A new theoretical study shows that in kernel ridge regression, generalization hinges on the alignment between learning targets and the eigenvectors of the kernel matrix, while training error is nearly ...