Chain recursion for the fast-band-mass envelope #
Deterministic engine for the uniform-in-m fast-band-mass envelope appearing in Ruelle's
proof of the multiplicative ergodic theorem (the proof of Lemma 1.4 in [Ruelle, Ergodic
theory of differentiable dynamical systems]): the slow/fast orthogonal decomposition over
an SVDData, the one-step band-leakage recursion (oneStep_recursion, an application of
oneStep_sandwich), the contraction-free chain solver (chain_geometric_sum), and the
band↔SVD adapter identifying bandProjector with the explicit fast projection
(toEuclideanLin_bandProjector_eq_fastProj), plus band-mass monotonicity in the cut and
the lower-bound lemma bandProjector_mass_ge_abs_inner_of_fix.
NOTE: the lower-bound lemma shows the envelope at an arbitrary cut
c₀ ∈ (exp λ_a, exp λ_e) is unprovable when an intermediate stratum
λ_a < λ_c < λ_e lies above the cut (the band then contains the λ_c Oseledets
direction, whose overlap with u_a(n) decays only at rate λ_c − λ_a). The sound
statement restricts the cut to the top gap below λ_e; the engine here closes that
corrected envelope.
Main results #
Oseledets.RuelleCofactor.SVDData.oneStep_recursion: the deterministic one-step band-leakage recursion for the fast-band mass along an SVD chain.Oseledets.RuelleCofactor.SVDData.chain_geometric_sum: the contraction-free chain solver — a recursiona (i+1) ≤ a i + R·ρ^iwitha 0 = 0is bounded byR/(1−ρ)uniformly in the index.Oseledets.toEuclideanLin_bandProjector_eq_fastProj: the band projector equals the explicit fast projection over the SVD chainOseledets.chainSVD.Oseledets.norm_bandProjector_mono: band mass is monotone in the cut.Oseledets.bandProjector_mass_ge_abs_inner_of_fix: a uniform band-mass bound dominates the overlap with any unit vector fixed by the limit projector (the obstruction lemma for the arbitrary-cut envelope).
Implementation notes #
Fix x in the a.e.-good set. Write
σ_j(t) = (toEuclideanLin (cocycle A T t x)).singularValues j (antitone in j),
u_a(n) = sortedGramEigenbasis A T n x ⟨a⟩ (a unit right-singular vector of
cocycle A T n x), λ_j = lam0 j.
We are given a gap pair λ_a < λ_e, a cut c₀ with exp λ_a < c₀ < exp λ_e, and we
must bound, uniformly in m ≥ n,
‖P^{>c₀}_m u_a(n)‖ ≤ C·exp(−n(λ_e − λ_a − δ)),
where P^{>c₀}_m = bandProjector A T (indicator (Ioi c₀) 1) m x is the orthogonal
projector onto the span of the time-m Gram eigenvectors u_j(m) whose exp-scale
eigenvalue σ_j(m)^{1/m} exceeds c₀.
The deterministic SVD chain (Oseledets.RuelleCofactor.SVDData) #
Instantiate Oseledets.RuelleCofactor.SVDData (EuclideanSpace ℝ (Fin d)) (card (Fin d)) at the
point x by:
e t := sortedGramEigenbasis A T t x— the time-tGram eigenbasis (right-singular basis);σ t j := σ_j(t)— the time-tsingular values;apply t u := toEuclideanLin (cocycle A T t x) u. The Parseval field isnorm_sq_cocycle_apply_eq_sum_singularValues(withreal_inner_commto flip the inner-product order to the⟪e j, u⟫convention ofOseledets.Lyapunov.RuelleCore).
For this S:
S.fastProj m hi u = Σ_{j∈hi} ⟪e m j, u⟫ • e m jand, whenhi = hiBand m := {j : c₀ < σ_j(m)^{1/m}},‖S.fastProj m hi u‖ = ‖toEuclideanLin (P^{>c₀}_m) u‖(the band projector is the orthogonal projection onto exactly this span — proved intoEuclideanLin_bandProjector_eq_fastProj).
The recursion #
u := u_a(n) lies in the time-n slow span loBand n := {j : σ_j(n)^{1/n} ≤ c₀} (it
equals the single basis vector u_a(n), and eventually σ_a(n)^{1/n} < c₀ so
a ∈ loBand n). For each step t = n+k → t+1:
- the slow cap at time
tfor the slow span iss_t := c₀^t(every slow eigenvalueσ_j(t) ≤ c₀^t); - the fast floor at time
t+1for the fast band istt_{t+1} := c₀^{t+1}; - the one-step bound is
‖A⁽ᵗ⁺¹⁾u‖ ≤ b_t ‖A⁽ᵗ⁾u‖withb_t := ‖A(Tᵗx)‖.oneStep_sandwichthen givesc₀^{t+1}·‖fastProj(t+1) u‖ ≤ b_t · c₀^t · ‖u‖i.e.‖fastProj(t+1) u‖ ≤ (b_t/c₀)·‖u‖. This is too lossy on its own (theb_tare only tempered,b_t ≤ exp(tη)). Ruelle's improvement: the slow part ofuat timet— not all ofu— feeds the fast band att+1, and the slow part's mass is whatfastProj(t) ualready controls. The correct one-step recursion (his displayed computation) is, witha_k := ‖fastProj(n+k) u‖:a_{k+1} ≤ exp(−γ̄)·a_k + R·exp(−k γ'), whereγ̄ = λ_e − λ_a − δ*is the per-step gap survival andR·exp(−kγ')the freshly-injected slow leakage.chain_leakage_expsolves this:a_k ≤ exp(−kγ̄)·a_0 + R·k·exp(−(k−1)·min γ̄ γ'). Witha_0 = 0(atm = n,u_a(n)is orthogonal to the fast band, since eventuallyσ_a(n)^{1/n} < c₀),a_k ≤ R·k·exp(−(k−1)·min γ̄ γ'), and sincea_kis measured at absolute timem = n+k, this carries theexp(−n γ)prefactor. The polynomialkand theexp(−kγ̄)tail give a constant uniform ink = m − n.
The δ*/stratum-gap and c₀-endpoint subtleties are handled where this engine is
consumed. The deterministic recursion engine itself (oneStep_sandwich,
chain_leakage_exp) is proved in Oseledets.Lyapunov.RuelleCore; this file builds the
band / SVDData adapter on top of it.
References #
- D. Ruelle, Ergodic theory of differentiable dynamical systems, Publ. Math. IHÉS 50 (1979), 27–58 (the proof of Lemma 1.4).
The slow projection onto the complement of hi.
Instances For
u reconstructs as fast + slow.
The slow projection lies in the slow span (the span of e t j, j ∈ hiᶜ).
One-step recursion (deterministic). Fix consecutive times t, t+1 with fast
bands hi t, hi (t+1). Assume the slow cap s at time t (every σ t j ≤ s for
j ∈ (hi t)ᶜ), the fast floor tt > 0 at time t+1 (every σ (t+1) j ≥ tt for
j ∈ hi (t+1)), and the step bound ‖apply (t+1) w‖ ≤ b·‖apply t w‖ for the slow part
w = slowProj t (hi t) u. Then
‖fastProj (t+1) (hi (t+1)) u‖
≤ ‖fastProj t (hi t) u‖ + (b·s/tt)·‖slowProj t (hi t) u‖.
Contraction-free chain solution. If a 0 = 0 and a (i+1) ≤ a i + R·ρ^i for all
i with 0 ≤ R, 0 ≤ ρ < 1, then a k ≤ R/(1−ρ) for every k — uniformly in k.
(Geometric series; no per-step contraction needed because the source already decays.)
