Vol. 59, No. 5, July 2010, 661–676
Lipschitz continuity of the gradient of a one-parametric class of SOC merit functions
Jein-Shan Chena* and Shaohua Panb
aDepartment of Mathematics, National Taiwan Normal University, Taipei, Taiwan;bSchool of Mathematical Sciences, South China
University of Technology, Guangzhou, China
(Received 27 March 2007; final version received 11 March 2008)
In this article, we show that a one-parametric class of SOC merit functions has a Lipschitz continuous gradient; and moreover, the Lipschitz constant is related to the parameter in this class of SOC merit functions. This fact will lay a building block when the merit function approach as well as the Newton-type method are employed for solving the second-order cone complementarity problem with this class of merit functions.
Keywords: second-order cone; merit function; spectral factorization; Lipschitz continuity
AMS Subject Classifications: 26B05; 26B35; 90C33; 65K05
1. Introduction
A well-known approach to solving the non-linear complementarity problem (NCP) is to reformulate it as the global minimization via a certain merit function over IRn. For the approach to be effective, the choice of the merit function is crucial. A popular choice is the squared norm of the Fischer–Burmeister (FB) function FB: IRnIRn!IRþdefined by
FBða, bÞ :¼1 2
Xn
i¼1
FBðai, biÞ
½ 2 ð1Þ
for all a ¼ (a1, . . . , an)T2IRn and b ¼ (b1, . . . , bn)T2IRn, where FB: IR IR ! IR is the Fischer–Burmeister NCP function given as
FBðai, biÞ ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi a2i þb2i q
aibi: ð2Þ
It has been shown that FB enjoys many desirable properties [8,9], for example, smoothness (continuous differentiability). This merit function and its analysis were subsequently extended by Tseng [18] to the semidefinite complementarity problem (SDCP)
*Corresponding author. Email: jschen@math.ntnu.edu.tw
ISSN 0233–1934 print/ISSN 1029–4945 online
2010 Taylor & Francis DOI: 10.1080/02331930802180319 http://www.informaworld.com
although only differentiability, not continuous differentiability, was established. In fact, the FB function for the SDCP is the matrix-valued function FB: Sn Sn! Sndefined by
FBðX, YÞ :¼ ðX2þY2Þ1=2 ðX þ YÞ,
while the squared norm of the FB function for the SDCP is the function FB: Sn Sn! IRþgiven by
FBðX, YÞ :¼1
2kðX, YÞk2,
where Sn denotes the set of real n n symmetric matrices. The function FB has been proved to be strongly semismooth [17]. More recently, the squared norm of the matrix- valued FB function FB was reported in [16] to be smooth and its gradient is Lipschitz continuous.
The second-order cone (SOC), also called the Lorentz cone, in IRnis defined as Kn:¼ ðx1, x2Þ 2 R Rn1j kx2k x1
, ð3Þ
where k k denotes the Euclidean norm. By definition, K1 is the set of non-negative reals IRþ. The second-order cone complementarity problem (SOCCP) is to find x, y 2 IRn satisfying
x ¼ FðÞ, y ¼ GðÞ, hx, yi ¼ 0, x 2 Kn, y 2 Kn, ð4Þ where h, i is the Euclidean inner product and F, G : IRn!IRnare continuous (possibly non-linear) mappings. The merit function approach based on reformulating the NCP as an equivalent unconstrained minimization can be extended to the SOCCP case [6].
This approach aims to find a smooth function : IRnIRn!IRþsuch that
ðx, yÞ ¼ 0 () x 2 Kn, y 2 Kn, hx, yi ¼ 0: ð5Þ We call such a SOC merit function. Then the SOCCP can be expressed as an unconstrained smooth (global) minimization problem:
2IRminn f ðÞ:¼ ðFðÞ, GðÞÞ: ð6Þ Analogously, the squared norm of FB function can be considered in the SOCCP setting.
We define FB: IRnIRn!IRþassociated with the second-order cone Knas
FBðx, yÞ :¼ 1
2kFBðx, yÞk2, ð7Þ
where FB: IRnIRn!IRnis the FB function defined by
FBðx, yÞ :¼ ðx2þy2Þ1=2x y: ð8Þ More specifically, for any x ¼ (x1, x2), y ¼ ( y1, y2) 2 IR IRn1, we define their Jordan productassociated with Knas
x y:¼ ðhx, yi, y1x2þx1y2Þ: ð9Þ The Jordan product , unlike scalar or matrix multiplication, is not associative, which is a main source of complication in the analysis of SOCCP. The identity element under this
product is e :¼ (1, 0, . . . , 0)T2IRn. We write x2to mean x x and write x þ y to mean the usual componentwise addition of vectors. It is known that x22 Kn for all x 2 IRn. Moreover, if x 2 Kn, then there exists a unique vector in Kn, denoted by x1/2, such that (x1/2)2¼x1/2x1/2¼x. Thus, FBdefined as in (8) is well-defined for all (x, y) 2 IRnIRn and maps IRnIRnto IRn. It was shown in [10] that FBsatisfies the relation (5). Hence,
FBas defined in (7) is a merit function for the SOCCP. In the recent manuscript [5], this SOC merit function was shown to be an LC1function (a smooth function with its gradient being locally Lipschitz continuous).
