general_rademacher_matrix_2p_trace_moment_general_index
Provedcandes-rechtkhintchinelinear-algebramatrix-completionreferencerudelson
General-index Rademacher matrix -trace-moment bound. For a finite family of Hermitian matrices over an arbitrary finite index type , indexed by , with the variance matrix satisfying a quadratic-form domination for all (so ), the symmetric Rademacher trace moment is bounded: . This lifts the standard Tropp/Khintchine engine to an arbitrary finite index via a reindexing equivalence (trace and powers are reindex-invariant), and supplies the engine's eigenvalue hypothesis from the quadratic-form bound. Proof (reduction): reindex and each to , convert the quadratic-form bound to the eigenvalue bound via the Rayleigh bridge, apply the engine, and pull the trace moment back unchanged.
Preamble
import Mathlib.Analysis.Matrix.Spectrum import Mathlib.LinearAlgebra.Matrix.Hermitian import Mathlib.LinearAlgebra.Matrix.Trace import Mathlib.Data.Matrix.Reflection import Mathlib.Data.Nat.Factorial.Basic open Matrix open scoped BigOperators
Formal statement
theorem general_rademacher_matrix_2p_trace_moment_general_index {ι : Type*} [Fintype ι] [DecidableEq ι] {μ : Type*} [Fintype μ] [DecidableEq μ] (H : ι → Matrix μ μ ℝ) (hHerm : ∀ c, (H c).IsHermitian) (normV : ℝ) (hnormVnn : 0 ≤ normV) (hVHerm : (∑ c : ι, H c * H c).IsHermitian) (hquad : ∀ v : μ → ℝ, (star v ⬝ᵥ (∑ c : ι, H c * H c) *ᵥ v) ≤ normV * (star v ⬝ᵥ v)) (p : ℕ) : (∑ eps : Finset ι, ((1 : ℝ) / 2) ^ (Fintype.card ι) * Matrix.trace ((∑ c : ι, (if c ∈ eps then (1 : ℝ) else -1) • H c) ^ (2 * p))) ≤ ((Nat.factorial (2 * p) : ℝ) / ((2 ^ p : ℝ) * (Nat.factorial p : ℝ))) * normV ^ p * (Fintype.card μ : ℝ) := by sorrySource
Candes-Recht 2009 (arXiv:0805.4471) Sec 6.1 / Tropp matrix concentration. The Rademacher symmetric matrix trace-moment bound (Khintchine/Tropp engine) lifted from the Fin d index to an arbitrary finite index type μ via a reindexing equivalence μ ≃ Fin (card μ). The eigenvalue hypothesis of the Fin-d engine is supplied here from a quadratic-form domination of the variance matrix ∑ Hc² (which is how the block-diagonal variance matrix is bounded in the dilation application).