Abstract
In Hilbert space setting we prove local lipchitzness of projections onto parametric polyhedral sets represented as solutions to systems of inequalities and equations with parameters appearing both in left- and right-hand sides of the constraints. In deriving main results we assume that data are locally Lipschitz functions of parameter and the relaxed constant rank constraint qualification condition is satisfied.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
1 Introduction
Continuity of metric projections of a given \({\bar{v}}\) onto moving subsets have already been investigated in a number of instances. In the framework of Hilbert spaces, the projection \(P_{C}({\bar{v}})\) of \({\bar{v}}\) onto closed convex sets \(C,C^\prime \), i.e., solutions to optimization problems
are unique and Hölder continuous with the exponent 1/2 in the sense that there exists a constant \(\ell _{H}>0\) with
where \(d_\rho (\cdot ,\cdot )\) denotes the bounded Hausdorff distance (see [2] and also [8, Example 1.2]).
In the case where the sets, on which we project a given \({\bar{v}}\), are solution sets to systems of equations and inequalities, the problem Proj is a special case of a general parametric problem
where \(x\in {{\mathcal {H}}}\), \(p\in {{\mathcal {D}}}\subset {{\mathcal {G}}}\), \({{\mathcal {H}}}\)-Hilbert space, \({{\mathcal {G}}}\)-metric space, \(\varphi _i:\ {{\mathcal {D}}}\times {{\mathcal {H}}}\rightarrow {\mathbb {R}}\), \(i\in \{0\}\cup I_1\cup I_2\). There exist numerous results concerning continuity of solutions to problem (Par) in finite dimensional settings, see e.g., [6, 18, 24, 27] and the references therein. In a recent paper by Mordukhovich and Nghia [23], in the finite-dimensional setting, the Hölderness and the Lipschitzness of the local minimizers to problem Par with \(I_{1}=\emptyset \) are investigated for \(C^{2}\) functions \(\varphi _{i}\), \(i\in I_{2}\), under Mangasarian–Fromowitz (MFCQ) and constant rank (CRCQ) constraints qualifications.
Let \({{\mathcal {H}}}\) be a Hilbert space and let \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a nonempty set of a normed space \({{\mathcal {G}}}\), \(p\in {{\mathcal {D}}}\) and \(v\in {{\mathcal {H}}}\). We consider the norm topology induced on \({{\mathcal {D}}}\) by the topology of space \({{\mathcal {G}}}\), i.e., U is an open set in \({{\mathcal {D}}}\) if \(U={{\mathcal {D}}}\cap U^\prime \), where \(U^\prime \) is open in \({{\mathcal {G}}}\) (see e.g. [11]).
We consider the following parametric optimization problem
where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset \), \(I_1=\emptyset \vee \{1,\dots ,q\}\), \(I_2=\emptyset \vee \{q+1,\dots ,m\}\) are locally Lipschitz on \({{\mathcal {D}}}\).
When \(C(p)\ne \emptyset \) for \(p\in {{\mathcal {D}}}\), problem (M(v,p)) is uniquely solvable for any \(v\in {{\mathcal {H}}}\) and the solution P(v, p) to problem (M(v,p)) is the projection of v onto C(p) i.e.
Our aim is to prove local Lipschitzness of the projection mapping \(P:\ {{\mathcal {H}}}\times {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), given by (1) at an arbitrary fixed \(({\bar{v}},{\bar{p}})\). The following example shows that even if the functions \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\) are globally Lipschitz, the projection onto C(p), \(p\in {{\mathcal {D}}}\), given by (3), may not be continuous (in the strong topology). For other examples see e.g. [31].
Example 1
Let \(p\in {\mathbb {R}}\), \({\bar{p}}=0\), \({\bar{v}}=(-1,-1)\) and
The projection of \(v=(v_1,v_2)\) from a neighbourhood of \({\bar{v}}\) onto C(p), for p close to \({\bar{p}}\) is equal to
Hence, \(P(\cdot ,\cdot )\) is not continuous at \(({\bar{p}},{\bar{v}})\).
Our analysis is based on a recent results of [25] concerning Lipschitzness (and Hölderness) of solutions to a class of parametric variational inclusions.
Essential part of our considerations is based on the relaxed constant rank constraint qualification (RCRCQ) introduced in [19] and investigated in [3, 17, 20]. According to our knowledge, no result is known in the literature, in which this particular constraint qualification condition is used in the context of stability of solutions to parametric problems (Par) with \(I_1\ne \emptyset \). Moreover, we assume only local Lipschitzness of the right-hand side functions \(f_i\), \(i\in I_1\cup I_2\) and left-hand side functions \(g_i\), \(i\in I_1\cup I_2\).
Observe, that, in general, the existing continuity-type results for solutions of problem (Par) are representation-dependent in the sense that e.g. MFCQ condition is representation-dependent (see Example 2). Observe that RCRCQ (Definition 2) is also representation-dependent. We take this fact into account by introducing the concept of equivalent representation (Definition 4) and the concept of equivalent stable representation (Definition 5). In Theorem 5 we show that the under assumption (H1) the existence of a suitable equivalent representation is necessary for the continuity of projections onto sets C(p), \(p\in {{\mathcal {D}}}\), given by (3).
The organization of the paper is as follows. In Theorem 1 of Sect. 2 we recall Theorem 6.5 of [25] in the form which corresponds to our settings. Theorem 1 provides sufficient and necessary conditions for the estimate (25) which is stronger than local Lipschitzness of projection \(P(\cdot ,\cdot )\) (see (II) of Theorem 1). For convenience of the reader we provide the proof of the sufficiency part of Theorem 1. In Sect. 3 we recall the relaxed constant rank qualification (RCRCQ) and the results concerning Lipschitz-likeness of parametrized constrained sets C(p), \(p\in {{\mathcal {D}}}\), given by (3). In Sect. 4 we investigate Lagrange multipliers of problem (M(v,p)) under RCRCQ. In Sect. 5 we introduce the concept of equivalent stable representation of sets C(p), \(p\in {{\mathcal {D}}}\), given by (3). Main results of this section are Corollary 1 and Theorem 5. Section 6 contains the main result of the present paper (Theorem 6) together with a number of corollaries referring to several particular cases of problem (M(v,p)). Section 7 concludes.
2 Underlying Facts
In the Hilbert space setting, P(v, p) given by (1) is characterized as solution to the parametric variational inequality
where N(x; C(p)) stands for the normal cone (in the sense of convex analysis) to the set C(p) at \(x\in C(p)\) i.e.,
Equivalently,
which is the classical characterisation of projection in Hilbert spaces.
Local Lipschitzness of solutions to general parametric variational inequalities has been recently investigated by Mordukhovich, Nghia and Pham in [25, Theorem 6.5].
The formulation of Theorem 6.5 of [25] cast to our problem is given in Theorem 1.
From the point of view of applications it is also interesting to investigate the particular case of problem (M(v,p)) with \(v\equiv {\bar{v}}\), i.e.,
where \({\bar{v}}\in {{\mathcal {H}}}\), i.e. the problem (M(p)) does not depend on parameter v.
When \(g_i(p)\equiv g_i\in {{\mathcal {H}}}\), \(i\in I_1\cup I_2\), the sets C(p) take the form
and the stability of the respective variational system (PV(v,p)) has been investigated in [30].
For any multifunction \({{\mathcal {F}}}:\ X\rightrightarrows Y\) its domain and graph are defined as
An important assumption of Theorem 1, is Lipschitz-like behaviour of the multifunction\({{\mathcal {C}}}: {{\mathcal {P}}}\rightarrow {{\mathcal {H}}}\), \({{\mathcal {C}}}(p):=C(p)\). Below we recall this concept.
Definition 1
[21, Definition 1.40] Let X, Y be normed spaces. Let \({{\mathcal {F}}}:\ X\rightrightarrows Y\) be a multifunction with \(\text {dom }\,{{\mathcal {F}}}\ne \emptyset \). Given \(({\bar{u}},{\bar{y}})\in \text {gph}\,{{\mathcal {F}}}\), we say that \({{\mathcal {F}}}\) is Lipschitz-like (pseudo-Lipschitz, has the Aubin property) around \(({\bar{u}},{\bar{y}})\) with modulus \(\ell \ge 0\) if there are neighbourhoods \(U({\bar{u}})\) of \({\bar{u}}\) and \(V({\bar{y}})\) of \({\bar{y}}\) such that
where B(0, 1) is the open unit ball in Y.
This property is crucial in investigation of parametric problems, e.g., in variational systems [12], critical point set [14, 26]. It is extensively studied in recent monographs [9, 21, 22, 28].
Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {\mathcal {H}}\) be a set-valued mapping defined as \({{\mathbb {C}}}(p):=C(p)\),
where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset ,\) \(I_1=\emptyset \vee \{1,\dots ,q\}\), \(I_2=\emptyset \vee \{q+1,\dots ,m\}\) are locally Lipschitz on \({{\mathcal {D}}}\). The set C(p) is the feasible solution set of problem (M(v,p)).
In view of [25, Example 6.4], the fact [25, Lemma 6.2] applied to problem (PV(v,p)) takes the following form.
Proposition 1
[25, Lemma 6.2] Let \({\bar{x}}=P({\bar{v}},{\bar{p}})\), \({\bar{v}}\in {{\mathcal {H}}}\), \({\bar{p}}\in {{\mathcal {D}}}\). If \({{\mathbb {C}}}\) is Lipschitz-like around \(({\bar{p}},{\bar{x}})\), then there exist constants \(\kappa _0,\ell ^0>0\) and neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) that the estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\).
Let us note, that in view of Lemma 6.2 of [25], we have \(\kappa _0=1-\lambda r\), where in our case \(\lambda =1\) and \(r=0\) (Lemma 5.2 of [25] remain true for \({{\mathcal {R}}}=r=0\)), i.e., (4) takes the form
For problem (M(v,p)), or the equivalent problem (PV(v,p)) considered in the present paper, Theorem 6.5 of [25] takes the following form.
Theorem 1
[25, Theorem 6.5] Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\in {{\mathcal {H}}}\) and \({\bar{x}}=P({\bar{v}},{\bar{p}})\). Suppose that
-
(A)
\({{\mathbb {C}}}\) is Lipschitz-like around \(({\bar{p}},{\bar{x}})\).
The following conditions are equivalent.