The band / SVDData adapter #
bandProjector A T χ m x = cfc χ (qpow A T m x), and qpow = cfc (·^{1/2m}) (gram), so by
CFC composition bandProjector A T χ m x = cfc (χ ∘ (·^{1/2m})) (gram A T m x). Hence it
acts on the
sorted Gram eigenbasis u_j(m) = sortedGramEigenbasis A T m x j diagonally, with eigenvalue
χ (σ_j(m)^{1/m}) (the indicator of the exp-scale band).
The band projector acts diagonally on the sorted Gram eigenbasis with eigenvalue
χ (qpow-eigenvalue).
The SVD chain data at a point x: time-t Gram eigenbasis, singular values, and the cocycle
action. The Parseval field is norm_sq_cocycle_apply_eq_sum_singularValues.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The "fast band" finset at time m: indices whose exp-scale (qpow) eigenvalue exceeds c₀.
Equations
- Oseledets.hiBand A T m x c₀ = {j : Fin (Fintype.card (Fin d)) | c₀ < ⋯.eigenvalues₀ j}
Instances For
The band projector applied to u equals the explicit fast projection
S.fastProj m (hiBand …) u onto the time-m Gram eigenvectors above the cut.
Band monotonicity in the cut. If c₁ ≤ c₀ then the higher-cut band mass is
dominated by the lower-cut band mass: ‖P^{>c₀}_m u‖ ≤ ‖P^{>c₁}_m u‖. (Higher cut ⟹
smaller fast index set ⟹ smaller projection.)
Obstruction: the arbitrary-cut envelope is too strong under intermediate strata #
The deterministic chain (above) shows the fast-band mass ‖P^{>c₀}_m u_a(n)‖ decays at the
rate of the first lam0-stratum strictly above log c₀. An envelope claiming decay at
the rate λ_e − λ_a for an arbitrary cut c₀ ∈ (exp λ_a, exp λ_e) is strictly
stronger. When a third stratum λ_c satisfies λ_a < λ_c < λ_e, a cut
c₀ ∈ (exp λ_a, exp λ_c) produces a band that contains the λ_c Oseledets direction,
so the band mass is ≍ exp(−n(λ_c − λ_a)), which is asymptotically larger than the
claimed C·exp(−n(λ_e − λ_a − δ)) (for δ < λ_e − λ_c). Hence the envelope at that cut
is false.
The lemma below makes the obstruction precise: the uniform-in-m band-mass envelope
dominates the overlap with any unit vector that the limit projector Pinf fixes —
including an intermediate-stratum limit eigenvector b'_c (which Pinf fixes because
exp λ_c > c₀). So the arbitrary-cut envelope at this cut forces
|⟪b'_c, u_a(n)⟫| ≤ C · exp(−n(λ_e − λ_a − δ)) .
But the genuine graded-overlap rate between the adjacent strata λ_a < λ_c is only
λ_c − λ_a (this is exactly the conclusion forward_graded_overlap proves for the pair
(a,c), and the overlap is generically of that exact order — nonzero leading coefficient).
Whenever the adjacent overlap is non-degenerate (the generic case),
|⟪b'_c, u_a(n)⟫| ≍ exp(−n(λ_c − λ_a)), which exceeds C·exp(−n(λ_e − λ_a − δ)) for all
large n (take δ < λ_e − λ_c). Hence the arbitrary-cut envelope is not provable in
general: it is a strictly stronger claim than the band mass actually satisfies at cuts
c₀ ∈ (exp λ_a, exp λ_c) whenever a third stratum λ_c ∈ (λ_a, λ_e) exists.
NOTE. The conclusion of forward_graded_overlap is still correct (the
λ_e-eigenvector overlap |⟪b'_e, u_a(n)⟫| does decay at λ_e − λ_a); only the route
through the arbitrary-cut envelope — bounding that overlap by the band mass at an
arbitrary cut and claiming the band mass also decays at λ_e − λ_a — is unsound when
intermediate strata are present. A sound envelope must restrict the cut to lie above the
immediately preceding stratum below λ_e (i.e. c₀ ∈ (exp λ_{e−1}, exp λ_e)), or the
consumer must transfer to the band at the cut whose first stratum above is exactly
λ_e.