Principal Component Projections Achieve Frequency Decomposition on Resting-State FMRI Data

Yi-Ou Li¹, Pratik Mukherjee¹

In this work, we observe that principal component analysis (PCA) on fMRI data not only decomposes the signal fluctuations into principal components ranked by the variance contribution, but also decomposes their temporal dynamics into ordered frequency bands, even within the 0.01 to 0.1 Hz BOLD frequency range. This observation suggests that dimension reduction of fMRI data using PCA should be determined not only based on the variance distribution of the spatial domain principal components, but also based on the frequency distribution of their corresponding projection vectors in the temporal domain.