Professional Documents
Culture Documents
net/publication/44260181
CITATIONS READS
3 87
2 authors, including:
SEE PROFILE
All content following this page was uploaded by Milanka Gardašević - Filipović on 27 May 2014.
presented so far and they are often restricted to the convex gradient operator ⎜⎜ ∂ , ∂ ,..., ∂ ⎟⎟ . The Euclidean product
⎝ ∂x1 ∂x 2 ∂x n ⎠
unconstrained case. In general, the various approaches are
based on combinations of the following methods: is denoted by ⋅,⋅ and ⋅ is the associated norm; B ( x, ρ ) is
subgradient methods; bundle techniques and the the ball centred at x with radius ρ . For a given symmetric
Moreau-Yoshida regularization.
For a function f it is very important that its positive definite linear operator M we set ⋅,⋅ M
:= M ⋅,⋅ ;
2
Moreau-Yoshida regularization is a new function which has hence it is shortly denoted by x := x, x .The smallest
M M
the same set of minima as f and is differentiable with
and the largest eigenvalue of M we denote by λ and Λ
Lipschitz continuous gradient, even when f is not respectively.
The domain of a given function f : R → R ∪ {+ ∞} is
differentiable. In [10], [11] and [19] the second order n
properties of the Moreau-Yoshida regularization of a given
function f are considered. { }
the set dom( f ) = x ∈ R f ( x ) < +∞ . We say that f is
n
Having in mind that the Moreau-Yoshida regularization of proper if its domain is nonempty.
The point x = arg min f ( x ) refers to the minimum
1 *
a proper closed convex function is an LC function, we
present an optimization algorithm (using the second order x∈R n
z∈R
{ }
f ( x ) = maxn f ( z ) + g T ( x − z ) holds, where g ∈ ∂f ( z ) Proof. See [13].
Lemma 4. If a proper convex function
(see [4]). f : R → R ∪ {+ ∞} is a differentiable function at a point
n
[10]): without any further assumptions, F has a Lipschitzian the point x ∈ R in the direction d ∈ R , i.e.:
n n
n
gradient on the whole space R . More precisely, for all
g1 − g 2
x1 , x2 ∈ R n the next formula: FD′′ ( x, d ) = limsup d,
α ↓0 α
∇F ( x1 ) − ∇F (x2 ) ≤ Λ ∇F (x1 ) − ∇F ( x2 ), x1 − x2 (13)
2
g1 ∈ ∂f ( p( x + αd )), g 2 ∈ ∂f ( p ( x ))
holds (see in [10]), where ∇F ( x) has the following form: where F ( x ) is defined by (11).
Lemma 6: Let f : R → R be a closed convex proper
n
G := ∇F ( x ) = M ( x − p ( x)) ∈ ∂f ( p ( x )) (14)
function and F is its Moreau –Yoshida regularization in the
and p ( x ) is the unique minimum in (11). So, according to sense of definition 5. Then the next statements are valid.
(i) FD′′ ( x k , kd ) = k FD′′ ( xk , d )
2
above consideration and Definition 3, we conclude that F is
(ii) FD′′( xk , d1 + d 2 ) ≤ 2(FD′′ ( xk , d1 ) + FD′′( xk , d 2 ))
1
an LC function (see in [11]).
Note that the function f has nonempty subdifferential at
(iii) FD′′( xk , d ) ≤ K ⋅ d , where K is some constant.
2
p( x ) = x − M g where g ∈ ∂f ( p ( x )) .
−1
(15) function and let F be its Moreau –Yoshida regularization.
Then the next statements are valid.
In [10] it is also proved that for all x1 , x2 ∈ R the next (i) FD′′ ( x, d ) is upper semicontinous with respect to
n
formula:
(x, d ) i.e. if xi → x and d i → d , then
p( x1 ) − p( x2 ) M ≤ M ( x1 − x2 ), p( x1 ) − p( x2 ) (16) lim sup FD′′(xi , di ) ≤ FD′′(x, d )
2
i →∞
is valid, namely the mapping x → p ( x ) , where p ( x ) is {
(ii) FD′′ ( x, d ) = max d Vd V ∈ ∂ F ( x )
T 2
}
Λ Proof. See in [18] and [2].
defined by (12), is Lipschitzian with constant (see
λ
Proposition 2.3. in [10]).
