§3 Convergence Results With Strong Convexity
In the first part of this section, we prove the well-definedness of the sequence {x
k} generated by Algorithm 1 under some weak monotonicity assumptions on the operator T . Then we give the convergence results obtained under an additional continuous assumption.
The following existence result is due to Shih and Tan [23].
Proposition 3.1 [23, Theorem 2]. Let C be a nonempty closed convex subset of X and let F be a (l, w)-u.s.c. and monotone set-valued operator from C into X
∗such that F (x) is weakly compact and convex for each x ∈ C. If C is either compact or
kzk→∞
lim
z∈C
inf
y∈F (z)
hy, z − x
0i
kzk = ∞,
for some x
0∈ C, then there exists a solution to the problem V I(F, C).
Using this, we can establish an existence result for strongly monotone operators.
Proposition 3.2. Let C be a nonempty closed convex subset of X and let F be a (l, w)-u.s.c. and strongly monotone set-valued operator from C into X
∗such that F (x) is weakly compact and convex for each x ∈ C, then there exists a unique solution to the problem V I(F, C).
Proof : Let z, x ∈ C and y
1∈ F (z), y
2∈ F (x). Since F is strongly monotone, there exists a > 0 such that
hy
1, z − xi ≥ hy
2, z − xi + akz − xk
2≥ −ky
2kkz − xk + akz − xk
2= (akz − xk − ky
2k)kz − xk.
Thus,
inf
y1∈F (z)
hy
1, z − xi
kzk ≥ (akz − xk − ky
2k) kz − xk kzk
→ ∞, as kzk −→ ∞.
By Proposition 3.1, V I(F, C) has a solution, and it is trivial that the solution is unique,
since F is strongly monotone. 2
In [25], Yao obtained the existence theorem for pseudomonotone operator as follows.
Proposition 3.3 [25, Corollary 2.1]. Let C be a nonempty closed convex subset of X and let F be a u.s.c. pseudomonotone set-valued operator from C into X
∗such that F (x) is weakly compact and convex for each x ∈ C. If C is either compact or
lim inf
kzk→∞
z∈C
inf
y∈F (z)
hy, z − x
0i > 0,
for some x
0∈ C, then there exists a solution to the problem V I(F, C).
Now we make some basic assumptions on our Algorithm 1.
Assumption 1.
(i) T is (l, w)-u.s.c. and weakly montone with constant L on C.
(ii) M is differentiable and strongly convex with constant α =
2b, and M
0is (l, w)- continuous.
(iii) The sequence {ε
k} satisfies the following condition : ε
k> 0 ,
∞
X
k=0
ε
k= +∞ ,
∞
X
k=0
ε
2k< +∞.
The following result analyzes the convergence of the Algorithm 1 involving pseudo- Dunn property.
Theorem 3.4. Suppose that the problem (P 1) has a solution x
∗. Under Assumption 1, for each step k, there exists a unique solution x
k+1generated by Algorithm 1 to the auxiliary problem (P 2). Moreover, if T has the pseudo-Dunn property with constant t on C, and if
∀ k ∈ N, ε
k+1≤ ε
kand α < ε
k<
t+β2b, where α > 0, β > 0,
then the sequence {x
k} is bounded. In addition, if M
0is Lipschitz continuous with con- stant m, and if T is (w, s)-u.s.c. on C, then every weakly cluster point of the sequence {x
k} is a solution of the problem (P 1).
Proof : Under Assumption 1, the operator F defined by F (y) = εT (y) + M
0(y) − M
0(x)
is (l, w)-u.s.c. and strongly monotone with constant b − εL on C , by Lemma 2.3. Ac- cording to Proposition 3.2, we get that there exists a unique solution to the auxiliary problem (P 2).
Consider the function Λ
x∗defined by
Λ
x∗(x, ε) := Φ
x∗(x) + Ω
x∗(x, ε), ∀ x ∈ C, where
Φ
x∗(x) := M (x
∗) − M (x) − hM
0(x), x
∗− xi, Ω
x∗(x, ε) := εhy
∗, x − x
∗i.
Since M
0is strongly monotone with constant b, by Proposition 2.2, M is strongly convex with constant α =
b2, and hence
Φ
x∗(x
k) ≥ b
2 kx
k− x
∗k
2≥ 0.
Since x
∗is a solution of the problem (P 1), Ω
x∗(x, ε) is nonnegative, hence Λ
x∗(x, ε) ≥ 0.
