Chapter Summaries & Visual Program Specifications
A self-contained study companion: each chapter of the book condensed into its core ideas, with precise definitions of the central concepts, an explanation of why each concept matters to the subject, and the specification of the interactive programs that visualize it. Every program listed here is implemented in JavaScript and runnable in the guided tour. Best seen on Firefox.
Examples and Basic Concepts
Thirteen canonical examples that the rest of the book returns to again and again.
A dynamical system is simply a space X together with a self-map f : X → X (discrete time) or a flow φₜ (continuous time). The subject asks one question in many guises: what happens to points under repeated application of the map? The orbit of a point, O(x) = {x, f(x), f²(x), …}, is the fundamental object, and everything else — recurrence, entropy, ergodicity, chaos — is a way of describing how orbits behave collectively.
Rather than beginning with abstract theory, the book introduces a zoo of concrete examples that among them exhibit essentially every phenomenon studied later: circle rotations R_α(x) = x + α mod 1 (the simplest systems with nontrivial recurrence), expanding endomorphisms E_m(x) = mx mod 1 (the simplest chaotic systems), the full shift and subshifts (dynamics on sequence spaces), the quadratic family q_c(x) = x² + c (one-parameter route to chaos), the Gauss map G(x) = {1/x} (continued fractions as dynamics), hyperbolic toral automorphisms such as Arnold's cat map, Smale's horseshoe, the solenoid, and the strange attractors of Lorenz and Hénon. These examples are not decoration: nearly every theorem in the book is best understood as a statement about one of them.
Key concepts
- Orbit, fixed point, periodic point
- The orbit of x is O(x) = {fⁿ(x) : n ≥ 0}. A fixed point satisfies f(p) = p; a periodic point of period n satisfies fⁿ(p) = p with n minimal. A fixed point is attracting if |f′(p)| < 1, repelling if |f′(p)| > 1. Orbits are the atoms of the theory. Classifying which orbit types a system supports — and how they gain and lose stability — is the whole game of low-dimensional dynamics.
- Topological conjugacy
- Systems f : X → X and g : Y → Y are conjugate if there is a homeomorphism h : X → Y with h ∘ f = g ∘ h. Conjugate systems have identical dynamics up to a continuous change of coordinates. Conjugacy is the notion of "sameness" for dynamical systems. Invariants that survive conjugacy — number of periodic orbits, transitivity, entropy — are the properties that genuinely belong to the dynamics rather than to a coordinate choice.
- Hyperbolic toral automorphism
- An integer matrix A with det A = ±1 and no eigenvalues on the unit circle induces a map of the torus Tⁿ = ℝⁿ/ℤⁿ. The cat map A = [[2,1],[1,1]] has eigenvalues φ² ≈ 2.618 and φ⁻² ≈ 0.382 (φ the golden ratio): one direction stretches, the other contracts. The cleanest example of uniform hyperbolicity — every deep theorem of Chapters 5, 6 and 9 (shadowing, ergodicity, entropy formulas) can be seen concretely here first.
- Lyapunov exponent
- The exponential rate at which nearby orbits separate: λ(x) = lim (1/n) log |(fⁿ)′(x)|. Positive λ means errors of size ε grow like ε·e^{λn}. This is the quantitative meaning of "sensitive dependence on initial conditions." A positive Lyapunov exponent is the practical fingerprint of chaos, and it reappears in Chapter 9 tied to entropy via the Pesin-type identity.
Programs
Cobweb Diagram — Quadratic Family & Gauss Map
§1.5, §1.6Turns iteration into geometry: from x₀ draw vertically to the graph of f, horizontally to the diagonal y = x, and repeat. Convergence to fixed points, capture by periodic cycles, and the onset of chaos in the logistic family all become visually obvious. Includes the Gauss map as a preset.
How: plot y = f(x) and y = x; alternate vertical and horizontal segments from (x₀, 0), coloring segments by iteration index.
Arnold's Cat Map — Hyperbolic Toral Automorphism
§1.7, §1.8A cat face on a 64×64 pixel grid is scrambled by A = [[2,1],[1,1]] mod 1: recognizable for 2–3 steps, then apparently random — yet it returns exactly at step 48, because on a rational grid the map is a permutation of finitely many pixels. Stable and unstable eigendirections are overlaid.
How: all 48 frames are precomputed by applying A once per step as a pixel permutation; animation steps through them at 500 ms per frame.
Strange Attractors — Hénon & Lorenz
§1.12, §1.13The Lorenz butterfly and the Hénon attractor side by side, with two nearby initial conditions overlaid so their exponential divergence — and hence the impossibility of long-term prediction — is directly visible, together with a numerical estimate of the leading Lyapunov exponent.
