Stretch, Fold, Repeat: Arnold's Cat Map

A deceptively simple stretch-and-fold

Take a square with opposite edges identified (a torus), and apply this linear map to every point (x, y), taken modulo 1:

[x']   [2  1] [x]
[y'] = [1  1] [y]     (mod 1)

That single matrix is Arnold’s cat map, popularised by Vladimir Arnold, who famously illustrated it by applying it repeatedly to a drawing of a cat’s face and watching it shred into apparent noise. The matrix has determinant 1 (so it preserves area — nothing is lost, everything is rearranged) and its eigenvalues are λ = (3 ± √5)/2, one greater than 1 and one its reciprocal, less than 1. That pair of eigenvalues is the entire mechanism: along one eigendirection every distance is stretched by λ ≈ 2.618 each step, along the perpendicular eigendirection it is compressed by the same factor.

What makes it an Anosov map

A map is called Anosov (after Dmitri Anosov) when, at every point, the tangent space splits into a stable direction that contracts and an unstable direction that expands, uniformly and persistently under iteration, with no neutral directions in between. The cat map is the simplest possible example: linear, so the stable and unstable directions are the same everywhere (the two eigenvectors of the matrix), rather than curving from point to point as they would in a general nonlinear Anosov system like geodesic flow on a negatively curved surface.

Why folding the torus makes it chaotic, not just stretching

Stretching alone, on an infinite plane, would just spread points apart forever — interesting, but not chaotic in any bounded sense. The torus is what turns stretching into folding: every time a stretched coordinate exceeds 1, the modulo operation wraps it back into [0,1), which folds the expanding direction back on top of itself. This stretch-and-fold pattern is the same qualitative mechanism as kneading dough, and it is exactly what produces the two hallmark signatures of chaos: sensitive dependence on initial conditions (nearby points separate at rate λⁿ per n iterations — the map’s positive Lyapunov exponent is ln λ ≈ 0.962) and topological mixing (any small patch of the torus, iterated enough times, spreads its area evenly over the whole surface, which is what visually shreds the cat image into what looks like noise).

The unmixing trick: perfect recurrence, not real randomness

Despite looking like noise after a few iterations, the cat map on a torus made of N×N discrete pixels is a bijection with integer arithmetic, and Poincaré’s recurrence theorem guarantees it is periodic: iterate it enough times and the exact original image reappears pixel-for-pixel. The period depends on N in a famously irregular way tied to number theory (related to the order of the matrix modulo N), but it is always finite and always exact — unlike a continuous chaotic system, nothing is lost to rounding, because the discrete map is invertible. This is the cleanest demonstration available that deterministic chaos is not randomness: every bit of information about the initial state survives, just scrambled in a way that looks statistically indistinguishable from noise until it isn’t.

Stable and unstable manifolds

For a generic point, its stable manifold is the line through it in the contracting eigendirection — points on that exact line converge together under forward iteration — and its unstable manifold is the line in the expanding eigendirection, where points diverge going forward but converge going backward in time. Because the eigendirections have irrational slope (governed by √5), these manifolds wind around the torus infinitely many times without ever exactly closing, densely filling the whole surface — the geometric fact underlying the map’s ergodicity.

Frequently asked questions

Try it live

Everything above runs in your browser — open Anosov Map and change the parameters while it is running. Nothing is installed, nothing is uploaded, the whole model lives in one tab.