-
(I)
The graphical subdifferential mapping \(Gr:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\times {{\mathcal {H}}}\) defined as
$$\begin{aligned} Gr(p):=\{ (x,x^\prime )\ |\ x\in C(p),\ x^\prime \in N(x;C(p)) \}=\text {gph}\,N(\cdot ;C(p)) \end{aligned}$$(6)is Lipschitz-like around \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\).
-
(II)
There exist neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) such that the estimate
$$\begin{aligned} \begin{aligned} \Vert (v_1-v_2)-2[P(v_1,p_1)-P(v_2,p_2)]\Vert \le \Vert v_1-v_2\Vert +\ell ^0\Vert p_1-p_2\Vert \end{aligned} \end{aligned}$$holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\) with some positive constant \(\ell ^0\).
For convenience of the Reader, we provide a direct proof of (I) \(\implies \) (II) of Theorem 1.
Proof
By Proposition 1, there exist constant \(\ell ^0>0\) and neighbourhoods \(V({\bar{v}})\), \(Q({\bar{p}})\) such that the estimate
holds for all \((v_1,p_1),(v_2,p_2)\in V({\bar{v}})\times Q({\bar{p}})\). Moreover, by (I), there exist neighbourhoods \(Q_1({\bar{p}})\subset Q({\bar{p}})\), \(U_1({\bar{x}})\), \(V_1({\bar{v}}-{\bar{x}})\) for which \(v+u\in V({\bar{v}})\) whenever \((u,p,v)\in U_1({\bar{x}})\times Q_1({\bar{p}})\times V_1({\bar{v}}-{\bar{x}})\) such that
holds for all \(p_1,p_2\in Q_1({\bar{p}})\), where \(\ell _{Gr}>0\) is a constant (here B(0, 1) is the unit ball in \({{\mathcal {H}}}\times {{\mathcal {H}}}\)). By (5), we have
Hence
There exist neighbourhoods \(U_2({\bar{x}}), Q_2({\bar{p}}),V_2({\bar{v}})\), such that \(U_2({\bar{x}})\times Q_2({\bar{p}})\times V_2({\bar{v}})\subset U_1({\bar{x}})\times Q_1({\bar{p}})\times V({\bar{v}})\) and \(P(V_2({\bar{x}}),Q_2({\bar{p}}))\subset U_2({\bar{x}})\), and
for all \((u,p,v)\in U_2({\bar{x}})\times Q_2({\bar{p}})\times V_2({\bar{v}})\). Now pick \((v_1,p_1),(v_2,p_2)\in V_2({\bar{v}})\times Q_2({\bar{p}})\) and define \(u_1:=P(v_1,p_1)\in U_2({\bar{x}})\) and \(u_2:=P(v_2,p_2)\in U_2({\bar{x}})\). Therefore, we have \(v_1^\prime :=v_1-u_1\in N(u_1;C(p_1))\cap V_1({\bar{v}}-{\bar{x}})\), i.e., \((u_1,v_1^\prime )\in Gr(p_1)\cap (U_1({\bar{x}})\times V_1({\bar{v}}-{\bar{x}}))\). By (8), there is \((u,v)\in Gr(p_2)\) satisfying
Define \(v^\prime := u+v \in u + N(u;C(p_2))\), i.e., \(u=P(v^\prime , C(p_2))\). Then
Hence, \(V^\prime ({\bar{v}}) \subset V({\bar{v}})\) by choosing \(Q_2({\bar{p}})\) sufficiently small. By (7), for pairs \((v^\prime ,p_2)\) and \((v_2,p_2)\) we have
Hence, for any \((v_1,p_1),(v_2,p_2)\in V_2({\bar{v}})\times Q_2({\bar{p}})\),
\(\square \)
Remark 1
It follows from the proof that under the assumptions of Theorem 1 and condition (I), the estimate in (II) holds with constant \(\ell ^0=4\ell _{Gr}\).
Remark 2
Clearly, \( \text {dom}\, Gr=\{p\in {{\mathcal {G}}}\ |\ Gr(p)\ne \emptyset \}=\text {dom}\,{\mathbb {C}} \) and
for \(p\in \text {dom}\, Gr\). Consequently, by taking \({\bar{p}}\in \text {dom}\,{\mathbb {C}}\), \(0\ne {\bar{v}}-{\bar{x}}\in N({\bar{x}};C({\bar{p}}))\) and a neighbourhood V(0) in \({{\mathcal {H}}}\) such that \(0\not \in {\bar{v}}-{\bar{x}}+V(0)\)
for p close to \({\bar{p}}\).
In view of Theorem 1 to prove (II) we need to show (I) and the condition (A) . Condition (A) for \({{\mathcal {C}}}\) given by (2) was investigated in details in [3, 5] and it is discussed in Sect. 3. Condition (I) is proved in Proposition 9 in Sect. 6 with the help of a number of propositions proved in Sect. 4.
In the sequel we make an extensive use of the lower Kuratowski limit (Painlevé–Kuratowski inner/lower limit) for a multifunction \({{\mathcal {F}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) at \({\bar{p}}\) defined as
Equivalently, \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}}{{\mathcal {F}}}(p)\) if for every neighbourhood \(V({\bar{x}})\) of \({\bar{x}}\) there exists a neighbourhood \(U({\bar{p}})\) of \({\bar{p}}\) such that \(V({\bar{x}})\cap {{\mathcal {F}}}(p)\ne \emptyset \) for \(p\in U({\bar{p}})\).
The following condition related to the lower Kuratowski limit is necessary for the continuity of the projection mapping P.
Proposition 2
Let \({\bar{p}}\in {{\mathcal {D}}}\) and \({\bar{v}}\in {{\mathcal {H}}}\). If the mapping \(P:\ {{\mathcal {H}}}\times {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\) given by (1) is continuous at \(({\bar{v}},{\bar{p}})\in {{\mathcal {H}}}\times {{\mathcal {D}}}\) with \({\bar{x}}:=P({\bar{v}},{\bar{p}})\in C({\bar{p}})\), then \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}}{{\mathbb {C}}}(p)\).
Proof
Suppose, by contradiction, that \({\bar{x}}\not \in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}}{{\mathbb {C}}}(p)\) . By definition, there exists a neighbourhood \(V({\bar{x}})\) of \({\bar{x}}\) such that in every neighbourhood \(U({\bar{p}})\) of \({\bar{p}}\) there exists \(p_{U}\in U({\bar{p}})\) satisfying
In consequence, \(P({\bar{v}},p_{U})\not \in V({\bar{x}}),\) which contradicts the continuity of P at \(({\bar{v}},{\bar{p}})\).\(\square \)
3 RCRCQ and Lipschitz-Likeness of the Set-Valued Mapping \({{\mathbb {C}}}\)
In this section we discuss Lipschitz-likeness of set-valued mapping \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {\mathcal {H}}\) defined by (3).
A crucial assumption of Theorem 1 is the Lipschitz-likeness of the set-valued mapping \({\mathcal {C}}\) given by (3). In this section we present recent results of [4] where the the Lipschiz-likeness of \( {{\mathcal {C}}}\) has been investigated with the help of the relaxed constant rank constaint qualification (RCRCQ).
For any \((p,x)\in {{\mathcal {D}}}\times {{\mathcal {H}}}\), let \(I_{p}(x):= \{ i\in I_1\cup I_2\ |\ \langle x\ |\ g_i(p)\rangle =f_i(p) \}\) denote the active index set for \(p\in {{\mathcal {D}}}\) at \(x\in {{\mathcal {H}}}\).
In our main results (Proposition 5, Theorem 6) we use the relaxed constant rank constraint qualification as defined in [4, Definition 4].
Definition 2
The relaxed constant rank constraint qualification (RCRCQ) for multifunction \({{\mathbb {C}}}\) is satisfied at \(({\bar{p}},{\bar{x}})\), \({\bar{x}}\in C({\bar{p}})\), given by (3), if there exists a neighbourhood \(U({\bar{p}})\) of \({\bar{p}}\) such that, for any index set J, \(I_1\subset J\subset I_{{\bar{p}}}({\bar{x}})\), for every \(p\in U({\bar{p}})\) the system of vectors \(\{ g_i(p), i\in J \}\) has constant rank, i.e.,
Due to the linearity with respect to x of functions defining the set C(p), condition (10) does not depend upon x from a neighbourhood of \({\bar{x}}\). It is worth to mention that RCRCQ does not imply MFCQ, nor MFCQ imply RCRCQ (see e.g. [15])
Proposition 3 says that in a neighbourhood of \({\bar{p}}\in {{\mathcal {D}}}\) we can equivalently represent the set C(p), given by (3), in such way that the normal vectors of equality constraints are linearly independent. A finite-dimensional analogue of Proposition 3 has been established in [17, Lemma 2.2].
Proposition 3
[4, Proposition 11] Let \({\bar{p}}\in {{\mathcal {D}}}\). Assume RCRCQ holds at \(({\bar{p}},{\bar{x}})\), \({\bar{x}}\in C({\bar{p}})\) for multifunction \({{\mathbb {C}}}\) and \(C(p)\ne \emptyset \) for \(p\in U_0({\bar{p}})\). There exists a neighbourhood \(U({\bar{p}})\) such that for all \(p\in U({\bar{p}})\)
where \(I_1^\prime \subset I_1\), \(|I_1^\prime |=\text {rank} \{g_i({\bar{p}}),\ i \in I_1 \}\) and \(g_i(p)\), \(i\in I_1^\prime \) are linearly independent.
In view of Proposition 3, in the sequel we assume that for any \({\bar{p}} \in {{\mathcal {D}}}\), \(g_i(p)\), \(i\in I_1\) are linearly independent in some neighbourhood \(U({\bar{p}})\).
Remark 3
Let us note that for the set-valued mapping \({\hat{{{\mathbb {C}}}}}:\ {{\mathcal {D}}}\rightrightarrows {\mathcal {H}}\), \({\hat{{{\mathbb {C}}}}} (p):={\hat{C}}(p)\), with \({\hat{C}}(p)\) defined by (2), the relaxed constant rank constraint qualification condition RCRCQ is satisfied at any \((p,x)\in \text {gph}{\hat{{{\mathbb {C}}}}}\). In the case of absence of equality constraints in (3) the condition RCRCQ is equivalent to constant rank constraint qualification (CRCQ) (see [1, 13, 16]) which has been already used in [23] in proving Lipschitzness of projections.
