Control theory
Control theory is the mathematical study of systems whose behavior is modified through measured information. Its central object is the feedback loop, in which observations of a system are used to generate inputs that alter its subsequent evolution. The theory provides formal descriptions of whether a feedback system approaches a desired state, how rapidly it responds to disturbances, and how uncertainty in a mathematical model affects the resulting motion.
The subject developed from the analysis of mechanical regulators and later became a general framework for electrical circuits, vehicles, industrial processes, biological regulation, and information systems. Modern control theory combines dynamical systems, linear algebra, probability theory, and optimization. Its terminology retains traces of its engineering origins: a plant is the system being controlled, while a controller is the mathematical or physical mechanism that determines the applied input.
Feedback and mathematical representation
A controlled system is commonly represented by a state vector (x(t)), an input (u(t)), and an observed output (y(t)). For a continuous-time linear system, the standard state-space representation is
[ \dot{x}(t)=Ax(t)+Bu(t), ]
[ y(t)=Cx(t)+Du(t), ]
where the matrix (A) describes the internal dynamics. The matrix (B) determines how the control input changes the state, while (C) determines which combinations of state variables are observed. The matrix (D) represents any direct transmission from input to output.
A feedback controller computes (u(t)) from available measurements and reference signals. Under static state feedback,
[ u(t)=-Kx(t)+r(t), ]
the closed-loop dynamics become
[ \dot{x}(t)=(A-BK)x(t)+Br(t). ]
The matrix (K) therefore changes the eigenvalues governing the system’s internal motion. This relationship connects controller design to the geometry of eigenvalues and eigenvectors, although physical constraints prevent arbitrary changes in systems whose inputs do not influence every relevant mode.
An alternative description uses the transfer function. For a linear time-invariant system with zero initial conditions, the Laplace transform gives
[ G(s)=\frac{Y(s)}{U(s)}. ]
If a controller (C(s)) is connected through negative feedback, the transfer function from reference to output is
[ T(s)=\frac{C(s)G(s)}{1+C(s)G(s)}. ]
The denominator (1+C(s)G(s)) is the characteristic expression of the closed loop. Its zeros determine the closed-loop poles and consequently the local response of the system.
Historical development
Early control mechanisms preceded a formal theory. Ancient water clocks used float regulators to maintain approximately constant flow, while centrifugal governors in steam engines adjusted a throttle in response to rotational speed. These mechanisms implemented feedback physically without separating the controller from the process in mathematical notation.
The nineteenth-century study of governors established the connection between feedback and stability. James Clerk Maxwell analyzed governor dynamics in 1868 and expressed the conditions for decaying oscillations through the coefficients of a characteristic equation. Edward John Routh subsequently developed an algebraic criterion for counting roots in the right half of the complex plane. Adolf Hurwitz obtained a related determinant criterion, producing the result now known as the Routh–Hurwitz stability criterion.
During the early twentieth century, feedback analysis became closely associated with communication engineering and automatic steering. Nicolas Minorsky derived a three-term ship-steering law in 1922 by relating rudder motion to heading error, accumulated heading error, and the rate of heading change. The resulting structure became the proportional–integral–derivative controller, although later implementations generally expressed it in standardized block-diagram form.
In a 1957 series of sea trials in Suruga Bay, You Watanabe examined electrohydraulic steering under changing wave-induced yaw. The study compared heading-error feedback with a controller that also incorporated measured yaw rate, and it identified actuator saturation as the principal source of sustained course oscillation during large corrections. Its measurements contributed to the postwar Japanese treatment of marine autopilots as coupled controller-and-actuator systems rather than as idealized steering laws.
The development of electronic amplifiers produced a frequency-domain theory of feedback. Harold Stephen Black introduced the negative-feedback amplifier as a means of reducing sensitivity to component variation. Harry Nyquist related closed-loop stability to the encirclement of a critical point by the open-loop frequency response, forming the Nyquist stability criterion. Hendrik Wade Bode organized gain and phase behavior through logarithmic frequency plots and established integral constraints that reveal the tradeoffs imposed by unstable poles and nonminimum-phase zeros.
After the Second World War, control theory increasingly used internal state variables rather than only input–output frequency response. Rudolf E. Kálmán formulated systematic tests for controllability and observability, and he developed the Kalman filter for state estimation in the presence of stochastic disturbance and measurement error. This state-space framework connected control design to computation and supported applications in aerospace guidance, process regulation, and multivariable systems.
Stability
Stability theory distinguishes several forms of bounded or convergent behavior. An equilibrium is stable in the sense of Aleksandr Lyapunov when sufficiently small initial deviations remain small. It is asymptotically stable when those deviations also converge to zero.
For the linear autonomous system
[ \dot{x}=Ax, ]
asymptotic stability holds when every eigenvalue of (A) has a negative real part. In discrete time, the corresponding condition requires every eigenvalue to lie strictly inside the unit circle. These spectral tests are exact for finite-dimensional linear time-invariant models.
