A Continuous-Time and State-Space Relaxation of the Linear Threshold Model with Nonlinear Opinion Dynamics

1. Problem Statement

The Linear Threshold Model (LTM) is a foundational framework for studying the propagation of collective behaviors (complex contagions) in networks. However, standard LTM formulations suffer from two critical limitations when applied to real-world biological and engineered systems:

• Discrete State and Time: The model relies on binary states (active/inactive) and discrete timesteps. This abstraction fails to capture the continuous nature of individual decision-making processes and prevents the analysis of how internal cognitive dynamics interact with external stimuli over time.
• Inability to Model Subthreshold Signaling: In a fully inactive system, LTM nodes cannot transition to an active state unless the weighted sum of neighbors strictly exceeds a threshold. This excludes the possibility of "spontaneous false alarms" or cascades driven by the accumulation of weak, subthreshold signals, which are frequently observed in nature.
• Lack of Temporal Dynamics: Standard LTM cannot explore how the timing or distribution of inputs affects cascade emergence, as it treats time as a sequence of steps rather than a continuous flow.

2. Methodology

The authors propose a Continuous-Time and State-Space Relaxation of the LTM based on the Nonlinear Opinion Dynamics (NOD) framework.

• Model Formulation:
  The NOD model describes the evolution of continuous opinion states $z \in [0, 1]^n$ using the following differential equation:
  $$\dot{z} = -z + \text{sat}(\mu \odot z + k(z \odot z) + \gamma Az + b)$$
  Where:

  • $z$: Continuous opinion vector.
  • $A$: Weighted adjacency matrix (social interactions).
  • $\mu, k, \gamma$: Parameters for linear self-reinforcement, nonlinear self-reinforcement, and social coupling, respectively.
  • $b$: External input/bias vector.
  • $\text{sat}(\cdot)$: A piece-wise linear saturation function mapping to $[0, 1]$.

• Mapping LTM to NOD:
  The core methodology involves replacing the discontinuous step-function thresholds of the LTM with the smooth saddle-node bifurcations inherent in the NOD model.

  • The LTM activation condition $(A\zeta)_i > \tau_i$ is mapped to the NOD bifurcation threshold $b^*_i = \frac{(1-\mu_i)^2}{4k}$.
  • By setting the NOD parameters such that $\tau_i = \frac{b^*_i}{\gamma}$, the authors create a direct correspondence between the discrete threshold $\tau_i$ and the continuous bifurcation point.

• Analytical Approach:
  The authors utilize dynamical systems theory, specifically analyzing:

  • Invariant Sets: Proving that the unit hypercube is invariant and that the "fully active" state ($z=1$) is an asymptotically stable equilibrium.
  • Bifurcation Analysis: Characterizing the tipping point where an agent transitions from a neutral state to an active state.
  • Inductive Proofs: Using mathematical induction to relate the discrete cascade steps of LTM to the continuous flow of NOD.

3. Key Contributions

The paper makes four primary theoretical contributions:

1. Formal Relaxation Procedure: A rigorous method to construct a NOD instance from any given LTM instance, effectively mapping discrete cascade processes to continuous dynamical flows.
2. Guaranteed Activation: Proof that for any seed set, if a node activates in the discrete LTM, it is guaranteed to activate in the continuous NOD relaxation (under appropriate parameter choices).
3. Asymptotic Equivalence: Proof that by sufficiently bounding the social coupling parameter ($\gamma$), the continuous NOD dynamics exactly recover the discrete cascade set of the LTM. This establishes a computable condition where the continuous model behaves identically to the discrete model.
4. Expanded Analytical Capabilities: Demonstration that the NOD framework allows for the exploration of phenomena impossible in LTM, specifically:
   • Cascades driven by strictly subthreshold inputs (weak signals accumulating over time).
   • The critical role of temporal separation in distributed inputs.

4. Key Results

• Theorem 1 (Inclusion): For any seed set $I$, the set of nodes activated in LTM ($C_{LTM}$) is a subset of those activated in NOD ($C_{NOD}$). This confirms NOD is a valid relaxation.
• Theorem 2 (Exact Recovery): There exists a threshold $\gamma^*$ such that for all $\gamma < \gamma^*$, $C_{NOD}(I) = C_{LTM}(I)$. This means the continuous model can be tuned to perfectly mimic the discrete model's outcomes.
• Subthreshold Cascades: Simulations (Fig. 3) demonstrate that while LTM cannot trigger a cascade with inputs below the threshold, NOD can. If weak inputs are applied simultaneously or with specific temporal delays, they can feed back through the network, pushing the system past the tipping point.
• Temporal Sensitivity: The paper shows that the timing of inputs matters in NOD. Randomly delaying subthreshold inputs can prevent a cascade even if the total energy of the input is the same, a nuance lost in discrete LTM.
• Parameter Sensitivity: The parameter $k$ controls the "switch-like" nature of the dynamics (higher $k$ = sharper transition), while $\gamma$ controls the strength of social coupling and the magnitude of subthreshold effects.

5. Significance

This work provides a mathematically rigorous bridge between discrete contagion models and continuous networked dynamics.

• Theoretical Advancement: It resolves the limitations of discrete models by embedding them in a differentiable state-space, allowing for the application of powerful tools from nonlinear dynamical systems (e.g., bifurcation analysis, stability theory).
• Practical Application: The framework is highly relevant for modeling biological systems (e.g., animal group escape responses) and engineered systems (e.g., neural networks, power grids) where decision-making is continuous and subthreshold signaling is common.
• New Insights: It reveals that "weak" signals, when distributed and temporally structured, can trigger global cascades—a phenomenon that discrete models fundamentally miss. This suggests that collective behavior is more sensitive to the temporal structure of information than previously thought.

In conclusion, the authors successfully demonstrate that the NOD framework is not only a faithful relaxation of the LTM (recovering its results under specific bounds) but also a superior tool for exploring the rich, continuous, and time-dependent nature of real-world collective decision-making.