The following theorem has been proved in [3].
Theorem 2
[4, Theorem 9] Let \({\mathcal {H}}\) be a Hilbert space, \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a subset of a normed space \({{\mathcal {G}}}\) and let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be given by (3). Assume RCRCQ is satisfied at \(({\bar{p}},{\bar{x}})\in \text {gph}\,{{\mathbb {C}}}\) and \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}(p)\). Then \({{\mathbb {C}}}\) is Lipschitz-like at \(({\bar{p}},{\bar{x}})\).
By applying Proposition 1 and Theorem 2, we immediately obtain the following Hölder type estimate for solutions to problem (M(v,p)).
Theorem 3
Let \({\mathcal {H}}\) be a Hilbert space, \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a subset of a normed space \({{\mathcal {G}}}\) and let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be given by (3). Assume RCRCQ is satisfied at a point \(({\bar{p}},{\bar{x}})\in \text {gph}\,{{\mathbb {C}}}\) and \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}(p)\). Then there exist constant \(\ell ^0>0\) and neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) that the estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\).
In view of Remark 3, the following result is an immediate consequence of Theorem 2.
Theorem 4
Let \({\mathcal {H}}\) be a Hilbert space, \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a subset of a normed space \({{\mathcal {G}}}\) and let \({\hat{{{\mathbb {C}}}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\), \({\hat{{{\mathbb {C}}}}} (p):={\hat{C}}(p)\), with \({\hat{C}}(p)\) given by (2). Assume that \(({\bar{p}},{\bar{x}})\in \text {gph}\,{\hat{{{\mathbb {C}}}}}\) and \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {\hat{{{\mathbb {C}}}}}(p)\). Then \({\hat{{{\mathbb {C}}}}}\) is Lipschitz-like at \(({\bar{p}},{\bar{x}})\).
The Lipschitz-likeness of \({\hat{{{\mathbb {C}}}}}\) has already been investigated in the finite-dimensional case in [10] with the help of the Mangasarian–Fromowitz constraint qualification MFCQ.
Definition 3
We say that the Mangasarian–Fromowitz constraint qualification (MFCQ) holds for \(C({\bar{p}})\) at \({\bar{x}}\in C({\bar{p}})\) if vectors \(g_{i}({\bar{p}})\), \(i\in I_{1}\) are linearly independent and there exists \(h\in {{\mathcal {H}}}\) such that
The following fact relates the Mangasarian–Fromowitz constraint qualification MFCQ to the lower Kuratowski limit of the set-valued mapping \({\bar{{{\mathbb {C}}}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\), \({\bar{{{\mathbb {C}}}}} (p):={\bar{C}}(p)\), with \({\bar{C}}(p)\) given by (3) with \(I_{1}=\emptyset \), i.e.
Proposition 4
If MFCQ holds for \({\bar{C}}({\bar{p}})\) at \({\bar{x}}\in {\bar{C}}({\bar{p}})\), then \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {\bar{C}}(p)\).
Proof
By MFCQ, there exists \(h\in {{\mathcal {H}}}\) such that
Let \(V({\bar{x}})\) be any neighbourhood of \({\bar{x}}\). There exists \(\alpha >0\) such that
and \({\bar{x}}+\alpha h\in V({\bar{x}})\). Since the functions \(g_{i},f_{i}\), \(i\in I_1\cup I_2\) are assumed to be locally Lipschitz at \({\bar{p}}\) there exists a neighbourhood \(U_{i}({\bar{p}})\) of \({\bar{p}}\) such that
where \(\ell _{f_i}\), \(\ell _{g_i}\) are locally Lipschitz constants of functions \(f_i\), \(g_i\), \(i\in I_1\cup I_2\) at \({\bar{p}}\), respectively.
Take \(\varepsilon >0\), \(\varepsilon <f_{i}({\bar{p}})-\langle g_{i}({\bar{p}})|{\bar{x}}+\alpha h\rangle >0.\) By shrinking the neighbourhood \(U_{i}({\bar{p}})\), \(i\in I_2\), we can assume that
Consequently, for \(p\in U_{i}({\bar{p}})\), \(i\in I_2\) we have
By (11),
By taking \(U({\bar{p}})=\bigcap _{i\in I_2}U_{i}({\bar{p}})\), we obtain the assertion. \(\square \)
In view of Proposition 4, Theorem 4 is a stronger result than Theorem 4.1 of [10] when applied to the linear case.
The following example shows that MFCQ is not a necessary condition for Lipschitz continuity of projection of v onto C(p), \(p\in {{\mathcal {D}}}\), given in (3).
Example 2
Let \(p\in B((0,0),1)\subset {\mathbb {R}}^2\), \({\bar{p}}=(0,0)\), \({\bar{v}}=(1,1)\) and
Then for all \(p\in B((0,0),1)\) we have \(C(p)=\{p\}\). Hence \(P(v,p)=p\) for \(p\in B((0,0),1)\) and for any \(v\in {\mathbb {R}}^2\). Hence, \(P(\cdot ,\cdot )\) is locally Lipschitz in a neighbourhood of \(({\bar{v}},{\bar{p}})\) and MFCQ is not satisfied at \(P({\bar{v}},{\bar{p}})\).
4 RCRCQ and Lagrange Multipliers
In our investigations of Lipschitzness of the projection \(P({\bar{v}},{\bar{p}})\) we are making an extensive use of representations of elements of normal cones (optimality conditions) and the behaviour of these representations in a neighbourhood of a given parameter \({\bar{p}}\). This requires the detailed investigation of the behaviour of the Lagrange multipliers around \({\bar{p}}\) (Proposition 5 and Proposition 7).
In this section we investigate properties of Lagrange multipliers of problem (M(v,p)) under RCRCQ condition. We start with the following elementary observation.
Remark 4
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{x}}\in C({\bar{p}})\), where \({{\mathbb {C}}}\) is given by (3). By the continuity of \(f_i\), \(g_i\), \(i\in I_1\cup I_2\), at \({\bar{p}}\) and the continuity of the inner product, there exist a neighbourhood \(U({\bar{p}})\) and a neighbourhood V(0) of \(0\in {{\mathcal {H}}}\), such that
for all \(p\in U_2({\bar{p}})\), \(x\in ({\bar{x}}+V(0))\cap C(p)\). Hence, \(I_p(x)\subset I_{{\bar{p}}}({\bar{x}})\) for \(p\in U({\bar{p}})\) and \(x\in ({\bar{x}}+V(0))\cap C(p)\).
In Proposition 5 we investigate representations of elements of \(N({\bar{x}}; C({\bar{p}}))\) of the form (12) below in a neighbourhood of \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\), where \({\bar{v}}-{\bar{x}}\in N({\bar{x}}; C({\bar{p}}))\). We prove that for all p close to \({\bar{p}}\), for all \(x\in C(p)\) close to \({\bar{x}}\), for all \(x'\in N(x;C(p))\) close to \({\bar{v}}-{\bar{x}}\), there exists a representation
and the function \(\lambda :\ U({\bar{p}})\times ({\bar{x}}+V(0))\times ({\bar{v}}-{\bar{x}}+V(0)) \rightarrow {\mathbb {R}}^{|I_{{\bar{p}}}({\bar{x}})|}\),
is continuous at \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\).
Proposition 5
Let \({\bar{p}}\in {{\mathcal {D}}}\). Suppose that \({\bar{v}}\notin C({\bar{p}})\), \({\bar{x}}\in C({\bar{p}})\), \(I_{{\bar{p}}}({\bar{x}})\ne \emptyset \).Footnote 1 Assume that RCRCQ holds at \({\bar{p}}\) (with a neighbourhood \(U_0({\bar{p}})\)) for multifunction \({{\mathbb {C}}}\), given by (3), and \(C(p)\ne \emptyset \) for \(p\in U_0({\bar{p}})\). Let
and \(g_i({\bar{p}})\), \(i\in I_1\cup {\bar{K}}\) are linearly independentFootnote 2. Then the following conditions hold.
-
(i)
There exist neighbourhoods \(U({\bar{p}})\), V(0) such that for any \(p\in U({\bar{p}})\) and any \((x,x^\prime )\in Gr(p)\cap ({\bar{x}}+V(0),{\bar{v}}-{\bar{x}}+V(0))\), where Gr is given by (6), there exists \({L}\subset (I_{p}(x)\cap I_2)\setminus {\bar{K}}\subset (I_{{\bar{p}}}({\bar{x}})\cap I_2)\setminus {\bar{K}}\)Footnote 3 such that the element \(x^\prime \) can be represented as
$$\begin{aligned} \begin{aligned}&x^\prime = \sum _{i \in I_1} \lambda _i g_i(p)+\sum _{i \in {\bar{K}} } \lambda _i g_i(p)+\sum _{i \in L } \lambda _i g_i(p),\\&\lambda _i> 0,\ i \in {\bar{K}},\ \lambda _i\ge 0,\ i\in L, \end{aligned} \end{aligned}$$(13)where \(g_i(p)\), \(i\in I_1\cup {\bar{K}}\cup L\) are linearly independent.
-
(ii)
For any \(\varepsilon >0\) one can choose in (i) neighbourhoods \(U({\bar{p}})\), V(0) such that in the representation (13) we have
$$\begin{aligned} \begin{aligned}&{\bar{\lambda }}_i-\varepsilon \le {\lambda }_i \le {\bar{\lambda }}_i+\varepsilon \quad \forall i \in I_1\cup {\bar{K}},\\&0\le {\lambda }_i \le \varepsilon \quad \forall i \in {L}. \end{aligned} \end{aligned}$$
Proof
Since \(g_i({\bar{p}})\), \(i\in I_1\cup {\bar{K}}\) are linearly independent and \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup {\bar{K}}\) are continuous at \({\bar{p}}\), by Lemma 1 (see Appendix), there exists a neighbourhood \(U_1({\bar{p}})\) such that \(g_i(p)\), \(i\in I\cup {\bar{K}}\), \(p\in U_1({\bar{p}})\) are linearly independent.
