to appear in Journal of Mathematical Analysis and Applications, 2011
A merit function method for infinite-dimensional SOCCPs
Yungyen Chiang 1
Department of Applied Mathematics National Sun Yat-sen University
Kaohsiung 80424, Taiwan
Shaohua Pan 2
School of Mathematical Sciences South China University of Technology
Guangzhou 510640, China
Jein-Shan Chen 3 Department of Mathematics National Taiwan Normal University
Taipei 11677, Taiwan
July 25, 2009
(revised December 20, 2010)
Abstract. We introduce the Jordan product associated with the second-order cone IK into the real Hilbert spaceH, and then define a one-parametric class of complementarity functions Φt on H × H with the parameter t ∈ [0, 2). We show that the squared norm of Φt with t ∈ (0, 2) is a continuously F(r´echet)-differentiable merit function. By this, the second-order cone complementarity problem (SOCCP) in H can be converted into an unconstrained smooth minimization problem involving this class of merit functions, and furthermore, under the monotonicity assumption, every stationary point of this minimization problem is shown to be a solution of the SOCCP.
Key words: Hilbert space, complementarity, second-order cone, merit functions.
1The author’s work is partially supported by grants from the National Science Council of the Republic of China. E-mail : chiangyy@math.nsysu.edu.tw.
2The author’s work is supported by Guangdong Natural Science Foundation (No. 9251802902000001) and the Fundamental Research Funds for the Central Universities (SCUT). E-mail: shhpan@scut.edu.cn
3Member of Mathematics Division, National Center for Theoretical Sciences, Taipei Office.
The author’s work is partially supported by National Science Council of Taiwan. E-mail:
jschen@math.ntnu.edu.tw.
1 Introduction
LetH be a real Hilbert space endowed with an inner product ⟨·, ·⟩. The complementarity problem CP(K, T ) in H is, for any given closed convex cone K ⊆ H and a continuously F(r´echet)-differentiable mapping T :H → H, to find a vector x ∈ H such that
x∈ K, T (x) ∈ K∗ and ⟨x, T (x)⟩ = 0 (1) where K∗ := {x ∈ H | ⟨x, y⟩ ≥ 0 ∀y ∈ K} is the dual cone of K. A closed convex cone K in H is called self-dual if K coincides with its dual cone K∗; for example, the non-negative orthant cone IRn+ :={(x1, . . . xn) ∈ IRn | xj ≥ 0, j = 1, 2, . . . , n} and the second-order cone (also called Lorentz cone) IKn := {(r, x′) ∈ IR × IRn−1 | r ≥ ∥x′∥}.
This paper is concerned with the complementarity problem associated with the infinite- dimensional second-order cone IK inH which is closed, convex and self-dual (see Section 2 for its definition). The problem, denoted by CP(IK, T ), is to find an x∈ IK such that x∈ IK, T (x) ∈ IK and ⟨x, T (x)⟩ = 0. (2) This class of problems arises directly from the optimality conditions of certain types of infinite-dimensional optimization problems such as the one in [10], which is the refor- mulation of a min-max optimization problem with linear constraints in a Hilbert space.
Recently, nonlinear symmetric cone optimization and complementarity problems in finite-dimensional spaces such as semidefinite cone optimization and complementarity problems, second-order cone (SOC) optimization and complementarity problems, and general symmetric cone optimization and complementarity problems, become an active research field of mathematical programming. Taking SOC optimization and comple- mentarity problems for example, there have proposed many effective solution methods, including the interior point methods [2, 18, 21, 22], the smoothing Newton methods [5, 9, 11], the semismooth Newton methods [16, 19], and the merit function method [6, 3]. However, to our best knowledge, there are few works about nonlinear symmetric cone optimization and complementarity problems in infinite-dimensional spaces except [10], in which with the JB algebras of finite rank primal-dual interior-point methods are presented for some special type of infinite-dimensional cone optimization problems.
In this paper, we consider a merit function method for solving the problem CP(IK, T ).
The method aims to seek a smooth merit function Ψ :H × H → IR+ satisfying
Ψ(x, y) = 0 ⇐⇒ x ∈ IK, y ∈ IK, ⟨x, y⟩ = 0, (3) and reformulates the problem CP(IK, T ) as a smooth minimization problem
minx∈HΨ(x, T (x)) (4)
in the sense that x∗is a solution of CP(IK, T ) if and only if x∗solves (4) with zero optimal value. We call such Ψ a merit function associated with IK. Like handling complemen- tarity problems in finite-dimensional spaces, we seek a merit function associated with IK with a complementarity function (C-function for short) associated with IK. Specifically, a mapping Φ :H × H → H is called a C-function associated with IK if for any x, y ∈ H,
Φ(x, y) = 0 ⇐⇒ x ∈ IK, y ∈ IK and ⟨x, y⟩ = 0.
Clearly, the squared norm of Φ induces a merit function associated with IK.
