Design of Hierarchical Excitable Networks

This paper presents a systematic method to construct vector fields that realize hierarchical excitable networks, where lower-level dynamics form heteroclinic connection patterns and top-level transitions between these patterns are excitable with zero threshold, thereby extending the simplex realization method of Ashwin and Postlethwaite.

The Gist

Imagine you are watching a complex play with two different layers of action happening at the same time.

The Big Picture: A Play Within a Play

Think of a hierarchical excitable network like a theater production with a Stage Manager and several Acting Troupes.

1. The Stage Manager (The Top Level): This is the "Superstructure." The Stage Manager decides which Troupe gets to perform next. They walk around the stage, pointing to different groups and saying, "You, go!" and then "You, go!" in a specific order.
2. The Acting Troupes (The Lower Level): These are the groups of actors. Each Troupe has its own internal script. When the Stage Manager calls them, they perform a specific scene (a pattern of movement or dialogue) over and over again.

The Problem the Paper Solves

In the world of math and physics, scientists often want to build systems (like models of the brain, ecosystems, or social groups) that switch between different behaviors in a specific order.

• Sometimes, a system needs to stay in one "mode" for a while, then switch to another.
• The challenge is: How do you write the "rules" (math equations) for a system so that it naturally follows a complex, multi-level script?

The Solution: The "Simplex-Simplex" Method

The authors, Sören and Alexander, invented a new recipe to build these systems. They call it the Simplex-Simplex Method. Here is how it works using our theater analogy:

1. The "Zero-Threshold" Switch (Excitable Connections)

Usually, in math models, switching from one state to another is like a train on a track. Once it leaves the station, it must go to the next station. It's a hard, permanent connection.

The authors use something called an excitable connection.

• The Analogy: Imagine the Stage Manager doesn't just point; they whisper a secret. If an actor is standing very close to the edge of the stage (within a tiny distance), the whisper pushes them over the edge, and they run to the next troupe.
• The Magic: The "whisper" is so sensitive that even if the actor is just a tiny bit off-center, they get pushed. In math terms, this is a "zero threshold." It means the system is incredibly sensitive to small nudges, allowing it to jump between different "modes" (troupes) easily, but only when the Stage Manager is ready.

2. The "On/Off" Switch (The Bump Function)

How do we make sure Troupe A doesn't perform while Troupe B is on stage?

• The authors use a mathematical "dimmer switch" called a bump function.
• When the Stage Manager is near Troupe A, the dimmer is turned all the way up for Troupe A (they perform their script) and all the way down for everyone else (they freeze or disappear).
• As the Stage Manager walks to Troupe B, the dimmer for A fades out, and the dimmer for B fades in.

3. The Result: A Hierarchical Dance

By stacking these switches, they created a system where:

• Level 1 (The Troupes): Inside each active troupe, the actors cycle through their own complex patterns (like a heartbeat or a neural firing pattern).
• Level 2 (The Stage Manager): The Stage Manager slowly walks from one troupe to the next, triggering the switch.

Why is this useful?

Real life is full of these "plays within plays."

• In the Brain: You might be thinking about a specific memory (the lower level pattern), but then a loud noise happens, and your brain switches to a "fight or flight" mode (the higher level switch).
• In Nature: Animals might have a daily rhythm (eating, sleeping) that gets interrupted by seasonal changes (migration).
• In Society: A group of people might have a routine way of arguing, but a major event (like an election) forces them to switch to a completely different way of interacting.

The "Simulation" Proof

The authors didn't just write the theory; they built a computer simulation.

• They created a "Stage Manager" that cycles through three groups.
• Two groups were simple loops (A -> B -> C -> A).
• One group was a complex, switching pattern (the "Kirk-Silber" network).
• The Result: The computer ran exactly as predicted. The Stage Manager moved from group to group. When a group was active, it performed its specific dance. When the manager moved on, the group froze, and the next one started.

In Summary

This paper provides a blueprint for engineers and scientists to build complex, multi-layered systems. It shows how to take a simple list of "who talks to whom" (a graph) and turn it into a dynamic machine that switches between different behaviors automatically, just like a well-rehearsed play with a clever stage manager.

They proved that you can build these "plays" mathematically, ensuring that the transitions happen smoothly and that the system behaves exactly as the designer intended, even if the underlying rules are very complex.

Technical Summary

Here is a detailed technical summary of the paper "Design of Hierarchical Excitable Networks" by Sören von der Gracht and Alexander Lohse.

1. Problem Statement

The paper addresses the challenge of constructing dynamical systems that exhibit specific, prescribed connection structures defined by directed graphs (digraphs). While previous research has focused on realizing single digraphs as heteroclinic networks (where trajectories connect saddle equilibria) or excitable networks (where trajectories are triggered by noise or perturbations), there was a lack of systematic methods to construct hierarchical structures.

Specifically, the authors aim to model systems where:

• Lower Level: There are distinct, complex connection patterns (represented by digraphs $G_1, \dots, G_N$) that operate as invariant sets (heteroclinic networks).
• Top Level: There is a higher-order structure (represented by a digraph $\Gamma$) that dictates transitions between these lower-level patterns.
• The Gap: Existing methods do not easily allow for a system where the top-level transitions are excitable (zero threshold) while the lower-level dynamics remain heteroclinic. This is crucial for modeling phenomena like cognitive processes or biological rhythms where a "higher-level" process triggers a switch between "lower-level" states without the strict time-reversibility of heteroclinic connections.

2. Methodology: The Simplex-Simplex Method

The authors propose a construction method called the Simplex-Simplex method, which couples two layers of the "Simplex Method" (originally developed by Ashwin & Postlethwaite).

