arXiv subjects

Dominique Orban

Publications and source records attributed to Dominique Orban.

36 records · Page 2Linked to original sources

arXiv2023

The Indefinite Proximal Gradient Method

We introduce a variant of the proximal gradient method in which the quadratic term is diagonal but may be indefinite, and is safeguarded by a trust region. Our method is a special case of the proximal quasi-Newton trust-region method of arXiv:2103.15993v3. We provide closed-form solution of the step computation in certain cases where the nonsmooth term is separable and the trust region is defined in infinity norm, so that no iterative subproblem solver is required. Our analysis expands upon that of arXiv:2103.15993v3 by generalizing the trust-region approach to problems with bound constraints. We provide an efficient open-source implementation of our method, named TRDH, in the Julia language in which Hessians approximations are given by diagonal quasi-Newton updates. TRDH evaluates one standard proximal operator and one indefinite proximal operator per iteration. We also analyze and implement a variant named iTRDH that performs a single indefinite proximal operator evaluation per iteration. We establish that iTRDH enjoys the same asymptotic worst-case iteration complexity as TRDH. We report numerical experience on unconstrained and bound-constrained problems, where TRDH and iTRDH are used both as standalone and subproblem solvers. Our results illustrate that, as standalone solvers, TRDH and iTRDH improve upon the quadratic regularization method R2 of arXiv:2103.15993v3 but also sometimes upon their quasi-Newton trust-region method, referred to here as TR-R2, in terms of smooth objective value and gradient evaluations. On challenging nonnegative matrix factorization, binary classification and data fitting problems, TRDH and iTRDH used as subproblem solvers inside TR improve upon TR-R2 for at least one choice of diagonal approximation.

math.OC↗

arXiv2023

A Levenberg-Marquardt Method for Nonsmooth Regularized Least Squares

We develop a Levenberg-Marquardt method for minimizing the sum of a smooth nonlinear least-squar es term $f(x) = \tfrac{1}{2} \|F(x)\|_2^2$ and a nonsmooth term $h$. Both $f$ and $h$ may be nonconvex. Steps are computed by minimizing the sum of a regularized linear least-squares model and a model of $h$ using a first-order method such as the proximal gradient method. We establish global convergence to a first-order stationary point of both a trust-region and a regularization variant of the Levenberg-Marquardt method under the assumptions that $F$ and its Jacobian are Lipschitz continuous and $h$ is proper and lower semi-continuous. In the worst case, both methods perform $O(ε^{-2})$ iterations to bring a measure of stationarity below $ε\in (0, 1)$. We report numerical results on three examples: a group-lasso basis-pursuit denoise example, a nonlinear support vector machine, and parameter estimation in neuron firing. For those examples to be implementable, we describe in detail how to evaluate proximal operators for separable $h$ and for the group lasso with trust-region constraint. In all cases, the Levenberg-Marquardt methods perform fewer outer iterations than a proximal-gradient method with adaptive step length and a quasi-Newton trust-region method, neither of which exploit the least-squares structure of the problem. Our results also highlight the need for more sophisticated subproblem solvers than simple first-order methods.

math.OC↗

arXiv2022

On GSOR, the Generalized Successive Overrelaxation Method for Double Saddle-Point Problems

We consider the generalized successive overrelaxation (GSOR) method for solving a class of block three-by-three saddle-point problems. Based on the necessary and sufficient conditions for all roots of a real cubic polynomial to have modulus less than one, we derive convergence results under reasonable assumptions. We also analyze a class of block lower triangular preconditioners induced from GSOR and derive explicit and sharp spectral bounds for the preconditioned matrices. We report numerical experiments on test problems from the liquid crystal director model and the coupled Stokes-Darcy flow, demonstrating the usefulness of GSOR.

math.NA↗

arXiv2022

A Stochastic Proximal Method for Nonsmooth Regularized Finite Sum Optimization

We consider the problem of training a deep neural network with nonsmooth regularization to retrieve a sparse and efficient sub-structure. Our regularizer is only assumed to be lower semi-continuous and prox-bounded. We combine an adaptive quadratic regularization approach with proximal stochastic gradient principles to derive a new solver, called SR2, whose convergence and worst-case complexity are established without knowledge or approximation of the gradient's Lipschitz constant. We formulate a stopping criteria that ensures an appropriate first-order stationarity measure converges to zero under certain conditions. We establish a worst-case iteration complexity of $\mathcal{O}(ε^{-2})$ that matches those of related methods like ProxGEN, where the learning rate is assumed to be related to the Lipschitz constant. Our experiments on network instances trained on CIFAR-10 and CIFAR-100 with $\ell_1$ and $\ell_0$ regularizations show that SR2 consistently achieves higher sparsity and accuracy than related methods such as ProxGEN and ProxSGD.