How: Hénon by direct iteration of (x,y) ↦ (a − x² + by, x); Lorenz by RK4 integration; Lyapunov exponent by renormalized two-orbit divergence.
Topological Dynamics
What can be said about long-run behavior using continuity alone — no measure, no smoothness.
Topological dynamics studies a continuous map on a compact metric space and asks which properties of orbits are visible to topology alone. The central discovery is that a surprisingly rich classification is possible: one can say precisely where orbits accumulate (ω-limit sets), which points participate in recurrent behavior (nonwandering set), whether some orbit visits everywhere (transitivity), whether all regions eventually communicate (mixing), and — most importantly — how many distinguishable orbits the system has (topological entropy).
Entropy is the star of the chapter. It assigns to each system a single number h(f) ≥ 0 measuring the exponential growth rate of the number of orbit segments that can be told apart at finite resolution. It is invariant under topological conjugacy, which makes it the most useful single invariant in dynamics: if two systems have different entropies, no continuous change of coordinates can ever identify them.
Key concepts
- ω-limit set
- ω(x) is the set of accumulation points of the forward orbit: ω(x) = ⋂_{n≥0} closure{f^k(x) : k ≥ n}. It is closed, invariant, and non-empty when the space is compact. It separates transient behavior from permanent behavior. The attractor you "see" when you plot a long orbit is exactly its ω-limit set.
- Nonwandering set Ω(f)
- A point p is nonwandering if every neighborhood U ∋ p satisfies fⁿ(U) ∩ U ≠ ∅ for some n > 0. Ω(f) is closed, invariant, and contains all periodic points and all ω-limit sets. Ω(f) is where all the interesting dynamics lives; everything outside it is passed through once and never again. Smale's program for classifying dynamical systems is organized entirely around the structure of Ω.
- Topological transitivity & minimality
- Transitive: some orbit is dense (equivalently, any two open sets are connected by the dynamics: fⁿ(U) ∩ V ≠ ∅). Minimal: every orbit is dense. Irrational rotations are minimal; the cat map is transitive but not minimal (it has periodic points). Transitivity says the system cannot be decomposed into independent subsystems — it is one indivisible dynamical unit. It is one of the three ingredients of Devaney's definition of chaos.
- Topological mixing
- For all open U, V there is N with fⁿ(U) ∩ V ≠ ∅ for all n ≥ N. Strictly stronger than transitivity. Mixing is the topological shadow of stochastic behavior: any initial blob, no matter how small, eventually overlaps every region simultaneously and forever. The cat map's scrambling of the cat face is mixing made visible.
- Topological entropy h(f)
- Call points x, y (n,ε)-separated if their orbits differ by ≥ ε within the first n steps. With s(n,ε) the maximal cardinality of a separated set, h(f) = lim_{ε→0} limsup (1/n) log s(n,ε). Equivalently defined by minimal spanning sets. For E_m, h = log m; for the cat map, h = log φ² ≈ 0.962. Entropy counts distinguishable futures. h > 0 is the sharpest available dividing line between order and chaos, and its conjugacy-invariance makes it the primary tool for proving two systems are genuinely different.
- Expansiveness
- There is δ > 0 such that any two distinct orbits eventually separate by at least δ: if d(fⁿx, fⁿy) ≤ δ for all n, then x = y. Expansiveness means finite-resolution observations eventually determine the point exactly — the property that makes symbolic coding (Chapter 3) faithful.
Programs
Omega-Limit Sets & Non-Wandering Set
§2.1For 2-D flows (Van der Pol, Duffing) many initial conditions are integrated and their ω-limit sets plotted in color: transient spirals fall away and the limit cycles and equilibria — the nonwandering skeleton of the flow — stand out.
How: RK4 integration; the first half of each trajectory is discarded as transient, the remainder scatter-plotted per basin.
Topological Entropy via Spanning/Separating Sets
§2.5, §2.6Computes h(f_r) for the logistic family numerically by counting spanning sets, and plots the entropy curve directly beneath the bifurcation diagram — so one sees entropy switch on exactly where the diagram turns chaotic, reaching log 2 at r = 4.
How: for each r and orbit length n, a greedy set cover estimates the minimal (n,ε)-spanning cardinality; the slope of log S(n,ε) against n estimates h.
Symbolic Dynamics
Replace a chaotic map by a shift on sequences — and compute everything by linear algebra.
Symbolic dynamics is the strategy of trading geometry for combinatorics. Partition the phase space into finitely many pieces labeled 0, …, m−1 and record, for each point, the sequence of labels its orbit visits: its itinerary. The map f then acts on itineraries by simply shifting the sequence one place — and for hyperbolic systems this coding loses essentially nothing. Hard analytic questions (how many periodic orbits? what is the entropy?) become counting problems about sequences.
The central objects are subshifts of finite type, where admissible sequences are those whose consecutive symbols are allowed by a 0-1 transition matrix A. Their theory is completely governed by the Perron-Frobenius theorem: the largest eigenvalue of A determines the entropy, its eigenvector determines the natural invariant measure, and traces of powers of A count periodic orbits exactly.
Key concepts
- Full shift Σ_m
- The space of all (one- or two-sided) sequences over an alphabet of m symbols, with the shift map σ(x)ᵢ = xᵢ₊₁. Compact, and h(σ) = log m. The universal chaotic system: it contains periodic orbits of every period, dense orbits, and maximal entropy. Every expansive system embeds in some full shift.
- Subshift of finite type (SFT)
- Given a 0-1 matrix A, keep only sequences with A[xᵢ, xᵢ₊₁] = 1 for all i — i.e. bi-infinite walks in a directed graph. SFTs are the discrete models of hyperbolic dynamics: Markov partitions (Chapter 5) conjugate Anosov systems to SFTs, transporting every question about smooth chaos into finite linear algebra.
- Perron-Frobenius theorem
- An irreducible non-negative matrix A has a simple positive eigenvalue λ_max = ρ(A) strictly dominating all others in modulus, with strictly positive eigenvector. For an SFT, h(σ_A) = log ρ(A) and the number of period-n points is Tr(Aⁿ) ≈ ρ(A)ⁿ. One theorem, three payoffs: entropy, orbit counts, and the stationary measure all read off from a single eigenvalue problem. It also underlies Google's PageRank (Chapter 4).
- Zeta function
- ζ_A(z) = exp(Σ Tr(Aⁿ) zⁿ/n) = 1/det(I − zA) packages all periodic orbit counts into one rational function. A dynamical analogue of the Riemann zeta function: its poles encode the growth of periodic orbits, connecting dynamics to number theory and spectral theory.
- Substitution systems
- A rule replacing each letter by a word — Fibonacci: a → ab, b → a; Thue-Morse: a → ab, b → ba — iterated to an infinite self-similar sequence, whose orbit closure under the shift is a minimal, zero-entropy system. These systems are ordered but never periodic — the mathematics of quasicrystals. They show that zero entropy does not mean simple: Thue-Morse has purely singular continuous spectrum, a genuinely exotic middle ground between periodic and chaotic.
Programs
Subshift of Finite Type — Transition Graph & Perron-Frobenius
§3.2–§3.4Edit a transition matrix and watch everything update live: the directed graph, the spectral radius (hence entropy h = log ρ(A)), the Perron eigenvector as stationary probabilities, and the count of periodic orbits growing as ρ(A)ⁿ on a log-scale plot.
How: power iteration for the leading eigenpair; Tr(Aⁿ) by repeated matrix multiplication.
Substitution Systems — Fibonacci & Thue-Morse
§3.6Iterates substitution rules to thousands of letters, displays the sequence as a color strip, and computes its Fourier power spectrum: sharp golden-ratio-spaced peaks for Fibonacci (quasicrystal order), a fractal singular-continuous spectrum for Thue-Morse.
How: hand-written iterative FFT on the ±1 sequence; letter frequencies checked against the Perron eigenvector of the substitution matrix.
Ergodic Theory
The statistical viewpoint: what a typical orbit does, on average, forever.
Ergodic theory adds a measure to the picture and asks statistical questions. If a transformation preserves a probability measure — the map neither creates nor destroys "volume" — then astonishingly strong conclusions follow about the long-run behavior of almost every orbit. The two pillars are Poincaré recurrence (orbits return, always) and the Birkhoff ergodic theorem (time averages exist and, for ergodic systems, equal space averages). This is the rigorous foundation beneath statistical mechanics: it is the precise sense in which a single trajectory of a gas explores its entire energy surface.
The chapter builds the full vocabulary — invariant measures, ergodicity, unique ergodicity, weak and strong mixing, discrete spectrum — and closes with two spectacular applications: Furstenberg's ergodic-theoretic proof of Szemerédi's theorem on arithmetic progressions, and Google's PageRank, which is nothing other than the invariant measure of a random walk on the web graph.
Key concepts
- Invariant measure
- μ is T-invariant if μ(T⁻¹A) = μ(A) for all measurable A. Krylov-Bogolyubov: every continuous map of a compact metric space has at least one. An invariant measure is a self-consistent notion of "how often the orbit visits each region." It is the object that turns a deterministic map into a stationary stochastic process.
