Abstract
In this paper, we consider a block-structured convex optimization model, where in the objective the block variables are nonseparable and they are further linearly coupled in the constraint. For the 2-block case, we propose a number of first-order algorithms to solve this model. First, the alternating direction method of multipliers (ADMM) is extended, assuming that it is easy to optimize the augmented Lagrangian function with one block of variables at each time while fixing the other block. We prove that iteration complexity bound holds under suitable conditions, where t is the number of iterations. If the subroutines of the ADMM cannot be implemented, then we propose new alternative algorithms to be called alternating proximal gradient method of multipliers, alternating gradient projection method of multipliers, and the hybrids thereof. Under suitable conditions, the iteration complexity bound is shown to hold for all the newly proposed algorithms. Finally, we extend the analysis for the ADMM to the general multi-block case.
Similar content being viewed by others
References
James, G.M., Paulson, C., Rusmevichientong, P.: The constrained lasso. Technical report, University of Southern California (2013)
Alizadeh, M., Li, X., Wang, Z., Scaglione, A., Melton, R.: Demand-side management in the smart grid: information processing for the power switch. IEEE Sig. Process. Mag. 29(5), 55–67 (2012)
Chang, T.-H., Alizadeh, M., Scaglione A.: Coordinated home energy management for real-time power balancing. In: IEEE Power and Energy Society General Meeting, pp. 1–8 (2012)
Li, N., Chen, L., Low, S.H.: Optimal demand response based on utility maximization in power networks. In: IEEE Power and Energy Society General Meeting, pp. 1–8 (2011)
Paatero, J.V., Lund, P.D.: A model for generating household electricity load profiles. Int. J. Ener. Res. 30(5), 273–290 (2006)
Cui, Y., Li, X., Sun, D., Toh, K.-C.: On the convergence properties of a majorized ADMM for linearly constrained convex optimization problems with coupled objective functions. arXiv:1502.00098 (2015)
Hong, M., Chang, T.-H., Wang, X., Razaviyayn, M., Ma, S., Luo, Z.-Q.: A block successive upper bound minimization method of multipliers for linearly constrained convex optimization. arXiv:1401.7079 (2014)
Bertsekas, D.P., Tsitsiklis, J.N.: Parallel and Distributed Computation: Numerical Methods, vol. 23. Prentice Hall, Englewood Cliffs (1989)
Douglas, J., Rachford, H.: On the numerical solution of heat conduction problems in two and three space variables. Trans. Am. Math. Soc. 82(2), 421–439 (1956)
Eckstein, J.: Splitting methods for monotone operators with applications to parallel optimization. PhD dissertation, Massachusetts Institute of Technology (1989)
Eckstein, J., Bertsekas, D.: On the Douglas-Rachford splitting method and the proximal point algorithm for maximal monotone operators. Math. Program. 55(1–3), 293–318 (1992)
Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Found. Trends. Mach. Learn. 3(1), 1–122 (2011)
Feng, C., Xu, H., Li, B.: An alternating direction method approach to cloud traffic management. arXiv:1407.8309 (2014)
Scheinberg, K., Ma, S., Goldfarb, D.: Sparse inverse covariance selection via alternating linearization methods. In: Advances in neural information processing systems, pp. 2101–2109 (2010)
Yin, W., Osher, S., Goldfarb, D., Darbon, J.: Bregman iterative algorithms for \(l_1\)-minimization with applications to compressed sensing. SIAM J. Imag. Sci. 1(1), 143–168 (2008)
Glowinski, R., Le Tallec, P.: Augmented Lagrangian and Operator-Splitting Methods in Nonlinear Mechanics, vol. 9. SIAM, Philadelphia (1989)
He, B., Yuan, X.: On the \(O(1/n)\) convergence rate of the Douglas-Rachford alternating direction method. SIAM J. Numer. Anal. 50(2), 700–709 (2012)
