Local asymptotic normality for discretely observed McKean-Vlasov diffusions

Basic Information

• Paper ID: 2511.13366
• Title: Local asymptotic normality for discretely observed McKean-Vlasov diffusions
• Authors: Akram Heidari, Mark Podolskij (University of Luxembourg)
• Classification: math.ST, stat.TH (Statistical Theory)
• Submission Date: November 17, 2025
• Paper Link: https://arxiv.org/abs/2511.13366

Abstract

This paper investigates the local asymptotic normality (LAN) property of the likelihood function for d-dimensional McKean-Vlasov stochastic differential equations observed discretely over a fixed time interval. The model contains joint parameters in both drift and diffusion coefficients, introducing challenges due to dependence on the process distribution. The authors employ Malliavin calculus techniques to derive stochastic expansions of the log-likelihood ratio and establish the LAN property under appropriate conditions. The main technical challenge stems from the implicit nature of transition densities, addressed through integration by parts and Gaussian-type bounds. This work extends existing LAN results for interacting particle systems to the mean-field setting, contributing to statistical inference for nonlinear stochastic models.

Research Background and Motivation

1. Research Problem

This paper addresses parameter estimation for McKean-Vlasov stochastic differential equations (SDEs) with discrete-time observations, establishing local asymptotic normality (LAN) of the likelihood function. The McKean-Vlasov equation takes the form:

$dX^{i,\theta}_t = b_{\theta_1}(X^{i,\theta}_t, \mu^\theta_t)dt + a_{\theta_2}(X^{i,\theta}_t)dW^i_t$

where $\mu^\theta_t$ is the distribution of $X^{i,\theta}_t$, rendering the equation inherently nonlinear.

2. Problem Importance

• Broad Applicability: McKean-Vlasov equations have wide applications in statistical physics, finance, and mean-field games
• Theoretical Foundation: The LAN property, introduced by Le Cam, is a fundamental tool for asymptotic statistical inference and provides lower bounds for the asymptotic variance of estimators
• Mean-Field Theory: Connects statistical inference for microscopic particle systems to macroscopic mean-field limits

3. Limitations of Existing Methods

• Continuous vs. Discrete Observation: Existing LAN results primarily address continuous observation 13, where closed-form likelihood expressions are available via Girsanov's theorem
• Implicit Transition Density: Under discrete observation, transition densities lack explicit expressions, requiring new technical approaches
• Interacting Particle Systems Challenge: For interacting particle systems (3.18), handling joint transition densities in $dN$ dimensions currently lacks adequate bounds in the literature

4. Research Motivation

• Fill the gap in LAN theory for discretely observed McKean-Vlasov equations
• Develop Malliavin calculus techniques for handling implicit transition densities
• Provide theoretical foundations for statistical inference in mean-field models, establishing theoretical connections with recent estimation methods 1

Core Contributions

1. Establishing LAN Property: First establishment of LAN property for discretely observed McKean-Vlasov equations under the asymptotic regime $\Delta_n \to 0, N \to \infty$ with fixed time interval $T$
2. Malliavin Calculus Techniques: Use of integration by parts formulas in Malliavin calculus to derive explicit representations of log-derivatives of transition densities (Proposition 3.1)
3. Stochastic Expansion: Establishment of precise stochastic expansions of log-likelihood ratios (Proposition 3.2), identifying principal and remainder terms
4. Explicit Asymptotic Covariance Matrix: Derivation of explicit asymptotic covariance matrix $\Sigma^{\theta_0}$, featuring functional derivatives $\partial_\mu b_{\theta_1}$ unique to McKean-Vlasov models
5. Distinct Estimation Rates: Proof that drift parameters are estimated at rate $\sqrt{N}$ while diffusion parameters are estimated at rate $\sqrt{N/\Delta_n}$, consistent with contrast estimation methods in 1
6. Technical Innovation: Resolution of the main technical obstacle through Gaussian-type bounds (Proposition 4.2) and moment estimates (Lemma 4.1) for implicit transition densities