- Poincaré recurrence & Kac's lemma
- If μ(A) > 0, then μ-almost every point of A returns to A infinitely often. Kac: for ergodic systems the mean return time is exactly 1/μ(A). Recurrence from nothing but measure preservation — no dynamics assumptions at all. It resolves (and sharpens) Zermelo's objection to Boltzmann: recurrence is certain, but Kac shows the waiting time for a rare configuration is astronomically long.
- Ergodicity
- Every invariant set has measure 0 or 1 — the system cannot be split into two nontrivial independent parts. Ergodicity is exactly the condition under which statistics of one typical orbit reveal the whole system: time average = space average. It converts "simulate one long trajectory" into a legitimate method for computing equilibrium quantities.
- Birkhoff ergodic theorem
- For T measure-preserving and φ ∈ L¹, the time averages (1/N) Σ φ(Tᵏx) converge for a.e. x; if T is ergodic the limit is ∫φ dμ, independent of the starting point. The law of large numbers for deterministic systems, and arguably the most important single theorem in the book — Chapters 6 and 9 are both built on top of it.
- Unique ergodicity & Weyl equidistribution
- A system with exactly one invariant measure. Then Birkhoff averages converge for every point (not just a.e.). Weyl: for irrational α the sequence {nα mod 1} equidistributes in [0,1]. The strongest possible statistical regularity — and the dynamical engine behind classical number-theoretic facts, such as the distribution of leading digits of 2ⁿ (Benford's law).
- Mixing
- μ(T⁻ⁿA ∩ B) → μ(A)μ(B): any two events become asymptotically independent. Mixing ⟹ ergodic, but not conversely (rotations are ergodic, never mixing). Mixing is genuine randomness emerging from determinism — correlations with the initial condition decay to zero. It is why chaotic systems are statistically indistinguishable from noise.
- Gauss measure
- The measure dμ = dx / ((1+x) log 2) is invariant for the Gauss map G(x) = {1/x}, whose iteration generates continued-fraction digits. Ergodic theory applied to number theory: Birkhoff's theorem then yields the Gauss-Kuzmin statistics — e.g. the digit 1 appears in a typical continued fraction about 41.5% of the time — facts with no known non-dynamical proof of comparable simplicity.
Programs
Poincaré Recurrence — Return-Time Distributions
§4.2Picks a small interval A of measure ε and histograms first-return times for the irrational rotation and for the doubling map. Both obey Kac's mean 1/ε, but with opposite shapes: geometric spread for the mixing doubling map, tight concentration for the rigid rotation.
How: iterate thousands of starting points in A until first return; histogram against the Kac prediction line.
Birkhoff Ergodic Theorem — Time vs Space Averages
§4.3, §4.5, §4.7Running averages (1/N) Σ φ(Rⁿ_α x) for irrational (ergodic) versus rational (non-ergodic) rotations: the ergodic curves collapse onto the space average ∫φ from every start, the rational ones converge to start-dependent limits. Side panel shows Weyl equidistribution building up point by point.
How: direct orbit computation with a live running-average plot for several observables and starting points.
PageRank as Ergodic Theory
§4.12, §4.6Builds a random web graph, forms the Google matrix G = d·A + (1−d)/n · 𝟙, and power-iterates to the stationary distribution — the PageRank. Node sizes grow to their final ranks as the iteration converges geometrically at rate d.
How: column-stochastic matrix from a random digraph; power iteration with residual plot; spring-layout graph drawing.
Hyperbolic Dynamics
Stretching in some directions, contracting in others — the geometric mechanism of chaos.
Hyperbolicity is the property that makes chaos tractable. A set is hyperbolic when the tangent space along every orbit splits into two invariant families of directions, TM = Eˢ ⊕ Eᵘ, with vectors in Eˢ contracted and vectors in Eᵘ expanded at uniform exponential rates. This single geometric hypothesis yields, one after another, the deepest theorems in the book: pseudo-orbits are shadowed by true orbits, stable and unstable manifolds exist and are smooth, periodic orbits are dense, the dynamics is conjugate to a subshift of finite type via Markov partitions, and the whole picture is structurally stable — unchanged, up to conjugacy, by small perturbations.
The philosophical payoff is large: hyperbolic chaos is individually unpredictable but collectively rigid. You cannot forecast one orbit, yet the system as a whole is so robust that even its perturbations are copies of itself.
Key concepts
- Hyperbolic set
- A compact invariant set Λ with a continuous invariant splitting T_xM = Eˢ(x) ⊕ Eᵘ(x) and constants C, λ < 1 such that ‖Dfⁿv‖ ≤ Cλⁿ‖v‖ on Eˢ and symmetrically backward on Eᵘ. The uniform rates are what turn qualitative chaos into quantitative theorems — every result in this chapter fails without them.
