5.4 An inertial extrapolation method for solving
5.4.1 Main result
5.4 An inertial extrapolation method for solving
Step 2. Set
wn=xn+θn(xn−xn−1).
Then, compute
un=wn−τnGwn and zn=PC(un−A(un)). (5.96) Step 3. Computeyn:=un−ηn(un−zn),whereηn:=αmn andmn is the smallest nonnegative
integer satisfying
⟨A(yn), un−zn⟩ ≥ σ
2||un−zn||2. (5.97) Step 4. Compute
xn+1 = (1−βn−αn)un+βnPCn(un), (5.98) where Cn :=
x∈H1 :Fn(x)≤0 and
Fn(x) := ⟨A(yn), x−yn⟩. (5.99) Update:
τn+1 =
(minn δ||w
n−un||
||Gwn−Gun||, τno
, if Gwn̸=Gun,
τn, otherwise.
(5.100) Stopping criterion: If zn =wn =un =xn, then stop, otherwise, set n :=n+ 1 and go back to Step 1.
Remark 5.4.3.
(a) Assumption 5.4.1 (1) (b) requires that the operator A is pseudomonotone and uni- formly continuous which is much more, a weaker assumption than the monotonicity and Lipschitz continuity assumption used in [228,229], as well as the inverse strongly monotonicity assumptions used in other papers for solving split variational inequality problems (see for example [54, 119, 136, 137, 178]). Thus, our method is applicable for a much more general class of Pseudomonotone and uniformly continuous opera- tors.
(b) In addition, our method requires at each iteration, only one projection onto C and another projection onto the half-space Cn, which can be easily computed. In fact,
PCn(un) :=
(un− ⟨A(y||A(yn),un−yn⟩
n)||2 A(yn), if ⟨A(yn), un−yn⟩>0, un, if ⟨A(yn), un−yn⟩ ≤0.
Hence, our method is less computationally expensive than other methods in the lit- erature [54, 228, 229], for solving the split variational inequality problems.
Remark 5.4.4. If {yn} is bounded, then Fn is Lipschitz continuous.
Indeed, if{yn}is bounded, then by Lemma 2.5.50, there exists M >¯ 0such that∥A(yn)∥ ≤ M¯ for all n≥1. Thus, for each x, y ∈C, we obtain
∥Fn(x)−Fn(y)∥=∥⟨A(yn), x−yn⟩ − ⟨A(yn), y −yn⟩∥
=∥⟨A(yn), x−y⟩∥
≤ ∥A(yn)∥∥x−y∥
≤M∥x¯ −y∥.
Therefore, Fn is Lipschitz continuous.
We now show that the stopping criterion of Algorithm 5.4.2 is valid.
Lemma 5.4.5. If zn=wn =un=xn in Algorithm 5.4.2, then xn ∈Γ.
Proof. If zn=wn=un=xn, then from (5.96), it is clear that xn =PC(xn−A(xn)).
Also, we obtain that xn = xn − τnGxn = xn − τnT∗(I − S)T xn, which implies that T∗(I−S)T xn = 0. That is,
ST xn =T xn+ ¯z, where T∗z¯= 0. Now, let z ∈Γ, then we obtain that
∥T xn−T z∥2 =∥T xn−T z∥2+ 2⟨xn−z, T∗z⟩¯
=∥T xn−T z∥2+ 2⟨T xn−T z,z⟩¯
=∥T xn−T z+ ¯z∥2− ∥¯z∥2
=∥ST xn−T z∥2− ∥¯z∥2
≤ ∥T xn−T z∥2− ∥¯z∥2,
which implies that ∥¯z∥ = 0. That is, ¯z = 0. Hence ST xn = T xn, which gives that T xn∈F(S).Therefore, xn∈Γ.
Next, we show that the limit of the stepsize {τn} generated by (5.100) exists.
Lemma 5.4.6. The limit of the stepsize {τn} exists and lim
n→∞τn>0.
Proof. From (5.100), it is obvious that τn+1 ≤τn ∀n ∈N.
