. Soc. 153); Yomdin 1987 / Newhouse 1989 (the volume-growth = entropy theorems)

Intuition Beginner

Picture a continuous machine that takes a space and folds it back onto itself once per tick. You cannot see individual points — your eyesight is finite, so you can only tell two starting points apart if, at some moment, the machine has driven them far enough away from each other to resolve. Topological entropy is one number answering a single question: as you watch for longer and longer, how fast does the count of startings-points-you-can-tell-apart grow?

A tame machine, like a slow rotation, never spreads nearby points; after a thousand ticks you still cannot distinguish many more starts than you could after one tick. The count grows slowly, and the entropy is zero. A wild machine, like a map that stretches the space by a factor of two every tick, pulls apart neighbours fast: each extra tick of watching roughly doubles how many starts you can separate. The count grows like a power of two, and the entropy is the logarithm of that base — here, the logarithm of $2$ .

So entropy is a growth rate of distinguishability. It does not care about any measure or probability — only about the topology, about which points the dynamics eventually pries apart. Two machines that are the same after a continuous relabelling have the same entropy, which makes it a fingerprint: a single number that survives every change of coordinates and tells a clockwork system from a genuinely complicated one.

Visual Beginner

Watch the doubling map on a circle: each tick replaces an angle by twice the angle, wrapped around. A tiny arc gets stretched to double its length each tick, and a viewer who can resolve features down to one fixed size sees the number of separately-visible pieces double every tick.

A short table contrasts how fast different machines proliferate distinguishable starts.

machine	what it does each tick	distinguishable starts after $n$ ticks	entropy
rigid rotation	carries arcs rigidly, never spreads	bounded	$0$
doubling map	stretches by $2$ , wraps around	about $2^{n}$	$lo g 2$
full shift on $k$ symbols	reveals one of $k$ labels	exactly $k^{n}$	$lo g k$
stretch by $m$ , wrap	stretches by $m$	about $m^{n}$	$lo g m$

Worked example Beginner

We count distinguishable starts for the doubling map and read off its entropy. Each tick replaces $x$ by $2 x$ and drops the whole-number part. In binary, $x = 0. b_{1} b_{2} b_{3} \dots$ , and doubling deletes the leading digit: $0. b_{1} b_{2} b_{3} \dots \to 0. b_{2} b_{3} b_{4} \dots$ .

Step 1. Fix a viewing resolution. Say you can resolve points down to size $1/2$ — you can tell which half of the circle, left or right, a point sits in. One snapshot sorts every start into $2$ groups by its first binary digit.

Step 2. Watch for $n$ ticks. After each tick the machine has shifted the digits, so the digit you can currently read was originally the $1$ st, then the $2$ nd, then the $3$ rd, and so on. Watching $n$ ticks lets you read the first $n$ binary digits of the start. Two starts are distinguishable-by-watching exactly when their first $n$ digits differ.

Step 3. Count the groups. There are $2^{n}$ possible length- $n$ binary strings, so watching $n$ ticks sorts starts into $2^{n}$ distinguishable groups: $2, 4, 8, 16, \dots$ as $n = 1, 2, 3, 4, \dots$ .

Step 4. Read the growth rate. The count is $2^{n}$ , so its logarithm is $n lo g 2$ , and dividing by the number of ticks gives $lo g 2$ per tick. That per-tick growth rate of the logarithm of the count is the topological entropy: $lo g 2$ .

What this tells us: stretching by a factor of two each tick is exactly what makes the count of tellable-apart starts double each tick, and the entropy is the logarithm of the stretch factor. A map that stretched by $3$ instead would give $3^{n}$ groups and entropy $lo g 3$ .

Check your understanding Beginner

Exercise (easy, multiple choice).

Topological entropy of a continuous map measures:

A. The total number of points in the space. B. The exponential growth rate of the number of orbit segments a finite-resolution observer can tell apart. C. How long a single orbit takes to return near its start. D. The probability that two random points are close.

Hint

Think of watching longer and longer and counting how many starts become distinguishable.

Answer

B. Topological entropy is the per-tick growth rate of the logarithm of the number of length- $n$ orbit segments that a fixed-resolution observer can separate. Feedback-correct: this is exactly the growth-of-distinguishability picture. Feedback-wrong: A counts points, which is unrelated to a growth rate; C describes recurrence and return times, a different invariant; D mentions probability, but topological entropy uses no measure at all — only the topology of which points get pried apart.

Formal definition Intermediate+

Let $(X, f)$ be a continuous map of a compact metric space, as in 38.01.03. There are two equivalent routes to the topological entropy $h_{top} (f)$ : the open-cover route of Adler, Konheim, and McAndrew, and the metric route of Bowen and Dinaburg.

Open covers. For an open cover $U$ of $X$ , let $N (U)$ be the minimal cardinality of a subcover; compactness makes this finite. The join of covers is $U \lor V = {U \cap V : U \in U, V \in V}$ , and $f^{- 1} U = {f^{- 1} U : U \in U}$ is again an open cover. Write $$ \mathcal{U}0^{n-1} = \bigvee{k=0}^{n-1} f^{-k}\mathcal{U} = \mathcal{U} \vee f^{-1}\mathcal{U} \vee \cdots \vee f^{-(n-1)}\mathcal{U}. $$ An element of $U_{0}^{n - 1}$ is a set of points whose first $n$ iterates follow a prescribed itinerary through $U$ . The quantity $a_{n} = lo g N (U_{0}^{n - 1})$ is subadditive, $a_{m + n} \leq a_{m} + a_{n}$ , so the limit below exists by Fekete's lemma.

