The Caccioppoli ultrafunctions (Q2417250)

``Weak'' solutions, such as {distributions} (as introduced by Laurent Schwartz, though other notions have been introduced by Colombeau and Sato), are solutions to partial differential equations (PDEs) that have physical meaning but are not differentiable functions (in many variables over the real or complex numbers). For example, a Sobolev space is more natural a function space (which only need be weakly [in the sense of distributions] partially differentiable) of solutions to many PDEs than that of partially differentiable functions. For example, the differential equation \(x y' = 0\), for the function \(y\) in the variable \(x\), has only constant differentiable functions as solutions. However, the notion of a point is a mathematical abstraction; in physics, we would consider small intervals around \(0\): thus the function \(y\) with value \(a\) for \(x < 0\) and value \(b\) for \(x > 0\) can also be considered a solution, because, for a test function \(\phi\) defined around \(0\), integration by parts yields \[ \int_{-\infty}^{+\infty} xy' \phi = y(x\phi)|_{-\infty}^{+\infty} - \int_{-\infty}^{+\infty} y (x\phi)' = 0. \] To define distributions as a generalization of differentiable functions, \begin{itemize} \item a notion of distribution, \item an embedding of the classical differentiable functions into that of distributions, and \item a notion of its derivative \end{itemize} are defined as follows (and as suggested by the example): \begin{itemize} \item A {distribution} is a continuous linear form on a function space, called test space, a space of infinitely differentiable functions of compact support (say on \(\mathbb{R}\) and taking values in \(\mathbb{R}\)). \item The {embedding} is given via the integral \(f \mapsto \int f \cdot\). In particular, any (locally) integrable function has a distributional derivative. \item The {derivative} \(T'\) of a distribution \(T\) is implicitly given by \[ T'(\cdot) := T(\cdot'); \] likewise the partial derivatives. For example, the derivative of the {Heaviside} function (as a distribution) on \(\mathbb{R}\) (that is \(1\) for all positive numbers and vanishes otherwise) is the Dirac measure (that evaluates a function in \(0\)). \end{itemize} Still, there are PDEs which can be treated more coherently by an even more general approach using non-standard analysis, where the real numbers are replaced by the hyperreal numbers and functions by ultrafunctions, particular functions on the hyperreals \(\mathbb{R}^* \supset \mathbb{R}\), which have recently been introduced in [\textit{V.~Benci}, Adv. Nonlinear Stud.~13, No.~2, 461--486 (2013; Zbl 1273.26035)]. (In the following, a superscript star denotes the ``natural non-standard analytic extension'' of the classical analytic object.) The original formulation to non-standard analysis by Robinson was model-theoretic; the present article builds on the more recent and simpler set-theoretic approach by a so-called {superstructure} \(V_\infty(E)\) (due to Elias Zakon) which is the union over all \(V_{n+1}(E) = V \cup P(E)\) where \(V_0(E) = E\) and \(P(E)\) is the power set of \(E\). The set \(P_{\mathrm{fin}}(V_\infty(\mathbb{R}))\) (where \(P_{\mathrm{fin}}(\mathbb{R})\) denotes the family of all finite subsets of \(\mathbb{R}\)) is enlarged by a ``point at infinity'' \(\Lambda\) and this set \(L\) is given a topology defined by the choice of an ultrafilter. While the real numbers are the completion of \(\mathbb{Q}\), that is, \[ \mathbb{R} = \{\text{Cauchy sequences in } \mathbb{Q} \} / \{\text{zero sequences in } \mathbb{Q} \}, \] we have \[ \mathbb{R}^* = \{\text{nets over $L$ in } \mathbb{R} \} / \{\text{nets eventually $0$ over $L$ in } \mathbb{R} \}. \] Given a net of functions \(f_\lambda : E_\lambda \to \mathbb{R}\) over \(L\), the \textit{internal} ultrafunction \(f : E^* \to \mathbb{R}^*\) is defined pointwise by \[ f(\lim \xi_\lambda) = \lim f_\lambda(\xi_\lambda). \] With a view to applications to partial differential equations and the calculus of variations, the present paper introduces \textit{Caccioppoli ultrafunctions} (conceivably named after Renato Caccioppoli, 1904--1959) that are suitable for localization and differentiation. To this end, suitable notions of \begin{itemize} \item a generalized partial derivative \(D_i\) and \item a generalized (so-called pointwise) integral \(\int\) \end{itemize} are introduced such that \begin{itemize} \item the derivative is local, that is, open