Timescale Limits of Linear-Threshold Networks
William Retnaraj, Simone Betteti, Alexander Davydov, Francesco Bullo, Jorge Cortés · 2026
Read Timescale Limits of Linear-Threshold Networks by William Retnaraj, Simone Betteti, Alexander Davydov, Francesco Bullo, Jorge Cortés (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16710.txt.
Introduction
Retnaraj and colleagues study global asymptotic stability in linear networks with uneven links and loss. Such networks capture neuron group action and interest control experts for their richness. A one-knob family keeping LDS has a fixed rest set. The fast limit settles to a moving system, and the slow limit to a jumpy hard-pick system. Under diagonal stability, the fast limit is strongly stable and the slow limit stable in the long run. Settling calm at both endpoints offers a grounded path toward global calm for likely networks.
Catalog record
| CREATOR | William Retnaraj, Simone Betteti, Alexander Davydov, Francesco Bullo, Jorge Cortés |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47297 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.16710
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Timescale Limits of Linear-Threshold Networks William Retnaraj, Simone Betteti, Alexander Davydov, Francesco Bullo, Jorge Cortés Abstract Linear-threshold networks (LTNs) capture the mesoscale behavior of interacting populations of neurons and are of particular interest to control theorists due to their dynamical richness and relative ease of analysis. The aim of this paper is to advance the study of global asymptotic stability in LTNs with asymmetric neural interactions and heterogeneous dissipation under the structural Lyapunov diagonal stability (LDS) condition. To this end, we introduce a one-parameter family of LTNs that preserves the LDS condition and has a parameter-independent equilibrium set. In the fast limit, this family converges to a projected dynamical system (PDS), while in the slow limit, it converges to a discontinuous hard-selector system (HSS). Under LDS, we prove that the fast PDS limit is globally exponentially stable and that the HSS limit is globally asymptotically stable. This alignment suggests that the limiting systems capture essential mechanisms governing stability across the entire LTN family. Together with numerical evidence, these findings indicate that resolving stability at the fast and slow endpoints provides a promising and structurally grounded path toward establishing global stability for LTNs with biologically plausible recurrence and diagonal dissipation. I Introduction Autonomous dynamical systems provide a foundational framework for modeling neural computation, capturing recurrent interactions through a combination of a dissipative component that contracts activity and a synaptic component that encodes nonlinear recurrent coupling. Classical results for systems with symmetric synaptic matrices, originating from Grossberg [ 1 ] and Hopfield [ 2 , 3 ] , exploit symmetry to construct Lyapunov energy functions, ensuring dissipation and convergence to an equilibrium. However, symmetry imposes constraints often incompatible with biologically-realistic connectivity. The linear-threshold network (LTN) is a tractable model that allows for asymmetric synaptic capturing, resulting in a wide variety of dynamical behavior, including stable equilibria [ 4 ] , multistability, oscillations [ 5 ] , and chaos [ 6 ] . Stability results for the LTN remain limited and often rely on restrictive assumptions such as total \mathcal{L} -stability [ 4 ] . A more general structural condition is Lyapunov diagonal stability (LDS) [ 7 ] , which has been studied in the context of various dissipative network systems [ 8 ] . For the closely-related Hopfield network, LDS guarantees global stability via a Lyapunov function [ 9 ] . Since the appearance of [ 9 ] in 1995, it has been a long-standing conjecture that LDS is sufficient for the stability of firing-rate networks, of which LTN is a special case. Existing proofs that use LDS, however, rely on scalar dissipation and do not extend to the biologically relevant case of arbitrary diagonal dissipation. Recent numerical evidence [ 10 ] suggests that LDS may suffice to ensure global stability for firing-rate networks. In particular, we focus on two relevant and limiting cases of linear-threshold dynamics: the projected dynamical system (PDS; fast limit) and the hard-selector system (HSS; slow limit). Conceptually, PDS reveals the stabilizing role of componentwise dissipative dominance, while HSS reveals the stabilizing role of nonlinear switching-regime boundary orientations. In this work, we show that both systems inherit strong dissipativity and stability properties under LDS. In particular, we introduce suitable Lyapunov energy functions for both fast