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.
1500 JOURNAL OF COMPUTERS, VOL. 8, NO. 6, JUNE 2013<br />
In view of the boundedness of { x k } and J<br />
k<br />
be<strong>in</strong>g a<br />
subset of the f<strong>in</strong>ite set I = {1, 2, , m}<br />
as well as<br />
Lemma 2.1, we know that there exists an <strong>in</strong>f<strong>in</strong>ite <strong>in</strong>dex<br />
'<br />
set K ⊆ K such that<br />
' '<br />
T<br />
k k k k<br />
k<br />
x → x, J ≡ J , ∀k∈K , det( A A ) ≥ε, ε ≥ ε.<br />
As a result,<br />
T<br />
lim( ) ( T<br />
A A =∇g x) ∇g ( x),<br />
'<br />
k∈K<br />
k k '<br />
'<br />
J J<br />
T<br />
det( ∇g <br />
'( x) ∇g '( x)) ≥ ε > 0.<br />
J<br />
J<br />
T −1<br />
Hence, we obta<strong>in</strong> T<br />
|| ( A ) || || <br />
k<br />
Ak → ∇g '( x) ∇ g '( x) ||,<br />
J J<br />
T −1<br />
this contradict || ( Ak<br />
Ak) || →∞, ( k∈ K).<br />
So the first<br />
conclusion 1) follows.<br />
k<br />
2) We firstly show that lim d0<br />
= 0 .<br />
k →∞<br />
k<br />
We suppose by contradiction that lim d0<br />
≠ 0, then<br />
k →∞<br />
there exist an <strong>in</strong>f<strong>in</strong>ite <strong>in</strong>dex set K and a constant σ > 0<br />
k<br />
such that || d0<br />
|| > σ holds for all k∈<br />
K.<br />
Tak<strong>in</strong>g notice<br />
of the boundedness of { x<br />
k } , by tak<strong>in</strong>g a subsequence if<br />
necessary, we may suppose that<br />
k <br />
'<br />
x → x, Jk<br />
≡ J , ∀k∈<br />
K.<br />
Us<strong>in</strong>g Taylor expansion, we analyze the first search<br />
<strong>in</strong>equality of Step 5, comb<strong>in</strong><strong>in</strong>g the proof of Theorem 3.2,<br />
k<br />
the fact that x → x * ,( k →∞ ) implies that it is true.<br />
k<br />
k<br />
The proof of limd<br />
= 0; limd = 0 are elementary<br />
k→∞<br />
k→∞<br />
from the result of 1) as well as formulas (9) and (10).<br />
3) The proof of 3) is elementary from the formulas (9),<br />
(10) and assumption H2.1.<br />
Lemma 4.2. Let H2.1 to H4.1 holds,<br />
k+<br />
1 k<br />
lim || x − x || = 0 . Thereby, the entire sequence { x<br />
k }<br />
k →∞<br />
* k<br />
converges to x i.e. x → x * , k →∞ .<br />
Proof.<br />
From the Lemma 4.1, it is easy to see that<br />
k+<br />
1 k k 2 k<br />
lim || x − x || = lim(|| tkd + tkd<br />
||)<br />
k→∞<br />
k→∞<br />
k<br />
k<br />
≤ lim(|| d || + || d<br />
||) = 0<br />
k →∞<br />
Moreover, together with Theorem 1.1.5 <strong>in</strong> [4], it shows<br />
k<br />
that x → x * , k →∞<br />
Lemma 4.3 It holds, for k large enough, that<br />
* k<br />
* k *<br />
1) J<br />
k<br />
≡ I( x ) I* , b → uI = ( u ,<br />
* j<br />
j∈I* ), v →( uj, j∈I*<br />
)<br />
k k T k<br />
2) I + ⊆ Lk = { j∈ Jk : g<br />
j( x ) +∇ g<br />
j( x ) d0<br />
= 0} ⊆ Jk.<br />
Proof.<br />
1) Prove J<br />
k<br />
≡ I*<br />
.<br />
On one hand, from Lemma 2.1, we know, for k large<br />
enough, that I *<br />
⊆ J k<br />
. On the other hand, if it doesn’t<br />
hold that J<br />
k<br />
⊆ I*<br />
, then there exist constants j 0<br />
and<br />
β > 0 , such that<br />
*<br />
