Abstract
This paper focuses on an inexact block coordinate method designed for nonsmooth optimization, where each block-subproblem is solved by performing a bounded number of steps of a variable metric proximal–gradient method with linesearch. We improve on the existing analysis for this algorithm in the nonconvex setting, showing that the iterates converge to a stationary point of the objective function even when the proximal operator is computed inexactly, according to an implementable inexactness condition. The result is obtained by introducing an appropriate surrogate function that takes into account the inexact evaluation of the proximal operator, and assuming that such function satisfies the Kurdyka–Łojasiewicz inequality. The proof technique employed here may be applied to other new or existing block coordinate methods suited for the same class of optimization problems.
Similar content being viewed by others
Data availability statement
The micro test image is included in the published article [43], available at https://ieeexplore.ieee.org/document/1199635. The satellite test image can be found in the Matlab package IR Tools, available at http://people.compute.dtu.dk/pcha/IRtools/. The peppers test image can be downloaded from the USC-SIPI Image Database, available at https://sipi.usc.edu/database/.
References
Ayers, G.R., Dainty, J.C.: Iterative blind deconvolution method and its applications. Opt. Lett. 13(7), 547–549 (1988)
Levin, A., Weiss, Y., Durand, F., Freeman, W.T.: Understanding and evaluating blind deconvolution algorithms. In: Proceedings of the 27th IEEE Conference on Computer Vision and Pattern Recognition, pp. 1964–1971 (2009)
Prato, M., La Camera, A., Bonettini, S., Rebegoldi, S., Bertero, M., Boccacci, P.: A blind deconvolution method for ground based telescopes and Fizeau interferometers. New Astron. 40, 1–13 (2015)
Bonettini, S.: Inexact block coordinate descent methods with application to the nonnegative matrix factorization. IMA J. Numer. Anal. 31(4), 1431–1452 (2011)
Pock, T., Sabach, S.: Inertial Proximal Alternating Linearized Minimization (iPALM) for Nonconvex and Nonsmooth Problems. SIAM J. Optim. 9(4), 1756–1787 (2016)
Xu, Y., Yin, W.: A block coordinate descent method for regularized multiconvex optimization with applications to nonnegative tensor factorization and completion. SIAM J. Imaging Sci. 6(3), 1758–1789 (2013)
Yang, Y., Pesavento, M., Luo, Z.-Q., Ottersen, B.: Inexact block coordinate descent algorithms for nonsmooth nonconvex optimization. IEEE Trans. Signal Process. 68, 947–961 (2020)
Chouzenoux, E., Corbineau, M.-C., Pesquet, J.-C., Scrivanti, G.: A variational approach for joint image recovery-segmentation based on spatially varying generalised gaussian models. arXiv:2209.01375 (2022)
Brandoni, D., Porcelli, M., Simoncini, V.: A spectral PALM algorithm for matrix and tensor-train based Dictionary Learning. arXiv:2107.11644 (2021)
Bach, F., Jenatton, R., Mairal, J., Obozinski, G.: Structured sparsity through convex optimization. Stat. Sci. 27(4), 450–468 (2012)
Rudin, L.I., Osher, S., Fatemi, E.: Nonlinear total variation based noise removal algorithms. J. Phys. D. 60(1–4), 259–268 (1992)
Bolte, J., Sabach, S., Teboulle, M.: Proximal alternating linearized minimization for nonconvex and nonsmooth problems. Math. Program. 146(1–2), 459–494 (2014)
Grippo, L., Sciandrone, M.: Globally convergent block-coordinate techniques for unconstrained optimization. Optim. Method Softw. 10(4), 587–637 (1999)
Grippo, L., Sciandrone, M.: On the convergence of the block nonlinear Gauss-Seidel method under convex constraints. Oper. Res. Lett. 26(3), 127–136 (2000)
Powell, M.J.D.: On search directions for minimization algorithms. Math. Program. 4(1), 193–201 (1973)
Bonettini, S., Loris, I., Porta, F., Prato, M.: Variable metric inexact line-search-based methods for nonsmooth optimization. SIAM J. Optim. 26(2), 891–921 (2016)
Chouzenoux, E., Pesquet, J.-C., Repetti, A.: Variable metric forward-backward algorithm for minimizing the sum of a differentiable function and a convex function. J. Optim. Theory Appl. 162(1), 107–132 (2014)
Combettes, P.L., Wajs, V.R.: Signal recovery by proximal forward-backward splitting. Multiscale Model. Simul. 4(4), 1168–1200 (2005)
Bonettini, S., Prato, M., Rebegoldi, S.: A block coordinate variable metric linesearch based proximal gradient method. Comput. Optim. Appl. 71(1), 5–52 (2018)
