- Research
- Open access
- Published:
New descent LQP alternating direction methods for solving a class of structured variational inequalities
Fixed Point Theory and Applications volume 2015, Article number: 137 (2015)
Abstract
In this paper, we propose two descent alternating direction methods based on a logarithmic-quadratic proximal method for structured variational inequalities. The first method can be viewed as an extension of the method proposed by Bnouhachem and Xu (Comput. Math. Appl. 67:671-680, 2014) by performing an additional step at each iteration. The second method generates the new iterate by searching the optimal step size along a new descent direction, which can be viewed as a refinement and improvement of the first one. Under certain conditions, the global convergence of the both methods is proved.
1 Introduction
We consider the constrained convex programming problem with the following separate structure:
where \(\theta_{1}:\mathcal{R}_{+}^{n}\rightarrow\mathcal{R}\) and \(\theta_{2}:\mathcal{R}_{+}^{m}\rightarrow\mathcal{R}\) are closed proper convex functions, \(A\in{ \mathcal{R}}^{l\times n}\), \({B \in{ \mathcal{R}}^{l\times m}}\) are given matrices, and \(b\in{ \mathcal{R}}^{l}\) is a given vector.
A large number of problems can be modeled as problem (1.1). In practice, these classes of problems have very large size and, due to their practical importance, they have received a great deal of attention from many researchers. Various methods have been suggested to find the solution of problem (1.1). A popular approach is the alternating direction method (ADM) which was proposed by Gabay and Mercier [2] and Gabay [3]. The ADM can reduce the scale of variational inequalities by decomposing the original problem into a series of subproblems with a lower scale. To make the ADM more efficient and practical, some strategies have been studied; For further details, we refer to [4–13] and the references therein.
Let \(\partial(\cdot)\) denote the sub-gradient operator of a convex function, and \(f(x) \in\partial\theta_{1}(x)\) and \(g(y)\in\partial\theta_{2}(y)\) are the sub-gradient of \(\theta_{1}(x)\) and \(\theta_{2}(y)\), respectively. By attaching a Lagrange multiplier vector \(\lambda\in{ \mathcal{R}}^{l}\) to the linear constraint \(Ax + By =b\), problem (1.1) can be written in terms of finding \(w \in{ \mathcal{W}}\) such that
where
Problem (1.2)-(1.3) is referred to as structured variational inequalities (in short, SVI).
Very recently, Yuan and Li [14] developed the following logarithmic-quadratic proximal (LQP)-based decomposition method by applying the LQP terms to regularize the ADM subproblems: For a given \(w^{k}=(x^{k},y^{k},\lambda^{k})\in\mathcal{R}^{n}_{++}\times \mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\), and \(\mu\in(0,1)\), the new iterative \((x^{k+1}, y^{k+1}, \lambda^{k+1})\) is obtained via solving the following system:
where \(H \in{ \mathcal{R}}^{l \times l}\), \(R \in{ \mathcal{R}}^{n \times n}\), and \(S \in{ \mathcal{R}}^{m \times m}\) are symmetric positive definite.
Note that the LQP method was presented originally in [15]. It seems that it is easier to solve a series of systems of nonlinear equations than to solve a series of sub-variational inequalities in many cases. Later, Bnouhachem et al. [16], Bnouhachem and Ansari [17], and Li [18] proposed some LQP alternating direction methods and made the LQP alternating direction method more practical. Each iteration of the above methods contains a prediction and a correction. The predictor is obtained via solving (1.4)-(1.6) and the new iterate is obtained by a convex combination of the previous point and the one generated by a projection-type method along a descent direction for [17, 18]. In 2014, Bnouhachem and Xu [1] proposed the following LQP alternating direction: For a given \(w^{k}=(x^{k},y^{k},\lambda^{k})\in\mathcal{R}^{n}_{++}\times \mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\), and \(\mu\in(0,1)\), the predictor \(\tilde{w}^{k}=(\tilde{x}^{k},\tilde{y}^{k},\tilde{\lambda }^{k})\in \mathcal{R}^{n}_{++}\times\mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\) is obtained via solving the following system:
where
and
is a positive definite (block diagonal) matrix.
Take the new iterate \(w^{k+1}=(x^{k+1},y^{k+1},\lambda^{k+1})\) as the solution of the following system:
Each iteration of the above method contains a prediction and a correction. The predictor is obtained via solving (1.4)-(1.6) and the new iterate is computed directly by an explicit formula derived from the original LQP method.