Also, we know by Lemma 2.5.25 that G is 2∥T∥2-Lipschitz continuous. Thus, we get in the case of Gun̸=Gwn that
τn+1 = min
δ∥wn−un∥
∥Gwn−Gun∥, τn
≥min δ
2∥T∥2, τn
. Hence, by induction, we obtain that {τn} is bounded below by minn
δ
2∥T∥2, τ1o
. Hence, the limit of {τn} exists and lim
n→∞τn≥minn
τ1,2∥Tδ∥2
o
>0.
Lemma 5.4.7. Let Assumption 5.4.1 hold and let {xn} be a sequence generated by Algo- rithm 5.4.2. Assume that the subsequence {xnk} of {xn} converges weakly to a point x∗, and lim
k→∞∥xnk −unk∥= 0, lim
k→∞∥unk −wnk∥= 0 and lim
k→∞∥unk −znk||= 0, then x∗ ∈Γ.
Proof. By Lemma 2.5.48 we obtain
⟨unk −A(unk)−znk, x−znk⟩ ≤0, ∀ x∈C, which implies that
⟨unk −znk, x−znk⟩ ≤ ⟨A(unk), x−znk⟩, ∀ x∈C.
Hence
⟨unk−znk, x−znk⟩+⟨A(unk), znk−unk⟩ ≤ ⟨A(unk), x−unk⟩, ∀x∈C. (5.101) Fix x∈C and let k → ∞in (5.101). Since lim
k→∞∥unk−znk||= 0, we have 0≤lim inf
k→∞ ⟨A(unk), x−unk⟩ ∀x∈C. (5.102) Now, choose a sequence{δk}of positive numbers such thatδk+1 ≤δk, ∀ k ≥1 and δk → 0 as k → ∞.Then, for each δk,we denote by Nk (which exists as a result of (5.102)) the smallest positive integer such that
⟨A(unj), x−unj⟩+δk ≥0 ∀j ≥Nk. (5.103) Since {δk} is decreasing, we have that {Nk} is increasing. Furthermore, we set for each k ≥1, mNk = ∥A(uA(uNk)
Nk)∥2, providedA(uNk)̸= 0.Then it is easy to see that⟨A(uNk), mNk⟩= 1 for each k≥1. Thus, by (5.103), we have that
⟨A(uNk), x+δkmNk−uNk⟩ ≥0, which implies by the pseudo-monotonicity of A that
⟨A(x+δkmNk), x+δkmNk−uNk⟩ ≥0. (5.104) Since {xnk} converges weakly to x∗, we obtain by our hypothesis that {unk}, {znk} and {wnk} also converge weakly to x∗. Thus, by the sequentially weakly continuity of A, we have that {A(unk)}converges weakly to A(x∗). IfA(x∗) = 0, then x∗ ∈V I(C, A).On the other hand, if we suppose thatA(x∗)̸= 0,then by the weakly lower semicontinuity of∥ · ∥, we obtain that
0<∥A(x∗)∥ ≤lim inf
k→∞ ∥A(unk)∥.
Since {uNk} ⊂ {unk}, we obtain that 0≤lim sup
k→∞
∥δkmNk∥= lim sup
k→∞
δk
∥A(unk)∥
≤
lim sup
k→∞
δk lim inf
k→∞ ∥A(unk)∥
≤ 0
∥A(x∗)∥ = 0.
Therefore, lim
k→∞∥δkmNk∥= 0. Thus, lettingk → ∞ in (5.104) yields
⟨A(x), x−x∗⟩ ≥0 ∀x∈C, (5.105) which implies by Lemma 2.5.9 that x∗ ∈V I(C, A).
Now, to show that T x∗ ∈ F(S), recall that Lemma 5.4.6 gives that lim
n→∞τn = τ > 0.
Furthermore, sinceG≡T∗(I−S)T is Lipschitz continuous,{T∗(I−S)T wnk}is bounded.
