Machine learning interatomic potentials (MLIPs) based on a large data set obtained by density functional theory calculation have been developed recently. This study gives both conceptual and practical bases for the high accuracy of MLIPs, although MLIPs have been considered to be simply an accurate black-box description of atomic energy. We also construct the most accurate MLIP of elemental Ti ever reported using a linearized MLIP framework and many angular-dependent descriptors, which also corresponds to a generalization of the modified embedded atom method potential.