WhenH is the Euclidean space IRn, the Fischer-Burmeister (FB) and natural residual (NR) C-functions associated with the SOC IKn [9] are respectively defined as
ΦFB(x, y) := (x2+ y2)1/2− (x + y) ∀x, y ∈ IRn (5) and
ΦNR(x, y) := x− (x − y)+ ∀x, y ∈ IRn, (6) where x2 = x•x with “•” means the Jordan product in IRn, x1/2with x∈ IKn is a vector such that x1/2• x1/2 = x, and (x)+ denotes the projection onto IKn. The function ΦFB was well-studied in [6, 20], and particularly its squared norm was shown to be a smooth merit function in [6]. Since the squared norm of ΦNR is not differentiable, it is often involved in the smoothing methods for the SOCCPs [5, 11]. The above two C-functions are subsumed in Kanzow and Kleinmichel C-function associated with IKn:
Φt(x, y) :=[
(x− y)2+ 2tx• y]1/2
− (x + y) ∀x, y ∈ IRn (7) where t is an arbitrary but fixed real number from [0, 2). This function was studied in [4]
and its squared norm with t∈ (0, 2) was shown to be continuously differentiable. Note that, when n = 1, ΦFB, ΦNR and Φt reduce to the FB NCP-function [8], the minimum function [14], and the Kanzow and Kleinmichel NCP-function [12], respectively.
To define these C-functions in the Hilbert spaceH, we introduce the Jordan product associated with the cone IK, and extend the Kanzow and Kleinmichel C-function defined in (7) toH and show that it satisfies the property (3) for each t ∈ [0, 2). In Section 4, we prove that the squared norm of this class of C-functions with t∈ (0, 2) are continuously F-differentiable in H × H. Note that the corresponding results in [4, 6] were proved by the spectral factorization of vectors, but here we shall not formally use this concept. In Section 5, under the monotonicity assumption, we establish that every stationary point of the unconstrained minimization problem involving this class of merit functions is a solution of CP(IK, T ), which generalizes the results of [4, Prop.4.1] and [6, Prop.3].
Throughout this paper, ∥ · ∥ denote the norm induced by the inner product ⟨·, ·⟩ in H. For any given Banach spaces X and Y, let L(X , Y) denote the Banach space of all
continuous linear mappings from X into Y. We simply write L(X , X ) = L(X ) and de- note GL(X ) by the set of all invertible mappings in L(X ). The norm of any l ∈ L(X , Y) is defined by ∥l∥ := sup{∥l(x)∥ | x ∈ X and ∥x∥ = 1}. In addition, for any self-adjoint linear operator l from X → X , we write l ≻ 0 (respectively, l ≽ 0) to mean that l is positive definite (respectively, positive semidefinite).
2 Lorentz cone and Jordan product
This section is devoted to introducing the Lorentz cone IK mentioned above which is the unique self-dual cone in a family of pointed closed convex cones K in H. Every cone in K is the image of IK under some mapping in GL(H). Associated with the self-dual closed convex cone, the Jordan product is introduced into the Hilbert space H.
For every integer n > 1, the Lorentz cone IKn given in Section 1 can be written as IKn:=
{
x∈ IRn | ⟨x, e⟩ ≥ 1
√2∥x∥
}
with e = (1, 0)∈ IR × IRn−1.
This motivates us to consider the following closed convex cone in the Hilbert space H:
K(e, r) :=
{
x∈ H | ⟨x, e⟩ ≥ r∥x∥}
where e∈ H with ∥e∥ = 1 and r ∈ IR with 0 < r < 1. Observe that K(e, r) is pointed, i.e., K(e, r)∩ (−K(e, r)) = {0}. Let ⟨e⟩⊥ :={
x∈ H | ⟨x, e⟩ = 0}
. Then any x∈ H can be written as x = x′ + λe with x′ ∈ ⟨e⟩⊥ and λ ∈ IR. By noting that
⟨x, e⟩ ≥ r∥x∥ ⇐⇒ λ ≥ r(∥x′∥2+ λ2)1/2 ⇐⇒ λ ≥ r
√1− r2∥x′∥,
the closed convex cone K(e, r) can be expressed as K(e, r) =
{
x′+ λe∈ H | x′ ∈ ⟨e⟩⊥ and λ≥ r
√1− r2 ∥x′∥ }
.
Proposition 2.1 For any unit vector e∈ H and 0 < r < 1, the dual cone of K(e, r) is K(e,√
1− r2). Hence, the cone K (
e,√1 2
)
={
x′+ λe∈ H | λ ≥ ∥x′∥}
is self-dual.
Proof. Let x = x′+ λe∈ K(e,√
1− r2) and y = y′+ µe ∈ K(e, r) be arbitrary. Since λµ≥ ∥x′∥ · ∥y′∥, we have ⟨x, y⟩ ≥ ⟨x′, y′⟩ + ∥x′∥ · ∥y′∥ ≥ 0. This proves that
K(e,√
1− r2)⊂ K∗(e, r).
Conversely, let x = x′+ λe∈ K∗(e, r) be arbitrary, and we will prove x∈ K(e,√
1− r2), i.e., λ≥ r−1√
1− r2 ∥x′∥. This is trivial when x′ = 0. When x′ ̸= 0, by considering the element v =−r−1√
1− r2x′+∥x′∥ e of K(e, r), we have 0≤ ⟨x, v⟩ =(
λ− r−1√
1− r2∥x′∥)
∥x′∥, which implies the result. The proof is complete. 2
Note that the unit vector e ∈ H is not unique. Every unit vector e determines a Lorentz cone K(e,√1
2). In this work, we consider a fixed unit vector e and write IK = K
( e, 1
√2 )
= {
x′+ λe∈ H | λ ≥ ∥x′∥} .
