Timescale Limits of Linear-Threshold Networks

1. Problem Statement and Motivation

The paper addresses the global asymptotic stability (GAS) of Linear-Threshold Networks (LTNs), a tractable model for recurrent neural populations.

• Context: LTNs are defined by the dynamics $\dot{x} = -Dx + [Wx + u]_0^1$, where $x$ represents firing rates, $D$ is a diagonal dissipation matrix (allowing for heterogeneous leak rates), $W$ is an asymmetric synaptic matrix, and $[\cdot]_0^1$ is the linear-threshold saturation function.
• The Challenge: While stability is well-understood for symmetric networks (Hopfield networks) via Lyapunov energy functions, LTNs with asymmetric interactions and heterogeneous dissipation ($D \neq I$) are difficult to analyze.
• The Conjecture: The authors focus on Conjecture 1: If the matrix $A = W - D$ is Lyapunov Diagonally Stable (LDS) (i.e., there exists a positive diagonal matrix $\Lambda$ such that $A^\top \Lambda + \Lambda A \prec 0$), then the LTN admits a unique globally asymptotically stable equilibrium.
• Gap: Existing analytical proofs for LDS rely on scalar dissipation ($D=I$) and fail for the biologically relevant case of arbitrary diagonal $D$.

2. Methodology: The $\tau$-LTN Family

To bridge the gap between analytical tractability and biological realism, the authors introduce a one-parameter family of LTNs ($\tau$-LTN) that preserves the LDS condition and the equilibrium set across all time scales.

• Definition: The family is defined by the dynamics:
  $$\dot{x} = \frac{1}{\tau} \left( -Dx + [Dx + \tau(Ax + u)]_0^1 \right)$$
  where $\tau > 0$ is a timescale parameter and $A = W - D$.
• Key Property: All members of this family share the same unique equilibrium point (under the LDS assumption) and the same state space $X = \{x \mid Dx \in [0, 1]^n\}$.
• Limiting Regimes: The authors analyze the behavior of this family at two extreme limits:
  1. Fast Limit ($\tau \to 0^+$): Converges to a Projected Dynamical System (PDS). The dynamics are dominated by projection onto the invariant set $X$.
  2. Slow Limit ($\tau \to +\infty$): Converges to a Hard-Selector System (HSS), a discontinuous differential inclusion where the saturation function acts as a "hard switch" based on the sign of $Ax+u$.

3. Key Contributions

The paper makes three primary theoretical contributions:

1. Construction of the $\tau$-LTN Family: A parametrized deformation of the standard LTN that preserves the LDS structural property for all $\tau$. This allows the isolation of stability mechanisms (dissipation vs. switching) without altering the equilibrium.
2. Equilibrium Analysis: Proof that the equilibrium set is independent of $\tau$. Under the LDS condition, the family possesses a unique equilibrium for all $\tau$, including the limiting cases.
3. Stability Proofs for Endpoints:
   • Fast Limit (PDS): Proved to be Globally Exponentially Stable (GES). The authors construct a quadratic Lyapunov function $V_\Lambda(x) = \frac{1}{2}\|x - x^*\|_\Lambda^2$ and show that the projection operator enforces contraction in the $\Lambda$-weighted metric.
   • Slow Limit (HSS): Proved to be Strongly Globally Asymptotically Stable (GAS). The authors introduce a non-smooth Lyapunov function $V_\infty(x)$ based on the maximum of linear forms. They prove that this function decreases along Filippov solutions of the discontinuous system.

4. Key Results

• Theorem V.4 (Fast Limit): If $A \in \text{LDS}$, the projected dynamical system (PDS) has a unique equilibrium that is globally exponentially stable. The proof relies on the fact that the projection operator $\Pi_X$ ensures the vector field points inward or slides along the boundary in a way that respects the diagonal stability metric.
• Theorem VI.6 (Slow Limit): If $A \in \text{LDS}$, the hard-selector system (HSS) has a unique equilibrium that is strongly globally asymptotically stable. The proof utilizes a piecewise linear Lyapunov function that captures the "switching" logic of the saturation function.
• Numerical Evidence:
  • Simulations confirm that trajectories near the limits behave as predicted by the respective Lyapunov certificates.
  • Crucially, in non-LDS scenarios (where limit cycles exist), the $\tau$-LTN family preserves the oscillatory behavior across all $\tau$ and at both endpoints, suggesting the family faithfully captures the underlying dynamical skeleton.

5. Significance and Implications

• Decomposition of Stability: The work suggests that global stability in LTNs under LDS arises from two distinct but compatible mechanisms:
  1. Dissipative Dominance: Evident in the fast limit (PDS), where the linear drift is contracted by the diagonal dissipation.
  2. Coherent Switching: Evident in the slow limit (HSS), where the nonlinear saturation boundaries are oriented such that they guide trajectories toward the equilibrium.
• Pathway to Conjecture 1: While the paper does not prove GAS for the intermediate $\tau$ values (specifically $\tau=1$, the standard LTN), it establishes that both endpoints are stable. This strongly implies that the intermediate systems are stable, as they interpolate between two stable regimes governed by the same structural condition (LDS).
• Biological Plausibility: By allowing $D \neq I$, the model accommodates heterogeneous neuronal leak rates, making the stability results more applicable to real neural circuits than previous symmetric or scalar-dissipation models.
• Future Directions: The authors posit that a general stability proof for the full family could be achieved by interpolating the Lyapunov functions found for the endpoints or using continuation arguments.

In summary, the paper provides a rigorous structural framework to analyze LTN stability by decomposing the problem into analytically solvable fast and slow limits, offering strong evidence that Lyapunov Diagonal Stability is indeed sufficient for global stability in asymmetric, heterogeneous neural networks.