Graph Ornstein–Uhlenbeck Process (GrOU)

• GrOU process is a multivariate stochastic model that integrates node-level mean reversion with network-mediated interactions using graph structures.
• It employs rigorous likelihood theory and closed-form maximum-likelihood estimators to infer parameters from continuous and discretized data.
• Applications span high-frequency finance, phylogenetics, and networked data, demonstrating improved forecasting accuracy and computational efficiency.

The Graph Ornstein–Uhlenbeck (GrOU) process is a class of multivariate stochastic processes designed to model the temporal evolution of vector-valued signals under the influence of network (graph) structures. Incorporating effects of both node-specific dynamics (“momentum” or “mean-reversion”) and restricted, network-mediated interactions between nodes (“network effects”), GrOU processes generalize the classical Ornstein–Uhlenbeck framework to accommodate dependence patterns imposed by arbitrary graphs and allow for general Lévy (jump) noise. The GrOU paradigm is supported by rigorous likelihood theory, explicit maximum-likelihood estimators, and sparse penalized estimation, with extensions for stochastic volatility, adaptation to edge-indexed time series, and applications across phylogenetics, networked temporal data, and high-frequency finance (Courgeau et al., 2020, Courgeau et al., 2020, Chen et al., 15 May 2026, Bartoszek et al., 2016).

1. Stochastic Differential Equation Formulation and Parametrizations

Let $d$ denote the number of nodes (or components), and $A\in\{0,1\}^{d\times d}$ a (possibly normalized) adjacency matrix describing the static network structure, with $a_{ii}=0$. The canonical form is a $d$-dimensional process $Y_t$, evolving according to

$dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$

where $Q$ is a drift/interactions matrix tied to graph $A$ and $L_t$ a general $d$-dimensional Lévy process (characteristic triplet $A\in\{0,1\}^{d\times d}$0).

Two principal parametrizations are established (Courgeau et al., 2020, Courgeau et al., 2020):

• “$A\in\{0,1\}^{d\times d}$1-GrOU” (two-parameter): $A\in\{0,1\}^{d\times d}$2, with $A\in\{0,1\}^{d\times d}$3, and $A\in\{0,1\}^{d\times d}$4 denotes the row-normalized adjacency matrix ($A\in\{0,1\}^{d\times d}$5, $A\in\{0,1\}^{d\times d}$6). Here, $A\in\{0,1\}^{d\times d}$7 controls mean-reverting “momentum”; $A\in\{0,1\}^{d\times d}$8 embodies network propagation.
• “$A\in\{0,1\}^{d\times d}$9-GrOU” (node-wise full interaction): $a_{ii}=0$0, $a_{ii}=0$1, so each node $a_{ii}=0$2 has its own self-dynamics and neighbor weights.

Well-posedness and stationarity require $a_{ii}=0$3 to have all eigenvalues with positive real part and integrability of $a_{ii}=0$4 via $a_{ii}=0$5. In practice, for $a_{ii}=0$6-GrOU, stationarity holds when $a_{ii}=0$7 and, for $a_{ii}=0$8-GrOU, when $a_{ii}=0$9 is diagonally dominant (Courgeau et al., 2020).

In biological/network science applications, the drift is recast as $d$0, admitting arbitrary migration/interactions $d$1 among $d$2 lineages/nodes (Bartoszek et al., 2016).

2. Likelihood Functions and Inference Methods

For a continuously observed GrOU process, the (log-)likelihood for parameters $d$3 can be written as

$d$4

where $d$5 is a vector of martingale integrals over $d$6 (e.g., $d$7 in the $d$8-model), $d$9 its covariance (“information”) matrix (Courgeau et al., 2020).

The closed-form maximum-likelihood estimator (MLE) is

$Y_t$0

given $Y_t$1 is invertible, which is almost surely the case for large $Y_t$2 under ergodicity.

Under high-frequency, discrete-time non-uniform sampling (mesh size $Y_t$3), the MLE is constructed from discretized increments (with explicit jump filtering for Lévy processes). The estimator retains closed-form via sufficient statistics: $Y_t$4 with $Y_t$5, $Y_t$6, and $Y_t$7 as empirical sums involving the state vectors and filtered increments. Mapping to $Y_t$8 is linear (Courgeau et al., 2020).

