The scalar estimates behind Birkhoff's contraction theorem #
This file collects the scalar inequalities behind Birkhoff's contraction theorem for Hilbert's
projective metric (Matrix.hilbertProjectiveDist_mulVec_le in
TauCeti/Data/Matrix/BirkhoffContraction.lean). They contain no matrices.
For l ≥ 1, the function σ ↦ log ((1 + l * exp σ) / (l + exp σ)) vanishes at σ = 0 and has
derivative l * exp σ / (1 + l * exp σ) - exp σ / (l + exp σ), which is at most
(l - 1) / (l + 1) (with equality at σ = 0). It is therefore bounded by (l - 1) / (l + 1) * σ
for σ ≥ 0. Writing l = exp (Δ / 2), the slope is tanh (Δ / 4).
Combined with a polynomial inequality, this gives the following for R ≥ 1, l ≥ 1 and
positive α, β, α', β'. The logarithmic cross ratio of the pairs (R * α + β, α + β) and
(R * α' + β', α' + β') is at most (l - 1) / (l + 1) * log R whenever the pairs (α, β) and
(α', β') themselves have cross ratio α * β' / (β * α') at most l ^ 2. This is Birkhoff's
theorem in two dimensions.
Main results #
TauCeti.log_one_add_mul_exp_div_add_exp_le:log ((1 + l * exp σ) / (l + exp σ)) ≤ (l - 1) / (l + 1) * σfor1 ≤ land0 ≤ σ.TauCeti.birkhoff_cross_ratio_le: the polynomial cross-ratio bound by((1 + l * s) / (l + s)) ^ 2for1 ≤ sand1 ≤ l.TauCeti.log_birkhoff_cross_ratio_le: the two-dimensional form of Birkhoff's theorem, for1 ≤ Rand1 ≤ l.
References #
- G. Birkhoff, Extensions of Jentzsch's theorem, Trans. Amer. Math. Soc. 85 (1957), 219--227.
- E. Seneta, Non-negative Matrices and Markov Chains, Springer (2006), Chapter 3.
For 1 ≤ l and 0 ≤ σ, log ((1 + l * exp σ) / (l + exp σ)) ≤ (l - 1) / (l + 1) * σ.
Both sides vanish at σ = 0, and the derivative of the left side is at most (l - 1) / (l + 1).
The two-dimensional polynomial inequality behind Birkhoff's theorem. Let 1 ≤ s and
1 ≤ l, and let α, β, α', β' be positive with α * β' ≤ l ^ 2 * (β * α'). Then the cross ratio
of (s ^ 2 * α + β, α + β) and (s ^ 2 * α' + β', α' + β') is at most
((1 + l * s) / (l + s)) ^ 2.
The two-dimensional form of Birkhoff's theorem: for R ≥ 1, l ≥ 1 and positive
α, β, α', β' with α * β' ≤ l ^ 2 * (β * α'), the logarithmic cross ratio of
(R * α + β, α + β) and (R * α' + β', α' + β') is at most (l - 1) / (l + 1) * log R.