Based on the local patch concept, we proposed locally reconstructive patch alignment (LRPA) for dimensionality reduction. For each patch, LRPA aims to find the low-dimensional subspace in which the reconstruction error of the within-class nearest neighbors is minimized and the reconstruction error of the between-class nearest neighbors is maximized. LRPA preserves the local structure hidden in the high-dimensional space. More importantly, LRPA has natural connections with linear regression classification (LRC). While LRC uses reconstruction errors as the classification rule, a sample can be classified correctly when the within-class reconstruction error is minimal. The goal of LRPA makes it cooperate well with LRC. The experimental results on the extended Yale B (YALE-B), AR, PolyU finger knuckle print, and the palm print databases demonstrate LRPA plus LRC is an effective and robust pattern-recognition system.