📊 statistics

On goodness-of-fit testing for volatility in McKean-Vlasov models

This paper establishes a rigorous statistical framework for goodness-of-fit testing of volatility functions in McKean-Vlasov stochastic differential equations by proposing a test based on discrete particle observations and proving its asymptotic normality in a joint regime where both the number of particles and sampling frequency tend to infinity.

Original authors: Akram Heidari, Mark Podolskij

Published 2026-08-10

📖 8 min read🧠 Deep dive

CC BY 4.0

Original authors: Akram Heidari, Mark Podolskij

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a bustling city where millions of people are constantly moving, making decisions, and reacting to one another. In the world of science, this is often modeled using something called a "diffusion system." Think of it like a giant, invisible cloud of particles swirling around. In the old, classic way of studying these clouds, scientists assumed that each particle only cared about its own immediate neighborhood—like a person walking down a street who only looks at the ground right in front of them. But in the real world, things are more connected. A person's walk might change because they see a crowd gathering ahead, or because they hear a rumor spreading through the city. This is the realm of "McKean–Vlasov" models: a fancy way of saying that every single particle's movement depends not just on where it is, but on the entire shape and mood of the crowd it belongs to.

Now, every time these particles move, they jitter. Sometimes they take small, cautious steps; other times they lunge forward wildly. In math, this jitteriness is called "volatility." It's the measure of how chaotic or unpredictable the system is. If you want to predict where the crowd will go, or how risky a financial market might be, you need to know exactly how much that jitter happens. But here's the tricky part: in these complex, crowd-dependent systems, nobody had built a reliable "lie detector" to check if our guesses about the jitter were actually right. We could guess the rules, but we didn't have a way to rigorously test if our guess matched the reality of the data we collected.

This paper steps in to fix that gap. The authors, Akram Heidari and Mark Podolskij, have developed a new statistical "lie detector" specifically for these crowd-dependent systems. They created a test that looks at a massive collection of particles (imagine watching thousands of people simultaneously) and checks if the "jitteriness" of the crowd follows a specific pattern we think it does. They proved that if you watch enough people for long enough, and if you watch them very closely (high-frequency data), your test will tell you with a prescribed level of statistical confidence whether your model of the chaos is correct or if it's significantly off-base. It's like finally having a way to say, "Okay, we thought the crowd was jittering like a nervous cat, but the data shows they were actually jittering like a shaken soda can."

The Story of the Crowd and the Jitter

Let's dive into the details of this scientific adventure. The authors are working with a system of $N$ particles. To visualize this, imagine a giant stadium filled with $N$ people, where $N$ is a very large number. Each person is a "particle," and they are all moving around over a fixed period of time, let's say from the opening whistle to the final buzzer.

In this specific setup, the people are independent copies of each other, but they are all influenced by the same "crowd law." This means that while Person A doesn't directly talk to Person B, Person A's movement is influenced by the average behavior of everyone in the stadium. If the crowd starts to panic, Person A might start running, even if they didn't see the specific person who started the panic. This is the "McKean–Vlasov" magic: the individual is shaped by the collective.

The big question the authors wanted to answer was about the "volatility." In our stadium analogy, volatility is how much the people are shuffling, stumbling, or sprinting randomly. In the real world, this could be the price of a stock jumping around, or the temperature fluctuating in a chemical reaction. The authors wanted to test a specific hypothesis: "Does the way these people shuffle follow a specific formula we wrote down?"

Usually, in simpler models, you just look at one person's path to figure out the rules. But in this crowd-dependent world, looking at just one person isn't enough because their path is a mix of their own choices and the crowd's influence. So, the authors came up with a clever plan. They proposed a "Goodness-of-Fit" test. Think of this as a game of "Spot the Difference."

Here is how the game works:

The Guess: You have a theory about how the crowd shuffles. Maybe you think the shuffling is a simple mix of three different types of movements (like walking, jogging, and dancing). You write this down as a mathematical formula.
The Observation: You go to the stadium and take snapshots of everyone's position at very tiny intervals. You don't just watch one person; you watch all $N$ people, and you take these snapshots very frequently (high-frequency data).
The Test: You compare what you actually saw to what your formula predicted. You calculate a "distance" between your guess and reality. If the distance is zero (or very close to it), your guess is good. If the distance is huge, your guess is wrong.

