5.3 Two-level variational inequality and fixed point problems involving pseu-
5.3.1 Main result
0 2 4 6 8 10 12 14 Iteration number (n)
10-2 10-1 100
Errors
Appendix 8.1 Appendix 8.2 Appendix 8.3 Appendix 8.4 Algorithm 3.2
0 5 10 15
Iteration number (n) 10-2
10-1 100
Errors
Appendix 8.1 Appendix 8.2 Appendix 8.3 Appendix 8.4 Algorithm 3.2
0 2 4 6 8 10 12 14
Iteration number (n) 10-3
10-2 10-1 100
Errors
Appendix 8.1 Appendix 8.2 Appendix 8.3 Appendix 8.4 Algorithm 3.2
0 5 10 15
Iteration number (n) 10-2
10-1 100
Errors
Appendix 8.1 Appendix 8.2 Appendix 8.3 Appendix 8.4 Algorithm 3.2
Figure 5.2: Top left: Case I; Top right: Case II; Bottom left: Case III; Bottom right:
Case IV.
5.3 Two-level variational inequality and fixed point
(S3) Given I as the identity operator on H, let T : H → H be a uniformly continuous ϱ−demimetric operator with (I −T) demiclosed on zero. Also, let 0 < h ≤ tn ≤ (1−ϱ) for all n≥1.
(S4) The operatorf :H →H is an L−contraction with L∈(0,12).
(S5) The solution set of the variational inequality is nonempty. That is, V I(C, A) ̸= ∅.
Also, Ω̸=∅ where Ω :={x∈V I(C, A) : T(x) = x}.
(S6) Let {αn}n≥1 and {γn}n≥1 be positive sequences such that αn ∈ (0,1), lim
n→∞αn = 0, P
n≥1
αn =∞, and lim
n→∞
γn
αn = 0.
We introduce and study the following algorithm:
Algorithm 5.3.1. Calculation of the sequence {xn}.
Initialization: Let τ0 >0, µ∈(0,1), β∈(0,2)
Iterative Steps: Given the current iterate xn, calculate xn+1 as follows:
Step 1: Given xn−1, xn with n≥1, choose θn such that 0≤θn≤θ¯n where θ¯n=
(min{n−1+ϵn−1 ,∥x γn
n−xn−1∥}, if xn̸=xn−1 n−1
n−1+ϵ, otherwise
Step 2: Compute
wn =xn+θn(xn−xn−1) yn=PC(wn−τnAwn).
Step 3: Compute zn=wn−βηndn where
dn:=wn−yn−τn(Awn−Ayn) and
ηn:=
(⟨w
n−yn,dn⟩
∥dn∥2 , if dn ̸= 0,
0, otherwise.
Step 4: Compute
vn=zn−tn(zn−T zn),
xn+1 =αnf(vn) + (1−αn)vn. (5.42)
Update
τn+1 =
(min{µ∥Aw∥wn−yn∥
n−Ayn∥, τn},if Awn ̸=Ayn,
τn, otherwise. (5.43)
Step 5: Set n :=n+ 1 and go to Step 1.
Remark 5.3.2. The sequence {τn} generated by (5.43) is a non-increasing sequence and
n→∞lim τn =τ ≥min{τ0,µ L}.
Moreover, we have that
∥Awn−Ayn∥ ≤ µ
τn+1∥wn−yn∥, ∀n≥0.
We begin with the following lemmas which are critical in obtaining our strong convergence result.
Lemma 5.3.3. Let {xn}n≥0 be a sequence in H iteratively generated by Algorithm 5.3.1.
Assume that the conditions (S1)−(S6) hold, then {xn} is bounded.
Proof. Letx∗ ∈VI(C, A). We obtain from the definition of wn that
∥wn−x∗∥=∥xn+θn(xn−xn−1)−x∗∥
≤ ∥xn−x∗∥+θn∥xn−xn−1∥
=∥xn−x∗∥+αnθn
αn∥xn−xn−1∥.
