1: \begin{abstract}

2: In this paper we study a nonlinear dual space preconditioning approach for the relaxed \emph{Proximal Point Algorithm} (PPA) with application to monotone inclusions, called anisotropic PPA. The algorithm is an instance of Luque's nonlinear PPA wherein the nonlinear preconditioner is chosen as the gradient of a Legendre convex function. Since the preconditioned operator is nonmonotone in general, convergence cannot be shown using standard arguments, unless the preconditioner exhibits isotropy (preserves directions) as in existing literature. To address the broader applicability we show convergence along subsequences invoking a Bregman version of Fej\`er-monotonicity in the dual space.

3: Via a nonlinear generalization of Moreau's decomposition for operators, we provide a dual interpretation of the algorithm in terms of a forward iteration applied to a $D$-firmly nonexpansive mapping which involves the Bregman resolvent.

4: For a suitable preconditioner, superlinear convergence rates of arbitrary order are derived under a mild H\"older growth condition.

5: Finally, we discuss an anisotropic generalization of the proximal augmented Lagrangian method obtained via the propsed scheme. This has strong connections to Rockafellar's generalized and sharp Lagrangian functions.

6: \end{abstract}

7: