The level of self-organized criticality in oscillating Brownian motion: $n$-consistency and stable Poisson-type convergence of the MLE

This paper establishes that for discretely observed oscillating Brownian motion, the maximum likelihood estimator of the self-organized criticality level achieves $n$-consistency and converges stably to a bivariate Poisson-type distribution, despite the non-standard challenge posed by the discontinuity of the transition density at the true parameter.

The Gist

Imagine you are watching a tiny particle, like a speck of dust, floating in a fluid. Usually, we think of this particle moving randomly, like a drunk person stumbling in a straight line (this is called Brownian Motion).

But in this paper, the authors are studying a special, "bipolar" version of this particle. Imagine the fluid changes its thickness depending on where the particle is.

• If the particle is to the left of a specific invisible line (let's call it the Critical Level), the fluid is thick and slow (like honey).
• If the particle is to the right of that line, the fluid is thin and fast (like water).

The particle doesn't know where the line is; it just bumps around. The big mystery for the scientists is: Where exactly is that invisible line?

The Problem: A Broken Compass

The scientists have a camera taking thousands of photos of the particle's path. They want to use these photos to guess the location of the invisible line.

In most math problems, if you get a little closer to the answer, your "guessing tool" (called the Maximum Likelihood Estimator or MLE) gets smoother and more predictable, like a ball rolling gently down a hill toward the bottom.

But here, the hill is broken.
Because the fluid changes suddenly at the line, the math behind the guess jumps around wildly.

• If you guess the line is slightly to the left, the math says "No, that's wrong!"
• If you guess slightly to the right, it says "No, that's wrong!"
• The "score" for your guess doesn't form a smooth curve; it looks like a jagged, triangular sawblade with sharp spikes.

The Discovery: The "Poisson" Surprise

The authors, Johannes Brutsche and Angelika Rohde, figured out how to handle this jagged, broken math. They discovered something very strange and beautiful:

1. The "Triangle" Shape: When they zoomed in very close to the true line, the jagged math actually formed a perfect, sharp triangle pointing down. The bottom of the triangle is the true answer.
2. The "Jumping" Jumps: But right at the very tip of that triangle, the math doesn't just sit there. It jumps. It behaves like a Poisson process.

What is a Poisson process?
Think of a Poisson process like a rainstorm.

• In normal math (Gaussian), rain falls steadily and evenly.
• In this "Poisson" math, the rain falls in sudden, random cloudbursts. You might have a moment of silence, then CRASH a big drop hits, then silence, then CRASH again.

The authors found that the uncertainty in their guess isn't a smooth wave; it's a series of random, sudden "cloudbursts" (jumps) that happen exactly where the particle crosses the invisible line.

The Solution: Counting the Crossings

To make sense of these random jumps, the authors realized they needed to count how many times the particle crossed the invisible line. This count is called Local Time.

• The Metaphor: Imagine the invisible line is a busy street. The particle is a pedestrian. "Local Time" is simply counting how many times the pedestrian steps on that specific street.
• The more times the particle crosses the line, the more "data" the scientists have, and the sharper their guess becomes.
• If the particle never crosses the line, the scientists can't guess where it is (the math breaks down).

The Big Result: "n-Consistency"

The paper proves that if you take enough photos (let's say $n$ photos), your guess will be incredibly accurate.

• Standard Math: Usually, if you double your data, your error gets smaller by the square root of 2.
• This Paper: Because of the special "jagged" nature of the problem, the error shrinks linearly with the data. If you take 100 times more photos, your error becomes 100 times smaller. This is called $n$-consistency. It's a super-powerful result!

Why Does This Matter?

This isn't just about dust particles. This math applies to:

• Physics: How heat moves through materials that are patchy (some parts hot, some cold).
• Biology: How molecules move through cell membranes that have different densities.
• Finance: How stock prices might behave differently when they cross a certain psychological threshold.

