The Mathematics of Stability Under Chaos: When Do Noise and Perturbations Break Stable Systems?

The Fragile Promise of Stability

Imagine you've built a system — a robot, an algorithm, an economic model — that hovers perfectly at an equilibrium point. Small nudges cause it to return. Disturbances fade. Life is stable.

Now introduce noise. Real systems don't exist in isolation. Sensors drift. Wind gusts. Markets wobble. The question that haunts every control engineer, every systems biologist, every financial modeler is deceptively simple: How much chaos can stability absorb before it collapses?

For deterministic systems — those governed by clean equations with no randomness — this question has a satisfying answer. The theory of input-to-state stability, developed over decades by researchers like Eduardo Sontag, tells us precisely when a stable system remains stable under perturbation.

But the world is noisy. Brownian motion drives particle diffusion. Thermal fluctuations jiggle nanoscale machines. Network delays inject randomness into coordinated systems. When the governing equations include random terms — when we're dealing with stochastic differential equations — the old answers don't hold.

A new paper by Robert H. Moldenhauer, Dragan Nešić, Mathieu Granzotto, Romain Postoyan, and Andrew R. Teel, titled "On robustness, input-to-state stability and backstepping for stochastic differential equations," tackles exactly this gap. Their work establishes when and how stochastic systems — those buffeted by noise — remain robust to the perturbations that real life inevitably delivers.

The core finding is this: if you can prove stability of a nominal stochastic system using a Lyapunov function, then that stability survives small perturbations — but only under specific conditions that depend on how the perturbation vanishes (or doesn't) near the equilibrium.

This isn't merely a theoretical nicety. It determines whether the autopilot stays aloft, whether the power grid holds steady, whether the drug delivery system maintains its dose. Understanding robustness in stochastic settings is understanding why some engineered systems fail catastrophically under real-world conditions while others gracefully absorb the chaos.

The Mathematics of Motion and Chance

To understand what Moldenhauer and colleagues accomplished, we need to grasp the setting: stochastic differential equations, or SDEs.

A deterministic differential equation tells you exactly how a system evolves: given the current state and a small time step, the equations pin down the next state. The classic example is Newton's law: position and velocity determine where you'll be an instant later. No mystery. No randomness.

An SDE is different. Alongside the deterministic drift — the expected motion — there's a diffusion term scaled by a Brownian motion. Brownian motion, or the Wiener process as mathematicians prefer, is the mathematical embodiment of random jitter. A particle experiencing thermal fluctuations doesn't follow a precise path; it Diffuses. Its increments over time are independent of the past and normally distributed with variance proportional to the time elapsed. You can't predict exactly where a dust mote will be in ten seconds, but you can characterize the probability distribution.

The canonical SDE looks like this:

$$dx=f(x,u)dt+g(x,u)dw$$

Here, $x$ is the state vector, $u$ is an input or disturbance, $f$ is the drift function determining expected motion, $g$ captures the sensitivity to noise, and $w$ is a Brownian motion. The notation means that over a small time interval $dt$, the state changes by the deterministic increment $f, dt$ plus a random kick $g, dw$ drawn from a normal distribution.

When $u = 0$ and $g = 0$, we recover the familiar deterministic system $dx = f(x), dt$ with its equilibrium points and stability properties. But introduce noise — let $g$ be nonzero — and the mathematics fundamentally changes.

Stability in probability is the natural notion for SDEs. Rather than demanding that trajectories converge with certainty (which is too strict for stochastic systems), GASp asks that trajectories converge with probability approaching 1. The origin is globally asymptotically stable in probability if: (1) small initial conditions stay small with high probability (stability), and (2) regardless of where you start, trajectories converge to the origin with probability 1 (attractivity).

This is the setting Moldenhauer et al. work in.

The Problem: When Does Stability Survive Perturbation?

The paper's central question is robustness: given a stable stochastic system, what perturbations can it tolerate without losing stability?