Methodology Details

Task Definition

Observed Data: $\{X^{i,\theta}_{t_k}\}_{i=1,\ldots,N}^{k=1,\ldots,n}$ where $t_k = Tk/n$, $\Delta_n = T/n$ is the discretization step

Parameter Perturbation: $\theta^+ = (\theta_1^+, \theta_2^+) = \left(\theta_1^0 + \frac{u}{\sqrt{N}}, \theta_2^0 + \frac{v}{\sqrt{N/\Delta_n}}\right)$

Objective: Prove that the log-likelihood ratio $z(\theta_0, \theta^+) := \log \frac{dP_{\theta^+}}{dP_{\theta_0}}$ satisfies the LAN property, namely $z(\theta_0, \theta^+) \xrightarrow{law} \begin{pmatrix} u \\ v \end{pmatrix}^\top N_{\theta_0} - \frac{1}{2}\begin{pmatrix} u \\ v \end{pmatrix}^\top \Sigma_{\theta_0} \begin{pmatrix} u \\ v \end{pmatrix}$

where $N_{\theta_0} \sim N(0, \Sigma_{\theta_0})$.

Model Architecture

1. McKean-Vlasov Equation Structure

The model assumes $N$ i.i.d. particles, each satisfying: $dX^{i,\theta}_t = b_{\theta_1}(X^{i,\theta}_t, \mu^\theta_t)dt + a_{\theta_2}(X^{i,\theta}_t)dW^i_t$

Key features:

• Distribution Dependence: Drift term depends on marginal distribution $\mu^\theta_t = \text{Law}(X^{i,\theta}_t)$
• Parameter Separation: Drift parameter $\theta_1$ and diffusion parameter $\theta_2$ appear in different coefficients
• Independence: Brownian motions $(W^i)_{1\leq i \leq N}$ of different particles are mutually independent

2. Log-Likelihood Ratio Decomposition

Using the Markov property: $z(\theta_0, \theta^+) = \sum_{k=1}^n \sum_{i=1}^N \log \frac{p^{\theta^+}}{p^{\theta_0}}(t_k, t_{k+1}, X^i_{t_k}, X^i_{t_{k+1}})$

Further decomposed into drift and diffusion components: $z(\theta_0, \theta^+) = \sum_{k=1}^n \sum_{i=1}^N (\zeta^{i,\theta_1}_k + \zeta^{i,\theta_2}_k)$

where $\zeta^{i,\theta_1}_k = \frac{u}{\sqrt{N}} \int_0^1 \frac{\partial_{\theta_1} p^{\theta_1(l), \theta_2^+}}{p^{\theta_1(l), \theta_2^+}}(t_k, t_{k+1}, X^i_{t_k}, X^i_{t_{k+1}})dl$

$\zeta^{i,\theta_2}_k = \frac{v}{\sqrt{N/\Delta_n}} \int_0^1 \frac{\partial_{\theta_2} p^{\theta_1^0, \theta_2(l)}}{p^{\theta_1^0, \theta_2(l)}}(t_k, t_{k+1}, X^i_{t_k}, X^i_{t_{k+1}})dl$

3. Malliavin Calculus Representation (Proposition 3.1)

Key Technique: For parameter derivatives of transition densities, using Malliavin calculus yields:

$\frac{\partial_{\theta_1} p^\theta}{p^\theta}(t_k, t_{k+1}, x, y) = \frac{1}{\Delta_n} E^\theta_{t_k,x}\left[\sum_{r=1}^d \delta(\partial_{\theta_1} X^{i,\theta}_{r,\Delta_n} U^i_r) \Big| X^{i,\theta}_{t_{k+1}} = y\right]$

where:

• $\delta$ is the Skorohod integral (adjoint of Malliavin derivative)
• $U^i_s = a^{-1}_{\theta_2}(X^{i,\theta}_{t_k+s}) Y^{i,\theta}_s (Y^{i,\theta}_{\Delta_n})^{-1}$
• $Y^{i,\theta}_t$ is the process matrix satisfying linear SDE (3.14)