stat.ML↗

arXiv2022

A semi-conjugate gradient method for solving unsymmetric positive definite linear systems

The conjugate gradient (CG) method is a classic Krylov subspace method for solving symmetric positive definite linear systems. We introduce an analogous semi-conjugate gradient (SCG) method for unsymmetric positive definite linear systems. Unlike CG, SCG requires the solution of a lower triangular linear system to produce each semi-conjugate direction. We prove that SCG is theoretically equivalent to the full orthogonalization method (FOM), which is based on the Arnoldi process and converges in a finite number of steps. Because SCG's triangular system increases in size each iteration, we study a sliding window implementation (SWI) to improve efficiency, and show that the directions produced are still locally semi-conjugate. A counterexample illustrates that SWI is different from the direct incomplete orthogonalization method (DIOM), which is FOM with a sliding window. Numerical experiments from the convection-diffusion equation and other applications show that SCG is robust and that the sliding window implementation SWI allows SCG to solve large systems efficiently.

math.NA↗

arXiv2022

Computing a Sparse Projection into a Box

We describe a procedure to compute a projection of $w \in \mathbb{R}^n$ into the intersection of the so-called \emph{zero-norm} ball $k \mathbb{B}_0$ of radius $k$, i.e., the set of $k$-sparse vectors, with a box centered at a point of $k \mathbb{B}_0$. The need for such projection arises in the context of certain trust-region methods for nonsmooth regularized optimization. Although the set into which we wish to project is nonconvex, we show that a solution may be found in $O(n \log(n))$ operations. We describe our Julia implementation and illustrate our procedure in the context of two trust-region methods for nonsmooth regularized optimization.

math.OC↗

arXiv2021

Adaptive First- and Second-Order Algorithms for Large-Scale Machine Learning

In this paper, we consider both first- and second-order techniques to address continuous optimization problems arising in machine learning. In the first-order case, we propose a framework of transition from deterministic or semi-deterministic to stochastic quadratic regularization methods. We leverage the two-phase nature of stochastic optimization to propose a novel first-order algorithm with adaptive sampling and adaptive step size. In the second-order case, we propose a novel stochastic damped L-BFGS method that improves on previous algorithms in the highly nonconvex context of deep learning. Both algorithms are evaluated on well-known deep learning datasets and exhibit promising performance.

cs.LG↗

arXiv2021

GPMR: An Iterative Method for Unsymmetric Partitioned Linear Systems

We introduce an iterative method named GPMR for solving 2x2 block unsymmetric linear systems. GPMR is based on a new process that reduces simultaneously two rectangular matrices to upper Hessenberg form and that is closely related to the block-Arnoldi process. GPMR is tantamount to Block-GMRES with two right-hand sides in which the two approximate solutions are summed at each iteration, but requires less storage and work per iteration. We compare the performance of GPMR with GMRES and Block-GMRES on linear systems from the SuiteSparse Matrix Collection. In our experiments, GPMR terminates significantly earlier than GMRES on a residual-based stopping condition with an improvement ranging from around 10% up to 50% in terms of number of iterations. We also illustrate by experiment that GPMR appears more resilient to loss of orthogonality than Block-GMRES.

math.NA↗

arXiv2021

A Proximal Quasi-Newton Trust-Region Method for Nonsmooth Regularized Optimization

We develop a trust-region method for minimizing the sum of a smooth term $f$ and a nonsmooth term $h$), both of which can be nonconvex. Each iteration of our method minimizes a possibly nonconvex model of $f + h$ in a trust region. The model coincides with $f + h$ in value and subdifferential at the center. We establish global convergence to a first-order stationary point when $f$ satisfies a smoothness condition that holds, in particular, when it has Lipschitz-continuous gradient, and $h$ is proper and lower semi-continuous. The model of $h$ is required to be proper, lower-semi-continuous and prox-bounded. Under these weak assumptions, we establish a worst-case $O(1/ε^2)$ iteration complexity bound that matches the best known complexity bound of standard trust-region methods for smooth optimization. We detail a special instance, named TR-PG, in which we use a limited-memory quasi-Newton model of $f$ and compute a step with the proximal gradient method, resulting in a practical proximal quasi-Newton method. We establish similar convergence properties and complexity bound for a quadratic regularization variant, named R2, and provide an interpretation as a proximal gradient method with adaptive step size for nonconvex problems. R2 may also be used to compute steps inside the trust-region method, resulting in an implementation named TR-R2. We describe our Julia implementations and report numerical results on inverse problems from sparse optimization and signal processing. Both TR-PG and TR-R2 exhibit promising performance and compare favorably with two linesearch proximal quasi-Newton methods based on convex models.

