In　this　paper,　a　novel　approach　for　expression　invariant　face　recognition　is　proposed.　It＇s　implemented　by　common　vector　and　the　feature　extraction　is　based　on　Curvelet　transform,　which　is　one　of　the　multiscale　geometric　transforms　and　has　better　sparsity　than　the　other　transforms,　especially　for　the　edges.　We　assume　the　different　images　of　the　same　subject　as　an　interrelated　ensemble　with　expression　variations　which　could　be　viewed　as　sparse　respectively.　Then,　each　subject　can　be　represented　by　a　common　feature　vector,　capturing　its　essential　features,　and　an　innovation　feature　vector,　corresponding　to　the　expression　variations.　By　keeping　the　two　feature　vectors　of　each　subject,　we　can　use　them　to　classify　an　image　by　the　minimum　reconstruction　error.