In this paper, a two-step inertial Tseng extragradient method involving self-adaptive and Armijo-like step sizes is introduced for solving variational inequalities with a quasimonotone cost function in the setting of a real Hilbert space. Weak convergence of the sequence generated by the proposed algorithm is proved without assuming the Lipschitz condition. An interesting feature of the proposed algorithm is its ability to select the better step size between the self-adaptive and Armijo-like options at each iteration step. Moreover, removing the requirement for the Lipschitz condition on the cost function broadens the applicability of the proposed method. Finally, the algorithm accelerates and complements several existing iterative algorithms for solving variational inequalities in Hilbert spaces.
It is well-known that one of the drawbacks of the EGM arises from its requirement to perform two projections onto the closed convex set during each iteration. This requirement can be computationally expensive especially when the structure of C is not simple. Another drawback of this method is the fact that parameter \lambda depends on the explicit value of the Lipschitz constant L which is can be challenging to obtain.
It is important to note that all the improvements of the EGM and their accelerated versions mentioned above have not fully addressed Remark 1.1. They are yet to dispense with the dependency of the step-size on the explicit value of the Lipschitz constant.
Let H be a real Hilbert space. An operator A: H \rightarrow H is said to be:
(i) Lipschitz continuous on H, if there exist a constant L > 0 such that
(ii) uniformly continuous, if for every \epsilon > 0, there exists \delta = \delta(\epsilon)> 0, such that
(iii) sequentially weakly-strongly continuous, if for each sequence {x_k}, we have x_k\rightharpoonup x \in H implies that Ax_k \rightarrow Ax \in H;
(iv) sequentially weakly continuous, if for each sequence {x_n},we have x_n\rightharpoonup x \in H implies that Ax_k \rightarrow Ax \in H ;
(v) monotone, if \langle Ax-Ay,x-y \rangle \geq 0, \forall x,y \in H ;
(vi) \alpha-strongly pseudomonotone, if there exists \alpha > 0 such that
(vii) pseudomonotone, if
(viii) quasimonotone, if
Let x,y,z\in H and a,b\in\mathbb{R}. Then
Let C be a nonempty closed subset of a real Hilbert space H. \forall v\in H and \forall z\in C, we have
The following identities hold for all u,v\in H :
([lema60]) Suppose either
(a) A is pseudomonotone on C and S\neq\emptyset;
(b) A is the gradient of G, where G is a differential quasiconvex function on an open set K\supset C and attains its global minimum on C;
(c) A is quasimonotone on C, A\neq 0 on C and C is bounded;
(d) A is quasimonotone on C, A\neq 0 on C and there exists a positive number r such that, for every x\in C with ||x|| \geq r, there exists y \in C such that ||y|| \leq r and \langle Ax,y-x\rangle \leq 0;
(e) A is quasimonotone on C, int C \neq \emptyset and there exists x^* \in S such that Ax^*\neq 0,
([lema16]) Let H be a Hilbert space and A : H \rightarrow H be a uniformly continuous operator. Suppose x\in H and \psi \geq \sigma > 0. The following inequality holds:
(i) S_{D}\neq \emptyset;
(ii) A is Lipschitz continuous on H;
(iii) A satisfies the following condition: whenever {x_{k}}\subset C and x_{k}\rightharpoonup v^{*}, one has \left\| {A{v^*}} \right\| \le \lim \mathop {\inf }\limits_{n \to \infty } \left\| {A{x_k}} \right\|;
(iv) A is quasimonotone on H;
(v) The set \{v\in C: Av = 0\} \backslash S_{D} is a finite set.
Assume that \alpha, \beta and \mu meet the following conditions:
(a) 0\leq \alpha \leq \frac{1-\mu}{3+\mu};
(b) max\{2\alpha (\frac{1-\mu}{3+\mu})-(1-\alpha),\frac{1}{2}[\alpha(1+\mu)-(\frac{(1-\mu)(1-\alpha)^2}{1+\alpha})]\}<\beta\leq0;
(c) 2\alpha^2\mu-(1-3\alpha)+\mu(1-\alpha)-\beta(4\alpha+3-\mu)+2\mu\beta^2<0.
Under the (i)–(iv) of Assumption 1, the sequence {\lambda_{k}} generated Algorithm ? satisfying
Firstly, it is obvious that {\lambda^{(1)}_{k}} is a monotonically decreasing sequence. Since A is a Lipschitz continuous mapping with constant L>0, in the case of Aw_k-Ay_k\neq0, we have
Suppose Assumption 3.1 (i) and (ii) and Assumption 3.2 hold. Then, the sequence \{x_{k}\} generated by Algorithm ? is bounded.
