Small noise asymptotics for a class of jump-diffusions with heavy tails for large times

This paper establishes that for positive recurrent Lévy diffusions driven by scaled Brownian motion and $\alpha$-stable processes ($1<\alpha<2$) in the small noise regime, the large-time limiting behavior of the one-dimensional marginal distribution is determined by the optimal value of a deterministic control problem featuring both continuous and impulse controls.

The Gist

Imagine you are trying to guide a very stubborn, slightly drunk hiker (the diffusion process) back to a cozy campfire (the stable equilibrium) in the middle of a vast, foggy forest.

Usually, the hiker wanders around due to small, random gusts of wind (the Brownian motion). In the classic world of physics, if you make those wind gusts very tiny, the hiker will almost always stay right next to the fire. The probability of the hiker wandering far away drops off very quickly, like a smooth hill.

But this paper asks a different question: What happens if the hiker isn't just buffeted by gentle wind, but is also occasionally hit by giant, unpredictable boulders rolling down the hill? These boulders represent "heavy tails" or $\alpha$-stable processes. They are rare, but when they happen, they can knock the hiker miles away instantly.

The authors of this paper are trying to figure out: If we make the wind and the boulders very small (but the boulders still exist), where will the hiker end up after a very long time?

Here is the breakdown of their discovery using simple analogies:

1. The Two Types of "Noise"

The hiker is being pushed by two things:

• The Gentle Wind (Brownian Motion): This pushes the hiker continuously. It's like a constant, soft nudge.
• The Giant Boulders (Heavy Tails/Jumps): These are rare, massive jumps. The hiker might be sitting still, and BOOM, a boulder hits them, and they are suddenly 100 meters away.

The paper studies a specific scenario where the "wind" is scaled down by a factor of $\sqrt{\log n}$ and the "boulders" are scaled down by $1/n^\gamma$. As$n$ gets huge, both forces become tiny, but they behave differently.

2. The "Control" Problem: How to Get Back?

The core of the paper is about finding the cheapest way to get the hiker back to the campfire. The authors treat this as a game of "Optimal Control."

Imagine you are the hiker's coach. You have two tools to guide them back:

1. Continuous Steering: You can gently push them with your hand (like the wind). This costs energy based on how hard you push.
2. The Teleportation Button (Impulse Control): You can instantly teleport them a short distance. This is like the "jump" from the boulder.

The Big Twist:
In the classic "wind only" world, the cost to get back is based on how far you push.
In this "boulder" world, the cost of a jump depends only on how many times you jump, not how far you jump.

• Analogy: Imagine a taxi service where the price is $10 per ride, regardless of whether you go 1 block or 100 blocks. If you need to go far, it's cheaper to take one giant taxi ride than to take 100 tiny ones.

3. The Main Discovery

The authors prove that even though the boulders are rare, they change the rules of the game completely.

• The "Rate Function" (The Map of Danger): In the old world, the map of danger was a smooth hill. In this new world, the map is a staircase.
• The Strategy: To get the hiker back to the campfire efficiently, the optimal strategy is a mix:
  • Use continuous steering (the wind) to handle small adjustments.
  • Use impulse jumps (the boulders) only when the hiker is very far away.
  • Crucially: Because the "taxi" (jump) has a fixed price per ride, the optimal strategy involves taking very few, very large jumps rather than many small ones. You don't want to pay the "entry fee" for a jump too many times.

4. Why This Matters

The paper solves a mathematical puzzle that previous methods couldn't crack.

• The Problem: Standard math tools (Large Deviation Principles) work great for smooth wind but break down when giant boulders are involved because the math for "boulders" is too messy (it's a "Weak" Large Deviation Principle).
• The Solution: The authors used a clever trick from Control Theory (specifically Dynamic Programming). Instead of trying to predict the exact path of the hiker, they calculated the "cost" of the best possible path.
• The Result: They showed that even with these wild, heavy-tailed jumps, the long-term behavior of the system is still predictable. It's dictated by a specific "cost function" that balances the effort of walking (continuous control) against the fixed cost of jumping (impulse control).

