Lyapunov Exponents: Measuring the Rate of Chaos

What the theorem claims

Let (X, ℬ, μ) be a probability space, T: X → X a measure-preserving transformation, and A: X → GL(d, ℝ) a measurable matrix-valued function (the cocycle generator). Define the products Aⁿ(x) = A(Tⁿ⁻¹x)·⋯·A(Tx)·A(x). The integrability hypothesis is that log⁺‖A(x)‖ ∈ L¹(μ), where log⁺ = max(log, 0).

Oseledets' Multiplicative Ergodic Theorem. Under these hypotheses there is a T-invariant set of full μ-measure on which the limit

Λ(x) = limₙ→∞ (Aⁿ(x)ᵀ Aⁿ(x))^{1/(2n)}

exists as a positive-semidefinite symmetric matrix. Its logarithms of eigenvalues are the Lyapunov exponents λ₁ > λ₂ > ⋯ > λ_r(x) with multiplicities. For μ-a.e. x and every nonzero v in the corresponding subspace, limₙ (1/n) log ‖Aⁿ(x)v‖ = λᵢ(x). When T is ergodic the λᵢ and their multiplicities are constant a.e. If A takes values in GL(d,ℝ) (invertible, two-sided), the tangent space splits as ℝ^d = ⊕ᵢ Eᵢ(x) into an Oseledets splitting with A(x)Eᵢ(x) = Eᵢ(Tx).

The picture: a multiplicative Jordan form

For a single matrix A, the singular values σ₁ ≥ ⋯ ≥ σ_d control how A stretches the unit sphere into an ellipsoid: the i-th semi-axis has length σᵢ. Iterating A gives semi-axes σᵢⁿ, and (1/n) log σᵢ(Aⁿ) → log σᵢ. Lyapunov exponents are exactly this, but for a varying product where the matrix you apply depends on where the orbit currently sits.

The miracle is that even though the matrices A(Tˣ) don't commute and their eigenvalues are meaningless for the product, the rates stabilize. Picture an infinitesimal ball of initial conditions carried along the orbit: it deforms into an ellipsoid whose axes grow (or shrink) like e^{nλᵢ}. The Oseledets splitting names the directions Eᵢ(x): push a vector along Eᵢ and it grows precisely at rate λᵢ. This is a genuine, multiplicative, orbit-dependent analogue of diagonalizing a matrix — a Jordan/Lyapunov normal form for random and chaotic products.

Key idea of the proof

The engine is subadditivity. The sequence gₙ(x) = log ‖Aⁿ(x)‖ is subadditive along the orbit: because Aⁿ⁺ᵐ(x) = Aᵐ(Tⁿx)·Aⁿ(x) and ‖BC‖ ≤ ‖B‖‖C‖, we get gₙ₊ₘ(x) ≤ gₙ(x) + gₘ(Tⁿx). Kingman's Subadditive Ergodic Theorem (1968) then guarantees that gₙ(x)/n converges μ-a.e. to a limit λ₁(x) — the top exponent — provided g₁⁺ ∈ L¹, which is exactly the integrability hypothesis. Furstenberg and Kesten had proven this top-exponent statement for i.i.d. random matrices in 1960.

Getting the full spectrum and the splitting requires more: apply subadditivity to exterior powers Λᵏ A, whose top exponent is λ₁+⋯+λₖ, recovering each λᵢ by differences. The invariant subspaces come from a polar/QR (Gram–Schmidt) decomposition of the cocycle: writing Aⁿ = Qₙ Rₙ and passing to the limit of Rₙ^{1/n} produces the limiting matrix Λ(x); its eigenspaces, pulled back consistently along the orbit (using the invertible, two-sided dynamics), assemble into the Oseledets splitting. Raghunathan and Ruelle gave streamlined proofs in the 1970s–80s.

Worked example: Arnold's cat map

Take the linear Anosov map on the torus 𝕋² = ℝ²/ℤ² given by the matrix A = [[2,1],[1,1]], which is in SL(2,ℤ) (determinant 1, so area-preserving). Here the cocycle is constant — A(x) = A for all x — so Aⁿ(x) = Aⁿ and the Lyapunov exponents are just log|eigenvalue|.