A. Mathematical Framework

The system is defined by a vector field $\dot{x} = f(x)$ in a high-dimensional phase space. The construction involves:

1. Superstructure ($X$): A set of variables $X = (X_1, \dots, X_N)$ representing the top-level digraph $\Gamma$. This subsystem is decoupled and follows the standard simplex realization of $\Gamma$.
2. Substructures ($x^j$): For each node $V_j$ in $\Gamma$, there is a corresponding subspace of variables $x^j = (x^j_1, \dots, x^j_{n_j})$ representing the lower-level digraph $G_j$.
3. Coupling via Bump Functions: The dynamics of the substructures are modulated by the state of the superstructure using smooth transition functions (bump functions) $b^j_\varepsilon(X)$.
   • When $X$ is near the equilibrium $X_j$ (representing node $V_j$), the function $b^j_\varepsilon \approx 1$, activating the dynamics of the substructure $G_j$.
   • When $X$ moves away from $X_j$, the function decays to 0, causing the substructure variables to decay to zero (or a specific state), effectively "turning off" that sub-dynamics.

B. System Equations

The core system (Equation 4) is defined as:
$$\dot{X}_j = X_j \left( 1 - \|X\|^2 + \sum_{k=1}^N a_{jk} X_k^2 \right)$$
$$\dot{x}^j_i = x^j_i \left[ \left( 1 - \|x^j\|^2 + \sum_{k=1}^{n_j} \alpha^j_{ik} (x^j_k)^2 \right) b^j_\varepsilon(X) - (1 - b^j_\varepsilon(X)) \right]$$

• The first equation drives the heteroclinic cycle on the top level.
• The second equation creates a convex combination: it realizes the heteroclinic network $G_j$ when active ($b \approx 1$) and forces decay to zero when inactive ($b \approx 0$).

C. Time-Scale Adaptation

To ensure numerical observability of the hierarchical transitions (preventing the top level from switching too fast for the bottom level to react), the authors introduce scaling parameters $\Phi, \Psi, \Omega$ to adjust the velocity of the superstructure, the active substructure, and the decay of inactive substructures, respectively.

3. Key Contributions

1. Systematic Construction of Hierarchical Networks: The paper provides the first systematic algorithm to realize a collection of digraphs $G_1, \dots, G_N$ and a super-graph $\Gamma$ as a single dynamical system.
2. Hybrid Dynamics (Heteroclinic + Excitable):
   • Lower Level: The connections within each $G_j$ are heteroclinic (trajectories connect saddle equilibria).
   • Top Level: The connections between the invariant sets $N_j$ (realizations of $G_j$) are excitable with zero threshold.
   • Crucial Distinction: The top-level connections are not heteroclinic. The $\alpha$-limit set of a trajectory transitioning from $N_j$ to $N_k$ is empty (the trajectory comes from infinity in backward time), whereas a heteroclinic connection would limit to a specific equilibrium or cycle in backward time.
3. Theorem 3.3 (Realization Result): The authors prove that if the input digraphs contain no 1-cycles or 2-cycles, the constructed system (4) realizes the hierarchical structure for any amplitude $\delta > 0$.
4. Extension of Simplex Method: The work generalizes the Ashwin & Postlethwaite simplex method from realizing graphs as networks of equilibria to realizing graphs as networks of invariant sets (which are themselves networks).

4. Results

The authors validate their theoretical construction through two detailed numerical examples:

• Example 1 (Switching Substructure):

  • Structure: A top-level 3-cycle ($\Gamma$) driving three substructures: two 3-cycles ($G_1, G_2$) and one Kirk-Silber network ($G_3$).
  • Observation: The simulation shows the top-level variable $X$ cycling through states. As $X$ approaches a specific node, the corresponding sub-dynamics "wake up" and follow its specific cycle. When $X$ moves to the next node, the previous sub-dynamics decay to zero, and the new one activates.
  • Kirk-Silber Behavior: The simulation correctly reproduces the known "switching" behavior of the Kirk-Silber network (switching from an upper cycle to a lower cycle) only when that specific substructure is active.

• Example 2 (Switching Superstructure):

  • Structure: The top level is a Kirk-Silber network, driving four substructures (two 3-cycles, two 4-cycles).
  • Observation: The top-level dynamics exhibit the complex switching behavior of the Kirk-Silber network. Consequently, the substructures activate and deactivate in a complex, non-periodic sequence dictated by the top-level switches.

• Numerical Stability: The simulations confirm that the invariant sets are robust and that the zero-threshold excitable connections allow trajectories to transition between these sets in forward time, mimicking heteroclinic behavior without the strict backward-time constraints.

5. Significance and Implications

• Modeling Complex Systems: This method provides a powerful tool for modeling systems with intrinsic hierarchies, such as:
  • Neuroscience: Modeling cognitive processes where memory formation or learning (lower level) is modulated by higher-level attention or creativity mechanisms.
  • Biological Rhythms: Explaining polyrhythms in central pattern generators where different movement patterns are switched by a higher-level control signal.
  • Social/Biological Networks: Modeling networks where interaction rules change over time based on external factors (e.g., physical proximity).
• Temporal Networks: The construction effectively creates a dynamical realization of a temporal network, where the interaction structure itself evolves over time according to a prescribed graph.
• Theoretical Advancement: It bridges the gap between heteroclinic networks and excitable networks, demonstrating that a system can possess both types of dynamics simultaneously at different hierarchical levels.
• Future Directions: The authors suggest this method can be iterated to create multi-level hierarchies (3+ levels) and modified to produce true heteroclinic connections between invariant sets (though this requires relaxing the compactness of the phase space or accepting multiple copies of networks).

In conclusion, the paper successfully demonstrates that complex, multi-scale dynamical behaviors can be engineered systematically by coupling simplex realizations, offering a new paradigm for constructing dynamical systems with prescribed hierarchical topologies.