Operator-Valued LMI Optimization

• Operator-valued LMI optimization is a framework using self-adjoint operator constraints to define convex feasibility in both finite- and infinite-dimensional systems.
• It applies to advanced control theory and PDE stabilization by generalizing classical LMIs with powerful semidefinite programming relaxations.
• Computational approaches leverage matrix convexity, dual Riccati structures, and SOS methods to address scalability and approximation challenges.

Operator-valued linear matrix inequality (LMI) optimization is the study and solution of optimization problems where the feasibility or cost constraints are expressed as operator-valued linear matrix inequalities acting on infinite- or finite-dimensional spaces. Such problems generalize standard matrix-valued LMI optimization to the setting of bounded self-adjoint operators and are foundational in noncommutative and functional analysis, control theory, and real algebraic geometry. They encode convexity, feasibility, and performance conditions for both finite-dimensional and operator-theoretic systems, and their tractable relaxations form the backbone of contemporary semidefinite programming (SDP) for structured, possibly infinite-dimensional, problems.

1. Definitions and Mathematical Foundations

A Hermitian operator-valued linear matrix pencil is an affine map

$L(x) = A_0 + \sum_{i=1}^g A_i x_i,$

where $A_i = A_i^* \in B(\mathcal{H})$ are bounded self-adjoint operators on a complex Hilbert space $\mathcal{H}$, and $x = (x_1, \ldots, x_g)$ consists of formal noncommuting variables. For a $g$-tuple $X = (X_1, \ldots, X_g)$ of self-adjoint operators on a Hilbert space $\mathcal{K}$, the pencil is evaluated as

$$L(X) = A_0 \otimes I_\mathcal{K} + \sum_{i=1}^g A_i \otimes X_i \in B(\mathcal{H} \otimes \mathcal{K}).$$

The operator-valued LMI constraint is $L(X) \succeq 0$, i.e., $L(X)$ is positive semidefinite. Feasible sets for such constraints—so-called free spectrahedra—are structured by matrix convexity: closedness under direct sums and isometric conjugation. The Helton–McCullough theorem characterizes free spectrahedra as precisely those domains determined by such pencils, i.e., matrix-convex, noncommutative semialgebraic sets defined by operator-valued LMIs (Volčič, 2024).

In applications to partial differential equations or systems with operator-valued variables, problem data such as

$$F(x; \gamma) = F_0(x) + \sum_{i=1}^s \gamma_i F_i(x)$$

are symmetric polynomial matrix-valued functions, and the LMI constraint appears as

$$\forall w \in H, \quad \mathcal{F}_\gamma\{w\} := \int_{-1}^{1} (D^k w(x))^T F(x; \gamma) D^k w(x) dx \geq 0,$$

where $D^k w$ denotes the vector of derivatives and $H$ is a finite-energy function space (Fantuzzi et al., 2016).

2. Finite-Dimensional Special Cases and Explicit Formulas

Finite-dimensional LMI optimization can be viewed as a special case. For Hermitian matrices $A_i \in \mathbb{C}^{m_i \times m_i}$, $B_i \in \mathbb{C}^{m_i \times n}$, and $X \in \mathbb{C}^{n \times n}$ Hermitian, canonical problems include extremizing rank or inertia of

$f(X) = A_1 - B_1 X B_1^*$

subject to the Löwner-ordered operator LMI $B_2 X B_2^* \succeq A_2$, or the reversed inequality.

Closed-form solutions for maximal and minimal rank and inertia under such matrix-valued LMIs are given via block-matrix Schur complements, ranks, and inertia formulas:

• $$M \equiv \left[ \begin{smallmatrix} A_1 & B_1 & 0 \ B_1^* & 0 & B_2^* \ 0 & B_2 & -A_2 \end{smallmatrix} \right]$$
• $M_1 =$ Schur complement for the lower block
• The extremal values for $f(X)$ (such as $\max_{X} \mathrm{rank}(f(X))$) are explicit functions of $A_1, B_1, B_2, A_2$, and their blockwise ranks and inertias (Tian, 2013).

The algebraic toolkit includes congruence transformations, Sylvester-type rank identities, Moore–Penrose inverses, and structural feasibility checks. Although these results are matrix-theoretic, the analytic techniques such as block operator congruence and generalized inverses are directly relevant for certain classes of bounded operators on infinite-dimensional Hilbert spaces, subject to appropriate spectral assumptions.

3. Relaxations: From Infinite-Dimensional Operator LMIs to SDPs

Infinite-dimensional operator-valued LMI constraints arising in control of PDEs or dynamical systems (e.g., in stabilization or input-output analysis) are generally intractable in raw form. A standard approach is to approximate these constraints using finite-dimensional relaxations:

• Outer relaxations: Restrict the function space $H$ to finite-dimensional polynomial subspaces (e.g., $N$-degree polynomials) to obtain a finite LMI in the coefficients of $w$, hence a finite SDP. Under regularity, these outer SDPs yield a monotone non-decreasing sequence of lower bounds converging to the true optimum (Fantuzzi et al., 2016).
• Inner relaxations: Use truncated orthogonal expansions (e.g., Legendre series) with tail bounds, along with sum-of-squares (SOS) conditions and auxiliary variables, to ensure positivity of the entire quadratic form. These yield an arguably conservative inner SDP, providing upper bounds and strictly feasible candidate solutions.