The eigenvalues of A solve μ² − 3μ + 1 = 0, giving μ± = (3 ± √5)/2. Thus

λ₁ = log((3+√5)/2) ≈ 0.9624,   λ₂ = log((3−√5)/2) = −λ₁ ≈ −0.9624.

The sum λ₁+λ₂ = log(det A) = 0, confirming area preservation. The Oseledets subspaces E₁, E₂ are the eigenlines of A (irrational slopes (√5−1)/2 ≈ 0.618 and −(√5+1)/2 ≈ −1.618), and they coincide with the stable/unstable manifolds of the Anosov structure. Along E₁ lengths multiply by e^{λ₁} ≈ 2.618 each step (a stretch), along E₂ they contract by the reciprocal. Because λ₁ > 0, the cat map is chaotic — a textbook hyperbolic system whose mixing you can watch scramble any image on the torus.

Why the hypotheses matter

Integrability (log⁺‖A‖ ∈ L¹) is essential. Without it the Cesàro averages (1/n) log‖Aⁿ‖ can fail to converge or run to +∞; Kingman's theorem needs g₁⁺ integrable. If also log⁺‖A⁻¹‖ ∈ L¹ you additionally get the two-sided splitting rather than just a decreasing filtration V₁ ⊃ V₂ ⊃ ⋯; drop the invertibility/two-sided assumption and the clean direct-sum decomposition degrades to a flag.

Almost-everywhere, not everywhere. Exponents may genuinely differ on a measure-zero set — think of the unstable periodic orbits threaded through a chaotic attractor. And ergodicity is what makes the spectrum a set of constants rather than a measurable function of x. Regularity of the point matters too: Oseledets guarantees the limit only at Lyapunov-regular points, and Pesin theory builds on exactly these. Connections: this is the foundation of Pesin's entropy formula (h_μ(f) = ∫ Σ λᵢ⁺ dμ for smooth systems), of the Margulis–Ruelle inequality, and of stable-manifold theory in nonuniform hyperbolicity.

Applications and significance

Lyapunov exponents are the quantitative bedrock of chaos. A single positive exponent λ₁ > 0 with a bounded attractor is the standard definition of a chaotic (sensitive-dependence) system, and 1/λ₁ is the Lyapunov time — for the inner Solar System roughly 5 million years, which is why planetary orbits are ultimately unpredictable despite Newton's exact laws. The full spectrum feeds the Kaplan–Yorke (Lyapunov) dimension D_L = k + (λ₁+⋯+λ_k)/|λ_{k+1}|, an estimate of an attractor's fractal dimension.

Beyond dynamics, the theorem governs products of random matrices (Furstenberg's positivity theorem, wireless-channel and disordered-media models where λ₁ is the Lyapunov exponent of Anderson localization), and it underlies numerical algorithms — the QR/Gram–Schmidt Benettin method computes the whole spectrum by periodically re-orthonormalizing an evolving frame. In smooth ergodic theory it powers Pesin theory, entropy formulas, and rigidity results. Wherever a system multiplies non-commuting linear maps along an orbit, Oseledets tells you the exponential rates exist and how to find them.

The Lyapunov spectrum of canonical systems, and what the largest exponent λ₁ tells you

System	Lyapunov spectrum	Sign of λ₁	Dynamical meaning
Fixed point of contraction x↦x/2	λ = −log 2 < 0	negative	asymptotically stable; trajectories converge
Doubling map x↦2x mod 1	λ = log 2 ≈ 0.693	positive	chaotic; ~1 bit of precision lost per step
Cat map (Arnold), [[2,1],[1,1]]	λ₁ = log((3+√5)/2), λ₂ = −λ₁	positive	hyperbolic (Anosov); one expanding, one contracting direction
Rotation of the circle x↦x+α	λ = 0	zero	no divergence; predictable, non-chaotic
Lorenz attractor (classic params)	λ₁≈0.906, λ₂=0, λ₃≈−14.57	positive	strange attractor; λ₂=0 along the flow, sum<0 (dissipative)