Now we study the variation of Λ
x∗for Algorithm 1 :
Γ
k= Λ
x∗(x
k+1, ε
k) − Λ
x∗(x
k, ε
k) = s
1+ s
2+ s
3, where
s
1:= M (x
k) − M (x
k+1) − hM
0(x
k), x
k− x
k+1i, s
2:= hM
0(x
k) − M
0(x
k+1), x
∗− x
k+1i,
s
3:= ε
k+1hy
∗, x
k+1− x
∗i − ε
khy
∗, x
k− x
∗i.
Since M
0is strongly monotone,
s
1≤ − b
2 kx
k+1− x
kk
2.
Using (3) with z = x
∗, ε
khy
k, x
∗− x
k+1i + B(x
k+1, x
k, x
∗) ≥ 0, and therefore s
2≤ ε
khy
k, x
∗− x
k+1i = ε
khy
k, x
∗− x
ki + ε
khy
k, x
k− x
k+1i.
Using (1) with z = x
k, we have some y
∗∈ T (x
∗) such that hy
∗, x
k− x
∗i ≥ 0.
Since T has the pseudo-Dunn property with constant t and for some y
k∈ T (x
k), we have hy
k, x
k− x
∗i ≥ 1
t ky
k− y
∗k
2. Thus,
s
2≤ − ε
kt ky
k− y
∗k
2+ ε
khy
k, x
k− x
k+1i.
Since ε
k+1≤ ε
kfor each k, we obtain s
2+ s
3≤ − ε
kt ky
k− y
∗k
2+ ε
khy
k− y
∗, x
k− x
k+1i.
Therefore,
Γ
k≤ − b
2 kx
k+1− x
kk
2− ε
kt ky
k− y
∗k
2+ε
khy
k− y
∗, x
k− x
k+1i.
Since
ε
kky
k− y
∗k · kx
k+1− x
kk ≤ ε
2k2λ ky
k− y
∗k
2+ λ
2 kx
k+1− x
kk
2,
where λ > 0, we get Γ
k≤ −( b
2 − λ
2 )kx
k+1− x
kk
2− ε
k( 1 t − ε
k2λ )ky
k− y
∗k
2. Thus, this condition α < ε
k<
t+β2b, where α > 0, β > 0 yields
Γ
k≤ −( b 2 − λ
2 )kx
k+1− x
kk
2− αβ
t(t + β) ky
k− y
∗k
2,
and for λ < b, Γ
kis negative unless x
k+1= x
kand y
k= y
∗. Then, according to (3), {x
k} is a solution to the problem (P 1). Note that the sequence {Λ
x∗(x
k, ε
k)} is strictly decreasing, and since it is positive, it must be convergent and the difference between two consecutive terms tends to zero. Therefore, kx
k+1− x
kk and ky
k− y
∗k are convergent to zero. Moreover, since the sequence {Λ
x∗(x
k, ε
k)} converges, it is bounded, and so is {x
k}.
Let ¯ x be a weakly cluster point of the sequence {x
k}, and let {x
ki} be a subsequence weakly converging to ¯ x. By (3), since M
0is Lipschitz continuous with constant m and ε
k> α, we have some y
k∈ T (x
k),
hy
k, z − x
k+1i ≥ − m
α kx
k+1− x
kk · kz − x
k+1k, ∀z ∈ C.
Since kx
ki+1− x
kik converges to zero and y
kistrongly converges to y
∗, taking the limit for the subsequence {k
i} in the last inequality yields
hy
∗, z − ¯ xi ≥ 0, ∀z ∈ C. (4)
Moreover, h¯ y, x
ki− ¯ xi converges to 0, for each ¯ y ∈ T (¯ x). If 0 ∈ T (¯ x), then ¯ x is clearly a solution of the problem (P 1). Otherwise, we let ¯ y ∈ T (¯ x) with ¯ y 6= 0, and let
w
ki= x
ki− h¯ y, x
ki− ¯ xi k¯ yk
2y. ¯ Then,
h¯ y, w
ki− ¯ xi = 0, (5)
and kw
ki−x
kik converges to 0. Thus, w
kiweakly converges to ¯ x. On the other hand, since T is (w, s)-u.s.c., for u
ki∈ T (w
ki) , u
kistrongly converges to y
∗, due to ky
ki− y
∗k −→ 0 and ku
ki− y
kik −→ 0. By (5) and the pseudo-Dunn property of T , we get
hu
ki, w
ki− ¯ xi ≥ 1
t ku
ki− ¯ yk
2.
