38.03.02 · dynamics / hyperbolicity-structural-stability

The Smale Horseshoe and the Smale-Birkhoff Homoclinic Theorem

shipped3 tiersLean: none

Anchor (Master): Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Ch. 2 §2.5 (the horseshoe), Ch. 6 §6.5 (the inclination/λ-lemma), Ch. 6 §6.6 (transverse homoclinic points and the Smale-Birkhoff theorem); Smale 1965 *Diffeomorphisms with many periodic points* (in *Differential and Combinatorial Topology*, Princeton); Smale 1967 *Differentiable dynamical systems* (Bull. Amer. Math. Soc. 73); Guckenheimer-Holmes 1983 *Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields* (Springer) Ch. 4–5 (the homoclinic tangle and Melnikov's method)

Intuition Beginner

Take a tall rubber rectangle. Stretch it hard in the horizontal direction so it becomes long and thin, squeeze it in the vertical direction, then bend the long strip into a horseshoe shape and lay it back down across the original rectangle. The bent strip crosses the rectangle in two separate vertical bands. That single move — stretch, fold, lay back — is the Smale horseshoe, and repeating it forever is one of the simplest machines that produces genuine chaos.

Why care about a folded rubber sheet? Because almost everything chaotic that appears in real systems — a forced pendulum, a buckled beam, a planet tugged by two others — hides a horseshoe inside it. The horseshoe is the universal skeleton of chaos: wherever you find one, you find infinitely many repeating motions, motions that never repeat, and motions so sensitive that the tiniest nudge sends them somewhere completely different. The remarkable part is that all of this survives if you bump the rule a little. The chaos is not a fluke of the exact formula; it is sturdy.

The deepest fact is a bookkeeping miracle. If you track which of the two bands a point lands in at every step — forward and backward in time — you get an endless string of two symbols, like a coin-flip record that runs in both directions. Every such string is achievable, and the string completely names the point. So the tangled geometry of the fold becomes the plainest object imaginable: all possible sequences of two letters. That translation is what makes the chaos understandable rather than merely complicated.

Visual Beginner

Picture a square. Inside it, mark two thin vertical strips, a left one and a right one. The horseshoe rule takes the whole square, stretches it sideways and squashes it down, folds it into a horseshoe, and drops it back so that the two arms of the horseshoe lie exactly over those two vertical strips. A point that survives one step had to start in one of the two strips; a point that survives forever, forward and backward, had to thread the needle at every single step.

A small table fixes the vocabulary.

name	what it is	everyday image
the two strips	the left/right bands $0$ and $1$	the two arms of the horseshoe
the surviving set	points that stay in the square forever	a fine dust of points (a Cantor set)
itinerary	the record of which strip at each step	a two-way coin-flip log
homoclinic point	a point that came from a rest spot and returns to it	a loop that leaves home and crawls back

Worked example Beginner

Let us name surviving points by their itineraries and count the repeating ones. A point's itinerary is a doubly infinite string of $0$ s and $1$ s: the $n$ -th letter is $0$ if the $n$ -th step lands the point in the left strip, $1$ if the right strip. Reading negative steps tells you where the point came from; reading positive steps tells you where it goes.

Step 1. A point repeats with period $1$ (it is a fixed point) exactly when its itinerary is constant: all $0$ s, or all $1$ s. That is $2$ points.

Step 2. A point repeats with period dividing $2$ exactly when its itinerary repeats every $2$ letters: the repeating block is one of $00$ , $01$ , $10$ , $11$ . That is $4$ itineraries, i.e. $4$ points whose period divides $2$ .

Step 3. In general, a point repeats with period dividing $n$ exactly when its itinerary is built by repeating a block of length $n$ . There are $2$ choices for each of the $n$ slots, so $2^{n}$ such blocks, hence $2^{n}$ points.

Step 4. Plug in numbers: period dividing $3$ gives $2^{3} = 8$ points; period dividing $4$ gives $2^{4} = 16$ ; period dividing $10$ gives $2^{10} = 1024$ . The count explodes geometrically, doubling each time you allow one more step in the block.

What this tells us: because every string of two symbols is a legal itinerary, counting repeating motions of the horseshoe is the same as counting repeating coin-flip patterns — and there are $2^{n}$ of those for blocks of length $n$ . The folding has manufactured infinitely many periodic orbits, growing at the rate $2^{n}$ , which is the fingerprint of an entropy equal to the logarithm of $2$ .

Check your understanding Beginner

Exercise (easy, multiple choice).

What three operations, in order, describe one step of the Smale horseshoe map?

A. Rotate, reflect, translate. B. Stretch in one direction, squeeze in the other, then fold and lay back across the square. C. Shrink uniformly toward a single point. D. Swap the two halves of the square.

Hint

Think of pulling taffy: long and thin in one direction, then bent over itself.

Answer

B. The horseshoe stretches the square along one direction (uniform expansion), squeezes it along the perpendicular direction (uniform contraction), then folds the long thin strip into a horseshoe and lays it back across the original square, crossing it in two bands. Feedback-correct: this stretch-and-fold is exactly the mechanism that makes nearby points separate (the stretch) while pinning the surviving set onto a thin skeleton (the squeeze). Feedback-wrong: rotations, reflections, and uniform shrinking do not separate nearby trajectories, so they cannot produce chaos.

Exercise (easy, true-false).

True or false: a point of the horseshoe is completely determined by its itinerary — the doubly infinite record of which strip ( $0$ or $1$ ) it visits at each step.

Hint

Recall that the surviving set is a fine dust of points and that no two distinct surviving points share the same forward-and-backward strip record.

Answer

True. The itinerary map is a one-to-one correspondence between surviving points and doubly infinite strings of two symbols: each surviving point has exactly one itinerary, and each string of $0$ s and $1$ s is the itinerary of exactly one surviving point. Feedback-correct: this is the coding that turns the folded geometry into all sequences of two letters. Feedback-wrong: distinct surviving points always differ at some step, so the itinerary never fails to separate them.

Formal definition Intermediate+

Work in the unit square $S = [0, 1]^{2} \subseteq R^{2}$ (or a disc containing it). The Smale horseshoe is a diffeomorphism $f$ of the plane, mapping $S$ across itself, built from the stretch-and-fold picture and pinned down by the following geometry. There are two disjoint full-height vertical strips $V_{0}, V_{1} \subseteq S$ and two disjoint full-width horizontal strips $H_{0}, H_{1} \subseteq S$ such that $f (V_{i}) = H_{i}$ , and on each $V_{i}$ the map $f$ is affine: it expands the horizontal direction by a factor $μ > 2$ and contracts the vertical direction by a factor $0 < λ < 1/2$ , possibly composed with a reflection on $V_{1}$ to produce the fold. Thus $f$ stretches each vertical strip into a long thin horizontal strip laid across $S$ .

