When managing large-scale networked systems—such as power grids, autonomous vehicle swarms, or localized robotic networks—centralized control is often impossible. Instead, we require distributed control, where individual agents make decisions using only local information from their immediate neighbors.
Mathematically, this means the controller matrix $K$ must be sparse, containing structural zeros wherever communication links are absent. However, designing an optimal sparse controller is notoriously difficult. When a sparse controller is placed inside a feedback loop with a physical plant $P$, the closed-loop mapping from disturbances to outputs is typically governed by a non-linear relationship like $(I - PK)^{-1}$. Even if $K$ is sparse, its interaction with $P$ causes information to propagate throughout the entire network over time, making the resulting optimization problem highly non-convex and NP-hard.
For years, the standard workaround was a geometric condition known as Quadratic Invariance (QI), discussed in foundational contexts like Rotkowitz et al. (2010). QI dictates that a system can be convexified if its communication network allows information to flow at least as fast as physical signals propagate through the plant. But what happens when a system is not quadratically invariant—when communication is slower, sparser, or more constrained than the underlying physics?
A powerful research direction bypasses the limitations of QI by shifting sparsity constraints away from the raw controller matrix and onto its internal mathematical factorizations.
The Core Intuition: Internal Sparsity Structures
To understand why this works, consider how we represent a controller. Instead of optimizing the raw feedback matrix $K$ directly, we can express it through alternative parameterizations. If we can ensure that these internal representations preserve sparsity through the feedback loop, the optimization problem remains convex, even if the overall system lacks quadratic invariance.
Sparse Coprime Factorizations
One way to achieve this is by factoring the controller into a ratio of two stable matrices, known as a coprime factorization ($K = V M^{-1}$).
Sabau et al. (2015) demonstrated the power of this approach in autonomous vehicle platoons. They discovered that enforcing sparsity constraints on the left coprime factors of a controller, rather than on its transfer function matrix, is sufficient to guarantee a distributed implementation. Because the internal mathematical representation ensures that global objectives are met through localized interactions, vehicles can maintain string stability using only information from their immediate predecessors, matching the performance of a fully centralized system.
Term-Wise Power Series and Generalized Sparsity Subspaces
A more general framework emerges when evaluating the controller's mapping as a power series expansion. In non-QI systems, the feedback loop typically scrambles the sparsity pattern across successive terms of the expansion, violating communication constraints.
To solve this, Furieri and Kamgarpour (2018) introduced Generalized Sparsity Subspaces (GSS). By enforcing term-wise power series sparsity—meaning they constrain every individual component of the controller's expansion to maintain a specific structural pattern—they prevent the feedback loop from destroying the required information flow. This subspace decomposition keeps the optimization problem convex and computationally tractable.
Traditional Design (Non-QI):
Controller K (Sparse) ──> Feedback Loop (I-PK)⁻¹ ──> Sparsity Destroyed (Non-Convex)
Generalized Sparsity Subspaces:
Internal Expansion Terms ──>[Enforce Term-Wise Sparsity]──> Convex Optimization Retained
Building on this, Furieri et al. (2019) formulated the concept of Sparsity Invariance (SI). By utilizing a fractional controller representation, SI decouples the sparsity requirements of the controller's internal feedback and feedforward components. They introduced a polynomial-time algorithm to identify the optimal binary masks for these internal matrices based on the physical topology of the network. This maximizes the available search space for controller design, guaranteeing that if a system happens to be quadratically invariant, it recovers the globally optimal solution, while still providing high-performance, feasible controllers when QI fails.
Scaling Up via Algorithmic Modularity
Once a distributed control problem is cast into a convex form using these generalized subspaces or factorizations, the next challenge is scaling the computation to massive networks.
- Chordal Decomposition and ADMM: Zheng et al. (2020) showed that the inherent sparsity of system interconnections can be exploited using chordal decomposition. By breaking a large-scale decentralized problem into smaller, overlapping subproblems, subsystems can collaboratively design controllers using the Alternating Direction Method of Multipliers (ADMM) without requiring a centralized model. Similarly, Parsi et al. (2021) applied sparse structures to adaptive Model Predictive Control (MPC), ensuring that computational complexity scales linearly rather than exponentially with the number of agents.
- Operator Splitting: For massive "systems of systems," Tang et al. (2025) developed a framework called "Split-as-a-Pro." By combining behavioral systems theory with operator splitting, they decoupled local subsystem dynamics from global network constraints. Each node independently solves its own local projection and communicates exclusively to satisfy network-wide requirements.
- Sparsity-Promoting Norms: Instead of enforcing a fixed communication topology, researchers often want to discover the most efficient network structure. Ito et al. (2021) proved a surprising theoretical equivalence in continuous-time stochastic systems: under certain control-affine conditions, minimizing the non-convex $L^0$ cost (which counts active communication links) is mathematically identical to minimizing the convex $L^1$ norm. This allows practitioners to use efficient convex relaxations to find optimal sparse structures. This is echoed in work by Guo et al. (2024) and Miller et al. (2024), who combine sparsity-promoting norms with Linear Matrix Inequalities (LMIs) to synthesize structured controllers directly from data or under stochastic noise.
Physical Realities and Trade-offs
While generalized sparsity frameworks significantly expand the types of networks we can control, they introduce distinct engineering trade-offs:
- The Discretization Gap: Most physical systems operate in continuous time, but controllers run on digital, discrete-time hardware. Matni and Chen (2020) warned that standard discretization techniques often destroy the mathematical sparsity of continuous-time models. Preserving this sparsity requires specialized discretization methods with explicit error bounds to ensure that localized control strategies remain valid when deployed.
- Conservatism: Transforming an NP-hard problem into a tractable one always comes at a cost. As noted by Miller et al. (2024), the mathematical relaxations used to guarantee stability—especially when dealing with noisy, data-driven models—can introduce conservatism. This means the design algorithms might occasionally fail to find a valid distributed controller even if one exists within the unrelaxed, non-convex search space.
By shifting our focus from the external structure of the controller to its internal algebraic properties, generalized sparsity subspaces provide a rigorous path forward for complex networks, proving that we can achieve localized, scalable control far beyond the boundaries of quadratic invariance.
Key Foundational Work
Convexity of optimal control over networks with delays and arbitrary topology (2010) — contributed 1 concept elements to this direction.
Distributed robust stability analysis of interconnected uncertain systems (2012) — contributed 1 concept elements to this direction.