We consider a class of linearly constrained separable convex programming problems whose objective functions are the sum of three convex functions without coupled variables. For those problems, Han and Yuan (2012) have shown that the sequence generated by the alternating direction method of multipliers (ADMM) with three blocks converges globally to their KKT points under some technical conditions. In this paper, a new proof of this result is found under new conditions which are much weaker than Han and Yuan’s assumptions. Moreover, in order to accelerate the ADMM with three blocks, we also propose a relaxed ADMM involving an additional computation of optimal step size and establish its global convergence under mild conditions.
1. Introduction
In various fields of applied mathematics and engineering, many problems can be equivalently formulated as a separable convex optimization problem with two blocks; that is, given two closed convex functions fi:ℜni→ℜ∪{+∞},i=1,2, to find a solution pair (x1*,x2*) of the following problem:
(1)minf1(x1)+f2(x2)s.t.A1x1+A2x2=b,
where Ai is a matrix in ℜp×ni,i=1,2, and b is a vector in ℜp. The classical alternating direction method of multipliers (ADMM) [1, 2] applied to problem (1) yields the following scheme:
(2)x1k+1=argminx1∈ℜn1f1(x1)-〈A1Tλk,x1〉+β2∥A1x1+A2x2k-b∥2,x2k+1=argminx2∈ℜn2f2(x2)-〈A2Tλk,x2〉+β2∥A1x1k+1+A2x2-b∥2,λk+1=λk-β(A1x1k+1+A2x2k+1-b),
where λk is a Lagrangian multiplier and β>0 is a penalty parameter. Possibly due to its simplicity and effectiveness, the ADMM with two blocks has received continuous attention both in theoretical and application domains. We refer to [3–8] for theoretical results on ADMM with two blocks and [9–13] for its efficient applications in high-dimensional statistics, compressive sensing, finance, image processing, and engineering, to name just a few.
In this paper, we concentrate on the linearly constrained convex programming problem with three blocks:
(3)minf1(x1)+f2(x2)+f3(x3)s.t.A1x1+A2x2+A3x3=b,
where f3:ℜn3→ℜ∪{+∞} is a closed convex function and A3 is a matrix in ℜp×n3. For solving (3), a nature idea is to extend the ADMM with two blocks to the ADMM with three blocks in which the next iteration (x2k+1,x3k+1,λk+1) is updated by
(4)(x2k+1,x3k+1,λk+1):=(x~2k,x~3k,λ~k),
where
(5)x~1k=argminx1∈ℜn1f1(x1)-〈A1Tλk,x1〉+β2∥A1x1+A2x2k+A3x3k-b∥2,x~2k=argminx2∈ℜn2f2(x2)-〈A2Tλk,x2〉+β2∥A1x~1k+A2x2+A3xk-b∥2,x~3k=argminx3∈ℜn3f3(x3)-〈A3Tλk,x3〉+β2∥A1x~1k+A2x~2k+A3x3-b∥2,λ~k=λk-β(A1x~1k+A2x~2k+A3x~3k-b).
Similar to the ADMM with two blocks, the ADMM with three blocks has found numerous applications in a broad spectrum of areas, such as doubly nonnegative cone programming [14], high-dimensional statistics [15, 16], imaging science [17], and engineering [18]. Even though its numerical efficiency is clearly seen from those applications, the theoretical treatment of ADMM with three blocks is challenging and the convergence of the ADMM is still open given only the convex assumptions of the objective function. To alleviate this difficulty, the authors of [19, 20] proposed prediction-correction type methods to solve the general separable convex programming; however, numerical results show that the direct ADMM outperforms its variants substantially. Therefore, it is of great significance to investigate the theoretical performance of the ADMM with three blocks even only to provide sufficient conditions to guarantee the convergence. To the best of our knowledge, there exist only two works aiming to attack the convergence problem of the direct ADMM with three blocks. By using an error bound analysis method, Hong and Luo [21] proved the linear convergence of the ADMM with m blocks for sufficiently small β subject to some technical conditions. However, the sufficiently small requirement on β makes the algorithm difficult to implement. In [22], Han and Yuan employed a contractive analysis method to establish the convergence of ADMM under the strongly convex assumptions of fi and the parameter β less than a threshold depending on all the strongly convex moduli. In this paper, we firstly prove the convergence of ADMM with three blocks under two conditions weaker than those of [22]. In our conditions, the threshold on the parameter β only relies on the strongly convex moduli of f2 and f3, and furthermore f1 is not necessarily strongly convex in one of our conditions. Also, the restricted range of β in this paper is shown to be at least three times as big as that of [22].