Unless stated otherwise, we shall alternatively write any x ∈ H as x = x′ + λe with x′ ∈ ⟨e⟩⊥ and λ = ⟨x, e⟩. This expression is needed for stating many results and sim- plifying the computation in the subsequent analysis. In addition, for any x, y ∈ H, we shall write x≻IK y (respectively, x≽IK y) if x− y ∈ intIK (respectively, x − y ∈ IK).
Next we show that the solution sets of complementarity problems associated with any K(e, r) are related to those associated with IK via the mappings in GL(H).
Lemma 2.1 For any given 0 < r, s < 1, let Λ(r,s):H → H be the mapping defined by Λ(r,s)(x′+ λe) :=
√1− s2
√1− r2 x′+ sλ
r e ∀x′+ λe∈ H.
Then, the following statements hold.
(a) Λ(r,s) ∈ GL(H) with Λ−1(r,s)= Λ(s,r), and Λ(r,s) maps K(e, r) onto K(e, s).
(b) Let Λr := Λ(r,√1
2). If r2 + s2 = 1, then ⟨Λr(x), Λs(y)⟩ = 2rs1 ⟨x, y⟩ for all x, y ∈ H.
Proof. (a) It is clear that Λ(r,s) is linear and Λ−1(r,s) = Λ(s,r). For x′ ∈ ⟨e⟩⊥ and λ∈ IR,
∥Λ(r,s)(x′+ λe)∥2 = 1− s2
1− r2 ∥x′∥2+ s2
r2 λ2 ≤ max
{ 1− s2 1− r2, s2
r2 }
∥x′+ λe∥2. This proves the continuity of Λ(r,s). Also, Λ(r,s) maps K(e, r) onto K(e, s) by noting that
x′+ λe∈ Λ(r,s)(K(e, r)) ⇐⇒ Λ(s,r)(x′ + λe)∈ K(e, r)
⇐⇒ rλ
s ≥ r
√1− r2 ·
√1− r2
√1− s2 ∥x′∥
⇐⇒ λ ≥ s
√1− s2 ∥x′∥.
(b) We write x = x′+ λe and y = y′+ µe. Then, Λr(x′+ λe) = 1
√2(1− r2) x′ + λ
√2r e = 1
√2s x′+ λ
√2r e ;
Λs(y′ + µe) = 1
√2r y′+ µ
√2(1− r2) e = 1
√2r y′+ µ
√2s e . Now, the assertion follows immediately by a direct computation. 2
From Lemma 2.1, we immediately obtain the following proposition.
Proposition 2.2 Let 0 < r, s < 1 be such that r2+ s2 = 1, and T :H → H be given.
(a) A point x ∈ H solves the problem CP(K(e, r), T ) if and only if Λr(x) solves the problem CP(IK, Λs◦ T ◦ Λ−1r ).
(b) If Φ :H × H → H is a C-function associated with IK, then the mapping Φr(x, y) :=
Φ(Λr(x), Λs(y)) is a C-function associated with K(e, r).
Next we introduce the Jordan product associated with the Lorentz cone IK. For any x = x′+ λe∈ H and y = y′+ µe ∈ H, we define the Jordan product of x and y by
x• y := (µx′+ λy′) +⟨x, y⟩e, (8) and write x2 = x • x. Clearly, when H = IRn and e = (1, 0) ∈ IR × IRn−1, this definition is same as the one given by [7, Chapter II]. By the definition in (8) and a direct computation, it is easy to verify that the following properties hold.
Property 2.1 (i) x• y = y • x and x • e = x for all x, y ∈ H.
(ii) (x + y)• z = x • z + y • z for all x, y, z ∈ H.
(iii) ⟨x, y • z⟩ = ⟨y, x • z⟩ = ⟨z, x • y⟩ for all x, y, z ∈ H.
(iv) For any x = x′+ λe∈ H, x2 = x• x = 2λx′+∥x∥2e∈ IK and ⟨x2, e⟩ = ∥x∥2. (v) If x = x′+ λe∈ IK, then there is a unique x1/2 ∈ IK such that (x1/2)2 = x, where
x1/2 =
{ 0 if x = 0;
x′/(2τ ) + τ e othewise with τ =
√
λ +√
λ2− ∥x′∥2
2 . (9)
(vi) Every x = x′+ λe∈ H with λ2− ∥x′∥2 ̸= 0 is invertible w.r.t. the Jordan product, i.e., there is a unique point x−1 ∈ H such that x • x−1 = e, where
x−1 = −x′ + λe
λ2− ∥x′∥2. (10)
Moreover, x∈ intIK if and only if x−1 ∈ intIK.
Associated with every x ∈ H, we define a linear mapping Lx from H to H by
Lxy := x• y for any y ∈ H. (11)
Clearly, Lx ∈ L(H). Also, the mapping possesses the following favorable properties.
Lemma 2.2 For any x∈ H, let Lx ∈ L(H) be defined as above. Then, we have (a) x≻IK 0⇐⇒ Lx ≻ 0 and x ≽IK 0⇐⇒ Lx ≽ 0.