Definition (entropy of a cover; topological entropy). The entropy of $f$ relative to $U$ is $$ h(f, \mathcal{U}) = \lim_{n \to \infty} \frac{1}{n} \log N!\Big( \bigvee_{k=0}^{n-1} f^{-k}\mathcal{U} \Big) = \inf_{n \geq 1} \frac{1}{n} \log N!\Big( \bigvee_{k=0}^{n-1} f^{-k}\mathcal{U} \Big), $$ and the topological entropy of $f$ is the supremum over all open covers, $$ h_{\mathrm{top}}(f) = \sup_{\mathcal{U}} h(f, \mathcal{U}). \tag{AKM} $$ ^{[Adler 1965]}

Bowen-Dinaburg metrics. For $n \geq 1$ define the Bowen metric $$ d_n(x, y) = \max_{0 \leq k < n} d\big(f^k(x), f^k(y)\big), $$ the largest separation of the two orbits over the first $n$ steps. A set $E \subseteq X$ is * $(n, ε)$ -separated* if $d_{n} (x, y) > ε$ for all distinct $x, y \in E$ , and $(n, ε)$ -spanning if every point of $X$ is within $d_{n}$ -distance $ε$ of some point of $E$ . Let $s (n, ε)$ be the maximal cardinality of an $(n, ε)$ -separated set and $r (n, ε)$ the minimal cardinality of an $(n, ε)$ -spanning set; compactness makes both finite, and $r (n, ε) \leq s (n, ε) \leq r (n, ε /2)$ .

Definition (Bowen-Dinaburg entropy). Set $$ h_{\mathrm{sep}}(f) = \lim_{\varepsilon \to 0^+} \limsup_{n \to \infty} \frac{1}{n} \log s(n, \varepsilon), \qquad h_{\mathrm{span}}(f) = \lim_{\varepsilon \to 0^+} \limsup_{n \to \infty} \frac{1}{n} \log r(n, \varepsilon). $$ ^{[Bowen 1971]}, ^{[Dinaburg 1970]} The two limits coincide and equal the open-cover entropy, $h_{sep} (f) = h_{span} (f) = h_{top} (f)$ ; this equivalence is proved in the Key theorem below. The metrics $d_{n}$ are the Bowen-Dinaburg metrics; $s (n, ε)$ counts the orbit segments that are $ε$ -distinguishable over $n$ steps, the rigorous form of the Beginner-tier "distinguishable starts".

The entropy is unchanged under topological conjugacy: if $h \circ f = g \circ h$ with $h$ a homeomorphism of compact metric spaces, then $h_{top} (f) = h_{top} (g)$ , since $h$ carries covers to covers and, being uniformly continuous with uniformly continuous inverse, carries separated sets to separated sets up to a fixed change of scale.

Counterexamples to common slips

Topological entropy is not a measure-theoretic quantity. It uses no invariant measure; it depends only on the topology and the map. The Kolmogorov-Sinai entropy $h_{μ} (f)$ of 38.06.02 is attached to a chosen invariant measure $μ$ , and in general $h_{μ} (f) \leq h_{top} (f)$ , with equality for a measure of maximal entropy.
The cover entropy $h (f, U)$ for one cover is only a lower bound. A coarse cover sees little; the one-set cover ${X}$ gives $h (f, {X}) = 0$ for every $f$ . The supremum over covers (or the $ε \to 0$ limit over separated sets) is essential — the analogue of taking the finest practical partition in measure-theoretic entropy.
Separated and spanning counts differ by an order of refinement, not in their growth rate. The inequalities $r (n, ε) \leq s (n, ε) \leq r (n, ε /2)$ show the two count functions are interleaved, so after the $ε \to 0$ limit they produce the same number; conflating $s (n, ε)$ and $r (n, ε)$ at fixed $ε$ is the slip, taking the limit removes it.
Entropy is not the topological-mixing or transitivity of 38.01.03. A system can be minimal with zero entropy (irrational rotation) or have positive entropy without being minimal (doubling map). Positive entropy is a quantitative strengthening of "complicated", orthogonal to the connectivity hierarchy minimal/transitive/mixing.

Key theorem with proof Intermediate+

Theorem (Bowen-Dinaburg = Adler-Konheim-McAndrew). Let $f : X \to X$ be continuous on a compact metric space. Then $$ h_{\mathrm{top}}(f) = \lim_{\varepsilon \to 0^+} \limsup_{n \to \infty} \frac{1}{n} \log s(n, \varepsilon) = \lim_{\varepsilon \to 0^+} \limsup_{n \to \infty} \frac{1}{n} \log r(n, \varepsilon). $$ The open-cover entropy and the separated- and spanning-set entropies all agree. (See ^{[Walters Ch. 7]}, ^{[Bowen 1971]}, ^{[Katok-Hasselblatt 1995 §3.1]}.)