Let q\in S_{D}. First of all, we estimate ||x_{k+1}-q||^2. It can be obtained from the definition of x_{k+1} that:
From the definition of y_{k}, and the fact that S_{D}\subset C,
it can be obtained by Lemma 2.3 that:
that is
Appling (?) to (Proof), we will have
Since q\in S_{D}, \lambda_{k}>0 and y_{k}\in C, we have that
Therefore, from (?) we can have
From the definition of \lambda_{k}, if {Aw_{k}-Ay_{k}}\neq0, then we have
which means:
Also, we have \lambda_{k}^{(2)}\left\|A w_{k}-A y_{k}\right\| \leq \mu\left\|w_{k}-y_{k}\right\|. Therefore,
Substituting (?) into (?), we have
Observe that
and
Hence, using (?) into (?), we obtain that
Applying (?) into (?), we have
Also,
Using Lemma 2.2, we get
Furthermore, it can be obtained that
Applying (?) and (?) with (?) and note that \beta <0 we have
By rearranging the inequality above we obtain that:
Now, let
Then we can rewrite (?) as
Claim: \Gamma_{k}\geq0, \forall k>1. Since
From Assumption 2, we have that \alpha<\frac{1}{2},\beta\leq0,\frac{\alpha(3+\mu)}{1-\mu}-1\leq\beta and 0<\alpha<\frac{1-\mu}{3+\mu}, it can be conclude that \Gamma_{k}\geq 0, \forall k>1. Let
Thus, we deduce from (?) that
From Assumption 2(b),
which implies that
which can conclude that k_{1}>0.
By Assumption 2(c) we can have k_{2}>0, thus, (?) can be rewritten as
Letting \Gamma_k^{\prime}=\Gamma_k+k_1\|x_{k-1}-x_{k-2}\|^2. Then, \Gamma_k^{^{\prime}}\geq0, \forall k\geq1. Therefore, we deduce from (?) that
which implies that the sequence {\Gamma_{k+1}^{'}} is decreasing and bounded from below and thus \lim\limits_{n\rightarrow\infty}\Gamma_{k}^{'} exists.
Hence, by rearranging (?) and letting k approach infinity, we have
It is easy to see that
Consequently, by (?) we have
Furthermore,
Observe that
Using (? )we deduce from (?) that
Also,
Due to the existence of the limit \Gamma_{k}^{'} and (?), we can obtain that the limit of \Gamma_{k} also exists and therefore, the sequence \{\Gamma_{k}\} is bounded. Since \lim\limits_{k\rightarrow\infty}||x_{k+1}-x_{k}||=0, from the definition of \Gamma_{k} we can get that
exists. Due to the boundedness of \{\Gamma_{k}\}, we have from (?) that \{x_{k}\} is bounded and \{y_{k}\} and \{w_{k}\} are bounded.
Let \{x_{k}\} be the sequence generated by algorithm ?, satisfying Assumptions 1(i)-(iv) and Assumptions 2(a)-(c), and assuming that \lim \limits_{k\rightarrow\infty}w_{k}-y_{k} = 0. If v^*is one of the weakly clustered points of {y_k}, then we have at least one of the following: v^*\in S_D or Av^*= 0.
From the lemma (4.2) above, it can be concluded that \{y_{k}\} is bounded. Therefore let v^* be a weak cluster point of \{y_{k}\}. Hence, we denote \{y_{k_{j}}\} as a subsequence of \{y_{k}\} such that y_{k_{j}}\rightharpoonup v^*\in C.
Now we discuss in two cases.
First, we assume that \lim \limits_{k\rightarrow\infty}||Ay_{k_{j}}||=0. Consequently, \lim \limits_{k\rightarrow\infty}||Ay_{k_{j}}||=\liminf \limits_{k\rightarrow\infty}||Ay_{k_{j}}||=0
Through the assumption above that y_{k_{j}}\rightharpoonup v^*\in C and A satisfies Assumption 1(iii), that is
Which just implies that Av^*=0.
Now we consider the another situation.
