Contribution
It names a class, adaptive or coevolutionary networks, and argues that studies from Boolean networks, coupled maps, games and epidemics share four “hallmarks” that need a feedback loop between state and topology. It offers a first inventory of rules and their outcomes (Box 2) and a conjectured common mechanism: dynamics on the network make global topological information locally available, and local rewiring acts on it.
Key insight
In a static network the topology shapes the dynamics; in an adaptive one the dynamics can also read out a global property of the topology, such as whether connectivity is above a percolation point, at each node. Local rewiring that reacts to that reading drives the topology to the point where the dynamics is critical, robustly and without tuning.
Assumptions
None formal: it is a review. Its framing assumes a separation between state (“dynamics”) and topology (“evolution”), defined in Box 1, and distinguishes systems with separated timescales from those where the two change at the same rate (§VI).
Key results
As reported from the papers reviewed, not re-derived:
- Criticality (§III). Christensen et al.’s Bak–Sneppen variant drives the mean degree of the largest cluster to 2. Bornholdt and Rohlf’s threshold network, where frozen nodes gain links and active nodes lose them, reaches K = 2 + 12.4 N^(−0.47), so K_c = 2 for large N.
- Division of labour (§IV). Ito and Kaneko’s coupled logistic maps, with weights strengthened between similar states, split into high- and low-out-degree classes that persist while individual links turn over; Zhou and Kurths’s opposite rule (strengthen links between different states) homogenizes the network and aids synchronization.
- Games (§V). Adaptive prisoner’s dilemmas raise cooperation (“topological punishment”); Pacheco et al. show that with fast rewiring the game maps onto a coordination game. Holme and Ghoshal’s agents seeking centrality at low degree form a transient “VIP” class.
- Epidemics and opinions (§VI). In Gross et al.’s adaptive SIS model, susceptibles rewiring away from infected neighbours give discontinuous transitions, bistability and oscillation; a three-variable moment closure captures what one variable captures in the static model. Holme and Newman’s model (LIT-1128) has a continuous transition.
- Inventory (Box 2). “Activity disconnects” leads to self-organized percolation; “like-and-like” to heterogeneous topologies and distinct node classes; “differences attract” to homogeneous ones.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Adaptive networks recurrently self-organize to critical states by local rules | moderate | several models (§III), each simulated |
| C2 | Like-with-like reinforcement yields heterogeneous topologies and classes of nodes | weak | a handful of models; “quite universally” is the authors’ impression |
| C3 | Dynamics on a network can make global topological information locally accessible | moderate | Bornholdt–Rohlf argument and a toy model (end of §III) |
| C4 | The four hallmarks have one underlying mechanism | weak | conjecture (§VII) |
| C5 | In a timescale-separated system the fast dynamics must have a discontinuous transition for the adaptive interplay to persist | weak | asserted as a “mathematical insight”, not shown |
Concepts
- Adaptive (coevolutionary) network: one whose topology evolves as a function of node states, which evolve on that topology.
- Dynamics / evolution: in this paper, change of state and change of topology respectively (Box 1).
- Division of labour: emergence of topologically and functionally distinct node classes from a homogeneous start; “de-mixing” when nodes keep their class.
Connections
It surveys, among others, Christensen et al. 1998, Bornholdt and Rohlf 2000, Ito and Kaneko 2002–03, Zhou and Kurths 2006, Pacheco et al. 2006, Holme and Ghoshal 2006, Gross et al. 2006 and Holme and Newman 2006 (LIT-1128). None of these except Holme and Newman is registered here.
Bearing on the record
- QUESTION-117. A15 named it for “the general tools for the endorsement canvas read as a directed weighted adaptive network”. The review supplies vocabulary and a list of mechanisms, not tools: no model in it has authority or role, and the only weighted, directed cases (Ito and Kaneko, Zhou and Kurths) are coupled oscillators. Two items bear directly. Like-and-like reinforcement producing persistent leader and follower classes is a predecessor the question should beat if endorsement concentration is offered as new. And the moment-closure approach, a few variables for state plus topology, is the kind of economical reduction the question’s “as economically” invites. Proposed dated note below.
- LIT-1128. The review’s summary of Holme and Newman is wrong in a way that matters for anyone using it as a pointer (Limitations).
No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Reported, not checked. Results are taken from the cited papers; the review adds no analysis.
- Errors. Its account of Holme and Newman (§VI) has the parameter backwards: it says disagreeing neighbours convince each other with probability φ and rewire with 1 − φ, so that at φ = 0 “opinions never change” and at φ = 1 “no connections are rewired”. In the paper φ is the probability of rewiring (LIT-1128, Fig. 1 and §II). It also says the final distribution of followers “approaches a power-law”, which the paper finds only at the critical point. Its account of Bornholdt and Rohlf says that below K = 2 “all nodes change their state” and above it “almost no node changes”, the reverse of the next sentence’s rule (links are added when K < 2, which requires frozen nodes there). The bibliography dates Holme and Newman 2007 (it is 2006) and gives Holme and Ghoshal’s article number as 908701 (098701); “Christiansen” and “Christensen” alternate.
- Exponent sign. Zhou and Kurths’s V(k) ∼ k^(−θ) with θ = −0.48 is a weight growing with degree; the double negative is as printed.
Open questions
- Is C5 true, and what would a proof need?
- Which adaptive-network rules produce persistent role classes in a directed, weighted network, as an endorsement network is?
Note of 2026-10-29: Ito and Kaneko, registered and read
Of the works Connections names as not registered, Ito and Kaneko is now registered and read: their 2002 letter, LIT-1842, NOTE-1414. It answers this note’s second open question in one case: Hebbian, similarity-based reinforcement with normalization of incoming weights produces persistent role classes in a directed, weighted network. The review’s summary of it (§IV) is accurate on the classes and on their persistence while links turn over. It leaves out that the classes are defined by outgoing weight, that incoming weight is fixed at 1 by construction, and that an ablation shows the coupling of state to allocation is needed. The opposite rule, Zhou and Kurths’s, is still held only through the review.