Lemma 5: The following statements are equivalent: V. SEMI-SMOOTH FUNCTIONS AND OPTIMALITY CONDITIONS
(i) x minimizes f ; (iv) x minimizes F ;
Definition 7: A function ∇F : R → R
n n
is said to be
(ii) p ( x ) = x (v) f ( p( x )) = f ( x ) ; semi-smooth at the point x ∈ R
n
if ∇F is locally
(iii) ∇F ( x ) = 0 (vi) F ( x ) = f ( x ) Lipschitzian x∈R at and the limit
n
IV. DINI SECOND UPPER DIRECTIONAL DERIVATIVE Note that for a closed convex proper function, the gradient
of its Moreau-Yoshida regularization is a semi-smooth
We shall give some preliminaries that will be used in the
function.
remainder of the paper.
Lemma 8. [18]: If ∇F : R → R is semi-smooth at the
n n
Definition 6. [18] The second order Dini upper directional
point x ∈ R then ∇F is directionally differentiable at
n
derivative of the function f ∈ LC at the point x ∈ R in
1 n
′
[∇f (x + αd ) − ∇f (x )]T ⋅ d . Vh − (∇F ) ( x, h ) = o( h ) . Similarly we have that
f D′′ ( x, d ) = lim sup
α ↓0 α h T Vh − F ′′( x, h ) = o h ( ). 2
to Lemma 5, theorem 23.1 in [15] and the fact that the next
1 ( ) 1
F xk + q i (k )d k − F (xk ) ≤ − q i (k )σ (FD′′( xk , d k )) (18)
inequalities F ′( x + td , d ) ≥ (F ( x + td ) − F ( x )) ≥ 0 2
t
hold we have where σ : [0,+∞) →[0,+∞) is a continuous function
F ′( x + td , d ) − F ′( x, d ) satisfying δ1t ≤ σ (t ) ≤ δ 2t and 0 < δ1 < δ 2 < 1 .
FD′′ ( x, d ) = limsup ≥0■
t ↓0 t We suppose that
Lemma 10. Let f : R → R be a proper closed convex
n
≤ FD′′( xk , d ) ≤ c2 d
2 2
c1 d (19)
function, F its Moreau-Yoshida regularization, and x a
hold for some c1 and c2 such that 0 < c1 < c2 .
point from R . If F ′( x, d ) = 0 and FD′′ ( x, d ) > 0 for all
n
is not a strict minimum of the function F, nor a proximal point Since ∇F is locally Lipschitzian we have that
of the function F. Then there exists a sequence
∇F ( xk ) d ≤ Λ d
T
holds (see in [17]), therefore
{xk }, xk → x such that F (xk ) ≤ F (x ) holds for every k .
1
If we define the sequence {x k }, xk → x by xk = x + t k d , ∇F ( xk ) d + ⋅ 0 ≤ Λ d holds and hence we have that:
T
2
xk − x
where t k = then by Lemma 8 and Lemma 6 it 1 1
∇F (xk ) d + FD′′(x, d ) − K d ≤ Λ d . Hence, we have:
T 2
d
2 2
follows that
1 1 1
( ) − Λ d + K d ≤ ∇F(xk ) d + FD′′(x, d )≤ Λ d + K d
2 T 2
1
F ( xk ) − F (x ) − t k ∇F ( x ) d = t k2 FD′′ ( x, d ) + o d
T 2
2 2 2
2
1 Φ (d ) 1
holds. Since ∇F ( x ) = 0 (from assumption of Lemma 10) it and − Λ + K d ≤ k ≤Λ+ K d .
2 d 2
1 2
follows that t k FD′′ (x, d ) ≤ 0 , which contradicts the This establish coercivity of the function Φ k .■
2
assumption.■
Remark. Coercivity of the function Φ k assures that the
optimal solution of problem (17) exists (see in [18]). It also
VI. A MODEL ALGORITHM means that, under the assumption (19) the direction sequence
In this section an algorithm for solving the problem (1) is {d k } is bounded sequence on R n (proof is analogous to the
introduced. We suppose that at each x ∈ R it is possible to
n
proof in [18]).
