Generalized Twice Differentiability and Quadratic Bundles

in Second-Order Variational Analysis

Pham Duy Khanh∗ Boris S. Mordukhovich† Vo Thanh Phat‡ Le Duc Viet§

January 7, 2025
arXiv:2501.02067v1 [[Link]] 3 Jan 2025

Abstract. In this paper, we investigate the concepts of generalized twice differentiability and quadratic bundles
of nonsmooth functions that have been very recently proposed by Rockafellar in the framework of second-order
variational analysis. These constructions, in contrast to second-order subdifferentials, are defined in primal spaces.
We develop new techniques to study generalized twice differentiability for a broad class of prox-regular functions,
establish their novel characterizations. Subsequently, quadratic bundles of prox-regular functions are shown to be
nonempty, which provides the ground of potential applications in variational analysis and optimization.
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
Starting from the original works [24, 25], second-order epi-derivatives/subderivatives have been designed as
neoclassical counterparts of the Hessians, while concerning epigraphs of second-order difference quotients
in contrast to their graph in the classical approach. The corresponding property, labeled as twice epi-
differentiability, holds in broad settings such as maxima of C 2 -smooth functions, compositions of piecewise
linear-quadratic and smooth functions [30], and—more generally—parabolically regular functions; see [13,
14]. However, generalized second-order derivatives of this primal type suffer from the lack of robustness,
which prevents developing satisfactory calculus rules and broad applications. This is contrary to the
second-order subdifferential by Mordukhovich, a robust construction of the dual type; see [17] and the
references therein. In particular, second-order subderivatives are capable of characterizing strong local
minimizers yet fail to do so for tilt-stable local minimizers, a variational stability property requiring stable
behavior of local minimizers under tilted/linear shifts of objective functions; see the landmark work [23]
and many subsequent publications as, e.g., [3, 5, 17, 18, 19] among others.
This calls for the need of more robust generalized second-order constructions of primal and/or primal-
dual types to deal with robust stability and numerical issues of variational analysis and optimization.
Quite recently [27], Rockafellar introduces novel generalized second-order derivatives, called quadratic
bundles, which are defined as collections of epigraphical limits of sequences of second-order subderivatives
of primal-dual pairs converging to the point in question, with the objective function being generalized twice
∗
Group of Analysis and Applied Mathematics, Department of Mathematics, Ho Chi Minh City University of Education,
Ho Chi Minh City, Vietnam. E-mail: khanhpd@[Link].
†
Department of Mathematics, Wayne State University, Detroit, Michigan, USA. E-mail: aa1086@[Link]. Research
of this author was partly supported by the US National Science Foundation under grant DMS-2204519, by the Australian
Research Council under Discovery Project DP-190100555, and by Project 111 of China under grant D21024.
‡
Department of Mathematics and Statistics, University of North Dakota, Grand Forks, North Dakota, USA. E-mail:
[Link].1@[Link].
§
Department of Mathematics, Wayne State University, Detroit, Michigan, USA. E-mail: vietle@[Link]. Research of
this author was partly supported by the US National Science Foundation under grant DMS-2204519.

1
differentiable along these sequences. His motivations come from the study of variational sufficiency in
problems of parametric composite optimization with applications to convergence of numerical algorithms
of proximal and augmented Lagrangian types.
In this paper, we explore the concept of quadratic bundles as the basis for broad upcoming applications.
Our first goal is to provide a systematic study of generalized twice differentiability, a concept that requires
both twice epi-differentiability and the derivative itself being a generalized quadratic form, which serves as
a crucial building block for our upcoming endeavors. Although the fundamental papers [21, 24] provide
much insight into these two concepts, the precise formulation and name are coined recently in [27]. We now
systematically investigate quadratic bundles, with a revision of this construction that operates properly
in the absence of the subdifferential continuity inherent in convex functions. As the cumulation of the
study below, prox-regular functions are shown to always have nonempty quadratic bundles, which testifies
to the importance and usefulness of this concept.
To achieve our goals in the general setting of prox-regular functions, the notions of Moreau envelopes and
associated proximal mappings are of paramount importance. Originally introduced by Moreau in [20],
proximal mappings of convex functions were viewed as generalizations of projections. The associated epi-
addition (infimal convolution) operation in Moreau envelopes has been also implemented but not yet as
an approximation and regularization tool for the function in question. To the best of our knowledge, this
crucial interpretation appeared for the first time in the subsequent works by Attouch and Wets for convex
and nonconvex functions [1, 2]. Moreau envelopes play a significant role in our developments below.
In [22], Poliquin and Rockafellar introduce the class of extended-real-valued prox-regular functions, which
by now has been recognized as the main class of functions in second-order variational analysis. This
class encompasses, in particular, C 1,1 -smooth functions and its lower-C 2 counterparts, strongly amenable
functions, convex and variationally ones, etc. It is verified in the same paper [22] that Moreau envelopes
of prox-regular functions are (under mild additional assumptions) of class C 1,1 around the reference point,
i.e., C 1 -smooth with locally continuous gradients. The recent paper [27] reveals certain connections
between the newly introduced quadratic bundles of functions and the Hessian bundles of the associated
Moreau envelopes, which are conveniently guaranteed to be nonempty with the C 1,1 -smoothness of the
envelopes. These facts and observations confirm the irreplaceable role of Moreau envelopes and the
underlying assumption of prox-regularity, which we employ and develop throughout our work.
The rest of the paper is organized as follows. Section 2 equips the reader with some basic material of
variational analysis and generalized differentiation that is broadly exploit in deriving the major results
below. Section 3 revisits the very recent unification of variational convexity and prox-regularity and their
connection with second-order subderivatives while deriving some new facts in this direction. Section 4
addresses generalized twice differentiable functions with establishing their important properties. This
study is continued in Section 5 by involving prox-regularity and Moreau envelopes. Here we obtain the
crucial characterization of generalized twice differentiability for prox-regular functions via classical twice
differentiability of Moreau envelopes. The obtained result serves as the central building block for the
subsequent sections. Section 6 offers a revised version of quadratic bundles for nonconvex functions and
presents new facts and instructive examples for this second-order generalized derivative construction.
Among the most significant results of this section is establishing the nonemptiness of quadratic bundles
associated with any extended-real-valued prox-regular function. Section 7 recaps the major ideas and
results achieved in the paper lists some directions of our future research.

2 Preliminaries from Variational Analysis

Throughout the paper, we use the standard terminology and notation of variational analysis and gener-
alized differentiation; see [15, 16, 30]. Recall that IRn stands for the n-dimensional space with the inner
product !·, ·" and the Euclidean (unless otherwise stated) norm (norm-2) written as # · #. The symbol
C
B(x̄, r) signifies the open ball centered at x̄ with radius r > 0. We write xk −
→ x̄ when xk → x̄ as k → ∞

2
with xk ∈ C for all k ∈ IN := {1, 2, . . .}. For a (linear) subspace L ⊂ IRn , recall the Pythagorean formula

#w#2 = #PL w#2 + #PL⊥ w#2 whenever w ∈ IRn , (2.1)

where the orthogonal projector PL is

! the self-adjoint linear operator whose range is L and whose kernel is
L⊥ := {w ∈ IRn | !w, x" = 0, x ∈ L .
Given now an extended-real-value function f : IRn → IR ∪ {∞} := IR, the domain and epigraph of f are
defined, respectively, by
" # ! " # !
dom f := x ∈ IRn # f (x) < ∞ , epi f := (x, α) ∈ IRn × IR # α ≥ f (x) .

The properness of f means that dom f ,= ∅, which we always assume. Recall that f is lower semicontinuous
(l.s.c.) at x̄ if lim inf f (x) ≥ f (x̄). This property holds on a set C if f is l.s.c. at any x̄ ∈ C.
x→x̄
The (general/basic, limiting, Mordukhovich) subdifferential of f : IRn → IR at x̄ ∈ dom f is defined by
$ # f (x) − f (xk ) − !vk , x − xk " %
#
∂f (x̄) := v ∈ IRn # ∃ xk → x̄, vk → v with lim inf ≥0 . (2.2)
x→xk #x − xk #

When f is convex, (2.2) reduces to the classical subdifferential of convex analysis, while otherwise ∂f (x̄)
is often nonconvex; e.g., we have ∂f (0) = {−1, 1} for the function f (x) = −|x|, x ∈ IR. Nevertheless,
the subdifferential ∂f satisfies comprehensive calculus rules for the general class of l.s.c. functions. This
calculus is based on variational/extremal principles and techniques of variational analysis; see [15, 16, 30].
Let F : IRn ⇒ IRm be a set-valued mapping/multifunction with the graph gph F := {(x, y) ∈ IRn ×
IRm | y ∈ F (x)}. The graphical derivative of F at (x̄, ȳ) ∈ gph F is
" # !
DF (x̄, ȳ)(u) := v ∈ IRn # (u, v) ∈ Tgph F (x̄, ȳ) , u ∈ IRn , (2.3)

where the tangent/contingent cone to a set Ω ⊂ IRd at z̄ ∈ Ω is defined by

" # Ω !
TΩ (z̄) := w ∈ IRd # ∃ zk −
→ x̄, tk ↓ 0 with (xk − x̄)/tk → w

Next we discuss the concept of epi-convergence for sequences of functions, which plays a crucial role in
defining first-order and second-order subderivatives; more details can be found in [1, 30]. To begin with,
recall the outer limit and inner limit of the set sequence {Ck } ⊂ IRn defined, respectively, by
" # !
Lim sup Ck := x # ∃ xk ∈ Ck and a subsequence {xkj } with xkj → x as j → ∞ ,
k→∞
" # !
Lim inf Ck := x ∈ IRn # ∃ xk ∈ Ck with xk → x as k → ∞ .
k→∞

It is said that {Ck } is convergent to some C (denoted as Ck → C) if

Lim sup Ck = Lim inf Ck =: Lim Ck = C.

x→∞ x→∞ x→∞

When F : IRn → IRm is norm-coercive, in the sense that #F (x)# → ∞ as #x# → ∞, the convergence
Ck → C implies that F (Ck ) → F (C); see [30, Theorem 4.26].
Now we recall epigraphical limits for sequences of functions and graphical limits for sequences of multi-
functions by following the book [30]. For a sequence of functions fk : IRn → IR, the lower epigraphical
limit e- liminf fk and the upper epigraphical limit e- limsup fk are defined via their epigraphs as
k→∞ k→∞
& ' & '
epi e- liminf fk = Lim sup(epi fk ), epi e- limsup fk = Lim inf (epi fk ).
k→∞ k→∞ k→∞ k→∞

3
When these limits agree, we say that (fk ) epigraphically converges to f and denote this by either e- lim fk =
k→∞
e e c
f , or fk −
→ f . If in addition −fk −
→ −f , then fk continuously converges to f , which is denoted by fk −
→f
n
and means that for all xk → x and x ∈ IR , we have fk (xk ) → f (x); see [30, Theorem 7.11]. An alternative
statement to determine the lower and upper epigraphical limits together with a characterization of epi-
convergence is given in the following result taken from [30, Proposition 7.2].

Proposition 2.1. Let {fk } be a sequence of functions on IRn , and let x ∈ IRn . Then we have
& ' " # !
e- liminf fk (x) = min α ∈ IR # ∃ xk → x, lim inf fk (xk ) = α ,
k→∞
& ' " # !
e- limsup fk (x) = min α ∈ IR # ∃ xk → x, lim sup fk (xk ) = α .
k→∞

Therefore, {fk } epigraphically converges to f if and only if at each point x it holds

(
lim inf fk (xk ) ≥ f (x) for every sequence xk → x,
lim sup fk (xk ) ≤ f (x) for some sequence xk → x.

In particular, if {fk } epigraphically converges to f , then for all x there exists a sequence xk → x such
that fk (xk ) → f (x) as k → ∞.

Epi-convergence is instrumental for the introduction of second-order subderivatives of extended-real-

valued functions; see [30, Definition 13.3]. Consider first the second-order difference quotients of f : IRn →
IR at x̄ ∈ dom f for v̄ ∈ IRn given as

f (x̄ + tw) − f (x̄) − t!v̄, w"

∆2t f (x̄|v̄)(w) := 1 2 whenever w ∈ IRn and t > 0
2 t

and then define and second-order subderivative of f at x̄ for v̄ by

& '
d2 f (x̄|v̄)(w) := e- liminf ∆2t f (x̄|v̄) (w) = lim inf ∆2t f (x̄|v̄)(w% ), w ∈ IRn . (2.4)
t↓0 t↓0
w " →w

Note that d2 f (x̄|v̄) may take the value −∞. When the epigraphical lower limit in (2.4) is actually a
full limit, we say that of f is twice epi-differentiable at x̄ for v̄. If in addition d2 f (x̄|v̄) is proper, then
f is properly twice epi-differentiable at x̄ for v̄. A major subclass of nonsmooth twice epi-differentiable
functions consists of parabolically regular ones; see [13, 14, 30]. Observe that a similar formula to (2.4)
for the upper epigraphical limit of ∆2t f (x̄|v̄) as t ↓ 0 is not available, and we usually have to appeal to
Proposition 2.1 providing the expression
& ' " # !
e- limsup ∆2t f (x̄|v̄) (w) = min α ∈ IR # ∃ tk ↓ 0, wk → w, lim sup ∆2tk f (x̄|v̄)(wk ) = α
t↓0

when calculating this part of twice epi-differentiability, see, e.g., the function in Example 4.6 below. In the
case of classical twice differentiability, second-order subderivatives reduce to quadratic forms associated
with the Hessian matrices at x̄ (see [30, Proposition 13.8]):

d2 f (x̄|∇f (x̄))(w) = !w, ∇2 f (x̄)w", w ∈ IRn .

