This paper presents a novel learning method for nonlinear mapping between arbitrary dimensional spaces. Unlike artificial neural nets, GMDH, and other methods, our method doesn't require complicated control parameters. Providing a feasible error threshold and training samples, it automatically divides the objective mapping into partially linear mappings. Since decomposed mappings are maintained by a binary tree, the linear mapping corresponding to an input is quickly selected. We call this method Partially Linear Mapping tree (PaLM-tree) . In order to estimate the most reliable linear mappings satisfying the feasible error criterion, we employ split-and-merge strategy for the decomposition. Through the experiments on function estimation, image segmentation, and camera calibration problems, we confirmed the advantages of PaLM-tree.