In which emergence turns out to be a theorem.
Three constructions from your existing toolkit share a shape you may not have named.
Generated subgroups. Take elements a, b in a group G. The subgroup ⟨a, b⟩ is not assigned—it is forced. You declare generators; the group operation determines what must be included. The subgroup is the least fixed point of closure under multiplication and inversion. You specified a rule; the domain assembled itself.
Eigenspaces. Given a linear operator T : V → V and a scalar λ, the eigenspace Eλ = ker(T − λI) is compelled by the operator, not chosen by the analyst. You wrote down a relation between T and λ; the subspace satisfying it materialized as the solution set. Again: relation first, domain second.
Algebraic closure of Q. No finite list of adjoined roots produces it. You state a universal property—every non-constant polynomial splits—and the field satisfying that condition is the minimal such extension. The property dictates the domain; the domain does not precede the property.
In each case, you specified a relation or property, and the domain fell out as a fixed pointA value unchanged by an operation — like a config that a convergence loop returns unmodified.
Consider a small reaction network built from six molecule types and five reactions. Two molecules are designated food—available in unlimited supply from outside the system. The remaining four must be produced internally or not at all.
Molecules: f1, f2 (food, shown in blue); a, b, c, d (non-food, shown in green or grey).
Visual conventions: circles represent molecules, diamonds represent reactions. Solid arrows connect reactants to reactions and reactions to products. Dashed red arrows indicate catalysis—which molecule enables which reaction without being consumed.
| Reaction | Reactants | Product | Catalyst |
|---|---|---|---|
| r1 | f1 + f2 | a | b |
| r2 | a | b | a |
| r3 | f1 + a | c | b |
| r4 | a + b | c | c |
| r5 | d + f2 | a | d |
Two features deserve attention. First, d is an orphan: it appears as reactant and catalyst in r5, but no reaction in the network produces it. Second, r1 and r2 form a mutual dependency: r1 produces a but requires b as catalyst, while r2 produces b but requires a as both reactant and catalyst. Neither can fire unless the other already has.
Which reactions can sustain themselves from the food set alone?
Your training is clear on procedure: fix the domain, then derive consequences. Choose a group G, then classify its normal subgroups. Pick a manifold M, then compute its cohomology. The domain is always the starting point; the results follow from it.
This reaction network refuses that workflow. What molecules are present depends on which reactions can fire. Which reactions can fire depends on whether their reactants and catalysts are available. Whether those catalysts are available depends on which other reactions have already produced them—which brings you back to the first question. You feel a circularity, and your instinct says: ill-defined.
But reframe the question. You are not looking for a domain to work over. You are looking for a subset R′ ⊆ R that is a fixed point of a specific operator on ℘(R). The apparent circularity is simply the self-referential character of any fixed-point equation. You solve R′ = Φ(R′).
Enumerating all 2|R| subsets of R to find R′ = Φ(R′) is exponential. But Φ is monotoneOrder-preserving: larger input always produces larger-or-equal output. Guarantees convergence when iterating on a finite latticeA set with a partial ordering where any two elements have a unique max (join) and min (meet) — a generalized hierarchy with guaranteed merge operations, so we can compute the greatest fixed point by iterating from the top: start with R′ = R (all reactions), compute the closureAll elements reachable from a starting set under given operations — like computing all reachable nodes in a dependency graph clR′(F), remove every reaction whose reactants or catalyst fall outside that closure, and repeat. Each step either removes at least one reaction or converges. The process terminates in at most |R| iterations.
The key observation: if R′ ⊇ R″, then clR′(F) ⊇ clR″(F), so Φ(R′) ⊇ Φ(R″). This makes Φ a monotone map on (℘(R), ⊇). Tarski's fixed-point theoremGuarantees that any order-preserving function on a complete lattice has both a greatest and a least fixed point — existence by structure, not by search guarantees a unique greatest fixed point, which is exactly what the pruning algorithm converges to.
Step through it on the puzzle from Section 2:
The operator Φ(R′) = {r ∈ R′ : ρ(r) ⊆ clR′(F) ∧ ∃x ∈ clR′(F), (x,r) ∈ C} is monotone on (℘(R), ⊇). Because the lattice is finite, Tarski guarantees not merely existence but uniqueness of the greatest fixed point. The pruning algorithm computes it by iterating Φ downward from R. Each application either strictly shrinks the set (removing at least one reaction) or returns the same set (convergence). The result is the maxRAF—the largest self-sustaining, autocatalytic subset of the network.
A Catalytic Reaction System (CRS) is a triple Q = (X, R, C), where X is a finite set of types, R is a set of reactions (each r having a reactant set ρ(r) ⊆ X and a product set π(r) ⊆ X), and C ⊆ X × R is the catalysis relation.