Definition (the invariant set). The maximal $f$ -invariant subset of $S$ is $$ \Lambda = \bigcap_{n \in \mathbb{Z}} f^n(S) = { x : f^n(x) \in S \text{ for all } n \in \mathbb{Z} }. $$ The forward intersection $⋂_{n \geq 0} f^{- n} (S)$ is a product $C \times [0, 1]$ with $C$ a Cantor set in the horizontal coordinate; the backward intersection $⋂_{n \geq 0} f^{n} (S)$ is $[0, 1] \times C^{'}$ ; their intersection $Λ = C \times C^{'}$ is a product of two Cantor sets, a compact, perfect, totally disconnected set.

Definition (itinerary / coding map). For $x \in Λ$ define the itinerary $h (x) \in {0, 1}^{Z}$ by $$ h(x)_n = i \iff f^n(x) \in V_i \quad (i \in {0,1}), $$ which is well defined because each $f^{n} (x) \in Λ \subseteq V_{0} ⊔ V_{1}$ . Write $σ : {0, 1}^{Z} \to {0, 1}^{Z}$ for the full two-sided $2$ -shift $(σ a)_{n} = a_{n + 1}$ of 38.02.02; here the transition matrix is the all-ones $2 \times 2$ matrix, so every pair of symbols is allowed and ${0, 1}^{Z}$ is the full shift, not a proper subshift.

Sign / convention. Throughout, $Λ$ is hyperbolic in the sense of 38.03.01 with the horizontal direction unstable ( $E^{u}$ , expanded by $μ$ ) and the vertical direction stable ( $E^{s}$ , contracted by $λ$ ); the constants are uniform, $(C, λ)$ -hyperbolic with $C = 1$ in the affine model and adapted metric. This matches Katok-Hasselblatt ^{[Katok-Hasselblatt 1995 §2.5]}.

Definition (homoclinic point). Let $p$ be a hyperbolic periodic point of a diffeomorphism $g$ , with stable and unstable manifolds $W^{s} (p), W^{u} (p)$ (the global invariant manifolds of 38.03.01). A point $q \neq = p$ is homoclinic to $p$ if $q \in W^{s} (p) \cap W^{u} (p)$ , i.e. $g^{n} q \to p$ as $n \to + \infty$ and as $n \to - \infty$ (along the periodic orbit). The homoclinic point $q$ is transverse if the intersection is transverse at $q$ : $$ T_q W^s(p) \oplus T_q W^u(p) = T_q M, $$ and non-degenerate tangential (a homoclinic tangency) if the manifolds touch without crossing transversally.

Definition (the homoclinic tangle). Near a transverse homoclinic point the invariant manifolds $W^{s} (p)$ and $W^{u} (p)$ intersect infinitely often: since both are invariant, the single intersection $q$ forces intersections at all $g^{n} q$ , and the manifolds, unable to self-intersect, fold into an intricate lattice of "lobes" — the homoclinic tangle Poincaré discovered in the three-body problem.

Counterexamples to common slips

The horseshoe shift is the full shift, not a proper SFT. Both symbols may follow either symbol, so the transition matrix is $(1111)$ and the system is the full $2$ -shift with entropy $lo g 2$ . A horseshoe with a forbidden transition is a different (sub)shift; the standard horseshoe forbids nothing.
Homoclinic is not the same as heteroclinic. A homoclinic point connects a periodic orbit to itself ( $W^{s}$ and $W^{u}$ of the same $p$ ); a heteroclinic point connects two different orbits. The Smale-Birkhoff theorem is about the homoclinic (self-) case.
Transversality is essential. A homoclinic tangency (non-transverse intersection) does not directly yield a horseshoe; tangencies are the seeds of Newhouse phenomena and bifurcation cascades, a strictly more delicate theory. The clean horseshoe conclusion needs the transverse crossing.
The invariant set is a Cantor set, not a curve or a region. $Λ$ has empty interior and no isolated points; it is $C \times C^{'}$ , dust, not a submanifold. Expecting $Λ$ to be one-dimensional is a frequent error.

Key theorem with proof Intermediate+

Theorem (Smale: $f ∣_{Λ}$ is conjugate to the full $2$ -shift). For the Smale horseshoe $f$ with invariant set $Λ = ⋂_{n \in Z} f^{n} (S)$ , the itinerary map $h : Λ \to {0, 1}^{Z}$ is a homeomorphism with $h \circ f = σ \circ h$ . Consequently $Λ$ is a hyperbolic invariant set on which $f$ has $2^{n}$ points of period dividing $n$ , a dense set of periodic points, a dense orbit, sensitive dependence on initial conditions, and topological entropy $lo g 2$ . (Smale 1967 ^{[Smale 1967]}; Katok-Hasselblatt ^{[Katok-Hasselblatt 1995 §2.5]}; Devaney ^{[Devaney 1989 §2.3]}.)

Proof. The coding is well defined and conjugating. For $x \in Λ$ every iterate $f^{n} (x)$ lies in $Λ \subseteq V_{0} ⊔ V_{1}$ , so $h (x)_{n} \in {0, 1}$ is defined. The intertwining is immediate from the definition: $h (f (x))_{n} = i ⟺ f^{n} (f (x)) = f^{n + 1} (x) \in V_{i} ⟺ h (x)_{n + 1} = i$ , so $h (f (x)) = σ (h (x))$ .

Surjectivity. Fix a target sequence $a \in {0, 1}^{Z}$ . The set of points with prescribed itinerary on the window $[- m, m]$ is $$ R_{a,m} = \bigcap_{n=-m}^{m} f^{-n}(V_{a_n}), $$ which is a non-empty closed rectangle: applying $f^{- n}$ to the full-height strip $V_{a_{n}}$ and intersecting, the forward conditions ( $n \geq 0$ ) cut out a thin vertical sub-strip and the backward conditions ( $n < 0$ ) cut out a thin *horizontal* sub-strip, and because $μ > 2$ and $λ < 1/2$ each added condition shrinks the relevant width by a definite factor while keeping the rectangle non-empty (the strips are full height/width, so the intersections always meet). The nested sequence $R_{a, m}$ of non-empty compact rectangles has horizontal width $\leq μ^{- m}$ and vertical height $\leq λ^{m}$ , both $\to 0$ , so by the nested-compact-set theorem $⋂_{m} R_{a, m}$ is a single point $x$ , and $h (x) = a$ .

Injectivity. If $h (x) = h (y) = a$ then $x, y \in ⋂_{m} R_{a, m}$ , which we just showed is a single point, so $x = y$ .

Continuity and homeomorphism. $h$ is continuous: two points with the same itinerary on $[- m, m]$ lie in a common rectangle $R_{a, m}$ of diameter $\to 0$ , so closeness of itineraries (agreement on a long central window) forces closeness of points, and conversely the strips $V_{0}, V_{1}$ are a fixed positive distance apart so nearby points share a long central itinerary block. Thus $h$ and $h^{- 1}$ are continuous; $Λ$ compact and ${0, 1}^{Z}$ Hausdorff make $h$ a homeomorphism.

Dynamical consequences via the conjugacy. Transport the shift's properties of 38.02.02 across $h$ . The full $2$ -shift has $# Fix (σ^{n}) = tr (A^{n})$ with $A = (1111)$ , giving $tr (A^{n}) = 2^{n}$ , so $f ∣_{Λ}$ has $2^{n}$ points of period dividing $n$ ; periodic sequences are dense, hence periodic points are dense in $Λ$ ; the sequence concatenating all finite words has dense orbit, so $f ∣_{Λ}$ is topologically transitive; sensitive dependence holds because two sequences agreeing on a long central block but differing far out are shifted until they differ centrally. Finally the topological entropy is a conjugacy invariant and equals $lo g$ of the Perron eigenvalue $2$ of $A$ , so $h_{top} (f ∣_{Λ}) = lo g 2$ . $□$

Bridge. This conjugacy builds toward the entire theory of chaotic invariant sets and appears again in every concrete chaos verification, because it realises the abstract hyperbolic basic set of 38.03.01 as the single most explicit model: a product of Cantor sets carrying exactly the full $2$ -shift of 38.02.02. The foundational reason the coding works is that uniform stretch-and-fold makes the forward itinerary pin the horizontal (unstable) coordinate to a point and the backward itinerary pin the vertical (stable) coordinate — this is exactly the local product structure $W^{u} \times W^{s}$ of a hyperbolic set, read symbolically. The horseshoe generalises the cat map: where the toral automorphism codes to a proper subshift through a Markov partition, the horseshoe codes to the full shift because nothing is forbidden, and the central insight is that hyperbolicity plus the fold geometry turns a planar diffeomorphism into pure combinatorics. The horseshoe is dual to the spectral decomposition — there a hyperbolic set is resolved into transitive atoms, here a single atom is displayed as a full shift — and putting these together the horseshoe becomes the universal local model that the next theorem shows is forced wherever stable and unstable manifolds cross transversally. The bridge is the recognition that a Cantor set of itineraries is the geometric face of the full shift.

Exercises Intermediate+

Exercise 4 (medium, symbolic).

Show directly that $Λ = ⋂_{n \in Z} f^{n} (S)$ is a hyperbolic set for $f$ , identifying $E^{s}$ and $E^{u}$ and the constants, using the affine form of $f$ on the strips ( $μ > 2$ horizontal expansion, $λ < 1/2$ vertical contraction).

Hint

On each strip $D f$ is the diagonal matrix $diag (\pm μ, \pm λ)$ (sign from the fold). The horizontal axis is expanded, the vertical axis contracted, uniformly over $Λ$ .

Answer

On $V_{0}$ and $V_{1}$ the map is affine with derivative $D f = diag (μ, λ)$ up to the reflection on $V_{1}$ (which sends $λ \mapsto - λ$ , irrelevant to norms). Take $E_{x}^{u} = R \cdot (1, 0)$ (horizontal) and $E_{x}^{s} = R \cdot (0, 1)$ (vertical) at every $x \in Λ$ ; these are constant, continuous, and $D f$ -invariant since $D f$ is diagonal. For $v \in E_{x}^{s}$ , $∥ D f^{n} v ∥ = λ^{n} ∥ v ∥$ with $0 < λ < 1/2 < 1$ ; for $v \in E_{x}^{u}$ , $∥ D f^{- n} v ∥ = μ^{- n} ∥ v ∥ \leq 2^{- n} ∥ v ∥$ with $μ^{- 1} < 1/2 < 1$ . Thus $Λ$ is hyperbolic with $C = 1$ and contraction rate $max (λ, μ^{- 1}) < 1/2$ , the horizontal direction unstable and the vertical stable. Rubric: full credit for the constant splitting, the diagonal derivative, and the uniform $(C, λ)$ bounds.

Exercise 5 (medium, short-answer).

State the $λ$ -lemma (inclination lemma) precisely and explain, in one paragraph, why it is the mechanism that lets a transverse homoclinic point produce a horseshoe.

Hint

The lemma says forward iterates of a disc transverse to $W^{s} (p)$ converge $C^{1}$ to $W^{u} (p)$ . Use it to make a strip near $W^{u} (p)$ fold back across a strip near $W^{s} (p)$ .

Answer

$λ$ -lemma. Let $p$ be a hyperbolic fixed point of a $C^{1}$ diffeomorphism $g$ , with local stable and unstable manifolds $W_{loc}^{s} (p), W_{loc}^{u} (p)$ . Let $D$ be a disc of dimension $dim W^{u} (p)$ meeting $W_{loc}^{s} (p)$ transversally at one point. Then for every $ε > 0$ there is $N$ such that for $n \geq N$ the iterate $g^{n} (D)$ contains a disc that is $ε$ - $C^{1}$ -close to $W_{loc}^{u} (p)$ . Why it forces a horseshoe. A transverse homoclinic point $q \in W^{s} (p) ⋔ W^{u} (p)$ gives a small piece of $W^{u} (p)$ passing transversally through $W^{s} (p)$ near $q$ . Iterating forward, the $λ$ -lemma makes that piece stretch along and accumulate on $W^{u} (p)$ , carrying it back near $p$ and across $W^{s} (p)$ again; iterating the stable side backward does the dual. The result is a thin neighbourhood that $g^{N}$ stretches in the unstable direction, contracts in the stable direction, and folds back across itself — precisely the Markov/Conley-Moser strip geometry of a horseshoe for the iterate $g^{N}$ . Rubric: full credit for the correct statement and the stretch-fold-from-accumulation explanation.

Exercise 6 (medium, short-answer).

Explain why the existence of a single transverse homoclinic point forces infinitely many of them, and describe the resulting homoclinic tangle.

Hint

$W^{s} (p)$ and $W^{u} (p)$ are both invariant. If $q$ lies on both, where do all the iterates $g^{n} q$ lie?