Proof. First, $s (n, ε)$ and $r (n, ε)$ produce the same limit. From $r (n, ε) \leq s (n, ε) \leq r (n, ε /2)$ — the left because a maximal $(n, ε)$ -separated set is $(n, ε)$ -spanning (else a point at $d_{n}$ -distance $> ε$ from all of $E$ could be added, contradicting maximality), the right because an $ε /2$ -spanning set lets at most one separated point lie in each $d_{n}$ -ball of radius $ε /2$ — both count functions have the same exponential growth rate after letting $ε \to 0$ . Call the common value $h_{d} (f)$ .

Now relate $h_{d} (f)$ to covers. Fix $ε > 0$ and let $U$ be a finite open cover by sets of diameter $< ε$ , with Lebesgue number $2 δ$ (every $d$ -ball of radius $δ$ lies in some $U \in U$ ). The atoms of $U_{0}^{n - 1}$ have $d_{n}$ -diameter $< ε$ , since membership in a single atom forces $f^{k} (x), f^{k} (y)$ to share a set of diameter $< ε$ for each $k < n$ . An $(n, ε)$ -separated set meets each atom at most once, so $s (n, ε) \leq N (U_{0}^{n - 1})$ ; dividing by $n$ and taking limits, $h_{d} (f) \leq sup_{U} h (f, U) = h_{top} (f)$ .

Conversely, with the same $U$ of Lebesgue number $2 δ$ , an $(n, δ)$ -spanning set $E$ has the property that the $d_{n}$ -balls of radius $δ$ centred at its points cover $X$ , and each such ball lies inside a single atom of $U_{0}^{n - 1}$ (by the Lebesgue-number property applied coordinatewise). Hence ${$ atoms meeting these balls $}$ is a subcover of size $\leq r (n, δ)$ , giving $N (U_{0}^{n - 1}) \leq r (n, δ)$ . Dividing by $n$ , taking $n \to \infty$ then $δ \to 0$ , $h (f, U) \leq h_{d} (f)$ for every $U$ , so $h_{top} (f) \leq h_{d} (f)$ . The two inequalities give $h_{top} (f) = h_{d} (f)$ . $□$

Bridge. This equivalence builds toward every concrete computation and appears again in the model evaluations of the Advanced results, where the separated-set count is the workable side — for the full shift one simply counts cylinder words, for an expanding map one counts inverse branches. The foundational reason the two definitions agree is that an open cover of small diameter and a Bowen ball of small radius measure the same thing — orbit segments that stay close for $n$ steps — so the supremum over covers and the limit over resolutions are two readings of one exponential growth rate; this is exactly the topological shadow of the partition-refinement that defines Kolmogorov-Sinai entropy in 38.06.02, where the join of partitions plays the role of the join of covers. The metric form generalises the Beginner-tier counting of distinguishable starts into a definition valid on any compact metric space, and putting these together with the variational principle shows topological entropy to be the supremum over invariant measures of the metric entropies, so the central insight is that one number certified by counting separated orbit segments is simultaneously the maximal rate of measure-theoretic information production. The bridge is the recognition that distinguishability over $n$ steps, measured topologically by covers or metrically by Bowen balls, is the single invariant the whole entropy theory turns on.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Prove that $a_{n} = lo g N (⋁_{k = 0}^{n - 1} f^{- k} U)$ is subadditive, $a_{m + n} \leq a_{m} + a_{n}$ , so that $h (f, U) = lim_{n} a_{n} / n$ exists.

Hint

Use $N (A \lor B) \leq N (A) N (B)$ and $N (f^{- 1} A) \leq N (A)$ .

Answer

If $A$ has a subcover of size $N (A)$ and $B$ one of size $N (B)$ , the pairwise intersections of their members form a subcover of $A \lor B$ of size $\leq N (A) N (B)$ , so $N (A \lor B) \leq N (A) N (B)$ . Also a subcover of $A$ pulls back to a subcover of $f^{- 1} A$ of the same size, so $N (f^{- 1} A) \leq N (A)$ . Now split the window: $⋁_{k = 0}^{m + n - 1} f^{- k} U = (⋁_{k = 0}^{m - 1} f^{- k} U) \lor f^{- m} (⋁_{k = 0}^{n - 1} f^{- k} U)$ . Taking $lo g N$ of both sides and applying the two inequalities, $a_{m + n} \leq lo g N (k = 0 ⋁ m - 1 f^{- k} U) + lo g N (f^{- m} k = 0 ⋁ n - 1 f^{- k} U) \leq a_{m} + a_{n} .$ Fekete's subadditivity lemma gives $a_{n} / n \to in f_{n} a_{n} / n = h (f, U)$ . Rubric: full credit for the two cover inequalities and the window split.

Exercise 4 (medium, symbolic).

Prove that for every $m \geq 1$ , $h_{top} (f^{m}) = m h_{top} (f)$ .

Hint

Work with separated sets. An $(n, ε)$ -separated set for $f^{m}$ relates to an $(nm, ε)$ -separated set for $f$ ; use the Bowen metric over windows of length $m$ versus $1$ .

