Theorem 1.1: and
ProvedRGMISFast.theorem_1_1There is an absolute constant with the following property. Let be a finite simple undirected graph with maximum degree at most , with independent Exp(1) ranks. For every vertex fixed before the ranks are sampled and every , the number of fresh calls of the memoized, pruned Procedure 1.1 started at satisfies
and consequently
This answers the question of Nguyen and Onak on the local query complexity of randomized greedy MIS from a worst-case vertex: polynomial in , independent of the size of the graph.
Formalization Note is chosen before the vertex type, the graph, , the root and , and the same serves both clauses. is the fresh-call counter of the explicit procedure; rank ties (probability zero) are broken by a fixed order on , and the theorem holds for every such order.
Differences from the source:
- encoding: Q(u∗), the number of fresh calls of RGMIS(G, r, u∗) →
Q G u r, the fresh-call counter of the explicit memoized recursionrun, rank ties broken by the order on V (any LinearOrder V); reason: ties have probability zero; for distinct ranks the order on V plays no role - generalization: the tail and expectation bounds → the bounds plus the clause
Measurable (Q G u); reason: a true added clause that makes the lower integral the expectation; dropping it gives the printed theorem - encoding: Pr[·] and E[Q(u∗)] → outer measure of {Q > C∆^90(1+h)} under
rankLaw V; lower Lebesgue integral of Q as an extended nonnegative real; reason: with Q measurable these are the probability and the expectation
import Mathlib import Definitions.Def_RGMISFast_run import Definitions.Def_RGMISFast_rankLaw set_option autoImplicit false open MeasureTheory ProbabilityTheory open scoped ENNReal
namespace RGMISFast
/-- Theorem 1.1 (Dalirrooyfard, Makarychev, Mitrović, Ramachandran, *Randomized Greedy MIS Is Fast at Every Node*,
STOC 2027 submission (manuscript sha256 e93cdfec…), p.3): for a finite simple graph with maximum degree at most `∆ ≥ 2` and
i.i.d. Exp(1) ranks, for every root `u∗` fixed before the ranks are sampled and every `h ≥ 0`,
`Pr[Q(u∗) > C∆^90(1 + h)] ≤ e^{−h}`, where `C` is an absolute constant; consequently
`E[Q(u∗)] ≤ 2C∆^90`.
**Formalization Note** `C` is chosen before the vertex type, the graph, `∆`, the root and `h`,
and the same `C` serves both clauses. `∆` is a bound on every degree. `Q` is the fresh-call
counter of the memoized, pruned Procedure 1.1 (`RGMISFast.Q`); rank ties, which have probability
zero, are broken by the order on `V`, and the statement holds for every such order. The law of the
ranks is `rankLaw V`, the product of Exp(1) laws. The tail is the (outer) measure of the set
`{Q > C∆^90(1+h)}`; the expectation is the lower Lebesgue integral of `Q` as an extended
nonnegative real, made the honest expectation by the added clause `Measurable Q`. -/
theorem theorem_1_1 : ∃ C : ℝ, 0 < C ∧
∀ (V : Type) [Fintype V] [LinearOrder V] (G : SimpleGraph V) [DecidableRel G.Adj] (Δ : ℕ),
2 ≤ Δ → (∀ v, G.degree v ≤ Δ) → ∀ u : V,
Measurable (fun r : V → ℝ => Q G u r) ∧
(∀ h : ℝ, 0 ≤ h →
rankLaw V {r | C * (Δ : ℝ) ^ 90 * (1 + h) < (Q G u r : ℝ)} ≤
ENNReal.ofReal (Real.exp (-h))) ∧
∫⁻ r, (Q G u r : ℝ≥0∞) ∂(rankLaw V) ≤ ENNReal.ofReal (2 * C * (Δ : ℝ) ^ 90) := by sorry
end RGMISFast
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.