Monteiro, R.D., Svaiter, B.F.: Iteration-complexity of block-decomposition algorithms and the alternating direction method of multipliers. SIAM J. Optim. 23(1), 475–507 (2013)
Boley, D.: Local linear convergence of the alternating direction method of multipliers on quadratic or linear programs. SIAM J. Optim. 23(4), 2183–2207 (2013)
Deng, W., and Yin, W.: On the global and linear convergence of the generalized alternating direction method of multipliers. J. Sci. Comput. 1–28 (2012)
Hong, M., Luo, Z.: On the linear convergence of the alternating direction method of multipliers. arXiv:1208.3922 (2012)
Lin, T., Ma, S., Zhang, S.: On the global linear convergence of the ADMM with multi-block variables. arXiv:1408.4266 (2014)
Chen, C., He, B., Ye, Y., Yuan, X.: The direct extension of admm for multi-block convex minimization problems is not necessarily convergent. Math. Program. 155(1–2), 57–79 (2016)
Deng, W., Lai, M.-J., Peng, Z., Yin, W.: Parallel multi-block admm with o (1/k) convergence. arXiv:1312.3040 (2013)
He, B., Hou, L., Yuan, X.: On full jacobian decomposition of the augmented lagrangian method for separable convex programming. SIAM J. Optim. 25(4), 2274–2312 (2015)
He, B., Tao, M., Yuan, X.: Convergence rate and iteration complexity on the alternating direction method of multipliers with a substitution procedure for separable convex programming. Optimization Online (2012)
Lin, T., Ma, S., Zhang, S.: Iteration complexity analysis of multi-block ADMM for a family of convex minimization without strong convexity. J. Sci. Comput. pp. 1–30 (2016)
Hong, M., Luo, Z.-Q., Razaviyayn, M.: Convergence analysis of alternating direction method of multipliers for a family of nonconvex problems. In: IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 3836–3840 (2015)
Ortega, J., Rheinboldt, W.: Iterative Solution of Nonlinear Equations in Several Variables. Classics in applied mathematics, vol. 30. SIAM, Philadelphia (2000)
Drori, Y., Sabach, S., Teboulle, M.: A simple algorithm for a class of nonsmooth convex-concave saddle-point problems. Oper. Res. Lett. 43(2), 209–214 (2015)
Gao, X., Jiang, B., Zhang, S.: On the information-adaptive variants of the ADMM: an iteration complexity perspective. Optimization Online (2014)
Liu, J., Chen, J., Ye, J.: Large-scale sparse logistic regression. In: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 547–556. ACM (2009)
Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J. Imag. Sci. 2(1), 183–202 (2009)
Lin, T., Ma, S., Zhang, S.: An extragradient-based alternating direction method for convex minimization. Foundations of Computational Mathematics, pp. 1–25 (2015)
Robinson, D.P., Tappenden, R.E.: A flexible admm algorithm for big data applications. arXiv:1502.04391 (2015)
Bolte, J., Sabach, S., Teboulle, M.: Proximal alternating linearized minimization for nonconvex and nonsmooth problems. Math. Program. 146, 459–494 (2014)
Author information
Authors and Affiliations
Corresponding author
Additional information
This paper is dedicated to Professor Lian-Sheng Zhang in celebration of his 80th birthday.
This paper was supported in part by National Science Foundation (No. CMMI-1462408).
Appendix: Proofs of the Convergence Theorems
Appendix: Proofs of the Convergence Theorems
1.1 Appendix 1: Proof of Theorem 3.1
We have \(F(w)=\left( \begin{array}{lll} 0 &{}\quad 0 &{}\quad -A^\mathrm{T} \\ 0 &{}\quad 0 &{}\quad -B^\mathrm{T} \\ A &{}\quad B&{}\quad 0 \\ \end{array} \right) \left( \begin{array}{l} x \\ y \\ \lambda \\ \end{array} \right) - \left( \begin{array}{l} 0 \\ 0 \\ b \\ \end{array} \right) ,\) for any \(w_1\) and \(w_2\), and so
Expanding on this identity, we have for any \(w^0,w^1,\cdots ,w^{t-1}\) and \({\bar{w}} = \frac{1}{t} \sum \limits _{k=0}^{t-1} w^k\), that
We begin our analysis with the following property of the ADMM algorithm.
