Sturdiness by conditioning: defect zero at depth for of mass
ProvedKServer.chunk_restrict_sturdySturdiness by conditioning.
Let be a chunk system with online escapes on a metric space between the marked points , with size floor , size ceiling , escape price and expected total mass at least , in the sense of KServer.ChunkSystemB. Recall that is -sturdy at depth with defect (KServer.ChunkSystemB.SturdyL1) when the Doob martingale of the total mass satisfies
i.e. the conditional expected total mass never drops appreciably below its mean during the first steps of the filtration.
Statement. For every depth there is a chunk system with the same size floor, size ceiling, escape price and chunk-count bound, with the same number of chunks, whose expected total mass is at least
which is -sturdy at depth with defect , whose initial information is constant, and which has no empty chunk whenever has none.
Content. Sturdiness with defect at depth says that the Doob martingale of the total mass is constant up to time ; the statement is that this can always be arranged, for a price of chunk masses. Condition on the time- atom on which the conditional expected total mass is largest — by the tower property that conditional expectation is at least the unconditional mean — and forget the mass of the first chunks, whose sizes are already known on and total at most there. On the conditioned system the filtration is trivial up to time , so the Doob martingale is constant up to time and the defect vanishes.
Conditioning is legitimate because the cost axiom of a chunk system is stated atom by atom: for a chunk index the time- atoms of the conditioned system are exactly the time- atoms of the original that lie inside , and both sides of the axiom are simply rescaled by the mass of ; for the size has been set to and the axiom is the nonnegativity of the escape-augmented evader cost.
Role. In the induction of Bubeck--Coester--Rabani's Lemma 12 the level step KServer.level_step_sturdy consumes a sturdiness defect at a depth at least the number of race steps and charges against the mass it produces. The statement above quantifies the price of manufacturing that hypothesis from nothing: of mass for depth , i.e. one chunk mass per unit of depth, to be compared with the anti-concentration gain of a -step race, which is of order chunk masses. So conditioning alone cannot supply the sturdiness needed by a long race, and a regularity invariant has to be carried through the construction instead — this is why the level invariant of the induction has to record more than a two-sided size window.
import Mathlib import Definitions.Def_KServer_model import Definitions.Def_KServer_evader import Definitions.Def_KServer_evader_bail import Definitions.Def_KServer_chunk_system_b import Definitions.Def_KServer_chunk_cond import Definitions.Def_KServer_chunk_adjust import Definitions.Def_KServer_sturdy
namespace KServer
/-- **Sturdiness by conditioning.** Every chunk system with size floor zero and
nonnegative escape price can be turned, for any depth `n`, into a chunk system
that is `L¹`-sturdy at depth `n` with defect `0` — its Doob martingale of the
total mass is constant up to time `n` — at the cost of at most `n · cB` of
expected total mass. One conditions on the time-`n` atom on which the
conditional expected total is largest and forgets the mass of the first `n`
chunks. -/
theorem chunk_restrict_sturdy {X : Type*} [MetricSpace X] {s t : X}
{cB T pe : ℝ} {mL : ℕ}
(C : ChunkSystemB X s t 0 cB T pe mL)
(hpe : 0 ≤ pe) (hcB : 0 ≤ cB) (n : ℕ) :
∃ C' : ChunkSystemB X s t 0 cB (T - n * cB) pe mL,
C'.m = C.m ∧
C'.SturdyL1 n 0 ∧
(∀ ω₁ ω₂ : C'.Ω, C'.hist 0 ω₁ = C'.hist 0 ω₂) ∧
((∀ (ω : C.Ω) (i : Fin C.m), C.chunk ω i ≠ []) →
∀ (ω : C'.Ω) (i : Fin C'.m), C'.chunk ω i ≠ []) := by sorry
end KServer