Discretization error cancellation in electronic structure calculation: toward a quantitative study^∗,∗∗

¹ CERMICS, Ecole des Ponts and INRIA Paris, 6 & 8 Avenue Blaise Pascal, 77455 Marne-la-Vallée, France.
cances@cermics.enpc.fr
² Sorbonne Universités, UPMC Univ. Paris 06 and CNRS, UMR 7598, Laboratoire Jacques-Louis Lions, 75005, Paris, France, and Sorbonne Universités, UPMC Univ. Paris 06, Institut du Calcul et de la Simulation, 75005, Paris, France.
dusson@ljll.math.upmc.fr

Received: 27 January 2017
Revised: 2 June 2017
Accepted: 4 July 2017

Abstract

It is often claimed that error cancellation plays an essential role in quantum chemistry and first-principle simulation for condensed matter physics and materials science. Indeed, while the energy of a large, or even medium-size, molecular system cannot be estimated numerically within chemical accuracy (typically 1 kcal/mol or 1 mHa), it is considered that the energy difference between two configurations of the same system can be computed in practice within the desired accuracy. The purpose of this paper is to initiate the quantitative study of discretization error cancellation. Discretization error is the error component due to the fact that the model used in the calculation (e.g. Kohn−Sham LDA) must be discretized in a finite basis set to be solved by a computer. We first report comprehensive numerical simulations performed with Abinit [X. Gonze, B. Amadon, P.-M. Anglade et al., Comput. Phys. Commun. 180 (2009) 2582–2615; X. Gonze, J.-M. Beuken, R. Caracas et al., Comput. Materials Sci. 25 (2002) 478–492] on two simple chemical systems, the hydrogen molecule on the one hand, and a system consisting of two oxygen atoms and four hydrogen atoms on the other hand. We observe that errors on energy differences are indeed significantly smaller than errors on energies, but that these two quantities asymptotically converge at the same rate when the energy cut-off goes to infinity. We then analyze a simple one-dimensional periodic Schrödinger equation with Dirac potentials, for which analytic solutions are available. This allows us to explain the discretization error cancellation phenomenon on this test case with quantitative mathematical arguments.