Summary in a Nutshell

The authors solved a puzzle where the usual rules of guessing didn't work because the "terrain" was broken. They found that instead of a smooth slide to the answer, the solution looks like a jagged triangle with random lightning strikes (jumps) at the bottom. By counting how often the particle crosses the invisible line, they proved that with enough data, we can find that invisible line with incredible, lightning-fast precision.

Technical Summary

Here is a detailed technical summary of the paper "The Level of Self-Organized Criticality in Oscillating Brownian Motion: $n$-Consistency and Stable Poisson-Type Convergence of the MLE" by Johannes Brutsche and Angelika Rohde.

1. Problem Statement

The paper addresses the statistical inference problem for the level of self-organized criticality ($\rho$) in an Oscillating Brownian Motion (OBM). The process $X = (X_t)_{t \in [0,1]}$ is defined as the unique strong solution to the homogeneous stochastic differential equation (SDE):
$$dX_t = \sigma_\rho(X_t) dW_t, \quad X_0 = x_0$$
where $W$ is a standard Brownian motion and the diffusion coefficient $\sigma_\rho(x)$ is a step function:
$$\sigma_\rho(x) = \begin{cases} \alpha & \text{if } x < \rho \\ \beta & \text{if } x \ge \rho \end{cases}$$
with known constants $\alpha, \beta > 0$ and $\alpha \neq \beta$. The parameter $\rho$ represents the threshold where the volatility switches.

Key Challenges:

• Discontinuity: Unlike standard change-point models where the break occurs in time, here the break depends on the state of the process. Consequently, the transition density $p^\rho_t(x,y)$ is discontinuous with respect to the parameter $\rho$.
• Non-Standard Asymptotics: The discontinuity implies the log-likelihood function is not continuous (it is only c`adl`ag) and not even upper semi-continuous. This violates standard regularity conditions required for classical Maximum Likelihood Estimator (MLE) theory (e.g., Taylor expansions, standard Gaussian limits).
• Infill Asymptotics: The study considers discrete observations $X_{i/n}$ for $i=1, \dots, n$ as $n \to \infty$ (high-frequency data), rather than long-time asymptotics.

2. Methodology

The authors develop a rigorous asymptotic theory for the MLE $\hat{\rho}_n$ by decomposing the log-likelihood function and exploiting the specific structure of the OBM.

A. Decomposition of the Log-Likelihood
The log-likelihood ratio $\ell_n(\theta)$ (where $\theta = \rho - \rho_0$) is decomposed based on the nine possible regimes of the transition density relative to the true parameter $\rho_0$ and the candidate $\rho_0 + \theta$. These regimes depend on whether the path segments $(X_{(k-1)/n}, X_{k/n})$ lie entirely below $\rho_0$, entirely above, or cross the threshold.
$$\ell_n(\theta) = \sum_{j=1}^9 I_j(\theta)$$
The authors show that the behavior of $\ell_n(\theta)$ changes drastically depending on the scale of $\theta$:

• Global scale ($|\theta| \gg 1/\sqrt{n}$): The likelihood is dominated by terms related to the occupation time of the process.
• $\sqrt{n}$-scale: The likelihood exhibits a "triangular" shape near the true parameter, driven by a negative drift.
• $1/n$-scale: The likelihood exhibits jumps and piecewise linear behavior.

B. Martingale-Drift Decomposition
The sequential log-likelihood process is decomposed into a martingale part $M_n$ and a drift part $B_n$:
$$\ell_n(\theta) = M_n(\theta) + B_n(\theta)$$

• Drift ($B_n$): Proven to be negative and dominant outside a $1/n$-neighborhood of$\rho_0$. It scales linearly with$|\theta|$(triangular shape) in the local neighborhood, driven by the local time of the process at$\rho_0$.
• Martingale ($M_n$): The stochastic fluctuations are driven by rare events where the process crosses the threshold. These events violate the classical Lindeberg condition, leading to non-Gaussian limits.

C. Stable Convergence Framework
Because the limit distribution depends on the random local time $L^{\rho_0}_1(X)$, the authors prove $\mathcal{F}$-stable convergence (convergence in law conditional on the filtration generated by the Brownian motion) rather than simple convergence in distribution. This allows the limit to be multiplied by the observable local time estimator.