subsets in the domain of the function are preserved, \item integration by parts works, and \item a generalized Gauss theorem holds. \end{itemize} While the ultrafunctions have a natural multiplication, the construction of the derivative does not violate Schwartz's impossibility theorem of a differential algebra containing all distributions, because the derivative does \textit{not} satisfy the Leibniz rule in general. In Section~2, the notions of non-standard analysis, such as the field of hyperreal numbers \(\mathbb{R}^*\), are introduced. Let \(\Omega\) be an open bounded subset in \(\mathbb{R}^N\). Section~3 introduces the \textit{Caccioppoli functions} on \(\mathbb{R}\) as linear combinations of certain ``indicator functions'' of \textit{Caccipoli sets} that are open, bounded sets whose boundary has measure zero. In particular, they are locally integrable. (In fact, the Caccioppoli functions are locally integrable functions that lend themselves towards modelling ultrafunctions from them.) The space of Caccioppoli ultrafunctions \(V_\Lambda(\Omega)\) on \(\mathbb{R}^*\) \begin{itemize} \item is {modelled} on the space of Caccioppoli functions \(V(\Omega)\) (that is, every function \(f\) in \(V_\Lambda(\Omega)\) is the internal limit \(f = \lim f_\lambda\) for \(f_\lambda\) in a finite-dimensional subspace \(V_\lambda(\Omega)\) of \(V(\Omega)\)), and \item is determined by the existence of a particular \(\sigma\)-\textit{basis} (that consists of internal ultrafunctions of which every ultrafunction is a hyperfinite sum). \end{itemize} The technical construction of the space \(V_\Lambda(\Omega)\) is then postponed to Section 5.1. Because the Caccioppoli functions are locally integrable, a natural integral on the Caccioppoli ultrafunctions can be defined; in particular, the integral \(\int f\) of the internal ultrafunctions \(f = \lim f_\lambda\) in the \(\sigma\)-basis is given by \(\lim \int f_\lambda\). In Section 4, the notion of generalized (directional) derivative \(D_i\) (for \(i = 1, \dots, N\)) is introduced that should satisfy \begin{itemize} \item on an internal ultrafunction, \(u = \lim u_\lambda\) in \(V_\Lambda(\Omega)\) for differentiable \(u_\lambda\) in \(V_\lambda(\Omega)\) by \(D_i u = \lim \partial_i u\), \item integration by parts \(\int D_i uv \, dx = - \int u D_i v \, dx\), and \item a Gauss theorem. \end{itemize} Its technical construction is then postponed to Section 5.2. Section 6.2 cites the example of the Poisson problem \(-\Delta F = f\) on a domain \(\Omega\) in \(\mathbb{R}^N\) such that \(F = 0\) on \(\partial \Omega\). Its solutions are given by the convolution product \(f \ast \phi\) of \(f\) with the fundamental solution \(\phi\) of the Laplace equation given, up to known constant, by \(\log |\cdot|\) for \(n = 2\) and \(|\cdot|^{n-2}\) for \(n > 2\). The solutions in dimension \(N = 2\) together with \(N > 2\) cannot be treated coherently over the reals, but over the hyperreal numbers. (Just like the differential equation \(y' = x^{1-n}\) is solved for \(n = 2\) by \(\log |\cdot|\), but by \((1-n)^{-1} \cdot x^{2-n}\) otherwise.) Specifically, the solution for \(N > 2\) minimizes a certain Dirichlet integral (defined via the Laplacian operator), but for \(N = 2\) such a minimum does not exist. However, Section 6.1 shows for all \(N\) (under certain general conditions, which are satisfied here) by a standard approximation over finite-dimensional spaces (that is, the solution is the limit of the net of solutions over all finite-dimensional vector spaces) that the corresponding Dirichlet integral (where the Laplacian is replaced by the generalized Laplacian) can be minimized among ultrafunctions. This minimizer can be made explicit under certain conditions. Section 6.3 shows, in dimension one, that the minimizer of \[ J(u)=\int_0^1 \frac{1}{2} a(u) |u'(x)|^2 - \gamma u(x)\,dx, \] for \(a\) on \(\mathbb{R}\) equal to \(0\) on \([1,2]\) and \(1\) elsewhere, is a step function that is a distribution with a physical meaning, but is not weakly differentiable, in particular, it is not in a Sobolev space. Reviewer's remarks. The section ``Notations'' defines some notations, such as \(\mathcal{C}(\Omega)\) for the set of continuous functions defined on a set \(\Omega\) in \(\mathbb{R}\), but other notations such as, supposedly, the set of locally integrable functions \(L^1_{\mathrm{loc}}(\Omega)\), are expected to be known. In Section 2.2, it is not clear which norm the infinite set \(E\) carries.