Answer

Both $W^{s} (p)$ and $W^{u} (p)$ are $g$ -invariant: $g (W^{s} (p)) = W^{s} (p)$ and likewise for $W^{u} (p)$ . If $q \in W^{s} (p) \cap W^{u} (p)$ , then $g^{n} q \in W^{s} (p) \cap W^{u} (p)$ for every $n \in Z$ , and since $q$ is homoclinic, $g^{n} q \to p$ as $n \to \pm \infty$ but the $g^{n} q$ are distinct (away from $p$ ), so the intersection is infinite, accumulating on $p$ . Transversality at $q$ is preserved by the diffeomorphism, so each $g^{n} q$ is again a transverse intersection. Because $W^{s} (p)$ and $W^{u} (p)$ are injectively immersed curves that cannot self-intersect, they accommodate infinitely many mutual crossings only by oscillating wildly: $W^{u} (p)$ develops longer and thinner "lobes" stretching along $W^{s} (p)$ (and dually), interlacing into the homoclinic tangle. This is the picture Poincaré found in the restricted three-body problem and called too complicated to draw. Rubric: full credit for the invariance-forces-infinitely-many argument and the lobe/tangle description.

Exercise 7 (hard, short-answer).

Prove that the Smale horseshoe is structurally stable on a neighbourhood of $Λ$ : any $g$ that is $C^{1}$ -close to $f$ has an invariant set $Λ_{g}$ near $Λ$ with $g ∣_{Λ_{g}}$ topologically conjugate to $f ∣_{Λ}$ (hence to the full $2$ -shift). You may cite the $C^{1}$ -openness of hyperbolicity from 38.03.01.

Hint

Hyperbolicity is $C^{1}$ -open and the strip-crossing geometry (Conley-Moser conditions) is an open condition; both persist under small perturbation, and the coding survives.

Answer

The horseshoe geometry is governed by the Conley-Moser conditions: $g$ maps two vertical strips across two horizontal strips with uniform horizontal expansion and vertical contraction and with the correct boundary behaviour. These are open conditions in the $C^{1}$ topology: each is a strict inequality on $g$ and $D g$ over the compact square. Hence for $g$ in a $C^{1}$ -neighbourhood of $f$ , the perturbed map $g$ still maps two vertical strips $V_{0}^{g}, V_{1}^{g}$ across two horizontal strips with expansion $μ^{'} > 2$ and contraction $λ^{'} < 1/2$ close to the originals. The same nested-rectangle argument as in the Key theorem then produces a coding map $h_{g} : Λ_{g} \to {0, 1}^{Z}$ , where $Λ_{g} = ⋂_{n} g^{n} (S)$ , which is again a homeomorphism conjugating $g ∣_{Λ_{g}}$ to the full $2$ -shift $σ$ . Composing, $h_{g}^{- 1} \circ h : Λ \to Λ_{g}$ is a homeomorphism with $(h_{g}^{- 1} \circ h) \circ f = g \circ (h_{g}^{- 1} \circ h)$ , the desired conjugacy. Equivalently, $Λ$ is a hyperbolic set, hyperbolicity is $C^{1}$ -open by the cone criterion of 38.03.01, and the local product structure plus expansiveness give structural stability of the basic set via shadowing. Rubric: full credit for invoking openness of the Conley-Moser/cone conditions, reconstructing the coding for $g$ , and assembling the conjugacy.

Exercise 8 (hard, short-answer).

State the Melnikov criterion for the existence of transverse homoclinic points in a periodically forced planar system $\overset{z}{˙} = X_{0} (z) + ε X_{1} (z, t)$ , where the unperturbed system has a homoclinic orbit $z_{0} (t)$ to a hyperbolic saddle. Explain how a simple zero of the Melnikov function yields, via the Smale-Birkhoff theorem, a horseshoe.

Hint

The Melnikov function $M (t_{0})$ is the leading-order signed distance between the perturbed stable and unstable manifolds, an integral of the wedge $X_{0} \land X_{1}$ along the unperturbed homoclinic orbit. A simple zero means the manifolds cross transversally.

Answer

Melnikov function. For $\overset{z}{˙} = X_{0} (z) + ε X_{1} (z, t)$ with $X_{1}$ $T$ -periodic in $t$ and an unperturbed homoclinic orbit $z_{0} (t)$ to a hyperbolic saddle of $X_{0}$ , define $$ M(t_0) = \int_{-\infty}^{\infty} X_0\big(z_0(t)\big) \wedge X_1\big(z_0(t),, t + t_0\big), dt, $$ where $a \land b = a_{1} b_{2} - a_{2} b_{1}$ . To first order in $ε$ , the distance between the perturbed stable and unstable manifolds of the saddle (now a hyperbolic periodic orbit of the time- $T$ map / suspended flow) at the section parameter $t_{0}$ is $\frac{ε M ( t _{0} )}{∥ X _{0} ( z _{0} ( 0 )) ∥} + O (ε^{2})$ . Simple zero $\Rightarrow$ horseshoe. If $M (t_{0})$ has a simple zero $M (t_{0}^{*}) = 0$ , $M^{'} (t_{0}^{*}) \neq = 0$ , then for small $ε \neq = 0$ the perturbed stable and unstable manifolds intersect transversally, producing a transverse homoclinic point of the Poincaré (time- $T$ ) map. The Smale-Birkhoff homoclinic theorem then gives a horseshoe in some iterate of the Poincaré map: the system has a hyperbolic invariant set conjugate to the full shift, hence chaos, for all sufficiently small $ε \neq = 0$ . Rubric: full credit for the Melnikov integral, the simple-zero transversality condition, and chaining to Smale-Birkhoff.

Advanced results Master

Theorem (Smale-Birkhoff homoclinic theorem). Let $g$ be a diffeomorphism of a manifold with a hyperbolic periodic point $p$ possessing a transverse homoclinic point $q \in W^{s} (p) ⋔ W^{u} (p)$ . Then there is an integer $N$ and a compact $g^{N}$ -invariant set $Λ ∋ p, q$ such that $g^{N} ∣_{Λ}$ is topologically conjugate to the full two-sided shift on finitely many symbols; in particular $g$ has infinitely many periodic orbits, positive topological entropy, and sensitive dependence on initial conditions in a neighbourhood of the homoclinic orbit. (Smale 1965, 1967 ^{[Smale 1967]}; Katok-Hasselblatt ^{[Katok-Hasselblatt 1995 §6.6]}; Moser ^{[Moser 1973]}.)

