What does PCA do, and what are its limitations?

Principal component analysis creates orthogonal directions that capture decreasing amounts of variance in the fitted data. It can reduce dimensionality and correlated numeric features, often after suitable scaling. High variance is not the same as predictive usefulness, so dropping low-variance directions can remove target-relevant information. Components may be hard to interpret and ordinary PCA captures linear structure. Fit PCA only on training data, choose dimensionality through appropriate validation, and compare with a simpler baseline. Explained variance alone does not establish that downstream predictions improve.