In edge-indexed GrOU, the likelihood involves a quadratic form in the martingale increments of edge processes and their network-lagged values, with closed-form MLE in both continuous and discretized settings (Chen et al., 15 May 2026).

3. Asymptotic Theory for Parameter Estimators

Key asymptotic properties, established via martingale central limit theorems and local asymptotic normality (LAN), are as follows:

• Consistency: $Y_t$9 in probability as $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$0, under ergodicity and regularity.
• Central Limit Theorem: For both stationary continuous and discretized observation,

$dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$1

analogously for discrete estimators as $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$2,

$dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$3

for both finite and infinite-activity jump regimes (Blumenthal–Getoor index $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$4) (Courgeau et al., 2020, Courgeau et al., 2020, Chen et al., 15 May 2026).

In the adaptive Lasso regime, oracle properties hold: model selection consistency and asymptotic normality on the true support. Stable convergence results enable joint inference on both $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$5 and the network topology (Courgeau et al., 2020, Courgeau et al., 2020).

4. Graph Recovery and Sparse Estimation

When the underlying adjacency $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$6 or interaction structure is unknown, penalized likelihood methods are employed to infer both $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$7 and the graph topology. The Adaptive Lasso objective penalizes off-diagonal entries with data-driven weights: $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$8 with adaptive weights $dY_t = -Q\,Y_{t-}\,dt + dL_t, \qquad t\ge0,$9 based on an initial estimator, and tuning $Q$0 controlling sparsity. Under suitable scaling ($Q$1 and $Q$2), recovery is consistent in both support and nonzero parameter estimation (Courgeau et al., 2020, Courgeau et al., 2020). The approach is robust under both finite and infinite jump activity, and thresholding the increments (with exponent $Q$3) optimally trades off noise suppression and jump retention.

5. Extensions: Edge-Indexed Models, Stochastic Volatility, and Applications

Recent work generalizes GrOU processes to edge-indexed time series: each process is indexed by an edge rather than a node, capturing dynamics of edge-level interactions in networks. The SDE is specified in a block-companion form with network-lagged contributions and Lévy noise, supporting multi-lag dependencies and neighborhood-weighted propagation. Edge-GrOU models admit explicit transition laws, closed-form likelihoods, and rigorous asymptotic theory in both continuous and high-frequency discrete time (Chen et al., 15 May 2026).

Stochastic volatility extension: The driving noise can be modulated by a positive-semidefinite OU process and random time-change, accommodating time-varying volatility and jump intensity while preserving stationarity and the likelihood structure conditionally (Courgeau et al., 2020).

Notable applications:

• Wind capacity factor measurements across spatially distributed networks, demonstrating improved bias and variance properties over least-squares AR(1) estimators, especially in dense networks and under high noise (Courgeau et al., 2020).
• High-frequency financial data, specifically pairwise covariances from limit order books, where edge-GrOU offers superior forecasting accuracy and computational efficiency, with model selection via BIC and empirical screening over candidate networks (Chen et al., 15 May 2026).
• Evolutionary biology (phylogenetic trait models) incorporating migration/hybridization and ecological interactions, with explicit moment formulas and proper handling of non-tree dependency structures (Bartoszek et al., 2016).

6. Stationarity, Mixing, and Theoretical Properties

For the SDE to be well-posed and the process to be strictly stationary and ergodic, $Q$4 (or $Q$5 in the edge case) must be Hurwitz (i.e., $Q$6 or $Q$7). Lévy processes must have logarithmic moment integrability of jumps: $Q$8. Under these conditions, GrOU admits a unique invariant law and mixing properties, facilitating asymptotic likelihood theory (Courgeau et al., 2020, Chen et al., 15 May 2026).

7. Empirical and Simulation Evidence

Empirical studies on real and synthetic networks confirm:

• Finite-sample consistency and approximate normality of estimators.
• Superiority of GrOU MLE over standard least-squares estimators, especially under informative noise and network effects.
• Estimation variability depends on graph density; denser graphs yield higher variance for network effects.
• Edge-GrOU models achieve computational efficiency and forecasting accuracy, with robustness to model selection and network inference (Courgeau et al., 2020, Chen et al., 15 May 2026).

Simulation and real-world applications underscore the versatility and effectiveness of the GrOU framework for dynamic network-structured data, supporting both fundamental theory and a broad range of implementations.