Abstract
Mapping the effective connectivity of the human brain from functional MRI (fMRI) data is a significant challenge due to the high dimensionality of neural networks and the limited availability of time samples. This paper investigates the performance of different estimators for the multivariate Ornstein-Uhlenbeck (mOU) model. We demonstrate that while standard Bayesian estimation with a uniform prior is mathematically equivalent to the method of moments, it suffers from numerical instability in high-dimensional settings. In contrast, an iterative Lyapunov optimization approach proves to be significantly more robust and data-efficient, achieving high accuracy with up to eight times fewer samples than Bayesian methods.
Introduction
Understanding the "wiring diagram" or effective connectivity of the brain is crucial for diagnosing and treating neuropsychiatric disorders. Researchers typically model these dynamics using the multivariate Ornstein-Uhlenbeck (mOU) process. However, estimating the parameters of this model—specifically the connectivity matrix $C$ and the noise covariance $\Sigma$—is hindered by the "curse of dimensionality." With over 100 brain regions and often fewer than a few hundred fMRI time points, traditional estimators frequently fail to provide reliable results.
Methodology
The study compares three primary estimation procedures:
- Moments Method: An analytical approach matching theoretical and empirical covariances.
- Bayesian Approach: Derived using a uniform prior. We show that the posterior mean in this case is identical to the Moments estimator.
- Lyapunov Optimization: An iterative gradient-descent method that fits the model's predicted covariance to the observed data without requiring a matrix logarithm.
A critical finding is that the Bayesian and Moments methods rely on a matrix logarithm to recover connectivity. In high dimensions, small errors in covariance estimation lead to large numerical instabilities in the logarithm, often introducing significant imaginary components that degrade the real-valued connectivity estimate.
Results and Discussion
Our evaluations using synthetic random networks and brain-like topologies reveal a stark contrast in performance.
Scaling and Efficiency
As the number of nodes increases, the accuracy of Bayesian/Moments estimation declines sharply. Conversely, the Lyapunov method maintains high reliability across larger networks. In terms of sample efficiency, the Lyapunov approach requires approximately eight times fewer time samples to achieve the same accuracy as the Bayesian method for a standard 100-node network.
Robustness in Brain-like Networks
When applied to synthetic networks modeled after real human brain topology, the Lyapunov method consistently outperforms its counterparts. The iterative nature of Lyapunov optimization acts as an implicit regularizer, avoiding the "matrix log trap" that plagues analytical methods.
Conclusion
While Bayesian methods are often preferred for their ability to incorporate prior knowledge, this study shows that a standard Bayesian framework with a flat prior offers no advantage over simpler analytical methods and is numerically inferior to iterative optimization in high-dimensional brain modeling. Future work will explore whether more sophisticated, non-uniform priors—such as those derived from structural DTI scans—can bridge this performance gap. For current large-scale fMRI analysis, Lyapunov optimization remains the most robust and efficient choice.
Summary updated