Principal Component A nalysis and Fisher Linear Discriminant methods have demonstrated their success in fac edete ction, r ecognition and atrcking. The repr e- sentations in these subspac e methods are based on sec- ond order statistics of the image set, and do not address higher order statistic al dep endencies such as the rela- tionships among three or more pixels. R ecently Higher Order Statistics and Independent Component Analysis (ICA) have been used as informative repr esentations for visual recognition. In this paper, we investigate the use of Kernel Principal Component Analysis and Ker- nel Fisher Linear Discriminant for learning low dimen- sional repr esentations for fac e recognition, which we call Kernel Eigenface and Kernel Fisherface methods. While Eigenface and Fisherface methods aim to find proje ction directions b ased on second order corr elation of samples, Kernel Eigenface and Kernel Fisherface methods provide gener alizationswhich take higher or- der correlations into account. We compare the perfor- mance of kernel methods with classical algorithms such as Eigenface, Fisherface, ICA, and Support Vector Ma- chine (SVM) within the context of appearance-based face recognition problem using two data sets where im- ages vary in pose, scale, lighting and expr ession. Ex- perimental results show that kernel methods provide better r epr esentations and achieve lower error rates for face recognition.