math / data /lean /OAI /Combinatorics /ContingencyTables /Transport /CompletionChainFlow.lean
introvoyz045's picture
Migrated from GitHub
4a56cd6 verified
Raw History Blame Contribute Delete
3.35 kB
import OAI.Combinatorics.ContingencyTables.Transport.TranslatedCompletionChain
namespace OAI
/-! # Exact edge flow of the completion-translation chain -/
namespace ContingencyTables.CompletionTranslation
open scoped BigOperators Classical
open CommonBasesFPRAS
variable {X I J : Type*} [Fintype X] [Nonempty X] [Fintype I] [Fintype J]
omit [Nonempty X] in
lemma completionMove_flow (R : X β†’ Option I β†’ β„•) (P : X β†’ Option J β†’ β„•)
(htotal : βˆ€ x,(βˆ‘ i,R x i)=βˆ‘ j,P x j)
(Adj : X β†’ X β†’ Prop) (Ξ² : ℝ) (x y : X) :
((blockCount R P x:ℝ)/MetropolisChain.normalizer (fun z => (blockCount R P z:ℝ)))*
completionMove R P Adj Ξ² x y =
if Adj x y then Ξ²*(acceptedCount R P x y:ℝ)/
MetropolisChain.normalizer (fun z => (blockCount R P z:ℝ)) else 0 := by
have hx : (blockCount R P x:ℝ)β‰ 0 := ne_of_gt (by exact_mod_cast blockCount_pos R P htotal x)
unfold completionMove
split_ifs
Β· exact div_mul_div_cancelβ‚€' hx _ _
Β· exact mul_zero _
/-- The flow formula includes the holding terms: those contribute zero
energy. The numerator counts literal acceptable completion translations. -/
theorem completionChain_energy (R : X β†’ Option I β†’ β„•) (P : X β†’ Option J β†’ β„•)
(htotal : βˆ€ x,(βˆ‘ i,R x i)=βˆ‘ j,P x j)
(Adj : X β†’ X β†’ Prop) (hsym : βˆ€ x y,Adj x y ↔ Adj y x)
(Ξ² : ℝ) (hΞ² : 0 ≀ Ξ²) (D : β„•)
(hdegree : βˆ€ x,(Finset.univ.filter (Adj x)).card ≀ D)
(hproposal : (D:ℝ)*Ξ² ≀ 1/2) (H : X β†’ ℝ) :
(completionChain R P htotal Adj hsym Ξ² hΞ² D hdegree hproposal).energy H=
(βˆ‘ x,βˆ‘ y,if Adj x y then
Ξ²*(acceptedCount R P x y:ℝ)*(H x-H y)^2 else 0)/
(2*MetropolisChain.normalizer (fun z => (blockCount R P z:ℝ))) := by
unfold completionChain
rw [FiniteChain.ofMoves_energy]
simp_rw [completionMove_flow R P htotal Adj Ξ²]
have hterm (x y : X) :
(if Adj x y then Ξ²*(acceptedCount R P x y:ℝ)/
MetropolisChain.normalizer (fun z => (blockCount R P z:ℝ)) else 0)*(H x-H y)^2 =
(if Adj x y then Ξ²*(acceptedCount R P x y:ℝ)*(H x-H y)^2 else 0)/
MetropolisChain.normalizer (fun z => (blockCount R P z:ℝ)) := by
split_ifs <;> ring
simp_rw [hterm,← Finset.sum_div]
ring
omit [Fintype X] [Nonempty X] in
/-- A half-acceptance estimate yields the min-weight comparison on every
actual edge, with no request for completion-count evaluation. -/
lemma completionMove_lower (R : X β†’ Option I β†’ β„•) (P : X β†’ Option J β†’ β„•)
(htotal : βˆ€ x,(βˆ‘ i,R x i)=βˆ‘ j,P x j)
(Adj : X β†’ X β†’ Prop) (Ξ² : ℝ) (hΞ² : 0 ≀ Ξ²) (x y : X)
(hxy : Adj x y)
(haccept : (blockCount R P x:ℝ)/2 ≀ acceptedCount R P x y) :
Ξ²/2*min 1 ((blockCount R P y:ℝ)/blockCount R P x) ≀ completionMove R P Adj Ξ² x y := by
have hx : (0:ℝ)<blockCount R P x := by exact_mod_cast blockCount_pos R P htotal x
have hm : Ξ²/2*min 1 ((blockCount R P y:ℝ)/blockCount R P x) ≀ Ξ²/2 := by
simpa only [mul_one] using mul_le_mul_of_nonneg_left
(min_le_left (1:ℝ) ((blockCount R P y:ℝ)/blockCount R P x)) (by positivity : 0 ≀ Ξ²/2)
apply hm.trans
unfold completionMove
rw [ite_eq_left hxy]
apply (le_div_iffβ‚€ hx).mpr
nlinarith [mul_nonneg hΞ² (sub_nonneg.mpr haccept)]
end ContingencyTables.CompletionTranslation
end OAI