- Shadowing lemma
- Near a hyperbolic set, for every ε there is δ such that every δ-pseudo-orbit (a sequence with d(f(xₙ), xₙ₊₁) ≤ δ) lies within ε of a genuine orbit. For the cat map the shadowing constant is C = 1/(λ⁺−1) + 1/(1−λ⁻) = √5. This is why numerical simulation of chaos is trustworthy at all: your computed orbit, corrupted by round-off at every step, is provably close to some true orbit of the system. Errors do not compound — the expanding and contracting directions each absorb their share.
- Stable and unstable manifolds (Hadamard-Perron)
- Wˢ(p) = {x : d(fⁿx, fⁿp) → 0} and Wᵘ(p) = {x : d(f⁻ⁿx, f⁻ⁿp) → 0} are injectively immersed smooth submanifolds tangent to Eˢ, Eᵘ. These are the geometric skeleton of chaos. For the Hénon map, the strange attractor is the closure of an unstable manifold; and a transverse intersection of Wˢ with Wᵘ (a homoclinic point) forces a horseshoe — Poincaré's original glimpse of chaos.
- Smale horseshoe
- A map that stretches a square, folds it, and lays it back across itself. The invariant set is a Cantor set on which the dynamics is conjugate to the full 2-shift. The horseshoe is chaos in its purest, fully-solved form — and it is universal: every transverse homoclinic point contains one. Chaos is not exotic; it is generic near homoclinic behavior.
- Anosov diffeomorphism & structural stability
- A diffeomorphism hyperbolic on its entire manifold (the cat map is the standard example). Anosov systems are structurally stable: any C¹-small perturbation is topologically conjugate to the original. Structural stability is the strongest possible robustness — the qualitative dynamics survives any small modeling error. This is the mathematical justification for trusting idealized models of chaotic phenomena.
- Markov partitions
- A partition of a hyperbolic set into "rectangles" aligned with the stable/unstable directions, whose itinerary coding conjugates the map to a subshift of finite type. The bridge between Chapters 3 and 5: smooth hyperbolic chaos is combinatorics in disguise. All the counting machinery of symbolic dynamics — entropy, zeta functions, orbit counts — transfers to smooth systems through this construction.
Programs
Shadowing Lemma — Pseudo-orbit vs True Orbit
§5.3Three columns at δ = 0.001, 0.01, 0.1: a noisy pseudo-orbit of the cat map (dashed) with the true orbit shadowing it (solid), plus error plots against the theoretical bound C·δ with C = √5. Small noise is invisibly absorbed; large noise visibly escapes the tube.
How: exact iteration plus per-step uniform noise; toroidal distance tracks the shadowing error per step.
Stable & Unstable Manifolds — Hénon Fixed Point
§5.6, §5.8Computes both manifolds of the Hénon saddle point: Wᵘ (red) by forward-iterating a tiny segment along the unstable eigendirection — folding into the familiar strange attractor — and Wˢ (cyan) via the exact inverse map. Their crossings are homoclinic points, each one concealing a horseshoe.
How: closed-form fixed point and eigenvectors; forward iteration for Wᵘ, iteration of f⁻¹(X,Y) = (Y, (X − a + Y²)/b) for Wˢ.
Ergodicity of Anosov Diffeomorphisms
One landmark theorem, proved in full: volume-preserving Anosov systems are ergodic.
This chapter is a case study in how geometry and measure theory combine. The theorem — every C² volume-preserving Anosov diffeomorphism is ergodic — is proved by Hopf's argument, one of the most beautiful proof strategies in dynamics. By Birkhoff's theorem, a forward time average is constant along stable manifolds (points on the same stable leaf have asymptotic futures); a backward time average is constant along unstable manifolds. Since forward and backward averages agree almost everywhere, an invariant function must be constant along both foliations — and since Eˢ and Eᵘ together span every tangent space, it must be constant almost everywhere. Hence ergodicity.
The subtlety — and the chapter's real technical content — is that the stable and unstable foliations are generally only Hölder continuous, not smooth, so the final "Fubini" step requires proving the foliations are absolutely continuous: they do not concentrate sets of positive volume onto sets of zero volume. This is the key lemma, and its failure in non-uniformly hyperbolic settings ("Fubini's nightmare") is a celebrated pathology of modern dynamics.
Key concepts
- Hopf argument
- Invariant L² functions are a.e. constant along stable leaves (forward averages) and along unstable leaves (backward averages); absolute continuity plus a Fubini-type argument forces such functions to be a.e. constant, which is precisely ergodicity. A purely geometric mechanism for a measure-theoretic conclusion — the template extended by Anosov, Sinai and, decades later, by the Pugh-Shub program on partial hyperbolicity ("ergodicity is stable").
