Documentation

Oseledets.Lyapunov.ForwardGradedOverlap

The n-scaled forward graded overlap bound #

This module proves forward_graded_overlap: almost everywhere, for every δ > 0 there is a constant c ≥ 1 such that, eventually in n, for all sorted-eigenbasis indices a and limit-eigenbasis indices e,

|⟪b' x e, u_a(n)⟫| ≤ c · exp(−n·(max(λ_e − λ_a, 0) − δ)),

where u_a(n) = sortedGramEigenbasis A T n x a and b' x e is the limit-eigenbasis vector at eigenvalue exp(λ_e). This is the a.e. form of the graded-overlap estimate underlying the forward Oseledets filtration.

Main results #

Implementation notes #

Fix x in the full-measure set on which all the hypotheses hold. Write λ_a = lam0 a and u_a(n) = sortedGramEigenbasis A T n x a.

For a trivial pair λ_e ≤ λ_a we have max(λ_e − λ_a, 0) = 0, so the right-hand side is c · exp(nδ) ≥ 1 ≥ |⟪b'_e, u_a(n)⟫| by Cauchy–Schwarz (both are unit vectors); the bound holds for all n with c = 1. The content is the gap pairs λ_e > λ_a, where the time-n slow vector u_a(n) has overlap with the limit fast band decaying like exp(−n(λ_e − λ_a − δ)). There the proof proceeds in three steps:

  1. Gap cut. The finitely many values {exp(λ_j) : j < d} form a finite set, so one may choose a cut c₀ with exp(λ_a) < c₀ < exp(λ_e) avoiding all of them; hident applies at this cut.
  2. Two-time chain. The time-n slow vector has fast-band mass at the time-m cut decaying like exp(−n(λ_e − λ_a − δ)), uniformly in m ≥ n. This is the analytic content packaged here as the hypothesis hchain (see below).
  3. Limit transfer. The limit-eigenbasis vector b'_e at level exp(λ_e) > c₀ is fixed by the limit band projector Pinf = cfc (indicator (Ioi c₀)) (lambdaHat A T x). Self-adjointness of Pinf and convergence of the time-m band projectors to Pinf then bound |⟪b'_e, u_a(n)⟫| by the step-2 mass bound, with no rate needed for the vanishing term.

The finitely many pairs are combined via eventually_all, and c is the maximum of the step-2 constants over the gap pairs (independent of n, a, e).

The two-time chain envelope is supplied as the hypothesis hchain: for a gap pair λ_a < λ_e and any cut c₀ strictly between the two exp-levels, the fast-band mass of the time-n slow eigenvector, measured by the time-m band projector at c₀, decays like exp(−n(λ_e − λ_a − δ)) uniformly in m ≥ n. A single deterministic operator-norm step over [n, m] is too lossy (it gives exp((m − n)λ₀ − mλ_e), which diverges as m → ∞ in the gap direction), so the per-step recursion that keeps the vector in the slow cone is genuinely required.

References #

Deterministic helper lemmas #

theorem Oseledets.toEuclideanLin_cfc_fix_eigenvector {d : } [NeZero d] (M : Matrix (Fin d) (Fin d) ) (hM : M.IsHermitian) (f : ) (v : EuclideanSpace (Fin d)) {lam : } (hev : (Matrix.toEuclideanLin M) v = lam v) (hf : f lam = 1) :

CFC fixes an arbitrary eigenvector at an eigenvalue where f = 1. If M is Hermitian, toEuclideanLin M v = lam • v, and f lam = 1, then toEuclideanLin (cfc f M) v = v. Proof: expand v in the (orthonormal) eigenvector basis; the component ⟪b_j, v⟫ is nonzero only for eigenvalues eigenvalues j = lam (eigenvectors at distinct eigenvalues are orthogonal because M is self-adjoint), and there f (eigenvalues j) = f lam = 1.

theorem Oseledets.tendsto_toEuclideanLin_apply {d : } {P : Matrix (Fin d) (Fin d) } {Pinf : Matrix (Fin d) (Fin d) } (hP : Filter.Tendsto P Filter.atTop (nhds Pinf)) (u : EuclideanSpace (Fin d)) :

Continuity of matrix→vector application. If P_m → Pinf (in the finite-dimensional matrix norm) then toEuclideanLin (P_m) u → toEuclideanLin Pinf u.

