A duality method for non-convex problems in Calculus of Variations

Overview & Synopsis

In this thesis, we study a general principle of convexification to treat certain non convex variationalproblems in Rd. Thanks to this principle we are able to enforce the powerful duality techniques andbring back such problems to primal-dual formulations in Rd+1, thus making efficient the numericalsearch of a global minimizer. A theory of duality and calibration fields is reformulated in the caseof linear-growth functionals. Under suitable assumptions, this allows us to revisit and extend anexclusion principle discovered by Visintin in the 1990s and to reduce the original problem to theminimization of a convex functional in Rd. This result is then applied successfully to a class offree boundary or multiphase problems that we treat numerically obtaining very accurate interfaces.On the other hand we apply the theory of calibrations to a classical problem of minimal surfaceswith free boundary and establish new results related to the comparison with its variant where theminimal surfaces functional is replaced by the total variation. We generalize the notion of calibrabilityintroduced by Caselles-Chambolle and Al. and construct explicitly a dual solution for the problemassociated with the second functional by using a locally Lipschitzian potential related to the distanceto the cut-locus. The last part of the thesis is devoted to primal-dual optimization algorithms forthe search of saddle points, introducing new more efficient variants in precision and computationtime. In particular, we experiment a semi-implicit variant of the Arrow-Hurwicz method whichallows to reduce drastically the number of iterations necessary to obtain a sharp accuracy of theinterfaces. Eventually we tackle the structural non-differentiability of the Lagrangian arising fromour method by means of a geometric projection method on the epigraph, thus offering an alternativeto all classical regularization methods.