Another popular SOC merit function is the natural residual merit function
NRðx, yÞ :¼ kNRðx, yÞk2, ð10Þ
which is induced by the natural residual function NR: IRnIRn!IRn
NRðx, yÞ :¼ x ½x yþ, ð11Þ
where [ ]þmeans the projection onto Kn. The natural residual function NRwas studied in [10,11] which is involved in smoothing methods for the SOCCP. A drawback of NRis its non-differentiability compared to FB. Some other classes of SOC merit functions for the SOCCP are also recently studied in [1,2].
In this article, we consider the following one-parametric class of SOC merit functions which was originally proposed in [12] for the NCP case:
ðx, yÞ :¼1
2kðx, yÞk2 ð12Þ
where : IRnIRn!IRnis a family of functions associated with the SOC, defined by
ðx, yÞ :¼ ðx yÞ 2þðx yÞ1=2
ðx þ yÞ, ð13Þ
and is a fixed parameter such that 2 (0, 4). It can be verified that for any x, y 2 IRn ðx yÞ2þðx yÞ ¼ x þ 2
2 y
2
þð4 Þ 4 y2
¼ y þ 2 2 x
2
þð4 Þ 4 x2
Kn 0, ð14Þ
where the inequality holds because 2 (0, 4). Therefore, in (13) is well-defined. Notice that reduces to the FB function FBwhen ¼ 2, whereas it becomes a multiple of the natural residual function NR when ! 0. Thus, this class of SOC complementarity functions covers the current two most important SOC complementarity functions so that a closer look and study for this new class of functions is worthwhile.
In fact, as mentioned in [5], when solving the equivalent unconstrained minimization via SOC merit functions, it is very important to show that the gradient of the employed SOC merit function is sufficiently smooth so as to warrant the convergence of appropriate computational methods. Here, we are particularly concerned with the conjugate gradient method. The method generally requires the Lipschitz continuity of the gradient ( f 2 LC1by our notation). The main purpose of this article is to show that the function as defined in (12) has a globally Lipschitz continuous gradient. Thus, this article can be regarded as
a follow-up of [5] since it extends the LC1property of FBto the class of SOC merit functions
. Nonetheless, the technique used here is a bit different and the analysis is more tedious and subtle. In particular, from the extension work, we see that the Lipschitz continuity of the gradient of becomes worse when ! 0. This fact will provide an instructional help for the design of algorithms with this class of SOC merit functions.
Throughout this article, IRndenotes the space of n-dimensional real column vectors and the supscript ‘T’ represents the transpose. For any differentiable function f : IRn!IR, rf(x) denotes the gradient of f at x. For any differentiable mapping F ¼(F1, . . . , Fm)T: IRn!IRm, rF(x) ¼ [rF1(x) rFm(x)] is a n m matrix denoting the transposed Jacobian of F at x. For non-negative scalars and , we write ¼ O() to mean C, with C independent of and .
2. Preliminaries
It is known that Knis a closed convex self-dual cone with non-empty interior given as Kn:¼ ðx1, x2Þ 2 R Rn1j kx2k x1
:
For any x ¼ (x1, x2) 2 IR IRn1, we define the determinant and the trace of x as follows:
detðxÞ :¼ x21 kx2k2, trðxÞ ¼ 2x1:
In general, det(x y) 6¼ det(x)det( y) unless x and y are collinear, i.e. x ¼ y for some 2 IR.
A vector x ¼ (x1, x2) 2 IR IRn1is said to be invertible if det(x) 6¼ 0. If x is invertible, then there exists a unique y ¼ ( y1, y2) 2 IR IRn1satisfying x y ¼ y x ¼ e. We call this y the inverse of x and denote it by x1. In fact, we have
x1¼ 1
x21 kx2k2ðx1, x2Þ ¼ 1 detðxÞ
trðxÞe x
:
It is not difficult to see that x 2 int(Kn) if and only if x12int(Kn). For any x ¼(x1, x2) 2 IR IRn1, we define the matrix Lxby
Lx:¼ x1 xT2 x2 x1I
:
It is easily verified that Lxþy¼LxþLyand x y ¼ Lxyfor any x, y 2 IRn, and Lxis positive definite (and hence invertible) if and only if x 2 int(Kn). However, L1x y 6¼ x1y, for some x 2int(Kn) and y 2 IRn, i.e. L1x 6¼Lx1.
We next recall from [10] that each x ¼ (x1, x2) 2 IR IRn1 admits a spectral factorization, associated with Kn, of the form
x ¼ 1ðxÞuð1Þx þ2ðxÞuð2Þx , ð15Þ where 1(x), 2(x) and uð1Þx , uð2Þx are the spectral values and the associated spectral vectors of x, with respect to Kn, given by
iðxÞ ¼ x1þ ð1Þikx2k, ð16Þ
uðiÞx ¼ 1 2
1, ð1Þi x2
kx2k
, if x26¼0, 1
2 1, ð1Þiw2
, if x2¼0, 8>
><
>>
:
ð17Þ
for i ¼ 1, 2, with w2 being any vector in IRn1 satisfying k w2k ¼1. The spectral factorization of x, x2and x1/2as well as the matrix Lxhave various interesting properties (cf. [10]). We list some properties that we will use later.
