Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Download Full Issue in PDF - Academy Publisher
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013 1499<br />
∇ f ( x ) d =∇f( x ) d −δ<br />
∇f( x ) A ( A A ) e<br />
k T k k T k k T T −1<br />
0 k k k k k<br />
1 k T k kT k 1 k T k<br />
≤− ( d0 ) Hkd0 + b p ≤− ( d0)<br />
Hkd0 < 0.<br />
2 2<br />
Thereby, we know that d<br />
k is a feasible descent<br />
k<br />
direction of (1) at x .<br />
III. GLOBAL CONVERGENCE OF ALGORITHM<br />
In this section, firstly, it is shown that Algorithm A<br />
given <strong>in</strong> section 2 is well-def<strong>in</strong>ed, that is to say, for every<br />
k, that the l<strong>in</strong>e search at Step 5 is always successful<br />
Lemma 3.1 The l<strong>in</strong>e search <strong>in</strong> step 5 yields a stepsize<br />
1 i<br />
t<br />
k<br />
= ( ) for some f<strong>in</strong>ite i = i( k)<br />
.<br />
2<br />
Proof.<br />
It is a well-known result accord<strong>in</strong>g to Lemma 2.2. For<br />
(11),<br />
k k 2 k k k T k<br />
s f( x + td + t d<br />
) − f( x ) −αt∇f( x ) d<br />
k T k 2 k k T k<br />
=∇ f ( x ) ( td + t d<br />
) + o( t) −αt∇f( x ) d<br />
k T k<br />
= (1 −α) t∇ f( x ) d + o( t).<br />
For (12), if<br />
k k<br />
j∉ I( x ), gj<br />
( x ) < 0;<br />
k k k T k<br />
j∈ I( x ), gj( x ) = 0, ∇ gj( x ) d < 0,<br />
so we have<br />
k k 2 k k T k 2 k<br />
g<br />
j<br />
( x + td + t d<br />
) =∇ f( x ) ( td + t d<br />
) + o( t)<br />
k T k<br />
= αt∇ g<br />
j<br />
( x ) d + O( t).<br />
In the sequel, the global convergence of Algorithm A<br />
is shown. For this reason, we make the follow<strong>in</strong>g<br />
additional assumption.<br />
H3.1 { x k } is bounded, which is the sequence<br />
generated by the algorithm, and there exist constants<br />
2 T<br />
2<br />
b≥ a > 0 , such that a|| y|| ≤ y Hk<br />
y ≤ b|| y||<br />
, for all k<br />
n<br />
and all y∈ R .<br />
S<strong>in</strong>ce there are only f<strong>in</strong>itely many choices for sets<br />
k k k k k<br />
Jk<br />
⊆ I , and the sequence { d0, d1<br />
, d , v , b } is bounded,<br />
we can assume without loss of generality that there exists<br />
a subsequence K, such that<br />
x → x , H → H , d →d , d →d , d<br />
→d<br />
,<br />
b b v v J J k K<br />
k * k * k * k *<br />
k * 0 0<br />
k * k *<br />
→ , → ,<br />
k<br />
≡ ≠ Φ, ∈ ,<br />
where J is a constant set.<br />
(15)<br />
Theorem 3.2 The algorithm either stops at the KKT<br />
k<br />
po<strong>in</strong>t x of the problem (1) <strong>in</strong> f<strong>in</strong>ite number of steps, or<br />
generates an <strong>in</strong>f<strong>in</strong>ite sequence { x k } any accumulation<br />
*<br />
po<strong>in</strong>t x of which is a KKT po<strong>in</strong>t of the problem (1).<br />
Proof.<br />
The first statement is easy to show, s<strong>in</strong>ce the only<br />
stopp<strong>in</strong>g po<strong>in</strong>t is <strong>in</strong> step 3. Thus, assume that the<br />