(b) If x = x′+ λe with λ̸= 0 and |λ| ̸=∥x′∥, then Lx∈ GL(H) with the inverse given by L−1x y = λ−1(
y′− ⟨x−1, y⟩x′)
+⟨x−1, y⟩e for any y = y′+ µe ∈ H. (12) Proof. (a) Fix any x = x′+ λe∈ H. It suffices to prove the first equivalence, and the second equivalence follow from the first equivalence and the closedness of IK. Note that Lx ≻ 0 if and only if ⟨h, Lxh⟩ > 0 for any h = h′+ ξe ∈ H\{0}, whereas
⟨h, Lxh⟩ > 0 ⇐⇒ λ∥h′∥2+ 2ξ⟨x′, h′⟩ + λξ2 > 0
⇐⇒ λ > 0 and 4⟨x′, h′⟩2− 4λ2∥h′∥2 < 0
⇐⇒ λ > 0 and ∥x′∥ < λ.
(b) To prove Lx∈ GL(H), it suffices to prove that Lxy = 0 for some y = y′+ µe ∈ H implies y = 0. Indeed, since Lxy = 0 implies ∥x • y∥2 = 0, which is equivalent to
λy′+ µy′ = 0 and ⟨x′, y′⟩ + λµ = 0.
Since λ ̸= 0, from the first equality we have y′ = −λ−1µx′. Substituting it into the second equality yields µ = 0, and so y′ = 0. A direct computation verifies (12). 2
3 Kanzow-Kleinmichel merit function
In this section, we will extend Kanzow-Kleinmichel C-function in (7) to the real Hilbert spaceH, and present some technical lemmas that will be used in the subsequent analysis.
Let t be an arbitrary real number in [0, 2). Define the mapping Φt:H × H → H by Φt(x, y) := [
(x− y)2+ 2t(x• y)]1/2
− (x + y). (13)
Note that, for any t∈ [0, 2) and any x, y ∈ H,
(x− y)2+ 2t(x• y) = (x + (t − 1)y)2 + t(2− t)y2 ∈ IK. (14)
Hence, the function Φt is well-defined. It is easy to see that when t = 1 and t = 0, Φt reduces to the FB and the NR C-function associated with IK, respectively.
To show that each Φtis a C-function associated with IK, we need the following result which is an infinitely dimensional version of [9, Prop.2.1]. The proof given in [9] was based on the geometry of vectors in Euclidean spaces, that is, the notion of an angle between vectors. We here give another proof without using this notion.
Lemma 3.1 For any x, y∈ H, the following statements are equivalent:
(a) x∈ IK, y ∈ IK and ⟨x, y⟩ = 0;
(b) x∈ IK, y ∈ IK and x • y = 0;
(c) x + y∈ IK and x • y = 0.
(d) It holds that (i) x = 0, y ∈ IK; or (ii) x ∈ IK, y = 0; or (iii) x ∈ ∂IK, y ∈ ∂IK and
⟨x, y⟩ = 0, where ∂IK := {x′+ λe∈ H | λ = ∥x′∥} denotes the boundary of IK.
Proof. Clearly, (b)⇒ (c) and (d) ⇒ (a). We need to prove (a) ⇒ (b) and (c) ⇒ (d).
(a) ⇒ (b). Write x = x′ + λe and y = y′ + µe. By (8) and ⟨x, y⟩ = 0, we have x• y = (µx′+ λy′). Since λ≥ ∥x′∥ and µ ≥ ∥y′∥ by x, y ∈ IK, it follows that
∥µx′+ λy′∥2 = µ2∥x′∥2− 2λ2µ2+ λ2∥y′∥2 ≤ 0,
and µx′+ λy′ = 0 follows. Thus, we obtain x• y = 0, and hence (a) implies (b).
(c) ⇒ (d). Since x • y = 0 implies ∥(µx′+ λy′) +⟨x, y⟩e∥2 =∥µx′+ λy′∥2+⟨x, y⟩2 = 0, we have ⟨x, y⟩ = 0 and µx′ + λy′ = 0. If λ = 0, µ ̸= 0, then from µx′ + λy′ = 0 and
⟨x, y⟩ = 0, we get x′ = 0, and then x = 0. Together with x + y ∈ IK, we obtain y ∈ IK, and so Case (i) holds. If λ ̸= 0, µ = 0, a similar argument yields that Case (ii) holds.
If λ = µ = 0, then from x + y ∈ IK it follows that ∥x′ + y′∥ = 0. This along with
⟨x, y⟩ = 0 and λ = 0, µ = 0 yields that x′ = 0 and y′ = 0, and consequently, x = y = 0.
Hence, Cases (i), (ii) and (iii) hold. Now, assume that λµ ̸= 0. From µx′ + λy′ = 0 and ⟨x, y⟩ = 0, we obtain λ2 = ∥x′∥2 and µ2 = ∥y′∥2. This, together with x + y ∈ IK, i.e. (λ + µ)2 ≥ ∥x′ + y′∥2, implies λµ ≥ ⟨x′, y′⟩ = −λµ, and hence λµ > 0. Since λ + µ≥ ∥x′+ y′∥, we get λ > 0 and µ > 0. Thus, λ = ∥x′∥ and µ = ∥y′∥, which implies that x, y ∈ IK. That is, Case (iii) follows. 2
Let Ψt:H × H → IR+ denote the squared norm of the function Φt, that is,
Ψt(x, y) :=∥Φt(x, y)∥2 ∀x, y ∈ H. (15)
From the expression of Φt and Lemma 3.1, it follows that
Ψt(x, y) = 0⇔ Φt(x, y) = 0 ⇔ x + y ∈ IK and (x − y)2+ 2t(x• y) = (x + y)2
⇔ x + y ∈ IK and x • y = 0
⇔ x ∈ IK, y ∈ IK, ⟨x, y⟩ = 0.