Property 2.1 For any x ¼(x1, x2) 2 IR IRn1with spectral values 1(x), 2(x) and spectral vectors uð1Þx , uð2Þx , the following results hold.
(a) x2¼21ðxÞuð1Þx þ22ðxÞuð2Þx 2 Kn.
(b) If x 2 Kn, then 0 1(x) 2(x) and x1=2 ¼ ffiffiffiffiffiffiffiffiffiffiffi
1ðxÞ
p uð1Þx þ ffiffiffiffiffiffiffiffiffiffiffi
2ðxÞ p uð2Þx . (c) If x 2 int(Kn, then 0 5 1(x) 2(x), and Lxis invertible with
L1x ¼ 1 detðxÞ
x1 xT2
x2
detðxÞ x1
I þ 1 x1
x2xT2 2
4
3 5:
(d) The determinant, the trace and the Euclidean norm of x can be denoted by 1(x),
2(x):
detðxÞ ¼ 1ðxÞ2ðxÞ, trðxÞ ¼ 1ðxÞ þ 2ðxÞ, kxk2¼21ðxÞ þ 22ðxÞ
2 :
Before giving out several technical lemmas that will be applied in the next section, we introduce some notations that will be frequently used in the subsequent analysis.
Unless otherwise stated, in this article, we always write
w ¼ wðx, yÞ :¼ ðx yÞ2þðx yÞ and z ¼ zðx, yÞ :¼ ½ðx yÞ2þðx yÞ1=2: ð18Þ Since (x y)2þ(x y) ¼ x2þy2þ( 2) (x y) 2 Knfor any x, y 2 IRn, we have
w:¼ w1
w2
¼ kxk2þ kyk2þ ð 2ÞxTy 2ðx1x2þy1y2Þ þ ð 2Þðx1y2þy1x2Þ
!
2 Kn: ð19Þ
From this, it follows that the spectral values of w are given by
1ðwÞ:¼ kxk2þ kyk2þ ð 2ÞxTy k2ðx1x2þy1y2Þ þ ð 2Þðx1y2þy1x2Þk,
2ðwÞ:¼ kxk2þ kyk2þ ð 2ÞxTy þ k2ðx1x2þy1y2Þ þ ð 2Þðx1y2þy1x2Þk: ð20Þ By Property 2.1(b), the vector z has the spectral values ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ
p , ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p and
z:¼ ðz1, z2Þ ¼
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p
2 ,
ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ p
2 w2
!
, ð21Þ
where w2 :¼ ðw2=kw2kÞ if w26¼0 and otherwise w2 is any vector in IRn1 satisfying kw2k ¼1.
The following four technical lemmas are crucial in proving our main results.
Lemma 2.1 measures how close w comes to the boundary of Kn, and Lemma 2.2 describes the behaviour of (x, y) when w lies on the boundary of Kn. Lemma 2.3 talks about the differential rule for the Jordan product function. Lemma 2.4 gives the gradient of the function z(x, y).
LEMMA 2.1 [4, Lemma 3.4] For any x ¼(x1, x2), y ¼ ( y1, y2) 2 IR IRn1and 2(0, 4), if w2¼2(x1x2þy1y2) þ ( 2)(x1y2þy1x2) 6¼ 0, then we have
x1þ 2 2 y1
þ ð1Þi
x2þ 2
2 y2T w2
kw2k
2
x2þ 2 2 y2
þ ð1Þi
x1þ 2 2 y1
w2 kw2k
2
kxk2þ kyk2þ ð 2Þhx, yi þ ð1Þikw2k
iðwÞ
for i ¼1, 2, and furthermore these relations also hold when interchanging x and y.
