Symbolic dynamics and horseshoes

Let Sigma be the set of all bi-infinite sequences of zeros and ones, indexed by every integer. Give it the product topology, for example through a weighted distance in which disagreement near index zero matters more than disagreement far away. The left shift sends the symbol at index j+1 into index j. It is a homeomorphism because shifting right reverses it. A one-sided shift instead forgets its earliest symbol and is not invertible, so the two models encode different amounts of past information.

For a diffeomorphism F on a geometric invariant set Lambda, an itinerary map h is a topological conjugacy when it is a homeomorphism and h after F equals shift after h. Every allowed symbolic orbit then corresponds to exactly one geometric orbit with the same temporal organization. A noninjective coding may give only a semiconjugacy, which transfers fewer properties without further argument. Boundaries and overlapping coding rectangles must be handled rather than silently assuming every symbolic label identifies one point.

In an ideal two-strip horseshoe, a smooth invertible planar map contracts one direction, expands the other, folds, and crosses a reference square in two separated strips with the standard full-crossing geometry. Keep only points that stay in the square for every integer iterate. Future strip choices narrow one coordinate; past choices narrow the other. Uniform contraction of the nested rectangles then identifies one point for every bi-infinite itinerary. This is the mechanism behind the invariant Cantor set and its conjugacy, not a consequence of a decorative U-shaped image alone.

The invariant set supports dense periodic points, topological transitivity, sensitive dependence, and positive topological entropy. Finite binary words generate exponentially many distinguishable orbit segments. A sequence repeating a word gives a periodic point, while nonperiodic sequences give nonperiodic orbits. The set is usually a saddle-type invariant Cantor set rather than an attractor: many nearby points leave the chosen square. Area contraction or one expanding singular value by itself establishes neither the required return geometry nor the full symbolic correspondence.

The Smale–Birkhoff homoclinic theorem connects this model to less explicitly constructed maps. For a continuously differentiable planar diffeomorphism, a transverse intersection away from a hyperbolic saddle fixed point between its stable and unstable manifolds produces a horseshoe for a suitable iterate in the nearby homoclinic dynamics. A hyperbolic periodic point can be treated through its period iterate. Transverse means the tangent stable and unstable directions span the plane at the intersection; merely drawing two curves close together or tangent is insufficient.

This is a theorem about an invariant set for some iterate, so time units must be preserved. If F to the k has a full two-symbol horseshoe, the corresponding entropy lower bound per original iteration is log two divided by k. It does not follow that the whole phase space has one mixing attractor, that almost every initial point visits the horseshoe, or that its symbolic probability law is automatically a fair independent coin process. Those are additional topological, measure-theoretic, or physical-observability questions. The rigorous gain is already substantial: complicated recurrence is forced on a precisely identified invariant subset.