bernoulli_expectation_sqrt_le_sqrt_expectation
ProvedSquare-root Jensen (Cauchy–Schwarz) for the Bernoulli expectation.
Fix dimensions and an inclusion probability . Let denote the Bernoulli expectation, where and . For every pointwise-nonnegative statistic ,
This is the concavity of against a probability measure, equivalently the Cauchy–Schwarz inequality
using that the weights are nonnegative (because ) and sum to (binomial theorem, ).
It is the -versus- power-mean bridge in the conditional-Khintchine matrix-moment estimates: it converts an bound into the bound needed for the Rademacher symmetrization scale.
Preamble
import Definitions.Def_matrix_completion_bernoulli open MatrixCompletion
Formal statement
theorem bernoulli_expectation_sqrt_le_sqrt_expectation
{n1 n2 : ℕ} (p : ℝ)
(F : Finset (Fin n1 × Fin n2) → ℝ) :
0 ≤ p → p ≤ 1 → (∀ Ω, 0 ≤ F Ω) →
bernoulliExpectation p (fun Ω => Real.sqrt (F Ω)) ≤
Real.sqrt (bernoulliExpectation p F) := by sorrySource
Candès–Recht, Exact Matrix Completion via Convex Optimization (symmetrization / matrix-moment machinery)