Core Concepts
This paper presents an improved predictive online primal-dual method (POPD2) for efficiently solving dynamic inverse problems in applications such as image stabilization, fluid flow monitoring, and dynamic medical imaging. The method provides a more concise analysis, relaxes restrictive conditions on the dual predictor, and develops several improved dual predictors.
Abstract
The paper addresses the challenge of processing information that evolves over time in real-world applications, such as computational image stabilization, fluid flow monitoring, and dynamic medical imaging. When the monitoring period is long and results are needed immediately, online reconstruction techniques are required.
The authors present the Predictive Online Primal-Dual Proximal Splitting (POPD2) method, which extends previous work on predictive online primal-dual methods in two ways:
Provides a more concise analysis that symmetrizes previously unsymmetric regret bounds and relaxes restrictive conditions on the dual predictor.
Develops several improved dual predictors based on the relaxed conditions.
The key aspects of the paper are:
Formulation of the online optimization problem as minimizing the sum of time-varying convex functions.
Presentation of the POPD2 algorithm, which incorporates a forward step with respect to the smooth term and simplifies the dual prediction.
Derivation of a symmetric dynamic regret bound for the POPD2 algorithm, which improves upon the previous unsymmetric regret bounds.
Analysis of a broad class of "pseudo-affine" primal-dual predictors and examples in the context of optical flow and dynamic PET reconstruction.
Numerical demonstrations of the efficacy of the proposed method and predictors in image stabilization and dynamic PET reconstruction.
The paper provides a rigorous theoretical analysis and practical solutions for efficiently processing dynamic data in various applications.