paper
active
1991
ⓒ 725
paper:doi-10-1007-978-1-4899-0718-9-37

Every Good Regulator of a System Must Be a Model of That System

ByRoger C. Conant·W. Ross Ashby
DOI 10.1007/978-1-4899-0718-9_37OpenAlex W4237718205
Active inference & free energy principleActive inference & free energy principlecomplex dynamic systemModel-based regulation

TL;DR

Any regulator that optimally controls a system must, as a mathematical necessity, contain a model of that system — this is the central theorem proved by Conant and Ashby, originally published in 1970 and reprinted in the 1991 volume *Facets of Systems Science* (IFSR Series, volume 7, pp. 511–519). Using set-theoretic machinery drawn from Bourbaki's 1958 *Théorie des Ensembles* and Riguet's 1948 work on binary relations (*Bull. Soc. Math. Fr.*, vol. 76, p. 114), the proof formalizes regulation as a mapping problem: a regulator R acts on a system S to produce outcomes in a goal set G, and the theorem establishes that the optimal such mapping is homomorphic to S — that is, the regulator's internal structure must mirror the causal structure of the regulated system. The method introduced is a formal set-theoretic proof of isomorphism between regulator and system, grounded in partition lattice theory as developed by Hartmanis and Stearns (1966) and prior information-theoretic work by Conant in *IEEE Trans. Systems Sci.*, vol. 5 (1969). Applied domains motivating the work include air traffic flows around New York, endocrine balances of the pregnant sheep, and money flows among banking centres. The paper argues this implies that effective control of any complex dynamic system cannot be achieved without the regulator embodying, explicitly or implicitly, a working model of that system — making model-building not merely useful but logically obligatory for good regulation.

What to take away

  1. 1. The Good Regulator Theorem states that any regulator achieving optimal (or even merely 'good') regulation of a system S must contain a model of S as a necessary mathematical consequence, not a design choice.
  2. 2. The proof is constructed using set-theoretic formalism from Bourbaki's 1958 *Théorie des Ensembles* (3rd edition, Hermann, Paris) combined with Riguet's 1948 theory of binary relations published in *Bull. Soc. Math. Fr.*, vol. 76, p. 114.
  3. 3. Conant's own 1969 paper in *IEEE Trans. Systems Sci.*, vol. 5, p. 334, on information-theoretic aspects of regulation provides a direct precursor to the algebraic formulation developed here.
  4. 4. The theorem is framed using partition lattice theory as developed by Hartmanis and Stearns in their 1966 *Algebraic Structure Theory of Sequential Machines* (Prentice-Hall), allowing the model relationship to be expressed as a homomorphism between state-transition structures.
  5. 5. Three motivating empirical substrates are cited in the abstract: air traffic flows around New York, endocrine balances of the pregnant sheep, and money flows among banking centres, illustrating the theorem's claimed universality across physical, biological, and economic systems.
  6. 6. The chapter spans pages 511–519 of the 1991 *Facets of Systems Science* volume (IFSR Series, vol. 7, Springer, Boston), making it a reprint/consolidation rather than the original 1970 journal publication, which affects citation context.
  7. 7. The proof's load-bearing methodological choice is to represent regulation formally as a set-valued mapping from system states to regulator outputs constrained to a goal set G, then show the optimal such mapping must be homomorphic to the system's own transition structure.
  8. 8. An open question raised by the theorem is whether 'partial' or approximate models suffice for near-optimal regulation, and if so, what the quantitative relationship is between model fidelity and regulatory performance degradation.
  9. 9. Sommerhoff's 1950 *Analytical Biology* (Oxford University Press) is cited as a biological grounding for the goal-directed regulation framework, linking the theorem to teleological theories of organism function.
  10. 10. The theorem implies that any learning or adaptive controller — including modern machine-learning-based controllers — must implicitly acquire a model of the controlled system during training, providing a theoretical basis for why model-based reinforcement learning outperforms model-free methods when models are accurate.

Peer brief — for seminar discussion

Conant and Ashby's contribution, reprinted in the 1991 *Facets of Systems Science* (IFSR Series, vol. 7, pp. 511–519), proves what is now called the Good Regulator Theorem: any regulator that successfully controls a system must contain, embedded in its structure, a model of that system. The theorem is derived formally using Bourbaki's 1958 set-theoretic apparatus and Riguet's 1948 binary-relation theory (*Bull. Soc. Math. Fr.*, vol. 76), with regulation represented as a mapping from the system's state space into a set of regulatory actions constrained to produce outcomes within a goal set G. The method introduced is a partition-homomorphism proof: the authors show that the optimal mapping from system states to regulator outputs induces a homomorphism between the system's transition structure and the regulator's internal structure, meaning the regulator's partitioning of its own state space must mirror that of the system. The key implication is that model-building is not a pragmatic engineering convenience but a logical necessity — a regulator without an isomorphic internal model cannot, in principle, be optimal. This extends to biological regulators (the pregnant sheep's endocrine system), engineered regulators (air traffic control around New York), and economic regulators (money flows among banking centres), all cited as motivating cases. An alternative approach that could have been used is information-theoretic: Conant's own 1969 *IEEE Trans. Systems Sci.*, vol. 5 paper frames regulation in terms of transmitted information, and one could derive a weaker version of the theorem by showing that zero-error regulation requires the regulator's channel capacity to match the system's entropy rate — though this would not establish the structural isomorphism claim. The prediction the theorem makes is that any system achieving sustained optimal regulation will, upon inspection, reveal internal structure homomorphic to the regulated system, a falsifiable hypothesis that has driven subsequent work on internal models in motor control and adaptive control theory. The most pointed critical pressure point is the scope of the optimality assumption: the theorem holds for optimal regulators, but real-world regulators are satisficing, not optimizing, and it is not shown how much structural correspondence is required for near-good-enough regulation. A critical reader would push back on whether the homomorphism requirement degrades gracefully — that is, whether a regulator with a coarse or partially incorrect model is provably worse in proportion to its model error, or whether the theorem is essentially all-or-nothing and therefore less actionable for engineering approximate controllers. The 2,101 access count on the Springer page, against only 7 formal citations in that volume, also suggests the theorem circulates more through informal influence than direct citation chains, which complicates assessing its empirical uptake.

Findings (1)

Claims (1)

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

+4 more

Similar preprints — Semantic Scholar

Cited by (3)

  • A tale of two densities: active inference is enactive inference

    Ramstead, Kirchhoff, and Friston argue that generative models in active inference under the free energy principle (FEP) are control systems—not structural representations—and that this distinction has

  • Active inference on discrete state-spaces: a synthesis

    Active inference on discrete state-spaces, formalized as partially observable Markov decision processes (POMDPs) with likelihood matrix A, transition matrix B, and prior D, unifies perception, plannin

  • Life as we know it

    Any ergodic random dynamical system possessing a Markov blanket will, almost surely, appear to engage in active inference and maintain autopoietic integrity—making biological self-organization not a r