File size: 7,946 Bytes
dd14c26 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 | import OAI.Probability.MatroidProphet.Main
namespace OAI
/-!
# Source-exact accounting contracts
The conditioning in `lem:all-orders` fixes every test bit outside the focal
listed group. Averaging only the remaining group's bits is stronger than
merely averaging all test bits. We make that finite conditional expectation
explicit below, without conditioning on any arrival order.
-/
namespace MatroidProphet.AccountingIndependent
open Set Finset Pivots
variable {α : Type*} [Fintype α] [DecidableEq α]
/-- The manuscript's test-thinning inequality for each fixed outside-mask
realization. `A` is exactly the fixed positive test set outside `G h`; only
`B ⊆ G h` is averaged. In particular, no selected arrival order is conditioned
on, and the competition set may depend on all outside bits. -/
theorem thinning_finalRankStatistic_conditioned
(M : Matroid α) (hE : M.E = Set.univ)
(κ : ℕ) (D C : ℕ → Set α) (G : ℕ → Finset α)
(hG : Pairwise (fun i j => Disjoint (G i) (G j)))
(h final : ℕ) (O : Set α) (ε : Fin 2) (t : ℝ)
(ht0 : 0 ≤ t) (ht1 : t ≤ 1)
(A : Finset α) (hA : Disjoint A (G h)) :
t * (finalRankStatistic M hE κ D C
(fun j => (A : Set α) ∩ (G j : Set α)) h final (G h) O Set.univ ε : ℝ) ≤
bitsExpectation (fun _ : α => t) (G h) (fun B =>
(finalRankStatistic M hE κ D C
(fun j => ((A ∪ B : Finset α) : Set α) ∩ (G j : Set α))
h final (G h) O ((A ∪ B : Finset α) : Set α) ε : ℝ)) := by
classical
simp only [finalRankStatistic, Set.inter_univ, Nat.cast_sum,
bitsExpectation_sum, Finset.mul_sum]
apply Finset.sum_le_sum
intro b hb
let S := nominalLayerSet M hE κ D C h (G h) ε b \ O
have hS : S ⊆ (G h : Set α) := fun e he =>
nominalLayerSet_subset M hE κ D C h (G h) ε b he.1
have hdis : S ∩ (A : Set α) = ∅ := by
apply Set.eq_empty_iff_forall_notMem.mpr
rintro e ⟨heS, heA⟩
exact Finset.disjoint_left.mp hA heA (hS heS)
have ht := independent_thinning_rank M hE (G h) S.toFinset
(fun e he => hS (by simpa only [Set.mem_toFinset] using he))
(guardedPath M hE κ D C final (b.val.val - 2) ∪
lowerCompetition M hE κ D C
(fun j => (A : Set α) ∩ (G j : Set α)) b.val.val h) t ht0 ht1
calc
_ ≤ bitsExpectation (fun _ : α => t) (G h) (fun B =>
(conditionalRank M (S ∩ (B : Set α))
(guardedPath M hE κ D C final (b.val.val - 2) ∪
lowerCompetition M hE κ D C
(fun j => (A : Set α) ∩ (G j : Set α)) b.val.val h) : ℝ)) := by
simpa only [Finset.coe_inter, Set.coe_toFinset] using ht
_ = _ := by
apply bitsExpectation_congr
intro B hB
rw [lowerCompetition_ignore_focal M hE κ D C G hG b.val.val h A B hB]
have heq : S ∩ ((A ∪ B : Finset α) : Set α) = S ∩ (B : Set α) := by
rw [Finset.coe_union, Set.inter_union_distrib_left, hdis, Set.empty_union]
change (conditionalRank M (S ∩ (B : Set α)) _ : ℝ) =
(conditionalRank M (S ∩ ((A ∪ B : Finset α) : Set α)) _ : ℝ)
rw [heq]
/-- The unscaled real-valued form of `lem:rank-cost`. There is a single
sum of later-group costs for the entire parity timeline, not one cost per
layer. The density contribution is real division, never truncated division. -/
theorem rank_cost_real
(M : Matroid α) (hE : M.E = Set.univ)
(κ : ℕ) (hκ : 0 < κ) (D C T : ℕ → Set α) (h later : ℕ)
(U O : Set α) (ε : Fin 2) :
(nominalRankStatistic M hE κ D C T h U O ε : ℝ) -
((C h).ncard + ∑ j ∈ Finset.range later,
((C (h+1+j)).ncard + (D (h+1+j)).ncard / (κ : ℝ))) ≤
(finalRankStatistic M hE κ D C T h (h+1+later) U O Set.univ ε : ℝ) := by
have hn := testRankStatistic_le_final_cost M hE κ hκ D C T h later U O Set.univ ε
have heq : testRankStatistic M hE κ D C T h U O Set.univ ε =