If \limsup\limits_{k\to\infty}\|Ay_{k_j}\|>0. Without loss of generality, we take \lim\limits_{k\to\infty}\|Ay_{k_j}\|=M_{1}>0. It then follows that there exists a \mathbf{K}\in\mathbb{K} such that \|Ay_{k_j}\|>\frac{M_{1}}{2} for all k\geq \mathbf{K} Since y_{k_{j}}=P_C(w_{k_{j}}-\lambda_{k_{j}}Aw_{k_{j}}), we have
For the weak convergence of w_{k_{j}}, w_{k_{j}} is bounded. Then, for the A Lipschitz continuous, Aw_{k_{j}} is bounded. By (?) we can get that
therefore \|y_{k_{j}}\| is also bounded and through the Lemma(4.1) we get that \lambda_{k_{j}}\geq min\{\frac{l\mu}{L},\lambda_{0}\}. Passing (?) to the limit as k\rightarrow\infty, that is
Observe that
Thus we can get \lim \limits_{k\rightarrow\infty}||Aw_{k_{j}}-Ay_{k_{j}}||=0, for the \lim \limits_{k\rightarrow\infty}||w_{k_{j}}-y_{k_{j}}||=0 and the L -Lipschitz continuity on H of A. Together with (?) and (?) which implies that
If we suppose that \limsup\limits_{k\rightarrow\infty}\langle Ay_{k_{j}},x-y_{k_{j}}\rangle>0, then there exists a subsequence denoted by y_{n_{k_{j}}} such that \lim\limits_{j\rightarrow\infty}\langle Ay_{n_{k_{j}}},x-y_{n_{k_{j}}} \rangle > 0, that is there exists j_{0}\in \mathbb{K} such that
Through the quasimonotonicity(2.1) of A, we get \langle Ax,x-y_{n_{k_{j}}}\rangle\geq0, as j\rightarrow\infty, we conclude that v^*\in S_{D} On the other hand we suppose that if \limsup\limits_{k\rightarrow\infty}\langle Ay_{k_{j}},x-y_{k_{j}}\rangle=0, plus (?) implies that
Let \epsilon_{k}:= |Ay_{k_{j}},x-y_{k_{j}}|+\frac{1}{k+1}. Thus we obtain that
Furthermore, for each k\geq \mathbf{K} we can get Ay_{k_{j}}\neq0. Defining that
then \langle Ay_{k_{j}},r_{k_{j}}\rangle=1 for each k\geq \mathbf{K}. Thus we can conclude from (?) that,
Since A is quasimonotone on H, we get
Thus
Observe that, tending k\rightarrow\infty, for \{x+\epsilon_{k}r_{k_{j}}-y_{k_{j}}\} is bounded and \lim\limits_{k\rightarrow\infty}\epsilon_{k}=0, we can conclude that \langle Ax,x-v^*\rangle\geq0,~\forall x\in C. This implies that v^*\in S_D.
Suppose \{x_{k}\} is a sequence generated by Algorithm ?. Then under Assumptions 1 and Assumptions 2 and Ax\neq 0, \forall x \in C. Then {x_{k}} converges weakly to an element of S_{D} \subset S.
Suppose w_{w}(x_{k}) is a set of weak cluster points of \{x_{k}\}. Which is
We take v^*\in w_{w}(x_{k}).Therefore, there exists a subsequence \{x_{n_{k}}\}\subset\{x_{k}\} such that \{x_{n_{k}}\}\rightarrow v^*, k \rightarrow\infty. Since C is weakly closed, we have that v^*\in C. Furthermore, we can conclude that Av^*\neq0 for Ax\neq0, ~\forall x\in C. By Lemma (4.3), we have v^*\in S_{D}. Hence w_{w}(x_{k})\subset S_{D}. By Lemma (4.2), \lim\limits_{n\rightarrow\infty}\Gamma_{k} exists and \lim\limits_{n\rightarrow\infty}||x_{k+1}-x_{k}||=0, we have
exists for \forall q \in S_{D}.
Next we show that x_{k}\rightharpoonup x^*\in S_{D}. Let \{x_{n_{j}}\} and \{x_{n_{m}}\} both \subset \{x_{k}\} such that x_{n_{j}}\rightharpoonup v^*, j\rightarrow\infty and x_{n_{m}}\rightharpoonup x^*, m\rightarrow\infty.
Then we show that x^*=v^*, using (2.4) observe that
and
Therefore,
and
Addition of (?), (?) and (?) gives
Through (?) we know that
exists, and
also exists, which implies with (?) that \lim\limits_{n\to\infty}\langle x_{k}-\alpha x_{k-1}-\beta x_{k-2},x^*-v^* \rangle exists.
Consequently,
Thus,
For \beta\leq0<1-\alpha, 1-\alpha-\beta \neq0, we can conclude that x^*=v^*.
Hence, we deduce that \{x_{k}\} converges weakly to a point in S_{D}. This completes the proof.
The text on this page is from arXiv:2601.10370v1 by Jian-Wen Peng, Jun-Jie Luo and Abubakar Adamu, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.
The record of each item is in analysis/objects/2601/2601.10370.jsonl.