Hence, we get that
∥(I−τnkT∗(I−S)T)wnk−(I−τ T∗(I−S)T)wnk∥=|τnk−τ|∥T∗(I−S)T wnk∥ →0, as k → ∞.
That is, lim
k→∞∥unk−(I−τ T∗(I −S)T)wnk∥= 0, which implies by our hypothesis that
k→∞lim ||wnk− I−τ T∗(I−S)T
wnk||= 0. (5.106) Thus, we obtain from Lemma2.5.55, thatx∗ ∈F I−τ T∗(I−S)T
.Hence, using the same line of argument as in the proof of Lemma 5.4.5, we obtain that T x∗ ∈ F(S). Therefore x∗ ∈Γ.
Lemma 5.4.8. Suppose Assumption 5.4.1 holds. Then, the sequence {xn} generated by Algorithm 5.4.2is bounded.
Proof. From Lemma 2.5.16 (ii), (iii), and (iv), un can be written in the form
un = (1−βn)wn+βnVnwn, (5.107) where βn=τn||T||2 and Vn is a nonexpansive mapping for eachn ∈N. That is, (I−τnG) is τn||T||2-averaged.
Now, let x∗ ∈Γ, we obtain from (5.107) that
||un−x∗||2 =||(1−βn)wn+βnVnwn−x∗||2
= (1−βn)||wn−x∗||2+βn||Vnwn−x∗||2−βn(1−βn)||Vnwn−wn||2
≤ ||wn−x∗||2−βn(1−βn)||Vnwn−wn||2 (5.108)
≤ ||wn−x∗||2.
On the other hand,
||wn−x∗||=||xn+θn(xn−xn−1)−x∗||
≤ ||xn−x∗||+θn||xn−xn−1||
=||xn−x∗||+αnθn αn
||xn−xn−1||.
Now from (5.95), we observe that θn
αn||xn−xn−1|| ≤ ϵn
αn → 0. Thus there exists M1 >0 such that θn
αn||xn−xn−1|| ≤M1 ∀n ≥1, and so,
||un−x∗|| ≤ ||wn−x∗|| ≤ ||xn−x∗||+αnM1. (5.109) From (5.98), we obtain
||xn+1−x∗||=||(1−βn−αn)un+βnPCnun−x∗||
=||(1−βn−αn)(un−x∗) +βn(PCnun−x∗)−αnx∗||
≤ ||(1−βn−αn)(un−x∗) +βn(PCnun−x∗)||+αn||x∗||. (5.110) Note that
||(1−βn−αn)(un−x∗) +βn(PCnun−x∗)||2
= (1−βn−αn)2||un−x∗||2+ 2(1−βn−αn)βn⟨un−x∗, PCnun−x∗⟩ +βn2||PCnun−x∗||2
≤(1−βn−αn)2||un−x∗||2+ 2(1−βn−αn)βn||un−x∗||||PCnun−x∗||
+βn2||PCnun−x∗||2
≤(1−βn−αn)2||un−x∗||2+ (1−βn−αn)βn||un−x∗||2 + (1−βn−αn)βn||PCnun−x∗||2+βn2||PCnun−x∗||2
≤(1−βn−αn)(1−αn)||un−x∗||2+ (1−αn)βn||PCnun−x∗||2 (5.111)
≤(1−βn−αn)(1−αn)||un−x∗||2+ (1−αn)βn||un−x∗||2
= (1−αn)2||un−x∗||2, which implies from (5.109) that
||(1−βn−αn)(un−x∗) +βn(PCnun−x∗)||= (1−αn)||xn−x∗||+ (1−αn)αnM1
≤(1−αn)||xn−x∗||+αnM1. (5.112) Now, combining (5.110) and (5.112) gives
||xn+1−x∗|| ≤(1−αn)||xn−x∗||+αn||x∗||+αnM1
= (1−αn)||xn−x∗||+αn(||x∗||+M1).
It follows from Lemma 2.5.24 that {xn} is bounded. Consequently, {A(un)},{un},{wn}, {zn} and {A(yn)}are also bounded.