Inspired and motivated by the above research, we present two methods to solve SVI. The first one can be viewed as an extension of the method proposed in [1] by performing an additional step at each iteration. The second method can be viewed as an extension of the first one by using a new descent direction which provides a significant refinement and improvement of the first one. We also study the global convergence of the both methods method under certain conditions.
2 Iterative methods and convergence results
In this section, we suggest and analyze two new modified logarithmic-quadratic proximal alternating direction methods for solving structured variational inequalities. The following lemma provides some basic properties of the projection onto Ω.
Lemma 2.1
Let G be a symmetry positive definite matrix and Ω be a nonempty closed convex subset of \(R^{l}\). We denote by \(P_{\Omega,G}(\cdot)\) the projection under the G-norm, that is,
Then
For convenience, we make the following standard assumptions to guarantee that the problem under consideration is solvable and the proposed methods are well defined.
Assumption A
\(f(x)\) is monotone with respect to \(\mathcal{R}^{n}_{++}\) and \(g(y)\) is monotone with respect to \(\mathcal{R}^{m}_{++}\).
Assumption B
The solution set of SVI, denoted by \({\mathcal{W}}^{*}\), is nonempty.
Then the iterative scheme of the first method is given as follows.
Algorithm 2.1
-
Step 0.
The initial step:
Given \(\varepsilon>0\), \(\mu\in(0,1)\) and \(w^{1}=(x^{1}, y^{1}, \lambda^{1}) \in\mathcal{R}^{n}_{++}\times \mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\). Set \(k=1\).
-
Step 1.
Prediction step:
Compute \(\tilde{w}^{k}=(\tilde{x}^{k},\tilde{y}^{k},\tilde{\lambda }^{k})\in\mathcal{R}^{n}_{++}\times \mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\) by solving the system (1.7a)-(1.7c).
-
Step 2.
Convergence verification:
If \(\{\max{ \Vert x^{k}-\tilde{x}^{k}\Vert _{\infty}, \Vert y^{k}-\tilde{y}^{k}\Vert _{\infty}}, \Vert \lambda^{k}-\tilde{\lambda}^{k}\Vert _{\infty}\}<\epsilon\), then stop.
-
Step 3.
Correction step:
Compute \(\bar{w}^{k}=(\bar{x}^{k},\bar{y}^{k},\bar{\lambda}^{k})\) by solving the following system:
$$\begin{aligned}& \frac{1-\mu}{1+\mu}\alpha_{k}\bigl[f\bigl( \tilde{x}^{k}\bigr)-A^{\top}\tilde{\lambda }^{k}\bigr] + R\bigl[\bigl(x-x^{k}\bigr) + \mu\bigl(x^{k} - X_{k}^{2} x^{-1}\bigr)\bigr]= 0, \end{aligned}$$(2.4a)$$\begin{aligned}& \frac{1-\mu}{1+\mu}\alpha_{k}\bigl[g\bigl( \tilde{y}^{k}\bigr)-B^{\top}\tilde{\lambda }^{k}\bigr] + S\bigl[\bigl(y-y^{k}\bigr) + \mu\bigl(y^{k} - Y_{k}^{2} y^{-1}\bigr)\bigr] = 0, \end{aligned}$$(2.4b)$$\begin{aligned}& \lambda^{k+1} = \lambda^{k} - \frac{1-\mu}{1+\mu}\alpha_{k} H \bigl(A \tilde{x}^{k} + B \tilde{y}^{k} - b\bigr), \end{aligned}$$(2.4c)where
$$\begin{aligned}& \alpha_{k}={\frac{\varphi_{k}}{\Vert d_{k}\Vert _{G}^{2}}}, \end{aligned}$$(2.5)$$\begin{aligned}& \varphi_{k}:= \bigl\Vert x^{k} - \tilde{x}^{k}\bigr\Vert _{R}^{2}+ \bigl\Vert y^{k} -\tilde{y}^{k}\bigr\Vert _{S}^{2}+ \bigl\Vert \lambda^{k} -\lambda^{k+1}\bigr\Vert _{H^{-1}}^{2}+\bigl(w^{k}- \tilde{w}^{k} \bigr)^{\top}\xi^{k}, \end{aligned}$$(2.6)and
$$ d_{k}:=w^{k}- \tilde{w}^{k} +G^{-1}\xi^{k}. $$(2.7)The new iterate \(w^{k+1}(\tau_{k})=(x^{k+1},y^{k+1},\lambda^{k+1})\) is given by
$$ w^{k+1}(\tau_{k})=\rho w^{k}+(1- \rho)P_{\mathcal{W}} \bigl[w^{k}-\tau_{k} d'_{k} \bigr], \quad\rho\in(0,1), $$(2.8)where
$$ d'_{k}:=w^{k}-\bar{w}^{k}. $$(2.9)Set \(k:= k + 1\) and go to Step 1.
