Turing patterns on adaptive networks

Technical Summary: Turing Patterns on Adaptive Networks

Problem Statement
Turing instability, a mechanism explaining the spontaneous formation of spatial patterns from homogeneous states via reaction-diffusion dynamics, has been extensively studied on static networks. In these traditional frameworks, the network topology (nodes and link weights) is fixed, or its evolution is exogenous (prescribed a priori). However, many real-world systems—ranging from neural synapses and social networks to chemical systems—exhibit adaptive networks, where link weights evolve endogenously as a function of the node states. This creates a feedback loop: node dynamics influence connectivity, and evolving connectivity influences node dynamics. Despite the prevalence of such systems, a general theoretical framework for the emergence of Turing patterns on adaptive networks has been lacking. This work addresses the gap by providing a general theory for Turing instability on adaptive symmetric networks with positive weights.

Methodology
The authors propose a mathematical framework combining nonlinear reaction-diffusion systems with adaptive link dynamics:

1. Reaction-Diffusion System: Two species, $u$ and $v$, react on network nodes according to nonlinear functions $f$ and $g$, and diffuse across links. The diffusion is governed by a time-dependent weighted Laplacian matrix $L(t)$, where the weights $w_{ij}(t)$ evolve over time.
2. Adaptive Mechanism: The evolution of link weights is driven by a local pattern amplitude, $A_{loc}$, which measures the deviation of connected nodes from the homogeneous equilibrium. The weight dynamics follow an ODE:
   $$\frac{dw_{ij}}{dt} = a_{ij} (A^* - A_{loc}) k(w_{ij})$$
   where $A^*$ is a control parameter and $k(w)$ is a nonlinear response function. The authors analyze two specific response functions inspired by biological learning rules:
   • Parabola-like: $k(w) = w(w^*_1 - w)$, resembling Hebbian learning.
   • Cubic-like: $k(w) = w(w^*_1 - w)(w^*_2 - w)$, resembling Spike-Timing Dependent Plasticity (STDP).
3. Stability Analysis: The authors perform a linear stability analysis around the homogeneous equilibrium $(u^*, v^*, w^*_1)$. By linearizing the coupled system of reaction-diffusion and weight evolution equations, they derive a block matrix $M$. The stability is determined by the spectrum of this matrix, which decouples into the spectrum of the reaction-diffusion block $S$ and the weight evolution block.
4. Dispersion Relation: The condition for Turing instability is derived by analyzing the eigenvalues of the Laplacian matrix of the underlying static network. The authors establish that instability occurs when the dispersion relation $\rho(\Lambda(\alpha))$ becomes positive for certain eigenvalues $\Lambda(\alpha)$, dependent on the diffusion coefficients and the Jacobian of the reaction terms.

Key Contributions

• General Theory: The paper establishes the validity of the Turing mechanism on adaptive networks, proving that instability conditions rely on the spectral properties of the Laplacian matrix and model parameters, thereby strengthening the interplay between dynamics and topology.
• Endogenous Adaptation: Unlike previous studies on time-varying networks where link dynamics were exogenous, this work treats the network evolution as an endogenous process driven by node states.
• Novel Pattern Types: The analysis reveals that adaptation can lead to:
  • Stationary and Wave-like Patterns: Similar to static networks but modulated by weight evolution.
  • "Bursty" Patterns: A regime where the system oscillates between a homogeneous state (no pattern) and a patterned state. This occurs when pattern formation increases local amplitude, causing weights to shift and suppress diffusion, eventually collapsing the pattern, which then allows weights to recover and restart the cycle.
  • Fragmentation-Induced Patterns: In specific regimes (particularly with cubic-like responses), adaptation can drive link weights to zero, effectively breaking the network into small, disconnected subgraphs. Surprisingly, these small modules can sustain Turing patterns that would be impossible on the original small groups of nodes if they were isolated from the start.
• Induction of Instability: The study demonstrates that adaptation can "force" a system into a Turing-unstable regime even if the initial static network configuration (with initial weights) would not support patterns.

Results
Numerical simulations using the Brusselator and FitzHugh-Nagumo models on networks of $N=20$ nodes corroborate the theoretical findings:

• Parabola-like Response: Depending on the threshold $A^*$, the system exhibits stationary patterns (if $A^*$ is large), transient patterns that decay (if $A^*=0$), or irregular "bursty" oscillations (for intermediate $A^*$).
• Cubic-like Response: This response allows for more complex behaviors. For certain $A^*$ values, the network fragments into small clusters (e.g., pairs or triplets of nodes). Despite the small size of these clusters, Turing patterns emerge and persist because the slow adaptation "traps" the species in a differentiated state before the links vanish.
• Overcoming Static Limitations: Simulations show that a network initialized with weights that yield a negative dispersion relation (stable homogeneous state) can evolve via adaptation to a state with a positive dispersion relation, thereby inducing pattern formation that was impossible in the static counterpart.

Significance
The authors claim that this work provides a necessary theoretical foundation for understanding self-organization in real-world systems where network structure and node dynamics co-evolve. By demonstrating that adaptive mechanisms can not only sustain but also generate Turing patterns in regimes where static networks fail, the paper expands the applicability of Turing theory. The findings suggest that the feedback between state and connectivity is a critical factor in the emergence of complex spatio-temporal motifs in biological, social, and chemical systems. The work opens avenues for studying higher-order structures, directed networks, and the role of noise in adaptive Turing systems.