Controllability is a property of a dynamical system that describes whether external inputs can move its internal state from one specified condition to another in a finite time. In control theory, it establishes which motions or operating conditions are achievable through the available control inputs. Controllability depends on the system dynamics, the placement and action of actuators, the permitted inputs, and the time available; it does not simply mean that a system responds to a command.
States, inputs, and reachability
A system's state is a collection of variables sufficient, together with future inputs, to determine its future evolution. For a moving object, these might include position and velocity; for an electrical circuit, capacitor voltages and inductor currents.
A continuous-time state-space model can be written as
where is the state and is the input.
A target state is reachable from an initial state over a specified interval if there is an admissible input that takes the system from to . A system is completely state-controllable if every initial state can be transferred to every target state in a finite time, under the stated assumptions about inputs and timing.
Terminology varies between texts:
- Reachability often refers specifically to reaching arbitrary states from the zero state.
- Null controllability means that every initial state can be driven to zero.
- Complete controllability usually means that arbitrary state-to-state transfers are possible.
These properties coincide for finite-dimensional continuous-time linear time-invariant systems with unrestricted inputs. They need not coincide in more general settings, particularly discrete-time systems with singular state-transition matrices.
Controllability concerns the internal state, not merely a measured output. Output controllability, the ability to attain arbitrary output values, is a different and potentially weaker property.
Linear time-invariant systems
For a finite-dimensional linear time-invariant system, the standard continuous-time model is
where is an system matrix and is an input matrix. Its solution at time , taking the initial time as zero, is
The first term describes unforced evolution. The integral describes the changes that inputs can produce. Controllability asks whether these changes span the entire state space.
The standard results assume real-valued inputs without amplitude bounds, normally drawn from a suitable class such as piecewise-continuous or square-integrable functions. Physical constraints can substantially change what is achievable.
Kalman rank criterion
The controllability matrix is
The pair is controllable if and only if
The columns of describe directions in which inputs act directly. The columns of describe additional directions accessible through the system's internal dynamics. A single input can therefore control multiple state variables.
If the rank is , the states reachable from zero form an -dimensional subspace of the state space. This subspace is invariant under .
Example: position and velocity
An idealized object with acceleration as its control input satisfies
With ,
Its controllability matrix is
which has rank two. Although the input acts directly only on velocity, its accumulated effect changes position. Both variables can be assigned final values through a suitable input over any positive time interval, provided input magnitude is unrestricted.
By contrast, the system
is not controllable: no input can change .
Popov–Belevitch–Hautus test
The Popov–Belevitch–Hautus, or PBH, test gives an equivalent condition:
for every complex eigenvalue of .
Equivalently, no nonzero left eigenvector of may be orthogonal to every column of . This connects controllability to individual dynamical modes: an uncontrollable mode cannot be influenced by the inputs.
Controllability Gramian and control energy
For a continuous-time linear time-invariant system, the finite-horizon controllability Gramian is
For every , the system is controllable if and only if is positive definite.
The Gramian also measures how much input effort is required. Define the displacement from the uncontrolled final state as
When the Gramian is invertible, the minimum value of
among inputs that achieve the transfer is
An input achieving this minimum is
This quadratic cost represents input effort; it equals physical energy only when the model and units justify that interpretation.
Small Gramian eigenvalues identify state-space directions that are expensive to reach. Consequently, exact controllability is not a guarantee of practical maneuverability: a mathematically controllable system may require excessive force, voltage, fuel, or time.
For an asymptotically stable , the infinite-horizon Gramian exists and satisfies the Lyapunov equation
Discrete-time and time-varying systems
For the discrete-time model
the state after steps is
Arbitrary state-to-state transfers over that horizon are possible exactly when the displayed input-to-state matrix has full row rank. Reachability in at most steps is characterized by the same Kalman matrix used in continuous time.
Null controllability alone is weaker. For example, if and , every initial state becomes zero after one step, but no nonzero state can be reached from zero.
