Contribution
Friedkin (2011) proposed that a group member’s self-weight on the next issue follows their control over the last, and studied it by simulation. This paper takes the simplest version, DeGroot averaging on each issue with fixed relative interpersonal weights, writes it as an explicit map on the simplex, and proves where it goes from every starting point. The answer depends on one vector: the eigenvector centralities of the matrix of relative weights.
Key insight
“the evolution and limiting values of the self-weights depend only upon the network structure” (p. 374). Who ends up with how much power is fixed by who listens to whom, through the eigenvector centrality c, and not by anyone’s starting confidence. The reflected-appraisal loop amplifies centrality: those above a threshold end with more power than their centrality, those below with less.
Assumptions
- DeGroot on each issue (Eq. 2.1, p. 372). y(s, t+1) = W(s)y(s, t), which reaches consensus w(x)ᵀy(s) on every issue. A member’s social power is their weight w_i in that consensus.
- Decomposition (Eq. 2.3, p. 373). W(x) = diag(x) + (I − diag(x))C, where x is the vector of self-weights and C, the relative interaction matrix, is row-stochastic with zero diagonal and constant across issues.
- Reflected appraisal (Eq. 2.6, p. 374). x(s+1) = w(x(s)): the self-weight on the next issue is the control exerted over the last.
- Separated time scales. Consensus on an issue is reached before self-weights update (p. 374).
- Perceived power is accurate. Members know their control; for large groups the authors propose a distributed process by which each could estimate it (Eq. 2.7, p. 375).
- C irreducible (strongly connected) for the main theorem; Section 5 treats reducible C with at least two globally reachable nodes and an aperiodic subgraph among them.
Key results
- Lemma 2.2 (p. 376). The model is x(s+1) = F(x(s)), with F_i(x) = α(x) c_i / (1 − x_i) off the vertices, α normalising. The only parameter is c, the dominant left eigenvector of C.
- Lemma 2.3 (p. 377). No centrality exceeds 1/2; one reaches 1/2 if and only if the network is a star with that node at its centre.
- Lemma 3.1 (p. 378), democracy. If C is doubly stochastic (all c_i = 1/n), every non-autocratic start converges to equal self-weights 1/n, and each consensus is the plain average. “If such structures are unusual, then so is a democratic equilibrium” (p. 376).
- Lemma 3.2 (pp. 378–379), autocracy. If C is a star, every non-autocratic start converges to the centre’s full power; the consensus becomes the centre’s initial opinion.
- Theorem 4.1 (p. 380). Otherwise, the equilibria are the n autocratic vertices and one interior point x*, whose entries are ordered as the centralities c; every non-autocratic start converges to x*.
- Proposition 4.2 (pp. 380–381), accumulation. There is a threshold c_thrshld: members above it end with power above their centrality, those below with power below it.
- Reducible networks (Section 5, pp. 383–385). Members not globally reachable, whom no path of influence from others reaches, have centrality 0, and “the individuals who are not globally reachable in G(C) have zero social power in the equilibrium influence networks” (p. 384).
- Examples (pp. 381–385). Krackhardt’s advice network of 21 managers: the four managers nobody asks for advice end with zero power; the top centrality, 0.205, becomes power 0.236. Simulations on networks of 101, 141 and 4,031 nodes converge in four or five issues.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Under DeGroot–Friedkin dynamics with fixed C, self-weights converge to a unique equilibrium ordered by eigenvector centrality | strong: theorem | Theorem 4.1, Appendix F |
| C2 | Power accumulates above a centrality threshold | strong: theorem | Proposition 4.2 |
| C3 | Autocracy arises exactly for star networks, democracy exactly for doubly stochastic ones | strong: lemmas | Lemmas 2.3, 3.1, 3.2 |
| C4 | Members whom no path reaches end with zero power | strong for the stated reducible case | Section 5 |
| C5 | The mechanism describes real groups | weak: one small experiment cited (Friedkin 2011) | p. 386 |
| C6 | It explains Michels’ iron law of oligarchy | weak: suggestive | p. 371 |
Concepts
- Social power: “relative control over issue outcomes” (p. 370), the weight of a member’s initial opinion in the consensus, after Cartwright.
- Self-weight, self-appraisal: x_i = w_ii, closure to interpersonal influence.
- Relative interaction matrix C: who accords how much of their openness to whom, held fixed.
- Eigenvector centrality c: the dominant left eigenvector of C.
- Autocratic and democratic configurations: a simplex vertex (one member with all power) and equal power 1/n.
Connections
It formalises Friedkin’s reflected-appraisal mechanism (2011, LIT-1404, unread) and the Cooley hypothesis (NOTE-1080). Its empirical companion is Friedkin, Jia & Bullo 2016 (LIT-1482, NOTE-1126), which uses Friedkin–Johnsen rather than DeGroot dynamics and a matrix that changes from issue to issue. It belongs with the record’s DeGroot readings, Golub & Jackson (NOTE-840; THEORY-321), where a largest influence weight that does not vanish keeps the consensus from the truth.
Bearing on the record
- CLAIM-448 (coherence is not integration). The claim names
this model as the baseline: “Self-reinforcing authority is not itself
new”. The reading confirms that and gives the baseline two sharper
edges.
- Consensus is built in. Under DeGroot dynamics every issue ends in consensus whatever the distribution of power, so a coherent record is guaranteed and says nothing about whose evaluations it carries. The content of the consensus is the power-weighted average, and under the reflection loop that weighting moves toward the most central.
- Those nobody listens to vanish. A member whom no path of influence reaches has zero power at equilibrium (Section 5). Their opinion is voiced and discarded, and the record of outcomes is unchanged by it. The paper’s own reading: “the mechanism dampens protest movements assembled by individuals with low levels of relative control” (p. 386). This is the claim’s thesis in a proved special case, with one difference the claim should keep: here the exclusion is of uptake by the network, not of admission by a regime, and nothing chooses it.
- Where the claim adds something. In the model C is fixed and exogenous; nothing maintains or revises who is heard. The claim’s maintained legibility regime is what the model leaves out, as the claim says.
- Filed as THEORY-916.
- CLAIM-440. A prevailing ordering read off this kind of record would be the most central member’s, not the organization’s.
- QUESTION-117 and CASE-092. The LIT was filed as the baseline these ask to be compared with. The comparison now has the model’s exact form: fixed C, DeGroot consensus, x(s+1) = w(x(s)).
- CLAIM-277 and CLAIM-364. Authority made durable by its own exercise, in a social setting, and proved to concentrate.
- No instruction for machine-learning practice.
Limitations
- A model, with little test. “an assessment of the validity of the model has just begun”; “The strength of tendency for larger groups and longer issue sequences is presently unknown” (p. 386).
- Robustness unshown. For anchored (Friedkin–Johnsen) dynamics the simulations show “a more pronounced tendency to autocracy”; “At the present time, we cannot assert robustness” (p. 386).
- C is fixed. Real groups change whom they listen to; the companion paper relaxes this in simulation only.
- Accurate perception of one’s power is assumed: members “know their place” (p. 385).
- Proofs skimmed, not checked in this reading.
Open questions
- Does the interior equilibrium survive anchoring, or do anchored groups always tend to autocracy, as the companion simulations suggest?
- What happens when C itself evolves, for example when members shift weight toward the powerful?