Chapter 52
Assembly Index and the Depth of the Fractal Learner
Two constructions in this book measure the same thing and have not yet been introduced to each other.
Chapter 48 identified the assembly index of a term with the minimum path length in its open synchronization tree: the fewest environment-assisted joining steps that suffice to build it from elementary material. Chapter 24 built a tower — the outside of one level is the inside of the next — and observed that an individual is a fixed point of a partition whose multiple solutions are the levels of that tower. The first is a number attached to an object by a chemist. The second is a structural fact about a mind assembled out of learners.
This chapter argues that the second is bounded by the first, and draws the consequence: assembly theory, taken seriously, gives an upper bound on what an ecology-as-mind can know.
52.1 Every level of the tower costs a joining step
Recall the generating step of Chapter 24. A medium is a channel over which the parts of a region communicate, and \(@C\) — the quotation of a medium — sits one level up the reflection tower from the names it carries, because \(C\)’s free names are the endpoints it joins. Proposition 24.6 says that the media of one level are the channels of the next, so ascending the tower means constructing, at each level, a carrier for the names of the level below.
That construction is not free, and its cost is exactly the kind of cost the assembly index counts.
Let \(E\) be an ecology assembled from elementary terms, and let \(h(E)\) be the height of its medium tower in the sense of Proposition 24.6. Then \[h(E) \;\leq\; \AI(E),\] where \(\AI\) is the assembly index of Proposition 48.1.
Ascending one level of the tower requires exhibiting a medium \(C\) whose endpoints are names of the level below. By the no-self-code theorem a name codes a strictly smaller term, so \(C\) is not among the terms it carries and must be constructed. Its construction is a transition in the open synchronization tree: it joins previously present material at a minimal context, which is what an edge of \(\mathcal{ST}_o\) is. Hence each level contributes at least one edge to any path from the elementary terms to \(E\), and in particular to a minimum-length such path. Summing over levels gives \(h(E) \leq \AI(E)\).
The inequality is not tight, and the slack is where the biology is. An organism may spend a great many joining steps on structure that contributes no new level — redundancy, repair machinery, sheer bulk — and Proposition 24.12 says it must spend some, because a medium beyond the profitable radius carries no traffic and a level with no traffic is not a level. So \(\AI\) counts everything built and \(h\) counts only the building that bought a new inside. The ratio \(h(E)/\AI(E)\) is therefore a measure of architectural efficiency, and one would expect it to be small for a crystal, larger for a bacterium, and larger still for a nervous system. We do not know how to compute it for any real organism, and we note that the quantity is at least in principle measurable, which is more than can be said for most proposed complexity measures.
52.2 Depth bounds what can be told apart
The bound becomes interesting when combined with what the learner’s hypothesis language can express.
Chapter 21 established that a mortal scientist’s hypotheses are formulae of a logic of namespaces rather than of individuals, because a specimen affords exactly one measurement and the measurement ends it. Chapter 16 gave that logic its modal operators, indexed by minimal contexts. The two facts together mean that a hypothesis is a modal formula whose nesting alternates over namespace boundaries.
A learner in an ecology \(E\) can form and evaluate hypotheses of namespace nesting depth at most \(h(E)\).
A formula of nesting depth \(d\) quantifies over \(d\) distinguishable namespace levels. Evaluating it requires access to a percept that has traversed those levels, which by Definition 24.7 is a percept of integration depth \(d\). By Proposition 24.9 such a percept survives with probability \(p^d\) and costs \(d\) metered rendezvous. If \(d\) exceeds \(h(E)\) there are no further levels for it to traverse, and the formula distinguishes nothing the shallower formulae did not.
Now recall the standard fact about modal logic that makes this bite: formulae of Hennessy–Milner nesting depth \(n\) characterize exactly \(n\)-step bisimulation. Two processes agreeing on all formulae of depth \(n\) are indistinguishable by any experiment of \(n\) rounds, however cleverly designed.
Let \(E\) be an ecology-as-mind with assembly index \(\AI(E)\). Then \(E\) cannot distinguish two environments that are \(\AI(E)\)-step bisimilar. Equivalently, the ontology available to \(E\) is at best the quotient of its world by \(\AI(E)\)-step bisimulation.
Compose Proposition 52.1, Proposition 52.2, and the Hennessy–Milner characterization.
Theorem 52.1 is not a statement about Turing degree, and it would be a serious misreading to take it as one. The rho calculus is Turing complete, and so is every ecology built in it, at every assembly index above the trivial: a single sufficiently clever term computes everything computable, given unbounded time and unbounded tokens. Assembly index does not bound what \(E\) could compute in principle.
What it bounds is what \(E\) can tell apart, and that is the notion of power the rest of this book has been using. Chapter 21 showed that the equivalence a budgeted learner can afford is coarser than the one that is true, and that the gap is economics rather than ignorance. Theorem 52.1 says how much coarser, in terms of a quantity a chemist can in principle measure. The bound is on resolution, not on reach.