The proof is the geometric realisation of Exercise 5. Take a small "box" neighbourhood $B$ straddling the local stable manifold near $p$ and the homoclinic point $q$ . The $λ$ -lemma guarantees that for large $N$ the image $g^{N} (B)$ returns and crosses $B$ in the stretch-fold horseshoe configuration: $g^{N}$ expands along the $W^{u}$ -direction, contracts along the $W^{s}$ -direction, and laps $g^{N} (B)$ back over $B$ in at least two disjoint strips. The Conley-Moser conditions are then verified for $g^{N}$ on $B$ , so $g^{N} ∣_{Λ}$ with $Λ = ⋂_{n} g^{N n} (B)$ is conjugate to a full shift by the Key theorem's coding argument. This converts the qualitative datum " $W^{s}$ and $W^{u}$ cross transversally" into the quantitative chaos of a shift, and it is the theorem that makes the horseshoe the universal local model of chaos.

Theorem ( $λ$ -lemma / inclination lemma; Palis). Let $p$ be a hyperbolic fixed point of a $C^{1}$ diffeomorphism $g$ of $R^{m}$ , with $dim W^{u} (p) = u$ . Let $D^{u}$ be a $u$ -dimensional disc transverse to $W_{loc}^{s} (p)$ . Then $⋃_{n \geq 0} g^{n} (D^{u})$ contains discs converging in the $C^{1}$ topology to $W_{loc}^{u} (p)$ : for every $ε > 0$ there is $N$ with $g^{n} (D^{u}) \supseteq D_{n}^{u}$ , $D_{n}^{u}$ being $ε$ - $C^{1}$ -close to $W_{loc}^{u} (p)$ for $n \geq N$ . (Palis 1969; Katok-Hasselblatt ^{[Katok-Hasselblatt 1995 §6.5]}.)

The inclination lemma is the engine driving every homoclinic construction. Its proof is a graph-transform / cone-field argument identical in spirit to the stable-manifold theorem of 38.03.01: write $D^{u}$ as a graph over $E^{u}$ with bounded "inclination" (tangent slope toward $E^{s}$ ); the hyperbolic linear part contracts inclinations toward $0$ under forward iteration, so the iterated discs flatten onto the unstable directions and, by uniform expansion along $E^{u}$ , grow to fill out $W_{loc}^{u} (p)$ . The $C^{1}$ -closeness — not merely $C^{0}$ — is what makes transversality persist along the construction, which is why the Smale-Birkhoff conclusion is robust.

Theorem (Birkhoff: accumulation of periodic orbits at a homoclinic point). Near a transverse homoclinic point of a surface diffeomorphism there are infinitely many periodic points, of unboundedly large period, accumulating on the homoclinic orbit. (Birkhoff 1935; subsumed by Smale-Birkhoff.)

Birkhoff proved, before the symbolic-dynamics formulation, that transverse homoclinic points are accumulated by periodic points — the analytic discovery that Smale's horseshoe later explained structurally. In the coding, a periodic orbit shadowing the homoclinic excursion corresponds to a periodic itinerary $\overline{0^{k} 1^{j}}$ that spends a long time near $p$ (the symbol for the box near $p$ ) and makes the homoclinic loop (the symbol for the excursion box); letting the dwell time $k$ grow produces periodic orbits of unbounded period limiting onto the homoclinic orbit. This is the historical seed of the whole subject.

Theorem (Melnikov's method; existence of transverse homoclinic orbits). For $\overset{z}{˙} = X_{0} (z) + ε X_{1} (z, t)$ with $X_{1}$ $T$ -periodic and an unperturbed homoclinic loop $z_{0}$ to a hyperbolic saddle, if the Melnikov function $M (t_{0}) = \int_{- \infty}^{\infty} X_{0} (z_{0} (t)) \land X_{1} (z_{0} (t), t + t_{0}) d t$ has a simple zero, then for all small $ε \neq = 0$ the Poincaré map has a transverse homoclinic point, hence (by Smale-Birkhoff) a horseshoe. (Melnikov 1963; Guckenheimer-Holmes ^{[Guckenheimer-Holmes 1983 §4.5–§5.3]}.)

Melnikov's method makes the abstract hypothesis "transverse homoclinic point" checkable in concrete forced oscillators: one computes a single convergent integral along the known unperturbed orbit and checks it for simple zeros. Applied to the forced Duffing equation $\overset{x}{¨} - x + x^{3} = ε (γ cos ω t - δ \overset{x}{˙})$ it yields the threshold $γ / δ >$ (an explicit ratio of hyperbolic functions of $ω$ ) above which the system has a horseshoe, the first rigorous chaos criterion for a physical differential equation, and it is the bridge from the geometric Smale-Birkhoff theorem to applied nonlinear dynamics.

Synthesis. These four theorems are one mechanism viewed at four altitudes, and putting these together the qualitative fact of a transverse manifold crossing becomes the quantitative chaos of the full shift. The central insight is that the $λ$ -lemma converts a single transverse homoclinic point into the stretch-fold geometry of a horseshoe, so the foundational reason chaos appears near any transverse homoclinic orbit is that uniform hyperbolicity at $p$ forces iterated discs to accumulate on $W^{u} (p)$ , manufacturing the Conley-Moser strips for an iterate $g^{N}$ . This is exactly the principle that the horseshoe of the Key theorem is forced, not assumed: where that theorem built a horseshoe by hand, Smale-Birkhoff shows the same structure is compelled by transversality, and the coding of 38.02.02 then certifies entropy $lo g (# symbols)$ , infinitely many periodic orbits, and sensitive dependence. The horseshoe is dual to the spectral decomposition of 38.03.01 — that theorem disassembles a hyperbolic non-wandering set into transitive atoms, this one shows each homoclinic tangle contains such an atom — and Melnikov's method makes the whole chain effective, generalising the abstract theory to the forced Duffing and pendulum equations where a single integral certifies the tangle. The bridge is the realisation that Poincaré's intractable homoclinic tangle and Smale's transparent symbolic horseshoe are the same object: the tangle is the geometry, the full shift is its arithmetic.

Full proof set Master

Proposition 1 (the invariant set is a product of two Cantor sets). For the affine Smale horseshoe with $μ > 2$ , $λ < 1/2$ , the set $Λ = ⋂_{n \in Z} f^{n} (S)$ equals $C \times C^{'}$ where $C, C^{'}$ are Cantor sets.