The authors didn't just invent a test; they proved that this test actually works. They showed that as you increase the number of people ( $N$ ) and the speed at which you take snapshots (making the time between snapshots, $\Delta_n$ , smaller and smaller), your test becomes incredibly accurate.

There is a catch, though. The authors found that you can't just crank up the number of people and the speed of observation however you like. There is a delicate balance. They proved that for the test to work perfectly, the number of people ( $N$ ) and the square of the time between snapshots ( $\Delta_n^2$ ) must multiply to a very small number. In plain English, this means that if you have a huge crowd, you need to be extremely fast with your observations to avoid errors. If you are too slow, the "noise" in your data will mess up the test, and you might think your formula is right when it's actually wrong.

What They Found

The paper's main discovery is a mathematical proof that this testing method is valid. They established that under the right conditions (lots of particles, very fast observations, and that specific balance between the two), the "distance" your test calculates behaves in a predictable way.

Specifically, they proved that if your guess about the crowd's jitter is correct, the test statistic will follow a "bell curve" (a normal distribution) centered around zero. This is huge because it means you can set a "confidence level." For example, if you want to be 95% sure, you can set a threshold. If your test result is way outside that threshold, you know with 95% confidence that your guess about the volatility is wrong.

They also tackled a practical problem: how do you know what the "bell curve" looks like if you don't know the exact rules of the crowd? They proposed a method called "sub-sampling." Imagine you have a stadium of 10,000 people. Instead of using all of them to figure out the curve, you split them into 100 smaller groups of 100. You run the test on each group separately. Because the groups are independent, you can use the results from these smaller groups to estimate the shape of the big curve. They proved that this method works and gives you a reliable way to decide if your model is good or bad.

Why This Matters

Before this paper, scientists had great tools for testing simple systems where particles acted alone. But for complex systems where everyone influences everyone else—like financial markets, biological populations, or large-scale engineering systems—there was no rigorous way to check if the "jitter" part of the model was correct. You could build a beautiful model, but you couldn't be sure if the part that described the chaos was accurate.

This paper provides the first rigorous framework to answer that question. It doesn't just say, "Hey, this looks like it might work." It says, "Here is the math that proves this test works, here is how you calculate the error, and here is how you know when to reject a bad model."

The authors are very clear about what they didn't do, too. They didn't claim to solve the problem of predicting the future. They didn't say their test works for every single type of crowd behavior imaginable. They specifically focused on systems where the particles are independent copies of each other (i.i.d.) and where the observations are discrete (snapshots, not a continuous video). They also noted that if the balance between the number of particles and the observation speed isn't right, the test might have a bias, meaning it could give you a false sense of security.

In the end, this paper is like handing a scientist a new, calibrated ruler. Before, they were trying to measure the chaos of a complex crowd with a ruler made of rubber. Now, they have a rigid, mathematically proven tool that tells them exactly how much their model of the chaos deviates from reality. It's a foundational step that allows researchers to trust their models more, knowing they have a way to check the most unpredictable part of the system: the volatility.

Technical Summary: Goodness-of-Fit Testing for Volatility in McKean–Vlasov Models

Problem Statement
This paper addresses the lack of statistical methodology for validating the structural assumptions of volatility functions in independent and identically distributed (i.i.d.) McKean–Vlasov stochastic differential equations (SDEs). While integrated volatility estimation for classical SDEs is well-established, formal model validation and goodness-of-fit testing for McKean–Vlasov systems—where dynamics depend on the probability distribution of the solution—remain largely unexplored. This gap is particularly critical in settings combining large particle limits with high-frequency observations. The specific challenge lies in the distributional dependence of the volatility coefficient, which introduces nonlinear interactions and path-dependent effects that prevent the direct application of standard high-frequency statistical techniques.

Methodology
The authors propose a goodness-of-fit testing framework based on discrete observations of a particle system consisting of $N$ particles observed at $n$ time points over a fixed interval $[0, T]$ . The analysis operates in a joint asymptotic regime where both the number of particles $N$ and the sampling frequency $n$ tend to infinity, while the time horizon $T$ remains fixed.