\(\tau_{k}\) is a positive scalar.
How to choose a suitable step length \(\tau_{k}\) to force convergence will be discussed later.
Theorem 2.1
[1]
For given \(w^{k}=(x^{k}, y^{k},\lambda^{k}) \in\mathcal{R}_{++}^{n} \times \mathcal{R}_{++}^{m}\times{ \mathcal{R}}^{l}\), let \(\bar{w}^{k}=(\bar{x}^{k},\bar{y}^{k},\bar{\lambda}^{k})\) be generated by (2.4a)-(2.4c). Then for any \(w^{*}=(x^{*}, y^{*}, \lambda^{*}) \in{ \mathcal{W}}^{*}\), we have
where
Theorem 2.2
[1]
For given \(w^{k}\in\mathcal{R}_{++}^{n} \times\mathcal{R}_{++}^{m}\times {\mathcal{R}}^{l}\), let \(\tilde{w}^{k}\) be generated by (1.7b)-(1.7c), then we have the following:
and
How should one choose values of \(\tau_{k}\) to ensure that \(w^{k+1}(\tau_{k})\) is closer to the solution set than \(w^{k}\)? For this purpose, we define
Theorem 2.3
Let \(w^{*}=(x^{*},y^{*}, \lambda^{*}) \in{ \mathcal{W}}^{*}\), then we have
Proof
Since \(w^{*}=(x^{*},y^{*}, \lambda^{*}) \in{ \mathcal{W}}^{*}\) and \(w_{*}^{k}(\tau_{k})=P_{\mathcal{W}} [w^{k}-\tau_{k} d'_{k} ]\), it follows from (2.3) that
On the other hand, we have
Using the identity
for \(a=w^{k}-w^{k}_{*}(\tau_{k})\), \(b=w^{k}_{*}(\tau_{k})-w^{*}\), and (2.16), we obtain
Using the definition of \(\Theta_{1}(\tau_{k})\), we get
The identity
implies
Substituting (2.19) in (2.17), we get the assertion of this lemma. □
Using Theorem 2.1 and Theorem 2.3, we get
where
\(\Lambda_{1}(\tau_{k})\) measures the progress obtained in the kth iteration. It is natural to choose a step length \(\tau_{k}\) which maximizes the progress. Note that \(\Lambda_{1}(\tau_{k})\) is a quadratic function of \(\tau_{k}\) and it reaches its maximum at
and
Using (2.13), we have
which implies
Remark 2.1
If \(\rho=0\) and \(\tau^{*}_{k}=1\), Algorithm 2.1 reduces to the method proposed in [1]. Since \(\tau^{*}_{k}\) is to maximize the profit function \(\Lambda_{1}(\tau)\), we have
Inequalities (2.20) and (2.25) show theoretically that Algorithm 2.1 is expected to make more progress than the method proposed in [1] at each iteration, and so it explains theoretically that Algorithm 2.1 outperforms the method proposed in [1].
Inspired and motivated by the first method, we next propose the following new LQP alternating direction method for solving SVI.
Algorithm 2.2
-
Step 0.
The initial step:
Given \(\varepsilon>0\), \(\mu\in(0,1)\), and \(w^{1}=(x^{1}, y^{1}, \lambda^{1}) \in\mathcal{R}^{n}_{++}\times\mathcal {R}^{m}_{++}\times\mathcal{R}^{l}\) and \(d''_{0}=(0,0,0)\). Set \(k=1\).
-
Step 1.