Therefore, by taking the limit, we conclude that u
kistrongly converges to ¯ y. Then, y
∗=
¯
y ∈ T (¯ x) and hence ¯ x is a solution to the problem (P 1). 2
Corollary 3.5. Under all the conditions of Theorem 3.4, if M
0is continuous from X equipped with the weak topology to X
∗equipped with the weak topology, then the sequence {x
k} weakly converges to a unique solution of the problem (P 1).
Proof : Following the proof of Theorem 3.4, we need only to prove the uniqueness of
¯
x. Assume that the sequence {x
k} has two weakly cluster points ¯ x and ˆ x. Then, both cluster points can be used as x
∗to defined the function Λ
x∗. Consider the subsequence {k
i} and {l
j} such that x
kiand x
ljweakly converge to ¯ x and ˆ x, respectively. Then, by Lemma 1.4 (i) and the fact ¯ y = ˆ y, we get
Λ ˆ
x∗(x
ki, ε
ki) = ¯ Λ
x∗(x
ki, ε
ki) + R
x∗(x
ki), where
R
x∗(x
ki) = M (ˆ x) − M (¯ x) − hM
0(x
ki), ˆ x − ¯ xi.
Note that ¯ Λ
x∗(x
ki, ε
ki) and ˆ Λ
x∗(x
ki, ε
ki) tend to some ¯ l and ˆ l, respectively. Since M
0is (w, w)-continuous, R
x∗(x
ki) converges to ˆ l − ¯ l. Since M is strongly convex with constant
b
2
, by Proposition 2.2, we obtain
ˆ l ≥ ¯ l + b
2 kˆ x − ¯ xk
2. By interchangeing the role of ¯ x, the same calculations yield
¯ l ≥ ˆ l + b
2 kˆ x − ¯ xk
2. Then
0 ≤ b
2 kˆ x − ¯ xk
2≤ min{¯ l − ˆ l, ˆ l − ¯ l} ≤ 0.
Hence, ¯ x = ˆ x. 2
We note that, if T is strongly monotone with constant a and Lipschitz continuous with constant µ on C, then it has the Dunn property with constant
µa2. Without monotonicity, we shall deal with the pseudomonotone case as follows.
Lemma 3.6. If T is strongly pseudomonotone with constant e and Lipschitz con- tinuous with constant µ on C, then it has the pseudo-Dunn property with constant
µe2.
Proof : Given any (x
1, y
1) and (x
2, y
2) ∈ G(T ), since T is strongly pseudomonotone with constant e,
hy
1, x
2− x
1i ≥ 0 =⇒ hy
2, x
2− x
1i ≥ ekx
2− x
1k
2. Since T is Lipschitz continuous with constant µ,
ky
2− y
1k ≤ µkx
2− x
1k,
which implies
kx
2− x
1k
2≥ 1
µ
2ky
2− y
1k
2. So we have
hy
2, x
2− x
1i ≥ e · 1
µ
2ky
2− y
1k
2.
We complete the proof. 2
We next analyze the convergence of Algorithm 1 invloving strongly monotone operator with Lipschitz property.
Theorem 3.7. Suppose that the problem (P 1) has a solution x
∗∈ C. Under As- sumption 1, for each step k, there exists a unique solution x
k+1generated by Algorithm 1 to the auxiliary problem (P 2). If T is strongly pseudomonotone with constant e on C (x
∗is then unique) and Lipschitz continuous with constant µ on C, and if
∀ k ∈ N, α < ε
k<
µ2eb2+β, where α > 0, β > 0,
then the sequence {x
k} strongly converges to x
∗. Moreover, if M
0is Lipschitz continuous with constant m on C and t =
eαm+
µe< 1, then the error estimation of solution x
∗is
kx
k+1− x
∗k ≤ t
1 − t kx
k− x
∗k.
Proof : By using Theorem 3.4 and Lemma 3.6, {x
k} weakly converges to x
∗. Further- more, {x
k} is strongly convergent. We illustrate as follows : As before, we consider Φ
x∗(x), but ignore Ω
x∗(x, ε). Note that Φ
x∗(x) ≥ 0, and define ∆
k= Φ
x∗(x
k+1)−Φ
x∗(x
k) = s
1+s
2, where the definitions of s
1and s
2are same as Theorem 3.4. Thus,
s
1≤ − b
2 kx
k+1− x
kk
2, and
s
2≤ ε
khy
k, x
∗− x
k+1i
= ε
khy
k− y
k+1, x
∗− x
k+1i + ε
khy
k+1, x
∗− x
k+1i.
