Captures only linear relationships; nonlinear structure requires t-SNE, UMAP, or autoencoders - Assumes variance equals importance (not always true---a low-variance feature could be the most predictive) - Components can be difficult to interpret, especially when many features contribute to each comp