Prediction step:
Compute \(\tilde{w}^{k}=(\tilde{x}^{k},\tilde{y}^{k},\tilde{\lambda }^{k})\in\mathcal{R}^{n}_{++}\times \mathcal{R}^{m}_{++}\times\mathcal{R}^{l}\) by solving the system (1.7a)-(1.7c).
-
Step 2.
Convergence verification:
If \(\{\max{ \Vert x^{k}-\tilde{x}^{k}\Vert _{\infty}, \Vert y^{k}-\tilde{y}^{k}\Vert _{\infty}}, \Vert \lambda^{k}-\tilde{\lambda}^{k}\Vert _{\infty}\}<\epsilon\), then stop.
-
Step 3.
Correction step:
Compute \(\bar{w}^{k}=(\bar{x}^{k},\bar{y}^{k},\bar{\lambda}^{k})\) by solving the system (2.4a)-(2.4c). The new iterate \(w^{k+1}(\beta_{k} )\) is given by
$$ w^{k+1}(\beta_{k})=\rho w^{k}+(1- \rho)P_{\mathcal{W}} \bigl[w^{k}-\beta_{k} d''_{k} \bigr] , \quad\rho\in(0,1), $$(2.26)where
$$\begin{aligned}& d''_{k}:=d'_{k}+ \sigma_{k}{d''_{k-1}}, \\& \sigma_{k}=\operatorname{max} \biggl(0,\frac{-{d^{\prime\top}_{k}}G{d''_{k-1}}}{\Vert {d''_{k-1}}\Vert _{G}^{2}} \biggr) , \end{aligned}$$(2.27)and
$$ \beta_{k} =\frac{\Vert d'_{k}\Vert _{G}^{2}+ (\frac{1-\mu}{1+\mu} ) \Phi(\alpha_{k})}{2\Vert d''_{k}\Vert _{G}^{2}}. $$(2.28)Set \(k:= k + 1\) and go to Step 1.
The following lemma has motivated us to choose the new descent direction as \(d''_{k}\).
Lemma 2.2
For any \(k\geq1\), we have
Proof
From (2.18) and (2.10) it is easy to show that
Note that (2.29) is trivially true for \(k=1\) since \(d''_{0}=(0,0,0)\). By induction, consider any \(k\geq2\) and assume that
Using the definition of \(d''_{k-1}\), we have
where the second inequality follows from (2.30). Hence, the lemma is proved. □
Remark 2.2
From (2.30) and Lemma 2.2, we get
Then \(-d''_{k} \) is a descent direction of \(\frac{1}{2}\Vert w^{k}-w^{*}\Vert ^{2}\) at the point \(w^{k}\). Since \(-d''_{k} \) is a descent direction of the distance function at \(w^{k}\), along \(-d''_{k}\), one can find a new iterate which is closer to the solution set. Due to this fact, we construct (2.26).
Lemma 2.3
Let \(w^{*}=(x^{*},y^{*}, \lambda^{*}) \in{ \mathcal{W}}^{*}\), then we have
where
Proof
Similar to (2.17), we have
Using (2.19) and the definition of \(\Theta_{2}(\beta_{k})\), we get the assertion of this lemma. □
Using Theorem 2.1 and Lemma 2.3, we get
where
Remark 2.3
From (2.27), it is easy to prove that
which implies that
From (2.20), (2.22), (2.33), and (2.34), we have
and
From the above inequalities, we obtain
Inequality (2.35) shows theoretically that Algorithm 2.2 can be expected to make more progress than Algorithm 2.1 at each iteration, and so it explains theoretically that Algorithm 2.2 outperforms Algorithm 2.1.
Now, we mainly focus on investigating the convergence of the proposed methods. The following theorem plays a crucial role in the convergence of the proposed methods.
Theorem 2.4
Let \(w^{k+1}\) be the new iterate generated by
Then, for any \(w^{*} \in{ \mathcal{W}}^{*}\), \(w^{k}\) and \(\tilde{w}^{k}\) are bounded, and
where
Proof
Inequality (2.36) follows from (2.20), (2.24), (2.33), and (2.35) immediately. It follows from (2.36) that
and thus, \(\{ w^{k}\}\) is a bounded sequence.
It follows from (2.36) that
which means that
Since \(\{ w^{k}\}\) is a bounded sequence, we conclude that \(\{\tilde {w}^{k}\}\) is also bounded. □
The convergence of the proposed method can be proved by using similar arguments to [1]. Hence the proof is omitted.