Chouzenoux, E., Pesquet, J.-C., Repetti, A.: A block coordinate variable metric forward–backward algorithm. J. Glob. Optim. 66(3), 457–485 (2016)
Frankel, P., Garrigos, G., Peypouquet, J.: Splitting methods with variable metric for Kurdyka–Łojasiewicz functions and general convergence rates. J. Opt. Theory Appl. 165, 874–900 (2015)
Ochs, P.: Unifying abstract inexact convergence theorems and block coordinate variable metric iPiano. SIAM J. Optim. 29(1), 541–570 (2019)
Barzilai, J., Borwein, J.M.: Two-point step size gradient methods. IMA J. Numer. Anal. 8(1), 141–148 (1988)
Pierro, A.R.D.: A modified expectation maximization algorithm for penalized likelihood estimation in emission tomography. IEEE Trans. Med. Imaging 14(1), 132–137 (1995)
Attouch, H., Bolte, J., Svaiter, B.F.: Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward-backward splitting, and regularized Gauss-Seidel methods. Math. Program. 137(1–2), 91–129 (2013)
Gur, E., Sabach, S., Shtern, S.: Convergent nested alternating minimization algorithms for nonconvex optimization problems. Math. Oper. Res. (2022)
Ma, S., Goldfarb, D., Chen, L.: Fixed point and Bregman iterative methods for matrix rank minimization. Math. Program. 128, 321–353 (2011)
Bonettini, S., Prato, M., Rebegoldi, S.: Convergence of inexact forward-backward algorithms using the forward–backward envelope. SIAM J. Optim. 30(4), 3069–3097 (2020)
Bonettini, S., Prato, M., Rebegoldi, S.: New convergence results for the inexact variable metric forward-backward method. Appl. Math. Comput. 392, 125719 (2021)
Rockafellar, R.T., Wets, R.J.-B., Wets, M.: Variational Analysis. Grundlehren der Mathematischen Wissenschaften, vol. 317. Springer, Berlin (1998)
Zalinescu, A.: Convex Analysis in General Vector Spaces. World Scientific, River Edge (2002)
Birgin, E.G., Martinez, J.M., Raydan, M.: Inexact spectral projected gradient methods on convex sets. IMA J. Numer. Anal. 23(4), 539–559 (2003)
Bonettini, S., Loris, I., Porta, F., Prato, M., Rebegoldi, S.: On the convergence of a linesearch based proximal-gradient method for nonconvex optimization. Inverse Probl. 33(5), 055005 (2017)
Lee, C.-P., Wright, S.J.: Inexact Successive quadratic approximation for regularized optimization. Comput. Optim. Appl. 72(3), 641–674 (2019)
Repetti, A., Wiaux, Y.: Variable metric forward–backward algorithm for composite minimization problems. SIAM J. Optim. 31(2), 1215–1241 (2020)
Bertsekas, D.: Nonlinear Programming. Athena Scientific, Belmont (1999)
Perrone, D., Favaro, P.: A clearer picture of total variation blind deconvolution. IEEE Trans. Pattern Anal. Mach. Intell. 38(6), 1041–1055 (2016)
Rebegoldi, S., Bonettini, S., Prato, M.: Efficient Block Coordinate Methods for Blind Cauchy Denoising. In: Sergeyev, Y.D., Kvasov, D.E. (eds.) Numerical Computations: Theory and Algorithms, pp. 198–211. Springer, Cham (2020)
Lantéri, H., Roche, M., Cuevas, O., Aime, C.: A general method to devise maximum likelihood signal restoration multiplicative algorithms with non-negativity constraints. Signal Process. 81(5), 945–974 (2001)
Bonettini, S., Zanella, R., Zanni, L.: A scaled gradient projection method for constrained image deblurring. Inverse Probl. 25(1), 015002 (2009)
Beck, A., Teboulle, M.: A fast iterative shrinkage-thresholding algorithm for linear inverse problems. SIAM J. Imaging Sci. 2(1), 183–202 (2009)
Ochs, P., Dosovitskiy, A., Brox, T., Pock, T.: On iteratively reweighted algorithms for nonsmooth nonconvex optimization in computer vision. SIAM J. Imaging Sci. 8(1), 331–372 (2015)
Willett, R.M., Nowak, R.D.: Platelets: a multiscale approach for recovering edges and surfaces in photon limited medical imaging. IEEE Trans. Med. Imaging 22, 332–350 (2003)
Funding
Simone Rebegoldi is a member of the INdAM research group GNCS. This work was partially supported by INdAM research group GNCS, under the GNCS project CUP_E55F22000270001.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
All authors certify that they have no affiliations with or involvement in any organization or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix A Proof of Lemma 8
Proof of Lemma 8
Applying the descent lemma [36, Proposition A.24] yields