By Remark 2 and Remark 4, there exist neighbourhoods \(U_2({\bar{p}})\subset U_1({\bar{p}})\) and \(V_1(0)\), such that for all \(p\in U_2({\bar{p}})\) if \((x,x^\prime )\in Gr(p)\cap ({\bar{x}}+V_1(0),{\bar{v}}-{\bar{x}}+V_1(0))\), then \(I_p(x)\subset I_{{\bar{p}}}({\bar{x}})\), \(x\in \text {bd}\, C(p)\) and \(x^\prime \ne 0\). Moreover, by [7, Theorem 6.40]Footnote 4 we have
Now, on the contrary suppose that the assertion of the proposition does not hold, i.e., there exist sequences \(p_n\rightarrow {\bar{p}}\), \(x_n\rightarrow {\bar{x}}\), \(x_n\in C(p_n)\), \(x_n^\prime \rightarrow {\bar{v}}-{\bar{x}}\), \(x_n^\prime \in N(x_n;C(p_n))\) such that
By (14), for all \(n\in {\mathbb {N}}\), sufficiently large, \(x_n^\prime \) can be represented in the form
where \(\lambda _i^n\ge 0\), \(i\in I_{p_n}(x_n)\cap I_2\). We can rewrite (16) as
and, by putting \(\lambda _i^n=0\) for \(i \in {\bar{K}}\setminus I_{p_n}(x_n)\), \(n\in {\mathbb {N}}\), we get
where \(\lambda _i^n\ge 0\), \(i\in (I_{p_n}(x_n)\cap I_2)\cup {\bar{K}}\). By Lemma 3 (see Appendix), for all \(n\in {\mathbb {N}}\), sufficiently large, there exist \({\hat{I}}_2^n\subset (I_{p_n}(x_n)\setminus {\bar{K}})\cap I_2\) and \({\tilde{\lambda }}_i^n\), \(i\in I_1\cup {\bar{K}}\cup {\hat{I}}_2^n\), \({\tilde{\lambda }}_i^n\in {\mathbb {R}}\), \(i\in I_1\cup {\bar{K}}\), \({\tilde{\lambda }}_i^n>0\), \(i\in {\hat{I}}_2^n\) such that
where \({\tilde{\lambda }}_i\in {\mathbb {R}}\), \(i\in I_1\cup {\bar{K}}\), \({\tilde{\lambda }}_i> 0 \), \(i\in {\hat{I}}_2^n\) and \(g_i(p_n)\), \(i\in I_1\cup {\bar{K}}\cup {\hat{I}}_2^n\) are linearly independent.
By passing to a subsequence, if necessary, we can assume that \({\hat{I}}_2^n=:I_2^\prime \) and
where \({\tilde{\lambda }}_i^n\), \({\tilde{\lambda }}_i^n\in {\mathbb {R}}\), \(i\in I_1\cup {\bar{K}}\), \({\tilde{\lambda }}_i^n> 0\), \(i\in I_2^\prime \) and \(g_i(p_n)\), \(i\in I_1\cup {\bar{K}}\cup I_2^\prime \) are linearly independent.
By (15), it must be \({\tilde{\lambda }}_i^n\le 0\) for some \(i_n\in {\bar{K}}\). Passing again to the subsequence in (17), if necessary, we conclude that there exists \(i\in {\bar{K}}\) such that \({\tilde{\lambda }}_i^n\le 0\).
On the other hand, by Lemma 4 (see Appendix), we have \({\tilde{\lambda }}_i^n\rightarrow {\bar{\lambda }}_i>0\), \(i\in {\bar{K}}\), which leads to a contradiction. This proves (i).
To prove (ii) suppose there exist \(\varepsilon >0\) and a sequence \(\{i_n\}_{n\in {\mathbb {N}}}\subset I\cup {\bar{K}}\cup I_2^\prime \), such that in the representation (17) for each \(n\in {\mathbb {N}}\) one of the following holds:
-
1.
\({\tilde{\lambda }}_{i_n}>{\bar{\lambda }}_i+\varepsilon \) and \(i_n\in I_1\cup {\bar{K}}\),
-
2.
\({\tilde{\lambda }}_{i_n}<{\bar{\lambda }}_i-\varepsilon \) and \(i_n\in I_1\cup {\bar{K}}\),
-
3.
\({\tilde{\lambda }}_{i_n}>\varepsilon \) and \(i_n\in I_2^\prime \).
By taking a subsequence of \(\{x_n^\prime \}_{n\in {\mathbb {N}}}\), if necessary, one can assume that only one of the cases 1., 2., 3. holds for all \(n\in {\mathbb {N}}\). On the other hand, by Lemma 4 (see Appendix), we have \({\tilde{\lambda }}_i^n\rightarrow {\bar{\lambda }}_i\), \(i\in I_1\cup {\bar{K}}\) and \({\tilde{\lambda }}_i^n\rightarrow 0\), \(i\in I_2^\prime \), which leads to a contradiction. This proves (ii). \(\square \)
It is clear that even if an element \({\bar{v}}-{\bar{x}}\) from the normal cone \(N({\bar{x}}; C({\bar{p}}))\) has a unique representation as a combination of some vectors \(g_{i}({\bar{p}})\), \(i\in I_{{\bar{p}}}({\bar{x}})\), then, in a neighbourhood of \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\), where \({\bar{v}}-{\bar{x}}\in N({\bar{x}};C({\bar{p}}))\), the elements \(x^\prime \in N(x; C(p))\) may not have unique representations in terms of combinations of vectors \(g_{i}(p)\), \(i\in I_{p}(x)\).
The example below illustrates the situation when the representation of \({\bar{v}}-{\bar{x}}\) is not unique.
Example 3
Let \(p\in {\mathbb {R}}\), \({\bar{v}}=(0,1)\) and
In this case \(P({\bar{v}},p)=(0,0):={\bar{x}}\) for all \(p\in {\mathbb {R}}\). However, for \({\bar{p}}=0\) we have \(I_{{\bar{p}}}({\bar{x}})=\{1,2,3\}\) and
and for any \(p\ne 0\), \(x=(p,0)\), \(x^\prime \in N(x;C(p))\) we have
In the proposition below we show, that under assumptions appearing in Theorem 6 we have \(Gr(p)\cap V({\bar{x}},{\bar{v}}-{\bar{x}})\ne \emptyset \) for \((p,x,x^\prime )\) in some neighbourhood of \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\).
Proposition 6
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\notin C({\bar{p}})\) and \({\bar{x}}=P({\bar{v}},{\bar{p}})\). Assume that RCRCQ holds at \(({\bar{p}},{\bar{x}})\) (with a neighbourhood \(U_0({\bar{p}})\)) for multifunction \({{\mathbb {C}}}\), given by (3), and \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}(p)\). Then
i.e., \(({\bar{x}},{\bar{v}}-{\bar{x}})\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} Gr(p).\)
Proof
On the contrary, suppose that there exist a neighbourhood \(V({\bar{x}},{\bar{v}}-{\bar{x}})\) and a sequence \(p_n\rightarrow {\bar{p}}\) such that
Without loss of generality we may assume that \(p_n\in U_0({\bar{p}})\). Let \(x_n:=P({\bar{v}},p_n)\). Then, by Theorem 3, we have \(x_n\rightarrow {\bar{x}}\). Moreover, \(x_n^\prime :={\bar{v}}-x_n\in N(x_n;C(p_n))\) and \(x_n^\prime \rightarrow {\bar{v}}-{\bar{x}}\). Thus, \((x_n,x_n^\prime )\in Gr(p_n)\cap V({\bar{x}},{\bar{v}}-{\bar{x}})\) for large n. \(\square \)
Let us recall that the set of Lagrange multipliers associated with problem (M(v,p)) is defined as
and for \(M>0\) let
In proposition below we show that \({\varLambda }_{{v}}^M(p,P(v,p))\ne \emptyset \) in some neighbourhood of \(({\bar{p}},{\bar{v}})\) under RCRCQ and the Kuratowski limit conditions.
Proposition 7
Suppose that \({\bar{v}}\notin C({\bar{p}})\). Assume that RCRCQ holds at \(({\bar{p}},P({\bar{v}},{\bar{p}}))\) (with a neighbourhood \(U_0({\bar{p}})\)) for multifunction \({{\mathbb {C}}}\), given by (3), and \(P({\bar{v}},{\bar{p}})\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}(p)\). Let the formula (12) hold, i.e.,
and \(g_i({\bar{p}})\), \(i\in I_1\cup {\bar{K}}\) are linearly independent. There exist neighbourhoods \(U_1({\bar{p}})\), \(V_1(0)\) and \(M>0\) such that for all \(p\in U_1({\bar{p}})\) and \(v\in {\bar{v}}+V_1(0)\) we have
Proof
Let \(\varepsilon >0\). By Proposition 5, there exist neighbourhoods \(U({\bar{p}})\), V(0) such that for every \(p\in U({\bar{p}})\) and any \((x,x^\prime )\in Gr(p)\cap ({\bar{x}}+V(0),{\bar{v}}-{\bar{x}}+V(0))\), there exists \({L}\subset (I_p(x)\cap I_2)\setminus K \subset (I_{{\bar{p}}}({\bar{x}})\cap I_2)\setminus K\) such that the formula (13) holds i.e.,
where \(g_i(p)\) \(i\in I_1\cup {\bar{K}}\cup {L}\) are linearly independent and additionally
Let \(V_1(0)\) be such that \(V_1(0)\subset \frac{1}{2}V(0)\). By the continuity of \(P(\cdot ,\cdot )\) at \(({\bar{v}},{\bar{p}})\) (see Theorem 2 and Proposition 1), there exist neighbourhoods \(U_2({\bar{p}})\), \(V_2(0)\subset \frac{1}{2}V(0)\) such that
for all \(p\in U_2({\bar{p}})\), \(v\in {\bar{v}}+V_2(0)\). Hence,
Let \(U_1({\bar{p}}):=U({\bar{p}})\cap U_2({\bar{p}})\). Then for all \(p\in U_1({\bar{p}})\), \(v\in {\bar{v}}+V(0)\) there exists \({L}\subset I_{{\bar{p}}}(P({\bar{v}},{\bar{p}}))\cap I_2\) (\(L\subset (I_p(v)\cap I_2)\setminus K\)) such that
and (19) holds. This means that for all \(p\in U_1({\bar{p}})\), \(v\in {\bar{v}}+V(0)\),
i.e. (18) holds.\(\square \)
5 Stable Representations
As already noted in Example 3, a number of different index sets \({\bar{K}}\) could be used in (12). On the other hand, the set of those index sets \({\bar{K}}\) for which (12) holds is nonempty (may consists of the empty set only).