Theorem 2.5
[1]
The sequence \(\{w^{k}\}\) generated by the proposed methods converges to some \(w^{\infty}\) which is a solution of SVI.
3 Conclusions
In this paper, we proposed two new modified logarithmic-quadratic proximal alternating direction methods for solving structured variational inequalities. The first one can be viewed as an extension of the method proposed in [1] by performing an additional step at each iteration and another optimal step length is employed to reach substantial progress in each iteration while the second method generates the new iterate by searching the optimal step size along a new descent direction. It is proved theoretically that the lower-bound of the progress obtained by the second method is greater than that by the first one. Global convergence of the proposed method is proved under mild assumptions.
References
Bnouhachem, A, Xu, MH: An inexact LQP alternating direction method for solving a class of structured variational inequalities. Comput. Appl. Math. 67, 671-680 (2014)
Gabay, D, Mercier, B: A dual algorithm for the solution of nonlinear variational problems via finite-element approximations. Comput. Appl. Math. 2, 17-40 (1976)
Gabay, D: Applications of the method of multipliers to variational inequalities. In: Fortin, M, Glowinski, R (eds.) Augmented Lagrange Methods: Applications to the Solution of Boundary-Valued Problems, pp. 299-331. North-Holland, Amsterdam (1983)
Chen, G, Teboulle, M: A proximal-based decomposition method for convex minimization problems. Math. Program. 64, 81-101 (1994)
Eckstein, J: Some saddle-function splitting methods for convex programming. Optim. Methods Softw. 4, 75-83 (1994)
Kontogiorgis, S, Meyer, RR: A variable-penalty alternating directions method for convex optimization. Math. Program. 83, 29-53 (1998)
He, BS, Yang, H, Wang, SL: Alternating directions method with self-adaptive penalty parameters for monotone variational inequalities. J. Optim. Theory Appl. 106, 349-368 (2000)
He, BS, Zhou, J: A modified alternating direction method for convex minimization problems. Appl. Math. Lett. 13, 123-130 (2000)
He, BS, Liao, LZ, Han, DR, Yang, H: A new inexact alternating directions method for monotone variational inequalities. Math. Program. 92, 103-118 (2002)
Jiang, ZK, Bnouhachem, A: A projection-based prediction-correction method for structured monotone variational inequalities. Appl. Math. Comput. 202, 747-759 (2008)
Yang, J, Zang, Y: Alternating direction algorithms for \(l_{1}\) problems in compressive sensing. SIAM J. Sci. Comput. 33(1), 250-278 (2011)
Tao, M, Yuan, XM: On the \(O(1/t)\) convergence rate of alternating direction method with logarithmic-quadratic proximal regularization. SIAM J. Optim. 22(4), 1431-1448 (2012)
Wang, K, Xu, LL, Han, DR: A new parallel splitting descent method for structured variational inequalities. J. Ind. Manag. Optim. 10(2), 461-476 (2014)
Yuan, XM, Li, M: An LQP-based decomposition method for solving a class of variational inequalities. SIAM J. Optim. 21(4), 1309-1318 (2011)
Auslender, A, Teboulle, M, Ben-Tiba, S: A logarithmic-quadratic proximal method for variational inequalities. Comput. Optim. Appl. 12, 31-40 (1999)
Bnouhachem, A, Benazza, H, Khalfaoui, M: An inexact alternating direction method for solving a class of structured variational inequalities. Appl. Math. Comput. 219, 7837-7846 (2013)
Bnouhachem, A, Ansari, QH: A descent LQP alternating direction method for solving variational inequality problems with separable structure. Appl. Math. Comput. 246, 519-532 (2014)
Li, M: A hybrid LQP-based method for structured variational inequalities. Int. J. Comput. Math. 89(10), 1412-1425 (2012)
Acknowledgements
The authors are grateful to King Fahd University of Petroleum & Minerals for providing excellent research facilities.
Author information
Authors and Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
All authors contributed equally and significantly in writing this article. All authors read and approved the final manuscript.
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Bnouhachem, A., Al-Homidan, S. & Ansari, Q.H. New descent LQP alternating direction methods for solving a class of structured variational inequalities. Fixed Point Theory Appl 2015, 137 (2015). https://doi.org/10.1186/s13663-015-0387-1
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s13663-015-0387-1