Summary in One Sentence

This paper shows that when a system is disturbed by both gentle winds and rare, massive jumps, the most efficient way to return to stability is a hybrid strategy: walk most of the way, but when you are far off, take a few massive "teleportation" jumps, because the "price" of jumping is fixed regardless of distance.

The Takeaway: In a world of heavy-tailed risks (like financial crashes or extreme weather), the best recovery strategy isn't to react to every little fluctuation, but to endure the small stuff and make bold, calculated leaps when necessary.

Technical Summary

Here is a detailed technical summary of the paper "Small Noise Asymptotics for a Class of Jump-Diffusions with Heavy Tails for Large Times" by Sumith Reddy Anugu, Siva R. Athreya, and Vivek S. Borkar.

1. Problem Statement

The paper investigates the small noise asymptotics of the invariant measures (or long-term marginal distributions) for a specific class of Lévy diffusions driven by both a Brownian motion and an $\alpha$-stable process with heavy tails ($1 < \alpha < 2$).

The system is modeled by the following stochastic differential equation (SDE) in $\mathbb{R}^p$:
$$X_{n,\gamma}(t) = X_{n,\gamma}(0) + \int_0^t b(X_{n,\gamma}(s)) ds + \frac{1}{\sqrt{\log n}}W(t) + \frac{1}{n^\gamma}L_\alpha(t)$$
where:

• $b(\cdot)$ is a globally Lipschitz drift term such that the associated deterministic ODE $\dot{z} = b(z)$ has a unique, globally asymptotically stable equilibrium at the origin ($z=0$).
• $W(t)$ is a standard $p$-dimensional Brownian motion.
• $L_\alpha(t)$ is a $p$-dimensional $\alpha$-stable process ($1 < \alpha < 2$).
• $n \to \infty$ represents the vanishing noise limit.
• $\gamma > 0$ is a scaling parameter controlling the relative strength of the jump component.

The Core Challenge:
In classical Freidlin-Wentzell theory (for pure Brownian motion), the large deviation principle (LDP) for the invariant measure is governed by a rate function derived from an optimal control problem involving only continuous controls. However, for $\alpha$-stable processes with $1 < \alpha < 2$, the sample-path LDP **does not hold** under standard scaling; only a **Weak Large Deviation Principle (WLDP)** holds. Furthermore, the cost of jumps in the$\alpha$-stable regime is discrete (based on the number of jumps) rather than continuous (based on the magnitude of the path). The authors aim to characterize the limiting behavior of the one-dimensional marginal distribution$X_{n,\gamma}(T)$as$n \to \infty$followed by$T \to \infty$.

2. Methodology

The authors employ a combination of large deviation theory, control theory, and dynamic programming.

A. Weak Large Deviation Principle (WLDP)

Since the standard LDP fails for the scaled $\alpha$-stable process $\frac{1}{n^\gamma}L_\alpha$, the authors rely on the WLDP established in prior literature (specifically [14]). They establish that the family of processes satisfies a WLDP with a rate function that is discrete-valued, penalizing the number of jumps rather than their size.

B. Contraction Principle for WLDP

A major technical hurdle is that the standard contraction principle (which maps LDPs through continuous functions) does not automatically apply to WLDPs. The authors prove a weak contraction principle (Lemma 2.10) tailored for their setting. They leverage the fact that the $\alpha$-stable process is a pure jump process. By restricting the analysis to sets of piecewise constant functions and utilizing the compactness of level sets for the rate functions of the Brownian and initial components, they show that the WLDP is preserved under the continuous mapping defined by the SDE solution.

C. Optimal Control Formulation

The rate function for the limiting distribution is identified as the value function of a deterministic optimal control problem.