In order to accelerate the ADMM with three blocks, we also propose a relaxed ADMM with three blocks which involves an additional computation of optimal step size. Specifically, with the triple (x2k,x3k,λk), we first generate a predictor (x~2k,x~3k,λ~k) according to (5) and then obtain (x2k+1,x3k+1,λk+1) in the next iteration by
(6)x2k+1=x2k-γαk*(x2k-x~2k),x3k+1=x3k-γαk*(x3k-x~3k),λk+1=λk-γαk*(λk-λ~k),
where γ∈(0,2) and αk* is special step size defined in (43). The convergence of the relaxed ADMM is also established under mild conditions. We should mention that it is possible to modify the analyses given in this paper to be problems with more than three blocks of separability. But this is not the focus of this paper.
The remaining parts of this paper are organized as follows. In Section 2, we list some preliminaries on the strongly convex function, subdifferential, and the ADMM with three blocks. In Section 3, we first show the contractive property of the distance between the sequence generated by ADMM with three blocks and the solution set and then prove the convergence of ADMM under certain conditions. In Section 4, we extend the direct ADMM with three blocks to the relaxed ADMM with an optimal step size and establish its convergence under suitable conditions. We conclude our paper in Section 5.
Notation. For any positive integer m, let Im be the m×m identity matrix. We use ∥·∥ and ∥·∥2 to denote the vector Euclidean norm and the spectral norm of matrices (defined as the maximum singular value of matrices). For any symmetric matrix S∈ℜn×n, we write ∥x∥S2=xTSx for any x∈ℜn. G and M are two (n2+n3+p)×(n2+n3+p) matrices defined by
(7)G:=(βA2TA2000βA3TA3000Iβ),M:=(2βA2TA2000βA3TA3000Iβ),
respectively. For given x1∈ℜn1,x2∈ℜn2,x3∈ℜn3, and λ∈ℜp, we frequently use u and v to denote the joint vectors of x2,x3,λ and x1,x2,x3,λ, respectively; that is,
(8)u=[x2T,x3T,λT]T,v=[x1T,x2T,x3T,λT]T,
while u~ and v~ are the joint vectors corresponding to x~2,x~3,λ~ and x~1,x~2,x~3,λ~.
2. Preliminaries
Throughout this paper, we assume fi,i=1,2,3, are strongly convex functions with modulus μi≥0; that is
(9)fi((1-α)z+αz′)≤(1-α)fi(z)+αfi(z′)-12μiα(1-α)∥z-z′∥2,∀z,z′∈ℜni,
for each i. Note that fi is a strongly convex function with modulus 0 being equivalent to the convexity of fi. Let x be a point of dom(fi); the subdifferential of fi at x is defined by
(10)∂fi(x):={x*∣f(z)≥f(x)+〈x*,z-x〉,∀z}.
From Proposition 6 in [23], we know that, for each i, ∂fi is strongly monotone with modulus μi which means
(11)〈z1-z2,x1-x2〉≥μi∥z1-z2∥2≥0,∀x1,x2,z1∈∂fi(x1),z2∈∂fi(x2).
The next lemma introduced in [22] plays a key role in the convergence analysis of the ADMM and the relaxed ADMM with three blocks.
Lemma 1.
Let (x1*,x2*,x3*,λ*) be any KKT point of problem (3). Let v~k be generated by (5) from given uk. Then, one has
(12)〈u~k-u*,G(uk-u~k)〉≥∑i=13μi∥x~ik-xi*∥2+〈λk-λ~k,∑i=23Ai(xik-x~ik)〉+β〈A3(x~3k-x3*),A2(x~2k-x2k)〉.
3. The ADMM with Three Blocks
In this section, we first investigate the contractive property of the distance between the sequence generated by ADMM with three blocks and the solution set under the condition that 0<β≤min{μ2/∥A2∥22,μ3/∥A3∥22}.
Lemma 2.