To get d2 f (x̄|v̄)(w) ∈ IR, it is necessary that w belongs to the critical cone of f at x̄ for v̄ defined by
" # !
K(x̄, v̄) := w ∈ IRn # df (x̄)(w) = !v̄, w" , (2.5)

4
where the first-order subderivative of f at x̄ ∈ dom f is given by

f (x̄ + tw% ) − f (x̄)

df (x̄)(w) := lim inf ∆t f (x̄)(w% ) with ∆t f (x̄)(w% ) := , w ∈ IRn . (2.6)
t↓0 t
w " →w

Next we consider set-valued mappings Fk : IRn ⇒ IRm , k ∈ IN, and define for them the graphical outer
limit and graphical inner limit via the set convergence
& ' & '
gph g- limsup Fk := Lim sup(gph Fk ), gph g- liminf Fk := Lim inf (gph Fk ).
k→∞ k→∞ k→∞ k→∞

When these limits agree, we say that {Fk } graphically converges to some F denoted as F = g- lim Fk or
k→∞
g
Fk −
→ F . The usage of graphical convergence allows us represent the graphical derivative (2.3) as

F (x̄ + tw) − ȳ
DF (x̄|ȳ) = g- limsup ∆t F (x̄|ȳ), where ∆t F (x̄|ȳ)(w) := . (2.7)
t↓0 t

A multifunction F is proto-differentiable at x̄ for ȳ ∈ F (x̄) if the limit in (2.7) is a full graphical limit.

Proposition 2.2. If a single-valued mapping F : IRn → IRm is differentiable at x̄ ∈ IRn , then F is

proto-differentiable at x̄ for ȳ = F (x̄).

Proof. Assume that F : IRn → IRm is differentiable at x̄ ∈ IRn . Then we get by [30, Proposition 8.34]
that DF (x̄|ȳ) = ∇F (x̄). This gives us z̄ = ∇F (x̄)w̄ for any w̄ ∈ IRn and z̄ ∈ DF (x̄|ȳ)(w̄). By the
differentiability of F at x̄, for any tk ↓ 0 we have
F (x̄ + tk w̄) − F (x̄)
lim = ∇F (x̄)w̄.
k→∞ tk
F (x̄ + tk w̄) − F (x̄)
Letting zk := and wk := w̄ verifies the proto-differentiability of F at x̄ for ȳ.
tk
Finally in this section, we recall the notions of Moreau λ-envelopes and λ-proximal mappings associated
with proper l.s.c. functions f with a parameter λ > 0 defined by
$ 1 % " 1
eλ f (x) := inf n f (u) + #u − x#2 , Pλ f (x) := argmin f (u) + #u − x#2 }.
u∈IR 2λ u∈IR n 2λ

The function f is prox-bounded if there exists λ > 0 such that eλ f (x) > −∞ for some x ∈ IRn , which is
equivalent to saying that f is minorized by a quadratic function.

3 Variational s-Convexity and Second-Order Subderivatives

In the recent paper [28], Rockafellar unifies the variational convexity and prox-regularity of extended-real-
functions on finite-dimensional spaces into a single concept called variational s-convexity. Recall first that
f : IRn → IR is s-convex on IRn for a given number s ∈ IR if its quadratic shift f − 2s # · #2 is convex on the
entire space. This global property is called s-strong convexity when s > 0 and (−s)-weak convexity when
s < 0. Now we formulate the major notion.

Definition 3.1. Let s be any real number. A proper l.s.c. function f : IRn → IR is variationally s-convex
at x̄ for v̄ ∈ ∂f (x̄) if f is finite at x̄ and there exist an s-convex function f) as well as convex neighborhoods
U of x̄ and V of v̄ together with ε > 0 such that f) ≤ f on U and

(U × V ) ∩ gph ∂ f) = (Uε × V ) ∩ gph ∂f, f)(x) = f (x) at common elements (x, v),

5
where Uε := {x ∈ U | f (x) < f (x̄) + ε}. We can always select U := B(x̄, ε), V := B(v̄, ε) and then say
that f is variationally s-convex at x̄ for v̄ with corresponding radius ε > 0. When s ≤ 0, f is said to be
(−s)-level prox-regular at x̄ for v̄, while for s = 0 we say that f is variationally convex at x̄ for v̄. When
s > 0, f is called to be s-strongly variationally convex at x̄ for v̄.

Prox-regular functions have been around for a long time; see Section 1 and also the books [17, 31] in
infinite dimensions. On the other hand, the notion of variational convexity appears rather recently in [26]
in finite-dimensional spaces while being further studied, characterized, and applied in this framework in
[9, 10, 11, 17, 27]. Infinite-dimensional developments are presented in [8]. The novel unified version from
Definition 3.1 is utilized in [4, 28, 29].
Let us now discuss the notion of f -attentive ε-localizations of ∂f around a point. Given an l.s.c. function
f : IRn → IR, its subdifferential (2.2), and some number ε > 0, we say that a set-valued mapping
Tε : IRn ⇒ IRn is an f -attentive ε-localization of ∂f around (x̄, v̄) ∈ gph ∂f if
" # !
gph Tε = (x, v) ∈ gph ∂f # x ∈ B(x̄, ε), v ∈ B(v̄, ε), f (x) < f (x̄) + ε . (3.1)

By Definition 3.1, an r-level prox-regular function f at x̄ for v̄ agrees with an r-weakly convex function
f) on an f -attentive localization of ∂f around (x̄, v̄), while an (s-strongly) variational convex function
coincides with an (s-strongly) convex function in the similar manner. In light of this interpretation, we
get the following characterization of variationally s-convex functions f taken from [28, Theorem 1].

Theorem 3.2. Given level s ∈ IR and an extended-real-valued function f : IRn → IR, suppose that f is
finite and l.s.c. at x̄. Then f is variationally s-convex at x̄ for v̄ ∈ ∂f (x̄) if and only if there exist a
number ε > 0, a neighborhood U of x̄, and an f -attentive ε-localization of ∂f around (x̄, v̄) defined in
(3.1) such that for all x% ∈ U we have
s
f (x% ) ≥ f (x) + !v, x% − x" + #x% − x#2 whenever (x, v) ∈ gph Tε . (3.2)
2
gph ∂f
It is said [30, Definition 13.28] that f is subdifferentially continuous at x̄ for v̄ if for all (xk , vk ) −−−−−→ (x̄, v̄)
we have f (xk ) → f (x̄). The subdifferential continuity of f at x̄ for v̄ allows us to replace gph Tε by an
ordinary neighborhood of (x̄, v̄) on gph ∂f , i.e., the variationally s-convexity of f at x̄ for v̄ holds if there
exists ε > 0 such that for all (x, v) ∈ gph ∂f ∩ B((x̄, v̄), ε) we have
s
f (x% ) ≥ f (x) + !v, x% − x" + #x% − x#2 whenever x% ∈ B(x̄, ε).
2
In the next new result, important in what follows, we use the notation
" # !
x, v*) ∈ gph ∂f # x
gph Tε (x, v) := (* * ∈ B(x, ε), v* ∈ B(v, ε), f (*
x) < f (x) + ε .

Proposition 3.3. Let f : IRn → IR be a proper l.s.c. function, and let 0 < ε2 < ε1 . Then:

(i) For all (x, v) ∈ gph Tε2 (x̄, v̄), we have gph Tε1 −ε2 (x, v) ⊂ gph Tε1 (x̄, v̄).

(ii) Fix any s ∈ IR and assume that f is variationally s-convex at x̄ for v̄. Let ε1 be the corresponding
radius, i.e., (3.2) holds for all (x, v) ∈ gph Tε1 (x̄, v̄) and x% ∈ B(x̄, ε1 ). Then whenever (x, v) ∈
gph Tε2 (x̄, v̄), f is variationally s-convex at x for v with the radius ε1 − ε2 > 0.

Proof. (i) Fix (x, v) ∈ gph Tε2 (x̄, v̄) and get for any (* x, v*) ∈ gph Tε1 −ε2 (x, v) that x
* ∈ B(x, ε1 − ε2 ),
v* ∈ B(x, ε1 − ε2 ), which obviously yields x
* ∈ B(x̄, ε1 ) and v* ∈ B(v̄, ε1 ). Moreover, we have

x) < f (x) + ε1 − ε2 < [f (x̄) + ε2 ] + ε1 − ε2 = f (x̄) + ε1 ,

f (*

and hence gph Tε1 −ε2 (x, v) ⊂ gph Tε1 (x̄, v̄).

6
(ii) Assume that f : IRn → IR is variationally s-convex at x̄ for v̄ with corresponding radius ε1 . Pick
(x, v) ∈ gph Tε1 (x̄, v̄) and get by (i) that gph Tε1 −ε2 (x, v) ⊂ gph Tε1 (x̄, v̄). Take then any (* x, v*) ∈
gph Tε1 −ε2 (x, v) ⊂ gph Tε1 (x̄, v̄) and deduce from the variationally s-convexity of f at x̄ for v̄ with radius
ε1 > 0 the fulfillment of the inequality
s
f (x% ) ≥ f (* v , x% − x
x) + !* *" + #x% − x
*#2 whenever x% ∈ B(x̄, ε1 )
2
Since we obviously have B(x, ε1 − ε2 ) ⊂ B(x̄, ε1 ), it follows that
s
f (x% ) ≥ f (* v , x% − x
x) + !* *" + #x% − x
*#2 l for all x% ∈ B(x, ε1 − ε2 ).
2
Remembering that (* x, v*) was taken arbitrarily in gph Tε1 −ε2 (x, v), we arrive at the variational s-convexity
of f at x for v with the corresponding radius ε1 − ε2 .

The remaining part of this section addresses relationships between quadratic growth and second-order
subderivatives with applications to variational s-convexity of extended-real-valued functions.

Theorem 3.4. Let f : IRn → IR be an extended-real-valued l.s.c. function with x̄ ∈ dom f , and let v̄ ∈ IRn .
Consider the following assertions:

(i) There exists a neighborhood U of x̄ on which the second-order growth condition is satisfied
κ
f (x) ≥ f (x̄) + !v̄, x − x̄" + #x − x̄#2 , x ∈ U. (3.3)
2

(ii) We have the second-order subderivative estimate

d2 f (x̄|v̄)(w) ≥ µ#w#2 for all w ∈ IRn . (3.4)

Then implication (i) =⇒ (ii) holds with µ = κ, while the reverse one (ii) =⇒ (i) holds with µ > κ.

Proof. Assuming (i), take any (x̄, v̄) ∈ IRn × IRn , x̄ ∈ dom f , and a neighborhood U of x̄ on which (3.3)
is satisfied. We have x = x̄ + tw% ∈ U for any w ∈ IRn , any t > 0 sufficiently small, and any w% ∈ IRn
sufficiently close to w. This yields the estimate

f (x̄ + tw% ) − f (x̄) − !v̄, tw% "

1 2 ≥ κ#w% #2 .
2t

Passing to the limit on both sides above brings us to

f (x̄ + tw% ) − f (x̄) − !v, tw% "

d2 f (x̄|v̄)(w) = lim inf 1 2 ≥ lim inf κ#w% #2 = κ#w#2 ,
t↓0
2 t t↓0
w " →w "
w →w

which thus justifies (3.4) with µ = κ.

To verify the opposite implication, suppose that (ii) holds and that µ > κ. Then

d2 f (x̄|v̄)(u) ≥ µ for all u ∈ IRn with #u# = 1. (3.5)

Pick any u from (3.5) and deduce from definition (2.4) of the second-order subderivative that there exist
tu > 0 and a neighborhood Nu of u such that

∆2t f (x̄|v̄)(u% ) ≥ κ for all u% ∈ Nu and t ∈ (0, tu ).

7
Since the open sets Nu , #u# = 1 cover the compact set {u ∈ IRn | #u# = 1}, we find finitely many
u1 , . . . , um such that the latter set is covered by (Nui )i=1,...,m . Define
" !
t0 := min tu1 , . . . , tum

and observe that ∆2t f (x̄|v̄)(u) ≥ κ whenever #u# = 1 and t ∈ (0, t0 ). We have furthermore that
∆2t f (x̄|v̄)(0) = 0 and that
& w ' & w '
∆2t f (x̄|v̄)(w) = ∆2t f (x̄|v̄) #w# = #w#2 ∆2t(w( f (x̄|v̄) ≥ #w#2 κ
#w# #w#

for all w ∈ IRn \ {0} and all t > 0 with #tw# < t0 . Therefore,
κ
f (x̄ + tw) ≥ f (x̄) + !v̄, tw" + #tw#2 whenever w ∈ IRn , t > 0 with #tw# < t0 .
2
Denoting U := B(x̄, t0 ), we arrive at (3.3) and thus complete the proof of the theorem.