D. Limit Process Construction
The limit of the rescaled log-likelihood process is constructed as a bivariate Poisson process with a negative drift. The intensity of the Poisson jumps is proportional to the local time $L^{\rho_0}_1(X)$.

3. Key Contributions and Results

1. $n$-Consistency of the MLE
The paper proves that the MLE $\hat{\rho}_n$ converges to the true parameter $\rho_0$ at the rate of **$1/n$** (i.e.,$n(\hat{\rho}_n - \rho_0) = O_P(1)$) on the event that the process crosses the threshold ($L^{\rho_0}_1(X) > 0$).

• Significance: This is a super-efficient rate compared to the standard $\sqrt{n}$ rate in regular parametric models. The discontinuity of the likelihood allows for faster convergence.
• Proof Strategy: The proof involves a three-step exclusion argument, showing the MLE cannot lie outside a compact set, then a $1/\sqrt{n}$-neighborhood, and finally a$1/n$-neighborhood.

2. Stable Poisson-Type Limit Distribution
The main theorem establishes the stable convergence of the rescaled estimation error:
$$n(\hat{\rho}_n - \rho_0) \mathbb{1}_{\{L^{\rho_0}_1(X) > 0\}} \xrightarrow{\mathcal{F}\text{-st}} \text{arg sup}_{z \in \mathbb{R}} \ell(z L^{\rho_0}_1(X))$$
where the limit process $\ell(z)$ is defined as:
$$\ell(z) = \text{Drift}(z) + \text{Compensated Poisson Process}(z)$$

• The drift is linear and negative.
• The stochastic part is a sum of two independent compensated Poisson processes (one for $z>0$, one for $z<0$) with intensities proportional to the local time.
• The limit distribution is not Gaussian; it is determined by the argmax of a process with jumps.

3. Statistical Inference and Confidence Intervals
Since the local time $L^{\rho_0}_1(X)$ is unobservable, the authors utilize the stable convergence property to construct a practical confidence interval. By replacing the local time with a consistent estimator $\hat{L}^{\hat{\rho}_n}_n$, they derive an asymptotic $(1-\kappa)$ confidence interval for $\rho_0$:
$$\left[ \hat{\rho}_n - \frac{1}{n \hat{L}^{\hat{\rho}_n}_n} q_{1-\kappa/2}, \quad \hat{\rho}_n - \frac{1}{n \hat{L}^{\hat{\rho}_n}_n} q_{\kappa/2} \right]$$
where $q$ denotes quantiles of the limit distribution $\text{arg sup}_{z} \ell(z)$.

4. Significance and Impact

• Theoretical Advancement: This work extends the theory of MLEs to a class of models with discontinuous likelihoods and state-dependent structural breaks, filling a gap between standard change-point models and ergodic threshold diffusion models.
• Non-Gaussian Limits: It demonstrates that in high-frequency settings with discontinuous parameters, the limiting distribution of the MLE is inherently Poissonian rather than Gaussian, challenging the universality of the Central Limit Theorem in this context.
• Super-Efficiency: The $n$-consistency rate highlights that the "roughness" of the likelihood surface near the true parameter actually aids in precise estimation, provided the process visits the critical level.
• Methodological Innovation: The paper introduces a novel technique of analyzing the sequential log-likelihood as a semimartingale in time to bridge the gap between discrete filtration asymptotics and stable convergence results for processes with jumps (extending results by Jacod).

5. Conclusion

Brutsche and Rohde successfully establish that for Oscillating Brownian Motion, the MLE for the critical level $\rho$ is $n$-consistent and converges stably to a distribution characterized by the argmax of a process involving a negative drift and Poisson jumps. The intensity of these jumps is governed by the local time of the process, linking the estimation error directly to the amount of time the process spends near the critical threshold. This provides a complete statistical framework for inference in this non-standard, discontinuous setting.