Lemma 5.4.9. Let{xn}be generated by Algorithm5.4.2such that Assumption5.4.1holds.
Assume that lim
k→∞||PCnkunk−unk||= 0 = lim
k→∞ηnk||unk−znk||2, then lim
k→∞∥unk−znk∥= 0.
Proof. From Step 3, we have that ηn = αmn with α ∈ (0,1). Hence, {ηn} ⊂ (0,1) is bounded. Thus, there exists a subsequence {ηnk} of {ηn} such that lim inf
k→∞ ηnk ≥0.
Case 1. Suppose that lim inf
k→∞ ηnk >0. Then
0≤ ∥unk−znk∥2 = ηnk∥unk−znk∥2 ηnk , which implies that
lim sup
k→∞
∥unk −znk∥2 ≤lim sup
k→∞
ηnk∥unk−znk∥2
lim sup
k→∞
1 ηnk
=
lim sup
k→∞
ηnk∥unk −znk∥2 1 lim inf
k→∞ ηnk
= 0.
Hence, lim sup
k→∞
∥unk−znk∥= 0. Therefore, lim
k→∞∥unk −znk∥= 0.
Case 2 . Suppose that lim inf
k→∞ ηnk = 0.Then, we may assume without loss of generality that
k→∞lim ηnk = 0 and lim
k→∞∥unk −znk∥=a≥0.
Define ¯ynk−unk := 1
αηnk(znk−unk).Since {znk−unk}is bounded and lim
k→∞ηnk = 0, then
k→∞lim ∥¯ynk −unk∥= 0. (5.113) From the step-size rule and definition of ¯yk, we have
⟨A(¯ynk), unk−znk⟩< σ
2∥unk −znk∥2, ∀k ∈N. This implies that,
2⟨A(unk), unk−znk⟩+ 2⟨A(¯ynk)−A(unk), unk −znk⟩< σ∥unk −znk∥2, ∀k ∈N. Setting µnk :=unk−A(unk), we obtain
2⟨unk−µnk, unk −znk⟩+ 2⟨A(¯ynk)−A(unk), unk−znk⟩< σ∥unk−znk∥2, ∀k ∈N. Using Lemma 2.5.18(ii) we have
2⟨unk−µnk, unk−znk⟩=∥unk−znk∥2+∥unk −µnk∥2 − ∥znk −µnk∥2.
Therefore,
∥unk−µnk∥2− ∥znk−µnk∥2 <(σ−1)∥unk−znk∥2−2⟨A(¯ynk)−A(unk), unk −znk⟩,
∀k∈N. Since A is uniformly continuous on bounded subsets of C and (5.113), if a > 0 then the right hand side of the inequality above converges to (σ −1)a < 0 as k → ∞. From the last inequality, we obtain
lim sup
k→∞
∥unk−µnk∥2− ∥znk −µnk∥2
≤(σ−1)a <0.
For ϵ= −(σ−1)a
2 >0,there exists N ∈N such that
∥unk −µnk∥2− ∥znk −µnk∥2 ≤(σ−1)a+ϵ= (σ−1)a
2 ≤0 ∀k∈N, k ≥N; Thus
∥unk −µnk∥<∥znk −µnk∥ ∀k ∈N, k ≥N.
which is a contradiction to the definition of znk = PC unk −A(unk)
. Hence a = 0, the proof is complete.
Theorem 5.4.10. Suppose that Assumption 5.4.1 holds. Then, the sequence {xn} gener- ated by Algorithm 5.4.2 converges strongly to the point x∗ ∈Γ, where
∥x∗∥= min{∥¯x∥: ¯x∈Γ}.