Now from Step 1, observe that θn∥xn−xn−1∥ ≤γn, ∀ n≥1, which implies that θn
αn∥xn−xn−1∥ ≤ γn
αn →0, as n → ∞. (5.44)
Thus, there exists K1 >0 such that θn
αn∥xn−xn−1∥ ≤K1, ∀ n ≥1, (5.45) and so,
∥wn−x∗∥ ≤ ∥xn−x∗∥+αnK1, ∀n≥1. (5.46) Since yn=PC(wn−τnAwn), we obtain from properties of projection that,
⟨wn−τnAwn−yn, yn−x∗⟩ ≥0. (5.47) Also, since x∗ ∈ V I(C, A) and yn ∈ C, we have that ⟨Ax∗, yn−x∗⟩ ≥ 0. By the pseu- domonotonicity of A, we get that,
⟨Ayn, yn−x∗⟩ ≥0. (5.48)
Since τn>0, we have that
⟨τnA(yn), yn−x∗⟩ ≥0. (5.49) Using (5.47) and (5.49), we obtain that;
⟨yn−x∗, wn−yn−τn(Awn−Ayn)⟩ ≥0. (5.50) Equivalently,
⟨yn−x∗, dn⟩ ≥0. (5.51)
Observe from Step 3 and (5.51) that
⟨wn−x∗, dn⟩=⟨wn−yn, dn⟩+⟨yn−x∗, dn⟩
≥ ⟨wn−yn, dn⟩. (5.52)
Using Algorithm 5.3.1, (5.46) and (5.52), we get that (using the fact that∥dn∥ ̸= 0),
∥zn−x∗∥2 =∥wn−βηndn−x∗∥2
=∥wn−x∗∥2+β2ηn2∥dn∥2−2βηn⟨wn−x∗, dn⟩
≤ ∥wn−x∗∥2+β2ηn⟨wn−yn, dn⟩ −2βηn⟨wn−yn, dn⟩
=∥wn−x∗∥2−(2−β)βηn⟨wn−yn, dn⟩
=∥wn−x∗∥2−(2−β)β∥ηndn∥2
≤ ∥wn−x∗∥2− 1
β(2−β)∥wn−zn∥2
≤ ∥xn−x∗∥2− 1
β(2−β)∥wn−zn∥2+αnK1. (5.53) Using Algorithm 5.3.1, (5.53) and the definition of T, we obtain that,
∥vn−x∗∥2 =∥zn−tn(zn−T zn)−x∗∥2
=∥zn−x∗∥2+t2n∥zn−T zn∥2−2tn⟨zn−T zn, zn−x∗⟩
≤ ∥zn−x∗∥2−tn(1−ϱ−tn)∥zn−T zn∥2
≤ ∥xn−x∗∥2− 1
β(2−β)∥wn−zn∥2−tn(1−ϱ−tn)∥zn−T zn∥2+αnK1. (5.54) Again, using Algorithm 7.2.3, (5.54), and the fact thatf is a contraction, we obtain that;
∥xn+1−x∗∥2 =∥αnf(vn) + (1−αn)vn−x∗∥2
≤αn∥f(vn)−f(x∗) +f(x∗)−x∗∥2+ (1−αn)∥vn−x∗∥2
≤2αn[L∥vn−x∗∥2+∥f(x∗)−x∗∥2] + (1−αn)∥vn−x∗∥2
= [1−αn(1−2L)]∥vn−x∗∥2 + 2αn∥f(x∗)−x∗∥2
≤[1−αn(1−2L)]∥xn−x∗∥2+αn(1−2L) 2
1−2L∥f(x∗)−x∗∥2
−tn[1−αn(1−2L)](1−ϱ−tn)∥zn−T zn∥2
− 1
β(2−β)[1−αn(1−2L)]∥wn−zn∥2
+αn[1−αn(1−2L)]αnK1 (5.55)
Using the assumptions S3 and S6, and for some K2 >0, we have that;
∥xn+1−x∗∥2 ≤[1−αn(1−2L)] ∥xn−x∗∥2+K2
+αn(1−2L) 2
1−2L∥f(x∗)−x∗∥2. (5.56) If we define M := max
∥x0−x∗∥2+K2,1−2L2 ∥f(x∗)−x∗∥2 , then it is easy to see that
∥xn+1−x∗∥2 ≤M for all n≥0. (5.57) Thus, {xn}∞n=1 is bounded. Consequently, {wn},{Awn},{yn},{Ayn},{zn},{T zn}, and {vn} are all bounded. ■
Lemma 5.3.4. Suppose that conditions S1 and S2 hold, and {xn}∞n=1 is a sequence gen- erated by Algorithm 5.3.1, then, there exists n0 ∈N such that
∥wn−yn∥2 ≤
1 +µττn
n+1
2
h
1−µττn
n+1
βi2∥zn−wn∥2 ∀ n≥n0. (5.58) Proof. From Algorithm 3.1, we easily see that,
∥dn∥=∥wn−yn−τn(Awn−Ayn)∥ ≥ ∥wn−yn∥ −τn∥Awn−Ayn∥
≥ ∥wn−yn∥ −µ τn
τn+1∥wn−yn∥
=
1−µ τn τn+1
∥wn−yn∥. (5.59)
But lim
n→∞
1−µττn
n+1
= 1−µ >0. So, there exists n0 ∈N such that 1−µ τn
τn+1 > 1−µ
2 ∀n ≥n0.