52.3 Copy number is the other axis
Assembly theory carries two parameters, and so far only one has been used. The second is copy number, given a rigorous definition in Definition 48.4 as the count of bisimilar processes sitting at channels within a namespace region.
At first sight it does something orthogonal. Where assembly index bounds how many levels the tower has, copy number bounds how many independent samples the learner can take at any level — which is to say, how much statistical confidence it can buy. Chapter 21 required this without naming it: hypotheses are about kinds because a specimen affords one measurement, and confidence about a kind requires many specimens of it.
The orthogonality does not survive contact with the price of depth, and the next chapter is where it fails. The short version is that copies are not an independent choice a lineage makes alongside architecture. They are forced by it: a percept that must cross \(d\) levels arrives with probability \(p^{d}\), so the samples needed to sustain a hypothesis at that depth grow exponentially in the depth, and a deep architecture that nobody can afford to populate is not an architecture in use. Proposition 53.9 makes this a curve in the \((\AI, \CN)\) plane whose slope is fixed by medium reliability. The two parameters of the biosignature are coordinates on a frontier, not independent readings.
The two parameters have different economics, and the difference is familiar from Chapter 24. Depth is exponentially fragile: reliability \(p^d\), so each additional level costs geometrically more in retransmission and repair. Width is linear: twice the copies, twice the samples, twice the tokens. An ecology under resource pressure will therefore trade depth away first, and one predicts — as an empirical matter, not a theorem — that lineages entering a period of scarcity should lose architectural levels before they lose population. Assembly theory’s joint claim, that high assembly index together with high copy number is the biosignature, reads in this framework as the observation that only a system with a metabolism can afford to buy both at once — and, once Proposition 53.9 is available, as the stronger observation that buying the first obliges you to buy the second.
52.4 Apertures, and the connection to choice
The Prestige will take on an obligation, and this is the place where the Turn can supply what it needs.
The argument there — Chapter 57, which a reader may not have reached — is that whatever exceeds Turing computation in a physical system must be spent at the apertures: the cuts where contention is observable and a scheduler must pick. That argument asserts, and does not establish, that the number of apertures grows with the structural elaboration of a population rather than with its raw size, and that the thing which measures elaboration is the assembly index. Both halves are available here, so we establish them here and let the Prestige cite them.
Proposition 52.1 is what makes that precise. Apertures live at boundaries between levels, because a cut internal to a level with a single owner is resolved by that owner and carries no observable choice. So the count of aperture classes is bounded by the number of level boundaries, which is \(h(E) - 1\), which is bounded by \(\AI(E)\).
The number of distinguishable aperture classes in an ecology \(E\) is at most \(\AI(E) - 1\). Consequently, if Conjecture 57.1 holds, the choice strength an ecology can exhibit is bounded by a function of its assembly index.
This is the sharpest form of the framework’s claim about mind, and it is worth stating plainly. A thing that has been assembled by \(n\) joining steps can have at most \(n\) levels of inside, can afford hypotheses of at most \(n\) nested namespaces, can resolve its world at best to \(n\)-step bisimulation, and can host at most \(n-1\) kinds of genuine indeterminacy. Mind is bounded by manufacture.
52.5 Open gaps
Three, and they are not small.
The first is that Proposition 52.1 identifies a level of the medium tower with at least one joining step, but does not show that the joining steps for distinct levels are distinct edges of a single minimum-length path. If a single construction could serve two levels the count would collapse. Proposition 53.5 closes this whenever the levels are rooted at pairwise disjoint scopes, which by Proposition 23.9 is something a lineage can arrange; what remains open is the general case, where the strata overlap and a single construction might indeed be charged twice. We suspect that case is genuinely harder rather than merely unwritten.
The second is that Proposition 52.2 treats namespace nesting depth and modal nesting depth as the same number. They are not obviously the same number. The identification is exactly the sort of thing the OSLF construction of Chapter 19 should settle, and until the manufacture of the logic is written down rather than described, the identification is a plausible assumption rather than a fact.
The third is that assembly index is defined for an object and an ecology is not obviously an object. Chapter 24 argued that an individual is a fixed point of a partition, and that argument gives us the right to speak of \(\AI(E)\) only once a solution to the fixed-point condition has been chosen. Different solutions give different indices, and we have said nothing about how they compare. This is the same circularity that made individuation a fixed point rather than a construction, arriving now in a place where it costs a number. Chapter 53 adds one observation about it and no solution: the choice of a solution is the choice of a generator for the ecology’s scope, in the sense of Definition 18.2, and the levels of the tower are its strata. So the question of which \(\AI(E)\) one means is the question of which generator, and the two circularities — which region, and what can be afforded to survey it — are one circularity.