In classical control theory, proving that a system is stable usually requires finding a Lyapunov function—a mathematical expression of the system's internal "energy" or error. According to traditional Lyapunov stability theory, this energy must strictly and monotonically decrease at every single instant in time. If the energy increases even momentarily, the classical proof breaks down.
While mathematically convenient, this strict monotonicity requirement imposes a heavy tax on modern networked and digital control systems. If a controller must guarantee that energy decreases at every instant, it must sample sensors and update actuators at an aggressively high frequency. This leads to excessive communication overhead, rapid battery depletion in wireless nodes, and a lack of resilience against network imperfections like packet loss.
Non-monotonic Lyapunov stability relaxes this fundamental constraint. Instead of forcing the system's energy to decrease monotonically at every moment, it allows the energy to temporarily fluctuate, overshoot, or increase, provided that the long-term trend guarantees convergence. By letting the system "breathe" between updates, this paradigm unlocks massive savings in computational and network resources.
The Intuition: Trading Instantaneous Strictness for Long-Term Trends
To understand why non-monotonic stability works, imagine driving a car down a mountain. Traditional Lyapunov theory demands that your altitude must decrease at every meter of the journey. To guarantee this, you would have to ride the brakes continuously, reacting to every tiny bump or incline to prevent even a millimeter of upward movement.
Non-monotonic stability takes a more practical view: as long as you safely reach the valley at the bottom, it does not matter if you occasionally coast up a few small hills along the way.
In digital and networked control, updates happen at discrete intervals. Between these updates, the system runs "open-loop" on old control signals, meaning its energy can naturally fluctuate. Early work by Mazo et al. demonstrated that relaxing the requirement for continuous energy reduction allows self-triggered controllers to calculate significantly longer "sleep" intervals between updates while still maintaining rigorous stability guarantees.
Modern research has formalized this intuition into two primary methodologies: performance barriers and horizon-based averages.
Mechanism 1: Performance-Barrier Envelopes
One way to safely permit temporary energy increases is to trap the system's energy beneath a known, decaying ceiling. Instead of forcing the Lyapunov function $V(t)$ to satisfy $\dot{V}(t) < 0$ at all times, we allow $V(t)$ to grow, provided it never crosses an explicitly defined performance envelope.
Energy V(t)
^
| \ Performance Envelope (Exponential Decay)
| \---------------------------------------
| \ /\ \ \
| \ / \ /\ \ \
| \_/ \/ \ \----------------------\
| \ / \
| \/ \
+---------------------------------------------------> Time
As long as this envelope decays to zero over time, the underlying system is guaranteed to be asymptotically stable.
Ong et al. leveraged this concept to design event-triggered control systems where a new control signal is transmitted only when the system's energy threatens to breach the performance barrier. The gap between the current energy and the barrier acts as a buffer that absorbs estimation errors. In distributed network simulations, this mechanism reduced communication overhead by nearly 20-fold while theoretically ruling out "Zeno behavior" (the pathological condition of infinite triggers in finite time).
This framework is not limited to simple linear systems. Rathnayake et al. successfully extended performance-barrier event-triggered control to infinite-dimensional systems governed by reaction-diffusion partial differential equations (PDEs). By monitoring a "performance residual"—the explicit mathematical distance between the Lyapunov function and the decaying envelope—the authors demonstrated that global exponential stability can be maintained even with highly sparse boundary control updates.
Mechanism 2: Average and Horizon-Based Decay
A second approach to non-monotonic stability evaluates the system's behavior over a window of time rather than against an instantaneous barrier.
Instead of checking if the energy decreased in the last millisecond, these methods ensure that the energy decreases on average over a longer time horizon, or relative to a historical baseline. For example, Hertneck & Allgöwer proposed a dynamic self-triggered control framework where stability is maintained by ensuring a decrease over a discounted average of past Lyapunov values. By remembering its past trajectory, the controller can determine if it has accumulated enough "energy credit" to safely skip upcoming communication intervals.
This shifting perspective from point-wise constraints to asymptotic trends is highly effective for periodic event-triggered control (PETC). Hertneck et al. showed that allowing the Lyapunov function to fluctuate between sampling points can extend the maximum allowable sampling period (MASP) by more than double compared to traditional monotonic methods.
This aligns with broader theoretical developments in control synthesis. For instance, Kundu & Chatterjee proved that switched systems—which jump abruptly between different operational modes—can achieve asymptotic stability even if individual switching signals cause temporary instability, provided the overall energy balance converges over an infinite horizon. Similarly, Sivaranjani et al. utilized time integrals to evaluate input-output behavior within defined conic bounds, allowing transient violations without sacrificing feedback stability.
Why It Matters: Resilience to Real-World Imperfections
The true power of non-monotonic Lyapunov analysis emerges when control algorithms confront imperfect physical hardware and networks.
Network Dropouts and Weakly-Hard Constraints
In real-world networked control, packets are routinely dropped. If your stability proof requires strict monotonicity, a single lost packet can invalidate your mathematical guarantees. By utilizing non-monotonic Lyapunov functions, Hertneck et al. developed a less conservative stability analysis for systems operating under weakly hard real-time (WHRT) constraints. Because the framework explicitly permits temporary energy increases during communication dropouts, it can guarantee asymptotic stability under much larger nominal sampling periods.
Inaccurate Prediction Models
To save communication bandwidth, modern controllers often use a simplified model at the actuator to "guess" the system's state between updates. However, these models are rarely perfect. Hertneck et al. demonstrated that non-monotonic Lyapunov functions can absorb the state fluctuations and deviations caused by inaccurate or simplified discrete-time prediction models, ensuring long-term convergence without requiring frequent corrective transmissions.
Stochastic Disturbances
When systems face unbounded stochastic noise, enforcing strict step-by-step cost reduction can jeopardize the controller's feasibility. Hewing et al. showed that in stochastic Model Predictive Control (MPC), relaxing the nominal state update rule to allow non-monotonic cost trajectories enables the controller to prioritize feedback and feasibility, improving recovery from unmodeled disturbances while preserving closed-loop chance constraint satisfaction. At a deeper theoretical level, this relies on the fact that highly oscillating or volatile systems can still be proven to converge to a stable average over time, a principle formalized by Han et al. for multi-valued stochastic differential equations.
Summary
Non-monotonic Lyapunov stability bridges the gap between rigid mathematical ideals and the messy reality of networked hardware. By shifting the focus from instantaneous energy descent to average or bounded long-term convergence, it provides engineers with the mathematical tools to build control systems that are drastically more efficient, resilient, and adaptable.
Key Foundational Work
Stability analysis of deterministic and stochastic switched systems via a comparison principle and multiple Lyapunov functions (2006) — contributed 2 concept elements to this direction.
An ISS self-triggered implementation of linear controllers (2010) — contributed 1 concept elements to this direction.