Fast and accurate phase equilibrium calculations have been a hot research topic during the past decades due to the increased computational effort they impose in process simulation. Reduction methods have attracted significant interest as a means to accelerate calculations by decreasing the problem's dimensionality. However, when many binary interaction parameters (BIPs) exhibit non-zero values, reduction methods offer limited advantage and the introduction of approximation error is inevitable. In this work a new method is presented for replacing the conventional spectral decomposition basis vectors by new ones so that the approximation error of the energy parameter is minimized instead of that of BIP matrix. The new reduced variables set leads to improved flash calculations accuracy, thus allowing them to be performed at a given accuracy level using fewer reduced variables compared to the conventional approach. Moreover, it is shown that the formulation of the proposed matrix decomposition method is a generalization of the spectral analysis one. The extension of the proposed method to temperature-dependent BIPs is also discussed. A set of examples demonstrates the efficiency of the proposed method. The method can be readily applied to all existing phase behavior algorithms within the reduced variables framework as it intervenes solely on the treatment of the BIP matrix. (C) 2012 Elsevier B.V. All rights reserved.