math.OC↗

arXiv2021

Scalable adaptive cubic regularization methods

Adaptive cubic regularization (ARC) methods for unconstrained optimization compute steps from linear systems involving a shifted Hessian in the spirit of the Levenberg-Marquardt and trust-region methods. The standard approach consists in performing an iterative search for the shift akin to solving the secular equation in trust-region methods. Such search requires computing the Cholesky factorization of a tentative shifted Hessian at each iteration, which limits the size of problems that can be reasonably considered. We propose a scalable implementation of ARC named ARCqK in which we solve a set of shifted systems concurrently by way of an appropriate modification of the Lanczos formulation of the conjugate gradient (CG) method. At each iteration of ARCqK to solve a problem with n variables, a range of m << n shift parameters is selected. The computational overhead in CG beyond the Lanczos process is thirteen scalar operations to update five vectors of length m and two n-vector updates for each value of the shift. The CG variant only requires one Hessian-vector product and one dot product per iteration, independently of the number of shift parameters. Solves corresponding to inadequate shift parameters are interrupted early. All shifted systems are solved inexactly. Such modest cost makes our implementation scalable and appropriate for large-scale problems. We provide a new analysis of the inexact ARC method including its worst case evaluation complexity, global and asymptotic convergence. We describe our implementation and provide preliminary numerical observations that confirm that for problems of size at least 100, our implementation of ARCqK is more efficient than a classic Steihaug-Toint trust region method. Finally, we generalize our convergence results to inexact Hessians and nonlinear least-squares problems.

math.OC↗

arXiv2021

TriCG and TriMR: Two Iterative Methods for Symmetric Quasi-Definite Systems

We introduce iterative methods named TriCG and TriMR for solving symmetric quasi-definite systems based on the orthogonal tridiagonalization process proposed by Saunders, Simon and Yip in 1988. TriCG and TriMR are tantamount to preconditioned Block-CG and Block-MINRES with two right-hand sides in which the two approximate solutions are summed at each iteration, but require less storage and work per iteration. We evaluate the performance of TriCG and TriMR on linear systems generated from the SuiteSparse Matrix Collection and from discretized and stablized Stokes equations. We compare TriCG and TriMR with SYMMLQ and MINRES, the recommended Krylov methods for symmetric and indefinite systems. In all our experiments, TriCG and TriMR terminate earlier than SYMMLQ and MINRES on a residual-based stopping condition with an improvement of up to 50% in terms of number of iterations. They also terminate more reliably than Block-CG and Block-MINRES. Experiments in quadruple and octuple precision suggest that loss of orthogonality in the basis vectors is significantly less pronounced in TriCG and TriMR than in Block-CG and Block-MINRES.

math.NA↗

arXiv2021

A Julia implementation of Algorithm NCL for constrained optimization

Algorithm NCL is designed for general smooth optimization problems where first and second derivatives are available, including problems whose constraints may not be linearly independent at a solution (i.e., do not satisfy the LICQ). It is equivalent to the LANCELOT augmented Lagrangian method, reformulated as a short sequence of nonlinearly constrained subproblems that can be solved efficiently by IPOPT and KNITRO, with warm starts on each subproblem. We give numerical results from a Julia implementation of Algorithm NCL on tax policy models that do not satisfy the LICQ, and on nonlinear least-squares problems and general problems from the CUTEst test set.

math.OC↗

arXiv2021

Constraint-Preconditioned Krylov Solvers for Regularized Saddle-Point Systems

We consider the iterative solution of regularized saddle-point systems. When the leading block is symmetric and positive semi-definite on an appropriate subspace, Dollar, Gould, Schilders, and Wathen (2006) describe how to apply the conjugate gradient (CG) method coupled with a constraint preconditioner, a choice that has proved to be effective in optimization applications. We investigate the design of constraint-preconditioned variants of other Krylov methods for regularized systems by focusing on the underlying basis-generation process. We build upon principles laid out by Gould, Orban, and Rees (2014) to provide general guidelines that allow us to specialize any Krylov method to regularized saddle-point systems. In particular, we obtain constraint-preconditioned variants of Lanczos and Arnoldi-based methods, including the Lanczos version of CG, MINRES, SYMMLQ, GMRES(m) and DQGMRES. We also provide MATLAB implementations in hopes that they are useful as a basis for the development of more sophisticated software. Finally, we illustrate the numerical behavior of constraint-preconditioned Krylov solvers using symmetric and nonsymmetric systems arising from constrained optimization.

math.NA↗

arXiv2020

Stochastic Damped L-BFGS with Controlled Norm of the Hessian Approximation

We propose a new stochastic variance-reduced damped L-BFGS algorithm, where we leverage estimates of bounds on the largest and smallest eigenvalues of the Hessian approximation to balance its quality and conditioning. Our algorithm, VARCHEN, draws from previous work that proposed a novel stochastic damped L-BFGS algorithm called SdLBFGS. We establish almost sure convergence to a stationary point and a complexity bound. We empirically demonstrate that VARCHEN is more robust than SdLBFGS-VR and SVRG on a modified DavidNet problem -- a highly nonconvex and ill-conditioned problem that arises in the context of deep learning, and their performance is comparable on a logistic regression problem and a nonconvex support-vector machine problem.

