On strict proto-differentiability of set-valued mappings
Helmut Gfrerer · 2024
Read On strict proto-differentiability of set-valued mappings by Helmut Gfrerer (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2411-01346.txt.
Description
Abstract Abstract. We will show that a multifunction is strictly proto-differentiable at a point of its graph if and only if it is graphically strictly differentiable, i.e., the graph of the multifunction locally coincides, up to a change of coordinates, with the graph of a single-valued mapping, which is strictly differentiable at the transformed reference point.
Catalog record
| CREATOR | Helmut Gfrerer |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2411-01346
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
On strict proto-differentiability of set-valued mappings Helmut Gfrerer Abstract Abstract. We will show that a multifunction is strictly proto-differentiable at a point of its graph if and only if it is graphically strictly differentiable, i.e., the graph of the multifunction locally coincides, up to a change of coordinates, with the graph of a single-valued mapping, which is strictly differentiable at the transformed reference point. This result allows point-based characterizations of strict proto-differentiability in terms of various generalized derivatives. Further we will prove that under strict proto-differentiability the properties of strong metric regularity, metric regularity and strong metric subregularity are equivalent. Finally, under strict proto-differentiability of the subgradient mapping, we provide a novel second-order relation between function values and subgradients for prox-regular functions which constitutes a nonsmooth extension of the trapezoidal rule of numerical integration. Key words. Strict proto-differentiability, strict differentiability, graphically Lipschitzian, generalized derivatives, strong metric regularity. AMS Subject classification. 49J53, 90C31. 1 Introduction The notion of strict proto-differentiability was introduced by Poliquin and Rockafellar [ 15 ] in their study of the subgradient mapping of prox-regular functions. It was shown in [ 15 ] that for a prox-regular and subdifferentially continuous function \varphi:\mathbb{R}^{n}\to(-\infty,\infty] the graph of the subgradient mapping \partial\varphi locally coincides, up to a linear transformation of the coordinates, with the graph of the proximal mapping which is a single-valued and locally Lipschitzian mapping in this case. Then it was shown in [ 16 ] that the subgradient mapping for such functions is strictly proto-differentiable if and only if the proximal mapping is strictly differentiable. If the assumption of subdifferential continuity is omitted, the stated properties still hold for some \varphi -attentive \epsilon -localization of \partial\varphi . Since the introduction of strict proto-differentiability in [ 15 ] , most contributions deal only with subgradient mappings for special function classes and its relation with strict twice epi-differentiability, see, e.g., [ 9 , 10 , 11 ] for some recent papers. Very little investigations have been done for general multifunctions. In the very recent paper [ 8 ] it is shown that for a graphically Lipschitzian and SCD-semismooth ∗ multifunction, the set of all points belonging to its graph where strict proto-differentiability fails to hold, is negligible with respect to some Hausdorff measure. Strict proto-differentiability has its roots in the notion of strictly smooth sets as introduced by Rockafellar [ 17 ] . We will show that a strictly smooth set coincides locally around the reference point, up to a permutation of the coordinates, with the graph of a strictly differentiable