Contribution
It shows by simulation that a network of identical chaotic units with Hebbian, resource-limited weights organizes itself into two persistent classes while the weights themselves keep changing. The classes differ in whom they influence: a small core sends strong connections to everyone, and a large periphery sends almost none back. No synchronization of the units is needed.
Key insight
Normalizing each unit’s incoming weights makes the units compete to be attended to, and strengthening weights between similar states makes attention follow similarity. Together, in a chaotic regime with fast weight turnover, they concentrate attention on a few units, and the concentration is absorbing: a unit that drops out of the core never rejoins, and one that joins never leaves.
Assumptions
- Units (eq. 1). x_{n+1}^i = f[(1 − c)x_n^i + c Σ_j w_n^{ij} x_n^j], f(x) = ax(1 − x). w^{ij} is the weight from j to i.
- Weights (eq. 2). w_{n+1}^{ij} = [1 + δ g(x^i, x^j)]w_n^{ij} / Σ_j [1 + δ g(x^i, x^j)]w_n^{ij}, with g = 1 − 2|x^i − x^j|. Incoming weights sum to 1 for every unit, so the network is directed and asymmetric. g is negative when states differ by more than 0.5.
- Initial state. Self-weights 0, all other weights 1/(N − 1), states uniform on (0, 1). So the units start identical in role.
- Parameters. δ = 0.1 (similar behaviour “over a wide range”, but no structure for δ < 0.01; unit and weight timescales must be of the same order); N = 100; the desynchronized phase, about 3.7 < a < 4.0 and 0 < c < 0.2; detailed results at a = 3.97, c = 0.12.
Key results
- Phases. Of the four GCM phases, three appear: coherent, ordered (clusters) and desynchronized. Only the desynchronized phase has a changing network.
- Activity and variance (Fig. 1). Mean weight change per step is high in a band around c = 0.15(a − 3.7). The variance across units of total outgoing weight W_out^i is large just below that line. Low-activity regions hold fixed correlated pairs.
- Two classes (Fig. 2). Time-averaged W_out^i separates into a high and a low group. No unit migrates after about 3 × 10⁶ steps (note [10]: once a unit’s W_out is low enough it never rejoins the high group, and conversely). The high group fluctuates more, so the low group is “more stable”.
- Degrees (Fig. 3). With a connection counted when w^{ij} > 1/(N − 1), in-degree is unimodal, peak about 8. Out-degree is bimodal: an exponential part and a peak near 45.
- Core and periphery (Table I). Core: out-degree above 20, “typically 14” units. Connections: core→core 159 (density 0.6625), core→periphery 792 (0.5893), periphery→core 2 (0.0015), periphery→periphery 234 (0.0336). “The dynamics of the peripheral group have almost no influence on the core group.”
- Ablation. When weights follow the states but the states do not depend on the weights, “no structure ever appears”.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Identical units split into a persistent core and periphery by outgoing weight | moderate | simulation, N = 100, one parameter point in detail, “many initial conditions” for degrees |
| C2 | The split occurs without synchronization of the units | moderate | phase diagram; desynchronized phase only |
| C3 | Membership is absorbing after a transient | moderate | Fig. 2 and note [10], one trial shown |
| C4 | The structure requires the two-way coupling of states and weights | weak | one ablation, details “presented elsewhere” |
| C5 | The phenomenon is general for networks whose links follow relations among dynamic units | weak | argued from universality of GCMs |
Concepts
- Activity of the network: A = mean over pairs and steps of |w_n^{ij} − w_{n−1}^{ij}|, eq. (3).
- W_out^i: Σ_j W^{ji}, the time-averaged total weight unit i sends.
- Core / peripheral group: units with out-degree above / at most 20 in the thresholded graph.
