Technical Summary: A Unique Inverse Decomposition of Positive Definite Matrices Under Linear Constraints

1. Problem Statement

The paper investigates a nonlinear decomposition of a symmetric positive definite (SPD) matrix $A \in S^n_{++}$. The objective is to decompose $A$ into two components:
$$A = (B^*)^{-1} + C^*$$
where:

• $C^*$ belongs to a prescribed linear subspace $S \subset S^n$.
• $B^*$ belongs to the orthogonal complement of $S$ with respect to the trace inner product, denoted $S^\perp$, and is itself positive definite ($B^* \in S^\perp \cap S^n_{++}$).

The central mathematical question is whether such a decomposition exists, is unique, and possesses stable structural properties for any subspace $S$ satisfying a specific nondegeneracy condition.

2. Methodology and Theoretical Framework

2.1 Existence and Uniqueness via Variational Characterization

The authors establish the existence and uniqueness of the decomposition by framing the problem as a strictly convex optimization task.

• Primal Formulation: The component $C^*$ is characterized as the unique minimizer of the strictly convex log-determinant functional:
  $$\min_{C \in S, A-C \succ 0} -\log |A - C|$$
  The feasible set is defined as $\Omega_C = \{ C \in S : A - C \succ 0 \}$.
• Nondegeneracy Condition: The decomposition is guaranteed to exist and be unique if the subspace $S$ satisfies the "sharp nondegeneracy condition":
  $$S \cap S^n_+ = \{0\}$$
  This condition ensures that $S$ contains no nonzero positive semidefinite matrices. The authors prove this condition is essentially necessary; if violated, the feasible set becomes unbounded, and the objective function is unbounded below.
• Dual Formulation: The problem admits a dual interpretation where $B^*$ is the unique maximizer of a strictly concave function over $S^\perp \cap S^n_{++}$:
  $$\max_{B \in S^\perp, B \succ 0} \log |B| - \text{tr}(BA)$$
  Strong duality holds between the primal and dual problems.

2.2 Structural Properties

The paper derives several algebraic and geometric properties of the decomposition:

• Symmetry Inheritance: If the input matrix $A$ and the subspace $S$ are invariant under a finite orthogonal group action (via conjugation), the resulting components $B^*$ and $C^*$ inherit this invariance. This allows for dimension reduction in symmetric settings.
• Inverse Relationship: The decomposition of the inverse matrix $A^{-1}$ is directly related to the decomposition of $A$. Specifically, if $A = (B^*)^{-1} + C^*$, then $A^{-1}$ admits a unique decomposition with respect to the transformed subspace $\hat{S} = A^{-1}SA^{-1}$.
• Stability: The mapping from the input matrix $A$ to the components $(B^*, C^*)$ is Lipschitz continuous on open sets where the nondegeneracy condition holds, provided the distance between $A$ and the subspace $S$ is bounded away from zero.

2.3 Algorithmic Approach

To compute the decomposition, the authors propose Newton-type algorithms with feasibility-preserving line searches.

• Newton-CG: For large-scale problems, a Newton-Conjugate Gradient (Newton-CG) method is employed. This approach avoids explicit formation of the Hessian matrix, relying instead on Hessian-vector products.
• Complexity: The per-iteration complexity depends on the structure of $A$ and $S$. The framework accommodates structured matrices (e.g., banded, Toeplitz, or group-invariant), where linear solves and trace evaluations can be performed more efficiently than in the dense case ($O(n^3)$).
• Convergence: The algorithm is proven to be globally convergent with local quadratic convergence rates, leveraging the self-concordance of the log-determinant barrier.

3. Key Results

1. Global Existence and Uniqueness: Theorem 2.1 proves that for any $A \in S^n_{++}$ and any subspace $S$ satisfying $S \cap S^n_+ = \{0\}$, there exists a unique pair $(B^*, C^*)$ satisfying the decomposition.
2. Variational Equivalence: The decomposition is rigorously linked to the minimization of a log-determinant objective, providing a natural dual formulation.
3. Algorithmic Efficiency: The proposed Newton-CG methods provide provable convergence guarantees. The paper details how algebraic structures (bandedness, Toeplitz, symmetry) can be exploited to reduce computational costs, offering specific complexity bounds for dense, banded, Toeplitz, and group-invariant settings.
4. Application to Finance: The decomposition is applied to exponential utility maximization in a discrete-time Gaussian market model.
   • The problem involves maximizing expected utility $E_P[-\exp(-\alpha V^\gamma)]$ subject to information constraints.
   • The optimal portfolio is shown to be linear and characterized by the unique decomposition of the precision matrix $\Lambda = \Sigma^{-1}$.
   • The component $C^*$ (or $\hat{\Gamma}$ in the application) represents the hedging structure constrained by available information, while $B^*$ (or $\hat{Q}$) relates to the residual risk.
   • Numerical experiments on a mixed fractional Brownian motion model demonstrate the utility of the decomposition under different information structures (full information vs. Markovian strategies vs. observation of a single index).

4. Significance and Claims

The paper claims to provide a canonical algebraic decomposition that serves as a structured, non-Euclidean analogue to the classical orthogonal decomposition of symmetric matrices. Its significance is framed in three areas:

• Theoretical Contribution: It establishes a global existence and uniqueness theorem for a nonlinear inverse-based decomposition under linear constraints, a result that goes beyond the scope of existing literature on covariance selection and Gaussian maximum-likelihood estimation (which typically focus on the dual problem or specific statistical settings).
• Algorithmic Contribution: It develops feasibility-preserving Newton-type algorithms with rigorous convergence guarantees and complexity analysis tailored to structured matrices, offering a practical computational tool for solving these decomposition problems.
• Application Contribution: It demonstrates that this abstract decomposition arises naturally in mathematical finance, specifically in exponential utility maximization with partial information. The decomposition provides a clear economic interpretation of the optimal portfolio in terms of information constraints and hedging structures.

The authors emphasize that while the variational characterization is rooted in classical convex analysis, the specific combination of matrix inversion with dual linear constraints yields a unique structural result with broad applicability, distinct from standard statistical estimation problems.