Parameter Derivative Process: $\partial_{\theta_1} X^{i,\theta}_t$ satisfies the SDE: $\partial_{\theta_1} X^{i,\theta}_t = \int_0^t \left(\partial_{\theta_1} b_{\theta_1} + \nabla_x b_{\theta_1} \partial_{\theta_1} X^{i,\theta}_s + \int_{\mathbb{R}^d} \partial_\mu b_{\theta_1}(X^{i,\theta}_{t_k+s}, y, \mu^\theta_{t_k+s}) \partial_{\theta_1}\mu^\theta_{t_k+s}(dy)\right)ds + \ldots$

Note the third term contains functional derivative $\partial_\mu b_{\theta_1}$, unique to McKean-Vlasov models.

Technical Innovation Points

1. Skorohod Integral Stochastic Expansion (Proposition 3.2)

Drift Component: Prove $\delta(\partial_{\theta_1} X^{i,\theta}_{r,\Delta_n} U^i_r) = \Delta_n z^{\theta}_{r,\theta_1}(X^{i,\theta}_{t_k}) [a^{-2}_{\theta_2}(X^{i,\theta}_{t_k})(X^{i,\theta}_{t_{k+1}} - m^\theta_{t_k,t_{k+1}}(X^{i,\theta}_{t_k}))]_r + H^i_{t_{k+1}}$

where $H^i_{t_{k+1}}$ is the remainder term satisfying $(E^\theta_{t_k,x}|H^i_{t_{k+1}}|^\tau)^{1/\tau} = R^i_{t_k}(\Delta_n^2)$.

Key Quantity: $z^\theta_t(x) := \partial_{\theta_1} b_{\theta_1}(x, \mu^\theta_t) + \int_{\mathbb{R}^d} \partial_\mu b_{\theta_1}(x, y, \mu^\theta_t) \partial_{\theta_1}\mu^\theta_t(dy)$

This quantity plays a central role in the asymptotic covariance matrix.

Technical Path:

• Use integration by parts formula (2.7): $\delta(Fu) = F\delta(u) - \langle DF, u\rangle_H$
• Approximate $U^i_r$ by $\hat{U}^i_r = a^{-1}_{\theta_2}(X^{i,\theta}_{t_k+r})$
• Prove each remainder term $H^{i,j}_n, j=1,2,3$ is of order $\Delta_n^2$

Diffusion Component: Similarly prove $\delta(\partial_{\theta_2} X^{i,\theta}_{r,\Delta_n} U^i_r) = [\partial_{\theta_2} a_{\theta_2}(X^i_{t_k}) a^{-1}_{\theta_2}(X^i_{t_k})(X^i_{t_{k+1}} - m^\theta_{t_k,t_{k+1}}(X^i_{t_k}))]_r \times [\ldots] + \text{remainder}$

Remainder satisfies order $\Delta_n^{3/2}$.

2. Distinction from Classical SDE Methods

• Gobet Method 19,20: Original method for ergodic diffusion processes, relying on ergodicity from long-time observation
• This Paper: No ergodicity assumption needed; asymptotics driven by particle number $N \to \infty$
• Functional Derivatives: McKean-Vlasov models feature $\partial_\mu b_{\theta_1}$ terms absent in classical SDEs

3. Application of Transition Density Bounds (Proposition 4.2)

Aronson-type Upper and Lower Bounds: $\frac{1}{L\Delta_n^{d/2}} \exp\left(-c\frac{\|x-y\|^2}{\Delta_n}\right) \exp(-c\Delta_n\|x\|^2) \leq p^\theta(t_k, t_{k+1}, x, y)$$\leq \frac{L}{\Delta_n^{d/2}} \exp\left(-\frac{\|x-y\|^2}{c\Delta_n}\right) \exp(c\Delta_n\|x\|^2)$

Parameter Derivative Bounds: $E^{\bar{\theta}}_{t_k,x}\left[\left|\frac{\partial_{\theta_1} p^\theta}{p^\theta}(t_k, t_{k+1}, x, X^i_{t_{k+1}})\right|^m\right] \leq \frac{L}{\Delta_n^{m/2}} \exp(c\Delta_n\|x\|^2)(1+\|x\|)^q$

These bounds are crucial for proving negligibility of remainder terms (Proposition 4.4).