cs.LG↗

arXiv2020

Scaled Projected-Directions Methods with Application to Transmission Tomography

Statistical image reconstruction in X-Ray computed tomography yields large-scale regularized linear least-squares problems with nonnegativity bounds, where the memory footprint of the operator is a concern. Discretizing images in cylindrical coordinates results in significant memory savings, and allows parallel operator-vector products without on-the-fly computation of the operator, without necessarily decreasing image quality. However, it deteriorates the conditioning of the operator. We improve the Hessian conditioning by way of a block-circulant scaling operator and we propose a strategy to handle nondiagonal scaling in the context of projected-directions methods for bound-constrained problems. We describe our implementation of the scaling strategy using two algorithms: TRON, a trust-region method with exact second derivatives, and L-BFGS-B, a linesearch method with a limited-memory quasi-Newton Hessian approximation. We compare our approach with one where a change of variable is made in the problem. On two reconstruction problems, our approach converges faster than the change of variable approach, and achieves much tighter accuracy in terms of optimality residual than a first-order method.

math.OC↗

arXiv2019

Implementing a smooth exact penalty function for general constrained nonlinear optimization

We build upon Estrin et al. (2019) to develop a general constrained nonlinear optimization algorithm based on a smooth penalty function proposed by Fletcher (1970, 1973b). Although Fletcher's approach has historically been considered impractical, we show that the computational kernels required are no more expensive than those in other widely accepted methods for nonlinear optimization. The main kernel for evaluating the penalty function and its derivatives solves structured linear systems. When the matrices are available explicitly, we store a single factorization each iteration. Otherwise, we obtain a factorization-free optimization algorithm by solving each linear system iteratively. The penalty function shows promise in cases where the linear systems can be solved efficiently, e.g., PDE-constrained optimization problems when efficient preconditioners exist. We demonstrate the merits of the approach, and give numerical results on several PDE-constrained and standard test problems.

math.OC↗

arXiv2019

Implementing a smooth exact penalty function for equality-constrained nonlinear optimization

We develop a general equality-constrained nonlinear optimization algorithm based on a smooth penalty function proposed by Fletcher (1970). Although it was historically considered to be computationally prohibitive in practice, we demonstrate that the computational kernels required are no more expensive than other widely accepted methods for nonlinear optimization. The main kernel required to evaluate the penalty function and its derivatives is solving a structured linear system. We show how to solve this system efficiently by storing a single factorization each iteration when the matrices are available explicitly. We further show how to adapt the penalty function to the class of factorization-free algorithms by solving the linear system iteratively. The penalty function therefore has promise when the linear system can be solved efficiently, e.g., for PDE-constrained optimization problems where efficient preconditioners exist. We discuss extensions including handling simple constraints explicitly, regularizing the penalty function, and inexact evaluation of the penalty function and its gradients. We demonstrate the merits of the approach and its various features on some nonlinear programs from a standard test set, and some PDE-constrained optimization problems.

math.OC↗

arXiv2019

BiLQ: An Iterative Method for Nonsymmetric Linear Systems with a Quasi-Minimum Error Property

We introduce an iterative method named BiLQ for solving general square linear systems Ax = b based on the Lanczos biorthogonalization process defined by least-norm subproblems, and that is a natural companion to BiCG and QMR. Whereas the BiCG (Fletcher, 1976), CGS (Sonneveld, 1989) and BiCGSTAB (van der Vorst, 1992) iterates may not exist when the tridiagonal projection of A is singular, BiLQ is reliable on compatible systems even if A is ill-conditioned or rank deficient. As in the symmetric case, the BiCG residual is often smaller than the BiLQ residual and, when the BiCG iterate exists, an inexpensive transfer from the BiLQ iterate is possible. Although the Euclidean norm of the BiLQ error is usually not monotonic, it is monotonic in a different norm that depends on the Lanczos vectors. We establish a similar property for the QMR (Freund and Nachtigal, 1991) residual. BiLQ combines with QMR to take advantage of two initial vectors and solve a system and an adjoint system simultaneously at a cost similar to that of applying either method. We derive an analogous combination of USYMLQ and USYMQR based on the orthogonal tridiagonalization process (Saunders, Simon, and Yip, 1988). The resulting combinations, named BiLQR and TriLQR, may be used to estimate integral functionals involving the solution of a primal and an adjoint system. We compare BiLQR and TriLQR with Minres-qlp on a related augmented system, which performs a comparable amount of work and requires comparable storage. In our experiments, BiLQR terminates earlier than TriLQR and MINRES-QLP in terms of residual and error of the primal and adjoint systems.

math.NA↗