roseking's picture
Add lean source and binary files (part 124)
a89d282 verified
Raw History Blame Contribute Delete
2.1 kB
import OAI.Probability.InvariantIsing.Core.Ward
namespace OAI
/-! Endpoint Gibbs averages bound a finite log-partition difference. These
are the convex secant inequalities used to change spectral multiplicities. -/
noncomputable section
open IsingPerceptron
open scoped BigOperators
namespace InvariantIsing
lemma log_finitePartition_sub_ge_gibbsAverage
{S : Type*} [Fintype S] {w : S → ℝ} (hw : GibbsReference w)
(H K : S → ℝ) :
gibbsAverage w H (fun s => K s - H s) ≤
Real.log (finitePartition w K) - Real.log (finitePartition w H) := by
let Y := fun s => K s - H s
have hd := hasDerivAt_log_finitePartition hw (t := (0 : ℝ))
(fun s => (hasDerivAt_const 0 (H s)).add ((hasDerivAt_id 0).mul_const (Y s)))
have hc := convexOn_log_finitePartition hw H Y
have hs := hc.le_slope_of_hasDerivAt (Set.mem_univ (0 : ℝ))
(Set.mem_univ (1 : ℝ)) (by norm_num) hd
simpa only [slope_def_field, sub_zero, div_one, zero_add, zero_mul, one_mul,
add_sub_cancel, Y, gibbsAverage, Pi.add_apply, id_eq, add_zero] using hs
lemma log_finitePartition_sub_le_gibbsAverage
{S : Type*} [Fintype S] {w : S → ℝ} (hw : GibbsReference w)
(H K : S → ℝ) :
Real.log (finitePartition w K) - Real.log (finitePartition w H) ≤
gibbsAverage w K (fun s => K s - H s) := by
have h := log_finitePartition_sub_ge_gibbsAverage hw K H
have he : gibbsAverage w K (fun s => H s - K s) =
-gibbsAverage w K (fun s => K s - H s) := by
simp only [gibbsAverage, mul_sub, Finset.sum_sub_distrib]
ring
rw [he] at h
linarith
lemma log_finitePartition_secant_bounds
{S : Type*} [Fintype S] {w : S → ℝ} (hw : GibbsReference w)
(H K : S → ℝ) :
gibbsAverage w H (fun s => K s - H s) ≤
Real.log (finitePartition w K) - Real.log (finitePartition w H) ∧
Real.log (finitePartition w K) - Real.log (finitePartition w H) ≤
gibbsAverage w K (fun s => K s - H s) :=
⟨log_finitePartition_sub_ge_gibbsAverage hw H K,
log_finitePartition_sub_le_gibbsAverage hw H K⟩
end InvariantIsing
end
end OAI