- Stable/unstable foliations
- The families {Wˢ(x)} and {Wᵘ(x)} partition the manifold into smooth leaves that fit together only Hölder-continuously. Individually perfect, collectively rough: this tension is where the real difficulty of smooth ergodic theory lives.
- Absolute continuity of foliations
- Holonomy maps (sliding along leaves between transversals) send zero-volume sets to zero-volume sets, with Jacobians controlled by the dynamics. The precise hypothesis that legitimizes integrating "leaf by leaf." It is the technical heart of the ergodicity proof, and the property whose failure produces measure-theoretic pathologies in weaker settings.
Programs
Ergodicity of the Cat Map — Averages Along Foliations
§6.3Hopf's argument made visible: heatmaps of the Birkhoff averages A_N⁺φ on a 64×64 grid for N = 1, 10, 50, and beyond. The initial pattern of φ progressively flattens to the uniform gray of ∫φ dμ — ergodicity happening on screen — alongside a plot of the variance along stable leaves decaying to zero.
How: per-cell orbit averaging over the grid; diverging colormap with scale fixed at N = 1 so the flattening is honest.
Low-Dimensional Dynamics
In one dimension, chaos can be completely classified — and it is universal.
Dimension one is special: orbits cannot cross each other, and this single topological constraint produces a complete and stunningly detailed theory. For circle homeomorphisms, a single number — the rotation number ρ(f) — governs everything: rational ρ means periodic orbits, irrational ρ means the map is (semi-)conjugate to the rigid rotation R_ρ (Poincaré's classification, sharpened by Denjoy's theorem for C² maps).
For interval maps, the surprises are Sharkovsky's theorem — the existence of a single period-3 orbit forces periodic orbits of every period, via a strange total ordering of the natural numbers — and Feigenbaum's discovery that the period-doubling route to chaos is quantitatively universal: the constant δ ≈ 4.669 governs the bifurcation cascade of every unimodal map, from the logistic equation to dripping faucets and Rayleigh-Bénard convection experiments. Universality means the fine structure of the transition to chaos does not depend on the equations — a renormalization phenomenon directly analogous to critical exponents in statistical physics.
Key concepts
- Rotation number
- For a lift F of an orientation-preserving circle homeomorphism, ρ(f) = lim Fⁿ(x)/n mod 1 — the limit exists and is independent of x. Rational ρ = p/q ⟺ f has a period-q orbit. A complete invariant computed from any single orbit: one number decides the entire qualitative behavior. The prototype of what a classification theorem should look like.
- Poincaré classification & Denjoy's theorem
- Irrational ρ: f is semi-conjugate to R_ρ; if f is C² (Denjoy), fully conjugate. Denjoy also built C¹ counterexamples with wandering intervals. A sharp smoothness threshold: at C² the dynamics is rigid, at C¹ pathologies appear. One of the first results showing that regularity hypotheses in dynamics are not technicalities but genuine dividing lines.
- Arnold tongues & the devil's staircase
- In the standard family f(θ) = θ + Ω − (K/2π) sin 2πθ, the parameter regions where ρ locks to each rational p/q form wedges ("tongues") of width ~K^q; at K = 1 the graph of ρ(Ω) is a continuous function that is constant on a full-measure union of intervals. Mode-locking is ubiquitous in nature — coupled oscillators, heartbeats, the 3:2 resonance of Mercury. The tongues explain why coupled systems prefer rational frequency ratios.
- Sharkovsky's theorem
- Order the naturals: 3 ≻ 5 ≻ 7 ≻ … ≻ 2·3 ≻ 2·5 ≻ … ≻ 2² ≻ 2 ≻ 1. If a continuous interval map has a periodic point of period n, it has one of every period m with n ≻ m. In particular, period 3 forces all periods. A complete combinatorial law of period coexistence from continuity alone — no smoothness, no hyperbolicity. "Period three implies chaos" (Li-Yorke) is its most famous corollary.
- Feigenbaum universality
- Period-doubling bifurcation parameters of any unimodal family accumulate geometrically: (rₙ − rₙ₋₁)/(rₙ₊₁ − rₙ) → δ = 4.6692…, with a companion spatial scaling α ≈ 2.5029. Explained by a fixed point of the renormalization (doubling) operator with a single unstable eigenvalue δ. The measured route to chaos is the same in equations and in experiments because it is a property of renormalization, not of any particular system — dynamics' analogue of universality classes in phase transitions.
