The rectangular Bregman objective attains its minimum on the Schatten power sphere
ProvedHlawkaSchatten.rectangularBregmanObjective_isGlobalMinimumValueLet be finite-dimensional complex inner-product spaces, let be a nonempty finite index type, let with , and let (no sign condition on ). For a complex-linear map with singular values , the Schatten -power sphere is (schattenPowerSphere), and let (that sphere).
For , write for its Hermitian dilation, the self-adjoint operator on with (hermitianDilation). For a self-adjoint operator and , write for the result of applying to 's eigenvalues in any orthonormal eigenbasis of ; this operator itself does not depend on which eigenbasis is chosen, since it acts as on every -eigenvector of , not only on the vectors of one particular chosen eigenbasis — which is what lets and below be single well-defined operators. For , the dilated Bregman trace is
(dilatedBregmanTrace), with and . The rectangular Bregman objective is the half-dilated weighted sum
as a function of ranging over the Schatten -power sphere, and is the (unnormalized) weighted sum of the underlying maps. Writing for the Schatten -norm (schattenPNorm), the theorem states that attains the global minimum value
over the Schatten -power sphere.
This is the operator-level analogue, for the Schatten -norm on rectangular maps, of the exact Hilbert-space minimum of a weighted squared-distance objective: it is one of the two variational facts (the other being the analogous minimum for the paired Mazur-distance objective) whose comparison produces the two-sided bound between triple and pair deficits that this construction needs.
Formalization Note The Schatten -power sphere is defined by ; for this is equivalent to (schattenPNorm_eq_one_iff). In this theorem's range , is convex and differentiable with derivative . The proof uses the resulting scalar Bregman nonnegativity through a supporting theorem, then lifts it to by the spectral trace identity.
import Definitions.Def_HlawkaSchatten_SchattenNorm
import Definitions.Def_HlawkaSchatten_Variational
import Mathlib.Analysis.Calculus.LHopital
import Mathlib.Analysis.Convex.Deriv
import Mathlib.Analysis.Convex.SpecificFunctions.Basic
import Mathlib.Analysis.InnerProductSpace.Adjoint
import Mathlib.Analysis.InnerProductSpace.Basic
import Mathlib.Analysis.InnerProductSpace.Dual
import Mathlib.Analysis.InnerProductSpace.NormPow
import Mathlib.Analysis.InnerProductSpace.PiL2
import Mathlib.Analysis.InnerProductSpace.ProdL2
import Mathlib.Analysis.InnerProductSpace.SingularValues
import Mathlib.Analysis.InnerProductSpace.Trace
import Mathlib.Data.Sign.Basic
import Mathlib.LinearAlgebra.Matrix.Charpoly.Basic
import Mathlib.Topology.Compactification.OnePoint.Basic
import Mathlib.Topology.Instances.Sign
/-
Copyright (c) 2026 Ezzeri Esa. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Authors: Ezzeri Esa
-/
/-!
# Variational minima for the Bregman--Mazur argument
This file proves two reusable parts of the variational layer. First, a
pointwise two-sided comparison transports to attained global minima, even
when the two objectives are indexed by different but equivalent spheres.
Second, the weighted squared-distance objective on a Hilbert unit sphere has
the exact minimum used in the Schatten argument.
-/
open scoped InnerProductSpace ComplexConjugate
variable {𝕜 H ι : Type*} [RCLike 𝕜] [Fintype ι]
[NormedAddCommGroup H] [InnerProductSpace 𝕜 H]
section RectangularMazur
variable {E F κ : Type*} [Fintype κ]
[NormedAddCommGroup E] [InnerProductSpace ℂ E] [FiniteDimensional ℂ E]
[NormedAddCommGroup F] [InnerProductSpace ℂ F] [FiniteDimensional ℂ F]
open HlawkaSchatten
theorem HlawkaSchatten.rectangularBregmanObjective_isGlobalMinimumValue
[Nonempty ι] {p : ℝ} (hp : 1 < p) (a : ι → ℝ)
(u : ι → schattenPowerSphere (𝕜 := ℂ) (E := E) (F := F) p) :
IsGlobalMinimumValue (rectangularBregmanObjective p a u)
((∑ i, a i) - schattenPNorm p (rectangularWeightedSum a u)) := by sorry