single-valued function, which implies in particular that the set is a Lipschitz manifold in the sense of [ 17 ] . This property carries over to strict proto-differentiability of multifunctions and allows its characterization by several generalized derivatives together with the graphical Lipschitz property. These characterizations can be used to show that under strict proto-differentiability and some condition on the dimension of the graphic Lipschitz property, the properties of strong metric regularity, metric regularity and strong metric subregularity are equivalent. We will also consider the special case of the subgradient mapping of prox-regular functions. Besides various characterizations of strict proto-differentiability by means of \varphi -attentive generalized derivatives as introduced in the very recent paper [ 4 ] , we also provide a new second-order relationship between function values and subgradients. The paper is organized in the following way. In Section 2, which is divided into several subsections, we provide all the necessary definitions together with some easy consequences and well-known results. In Section 3 we state the characterizations of strictly smooth sets and strict proto-differentiability of mappings together with the equivalences between various regularity properties. The final Section 4 is devoted to the subgradient mapping of prox-regular functions. Given a set-valued mapping F:\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{m} , we denote its domain by \mathrm{dom}\,F:=\{x\in\mathbb{R}^{n}\,\mid\,F(x)\not=\emptyset\} and its graph by \mathrm{gph}\,F:=\{(x,y)\in\mathbb{R}^{n}\times\mathbb{R}^{m}\,\mid\,y\in F(x)\} . The space of real m\times n matrices is denoted by \mathbb{R}^{m\times n} and for a nonsingular matrix A\in\mathbb{R}^{n\times n} we define A^{-T}:=(A^{-1})^{T} . For A\in\mathbb{R}^{n\times d} , B\in\mathbb{R}^{m\times d} we denote by {\rm rge\;}A:=\{Ap\,\mid\,p\in\mathbb{R}^{d}\},\quad{\rm rge\;}(A,B):=\{(Ap,Bp)\,\mid\,p\in\mathbb{R}^{d}\} the subspaces of \mathbb{R}^{n} and \mathbb{R}^{n}\times\mathbb{R}^{m} generated by the ranges of A and (A,B) , respectively. The dimension of a subspace L of a finite-dimensional Hilbert space is denoted by \dim L and L^{\perp} signifies its orthogonal complement. Given a differentiable mapping f:\mathbb{R}^{n}\to\mathbb{R}^{m} , we write \nabla f(x) for the Jacobian of f at x . Hence, in case of a function ( m=1 ) \nabla f(x) is a row vector. The open ball with radius r>0 around a point x\in\mathbb{R}^{n} is denoted by {\cal B}_{r}(x) . The polar cone to a set A\subset\mathbb{R}^{n} is denoted by A^{\circ} . Given a parameterized family C_{t} of subsets of a metric space, where t also belongs to a metric space, we define the upper (outer) and lower (inner) limits in the sense of Painlevé–Kuratowski as \displaystyle\limsup_{t\to\bar{t}}C_{t}:=\{x\,\mid\,\liminf_{t\to\bar{t}}\,{\rm dist}(x,C_{t})=0\},\ \liminf_{t\to\bar{t}}C_{t}:=\{x\,\mid\,\limsup_{t\to\bar{t}}\,{\rm dist}(x,C_{t})=0\}. 2 Basic definitions and preliminary results 2.1 Variational geometry and strictly smooth sets Definition 2.1 . Consider a set \Omega\subset\mathbb{R}^{n} and a point \bar{z}\in\Omega . 1. The tangent cone, the regular (Clarke) tangent cone and the paratingent cone to \Omega at \bar{z} are given by \displaystyle T_{\Omega}(\bar{z}):=\limsup_{t\downarrow 0}\frac{\Omega-\bar{z}}{t},\quad\widehat{T}_{\Omega}(\bar{z}):=\liminf_{{\textstyle{z\mathop{\rightarrow}\limits^{\Omega}\bar{z}\atop t\downarrow 0}}}\frac{\Omega-z}{t},\quad T^{P}_{\Omega}(\bar{z}):=\limsup_{{\textstyle{z\mathop{\rightarrow}\limits^{\Omega}\bar{z}\atop t\downarrow 0}}}\frac{\Omega-z}{t}. The set \Omega is called geometrically derivable at \bar{z} if T_{\Omega}=\lim_{t\downarrow 0}(\Omega-\bar{z})/t and \Omega is said to be strictly smooth at \bar{z} if \lim_{{\textstyle{z\mathop{\rightarrow}\limits^{\Omega}\bar{z}\atop t\downarrow 0}}}\frac{\Omega-z}{t} exists, i.e., when \widehat{T}_{\Omega}(\bar{z})=T^{P}_{\Omega}(\bar{z}) . 