Variational Analysis in Spectral Decomposition Systems

Basic Information

• Paper ID: 2510.11433
• Title: Variational Analysis in Spectral Decomposition Systems
• Authors: Hòa T. Bùi (Curtin University), Minh N. Bùi (University of Graz), Christian Clason (University of Graz)
• Classification: math.OC (Optimization and Control)
• Publication Date: October 13, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.11433

Abstract

This paper investigates the variational analysis of spectral functions and spectral sets that depend solely on the eigenvalues of matrices. Building upon the authors' previous work on convex analysis of such functions, the paper addresses this problem within an abstract framework of spectral decomposition systems, which encompasses a broad range of previously studied settings, including eigenvalue decomposition of Hermitian matrices and singular value decomposition of rectangular matrices. The framework also enables the derivation of new results in more general settings, such as normal decomposition systems and signed singular value decompositions. The main results characterize the Fréchet and limiting normal cones of spectral sets, as well as the Fréchet, limiting, and Clarke subdifferentials of spectral functions through reduced functions. For the latter, the authors also characterize Fréchet differentiability. Finally, a generalization of the Lidskiĭ theorem concerning the spectrum of additive perturbations of Hermitian matrices to arbitrary spectral decomposition systems is obtained.

Research Background and Motivation

Problem Background

1. Practical Requirements: Many practically relevant optimization problems naturally arise in matrix form rather than vector form, such as nonnegative matrix factorization, matrix completion, low-rank approximation, and operator learning.
2. Basis Invariance: In applications such as operator learning, the concern is optimizing finite-dimensional linear operators rather than their specific matrix representations, which requires objective functions to remain invariant under basis changes.
3. Universality of Spectral Functions: Under appropriate assumptions, such invariant functions are completely characterized by their dependence on eigenvalues (or singular values), as exemplified by the nuclear norm of matrices.

Research Significance

Spectral functions are ubiquitous in numerous applications:

• Robust matrix estimation
• Signal processing
• Conic programming
• Semidefinite programming
• Nonlinear elasticity
• Brain network analysis

Limitations of Existing Approaches

While many such problems can be formulated as convex problems, this is not always the case. For example:

• Low-rank matrix completion via Schatten p-norm minimization (0 < p < 1)
• Mathematical programming with semidefinite cone complementarity constraints (SDCMPCC)

The main challenge in existing work is that invariant functions depend only on the set of eigenvalues rather than their ordering, and various works address specific settings in isolation.

Research Motivation

This paper aims to unify results on Fréchet, limiting, and Clarke subdifferentials as well as Fréchet differentiability into a unified framework that encompasses all these settings and, more importantly, enables easier derivation of results for settings and objects not yet covered.

Core Contributions

1. Unified Theoretical Framework: Establishes an abstract framework for spectral decomposition systems that unifies various settings previously studied in isolation (Hermitian matrices, rectangular matrices, Euclidean Jordan algebras, etc.).
2. Variational Geometric Characterization: Completely characterizes the Fréchet and limiting normal cones of spectral sets: $N^\#(X; \gamma^{-1}(D)) = \{\Lambda_a y \mid y \in N^F(\gamma(X); D) \text{ and } a \in A_X\}$
3. Subdifferential Formulas: Establishes complete characterization of Fréchet and limiting subdifferentials of spectral functions: $\partial^\#(\varphi \circ \gamma)(X) = \{\Lambda_a y \mid y \in \partial^\# \varphi(\gamma(X)) \text{ and } a \in A_X\}$
4. Differentiability Characterization: Completely characterizes Fréchet differentiability of spectral functions and their derivatives.
5. Clarke Subdifferential: Provides a representation of the Clarke subdifferential of spectral functions.
6. Generalized Lidskiĭ Theorem: Extends the classical Lidskiĭ theorem to arbitrary spectral decomposition systems.

Methodology Details

Definition of Spectral Decomposition Systems

Definition 2.1 (Spectral Decomposition System): A spectral decomposition system of a Euclidean space ℌ is a quadruple 𝔖 = (𝒳, S, γ, (Λₐ)ₐ∈A), where:

• 𝒳 is a Euclidean space
• S is a group acting on 𝒳 through linear isometries
• γ：ℌ → 𝒳 is the spectral map
• (Λₐ)ₐ∈A is a family of linear isometries from 𝒳 to ℌ

satisfying three compatibility conditions:

• A There exists an S-invariant map τ：𝒳 → 𝒳 such that τ(x) ∈ S·x and γ ∘ Λₐ = τ
• B For each X ∈ ℌ, there exists a ∈ A such that X = Λₐγ(X)
• C Generalized von Neumann trace inequality: ⟨X|Y⟩ ≤ ⟨γ(X)|γ(Y)⟩

Spectral Functions and Spectral Sets

Definition 2.3:

• Spectral Function: Φ：ℌ → -∞,+∞ is a spectral function if and only if γ(X) = γ(Y) ⇒ Φ(X) = Φ(Y)
• Spectral Set: D ⊂ ℌ is a spectral set if and only if its indicator function is a spectral function

Key Characterization (Proposition 2.4): A function Φ is a spectral function if and only if there exists an S-invariant function φ such that Φ = φ ∘ γ.

Core Technical Methods

1. Geometric Method: Transfers results from normal cones of spectral sets to subdifferentials of spectral functions through standard techniques involving normal cones to epigraphs.
2. Product Space Technique: Utilizes product space spectral decomposition systems (Example 2.14) for handling epigraphs.
3. Compactness Arguments: Leverages the closedness of {Λₐ}ₐ∈A in Assumption 2.2 to establish compactness through Proposition 2.18.