Answer

Write $d_{n}^{(g)}$ for the Bowen metric of a map $g$ . For $g = f^{m}$ , $d_{n}^{(f^{m})} (x, y) = max_{0 \leq k < n} d (f^{mk} x, f^{mk} y) \leq max_{0 \leq j < nm} d (f^{j} x, f^{j} y) = d_{nm}^{(f)} (x, y)$ , so the metric $d_{n}^{(f^{m})}$ is coarser than $d_{nm}^{(f)}$ : any two points separated by more than $ε$ under $d_{n}^{(f^{m})}$ are separated under $d_{nm}^{(f)}$ too, giving $s_{f^{m}} (n, ε) \leq s_{f} (nm, ε)$ . Conversely, by uniform continuity of $f, \dots, f^{m - 1}$ , for each $ε$ there is $ε^{'} \leq ε$ with $d_{nm}^{(f)} (x, y) > ε \Rightarrow d_{n}^{(f^{m})} (x, y) > ε^{'}$ , so $s_{f} (nm, ε) \leq s_{f^{m}} (n, ε^{'})$ . Thus $\frac{1}{n} lo g s_{f^{m}} (n, \cdot)$ and $\frac{m}{nm} lo g s_{f} (nm, \cdot)$ are sandwiched together; taking $lim sup_{n}$ then $ε \to 0$ gives $h_{top} (f^{m}) = m h_{top} (f)$ . Rubric: full credit for the Bowen-metric comparison over windows of length $m$ and the uniform-continuity reverse bound, yielding the factor $m$ .

Exercise 5 (medium, symbolic).

Show that the full two-sided shift $σ$ on $Σ_{k} = {0, \dots, k - 1}^{Z}$ (with metric $d (x, y) = 2^{- m i n {∣ j ∣ : x_{j} \neq = y_{j}}}$ ) has $h_{top} (σ) = lo g k$ .

Hint

Two sequences agreeing on the central block of length $\approx n$ are $d_{n}$ -close; count central blocks.

Answer

Fix $ε = 2^{- 1}$ . Two sequences $x, y$ have $d_{n} (x, y) = max_{0 \leq j < n} d (σ^{j} x, σ^{j} y) \leq ε$ iff for each $j < n$ the sequences $σ^{j} x, σ^{j} y$ agree on coordinate $0$ , i.e. $x$ and $y$ agree on coordinates $0, 1, \dots, n - 1$ . Distinct words on ${0, \dots, k - 1}^{n}$ give $d_{n}$ -distance $> ε$ , so the $k^{n}$ sequences with all-zero tails outside $[0, n)$ form an $(n, ε)$ -separated set: $s (n, ε) \geq k^{n}$ . Conversely any two sequences agreeing on $[0, n)$ satisfy $d_{n} \leq 2^{- n} \leq ε$ , so a set of representatives, one per length- $n$ central word, is $(n, ε)$ -spanning of size $k^{n}$ : $r (n, ε) \leq k^{n}$ . Therefore $s (n, ε) = r (n, ε) = k^{n}$ and $h_{top} (σ) = lim_{ε} lim sup_{n} \frac{1}{n} lo g k^{n} = lo g k$ . Rubric: full credit for identifying $d_{n}$ -closeness with central-block agreement and counting $k^{n}$ words.

Exercise 7 (hard, symbolic).

Let $A$ be an irreducible $0/1$ transition matrix and $X_{A} \subseteq {1, \dots, k}^{Z}$ the associated subshift of finite type. Prove that $h_{top} (X_{A}, σ) = lo g λ (A)$ , where $λ (A)$ is the Perron eigenvalue.

Hint

The number of admissible length- $n$ words equals $\sum_{i, j} (A^{n - 1})_{ij}$ ; use Perron-Frobenius to find its growth rate.

Answer

As in Exercise 5, the $(n, ε)$ -separated count for $ε = 2^{- 1}$ equals the number $W_{n}$ of admissible words of length $n$ in $X_{A}$ , since $d_{n}$ -closeness is agreement on the central length- $n$ block. An admissible word $i_{0} i_{1} \dots i_{n - 1}$ requires $A_{i_{0} i_{1}} = A_{i_{1} i_{2}} = \dots = 1$ , so $W_{n} = \sum_{i, j} (A^{n - 1})_{ij} = 1^{⊤} A^{n - 1} 1$ , the sum of all entries of $A^{n - 1}$ . Since $A$ is irreducible with non-negative integer entries, the Perron-Frobenius theorem 38.02.02 gives a simple positive eigenvalue $λ = λ (A)$ strictly dominating all others, with strictly positive left and right eigenvectors $u, v$ . Then $A^{n - 1} = λ^{n - 1} v u^{⊤} + O ((λ^{'})^{n - 1})$ with $∣ λ^{'} ∣ < λ$ , so $W_{n} = λ^{n - 1} (1^{⊤} v) (u^{⊤} 1) + O ((λ^{'})^{n - 1})$ , which grows like $C λ^{n}$ with $C > 0$ . Hence $h_{top} (X_{A}) = lim_{n} \frac{1}{n} lo g W_{n} = lo g λ (A)$ . Rubric: full credit for $W_{n} = 1^{⊤} A^{n - 1} 1$ , the Perron-Frobenius dominant-eigenvalue expansion, and the growth rate $lo g λ$ .

Exercise 8 (hard, symbolic).

Let $T_{A}$ be a hyperbolic toral automorphism of $T^{2}$ induced by $A \in SL_{2} (Z)$ with eigenvalues $λ, λ^{- 1}$ , $∣ λ ∣ > 1$ . Show $h_{top} (T_{A}) = lo g ∣ λ ∣ = \sum_{∣ λ_{i} ∣ > 1} lo g ∣ λ_{i} ∣$ .

