Generalized Second-Order Derivatives of Convex Functions in Reflexive Banach Spaces (Q4032453)

Let \(E\) be a reflexive Banach space. The author considers an extended- real-valued convex functional \(f\) defined on \(E\), and a point \(x\) where \(f(x)\) is finite. Following Rockafellar, he defines for every fixed \(\xi\) in \(E^*\) a family of functions on \(E\) indexed by \(t>0\): \[ \left(\Delta_ t^{(2)} f\right)_{x,\xi}(v) := {2\over t^ 2}\left[ f(x+tv) - f(x) - t\left\langle \xi,v \right\rangle \right], \;v\in E. \] If, as \(t\to 0^ +\), these functions epi-converge (Mosco) to some limit function \(\phi\) obeying \(\phi(0)>-\infty\), then the function \(f\) is called twice epi-differentiable at \(x\) relative to \(\xi\), and the function \(\phi\) is denoted by \(f''_{x,\xi}\). Section 2 of the paper describes the elementary properties of this construction -- in particular, the role of the Legendre-Fenchel transform -- and presents some simple examples of twice epi-differentiable convex functions. Note that for \(f''_{x,\xi}\) to exist, it is necessary that \(\xi\) belong to \(\partial f(x)\), the usual convex subgradient of \(f\) at the point \(x\). Section 3 deals with proto-differentiation, a known notion of first-order differentiation for set-valued mappings again based on set convergence. Several results of independent interest are given, but the main one asserts that in the situation of the previous paragraph, the function \(f\) is twice epi-differentiable at \(x\) relative to \(\xi\) if and only if \(\xi\) lies in \(\partial f(x)\) and \(\partial f\) is protodifferentiable at \(x\) relative to \(\xi\), with its proto-derivative being a maximal monotone operator. The proof is built on a theorem of Attouch relating the epi- convergence of a family of convex functions to the graphical convergence of the corresponding subgradient multifunctions. Some calculus of second epiderivatives is described in Section 4. The highlight here is a careful study of the relationship between differentiability properties of \(f\) and those of the corresponding Moreau-Yosida regularization \(f^ \lambda(x) := \inf\left\{ f(u) + {1\over 2\lambda} \left\| u - x \right\|^ 2 : u\in E \right\}\). This is the technical advance supporting Section 5, where the author shows that the twice epi-differentiability of an extended-valued integrand \(f(t,x)\) on \([0,1]\times{\mathbf R}^ N\) implies the same property for the corresponding integral functional \(I_ f(x) = \int_ 0^ 1 f(s,x(s)) ds\) on \(L^ p[0,1]\) (\(1<p<\infty\)). The integrand must be ``normal, proper, and convex,'' as defined by Rockafellar. Under this condition, when one has a selection \(\xi(s)\in\partial f(s,x(s))\) a.e.\ such that \(f(s,\cdot)\) is twice epi-differentiable at \(x(s)\) relative to \(\xi(s)\) a.e., then it follows that \(I_ f\) is twice epi-differentiable at \(x\) relative to \(\xi\), and \[ (I_ f)''_{x,\xi}(v) = \int_ 0^ 1 f''_{x(s),\xi(s)}(s,v(s)) ds, \;v\in L^ p[0,1]. \]