In introductory physics, economics, or epidemiology, we often study systems that eventually forget their past. A pendulum with friction always comes to rest at its lowest point, regardless of how hard it was initially pushed. A basic disease model typically converges to a single, predictable endemic equilibrium determined entirely by biological parameters like transmission and recovery rates. This property—where a system possesses a unique global attractor—is mathematically comforting but physically fragile.
When dynamical systems become highly non-linear, involve multiple competing agents, or incorporate higher-order group interactions, this uniqueness evaporates. Instead, the system enters a regime of multistability, where multiple distinct steady states coexist. In these landscapes, a system's history is its destiny: the exact initial conditions dictate which long-term equilibrium the system will select.
Continuous Manifolds and the Power of a Head Start
To understand how initial conditions lock in long-term outcomes, consider a symmetric competitive system. Imagine three distinct viruses spreading through a social network, where every virus happens to share the exact same transmission and recovery rates. Intuitively, one might expect them to settle into an equal split, or perhaps for one to randomly eliminate the others.
However, non-linear network dynamics reveal a different mechanism. As shown by Gracy et al. in On the Endemic Behavior of a Competitive Tri-Virus SIS Networked Model, such symmetric systems do not converge to a single isolated point. Instead, they collapse onto a continuous manifold of equilibria—a line or a plane embedded within the state space.
Within this continuous set of choices, the total infection level across the population remains constant, but the relative market share of each virus is a direct projection of its initial state. A slight early-stage advantage or a higher initial dosage allows one strain to permanently capture a larger share of the population. This phenomenon, further verified via Lyapunov-like functions by Gracy et al. in Towards Understanding the Endemic Behavior of a Competitive Tri-Virus SIS Networked Model, demonstrates how structural symmetries in a system can preserve initial variations as permanent, long-term outcomes.
Higher-Order Interactions and Discrete Multistability
The nature of initial condition dependence shifts dramatically when we move from pairwise interactions (edges on a graph) to group-based interactions (simplicial complexes or hypergraphs). In the real world, ideas and pathogens do not just spread from person to person; they accelerate through group settings like dinner parties or boardroom meetings. Group dynamics introduce higher-order non-linearities that break continuous manifolds apart into distinct, competing discrete states.
For instance, in a network where individuals interact in groups, a system can exhibit tri-stability. As discovered by Gracy et al. in Competitive Networked Bivirus SIS spread over Hypergraphs and later extended in Networked competitive bivirus SIS spread with higher order interactions, higher-order interactions allow a disease-free state and multiple distinct single-pathogen outbreak states to be stable simultaneously.
Initial Density of Agent A High ---> Converges to State 1 (Agent A Dominates)
Initial Density of Both Low ---> Converges to State 2 (Disease-Free)
Initial Density of Agent B High ---> Converges to State 3 (Agent B Dominates)
This discrete fragmentation introduces hysteresis and alternative dominance. As Pan demonstrated in Competing spreading dynamics in simplicial complex, the survival of one competing epidemic over another becomes entirely contingent on the historical trajectory of the spread. If you alter the system parameters and then reverse them, the system will not return to its original state; it remains trapped in whichever discrete basin of attraction its history pushed it into. This extreme sensitivity to initial configurations is a universal feature of higher-order structures, appearing equally in physical synchronization models of $D$-dimensional oscillators on simplicial complexes, as detailed by Dai et al. in D-dimensional oscillators in simplicial structures: Odd and even dimensions display different synchronization scenarios.
Navigating Basins of Attraction
When a system's final state is determined by its starting point, the core analytical challenge shifts: we no longer care just about where the equilibria are, but about mapping their basins of attraction—the regions of state space that flow into each specific steady state.
To tackle this, researchers often pivot away from local stability analysis toward Monotone Dynamical Systems (MDS) theory. For example, Doshi et al. leveraged MDS theory in Competing Epidemics on Graphs - Global Convergence and Coexistence to establish global convergence boundaries for competing contagions. Yet, as the authors later noted in Convergence of Bi-Virus Epidemic Models with Non-Linear Rates on Networks - A Monotone Dynamical Systems Approach, while MDS can prove that a non-linear system will eventually settle into a set of coexistence states, explicitly defining the exact geometric boundaries separating these basins remains an open mathematical challenge.
This difficulty spans multiple disciplines:
- Coupled Oscillators: In physical systems, Zhou et al. mapped the intricate basins of attraction for Kuramoto oscillators in Critical points, stability, and basins of attraction of three Kuramoto oscillators with isosceles triangle network, showing how complex coupling forces make predicting the final synchronized state highly sensitive to initial phase configurations.
- Social and Strategic Systems: In pandemic management models governed by strategic game theory, Xuan et al. demonstrated in Optimal Policies for a Pandemic: A Stochastic Game Approach and a Deep Learning Algorithm that systems can settle into entirely different Nash equilibria (e.g., coordinated control vs. escalation) based purely on the initial expectations and starting conditions of the participants.
- Macroeconomics: Even in neoclassical economic growth models with production delays, Wang highlighted in Global exponential stability for a delay differential neoclassical growth model that standard global stability proofs break down when complex, non-unique equilibrium points emerge under varying initial capital-per-labor ratios.
The Non-Equilibrium Frontier
The challenge of non-unique steady states stretches deep into the foundations of statistical mechanics and probability theory. In standard equilibrium thermodynamics, systems are well-behaved and follow universal distributions (like the Gibbs distribution) that forget their initial states. However, when a system is driven out of equilibrium—such as a non-linear chain of oscillators connected to two heat baths at different temperatures—the classical proofs of uniqueness collapse.
In a pair of foundational papers, Eckmann et al. proved that for small temperature differences, a unique, mixing stationary state exists (Non-Equilibrium Statistical Mechanics of Anharmonic Chains Coupled to Two Heat Baths at Different Temperatures). However, the authors established that under large temperature gradients, it remains an open question whether non-linear interactions will cause the system to fragment into multiple coexisting configurations (Entropy Production in Nonlinear, Thermally Driven Hamiltonian Systems).
This lack of a universal framework for non-equilibrium systems is a major open problem. As Lancia et al. observed in Equilibrium and Non-equilibrium Ising Models by Means of PCA, irreversible systems lack a general formula for calculating stationary measures, making their long-term states highly dependent on specific local dynamics and initial conditions.
This fundamental gap is mirrored in pure probability theory. When analyzing stochastic wave equations, Thomas noted in Low regularity solutions to a gently stochastic nonlinear wave equation in nonequilibrium statistical mechanics that characterizing non-equilibrium invariant measures remains unaddressed. Ultimately, as Lasserre et al. pointed out in On the classification of Markov chains via occupation measures, a unified classification for Markov chains on general state spaces with non-unique invariant measures remains an unsolved challenge. We are left with a profound realization: across physical, biological, and social systems, when uniqueness is lost, history matters.
Key Foundational Work
Non-Equilibrium Statistical Mechanics of Anharmonic Chains Coupled to Two Heat Baths at Different Temperatures (1999) — contributed 3 concept elements to this direction.
Entropy Production in Nonlinear, Thermally Driven Hamiltonian Systems (1999) — contributed 1 concept elements to this direction.
On the classification of Markov chains via occupation measures (2000) — contributed 1 concept elements to this direction.