Active Shape Models (ASM) is a powerful statistical tool for extracting objects, e.g. face, from images. It is composed of two parts: ASM model and ASM search. In ASM, these two parts are treated separately. First, ASM model is trained. Then, ASM search is performed using this model. However, we find that these two parts are closely interrelated. The performance of ASM depends on both of them. Improvement on one of them does not consequentially improve the overall performance, for it may worsen the other. In this paper, we find the key parameter that relates these two parts: subspace explanation proportion. By optimizing subspace explanation proportion, the overall performance of ASM can improve by a percentage of about 20 in our experiments. Furthermore, this paper proposes to decompose the ASM overall error into ASM model subspace reconstruction error and ASM search error, proving that the square of the subspace reconstruction error is linearly related with the subspace explanation proportion and finding that the square of the search error is a piecewise function of the explanation proportion. This decomposition is a new method for further analysis and possible improvement. Based on this decomposition, we propose a method to estimate the optimal explanation proportion. Experiments show that the estimation is satisfactory.