Let v*=(x1*,x2*,x3*,λ*) be a KKT point of problem (3) and let the sequence {vk=(x1k,x2k,x3k,λk)} be generated by the ADMM with three blocks. Then, it holds that
(13)∥uk+1-u*∥M2≤∥uk-u*∥M2-β∥A3(x3k+1-x3k)∥2-β∥A1x1k+1+A2x2k+A3x3k+1-b∥2-2μ1∥x1k+1-x1*∥2-2∥x2k+1-x2*∥μ2In2-βA2TA22-2∥x3k+1-x3*∥μ3In3-βA3TA32.
Proof.
Since x3j minimizes f3(·)-〈A3Tλj,·〉, we deduce from the first order optimality condition that
(14)A3Tλj∈∂f3(x3j),j=0,1,…,k.
By (14) and the monotonicity of ∂f3(·) (11), it is easily seen that
(15)〈x3k-x3k+1,A3Tλk-A3Tλk+1〉≥0.
Then for each k,
(16)〈uk+1-u*,G(uk-uk+1)〉≥∑i=13μi∥xik+1-xi*∥2+〈λk-λk+1,A2(xik-x2k+1)〉+β〈A3(x3k+1-x3*),A2(x2k+1-x2k)〉≥∑i=12μi∥xik+1-xi*∥2+∥x3k+1-x3*∥μ3In3-βA3TA32+〈λk-λk+1,A2(x2k-x2k+1)〉-β4∥A2(x2k+1-x2k)∥2,
where the last “≥” follows from the elementary inequality
(17)〈x,y〉≥-∥x∥2-14∥y∥2.
Since
(18)∥A3(x3k+1-x3k)∥2≤2∥A3(x3k+1-x3*)∥2+2∥A3(x3k-x3*)∥2,
by direct computations, we further obtain that
(19)∥uk-u*∥G2≥∥uk+1-u*∥G2+∥uk+1-uk∥G2+2μ1∥x1k+1-x1*∥2+2∥x2k+1-x2*∥μ2In2-(β/2)A2TA22+2∥x3k+1-x3*∥μ3In3-βA3TA32+2〈λk-λk+1,A2(x2k-x2k+1)〉-β∥A2(x2k-x2*)∥2,
which, together with G=M-(βA2TA200), implies
(20)∥uk-u*∥M2≥∥uk+1-uk∥G2+∥uk+1-u*∥M2+2μ1∥x1k+1-x1*∥2+2∥x2k+1-x2*∥μ2In2-βA2TA22+2∥x3k+1-x3*∥μ3In3-βA3TA32+2〈λk-λk+1,A2(x2k-x2k+1)〉.
Note that
(21)∥x2k-x2k+1∥βA2TA22+2〈λk-λk+1,A2(x2k-x2k+1)〉+1β∥λk-λk+1∥2=β∥A1x1k+1+A2x2k+A3x3k+1-b∥2.
We complete the proof of this lemma.
With the above preparation, we are ready to prove the convergence of the ADMM with three blocks for solving (3) given the following conditions.
Theorem 3.
Let {vk=(x1k,x2k,x3k,λk)} be the sequence generated by the ADMM with three blocks. Then {vk} converges to a KKT point of problem (3) if either of the following conditions holds:
μ1>0 and 0<β≤min{μ2/∥A2∥22,μ3/∥A3∥22};
A1 is of full column rank, 0<β<μ2∥A2∥22, and β≤μ3∥A3∥22.
Proof.
By the inequality (13), it follows that the sequence {A2x2k,A3x3k,λk} is bounded. Recall that
(22)A1x1k+1=λk-λk+1β-A2x2k+1-A3x3k+1+b.
Hence {A1x1k} is also bounded. Moreover, from (13) we see immediately that
(23)+∞>∑k=1∞β∥A3(x3k+1-x3k)∥2+β∥A1x1k+1+A2x2k+A3x3k+1-b∥2+∑k=1∞2μ1∥x1k+1-x1*∥2+2∥x2k+1-x2*∥μ2In2-βA2TA22+2∥x3k+1-x3*∥μ3In3-βA3TA32.
According to the condition that 0<β≤min{μ2/∥A2∥22,μ3/∥A3∥22}, we know
(24)∑k=1∞∥A3(x3k+1-x3k)∥2<∞,∑k=1∞∥A1x1k+1+A2x2k+A3x3k+1-b∥2<+∞,∑k=1∞μ1∥x1k+1-x1*∥2<+∞,∑k=1∞∥x2k+1-x2*∥μ2In2-βA2TA22<+∞,∑k=1∞∥x3k+1-x3*∥μ3In3-βA3TA32<+∞.