g<br />
j<br />
( x ) ≤− β < 0, j<br />
0<br />
0<br />
∈ Jk.<br />
k<br />
So, accord<strong>in</strong>g to d 0<br />
→ 0 and the functions g ( x ),<br />
j<br />
( j∈<br />
I ) are cont<strong>in</strong>uously differentiable, for k large<br />
k<br />
enough, if v < 0 , we have<br />
j0<br />
k * T k k * T k<br />
+∇<br />
j0 0<br />
=−<br />
j<br />
+∇<br />
0 j0<br />
0<br />
p ( x ) g ( x ) d v g ( x ) d<br />
j0<br />
1 k<br />
≥− v<br />
j<br />
> 0.<br />
0<br />
2<br />
Otherwise,<br />
k * T k<br />
p ( x ) +∇g ( )<br />
j<br />
j<br />
x d<br />
0<br />
0<br />
0<br />
k * T k 1<br />
k<br />
= g<br />
j<br />
( x ) +∇g ( )<br />
0 j<br />
x d<br />
0 0<br />
≤− β < 0, ( vj<br />
≥0)<br />
0<br />
2<br />
which is contradictory with (8) and the fact j 0<br />
∈ J k<br />
. So,<br />
J<br />
k<br />
≡ I*<br />
(for k large enough).<br />
k<br />
* k *<br />
Prove that b → uI = ( u ,<br />
* j<br />
j∈I* ), v →( uj, j∈ I*<br />
).<br />
k *<br />
For the v →( uj<br />
, j∈I*<br />
) statement, we have the<br />
k<br />
follow<strong>in</strong>g results from the def<strong>in</strong>ition of v ,<br />
k * T −1 T *<br />
v →−−B∇ f( x ) =−( A A ) A ∇ f( x )<br />
* * * *<br />
In addition, s<strong>in</strong>ce x * is a KKT po<strong>in</strong>t of (1), it is<br />
evident that<br />
* *<br />
∇ f( x ) + Au .<br />
* I<br />
= 0, u<br />
* I<br />
= −B * *<br />
∇ f( x )<br />
T −1 T *<br />
i.e. uI<br />
=−( A<br />
* *<br />
A* ) A*<br />
∇ f( x ).<br />
k<br />
Otherwise, from (8), the fact that d0 → 0 implies that<br />
k k k k<br />
*<br />
∇ f ( x ) + Hkd0 + Akb = 0, b →−B*<br />
∇ f( x ) = uI<br />
.<br />
*<br />
The claim holds.<br />
2) For<br />
Furthermore, it has<br />
*<br />
lim( k k<br />
x , d ) ( x , 0)<br />
0<br />
k →∞<br />
k *<br />
uI+ uI+<br />
x→∞<br />
= , we have<br />
Lk<br />
*<br />
⊆ I( x ) .<br />
lim = > 0 , so the proof is<br />
f<strong>in</strong>ished.<br />
In order to obta<strong>in</strong> super-l<strong>in</strong>ear convergence, a crucial<br />
requirement is that a unit step size is used <strong>in</strong> a<br />
neighborhood of the solution. This can be achieved if the<br />
follow<strong>in</strong>g assumption is satisfied.<br />
H4.2 Let<br />
2 k k k k<br />
|| (<br />
xxLx ( , uJ ) H) || (|| ||)<br />
k k<br />
d o d<br />
∇ − = , where<br />
= +∑ .<br />
k<br />
k<br />
Lxu ( , ) f( x) u g( x)<br />
Jk<br />
Jk<br />
j<br />
j∈Jk<br />
Lemma 4.4 Suppose that Assumption H 2.1 to H 4.2<br />
are all satisfied. Then, the step size <strong>in</strong> Algorithm A<br />
always one, i.e. tk<br />
≡ 1, if k is large enough.<br />
Proof.<br />
It is only necessary to prove that<br />
f ( x k + d k + d k ) ≤ f( x k ) + α∇f( x k ) T d<br />
k , (19)<br />
k k k<br />
g ( x + d + d ) ≤0, j∈I.<br />
(20)<br />
j<br />
*<br />
For (12) if j ∈ I \ I we have g ( ) 0<br />
*<br />
j<br />
x < ,<br />
k k<br />
k<br />
*<br />
( x , d , d ) →( x , 0, 0)( k →∞ ), then, it is easy to<br />
obta<strong>in</strong> g ( k k k<br />
j<br />
x + d + d ) ≤0<br />
holds.<br />
If j ∈ I*<br />
we have<br />
© 2013 ACADEMY PUBLISHER