Characterizations of Variational Convexity and Tilt Stability via Quadratic Bundles

Pham Duy Khanh, Boris S. Mordukhovich, Vo Thanh Phat, Le Duc Viet · 2025

Read Characterizations of Variational Convexity and Tilt Stability via Quadratic Bundles by Pham Duy Khanh, Boris S. Mordukhovich, Vo Thanh Phat, Le Duc Viet (2025) 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-2501-04629.txt.

Introduction

Optimization judges a solution's stability by how it wobbles under small shifts, a property called tilt stability, and its cousin variational convexity. Pham Duy Khanh and three co-authors characterize both in 2025 for a broad class of rough functions, using a new second order tool called quadratic bundles. The setting drops an assumption earlier work needed, that the subgradient map behaves continuously.

Their characterizations come in point based form, checkable at a single candidate point. The bridge is a smoothing of the function, the Moreau envelope, whose classical curvature encodes the generalized one. The quadratic bundle machinery itself gets a revision along the way. The authors present the results as sharper tests for stability in nonsmooth problems.

Catalog record

CREATORPham Duy Khanh, Boris S. Mordukhovich, Vo Thanh Phat, Le Duc Viet
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-04629

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Characterizations of Variational Convexity and Tilt Stability via Quadratic Bundles Pham Duy Khanh, Boris S. Mordukhovich, Vo Thanh Phat, Le Duc Viet Abstract Dedicated to the memory of Hedy Attouch, an outstanding mathematician and very nice human being Abstract . In this paper, we establish characterizations of variational s -convexity and tilt stability for prox-regular functions in the absence of subdifferential continuity via quadratic bundles, a kind of primal-dual generalized second-order derivatives recently introduced by Rockafellar. Deriving such characterizations in the effective pointbased form requires a certain revision of quadratic bundles investigated below. Our device is based on the notion of generalized twice differentiability and its novel characterization via classical twice differentiability of the associated Moreau envelopes combined with various limiting procedures for functions and sets. Key words . Set-valued and variational analysis, prox-regularity, variational convexity, tilt stability, epi-convergence, subderivatives, generalized twice differentiability, quadratic bundles, Moreau envelopes Mathematics Subject Classification (2020) 49J52, 49J53, 90C31 1 Introduction In the recent years, the notion of ( strong ) variational convexity of extended-real-valued functions whose subdifferential graphs are not distinguishable from those for locally (strongly) convex functions, emerged as an attempt to study local optimality and numerical algorithms of the proximal and augmented Lagrangian types; see the series of Rockafellar’s papers [ 28 , 29 , 30 , 31 ] , which are mainly based on second-order subderivative constructions of the primal type in finite-dimensional spaces. Another dual-type approach to these fundamental issues, in both finite and infinite dimensions, was developed in [ 10 , 11 , 12 , 20 ] with applications [ 13 ] to generalized Newtonian methods for solving problems of nonsmooth nonconvex optimization and machine learning. The latter approach relays on the usage of coderivative-based second-order subdifferentials (or generalized Hessians) introduced by Mordukhovich in [ 17 ] and then largely developed in the literature over the years; see the new book [ 20 ] and the vast bibliographies therein for numerous theoretical and computational results involving such dual-type constructions with various applications to optimization, control, variational stability, and practical modeling. Quite recently [ 29 ] , Rockafellar suggested yet another, primal-dual approach to variational convexity and its strong counterpart based on the newly introduced notion of quadratic bundles for convex 1 1 1 When the results of our study in this paper and in the preceding one [ 14 ] have been obtained, we got familiar with the very fresh Rockafellar’s manuscript [ 32 ] containing extensions of quadratic bundles to nonconvex functions and their applications to some issues discussed in [ 14 ] and in the current paper; see more comments below. extended-real-valued functions. This approach is associated with the concept of epigraphical convergence of sequences of functions and sets, which was originated by Wijsman in [ 35 ] for convex functions and then largely grew outside of convexity. The fundamental contributions by Hedy Attouch into the study and applications of epi-convergence are difficult to overstate; see his monograph [ 1 ] as well as the book by Rockafellar and Wets [ 33 ] with the references therein. The notion of quadratic bundles is closely related to generalized twice differentiability of extended-real-valued functions defined in [ 29 ] and then systematically studied in our preceding work [ 14 ] , where this notion was equivalently characterized, for the general class of prox-regular functions, via the classical twice differentiability of the associated Moreau envelopes . The obtained characterization played a crucial role in [ 14 ] for our investigation of generalized twice differentiability of