Experimental Setup

Examples for Theoretical Verification

This is primarily a theoretical work, with verification of the framework's universality through multiple concrete examples:

1. Eigenvalue Decomposition of Hermitian Matrices (Example 2.11)
2. Singular Value Decomposition of Rectangular Matrices (Example 2.12)
3. Euclidean Jordan Algebras (Example 2.10)
4. Normal Decomposition Systems (Example 2.8)
5. Signed Singular Value Decomposition (Example 2.13)

Verification Methods

• Prove that each concrete example satisfies the definition of spectral decomposition systems
• Verify that the closedness assumption (Assumption 2.2) holds in all examples
• Demonstrate how to recover known specific results from the general theory

Experimental Results

Main Theoretical Results

Theorem 4.1 (Subdifferential Characterization): Let φ：𝒳 → -∞,+∞ be S-invariant, then:

• Fréchet subdifferential: ∂F(φ ∘ γ)(X) = {Λₐy | y ∈ ∂Fφ(γ(X)) and a ∈ A_X}
• Limiting subdifferential: ∂L(φ ∘ γ)(X) = {Λₐy | y ∈ ∂Lφ(γ(X)) and a ∈ A_X}

Corollary 4.8 (Differentiability Characterization): φ ∘ γ is Fréchet differentiable at Λₐx if and only if φ is Fréchet differentiable at x, in which case: $∇(φ ∘ γ)(Λₐx) = Λₐ(∇φ(x))$

Proposition 4.11 (Clarke Subdifferential): $∂C(φ ∘ γ)(X) = \text{conv}\{Λₐy | y ∈ ∂Cφ(γ(X)) \text{ and } a ∈ A_X\}$

Theorem 5.1 (Generalized Lidskiĭ Theorem): If S is a finite group, then: $γ(X + Y) - γ(X) ∈ \text{conv}(S · γ(Y))$

Unification of Existing Results

The paper's results unify multiple known results:

• Lewis 31 results on Hermitian matrices
• Lewis-Sendov 34,35 results on singular values
• Lourenço-Takeda 37 results on Euclidean Jordan algebras
• Classical Lidskiĭ theorem and its various generalizations

New Results

• Variational analysis in normal decomposition systems (resolving an open problem from 33)
• Non-convex case of signed singular value decomposition

Related Work

Main Research Directions

1. Eigenvalues of Symmetric Functions: Lewis 28,31 analyzed Fréchet differentiability and Clarke subdifferentials
2. Signed Symmetric Functions of Singular Values: Lewis-Sendov 34,35 series of works
3. Euclidean Jordan Algebras: Works 52,1,37,51 and others
4. Lie Theory Framework: Berezin-Gel'fand 5, Tam 53, and others

Advantages of This Paper

1. Unification: First to unify all these settings into a single framework
2. Generality: Capable of handling settings not previously covered
3. Simplification: Bypasses matrix-dependent proof techniques through geometric methods
4. Completeness: Simultaneously addresses Fréchet, limiting, and Clarke subdifferentials

Conclusions and Discussion

Main Conclusions

1. Theoretical Unification: Successfully establishes an abstract framework for spectral decomposition systems that unifies various settings in variational analysis.
2. Complete Characterization: Provides complete characterizations of various variational analysis objects for spectral functions and spectral sets.
3. Methodological Innovation: Through geometric methods and product space techniques, provides more concise proofs than existing work.

Limitations

1. Finiteness Assumption: The generalization of the Lidskiĭ theorem requires the group S to be finite.
2. Closedness Condition: Requires an additional closedness assumption (Assumption 2.2).
3. Clarke Subdifferential: Uncertain whether the convex hull operation can be omitted.

Future Directions

1. Applications: Utilize subdifferential representations for necessary optimality conditions in concrete matrix optimization problems.
2. Metric Regularity: Characterize Lipschitz properties of solution mappings.
3. Second-Order Analysis: Extend second-order variational analysis results for spectral functions.

In-Depth Evaluation

Strengths

1. Theoretical Depth: Establishes a profound abstract framework with strong theoretical value.
2. Strong Unification: Successfully unifies multiple seemingly unrelated research areas.
3. Technical Innovation: Use of geometric methods simplifies complex technical proofs.
4. Complete Results: Systematically addresses various subdifferential concepts.
5. Clear Presentation: Well-organized paper with clear definitions and theorem statements.

Weaknesses

1. High Abstraction: The theoretical framework is quite abstract and may be difficult to understand and apply.
2. Computational Complexity: Computing subdifferentials in practice may still be difficult.
3. Few Application Examples: Lacks concrete numerical examples and applications.

Impact

1. Academic Value: Provides important theoretical tools for variational analysis and matrix optimization.
2. Unifying Role: Consolidates scattered research results into a unified framework.
3. Inspirational Value: May inspire similar unification efforts in other fields.

Applicable Scenarios

1. Matrix Optimization: Theoretical analysis of semidefinite programming, matrix completion, and related problems.
2. Nonsmooth Optimization: Non-convex optimization problems involving spectral functions.
3. Variational Analysis: Theoretical research requiring computation of subdifferentials of complex functions.

References

The paper cites 55 related references, primarily including:

• Classical variational analysis texts: Rockafellar & Wets 48, Mordukhovich 42
• Spectral function analysis: Lewis series of works 28,31,34,35
• Euclidean Jordan algebras: Faraut & Korányi 19, related application works
• Matrix analysis: Horn & Johnson 22, Bhatia 6

This paper represents an important theoretical contribution to the field of variational analysis. Through the abstract framework of spectral decomposition systems, it successfully unifies multiple research directions and establishes a solid theoretical foundation for further development in this field.