They formalize this in two ways. The first is a state-dependent perturbation bound: the allowed perturbation size can vary with the state, vanishing at the origin but positive everywhere else. Near equilibrium, you can tolerate only tiny disturbances; far from equilibrium, larger kicks are acceptable. This matches intuition: a pendulum near its resting point is sensitive to small forces, while the same pendulum near the top of its swing is unstable regardless.

The second formalization is stochastic input-to-state stability (ISS). Here, non-vanishing perturbations are allowed everywhere — even at equilibrium — but the system must respond in a bounded way proportional to the input magnitude. A small disturbance yields small deviations; a large disturbance causes larger but still controlled excursions. This is the stochastic generalization of Sontag's classical ISS framework.

The challenge is that existing theory doesn't tell us when these properties hold for SDEs. The deterministic case was largely settled by work in the 1990s, but the stochastic extension has remained open.

Proving Robustness from Lyapunov Functions

The paper's first main result shows how to establish robustness from a Lyapunov function. A Lyapunov function is a kind of energy landscape: if the system's "energy" always decreases, the system must settle to rest. For SDEs, the appropriate object is a stochastic Lyapunov function (SLF) — a twice-differentiable function $V$ satisfying:

• Lower and upper bounds: $\alpha_1(|x|) \leq V(x) \leq \alpha_2(|x|)$, where $\alpha_1, \alpha_2 \in \mathcal{K}_\infty$ (continuous, strictly increasing, zero at origin, unbounded)
• Negative drift: $\mathcal{L}_0 V(x) \leq -\rho(x) < 0$ for $x \neq 0$, where $\mathcal{L}_0$ is the infinitesimal generator accounting for both drift and diffusion

The operator $\mathcal{L}_u$ is the stochastic analogue of the Lie derivative. For a function $V$, it computes the expected rate of change:

$${\mathcal{L}}_{u}V(x)=\frac{\partial V}{\partial x}f(x,u)+\frac{1}{2}\text{Tr}\left(g(x,u{)}^{\top }\frac{{\partial }^{2}V}{\partial x^2}g(x,u)\right)$$

The first term captures deterministic dynamics; the trace term captures how diffusion spreads probability mass.

The key insight is Proposition 1: given a stochastic Lyapunov function for the nominal system (where $u = 0$), you can characterize exactly how large perturbations can be and still maintain negative drift. Specifically, for any desired decay rate $\rho$ weaker than the nominal $\varrho$, there exists a state-dependent bound $\delta \in \mathcal{PD}$ (positive definite, meaning $\delta(0) = 0$ and $\delta(x) > 0$ for $x \neq 0$) such that:

$$\underset{|u|\le \delta (x)}{\max }{\mathcal{L}}_{u}V(x)\le -\rho (x)$$

In words: if the perturbation magnitude stays below $\delta(x)$ everywhere, the Lyapunov function still decays. This immediately implies global asymptotic stability in probability for the perturbed system — Theorem 1.

The result is clean but comes with a caveat. It requires the existence of a $\mathcal{C}_0^2$ Lyapunov function (twice differentiable everywhere except possibly at the origin). This isn't automatically guaranteed by nominal stability; there exist SDEs that are globally asymptotically stable in probability but admit no smooth Lyapunov function at the origin.

The authors acknowledge this gap honestly. A full converse theorem — one that derives a Lyapunov function from stability — remains an open problem for general GASp. But they sidestep this by assuming the Lyapunov function exists, which is the standard approach in the literature.

An important special case resolves the concern. When the nominal system is deterministic (the diffusion term $g(x, 0) \equiv 0$), the classical smooth converse theorem of Lin, Bryne, and Sontag applies. Deterministic global asymptotic stability implies the existence of a smooth Lyapunov function, which then implies robustness to stochastic perturbations. This is stated as Corollary 1.

From Local to Proportional Perturbations

The first result handles perturbations that vanish at the origin. But what about larger perturbations? Can we tolerate non-vanishing disturbances everywhere?