For a continuous-time linear time-varying system,
let be its state-transition matrix. Controllability over is characterized by the positive definiteness of
Testing a Kalman matrix at individual instants is not generally sufficient: the evolution of the input directions over the whole interval matters.
Relation to observability and stabilization
Observability asks whether the internal state can be reconstructed from measured outputs. It is distinct from controllability: an input may influence a state that no sensor reveals, or a sensor may reveal a state that no input can influence.
For a linear system with output equation , observability of is mathematically dual to controllability of .
A suitable change of coordinates separates controllable and uncontrollable parts of a linear system. Under state feedback
the closed-loop system matrix becomes . For unrestricted static full-state feedback, controllability permits arbitrary placement of its eigenvalues, subject to complex-conjugate pairing for real matrices. Uncontrollable modes cannot be reassigned by such feedback.
Stabilizability is weaker than controllability. A system is stabilizable when every uncontrollable mode is already asymptotically stable. Thus feedback stabilization can be possible even when arbitrary state transfers are not. In continuous time, uncontrollable eigenvalues must have strictly negative real parts; in discrete time, they must lie strictly inside the unit circle.
A transfer function alone does not necessarily expose uncontrollable internal dynamics. Transfer functions describe input–output behavior, and nonminimal state-space realizations can contain internal modes absent from that description.
Nonlinear controllability
For nonlinear systems, controllability can depend on the initial state, target, operating region, time horizon, and permitted input signs or magnitudes.
Common distinctions include:
- Global controllability: transfers are possible between any states in the specified state domain.
- Local controllability: transfers are possible within a neighborhood, with precise definitions differing in how trajectories are restricted.
- Small-time local controllability: the reachable set contains a neighborhood of the initial state for every positive time bound.
- Accessibility: the reachable set has nonempty interior; it need not contain a neighborhood of the initial state.
For a control-affine system,
the vector fields describe direct control directions. Their Lie brackets describe additional directions obtainable through appropriately sequenced motions. Under suitable assumptions, rank conditions involving these fields and their brackets establish accessibility; stronger conclusions about controllability require further hypotheses.
This distinction is important in robotics and motion planning. A wheeled vehicle may be unable to move sideways instantaneously yet still achieve a lateral displacement through a sequence of forward, backward, and turning motions.
Controllability of a smooth system's linearization at an equilibrium is a sufficient condition for local controllability under standard assumptions, including inputs available around the equilibrium value. An uncontrollable linearization does not establish that the nonlinear system is uncontrollable, because nonlinear effects can provide additional motion directions.
Engineering applications and limitations
Controllability analysis informs actuator selection, placement, and redundancy in aircraft, robots, electrical networks, and process control. It also helps determine whether desired transfers are feasible before an optimal controller or model predictive controller is designed.
Several qualifications matter in practice:
- Input and state constraints. Saturation, unilateral forces, temperature limits, and obstacles restrict the reachable set. Unconstrained controllability does not establish constrained feasibility.
- Sampling. A controllable continuous-time plant can yield an uncontrollable sampled model at particular sampling periods.
- Numerical conditioning. Directly forming high powers of can make the Kalman rank test unreliable in floating-point arithmetic. Orthogonal decompositions and controllability staircase methods are generally more dependable.
- Model uncertainty. A nominal full-rank model may be close to losing controllability. Robust control considers uncertainty and performance questions beyond the binary rank condition.
- Network structure. Structural controllability asks whether a specified pattern of zero and potentially nonzero matrix entries permits controllability for almost all choices of its free parameters. It does not guarantee controllability for every numerical realization or acceptable control effort.
- Distributed systems. For systems governed by partial differential equations, exact controllability, approximate controllability, and null controllability may differ substantially. Finite-dimensional rank tests do not transfer directly.
Historical development
Controllability became a central concept in modern control theory through Rudolf E. Kalman's state-space work around 1960. Its development shifted attention from input–output response alone to the internal state of a system and the directions in which inputs can move it.
The associated rank criteria, duality with observability, and state-space decompositions became foundations for multivariable control. Subsequent work extended controllability theory to nonlinear, time-varying, constrained, stochastic, networked, and infinite-dimensional systems.