- Schwarzian derivative
- Sf = f‴/f′ − (3/2)(f″/f′)². Negative Schwarzian is preserved under composition and limits each attracting cycle to attract a critical point. The technical device that makes the unimodal theory clean: a map with one critical point and Sf < 0 has at most one attracting periodic orbit — so the bifurcation diagram shows the attractor, not one of many.
Programs
Rotation Number — Arnold Tongues
§7.1, §7.2A 200×200 heatmap of ρ(K, Ω) over the parameter square: rational tongues in fixed colors fan out from the Ω-axis and widen with K. A side strip shows the devil's-staircase graph of ρ(Ω) at the currently selected K.
How: the lift is iterated ~1000 steps per pixel (chunked rendering); ρ is classified against Farey fractions with q ≤ 8.
Sharkovsky's Theorem — Period Existence Cascade
§7.3A heat-strip over the tent-map family showing exactly which periods exist at each parameter: periods switch on in reverse Sharkovsky order as the parameter grows, and the moment period 3 appears, every row below it is already lit. A live cobweb shows the actual orbits.
How: period-n point counts via sign changes of fⁿ(x) − x, Möbius-corrected to count genuine least periods.
Bifurcation Diagram & Feigenbaum Universality
§7.6–§7.8A high-resolution, zoomable bifurcation diagram of the logistic family with the period-doubling parameters marked and a live table of ratios δₙ converging to 4.6692. Drag-zoom into the period-3 window at r ≈ 3.83 to find the whole cascade repeated in miniature.
How: density-accumulated ImageData rendering; per-column transient discard; animated r-sweep with a synchronized orbit strip.
Complex Dynamics
Iterating rational maps on the Riemann sphere — where chaos meets complex analysis and becomes fractal geometry.
Iterating a rational map R : ℂ̄ → ℂ̄ splits the sphere into two invariant sets with opposite characters: the Fatou set, where nearby initial conditions stay together (stable, predictable dynamics), and its complement the Julia set J(R), where the dynamics is chaotic — sensitive, transitive, with repelling periodic points dense. The rigidity of complex analysis (one derivative controls everything; Montel's theorem on normal families is the workhorse) makes this theory both more complete and more visually spectacular than its real counterpart: Julia sets are the canonical mathematical fractals.
For the quadratic family f_c(z) = z² + c, all questions funnel through the orbit of the single critical point 0, and the Mandelbrot set M = {c : the orbit of 0 stays bounded} becomes the master map of parameter space: c ∈ M exactly when J(f_c) is connected, and c ∉ M exactly when it is a Cantor dust. One picture in parameter space indexes the topology of infinitely many pictures in dynamical space.
Key concepts
- Fatou set / Julia set
- F(R): the largest open set on which the iterates {Rⁿ} form a normal (equicontinuous) family; J(R) its complement. J is closed, perfect, fully invariant, and is the closure of the repelling periodic points. A complete, canonical decomposition into "order" and "chaos" — sharper than anything available for real systems. All the fractal intricacy of complex dynamics lives exactly on J.
- Montel's theorem
- A family of holomorphic functions omitting three points of the sphere is normal. The engine of every proof in the chapter: wherever orbits avoid three values, dynamics is tame — so chaos is cornered onto the Julia set. A textbook example of one powerful analytic theorem organizing an entire dynamical theory.
- Mandelbrot set & the dichotomy
- M = {c ∈ ℂ : sup_n |f_cⁿ(0)| < ∞}. Fundamental dichotomy: J(f_c) is connected if c ∈ M, and totally disconnected (Cantor) if c ∉ M — because everything hinges on whether the critical orbit escapes. M is the parameter-space atlas of an entire family: its cardioid, bulbs and filaments encode which attracting cycles exist for which c. Douady-Hubbard proved M is connected; whether it is locally connected (MLC) remains the central open problem of the field.
- Critical orbits
- Every attracting cycle of a rational map attracts at least one critical point; a degree-d map has 2d − 2 of them, so finitely many critical orbits govern all stable behavior. This finiteness is why one-parameter pictures like M are possible at all: the quadratic family has a single critical point, so a single orbit decides the fate of the whole map.
Programs
Newton's Method Basins — Rational Map on ℂ̄
§8.4, §8.5Colors each starting point of the plane by which root of p(z) Newton's method finds, with brightness by convergence speed. The basin boundaries — the Julia set of the Newton map — are fractal: near them, arbitrarily small changes in the starting guess change which root you reach. The cubic z³ − 2z + 2 preset exhibits black regions attracted to a 2-cycle: Newton's method failing on an open set of starting points.
How: per-pixel Newton iteration with root classification, chunked strip rendering, zoom and pan.