Proof. Using Lemma 2.5.8 and Lemma 2.5.49, we obtain
||PCn(un)−x∗|| ≤ ||un−x∗||2− ||PCn(un)−un||2
=||un−x∗||2 −dist2(un, Cn)
≤ ||un−x∗||2−1
mFn(un)2
=||un−x∗||2 −1
m⟨A(yn), un−yn⟩2
=||un−x∗||2 −ηn
m⟨A(yn), un−zn⟩2
≤ ||un−x∗||2−σηn
2m||un−zn||22
. (5.114)
Furthermore, we have
xn+1 = (1−βn−αn)un+βnPCn(un) = (1−βn)un+βnPCn(un)−αnun.
Let ϕn = (1−βn)un+βnPCn(un), then we obtain
||ϕn−x∗||2 =||(1−βn)un+βnPCn(un)−x∗||2
=||(1−βn)(un−x∗) +βn(PCn(un)−x∗)||2
= (1−βn)2||un−x∗||2 +βn2||PCn(un)−x∗||2+ 2(1−βn)βn⟨un−x∗, PCn(un)−x∗⟩
≤(1−βn)2||un−x∗||2+βn2||PCn(un)−x∗||2+ 2(1−βn)βn||un−x∗||||PCn(un)−x∗||
≤(1−βn)2||un−x∗||2+βn2||PCn(un)−x∗||2+ (1−βn)βn||un−x∗||2 (5.115) + (1−βn)βn||PCn(un)−x∗||2
= (1−βn)||un−x∗||2+βn||PCn(un)−x∗||2
≤ ||wn−x∗||2. (5.116)
On the other hand, we have
||wn−x∗||2 =||xn+θn(xn−xn−1)−x∗||2
≤ ||xn−x∗||2+ 2θn⟨xn−x∗, xn−xn−1⟩+θn2||xn−xn−1||2
≤ ||xn−x∗||2+ 2θn||xn−x∗||||xn−xn−1||+θn2||xn−xn−1||2
=||xn−x∗||2+θn||xn−xn−1||(2||xn−x∗||+θn||xn−xn−1||)
≤ ||xn−x∗||2+θn||xn−xn−1||M2, (5.117) for some M2 >0. Combining (5.116) and (5.117) we obtain
||ϕn−x∗||2 ≤ ||xn−x∗||2+θn||xn−xn−1||M2. (5.118) Since ϕn = (1−βn)un+βnPCn(un), we have un−ϕn = βn(un−PCn(un)). Therefore it follows that
xn+1 =ϕn−αnun = (1−αn)ϕn−αn(un−ϕn) = (1−αn)ϕn−αnβn(un−PCn(un)).
Thus, from Lemma 2.5.18(ii), we obtain
||xn+1−x∗||2 =||(1−αn)ϕn−αnβn(un−PCn(un))−x∗||2
=||(1−αn)(ϕn−x∗)− αnβn(un−PCn(un)) +αnx∗
||2
≤(1−αn)2||ϕn−x∗||2−2⟨αnβn(un−PCn(un)) +αnx∗, xn+1−x∗⟩
≤(1−αn)||ϕn−x∗||2+ 2⟨αnβn(un−PCn(un)), x∗−xn+1⟩ + 2αn⟨x∗, x∗−xn+1⟩
≤(1−αn)||ϕn−x∗||2+ 2αnβn||un−PCn(un)||||x∗−xn+1||
+ 2αn⟨x∗, x∗−xn+1⟩. (5.119)
Therefore, using (5.118) and (5.119), we have
||xn+1−x∗||2 ≤(1−αn)||xn−x∗||2+ (1−αn)θn||xn−xn−1||M2 + 2αnβn||un−PCn(un)||||x∗−xn+1||+ 2αn⟨x∗, x∗−xn+1⟩
= (1−αn)||xn−x∗||2+αnhθn
αn(1−αn)||xn−xn−1||M2+ 2βn||un−PCn(un)||||x∗−xn+1||
+ 2⟨x∗, x∗−xn+1⟩i
= (1−αn)||xn−x∗||2+αndn, (5.120)
wheredn :=hθn
αn(1−αn)||xn−xn−1||M2+2βn||un−PCn(un)||||x∗−xn+1||+2⟨x∗, x∗−xn+1⟩i . To show that the sequence n
||xn−x∗||o
converges to zero, by Lemma2.5.55, it is enough to show that lim sup
k→∞
dnk ≤0 (where{dnk}is a subsequence of{dn}), for every subsequence n||xnk−x∗||o
of n
||xn−x∗||o
satisfying lim inf
k→∞
||xnk+1−x∗|| − ||xnk −x∗||
≥0.
