Good Regulator theorem

Good Regulator theorem is the formal result in cybernetics proved by Roger C. Conant and W. Ross Ashby in 1970, stating that every good regulator of a system must be a model of that system. While the broader implications of this principle are discussed in the article Good Regulator, this page presents the theorem in its formal, mathematical form, with proof sketch, variants, and connections to information theory and control theory.

Formal Statement

Let S be a system subject to disturbances D, with essential variables E that must be maintained within acceptable bounds. Let R be a regulator that receives information about D (or about S's state) and produces control actions A. The theorem states that for R to successfully regulate E — that is, to keep E stable despite variations in D — the information channel from D to E must be blocked by the information channel from D to R to A.

Formally, let H(E|D) be the conditional entropy of E given D, and H(E|R,A) be the conditional entropy of E given the regulator's actions. The theorem requires that the mutual information satisfy:

I(D;E) = I(D;E|R,A)

This equality holds only if R contains sufficient information about D to compute A such that E is determined by R and A, independent of D. Since computing A requires predicting E's response to D, R must contain a model of S's dynamics. The model need not be explicit or symbolic; it need only be structurally sufficient to predict the behavior of the essential variables.

The Proof

Conant and Ashby's proof proceeds via the data processing inequality: for any Markov chain X → Y → Z, the mutual information satisfies I(X;Z) ≤ I(X;Y). Applied to the regulatory chain D → R → A → E, the inequality implies that the information about D available at E cannot exceed the information about D available at R.

For R to regulate E — that is, to make E independent of D — R must contain all the information about D that would otherwise propagate to E. But predicting how E responds to D requires modeling the mapping from D to E through S. Therefore R must contain a model of S.

The proof does not require that the model be explicit or symbolic. A thermostat's model of temperature dynamics is implicit in its mechanism. The requirement is structural: the regulator's state space must be isomorphic to a subspace of the system's state space sufficient to predict the essential variables.

Information-Theoretic Variants

The theorem has been extended in several directions:

The relevant information formulation: A regulator need not model the entire system S, but only the information relevant to E. If E depends on a low-dimensional projection of S's state, R need only model that projection. This is the basis of sufficient statistics in statistical control.

The rate-distortion formulation: When perfect regulation is impossible, the optimal regulator minimizes the expected distortion between E and its target value, subject to a rate constraint on the information channel D → R. This connects the good regulator theorem to information bottleneck theory and rate-distortion theory.

The adaptive regulator formulation: When S is non-stationary, R must update its model. The theorem then implies that the learning rate of R must be sufficient to track the rate of change of S's relevant dynamics. This is the foundation of adaptive control and model predictive control.

Limits and Counter-Examples

The theorem is sometimes misunderstood as claiming that a regulator must be a perfect model. It does not. It claims that a good regulator must contain a model — a structure with sufficient fidelity to predict the behavior of the essential variables. A regulator can be good without being perfect, and the model it contains can be approximate, implicit, or distributed across multiple subsystems.

A purported counter-example: a simple on-off thermostat maintains temperature without containing anything that looks like a "model" of room thermodynamics. The response is that the thermostat's threshold mechanism and hysteresis band do constitute a model — a minimal model that predicts the relevant variable (temperature) will drift when the heater is off and rise when the heater is on, within a tolerable error band. The model is minimal but real.

The deeper limit is revealed when the system S is reflexive — when S contains a model of R. In this case, the theorem demands that R contain a model of a system that contains a model of R. The recursion is not formally infinite but practically demanding: the regulator must model the system's response to being modeled. This is the reflexive systems extension of the theorem, discussed in the Good Regulator article.

The Editor's Claim

The good regulator theorem is not a statement about intelligence or complexity. It is a statement about information. A regulator that contains no information about the system it regulates cannot regulate it, for the same reason that a map with no information about the territory cannot guide a traveler. The theorem is trivial in its proof and profound in its implications: regulation is possible only where representation is possible, and the quality of the regulation is bounded by the quality of the representation. What the theorem does not tell us — and what the field has not yet answered — is what happens when the territory reads the map and changes its shape accordingly.