Proof. The forward-invariant set $Λ^{+} = ⋂_{n \geq 0} f^{- n} (S)$ is computed by intersecting preimages: $f^{- 1} (S) \cap S = V_{0} ⊔ V_{1}$ (two vertical strips, since $f$ maps each $V_{i}$ onto a full-width horizontal strip, so the points of $S$ whose image is in $S$ are exactly the two vertical strips). Inductively $⋂_{k = 0}^{n} f^{- k} (S)$ is a union of $2^{n}$ vertical sub-strips, each of horizontal width $μ^{- n} \to 0$ , nested and refining by the factor $μ^{- 1} < 1/2$ at each stage; the horizontal cross-section is therefore the standard middle-fraction Cantor construction, giving $Λ^{+} = C \times [0, 1]$ with $C \subseteq [0, 1]$ a Cantor set (compact, perfect, totally disconnected). Dually $Λ^{-} = ⋂_{n \geq 0} f^{n} (S) = [0, 1] \times C^{'}$ with $C^{'}$ a Cantor set in the vertical coordinate, from the contraction $λ < 1/2$ . Then $Λ = Λ^{+} \cap Λ^{-} = (C \times [0, 1]) \cap ([0, 1] \times C^{'}) = C \times C^{'}$ , a product of two Cantor sets, hence itself compact, perfect, and totally disconnected. $□$

Proposition 2 (the coding map is a homeomorphism). The itinerary map $h : Λ \to {0, 1}^{Z}$ is a bijection, continuous with continuous inverse, intertwining $f$ and $σ$ .

Proof. Intertwining $h \circ f = σ \circ h$ is the index shift $h (f x)_{n} = [f^{n + 1} x \in V_{∙}] = h (x)_{n + 1}$ . For bijectivity, fix $a \in {0, 1}^{Z}$ and set $R_{a, m} = ⋂_{n = - m}^{m} f^{- n} (V_{a_{n}})$ . Each forward condition $f^{- n} (V_{a_{n}})$ , $n \geq 0$ , is a vertical strip of width $μ^{- n}$ ; their intersection over $0 \leq n \leq m$ is one vertical sub-strip of width $μ^{- m}$ . Each backward condition, $n < 0$ , is a horizontal strip of height $λ^{∣ n ∣}$ ; their intersection is one horizontal sub-strip of height $λ^{m}$ . The rectangle $R_{a, m}$ is the (non-empty, since strips are full height/width and always meet) intersection, of horizontal width $μ^{- m}$ and vertical height $λ^{m}$ , both tending to $0$ . The nested compact rectangles $R_{a, m} \supseteq R_{a, m + 1} \supseteq \dots$ have intersection a single point $x = h^{- 1} (a)$ ; this gives surjectivity and injectivity at once. Continuity of $h$ : if $h (x), h (y)$ agree on $[- m, m]$ then $x, y \in R_{h (x), m}$ of diameter $\to 0$ , so $d (x, y) \to 0$ as $m \to \infty$ ; continuity of $h^{- 1}$ : the strips $V_{0}, V_{1}$ have a positive gap, so points within distance $δ_{m}$ share the central itinerary block $[- m, m]$ . Compact domain plus Hausdorff codomain upgrade the continuous bijection $h$ to a homeomorphism. $□$

Proposition 3 (the horseshoe is structurally stable). There is a $C^{1}$ -neighbourhood $U$ of the horseshoe $f$ such that every $g \in U$ has $Λ_{g} = ⋂_{n} g^{n} (S)$ with $g ∣_{Λ_{g}}$ topologically conjugate to $f ∣_{Λ}$ .

Proof. $Λ$ is a hyperbolic set for $f$ (Exercise 4: constant splitting, $D f = diag (\pm μ, \pm λ)$ , uniform $(1, max (λ, μ^{- 1}))$ -bounds). By the $C^{1}$ -openness of hyperbolicity 38.03.01, there is a $C^{1}$ -neighbourhood $U_{0}$ of $f$ and a neighbourhood $V \supseteq Λ$ such that for $g \in U_{0}$ the maximal $g$ -invariant set $Λ_{g}$ in $V$ is hyperbolic with cone fields close to those of $f$ . The defining strip geometry (Conley-Moser conditions: two vertical strips mapped across two horizontal strips, uniform expansion/contraction, correct boundary crossing) is a finite list of strict inequalities on $g, D g$ over the compact square, hence open; shrink to $U \subseteq U_{0}$ so these persist. The nested-rectangle construction of Proposition 2 applies verbatim to $g$ (widths shrink by $μ^{' - 1} < 1/2$ , heights by $λ^{'} < 1/2$ ), yielding a coding homeomorphism $h_{g} : Λ_{g} \to {0, 1}^{Z}$ with $h_{g} \circ g = σ \circ h_{g}$ . Then $Φ = h_{g}^{- 1} \circ h : Λ \to Λ_{g}$ is a homeomorphism and $Φ \circ f = h_{g}^{- 1} \circ σ \circ h = h_{g}^{- 1} \circ σ \circ h_{g} \circ Φ = g \circ Φ$ , the conjugacy. $□$

Proposition 4 ( $λ$ -lemma in the linear model). Let $g$ have hyperbolic fixed point $0$ with $D g_{0} = diag (μ, λ)$ , $μ > 1 > λ > 0$ in $R^{2}$ , $W^{u} = {y = 0}$ , $W^{s} = {x = 0}$ . Let $D$ be a curve ${(x, ϕ (x)) : ∣ x ∣ \leq r}$ transverse to $W^{s}$ at a point with $x$ -coordinate $x_{0} \neq = 0$ and bounded slope. Then $g^{n} (D)$ contains arcs $C^{1}$ -converging to $W_{loc}^{u}$ .

Proof. Restrict to the linear model $g (x, y) = (μx, λ y)$ (the nonlinear case follows by a graph-transform perturbation of this computation). A point $(x, ϕ (x))$ of $D$ maps to $(μx, λ ϕ (x))$ , so $g^{n} (D)$ is the curve $ξ \mapsto (μ^{n} x, λ^{n} ϕ (x))$ with $ξ = μ^{n} x$ ; as a graph over the $x$ -axis it is $y = ψ_{n} (ξ) := λ^{n} ϕ (μ^{- n} ξ)$ . The slope is $ψ_{n}^{'} (ξ) = (λ / μ)^{n} ϕ^{'} (μ^{- n} ξ) \to 0$ uniformly on $∣ ξ ∣ \leq R$ , since $(λ / μ)^{n} \to 0$ and $ϕ^{'}$ is bounded. The value $ψ_{n} (ξ) = λ^{n} ϕ (μ^{- n} ξ) \to 0$ uniformly as well. Meanwhile the $x$ -extent of $g^{n} (D)$ is $μ^{n} r \to \infty$ , so for any fixed window $∣ ξ ∣ \leq R$ the arc $g^{n} (D) \cap {∣ ξ ∣ \leq R}$ is non-empty for large $n$ (using $x_{0} \neq = 0$ so the transverse point's image has growing $∣ x ∣$ ). Thus on each fixed window the arcs converge in $C^{1}$ (value and slope both $\to 0$ ) to the segment ${y = 0} = W_{loc}^{u}$ . $□$

Proposition 5 (Smale-Birkhoff: transverse homoclinic point forces a full shift in an iterate). If $g$ has a hyperbolic fixed point $p$ and a transverse homoclinic point $q$ , there is $N$ and a $g^{N}$ -invariant Cantor set $Λ ∋ p, q$ with $g^{N} ∣_{Λ}$ conjugate to the full $2$ -shift.