theorem Oseledets.abs_inner_le_of_bandProjector_mass_bound {d : } {P : Matrix (Fin d) (Fin d) } {Pinf : Matrix (Fin d) (Fin d) } (hP : Filter.Tendsto P Filter.atTop (nhds Pinf)) (hPinfsa : IsSelfAdjoint Pinf) (w u : EuclideanSpace (Fin d)) (hwnorm : w 1) (hfix : (Matrix.toEuclideanLin Pinf) w = w) {B : } (hbound : ∀ᶠ (m : ) in Filter.atTop, (Matrix.toEuclideanLin (P m)) u B) :

Limit-transfer reduction (Ruelle Lemma 1.4, step 3). Let w be a unit vector that is fixed by the limit band projector Pinf (toEuclideanLin Pinf w = w), with Pinf self-adjoint, and suppose the time-m band projectors P m converge to Pinf. If the fast-band mass ‖toEuclideanLin (P m) u‖ is eventually bounded by B, then |⟪w, u⟫| ≤ B. (No rate is needed for the vanishing of Pinf − P m.)

theorem Oseledets.exists_spectral_cut {d : } (g : Fin d) {lo hi : } (hlohi : lo < hi) :
∃ (c₀ : ), Real.exp lo < c₀ c₀ < Real.exp hi ∀ (i : Fin d), Real.exp (g i) c₀

Gap-cut selection (Ruelle Lemma 1.4, step 1). For two levels lo < hi, the open interval (exp lo, exp hi) is infinite, so it contains a point c₀ avoiding the finite spectrum {exp (g i) : i}. This supplies a spectral cut strictly between the two exp-levels at which the band projector and hident may be evaluated.

theorem Oseledets.forward_graded_overlap {X : Type u_1} [MeasurableSpace X] {μ : MeasureTheory.Measure X} {d : } {T : XX} [MeasureTheory.IsProbabilityMeasure μ] [NeZero d] (_hT : Ergodic T μ) {A : XMatrix (Fin d) (Fin d) } (_hA : ∀ (x : X), (A x).det 0) (_hAmeas : Measurable A) (_hint : IntegrableLogNorm A μ) (_hint' : IntegrableLogNorm (fun (x : X) => (A x)⁻¹) μ) (lam0 : ) (hlam0 : i < d, ∀ᵐ (x : X) μ, Filter.Tendsto (fun (n : ) => (↑n)⁻¹ * Real.log ((Matrix.toEuclideanLin (cocycle A T n x)).singularValues i)) Filter.atTop (nhds (lam0 i))) (b' : XOrthonormalBasis (Fin d) (EuclideanSpace (Fin d))) (hb' : ∀ᵐ (x : X) μ, ∀ (e : Fin d), (Matrix.toEuclideanLin (lambdaHat A T x)) ((b' x) e) = Real.exp (lamSing A T x e) (b' x) e) (hident : ∀ᵐ (x : X) μ, ∀ (c : ), 0 < c(∀ (i : Fin d), Real.exp (lamSing A T x i) c)Filter.Tendsto (fun (n : ) => bandProjector A T ((Set.Ioi c).indicator 1) n x) Filter.atTop (nhds (cfc ((Set.Ioi c).indicator 1) (lambdaHat A T x)))) (hchain : ∀ᵐ (x : X) μ, ∀ (δ : ), 0 < δ∃ (C : ), 1 C ∀ (a e : Fin d), lam0 a < lam0 e∀ (c₀ : ), Real.exp (lam0 a) < c₀c₀ < Real.exp (lam0 e)∀ᶠ (n : ) in Filter.atTop, ∀ (m : ), n m(Matrix.toEuclideanLin (bandProjector A T ((Set.Ioi c₀).indicator 1) m x)) ((sortedGramEigenbasis A T n x) a, ) C * Real.exp (-n * (lam0 e - lam0 a - δ))) :
∀ᵐ (x : X) μ, ∀ (δ : ), 0 < δ∃ (c : ), 1 c ∀ᶠ (n : ) in Filter.atTop, ∀ (a e : Fin d), |inner ((b' x) e) ((sortedGramEigenbasis A T n x) a, )| c * Real.exp (-n * (max (lam0 e - lam0 a) 0 - δ))