An Accelerated Online PCA with O(1) Complexity for Learning Molecular Dynamics Data

Alakkari, SalaheddinDingliana, JohnJan Byska and Michael Krone and Björn Sommer2018-06-022018-06-022018978-3-03868-061-1https://doi.org/10.2312/molva.20181100https://diglib.eg.org:443/handle/10.2312/molva20181100In this paper, we discuss the problem of decomposing complex and large Molecular Dynamics trajectory data into simple low-resolution representation using Principal Component Analysis (PCA). Since applying standard PCA for such large data is expensive in terms of space and time complexity, we propose a novel online PCA algorithm with O(1) complexity per new timestep. Our approach is able to approximate the full dimensional eigenspace per new time-step of MD simulation. Experimental results indicate that our technique provides an effective approximation to the original eigenspace computed using standard PCA in batch mode.An Accelerated Online PCA with O(1) Complexity for Learning Molecular Dynamics Data10.2312/molva.201811001-8