By setting \(z_i^{(k,\bar{\ell }_i^{(k)})}=x_i^{(k,\bar{\ell }_i^{(k)})}-\alpha _i^{(k,\bar{\ell }_i^{(k)})} (D_i^{(k,\bar{\ell }_i^{(k)})})^{-1}\nabla _i f_0({\tilde{x}}_i^{(k,\bar{\ell }_i^{(k)})})\), \(\epsilon _i^{(k)}=\epsilon _i^{(k,\bar{\ell }_i^{(k)})}\), and applying some calculus rules for the \(\epsilon _i^{(k)}-\)subdifferential (see [33, Appendix A]), the inexactness condition (20) implies the existence of a vector \(e_i^{(k)}\in {\mathbb {R}}^{n_i}\) such that
By employing Definition 4 with \({{\mathcal {F}}}=f_i\), \(u = \hat{u}_i^{(k,\bar{\ell }_i^{(k)})}\), \(z = u_i^{(k,\bar{\ell }_i^{(k)})}\), \(\epsilon =\epsilon _i^{(k)}\), and w given by (A2), we have
Summing the previous inequality over \(i=1,\ldots ,p\) with (A1) and recalling the definition of the points \(\hat{u}^{(k)}\), \({\tilde{u}}^{(k)}\) leads to
By rearranging terms and applying the Cauchy-Schwarz inequality, we get
At this point, for all \(i=1,\ldots ,p\), we can write the following chain of inequalities:
where the first inequality follows by means of Assumption 1(iii) and the second one is obtained by applying the definition of points \({\tilde{x}}_i^{(k,\bar{\ell }_i^{(k)})},\hat{u}^{(k)}\), the basic inequality \(\sqrt{a+b}\le \sqrt{a}+\sqrt{b}\), and the triangular inequality. We can further proceed as follows
where again we have used the triangular inequality to get the first inequality. Summing (A5) over \(i=1,\ldots ,p\) leads to
We can now employ inequalities (22)–(23)–(25) to deduce the following upper bound
where \({\tilde{c}}_1=L\alpha _{\max }\mu \left( 4\sqrt{\tau (1+\tau )}+\sqrt{2\tau \left( 1+\frac{\tau }{2}\right) }\right) \ge 0\). Multiplying and dividing the quantity inside the round brackets by \(\sum _{i=1}^{p}L_i^{(k)}\), applying the Jensen inequality to the function \(\phi (t)=t^2\) and recalling that \(L_i^{(k)}\le L_i\) for all \(k\in {\mathbb {N}}\), \(i=1,\ldots ,p\), allows to obtain
where \(c_1={\tilde{c}}_1\sum _{i=1}^{p}L_i\ge 0\). Similarly, we can provide the following upper bound
where the first inequality follows again from the Cauchy-Schwarz inequality together with \(\alpha _i^{(k,\bar{\ell }_i^{(k)})}\ge \alpha _{\min }\) and \(D_i^{(k,\bar{\ell }_i^{(k)})}\in {\mathcal {M}}_\mu \), the second inequality is due to the application of (23)–(24) and \(c_2=\frac{2\mu ^2\alpha _{\max }}{\alpha _{\min }}\sqrt{\tau (1+\tau )}\ge 0\). Moreover, from relation (A3), we deduce
Then we have
where the second inequality follows from (23) and (A8), whereas \(c_3=\frac{\mu ^2\alpha _{\max }\tau }{\alpha _{\min }}\ge 0\).
Finally, plugging (A6)–(A7)–(A8) inside (A4) gives
Then the thesis follows by setting \(c=c_1+c_2+c_3+\frac{{M}\alpha _{\max }\mu \tau }{2}+\frac{\tau }{2}\ge 0\). \(\square \)
Appendix B Proof of Lemma 9
Proof of Lemma 9
Using the descent lemma, we obtain
We sum \(\sum _{i=1}^{p}f_i(\hat{u}_i^{(k,\bar{\ell }_i^{(k)})})\) to both sides and rearrange terms as follows:
By adding the quantity \(\sum _{i=1}^{p}(2\alpha _i^{(k,\bar{\ell }_i^{(k)})})^{-1}\Vert \hat{u}_i^{(k,\bar{\ell }_i^{(k)})}-x_i^{(k,\bar{\ell }_i^{(k)})}\Vert _{D_i^{(k,\bar{\ell }_i^{(k)})}}^2\), and summing and subtracting the term \(\sum _{i=1}^{p}(2\alpha _i^{(k,\bar{\ell }_i^{(k)})})^{-1}\Vert x_i^{(k)}-x_i^{(k,\bar{\ell }_i^{(k)})}\Vert _{D_i^{(k,\bar{\ell }_i^{(k)})}}^2\) to the right-hand side of the previous inequality, we directly get
By setting \(u=x_i^{(k,\bar{\ell }_i^{(k)})}\) and \(u=x_i^{(k)}\) in (15), respectively, we obtain
Plugging condition (18) inside the previous inequality leads to
By inserting (B11) inside (B9), we can continue as follows
A direct application of the triangular inequality yields
We now provide an upper bound for the third addend in the above inequality, proceeding as follows
where the third inequality follows from the definition of \({\tilde{x}}_i^{(k,\bar{\ell }_i^{(k)})}\) in (12), whereas we have used the triangular inequality, \(\bar{\ell }_i^{(k)}\le L_i^{(k)}\), and \(i\le p\), to get the last three inequalities. If we sum (B13) over \(i=1,\ldots ,p\) and apply it to (B12), we come to
By combining Lemma 5 with Lemma 7, we obtain the following limits as \(k\rightarrow \infty \)
Hence inequality (B14) yields the thesis. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Rebegoldi, S. Analysis of a variable metric block coordinate method under proximal errors. Ann Univ Ferrara 70, 23–61 (2024). https://doi.org/10.1007/s11565-022-00456-z
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11565-022-00456-z