The following example shows that implication (ii)=⇒(i) of Theorem 3.4 fails if µ = κ.
κ 2
Example 3.5. Letting f (x) := x3 + x on IR with x̄ = 0 and v̄ = 0, we see that f %% (x) = 6x + κ and
2
d2 f (0|0)(w) = f %% (0). w2 = κw2 .

Thus (3.4) holds. On the other hand, (3.3) is equivalent to the existence of a neighborhood U of 0 such
that x3 ≥ 0 for all x ∈ U , which cannot happen.

Note that Theorem 3.4 is an extension of [30, Proposition 13.24(c)] giving us a characterization of strong
local minimizers of f when v̄ = 0 via second-order subderivatives. Contrary to the latter, we do not
require that v̄ ∈ ∂f (x̄) for this equivalence to hold, and our result explicitly establish a relationship
between moduli of the positive-definiteness of d2 f (x̄|v̄) of the strong local minimizer in question.
The next assessment establishes an important quadratic estimate of second-order subderivatives for vari-
ationally s-convex extended-real-valued functions.

Corollary 3.6. Given s ∈ IR, let f : IRn → IR be variationally s-convex at x̄ ∈ dom f for v̄ ∈ ∂f (x̄).
Then there exists an f -attentive ε-localization Tε of ∂f around (x̄, v̄) with some ε > 0 such that

d2 f (x|v)(w) ≥ s#w#2 for any (x, v) ∈ gph Tε and w ∈ IRn .

Proof. Assume that f is variationally s-convex at x̄ for v̄ for some s ∈ IR. It follows from Theorem 3.2
that there exist an f -attentive εlocalization Tε of ∂f around (x̄, v̄) and a neighborhood U of x̄ such that
s
f (x% ) ≥ f (x) + !v, x% − x" + #x% − x#2 whenever (x, v) ∈ gph Tε and x% ∈ U. (3.6)
2
For any (x, v) ∈ gph Tε , the set U is also a neighborhood of x, and so (3.6) is a second-order growth
condition around x. Applying Theorem 3.4 for such (x, v) tells us that

d2 f (x|v)(w) ≥ s#w#2 for all w ∈ IRn ,

which therefore completes the proof of the corollary.

8
4 Properties of Generalized Twice Differentiable Functions
In this section, we study the notion of generalized twice differentiability of extended-real-valued function,
the terms that has been recently coined by Rockafellar in [27]. This concept acts as a bridge between
merely twice-epi differentiability and classical twice differentiability, requiring the good behavior of second-
order subderivatives while not restricting the function too much. The history of this notion can be traced
back to the paper [21], where a characterization for global minimizers was established in Theorem 3.8
therein. Following [21, Definition 3.7], we have the following extension of ordinary quadratic forms. For
any matrix A ∈ IRn×n , consider the quadratic function

qA (x) := !x, Ax", x ∈ IRn .

Definition 4.1. A function q : IRn → (−∞, ∞] is called a generalized quadratic form if it is expressible
1
as q = qA + δL , where L is a linear subspace of IRn , and where A ∈ IRn×n is a symmetric matrix.
2
It follows from Definition 4.1 that q(0) = 0 and ∂q is a generalized linear mapping, i.e., gph ∂q is a subspace
of IRn × IRn . Generalized quadratic forms allow us to formulate the aforementioned notion from [27].

Definition 4.2. A function f : IRn → IR is generalized twice differentiable at x̄ for v̄ ∈ ∂f (x̄) if it is twice
epi-differentiable at x̄ for v̄ and its second subderivative d2 f (x̄|v̄) is a generalized quadratic form.

It is easy to see that the classical twice differentiability of f at x̄ yields the generalized one with

d2 f (x̄|∇f (x̄))(w) = !w, ∇2 f (x̄)w" for all w ∈ IRn

The next theorem shows that generalized twice differentiability may hold for (first-order) nonsmooth func-
tions and fully clarifies this issue for any norm function on IRn with calculating its second subderivative.

Theorem 4.3. Let f = # · # be a norm (not necessary Euclidean) function on IRn . Then we have:

(i) Whenever v ∈ ∂f (0), f is twice epi-differentiable at 0 for v and the corresponding second-order
subderivative is calculated by

d2 f (0|v) = δK(0,v) as v ∈ ∂f (0) (4.1)

via the critical cone K(0, v) of f at 0 for v taken from (2.5).

(ii) f is generalized twice differentiable at 0 for v if and only if v ∈ int ∂f (0).

Proof. It follows from (2.6) the first subderivative expressions

f (0 + tw% ) − f (0) t#w% #

df (0)(w) = lim inf = lim inf = lim inf #w% # = #w#, w ∈ IRn ,
t↓0 t t↓0 t t↓0
w " →w w " →w w " →w

which yield the critical cone representation

" # ! " # !
K(0, v) = w ∈ IRn # df (0)(w) = !v, w" = w ∈ IRn # #w# = !v, w" for any v ∈ ∂f (0). (4.2)

To verify (i), recall that the dual norm of # · #, denoted by # · #∗ , is defined by

" # !
#z#∗ := sup !z, x" # #x# = 1 .
" # !
We know that ∂f (0) = v ∈ IRn # #v#∗ ≤ 1 , which implies by the Cauchy-Schwarz inequality that

!v, w" ≤ #v#∗ #w# ≤ #w# for all v ∈ ∂f (0), w ∈ IRn . (4.3)

9
Therefore, whenever w ∈ IRn and t > 0, we get
f (0 + tw) − f (0) − t!v, w" #w# − !v, w"
∆2t f (0|v)(w) = 1 2 = 1 .
2t 2t

It follows from (4.3) that the family of functions {∆2t f (0|v)}t∈(0,∞) is nondecreasing as t ↓ 0, which yields
its epigraphical convergence by [30, Proposition 7.4(d)] and hence justifies the twice epi-differentiability
of f at 0 for any v ∈ ∂f (0). To prove (i), it remains verifying (4.1). Indeed, it follows from ∆2t f (0|v) ≥ 0
as t > 0 that d2 f (0|v) ≥ 0. Pick any w ∈ K(0, v) and deduce from Proposition 2.1 that

#w# − !v, w"

d2 f (0|v)(w) ≤ 1 = 0,
2t

which tells us that d2 f (0|v)(w) = 0 by (4.3). In the reverse direction, taking w ∈

/ K(0, v) gives us
#w# − !v, w" > 0, which amounts to saying that

#w% # − !v, w% "

d2 f (0|v)(w) = lim inf ∆2t f (0|v)(w% ) = lim inf 1 = ∞.
2t
t↓0 t↓0
w " →w w " →w

This brings us to (4.1) and completes the proof of (i).

To proceed with the proof of (ii), we obviously have int ∂f (0) = {v ∈ IRn | #v#∗ < 1}. Pick any v ∈ IRn
with #v#∗ < 1 and fix w ∈ K(0, v). It follows from (4.3) that

#w# = !v, w" ≤ #v#∗ #w#,

which yields (1 − #v#∗ ).#w# ≤ 0. Since 1 − #v#∗ > 0, the latter tells us that w = 0. Therefore, for
all v ∈ int ∂f (0), we get K(0, v) = {0}, which leads us to d2 f (0|v) = δ{0} . This readily justifies the
generalized twice differentiability of f is at 0 for v for any v ∈ int ∂f (0).
Now we show that for any v ∈ IRn with #v#∗ = 1, the cone K(0, v) cannot be a linear subspace and
hence the generalized twice differentiability of f = # · # at 0 for v fails due to (4.1). Indeed, take any
v ∈ IRn with #v#∗ = 1 and deduce from the dual norm definition the existence of x̄ such that #x̄# = 1
and #v#∗ = !v, x̄". Since #v#∗ = 1, it follows that

#x̄# = 1 = !v, x̄".

By (4.2), the latter amounts to x̄ ∈ K(0, v). Assuming by the contrary that K(0, v) is a linear subspace
ensures that −x̄ ∈ K(0, v), and hence

−# − x̄# = !v, x̄" = #x̄#.

This yields #x̄# = 0, which is a contradiction. Therefore, for each v ∈ IRn with #v#∗ = 1, the cone K(0, v)
cannot be a linear subspace, and thus f is not generalized twice differentiable at 0 for v.

On the other hand, the next theorem shows that generalized twice differentiability for C 1,1 functions
reduces to their classical twice differentiability at the point in question. The following technical result
taken from [7, Lemma A.11] is useful in what follows.

Lemma 4.4. Let f : IRn → IR be differentiable on the given interval [x1 , x2 ] ⊂ IRn , and let the gradient
∇f be Lipschitz continuous on that interval with constant L > 0. Then we have
L
|f (x2 ) − f (x1 ) − !∇f (x1 ), x2 − x1 "| ≤ #x2 − x1 #2 .
2
Here is the aforementioned theorem whose both parts are important for the subsequent applications.

10
Theorem 4.5. Let f : IRn → IR be differentiable on a neighborhood U of x̄ having the Lipschitz continuous
gradient ∇f on U with modulus L > 0. The following assertions hold:

(i) For all x ∈ U , we have the estimate

# 2 + , #
#d f x|∇f (x) (w)# ≤ L#w#2 whenever w ∈ IRn , (4.4)

which ensures, in particular, that d2 f (x|∇f (x)) is finite on IRn .

(ii) For any point x ∈ U , the generalized twice differentiability of f at x for ∇f (x) is equivalent to the
twice differentiability of f at x in the classical sense.

Proof. To verify (i), fix any vector w ∈ IRn and apply Lemma 4.4, which tells us that for all w% sufficiently
close to w and all t > 0 sufficiently small with x + tw% ∈ U , we get
# L
|f (x + tw% ) − f (x) − !∇f (x), tw% "# ≤ t2 #w% #2 .
2
This yields, for all such w% and t, the estimate
# 2 + ,
#∆ f x|∇f (x) (w% )| ≤ L#w% #2 .
t

Passing to the limits inferior when t ↓ 0, w% → w and using (2.4) bring us to

|d2 f (x|∇f (x))(w)| ≤ L#w#2 ,

which clearly implies that d2 f (x|∇f (x)) is finite.

To proceed with the proof of (ii), we only need to show that if f is generalized twice differentiable at some
x ∈ U for ∇f (x), then f is twice differentiable at x in the classical sense. Assuming that f is generalized
twice differentiable at x for ∇f (x), we find a symmetric matrix A and a linear subspace L such that

d2 f (x|∇f (x)) = qA + δL . (4.5)

Since f is of class C 1,1 around x, it is prox-regular everywhere near x by [30, Proposition 13.34]. It follows
from assertion (i) that d2 f (x|∇f (x)) is finite everywhere, i.e., the linear subspace L in (4.5) is the entire
IRn . Applying [22, Theorem 6.7] tells us that f has a quadratic expansion at x meaning that
1
f (y) = f (x) + !∇f (x), y − x" + d2 f (x|∇f (x))(y − x) + o(|y − x|2 )
2
1
= f (x) + !∇f (x), y − x" + !y − x, A(y − x)" + o(|y − x|2 ) for all y near x.
2
Combining finally [30, Corollary 13.42] with the differentiability f on a neighborhood of x, we conclude
that f is twice differentiable at x in the classical sense and thus complete the proof of the theorem.

Now we present two examples showing that the assumptions of Theorem 4.5 are essential for the conclusion.

Example 4.6. The conclusion of Theorem 4.5(ii) fails if f is C 1 -smooth but not of class C 1,1 around x̄.

Proof. Consider the function f : IR → IR defined by f (x) := |x|3/2 , x ∈ IR. Then f is C 1 -smooth around
x̄ = 0 but is not of class C 1,1 around that point. This function is obviously not twice differentiable at x̄ = 0
in the classical sense. Let us check that f is generalized twice differentiable at x̄ = 0 for v̄ := f (x̄) = 0 .
To calculate d2 f (0|0), we first get that

f (0 + tw) − f (0) − t.0.w |tw|3/2 |w|3/2

∆2t f (0|0)(w) = 1 2 = 1 2 = 1
√ for all w ∈ IRn and t > 0.
2t 2t 2 t

11
It follows from Proposition 2.1 that
- . - .
e- liminf ∆2t f (0|0) (0) = e- limsup ∆2t f (0|0) (0) = 0, (4.6)
t↓0 t↓0

which readily tells us by (2.4) that

d2 f (0|0)(1) = d2 f (0|0)(−1) = ∞.
The positive homogeneity of degree 2 of d2 f (0|0) ensures that d2 f (0|0) = δ{0} , which is a generalized
quadratic form on IR. Furthermore, observe the relationships
- . - .
e- liminf ∆2t f (0|0) (w) = e- limsup ∆2t f (0|0) (w) = ∞, w ∈ IR, (4.7)
t↓0 t↓0

which being combined with (4.6) and (4.7) establish twice epi-differentiability of f at 0 for 0.
Example 4.7. The conclusion of Theorem 4.5(ii) fails if the generalized twice differentiability therein is
replaced by merely twice epi-differentiability.
Proof. Consider the function f : IR → IR defined by f (x) := x2 sgn(x) with ∇f (x) = |x| for all x ∈ IR,
which is of class C 1,1 around x̄ = 0. To check that f is twice epi-differentiable at 0 for 0, we have
t2 w2 sgn(tw) − 0 − t · 0 · w
∆2t f (0|0)(w) = 1 2 = 2w2 sgn(w) for all t > 0 and w ∈ IRn .
2t
Observing that the function ∆2t f (0|0)(w) is continuous everywhere yields ∆2t f (0|0)(w% ) → ∆2t f (0|0)(w)
as w% → w. Apply Proposition 2.1 to get the expressions
- . - .
e- liminf ∆2t f (0|0) (w) = e- limsup ∆2t f (0|0) (w) = 2w2 sgn(w), w ∈ IR.
t↓0 t↓0

Thus ∆2t f (0|0)(w) epigraphically converges as t ↓ 0 to the function d2 f (0|0)(w) = 2w2 sgn(w), w ∈ IR,
which justifies the twice epi-differentiability of f at 0 for 0. On the other hand, d2 f (0|0) is not a generalized
quadratic form on IR because any generalized quadratic form q on IR must either be q(w) = aw2 for all
w ∈ IR with a ∈ IR, or q = δ{0} . Therefore, f is not generalized twice differentiable at 0 for 0.
The next proposition establishes the preservation of generalized twice differentiability for extended-real-
valued functions under twice differentiable additions.
Proposition 4.8. Let g : IRn → IR be l.s.c. with v̄ ∈ ∂g(x̄), and let f : IRn → IR be strictly differentiable
at x̄ ∈ IRn and twice differentiable at this point in the classical senses. The following assertions hold:
(i) For all w ∈ IRn , we have the equality
d2 (f + g)(x̄|∇f (x̄) + v̄)(w) = !w, ∇2 f (x̄)w" + d2 g(x̄|v̄)(w).