QUINOPT implements these schemes in the YALMIP environment, automating both outer and inner SDP construction. Convergence is observed for prototypical PDE-stability benchmarks, with outer and inner relaxations approaching the true optimum as polynomial degree increases.

Relaxation Scheme	Result Type	Description
Outer (Polynomial)	Lower Bound	Feasibility on finite-dimensional polynomial subspace
Inner (SOS + expansion)	Upper Bound	Sufficient positivity on the whole space via SOS and tail estimates

4. Operator-Valued LMI Duality and Riccati Structure

In continuous-time and infinite-dimensional control, as exemplified by the linear-quadratic regulator (LQR) and general IQC (Integral Quadratic Constraints) analysis, operator-valued LMI constraints are naturally formulated in terms of covariance operators.

Given state-input trajectories $x(t), u(t)$ on $[0,T]$, one defines a time-varying joint covariance

$$\Sigma(t) = \mathbb{E}\begin{pmatrix} x(t) \ u(t) \end{pmatrix}\begin{pmatrix} x(t) \ u(t) \end{pmatrix}^T,$$

which satisfies a linear operator ODE. The convex cone $\Sigma(t) \succeq 0$ is an LMI at each $t$, i.e., a family of operator-valued LMIs indexed by time.

The primal optimal control problem (minimizing weighted $L^2$ norm of state and control) becomes a convex SDP over these operator-valued signals, subject to the operator LMI and linear equality constraints. Linear-conic duality in Banach spaces provides the dual formulation: time-varying, operator-valued LMIs (with dual variables in $L^\infty$). Riccati differential equations arise as extremal (slack) solutions of these dual LMIs, and the state-feedback structure is a consequence of complimentary slackness (Bamieh, 2024).

Notably:

• Operator-valued LMIs precisely encode optimality conditions for LQR, $H^\infty$ analysis, and IQCs in this framework.
• The appearance of Riccati equations is a dual phenomenon, independent of square-completion arguments.

5. Structural and Algebraic Properties: Matrix Convexity and Free Spectrahedra

Operator-valued LMIs encapsulate noncommutative convexity. A central result is that free semialgebraic sets defined by noncommutative Hermitian polynomials are matrix-convex if and only if they are operator-valued LMI solution sets (free spectrahedra) (Volčič, 2024). The operator-valued Helton–McCullough theorem confirms that all matrix-convex domains can be realized as operator-valued LMI feasible sets.

The algebraic structure is codified by the Perfect Positivstellensatz: any noncommutative polynomial that is positive semidefinite on a free spectrahedron admits a sum-of-squares plus LMI multiplier certification, with explicit degree bounds. This guarantees that noncommutative eigenvalue optimization over operator-valued LMI domains reduces to an SDP in finitely many variables and coefficients, even in the operator-valued case:

$\sup_{X \in \mathcal{D}_L} \lambda_{\max}(f(X))$

subject to $L(X) \succeq 0$ is equivalent to an SDP via the sum-of-squares representation.

6. Computational Methods and Algorithmic Aspects

In the finite-dimensional setting, convex LMIs subject to affine operator constraints can be efficiently solved by interior-point methods via modeling toolboxes (CVX, YALMIP). Complexity typically scales as $O(N^6)$ in the dimension of the largest semidefinite block; practical problems up to a few hundred dimensions are tractable (Dahdah et al., 2021).

When more complex constraints (e.g., Lyapunov stability, $H_\infty$ norm bounds) introduce bilinear or even operator-valued constraints, the problem becomes a bilinear matrix inequality (BMI). A practical approach, as recommended in Koopman operator regression, is an alternating descent: coordinatewise convex-SDP optimization alternately in the operator and slack variables, yielding locally optimal but not globally certified solutions in practice.

For operator-valued or infinite-dimensional problems, polynomial/trigonometric expansions and SOS relaxations are essential for practical computation. QUINOPT provides an automated environment for the construction and solution of these operator-valued SDPs and is demonstrated to achieve high efficiency and accuracy on prototypical stability and control tasks in PDEs (Fantuzzi et al., 2016).

7. Extensions, Limitations, and Outlook

Limitations include scalability: the size of SDPs from operator-valued relaxations grows rapidly with polynomial degree or dimensionality, especially for inner (SOS-based) relaxations, which are bottlenecked by the state dimension squared. Inner relaxations relying on conservative tail bounds may produce infeasibility in non-elliptic or degenerate cases, and there is not always a guarantee of asymptotic tightness for upper bounds as polynomial degree increases (Fantuzzi et al., 2016).

Generalizations are in active development, including time-dependent operator-valued LMIs, higher-dimensional domain expansions, and frameworks incorporating nonhomogeneous or higher-than-quadratic integral inequalities.

From a functional-analytic perspective, operator-valued LMI optimization provides the theoretical underpinning for convexity in infinite-dimensional noncommutative spaces. The approach naturally unifies systems-theoretic, control, and operator-algebraic methods, and the operator-valued Positivstellensatz (in $C^*$-algebraic settings) extends strong duality and certification results to the operator domain (Volčič, 2024). A plausible implication is that further technical advances in scalable, structured SDP relaxations and operator-system duality will be central to future work in PDE control, quantum information, and noncommutative optimization.