Snapshot data-driven covariance steering for linear dynamical systems

Yosuke Inoue, Mika Inoue, Kenji Kashima, Yasuhiro OnogiNonlinear 2025 · DOI

Abstract

This paper addresses a covariance steering problem for linear dynamical systems where the system model is unknown and only sparse "snapshot" data points are available. The proposed framework decomposes the challenge into two parts: data-driven modeling using a Schrödinger bridge formulation and a model-based covariance steering problem solved via a novel gradient-based algorithm. Numerical simulations demonstrate the effectiveness of this approach in steering state distributions toward desired targets, offering a promising path for personalized medical interventions in complex biological systems.

Introduction: Health as a Statistical Distribution

In complex biological systems, such as Gene Regulatory Networks (GRNs), the "state" of a system is often better characterized by a distribution across a population of cells rather than a single point. Diseases like obesity or cancer frequently manifest as a shift in these statistical distributions. For instance, single-cell RNA sequencing (scRNA-seq) data reveals distinct clustering patterns between healthy and diseased states.

UMAP visualization of gene expression distributions in normal vs. obese mice, highlighting the shift in cell populations.

Traditional control theory typically requires a precise mathematical model or dense time-series data. However, in clinical settings, we often face two major hurdles:

  1. Unknown Dynamics: The specific interactions within a patient's GRN are not fully known.
  2. Sparse Snapshot Data: We cannot track the same cell over time; instead, we obtain "snapshots"—measurements of different cells at discrete time points.

This paper introduces a Data-driven Covariance Steering (D-CS) framework designed to overcome these limitations by steering a "diseased" distribution back to a "healthy" target using limited snapshot data.

Methodology: The D-CS Framework

The proposed approach operates in a closed-loop fashion, alternating between system identification and intervention decision-making.

The proposed control logic framework, illustrating the cycle of data collection, system identification, and intervention decision across multiple stages.

1. High-accuracy System IDentification (HiSID)

To estimate the unknown system matrix ($A$), the authors employ a Schrödinger Bridge formulation. This method identifies the most probable stochastic process connecting two observed snapshot distributions. By iteratively solving the Schrödinger Bridge problem and a subsequent identification problem, the HiSID algorithm refines its estimate of the system's internal dynamics, effectively "filling in the gaps" between sparse data points.

2. Gradient-based Covariance Steering (CS)

With an estimated model in hand, the goal shifts to finding an optimal control input ($U$)—representing a therapeutic intervention like CRISPR or mRNA delivery—to move the current state distribution toward the target.

The objective is to minimize the Kullback–Leibler (KL) divergence between the system's state and the reference distribution. A key contribution of this work is a computationally efficient gradient-based optimization. The authors derived a method to compute the gradient of the cost function using the solutions to two Lyapunov equations, avoiding the high computational costs associated with "black-box" optimization methods.

Results and Discussion

The D-CS algorithm was evaluated using a synthetic 5-dimensional linear dynamical system designed to mimic a GRN. In simulations where the system matrix $A$ was entirely unknown to the controller, the results were significant:

  • Statistical Convergence: The cost function ($J_P$), measuring the KL divergence from the healthy state, was reduced from 1.47 (uncontrolled) to 0.108 (controlled).
  • Distributional Control: Visualization through Principal Component Analysis (PCA) confirmed that the controlled system successfully compressed the state distribution to fit within the target 99% confidence ellipse, whereas the uncontrolled system drifted significantly.
  • Model Accuracy: The HiSID algorithm demonstrated high precision in estimating the elements of the unknown matrix $A$, providing a robust foundation for the steering algorithm.

Conclusion and Future Directions

This research provides a theoretical and algorithmic bridge between sparse biological data and precise medical intervention. By treating health as a statistical property and utilizing optimal transport theory (Schrödinger Bridges), the D-CS framework offers a path toward personalized, data-driven therapies.

Future work will focus on extending this framework to nonlinear dynamics, which more accurately reflect complex biological interactions. Additionally, the authors aim to address the combinatorial challenge of intervention site selection—identifying which specific genes to target among thousands—using advanced subgradient-based methods.

Summary updated

Abstract

This paper addresses a covariance steering problem for linear dynamical systems. A key feature of the problem is that the system model is unknown, and only a few snapshot data points on state variations are available. To address this challenge, the problem is decomposed into two parts: data-driven modeling based on snapshot data and a covariance steering problem toward the desired distribution using the estimated model. For the data-driven modeling problem, we propose a solution method based on the Schrödinger bridge formulation, as studied in prior work. For the model-based covariance steering problem, we develop a gradient-based algorithm in which the gradient is computed using two solutions of Lyapunov equations. Finally, we present a numerical simulation to demonstrate the effectiveness of the proposed approach.
BibTeX
@inproceedings{inoue2025snapshot,
  author = {Yosuke Inoue and Mika Inoue and Kenji Kashima and Yasuhiro Onogi},
  title = {{Snapshot data-driven covariance steering for linear dynamical systems}},
  booktitle = {Nonlinear},
  year = {2025},
  doi = {10.1587/nolta.17.138},
}
Figure