In classical control theory, we typically analyze a system by following its trajectory—a single point moving through state space over time. When the system's dynamics are nonlinear, finding the optimal trajectory or proving that the system stays within a "safe" region becomes a non-convex optimization problem. These problems are notoriously difficult to solve globally because they are riddled with local minima.
A powerful alternative, championed across the work of Lasserre et al. and Korda et al., is to stop looking at individual points and start looking at distributions. By "lifting" the problem into the space of measures, complex nonlinear dynamics can be reformulated as linear constraints.
The Intuition of Occupation Measures
Imagine a robot moving through a room. Instead of asking "Where is the robot at time $t$?", we ask "How much total time does the robot spend in this specific corner of the room?"
This distribution of time is called an occupation measure. As Lasserre et al. explain, this measure represents the time-average distribution of system states and controls. The beauty of this shift is that while the robot’s path might be a jagged, nonlinear mess, the way the occupation measure evolves is governed by linear equations. Specifically, the physics of the system (the differential equations) are transformed into linear relationships between the moments of these measures.
Linearizing the Impossible
This transformation allows researchers to tackle problems that are otherwise intractable:
- Global Optimality: Because the reformulated problem is a linear program (LP), any solution we find is globally optimal. Lasserre et al. demonstrate that this allows for the synthesis of near-optimal feedback control laws even for highly non-convex polynomial systems.
- Safety and Reachability: To prove a system is safe, we need to know its Region of Attraction (ROA)—the set of all starting points that end up at a desired goal. Korda et al. show that by tracking the evolution of state distributions rather than trajectories, one can compute guaranteed "safe" inner approximations of these regions using convex optimization.
- Hybrid Systems: Systems that switch between different behaviors (like a walking robot making and breaking contact with the ground) are usually NP-hard to optimize. However, Han et al. and Buehrle et al. show that occupation measures can represent these discrete mode transitions as part of a continuous distribution, effectively "smoothing" the discrete jumps into a convex framework.
The Moment-SOS Hierarchy
While the occupation measure reformulation results in a linear program, that program is infinite-dimensional because it lives in the space of all possible measures. To solve it numerically, researchers use the Moment-SOS (Sum-of-Squares) hierarchy.
As detailed in a survey by Lasserre, this hierarchy breaks the infinite-dimensional problem into a sequence of finite-dimensional Semidefinite Programs (SDPs). Each step in the hierarchy considers more "moments" of the measure (like mean, variance, skewness, and so on). As you include more moments, the approximation becomes tighter, providing increasingly accurate global bounds on the optimal cost or the system's peak values (Miller et al.).
A Unified Framework for Control
The power of the occupation measure approach is its versatility. It has been successfully applied to:
- Inverse Optimal Control: Deducing the underlying cost function from observed behavior by treating the problem as a conic dual to the measure-based LP (Pauwels et al.).
- Stochastic and Online Learning: Managing unpredictable environments by modeling policies as stationary occupation measures, which maps expected costs to linear functions over a convex polytope (Dick et al.; Kamgarpour et al.).
- Complex Dynamics: Handling systems with time delays (Miller et al.) or impulsive actions like rocket burns (Claeys et al.).
By shifting the perspective from the "where" of a single trajectory to the "how much" of a measure, this research direction provides a mathematically rigorous bridge between nonlinear dynamics and the powerful world of convex optimization.
Key Foundational Work
Nonlinear optimal control synthesis via occupation measures (2008) — contributed 2 concept elements to this direction.
Nonlinear Optimal Control via Occupation Measures and LMI-Relaxations (2008) — contributed 1 concept elements to this direction.
POCP: A package for polynomial optimal control problems (2009) — contributed 1 concept elements to this direction.