Julia Set & Mandelbrot Set — z² + c
§8.5, §8.6Dual panels: click any parameter c in the Mandelbrot set on the left, and the corresponding Julia set renders on the right — connected Jordan-curve-like sets inside the cardioid, "rabbits" in the bulbs, Cantor dust outside M. Deep zoom reveals miniature copies of M along the boundary.
How: escape-time iteration with smooth (n − log log|z| / log 2) coloring; chunked rendering with progress bars; scroll-zoom and drag-pan on both panels.
Measure-Theoretic Entropy
Entropy as information: how many bits per step a system generates — and the principle unifying both entropies.
The Kolmogorov-Sinai entropy h_μ(T) measures the average information produced per iteration: observe the system through a finite partition ξ (a coarse-grained measurement), and ask how much new information each successive observation delivers. For an isometry like a rotation, the answer is zero — the future is determined, no information is ever generated. For the doubling map, each step reveals exactly one new binary digit: one bit per iteration, h = log 2.
Three theorems structure the chapter. The Kolmogorov-Sinai theorem says a single generating partition suffices to compute the entropy (no supremum needed). The Shannon-McMillan-Breiman theorem localizes entropy at almost every point: the measure of the n-step cell containing x decays like e^{−n h_μ} — entropy is the exponential rate at which observations pin down the initial condition. And the variational principle closes the book's grand arc: h_top(f) = sup_μ h_μ(f) — topological entropy, defined with no measure at all in Chapter 2, is exactly the maximum information rate over every stationary way of observing the system. Topological and measurable dynamics are two views of one subject.
Key concepts
- Entropy of a partition
- H(ξ) = −Σ μ(C) log μ(C) — the Shannon information of the measurement "which cell of ξ does x lie in?" The refinement ξⁿ = ξ ∨ T⁻¹ξ ∨ … ∨ T^{−(n−1)}ξ records n consecutive observations. Imports Shannon's information theory into dynamics wholesale: a deterministic map plus a partition is a stationary information source, and dynamics can ask how many bits it emits.
- Kolmogorov-Sinai entropy
- h_μ(T, ξ) = lim H(ξⁿ)/n (the limit exists by subadditivity), and h_μ(T) = sup_ξ h_μ(T, ξ). KS theorem: if ξ is a generator (its refinements separate points), the supremum is already attained at ξ. Invented to answer a concrete question — are the 2-shift and 3-shift isomorphic? Entropy says no (log 2 ≠ log 3), the first invariant ever able to distinguish them, and the founding success of the isomorphism theory of measurable dynamics.
- Shannon-McMillan-Breiman theorem
- For ergodic T and a generating partition, −(1/n) log μ(ξⁿ(x)) → h_μ(T) for a.e. x, where ξⁿ(x) is the n-step cell containing x. Entropy made pointwise: typical orbit segments of length n all have probability ≈ e^{−nh}, so there are effectively e^{nh} of them. This "asymptotic equipartition" is simultaneously the foundation of data compression and the reason entropy is computable from one long simulated orbit.
- Variational principle
- h_top(f) = sup { h_μ(f) : μ an f-invariant probability measure }. A measure attaining the supremum is a measure of maximal entropy — e.g. for the full tent map / logistic at r = 4 it is the arcsine distribution dx / (π√(x(1−x))). The unification of the book's two halves: the purely topological orbit count of Chapter 2 equals the best possible information rate over all statistics. It is also the gateway to the thermodynamic formalism — pressure, equilibrium states, SRB measures — that powers modern smooth ergodic theory.
- Ornstein's theorem (stated)
- Two Bernoulli shifts are measure-isomorphic if and only if they have equal entropy. For the most random systems, one number is a complete invariant — the endpoint of the classification program that entropy began.
Programs
Measure-Theoretic Entropy — Partition Refinement
§9.1, §9.3, §9.4Plots H(ξⁿ) against n for the doubling map (exactly linear, slope log 2 — one bit per step) and the irrational rotation (sublinear, slope → 0 — zero entropy), alongside a drawing of the actual partition atoms refining as n grows: exponentially many for doubling, only ~2n arcs for the rotation (three-distance theorem).
How: exact atom computation for both maps; Shannon entropy of the atom measures; adjustable rotation angle α.
Variational Principle — h_top vs h_μ
§9.5For the logistic family, computes topological entropy (by lap counting of fⁿ) and the metric entropy of the natural SRB measure (via Shannon-McMillan-Breiman word frequencies) and plots both against r. At r = 4 they meet at log 2 — where the SRB measure happens to be the measure of maximal entropy; for r < 4, h_μ < h_top, exactly as the variational principle requires.
How: lap counting for h_top; long-orbit itinerary word frequencies for the SMB estimate; incremental computation with a progress bar.