Hint

A Bowen ball $B_{n} (x, ε)$ for $T_{A}$ is roughly a rectangle of unstable length $ε ∣ λ ∣^{- n}$ and stable length $ε$ ; count how many tile the torus.

Answer

Use coordinates along the stable eigendirection $E^{s}$ (contraction $∣ λ ∣^{- 1}$ ) and unstable eigendirection $E^{u}$ (expansion $∣ λ ∣$ ). For $y$ near $x$ , $T_{A}^{k}$ multiplies the $E^{u}$ -component of $y - x$ by $λ^{k}$ and the $E^{s}$ -component by $λ^{- k}$ . The Bowen condition $d_{n} (x, y) \leq ε$ forces $∣ E^{u} -component ∣ \leq ε ∣ λ ∣^{- (n - 1)}$ (the last iterate $k = n - 1$ stretches it most) and $∣ E^{s} -component ∣ \leq ε$ (the component $k = 0$ , the largest, since it only contracts). So $B_{n} (x, ε)$ is essentially a rectangle of dimensions $ε ∣ λ ∣^{- (n - 1)} \times ε$ , area $\sim ε^{2} ∣ λ ∣^{- n}$ . An $(n, ε)$ -spanning set needs about (torus area) $/$ (Bowen-ball area) $\sim ∣ λ ∣^{n} / ε^{2}$ points, and a separated set has the same order, so $s (n, ε), r (n, ε) ≍ ∣ λ ∣^{n}$ . Therefore $h_{top} (T_{A}) = lim_{ε} lim sup_{n} \frac{1}{n} lo g (∣ λ ∣^{n} / ε^{2}) = lo g ∣ λ ∣$ , which equals $\sum_{∣ λ_{i} ∣ > 1} lo g ∣ λ_{i} ∣$ since $λ$ is the only eigenvalue of modulus $> 1$ . Rubric: full credit for the Bowen-ball-as-rectangle estimate, the area count, and the conclusion $lo g ∣ λ ∣$ .

Advanced results Master

Theorem (basic functorial properties). Topological entropy is a topological-conjugacy invariant. It satisfies $h_{top} (f^{m}) = m h_{top} (f)$ for $m \geq 1$ , and $h_{top} (f^{- 1}) = h_{top} (f)$ when $f$ is a homeomorphism; for a product, $h_{top} (f \times g) = h_{top} (f) + h_{top} (g)$ ; and if $π : (X, f) \to (Y, g)$ is a factor map (a continuous surjection with $π f = g π$ ), then $h_{top} (g) \leq h_{top} (f)$ , with equality when $π$ is finite-to-one with uniformly bounded fibres. (See ^{[Walters Ch. 7]}, ^{[Bowen 1971]}.)

These are the structural laws that make entropy a usable invariant. The product law expresses that independent systems add their complexities; the power law that watching $m$ ticks at a time multiplies the per-tick rate; the factor inequality that a quotient cannot manufacture distinguishability the cover did not already see. Invariance under conjugacy is what licenses entropy as a coordinate-free fingerprint, and it is the property that originally let Adler, Konheim, and McAndrew separate maps that no prior topological invariant distinguished.

Theorem (model evaluations). The full shift on $k$ symbols has $h_{top} = lo g k$ . A subshift of finite type $X_{A}$ with irreducible transition matrix $A$ has $h_{top} = lo g λ (A)$ , the logarithm of the Perron eigenvalue 38.02.02. The expanding map $x \mapsto m x mod 1$ has $h_{top} = lo g m = lo g (de g)$ , and more generally a $C^{1}$ expanding map of a compact manifold has $h_{top} = lo g (topological degree)$ when the degree controls the inverse-branch count. A hyperbolic toral automorphism $T_{A}$ has $h_{top} = \sum_{∣ λ_{i} ∣ > 1} lo g ∣ λ_{i} ∣$ , the sum over eigenvalues of modulus exceeding one. (See ^{[Katok-Hasselblatt 1995 §3.1]}, ^{[Walters Ch. 7]}.)

Each evaluation reads the entropy off a counting problem the separated-set definition makes transparent: words for the shift, admissible paths for the SFT, inverse branches for the expanding map, Bowen rectangles for the automorphism. The toral case is the linear, constant-exponent prototype of the Pesin picture, where entropy equals integrated positive Lyapunov exponents; here the exponents are the constants $lo g ∣ λ_{i} ∣$ , and their sum over the expanding directions is the entire entropy.

Theorem (entropy and volume growth; Yomdin-Newhouse). For a $C^{\infty}$ self-map $f$ of a compact manifold, the topological entropy equals the exponential growth rate of the volume of iterated submanifolds: $$ h_{\mathrm{top}}(f) = \max_{0 \leq i \leq \dim M} \limsup_{n \to \infty} \frac{1}{n} \log \sup_{\sigma} \operatorname{vol}\big(f^n \circ \sigma\big), $$ the supremum over $C^{\infty}$ singular $i$ -cubes $σ$ of bounded volume. Newhouse proved $h_{top} (f) \leq$ the volume-growth rate for $C^{\infty}$ maps, and Yomdin the reverse, so the two coincide. (See ^{[Yomdin 1987]}.)