Therefore, for each n≥n0, we have that ∥dn∥> 1−µ2 ∥wn−yn∥>0.
Moreover,
∥dn∥ ≤ ∥wn−yn∥+τn∥Awn−Ayn∥
≤ ∥wn−yn∥+µ τn
τn+1∥wn−yn∥
=
1 +µ τn τn+1
∥wn−yn∥ ∀ n ≥n0. (5.60) Thus, for each n≥n0, we obtain that,
∥dn∥2 ≤
1 +µ τn
τn+1 2
∥wn−yn∥2. (5.61)
Also, from Algorithm 5.3.1, we get that,
⟨wn−yn, dn⟩=⟨wn−yn, wn−yn−τn(Awn−Ayn)⟩
=∥wn−yn∥2−τn⟨wn−yn, Awn−Ayn⟩
≥
1−µ τn τn+1
∥wn−yn∥2. (5.62)
So, for each n≥n0, we obtain from (5.61) and (5.62);
ηn = ⟨wn−yn, dn⟩
∥dn∥2 ≥
1−µττn
n+1
1 +µττn
n+1
2. (5.63)
Also, from Algorithm 5.3.1 and (5.62), we obtain that;
ηn∥dn∥2 =⟨wn−yn, dn⟩ ≥
1−µ τn τn+1
∥wn−yn∥2. (5.64)
Thus, from (5.64) and Algorithm 5.3.1,
∥wn−yn∥2 ≤
1−µ τn τn+1
−1
ηn∥dn∥2
=
1−µ τn τn+1
−1
∥βηndn∥2 1 β2
1 ηn
=
1−µ τn τn+1
−1
∥zn−wn∥2 1 β2
1
ηn. (5.65)
Thus, using (5.63) in (5.65), we obtain that,
∥wn−yn∥2 ≤
1 +µττn
n+1
2
h
1−µττn
n+1
βi2∥zn−wn∥2 ∀ n ≥n0.
■
Lemma 5.3.5. Let {xn} be sequence generated by Algorithm5.3.1 under Assumptions S.
Suppose that there exists a subsequence {xnb} of {xn} which converges weakly to x¯ ∈ H and lim
b→∞∥wnb −ynb∥= 0, then x¯∈V I(C, A).
Proof. Clearly, from Algorithm 5.3.1 and AssumptionS6, we obtain that,
∥wn−xn∥=αn
θn
αn∥xn−xn−1∥ ≤ γn
αn →0 as n→ ∞. (5.66) Since {xn} is bounded, there exists a subsequence {xnb} of {xn} which converges weakly to ¯x∈H. From (2.9), we have that;
⟨wnb−τnbAwnb −ynb, x−ynb⟩ ≤0, ∀ x∈C.
Rearranging, we obtain that, 1
τnb⟨wnb −ynb, x−ynb⟩ ≤ ⟨Awnb, x−ynb⟩, ∀ x∈C.
Thus, 1 τnb
⟨wnb −ynb, x−ynb⟩+⟨Awnb, ynb −wnb⟩ ≤ ⟨Awnb, x−wnb⟩, ∀ x∈C. (5.67) Fixx∈C. Since{wnb}is bounded andAis Lipschitz continuous, then{Awnb}is bounded.