Proof. Choose linearising coordinates near $p$ with $W_{loc}^{s}, W_{loc}^{u}$ the coordinate axes. Pick a small rectangle $R_{0}$ around $p$ (the "home" box) with sides parallel to $W_{loc}^{s}, W_{loc}^{u}$ . The homoclinic point $q$ has $g^{- k} q \to p$ and $g^{k} q \to p$ ; choose $k$ so that $g^{- k} q \in R_{0}$ near $W_{loc}^{u}$ and a rectangle $R_{1}$ around $q$ 's excursion with $g^{m} (R_{1}) \subseteq R_{0}$ for suitable $m$ . By the $λ$ -lemma (Proposition 4) applied to the unstable side of $R_{0}$ , for large $N$ the image $g^{N} (R_{0})$ stretches along $W^{u} (p)$ , follows the homoclinic loop through $R_{1}$ , and returns to lie across $R_{0}$ , transverse to $W_{loc}^{s}$ near both $p$ and $q$ . Concretely $g^{N} (R_{0})$ meets $R_{0}$ in (at least) two disjoint full-height vertical sub-strips — one near $p$ , one near the homoclinic excursion — while $g^{N}$ expands the $W^{u}$ -direction and contracts the $W^{s}$ -direction. These are exactly the Conley-Moser horseshoe conditions for the map $G = g^{N}$ on $R_{0}$ with two symbols. The coding argument of Proposition 2 produces $Λ = ⋂_{n} G^{n} (R_{0})$ and a homeomorphism $Λ \to {0, 1}^{Z}$ conjugating $G ∣_{Λ}$ to $σ$ ; $p$ corresponds to $\overline{0}$ and the homoclinic orbit to a sequence asymptotic to $\overline{0}$ in both directions. $□$

Proposition 6 (entropy and periodic-orbit growth at a transverse homoclinic point). Under the hypotheses of Proposition 5, $g$ has positive topological entropy $h_{top} (g) \geq \frac{1}{N} lo g 2 > 0$ , and the number of periodic points of $g$ of period $\leq n$ grows at least exponentially.

Proof. From Proposition 5, $G = g^{N}$ restricted to $Λ$ is conjugate to the full $2$ -shift, so $h_{top} (G ∣_{Λ}) = lo g 2$ . Topological entropy scales under iteration as $h_{top} (g^{N}) = N h_{top} (g)$ and is monotone under restriction to invariant subsets, so $N h_{top} (g) \geq h_{top} (g^{N} ∣_{Λ}) = lo g 2$ , giving $h_{top} (g) \geq \frac{1}{N} lo g 2 > 0$ . For periodic orbits, $G ∣_{Λ}$ has $# Fix (G^{j}) = 2^{j}$ by the full-shift count of 38.02.02, so $g$ has at least $2^{j}$ points of period dividing $N j$ ; the count of periodic points up to period $n$ is therefore $\geq 2^{⌊ n / N ⌋}$ , exponential in $n$ . $□$

Connections Master

Hyperbolic sets, Anosov and Axiom-A systems, and the Smale spectral decomposition 38.03.01. The horseshoe's invariant set $Λ$ is the model hyperbolic basic set: a single transitive piece of an Axiom-A spectral decomposition whose internal dynamics is the full shift. This unit makes concrete the abstract structures of the prerequisite — the splitting $E^{s} \oplus E^{u}$ is the vertical/horizontal directions, the stable/unstable manifolds are the Cantor-set fibres, structural stability is the $C^{1}$ -robustness of the coding, and the $λ$ -lemma is the inclination engine behind the spectral-decomposition transitivity argument there. Where the prereq disassembles recurrence into atoms, the horseshoe exhibits one atom in full detail.
Shifts of finite type, transition matrices, and coding 38.02.02. The conjugacy $f ∣_{Λ} ≅ σ$ to the full two-sided $2$ -shift is the bridge by which all symbolic-dynamics machinery applies to the horseshoe: the periodic-point count $# Fix = tr (A^{n}) = 2^{n}$ , the entropy $lo g 2 = lo g$ (Perron eigenvalue), topological transitivity and mixing of the shift, all transport across $h$ . The horseshoe is the case where the transition matrix is all-ones (nothing forbidden), so the SFT is the full shift; modified horseshoes with forbidden transitions realise proper subshifts of finite type, making this the geometric source of the symbolic models.
Dynamical systems, orbits, and limit sets 38.01.01. The horseshoe supplies the canonical example of every chaos notion defined abstractly in the foundational unit — sensitive dependence, a dense orbit, dense periodic points, a non-wandering set equal to a Cantor set — and the homoclinic point is the concrete mechanism producing the complicated $ω$ -limit and non-wandering structure catalogued there. The conjugacy invariants of the foundational unit (entropy, periodic-orbit counts) take their first nontrivial positive values on the horseshoe.
Topological entropy and its computation 38.06.01. The horseshoe's entropy $lo g 2$ is the prototypical positive topological entropy, computed via the conjugacy to the full $2$ -shift; the entropy chapter measures exactly the exponential orbit-growth rate that the horseshoe's $2^{n}$ periodic-point count exhibits, and the Smale-Birkhoff theorem is the standard route to proving a smooth system has positive entropy by locating a transverse homoclinic point.

Historical & philosophical context Master

The homoclinic phenomenon was discovered by Henri Poincaré in his prize memoir on the three-body problem (1890), where he found that the stable and unstable curves of a periodic solution, if they meet once transversally, must meet infinitely often, weaving a tangle he described as so intricate he would not attempt to draw it — the first sighting of deterministic chaos. George Birkhoff, in the 1930s, proved that such transverse homoclinic points are accumulated by infinitely many periodic orbits of growing period (Birkhoff 1935), the analytic core of what is now the Smale-Birkhoff theorem. Norman Levinson's 1949 analysis of the forced van der Pol equation exhibited an explicit map with infinitely many periodic points, and it was this example that prompted Stephen Smale to abstract the horseshoe in his 1965 paper Diffeomorphisms with many periodic points ^{[Smale 1965]} and to recognise it, in the 1967 survey Differentiable dynamical systems ^{[Smale 1967]}, as a structurally stable source of a hyperbolic invariant set conjugate to the full shift.