LEMMA 2.2 [5, Lemma 3.2] For any x ¼(x1, x2), y ¼ ( y1, y2) 2 IR IRn1and 2(0, 4), if w ¼(x y)2þ(x y) =2int(Kn), then there always holds that
x21¼ kx2k2, y21¼ ky2k2, x1y1¼xT2y2, x1y2 ¼y1x2; x21þy21þ ð 2Þx1y1¼ kx1x2þy1y2þ ð 2Þx1y2k
¼ kx2k2þ ky2k2þ ð 2ÞxT2y2:
If, in addition, (x, y) 6¼ (0, 0), then w2¼2(x1x2þy1y2þ( 2)x1y2) 6¼ 0, and furthermore, xT2 w2
kw2k¼x1, x1
w2
kw2k¼x2, yT2 w2
kw2k¼y1, y1
w2
kw2k¼y2:
LEMMA 2.3 [6, Lemma 3.1] Let !: IRnIRn!IRmbe given by !(x, y) :¼ u(x, y) v(x, y), where u, v : IRnIRn!IRmare differentiable mappings. Then, ! is differentiable and
rx!ðx, yÞ ¼ rxuðx, yÞLvðx,yÞþ rxvðx, yÞLuðx,yÞ, ry!ðx, yÞ ¼ ryuðx, yÞLvðx,yÞþ ryvðx, yÞLuðx,yÞ: In particular, when !(x, y) ¼ x y, there holds
rx!ðx, yÞ ¼ Ly, ry!ðx, yÞ ¼ Lx; and when !(x, y) ¼ x2y2, there holds
rx!ðx, yÞ ¼ 2LxLy2, ry!ðx, yÞ ¼ 2LyLx2:
LEMMA 2.4 For any x, y 2 IRn and 2(0, 4), let z(x, y) be defined as in (18). Then the function z(x, y) is continuously differentiable at a point (x, y) satisfying (x y)2þ(x y) 2 int(Kn). Moreover, we have that
rxzðx, yÞ ¼ Lxþ 2 2 Ly
L1zðx,yÞ, ryzðx, yÞ ¼ Lyþ 2
2 Lx
L1zðx,yÞ:
Proof The differentiability of z(x, y) is an immediate consequence of [13], see also [3, Prop. 4]. Since z2(x, y) ¼ (x y)2þ(x y), applying Lemma 2.3 yields
2rxzðx, yÞLzðx,yÞ¼2LxyþLy ¼2Lxþ ð 2ÞLy:
Hence, rxzðx, yÞ ¼ ðLxþ ð 2=2ÞLyÞL1zðx,yÞ. In view of the symmetry of x and y in the z(x, y), there also holds that ryzðx, yÞ ¼ ðLyþ ð 2=2ÞLxÞL1zðx,yÞ. œ
3. Main results
In this section, we present the proof showing that the gradient function of is Lipschitz continuous. By the notation in Section 2, the function can be rewritten as
ðx, yÞ ¼1
2½ðx yÞ2þðx yÞ1=2 ðx þ yÞ2
¼1
2zðx, yÞ2ðx þ yÞT½ðx yÞ2þðx yÞ1=2þ1
2kx þ yk2
¼1 2
2ðwÞ þ 1ðwÞ
2 þ kx þ yk2
ðx þ yÞT½ðx yÞ2þðx yÞ1=2
¼1
22kxk2þ2kyk2þðx yÞ
ðx þ yÞT½ðx yÞ2þðx yÞ1=2
¼ kxk2þ kyk2þ
2ðx yÞ ðx þ yÞT½ðx yÞ2þðx yÞ1=2, ð22Þ where the third equality is due to Property 2.1(d). Clearly, the gradient of the function kxk2þ kyk2þ ð=2Þðx yÞ is globally Lipschitz continuous. Therefore, to show that the gradient of is globally Lipschitz continuous, we only need to show that following function
Fðx, yÞ :¼ ðx þ yÞT½ðx yÞ2þðx yÞ1=2 ð23Þ has a Lipschitz continuous gradient. The following lemma states the gradient of F(x, y).
LEMMA 3.1 For any x, y 2 IRnand 2(0, 4), let F : IRnIRn!IR be defined as in (23).
Then, the function F(x, y) is continuously differentiable everywhere. Moreover, rxF(0, 0) ¼ ryF(0, 0) ¼ 0. If (x, y) 6¼ (0, 0) and (x y)2þ(x y) 2 int(Kn), then
rxFðx, yÞ ¼ zðx, yÞ þ Lxþ 2 2 Ly
L1zðx,yÞðx þ yÞ ryFðx, yÞ ¼ zðx, yÞ þ Lyþ 2
2 Lx
L1zðx,yÞðx þ yÞ: ð24Þ
If(x, y) 6¼ (0, 0) and (x y)2þ(x y) =2int(Kn), then rxFðx, yÞ ¼ zðx, yÞ þ x1þ22 y1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21þy21þ ð 2Þx1y1
q ðx þ yÞ
ryFðx, yÞ ¼ zðx, yÞ þ y1þ22 x1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21þy21þ ð 2Þx1y1
q ðx þ yÞ: ð25Þ
Proof Case 1 x ¼ y ¼0. For any h, k 2 IRn, let 12be the spectral values and v(1)and v(2)be the corresponding vectors of (h k)2þ(h k). Then, by Property 2.1(b),
k½ðh kÞ2þðh kÞ1=2k ¼ k ffiffiffiffiffiffi
1
p vð1Þþ ffiffiffiffiffiffi
2
p vð2Þk ð ffiffiffiffiffiffi
1
p þ ffiffiffiffiffiffi
2
p Þ= ffiffiffi 2 p
ffiffiffiffiffiffiffiffi 22
p :
Therefore,
Fðh, kÞ Fð0, 0Þ ¼ ðh þ kÞT½ðh kÞ2þðh kÞ1=2
k½ðh kÞ2þðh kÞ1=2kkh þ kk
ffiffiffiffiffiffiffiffi 22
p ðkhk þ kkkÞ
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2ðkhk2þ kkk2þ ð 2ÞhTkÞ q
ðkhk þ kkkÞ
¼Oðkhk2þ kkk2Þ:
This shows that F(x, y) is differentiable at (0, 0) with rxF(0, 0) ¼ ryF(0, 0) ¼ 0.
Case 2 (x, y) 6¼ (0, 0) and (x y)2þ(x y) 2 int(Kn). Using Lemma 2.4, we readily obtain the formula in (24). Clearly, in this case, rxF(x, y) and ryF(x, y) are continuous, and consequently, F(x, y) is continuously differentiable at such points.