and slow limits, and prove global stability of the respective dynamics. The fact that both limiting mechanisms are stable under the same diagonal condition strongly suggests that the intermediate LTNs might also be stable. Thus, the analysis of the limiting systems is not only analytically convenient, but it gives us conceptually fundamental building blocks toward a general stability proof. Contributions: In this paper, we contribute to the analysis of global stability in asymmetric neural networks as follows: (i) we introduce a parametrized family of linear-threshold networks, named the \tau -LTN family, preserving Lyapunov diagonal stability for all \tau ; (ii) we show how the limiting endpoint systems, fast and slow, of the family correspond to projected dynamics and hard-selector dynamics, respectively, and establish uniqueness of equilibria for all systems; (iii) we prove global stability of the unique equilibrium of the fast-limit and slow-limit systems, respectively, by introducing suitable Lyapunov functions. Section II introduces notation and preliminaries. Section III presents the \tau -LTN family and formulates the stability problem. Section IV establishes equilibrium properties and preservation of diagonal stability. Sections V and VI analyze the fast and slow-limit systems. Section VII is a discussion with some numerical exploration, and Section VIII concludes with a summary of how the \tau -LTN family and the endpoints study helps us make progress towards the global stability of LTNs under Lyapunov diagonal stability, and we posit future research directions. II Preliminaries II-A Notation Define the set [n]=\{1,\dots,n\} for n>0 . Denote the set of all reals as \mathbb{R} , the set of all positive reals as \mathbb{R}_{+} , and the n -dimensional Euclidean space by \mathbb{R}^{n} . For this section, let z\in\mathbb{R} be a scalar and let x,y\in\mathbb{R}^{n} be vectors. Denote x=[x_{1},\dots,x_{n}]^{\top} . Let [z]_{0}^{m}=\max{(0,\min{(z,m)})} with m>0 . When applied to x\in\mathbb{R}^{n} , [x]_{0}^{m} is to be understood elementwise. Define the positive (resp. negative) part of z as (z)_{+}=\max(0,\,z) (resp. (z)_{-}=-\min(0,\,z) ). Define \bf 1_{n} to be the n -dimensional all-ones vector. Let \mathcal{X}\subset\mathbb{R}^{n} . Define {\bf 1_{\mathcal{X}}}:\mathbb{R}^{n}\rightarrow\{0,1\} be the 0 – 1 indicator of \mathcal{X} , so that {\bf 1_{\mathcal{X}}}(x)=1 if x\in\mathcal{X} , and 0 if not. Let \mathbb{D}^{n}_{+} be the space of all positive diagonal matrices. For this section, let A\in\mathbb{R}^{n\times n} be a real-valued square matrix. Let A_{i}\in\mathbb{R}^{n} represent the i th row of A so that A=[A_{1},\dots A_{n}]^{\top} . Let A_{ij}\in\mathbb{R} represent the entry in the i th row and j th column of A . Denote \mathrm{diag}(d_{1},\dots,d_{n}) as the diagonal matrix with scalars d_{1},\dots,d_{n} on the diagonal. We denote the identity matrix in \mathbb{R}^{n\times n} by I_{n} , and drop the subscript when the dimension is clear. By A\succ 0 (resp. A\prec 0 ), we mean that A is positive definite (resp. negative definite). Let \lVert x\rVert_{P} be the P -weighted Euclidean 2-norm of x for P\succ 0 . If P=I , the above becomes \lVert x\rVert_{2} or simply \lVert x\rVert , the Euclidean 2-norm. We say that A is Lyapunov diagonally stable ( A\in\mathcal{L}DS ) if A^{\top}\Lambda+\Lambda A\prec 0 holds for some \Lambda\in\mathbb{D}^{n}_{+} , which is called the diagonal certificate. Let \mathfrak{B}(X) be the collection of all subsets of set X . Define the convexified hard-selector map h:\mathbb{R}\rightarrow\mathfrak{B}([0,1]) and its vector equivalent \mathcal{H}:\mathbb{R}^{n}\rightarrow\mathfrak{B}([0,1]^{n}) as h(z)=\begin{cases}\{0\},&z<0,\\ [0,1],&z=0,\\ \{1\},&z>0;\end{cases}\qquad\mathcal{H}(x)=\begin{bmatrix}h(x_{1})\\ \vdots\\ h(x_{n})\end{bmatrix}. (1) II-B Projected dynamical systems Here, we describe basic notions on projected dynamical systems following [ 11 , 12 ] . Let \mathcal{X}\subset\mathbb{R}^{n} be a closed convex polyhedron and let x,v\in\mathbb{R}^{n} be two vectors. If x\in\mathcal{X} , we define the tangent cone of \mathcal{X} at x\ \in\mathcal{X} as T_{\mathcal{X}}(x) . Intutively, the tangent cone is the set of all directions that do not leave \mathcal{X} , if one steps in their direction infinitesimally from x\in\mathcal{X} . Define the projection operator \Pi_{\mathcal{X}}:\mathbb{R}^{n}\times\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{n} by \Pi_{\mathcal{X}}(x,v)=\mathrm{arg}\min_{y\in T_{\mathcal{X}}(x)}\lVert y-v\rVert^{2}. (2) Notice that the projection onto a closed convex cone (arising from the convexity of \mathcal{X} ) is single-valued. For any vector field X:\mathbb{R}^{n}\rightarrow\mathbb{R}^{n} , with slight abuse of notation, construct the following dynamical system that is a