Multidimensional Integral Fractional Ornstein--Uhlenbeck Process with an Application to Animal Movement

Technical Summary: Multidimensional Integral Fractional Ornstein–Uhlenbeck Process with an Application to Animal Movement

Problem Statement
Animal telemetry data presents a statistical challenge: locations are recorded at discrete, often irregular times, while movement evolves continuously. While continuous-time velocity models (such as the Ornstein–Uhlenbeck process) address this by treating position as the integral of velocity, classical models assume exponential decay of velocity covariance, implying only short-range temporal dependence. This limitation is problematic for movement patterns influenced by memory or long-range dependence. Previous work by Ramirez-Gonzalez and Sun (2026) introduced the integral fractional Ornstein–Uhlenbeck (ifOU) process to capture long-range dependence via a Hurst parameter ($H$). However, these analyses treated spatial coordinates (e.g., Longitude and Latitude) independently, failing to model the dependence between coordinates. Furthermore, extending ifOU to a multidimensional setting is non-trivial; arbitrary pairwise correlations between coordinates with different Hurst parameters can lead to invalid covariance structures.

Methodology
The authors propose the multidimensional integral fractional Ornstein–Uhlenbeck (mifOU) process. The model defines the position vector $\mu_H(t) \in \mathbb{R}^p$ as the temporal integral of a velocity vector $v_H(t)$, where each coordinate's velocity follows a fractional Ornstein–Uhlenbeck process driven by a multivariate fractional Brownian motion (mfBm).

Key methodological components include:

1. Covariance Construction: The model utilizes a "well-balanced" specification for the driving mfBm, ensuring cross-covariances are symmetric in their time arguments. The position cross-covariance is derived via an integral transformation of the driving noise covariance.
2. Admissible Correlation Region: The authors establish that for the joint covariance to be valid (positive semidefinite), the correlation parameters must satisfy specific finite-dimensional constraints dependent on the Hurst parameters ($H_i$) of each coordinate. They characterize this admissible region using a matrix condition involving Gamma functions and sine terms.
3. Parametrization: To facilitate likelihood optimization without violating positive-definiteness constraints, the authors introduce a bijective mapping (an LKJ-type partial-correlation parametrization) that transforms unconstrained variables into the admissible correlation region.
4. Inference and Simulation: The finite-dimensional distribution of the position process is Gaussian. This allows for:
   • Maximum likelihood estimation (MLE) of the complete parameter vector (damping $\beta$, scale $\sigma$, Hurst $H$, and correlations $\rho$).
   • Finite-dimensional simulation via a recursive algorithm.
   • Conditional velocity reconstruction from observed positions using the conditional Gaussian distribution.
5. Dependence Properties: Theoretical analysis shows the process is neither stationary nor intrinsically stationary. Long-range dependence (LRD) for separated increments occurs when the sum of the Hurst parameters of a coordinate pair ($H_i + H_j$) differs from 1. Specifically, if $H_i + H_j > 1$, the pair exhibits long memory; if $H_i + H_j \leq 1$, it exhibits short memory.

Results
The paper presents two main sets of results: numerical experiments and an application to animal telemetry.

• Numerical Experiments:

  • Joint estimation of all parameters (scale, damping, Hurst, and cross-correlations) from a single simulated 3D trajectory yielded accurate estimates with 99% confidence intervals containing the true values.
  • Conditional velocity reconstruction successfully recovered latent velocity trajectories from noisy position data.
  • Correlation estimation studies showed that while moderate correlations are estimated with low bias, weak correlations remain less precise. Under fixed-horizon grid refinement, the root mean squared error (RMSE) of correlation estimates decreased as the observation grid became finer.

• Application to Bat Migration:
  The model was applied to telemetry data from five common noctule bats migrating in Germany (three of which had altitude measurements).

  • Marginal Analysis: The best-fitting marginal models varied by individual and coordinate. For example, Bat 3's altitude was best fit by fractional Brownian motion (fBM), while Bat 4's altitude was best fit by a non-stationary $\zeta$ process.
  • Dependence Analysis:
    • Bats 1 & 2: Independent models (products of univariate fits) were preferred over correlated models for the horizontal coordinates.
    • Bats 3 & 4: Strong evidence supported a correlated ifOU model for the Longitude–Latitude pair, with estimated correlations of $\approx 0.64$ and $\approx 0.58$, respectively. The full 3D models remained competitive for Bat 3, suggesting potential dependence involving altitude.
    • Bat 5: The independent model was preferred, though a correlated model remained competitive ($\Delta$AIC $\approx 2$).
  • Sensitivity: Model rankings were generally stable across different time resolutions (30s, 1min, 2min), though specific marginal family choices (e.g., Confluent vs. fBM) showed sensitivity to the penalty criterion (AIC vs. BIC).

Key Contributions

1. Theoretical Extension: The paper extends the univariate ifOU process to a multidimensional setting, providing a rigorous framework for modeling dependent animal trajectories with long-range dependence.
2. Validity and Parametrization: It establishes the conditions for the validity of the joint covariance kernel and provides a practical reparametrization that ensures valid covariance matrices during likelihood optimization, eliminating the need for post-hoc positive-definiteness checks.
3. Computational Framework: The authors develop algorithms for simulation, likelihood inference, and conditional velocity reconstruction, making the model applicable to real-world, sparsely sampled telemetry data.
4. Empirical Insight: The application demonstrates that marginal temporal structures and cross-coordinate dependencies should be assessed separately. The study reveals that while some bats exhibit strong coordinated movement (correlated ifOU), others move independently, and that altitude dependence is generally weaker than horizontal dependence.

Significance and Claims
The authors claim that the mifOU process provides a flexible, statistically valid tool for analyzing animal movement that captures both long-range temporal dependence and inter-coordinate correlations. They emphasize that their approach allows for the joint estimation of marginal dynamics and dependence structures without imposing uniform assumptions across coordinates. The paper concludes that while the model successfully identifies strong dependence in specific cases (Bats 3 and 4), the evidence for dependence is not universal, and the choice of marginal covariance families varies significantly across individuals. The work supports the utility of fractional processes in animal telemetry but cautions against assuming stationarity or uniform memory regimes across all spatial dimensions. The authors modestly note that their numerical experiments describe finite-sample performance and do not establish asymptotic properties of the joint estimator.