compute f ( x), F ( x ), ∇F ( x ) and FD′′ ( x, d ) for a given Proposition 3. If the Moreau-Yoshida regularization F (⋅) of
d ∈R . n
the proper closed convex function f (⋅) satisfies the
At the k-th iteration we consider the following problem
condition (19), then:
1 (i) the function F (⋅) is uniformly and, hence, strictly
minn Φ k (d ), Φ k (d ) = ∇F ( xk ) d + FD′′ ( xk , d ) (17)
T
d ∈R 2 convex;
where FD′′ ( xk , d ) stands for the second order Dini upper {
(ii) the level set L( x0 ) = x ∈ R : F ( x ) ≤ F ( x0 ) is a
n
}
compact convex set, and
directional derivative at xk in the direction d . Note that if
(iii) there exists a unique point x* such that
Λ is a Lipschitzian constant for F it is also a Lipschitzian
constant for ∇F . The function Φ k (d ) is called an
( )
F x* = min F ( x ) .
x∈ L ( x 0 )
iteration function. It is easy to see that Φ k (0 ) = 0 and Proof. (i) From the assumption (19) and the mean value
theorem it follows that for all x ∈ L( x0 ) ( x ≠ x0 ) there
Φ k (d ) is Lipschitzian on R n . We generate the sequence
exists θ ∈ (0,1) such that:
{xk } of the form xk +1 = xk + α k d k , where the direction
1
F ( x) − F (x0 ) == ∇F ( x0 ) ( x − x0 ) + FD′′(x0 + θ (x − x0 ), x − x0 )
T
vector d k is a solution of the problem (17), and the step-size 2
αk is a number satisfying α k = q i (k ) , 0 < q < 1 , where 1
≥ ∇F (x0 ) (x − x0 ) + c1 x − x0 > ∇F ( x0 ) ( x − x0 )
T 2 T
2
i * (k ) = log q
[Kt ] we have that the claim of the theorem
1 2
= λ∇F (xk ) d + λ FD′′ ( xk , d ) q
T
2
where the last equality holds by Lemma 6. Hence dividing
holds for all q
i (k )
(
∈ 0, q i
*
(k )
).■
both sides by λ and letting λ↓0 we have that Convergence theorem. Suppose that f is a proper closed
∇F ( xk ) d ≥ 0 holds for any d ∈ R , and consequently
T n
convex function and F is its Moreau-Yoshida regularization
satisfies (19). Then for any initial point x0 ∈ R , xk → x∞ ,
n
it follows that xk is a stationary point of the function F , i.e.
∇F ( xk ) = 0 . Hence, by Lemma 5 it follows that xk is a as k → +∞ , where x∞ is a unique minimal point of the
(iii) ⇒ (i): Let xk be a point such that 0 ∈ ∂f ( x k ) . Then Proof. . If d k ≠ 0 is a solution of (17), it follows that
by (14), Lemma 5 and Lemma 9 it follows that Φ k (d k ) ≤ 0 = Φ k (0) . Consequently, we have by the
condition (19) that
∇F ( x k ) d ≥ 0
T
(20)
1 1
∇F ( xk ) d k ≤ − FD′′(xk , d k ) ≤ − c1 d k
T 2
<0
for any d ∈ R . Suppose that d ≠ 0 is the optimal solution
n 2 2
of the problem (17). From the property of the iterative From the above inequality it follows that the vector d k is a
function Φ it follows that: descent direction at xk , i.e. from the relations (18) and (19)
1
∇F ( xk ) d k ≤ − FD′′ ( xk , d k ) we get
T
(21)
2 ( ) 1
F( xk +1 ) − F(xk ) = F xk + qi(k )dk − F (xk ) ≤ − qi(k )σ (FD′′(xk , dk ))
2
Hence, (21) implies:
1 i(k ) 1 i(k )
∇F ( x k ) d k < 0 . ≤ − q δ1FD′′(xk , dk ) ≤ − q c1 dk
T 2
(22)
2 2
The above two inequalities (20) and (22) are for every d k ≠ 0 . Hence the sequence {F ( xk )} has the
contradictory.■
descent property, and, consequently, the sequence
Now we shall prove that there exists a finite i (k ) , i.e.
{xk } ⊂ L(x0 ) . Since L( x0 ) is by the Proposition 3 a
since d k is defined by (17), that the algorithm is
compact convex set, it follows that the sequence {xk } is
bounded. Therefore there exist accumulation points of the But, from the property of the iterative function in (17), we
sequence {xk } . have ∇ F ( x∞ ) d ∞ ≤ −
1
FD′′ ( x∞ , d ∞ ) . Therefore, we get a
T
2 2 2