Experimental Setup

Note: This is a pure theory paper without numerical experiments. Main results are theoretical theorem proofs.

Theoretical Verification Framework

Although lacking numerical experiments, the paper validates theoretical reasonableness through:

1. Consistency with Existing Results: Asymptotic covariance matrix $\Sigma^{\theta_0}$ under condition $N\Delta_n \to 0$ matches asymptotic variance of contrast estimation method in 1
2. Estimation Rates:
   • Drift parameters: $\sqrt{N}$ rate
   • Diffusion parameters: $\sqrt{N/\Delta_n}$ rate
   
   Consistent with classical SDE theory and recent literature 1
3. Special Cases: When $\partial_\mu b_{\theta_1} = 0$ (no distribution dependence), results reduce to LAN for classical diffusion processes

Experimental Results

Main Theoretical Result (Theorem 3.4)

LAN Property: Under assumptions A1-A5, $z(\theta_0, \theta^+) \xrightarrow{P^{\theta_0}-law} \begin{pmatrix} u \\ v \end{pmatrix}^\top N_{\theta_0} - \frac{1}{2}\begin{pmatrix} u \\ v \end{pmatrix}^\top \Sigma_{\theta_0} \begin{pmatrix} u \\ v \end{pmatrix}$

Asymptotic Covariance Matrix: $\Sigma_{\theta_0} = \begin{pmatrix} \Sigma^{\theta_0}_b & 0 \\ 0 & \Sigma^{\theta_0}_a \end{pmatrix}$

where $\Sigma^{\theta_0}_b = \int_0^T \int_{\mathbb{R}^d} z^{\theta_0}_s(x)^\top a^{-2}_{\theta_2^0}(x) z^{\theta_0}_s(x) \mu^{\theta_0}_s(dx)ds$

$\Sigma^{\theta_0}_a = 2\int_0^T \int_{\mathbb{R}^d} \text{tr}(\partial_{\theta_2} a_{\theta_2^0}(x) a^{-1}_{\theta_2^0}(x) \partial_{\theta_2} a_{\theta_2^0}(x) a^{-1}_{\theta_2^0}(x)) \mu^{\theta_0}_s(dx)ds$

Key Findings

1. Diagonal Structure: $\Sigma_{\theta_0}$ is diagonal, indicating asymptotic independence of drift and diffusion parameters
2. Role of Functional Derivatives: $z^{\theta_0}_s(x)$ contains $\partial_\mu b_{\theta_1}$ term, unique to McKean-Vlasov models, reflecting distribution dependence effects
3. Distinction from Interacting Particle Systems:
   • McKean-Vlasov model (1.1): Covariance includes $\partial_\mu b_{\theta_1}$
   • Interacting particle system (3.18): Covariance $z^\theta_t(x)$ simplifies to $\partial_{\theta_1} b_{\theta_1}(x, \mu^\theta_t)$

Proof Strategy Verification

Theorem 3.4 proof verified through six convergence conditions (4.30)-(4.36):

Condition (4.30): First moment of drift component $\sum_{k=1}^n \sum_{i=1}^N E^{\theta_0}_{t_k}[\hat{\zeta}^{i,\theta_1}_k] \xrightarrow{P^{\theta_0}} -\frac{1}{2}u^2 \Sigma^{\theta_0}_b$

Key step: Taylor expansion $m^{\theta_0}_{t_k,t_{k+1}}(X^i_{t_k}) - m^{\theta_1(l),\theta_2^+}_{t_k,t_{k+1}}(X^i_{t_k}) = -\frac{lu\Delta_n}{\sqrt{N}} z^{\theta_0}_{t_k}(X^i_{t_k}) + R^i_{t_k}(\varepsilon_{n,N}\Delta_n/\sqrt{N})$

Condition (4.31): Drift component second moment convergence to $u^2\Sigma^{\theta_0}_b$