The Yomdin-Newhouse identity ties the combinatorial entropy to differential geometry: entropy is exactly how fast $f$ stretches the area of surfaces, the length of curves, or the volume of higher cells, whichever grows fastest. The $C^{\infty}$ hypothesis is necessary — Yomdin's argument controls the loss of smoothness through Newton-polytope estimates on the growth of derivatives, and the theorem fails for merely $C^{r}$ maps with a defect term of order $dim M / r$ . It is the bridge from the measure-free, topological counting of Bowen balls to the volume geometry that drives smooth ergodic theory.

Synthesis. Topological entropy is the foundational reason the topological and measurable theories of complexity meet: it is the measure-free growth rate of distinguishable orbit segments, computed equivalently by open covers or Bowen balls, and the central insight of this unit is that this single number governs both the combinatorics of the shift and the geometry of smooth maps. This is exactly the content of the variational principle, which states $h_{top} (f) = sup_{μ} h_{μ} (f)$ over invariant measures, so the topological invariant dominates every Kolmogorov-Sinai entropy of 38.06.02 and is attained by a measure of maximal entropy; putting these together with the Yomdin-Newhouse theorem, entropy is simultaneously a counting exponent, a supremum of information rates, and a volume-growth rate, the three faces of one invariant. The construction is dual to the measure-theoretic entropy of 38.06.02 exactly as the join of covers is dual to the join of partitions: where partitions carry the measure, covers carry only the topology, and the supremum over covers replaces the supremum over partitions. The bridge is the recognition that the doubling map's $lo g 2$ , the golden-mean shift's $lo g \frac{1 + 5}{2}$ , and the cat map's $lo g \frac{3 + 5}{2}$ are one quantity wearing three computational disguises — symbolic word-counting, Perron-Frobenius spectral radius, and Lyapunov-exponent sum — and this same hierarchy generalises into the thermodynamic formalism, where topological entropy is the pressure of the zero potential and the model evaluations become special cases of the variational principle for pressure.

Full proof set Master

Proposition 1 (entropy is a conjugacy invariant). If $ϕ : X \to Y$ is a homeomorphism of compact metric spaces with $ϕ \circ f = g \circ ϕ$ , then $h_{top} (f) = h_{top} (g)$ .

Proof. If $U$ is an open cover of $Y$ then $ϕ^{- 1} U$ is an open cover of $X$ , and $ϕ^{- 1} (g^{- k} U) = f^{- k} (ϕ^{- 1} U)$ by the intertwining, so $ϕ^{- 1} (⋁_{k < n} g^{- k} U) = ⋁_{k < n} f^{- k} (ϕ^{- 1} U)$ . A homeomorphism carries subcovers to subcovers bijectively, so $N (⋁_{k < n} g^{- k} U) = N (⋁_{k < n} f^{- k} (ϕ^{- 1} U))$ , whence $h (g, U) = h (f, ϕ^{- 1} U)$ . Since $ϕ^{- 1}$ is a bijection between the open covers of $Y$ and those of $X$ , taking suprema gives $h_{top} (g) = h_{top} (f)$ . $□$

Proposition 2 (the full shift on $k$ symbols). The shift $σ$ on $Σ_{k} = {0, \dots, k - 1}^{Z}$ has $h_{top} (σ) = lo g k$ .

Proof. With the metric $d (x, y) = 2^{- m i n {∣ j ∣ : x_{j} \neq = y_{j}}}$ and $ε = 2^{- 1}$ , $d_{n} (x, y) \leq ε$ holds iff $x$ and $y$ agree on coordinates $0, \dots, n - 1$ (each $σ^{j}$ , $0 \leq j < n$ , must leave the zeroth coordinates equal). The $k^{n}$ sequences supported on $[0, n)$ are pairwise $d_{n}$ -separated by more than $ε$ , so $s (n, ε) \geq k^{n}$ ; and any two agreeing on $[0, n)$ are within $d_{n}$ -distance $2^{- n} \leq ε$ , so one representative per length- $n$ central word is $(n, ε)$ -spanning, $r (n, ε) \leq k^{n}$ . Hence $s (n, ε) = r (n, ε) = k^{n}$ and $h_{top} (σ) = lim_{ε \to 0} lim sup_{n} \frac{1}{n} lo g k^{n} = lo g k$ . $□$

Proposition 3 (subshift of finite type). For an irreducible $0/1$ matrix $A$ , the SFT $X_{A}$ has $h_{top} (σ ∣_{X_{A}}) = lo g λ (A)$ .

Proof. By the same $ε = 2^{- 1}$ computation, $s (n, ε)$ equals the number $W_{n}$ of admissible length- $n$ words, and an admissible word $i_{0} \dots i_{n - 1}$ is one with $A_{i_{l} i_{l + 1}} = 1$ for all $l$ , so $W_{n} = 1^{⊤} A^{n - 1} 1$ . Irreducibility and non-negativity invoke Perron-Frobenius 38.02.02: $A$ has a simple dominant eigenvalue $λ = λ (A) > 0$ with strictly positive eigenvectors, and $A^{n - 1} = λ^{n - 1} v u^{⊤} + R_{n}$ with $∥ R_{n} ∥ = O ((λ^{'})^{n - 1})$ , $∣ λ^{'} ∣ < λ$ . Then $W_{n} = λ^{n - 1} (1^{⊤} v) (u^{⊤} 1) + O ((λ^{'})^{n - 1}) ≍ λ^{n}$ , so $\frac{1}{n} lo g W_{n} \to lo g λ$ and $h_{top} (X_{A}) = lo g λ (A)$ . $□$

Proposition 4 (the times- $m$ map). The expanding map $E_{m} (x) = m x mod 1$ on the circle has $h_{top} (E_{m}) = lo g m$ .

