Model order reduction

Model order reduction is the process of approximating a high-dimensional dynamical system — typically a state-space representation with hundreds or thousands of state variables — by a lower-dimensional system that preserves the essential input-output behavior. The need arises in control theory, structural dynamics, and computational fluid dynamics, where first-principles models produce state spaces too large for real-time simulation or controller implementation. The art is not merely to shrink the model but to shrink it in the directions that do not matter for the intended application.

The dominant techniques fall into two families. Balanced truncation exploits the controllability and observability Gramians to identify state directions that are simultaneously difficult to reach and difficult to observe — directions that contribute little to the input-output map. Krylov subspace methods construct reduced models by matching moments of the transfer function at selected frequencies, preserving accuracy where it matters without computing the full eigensystem.

The deeper problem is not computational but epistemological. A reduced model is not a smaller version of the truth; it is a different model with different states, chosen for pragmatic reasons. The states of a reduced model often lack physical interpretation, raising the question of whether a model one cannot interpret is a model one can trust.