(ii) If g is generalized twice differentiable at x̄ for v̄, then the summation function f + g is generalized
twice differentiable at x̄ for ∇f (x̄) + v̄.
Proof. (i) Since f is strictly differentiable at x̄, it follows that ∇f (x̄) + v̄ ∈ ∂(f + g)(x̄) by [15, Proposi-
tion 1.107(ii)]. Since f is twice differentiable at x̄, we have
(f + g)(x̄ + tw% ) − (f + g)(x̄) − t!∇f (x̄) + v̄, w% "
d2 (f + g)(x̄|∇f (x̄) + v̄)(w) = lim inf 1 2
2t
t↓0
w " →w
/ 0
f (x̄ + tw% ) − f (x̄) − t!∇f (x̄), w% " g(x̄ + tw% ) − g(x̄) − t!v̄, w% "
= lim inf 1 2 + 1 2
2t 2t
t↓0
"
w →w
f (x̄ + tw% ) − f (x̄) − t!∇f (x̄), w% " g(x̄ + tw% ) − g(x̄) − t!v̄, w% "
= lim 1 2 + lim inf 1 2
2t 2t
t↓0 t↓0
w " →w "
w →w
2 2 n
= !w, ∇ f (x̄)w" + d g(x̄|v̄)(w), w ∈ IR .

12
(ii) By the generalized twice differentiability of g at x̄ for v̄, there exist a symmetric matrix A and a
subspace L of IRn such that d2 g(x̄|v̄) = qA + δL . It follows from (i) that

d2 (f + g)(x̄|∇f (x̄) + v̄)(w) = !w, (∇2 f (x̄) + A)w" + δL (w), w ∈ IRn . (4.8)

Since g is properly twice epi-differentiable at x̄ for v̄ and f is twice differentiable at x̄, we have that
e c
∆2t g(x̄|v̄) −
→ d2 g(x̄|v̄) and ∆2t f (x̄|∇f (x̄)) −
→ q∇2 f (x̄) as t ↓ 0. Applying [30, Theorem 7.46(b)] together
with (4.8) gives us as t ↓ 0 that
e
∆2t (f + g)(x̄|∇f (x̄) + v̄) = ∆2t f (x̄|∇f (x̄)) + ∆2t g(x̄|v̄) −
→ d2 (f + g)(x̄|∇f (x̄) + v̄)

thus verifying the twice epi-differentiability of f + g at x̄ for ∇f (x̄) + v̄. Therefore, f + g is generalized
twice differentiable at x̄ for ∇f (x̄) + v̄ as claimed.

5 Generalized Twice Differentiability via Moreau Envelopes

The main result of this section establishes the equivalence between generalized twice differentiability for
the broad class of (extended-real-valued) prox-regular and prox-bounded functions and its classical twice
differentiability counterpart for the associated Moreau envelopes. To achieve this goal, we present several
auxiliary statements of their own interest. The first lemma is extracted from [30, Proposition 13.37].
Lemma 5.1. Let f : IRn → IR be prox-bounded on IRn and r-level prox-regular at x̄ for v̄ ∈ ∂f (x̄) with
the corresponding radius ε > 0. Then for all λ ∈ (0, 1/r), there exists a neighborhood of x̄ + λv̄ on which:
(i) The proximal mappings Pλ f is single-valued and Lipschitz continuous with Pλ f (x̄ + λv̄) = x̄ + λv̄.

(ii) The Moreau envelope eλ f is of class C 1,1 and such that

∇eλ f = λ−1 [I − Pλ f ] = [λI + Tε−1 ]−1 , (5.1)

where Tε is the f -attentive ε-localization of ∂f around (x̄, v̄).

Moreover, the set-valued mapping Tε in (ii) can be chosen so that the set Uλ := rge (I + λTε ) serves for
all λ > 0 sufficiently small as a neighborhood of x̄ + λv̄ on which these properties hold.
The next technical observation is useful in what follows.
Lemma 5.2. For any set-valued mapping F : IRn ⇒ IRn and any λ ∈ IR, we have

(u, v) ∈ gph [λI + F −1 ]−1 ⇐⇒ (u − λv, v) ∈ gph F, i.e.,

gph F = Aλ [gph [λI + F −1 ]−1 ] for all x, y ∈ IRn ,

where Aλ : IRn × IRn → IRn × IRn is defined by

Aλ (x, y) := (x − λy, y),

which is an invertible linear mapping with A−1

λ (x, y) = (x + λy, y).

Proof. Based on the definitions, we have the equivalences

(u, v) ∈ gph [λI + F −1 ]−1 ⇐⇒ (v, u) ∈ gph [λI + F −1 ]

⇐⇒ (v, u − λv) ∈ gph F −1
⇐⇒ (u − λv, v) ∈ gph F,

which readily yield the other statements of the lemma.

13
Yet another lemma provides a desired ingredient for subsequent limiting procedures.
Lemma 5.3. Let f : IRn → IR be prox-regular at x̄ for v̄ ∈ ∂f (x̄). Take λ > 0 to be so small that all the
assertions in Lemma 5.1 hold; in particular, eλ f is of class C 1,1 around z̄ := x̄ + λv̄. For any sequence
zk → z̄, define the vectors

vk := ∇eλ f (zk ), xk := zk − λvk as k ∈ IN, (5.2)

Then we have the sequence of subgradients vk ∈ ∂f (xk ) such that

(xk , vk ) → (x̄, v̄), f (xk ) → f (x̄) as k → ∞.

Proof. Assuming that f : IRn → IR is prox-regular at x̄ for v̄ ∈ ∂f (x̄) with the corresponding radius
ε0 > 0, we fix any ε ∈ (0, ε0 ) and deduce from Lemma 5.1 that the representations in (5.1) hold on the
neighborhood Uλε := rge (I + λTε ) of z̄. Take a sequence zk → z̄ and construct vk , xk as in (5.2). We get
zk ∈ Uλε for all k sufficiently large and thus deduce from (5.1) that

vk = [λI + Tε−1 ]−1 (zk ),

which brings us to the equivalences

(vk , zk ) ∈ gph [λI + Tε−1 ] ⇐⇒ (vk , xk ) ∈ gph Tε−1 ⇐⇒ (xk , vk ) ∈ gph Tε .

Therefore, it holds for all k sufficiently large that

vk ∈ ∂f (xk ), xk ∈ B(x̄, ε), vk ∈ B(v̄, ε), f (xk ) < f (x̄) + ε.

Since ε > 0 was chosen arbitrarily small, we arrive at the claim conclusion.

The following proposition, extending with the modulus interplay the corresponding statement of [30,
Exercise 13.45], provides the equivalence between twice epi-differentiability of a prox-regular function and
the associated Moreau envelope.
Proposition 5.4. Let f : IRn → IR be prox-bounded on IRn and is r-level prox-regular at x̄ for v̄ ∈ ∂f (x̄).
Then for all λ ∈ (0, 1/r), the following assertions are equivalent:
(i) f is twice epi-differentiable at x̄ for v̄.

(ii) The Moreau envelope eλ f is twice epi-differentiable at x̄ + λv̄ for v̄.

Moreover, we have under the assumptions above that
-1 . -1 .& #1 '
#
eλ d2 f (x̄|v̄) = d2 eλ f x̄ + λv̄ # v̄ . (5.3)
2 2 2
Proof. Consider the shifted function

g(x) := f (x) − !v̄, x − x̄", x ∈ IRn

and observe by [30, Exercise 13.35] that g is r-level prox-regular at x̄ for 0. Since f is prox-bounded and
g differs from f by a linear term, g is also prox-bounded. Deduce now from by [6, Lemma 2.2] that
λ
eλ f (x̄ + λv̄ + x) = eλ g(x̄ + x) + #v̄#2 + !v̄, x" for all λ > 0, x ∈ IRn . (5.4)
2
It follows from the proof of Proposition 4.8 that g is twice epi-differentiable at x̄ for 0 with

d2 g(x̄|0) = d2 f (x̄|v̄). (5.5)

14
Similarly we get by the usage of (5.4) that the twice epi-differentiability of eλ f at x̄+λv̄ for v̄ is equivalent
to the twice epi-differentiability of eλ g at x̄ for 0 with the fulfillment of
-1 .& #1 ' -1 .
#
d2 eλ f x̄ + λv̄ # v̄ = d2 eλ g (x̄|0). (5.6)
2 2 2
Applying [30, Exercise 13.45] to g, and combining this with (5.5) and (5.6) establishes the equivalence
between (i) and (ii). The claimed equality (5.3) also follows from the above due to
- .& #
2 1
#1 ' -1 . -1 . -1 .
d eλ f x̄ + λv̄ ## v̄ = d2 eλ g (x̄|0) = eλ d2 g(x̄|0) = eλ d2 f (x̄|v̄) ,
2 2 2 2 2
which therefore completes the proof of the proposition.

If in addition the second-order subderivative d2 f (x̄|v̄) is a generalized quadratic form, i.e., f is generalized
twice differentiable at x̄ for v̄, then we establish below much stronger conclusion telling us that the
associated Moreau is not merely twice epi-differentiable but twice differentiable in the classical sense. To
proceed, we need to derive three more auxiliary results as follows.
Lemma 5.5. Let L ,= {0} be a linear subspace of IRn with dimension m ≤ n, let M ∈ IRn×n be a
positive-definite matrix, and let w ∈ IRn be a fixed vector. Given an orthogonal basis {v1 , . . . , vm } of L
and an m × n matrix B with columns vi , i = 1, . . . , m, we claim that the matrix B T M B is invertible and
B(B T M B)−1 B T w is a unique solution to the system
1
x ∈ L,
(5.7)
M x − w ∈ L⊥ .

Proof. To verify the invertibility of B T M B, take any u ∈ IRm such that B T M Bu = 0 and get
(Bu)T M Bu = 0. Since M is positive-definite, it follows that Bu = 0, which yields u = 0 by the
linear independence of the columns of B. To check further that B(B T M B)−1 B T w is a solution to (5.7),
observe that B(B T M B)−1 B T w ∈ L, which amounts to the fulfillment of the first inclusion in (5.7). For
the second one in (5.7), pick any u ∈ IRm and get

(Bu)T (M B(B T M B)−1 B T w − w) = uT (B T M B(B T M B)−1 B T w − B T w) = uT (B T w − B T w) = 0,

which shows that B(B T M B)−1 B T w satisfies (5.7). To verify the uniqueness of solutions to (5.7), take
two vectors x1 , x2 for which x1 − x2 ∈ L and M (x1 − x2 ) ∈ L⊥ . This yields (x1 − x2 )T M (x1 − x2 ) = 0
telling us that x1 − x2 = 0 by the positive-definiteness of M and thus completing the proof.

The next lemma concerns relationships between proto-differentiability of mappings and their graphical
derivatives under invertible linear operators.
Lemma 5.6. Let F and G be set-valued mappings from IRn to IRm such that gph G = A(gph F ), where
A : IRn × IRm → IRn × IRm is a linear invertible operator. Then given (x̄, v̄) ∈ IRn × IRm and (ū, w̄) :=
A(x̄, v̄), the proto-differentiability of F at x̄ for v̄ is equivalent to the proto-differentiability of G at ū for
w̄. In that case, we have the graphical derivative relationship

gph DG(ū|w̄) = A[gph DF (x̄|v̄)]. (5.8)

Proof. For all positive numbers t, we obviously get

gph G − (ū, w̄) A(gph F ) − A(x̄, v̄)
gph ∆t G(ū|w̄) = = .
t t
which readily yields the equalities
gph F − (x̄, v̄)
A−1 [gph ∆t G(ū|v̄)] = = gph ∆t F (x̄|v̄),
t

15
 gph ∆t G(ū|w̄) = A[gph ∆t F (x̄|v̄)].
Assuming now that G is proto-differentiable at ū for w̄ and taking the limits on both sides of the above
equation as t ↓ 0 bring us to the expression

gph DG(ū|w̄) = Lim A[gph ∆t F (x̄|v̄)]. (5.9)

t↓0

Check that A is norm-coercive. Indeed, the invertibility of A ensures that #A−1 # =

, 0 and

#x# = #A−1 Ax# ≤ #A−1 # · #A(x)# for all x ∈ IRn×m .