Proposition 7.1
Suppose \(h_2\) is strongly convex with parameter \(\sigma >0\). Let \(\{{\tilde{w}}^k\}\) be defined by (2.6), and the matrices Q, M, P be given in (2.5). First of all, for any \(w\in \Omega \), we have
Furthermore,
Proof
By the optimality condition of the two subproblems in ADMM, we have
where \(h'_1(x^{k+1})\in \partial h_1(x^{k+1})\), and
where \(h'_2(x^{k+1})\in \partial h_2(x^{k+1})\).
Note that \({\tilde{\lambda }}^k=\lambda ^k-\gamma (Ax^{k+1}+By^{k}-b)\). The above two inequalities can be rewritten as
and
Observe the following chain of inequalities
Since
we have
By the strong convexity of the function \(h_2(y)\), we have
Because of the convexity of \(h_1(x)\) and combining (7.8), (7.7), (7.6), (7.5), and (7.4), we have
for any \(w\in \Omega \) and \({\tilde{w}}^k\).
By definition of Q, (7.2) of Proposition 7.1 follows. For (7.3), due to the similarity, we refer to Lemma 3.2 in [17] (noting the matrices Q, P, and M).
The following theorem exhibits an important relationship between two consecutive iterates \(w^k\) and \(w^{k+1}\) from which the convergence would follow.
Proposition 7.2
Let \({w^k}\) be the sequence generated by the ADMM, \({{\tilde{w}}^k}\) be defined as in (2.6) and H satisfies \(H_s:=H-\left( L+\frac{L^2}{\sigma }\right) I_{q}\succeq 0\). Then the following holds:
where
Proof
It follows from Proposition 7.1 that
Note that \(H_{s}:=H-(L+\frac{L^2}{\sigma })I_{q}\succeq 0\), we have the following
Thus, combining (7.11) and (7.12), we have
By the definition of \(\hat{M}\) and \(H_d\) according to (7.10), it follows from (7.13) that
Letting \(w=w^*\) in (7.14), we have
By the monotonicity of F and using the optimality of \(w^*\), we have
which completes the proof.
1.2 Appendix 2: Proof of Theorem 3.1.
Proof
First, according to (7.9), it holds that \(\{w^k\}\) is bounded and
Thus, those two sequences have the same cluster points: For any \(w^{k_n}\rightarrow w^\infty \), by (7.16) we also have \({\tilde{w}}^{k_n}\rightarrow w^\infty \). Applying inequality (7.2) to \(\{w^{k_n}\},\{{\tilde{w}}^{k_n}\}\) and taking the limit, it yields that
Consequently, the cluster point \(w^\infty \) is an optimal solution. Since (7.9) is true for any optimal solution \(w^*\), it also holds for \(w^\infty \), and that implies \({w^k}\) will converge to \(w^\infty \).
Recall (7.2) and (7.3) in Proposition 7.1, those would imply that
Furthermore, since \(H-\left( L+\frac{L^2}{\sigma }\right) I_{q}\succeq 0\), we have
Thus, combining (7.18) and (7.19) leads to
By the definition of M in (2.5) and denoting \(\hat{H}=\gamma B^{\top }B+H\), (7.20) leads to
Before proceeding, let us introduce \({\bar{w}}_n:=\frac{1}{n}\sum \limits _{k=0}^{n-1} {\tilde{w}}^k\). Moreover, recall the definition of \({\bar{u}}_n\) in (3.2), we have
Now, summing the inequality (7.21) over \(k=0,1,\cdots ,t-1\) yields
where the first inequality is due to the convexity of h and (7.1).