Theorem 2 pushes further. Under stronger assumptions — namely, that $f$ and $g$ are continuously differentiable and the Lyapunov function satisfies certain scaling conditions near the origin — the perturbation bound can be chosen to be at least proportional to the state magnitude:

$$\underset{x\to 0}{\mathrm{liminf}}\frac{\delta (x)}{|x|}>0$$

This means that close to the equilibrium, the allowed perturbation scales linearly with the state. Small states allow proportionally small perturbations; the ratio doesn't vanish.

The proof technique reveals the underlying structure. Writing the drift and diffusion in a specialized form reveals that the perturbation terms enter the Lyapunov drift through expressions scaling as:

$${\mathcal{L}}_{u}V(x)={\mathcal{L}}_{0}V(x)+O\left(\rho (x)\left(\frac{|u|}{|x|}+\frac{|u{|}^2}{|x{|}^2}\right)\right)$$

Near the origin, the perturbation effect is amplified by factors of $|u|/|x|$. If $u$ scales like $x$ — meaning $|u| \leq c|x|$ for some constant $c$ — then these terms remain bounded. This is the mechanism enabling the proportional bound.

This local proportional result is more satisfying. It says that if you're close to equilibrium, small mistakes in your control input (or small disturbances) scaled by a proportional constant won't destabilize the system, as long as that constant is small enough.

Exponential Stability and Global Robustness

The progression continues. Theorem 3 (not fully shown in the excerpt but referenced in the paper) establishes robustness for exponential $p$-stability. This is a stronger property: not only does the system converge, it converges exponentially fast, with moments decaying at a rate $e^{-\alpha t}$.

Under exponential stability, the perturbation bound can be taken global and proportional: $|u| \leq d|x|$ for all $x$, with $d > 0$ sufficiently small. The allowed disturbance doesn't vanish near the origin; it scales with state everywhere.

The key here is leveraging a known converse theorem for exponential stability from R. Z. Khasminskii's classical work. This gives a $\mathcal{C}^2$ Lyapunov function with exponential decay properties, enabling the analysis to go through.

The result also establishes stochastic exponential ISS — the full generalization of ISS to exponential SDEs, without needing state-dependent scaling. If the nominal system is exponentially stable, it automatically absorbs proportional perturbations while maintaining ISS.

The Backstepping Connection

The paper's final contribution might be its most surprising. The authors apply their robustness framework to stochastic backstepping — a control design technique for cascaded nonlinear systems.

Backstepping builds controllers recursively. Start with a stable inner system. Treat its output as a virtual control for the next layer. Design the actual control to make the combined system stable. The technique has been wildly successful for deterministic systems, but the stochastic case is harder.

The problem, as the authors note, is that the cascade of a stable deterministic system and a stochastically ISS system need not be stable — a phenomenon documented in discrete time by Teel and colleagues and adaptable to continuous time. The stochastic nature breaks the standard small-gain arguments that work in deterministic settings.

Moldenhauer et al. solve this by using the same mean-value theorem approach from their robustness analysis. They construct a Lyapunov function for the cascade directly, employing state-dependent scaling similar to the weak ISS framework. The result is a novel backstepping design for pure-feedback form systems, where virtual controls enter the next subsystem nonlinearly — more general than the affine case tackled in earlier work by Deng, Krstić, and others.

The design doesn't directly invoke the paper's robustness theorems, but it shares the same intellectual DNA: a careful accounting of how perturbation terms enter the Lyapunov drift, and a structural decomposition that enables constructive analysis.

The Broader Significance

What does all this mean, practically speaking?

Consider an autonomous vehicle navigating with noisy sensor data. The controller is designed assuming clean measurements, but real sensors deliver signals corrupted by thermal noise, quantization error, and intermittent drops. The vehicle must remain stable despite these imperfections.

The classical approach is to add margins, over-engineer for worst-case disturbances, and test extensively. The Moldenhauer et al. framework offers something more: a principled characterization of exactly how much perturbation the system can tolerate. If your sensor noise stays within the state-dependent bound $\delta(x)$, stability is guaranteed by construction.

Or consider a power grid operating with renewable generation. Wind speed varies randomly; solar output fluctuates with cloud cover. The grid controller must maintain voltage and frequency stability despite stochastic input. The exponential ISS result tells you: if the nominal grid is exponentially stable and the renewable fluctuations scale proportionally with the deviation from nominal operating points, stability persists.