Case 3 (x, y) 6¼ (0, 0) and (x y)2þ(x y) =2int(Kn). Since 1(w) ¼ 0, now it follows from (18) and (21) that
zðx, yÞ ¼ ½ðx yÞ2þðx yÞ1=2¼1 2
ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p , ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p w2
T
, Moreover, by Lemma 2.2, we can compute that
w2¼2ðx1x2þy1y2þ ð 2Þx1y2Þ, 2ðwÞ ¼4ðx21þy21þ ð 2Þx1y1Þ ¼2kw2k:
Therefore, under this case, we have that
zðx, yÞ ¼ ½ðx yÞ2þðx yÞ1=2 ¼
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21þy21þ ð 2Þx1y1 q
x1x2þy1y2þ ð 2Þx1y2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21þy21þ ð 2Þx1y1
q 2 66 64
3 77 75:
By this, it is easy to verify that the formula (25) holds.
Note that z(x, y) is a continuous function. This together with the proof of Case (i) and Case (iii) of [4, Proposition 3.3] means that the gradient functions rxF(x, y) and ryF(x, y) are continuous at every (x, y) 2 IRnIRn. Hence, F(x, y) is continuously differentiable
everywhere. Thus, we complete the proof. œ
For the symmetry of x and y in rxF(x, y) and ryF(x, y), in the rest of this section, we concentrate on the proof of globally Lipschitz continuity of rxF(x, y). We first define a smooth approximation of rxF(x, y). For any 4 0, we let
^
w ¼ ^wðx, y, Þ :¼ ðx yÞ2þðx yÞ þ e,
^z ¼ ^zðx, y, Þ :¼ ½ðx yÞ2þðx yÞ þ e1=2, ð26Þ where e is the identity element under the Jordan product. It is not hard to see that
^
w1¼w1þ , w^2¼w2, 1ðwÞ ¼ ^ 1ðwÞ þ , 2ðwÞ ¼ ^ 2ðwÞ þ , ð27Þ and furthermore,
^z ¼ ð ^z1, ^z2Þ ¼1 2
ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p , ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
w2
ð28Þ
where w2 :¼ ðw2=kw2kÞ if w26¼0 and otherwise w2 is any vector in IRn1 satisfying kw2k ¼1. We define the mapping G(, , ) : IRnIRn!IRnby
Gðx, y, Þ ¼ ^zðx, y, Þ þ Lxþ 2 2 Ly
L1^zðx,y, Þðx þ yÞ: ð29Þ
By Lemmas 3.2 and 3.4 below, G(x, y, ) is actually a smooth approximation of rxF(x, y).
Based on the relation, in the sequel, we will prove the Lipschitz continuity of rxF(x, y) through arguing that G(x, y, ) is globally Lipschitz continuous.
LEMMA 3.2 For any x, y 2 IRnand 4 0, let G(x, y, ) be defined as in (29). Then,
!0limþGðx, y, Þ ¼ rxFðx, yÞ:
Proof If (x, y) ¼ (0, 0), then G(x, y, ) ¼ ( e)1/2for any 4 0. Therefore,
!0limþGð0, 0, Þ ¼ rxFð0, 0Þ ¼ 0:
If (x, y) 6¼ (0, 0) and (x y)2þ(x y) 2 int(Kn), then by (27), (28) and Property 2.1(c), it is easy to verify that lim !0þL1^zðx,y, Þ¼L1zðx,yÞ. This together with lim !0þL^zðx,y, Þ¼Lzðx,yÞ
implies that the conclusion holds.
Next, we consider (x, y) 6¼ (0, 0) and (x y)2þ(x y) =2int(Kn). For convenience, let
g:¼ x þ 2
2 y, h :¼ y þ 2
2 x, uðx, y, Þ :¼ L1^zðx,y, Þg, vðx, y, Þ :¼ L1^zðx,y, Þh: ð30Þ By Property 2.1(c), it is easy to compute that
u ¼ uðx, y, Þ ¼ 1 detð ^zÞ
^z1 ^z2
^z2
detð ^zÞ
^z1 I þ1
^z1^z2^z2 2
4
3 5 g½ 1g2
¼ 1
detð ^zÞ
g1^z1g2^z2
g1^z2þdetð ^zÞ
^z1 g2þg2^z2
^z1 ^z2
2 4
3 5
:¼ u1
u2
, ð31Þ
v ¼ vðx, y, Þ ¼ 1 detð ^zÞ
^z1 ^zT2
^z2 detð ^zÞ
^z1 I þ1
^z1 ^z2^zT2 2
4
3 5 h1
h2
¼ 1
detð ^zÞ
h1^z1hT2^z2
h1^z2þdetð ^zÞ
^z1 h2þhT2^z2
^z1 ^z2
2 4
3 5
:¼ v1
v2
: ð32Þ
Using (28) and Lemma 2.2, we have that
u1¼ 1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^ 2ðwÞ^ p
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p
2 g1þ
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p
2 gT2w2
" #
¼ 1
2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
g1þgT2w2
þ 1
2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
g1gT2w2
¼ g1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ þ
p , ðsince gT2w2 ¼g1Þ ð33Þ
u2¼ 1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^ 2ðwÞ^ p
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p
2 g1w2þ 2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^ p ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p g2
2 64
þ
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p 2
2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p gT2w2w2 3 75
¼ðg1þgT2w2Þw2 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p ðg1gT2w2Þw2 2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ 2g22gT2w2w2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p