Therefore, the convergence #x# → ∞ obviously yields #A(x)# → ∞. The norm-coercivity of A−1 can be
checked similarly. Further, rewrite the right-hand side of (5.9) in the form
- .
gph DG(ū|w̄) = A Lim gph ∆t F (x̄|v̄)
t↓0

and conclude that the outer limit of gph ∆t F (x̄|v̄) exists as t ↓ 0, which means the proto-differentiability
of F at x̄ for v̄. Moreover, this gives us (5.8). Assuming on the other hand that F is proto-differentiable
at x̄ for v̄, we write gph F = A−1 (gph G) and deduce from the above that G is proto-differentiable at ū
for w̄ together with the fulfillment of the graphical derivative formula (5.8).

The final lemma concerns eigenvalue lower-boundedness of self-adjoint linear operators on subspaces.

Lemma 5.7. Let L be a subspace of IRn , let A : IRn → IRn be a self-adjoint linear operator, and let
σ ∈ IR. Impose on L the eigenvalue lower-boundedness

!w, Aw" ≥ σ#w#2 for all w ∈ L.

Then there exists a self-adjoint linear operator B : IRn → IRn such that the eigenvalue lower-boundedness
property of A on L ix extended to to the same property for B on the entire space, i.e..

!w, Bw" = !w, Aw" for all w ∈ L,

!w, Bw" ≥ σ#w#2 for all w ∈ IRn .

Proof. Recalling that both orthogonal projector operators PL and PL⊥ are self-adjoint, define the self-
adjoint linear operator B by
B := PL APL + σPL⊥ .
For any w ∈ IRn , we get the equalities

!w, Bw" = !w, PL APL w" + σ!w, PL⊥ w"

= !PL w, APL w" + σ!w, PL⊥ PL⊥ w"
= !PL w, APL w" + σ#PL⊥ w#2 .

Since PL⊥ (w) = 0 and PL (w) = w when w ∈ L, kit follows that !w, Bw" = !w, Aw". Moreover, by
PL w ∈ L and (2.1), the latter implies that

!w, Bw" = !PL w, APL w" + σ#PL⊥ w#2 ≥ σ#PL w#2 + σ#PL⊥ w#2 = σ#w#2 ,

which therefore verify our claims.

Now we are ready to establish the main result of this section.

16
Theorem 5.8. Let f : IRn → IR be prox-bounded on IRn and r-level prox-regular at x̄ for v̄ ∈ ∂f (x̄).
Then the following are equivalent for all λ ∈ (0, 1/r):
(i) f is generalized twice differentiable at x̄ for v̄.

(ii) eλ f is twice differentiable at x̄ + λv̄.

Proof. Fix λ ∈ (0, 1/r) and first verify implication (i)=⇒(ii). By the generalized twice differentiability
of f at x̄ for v̄, there exist a symmetric matrix A ∈ IRn×n and a subspace L ⊂ IRn such that

d2 f (x̄|v̄) = qA + δL . (5.10)

Since f is r-level prox-regular at x̄ for v̄ ∈ ∂f (x̄), it follows from Corollary 3.6 combined with (5.10) that

!x, Ax" ≥ −r#x#2 for all x ∈ L. (5.11)

Using (5.11), apply Lemma 5.7 and find a matrix B such that

!x, Bx" = !x, Ax" for all x ∈ L and !x, Bx" ≥ −r#x#2 for all x ∈ IRn , (5.12)

This allows us to rewrite (5.10) in the form

d2 f (x̄|v̄) = qB + δL , (5.13)

which being combined with (5.3) and (5.13) gives us

-1 .& #1 '
#
d2 eλ f (x̄ + λv̄|v̄)(w) = 2d2 eλ f x̄ + λv̄ # v̄ (w)
2 2
-1 .
= 2eλ d2 f (x̄|v̄) (w)
2
$1 1 %
= 2 inf n !x, Bx" + δL (x) + #x − w#2
x∈IR 2 2λ
$ 1 %
= inf !x, Bx" + #x − w#2 , w ∈ IRn .
x∈L λ
If L = {0}, we obviously have
1 1
d2 eλ f (x̄ + λv̄|v̄)(w) = !0, B0" + #0 − w#2 = #w#2 , w ∈ IRn ,
λ λ
so this is a quadratic form. Otherwise, consider the smooth function
1
ϕ(x) := !x, Bx" + #x − w#2 , x ∈ L,
λ
Combining λ < 1/r with (5.12) tells us that ϕ is strongly convex on L, and thus ϕ admits a unique global
minimizer x
*. Then we get by the subdifferential Fermat and elementary sum rules that
2
0 ∈ ∇ϕ(*
x) + NL (*
x) = 2B*
x+ x − w) + L⊥ ,
(*
λ
which can be rewritten in the form
2
2B*
x+ x − w) ∈ −L⊥ = L⊥ .
(*
λ
This ensures that x
* satisfies the system
(
x ∈ L,
& 1 ' 2 (5.14)
2 B + I x − w ∈ L⊥ .
λ λ

17
Take any orthogonal basis of L, and let C be the matrix whose columns are the vectors of the basis. Since
1/λ < r, it follows from (5.12) that B + λ1 I is a positive-definite matrix. By Lemma 5.5, x
* is the unique
solution to system (5.14) that is given by the formula
2 2 3 3−1
1 T 1
x
*= C C B+ I C C T w.
λ λ

Define now the matrices

1 & T& 1 ' '−1 T 1& '2
M := C C B+ I C C , Q := M BM + M −I
λ λ λ
and observe that both M and Q are symmetric and that x
* = M w. Then we get
1
d2 eλ f (x̄ + λv̄|v̄)(w) = ϕ(*
x) = !M w, BM w" + #M w − w#2 = !w, Qw".
λ
This tells us that if f is generalized twice differentiable at x̄ for v̄, then d2 eλ f (x̄ + λv̄|v̄)(w) is an ordinary
quadratic form. It follows from Proposition 5.4, due to the twice epi-differentiability of f at x̄ for v̄, that
eλ f is twice epi-differentiable at x̄ + λv̄ for v̄. Since d2 eλ f (x̄ + λv̄|v̄) is a quadratic form, eλ f is generalized
twice differentiable at x̄ + λv̄ for v̄. Remembering that f is prox-regular at x̄ for v̄, we conclude that eλ f
is of class C 1,1 around x̄ + λv̄. Therefore, Proposition 4.5 ensures that eλ f is twice differentiable at x̄ + λv̄.
To verify now the reverse implication (ii)=⇒(i), assume that eλ f is twice differentiable at x̄ + λv̄. As
follows from Proposition 5.4, f is twice epi-differentiable at x̄ for v̄. It remains to show that d2 f (x̄|v̄) is a
generalized quadratic form. Since ∇eλ f is differentiable at x̄ + λv̄, it is proto-differentiable at x̄ + λv̄ for
v̄ by Proposition 2.2. Then Lemma 5.1 tells is that

∇eλ f = [λI + Tε−1 ]−1 in a neighborhood U of x̄ + λv̄

for some mapping f -attentive ε-localization of ∂f around (x̄, v̄), and hence [λI + Tε−1 ]−1 is also proto-
differentiable at x̄ + λv̄ for v̄. Since ∇eλ f is differentiable at x̄ + λv̄, we deduce from [30, Example 8.34]
that the proto-derivative of ∇eλ f at x̄ + λv̄ for v̄ is given by

D[λI + Tε−1 ]−1 (x̄ + λv̄|v̄)(w) = D∇eλ f (x̄ + λv̄|v̄)(w) = ∇2 eλ f (x̄ + λv̄)w, w ∈ IRn . (5.15)

It follows now from Lemma 5.2 that

gph Tε = Aλ [gph [λI + Tε−1 ]−1 ]

with the mapping Aλ defined therein. Applying Lemma 5.6 with A := Aλ , F := [λI + Tε−1 ]−1 , and
G := Tε ensures that Tε is proto-differentiable at x̄ for v̄ and

gph DTε (x̄|v̄) = Aλ [gph D[λI + Tε−1 ]−1 ](x̄ + λv̄|v̄)]. (5.16)

By (5.15), gph D[λI + Tε−1 ]−1 ](x̄ + λv̄|v̄) is a linear subspace of dimension n in IRn × IRn . This implies
that Aλ [gph D[λI + Tε−1 ]−1 ](x̄ + λv̄|v̄)] is also a linear subspace of dimension n due to the invertibility of
Aλ . Combining the latter with (5.16), we conclude that gph DTε (x̄|v̄) is a n-dimensional linear subspace
of IRn × IRn . Applying finally [30, Proposition 13.40] tells us that
&1 '
∂ d2 f (x̄|v̄) = DTε (x̄|v̄),
2
+1 , 1
and so the set gph ∂ d2 f (x̄|v̄) is a linear subspace of dimension n. This means that d2 f (x̄|v̄) is a
2 2
generalized quadratic form and thus completes the proof of the theorem.

18
To conclude the section, we provide a uniform characterization of generalized twice differentiability along
f -attentive ε-localizations of ∂f around the reference point via the classical twice differentiability of the
associated Moreau envelopes. The uniformity here refers to the fact that we can fix a single sufficiently
small parameter for every points in the aforementioned f -attentive localization.

Corollary 5.9. Let f : IRn → IR be prox-bounded on IRn and r-level prox-regular at x̄ for v̄ ∈ ∂f (x̄).
Then for any small ε > 0, there exists an f -attentive ε-localization Tε of ∂f around (x̄, v̄) such that the
following assertions are equivalent for all λ ∈ (0, 1/r) and all (x, v) ∈ gph Tε :

(i) f is generalized twice differentiable at x for v.

(ii) eλ f is twice differentiable at x + λv.

Furthermore, we have for all (x, v) ∈ gph Tε that

-1 . -1 .& #1 '
#
eλ d2 f (x|v) = d2 eλ f x + λv # v . (5.17)
2 2 2
Proof. Due to the imposed prox-boundedness and prox-regularity assumptions, it follows from Propo-
sition 3.3 that there exists an f -attentive ε-localization Tε of ∂f around (x̄, v̄) such that f is r-level
prox-regular at x for v whenever (x, v) ∈ gph f . Applying Theorem 5.8 tells us that for all λ ∈ (0, 1/r),
f is generalized twice differentiable at x for v if and only if eλ f is twice differentiable at x + λv. Since f
is r-level prox-regular at x for v, the equality in (5.17) also follows from (5.3) in Theorem 5.8.

6 Quadratic Bundles
This section addresses the notion of quadratic bundles for extended-real-valued functions, which involves
the two main components: generalized twice differentiability of functions and the epigraphical limits of
the corresponding second-order epi-derivatives as generalized quadratic forms.
For any proper l.s.c, function f : IRn → IR, consider the set
" # !
Ωf := (x, v) ∈ gph ∂f # f is generalized twice differentiable at x for v . (6.1)

As an example, let f be the norm-2 function on IRn , which we already studied from the viewpoint of
generalized twice differentiability in Theorem 4.3. Using the calculations therein and taking into account
that the dual norm of norm-2 is itself give us the representation
" # ! " # !
Ωf = (x, ∇f (x)) # x ,= 0 ∪ (0, v) # #v# < 1 .

On the other hand, we have

" # ! " # !
gph ∂f = (x, ∇f (x)) # x ,= 0 ∪ (0, v) # #v# ≤ 1 .

This shows us that, in this particular example, the set Ωf is not necessarily closed while it is dense in
the graph gph ∂f . It follows from the results below that this density property holds for every l.s.c. convex
function. Moreover, we show that the set Ωf in (6.1) is locally dense in the graph of ∂f if f is prox-regular.
To proceed in this way, consider the Hessian bundle of f at a point x̄ ∈ dom f is defined by
2 " # !
∇ f (x̄) := H ∈ IRn×n # ∃xk → x̄ such that f is twice differentiable at xk and ∇2 f (xk ) → H . (6.2)
2
When f is of class C 1,1 around x̄, the Hessian bundle ∇ f (x̄) is a nonempty and compact set that consists
of symmetric matrices; [30, Theorem 13.52].

Theorem 6.1. Let f : IRn → IR be prox-regular at x̄ for v̄ ∈ ∂f (x̄), and let ε > 0. Then we have:

19
 (i) There is an f -attentive ε-localization Tε of ∂f around (x̄, v̄) such that Ωf ∩gph Tε is dense in gph Tε .

(ii) If in addition f is subdifferentially continuous at x̄ for v̄, then there exists a neighborhood W of
(x̄, v̄) such that the set Ωf ∩ W is dense in W ∩ gph ∂f .

