Imagine a system that abruptly changes its behavior over time, jumping between different operational modes. A classic example is a walking robot making and breaking contact with the ground, or a networked control system experiencing intermittent data packet drops. Even if every individual mode is perfectly stable on its own, switching between them can cause the system's state to diverge dramatically. Conversely, carefully orchestrated switching can stabilize a collection of entirely unstable modes.
To analyze these systems, control theorists rely on Lyapunov stability theory, which generalizes the notion of energy dissipation. If a system's total "energy" strictly decreases over time, it must eventually come to rest at an equilibrium point. However, finding a single, global Lyapunov function that decreases across all possible modes and transitions is often mathematically impossible. The solution is to use Multiple Lyapunov Functions (MLFs)—assigning a distinct, tailored energy function to each individual subsystem or mode, and tracking how these virtual energy levels evolve across transitions.
Tracking Virtual Energy Mode-by-Mode
The core intuition behind the MLF approach is that we do not need a single energy function to decrease continuously at every single moment. Instead, we can allow the system's active energy function to change whenever the system switches modes. As Atreyee Kundu et al. demonstrate in Generalized switching signals for input-to-state stability of switched systems, assigning multiple Lyapunov-like functions to individual subsystems allows for an independent assessment of stability for each component within a switched system.
However, if the energy function changes at every switch, what prevents the system from "ratcheting" up in energy over time? In Stability analysis of deterministic and stochastic switched systems via a comparison principle and multiple Lyapunov functions, Daniel Liberzon et al. show that overall stability depends heavily on the behavior of these functions specifically at the switching instants. To guarantee stability, the energy of a specific mode must be lower each time that mode is reactivated than it was when the mode was last turned off.
Atreyee Kundu et al. expand on this in Stabilizing switching signals: A transition from point-wise to asymptotic conditions, showing that stability analysis must account for both the energy dissipation within specific modes and the potential energy increases that occur during transitions between modes. This mode-specific verification framework, as noted by Majid Zamani et al. in Symbolic models for stochastic switched systems: A discretization and a discretization-free approach, allows researchers to verify stability and construct symbolic models for complex systems where a single global function cannot be found.
Managing Boundaries and Discontinuous Transitions
In many real-world applications, switching is state-dependent, triggered when the system state hits a specific boundary in space. For example, in robotics, physical contact introduces sudden changes in dynamics. Alp Aydinoglu et al. show in Stability Analysis of Complementarity Systems with Neural Network Controllers that piecewise quadratic energy functions are essential for modeling the distinct energy profiles of different physical contact modes, providing a precise tool for identifying regions of attraction where friction or impacts prevent simple convergence.
When these boundaries partition the state space into complex regions, Rafael Wisniewski et al. in Local analysis of hybrid systems on polyhedral sets with state-dependent switching utilize differential inclusions and convexification to apply Lyapunov-like functions across partitioned state spaces. This geometric approach becomes vital when systems exhibit infinite switching or sliding modes along boundaries. To handle these complex behaviors, Hamed Mojallali et al. formulate a generalized Lyapunov stability theorem in On robust stability of switched systems in the context of Filippov solutions, which explicitly ensures a decrease in the subsystem-specific Lyapunov functions across the switching surfaces.
For nonlinear systems, the mathematical structure of the derivative matters immensely. Kuize Zhang et al. demonstrate in High-Order S-Lemma with Application to Stability of a Class of Switched Nonlinear Systems that stability can be evaluated by determining if the system admits a Lyapunov function with a homogeneous derivative, establishing conditions under which a convex combination of individual modes translates to the stability of the entire system. Furthermore, when designing controllers via Model Predictive Control (MPC), Pascal Martius et al. show in Model predictive control of switched nonlinear systems under average dwell-time that the optimal value functions of individual subsystems must satisfy strict compatibility conditions to ensure that the cumulative cost across switching events does not lead to instability.
Embracing Stochasticity and Random Switching
Systems rarely switch in a perfectly predictable, deterministic fashion. Environmental noise, random network delays, and probabilistic failures mean that mode transitions are frequently stochastic. To analyze these systems, the MLF framework must be extended to evaluate the expected values or statistical moments of the Lyapunov functions.