The symbolic conjugacy reframed Poincaré's tangle as the cleanest object in dynamics, the full two-symbol shift, and the inclination lemma supplying the mechanism was proved by Jacob Palis (1969). Jürgen Moser's 1973 Stable and Random Motions ^{[Moser 1973]} gave the analytic Hamiltonian account and the Conley-Moser conditions, and Vladimir Melnikov's 1963 perturbative method, popularised by John Guckenheimer and Philip Holmes ^{[Guckenheimer-Holmes 1983]}, made the transverse-homoclinic hypothesis computable in physical forced oscillators such as the Duffing equation. Katok and Hasselblatt's 1995 treatise ^{[Katok-Hasselblatt 1995]} is the canonical modern reference for the horseshoe, the λ-lemma, and the Smale-Birkhoff theorem.

Bibliography Master

@incollection{Smale1965horseshoe,
  author    = {Smale, Stephen},
  title     = {Diffeomorphisms with many periodic points},
  booktitle = {Differential and Combinatorial Topology (A Symposium in Honor of Marston Morse)},
  editor    = {Cairns, Stewart S.},
  publisher = {Princeton University Press},
  year      = {1965},
  pages     = {63--80}
}

@article{Smale1967horseshoe,
  author  = {Smale, Stephen},
  title   = {Differentiable dynamical systems},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {73},
  year    = {1967},
  pages   = {747--817}
}

@book{KatokHasselblatt1995horseshoe,
  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}
}

@book{Devaney1989,
  author    = {Devaney, Robert L.},
  title     = {An Introduction to Chaotic Dynamical Systems},
  publisher = {Addison-Wesley},
  year      = {1989}
}

@book{Moser1973,
  author    = {Moser, J{\"u}rgen},
  title     = {Stable and Random Motions in Dynamical Systems},
  publisher = {Princeton University Press},
  series    = {Annals of Mathematics Studies},
  volume    = {77},
  year      = {1973}
}

@book{GuckenheimerHolmes1983,
  author    = {Guckenheimer, John and Holmes, Philip},
  title     = {Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields},
  publisher = {Springer-Verlag},
  series    = {Applied Mathematical Sciences},
  volume    = {42},
  year      = {1983}
}

@article{Palis1969,
  author  = {Palis, Jacob},
  title   = {On {M}orse-{S}male dynamical systems},
  journal = {Topology},
  volume  = {8},
  year    = {1969},
  pages   = {385--404}
}

@article{Melnikov1963,
  author  = {Melnikov, Vladimir K.},
  title   = {On the stability of a center for time-periodic perturbations},
  journal = {Transactions of the Moscow Mathematical Society},
  volume  = {12},
  year    = {1963},
  pages   = {1--57}
}

Prerequisites

38.03.01
38.02.02
38.01.01

Tier anchors

beginner: Strogatz 1994 *Nonlinear Dynamics and Chaos* (Addison-Wesley) Ch. 12 §12.2 (the Lorenz map and stretching-and-folding); Devaney 1989 *An Introduction to Chaotic Dynamical Systems* (Addison-Wesley) Ch. 2 §2.3 (the horseshoe drawn step by step); 3Blue1Brown *Chaos* and the Smale-horseshoe visual essays for the fold-and-lay-back animation
intermediate: Devaney 1989 *An Introduction to Chaotic Dynamical Systems* (Addison-Wesley) Ch. 2 §2.3 (the horseshoe and its conjugacy to the shift); Robinson 1999 *Dynamical Systems* (CRC) Ch. 8 §8.4–§8.7 (the horseshoe, transverse homoclinic points, the λ-lemma, the Smale-Birkhoff theorem); Wiggins 2003 *Introduction to Applied Nonlinear Dynamical Systems and Chaos* (Springer) Ch. 23–26 (Melnikov's method)
master: Katok-Hasselblatt 1995 *Introduction to the Modern Theory of Dynamical Systems* (Cambridge) Ch. 2 §2.5 (the horseshoe), Ch. 6 §6.5 (the inclination/λ-lemma), Ch. 6 §6.6 (transverse homoclinic points and the Smale-Birkhoff theorem); Smale 1965 *Diffeomorphisms with many periodic points* (in *Differential and Combinatorial Topology*, Princeton); Smale 1967 *Differentiable dynamical systems* (Bull. Amer. Math. Soc. 73); Guckenheimer-Holmes 1983 *Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields* (Springer) Ch. 4–5 (the homoclinic tangle and Melnikov's method)

References

Katok, Hasselblatt 1995 — Introduction to the Modern Theory of Dynamical Systems · Ch. 2 §2.5 (the Smale horseshoe, the invariant Cantor set $\Lambda=\bigcap_n f^n(\Delta)$, conjugacy of $f|_\Lambda$ to the full two-sided 2-shift), Ch. 6 §6.5 (the inclination/λ-lemma), Ch. 6 §6.6 (transverse homoclinic points, the horseshoe embedded in an iterate, the Smale-Birkhoff homoclinic theorem); Encyclopedia of Mathematics and its Applications 54, Cambridge University Press 1995
Smale 1967 — Differentiable dynamical systems · Bulletin of the American Mathematical Society 73 (1967) 747–817 §I.5 (the horseshoe as the model structurally stable hyperbolic basic set conjugate to the shift, infinitely many periodic orbits surviving perturbation)
Smale 1965 — Diffeomorphisms with many periodic points · in Differential and Combinatorial Topology (S. Cairns, ed.), Princeton University Press 1965, 63–80; the original construction of the horseshoe diffeomorphism, motivated by Levinson's work on the forced van der Pol equation, exhibiting a structurally stable diffeomorphism with infinitely many periodic points
Devaney 1989 — An Introduction to Chaotic Dynamical Systems · Addison-Wesley 1989, Ch. 2 §2.3 (the horseshoe map, the invariant Cantor set as a product of Cantor sets, the itinerary map and its conjugacy to the two-sided shift on two symbols)
Guckenheimer, Holmes 1983 — Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields · Springer Applied Mathematical Sciences 42, Ch. 5 §5.1–§5.3 (the homoclinic tangle, the Smale-Birkhoff homoclinic theorem) and §4.5 (the Melnikov function as the leading-order distance between perturbed stable and unstable manifolds, with simple zeros forcing transverse homoclinic points)
Moser 1973 — Stable and Random Motions in Dynamical Systems · Annals of Mathematics Studies 77, Princeton University Press 1973, Ch. III (the horseshoe, symbolic dynamics, and the structure of homoclinic points near a transverse homoclinic orbit; the Birkhoff-Smale theory in the analytic Hamiltonian setting)

Estimated time

beginner: 19m
intermediate: 45m
master: 87m