(iii) If f is convex, then the aforementioned property becomes global, i.e., Ωf is dense in gph ∂f .
Proof. (i) By Proposition 3.3, there exists an f -attentive ε-localization Tε of ∂f around (x̄, v̄) such that
f is prox-regular at x for v whenever (x, v) ∈ gph Tε . Fix (x, v) ∈ gph Tε and deduce from the prox-
regularity of f at x for v by using Lemma 5.1 that there exists λ > 0 so small that eλ f is of class C 1,1
around x + λv. Furthermore, we have on a neighborhood of x + λv that

∇eλ f = [λI + (Tε% )−1 ]−1 , (6.3)

for an f -attentive ε-localization Tε% of ∂f around (x, v). It follows from Proposition 3.3 that we can shrink
2
gph Tε% to get gph Tε% ⊂ gph Tε . By [30, Theorem 13.52], the Hessian bundle ∇ eλ f (x + λv) is nonempty.
2
Pick any H ∈ ∇ eλ f (x + λv) and find a sequence zk → x + λv such that eλ f is twice differentiable in the
classical sense at zk and ∇2 eλ f (zk ) → H as k → ∞. Fix k ∈ IN and define

vk := ∇eλ f (zk ), xk := zk − λvk .

By Lemma 5.3, we have vk ∈ Tε% (xk ) ⊂ Tε (xk ) for all large k with vk → v and xk → x as k → ∞. Since
eλ f is twice differentiable at zk , Theorem 5.8 tells us that f is generalized twice differentiable at xk for vk
meaning that (xk , vk ) ∈ Ωf . Therefore, the sequence (xk , vk ) belongs to the set gph Tε ∩ Ωf for all large
k and converges to (x, v). This readily implies that the set gph Tε ∩ Ωf is dense in gph Tε as claimed.
(ii) Assume in addition that f : IRn → IR is subdifferentially continuous at x̄ for v̄. Combining this
with Proposition (3.3) allows us to find a neighborhood W of (x̄, v̄) such that f is prox-regular at x for v
(x, v) ∈ gph Tε . Then we come to the claimed conclusion by repeating the arguments in the proof of (i).
(iii) When f is convex, we can utilize a stronger version of Lemma 5.1, where (6.3) holds everywhere
with Tε% replaced by ∂f ; see [30, Theorem 2.26]. This brings us to the desired conclusion and therefore
completes the proof of the theorem.

Our next goal is to study epi-convergence of sequences of generalized quadratic forms. Observe that if
a sequence of linear subspaces Lk converges, then it converges to a linear subspace L. Indeed, it follows
from [30, Exercise 4.14 and Proposition 4.15] that L is a convex cone. Since Lk = −Lk for all k, we get
−L = L as claimed. The following proposition presents the facts needed in our subsequent analysis; its
second part is also mentioned in [29, Proposition 4.3].
Proposition 6.2. Let {qk } be a sequence of generalized quadratic forms. Then we have:
(i) We can extract a subsequence of {qk } that epigraphically converges to q ,≡ ∞.

(ii) Assume that {qk } epigraphically converges to q and there exists a number r ≥ 0 such that

qk (w) ≥ −r#w#2 , w ∈ IRn , (6.4)

for all sufficiently large k. Then q is a generalized quadratic form satisfying

q(w) ≥ −r#w#2 whenever w ∈ IRn .

Proof. (i) Since each function qk is a generalized quadratic form, we have qk (0) = 0 for all k ∈ IN. It
follows from Proposition 2.1 that
& '
e- liminf qk (0) ≤ lim inf qk (0) = 0.
k→∞ k→∞

20
Hence e- liminf qk ,≡ ∞, which shows that the sequence {qk } does not escape epigraphically to the horizon,
k→∞
i.e., epi qk ,→ ∅ by [30, Proposition 7.5]. This ensures by [30, Proposition 7.6] that {qk } contains a
subsequence that epigraphically converges to a function q ,≡ ∞.
(ii) Let us show that the assumptions on {qk } and q yield q(0) = 0. Indeed, for any wk → 0 we have
lim inf qk (wk ) ≥ lim inf (−r#wk #2 ) = 0,
k→∞ k→∞

which implies that qk (wk ) = 0 when wk = 0, k ∈ IN, and thus q(0) = 0 by Proposition 2.1. Define further
the sequence of generalized quadratic forms {pk } by
pk (w) := qk (w) + r#w#2 , w ∈ IRn .
We obviously have that all pk are generalized quadratic form, and it follows from (6.4) that pk (w) ≥ 0
whenever w ∈ IRn . Therefore, each pk is convex for large k. Applying further [30, Theorem 7.46(ii)]
ensures that the sequence {pk } epigraphically converges to q + r# · #2 . Furthermore, [30, Theorem 12.35]
tells us that {∂pk } graphically converges to ∂(q + r# · #2 ) = ∂q + 2rI. Since each set gph ∂qk is a linear
subspace in IRn × IRn for all k, the limiting one gph (∂q + 2rI) is also a subspace, and so is gph ∂q. By
q(0) = 0, this amounts to q being a generalized quadratic form.

Now we are in a position to present the definition of quadratic bundles of extended-real-valued functions,
which was first introduced in [27, p. 187] in order to work with convex functions. Here we revise this
concept when moving to the general case. Note that when losing the innate subdifferential continuity of
convex functions, we need to provide an appropriate modification to obtain a reasonable construction. In
the original work, Rockafellar takes all the generalized quadratic forms that are epigraphical limits of a
sequence of second-order subderivatives, where the primal-dual pair tends to a specified pair while making
sure that the function is generalized twice differentiable all along. In the process of this work, we learn
from the very recent manuscript [29], where a generalized version of quadratic bundles is proposed for
nonconvex functions with incorporating the f -attentiveness. This change is crucial to produce effective
results as given below.
Definition 6.3. Let f : IRn → IR be a proper extended-real-valued function. The quadratic bundle of f
at x̄ ∈ dom f for v̄ ∈ ∂f (x̄), denoted by quadf (x̄|v̄), is defined as the collection of generalized quadratic
Ωf
forms q for which there exists (xk , vk ) −−→ (x̄, v̄) such that f (xk ) → f (x̄) and the sequence of generalized
quadratic forms qk = 21 d2 f (xk |vk ) converges epigraphically to q.
The following remark collects some important observations about quadratic bundles.
Remark 6.4.
(i) When f is convex (or more generally, when f is subdifferentially continuous at x̄ for v̄), the revised
definition of quadf (x̄|v̄) agrees with the original construction by Rockafellar in [27, p. 187].
(ii) It is necessary to impose the additional condition f (xk ) → f (x̄) for the primal-dual sequence
{(xk , vk )}, because the original construction fails to capture the strong variational convexity in
the absence of subdifferential continuity. Various results and discussions in this direction are given
in our forthcoming manuscript [12].
(iii) In Definition 6.3, along with the requirements imposed on the sequences {(xk , vk )}, the functions
1 2
d f (xk |vk ) need to be epigraphically convergent as well. Actually, as long as we can find a sequence
2
Ωf
of (xk , vk ) such that (xk , vk ) −−→ (x̄, v̄) and f (xk ) → f (x̄), we can assume without loss of generality
1
that the functions d2 f (xk |vk ) epigraphically converge without any additional information. In other
2
words, the construction of quadratic bundles relies solely on the sequences {(xk , vk )} that fulfill the
imposed requirements. This observation follows from Proposition 6.2.

21
 Ωf
(iv) Suppose that f : IRn → IR is prox-regular at x̄ for v̄ ∈ ∂f (x̄). Then whenever (xk , vk ) −−→ (x̄, v̄) and
1
f (xk ) → f (x̄), we see that if the functions d2 f (xk |vk ) epigraphically converge, then they converge
2
to a generalized quadratic form q. In other words, quadf (x̄|v̄) contains all the possible epi-limits q
1 Ωf
of d2 f (xk |vk ) with (xk , vk ) −−→ (x̄, v̄) and f (xk ) → f (x̄) as k → ∞. To verify this, suppose that
2
f : IRn → IR is r-level prox-regular at x̄ for v̄ with some r ≥ 0. It follows from Proposition 3.3 that f
is r-level prox-regular at x for v whenever (x, v) belongs to the graph of an f -attentive ε-localization
Tε of ∂f around (x̄, v̄). Moreover, Corollary 3.6 tells us that

d2 f (x|v)(w) ≥ −r#w#2 whenever w ∈ IRn

Ωf
for all such pairs of (x, v). Take further any sequences (xk , vk ) −−→ (x̄, v̄) and f (xk ) → f (x̄) such
that the functions 12 d2 f (xk |vk ) epigraphically converge to q. Then (xk , vk ) ∈ gph Tε for all k large
enough. Consequently, we have

d2 f (xk |vk )(w) ≥ −r#w#2 , w ∈ IRn .

Applying Proposition 6.2(ii) confirms that q is a generalized quadratic form.

(v) Iff is generalized twice differentiable at x̄ for v̄, then

1 2
d f (x̄|v̄) ∈ quadf (x̄|v̄).
2
Indeed, let (xk , vk ) := (x̄, v̄) for all k. It is clear that f is generalized twice differentiable at xk for
vk and that the sequence of 12 d2 f (xk |vk ) converges epigraphically to 21 d2 f (x̄|v̄), which readily shows
that 12 d2 f (x̄|v̄) ∈ quadf (x̄|v̄).
If f is subdifferentially continuous at x̄ for v̄, then the requirement f (xk ) → f (x̄) in Definition 6.3 is
superfluous. However, the following example shows that our modification in the absence of subdifferential
continuity makes a difference. For convenience, we call the construction of quadratic bundles, where the
addition condition f (xk ) → f (x̄) is not imposed, the “old” quadratic bundle with the notation quado f .
This means that for any (x̄, v̄) ∈ gph ∂f , the “old” quado f (x̄|v̄) is defined as the collection of generalized
Ωf
quadratic forms q for which there exists (xk , vk ) −−→ (x̄, v̄) such that the sequence of generalized quadratic
forms qk = 21 d2 f (xk |vk ) converges epigraphically to q.
Example 6.5. Consider the function f : IR → IR defined by
1 2
x , x ≥ 0,
f (x) :=
1, x < 0.

Then ∂f (0) = (−∞, 0], f is variationally 2-strongly convex at 0 for 0 (in particular, it is prox-regular at
0 for 0) but not subdifferentially continuous at 0 for 0. Moreover, we have

quado f (0|0) = {q[0] , q[1] , δ{0} }, quadf (0|0) = {q[1] , δ{0} }. (6.5)

Proof. First we check that the function f is variationally 2-strongly convex by Definition 3.1. Indeed,
direct calculations bring us 
 {2x}, x > 0,
)
∂f (x) = ∂f (x) = (−∞, 0], x = 0,

{0}, x < 0.
Take ε := 1/4 together with the neighborhoods U := (−1/2, 1/2) of 0 and V := (−1/2, 1/2) of 0. Then
" # ! " # !
(Uε × V ) ∩ gph ∂f = (0, v) # v ∈ (−1/2, 0] ∪ (u, 2u) # u ∈ (0, 1/4) . (6.6)

22
Consider further the function f) : IR → IR defined by
1 2
) x , x ≥ 0,
f (x) := 2
x − x, x < 0.
It is straightforward to check that

 {2x}, x > 0,
∂)f)(x) = ∂ f)(x) = [ − 1, 0], x = 0,

{2x − 1}, x < 0.
Consequently, we have the expression
" # ! " # !
(U × V ) ∩ gph ∂ f) = (0, v) # v ∈ (−1/2, 0] ∪ (u, 2u) # u ∈ (0, 1/4) . (6.7)
It follows from (6.6) and (6.7) that
(U × V ) ∩ gph ∂ f) = (Uε × V ) ∩ gph ∂f
for the chosen ε, U , and V . Moreover, for all x ≥ 0 we get f)(x) = f (x) = x2 and f) ≤ f on U .
Since f)(x) − x2 is convex, the function f)(x) is 2-convex, and thus f is variationally 2-strongly convex by
definition. On the other hand, f is not subdifferentially continuous at 0 for 0. To see this, take a sequence
gph ∂f
{(xk , 0)} with xk ↑ 0, which fulfills the condition (xk , 0) −−−−−→ (0, 0) while f (xk ) ,→ f (0) as k → ∞.
To proceed with the verification of our claims in this example, now we provide the calculations of the
quadratic bundle and “old” quadratic bundle of f at 0 for 0. Let us first calculate d2 f (x|v) for all pairs
(x, v) ∈ gph ∂f . We see that

 2 − 2tvw
2 f (0 + tw) − f (0) − vtw  t2
, w < 0,
∆t f (0|v)(w) = 1 2 = 2vw
2t

 2w2 − , w ≥ 0,
t
for all v ∈ (−∞, 0], all t > 0, and all w ∈ IRn . It easily follows that
1
2 2 % δ{0} (w), v < 0,
d f (0|v)(w) = lim inf ∆t f (0|v)(w ) =
t↓0 δ[0,∞) (w) + 2w2 , v = 0,
w " →w

whenever w ∈ IRn . Observe that f is twice differentiable at all x ∈ IR \ {0}, and so

1
2w2 , x > 0,
d2 f (x|f % (x))(w) =
0, x < 0,
for all x ∈ IR \ {0} all w ∈ IR, and all w ∈ IR. Next we determine the set Ωf from (6.1), which contains all
the pairs (x, v) ∈ gph ∂f where f is generalized twice differentiable at x for v. To this end, observe that
f is not generalized twice differentiable at 0 for 0 because d2 f (0|0) is not a generalized quadratic form.
Since f is twice differentiable at all x ,= 0, f is also generalized twice differentiable at x for f % (x) for all
x ,= 0. When v < 0, Proposition 2.1 tells us that
& '
e- limsup ∆2t f (0|v) (w) = δ{0} (w), w ∈ IRn ,
t↓0

which entails the twice epi-differentiability of f at 0 for such v. Since the function d2 f (0|v) is also a
generalized quadratic form, f is generalized twice differentiable at 0 for v for all v < 0. Therefore, the set
Ωf in (6.1) is calculated by
" # ! " # !
Ωf = (0, v) # v < 0 ∪ (x, f % (x)) # x ,= 0 .
Ωf
To determine further the “old” quadratic bundles, pick q ∈ quado f (0|0) and find (xk , vk ) −−→ (0, 0) and
1 2 e
d f (xk |vk ) −
→ q as k → ∞. Consider the following three possibilities:
2

23
 (i) If xk = 0 for infinitely many indexes of k, then there exists a subsequence km such that xkm = 0 for
all m, and so vkm < 0 for all m. Consequently,
1 2 e
d f (xkm |vkm ) = δ{0} →
− δ{0} ,
2
and hence we have q = δ{0} .

