State-space representation

A state-space representation is a mathematical model of a physical system as a set of first-order differential equations in the system's internal state variables. Rather than describing a system by its input-output relationship alone — as in the transfer function or frequency response approaches — state-space methods expose the internal dynamics, revealing how the current state evolves under the influence of inputs and how outputs are generated from that state.

The standard form is a pair of equations: the state equation \(\dot{x} = Ax + Bu\), which governs the evolution of the state vector \(x\) under input \(u\), and the output equation \(y = Cx + Du\), which maps the state to the observed output \(y\). The matrices \(A\), \(B\), \(C\), and \(D\) encode the system dynamics, input coupling, observation model, and direct feedthrough respectively. This representation is not unique: any invertible change of coordinates \(x' = Tx\) produces an equivalent description with transformed matrices, a fact that is exploited in canonical forms and model order reduction.

From Transfer Functions to State Space

Every linear time-invariant system described by a transfer function has infinitely many state-space realizations. The conversion from transfer function to state-space is called realization, and the reverse — computing the transfer function from state-space matrices — is straightforward: \(G(s) = C(sI - A)^{-1}B + D\). The controllability and observability of a realization determine whether it is minimal — that is, whether it uses the smallest possible number of state variables to capture the input-output behavior.

The state-space framework subsumes the transfer function approach. Where classical control theory asks whether the poles of a transfer function lie in the left half-plane, state-space theory asks whether the eigenvalues of \(A\) have negative real parts — the same condition, but expressed in a representation that also reveals what happens inside the black box. The Nyquist stability criterion and Bode plot techniques of classical design have state-space counterparts in Lyapunov stability theory and linear matrix inequalities.

Why State Space Won

State-space methods displaced classical frequency-domain techniques as the dominant framework for control design beginning in the 1960s, driven by the demands of the space program and the development of digital computers. The reasons are structural. Transfer functions struggle with multi-input multi-output systems: the transfer function becomes a matrix of scalar functions, and the intuitive graphical tools of SISO design lose their geometric clarity. State-space methods handle MIMO systems natively: the matrices simply grow in dimension.

More fundamentally, state-space theory provides the conceptual machinery for optimal control. The linear quadratic regulator (LQR) designs a feedback controller by minimizing a quadratic cost function over state and input trajectories. The Kalman filter estimates the state of a system from noisy measurements. Together, LQR and Kalman filtering form the separation principle: under suitable conditions, optimal estimation and optimal control can be designed independently and combined to produce an optimal output-feedback controller. No comparable result exists in the transfer function framework.

The Hidden Cost

The power of state-space methods comes with a cost that is rarely acknowledged. A state-space model of order \(n\) requires \(n^2\) parameters in the \(A\) matrix alone, and the number of parameters grows quadratically with system dimension. For large systems — power grids, climate models, biological networks — this parameter explosion makes identification, simulation, and control computationally prohibitive.

There is a deeper epistemological problem. The state vector \(x\) in a state-space model is often not directly measurable. It is a mathematical construct, a coordinate system chosen for computational convenience. When engineers speak of "estimating the state," they are estimating a fiction — a useful fiction, but a fiction nonetheless. The triumph of state-space methods is also the triumph of a particular mathematical ontology: one in which internal variables are granted causal power despite being unobservable in principle. This is not a bug to be fixed; it is a structural feature of a framework that has traded empirical transparency for computational tractability, and the trade is not always in our favor.