It therefore holds that
(25)limk→∞∥A3(x3k+1-x3k)∥2=0,limk→∞∥A1x1k+1+A2x2k+A3x3k+1-b∥2=0,(26)limk→∞μ1∥x1k+1-x1*∥2=0,limk→∞∥x2k+1-x2*∥μ2In2-βA2TA22=0,limk→∞∥x3k+1-x3*∥μ3In3-βA3TA32=0.
Therefore, the sequence {μ1∥x1k∥2, ∥x2k∥μ2In2-βA2TA22, ∥x3k∥μ3In3-βA2TA22} is bounded, which, together with the boundedness of {A1x1k,A2x2k,A3x3k,λk}, implies that {x2k,x3k,λk} is bounded, and {x1k} is bounded given the condition μ1>0 or A1 is of full column rank. Moreover, since
(27)∥x3k+1-x3k∥2=∥A3x3k+1-A3x3k∥2+∥x3k+1-x3k∥μ3In3-A3TA32,
by the first equality in (25) and the third equality in (26), it holds that
(28)limk→∞∥x3k+1-x3k∥=0.
We proceed to prove the convergence of ADMM by considering the following two cases.
Case 1 (μ1>0 and β≤min(μ2/∥A2∥22,μ3/∥A3∥22)). In this case, the sequence {x1k} converges to x1* and then
(29)limk→∞∥A2x2k+1-A2x2k∥=0,limk→∞∥λk+1-λk∥=0.
By the second equality in (26), we deduce from (29) that
(30)limk→∞∥x2k+1-x2k∥=0.
Since {x2k,x3k,λk} is bounded, there exist a triple (x2∞,x3∞,λ∞) and a subsequence {nk} such that
(31)limk→∞x2nk=x2∞,limk→∞x2nk=x2∞,limk→∞λnk=λ∞,which by combining (25), (29) with given conditions, implies
(32)limk→∞x2nk+1=x2∞,limk→∞x2nk+1=x2∞,limk→∞λnk+1=λ∞.
Note that
(33)0∈∂f1(x1k+1)-A1Tλk+1+A1TA2(x2k-x2k+1)+A1TA3(x3k-x3k+1),0∈∂f2(x2k+1)-A2Tλk+1+A2TA3(x3k-x3k+1),0∈∂f3(x3k+1)-A3Tλk+1,λk+1=λk-β(A1x1k+1+A2x2k+1+A3x3k+1).
Then, by taking the limits on both sides of (33), using (25) and (29), and invoking the upper semicontinuous of ∂f1(·), ∂f2(·), and ∂f3(·) [24], one can immediately write
(34)0∈∂f1(x*)-A1Tλ∞,0∈∂f2(x2∞)-A2Tλ∞,0∈∂f3(x3∞)-A3Tλ∞,A1x*+A2x2∞+A3x3∞=b,
which indicates (x1*,x2∞,x3∞,λ∞) is a KKT point of problem (3). Hence, the inequality (13) is also valid if (x1*,x2*,x3*,λ*) is replaced by (x1*,x2∞,x3∞,λ∞). Then it holds that
(35)2β∥A2x2k+1-A2x2∞∥2+β∥A3x3k+1-A3x3∞∥2+1β∥λk+1-λ∞∥2≤2β∥A2x2k-A2x2∞∥2+β∥A3x3k-A3x3∞∥2+1β∥λk-λ∞∥2,
which yields
(36)limk→∞∥x2k-x2∞∥A2TA22=0,limk→∞∥x3k-x3∞∥A3TA32=0,(37)limk→∞λk=λ∞.
By adding the last two equalities in (26) to (36), we know
(38)limk→∞x2k=x2∞,limk→∞x3k=x3∞.
Therefore, we have shown that the whole sequence {(x1k,x2k,x3k,λk)} converges to (x1*,x2∞,x3∞,λ∞) under condition (i) in Theorem 3.
Case 2 (A1 is of full column rank, 0<β<μ2/∥A2∥22, and β≤μ3/∥A3∥22). In this case, the sequence {x2k} converges to x2* and {x1k} is bounded. From the second equality in (25) and (28), we have
(39)limk→∞∥A1x1k+1-A1x1k∥=0,limk→∞∥λk-λk+1∥=0.