By (1) with z = x
k+1, and the strong pseudomonotonicity of T , we get s
2≤ −eε
kkx
k+1− x
∗k
2+ ε
khy
k− y
k+1, x
∗− x
k+1i.
Since T is Lipschitz continuous with constant µ, it follows that
∆
k≤ − b
2 kx
k+1− x
kk
2− eε
kkx
k+1− x
∗k
2+ ε
kµkx
k+1− x
kk · kx
k+1− x
∗k.
By the inequality:
ε
kµkx
k+1− x
kk · kx
k+1− x
∗k ≤ λ
2 kx
k+1− x
kk
2+ ε
2kµ
22λ kx
k+1− x
∗k
2, with λ = b, we have
∆
k≤ ε
2k( −e ε
k+ µ
22b )kx
k+1− x
∗k
2≤ −αβ
2b kx
k+1− x
∗k
2≤ 0.
So the sequence {Φ
x∗(x
k)} is strictly decreasing (unless x
k+1= x
∗) and bounded below by zero. Thus, {Φ
x∗(x
k)} converges. This yields ∆
ktends to 0 and hence {x
k} strongly converges to x
∗.
Now, by (3) with z = x
∗,
hε
ky
k+ M
0(x
k+1) − M
0(x
k), x
∗− x
k+1i ≥ 0.
Since T is strongly pseudomonotone with constnat e, by (1) with z = x
k+1, hy
k+1, x
k+1− x
∗i ≥ ekx
k+1− x
∗k
2.
It follows that
hM
0(x
k+1) − M
0(x
k), x
∗− x
k+1i + ε
khy
k− y
k+1, x
∗− x
k+1i ≥ eε
kkx
k+1− x
∗k
2. (6) On the other hand, since M
0is Lipschitz continuous with constant m, we have
hM
0(x
k+1) − M
0(x
k), x
∗− x
k+1i + ε
khy
k− y
k+1, x
∗− x
k+1i
≤ mkx
k+1− x
kk · kx
k+1− x
∗k + ε
kµkx
k+1− x
kk · kx
k+1− x
∗k
= (m + ε
kµ)kx
k+1− x
kk · kx
k+1− x
∗k. (7) Thus, combining (6) with (7), we obtain
(m + ε
kµ)kx
k+1− x
kk ≥ eε
kkx
k+1− x
∗k.
Hence,
kx
k+1− x
∗k ≤ ( m eε
k+ µ
e )kx
k+1− x
kk
≤ ( m eα + µ
e )kx
k+1− x
kk
≤ t(kx
k+1− x
∗k + kx
k− x
∗k).
It follows that kx
k+1−x
∗k ≤
1−ttkx
k−x
∗k. 2
The following two technical Lemmas will be used to the later theorem.
Lemma 3.8 [2, Lemma 5]. Let {x
n}, {µ
n} and {η
n} be three sequences such that
∞
X
k=1
µ
k< +∞,
x
n≤ η
n+
n−1
X
k=1
µ
ksup
l≤k+1
x
l, ∀ n = 1, 2, · · · .
If the sequence {η
n} is bounded above, then the sequence {x
n} is also bunded above.
Lemma 3.9 [2, Lemma 4]. Let f be a Lipschitz function on a Banach space X and consider the sequence {x
k} ⊂ X and {ε
k} are positive sequences such that
X
k∈N
ε
k= +∞,
∃ ζ, ∀ k ∈ N , kx
k+1− x
kk ≤ ζε
k,
∃ ξ,
Xk∈N
ε
k|f (x
k) − ξ| < +∞.
Then,
k→∞
lim f (x
k) = ξ.
Under the condition that the norm of T does not increase faster than linearly with the norm of x; that is,
∃ p, q > 0 s.t. kx
∗k ≤ pkxk + q, ∀ (x, x
∗) ∈ G(T ), (?) the following result deals with the convergence of the Algorithm 1 involving α-strongly pseudomonotone operators.
Theorem 3.10. Suppose that the problem (P 1) has a solution x
∗∈ C. Assume T and M
0satisfy Assumption 1. If T is α-strongly pseudomonotone with constant e on C, with the above property (?), then the sequence {x
k} strongly converges to the unique solution x
∗.