The backstepping result extends these guarantees to more complex architectures. Pure-feedback systems include many practically important cases: neural feedback loops, certain adaptive control schemes, systems with nonlinear input coupling. Being able to design stabilizing controllers for such systems under stochastic perturbations opens new design spaces.

Limitations and Open Questions

The paper is honest about its boundaries.

The main limitation is the dependence on a $\mathcal{C}_0^2$ Lyapunov function. This is assumed, not derived from nominal stability. A full converse theorem for stochastic GASp would tighten the results — nominal stability would directly imply robustness, as in the deterministic case. The authors note this is an open problem.

The analysis also relies on regularity assumptions: local Lipschitz continuity, continuous differentiability of $f$ and $g$ for the stronger results. Real systems don't always cooperate. Singular perturbations, discontinuous dynamics, and non-Lipschitz nonlinearities appear in practice. Extending the framework to these settings is a direction for future work.

The backstepping result is constructive but specialized to pure-feedback form. Strict-feedback systems (where virtual controls enter affinely) are handled differently in the literature; the relationship between these approaches deserves further exploration.

Finally, there's the question of computation. The theoretical guarantees are existentially quantified: "there exists" a perturbation bound $\delta$ such that stability holds. Computing this bound explicitly for a given system requires further analysis. Practical engineers need not just existence proofs but usable formulas.

What Comes Next

The most immediate extension is the missing converse theorem. If someone proves that stochastic global asymptotic stability in probability implies the existence of a $\mathcal{C}_0^2$ Lyapunov function under mild regularity conditions, the entire robustness framework collapses the assumption and becomes a direct consequence of nominal stability — exactly as in the deterministic case.

Numerical methods offer another frontier. The perturbation bounds $\delta$ in the paper are characterized abstractly. Algorithms to compute (or approximate) these bounds from system dynamics would make the theory actionable. Sum-of-squares programming has made progress on this for deterministic Lyapunov functions; a stochastic extension could follow similar lines.

Applications beckon. The framework applies wherever noise meets control: robot motion planning with sensor noise, financial models with stochastic volatility, biological systems with demographic stochasticity, power systems with renewable intermittency. Each domain will reveal new challenges and opportunities.

The backstepping connection is particularly rich. The paper shows that robustness analysis and control design can share tools and techniques. A unified treatment — where robustness guarantees feed directly into controller synthesis — could streamline the design process for stochastic systems.

The Larger Message

Underneath the technical results is a deeper insight about the nature of stability.

In deterministic systems, stability is a binary condition: either trajectories return to equilibrium or they don't. Perturbations either preserve stability or they don't.

In stochastic systems, the picture is richer. Stability becomes probabilistic. Perturbations can be tolerated to varying degrees depending on their structure. The state-dependent bounds in Moldenhauer's work reveal that robustness isn't uniform — it varies with position in state space, stronger far from equilibrium and fragile near it.

This matches lived experience. Systems near operating points are often more sensitive than they appear. The same autopilot that gracefully handles turbulence at altitude can diverge catastrophically if initialized slightly off-nominal. The same power grid that shrugs off a moderate load spike might black out if the spike arrives during a transient.

Understanding these nuances — characterizing exactly how robustness depends on state and perturbation structure — is what makes the stochastic stability framework valuable. It's not just mathematics for its own sake. It's the theoretical foundation for designing systems that survive the chaos of the real world.

The paper by Moldenhauer, Nešić, Granzotto, Postoyan, and Teel advances this foundation. It provides new tools for analyzing stochastic robustness, new techniques for stochastic control design, and new connections between previously separate threads of theory. What's needed next is implementation: turning these existence proofs into computable methods, applying them to real systems, and discovering what new capabilities they enable.

The stability of our engineered world depends on questions like these. Every autopilot, every grid controller, every adaptive algorithm relies on implicit answers to the problems Moldenhauer and colleagues address explicitly. Making those answers rigorous and applicable is the work — and this paper is a significant step forward.