Proper Laplacian Representation Learning