(ii) If xk = 0 for finitely many indexes and there are infinitely many indexes k with xk > 0, take a
1 e
subsequence {km } such that xkm > 0. Then vkm = f % (xkm ) and d2 f (xkm |vkm ) = q[1] −
→ q[1] , which
2
therefore yield q = q[1] .

(iii) If xk ≥ 0 for finitely many indexes and xk < 0 for infinitely many indexes, then there exists a
1 e
subsequence {km } such that xkm < 0 for all m. Then vkm = f % (xkm ) and d2 f (xkm |vkm ) = q[0] −
→
2
q[0] , which verify that q = q[0] .

Combining all the above allows us to conclude that

quado f (0|0) ⊂ {δ{0} , q[0] , q[1] }.

It is straightforward to check the reverse inclusion, which shows that the first part of (6.5) holds. Finally,
note that if we impose in addition the condition f (xk ) → f (0), then it follows that xk ≥ 0 for all large k.
Then we continue with the arguments similar to the parts (i) and (ii), which brings us to the inclusion
" !
quadf (0|0) ⊂ δ{0} , q[1] .

The opposite inclusion can be checked easily, which readily verifies the second part of (6.5).

Our next result establishes a sum rule for quadratic bundles.

Theorem 6.6. Assume that f : IRn → IR is C 2 -smooth around x̄ and that g : IRn → IR is proper l.s.c.
Then for any v̄ ∈ ∂g(x̄), we have ∇f (x̄) + v̄ ∈ ∂(f + g)(x̄) and
1
quad(f + g)(x̄|∇f (x̄) + v̄) = q∇2 f (x̄) + quadg(x̄|v̄).
2
In particular, the quadratic bundle of f is calculated by
$1 % $1 %
quadf (x̄|∇f (x̄)) = d2 f (x̄|∇f (x̄)) = q∇2 f (x̄) .
2 2
Proof. We easily get (see, e.g., [16, Proposition 1.30]) that ∇f (x̄) + v̄ ∈ ∂(f + g)(x̄). Pick

q ∈ quad(f + g)(x̄|∇f (x̄) + v̄)

Ωf +g e
and find (xk , vk ) −−−→ (x̄, ∇f (x̄) + v̄) with f (xk ) + g(xk ) → f (x̄) + g(x̄) and 12 d2 (f + g)(xk |vk ) −
→ q
2
as k → ∞. Since vk ∈ ∂(f + g)(xk ), the function f is C -smooth around xk for large k, and so we get
vk − ∇f (xk ) ∈ ∂g(xk ). Remembering that f + g is generalized twice differentiable at xk for vk and that
f is twice differentiable at xk for large k, we apply Proposition 4.8 to conclude that g = (f + g) + (−f )
is generalized twice differentiable at xk for vk − ∇f (xk ) and that
1
d2 g(xk |vk − ∇f (xk )) = d2 (f + g)(xk |vk ) − q∇2 f (xk ) . (6.8)
2

24
 Ωg e c
→ q and − 12 q∇2 f (xk ) −
−→ (x̄, v̄) as k → ∞. Since d2 (f + g)(xk |vk ) −
Hence (xk , vk − ∇f (xk )) − → − 21 q∇2 f (x̄) ,
it follows from [30, Theorem 7.46] that

e 1
d2 g(xk |vk − ∇f (xk )) −
→ q − q∇2 f (x̄) .
2
Furthermore, we see that g(xk ) → g(x̄) and thus deduce from Definition 6.3 that q− 21 q∇2 f (x̄) ∈ quadg(x̄|v̄).
In other words, this means that
1
q ∈ q∇2 f (x̄) + quadg(x̄|v̄),
2
which readily brings us to the inclusion
1
quad(f + g)(x̄|v̄ + ∇f (x̄)) ⊂ q∇2 f (x̄) + quadg(x̄|v̄). (6.9)
2
Take q such that q ∈ 12 q∇2 f (x̄) + quadg(x̄|v̄) telling us that

1
q − q∇2 f (x̄) ∈ quadg(x̄|v̄). (6.10)
2
Representing g = (−f ) + (f + g), we deduce from (6.10) and (6.9) that
1 1
q − q∇2 f (x̄) ∈ − q∇2 f (x̄) + quad(f + g)(x̄|v̄ + ∇f (x̄)),
2 2
The latter can be rewritten in the form
1
q 2 + quadg(x̄|v̄)" ⊂ quad(f + g)(x̄|v̄ + ∇f (x̄)),
2 ∇ f (x̄)
which completes verifying the claimed sum rule.

It is noteworthy that if f is merely twice differentiable at x̄ in the classical sense (but not C 2 -smooth),
then it does not guarantee that quadf (x̄|∇f (x̄)) be a singleton. Moreover, the quadratic bundle can
consist of uncountable many elements. The following example justifies this observation.

Example 6.7. Consider the function f : IR → IR defined by

1 4
x sin(1/x), x ,= 0,
f (x) :=
0, x = 0.

Then f is twice differentiable at 0 in the classical sense, but

" !
quadf (0|0) = q[a] | a ∈ [−1/2, 1/2] . (6.11)

Proof. While f % (0) = f %% (0) = 0, for all x ,= 0 we have

+ ,
f % (x) = x2 − cos(1/x) + 4x sin(1/x) , f %% (x) = −6x cos(1/x) + (−1 + 12x2 ) sin(1/x).

Furthermore, it follows from the construction of f that

d2 f (x|f % (x))(w) = f %% (x)w2 for all x ∈ IR and w ∈ IR.

As f %% is bounded around 0, for any sequence {xk } we have that the epi-convergence of the quadratic
functions fk (w) = 21 f %% (xk )w2 is equivalent to their pointwise convergence. Indeed, assume that fk (w) =

25
1 %%
2 f (xk )w epigraphically converges to f . Fix any w ∈ IRn and find by Proposition 2.1 a sequence wk → w
2

such that fk (wk ) → f (w). Consequently, we have

|fk (w) − f (w)| ≤ |fk (w) − fk (wk )| + |fk (wk ) − f (w)|

1# #
= #f %% (xk )# .|wk − w|.|wk + w| + |fk (wk ) − f (w)| → 0,
2
which entails the pointwise convergence of fk to f . Reversely, suppose that the sequence of fk (w) =
1 %% 2 %%
2 f (xk )w pointwise converges to f . Choose any w ,= 0 to see that f (xk ) → a for some a ∈ IR and thus
f (w) = 12 aw2 . It is easily to deduce from Proposition 2.1 fk epigraphically converges to f . Therefore, all
the epigraphical limits of such sequences must be quadratic forms q[a] , where 12 f %% (xk ) → a.
To proceed further, fix any a ∈ [−1/2, 1/2] and take xk → 0 such that sin(1/xk ) = −2a for all k. Then

f %% (xk ) = −6xk cos(1/xk ) + (−1 + 12x2k ) sin(1/xk ) → 2a,

which shows that the function q[a] belongs to quadf (0|0). Pick any q ∈ quadf (0|0) and deduce from the
twice differentiable of f in the classical sense on IR and the boundedness of f %% around 0 that there exists
a sequence xk → 0 such that the sequence of qk = 12 q[f "" (xk )] epigraphically converges to q[a] , a ∈ IR. It
follows from our observation above that 12 f %% (xk ) → a. Therefore,
+ ,
2|a| = lim sup |f %% (xk )| ≤ lim sup 6|xk | · | cos(1/xk )| + 12x2k .| sin(1/xk )| + | sin(1/xk )| ≤ 1,

and so a ∈ [−1/2, 1/2]. This confirms that the quadratic bundle of f at 0 for 0 is calculated by (6.11).

As shown by Example 6.7, the merely twice differentiability of a function everywhere is not sufficient
to conclude that the quadratic bundle reduces to the quadratic form associated with the Hessian. To
reach such a conclusion in (6.11), we converted the epi-convergence of quadratic forms into the pointwise
convergence of the corresponding Hessians. This suggests that there might be a connection between the
quadratic bundle and the Hessian bundle defined as in (6.2) for classes of functions where the second-order
subderivative d2 f (x|v) is finite. Indeed, the next theorem reveals the relationship between the Hessian
bundle of f and the quadratic bundles of f for functions of class C 1,1 .
Theorem 6.8. Let f : IRn → IR be an l.s.c. function with x̄ ∈ dom f and v̄ ∈ ∂f (x̄). Suppose that f is
subdifferentially continuous at x̄ for v̄. Then we have
$1 # %
# 2
quadf (x̄|v̄) ⊃ qH # H ∈ ∇ f (x̄) . (6.12)
2
If moreover f is of class C 1,1 around x̄ then (6.12) holds as an equality.
2
Proof. Pick H ∈ ∇ f (x̄) and find a sequence {xk } such that f is twice differentiable at xk in the classical
sense for all k, xk → x̄, and ∇2 f (xk ) → H as k → ∞. This implies that {q∇2 f (xk ) } epigraphically converges
to qH . Since f is subdifferentially continuous at x̄ for v̄, the condition f (xk ) → f (x̄) automatically satisfies.
It follows from [30, Proposition 13.8] that 12 d2 f (xk |∇f (xk )) = 21 q∇2 f (xk ) , and so

1
qH ∈ quadf (x̄|v̄) as k → ∞.
2
Ωf
Now assume that f is of class C 1,1 around x̄. Taking any q ∈ quadf (x̄|∇f (x̄)), find a sequence (xk , vk ) −−→
e
(x̄, ∇f (x̄)) such that 12 d2 f (xk |vk ) −
→ q, which ensures that f is generalized twice differentiable at xk for
vk whenever k ∈ IN. Proposition 4.5 tells us that f is twice differentiable at xk in the classical sense, and
then we get by [30, Proposition 13.8] that
1 2 1
d f (xk |vk )(w) = !w, ∇2 f (xk )w" for all w ∈ IRn .
2 2

26
The C 1,1 property of f around x̄ yields the boundedness of ∇2 f around this point, while [30, Theo-
rem 13.51] ensures that ∇2 f (xk ) is symmetric. Also it easily follows from Proposition 2.1 that q is finite
everywhere. Since functions of class C 1,1 are prox-regular, we deduce from Remark 6.4(ii) that q is an
ordinary quadratic form denoted by 12 qH , where the matrix H is symmetric. It remains to show that
2 e
{∇2 f (xk )} converges to H, which ensures that H ∈ ∇ f (x̄). Indeed, since 12 q∇2 f (xk ) −
→ 12 qH , we deduce
from Proposition 2.1 that there exists a sequence wk → w with
1 1
!wk , ∇2 f (xk )wk " = q∇2 f (xk ) (wk ) → qH (w) = !w, Hw", w ∈ IRn .
2 2
To this end, observe that

!wk , ∇2 f (xk )wk " − !w, Hw" = !wk − w, ∇2 f (xk )(wk + w)" + !w, (∇2 f (xk ) − H)w".

Since the sequence {∇2 f (xk )} is bounded, we have that

|!wk − w, ∇2 f (xk )(wk + w)"| ≤ #wk − w# · #wk + w#.#∇2 f (xk )# → 0.

Combining the latter with the the convergence !wk , ∇2 f (xk )wk " → !w, Hw" brings us to

!w, (∇2 f (xk ) − H)w" → 0 for all w ∈ IRn .

Choosing w := ei , i = 1, . . . , n, tells us that all the entries on the diagonal of ∇2 f (xk ) − H converge to 0.
Select finally w := ei + ej for i, j = 1, . . . , n, we see that all other entries also converge to 0 as k → ∞,
which therefore completes the proof of the theorem.