Techniques:

• Cross terms $\sum_{i_1 \neq i_2} \hat{\zeta}^{i_1,\theta_1}_k \hat{\zeta}^{i_2,\theta_1}_k$ asymptotically negligible
• Principal terms from $\sum_{i=1}^N (\hat{\zeta}^{i,\theta_1}_k)^2$

Condition (4.32): Fourth moment condition $\sum_{k=1}^n E^{\theta_0}_{t_k}\left[\left|\sum_{i=1}^N \hat{\zeta}^{i,\theta_1}_k\right|^4\right] \xrightarrow{P^{\theta_0}} 0$

Decompose fourth-order terms into different index combinations (all distinct, two pairs, all identical) and prove each converges to 0.

Conditions (4.33)-(4.35): Similar conditions for diffusion component, using conditional variance expansion $V^{\theta_0}_{t_k,t_{k+1}}(x) - V^{\theta_1^0,\theta_2(l)}_{t_k,t_{k+1}}(x) = -\frac{2lv\Delta_n^{3/2}}{\sqrt{N}} \partial_{\theta_2} a_{\theta_2^0}(X^i_{t_k}) a_{\theta_2^0}(X^i_{t_k}) + \ldots$

Condition (4.36): Asymptotic independence of drift and diffusion, proved by showing cross terms asymptotically zero

Related Work

1. Parameter Estimation for McKean-Vlasov Equations

Discrete Observation:

• 1 Amorino et al. (2023): Contrast estimation for interacting particle systems, establishing consistency and asymptotic normality
• 6 Bishwal (2011): Estimation for interactive diffusions
• 9 Chen (2021): Potential maximum likelihood estimation from single trajectory data
• 16,17 Genon-Catalot & Larédo (2021): Small variance and long-time McKean-Vlasov models
• 27 Liu & Qiao (2022): Path-dependent McKean-Vlasov SDEs
• 31 Sharrock et al. (2021): Online parameter estimation

Continuous Observation:

• 13 Della Maestra & Hoffmann (2023): LAN property for McKean-Vlasov models under mean-field regime (directly related to this paper)
  • Distinction: Continuous observation allows Girsanov theorem, closed-form likelihood expressions

2. Nonparametric Methods

• 2 Amorino et al. (2024): Polynomial rates via deconvolution
• 4 Belomestny et al. (2022): Semiparametric estimation for McKean-Vlasov SDEs
• 11 Comte et al. (2024): Nonparametric moment methods
• 12 Della Maestra & Hoffmann (2022): Nonparametric estimation for interacting particle systems
• 29 Nickl et al. (2025): Bayesian nonparametric inference for McKean-Vlasov models

3. Malliavin Calculus Applications in Statistics

• 19,20 Gobet (2001,2002):
  • Local asymptotic mixed normality for elliptic diffusions
  • LAN property for ergodic diffusions with discrete observations
  • Foundation of this paper's method: Using Malliavin calculus to derive transition density derivative representations

4. Advantages of This Paper Over Related Work

1. vs 13 (Continuous Observation):
   • Handles technical challenges of discrete observation
   • Does not rely on Girsanov theorem
2. vs 1 (Contrast Estimation):
   • Provides theoretical foundation for likelihood methods
   • Establishes LAN property, enabling derivation of asymptotic optimality for estimators
3. vs 20 (Classical SDE):
   • Extends to McKean-Vlasov setting
   • No ergodicity assumption required
   • Handles functional derivatives $\partial_\mu b$
4. vs Interacting Particle Systems:
   • Avoids bounds on high-dimensional joint transition densities (Note 3.3 identifies this as main obstacle for interacting particle system LAN)
   • Exploits i.i.d. structure to simplify analysis

Conclusions and Discussion

Main Conclusions

1. Establishment of LAN Property: First establishment of LAN property for discretely observed McKean-Vlasov equations, filling theoretical gap in the field
2. Explicit Form of Asymptotic Covariance Matrix: $\Sigma_{\theta_0} = \text{diag}(\Sigma^{\theta_0}_b, \Sigma^{\theta_0}_a)$ where drift component contains functional derivative $\partial_\mu b_{\theta_1}$, reflecting distribution dependence
3. Confirmation of Estimation Rates:
   • Drift: $\sqrt{N}$
   • Diffusion: $\sqrt{N/\Delta_n}$
   