Hypothesis Formulation: The null hypothesis ( $H_0$ ) posits that the squared volatility function $a^2(x, \mu_t)$ belongs to a linear span of known basis functions $\{a_1^2, \dots, a_d^2\}$ . The test statistic is constructed based on the $L^2$ -distance $L$ between the true squared volatility and the subspace generated by the basis functions, measured with respect to the law $\mu_t$ of the particles.
Estimator Construction: Since the true distribution $\mu_t$ $μ_{t}$ is unobservable, the authors replace it with the empirical measure $\mu_t^N$ $μ_{t}^{N}$ . They construct empirical estimators for the quantities defining the distance measure $L$ $L$ :
- $\hat{B}$ : An estimator for the integrated fourth moment of the volatility.
- $\hat{\Gamma}_k$ : Cross-moment estimators between the basis functions and the true volatility.
- $\hat{\Lambda}_{k,l}$ : Gram matrix estimators for the basis functions.
  The test statistic $\hat{S}_N$ is defined as the empirical analogue of the closed-form expression for $L$ , specifically $\hat{S}_N = \hat{B} - \hat{\Gamma}^\top \hat{\Lambda}^{-1} \hat{\Gamma}$ .
Asymptotic Analysis: The derivation of the limiting distribution relies on a joint limit where $N \to \infty$ and $\Delta_n = T/n \to 0$ , subject to the condition $N\Delta_n^2 \to 0$ . This condition is crucial to ensure that the bias arising from the Riemann-sum approximation and the error from replacing the true measure with the empirical measure do not dominate the asymptotic variance.

Key Results
The paper establishes the following theoretical results under Assumptions 1–3 (covering initial density, Lipschitz continuity, linear growth, and linear functional differentiability of coefficients):

Asymptotic Normality of Components: The authors prove stochastic expansions for the estimators $\hat{\Lambda}$ , $\hat{B}$ , and $\hat{\Gamma}_k$ . They demonstrate that $\sqrt{N}(\hat{\Lambda} - \Lambda)$ , $\sqrt{N}(\hat{B} - B)$ , and $\sqrt{N}(\hat{\Gamma}_k - \Gamma_k)$ converge to centered Gaussian processes. A key technical finding is that the term involving the empirical measure $\hat{\Lambda}$ introduces an additional asymptotically normal component due to the randomness of replacing $\mu_t$ with $\mu_t^N$ .
Central Limit Theorem for the Test Statistic: Under the null hypothesis, the test statistic satisfies a central limit theorem:
$\sqrt{N}\hat{S}_N \xrightarrow{L} \mathcal{N}(0, \tau^2)$
where $\tau^2$ is an asymptotic variance determined by the gradient of the distance function and the covariance structure of the underlying estimators.
Consistency and Power: The procedure is shown to be consistent against fixed alternatives (where $L > 0$ ). Furthermore, the authors analyze local alternatives of the form $a^2 = N^{-1/4}a_0^2 + \sum \lambda_k a_k^2$ , showing that the limiting distribution acquires a non-zero mean proportional to the distance of the perturbation from the null space, thereby allowing for the derivation of asymptotic power.
Variance Estimation: Due to the complex dependence on the empirical measure, the authors propose a subsampling procedure to estimate the asymptotic covariance matrix $\Sigma$ . They prove that this subsampling estimator is weakly consistent, enabling the construction of a feasible test that attains the prescribed significance level.

Significance and Claims
The paper claims to provide the first rigorous goodness-of-fit testing framework for volatility functions in i.i.d. McKean–Vlasov models based on discrete observations of interacting particle systems. The significance of this work lies in:

Extending Statistical Validation: It extends the scope of statistical model validation to diffusion models with distribution-dependent dynamics, a class of models increasingly used in economics, finance, and physics to model systemic risk and mean-field interactions.
Methodological Innovation: It overcomes the methodological challenge of distributional dependence by combining high-frequency asymptotics with mean-field arguments, specifically addressing the interaction between the number of particles and the sampling frequency.
Practical Applicability: By establishing a central limit theorem and providing a consistent variance estimator via subsampling, the paper offers a practically implementable procedure for assessing volatility specifications in high-dimensional mean-field diffusion models.

The authors explicitly note that their results rely on the condition $N\Delta_n^2 \to 0$ ; if this condition is violated (e.g., $N\Delta_n^2 = O(1)$ ), a non-negligible asymptotic bias arises, which would complicate the construction of a valid test without further knowledge of the drift and volatility coefficients.

The Story of the Crowd and the Jitter

What They Found

Why This Matters

More like this