Definition 4
Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\), be given by (3). We say that \({{\mathcal {R}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) is an equivalent representation of \({{\mathbb {C}}}\), if \(R(p)=C(p)\) for all \(p\in {{\mathcal {D}}}\) and R is given as
where \({\tilde{f}}_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \({\tilde{g}}_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in {\tilde{I}}_1\cup {\tilde{I}}_2\) are locally Lipschitz on \({{\mathcal {D}}}\) and \({\tilde{I}}_1\cup {\tilde{I}}_2\) is a finite, nonempty set. For a given representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\), we define \(I_{{p}}^{{\mathcal {R}}}(x)=\{ i\in {\tilde{I}}_1\cup {\tilde{I}}_2 \mid \langle x \mid {\tilde{g}}_i(p) \rangle = {\tilde{f}}_i(p) \}\), \(p\in {{\mathcal {D}}}\), \(x\in {{\mathcal {H}}}\).
Observe that \(N(x,C(p))=N(x,R(p))\) for all \(p\in {{\mathcal {D}}}\), \(x\in C(p)\). On the other hand the representations of elements from the normal cone may differ depending on the equivalent representation considered.
Consider any equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\). Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{x}}\in R({\bar{p}})=C({\bar{p}})\), \(I_{{\bar{p}}}^{{\mathcal {R}}}({\bar{x}})\ne \emptyset \) and let the following formula (c.f., formula (12)) hold
where \({\tilde{g}}_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\bar{K}}\) are linearly independent.
For a given representation \({{\mathcal {R}}}\), for any index set LFootnote 5 satisfying
we define multifunction \({{\mathcal {R}}}_{{\bar{K}},L}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) as \({{\mathcal {R}}}_{{\bar{K}},L}(p):=R_{{\bar{K}},L}(p)\),
Note that \({\bar{x}}\in R_{{\bar{K}},L}({\bar{p}})\) for any index set L satisfying (21) and, in general, \(R_{{\bar{K}},L}(p)\ne C_{{\bar{K}},L}(p)\), \(p\in {{\mathcal {D}}}\).
Definition 5
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\notin C({\bar{p}})\) and \({\bar{x}}\in C({\bar{p}})\). We say that multifunction \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\), given by (3), has a stable representation (in the sense of Kuratowski limit) at \(({\bar{p}},{\bar{v}},{\bar{x}})\) if there exists an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) for which there exists \({\bar{K}}\subset I_{{\bar{p}}}^{{\mathcal {R}}}({\bar{x}})\cap {\tilde{I}}_2\) such that (20) holds and
We say that \({{\mathcal {R}}}\) is a stable representation of \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{v}},{\bar{x}})\) if there exists \({\bar{K}}\) such that (20) and (22) hold.
Let us note that if multifunction \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\), given by (3), has a stable representation \({{\mathcal {R}}}\) at \(({\bar{p}},{\bar{v}},{\bar{x}})\) then \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} C(p)=\liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} R(p)\).
By Proposition 5 and Theorem 2, we obtain the following corollary.
Corollary 1
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{x}}\in C({\bar{p}})\), \(I_{{\bar{p}}}({\bar{x}})\ne \emptyset \) and \({\bar{v}}-{\bar{x}}\in N({\bar{x}};C({\bar{p}}))\). Assume that there exists an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) satisfying
-
RCRCQ holds for multifunction \({{\mathcal {R}}}\) at \(({\bar{p}},{\bar{x}})\),
-
\({{\mathcal {R}}}\) is a stable representation at \(({\bar{p}},{\bar{v}},{\bar{x}})\) with some set \({\bar{K}}\subset I_{{\bar{p}}}^{{\mathcal {R}}}({\bar{x}})\cap {\tilde{I}}_2\).
There exists a constant \(\ell >0\) such that for all \(\varepsilon >0\) one can find neighbourhoods \(U({\bar{p}})\) and V(0) satisfying
-
(1)
for any \(p\in U({\bar{p}})\) and any \((x,x^\prime )\in Gr(p)\cap ({\bar{x}}+V(0),{\bar{v}}-{\bar{x}}+V(0))\), it exists \({\hat{L}}\subset (I_{{p}}^{{\mathcal {R}}}({x})\cap {\tilde{I}}_2)\setminus {\bar{K}} \subset (I_{{\bar{p}}}^{{\mathcal {R}}}({\bar{x}})\cap {\tilde{I}}_2)\setminus {\bar{K}}\) satisfying (21) such that
$$\begin{aligned}&\exists \, {\lambda }_i,\ i \in {\tilde{I}}_1\cup {\bar{K}}\cup {{\hat{L}}}, \quad x^\prime = \sum _{i \in {\tilde{I}}_1\cup {\bar{K}}\cup {{\hat{L}}}} \lambda _i {\tilde{g}}_i(p),\\&{\bar{\lambda }}_i-\varepsilon \le {\lambda }_i \le {\bar{\lambda }}_i+\varepsilon \quad \forall i \in {\tilde{I}}_1\cup {\bar{K}},\\&0< {\lambda }_i \le \varepsilon \quad \forall i \in {{\hat{L}}}, \end{aligned}$$ -
(2)
for every L satisfying (21), every \(p_1,p_2\in U({\bar{p}})\) and every \(x_1\in ({\bar{x}}+V(0))\cap R_{{\bar{K}},L}(p_1)\) there exists \(x_2\in R_{{\bar{K}},{L}}(p_2)\) such that
$$\begin{aligned}&\Vert x_1-x_2\Vert \le \ell \Vert p_1-p_2\Vert , \end{aligned}$$i.e., the set-valued mapping \({{\mathcal {R}}}_{{\bar{K}},L}\) is Lipschitz-like at \(({\bar{p}},{\bar{x}})\).
Proof
Clearly, RCRCQ holds for any \({{\mathcal {R}}}_{{\bar{K}},L}\) at \(({\bar{p}},{\bar{x}})\), with L satisfying (21).
By Proposition 5 applied to \({{\mathcal {R}}}\), there exist neighbourhoods \(U_1({\bar{p}})\), \(V_1(0)\) such that assertion (1) holds.
By Theorem 2 applied to \({{\mathcal {R}}}\) at \(({\bar{p}},{\bar{x}})\), for any L satisfying (21), the multifunction \({{\mathcal {R}}}_{{\bar{K}},L}\) is Lipschitz-like at \(({\bar{p}},{\bar{x}})\) with neighbourhoods \(U_L({\bar{p}})\), \(V_L(0)\) and constant \(\ell _L>0\), i.e., assertion (2) holds.
The existence of neighbourhoods \(U({\bar{p}})\), V(0) and constant \(\ell >0\) satisfying the assertion follows from the fact that there is a finite number of sets L satisfying (21).\(\square \)
In Theorem 5 we use the following assumption (H1).
-
(H1)
There exist an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\), given by (3), with
$$\begin{aligned} {\bar{v}}-P({\bar{v}},{\bar{p}})=\sum _{i\in {\tilde{I}}_1\cup {\bar{K}}} {\bar{\lambda }}_i {\tilde{g}}_i({\bar{p}}) \end{aligned}$$where \({\bar{\lambda }}_i>0\), \(i\in {\bar{K}}\), \({\tilde{g}}_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\bar{K}}\) are linearly independent (\({\bar{K}}\subset I_2\cap I_{{\bar{p}}}^{{\mathcal {R}}}(P({\bar{v}},{\bar{p}}))\)), and neighbourhoods \(U({\bar{p}})\), \(W({\bar{v}})\) such that
-
(a)
\({\bar{K}}\subset I_{p}^{{\mathcal {R}}}(P(v,p))\) for all \(p\in U({\bar{p}})\), \(v\in W({\bar{v}})\),
-
(b)
for any \(p_n\rightarrow {\bar{p}}\) and any \(L\subset ({\tilde{I}}_2\cap I_{{\bar{p}}}^{{\mathcal {R}}}(P({\bar{v}},{\bar{p}})))\setminus {\bar{K}}\) such that \({\tilde{g}}_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\bar{K}}\cup L\) are linearly independent there exist sequence \(v_n\rightarrow {\bar{v}}\), such that \( L \subset I_{p_n}^{{\mathcal {R}}}(P(v_n,p_n))\) for n sufficiently large.
-
(a)
Below we show that if an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) satisfies assumption (H1), then the stability of \({{\mathcal {R}}}\) (in the sense of Definition 5) is necessary for continuity of projection operator P.
Theorem 5
Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be given by (3). Suppose that (H1) holds, i.e., there exists an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) satisfying conditions (a) and (b). If projection \(P:\ {{\mathcal {G}}}\times {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), with \(P(\cdot ,\cdot )\) given by (1), is continuous at \(({\bar{v}},{\bar{p}})\in {{\mathcal {H}}}\times {{\mathcal {D}}}\), then the representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) is stable at \(({\bar{p}},{\bar{v}},P({\bar{v}},{\bar{p}}))\).
Proof
By contradiction suppose, that representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) is not stable at \(({\bar{p}},{\bar{v}},P({\bar{v}},{\bar{p}}))\), i.e, for any \({\tilde{K}}\) such that
and \({\tilde{g}}_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\tilde{K}}\) are linearly independent, there exists \({\tilde{L}}\) satisfying (21) such that \(P({\bar{v}},{\bar{p}})\notin \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} R_{{\tilde{K}},{\tilde{L}}}(p)\). In particular, (23) holds for \({\tilde{K}}={\bar{K}}\) and for any \({\tilde{L}}\subset (I_2\cap I_{{\bar{p}}}^{{\mathcal {R}}}(P({\bar{v}},{\bar{p}})))\setminus {\bar{K}}\) such that \(g_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\bar{K}}\cup {\tilde{L}}\) are linearly independent.
By assumption that \(P({\bar{v}},{\bar{p}})\notin \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} R_{{\bar{K}},{\tilde{L}}}(p)\), there exists a neighbourhood V(0) such that in every neighbourhood of \({\bar{p}}\) one can find element p such that \((P({\bar{v}},{\bar{p}})+V(0))\cap R_{{\bar{K}},{\tilde{L}}}(p)=\emptyset \), i.e., there exists a sequence \(p_n\rightarrow {\bar{p}}\) such that \((P({\bar{v}},{\bar{p}})+V(0))\cap R_{{\bar{K}},{\tilde{L}}}(p_n)=\emptyset \).