Proof. The Bowen ball $B_{n} (x, ε) = {y : ∣ E_{m}^{k} x - E_{m}^{k} y ∣ < ε for k < n}$ (circle distance) is an arc of length $ε m^{- (n - 1)}$ for $ε < \frac{1}{2 m}$ , since $E_{m}$ multiplies short-arc lengths by $m$ until wrap-around and the tightest constraint is the last iterate $k = n - 1$ . The circle of length $1$ is covered by $\sim m^{n - 1} / ε$ such arcs, so $r (n, ε) ≍ m^{n}$ ; a maximal separated set has the same order. Therefore $h_{top} (E_{m}) = lim_{ε \to 0} lim sup_{n} \frac{1}{n} lo g (m^{n} / ε) = lo g m$ . For $m = 2$ this recovers the doubling map's $lo g 2$ . $□$

Proposition 5 (the power law). For every $m \geq 1$ , $h_{top} (f^{m}) = m h_{top} (f)$ .

Proof. Let $U$ be an open cover and $V = ⋁_{k = 0}^{m - 1} f^{- k} U$ . Then $⋁_{j = 0}^{n - 1} (f^{m})^{- j} V = ⋁_{k = 0}^{nm - 1} f^{- k} U$ , so $h (f^{m}, V) = lim_{n} \frac{1}{n} lo g N (⋁_{k < nm} f^{- k} U) = m lim_{n} \frac{1}{nm} lo g N (\dots) = m h (f, U)$ . Taking the supremum over $U$ (covers of the form $V$ are cofinal, being finer than $U$ , and refining can only increase cover entropy up to the limit) gives $h_{top} (f^{m}) \geq m h_{top} (f)$ ; the reverse follows because every cover $W$ for $f^{m}$ satisfies $h (f^{m}, W) \leq h (f^{m}, W \lor f^{- 1} W \lor \dots) = m h (f, W^{'})$ for a refining $W^{'}$ . Hence $h_{top} (f^{m}) = m h_{top} (f)$ . $□$

Connections Master

Topological transitivity, mixing, and Devaney chaos 38.01.03. Entropy quantifies the complexity that the connectivity hierarchy of that unit only classifies qualitatively. The doubling map and full shift are exhibited there as the canonical mixing, chaotic systems, and here they acquire their numerical fingerprints $lo g 2$ and $lo g k$ ; positive topological entropy is a quantitative sufficient condition for the chaotic regime, sharper than transitivity, and the Bowen-metric distinguishability of this unit is the metric refinement of the sensitive dependence defined there.
Shifts of finite type, transition matrices, and coding 38.02.02. The SFT entropy $lo g λ (A)$ is computed from the Perron-Frobenius spectral radius of the transition matrix introduced there, so that unit supplies the linear algebra — irreducibility, the dominant positive eigenvalue, the count $W_{n} = 1^{⊤} A^{n - 1} 1$ — on which this unit's symbolic entropy computations rest. Coding a smooth hyperbolic system to an SFT transfers its entropy to a Perron-root computation, the standard route to evaluating $h_{top}$ in practice.
Kolmogorov-Sinai entropy and the generator theorem 38.06.02. Measure-theoretic entropy is the partition-based companion to the cover-based topological entropy here; the variational principle $h_{top} (f) = sup_{μ} h_{μ} (f)$ makes the topological invariant the supremum of the measure-theoretic ones over invariant measures, with the join of partitions playing exactly the role the join of covers plays in this unit. A measure of maximal entropy realises the supremum and ties the two definitions into one.
The Shannon-McMillan-Breiman theorem 38.06.03. The asymptotic equipartition property there counts typical measure-theoretic names at rate $e^{n h_{μ}}$ , the measure-side analogue of this unit's counting of $(n, ε)$ -separated orbit segments at rate $e^{n h_{top}}$ ; through the variational principle the topological separated-set count dominates every measure-theoretic typical-name count.
Hyperbolic sets, Anosov and Axiom-A systems, the Smale decomposition 38.03.01. The Smale spectral decomposition splits an Axiom-A non-wandering set into transitive basic sets, each coded by a subshift of finite type; the topological entropy of the system is then the maximum of the SFT entropies $lo g λ (A)$ of its pieces, and the hyperbolic toral automorphism's $\sum_{∣ λ_{i} ∣ > 1} lo g ∣ λ_{i} ∣$ is the constant-exponent prototype of the Pesin entropy formula for the smooth hyperbolic theory.

Historical & philosophical context Master

Topological entropy was introduced by Roy Adler, Alan Konheim, and M. H. McAndrew in their 1965 Transactions of the American Mathematical Society paper ^{[Adler 1965]}, who transported Kolmogorov and Sinai's measure-theoretic invariant of 1958-59 to the purely topological setting by replacing measurable partitions with open covers and the partition entropy $- \sum μ (P_{i}) lo g μ (P_{i})$ with the cover quantity $lo g N (U)$ . Their definition was a conjugacy invariant computable without reference to any invariant measure, and it immediately distinguished maps that the existing topological theory could not.

