Action functional as an early warning indicator in the space of probability measures via Schrödinger bridge

Peng Zhang, Ting Gao, Jin Guo, Jinqiao DuanQuant. Biol 2025 · OpenReview

TL;DR

This paper introduces a novel method for detecting critical transitions in complex systems by using an "action functional" within the framework of Schrödinger bridges. This approach allows for the identification of early warning signals in the space of probability measures, offering a new perspective on predicting abrupt changes, particularly in the context of neurodegenerative diseases like Alzheimer's.

What Problem Does This Paper Address?

Many complex systems, from ecosystems to the human brain, exhibit critical transitions or "tipping points" where small changes can lead to drastic, often irreversible, shifts. Identifying these tipping points before they occur is crucial for intervention and understanding system dynamics. Traditional methods for detecting these early warning signals (EWS) often rely on statistical analysis of time series data. However, these methods can be limited, especially when dealing with high-dimensional, complex systems like brain activity. This paper aims to bridge this gap by developing a method that operates directly in the space of probability measures, leveraging the power of Schrödinger bridges to model the transition pathways between different system states.

Methods

The core of this work lies in the application of Schrödinger bridges to model the most probable transition pathways between two probability distributions (representing different system states). A Schrödinger bridge problem essentially finds the optimal path that connects two given probability measures, regularized by an entropy term.

The authors introduce a novel action functional as an early warning indicator. This functional is defined within the probability measure space, specifically using the Wasserstein distance in the Benamou-Brenier representation. This allows them to quantify the "effort" or "cost" associated with transitioning between states, and changes in this functional can signal an impending critical transition.

They also utilize forward-backward stochastic differential equations (FBSDEs) as a computational framework to solve the Schrödinger bridge problem and estimate the action functional. This approach allows for the analysis of high-dimensional data and complex dynamics.

Datasets

The paper uses two main types of data:

  1. Morris-Lecar Model: A well-established two-dimensional biophysical model of neuronal action potential generation. This serves as a controlled environment to test the method's ability to detect transitions between different dynamical regimes (e.g., from a stable equilibrium to a limit cycle).
  2. Alzheimer's Disease Neuroimaging Initiative (ADNI) Data: This dataset consists of structural MRI images from healthy individuals and patients with mild Alzheimer's disease. The authors use this real-world data to demonstrate the method's potential in identifying early-stage transitions from a healthy brain state to a pre-AD state. For this dataset, they employ Variational Autoencoders (VAE) for dimensionality reduction and t-distributed Stochastic Neighbor Embedding (t-SNE) for visualization.

Key Insights and Ideas

  • Action Functional as an EWS: The paper proposes that the action functional, when viewed as a measure of transition cost in probability space, can serve as a powerful early warning indicator for critical transitions. Changes in this functional can signal an impending shift in system dynamics.
  • Schrödinger Bridges for Transition Pathways: By framing the problem as a Schrödinger bridge, the authors can explicitly model and analyze the most probable pathways between different system states (represented by probability distributions). This offers a more nuanced understanding than traditional time-series analysis.
  • Bridging Theory and Application: The work successfully connects theoretical concepts from stochastic processes and optimal transport (Schrödinger bridges, action functionals) to practical applications in neuroscience and disease detection.
  • Detecting Subtle Changes: The method's ability to operate in probability space allows it to potentially detect subtle shifts that might be masked in raw time-series data, especially in high-dimensional systems.

Key Results

  • Morris-Lecar Model Validation: The method successfully identified transition dynamics in the Morris-Lecar model, showing how increasing noise or changing parameters affects the transition pathways. The evolutionary densities of these paths were visualized, demonstrating the convergence of their numerical solutions.
  • Alzheimer's Data Analysis: When applied to ADNI data, the method identified a critical transition point around the 25th step in the evolutionary density from healthy to early Alzheimer's states. This aligns with the concept of tipping points in disease progression and suggests the potential for early detection. The action functional indicator showed a significant disorder around this 25th step.

Limitations and Future Work

  • Geometric Representation: The authors acknowledge that the geometric representation of data (especially in low-dimensional manifolds) can significantly impact the algorithm's convergence. Developing better algorithms that account for the geometric properties of image distributions is a key area for future work.
  • Computational Cost: While the method is computationally efficient compared to some alternatives, dealing with high-dimensional data and complex dynamics can still be computationally intensive.
  • Generalizability: Further validation on diverse datasets and different types of complex systems would be beneficial to establish the broad applicability of this approach.
  • Intervention Strategies: While the method can detect tipping points, translating this detection into concrete intervention strategies requires further research.

Summary updated

BibTeX
@inproceedings{zhang2025action,
  author = {Peng Zhang and Ting Gao and Jin Guo and Jinqiao Duan},
  title = {{Action functional as an early warning indicator in the space of probability measures via Schrödinger bridge}},
  booktitle = {Quant. Biol},
  year = {2025},
  url = {https://openreview.net/forum?id=ZmNnIDt3lZ},
}
Figure