Consider a sequence \(v_n\rightarrow {\bar{v}}\) satisfying condition (b) of (H1). Then
This formula implies that \( {\tilde{I}}_1\cup {\bar{K}}\cup {\tilde{L}}\subset I_{p_n}^{{\mathcal {R}}}(P(v_n,p_n))\), \(P(v_n,p_n)\in R_{{\bar{K}},{\tilde{L}}}(p_n)\) and \(v_n-P(v_n,p_n)\in N(x_n,R_{{\tilde{L}}}(p_n))\). Thus \(P(v_n,p_n)=P_{R_{{\bar{K}},{\tilde{L}}}(p_n)}(v_n)\). Hence, \(P(v_n,p_n)\notin P({\bar{v}},{\bar{p}})+ V(0)\), which means that \(P(\cdot ,\cdot )\) is not continuous at \(({\bar{v}},{\bar{p}})\). \(\square \)
6 Main Results
In this section we prove local Lipschitzness of projections onto moving closed convex sets C(p) defined by (3). In view of Theorem 2 in order to apply Theorem 1 we need to investigate Lipschitz-likeness of the graphical subdifferential mapping Gr given by (6).
We start with the following technical fact.
Proposition 8
Let \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\) be such that \({\bar{x}}\in C({\bar{p}})\), \({\bar{v}}-{\bar{x}}\in N({\bar{x}};C({\bar{p}}))\), and \(I_{{\bar{p}}}({\bar{x}})=\{i \in I_1\cup I_2\ |\ \langle {\bar{x}} \ |\ g_i({\bar{p}})\rangle =f_i({\bar{p}}) \}\ne \emptyset \). The following conditions are equivalent:
-
(i)
The graphical subdifferential mapping \(Gr:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\times {{\mathcal {H}}}\) defined as
$$\begin{aligned} Gr(p)=\{ (x,x^\prime )\ |\ x\in C(p),\ x^\prime \in N(x;C(p)) \} \end{aligned}$$is Lipschitz-like around \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\)
-
(ii)
There exist \(\ell >0\) and neighbourhoods \(U({\bar{p}})\), V(0) in \({{\mathcal {H}}}\) such that
$$\begin{aligned}&\forall \ p_1,p_2\in U({\bar{p}}) \\&\forall \ x_1\in C(p_1)\cap ({\bar{x}}+V(0)),\ x_1^\prime \in N(x_1;C(p_1))\cap ({\bar{v}}-{\bar{x}}+V(0)) \\&\exists \ x_2\in C(p_2),\ x_2^\prime \in N(x_2;C(p_2))\ \text {satisfying} \\&\Vert x_1-x_2\Vert \le \ell \Vert p_1-p_2\Vert , \end{aligned}$$(a)$$\begin{aligned}&\Vert x_1^\prime -x_2^\prime \Vert \le \ell \Vert p_1-p_2\Vert . \end{aligned}$$(b)
Proof
By (i), there exist neighbourhoods \(U({\bar{p}})\), V(0) such that for every \((p_1,p_2)\in U({\bar{p}})\)
i.e., for all \((x_1,x_1^\prime ) \in \text {gph}\,Gr(p_1)\cap ({\bar{x}}+V(0),{\bar{v}}-{\bar{x}}+V(0))\) there exists \((x_2,x_2^\prime )\in Gr(p_2) \) such that
where \(B(0,1)\subset {{\mathcal {H}}}\times {{\mathcal {H}}}\) is the open unit ball in \({{\mathcal {H}}}\times {{\mathcal {H}}}\). Hence,
where \(x_{1}\in C(p_{1})\) and \(x_{2}\in C(p_{2})\), \(x_{1}'\in N(x_{1};C(p_{1}))\), \(x_{2}'\in N(x_{2};C(p_{2}))\), which implies (a) and (b). The converse implication is immediate. \(\square \)
Remark 5
Let us note that in (24) we use the norm \(\Vert \cdot \Vert _1\) in the Cartesian product \({{\mathcal {H}}}\times {{\mathcal {H}}}\). Clearly, any other equivalent norm can be used at this point.
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\notin C({\bar{p}})\) and \({\bar{x}}=P({\bar{v}},{\bar{p}})\). From [7, Theorem 6.41] (see also [29]) the following representation holds
In the proposition below we give sufficient conditions for the graphical subdifferential mapping Gr given by (6) to be Lipschitz-like at \(({\bar{p}},{\bar{x}},{\bar{v}}-{\bar{x}})\).
Proposition 9
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\notin C({\bar{p}})\). Assume that there exists an equivalent stable representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{v}},P({\bar{v}},{\bar{p}}))\), given by (3), (with set \({\bar{K}}\)) and RCRCQ holds for \({{\mathcal {R}}}\) at \(({\bar{p}},P({\bar{v}},{\bar{p}}))\) Then the graphical subdifferential mapping Gr, given by (6), is Lipschitz-like at \(({\bar{p}},P({\bar{v}},{\bar{p}}),{\bar{v}}-P({\bar{v}},{\bar{p}}))\).
Proof
Let \(\varepsilon >0\). Let \(U({\bar{p}})\), V(0) be as in Corollary 1. We have
where \({\tilde{g}}_i({\bar{p}})\), \(i\in {\tilde{I}}_1\cup {\bar{K}}\) are linearly independent.
Now, let \(p_1\in U({\bar{p}})\) and \(x_1\in (P({\bar{v}},{\bar{p}})+V(0))\cap C(p_1)\), \(x_1^\prime \in N(x_1;C(p_1))\cap ({\bar{v}}-P({\bar{v}},{\bar{p}})+V(0))\). By Corollary 1, there exists \({\tilde{L}}\subset I_{{\bar{p}}}^{{\mathcal {R}}}(P({\bar{v}},{\bar{p}}))\cap {\tilde{I}}_2\setminus {\bar{K}}\) such that \(x_1\in R_{{\bar{K}},{\tilde{L}}}(p_1)\) and
and for any \(p_2\in U({\bar{p}})\) there exists \(x_2\in R_{{\bar{K}},{\tilde{L}}}(p_2)\subset C(p_2)\) such that
for some \(\ell ^1>0\). Since \(x_2\in R_{{\bar{K}},{\tilde{L}}}(p_2)\) we have
Then
where we put \(\ell ^2:=\sum _{i \in I_{{\bar{p}}}^{{\mathcal {R}}}(P({\bar{v}},{\bar{p}}))} (|{\bar{\lambda }}_i|+\varepsilon ) \ell _{{\tilde{g}}_i} \). \(\square \)
Now we are ready to establish our main theorem.
Theorem 6
Let \({{\mathcal {H}}}\) be a Hilbert space and let \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a nonempty set of a normed space \({{\mathcal {G}}}\) and \({\bar{p}}\in {{\mathcal {D}}}\). Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be as in (3), where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset ,\) \(I_1=\emptyset \vee \{1,\dots ,q\}\), \(I_2=\emptyset \vee \{q+1,\dots ,m\}\) are locally Lipschitz on \({{\mathcal {D}}}\). Let \({\bar{v}}\notin C({\bar{p}})\). Assume that there exists an equivalent representation \({{\mathcal {R}}}\) of \({{\mathbb {C}}}\) such that
-
(1)
RCRCQ holds for multifunction \({{\mathcal {R}}}\) at \(({\bar{p}},{\bar{x}})\),
-
(2)
\({{\mathcal {R}}}\) is a stable representation of \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{v}},{\bar{x}})\) with some set \({\bar{K}}\subset I_{{\bar{p}}}^{{\mathcal {R}}}({\bar{x}})\cap {\tilde{I}}_2\).
There exist neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) such that the Lipschitzian estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\) with some positive constant \(\ell ^0\).
In particular, we get the following result.
Theorem 7
Let \({{\mathcal {H}}}\) be a Hilbert space and let \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a nonempty set of a normed space \({{\mathcal {G}}}\) and \({\bar{p}}\in {{\mathcal {D}}}\). Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be as in (3), where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset ,\) \(I_1=\emptyset \vee \{1,\dots ,q\}\), \(I_2=\emptyset \vee \{q+1,\dots ,m\}\) are locally Lipschitz on \({{\mathcal {D}}}\). Let \({\bar{v}}\notin C({\bar{p}})\). Assume that
-
(1)
RCRCQ holds for multifunction \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{x}})\),
-
(2)
\({{\mathbb {C}}}\) is a stable representation at \(({\bar{p}},{\bar{v}},{\bar{x}})\) with some set \({\bar{K}}\subset I_{{\bar{p}}}({\bar{x}})\cap {\tilde{I}}_2\).
There exist neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) such that the Lipschitzian estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\) with some positive constant \(\ell ^0\).
Proof of Theorem 6
The proof follows directly from Theorem 1, Theorem 4 and Proposition 9.\(\square \)
Clearly, by (25),
If the multifunction \({{\mathbb {C}}}\) is constant around \({\bar{p}}\), then assumptions of Theorem 6 are satisfied.
Corollary 2
Under assumptions of Theorem 6, projection of a given fixed \({\bar{v}}\) onto C(p), \(p\in {{\mathcal {D}}}\), i.e.,
is locally Lipschitz at \({\bar{p}}\).
Example 4 shows that for a given representation of \({{\mathbb {C}}}\) one can not expect that there exists \({\bar{K}}\) such that (20) and (21) holds.
Example 4
Let \({{\mathcal {H}}}={\mathbb {R}}^2\), \({{\mathcal {G}}}={\mathbb {R}}\), \({\bar{p}}=0\), \({\bar{v}}=(1,1)\) and
In this case we have
and \(P({\bar{v}},p)\) is a Lipschitz function of p. On the other hand,
Hence, the representation (27) of \({{\mathbb {C}}}\) is not stable at (0, (0, 0), (1, 1)) (see Fig. 1).
Remark 6
Let us note that multifunction \({{\mathbb {C}}}\), given by (27), can be equivalently represented as
where
The representation given in (28) is stable at \(({\bar{p}},{\bar{v}},P({\bar{v}},{\bar{p}}))\) and RCRCQ holds at \(({\bar{p}},P({\bar{p}},{\bar{v}}))\).