Since A1 is of full column rank, it therefore holds that
(40)limk→∞∥x1k+1-x1k∥=0.
Let (x1∞,x3∞,λ∞) be a cluster point of the sequence {x1k,x3k,λk}. Following a similar proof in Case 1, we are able to show (x1∞,x2*,x3∞,λ∞) is a KKT point of problem (3) and the whole sequence {(x1k,x2k,x3k,λk)} converges to this point.
Remark 4 (see [22]).
the authors proved the convergence of the ADMM under the conditions that f1,f2, and f3 are strongly convex and 0<β<min1≤i≤3{μi/3∥Ai∥22}. Our result improves the upper bound min1≤i≤3{μi/3∥Ai∥22} by min{μ2/∥A2∥22,μ3/∥A3∥22}. Moreover, in our condition (ii), the strongly convexity assumption is only imposed on f2 and f3 while f1 is not necessarily strongly convex with positive modulus.
4. The Relaxed ADMM with Three Blocks
For the ADMM with two blocks, Ye and Yuan [25] developed a variant of alternating direction method with an optimal step size. Numerical results demonstrated that an additional computation on the optimal step size would improve the efficiency of the new variant of ADMM. In this section, by adopting the essential idea of Ye and Yuan [25], we propose a relaxed ADMM with three blocks to accelerate the ADMM via an optimal step size. For notational simplicity, we write
(41)Φ(uk,u~k)≔3β4∥A2(x2k-x~2k)∥2+β∥A3(x3-x~3k)∥2+1β∥λk-λ~k∥2+〈λk-λ~k,A2(x2k-x~2k)+A3(x3k-x~3k)〉.
With uk=(x2k,x3k,λk), the new iterate of extended ADMM is produced by
(42)uk+1=uk-γα*(uk-u~k),γ∈(0,2),
where u~k is the solution of (5) and α* is defined by
(43)α*:=Φ(uk,u~k)∥uk-u~k∥G2.
Lemma 5.
Let the sequence {uk} be generated by the relaxed ADMM with three blocks. Then, if 0<β≤μ3/∥A3∥22, the following statements are valid:
By direct computations to Φ(uk,u~k), we know that
(44)Φ(uk,u~k)=3β4∥A2(x2k-x~2k)∥2+β∥A3(x3-x~3k)∥2+1β∥λk-λ~k∥2+〈λk-λ~k,A2(x2k-x~2k)+A3(x3k-x~3k)〉≥3β4∥A2(x2k-x~2k)∥2+β∥A3(x3-x~3k)∥2+1β∥λk-λ~k∥2-β2∥A2(x2k-x~2k)∥2-12β∥λk-λ~k∥2-3β4∥A3(x3k-x~3k)∥2-13β∥λk-λ~k∥2=β4∥A2(x2k-x~2k)∥2+β4∥A3(x3k-x~3k)∥2+16β∥λk-λ~k∥2,
where the second inequality follows Cauchy inequality. It therefore holds that
(45)Φ(uk,u~k)≥16∥uk-u~k∥G2,
which completes the proof of the first part. By Lemma 1 and the elementary inequality (17), it can be easily verified that
(46)〈uk-u*,G(uk-u~k)〉≥Φ(uk,u~k)+μ1∥x~1k-x1*∥2+μ2∥x~2k-x2*∥2+∥x~3k-x3*∥μ3In3-βA3TA32
and then
(47)∥uk+1-u*∥G2=∥uk-u*-γα*(uk-u~k)∥G2≤∥uk-u*∥G2-γ(2-γ)(α*)2×∥uk-u~k∥G2-2γα*μ1∥x~1k-x1*∥2-2μ2γα*∥x~2k-x2*∥2-2γα*∥x~3k-x3*∥μ3In3-βA3TA32.
This, together with the fact that α*≥1/6, completes the proof.
Based on the above inequality, we are able to prove the following convergence result of the relaxed ADMM with three blocks. Since the proof is in line with that of Theorem 3, we omit it.
Theorem 6.
Let {vk=(x1k,x2k,x3k,λk)} be the sequence generated by the relaxed ADMM. Then {vk} converges to a KKT point of problem (3) under the conditions that 0<β≤μ3/∥A3∥22 and A1,A2, and A3 are of full column rank.
