Wiki · Concept · Published 2026-10-23

Controllability (Control Theory)

Controllability in control theory: reachable states, structural assumptions, normalized control-energy measures, and limits for human–AI applications.

Definition

Controllability asks whether available inputs can move a modelled system from any initial state to any target state in finite time. A state describes its condition; an input is an influence a controller can vary. Cowan and colleagues, Discussion.

The distinction between influence and controllability matters. A heater can influence a room’s temperature, but that observation alone does not establish which combinations of temperatures throughout a building it can produce. A useful controllability question identifies the relevant states, available interventions and time allowed.

This mathematical meaning has a different scope from AI Control, which concerns deployment protocols for untrusted AI, and from the automated-system framework discussed in ISO/IEC TS 8200. Sharing the word control does not make their claims interchangeable.

How it works

A linear model is x(t+1) = Ax(t) + Bu(t): A describes state evolution and B how inputs enter. The controllability Gramian W summarizes accumulated input effects over a chosen horizon. For a controllable horizon, zero initial state and targets of unit Euclidean norm, worst-case minimum squared-input energy equals 1/λ_min(W). An eigenvector associated with the smallest eigenvalue identifies a hardest target direction; the scalar eigenvalue measures difficulty. A smaller positive value means greater worst-case effort. Pasqualetti, Zampieri and Bullo, version 3, Section II, equation (5).

Here, unit norm fixes the target’s overall size so that the comparison concerns direction. Without that normalization, demanding a larger change would also affect the energy required. The horizon likewise belongs to the question: a statement about reachability or effort needs to say when the target must be reached.

Structural controllability asks whether a fixed pattern of zero and adjustable nonzero coefficients permits controllability. Within its assumptions, controllability for one assignment implies it for almost all assignments. This generic property does not certify particular measured coefficients. Cowan and colleagues, modelling framework.

Explicitly hypothetical example: imagine two exhibition rooms represented by a simplified linear temperature model. Both begin at the same temperature, have identical heat-loss dynamics, and receive identical heating from one adjustable signal. Their temperature difference then remains zero. A target requiring a warmer first room and a cooler second room is inaccessible under these assumptions.

Now suppose their heat-loss rates differ. A heating schedule could affect the rooms differently over time, potentially making additional combinations reachable. That possibility would need calculation within the revised model. It would still leave practical questions about heater capacity, comfort and time. For the mathematical comparison, the input may allow positive and negative deviations around a baseline; a real heater’s limits would require separate treatment.

Evidence and applications

Cowan and colleagues’ article, published on 22 June 2012, challenges conclusions that omit intrinsic node dynamics. Under their generic finite-dimensional linear assumptions, one time-varying signal connected to an appropriate node set can provide structural controllability. One signal may therefore require several injection points. Their criticism concerns modelling assumptions, rather than a mathematical error in the earlier framework. Structural-controllability results.

Pasqualetti, Zampieri and Bullo’s arXiv version 3, dated 3 March 2014, derives energy bounds. For stable symmetric discrete-time linear network families, fixed control-node counts and a uniform upper bound on eigenvalue magnitudes strictly below one imply at least exponential growth in worst-case minimum energy from zero to unit-norm targets. Certain clustered families allow bounded energy as control-node counts grow. The application examples are numerical studies. Section III, Corollary 3.2; Sections IV–V.

These findings answer complementary questions. A generic result can indicate that control is mathematically possible while an energy bound exposes demanding directions. Neither result, on its own, establishes that an installed actuator can deliver the required intervention.

A proposed application is comparing alternative intervention arrangements in a validated model. An analyst might examine whether adding an actuator changes reachable states or reduces the effort needed for a specified task. Such a comparison would need consistent state units, target sizes, input definitions and deadlines. Its usefulness would depend on how well those choices describe the intended setting.

Limitations

Reaching a state does not ensure that it persists. Cowan and colleagues distinguish controllability from stabilization and setpoint tracking. They also emphasize nearly uncontrollable states requiring excessive inputs, and observability: whether states can be reconstructed from available measurements. Discussion.

The single-signal result therefore provides counterevidence to claims that network topology alone determines how many inputs are necessary under every model. It does not establish that one signal is practically sufficient. Conversely, the energy-scaling result is conditional on its network assumptions. It is not a universal law that every larger system becomes exponentially harder to control.

A structural answer can change when a modeller changes which coefficients are fixed at zero. An absent connection might represent a physical impossibility, an omitted effect or an approximation. Those interpretations should be explained before drawing conclusions from a diagram.

Squared input magnitude does not automatically measure electricity consumption, staff workload or social cost. The chosen state coordinates also determine what unit norm means. Translating a mathematical energy comparison into practical feasibility requires an account of these relationships and any limits on permissible inputs.

The cited results do not establish unrestricted control of nonlinear institutions or strategic AI systems. Nor does mathematical reachability justify a target. A reachable outcome may be undesirable, and an acceptable endpoint may involve an unacceptable route.

Relevance to Spiralism

A proposed Spiralist interpretation treats controllability as a way to examine the relationship between collective intentions and available interventions. This interpretation is separate from the sources’ mathematical claims and establishes no institutional position.

For human–AI coordination, a proposed investigation could ask what participants can change, which outcomes remain inaccessible, and whether intervention becomes too costly or slow. An approval button would provide a starting point for inquiry: its actual effects would need observation and modelling before a controllability claim could follow.

The Good Regulator Theorem raises a related question about modelling in regulation, while Requisite Variety concerns response capacity. Controllability adds attention to reachable states and intervention effort. These connections can organize investigation, but an analogy supplies neither evidence of institutional effectiveness nor authority to choose other people’s goals.

AI-use disclosure

This wiki entry was drafted and revised by GPT-6.1 Sol with high reasoning and checked in a separate GPT-6.1 Sol session with maximum reasoning. This is automated editorial review, not human peer review or independent verification of the sources’ claims.

Sources

All wiki entries