¼ g2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ þ
p : ðsince g1w2¼g2, gT2w2¼g1Þ ð34Þ
Similarly, we can also obtain that v1¼ h1
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ þ
p , v2¼ h2
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ þ
p :
From the above expressions of u and v, it then follows that
!0limþL1^zðx,y, Þðx þ yÞ ¼2
lim
!0þL1^zðx,y, Þ
x þ 2
2 y þ y þ 2 2 x
¼2
lim
!0þ
uðx, y, Þ þ vðx, y, Þ
¼ 1
ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p
x1þy1, x2þy2
T
:
This together with Lemma 2.2 yields that
!0limþGðx, y, Þ ¼ lim
!0þ^zðx, y, Þ þ
Lxþ 2 2 Ly
!0limþL1^zðx,y, Þðx þ yÞ
¼zðx, yÞ þ
x1þ 2
2 y1 xT2 þ 2 2 yT2 x2þ 2
2 y2
x1þ 2 2 y1
I 2
64
3 75
x1þy1 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p x2þy2
ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p 0 BB B@
1 CC CA
¼zðx, yÞ þ 1 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p
2
x1þ 2 2 y1
ðx1þy1Þ 2
x2þ 2 2 y2
ðx1þy1Þ 0
B@
1 CA
¼zðx, yÞ þ 1 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ p
2
x1þ 2 2 y1
ðx1þy1Þ 2
x1þ 2 2 y1
ðx2þy2Þ 0
B@
1 CA
¼zðx, yÞ þ x1þ22 y1 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x21þy21þðx1y1Þ
q ðx þ yÞ:
Thus, we complete the proof. œ
In what follows, we argue that the gradient function of G(x, y, ) is uniformly bounded, and then, by applying the mean-value theorem for vector-valued functions, we conclude that G(x, y, ) is globally Lipschitz continuous. The following two lemmas are crucial in proving our main result.
LEMMA 3.3 For any x, y 2 IRnand 4 0, let ^z : Rn Rn! Rnbe given as in (26). Then, the function ^zðx, y, Þ is continuously differentiable everywhere. Moreover, there exists a constant C independent of x, y and , such that
krx^zðx, y, Þk ¼
Lxþ 2 2 Ly
L1^zðx,y, Þ
C, kry^zðx, y, Þk ¼
Lyþ 2 2 Lx
L1^zðx,y, Þ
C:
Proof The first part of the conclusion follows directly from [10, Proposition 5.2].
The second part is implied by the proof of [15, Proposition 3.1]. œ LEMMA 3.4 For any x, y 2 IRn and 4 0, let G : IRnIRn!IRn be defined as in (29).
Then, the function G(x, y, ) is continuously differentiable everywhere. Moreover, there exists a constant C such that krxG(x, y, )k C(1 þ 1) and kryG(x, y, )k C(1 þ 1).
Proof The first part of the conclusion is due to [10, Proposition 5.2]. For the second part, by Lemma 3.3, it suffices to prove that the gradient of the following function:
Hðx, y, Þ :¼ Lxþ 2 2 Ly
L1^zðx,y, Þðx þ yÞ is uniformly bounded. From the definition of H(x, y, ), we notice that
Hðx, y, Þ ¼2
Lxþ 2 2 Ly
L1^zðx,y, Þ x þ 2
2 y þ y þ 2 2 x
¼2
x ðu þ vÞ þ 2
2 y ðu þ vÞ
where u and v are defined as in (30). Therefore, applying Lemma 2.3 yields that rxHðx, y, Þ ¼2
Luþvþ
rxuðx, y, Þ þ rxvðx, y, Þ
Lxþ 2 2 Ly
,
ryHðx, y, Þ ¼2
2 2 Luþvþ
ryuðx, y, Þ þ ryvðx, y, Þ
Lxþ 2 2 Ly
:
ð35Þ
To show that krxH(x, y, )k is uniformly bounded, we shall verify that both kLuþvk and kðrxuðx, y, Þ þ rxvðx, y, ÞÞðLxþ ð 2=2ÞLyÞk are uniformly bounded.
(i) To prove that kLuþvkis uniformly bounded, it is sufficient to argue that ju1j, ku2k and jv1j, ku2k are both uniformly bounded. First, we argue that ju1jand jv1jare uniformly bounded. From (33), we have that
u1¼ 1 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
g1þgT2w2
þ 1
2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
g1gT2w2
,
v1¼ 1 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
h1þhT2w2
þ 1
2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
h1hT2w2 :
Note that g1¼x1þ ð 2=2Þy1, g2¼x2þ ð 2=2Þy2and h1¼y1þ ð 2=2Þx1, h2¼y2þ ð 2=2Þx2, and w2¼ ðw2=kw2kÞ. Therefore, applying Lemma 2.1 yields that
g1gT2w2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p , jg1þgT2w2j ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p ð36Þ
and
h1hT2w2
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p , jh1þhT2w2j ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ
p ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p : ð37Þ
Combing with the expressions of u1and v1 given as above, we get ju1j 1 and jv1j 1.