Connections
It belongs to Kaneko’s globally coupled maps (Physica D 1990, 1991) and extends Ito and Kaneko’s earlier neural-network model (Neural Networks 2000). Gross and Blasius (LIT-1139, NOTE-1394) review it as their example of spontaneous “division of labour”, and set it against Zhou and Kurths’s opposite rule (strengthen links between different states), which homogenizes the network. Holme and Newman (LIT-1128) is a different mechanism: rewiring toward agreement, with no weights and no influence asymmetry.
Bearing on the record
- QUESTION-117. The triage expected the like-and-like predecessor to beat. The reading confirms it and makes it sharper. The model is closer to the record’s endorsement circuit than Gross and Blasius’s summary suggests. Each unit allocates a fixed budget of attention (incoming weights sum to 1). It moves attention toward those whose state resembles its own, and its state is driven by those it attends to. That is an endorsement network with normalized trust and similarity-based reinforcement. It gives, from identical units and with no role primitive, a persistent asymmetry of influence (a core of about 15 attended by most, attending almost only to itself) with link turnover, and an ablation saying the coupling of state to allocation is needed. So an answer to QUESTION-117 that offers persistent authority classes, or their need for the state–authority coupling, as new is matched by this model. What it cannot express is what the question’s circuit adds: states with content that can be right or wrong, any evaluation of outcomes, and authority indexed by role or issue. The outcome to beat is “concentration without regard to accuracy”. That is also persuasion bias’s (LIT-1134) outcome, and this model gets it with moving weights and hysteresis, which persuasion bias lacks.
- QUESTION-083. It is a population of identical processes, under one update rule with no governance primitive, in which differentiated and durable influence forms. That is the part QUESTION-295 (Answered) already holds. It shows nothing on niches or mutual regulation. The periphery has almost no influence on the core (2 of about 1,344 possible links), so the regulation is one-way, and nothing in the model is a competence that shows in rare contexts.
- QUESTION-132. The split looks like the question’s third outcome, entrenched asymmetry, more than stabilizing coexistence: membership is absorbing and the core does not depend on the periphery. But the paper measures no recovery when rare, no perturbation and no removal of a core unit. So it is a candidate case for the question’s classification, not evidence for one outcome. Which outcome it is could be tested cheaply by removing core units and seeing whether the periphery refills the core.
- NOTE-1394’s open question. One answer: Hebbian, similarity-based reinforcement with normalization of incoming weights produces persistent role classes in a directed, weighted network. The opposite rule (Zhou and Kurths) does not, per the review. Gross and Blasius’s summary is accurate on the classes and their persistence under link turnover. It leaves out that the classes are defined by outgoing weight, that incoming weight is fixed at 1 by construction, and the coupling ablation.
- CLAIM-279 (roles are regimes, not kinds). A small case for it: the same units under the same rule end in two classes, set by the dynamics and not by anything built into the units. Here the regime does not dissolve once formed, which the claim’s word “regime” may not intend.
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- A four-page letter of simulations. There is no analysis of why the core forms, why its size is about 14 to 16, or why membership is absorbing. More detail is promised “elsewhere”.
- One system size (N = 100; larger N “do not change qualitatively … except for the existence of a longer transient”), one δ, one parameter point in detail. Fig. 2 shows a single trial.
- Classes are defined by two thresholds (1/(N − 1) for a link, 20 for the core), chosen by eye from the bimodal histogram.
- Errors and slips. Table I fits a core of 16 and a periphery of 84 (densities 159/240, 792/1,344, 2/1,344, 234/6,972), though the text says the core is “typically 14”, so the table is presumably one realization. Typing slips: “strenghened”, “competetion”, “eminating”, “reffered”, “nother”, “senn”. The preprint’s date line reads “October 28, 2018”, a compile date, not the submission.
Open questions
- If core units are removed, does the periphery refill the core (the core as a regime), or does the network stay without one?
- Does the core’s size scale with N, with δ, or with the chaotic regime’s correlation structure?
- If each unit’s state carried a signal plus noise, would the core be the more accurate units or merely the more typical ones?