nominalRankStatistic M hE κ D C T h U O ε := by
simp only [testRankStatistic, nominalRankStatistic, Set.inter_univ]
rw [heq] at hn
have hr : (κ : ℝ) * (nominalRankStatistic M hE κ D C T h U O ε : ℝ) ≤
(κ : ℝ) * (finalRankStatistic M hE κ D C T h (h+1+later) U O Set.univ ε : ℝ) +
(κ : ℝ) * (C h).ncard +
∑ j ∈ Finset.range later,
((κ : ℝ) * (C (h+1+j)).ncard + (D (h+1+j)).ncard) := by
exact_mod_cast hn
have hkpos : (0 : ℝ) < κ := Nat.cast_pos.mpr hκ
have hcost : (κ : ℝ) *
((C h).ncard + ∑ j ∈ Finset.range later,
((C (h+1+j)).ncard + (D (h+1+j)).ncard / (κ : ℝ))) =
(κ : ℝ) * (C h).ncard + ∑ j ∈ Finset.range later,
((κ : ℝ) * (C (h+1+j)).ncard + (D (h+1+j)).ncard) := by
rw [mul_add, Finset.mul_sum]
congr 1
apply Finset.sum_congr rfl
intro j hj
rw [mul_add, mul_div_cancel₀ _ (ne_of_gt hkpos)]
nlinarith
open MainAlgorithm
/-- Both assertions of source `lem:all-orders`, instantiated to the actual
listed groups. The second assertion is the finite conditional law obtained by
fixing all of `H,D,C`, parity, and every test bit outside the listed group. -/
theorem all_orders {n : ℕ} (M : Matroid (Fin n)) (hE : M.E = Set.univ)
(d : MainMasks n) (w : Fin n → Option ℤ) (h : ℕ) :
(∀ π : ArrivalOrder n, listedLambda M hE d w h ≤
((selection M hE d w π).filter (fun e =>
∃ i, w e = some i ∧ h ≤ (groups M d w).idxOf i)).card) ∧
(∀ A : Finset (Fin n), Disjoint A (pathGroups M d w h) →
thinningRate * (finalRankStatistic M hE (2^100)
(groupMask M d w d.D) (groupMask M d w d.C)
(fun j => (A : Set (Fin n)) ∩ (pathGroups M d w j : Set (Fin n)))
h (groups M d w).length (pathGroups M d w h)
((d.H ∪ d.D ∪ d.C : Finset (Fin n)) : Set (Fin n)) Set.univ
(boolParity d.odd) : ℝ) ≤
bitsExpectation (fun _ : Fin n => thinningRate) (pathGroups M d w h)
(fun B => (listedLambda M hE {d with T := A ∪ B} w h : ℝ))) := by
constructor
· intro π
exact listedLambda_le_count M hE d w π h
· intro A hA
have ht := thinning_finalRankStatistic_conditioned M hE (2^100)
(groupMask M d w d.D) (groupMask M d w d.C) (pathGroups M d w)
(pathGroups_disjoint M d w) h (groups M d w).length
((d.H ∪ d.D ∪ d.C : Finset (Fin n)) : Set (Fin n)) (boolParity d.odd)
thinningRate constants_positive.2.2.1.le (by norm_num [thinningRate]) A hA
simpa only [listedLambda_update_T] using ht
/-- Source `lem:rank-cost` for the concrete listed group and final guarded
path, with the displayed real-valued cost and without any probabilistic or
independence assumption on its common generators. -/
theorem listed_rank_cost {n : ℕ} (M : Matroid (Fin n)) (hE : M.E = Set.univ)
(d : MainMasks n) (w : Fin n → Option ℤ) (h : ℕ)
(hh : h < (groups M d w).length) :
(listedZ M hE d w h : ℝ) -
((groupMask M d w d.C h).ncard +
∑ j ∈ Finset.range ((groups M d w).length - (h+1)),
((groupMask M d w d.C (h+1+j)).ncard +
(groupMask M d w d.D (h+1+j)).ncard / densityThreshold)) ≤
(finalRankStatistic M hE (2^100)
(groupMask M d w d.D) (groupMask M d w d.C) (groupMask M d w d.T)
h (groups M d w).length (groupMask M d w Finset.univ h)
((d.H ∪ d.D ∪ d.C : Finset (Fin n)) : Set (Fin n)) Set.univ
(boolParity d.odd) : ℝ) := by
have hc := rank_cost_real M hE (2^100) (by positivity)
(groupMask M d w d.D) (groupMask M d w d.C) (groupMask M d w d.T)
h ((groups M d w).length - (h+1)) (groupMask M d w Finset.univ h)
((d.H ∪ d.D ∪ d.C : Finset (Fin n)) : Set (Fin n)) (boolParity d.odd)
have hlen : h+1+((groups M d w).length - (h+1)) = (groups M d w).length := by omega
rw [hlen] at hc
simpa only [listedZ, densityThreshold, Nat.cast_pow, Nat.cast_ofNat] using hc
end MatroidProphet.AccountingIndependent
end OAI
|