Second, we argue that ku2kand kv2k are also uniformly bounded. By (34),
u2¼
g1þgT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
g1gT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ 2g22gT2w2w2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p ,
v2¼
h1þhT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
h1hT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ 2h22hT2w2w2
ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p :
Using (36) and (37) and the fact that k w2k ¼1, we obtain that g1þgT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
g1gT2w2
w2 2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
1
2kw2k þ1
2kw2k ¼1, h1þhT2w2
w2
2 ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
h1hT2w2
w2
2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p
1
2kw2k þ1
2kw2k ¼1, and
2g22gT2w2wT2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^ p
4kg2k ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p 4kg2k
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi kxk2þ kyk2þ ð 2ÞxTy þ
p 4,
2h22hT2w2wT2 ffiffiffiffiffiffiffiffiffiffiffi
1ðwÞ^
p þ ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p
4kh2k ffiffiffiffiffiffiffiffiffiffiffi
2ðwÞ^
p 4kh2k
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi kxk2þ kyk2þ ð 2ÞxTy þ
p 4:
The above inequalities imply that ku2k and kv2k are uniformly bounded. This together with the uniform boundedness of ju1jand jv1j implies that
kLuþvk ¼ u1þv1 ðu2þv2ÞT u2þv2 ðu1þv1ÞI
is also uniformly bounded.
(ii) Now, it comes to show that kðrxuðx, y, Þ þ rxvðx, y, ÞÞðLxþ ð 2=2ÞLyÞk is uniformly bounded. From the definition of u(x, y, ) and v(x, y, ) as given in (30), we have
^zðx, y, Þ ðuðx, y, Þ þ vðx, y, ÞÞ ¼ðx þ yÞ
2 :
Applying Lemma 2.3 then gives that
rx^zðx, y, ÞLuþvþ ðrxuðx, y, Þ þ rxvðx, y, ÞÞL^zðx,y, Þ¼ ð=2ÞI:
This is equivalent to saying that
ðrxuðx, y, Þ þ rxvðx, yÞÞL^zðx,y, Þ ¼ ð=2ÞI rx^zðx, y, ÞLuþv
¼
2I Lxþ 2 2 Ly
L1^zðx,y, ÞLuþv, where the second equality is due to Lemma 2.4. Therefore,
rxuðx, y, Þ þ rxvðx, y, Þ
Lxþ 2 2 Ly
¼
2I Lxþ 2 2 Ly
L1^zðx,y, ÞLuþv
L1^zðx,y, Þ Lxþ 2 2 Ly
¼
2L1^zðx,y, Þ Lxþ 2 2 Ly
Lxþ 2 2 Ly
L1^zðx,y, ÞLuþvL1^zðx,y, Þ Lxþ 2 2 Ly
¼
2 Lxþ 2 2 Ly
L1^zðx,y, Þ
T
Lxþ 2 2 Ly
L1^zðx,y, Þ
Luþv Lxþ 2 2 Ly
L1^zðx,y, Þ
T
: Now we have that
rxuðx, y, Þ þ rxvðx, y, Þ
Lxþ 2 2 Ly
2 Lxþ 2 2 Ly
L1^zðx,y, Þ
T
þ Lxþ 2
2 Ly
L1^zðx,y, Þ
kLuþvk Lxþ 2 2 Ly
L1^zðx,y, Þ
T
:
From Lemma 3.3, k½ðLxþ ð 2=2ÞLyÞL1^zðx,y, ÞTk is uniformly bounded.
This together with the uniform boundedness of kLuþvk yields that kðrxuðx, y, Þ þ rxvðx, y, ÞÞðLxþ ð 2=2ÞLyÞkis uniformly bounded.
From (i), (ii) and (35), we conclude that krxH(x, y, )k is uniformly bounded with the bound related to 1. Using similar arguments, we can prove that kryH(x, y, )k is uniformly bounded with the bound related to 1. Combining with Lemma 3.3, we then show that there exists a constant C such that krxG(x, y, )k C(1 þ 1) and
kryG(x, y, )k C(1 þ 1). œ
THEOREM 3.1 For any x, y 2 IRnand 2(0, 4), let F(x, y) be given as in (23). Then, their gradient functions rxF(x, y) and ryF(x, y) are globally Lipschitz continuous, i.e. there exists a constant C such that
krxFðx, yÞ rxFða, bÞk Cð1 þ 1Þkðx, yÞ ða, bÞk,
kryFðx, yÞ ryFða, bÞk Cð1 þ 1Þkðx, yÞ ða, bÞk ð38Þ for all(x, y), (a, b) 2 IRnIRn.
Proof We first prove that the function G(x, y, ) defined by (29) is globally Lipschitz continuous for any 4 0. For any (x, y) 2 IRnIRnand (a, b) 2 IRnIRn, we have that
Gðx, y, Þ Gða, b, Þ ¼ Gðx, y, Þ Gða, y, Þ þ Gða, y, Þ Gða, b, Þ:
From Lemma 3.4, we know that G(x, y, ) is continuously differentiable everywhere.