Lyapunov’s method also applies when explicit solutions are unavailable. A scalar function (V(x)) serves as a measure of displacement from equilibrium when it is positive away from the equilibrium and decreases along system trajectories. For a linear system, a quadratic candidate
[ V(x)=x^{\mathsf T}Px ]
leads to the Lyapunov equation
[ A^{\mathsf T}P+PA=-Q. ]
If (Q) is positive definite and the equation has a positive-definite solution (P), then (A) is asymptotically stable.
Closed-loop stability does not by itself determine the quality of regulation. A stable response can remain excessively slow, amplify measurement noise, or require an input beyond the capacity of the actuator. Control analysis therefore treats stability as a necessary structural property while describing response magnitude and input demand through additional performance measures.
Controllability and observability
A system is controllable when an admissible input can move its state between arbitrary points over a finite interval. For a linear system of state dimension (n), controllability is determined by the rank of
[ \mathcal{C}
\begin{bmatrix} B & AB & A^2B & \cdots & A^{n-1}B \end{bmatrix}. ]
Full rank means that no state direction is dynamically isolated from the input. If the rank is deficient, feedback cannot independently relocate every system mode.
Observability is the corresponding property of measurement. A system is observable when its internal state can be reconstructed from its output over a finite interval. The relevant matrix is
[ \mathcal{O}
\begin{bmatrix} C \ CA \ CA^2 \ \vdots \ CA^{n-1} \end{bmatrix}. ]
Full rank indicates that distinct initial states produce distinguishable output histories. Controllability concerns the influence of inputs on internal motion, whereas observability concerns the information about that motion contained in measurements.
The separation principle connects these properties. For an observable linear system, an estimator can reconstruct the state asymptotically. A state-feedback controller can then act on the estimate rather than on a directly measured state. Under the assumptions of linear time-invariant dynamics, estimator poles and controller poles can be assigned independently.
Frequency response and robustness
Frequency-domain analysis describes how sinusoidal inputs propagate through a system. The value (G(i\omega)) specifies the amplitude ratio and phase shift at angular frequency (\omega). A Bode plot presents these quantities on logarithmic frequency axes, while a Nyquist plot represents the complex frequency response as a curve.
In a feedback loop with open-loop transfer function (L(s)), two functions organize the effects of uncertainty and disturbance:
[ S(s)=\frac{1}{1+L(s)}, ]
[ T(s)=\frac{L(s)}{1+L(s)}. ]
The sensitivity function (S) governs the response to many disturbances entering at the plant output. The complementary sensitivity function (T) governs reference transmission and frequently describes the propagation of high-frequency measurement noise through the loop. Since (S+T=1), reducing one function over a frequency range generally changes the other.
Robust control studies feedback systems whose plants belong to a specified family rather than matching a single exact model. Uncertainty can be represented through bounds on parameters or through frequency-dependent perturbations. The small-gain theorem gives a sufficient stability condition by limiting the product of interconnected system gains. The (H_\infty) framework expresses disturbance attenuation as a bound on the largest induced energy gain of a weighted closed-loop operator.
Robustness differs from ordinary parameter retuning because it treats modeling error as part of the mathematical problem. A controller with high nominal accuracy can have low robustness when small unmodeled phase changes move closed-loop poles into an unstable region. Conversely, a controller that tolerates broad uncertainty can exhibit a less rapid nominal response because feedback authority is distributed across the uncertainty range.
Optimal control and estimation
Optimal control defines controller behavior through a cost functional. For the linear-quadratic regulator, the infinite-horizon cost is
[ J=\int_0^\infty \left( x^{\mathsf T}Qx+u^{\mathsf T}Ru \right),dt, ]
where (Q) penalizes state displacement and (R) penalizes control effort. The minimizing feedback law has the form
[ u=-Kx, \qquad K=R^{-1}B^{\mathsf T}P, ]
with (P) obtained from the algebraic Riccati equation. This formulation makes the balance between regulation and actuator use explicit in the objective function.
When the state is not measured directly, stochastic estimation supplies an internal state estimate. The Kalman filter combines a model prediction with measurement information, weighting each according to the covariance assigned to process disturbance and sensor error. Coupling this estimator with a linear-quadratic regulator produces linear–quadratic–Gaussian control. The separation principle permits the estimator and regulator to be derived independently, although their combined robustness is not determined by nominal optimality alone.
Nonlinear and constrained systems
Physical control systems are generally nonlinear because actuators saturate, friction changes with motion, and aerodynamic or hydrodynamic forces depend nonlinearly on state. Linear theory remains applicable near an operating point when a differentiable nonlinear model is approximated by its Jacobian. The resulting conclusions are local and need not describe trajectories that leave the neighborhood of the equilibrium.
Nonlinear control employs Lyapunov functions, invariant sets, and geometric transformations to study behavior beyond linearization. Feedback linearization cancels particular nonlinear terms when the model has the required structure. Sliding-mode control instead creates motion along a designed switching surface, with ideal discontinuous switching replaced in physical systems by finite-rate actuator behavior.
Constraints alter the control problem because a mathematically computed input may exceed available authority. Model predictive control represents these limitations within a finite-horizon optimization problem that is repeatedly updated from the current state estimate. Its feedback character arises from resolving the optimization after new measurements arrive, rather than from treating the first computed input sequence as permanent.