5. Conclusion Remarks
In this paper, we take a step to investigate the ADMM for separable convex programming problems with three blocks. Based on the contractive analysis of the distance between the sequence and the solution set, we establish theoretical results to guarantee the global convergence of ADMM with three blocks under weaker conditions than those employed in [22]. By adopting the essential idea of [25], we also present a relaxed ADMM with an optimal step size to accelerate the ADMM and prove its convergence under mild assumptions.
Acknowledgment
The first author is supported by the Natural Science Foundation of Jiangsu Province and the National Natural Science Foundation of China under Project no. 71271112. The second author is supported by university natural science research fund of jiangsu province under grant no. 13KJD110002.
GabayD.MercierB.A dual algorithm for the solution of nonlinear variational problems via finite element approximations197621740GlowinskiR.MarroccoA.Sur l'approximation, par éléments finis d'ordre un, et la résolution, par pénalisation-dualité, d'une classe de problèmes de Dirichlet non linéaires19759R-24176MR0388811ZBL0368.65053EcksteinJ.BertsekasD.On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators199255329331810.1007/BF01581204MR1168183ZBL0765.90073GabayD.Chapter ix applications of the method of multipliers to variational inequalities198315299331GlowinskiR.1984New York, NY, USASpringerMR737005GlowinskiR.Le TallecP.19899Philadelphia, Pa, USASociety for Industrial and Applied Mathematics (SIAM)10.1137/1.9781611970838MR1060954HeB.LiaoL.HanD.YangH.A new inexact alternating directions method for monotone variational inequalities200292110311810.1007/s101070100280MR1892298ZBL1009.90108HeB.YuanX.On the O(1/n) convergence rate of the Douglas-Rachford alternating direction method201250270070910.1137/110836936MR2914282ZBL1245.90084ChenC.HeB.YuanX.Matrix completion via an alternating direction method201232122724510.1093/imanum/drq039MR2875250ZBL1236.65043FazelM.PongT. K.SunD. F.TsengP.Hankel matrix rank minimization with applications to system identification and realization201234394697710.1137/110853996HeB.XuM.YuanX.Solving large-scale least squares semidefinite programming by alternating direction methods201132113615210.1137/090761549MR2811295ZBL1243.49039YangJ.ZhangY.YinW.A fast alternating direction method for tvl1-l2 signal reconstruction from partial fourier data201042288297YangJ.ZhangY.Alternating direction method algorithms for l1-problems in compressive sensing2011331250278WenZ.GoldfarbD.YinW.Alternating direction augmented Lagrangian methods for semidefinite programming201023-420323010.1007/s12532-010-0017-1MR2741485ZBL1206.90088TaoM.YuanX.Recovering low-rank and sparse components of matrices from incomplete and noisy observations2011211578110.1137/100781894MR2765489ZBL1218.90115YangJ.SunD.TohK.A proximal point algorithm for logdeterminant optimization with group Lasso regularization2013232857293PengY.GaneshA.WrightJ.XuW.MaY.RASL: robust alignment by sparse and low-rank decomposition for linearly correlated imagesProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR '10)2010763770MohanK.LondonP.FazelM.WittenD.LeeS. I.Node-based learning of multiple gaussian graphical modelshttp://arxiv.org/abs/1303.5145HeB.TaoM.XuM. H.YuanX. M.Alternating directions based contraction method for generally separable linearly constrained convex programming problems2013624573596HeB.TaoM.YuanX.Alternating direction method with Gaussian back substitution for separable convex programming201222231334010.1137/110822347MR2968856ZBL06081226HongM.LuoZ.On the linear convergence of the alternating direction Method of multipliershttp://arxiv.org/abs/1208.3922HanD.YuanX.A note on the alternating direction method of multipliers2012155122723810.1007/s10957-012-0003-zMR2983116ZBL1255.90093RockafellarR.Monotone operators and the proximal point algorithm1976145877898MR041048310.1137/0314056ZBL0358.90053RockafellarR.1970Princeton, NJ, USAPrinceton University PressPrinceton Mathematical Series, no. 28MR0274683YeC.YuanX.-M.A descent method for structured monotone variational inequalities200722232933810.1080/10556780600552693MR2288768ZBL1196.90118