Remark 6.9. The result in (6.12) does not hold in general if we remove the subdifferential continuity of
2
f at x̄ for v̄. Indeed, taking the function f in Example 6.5 gives us ∇ f (0) = {0, 2}, which yields
$1 # % " !
# 2
qH # H ∈ ∇ f (x̄) = q[0] , q[1] ,
2
while we have q[0] ∈
/ quadf (0|0).
The next example shows that when f is not of class C 1,1 , the inclusion in (6.12) could be strict, even if f
is C 1 -smooth and strictly convex.
Example 6.10. Consider f : IR → IR defined by
1 2
x , x < 0,
f (x) :=
x3/2 , x ≥ 0,

with x̄ = 0 and ∇f (0) = 0. Then f is C 1 -smooth everywhere and strictly convex. The quadratic bundle
of f at 0 for 0 is calculated by " !
quadf (0|0) = δ{0} , q[1] ,
2
and the Hessian bundle is ∇ f (0) = {2}. Therefore, (6.12) holds as a strict inclusion in this case.
Proof. It is clear that f is C 1 -smooth everywhere with
1
% 2x, x ≤ 0,
∇f (x) = f (x) = 1/2
3/2x , x > 0,

but not of class C 1,1 around x̄ = 0 and not twice differentiable at this point. The strict convexity of f
follows from the fact that f % is strictly increasing on IR. The Hessian at all nonzero points are given by
1
2 %% 2, x < 0,
∇ f (x) = f (x) = −1/2 (6.13)
3/4x , x > 0.

27
 2
To find ∇ f (0), take any sequence xk → 0 such that ∇2 f (xk ) converges to a finite number. It easily
follows that xk < 0 for all large k, since ∇2 f (xk ) → ∞ for any xk ↓ 0. Thus the limit of ∇2 f (xk ) is 2. It
2 2
follows from the above that 2 ∈ ∇ f (0), and we conclude in fact that ∇ f (0) = {2}.
Let us now verify the generalized twice differentiability of f at 0 for 0. We have for all t > 0 that

f (0 + tw) − f (0) − t0w  2w2 , w ≤ 0,
2 3/2
∆t f (0|0)(w) = 1 2 = 2w
2t
 , w > 0.
t1/2
Passing to the limits as t ↓ 0 gives us
1
2 2w2 , w ≤ 0,
d f (0|0)(w) = lim inf ∆2t f (0|0)(w% ) =
t↓0 ∞, w > 0.
w " →w

Since d2 f (0|0) is not a generalized quadratic form, f is not generalized twice differentiable at 0 for 0.
However, f is twice differentiability of f in the classical (and hence generalized) for all x ,= 0, and hence
the set Ωf defined in (6.1) is given by
" # !
Ωf = (x, ∇f (x)) # x ,= 0 .

Ωf
Take any q ∈ quadf (0|0) and find (xk , vk ) −−→ (0, 0) such that the sequence of 12 d2 f (xk |vk ) epigraphically
converges to q. Then we have that xk → 0 with xk ,= 0 for all k. Consider the following two possibilities:

(i) There exists a subsequence {xkm } such that xkm < 0 for all m. Then we have the functions

1 2 1
d f (xkm |vkm ) = · q[2] = q[1] ,
2 2
which epigraphically converge as m → ∞ to the function q[1] , and so q = q[1] .

(ii) There exists no subsequence {xkm } with xkm < 0 for all m, and so xk > 0 for all large k. Let us
show that the functions
1 2 3
d f (xk |vk )(w) = √ w2
2 4 xk
epigraphically converge to q = δ{0} . If w ,= 0, then any sequence wk → w satisfies

3
lim d2 f (xk |vk )(wk ) = lim √ wk2 = ∞.
k→∞ k→∞ 4 xk

Appealing to Proposition 2.1 yields q(w) = ∞ for all w ,= 0, and thus

" # !
q(0) = min α ∈ IR # ∃ wk → 0 with lim inf d2 f (xk |vk )(wk ) = α = 0.

We get that q[1] and δ{0} are all possible candidates for q ∈ quadf (0|0). It is easy to check that these
functions indeed belong to quadf (0|0). Therefore, the formula for quadf (0|0) is verified.

In the rest of this section, we demonstrate that quadratic bundles provide efficient tools to study prox-
regular functions by showing that the functions from this class are always twice differentiable with respect
to quadratic bundles.

Theorem 6.11. Let f : IRn → IR be prox-regular at x̄ for v̄ ∈ ∂f (x̄). Then the quadratic bundle
quadf (x̄|v̄) is nonempty.

28
Proof. If f : IRn → IR is prox-regular at x̄ for v̄ ∈ ∂f (x̄), then Lemma 5.1 allows us to fix a small
λ > 0 so that the associated Moreau envelope eλ f is of class C 1,1 around z̄ := x̄ + λv̄. Consequently, the
2 2
set ∇ eλ f (z̄) is nonempty. Pick any H ∈ ∇ eλ f (z̄) and find a sequence zk → z̄ such that eλ f is twice
differentiable at zk with zk → H for all k ∈ IN. Define the vectors

vk := ∇eλ f (zk ), xk := zk − λvk whenever k ∈ IN.

gph ∂f
Lemma 5.3 tells us that (xk , vk ) −−−−−→ (x̄, v̄) and f (xk ) → f (x̄) as k → ∞. Furthermore, we deduce from
Corollary 5.9, by the twice differentiability of eλ f at zk = xk +λvk , that f is generalized twice differentiable
1
at xk for vk with large k. Remark 6.4(i) allows us to extract a subsequence of the functions d2 f (xk |vk )
2
that epigraphically converges to some function q. Observe that we keep the convergence properties along
gph Ωf
the aforementioned subsequence: (xkm , vkm ) −−−−−→ (x̄, v̄) and f (xkm ) → f (x̄) as m → ∞. It follows from
Remark 6.4(ii) that q is a generalized quadratic form. Finally, Definition 6.3 confirms that q ∈ quadf (x̄|v̄),
and so we are done with the proof of the theorem.

The last example here demonstrates that the conclusion of Theorem 6.11 fails if f is merely C 1 -smooth
and not prox-regular.

Example 6.12. Let f : IR → IR be given by f (x) := −|x|3/2 , x ∈ IR. Then f is C 1 -smooth around x = 0
and f % (0) = 0 but quadf (0|0) = ∅.
38 3
Proof. We get that f % (x) = − |x|, that f is not twice differentiable at 0, and that f %% (x) = − 8
2 4 |x|
when x ,= 0. This brings us to the formula
3
d2 f (x|f % (x))(w) = − 8 w2 for all x ,= 0 and w ∈ IR.
4 |x|

Moreover, we get for t > 0 and w ∈ IR that

f (0 + tw) − f (0) − [Link] −3|w|3/2

∆2t f (0|0)(w) = 1 2 = .
2t 2t1/2

It easily follows therefore that

d2 f (0|0)(w) = lim inf ∆2t f (0|0)(w% ) = −∞, w ∈ IR.

t↓0
w " →w

In this way, we arrive at the representation

" # !
Ωf = (x, f % (x))# x ,= 0 .

Assume further that q ∈ quadf (0|0). Then q is a generalized quadratic form, and there exists a sequence
Ωf 1 e
(xk , vk ) −−→ (0, 0) such that d2 f (xk |vk ) −
→ q; in particular, xk ,= 0 for all k. We intend to show that
2
there exists w ∈ IR such that q(w) = −∞. Indeed, for any w ,= 0 and all wk → w it follows that
& 3 '
lim d2 f (xk |vk )(wk ) = lim − 8 wk2 = −∞.
4 |xk |

Employing Proposition 2.1 brings us to q(w) = −∞, which is a contradiction. Therefore, the quadratic
bundle quadf (0|0) is empty.

29
7 Conclusions and Future Research
This paper contributes to the investigation of generalized twice differentiability and quadratic bundles of
extended-real-valued functions. These recently introduced second-order notions belong to the primal-dual
realm of variational analysis with great potential for applications. We reveal new properties of generalized
twice differentiable functions, establish their crucial characterization via classical twice differentiability
of the associated Moreau envelopes, and apply them to the study of quadratic bundles. The obtained
results on generalized twice differentiability of extended-real-valued functions lead us to establishing the
fundamental property of quadratic bundles asserting their nonemptiness for the broad class of prox-regular
functions, which are the most important in second-order variational analysis.
Our future research will concentrate on applications of the developed tools of generalized twice differen-
tiability and quadratic bundles to new primal-dual characterizations of (strong) variational convexity of
extended-real-valued functions and tilt stability of local minimizers with applications to (strong) varia-
tional sufficiency in optimization and numerical algorithms of the first and second orders. We also plan
to proceed with the study of infinite-dimensional extensions of the developed methods and results.

References
[1] A. Attouch, Variational Convergence for Functions and Operators, Pitman, London, 1984.

[2] H. Attouch and R. J-B. Wets, A convergence theory for saddle functions, Trans. Amer. Math. Soc.
280 (1983), 1–41.

[3] D. Drusvyatskiy, B. S. Mordukhovich and T. T. A. Nghia, Second-order growth, tilt stability, and
metric regularity of the subdifferential, J. Convex Anal. 21 (2014), 1165–1192.

[4] H. Gfrerer, On second-order variational analysis of variational convexity of prox-regular functions,

arXiv:2408.13795 (2024).

[5] H. Gfrerer and B. S. Mordukhovich, Complete characterizations of tilt stability in nonlinear pro-
gramming under weakest qualification conditions, SIAM J. Optim. 25 (2015), 2081–2119.

[6] W. L. Hare and R. A. Poliquin, Prox-regularity and stability of the proximal mapping, J. Convex
Anal. 14 (2007), 589–606.

[7] A. F. Izmailov and M. V. Solodov, Newton-Type Methods for Optimization and Variational Problems,
Springer, New York, 2014.

[8] P. D. Khanh, V. V. H. Khoa, B. S. Mordukhovich and V. T. Phat, Variational and strong variational
convexity in infinite-dimensional variational analysis, SIAM J. Optim. 34 (2024), 2756–2787.

[9] P. D. Khanh, V. V. H. Khoa, B. S. Mordukhovich, and V.T. Phat, Second-order subdifferential opti-
mality conditions in nonsmooth optimization, to appear in SIAM J. Optim. (23025) arXiv:2312.16277.

[10] P. D. Khanh, B. S. Mordukhovich and V. T. Phat, Coderivative-based Newton methods in structured

nonconvex and nonsmooth optimization, arXiv:2403.04262 (2024).

[11] P. D. Khanh, B. S. Mordukhovich and V. T. Phat, Variational convexity of functions and variational
sufficiency in optimization, SIAM J. Optim. 33 (2023). 1121–1158.

[12] P. D. Khanh, B. S. Mordukhovich, V. T. Phat and L. D. Viet, Characterizations of variational

convexity and tilt stability via quadratic bundles, preprint (2024).

30
[13] A. Mohammadi, B. S. Mordukhovich and M. E. Sarabi, Parabolic regularity in geometric variational
analysis, Trans. Amer. Soc. 374 (2021), 711–763.

[14] A. Mohammadi and M. E. Sarabi, Twice epi-differentiability of extended-real-valued functions with

applications to composite optimization, SIAM J. Optim. 30 (2020), 2379–2409.

[15] B. S. Mordukhovich, Variational Analysis and Generalized Differentiation, I: Basic Theory, Springer,
Berlin, 2006.

[16] B. S. Mordukhovich, Variational Analysis and Applications, Springer, Cham, Switzerland, 2018.

[17] B. S. Mordukhovich, Second-Order Variational Analysis in Optimization, Variational Stability, and

Control: Theory, Algorithms, Applications, Springer, Cham, Switzerland, 2024.

[18] B. S. Mordukhovich and T. T. A. Nghia, Second-order characterizations of tilt stability with appli-
cations to nonlinear programming, Math. Program 149 (2015), 83–104.

[19] B. S. Mordukhovich and R. T. Rockafellar, Second-order subdifferential calculus with applications

to tilt stability in optimization, SIAM J. Optim. 22 (2012), 953–986.

[20] J. J. Moreau, Fonctions convexes duales et points proximaux dans un espace hilbertien, C. R. Acad.
Sci. Paris 255 (1962), 2897–2899.

[21] R. A. Poliquin and R. T. Rockafellar, Generalized Hessian properties of regularized nonsmooth

functions, SIAM J. Optim. 6 (1996), 1121–1137.

[22] R. A. Poliquin and R. T. Rockafellar, Prox-regular functions in variational analysis, Trans. Amer.
Math. Soc. 348 (1996), 1805–1838.

[23] R. A. Poliquin and R. T. Rockafellar, Tilt stability of a local minimum, SIAM J. Optim. 8 (1998),
287–299.

[24] R. T. Rockafellar, Maximal monotone relations and the second derivatives of nonsmooth functions,
Ann. Inst. H. Poincaré: Anal. Non Linéaire 2 (1985), 167–184.

[25] R. T. Rockafellar, First and second-order epi-differentiability in nonlinear programming, Trans.

Amer. Math. Soc. 307 (1988), 75–108.

[26] R. T. Rockafellar, Variational convexity and the local monotonicity of subgradient mappings, Vietnam
J. Math. 47 (2019), 547–561.

[27] R. T. Rockafellar, Augmented Lagrangians and hidden convexity in sufficient conditions for local
optimality, Math. Program 198 (2023), 159–194.

[28] R. T. Rockafellar, Variational convexity and prox-regularity, to appear in J. Convex Anal.;

[Link] rtr/[Link].

[29] R. T. Rockafellar, Derivative tests for prox-regularity and the modulus of convexity,
[Link] rtr/[Link].

[30] R. T. Rockafellar and R. J-B. Wets, Variational analysis, Springer, Berlin, 1998.

[31] L. Thibault. Unilateral Variational analysis in Banach spaces, World Scientific, Singapore, 2023

31