From convergence theory to implementation
Contributed to an open-source Julia implementation of the Combettes–Eckstein asynchronous, block-iterative primal-dual projective splitting algorithm. The implementation supports custom proximal operators, matrix or functional linear operators, and configurable block activation and delay bounds.
Empirical study
Studied sparse classifier training and stereoscopic image recovery, comparing synchronization, processor counts, and hyperparameter schedules. The detailed research page provides the experimental setup, timing results, and scope of the conclusions.
Research output and results
A note on asynchronous Projective Splitting in Julia is available on Optimization Online (2025; updated May 2026). The public PDF reports three-worker runtime decreasing from 923.442 to 701.419 seconds (24.0%) under simulated proximal-computation latency. The study also investigates worker overhead, delay bounds, and decreasing proximal-parameter schedules.