Proof : Consider the Bregman distance
φ
x∗(x) := M (x
∗) − M (x) − hM
0(x), x
∗− xi, ∀x ∈ C. (8)
Since M
0is strongly monotone with constant b, by Proposition 2.2, M is strongly convex
with constant α =
b2. Thus,
φ
x∗(x
k) = M (x
∗) − M (x
k) − hM
0(x
k), x
∗− x
ki
≥ αkx
∗− x
kk
2= b
2 kx
∗− x
kk
2.
Now we study the variation of φ
x∗for Algorithm 1 : ∆
k= φ
x∗(x
k+1) − φ
x∗(x
k) for each k. Notices that
∆
k= M (x
k) − M (x
k+1) − hM
0(x
k), x
k− x
k+1i + hM
0(x
k) − M
0(x
k+1), x
∗− x
k+1i
= S
1+ S
2, where
S
1:= M (x
k) − M (x
k+1) − hM
0(x
k), x
k− x
k+1i
≤ − b
2 kx
k+1− x
kk
2, (9)
and
S
2:= hM
0(x
k) − M
0(x
k+1), x
∗− x
k+1i
= −B(x
k+1, x
k, x
∗).
By (3) with z = x
∗,
ε
khy
k, x
∗− x
k+1i + B(x
k+1, x
k, x
∗) ≥ 0.
So
S
2≤ ε
khy
k, x
∗− x
k+1i
= ε
khy
k, x
∗− x
ki + ε
khy
k, x
k− x
k+1i.
By (1) with z = x
k, we have some y
∗∈ T (x
∗) such that hy
∗, x
k− x
∗i ≥ 0.
Since T is α-strongly pseudomonotone with constant e, we get hy
k, x
∗−x
ki ≤ −ekx
k−x
∗k
α. Thus,
S
2≤ ε
khy
k, x
k− x
k+1i − eε
kkx
k− x
∗k
α≤ ε
kky
kkkx
k+1− x
kk − eε
kkx
k− x
∗k
α≤ (ε
k)
22b ky
kk
2+ b
2 kx
k+1− x
kk
2− eε
kkx
k− x
∗k
α. (10) Therefore, combining (9) and (10), we obtain
∆
k= S
1+ S
2≤ ε
2k2b ky
kk
2− eε
kkx
k− x
∗k
α.
On the other hand, by the condition of assumption, we have ky
kk ≤ pkx
kk + q ≤ pkx
k− x
∗k + pkx
∗k + q.
Then, by the simple inequaity :
(u + v)
2≤ 2(u
2+ v
2), we get
ky
kk
2≤ 2p
2kx
k− x
∗k
2+ 2(pkx
∗k + q)
2. Thus,
∆
k≤ ε
2k2b ky
kk
2− eε
kkx
k− x
∗k
α≤ ε
2k(γkx
k− x
∗k
2+ δ) − eε
kkx
k− x
∗k
α.
where γ =
pb2and δ =
(pkx∗bk+q)2. Now, summing up this inequaity from k = 0 to n − 1, we have
b
2 kx
n− x
∗k
2≤ φ
x∗(x
n) =
n−1
X
k=0
∆
k+ φ
x∗(x
0)
≤ φ
x∗(x
0) +
n−1
X
k=0
ε
2k(γkx
k− x
∗k
2+ δ) − e
n−1
X
k=0
ε
kkx
k− x
∗k
α≤ φ
x∗(x
0) +
n−1
X
k=0
ε
2k(γkx
k− x
∗k
2+ δ). (11) Then,
kx
n− x
∗k
2≤ η
n+
n−1
X
k=1
µ
kkx
k− x
∗k
2, (12)
where
η
n= 2
b φ
x∗(x
0) + 2γε
20b kx
0− x
∗k
2+
n−1
X
k=0
2δε
2kb , µ
k= 2γε
2kb , ∀ k = 1, 2, · · ·, n − 1.
Since
kx
k− x
∗k
2≤ sup
l≤k+1
kx
l− x
∗k
2, ∀ k = 1, 2, · · ·, it follows from (12) that
kx
n− x
∗k
2≤ η
n+
n−1
X
k=1
µ
ksup
l≤k+1
kx
l− x
∗k
2.
By Lemma 3.8, the sequence {kx
n− x
∗k
2} is bounded, and so is {x
n}. Moreover, from Assumption 1 (iii) and (11), we get
X
k∈N