These equivalence immediately implies the following result.
Proposition 3.1 The functions Φt and Ψt are respectively a C-function and a merit function associated with IK.
In what follows, we provide some necessary technical lemmas that will be used later.
Lemma 3.2 For any given 0 < t < 2, x = x′+ λe∈ H and y = y′+ µe ∈ H, we have (x− y)2+ 2t(x• y) ∈ ∂IK ⇐⇒ x2+ y2 ∈ ∂IK
⇐⇒ |λ| = ∥x′∥, |µ| = ∥y′∥, λµ = ⟨x′, y′⟩ (16)
=⇒ λy′ = µx′
Proof. Using|λ| = ∥x′∥, |µ| = ∥y′∥ and λµ = ⟨x′, y′⟩, it is easy to verify ∥λy′−µx′∥2 = 0.
So, the implication in (16) holds. Now we prove the second equivalence. Noting that x2+ y2 = 2(λx′ + µy′) + (∥x∥2+∥y∥2)e,
2∥λx′+ µy′∥ ≤ 2∥λx′∥ + 2∥µy′∥ ≤ ∥x∥2+∥y∥2,
we have x2+ y2 ∈ ∂IK if and only if ∥x∥2+∥y∥2 = 2∥λx′∥ + 2∥µy′∥ = 2∥λx′+ µy′∥, i.e., (|λ| − ∥x′∥)2+ (|µ| − ∥y′∥)2 = 0 and ∥λx′∥ + ∥µy′∥ = ∥λx′+ µy′∥. Thus, we have
x2+ y2 ∈ ∂IK ⇐⇒ |λ| = ∥x′∥, |µ| = ∥y′∥, λµ⟨x′, y′⟩ = |λµ| · ∥x′∥ · ∥y′∥.
We may argue that, when |λ| = ∥x′∥ and |µ| = ∥y′∥, there holds that λµ⟨x′, y′⟩ = |λµ| · ∥x′∥ · ∥y′∥ ⇐⇒ λµ = ⟨x′, y′⟩.
Indeed, if the equality on the right hand side holds, then λµ⟨x′, y′⟩ = λ2µ2 = |λµ| ·
∥x′∥ · ∥y′∥, which implies the equality of the left hand side. Assume that the equality of the left hand side holds. If λµ = 0 or ∥x′∥ · ∥y′∥ = 0, then x = 0 or y = 0, and thus λµ = 0 =⟨x′, y′⟩; and if λµ ̸= 0 and ∥x′∥·∥y′∥ ̸= 0, using λµ⟨x′, y′⟩ = |λµ|·∥x′∥·∥y′∥ > 0 then yields that |⟨x′, y′⟩| = ∥x′∥ · ∥y′∥ = |λµ| and ⟨x′, y′⟩ = λµ. This proves that the equality of the right hand side holds, and the second equivalence in (16) follows.
To establish the first equivalence in (16), it suffices to prove that
(x− y)2+ 2t(x• y) ∈ ∂IK ⇐⇒ |λ| = ∥x′∥, |µ| = ∥y′∥, λµ = ⟨x′, y′⟩. (17)
Recall that (x− y)2+ 2t(x• y) = (x + (t − 1)y)2+ (√
t(2− t) y)2. By the result above, (x− y)2+ 2t(x• y) ∈ ∂IK ⇐⇒ |λ + (t − 1)µ| = ∥x′+ (t− 1)y′∥, |µ| = ∥y′∥,
µ(λ + (t− 1)µ) = ⟨x′+ (t− 1)y′, y′⟩.
Taking into account that |µ| = ∥y′∥ implies the following equivalences
|λ + (t − 1)µ| = ∥x′+ (t− 1)y′∥ ⇐⇒ λ2 + 2(t− 1)λµ = ∥x′∥2+ 2(t− 1)⟨x′, y′⟩, µ(λ + (t− 1)µ) = ⟨x′+ (t− 1)y′, y′⟩ ⇐⇒ λµ = ⟨x′, y′⟩,
we immediately obtain (17). Thus, the proof is complete. 2
The following lemma is essentially proved in [6, Lemma 3]. We give a simpler proof.
Lemma 3.3 For j = 1, 2, let xj = x′j+ λje∈ H. If λ1x′1+ λ2x′2 ̸= 0, then for j = 1, 2, (
λj + (−1)j
⟨ λ1x′1+ λ2x′2
∥λ1x′1+ λ2x′2∥, x′j
⟩)2
≤
x′j + (−1)jλj λ1x′1+ λ2x′2
∥λ1x′1+ λ2x′2∥ 2
≤ ∥x1∥2+∥x2∥2+ 2(−1)j∥λ1x′1+ λ2x′2∥.
