Good Regulator, Goals, and Channels

Neil D. Lawrence

FW26, William Gates Building

This Session

  • Unmarked review of Worksheet 3 submissions
  • Good Regulator overview: \(D,S,A,Z\); \(\min H(Z)\); \(H(A\mid S)=0\); caveats
  • Goal as change of measure: \(r=-\ln(\mathrm{d}\nu/\mathrm{d}\lambda)\)
  • Shannon \(R\le C\); DPI statement (Week 6 overflow)

Submission Review

Unmarked review. Patterns across the cohort, not individual grades.

  • Curiosity vs slogan collapse (LLM as Grasshopper)
  • Coarse-graining Maxwell: compartments and two ledgers
  • What good Skepticus looks like: press existence claims and conflations
  • Worksheet 4 already out — due 1 December before Quiz 4

Good Regulator Overview

Good Regulator Overview

Notation: \(D\), \(S\), \(A\), \(Z\)

The CGF was \(\psi(\boldsymbol{\theta})\). After the break, \(A\) is action.

  • \(D\): disturbance
  • \(S\): system state
  • \(A\): action (the regulator’s response; Conant–Ashby often write \(R\))
  • \(Z\): reguland / outcome (Ashby’s essential variable)

Two Setups

  • Cause-controlled: act on \(D\) (or on a model of \(D\))
  • Error-controlled: act on the state \(S\) that \(D\) caused
  • DPI: \(I(A;D)\le I(S;D)\) when \(A\) only sees \(S\)

What ``Good’’ Means Today

Success is low entropy in the reguland — with two caveats.

  • Objective sketch: \(\min_{\pi(a\mid s)} H(Z)\)
  • Slogan: every good regulator must be a model of the system
  • Weak reading of ``model’’: \(H(A\mid S)=0\), i.e. \(A=h(S)\)
  • Full theorem and caveats: Week 8 review

Requisite Variety

Quantity bound, not the structure theorem.

  • \(H(Z)\ge H(D)-I(D;A)\)
  • Useless variety: large \(H(A)\) with small \(I(D;A)\)
  • Ashby: how much regulatory information is required
  • GRT (Week 8): how that information may be structured

Name the existence claim. Fence the slogan.

  • Among \(H(Z)\)-minimisers there exists \(\pi\) with \(H(A\mid S)=0\)
  • Caveat 1: constant \(A=a_0\) also has \(H(A\mid S)=0\) (weak ``model’’)
  • Caveat 2: low \(H(Z)\) is predictability, not desirability
  • Concavity proof stays notes — do not board \(g(\alpha)\) today

Goals as Radon–Nikodym

Week 5: relative entropy needs a reference. Goals need one too.

  • Channel: occupancy over \((S,A)\) pushes to an outcome law on \(Z\)
  • Criterion: \(\min_\pi H(Z)\) — or \(H_\nu(Z)\) under a goal measure \(\nu\)
  • Base \(\lambda\); goal \(\nu\ll\lambda\); density \(\mathrm{d}\nu/\mathrm{d}\lambda\)
  • Reward reading: \(r(z)=-\ln(\mathrm{d}\nu/\mathrm{d}\lambda)\)

The map recovers reward for a reward channel. It does not license general intelligence.

  • Reward channel: entropy criterion collapses to return maximisation (under a volume threshold)
  • Regulation channel: hold a set-point with no scalar reward
  • Existence of a deterministic organiser \(\neq\) a rich world-model
  • Worksheet 4: keep slogan, existence claim, and AGI talk apart

Shannon Capacity and DPI

Shannon’s Noisy Channel Coding Theorem

A channel is a conditional \(p(y\mid x)\). Rate cannot beat capacity.

  • Mutual information \(I(X;Y)\) is the channel currency
  • Capacity \(C = \max_{p(x)} I(X;Y)\)
  • No-go: reliable communication requires \(R \le C\)
  • Prescription: the input \(p(x)\) that achieves \(C\)

Scaffolding, Not Outcomes

Chain Rule, Mutual Information, Capacity

Four results we will need later — stated, not proved today.

  • Chain rule: \(H(X,Y) = H(X) + H(Y|X)\)
  • Mutual information: \(I(X;Y) = H(X)-H(X|Y) = H(X)+H(Y)-H(X,Y)\)
  • Capacity: no rate above \(C\); achieving \(C\) requires the capacity-achieving \(p(x)\)
  • Data processing: if \(X\to Y\to Z\) is Markov, \(I(X;Z)\le I(X;Y)\)

Data-Processing Inequality

If \(X\to Y\to Z\) is Markov, processing cannot create information about \(X\).

  • Markov: \(p(z\mid x,y)=p(z\mid y)\)
  • Statement: \(I(X;Z)\le I(X;Y)\)
  • Soft partitions (Week 5): leftover \(H(G\mid X)\) already spent information
  • Full proof: notes / Week 8 enrichment — not today’s board

Notes Overflow: Multi-Information, DPI Proof, IB

Optional overflow: Multi-Information

Sampling Two Dimensional Variables

Correlation

  • Correlation is when two variables are dependent

Sampling Two Dimensional Variables

Independent Gaussians

Correlated Gaussians

Anticorrelated Gaussians

The Classical Observer

The Classical Observer - Correlated

The Classical Observer - Anti-correlated

Optional overflow: DPI Proof and Information Bottleneck

Data-Processing Inequality

Data-Processing Inequality

If \(X\to Y\to Z\) is a Markov chain, processing cannot create information about \(X\).

  • Markov: \(p(z\mid x,y)=p(z\mid y)\), equivalently \(I(X;Z\mid Y)=0\)
  • \(I(X;Y,Z)=I(X;Y)+I(X;Z\mid Y)=I(X;Y)\)
  • \(I(X;Y,Z)=I(X;Z)+I(X;Y\mid Z)\ge I(X;Z)\)
  • Therefore \(I(X;Z)\le I(X;Y)\)

KL Decay and Grouping

Under Markov dynamics, \(\text{KL}\left( p_t\,\|\,\pi \right)\) falls — coarse-graining in time.

  • Physics often says ``relative entropy’’; same object as week-3 / week-6 KL
  • Data-processing for distributions: each step cannot increase divergence to equilibrium
  • Same grouping split as the die, now for approach to \(\pi\):

\[ \text{KL}\left( p\,\|\,\pi \right)=\text{KL}\left( w\,\|\,w_\pi \right)+\sum_i w_i\, \text{KL}\left( p_i\,\|\,\pi_i \right) \]

Arrow of Time from the Same Split

How much of equilibration is between groups, how much within?

  • Between: \(\text{KL}\left( w\,\|\,w_\pi \right)\) — redistribution of macro probability
  • Within: \(\sum_i w_i \text{KL}\left( p_i\,\|\,\pi_i \right)\) — relaxation to local \(\pi_i\)
  • Demon / Crooks / transport: all move mass among groups while baths handle the rest

Information Bottleneck

Information Bottleneck

Least-committal \(T\) given a relevance constraint — MaxEnt, now on mutual information.

  • Markov: \(Y \leftrightarrow X \leftrightarrow T\)
  • Minimise \(I(X;T)\) subject to \(I(T;Y)\ge R\)
  • Lagrangian: \(I(X;T)-\beta I(T;Y)\)
  • DPI bound: \(I(T;Y)\le I(X;Y)\)

The Inaccessible Game

  • System isolated from observation
  • External observer cannot extract information
  • Internal state is inaccessible
  • Zero-player game with information-theoretic rules

Why “Inaccessible?”

Why “Inaccessible?”

From BLF axioms: Info gained = \(H(p) - H(q)\)

Our axiom: \(\sum h_i = C\) (constant)

\[\Delta(\sum h_i) = 0 \Rightarrow \text{observer learns nothing!}\]

What Makes It a Game?

Game Characteristics:

  • Zero-player game
  • State = probability distribution \(p(\mathbf{x}|\boldsymbol{\theta})\)
  • Rule = maximize entropy production
  • Constraint = \(\sum h_i = C\)
  • Dynamics = emerge from information geometry

Connection to Physical Reality

Physical Connections:

  • GENERIC structure emerges
  • Energy ↔︎ Entropy equivalence
  • Landauer’s principle derivable
  • Bridge between information and physics

Define This Week

  • Name \(D\), \(S\), \(A\), \(Z\) — and what \(\min H(Z)\) means
  • Why \(H(A\mid S)=0\) is a weak reading of ``model’’
  • How does \(r=-\ln(\mathrm{d}\nu/\mathrm{d}\lambda)\) connect goals to Week 5’s RN/KL language?
  • What does \(R\le C\) forbid, and what does DPI state?

After This Lecture

  • Worksheet 4 due 1 December, before Quiz 4
  • Quiz 4: first ten minutes of Week 8
  • Week 8: limits synthesis; no-barber \(\to\) quantum motivation; course close

Further Reading

  • the Good Regulator Theorem of Conant and Ashby (1970)

  • requisite variety; Shannon form of Ashby (1956)

  • the Good Regulator Theorem of Conant and Ashby (1970)

  • requisite variety of Ashby (1956)

  • noisy channel coding / capacity of Shannon (1948)

  • Chapter 7 of Cover and Thomas (1991)

  • Theorem 2.8.1 of Cover and Thomas (1991)

  • the method; optional of Tishby et al. (1999)

Thanks!

References

Ashby, W.R., 1956. An introduction to cybernetics. Chapman & Hall, London.
Conant, R.C., Ashby, W.R., 1970. Every good regulator of a system must be a model of that system. International Journal of Systems Science 1, 89–97.
Cover, T.M., Thomas, J.A., 1991. Elements of information theory. Wiley, New York.
Shannon, C.E., 1948. A mathematical theory of communication. The Bell System Technical Journal 27, 379–423. https://doi.org/10.1002/j.1538-7305.1948.tb01338.x
Tishby, N., Pereira, F.C., Bialek, W., 1999. The information bottleneck method, in: Proceedings of the 37th Annual Allerton Conference on Communication, Control and Computing. pp. 368–377.