• Control Space: The control consists of two parts:
  1. Continuous Control ($u$): Corresponds to the Brownian noise, with a quadratic cost $\frac{1}{2}\int \|u(t)\|^2 dt$.
  2. Impulse Control ($v$): Corresponds to the jumps. The cost is linear in the number of jumps: $\gamma \cdot I_L(v)$, where $I_L(v)$ counts the total number of jumps across dimensions.
• Time Reversal: To handle the infinite time horizon ($T \to \infty$), the authors use a time-reversal technique. They transform the problem of finding the probability of the system being at state $x$ at time $T$ into a problem of steering the time-reversed dynamics from $x$ back to the origin (or a neighborhood of it) with minimum cost.

3. Key Results

The Main Theorem (Theorem 1.2)

The paper proves that the family of random variables $\{X_{n,\gamma}(T)\}$ satisfies a Weak Large Deviation Principle as $n \to \infty$ followed by $T \to \infty$. The rate function $V_\gamma(x)$ is given by:
$$V_\gamma(x) = \inf_{(u,v) \in \mathcal{R}_x} \left( \frac{1}{2} \int_0^\infty \|u(s)\|^2 ds + \gamma I_L(v) \right)$$
where $\mathcal{R}_x$ is the set of control pairs $(u, v)$ that drive the time-reversed system from $x$ to the origin asymptotically.

Key Properties of the Rate Function $V_\gamma(x)$:

1. Independence from Initial Distribution: Unlike finite-time marginals, the limiting rate function $V_\gamma(x)$ is independent of the initial condition's rate function (provided the initial condition satisfies an LDP with a minimum at 0).
2. Finite Impulses: A crucial finding is that for any $x$, the optimal control strategy involves only a finite number of impulses (jumps). Specifically, the optimal path uses at most $p$ jumps (one per dimension) to reach the origin efficiently.
3. Scaling Dependence: The parameter $\gamma$ acts as a penalty weight for jumps.
   • As $\gamma \to \infty$, the penalty for jumps becomes prohibitive, and the optimal strategy approaches a purely continuous control (resembling the classical Brownian case).
   • As $\gamma \to 0$, jumps become "cheap," and the system may utilize impulses to escape the basin of attraction or reach the origin faster.
4. Convergence: The authors prove that the finite-time rate function $V_{\gamma, T}(x)$ converges pointwise and uniformly on compact sets to $V_\gamma(x)$ as $T \to \infty$.

4. Significance and Novelty

• Bridging Heavy Tails and Large Deviations: This work extends the classical Freidlin-Wentzell theory (which assumes Gaussian noise) to systems with heavy-tailed noise ($\alpha$-stable). It demonstrates that even when the sample-path LDP fails, a meaningful WLDP can be derived for the marginals of the process.
• Mixed Control Problems: The paper introduces a novel optimal control framework that combines continuous controls (for diffusion) and impulse controls (for jumps) with a cost structure that depends on the count of jumps rather than their magnitude. This reflects the specific nature of $\alpha$-stable processes where large jumps are rare but significant.
• Technical Innovation: The proof of the weak contraction principle for WLDPs in the presence of pure jump processes is a significant technical contribution. It overcomes the lack of standard LDPs for $\alpha$-stable processes by exploiting the specific structure of their sample paths (piecewise constant with finite jumps on compact intervals).
• Physical Interpretation: The results provide insight into how systems with heavy-tailed noise behave in the long run. They suggest that rare events (large deviations) in such systems are driven by a combination of small continuous fluctuations and a finite number of large jumps, with the trade-off determined by the parameter $\gamma$.

5. Conclusion

The paper successfully characterizes the large-time, small-noise asymptotics of Lévy diffusions with heavy tails. It establishes that the limiting behavior is governed by a rate function derived from a mixed continuous-impulse optimal control problem. This result generalizes classical diffusion theory to heavy-tailed environments, offering a rigorous mathematical foundation for understanding rare events in systems subject to both Gaussian and $\alpha$-stable noise.