Theorem 2 — μ(A⋆)ξ(B⋆) < 1/6 gives exact recovery for γ in an explicit interval
ProvedRankSparsity.Recovery.theorem_2Throughout, are real square matrices, is a compact singular value decomposition with having orthonormal columns and , is the space of matrices supported inside , , and are the incoherence quantities (1.2) and (1.1), is the spectral norm and the largest entry in magnitude.
Theorem 2. Let with
Then for every with
the unique optimum of
is . Moreover, when and , the value lies strictly inside this interval.
This is the main theorem of the paper: a deterministic condition, depending only on the support of and the row and column spaces of , under which the convex program (1.3) recovers the sparse and low-rank components exactly.
Formalization Note "Unique optimum" is strict: every other feasible pair has strictly larger objective. The upper end of the interval is written multiplied out, ; this is the same condition when and means "no upper bound" when , where Lean's convention would otherwise empty the interval. The "specifically" clause divides by and needs , so it is stated under , , which the paper leaves implicit. , and are taken with respect to a given compact SVD of ; the theorem holds for every such SVD.
import Definitions.Def_matrix_completion_tangent import Definitions.Def_RankSparsity_Recovery_Setup open MatrixCompletion
namespace RankSparsity.Recovery
/-- Theorem 2 (p. 10): if `μ(A⋆) ξ(B⋆) < 1/6`, then for every
`γ ∈ (ξ(B⋆)/(1 - 4μ(A⋆)ξ(B⋆)), (1 - 3μ(A⋆)ξ(B⋆))/μ(A⋆))` the unique optimum of (1.3)
with `C = A⋆ + B⋆` is `(A⋆, B⋆)`; and (for `A⋆, B⋆ ≠ 0`) `γ = √(3ξ(B⋆)/(2μ(A⋆)))` lies in
that range. The upper end is written multiplied out, `γ μ(A⋆) < 1 - 3μ(A⋆)ξ(B⋆)`. -/
theorem theorem_2 {n r : ℕ} (Astar Bstar : RealMatrix n n) (S : SVD Bstar r)
(h : mu Astar * xi S < 1 / 6) :
(∀ γ : ℝ, xi S / (1 - 4 * mu Astar * xi S) < γ →
γ * mu Astar < 1 - 3 * mu Astar * xi S →
IsUniqueOptimum γ (Astar + Bstar) Astar Bstar) ∧
(Astar ≠ 0 → Bstar ≠ 0 →
xi S / (1 - 4 * mu Astar * xi S) < Real.sqrt (3 * xi S / (2 * mu Astar)) ∧
Real.sqrt (3 * xi S / (2 * mu Astar)) < (1 - 3 * mu Astar * xi S) / mu Astar) := by sorry
end RankSparsity.Recovery
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.