Now, suppose that n
||xnk−x∗||o
is a subsequence of n
||xn−x∗||o
such that lim inf
k→∞
||xnk+1−x∗|| − ||xnk −x∗||
≥0, then, lim inf
k→∞
||xnk+1−x∗||2− ||xnk−x∗||2
= lim
k→∞inf
(||xnk+1−x∗|| − ||xnk−x∗||)(||xnk+1−x∗||+||xnk−x∗||)
≥0. (5.121) Now,
||xn+1−x∗||2 =||(1−βn−αn)un+βnPCn(un)−x∗||2
=||(1−βn−αn)(un−x∗) +βn(PCn(un)−x∗)−αnx∗||2
=||(1−βn−αn)(un−x∗) +βn(PCn(un)−x∗)||2
−2αn⟨(1−βn−αn)(un−x∗) +βn(PCn(un)−x∗), x∗⟩+α2n||x∗||2
≤ ||(1−βn−αn)(un−x∗) +βn(PCn(un)−x∗)||2+αnM3, (5.122) for some M3 >0. Substituting (5.111) into (5.122), we have
||xn+1−x∗||2 ≤(1−βn−αn)(1−αn)||un−x∗||2+ (1−αn)βn||PCn(un)−x∗||2+αnM3. (5.123) Using (5.114), (5.117), and (5.123), we have
||xn+1−x∗||2 ≤(1−βn−αn)(1−αn)||un−x∗||2 + (1−αn)βn||un−x∗||2
−(1−αn)βnσηn
2m||un−zn||22
+αnM3
= (1−αn)2||un−x∗||2−(1−αn)βn
σηn
2m||un−zn||22
+αnM3
≤ ||wn−x∗||2−(1−αn)βn σηn
2m||un−zn||2 2
+αnM3
≤ ||xn−x∗||2+θn||xn−xn−1||M2 −(1−αn)βn σηn
2m||un−zn||2 2
+αnM3
=||xn−x∗||2+αn θn
αn∥xn−xn−1∥M2+M3
−(1−αn)βn σηn
2m||un−zn||2 2
.
Hence
(1−αn)βn
σηn
2m||un−zn||2 2
≤ ||xn−x∗||2− ||xn+1−x∗||2 +αn
θn
αn∥xn−xn−1∥M2+M3
. (5.124)
By (5.121) and (5.124), we obtain lim sup
k→∞
(1−αnk)βnk σηnk
2m ||unk −znk||2 2
≤lim sup
k→∞
n
||xnk −x∗||2− ||xnk+1−x∗||2o
+ lim sup
k→∞
αnk
M1M2+M3
= lim sup
k→∞
n||xnk −x∗||2− ||xnk+1−x∗||2o
=−lim inf
k→∞
n||xnk+1−x∗||2− ||xnk −x∗||2o
≤0, which implies that
k→∞lim ηnk||unk −znk||2 = 0. (5.125) Now, from Step 3, we have
ynk =unk −αmn
unk−znk
=unk−ηnk
unk−znk . Thus, by (5.125), we obtain that
k→∞lim ∥unk−ynk∥= 0. (5.126) Also, observe that equation (5.111) and (5.122), gives
∥xn+1−x∗∥2 ≤(1−αn)∥un−x∗∥2+αnM3. Substituting (5.108) into the last inequality above, we obtain
∥xn+1−x∗∥2 ≤(1−αn)h
∥wn−x∗∥2 −βn(1−βn)∥Vnwn−wn∥2i
+αnM3
≤(1−αn)h
∥xn−x∗∥2+θn∥xn−xn−1∥M2i
−βn(1−βn)(1−αn)∥Vnwn−wn∥2 +αnM3.