Also, using the assumption that lim
b→∞∥wnb−ynb∥= 0 and, taking lim inf of (5.67) asb→ ∞, we get
lim inf
b→∞ ⟨Awnb, x−wnb⟩ ≥0. (5.68) By assumptions that lim
b→∞∥wnb−ynb∥= 0, and A is Lipschitz continuous, we get that
b→∞lim ∥Awnb −Aynb∥= 0. (5.69) It is easy to see that,
⟨Aynb, x−ynb⟩=⟨Aynb −Awnb, x−wnb⟩+⟨Awnb, x−wnb⟩+⟨Aynb, wnb −ynb⟩ (5.70) Using (5.68) and (5.69) in (5.70), we get that;
lim inf
b→∞ ⟨Aynb, x−ynb⟩ ≥0. (5.71) Let us choose a decreasing sequence of positive numbers {βb} with lim
b→∞βb = 0. For each b ∈N, we denote byNb the smallest positive number such that
⟨Aynj, x−ynj⟩+βb ≥0 ∀ j ≥Nb. (5.72) Since {βb} is decreasing, then the sequence {Nb} is increasing. Furthermore, for each b ∈ N, since {ynb} ⊂ C, we assume thatAyNb ̸= 0 (else, yNb is a solution of of VI(C,A)).
Set
mNb = AyNb
∥AyNb∥2,
where ⟨AyNb, mNb⟩= 1 for eachb ∈N. From (5.72), for eachb ∈N, we obtain that,
⟨AyNb, x+βbmNb−yNb⟩ ≥0. (5.73) Since A is pseudomonotone, we obtain that,
⟨A(x+βbmNb), x+βbmNb−yNb⟩ ≥0.
Thus,
⟨Ax, x−yNb⟩ ≥ ⟨Ax−A(x+βbmNb), x+βbmNb−yNb⟩ −βb⟨Ax, mNb⟩. (5.74) Since {xnb} converges weakly to ¯x as b→ ∞, then, using (5.66) and the assumption that
b→∞lim ∥wnb −ynb∥ = 0, we easily obtain that ynb ⇀ x¯ as b → ∞. By sequentially weakly continuity of A, we have that Aynb ⇀ A¯x as b → ∞. Assume that A¯x̸= 0 (otherwise, ¯x is a solution), by the sequentially weakly lower semicontinuity of norm, we have that
0<∥A¯x∥ ≤lim inf
b→∞ ∥Aynb∥.
Since {yNb} ⊂ {ynb}and βb →0 as b→ ∞, we have that, 0≤lim sup
b→∞
∥βbmNb∥= lim sup
b→∞
βb
∥Aynb∥
≤ lim supb→∞βb
lim infb→∞∥Aynb∥ ≤ 0
∥A¯x∥ = 0.
Thus, ∥βbmNb∥ → 0 as b → ∞. Hence, taking the limit as b → ∞ in (5.74), we obtain that,
⟨Ax, x−x⟩ ≥¯ 0.
Since x∈H is arbitrary, we obtain from Lemma 2.5.10 that ¯x∈V I(C, A)■
Lemma 5.3.6. Let {xn}n≥0 be a sequence in H iteratively generated by Algorithm 5.3.1.
Assume that ζn = (1−L)αn, and assumptions S1−S6 hold, then
∥xn+1−x∗∥2 ≤(1−ζn)h
∥xn−x∗∥2+αnK1i
+ζn 2
1−L⟨f(x∗)−x∗, xn+1−x∗⟩. (5.75) Proof. From Algorithm 5.3.1, (5.54), Lemma 2.5.18 and Assumption S6, we obtain that,
∥xn+1−x∗∥2 =∥αnf(vn) + (1−αn)vn−x∗∥2
=∥αn[f(vn)−f(x∗)] + (1−αn)[vn−x∗] +αn[f(x∗)−x∗]∥2
≤ ∥αn[f(vn)−f(x∗)] + (1−αn)[vn−x∗]∥2+ 2αn⟨xn+1−x∗, f(x∗)−x∗⟩
≤[1−(1−L)αn]∥vn−x∗∥2+ 2αn⟨xn+1−x∗, f(x∗)−x∗⟩
≤[1−(1−L)αn]∥xn−x∗∥2+ 2αn⟨xn+1−x∗, f(x∗)−x∗⟩ + [1−(1−L)αn]αnK1− 1
β(2−β)[1−(1−L)αn]∥wn−zn∥2
−tn(1−ϱ−tn)[1−(1−L)αn]∥zn−T zn∥2.
