Face recognition is an active research in the field of biometrics due to its potential benefits to various security based applications. To make the results of face recognition unsusceptible to different kinds of variations in the image and to enhance the accuracy, robustness of two or more methods can be fused in a single framework. The fusion can be achieved at various levels. Objective of this paper is to suggest optimal fusion of subspace methods to achieve robust results for various test conditions. This is achieved by performing feature level fusion of popular subspace methods namely PCA, LDA, LPP and ICA1. The Fusion is performed by considering different combinations of set of two, three and four subspace methods. Experiments are conducted by using two different databases: ORL and Yale. Experimental results suggest that by the fusion of these subspace approaches; there is a significant improvement in the accuracy compared to performance of an individual subspace method. This work helped us to determine the optimal combination of subspace methods to achieve robust results for specific test conditions.