prox-regular function and the modified version of quadratic bundles without subdifferential continuity. In particular, we established in [ 14 ] the nonemptiness of quadratic bundles for the general class of prox-regular functions. The main goal of this paper is to provide pointbased primal-dual characterizations of the notion of variational s -convexity of prox-regular functions via quadratic bundles in the absence of subdifferential continuity . In this way, we also obtain new pointbased primal-dual characterizations of tilt-stable local minimizers for such functions with quantitative modulus relationships . An important role in our analysis is played the newly established connections between the quadratic bundle of the function in question and the Hessian bundle of the associated Moreau envelope. The rest of the paper is structured as follows. Section 2 equips the reader with some basic constructions and facts of variational analysis and generalized differentiation for sets, extended-real-valued functions and set-valued mappings. Section 3 deals with generalized twice differentiability and quadratic bundles of functions. Section 4 presents primal-dual characterization of variational s -convexity, while Section 5 addresses tilt stability of local minimizers. The concluding Section 6 summarizes the main achievements of the paper and discusses some directions of our future research. 2 Preliminaries from Variational Analysis The setting of this paper is the n -dimensional space {\rm I\!R}^{n} equipped (unless otherwise stated) with the Euclidean norm \|\cdot\| and the inner product \langle\cdot,\cdot\rangle . We signify by B(\bar{x},r) the open ball centered at \bar{x} with radius r>0 and write x_{k}\xrightarrow{C}\bar{x} when the sequence \{x_{k}\}\subset C converges to \bar{x} . For an extended-real-valued function f:{\rm I\!R}^{n}\to{\rm I\!R}\cup\{\infty\}:=\overline{{\rm I\!R}} , the domain of f and epigraph of f are defined, respectively, by \mbox{\rm dom}\,f:=\big\{x\in{\rm I\!R}^{n}\;\big|\;f(x)<\infty\big\}\;\mbox{ and }\;\mbox{\rm epi}\,f:=\big\{(x,\alpha)\in{\rm I\!R}^{n}\times{\rm I\!R}\;\big|\;\alpha\geq f(x)\big\}. It is said that f is proper if \mbox{\rm dom}\,f\neq\emptyset and f is lower semicontinuous (l.s.c.) at \bar{x} if \liminf\limits_{x\to\bar{x}}f(x)\geq f(\bar{x}) . The convex hull and the conic convex conic hull of a set C\subset{\rm I\!R}^{n} are denoted by \mbox{\rm co}\,C and \mbox{\rm cone}\,C , respectively. The indicator function of C is defined by \delta_{C}(x):=0 for x\in C and \delta_{C}(x):=\infty otherwise. For a set-valued mapping/multifunction F:{\rm I\!R}^{n}\rightrightarrows{\rm I\!R}^{m} , the domain and graph of F are \mbox{\rm dom}\,F:=\big\{x\in{\rm I\!R}^{n}\;\big|\;F(x)\neq\emptyset\big\}\;\mbox{ and }\;\mbox{\rm gph}\,F:=\big\{(x,y)\in{\rm I\!R}^{n}\times{\rm I\!R}^{m}\;\big|\;y\in F(x)\big\}. Given f:{\rm I\!R}^{n}\to\overline{{\rm I\!R}} , the (Fréchet) regular subdifferential of f at \bar{x}\in\mbox{\rm dom}\,f is defined by \widehat{\partial}f(\bar{x}):=\Big\{v\in{\rm I\!R}^{n}\;\Big|\;\liminf\limits_{x\to\bar{x}}\frac{f(x)-f(\bar{x})-\langle v,x-\bar{x}\rangle}{\|x-\bar{x}\|}\geq 0\Big\} and the (Mordukhovich) limiting subdifferential of f at \bar{x} is \partial f(\bar{x}):=\mathop{{\rm Lim}\,{\rm sup}}\limits_{x\xrightarrow{f}\bar{x}}\widehat{\partial}f(\bar{x}), (2.1) where \mathop{{\rm Lim}\,{\rm sup}}\limits_{x\to\bar{x}}F(x) stands for the (Painlevé-Kuratowski) outer limit of F:{\rm I\!R}^{n}\rightrightarrows{\rm I\!R}^{m} at \bar{x} defined by \mathop{{\rm Lim}\,{\rm sup}}\limits_{x\to\bar{x}}F(x):=\big\{y\in{\rm I\!R}^{m}\;\big|\;\exists\,x_{k}\to\bar{x},\;y_{k}\to y\;\mbox{ with }\;y_{k}\in F(x_{k})\;\mbox{ as }\;k\in{\rm I\!N}:=\{1,2,\ldots\}\Big\}. When f is convex, both sets \widehat{\partial}f and \partial f agree with the subdifferential of convex analysis \partial f(\bar{x})=\big\{v\in{\rm I\!R}^{n}\;\big|\;\langle v,x-\bar{x}\rangle\leq f(x)-f(\bar{x})\;\mbox{ whenever }\;x\in{\rm I\!R}^{n}\big\}. When f is nonconvex, the set \widehat{\partial}f(\bar{x}) is still convex, while \partial f(\bar{x}) is not so, even for simple functions as, e.g., f(x)=-|x| , where \widehat{\partial}f(0)=\emptyset while \partial f(0)=\{-1,1\} . Nevertheless, \partial f (but not \widehat{\partial}f ) enjoys full calculus in both finite and infinite dimensions, which is based on extremal/variational principles of variational analysis; see the books [ 18 , 19 , 20 , 33 , 34 ] and the references therein for more details. Next we briefly review the notions of epi-convergence for sequences of extended-real-valued functions as well as of second-order subderivatives for such functions. More details on these and related notions of variational analysis can be found in [ 33 ] . To begin with, recall that the outer limit and inner limit of a set sequence \{C_{k}\}\subset{\rm I\!R}^{n} are defined, respectively, by \mathop{{\rm Lim}\,{\rm sup}}\limits_{k\to\infty}C_{k}=\big\{x\;\big|\;\exists\,\{x_{k_{j}}\}\subset\{x_{k}\}\;\mbox{ with }\;x_{k_{j}}\to x,\;x_{k}\in C_{k}\big\},\quad\mathop{{\rm Lim}\,{\rm inf}}\limits_{k\to\infty}C_{k}=\big\{x\;\big|\;\exists\,x_{k}\stackrel{{\scriptstyle C_{k}}}{{\to}}x\big\}. If \mathop{{\rm Lim}\,{\rm sup}}\limits_{x\to\infty}C_{k}=\mathop{{\rm Lim}\,{\rm inf}}\limits_{x\to\infty}C_{k}=:C , then

Read the complete text