So, using assumption S4, we obtain that,
∥xn+1−x∗∥2 ≤[1−(1−L)αn]∥xn−x∗∥2+ [1−(1−L)αn]αnK1 + (1−L)αn 2
1−L⟨f(x∗)−x∗, xn+1−x∗⟩ (5.76) If we let ζn= (1−L)αn, then we obtain from (5.76) that,
∥xn+1−x∗∥2 ≤(1−ζn)h
∥xn−x∗∥2+αnK1i
+ζn 2
1−L⟨f(x∗)−x∗, xn+1−x∗⟩. (5.77)
■
Theorem 5.3.7. Let {xn}n≥0 be a sequence in H iteratively generated by Algorithm 5.3.1. Assume that the conditions (S1) − (S6) hold, then {xn} converges strongly to x∗ =PΩ(f(x∗)).
Proof.
From Lemma 5.3.3,{xn}∞n=1 is bounded. Letx∗ =PΩ(f(x∗)). Now, from Algorithm7.2.3, (5.54), and Lemma 2.5.18, we obtain that,
∥xn+1−x∗∥2 =∥αnf(vn) + (1−αn)vn−x∗∥2
≤ ∥αnf(vn) + (1−αn)vn−x∗−αn(f(vn)−x∗)∥2 + 2αn⟨f(vn)−x∗, xn+1−x∗⟩
≤(1−αn)∥vn−x∗∥2+ 2αn⟨f(vn)−x∗, xn+1−x∗⟩
≤(1−αn)∥xn−x∗∥2+ 2αn⟨f(vn)−x∗, xn+1−x∗⟩
− 1
β(2−β)(1−αn)∥wn−zn∥2+ (1−αn)αnK1
−(1−αn)tn(1−ϱ−tn)∥zn−T zn∥2
≤(1−αn)∥xn−x∗∥2+ 2αn⟨f(vn)−x∗, xn+1−x∗⟩
− 1
β(2−β)(1−αn)∥wn−zn∥2+αnK1
−(1−αn)tn(1−ϱ−tn)∥zn−T zn∥2. (5.78) To show that the sequence{∥xn+1−x∗∥}converges to zero, we consider two possible cases:
Case 1: Suppose there exists n0 ∈N such that the real sequence ∥xn−x∗∥ is decreasing for all n ≥n0. It then follows that ∥xn−x∗∥ is convergent. Since {xn} is bounded, then from (5.78), assumptionS3, and using the fact that αn→0 as n→ ∞, we obtain that
n→∞lim ∥zn−T zn∥= 0, lim
n→∞∥wn−zn∥= 0. (5.79) From Algorithm 5.3.1, we get that;
∥vn−zn∥=tn∥zn−T zn∥ →0 as n→ ∞,
∥xn+1−vn∥=αn∥f(vn)−vn∥ →0 as n → ∞. (5.80)
Thus, using (5.66), (5.79), and (5.80), we obtain that,
∥xn+1−xn∥ ≤ ∥xn+1−vn∥+∥vn−zn∥+∥zn−wn∥+∥wn−xn∥
→0 as n → ∞. (5.81)
From the results above, we can easily deduce that,
∥xn+1−xn∥ →0 as n → ∞. (5.82)
Claim:
lim sup
n→∞
D
f(x∗)−x∗, xn−x∗E
≤0. (5.83)
Proof of Claim: Let{xnq}q≥0 be a subsequence of{xn}n≥0 such that lim sup
n→∞
D
f(x∗)−x∗, xn−x∗ E
= lim
q→∞
D
f(x∗)−x∗, xnq −x∗ E
(5.84) Since{xnq}q≥0is a bounded sequence inH, there exists a subsequence{xnqb}b≥0of{xnq}q≥0
which converges weakly to ¯x in H. Hence, xnqb ⇀ x¯ as b → ∞. For convenience, and WLOG, we will represent {xnqb}b≥0 by {xnb}b≥0.