Proof. It suffices to prove the inequalities for j = 1. The first inequality holds trivially since |⟨v, w⟩| ≤ ∥v∥ · ∥w∥ for all v, w ∈ H. The second inequality is proved as follows.
x′1− λ1
λ1x′1+ λ2x′2
∥λ1x′1+ λ2x′2∥
2 = ∥x′1∥2 − 2
∥λ1x′1+ λ2x′2∥⟨λ1x′1, λ1x′1+ λ2x′2⟩ + λ21
= ∥x1∥2 − 2∥λ1x′1+ λ2x′2∥ +2⟨λ2x′2, λ1x′1 + λ2x′2⟩
∥λ1x′1+ λ2x′2∥
≤ ∥x1∥2 − 2∥λ1x′1+ λ2x′2∥ + 2|λ2|∥x′2∥
≤ ∥x1∥2 − 2∥λ1x′1+ λ2x′2∥ + ∥x2∥2,
where the last inequality is using∥x2∥2 = λ22+∥x′2∥2. Thus, the proof is complete. 2
To end the contents of this section, we recall the concept of F(r´echet)-differentiability and present some continuously F-differentiable mappings for later use. For given Banach spaces X and Y, a mapping f from a nonempty open subset X of X into Y is said to be F-differentiable at x∈ X if there exists lx ∈ L(X , Y) such that
h→0lim
f (x + h)− f(x) − lxh
∥h∥ = 0,
and lx is called the F-differential of f at x, written by f′(x). When f is F-differentiable at every point of X, we say that f is F-differentiable on X. If f is F-differentiable
on a neighborhood U ⊂ X of a point x0 ∈ X, and if, as a mapping from U into the Banach space L(X , Y), the mapping x 7−→ f′(x) is continuous at x0, then f is said to be continuously F-differentiable at x0. The mapping f is called continuously F-differentiable on X if it is continuously F-differentiable at every point of X. Note that if f ∈ L(X , Y), then f is continuously F-differentiable on X with f′(x) = f for every x ∈ X , i.e., f′(x)v = f (v) for all v∈ X . By the definition, it is easy to verify the continuous F-differentiability of the mappings given below.
Example 3.1 (i) f (x) =⟨x, e⟩ for any x ∈ H with f′(x)v =⟨v, e⟩ for all v ∈ H.
(ii) f (x) = x− ⟨x, e⟩e for any x ∈ H with f′(x)v = v− ⟨v, e⟩e for all v ∈ H.
(iii) f (x) = x2 = x• x for any x ∈ H with f′(x)v = 2x• v for all v ∈ H.
(iv) f (x) =∥x∥2 for any x∈ H with f′(x)v = 2⟨x, v⟩ for all x, v ∈ H.
(v) f (x) =∥x∥ = ⟨x, x⟩1/2 for any x∈ H. Such f is continuously F-differentiable only on H \ {0} with f′(x)v = ∥x∥1 ⟨x, v⟩ for all v ∈ H.
4 Smoothness of merit function
This section is devoted to establishing the continuous F-differentiability (smoothness) of Ψt. For this purpose, we first investigate the F-differentiability of two special mappings defined as in the following two lemmas, respectively.
Lemma 4.1 Let σ(x) := x1/2 for any x∈ IK. Then, the following statements hold.
(a) σ is continuously F-differentiable on intIK, and for all v∈ H,
σ′(x)v =
√λ2 − ∥x′∥2
2τ ⟨x−1/2, v⟩x−1/2+v− ⟨v, e⟩e 2τ where τ is given as in (9).
(b) For every x∈ intIK, 2σ′(x)v = L−1σ(x)v for all v ∈ H.
(c) For every x ∈ intIK, the F-differential σ′(x) is a self-adjoint operator in L(H), i.e., ⟨σ′(x)v, w⟩ = ⟨v, σ′(x)w⟩ for all v, w ∈ H.
Proof. (a) Recall that σ(x) = 2τx′ + τ e for x = x′+ λe∈ IK \ {0}. Since τ as a mapping of x ∈ IK \ {0} is F-differentiable on intIK, the function σ is F-differentiable on intIK.
The differential of σ is computed as follows. Taking into account 2τ2 = λ +√
λ2− ∥x′∥2, by Example 3.1 it is not hard to calculate that for all v ∈ H,
4τ τ′(x)v = ⟨v, e⟩ + λ⟨v, e⟩ − ⟨v, x√ ′⟩
λ2− ∥x′∥2 = ⟨v, 2τ√ 2e− x′⟩ λ2− ∥x′∥2, and consequently,
τ′(x)v = 1 2√
λ2− ∥x′∥2
⟨
τ e− x′ 2τ, v
⟩
= 1 2
⟨x−1/2, v⟩ .
Together with the expression σ(x) = 2τx′ + τ e, we obtain that σ′(x)v = −τ′(x)v
2τ2 x′+ 1
2τ (v− ⟨v, e⟩e) + (τ′(x)v) e
=
⟨x−1/2, v⟩ 2τ
(−x′ 2τ + τ e
) + 1
2τ (v− ⟨v, e⟩e) (18)
=
⟨x−1/2, v⟩ 2τ ·√
λ2− ∥x′∥2x−1/2+ 1
2τ (v− ⟨v, e⟩e) .