Note the above inequality is true for all \(x\in \mathcal {X}, y\in \mathcal {Y}\), and \(\lambda \in \mathbb R^m\), hence it is also true for any optimal solution \(x^*, y^*\), and \({\mathcal {B}}_\rho =\{\lambda : \Vert \lambda \Vert \leqslant \rho \}\). As a result,
which, combined with (7.22), implies that
and so by optimizing over \((x^*,y^*)\in \mathcal {X}^* \times \mathcal {Y}^*\), we have
This completes the proof.
1.3 Appendix 3: Proof of Theorem 4.1
Similar to the analysis for ADMM, we need the following proposition in the analysis of APGMM.
Proposition 7.3
Let \(\{{\tilde{w}}^k\}\) be defined by (2.6), and the matrices Q, M, P be given as in (2.5). For any \(w\in \Omega \), we have
Proof
First, by the optimality condition of the two subproblems in APGMM, we have
and
Note that \({\tilde{\lambda }}^k=\lambda ^k-\gamma (Ax^{k+1}+By^{k}-b)\), and by the definition of \({\tilde{w}}^k\), the above two inequalities are equivalent to
and
Notice that
Besides, we also have
Thus
By the convexity of \(h_1(x)\) and \(h_2(y)\), combining (7.30), (7.29), (7.28), and (7.27), we have
for any \(w\in \Omega \) and \({\tilde{w}}^k\).
By definition of Q, we have shown (7.25) in Proposition 7.3. Equality (7.26) directly follows from (7.3) in Proposition 7.1.
With Proposition 7.3 in place, we can show Theorem 4.1 by exactly following the same steps as in the proof of Theorem 3.1, noting of course the altered assumptions on the matrices G and H. In the meanwhile, we also point out the following proposition which is similar to Proposition 7.2. Since most steps of the proofs are almost identical to that of the previous theorems, we omit the details for succinctness.
Proposition 7.4
Let \({w^k}\) be the sequence generated by the APGMM, \({{\tilde{w}}^k}\) be as defined in (2.6), and H and G are chosen so as to satisfy \(H_s:=H-LI_{q}\succ 0\) and \(G_s:=G-LI_{p}\succ 0\). Then the following holds :
where
and \(\hat{H}=\gamma B^{\mathrm{T}}B+H\).
Theorem 4.1 follows from the above propositions.
1.4 Appendix 4: Proof of Theorem 4.2
Similar to the analysis for APGMM, we do not need any strong convexity here, but we do need to assume that the gradients \(\nabla _x h_1(x)\) and \(\nabla _y h_2(y)\) are Lipschitz continuous. Without loss of generality, we further assume that the Lipschitz constant is the same as \(\nabla f(x,y)\) which is L, that is,
Proposition 7.5
Let \(\{{\tilde{w}}^k\}\) be defined by (2.6), and the matrices Q, M, P be as given in (2.5), and \(G:=\gamma A^{\mathrm{T}}A+\frac{1}{\alpha }I_{p}, H:=\frac{1}{\alpha }I_{q}-\gamma B^{\mathrm{T}}B\succeq 0\). First of all, for any \(w\in \Omega \), we have
Proof
First, by the optimality condition of the two subproblems in AGPMM, we have
and
Noting \({\tilde{\lambda }}^k=\lambda ^k-\gamma (Ax^{k+1}+By^{k}-b)\) and the definition of \({\tilde{w}}^k\), the above two inequalities are, respectively, equivalent to
and
Similar to Proposition 7.3, we have
Moreover, by (2.4), we have
Besides,
Thus
Combining (7.38), (7.37), (7.36), (7.35), and (7.34), and noticing that \(G:=\gamma A^{\mathrm{T}}A+\frac{1}{\alpha }I_{p}, H:=\frac{1}{\alpha }I_{q}-\gamma B^{\mathrm{T}}B\), we have, for any \(w\in \Omega \) and \({\tilde{w}}^k\), that
Using the definition of Q, (7.32) follows. In view of (7.3) in Proposition 7.1, equality (7.33) also readily follows.