Corollary 3
Suppose that in the definition of the set C(p), \(p\in {{\mathcal {D}}}\), given in (3), \(I_2=\emptyset \), i.e.,
Let \({\bar{p}}\in {{\mathcal {D}}}\), \({\bar{v}}\notin C({\bar{p}})\), \({\bar{x}}=P({\bar{v}},{\bar{p}})\) and the following hold:
-
(1)
RCRCQ holds for multifunction \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{x}})\), i.e., there exists a neighbourhood \(U({\bar{p}})\) such that
$$\begin{aligned} \text {rank}\,\{g_i(p),\ i\in I_1 \}=\text {rank}\,\{ g_i({\bar{p}}),i\in I_1 \},\quad p\in U({\bar{p}}). \end{aligned}$$ -
(2)
\({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}(p)\).
Then the projection P(v, p) is locally Lipschitz at \(({\bar{v}},{\bar{p}})\).
When LICQ condition holds for set \(C({\bar{p}})\) at \(P({\bar{v}},{\bar{p}})\), i.e., when \(g_i({\bar{p}})\), \(i\in I_{{\bar{p}}}(P({\bar{v}},{\bar{p}}))\) are linearly independent, Theorem 7 can rewritten in a considerably simplified form.
Theorem 8
Let \({{\mathcal {H}}}\) be a Hilbert space and let \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a nonempty set of a normed space \({{\mathcal {G}}}\) and \({\bar{p}}\in {{\mathcal {D}}}\). Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be as in (3), where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset \), \(I_1=\emptyset \vee \{1,\dots ,q\} \), \(I_2= \{q+1 ,\dots ,m\}\vee \emptyset \) are locally Lipschitz on \({{\mathcal {D}}}\). Let \({\bar{v}}\notin C({\bar{p}})\) and LICQ hold for set \(C({\bar{p}})\) at \(P({\bar{v}},{\bar{p}})\). There exist neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) such that the Lipschitzian estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\) with some positive constant \(\ell ^0\).
Proof
We have
where \(g_i({\bar{p}})\), \({\bar{p}}\in {{\mathcal {D}}}\), \(i\in I_1\cup K\subset I_{{\bar{p}}}(P({\bar{v}},{\bar{p}}))\) are linearly independent. Thus (1) of Theorem 7 is satisfied.
Now we show (2) of Theorem 7. Observe that LICQ hold for set \(C_{{\bar{K}},L}({\bar{p}})\) at \(P({\bar{v}},{\bar{p}})\) with any L satisfying (21). Hence, MFCQ holds for set \(C_{{\bar{K}},L}({\bar{p}})\) at \(P({\bar{v}},{\bar{p}})\) with any L satisfying (21). Thus, by Theorem 2.87 of [6], for any L satisfying (21), there exist \(\alpha >0\) and a neighbourhood \(U({\bar{p}})\) of \({\bar{p}}\) such that for all \(p\in U({\bar{p}})\)
This implies that \({\bar{x}}\in \liminf \limits _{p\rightarrow {\bar{p}},\ p\in {{\mathcal {D}}}} {{\mathbb {C}}}_{{\bar{K}},L}(p)\) for any L satisfying (21), i.e. assumption (2) of Theorem 7 is satisfied, which proves the assertion.\(\square \)
In view of proof of Theorem 8 the following corollary holds.
Corollary 4
Let \({{\mathcal {H}}}\) be a Hilbert space and let \({{\mathcal {D}}}\subset {{\mathcal {G}}}\) be a nonempty set of a normed space \({{\mathcal {G}}}\) and \({\bar{p}}\in {{\mathcal {D}}}\). Let \({{\mathbb {C}}}:\ {{\mathcal {D}}}\rightrightarrows {{\mathcal {H}}}\) be as in (3), where \(f_i:\ {{\mathcal {D}}}\rightarrow {\mathbb {R}}\), \(g_i:\ {{\mathcal {D}}}\rightarrow {{\mathcal {H}}}\), \(i\in I_1\cup I_2\ne \emptyset ,\) \(I_1=\emptyset \vee \{1,\dots ,q\}\), \(I_2=\emptyset \vee \{q+1,\dots ,m\}\) are locally Lipschitz on \({{\mathcal {D}}}\). Let \({\bar{v}}\notin C({\bar{p}})\) and
where \(g_i({\bar{p}})\), \(i\in I_1\cup {\bar{K}}\) are linearly independent. Assume that
-
(1)
RCRCQ holds for multifunction \({{\mathbb {C}}}\) at \(({\bar{p}},{\bar{x}})\),
-
(2)
MFCQ holds for set \(C_{{\bar{K}},L}({\bar{p}})\) at \(P({\bar{v}},{\bar{p}})\) with any L satisfying (21).
There exist neighbourhoods \(W({\bar{v}})\), \(U({\bar{p}})\) such that the Lipschitzian estimate
holds for all \((v_1,p_1),(v_2,p_2)\in W({\bar{v}})\times U({\bar{p}})\) with some positive constant \(\ell ^0\).
7 Conclusion
In the present paper we proved Lipschitzian stability of projections (in the sense of (25)) onto parametric polyhedral sets in Hilbert space setting with parameters appearing both in left- and right-hand sides of constraints, which are assumed to be locally Lipschitz. The equality and inequality constraints are allowed. Basic tools for our main results are RCRCQ condition and the representation stability condition (see Definition 5).
In general, there is no relationship between RCRCQ and MFCQ (cf. [15]). Moreover, in Propositions 5, 7, Corollary 1, Theorem 6 the conclusions depend upon formula (12) and the representation stability condition in which the index set \({\bar{K}}\) may not be uniquely defined.
Notes
With this assumption we limit our attention to point \({\bar{x}}\) which lay on the boundary of \(C({\bar{p}})\), \({\bar{x}}\in \text {bd}\, C({\bar{p}})\), where \(\text {bd}\) denotes the boundary of a set.
We admit \(K=\emptyset \).
We admit \(L=\emptyset \).
Let us note that each equality constraint in the set \(C(\cdot )\) can be represented as two inequalities, namely \(\langle x \mid g(p) \rangle = f_i(p) \iff \langle x \mid g(p) \rangle \le f_i(p) \wedge \langle x \mid -g(p) \rangle \le f_i(p)\) for any \(x\in {{\mathcal {H}}}\), \(p\in {{\mathcal {D}}}\) .
We also admit the case \(L=\emptyset \).
References
Andreani, R., Silva, P.J.: Constant rank constraint qualifications: a geometric introduction. Pesqui. Operacional 34, 481–494 (2014)
Attouch, H., Wets, R.J.B.: Quantitative stability of variational systems. II. A framework for nonlinear conditioning. SIAM J. Optim. 3(2), 359–381 (1993). https://doi.org/10.1137/0803016
Bednarczuk, E.M., Rutkowski, K.E.: On Lipschitz-like property for polyhedral moving sets. arXiv e-prints arXiv:1811.05166 (2018). To appear in SIOPT
Bednarczuk, E., Rutkowski, K.: On lipschitz-like property for polyhedral moving sets. SIAM J. Optim. 29(4), 2504–2516 (2019). https://doi.org/10.1137/18M1226518
Bednarczuk, E.M., Minchenko, L.I., Rutkowski, K.E.: On Lipschitz-like continuity of a class of set-valued mappings. arXiv e-prints arXiv:1905.08173 (2019)
Bonnans, J.F., Shapiro, A.: Perturbation analysis of optimization problems. Springer Series in Operations Research. Springer, New York (2000).https://doi.org/10.1007/978-1-4612-1394-9
Deutsch, F.: Best approximation in inner product spaces. CMS Books in Mathematics/Ouvrages de Mathématiques de la SMC, vol. 7. Springer, New York (2001). https://doi.org/10.1007/978-1-4684-9298-9
Dontchev, A.L.: Perturbations, approximations and sensitivity analysis of optimal control systems. Lecture Notes in Control and Information Sciences, vol. 52. Springer, Berlin (1983). https://doi.org/10.1007/BFb0043612
Dontchev, A., Rockafellar, R.: Implicit functions and solution mappings. A view from variational analysis. 2nd updated ed (2014)
Dontchev, A.L., Quincampoix, M., Zlateva, N.: Aubin criterion for metric regularity. J. Convex Anal. 13(2), 281–297 (2006)
Engelking, R.: General topology. Sigma Series in Pure Mathematics, vol. 6, second edn. Heldermann, Berlin (1989). Translated from the Polish by the author
Gfrerer, H., Outrata, J.V.: On the Aubin property of solution maps to parameterized variational systems with implicit constraints. Optimization 12, 11–21 (2019). https://doi.org/10.1080/02331934.2019.1657427
Janin, R.: Directional derivative of the marginal function in nonlinear programming. Math. Program. Stud. 21, 110–126 (1984). https://doi.org/10.1007/bfb0121214. Sensitivity, stability and parametric analysis
Klatte, D., Kummer, B.: Aubin property and uniqueness of solutions in cone constrained optimization. Math. Methods Oper. Res. 77(3), 291–304 (2013). https://doi.org/10.1007/s00186-013-0429-6
Kruger, A.Y., Minchenko, L., Outrata, J.V.: On relaxing the Mangasarian–Fromovitz constraint qualification. Positivity 18(1), 171–189 (2014). https://doi.org/10.1007/s11117-013-0238-4
Lu, S.: Implications of the constant rank constraint qualification. Math. Program. 126(2, Ser. A), 365–392 (2011). https://doi.org/10.1007/s10107-009-0288-3
Lu, S.: Relation between the constant rank and the relaxed constant rank constraint qualifications. Optimization 61(5), 555–566 (2012). https://doi.org/10.1080/02331934.2010.527972
Minchenko, L.I., Sakolchik, P.P.: Hölder behavior of optimal solutions and directional differentiability of marginal functions in nonlinear programming. J. Optim. Theory Appl. 90(3), 555–580 (1996). https://doi.org/10.1007/BF02189796
Minchenko, L., Stakhovski, S.: On relaxed constant rank regularity condition in mathematical programming. Optimization 60(4), 429–440 (2011). https://doi.org/10.1080/02331930902971377
Minchenko, L., Stakhovski, S.: Parametric nonlinear programming problems under the relaxed constant rank condition. SIAM J. Optim. 21(1), 314–332 (2011). https://doi.org/10.1137/090761318
Mordukhovich, B.S.: Variational analysis and generalized differentiation. I. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 330. Springer, Berlin (2006). Basic theory
Mordukhovich, B.S.: Variational analysis and applications. Springer Monographs in Mathematics. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-92775-6
Mordukhovich, B.S., Nghia, T.T.A.: Full Lipschitzian and Hölderian stability in optimization with applications to mathematical programming and optimal control. SIAM J. Optim. 24(3), 1344–1381 (2014). https://doi.org/10.1137/130906878
Mordukhovich, B.S., Rockafellar, R.T., Sarabi, M.E.: Characterizations of full stability in constrained optimization. SIAM J. Optim. 23(3), 1810–1849 (2013). https://doi.org/10.1137/120887722
Mordukhovich, B.S., Nghia, T.T.A., Pham, D.T.: Full stability of general parametric variational systems. Set-Valued and Variational Analysis (2018). https://doi.org/10.1007/s11228-018-0474-7
Outrata, J., Ramírez, C.H.: On the Aubin property of critical points to perturbed second-order cone programs. SIAM J. Optim. 21(3), 798–823 (2011). https://doi.org/10.1137/100807168
Ralph, D., Dempe, S.: Directional derivatives of the solution of a parametric nonlinear program. Math. Program. 70(2, Ser. A), 159–172 (1995). https://doi.org/10.1007/BF01585934
Rockafellar, R.T., Wets, R.J.B.: Variational analysis. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 317. Springer, Berlin (1998). https://doi.org/10.1007/978-3-642-02431-3
Rutkowski, K.E.: Closed-form expressions for projectors onto polyhedral sets in hilbert spaces. SIAM J. Optim. 27(3), 1758–1771 (2017). https://doi.org/10.1137/16M1087540
Yen, N.D.: Lipschitz continuity of solutions of variational inequalities with a parametric polyhedral constraint. Math. Oper. Res. 20(3), 695–708 (1995). http://www.jstor.org/stable/3690178
Zarantonello, E.H.: Projections on convex sets in Hilbert space and spectral theory. I. Projections on convex sets. In: Contributions to nonlinear functional analysis (Proc. Sympos., Math. Res. Center, Univ. Wisconsin, Madison), pp. 237–341 (1971)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendix
Appendix
Lemma 1
[4, Lemma 10] Let \(J=\{1,\dots ,k\}\). Let \(g_i:\ {{\mathcal {H}}}\rightarrow {\mathcal {H}}\), \(i\in J\) be continuous operators and let \({\bar{p}}\) be such that \(g_i({\bar{p}})\), \(i\in J\) are linearly independent. Then there exists a neighbourhood \(U({\bar{p}})\) such that for all \(p\in U({\bar{p}})\), \(g_i(p)\), \(i\in J\) are linearly independent.