The metric reformulation, more convenient for computation, was given independently by Efim Dinaburg in 1970 ^{[Dinaburg 1970]} and Rufus Bowen in 1971 ^{[Bowen 1971]}, who replaced open covers by $(n, ε)$ -separated and spanning sets in the Bowen metric and proved the equivalence with the Adler-Konheim-McAndrew definition; Bowen's formulation became standard and his name attaches to the dynamical metric $d_{n}$ . The link to measure-theoretic entropy was completed by the variational principle, established in stages by Dinaburg, Goodman, and Goodwyn around 1969-71, asserting that topological entropy is the supremum of Kolmogorov-Sinai entropies over invariant measures. The geometric meaning of entropy as a volume-growth rate was proved by Sheldon Newhouse in 1988 (one inequality) and Yosef Yomdin in 1987 (the reverse) ^{[Yomdin 1987]}, whose semialgebraic estimates on the growth of derivatives identified the topological entropy of a $C^{\infty}$ map with the exponential growth rate of the volumes of iterated submanifolds.

Bibliography Master

@article{AdlerKonheimMcAndrew1965,
  author  = {Adler, Roy L. and Konheim, Alan G. and McAndrew, M. H.},
  title   = {Topological entropy},
  journal = {Transactions of the American Mathematical Society},
  volume  = {114},
  year    = {1965},
  pages   = {309--319}
}

@article{Bowen1971,
  author  = {Bowen, Rufus},
  title   = {Entropy for group endomorphisms and homogeneous spaces},
  journal = {Transactions of the American Mathematical Society},
  volume  = {153},
  year    = {1971},
  pages   = {401--414}
}

@article{Dinaburg1970,
  author  = {Dinaburg, Efim I.},
  title   = {The relation between topological entropy and metric entropy},
  journal = {Doklady Akademii Nauk SSSR},
  volume  = {190},
  year    = {1970},
  pages   = {19--22}
}

@article{Yomdin1987,
  author  = {Yomdin, Yosef},
  title   = {Volume growth and entropy},
  journal = {Israel Journal of Mathematics},
  volume  = {57},
  number  = {3},
  year    = {1987},
  pages   = {285--300}
}

@article{Newhouse1989,
  author  = {Newhouse, Sheldon E.},
  title   = {Continuity properties of entropy},
  journal = {Annals of Mathematics},
  volume  = {129},
  number  = {2},
  year    = {1989},
  pages   = {215--235}
}

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

@book{KatokHasselblatt1995,
  author    = {Katok, Anatole and Hasselblatt, Boris},
  title     = {Introduction to the Modern Theory of Dynamical Systems},
  publisher = {Cambridge University Press},
  series    = {Encyclopedia of Mathematics and its Applications},
  volume    = {54},
  year      = {1995}
}

Prerequisites

38.01.03
38.02.02

Tier anchors

beginner: Brin-Stuck 2002 *Introduction to Dynamical Systems* (Cambridge University Press) Ch. 2 (entropy as the growth rate of distinguishable orbits); Strogatz 1994 *Nonlinear Dynamics and Chaos* (Addison-Wesley) Ch. 10 (the doubling map, exponential proliferation of binary itineraries)
intermediate: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) §7.1-7.2 (open-cover entropy, the Bowen-Dinaburg metric definition, basic properties); Brin-Stuck 2002 *Introduction to Dynamical Systems* (Cambridge University Press) §2.5-2.6
master: Walters 1982 *An Introduction to Ergodic Theory* (Springer GTM 79) Ch. 7; Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge University Press) §3.1 and Suppl.; Bowen 1971 *Entropy for group endomorphisms and homogeneous spaces* (Trans. Amer. Math. Soc. 153); Yomdin 1987 / Newhouse 1989 (the volume-growth = entropy theorems)

References

Adler, Konheim, McAndrew 1965 — Topological entropy · Transactions of the American Mathematical Society 114 (1965) 309-319; the original open-cover definition h(f) = sup_U lim (1/n) log N(U vee f^{-1}U vee ... vee f^{-(n-1)}U)
Bowen 1971 — Entropy for group endomorphisms and homogeneous spaces · Transactions of the American Mathematical Society 153 (1971) 401-414; the (n,epsilon)-separated/spanning-set definition and the equivalence with the open-cover entropy on a compact metric space
Dinaburg 1970 — The relation between topological entropy and metric entropy · Doklady Akademii Nauk SSSR 190 (1970) 19-22; the metric (separated-set) characterisation of topological entropy
Walters 1982 — An Introduction to Ergodic Theory · Springer GTM 79, Ch. 7 (topological entropy: open covers, separated and spanning sets, properties, the variational principle)
Katok, Hasselblatt 1995 — Introduction to the Modern Theory of Dynamical Systems · Encyclopedia of Mathematics and its Applications 54, Cambridge University Press 1995, §3.1 (topological entropy, the Bowen metric, computations for the shift and expanding maps)
Yomdin 1987 / Newhouse 1989 — Volume growth and entropy · Y. Yomdin, Israel J. Math. 57 (1987) 285-300; S. Newhouse, Ann. of Math. 129 (1989) 215-235; topological entropy of a C^infty map equals its exponential volume-growth rate

Estimated time

beginner: 18m
intermediate: 52m
master: 90m