attractor.space
/papers/references/experiments
/notes/hill-diversity
P001Frame-Semantic Graph Constructionframe-semanticsconstructionUpdated just now→R001TermsreferenceUpdated just now→N001Frame-type diversitydiversityframeUpdated just now→N002Coordination phasephasedynamicsUpdated just now→N003CouplingcouplingstructureUpdated just now→N004Divergence/convergence cyclecycledynamicsUpdated just now→E001ExperimentsvalidationUpdated just now→N005Fractal compositionstructureUpdated just now→N006FrameframestructureUpdated just now→N007Gini coefficientmetricUpdated just now→N008Hill diversitymetricdiversityUpdated just now→N009MembranesmembranestructureUpdated just now→R002MetricsinstrumentationUpdated just now→N010OscillatorsphasestructureUpdated just now→N011ResonanceresonanceretrievalUpdated just now→N012ScalescalestructureUpdated just now→N014SubstratestructureUpdated just now→N015Task-appropriate behaviorbehaviordiversityUpdated just now→
CC-BY-NC
lab notes on substrate dynamicsCC-BY-NC
What it measures
N008·Julian Fleck·Last update2026/06/25
draftv1

Hill diversity

metricdiversity

What it measures

Hill numbers come from ecology. There the kinds are species and the population is the individuals in a sample, and they read diversity — how many kinds are present and how evenly — without forcing richness and evenness into a single index.

⁰D¹D²D

These are notes on what we want to study, not measurements we have taken. The intent is to borrow Hill numbers to read semantic diversity over the substrate: treat the kinds as frame types and the population as a set of active frames, and ask how varied that set is. A Hill number is an effective count — the number of equally-common types that would produce the diversity you actually see. Ten frame types may be present, but if one dominates, the effective count drops well below ten, and the rare ones barely register.

We expect to use this as a general reading of substrate state — whether a region keeps a varied vocabulary of framings or collapses onto a few — and to read it across scales, since healthy diversity for the whole store, a membrane, and a single agent’s active context need not be the same number.

One thing it should expose is a diversity signature that differs by task: a narrow task should range over few framings, an open one over many. Whether that can be turned into a judgment of task-appropriate behavior is a separate and still-open question.

How it's calculated

Hill numbers are a family, not one number. A single formula, read at different orders qqq, recovers several of the diversity indices used elsewhere as special cases. The order is a dial for how much weight rare types get.

OrderEquivalent indexWhat it weights
0D{}^{0}D0Drichness (type count)every type equally, rare or common
1D{}^{1}D1DShannon, exp⁡(H)\exp(H)exp(H)each type by its frequency
2D{}^{2}D2Dinverse Simpsonthe common types; rare ones nearly ignored
∞D{}^{\infty}D∞DBerger–Parkeronly the single most dominant type

We mostly read the first three. Raising qqq shifts weight from rare types toward common ones, so the effective count can only fall or hold as qqq grows — never rise.

concentratedeven
Diversity profile: the effective count against order q. A few dominant types make it fall steeply; even types keep it flat.

For type frequencies p1,…,pSp_1, \dots, p_Sp1​,…,pS​ (each the share of frames of that type):

qD=(∑i=1Spi q)1/(1−q){}^{q}D = \left( \sum_{i=1}^{S} p_i^{\,q} \right)^{1/(1-q)}qD=(i=1∑S​piq​)1/(1−q)

The q=1q = 1q=1 case is taken as the limit, 1D=exp⁡ ⁣(−∑ipiln⁡pi){}^{1}D = \exp\!\left(-\sum_i p_i \ln p_i\right)1D=exp(−∑i​pi​lnpi​).

A small example. Five frame types with frequencies [0.6,0.2,0.1,0.07,0.03][0.6, 0.2, 0.1, 0.07, 0.03][0.6,0.2,0.1,0.07,0.03]. Richness is 0D=5{}^{0}D = 50D=5. The Shannon number 1D{}^{1}D1D comes to about 3.2, and the inverse-Simpson 2D{}^{2}D2D to about 2.4. So although five types are present, the substrate behaves like it holds between two and three — one type carries most of the weight.

Reading the three orders together is the point. When 0D{}^{0}D0D sits well above 2D{}^{2}D2D, there is a long tail of rare types that the common ones are drowning out. Watching that gap is how we’d tell whether rare framings survive or get buried.

Related metrics

N007Gini coefficientmetricUpdated just now→