   Consistent with recent contrast estimation methods 1
4. Technical Contribution: Development of Malliavin calculus techniques for implicit transition densities, combined with Gaussian bounds and integration by parts

Limitations

1. Strong Assumption Conditions:
   • A2: Coefficients bounded and Lipschitz continuous
   • A3: High-order smoothness of coefficients ($C^2$ with polynomially growing derivatives)
   • A5: Uniform ellipticity of diffusion matrix
   
   These conditions may not hold in practical applications
2. One-Dimensional Parameter Restriction: While paper indicates extension to multiparameter case, only $\theta_1, \theta_2 \in \mathbb{R}$ detailed
3. Gap for Interacting Particle Systems:
   • Note 3.3 indicates that for interacting particle systems (3.18), lack of bounds on $dN$-dimensional transition densities prevents LAN establishment
   • This is an important open problem
4. Asymptotic Regime: Requires $\Delta_n \to 0, N \to \infty$ simultaneously, with constraints on relative rate of $N\Delta_n$
5. Initial Distribution: Assumption A1 requires initial distribution $\mu_0$ to be sub-Gaussian, limiting applicability

Future Directions

1. LAN for Interacting Particle Systems: Develop high-dimensional transition density bounds, establish LAN for model (3.18)
2. Relaxing Assumption Conditions:
   • Study non-elliptic diffusion cases
   • Allow unbounded or locally Lipschitz coefficients
3. Multiparameter Extension: Complete treatment of $\theta_1 \in \mathbb{R}^p, \theta_2 \in \mathbb{R}^q$ cases
4. Optimal Estimator Construction: Utilize LAN property to construct asymptotically efficient estimators
5. Hypothesis Testing: Develop hypothesis testing theory for McKean-Vlasov models based on LAN property
6. Non-Ergodic Cases: Extend to non-ergodic McKean-Vlasov processes
7. High-Frequency Data: Study asymptotics when $\Delta_n \to 0$ at faster rates

In-Depth Evaluation

Strengths

1. Theoretical Rigor:
   • Complete and detailed proofs (Section 4 comprises half the paper)
   • Clear argumentation for each technical step
   • Appropriate use of modern stochastic analysis tools (Malliavin calculus)
2. Methodological Innovation:
   • Clever Malliavin Calculus Application: Technique of expanding Skorohod integral into principal plus remainder terms (Proposition 3.2) is core innovation
   • Functional Derivative Handling: Correct identification and treatment of $\partial_\mu b_{\theta_1}$ term, unique to McKean-Vlasov models
   • Remainder Control: Proposition 4.4 uniformly handles negligibility of various remainder terms
3. Theoretical Contribution:
   • Fills gap in LAN theory for discretely observed McKean-Vlasov equations
   • Connects likelihood methods with contrast estimation methods (relation to 1)
   • Provides theoretical foundation for asymptotic statistical inference in McKean-Vlasov models
4. Writing Clarity:
   • Clear structure: assumptions → main results → proofs
   • Comprehensive notation system (Section 2.1)
   • Adequate explanation of key difficulties and solutions (Introduction and Note 3.3)
5. Comprehensive Literature Review: Accurately positions paper within McKean-Vlasov statistical inference literature

Weaknesses

1. Limited Practical Applicability:
   • Strong assumption conditions may not hold in real data
   • No numerical simulations to verify theoretical results
   • No discussion of verifying assumptions in practice
2. Readability of Technical Details:
   • Section 4 proofs highly technical, unfriendly to non-specialists
   • Some key inequalities (e.g., Proposition 4.2 proof) reference 20 without detailed explanation of adaptation to McKean-Vlasov setting
3. Result Limitations:
   • LAN for interacting particle systems (3.18) remains open (Note 3.3)
   • Only treats fixed time interval $T$; no discussion of $T \to \infty$
4. Multiparameter Case Treatment:
   • Claims extensibility to multiparameter case but provides only framework
   • Technical details for multiparameter case (especially non-diagonal elements of $\Sigma_{\theta_0}$) not fully developed
5. Disconnect from Applications:
   • No concrete application examples
   • No discussion of using results in finance, neuroscience, etc.