2. The regular normal cone and the limiting normal cone to \Omega at \bar{z} are given by \displaystyle\widehat{N}_{\Omega}(\bar{z}):=\big(T_{\Omega}(\bar{z})\big)^{\circ},\quad N_{\Omega}(\bar{z}):=\limsup_{z\mathop{\rightarrow}\limits^{\Omega}\bar{z}}\widehat{N}_{\Omega}(z). The set \Omega is called normally regular at \bar{z} if \widehat{N}_{\Omega}(\bar{z})=N_{\Omega}(\bar{z}) . It is easy to see that the inclusions \widehat{T}_{\Omega}(\bar{z})\subset T_{\Omega}(\bar{z})\subset T^{P}_{\Omega}(\bar{z}) always hold. Further, \widehat{T}_{\Omega}(\bar{z}) is a closed convex cone and T^{P}_{\Omega}(\bar{z}) is a closed cone having the property T^{P}_{\Omega}(\bar{z})=-T^{P}_{\Omega}(\bar{z}) . Thus, \Omega is strictly smooth at \bar{z} if and only if \widehat{T}_{\Omega}(\bar{z})=T^{P}_{\Omega}(\bar{z})(=T_{\Omega}(\bar{z})) (2.1) is a subspace. Clearly, \widehat{N}_{\Omega}(\bar{z})\subset N_{\Omega}(\bar{z}) . Further, if \Omega is locally closed at \bar{z} then the tangent-normal relation \widehat{T}_{\Omega}(\bar{z})=N_{\Omega}(\bar{z})^{\circ} (2.2) holds, cf. [ 20 , Theorem 6.28] . Lemma 2.2 . Assume that the set \Omega\subset\mathbb{R}^{n} is locally closed at \bar{z} . Then \Omega is strictly smooth at \bar{z} if and only if both T^{P}_{\Omega}(\bar{z}) and N_{\Omega}(\bar{z}) are subspaces satisfying T^{P}_{\Omega}(\bar{z})^{\perp}=N_{\Omega}(\bar{z}). (2.3) Moreover, if \Omega is strictly smooth at \bar{z} then \Omega is normally regular at \bar{z} . Proof. If \Omega is strictly smooth, then T_{\Omega}^{P}(\bar{z}) is a subspace and, by polarizing ( 2.1 ) and taking into account ( 2.2 ), we obtain that T^{P}_{\Omega}(\bar{z})^{\circ}=T^{P}_{\Omega}(\bar{z})^{\perp}=T_{\Omega}(\bar{z})^{\circ}=\widehat{N}_{\Omega}(\bar{z})=\widehat{T}_{\Omega}(\bar{z})^{\circ}={\rm conv\,}N_{\Omega}(\bar{z}) is a subspace. Since \widehat{N}_{\Omega}(\bar{z})\subset N_{\Omega}(\bar{z}) , we conclude that {\rm conv\,}N_{\Omega}(\bar{z})=N_{\Omega}(\bar{z})=\widehat{N}_{\Omega}(\bar{z})=T^{P}_{\Omega}(\bar{z})^{\perp} is a subspace and \Omega is normally regular at \bar{z} . Conversely, if T^{P}_{\Omega}(\bar{z}) and N_{\Omega}(\bar{z}) are subspaces satisfying ( 2.3 ), by polarizing ( 2.3 ) and using ( 2.2 ) we obtain that T^{P}_{\Omega}(\bar{z})=N_{\Omega}(\bar{z})^{\perp}=N_{\Omega}(\bar{z})^{\circ}=\widehat{T}_{\Omega}(\bar{z}) showing that \Omega is strictly smooth at \bar{z} . ∎ 2.2 Generalized differentiation and strict proto-differentiability of set-valued mappings Definition 2.3 . Let F:\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{m} be a mapping and let (\bar{x},\bar{y})\in\mathrm{gph}\,F . 1. The strict derivative D_{*}F(\bar{x},\bar{y}):\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{m} , the graphical derivative DF(\bar{x},\bar{y}):\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{m} and the limiting (Mordukhovich) coderivative D^{*}F(\bar{x},\bar{y}):\mathbb{R}^{n}\rightrightarrows\mathbb{R}^{m} at \bar{x} for \bar{y} are given by \displaystyle\mathrm{gph}\,D_{*}F(\bar{x},\bar{y})=T_{\mathrm{gph}\,F}^{P}(\bar{x},\bar{y}),\quad\mathrm{gph}\,DF(\bar{x},\bar{y})=T_{\mathrm{gph}\,F}(\bar{x},\bar{y}), \displaystyle\mathrm{gph}\,D^{*}F(\bar{x},\bar{y})=\{(y^{*},x^{*})\,\mid\,(x^{*},-y^{*})\in N_{\mathrm{gph}\,F}(\bar{x},\bar{y})\}. 2. The mapping F is called graphically regular at (\bar{x},\bar{y}) if \mathrm{gph}\,F is normally regular at (\bar{x},\bar{y}) .