Given a food set F ⊆ X, a non-empty subset R′ ⊆ R is a Reflexively Autocatalytic, F-generated set (RAF) if:
where clR′(F) is the closure of F under repeated application of reactions in R′ whose reactants are available.
Read this definition carefully. No intrinsic property of the elements of X appears. There is no mass, no charge, no color, no function. The entire definition is stated in terms of three maps: ρ (what goes in), π (what comes out), and C (what enables what). The identity of each x ∈ X is exhaustively determined by its relational position in the network—what it catalyzes, what reactions produce it, what reactions consume it. Nothing else participates in the definition.
This is 100% relational mathematics. And because it is purely relational, it is substrate-independentThe mathematics works regardless of physical implementation — like how TCP/IP runs identically over copper, fiber, or radio. Any system that instantiates the same graph structure—the same pattern of reactant, product, and catalysis relations—is formally the same CRS. The proof machinery transfers wholesale: same monotone operator, same lattice, same fixed-point theorem. Only the labels change.
| Abstract | Chemistry | Ecology | Knowledge | Governance |
|---|---|---|---|---|
| x ∈ X | Molecule type | Species | Concept / technique | Institution / role |
| r ∈ R | Chemical reaction | Ecological interaction | Derivation / synthesis | Policy process |
| ρ(r) | Reactants consumed | Resources required | Prerequisite concepts | Required inputs |
| π(r) | Products generated | Outputs produced | New results derived | Decisions / outputs |
| (x,r) ∈ C | x catalyzes r | x enables interaction | x enables derivation | x authorizes process |
| F | Food molecules | Abiotic resources | Axioms / observations | Constitutional basis |
| RAF | Self-sustaining set | Stable ecosystem | Self-sustaining field | Viable governance |
The word "emergence" carries philosophical baggage. Set it aside. What follows is a theorem about lattice structure.
A RAF R′ is irreducibleCannot be broken into smaller parts that each have the same property — like a prime number or an atomic service that can't be decomposed further (an irrRAF) if no proper non-empty subset of R′ is itself a RAF. Equivalently: removing any single reaction from R′ destroys the RAF property.
Let R′ be an irrRAF. The property "self-sustaining and autocatalytic" holds for R′ but fails for every proper subset of R′.
Proof sketch. By definition of irreducibility, no proper non-empty subset of R′ is a RAF. The empty set trivially fails to be a RAF (it sustains nothing). Therefore the self-sustenance property is held by R′ and by no R″ ⊊ R′. □
This is the precise algebraic content of emergence: a property P of a composite S that provably fails for every proper sub-composite. It is a statement about the lattice of subsets of R, verifiable by the pruning algorithm. No interpretation is required; the lattice structure does the work.
The collection of all RAFs within a CRS forms a partially ordered set under inclusion, with the maxRAF at the top. The irrRAFs sit at the bottom—the minimal non-empty elements. Each irrRAF marks the precise boundary where a collective property first crystallizes: below it, self-sustenance is impossible; at it, self-sustenance appears as an irreducible whole.
Kauffman (1986)2 conjectured that in sufficiently complex catalytic networks, self-sustaining subsets appear with near-certainty. Mossel & Steel (2005)3 proved the conjecture for the binary polymer model with independent catalysis: in a random CRS where each molecule catalyzes each reaction independently with probability p, the probability that a RAF exists undergoes a sharp transition near p ∼ 1/n, where n = |X|.
Below the threshold: no subset of R is self-sustaining. No amount of searching the lattice yields a RAF. Above: irreducible self-sustenance appears. The transition sharpens as n grows—a familiar signature of phase transitionsAbrupt qualitative change at a critical threshold — like water freezing, or a network flipping from fragmented to connected at a percolation point in combinatorial structures. The passage from "no collective property anywhere" to "irreducible collective property guaranteed" is not gradual. It is a threshold phenomenon in the lattice, as abrupt as percolationA threshold phenomenon where global connectivity suddenly appears in a network as enough local connections form on a graph.
You didn't need to define the domain first. You needed to specify the relations—reactant maps, product maps, catalysis pairs, food set—and let a monotone operator on a finite powerset lattice find the domain for you. The fixed point is the answer. Tarski guaranteed its existence; the pruning algorithm computed it in polynomial time. Irreducibility gave emergence an algebraic definition. The phase transition told you when to expect it. None of this was philosophy. It was the same lattice theory, the same fixed-point machinery, the same structural reasoning you already use—applied to a graph you hadn't seen before. That is a theorem, not an argument. The substrate does not matter. Same lattice, same fixed point, same emergence.
References.
Autocatalytic Closure for the Working Mathematician — B. Stephenson, 2026.