Impact

1. Contribution to Field:
   • Theoretical Foundation: Provides solid theoretical basis for statistical inference in McKean-Vlasov models
   • Methodology: Systematic application of Malliavin calculus to McKean-Vlasov statistics
   • Open Problems: Clearly identifies technical obstacles for interacting particle system LAN (high-dimensional transition density bounds), directing future research
2. Practical Value:
   • Estimator Evaluation: Can assess asymptotic efficiency of existing estimators (e.g., contrast estimation in 1)
   • Lower Bounds: LAN property provides Cramér-Rao-type lower bounds for estimator asymptotic variance
   • Optimal Estimation: Guides construction of asymptotically efficient estimators
3. Reproducibility:
   • ✅ Theoretical results fully verifiable (complete proofs)
   • ❌ No code or numerical experiments
   • ✅ Assumptions clearly specified
   • ⚠️ Some technical details require consulting references 19,20,30
4. Expected Citation Impact:
   • Short-term: Specialists in mean-field statistical inference will cite
   • Medium-term: Likely becomes standard reference for McKean-Vlasov statistical inference
   • Long-term: Impact expands further if interacting particle system problem resolved

Applicable Scenarios

1. Theoretical Research:
   • Statistical theory for McKean-Vlasov models
   • Parameter estimation in mean-field games
   • Asymptotic statistics for nonlinear SDEs
2. Potential Application Fields:
   • Finance: Systemic risk models 18, option pricing 21
   • Neuroscience: Neural network models 3
   • Statistical Physics: Mean-field limits of particle systems
   • Social Dynamics: Opinion dynamics models 8
3. Method Applicability:
   • ✅ Large samples ($N$ large)
   • ✅ High-frequency observation ($\Delta_n$ small)
   • ✅ Fixed time interval
   • ✅ Smooth coefficients
   • ❌ Small samples or low-frequency observation
   • ❌ Non-elliptic diffusions
4. Comparison with Other Methods:
   • vs Contrast Estimation 1: LAN provides theoretical optimality; contrast estimation more computationally tractable
   • vs Bayesian Methods 29: LAN is frequentist; Bayesian methods more flexible but computationally intensive
   • vs Nonparametric Methods 12: LAN for parametric models; nonparametric methods suit model uncertainty

Key References

1. 1 Amorino et al. (2023): Contrast estimation for interacting particle systems, direct comparison object for this paper
2. 13 Della Maestra & Hoffmann (2023): LAN for continuous observation McKean-Vlasov, direct predecessor
3. 19,20 Gobet (2001,2002): Original source of Malliavin calculus methods
4. 30 Nualart (1995): Standard reference for Malliavin calculus
5. 22,25,26 Le Cam series: Foundational LAN theory literature

Summary

This paper makes important theoretical contributions to statistical inference for McKean-Vlasov stochastic differential equations. Through clever application of Malliavin calculus, the authors successfully establish local asymptotic normality for discrete observation, filling a theoretical gap in the field. The paper demonstrates high technical level, rigorous proofs, and provides solid theoretical foundation for asymptotic statistical inference in mean-field models.

Main values: (1) Theoretical completeness: Systematic establishment of LAN theory for McKean-Vlasov models; (2) Methodological innovation: Development of techniques for implicit transition densities; (3) Theoretical guidance: Provides benchmarks for asymptotic optimality of estimators.

Main limitations: (1) Strong assumption conditions; (2) Lack of numerical verification; (3) LAN for interacting particle systems remains open.

For researchers working on McKean-Vlasov model statistical inference, this is essential reading. For applied researchers, practical applicability requires verifying whether assumption conditions hold for specific problems.