We next prove that the F-differential σ′ is continuous at any given point a = a′+ αe ∈ intIK. For any x = x′+ λe∈ intIK, we write
τ (x) = τ =
√
λ +√
λ2− ∥x′∥2
2 and p(x) =
√λ2− ∥x′∥2 2τ (x) . Then, from the last equality in (18), it follows that for all v ∈ H,
∥σ′(x)v− σ′(a)v∥
≤ p(x)⟨x−1/2, v⟩x−1/2− p(a)⟨a−1/2, v⟩a−1/2 + 1
2τ (x)− 1 2τ (a)
· ∥v − ⟨v, e⟩e∥
≤ |p(x) − p(a)| · ⟨x−1/2, v⟩ · x−1/2 + p(a)· ⟨x−1/2− a−1/2, v⟩ · x−1/2 +p(a)· ⟨a−1/2, v⟩ · x−1/2− a−1/2 +
1
2τ (x) − 1 2τ (a)
· ∥v∥
≤ |p(x) − p(a)| · ∥x−1/2∥2· ∥v∥ + 1
2τ (x) − 1 2τ (a)
· ∥v∥
+p(a)· x−1/2− a−1/2 ( x1/2 + a−1/2 ) · ∥v∥.
This implies that
∥σ′(x)− σ′(a)∥ ≤ |p(x) − p(a)| · x−1/2 2+ 1
2τ (x) − 1 2τ (a)
+p(a)· x−1/2− a−1/2 (∥x1/2∥ + ∥a−1/2∥)
,
and consequently ∥σ′(x)− σ′(a)∥ → 0 as x → a.
(b) From the second equality in (18) and equation (12), we obtain for any v = v′+θe∈ H, 2σ′(x)v = 1
τ⟨σ(x)−1, v⟩ · (−x′
2τ + τ e )
+v′ τ
= 1
τ (
v′− ⟨σ(x)−1, v⟩ · x′ 2τ
)
+⟨σ(x)−1, v⟩e = L−1σ(x)v.
(c) For any given v, w ∈ H, we write σ′(x)v = v1 and σ′(x)w = w1. Then, by part (b), we have v = 2σ(x)• v1 and w = 2σ(x)• w1, and consequently
⟨σ′(x)v, w⟩ = 2⟨v1, σ(x)• w1⟩ = 2⟨σ(x) • v1, w1⟩ = ⟨v, σ′(x)w⟩.
This shows that σ′(x) is a self-adjoint operator. The proof is completed. 2
Lemma 4.2 For any x, y ∈ H and r ∈ IR, let ψr(x, y) := 2⟨(x2+ y2)1/2, x + ry⟩. Then, (a) ψr is F-differentiable at every point (a, b)∈ H × H with a2+ b2 ∈ ∂IK.
(b) For any given x = x′ + λe∈ H and y = y′+ µe ∈ H with x2+ y2 ∈ ∂IK \ {0}, ψ′r(x, y)(v, w) = 2(λ + rµ)
√λ2+ µ2 · (⟨v, x⟩ + ⟨w, y⟩) + 2⟨(
x2+ y2)1/2
, v + rw
⟩
for all v, w∈ H. Furthermore, ∥ψ′r(x, y)∥ ≤ 4(1 + |r|)√
∥x∥2+∥y∥2.
Proof. (a) For any (x, y) ̸= (0, 0), it can be seen that ψr is F-differentiable at (0, 0) since
|ψr(x, y)− ψr(0, 0)| = 2 ⟨(x2 + y2)1/2, x + ry⟩ ≤ 2√
∥x∥2+∥y∥2· ∥x + ry∥.
Next, we consider the case where (a, b) ̸= (0, 0). Write a = a′ + αe and b = b′ + βe.
Since a2 + b2 ∈ ∂IK, we have 2∥αa′ + βb′∥ = ∥a∥2 +∥b∥2 > 0. So, there exist a convex and bounded open neighborhood U of (a, b) in H × H and a constant ρ > 0 such that
∥λx′+ βy′∥ ≥ ρ for any (x, y) ∈ U with x = x′+ λe and y = y′+ µe. Notice that
(x2+ y2)1/2 = λx′+ µy′
τ (x, y) + τ (x, y)e where
τ (x, y) =
√
∥x∥2+∥y∥2+√
(∥x∥2+∥y∥2)2− 4∥λx′ + µy′∥2
2 .
Write
τj = τj(x, y) := ∥x∥2 +∥y∥2+ 2(−1)j∥λx′+ µy′∥ for j = 1, 2.
It is not difficult to verify that τ (x, y) =
√τ1+√ τ2
2 and 1
τ (x, y) =
√τ2− √τ1
2∥λx′+ µy′∥. (19) Consequently,
ψr(x, y) = 2
⟨λx′+ µy′
τ (x, y) , x′ + ry′
⟩
+ 2τ (x, y)(λ + rµ)
= (√
τ2−√ τ1)
⟨ λx′+ µy′
∥λx′+ µy′∥, x′ + ry′
⟩ + (√
τ1+√
τ2) (λ + rµ) := φ1(x, y) + φ2(x, y)
where
φj(x, y) :=
√
τj(x, y) (
λ + rµ + (−1)j
⟨ λx′+ µy′
∥λx′+ µy′∥, x′+ ry′
⟩)
for j = 1, 2.
