An Extensive Study of Two-Node McCulloch-Pitts Networks

Networks with two nodes are previously grouped into either two classes (mutually interactive, master-slave) or five classes (mutualism, competition, predator-prey, commensalism, amensalism). By allowing self-loops, the number of signed regulatory graphs increases to 39. We provide a complete summary of dynamical behaviors of the 39 two-node McCulloch-Pitts models when the link weights are constrained to three values [$-1$,0,$+1$] and Boolean node variables. Depending on whether the Boolean values are [$-1,1$] (bipolar) or [0,1] (binary), we show that the dynamics could also be different with the same signed regulatory graphs. We demonstrate that slight variations in the McCulloch-Pitts model (called variants) may lead to fundamentally different dynamics. We study the full model space and three kinds of robustness or stability: a) of a rule against parameter change on its overall dynamics, b) for a given state against parameter change on its final state, and c) against an initial state change on its final state. All these stability properties are loosely related to a model’s limiting dynamics, with the fixed-point rules to be more stable in the first two types of robustness, but less stable in the third robustness type. These analyses pave the way towards a better understanding of a minimum complex system.

💡 Research Summary

This paper presents a comprehensive exploration of the dynamics of two‑node McCulloch‑Pitts (MP) networks under highly constrained conditions: each link weight can take only the values –1, 0, +1 and each node variable is Boolean, either bipolar (–1, +1) or binary (0, 1). By allowing self‑loops (auto‑catalysis or self‑regulation) the authors expand the classic classification of two‑node interaction graphs from the traditional two (master‑slave, mutual) or five ecological classes (mutualism, competition, predator‑prey, commensalism, amensalism) to a total of 39 distinct signed regulatory graphs.

The authors first enumerate the full rule space: with four possible directed links (including self‑links) each taking three values, there are 3⁴ = 81 possible MP rules. Some of these are disconnected one‑node models or are equivalent under node swapping, leaving 39 unique signed graphs. They then define the update equations

 xₜ₊₁ = f(a xₜ + b yₜ)  yₜ₊₁ = f(c xₜ + d yₜ)

where (a,b,c,d) are the four link weights and f is a threshold function. For bipolar models f = Sign, for binary models f = Step. Crucially, the behavior at the threshold (when the weighted sum equals zero) is not uniquely defined. To capture this ambiguity the authors introduce six variants (V1–V6):

• V1 (bipolar “as‑is”) – retain the previous state at the threshold.
• V2 (bipolar positive) – output +1 at the threshold.
• V3 (bipolar negative) – output –1 at the threshold (equivalent to V2 after sign inversion).
• V4 (binary “as‑is”) – retain previous binary state at the threshold.
• V5 (binary positive) – output 1 at the threshold.
• V6 (binary negative) – output 0 at the threshold.

These seemingly minor choices generate dramatically different dynamics even for the same underlying signed graph.

To analyze the dynamics the authors employ a spectral method that enumerates all four possible initial states (for each Boolean encoding) and determines the attractor reached (fixed point, 2‑cycle, or the maximal 4‑cycle). They find that the limiting dynamics fall into three categories: (i) a single fixed point, (ii) a 2‑state cycle, and (iii) a 4‑state cycle (the longest possible for a two‑node Boolean system).

The paper’s central contribution is a systematic study of three robustness (stability) notions:

1. Rule‑level robustness – how often a rule’s overall attractor type (fixed point vs cycle) persists when any single link weight is perturbed.
2. State‑level robustness to parameter change – for a given initial state, whether the final attractor remains the same when a link weight is altered.
3. State‑level robustness to initial‑state change – whether a small change in the initial condition leads to a different final attractor under a fixed rule.

Results reveal a clear trade‑off. Fixed‑point rules are most robust under (1) and (2): they tend to survive weight perturbations and keep the same final state for a given start. However, they are the least robust under (3): a tiny change in the initial condition often flips the system to a different fixed point or a cycle. Conversely, rules that generate 4‑cycles are comparatively fragile to weight changes but are more tolerant of variations in the initial state.

The discussion connects these findings to classic gene‑regulatory and neural network models (Hopfield, Kauffman, Wagner). The authors note that the inclusion of self‑loops creates feedback loops whose sign (overall positive or negative) influences the propensity for cycles versus multistability, echoing Thomas’ conjectures on regulatory circuits. They also propose extensions: replacing the discontinuous threshold with a sigmoid activation to eliminate the ambiguity at zero, exploring asynchronous updating, and mixing different variants for the two nodes.

In summary, the study demonstrates that even the simplest possible regulatory network—just two nodes—exhibits a rich repertoire of dynamical behaviors that depend sensitively on (i) the signed interaction graph, (ii) the Boolean encoding, and (iii) the precise rule used at the threshold. By exhaustively cataloguing all 39 signed graphs across six variants and quantifying three distinct robustness measures, the authors provide a detailed map of the “edge of chaos” for minimal complex systems, laying groundwork for understanding larger regulatory or neural networks built from such elementary motifs.