A bias-matched random coordinate shift does not increase approximation error
ProvedNarrowDNF.ShiftProbability.one_coordinate_error_lecombinatoricsnarrow-dnf-authors-importprobabilitystoc-2027
Let be finite, , , and let , with monotone. Write . Define and . Then
Thus an independent random choice of upper shift with weight and lower shift with weight preserves the error budget in expectation. No monotonicity of the approximator is assumed.
Preamble
import Mathlib.Algebra.BigOperators.Ring.Finset
import Mathlib.Algebra.Order.BigOperators.Ring.Finset
import Mathlib.Data.Bool.Basic
import Mathlib.Data.Finset.Lattice.Fold
import Mathlib.Data.Fintype.BigOperators
import Mathlib.Data.Fintype.Pi
import Mathlib.Data.Real.Basic
import Mathlib.Logic.Equiv.Prod
import Mathlib.Order.Monotone.Basic
import Mathlib.Tactic.Linarith
import Mathlib.Tactic.Ring
import Definitions.Def_NarrowDNF_Adaptive
import Definitions.Def_NarrowDNF_CoordinateShift
import Definitions.Def_NarrowDNF_FiberShift
import Definitions.Def_NarrowDNF_ProductMeasure
import Theorems.Thm_NarrowDNF_ProductMeasure_expectation_split
namespace NarrowDNF
end NarrowDNF
namespace NarrowDNF.FiberShift
end NarrowDNF.FiberShift
namespace NarrowDNF.ProductMeasure
end NarrowDNF.ProductMeasure
namespace NarrowDNF.ShiftProbability
end NarrowDNF.ShiftProbability
noncomputable section
open NarrowDNF
open NarrowDNF.ShiftProbability
open ProductMeasure
variable {ι : Type*} [DecidableEq ι]
variable [Fintype ι]
Formal statement
theorem NarrowDNF.ShiftProbability.one_coordinate_error_le (p : ℝ) (hp0 : 0 ≤ p) (hp1 : p ≤ 1)
(i : ι) (f v : Cube ι → Bool) (hf : Monotone f) :
(1-p) * ProductMeasure.error p f (upperShift i v) +
p * ProductMeasure.error p f (lowerShift i v) ≤ ProductMeasure.error p f v := by sorrySource