p. 16 — Step 1 makes every run of a pigeonhole design put exactly T/2 subjects in each group
ProvedPigeonholeDesign.MultiDim.assign_balancedLet be even. For every partition of the covariate space into pigeonholes, every arrival sequence and every outcome of the tie-breaking coins, the pigeonhole design (Definition 3) assigns exactly subjects to control and exactly to treatment:
This is the role of Step 1 of the design. It makes the discrepancy of every run a genuine minimum-weight perfect matching.
Formalization Note The pigeonhole partition is given as an arbitrary map hole from the covariate space to a type of labels with decidable equality; its fibres are the pigeonholes. The statement is about every coin vector, not on average. It guarantees that the junk value of the discrepancy of an unbalanced split never occurs for this design.
import Mathlib import Definitions.Def_PigeonholeDesign_MultiDim_Setting
namespace PigeonholeDesign.MultiDim
theorem assign_balanced {E ι : Type*} [DecidableEq ι] (T : ℕ) (hT : Even T) (hole : E → ι)
(x : Fin T → E) (coin : Fin T → Bool) :
(Finset.univ.filter (fun t => PigeonholeDesign.Lower.assign hole x coin t = false)).card = T / 2 ∧
(Finset.univ.filter (fun t => PigeonholeDesign.Lower.assign hole x coin t = true)).card = T / 2 := by sorry
end PigeonholeDesign.MultiDim
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.