Hence, from the mean-value theorem, it follows that Gðx, y, Þ Gða, y, Þ ¼
Z1 0
rxGða þ tðx aÞ, y, Þðx aÞdt, Gða, y, Þ Gða, b, Þ ¼
Z1 0
ryGða, b þ tð y bÞ, Þð y bÞdt Combining the last two equations and using Lemma 3.4, we then obtain that
kGðx, y, Þ Gða, b, Þk Z1
0
rxGða þ tðx aÞ, y, Þðx aÞdt
þ Z1
0
ryGða, b þ tð y bÞ, Þð y bÞdt
Z1
0
krxGða þ tðx aÞ, y, Þkkx akdt
þ Z1
0
krxGða þ tðx aÞ, y, Þkkðx aÞkdt
Cð1 þ 1Þkðx, yÞ ða, bÞk, ð39Þ where C is a constant independent of x, y and , . From Lemma 3.2, we know that
!0limþGðx, y, Þ ¼ rxFðx, yÞ
for any x, y 2 IRnIRn. This together with (39) immediately yields that krxFðx, yÞ rxFða, bÞk ¼ k lim
!0þGðx, y, Þ lim
!0þGða, b, Þk
¼ lim
!0þkGðx, y, Þ Gða, b, Þk
Cð1 þ 1Þkðx, yÞ ða, bÞk:
Thus, we prove that rxF(x, y) is globally Lipschitz continuous. Similarly, we may prove
that ryF(x, y) is also globally Lipchitz continuous. œ
From the above theorem, we immediately obtain the following corollary.
COROLLARY 3.1 Let with 2(0, 4) be defined as in (12). Then is an LC1function, i.e.
the gradient functions rx (x, y) and ry (x, y) are globally Liptchitz continuous with the Lipschitz constant being O(1 þ 1).
Note
1. Member 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.
References
[1] J.-S. Chen, A new merit function and its related properties for the second-order cone complementarity problem, Pacific J. Optim. 2 (2006), pp. 167–179.
[2] ———, Two classes of merit functions for the second-order cone complementarity problem, Math. Methods Oper. Res. 64 (2006), pp. 495–519.
[3] J.-S. Chen, X. Chen, and P. Tseng, Analysis of nonsmooth vector-valued functions associated with second-order cones, Math. Progr. 101 (2004), pp. 95–117.
[4] J.-S. Chen and S.-H. Pan, A one-parametric class of merit functions for the second-order cone complementarity problem, Comput. Optimi. Appl. (2008), doi: 10.1007/s10589-008-9182-9.
[5] J.-S. Chen, D. Sun, and J. Sun, The SC1 property of the squared norm of the SOC Fischer- Burmeister function, Oper. Res. Lett 36 (2008), pp. 385–392.
[6] J.-S. Chen and P. Tseng, An unconstrained smooth minimization reformulation of the second-order cone complementarity problem, Math. Progr. 104 (2005), pp. 293–327.
[7] U. Faraut and A. Kora´nyi, Analysis on Symmetric Cones. Oxford Mathematical Monographs, Oxford University Press, New York, 1994.
[8] A. Fischer, A special Newton-type optimization methods, Optimization 24 (1992), pp. 269–284.
[9] ———, An NCP-function and its use for the solution of complementarity problem, in Recent Advances in Nonsmooth Optimization, D.-Z. Du, L.-Q. Qi, and R. Womersley, eds., World Scientific, Singapore, 1995, pp. 88–105.
[10] M. Fukushima, Z.-Q. Luo, and P. Tseng, Smoothing functions for second-order cone complementarity problems, SIAM J. Optim. 12 (2002), pp. 436–460.
[11] S. Hayashi, N. Yamashita, and M. Fukushima, A combined smoothing and regularization method for monotone second-order cone complementarity problems, SIAM J. Optim. 15 (2005), pp. 593–615.
[12] H. Kleinmichel and C. Kanzow, A new class of semismooth Newton-type methods for nonlinear complementarity problems, Comput. Optim. Appl. 11 (1998), pp. 227–251.
[13] A. Kora´nyi, Monotone functions on formally real Jordan algebras, Math. Ann. 269 (1984), pp. 73–76.
[14] J. Ortega and W. Rheinboldt, Iterative Solution of Nonlinear Equations in Several Variables.
SIAM Classics Appl. Math., Society for Industrial and Applied Mathematics, Philadelphia, PA, 2000.
[15] S.-H. Pan and J.-S. Chen, A semismooth Newton method for SOCCPs based on one-parametric class of SOC complementarity functions. Comput. Optimi. Appl. (2008), doi: 10.1007/s10589- 008-9166-9.
[16] C.-K. Sim, J. Sun, and D. Ralph, A note on the Lipschitz continuity of the gradient of the squared norm of the matrix-valued Fischer-Burmeister function, Math. Progr. 107 (2006), pp. 547–553.
[17] D. Sun and J. Sun, Strong semismoothness of the Fischer-Burmeister SDC and SOC complementarity functions, Math. Progr. 103 (2005), pp. 575–581.
[18] P. Tseng, Merit function for semidefinite complementarity problems, Math. Progr. 83 (1998), pp. 159–185.