Modern Control Theory: State-Space Analysis & Synthesis

Modern control theory is not one controller. It is a way of reasoning about dynamics through an internal state $x$: first decide what state means, then determine how it evolves, which directions an actuator can move, which directions sensors reveal, and only then synthesize feedback or an estimator.

The common mathematical object #

For a continuous-time linear time-invariant system,

$$\dot x=Ax+Bu+Ed,\qquad y=Cx+Du,$$

where $x\in\mathbb R^n$ is state, $u\in\mathbb R^m$ is manipulated input, $d$ is disturbance, and $y\in\mathbb R^p$ is measured or regulated output. The discrete-time counterpart is

$$x_{k+1}=A_dx_k+B_du_k+E_dd_k,\qquad y_k=C_dx_k+D_du_k.$$

The matrices are not just a storage format. Their structure answers different questions:

Analysis must precede synthesis #

A pole-placement command can return a matrix even when the engineering design is poor. The meaningful sequence is:

1. define states, inputs, outputs, disturbances, units, operating point, and validity region;
2. choose or transform a realization without confusing coordinates with physics;
3. compute propagation and identify stable, unstable, slow, fast, and repeated modes;
4. test controllability or stabilizability and observability or detectability;
5. choose closed-loop objectives compatible with actuator authority, sensor noise, delay, and uncertainty;
6. synthesize state feedback, integral augmentation, and an observer where needed;
7. verify the complete loop, including saturation, sample/hold behavior, initialization, and model mismatch.

Continuous and discrete designs are parallel, not interchangeable #

For continuous state feedback $u=-Kx$, the closed-loop matrix is $A-BK$ and stability requires its eigenvalues in the open left half-plane. For a discrete model, it is $A_d-B_dK$ and stability requires eigenvalues strictly inside the unit circle.

The two are connected by the sampler and hold, not by renaming $s$ as $z$. With a zero-order-held input and sample interval $T_s$,

$$A_d=e^{AT_s},\qquad B_d=\int_0^{T_s}e^{A\tau}B\,d\tau.$$

This exact sampled model preserves sample-to-sample state under the stated hold assumption. Computation delay, jitter, quantization, and intersample motion remain separate implementation effects; see Discrete-Time Control.

Where the topic boundaries lie #

• State-Space Realizations, Similarity & Canonical Forms explains coordinate changes, transfer-function realizations, minimality, and conditioning.
• State Transition Matrix & Forced Response derives how initial state and input history produce a trajectory.
• Controllability, Observability & Structural Modes separates modes that can be actuated or inferred from those hidden at the interface.
• State Feedback & Pole Placement shows what feedback changes, when assignment exists, and why feasible poles are not automatically good poles.
• State Observers & the Separation Principle turns measured output and known input into state estimates and combines them with feedback.
• Tracking, Disturbance Rejection & the Internal Model Principle covers integral and other servo augmentations for persistent signals.
• Optimal-Control Problem Formulation defines dynamics, admissible processes, endpoints, and the cost before a solution method is chosen.
• Pontryagin Maximum Principle derives constrained extremals and switching laws from state-costate dynamics.
• Dynamic Programming & HJB solves for a value function whose pointwise minimizer is feedback.
• Linear-Quadratic Regulation & Riccati Equations specializes that optimal-control machinery to linear dynamics and quadratic costs.

Kalman filtering, robust control, nonlinear control, and model predictive control build on this foundation but add different objectives and assumptions. The durable core is the chain state definition → propagation → structural feasibility → synthesis → closed-loop verification.