Since λx′+ µy′ ̸= 0 for any (x, y) ∈ U, the mappings
(x, y)7−→ ∥λx′+ µy′∥ and (x, y) 7−→ ∥λx′ + µy′∥−1 are continuously F-differentiable on U , and then√
τ2(x, y) is continuously F-differentiable on U since τ2(x, y) > 0 for (x, y)̸= (0, 0). Hence, φ2is continuously Fr´echet differentiable on U . To prove that φ1 is F-differentiable at (a, b), we let
f (x, y) := λ + rµ, g(x, y) := λx′+ µy′, p(x, y) := g(x, y)
∥g(x, y)∥, h(x, y) := x′+ ry′, φ3(x, y) := f (x, y)− ⟨p(x, y), h(x, y)⟩
for any (x, y)∈ U with x = x′+ λe and y = y′+ µe. Then, τ1(x, y) = ∥x∥2+∥y∥2− 2∥g(x, y)∥ and φ1(x, y) = √
τ1(x, y) φ3(x, y). (20) By Example 3.1, it is not hard to calculate that for any (v, w)∈ H × H,
f′(x, y)(v, w) = ⟨v, e⟩ + r⟨w, e⟩;
g′(x, y)(v, w) = λv +⟨v, e⟩(x′− λe) + µw + ⟨w, e⟩(y′ − µe);
p′(x, y)(v, w) = g′(x, y)(v, w)
∥g(x, y)∥ −⟨g′(x, y)(v, w), g(x, y)⟩
∥g(x, y)∥3 g(x, y), h′(x, y)(v, w) = v− ⟨v, e⟩e + rw − r⟨w, e⟩e.
Note that ∥g(x, y)∥ ≥ ρ for all (x, y) ∈ U. By the boundedness of U, there is a constant c > 0 such that ∥x∥ + ∥y∥ ≤ c for all (x, y) ∈ U. Thus, for any (x, y) ∈ U and any (v, w)∈ H × H, from the last four equalities it follows that
∥f′(x, y)(v, w)∥ ≤ (|r| + 1)(∥v∥ + ∥w∥), ∥h′(x, y)(v, w)∥ ≤ (|r| + 1)(∥v∥ + ∥w∥),
∥g′(x, y)(v, w)∥ ≤ 2c(∥v∥ + ∥w∥), ∥p′(x, y)(v, w)∥ ≤ 4c(∥v∥ + ∥w∥)
ρ .
Consequently,
∥τ1′(x, y)(v, w)∥ =
2⟨x,v⟩ + 2⟨y,w⟩ − 2⟨g′(x, y)(v, w), g(x, y)⟩
∥g(x, y)∥
,
≤ 2c(∥v∥ + ∥w∥) + 2∥g′(x, y)(v, w)∥ ≤ 6c(∥v∥ + ∥w∥),
|φ′3(x, y)(v, w)| ≤ ∥f′(x, y)(v, w)∥ + ∥p′(x, y)(v, w)∥ · ∥h(x, y)∥ + ∥h′(x, y)(v, w)∥
≤ M1(∥v∥ + ∥w∥)
where M1 = 2(|r| + 1) + 4ρ−1c2(|r| + 1). By the mean-value theorem, for any given (x, y)∈ U, there exists (¯x, ¯y) ∈ U on the line segment joining (a, b) to (x, y) such that
|φ3(x, y)− φ3(a, b)| = |φ′3(¯x, ¯y)(x− a, y − b)| ≤ M1(∥x − a∥ + ∥y − b∥).
We claim that φ3(a, b) = 0. To see this, from Lemma 3.2, |α| = ∥a′∥, |β| = ∥b′∥ and αβ =⟨a′, b′⟩, which implies that ∥αa′+ βb′∥ = α2+ β2, and
φ3(a, b) = α + rβ− 1
α2+ β2(α∥a′∥2+ (rα + β)⟨a′, b′⟩ + rβ∥b′∥2)
= α + rβ− 1
α2+ β2(α3+ rα2β + αβ2 + rβ3) = 0.
This claim implies that
|φ3(x, y)| ≤ M1(∥x − a∥ + ∥y − b∥) for any (x, y) ∈ U. (21) In addition, noting that τ1(a, b) = 0 and applying the Mean Value Theorem to τ1,
√τ1(x, y)≤ M2·√
∥x − a∥ + ∥y − b∥ for any (x, y) ∈ U, (22) where M2 =√
6c. Now from equations (20)–(22) it follows that, for any (x, y)∈ U,
|φ1(x, y)− φ2(a, b)| = |φ1(x, y)| ≤ M1M2(∥x − a∥ + ∥y − b∥)3/2,
which says that φ1 is F-differentiable at (a, b) with φ′1(a, b) being the zero mapping in L(H × H, IR). So, ψr is F-differentiable at (a, b) with ψr′(a, b) = φ′2(a, b).
(b) From part (a), we know that ψ′r(x, y) = φ′2(x, y). To compute φ′2(x, y), we write φ4(x, y) := f (x, y) +⟨p(x, y), h(x, y)⟩ = λ + rµ +
⟨ λx′+ µy′
∥λx′+ µy′∥, x′+ ry′
⟩ .
