In　the　spectral　methods　of　manifold　learning,　the　manifold　unfolding　tasks　are　formulated　as　optimization　problems.　The　optimal　solutions　to　these　problems　will　embed　all　samples　into　one　point.　To　avoid　the　degenerate　solutions,　the　spectral　methods　impose　a　unit　covariance　constraint　to　the　embedding　coordinates.　However,　this　constraint　usually　causes　highly　distorted　embeddings.　A　new　manifold　unfolding　method　is　proposed　in　this　paper,　which　discards　the　unit　covariance　constraint　completely.　The　central　idea　is　to　embed　the　manifold　boundary　at　first,　then　the　inner　regions.　The　embedding　positions　of　inner　samples　will　be　pulled　out　by　the　embedded　boundary　to　avoid　collapsing　into　one　point.　The　embedding　of　inner　samples　is　obtained　by　solving　a　linear　system　that　reflects　local　isometry　requirement,　using　the　embedding　of　boundary　as　a　boundary　condition.　The　embedding　of　boundary　is　determined　by　a　simplified　version　of　manifold,　and　a　manifold　boundary　detection　algorithm　and　a　manifold　graph　simplification　algorithm　are　thus　also　proposed.　Experimental　results　on　synthetic　and　real　data　sets　demonstrate　the　effectiveness　of　our　method,　which　results　in　less　mapping　distortion　than　spectral　methods.　Copyright　©　2010　Acta　Automatica　Sinica.　All　rights　reserved.