When modeling physical systems, training recurrent neural networks, or designing feedback controllers, a fundamental requirement is stability. For a continuous-time linear system $\dot{x} = Ax$, stability requires all eigenvalues of the matrix $A$ to lie in the left half of the complex plane ($\text{Re}(\lambda) \le 0$). For a discrete-time system $x_{k+1} = Ax_k$, the eigenvalues must lie within the unit circle ($|\lambda| \le 1$).
However, optimizing over the space of stable matrices is notoriously difficult. The set of stable matrices is non-convex, meaning that a linear combination of two stable matrices can be unstable. If you try to learn a matrix $A$ from data or optimize it for a control task, naive gradient descent will frequently step into unstable territory. Standard fixes—such as projecting the matrix back to the stable set via eigenvalue clipping—are computationally expensive and destroy critical structural information, as noted by Sutter et al..
To solve this, a powerful research direction has emerged: parameterizing matrices such that they are stable by construction. Instead of checking and correcting eigenvalues during optimization, we change the coordinate system or decompose the matrix into structured components that mathematically prevent instability.
Continuous-Time Stability via Energy Dissipation
The core intuition behind guaranteeing continuous-time stability comes from physics. A physical system is stable if its total energy is bounded and continuously dissipates over time. We can embed this principle directly into a matrix's architecture using a Dissipative Hamiltonian (DH) or port-Hamiltonian parameterization.
As characterized by Sharma et al., any stable continuous-time matrix $A$ can be decomposed into the following structure:
$$A = (J - R)Q$$
Where:
- $J$ is a skew-symmetric matrix ($J = -J^T$), representing internal energy-preserving flux or rotation.
- $R$ is a positive semidefinite matrix ($R \succeq 0$), representing energy dissipation.
- $Q$ is a positive definite matrix ($Q \succ 0$), representing the system's energy storage.
Why It Works
To see why this parameterization guarantees stability without ever computing an eigenvalue, we can construct a quadratic Lyapunov (energy) function: $V(x) = \frac{1}{2}x^T Q x$. Because $Q$ is positive definite, $V(x) > 0$ for all $x \neq 0$.
If we look at how this energy changes over time along the system's trajectories ($\dot{x} = Ax$), we get:
$$\dot{V}(x) = x^T Q \dot{x} = x^T Q A x$$
Substituting the DH parameterization $A = (J - R)Q$:
$$\dot{V}(x) = x^T Q (J - R) Q x$$
Because any matrix quadratic form can be evaluated through its symmetric part, and $J$ is skew-symmetric ($x^T Q J Q x = 0$), the rotation component completely vanishes from the energy derivative. We are left only with the dissipation component:
$$\dot{V}(x) = -x^T Q R Q x \le 0$$
Because $R \succeq 0$, the energy derivative $\dot{V}(x)$ is guaranteed to be non-positive for any choice of $x$. This proves stability by construction.
By optimizing over the spaces of $J$, $R$, and $Q$ under simple convex constraints, the non-convex geometry of stable matrices is mapped to a convex feasible set. This transformation allows researchers to use fast, projection-based gradient methods, shifting the complexity from the constraints to the objective function (Sharma et al.). For instance, Gillis et al. leverage this mapping to find the nearest positive-real (passive) system by replacing abstract frequency-domain conditions with structured matrix equations.
Preserving Structural Constraints
In many applications, stability is not the only constraint; the matrix must also maintain a specific structural property. For example, "positive systems" require a Metzler matrix (where all off-diagonal elements are non-negative).
If you try to enforce both stability and the Metzler property simultaneously, the optimization problem becomes highly constrained. However, Anderson et al. showed that the DH parameterization decouples these constraints. By restricting the energy storage matrix $Q$ to be a diagonal matrix, the Metzler sign constraints can be enforced via simple linear constraints on the dissipative component $R$ without compromising the stability guarantee.
A similar philosophy applies to computer vision and graphics. Lee et al. decompose deformation energy predictions into separate cotangent Laplacian and inverse mass matrices. This structural constraint ensures positive semi-definiteness, yielding stable and physically plausible shape transformations.
Discrete-Time Stability and Contractions
For discrete-time systems ($x_{k+1} = Ax_k$), stability requires that the matrix does not magnify the length of state vectors over time, meaning its spectral radius must be less than or equal to one ($\rho(A) \le 1$).
To parameterize this constraint without checking eigenvalues, Karow et al. introduced the SUB factorization:
$$A = S^{-1} U B S$$
Where:
- $S$ is an invertible matrix representing a coordinate transformation.
- $U$ is an orthogonal matrix ($U^T U = I$), representing a rotation or reflection.
- $B$ is a positive semidefinite contraction ($\|B\|_2 \le 1$), representing a shrinking operation.
Why It Works
Orthogonal matrices preserve the Euclidean norm ($\|Ux\|_2 = \|x\|_2$), while contractions strictly reduce or maintain it ($\|Bx\|_2 \le \|x\|_2$). Therefore, the product $UB$ is guaranteed to have an operator norm of at most one.
By applying the similarity transformation $S$, the matrix $A$ inherits this property in the transformed space. This factorized structure completely decouples the non-convex spectral radius constraint into independent, easily manageable geometric projections. This strategy has been successfully extended to discrete-time descriptor systems by Karow et al. and adopted by Mamakoukas et al. to learn stable linear dynamical systems from data with exceptional memory efficiency.
Applications in Machine Learning and Control
By embedding these structural parameterizations directly into optimization loops, researchers have unlocked more robust learning and control paradigms:
- Learning Nonlinear Dynamics: When learning linear embeddings of complex, nonlinear systems (such as Koopman operator learning), Pan et al. parameterize the operator's matrix to ensure its eigenvalues possess non-positive real parts. Similarly, Serino et al. use a specialized layer to constrain the fast dynamics of singularly perturbed systems to the left-half complex plane, guaranteeing the existence of an attracting slow manifold without relying on unreliable soft loss penalties.
- Robust Controller Design: In control theory, verifying closed-loop stability often requires expensive, iterative eigenvalue checks. Schwerdtner et al. bypass this by parameterizing the controller itself as a port-Hamiltonian system. Because the interconnection of two port-Hamiltonian systems is inherently port-Hamiltonian, the entire closed-loop system is guaranteed to be passive and stable by construction.
Ultimately, whether dealing with continuous-time dissipation or discrete-time contractions, the core insight remains the same: let geometry do the work that optimization finds hard. By choosing coordinate systems that mirror physical laws, stability becomes an invariant property of the architecture rather than a constraint to be chased.
Key Foundational Work
Non-linear stability of singular relative periodic orbits in Hamiltonian systems with symmetry (1999) — contributed 2 concept elements to this direction.
Eigenvalue Density, Li's Positivity, and the Critical Strip (2009) — contributed 1 concept elements to this direction.