Debasish Chatterjee addresses this in Stabilizing randomly switched systems by deriving inequalities that directly connect the properties of subsystem Lyapunov functions with the statistical characteristics of the switching signal, even when the switching follows a general, non-Markovian random process. When the transitions can be modeled as a Markov chain, incorporating transition probabilities yields much tighter, less conservative bounds. For instance, Ioannis Kordonis et al. demonstrate in Stochastic Stability in Max-Product and Max-Plus Systems with Markovian Jumps that accounting for Markovian transition probabilities allows for higher operational throughput in discrete event systems without compromising stability.
This probabilistic framing also helps stabilize networked control systems; Li Xie shows in Stabilizing Sampled-data Linear Systems with Markovian Packet Losses and Random Sampling that systems with Markovian packet losses can be analyzed as equivalent randomly sampled systems, bridging stability concepts across continuous-time, discrete-time, and random sampling domains. When synthesizing control policies for these systems, Bo Wu et al. in Switched Linear Systems Meet Markov Decision Processes: Stability Guaranteed Policy Synthesis show that while deterministic policies often fail to stabilize systems with probabilistic mode transitions, randomized policies synthesized via coordinate descent can successfully guarantee stability within a stochastic framework.
Surprisingly, introducing randomness and asynchrony can sometimes stabilize an otherwise unstable system. Sahin Lale et al. demonstrate in Stability and Identification of Random Asynchronous Linear Time-Invariant Systems that stability in randomized linear systems can be evaluated through an extended Lyapunov equation utilizing a positive linear map, allowing researchers to analytically determine whether stochastic update patterns will stabilize or destabilize the underlying dynamics.
To move beyond specific noise assumptions like Markov or i.i.d. noise, Tomomichi Hagiwara developed a generalized framework in On Second-Moment Stability of Discrete-Time Linear Systems With General Stochastic Dynamics, deriving new Lyapunov inequalities that serve as necessary and sufficient conditions for various stability notions under arbitrary stochastic processes. This generalized view is echoed by Bruno Sinopoli et al. in A Random Dynamical Systems Approach to Filtering in Large-scale Networks, who advocate for a Random Dynamical Systems approach to capture long-term statistical behaviors and ergodicity that traditional moment-based stability might overlook.
Memory, Shocks, and Non-Integer Dynamics
The ultimate test of the multiple Lyapunov function approach is its applicability to systems governed by non-standard calculus, such as fractional derivatives or impulsive shocks. When a system experiences both continuous random noise and sudden, instantaneous shocks, standard stochastic calculus tools break down due to non-differentiability. To resolve this, Junwei Liu demonstrates in EXISTENCE AND STABILITY OF ALMOST PERIODIC SOLUTIONS TO IMPULSIVE STOCHASTIC DIFFERENTIAL EQUATIONS that exponential stability can still be rigorously established by verifying long-term system behavior through piecewise almost periodic processes.
For systems that incorporate long-range memory alongside heavy-tailed noise, Tingting Hou shows in The stability with general decay rate of hybrid stochastic fractional differential equations driven by Lévy noise with impulsive effects that almost sure stability is guaranteed if the energy dissipation defined via Lyapunov functions successfully compensates for the disruptions introduced by both continuous Lévy noise and discrete shocks.
To evaluate these non-integer order systems without explicitly solving complex fractional differential equations, Jerzy Baranowski et al. in Applications of Direct Lyapunov Method in Caputo Non-Integer Order Systems leverage specialized mathematical inequalities designed for fractional derivatives of quadratic and power functions to directly evaluate the rate of change of Lyapunov candidates. By tracking energy through these specialized mathematical lenses, multiple Lyapunov functions provide a unified framework for guaranteeing stability across the deterministic, stochastic, and hybrid domains.
Key Foundational Work
Stability analysis of deterministic and stochastic switched systems via a comparison principle and multiple Lyapunov functions (2006) — contributed 3 concept elements to this direction.
Computation of Piecewise Quadratic Lyapunov Functions for Hybrid Systems (1996) — contributed 2 concept elements to this direction.