Recall from Lemma 5.3.4,
∥wnb−ynb∥2 ≤
1 +µττnb
nb+1
2
h
1−µττnb
nb+1
αi2∥znb−wnb∥2 ∀ nb ≥n0.
Using (5.79), we have that ∥wnb −ynb∥ → 0 as n → ∞. Thus, using Lemma 5.3.5, and the fact that xnb ⇀x¯ asb → ∞, we have that ¯x∈V I(C, A).
Next, we show that ¯x∈ Ω. From (5.79) and the fact that (I−T) is demiclosed at zero, we can deduce that ¯x∈F(T).
Therefore, it follows from our argument above that ¯x∈Ω. Thus, we obtain from Lemma 2.5.26 and (5.83) that
lim sup
n→∞
D
f(x∗)−x∗, xn−x∗E
= lim
q→∞
D
f(x∗)−x∗, xnq −x∗E
≤0. (5.85) Also, using the fact that ∥xn+1−xn∥ →0 as n→ ∞, we get from (5.85) that,
lim sup
n→∞
D
f(x∗)−x∗, xn+1−x∗E
≤0. (5.86)
Recall that from Lemma 5.3.6, we obtain that,
∥xn+1−x∗∥2 ≤(1−ζn) h
∥xn−x∗∥2+αnK1
i +ζn
2
1−L⟨f(x∗)−x∗, xn+1−x∗⟩. (5.87)
Thus, by Lemma 2.5.26, it follows that ∥xn−x∗∥ converges strongly to zero as n → ∞.
Hence, {xn} converges to x∗ ∈Ω. This completes the proof for the first case.
Case 2: Suppose that there exists a subsequence{∥xnj−x∗∥}∞j=0 of {∥xn−x∗∥}n≥0 such that ∥xnj − x∗∥ < ∥xnj+1 −x∗∥ for all j ≥ 0, then we obtain by Lemma 2.5.41 that there exists a non-decreasing sequence {mk}∞k=1 ⊂ N such that mk → ∞ as k → ∞ and
∥xmk −x∗∥ ≤ ∥xmk+1−x∗∥ for all k ∈ N. Since the sequences {xmk}∞k=1 is bounded, we obtain from (5.66), (5.79), (5.80) (and using the arguments displayed in Case 1) that as k → ∞;
∥wmk −xmk∥, ∥wmk−zmk∥ →0
∥vmk −zmk∥, ∥xmk+1−vmk∥ →0. (5.88) Applying (5.88), we obtain that,
∥xmk+1−xmk∥ →0 as k → ∞. (5.89) Furthermore, following the arguments used in Case 1, we obtain that
lim sup
k→∞
D
f(x∗)−x∗, xmk−x∗E
≤0. (5.90)
Again, using (5.89) and the same arguments as case 1, we have that, lim sup
k→∞
D
f(x∗)−x∗, xmk+1−x∗E
≤0. (5.91)
Again, from Lemma 5.3.6 and using the fact that ∥xmk −x∗∥2 ≤ ∥xmk+1 −x∗∥2 for any k ∈N we obtain by rearranging that,
ζmk∥xmk −x∗∥2 ≤ ∥xmk −x∗∥2− ∥xmk+1−x∗∥2+ζmk 2
1−L⟨f(x∗)−x∗, xmk+1−x∗⟩
+ (1−ζmk)αmkK1. (5.92)
Applying Lemma 2.5.41, we obtain that,
ζmk∥xk−x∗∥2 ≤ ∥xmk −x∗∥2− ∥xmk+1−x∗∥2+ζmk 2
1−L⟨f(x∗)−x∗, xmk+1−x∗⟩
+(1−ζmk)αmkK1. (5.93)
Thus, it follows that lim sup
k→∞
∥xk−x∗∥= 0. Hence, lim
k→∞∥xk−x∗∥= 0.
Thus, xn→x∗ asn → ∞. This completes the proof. ■
Remark 5.3.8. Theorem 5.3.7provides an iterative scheme for approximation of the so- lution of two-level variational inequality and fixed point problem where the A is Lipschitz continuous and pseudo-monotone and T is ϱ−demimetric. The result extends to solu- tion of variational inequality problem of monotone operators since every pseudo-monotone operator is monotone.