algorithm generates an <strong>in</strong>f<strong>in</strong>ite sequence { x k }, and (15)<br />
holds. Accord<strong>in</strong>g to Lemma 2.2, it is only necessary to<br />
*<br />
prove that d =<br />
0<br />
0<br />
. Suppose by contradiction that<br />
*<br />
d0 ≠ 0 .<br />
Then, from Lemma 2.2, it is obvious that d * is welldef<strong>in</strong>ed,<br />
and it holds that<br />
* T * * T * *<br />
∇ f ( x ) d < 0, ∇ g<br />
j<br />
( x ) d < 0, j∈I( x ) ⊆ J (16)<br />
Thus, from (16), it is easy to see that the step-size t k<br />
obta<strong>in</strong>ed <strong>in</strong> step 5 are bounded away from zero on<br />
K , i.<br />
e.<br />
t ≥ t* = <strong>in</strong>f{ t , k∈ K} > 0, k∈ K.<br />
(17)<br />
k<br />
k<br />
In addition, from (11) and Lemma 2.2, it is obvious<br />
k<br />
that { f ( x )} is monotonous decreas<strong>in</strong>g. So, accord<strong>in</strong>g to<br />
*<br />
assumption H 2.1, the fact that { x<br />
k } → x implies that<br />
k<br />
*<br />
f( x ) → f( x ), k →∞ .<br />
(18)<br />
So, from (11), (16), (17), it holds that<br />
k *<br />
k T k<br />
0= lim( f ( x ) − f( x )) ≤lim( αt ∇f( x ) d )<br />
x∈K<br />
x∈K<br />
1<br />
* T *<br />
≤ αt*<br />
∇ f( x ) d < 0,<br />
2<br />
k<br />
which is a contradiction thus lim d0<br />
= 0 . Thus, x * is a<br />
x→∞<br />
KKT po<strong>in</strong>t of (1).<br />
IV. THE RATE OF CONVERGENCE<br />
Now we discuss the convergent rate of the algorithm,<br />
and prove that the sequence { x<br />
k } generated by the<br />
algorithm is one-step super-l<strong>in</strong>early convergent under<br />
some mild conditions without the strict complementarily.<br />
For this purpose, we add some regularity hypothesis.<br />
H 4.1 The sequence { x<br />
k } generated by Algorithm A is<br />
bounded, and possess an accumulation po<strong>in</strong>t x * , such<br />
* *<br />
that the KKT pair ( x , u ) satisfies the strong secondorder<br />
sufficiency conditions, i.e.,<br />
T 2 * *<br />
d ∇<br />
xxL( x , u ) d > 0,<br />
n<br />
* T<br />
∀d∈Ω<br />
{ d∈R : d ≠ 0, ∇ gI<br />
+<br />
( x ) d = 0},<br />
+<br />
*<br />
Lxu ( , ) = f( x) + ug( x), I = { j∈ I: u > 0}.<br />
∑<br />
j∈I<br />
j j j<br />
Lemma 4.1 Suppose that assumptions H 2.1-H 3.1 hold,<br />
then,<br />
1) There exists a constant ζ > 0 , such that<br />
T −1<br />
|| ( Ak<br />
Ak) || ≤ ζ ;<br />
2) lim d k 0<br />
= 0; lim d k = 0; lim d<br />
k = 0;<br />
k→∞ k→∞ k→∞<br />
3)<br />
k k k k 2<br />
|| d || ∼|| d0<br />
||, || d<br />
|| = O(|| d || ),<br />
.<br />
k k k 3 k k 2<br />
|| d − d0||= O(|| d0<br />
|| ), || d<br />
|| = O(|| d || ).<br />
Proof.<br />
1) By contradiction, suppose that sequence<br />
T 1<br />
{|| ( Ak<br />
Ak) − ||} is unbounded, then there exists an <strong>in</strong>f<strong>in</strong>ite<br />
subset K, such that<br />
T −1<br />
|| ( A A ) || →∞, ( k∈<br />
K).<br />
k<br />
k<br />
K<br />
k<br />
© 2013 ACADEMY PUBLISHER