With Proposition 7.5, similar as before, we can show Theorem 4.2 by following the same approach as in the proof of Theorem 3.1. We skip the details here for succinctness.
Proposition 7.6
Let \({w^k}\) be the sequence generated by the AGPMM, \({{\tilde{w}}^k}\) be defined in (2.6), and \(G:=\gamma A^{\mathrm{T}}A+\frac{1}{\alpha }I_{p}, H:=\frac{1}{\alpha }I_{q}-\gamma B^{\mathrm{T}}B\). Suppose that \(\alpha \) satisfies that \(H_s:=H-2 L I_{q}\succ 0\) and \(G_s:=G-2 L I_{p}\succ 0\). Then the following holds
where
and \({\hat{H}}=\gamma B^{\mathrm{T}}B+H\).
Theorem 4.2 now follows from the above propositions.
1.5 Appendix 5: Proofs of Theorems 4.3 and 4.4
Proposition 6.1
Let \(\{{\tilde{w}}^k\}\) be defined by (2.6), and the matrices Q, M, P be given in (2.5). For any \(w\in \Omega \), we have
Proof
First, by the optimality condition of the two subproblems in ADM-PG, we have
and
Noting \({\tilde{\lambda }}^k=\lambda ^k-\gamma (Ax^{k+1}+By^{k}-b)\) and the definition of \({\tilde{w}}^k\), the above two inequalities are equivalent to
and
Moreover,
Besides,
and so
By the convexity of \(h_1(x)\) and \(h_2(y)\), combining (7.43), (7.42), (7.41), and (7.40), we have
for any \(w\in \Omega \) and \({\tilde{w}}^k\).
By similar derivations as in the proofs for Proposition 7.5, (7.39) follows.
With Proposition 6.1 in place, we can prove Theorem 4.3 similarly as in the proof of Theorem 3.1. We skip the details here for succinctness.
For ADM-GP, we do not need strong convexity, but we do need to assume that the gradient \(\nabla _y h_2(y)\) of \(h_2(y)\) is Lipschitz continuous. Without loss of generality, we further assume that the Lipschitz constant of \(\nabla _y h_2(y)\) is the same as \(\nabla f(x,y)\) which is L:
Proposition 6.2
Let \(\{{\tilde{w}}^k\}\) be defined by (2.6), and the matrices Q, M, P be given in (2.5), and \(H:=\frac{1}{\alpha }I_{q}-\gamma B^{\mathrm{T}}B\succeq 0\). For any \(w\in \Omega \), we have
Proof
By the optimality condition of the two subproblems in ADMM, we have
and
Noting \({\tilde{\lambda }}^k=\lambda ^k-\gamma (Ax^{k+1}+By^{k}-b)\) and the definition of \({\tilde{w}}^k\), the above two inequalities are equivalent to
and
Therefore,
Moreover, by (2.4), we have
Since
we have
By the convexity of \(h_1(x)\), combining (7.50), (7.49), (7.48), (7.47), (7.46), and noticing \(H:=\frac{1}{\alpha }I_{q}-\gamma B^{\mathrm{T}}B\) for any \(w\in \Omega \) and \({\tilde{w}}^k\), we have
As a result, (7.45) follows.
The proof of Theorem 4.4 follows a similar line of derivations as in the proof of Theorem 3.1, and so we omit the details here.
Rights and permissions
About this article
Cite this article
Gao, X., Zhang, SZ. First-Order Algorithms for Convex Optimization with Nonseparable Objective and Coupled Constraints . J. Oper. Res. Soc. China 5, 131–159 (2017). https://doi.org/10.1007/s40305-016-0131-5
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s40305-016-0131-5