Hence
βn(1−βn)(1−αn)∥Vnwn−wn∥2 ≤ ∥xn−x∗∥2− ∥xn+1−x∗∥2 +αn
θn αn
∥xn−xn−1∥M2 +M3
. (5.127)
By (5.121) and (5.127), we obtain lim sup
k→∞
h
βnk(1−βnk)∥Vnkwnk −wnk∥2i
≤lim sup
k→∞
h∥xnk −x∗∥2− ∥xnk+1−x∗∥2i + lim sup
k→∞
αnk
M1M2+M3
=−lim inf
k→∞
h
∥xnk+1−x∗∥2− ∥xnk −x∗∥2i
≤0, which implies that
k→∞lim ∥Vnkwnk−wnk∥= 0. (5.128) Thus, we obtain from (5.107) that
k→∞lim ∥unk −wnk∥= lim
k→∞βnk∥Vnkwnk −wnk∥= 0. (5.129) Also, combining (5.117) and (5.123) we have
∥xn+1−x∗∥2 ≤(1−βn−αn)(1−αn)∥un−x∗∥2+ (1−αn)βn∥un−x∗∥2
−(1−αn)βn∥PCnun−un∥2+αnM3
= (1−αn)2∥un−x∗∥2−(1−αn)βn∥PCnun−un∥2+αnM3
≤ ∥xn−x∗∥2+θn∥xn−xn−1∥M2−(1−αn)βn∥PCnun−un∥2+αnM3
≤ ∥xn−x∗∥2+αn θn
αn∥xn−xn−1∥M2+M3
−(1−αn)βn∥PCnun−un∥2, which implies that
(1−αn)βn∥PCnun−un∥2 ≤ ∥xn−x∗∥2− ∥xn+1−x∗∥2+αn
M1M2+M3
. (5.130) Thus, we obtain by (5.121) that
k→∞lim ∥PCnkunk−unk∥= 0. (5.131) Combining this with (5.125), we obtain from Lemma 5.4.9 that
k→∞lim ∥unk −znk∥= 0. (5.132) Furthermore, from Step 2 we obtain
∥wnk −xnk∥=αnk θnk
αnk∥xnk −xnk−1∥
→0 as k→ ∞. (5.133) Using (5.129) and (5.133), we obtain
k→∞lim ∥unk −xnk∥= 0. (5.134)
From (5.98) and (5.131), we have
∥xnk+1−unk∥ ≤βnk∥PCnkunk −unk∥+αnk||unk|| →0.
Thus, we obtain from (5.134) that
k→∞lim ∥xnk+1 −xnk∥= 0.
Since
xnk is bounded, there exists a subsequence
xnkj of
xnk which converges weakly to some ¯x∈H1, such that
lim sup
k→∞
⟨x∗, x∗−xnk⟩= lim
j→∞⟨x∗, x∗−xnj⟩=⟨x∗, x∗ −x⟩.¯ Also, we obtain from (5.129), (5.132), (5.134) and Lemma 5.4.7 that ¯x∈Γ.
From x∗ =PΓ0, we obtain lim sup
k→∞
⟨x∗, x∗−xnk⟩=⟨x∗, x∗−x⟩ ≤¯ 0.
Since limk→∞∥xnk+1 −xnk∥ →0,we obtain lim sup
k→∞
⟨x∗, x∗−xnk+1⟩ ≤0. (5.135) Now, recall that dnk :=hθnk
αnk(1−αnk)∥xnk −xnk−1∥M2+ 2βnk∥unk −PCnk(unk)∥
∥x∗−xnk+1∥+ 2⟨x∗, x∗−xnk+1⟩i . Hence, by (5.135), limn→∞
θn
αn∥xn−xn−1∥= 0, (5.131) and Lemma2.5.55, we obtain
n→∞lim ∥xn−x∗∥= 0.
Therefore, {xn} converges strongly to x∗ =PΓ0.