Lemma 2
[7, Lemma 6.33] If a nonzero vector x is a positive linear combination of the nonzero vectors \(g_1,\dots ,g_n\), then x is a positive linear combination of linearly independent subset of \(\{g_1,\dots ,g_n\}\).
Remark 7
Let \(J=\{1,\dots ,k\}\), \(J=W_1\cup W_2\), \(W_1\cap W_2=\emptyset \) and let \(x=\sum _{i\in J} \lambda _i g_i \), \(\lambda _i\le 0\), \(i\in W_1\), \(\lambda _i\ge 0\), \(i\in W_2\), where \(g_i\), \(i\in J\) are nonzero vectors. Then there exists \({\bar{J}}\subset J\) and \({\bar{\lambda }}_i\), \(i\in {\bar{J}}\) such that
and \(g_i\), \(i\in {\bar{J}}\) are linearly independent.
Proof
We have \(x=\sum _{i\in J_1} \lambda _i g_i \), where \(J_1\subset J\) and \(\lambda _i< 0\), \(i\in J_1\cap W_1\), \(\lambda _i>0\), \(i\in J_1\cap W_2\). Let
Then \(x=\sum _{i\in J_1} {\tilde{\lambda }}_i \tilde{g_i}\) and \({\tilde{\lambda }}_i\), \(i\in J_1\) are positive. Applying Lemma 2 we have that there exists \({\bar{J}}\subset J_1\) and \({\hat{\lambda }}_i>0\), \(i\in {\bar{J}}\) such that \(x=\sum _{i\in {\bar{J}}} {\hat{\lambda }}_i \tilde{g_i}\) and \({\tilde{g}}_i\), \(i\in {\bar{J}}\) are linearly independent. Now let
Then \(x=\sum _{i\in {\bar{J}}} {\bar{\lambda }}_i g_i \), \({\bar{\lambda }}_i<0\), \(i\in {\bar{J}}\cap W_1\), \({\bar{\lambda }}_i>0\), \(i\in {\bar{J}}\cap W_2\).\(\square \)
Lemma 3
[4, Lemma 12] Let \(x=\sum _{i\in J_1} \lambda _i a_i + \sum _{i\in J_2} \lambda _i a_i\), \(J_1\cap J_2=\emptyset \), \(J_1,J_2\) finite sets, \(\lambda _i\in {\mathbb {R}}\), \(i\in J_1\), \(\lambda _i\ge 0\), \(i\in J_2\) and \(a_i\), \(i\in J_1\cup J_2\) are non-zero vectors. Assume that \(a_i\), \(i\in J_1\) are linearly independent. Then there exists \(J_2^\prime \subset J_2\) and \(\lambda _i^\prime \), \(i\in J_1\cup J_2^\prime \), \(\lambda _i^\prime \in {\mathbb {R}}\), \(i\in J_1\), \(\lambda _i^\prime >0\), \(i\in J_2^\prime \) such that
and \(a_i\), \(i\in J_1\cup J_2^\prime \) are linearly independent.
Lemma 4
Let \({{\mathcal {H}}}\) be a Hilbert space. Let \(\{u_i^n\}_{n\in {\mathbb {N}}}\), \(i\in 1,\dots ,K\) be a sequence of K-tuples of vectors from \({{\mathcal {H}}}\) such that for any \(n\in {\mathbb {N}}\), \(u_i^n\), \(i\in 1,\dots ,K\) are linearly independent. Assume that
-
(1)
\(u_i^n\rightarrow u_i\), for \(i=1\dots ,K\), where \(u_i\), \(i=1\dots ,K\), are linearly independent,
-
(2)
\(\sum _{i=1}^{K} \lambda _i^n u_i^n\rightarrow \sum _{i=1}^{K} {\bar{\lambda }}_i {u}_i\), where \(\lambda _i^n,{\bar{\lambda }}_i\in {\mathbb {R}}\), \(i=1,\dots K\), \(n\in {\mathbb {N}}\).
Then \(\lambda _i^n\rightarrow {\bar{\lambda }}_i\), \(i=1,\dots ,K\).
Proof
For any \(n\in {\mathbb {N}}\) let \(\{e_i^n\}_{n\in {\mathbb {N}}}\) \(i=1,\dots ,K\) be a sequence of orthogonal vectors obtained by the Gram-Schmidt orthogonalization of \(u_i^n\) \(i=1,\dots ,K\), i.e.,
and \(e_i\), \(i=1,\dots ,K\), be orthogonal vectors obtained by the Gram-Schmidt orthogonalization of \(u_i\) \(i=1,\dots ,K\). Since \(u_i^n\rightarrow u_i\), \(i=1,\dots ,K\) we have \(e_i^n\rightarrow e_i\), \(i=1,\dots ,K\).
Let \({\tilde{\lambda }}_i^n,{\tilde{\lambda }}_i\in {\mathbb {R}}\), \(i=1,\dots K\), \(n\in {\mathbb {N}}\) be such that \(\sum _{i=1}^{K} \lambda _i^n u_i^n=\sum _{i=1}^{K} {\tilde{\lambda }}_i^n e_i^n\) and \(\sum _{i=1}^{K} {\bar{\lambda }}_i u_i=\sum _{i=1}^{K} {\tilde{\lambda }}_i e_i\). Since \( \sum _{i=1}^K ( \lambda _i^n u_i^n-{\bar{\lambda }}_i u_i)\rightarrow 0 \) we have
We have
By taking scalar products with \(e_{k}\), \(k\in \{1,\dots ,K\}\) we obtain
Now we show that \({\tilde{\lambda }}_i^n\) are bounded, i.e. \(|{\tilde{\lambda }}_i^n|\le M\) for \(i=1,\dots ,K\), \(n\in {\mathbb {N}}\) for some \(M\ge 0\). We have
If there exists \(k \in \{1,\dots ,K\}\) such that \(|{\tilde{\lambda }}_n^k|\rightarrow +\infty \), then \(\Vert \sum _{i=1}^K{\tilde{\lambda }}_i^n e_i^n\Vert ^2\rightarrow +\infty \), which contradicts \(\Vert \sum _{i=1}^K{\tilde{\lambda }}_i^n e_i^n\Vert ^2\rightarrow \Vert \sum _{i=1}^K{\tilde{\lambda }}_i e_i\Vert ^2= \Vert \sum _{i=1}^K\lambda _i u_i\Vert ^2<+\infty \).
Since \({\tilde{\lambda }}_i^n\) are bounded and \(e_i^n-e_i\rightarrow 0\), \(i=1,\dots ,K\) from (29) and (30) we conclude that \({\tilde{\lambda }}_k^n\rightarrow {\tilde{\lambda }}_k\) for any \(k\in \{1,2\dots ,K\}\).
For each \(n\in {\mathbb {N}}\) we have
Since \(u_i^n\), \(i=1,\dots ,K\) are linearly independent we have
Hence,
Since \(e_i^n\rightarrow e_i\) , \(u_i^n\rightarrow u_i\) and \({\tilde{\lambda }}_i^n\rightarrow {\tilde{\lambda }}_i\) for any \(i\in \{1,\dots ,K\}\) we obtain that \(\lambda _i^n \) converges for any \(i\in \{1,\dots ,K\}\) and \(\lambda _i^n\rightarrow \lambda _i\), \(i=1,\dots ,K\). \(\square \)
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Bednarczuk, E.M., Rutkowski, K.E. On Lipschitz Continuity of Projections onto Polyhedral Moving Sets. Appl Math Optim 84, 2147–2175 (2021). https://doi.org/10.1007/s00245-020-09706-y
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00245-020-09706-y
Keywords
- Lipschitzness of projection
- Relaxed constant rank constraint qualification condition
- Lipschitz-likeness
- Graphical subdifferential mapping