“Relative Continuity” for Non-Lipschitz Nonsmooth Convex Optimization Using Stochastic (or Deterministic) Mirror Descent

Relative smoothness—a notion introduced in Birnbaum et al. (Proceedings of the 12th ACM conference on electronic commerce, ACM, pp 127–136, 2011) and recently rediscovered in Bauschke et al. (Math Oper Res 330–348, 2016) and Lu et al. (Relatively-smooth convex optimization by first-order methods, and applications, arXiv:1610.05708 , 2016)—generalizes the standard notion of smoothness typically used in the analysis of gradient type methods. In this work we are taking ideas from well studied field of stochastic convex optimization and using them in order to obtain faster algorithms for minimizing relatively smooth functions. We propose and analyze two new algorithms: Relative Randomized Coordinate Descent (relRCD) and Relative Stochastic Gradient Descent (relSGD), both generalizing famous algorithms in the standard smooth setting. The methods we propose can be in fact seen as particular instances of stochastic mirror descent algorithms, which has been usually analyzed under stronger assumptions: Lipschitzness of the objective and strong convexity of the reference function. As a consequence, one of the proposed methods, relRCD corresponds to the first stochastic variant of mirror descent algorithm with linear convergence rate.

The stochastic mirror descent (SMD) algorithm is a general class of training algorithms that utilizes a mirror potential to influence the implicit bias of the training algorithm and includes stochastic gradient descent (SGD) as a special case. In this paper, we explore the performance of the SMD on mean-field ensemble models and generalize earlier results obtained for SGD. The evolution of the distribution of parameters is mapped to a continuous time process in the space of probability distributions. Our main result gives a nonlinear partial differential equation (PDE) to which the continuous time process converges in the asymptotic of large networks. The impact of the mirror potential appears through a multiplicative term that is equal to the inverse of its Hessian and defines a gradient flow over an appropriate Riemannian manifold. We provide numerical simulations which allow us to study and characterize the effect of the mirror potential on the performance of networks trained with SMD for some binary classification problems.

In this paper we present a variant of the proximal forward-backward splitting iteration for solving nonsmooth optimization problems in Hilbert spaces, when the objective function is the sum of two nondifferentiable convex functions. The proposed iteration, which will be called Proximal Subgradient Splitting Method, extends the classical subgradient iteration for important classes of problems, exploiting the additive structure of the objective function. The weak convergence of the generated sequence was established using different stepsizes and under suitable assumptions. Moreover, we analyze the complexity of the iterates.

The standard assumption for proving linear convergence of first order methods for smooth convex optimization is the strong convexity of the objective function, an assumption which does not hold for many practical applications. In this paper, we derive linear convergence rates of several first order methods for solving smooth non-strongly convex constrained optimization problems, i.e. involving an objective function with a Lipschitz continuous gradient that satisfies some relaxed strong convexity condition. In particular, in the case of smooth constrained convex optimization, we provide several relaxations of the strong convexity conditions and prove that they are sufficient for getting linear convergence for several first order methods such as projected gradient, fast gradient and feasible descent methods. We also provide examples of functional classes that satisfy our proposed relaxations of strong convexity conditions. Finally, we show that the proposed relaxed strong convexity conditions cover important applications ranging from solving linear systems, Linear Programming, and dual formulations of linearly constrained convex problems.

ABSTRACTIn the late sixties, N. Shor and B. Polyak independently proposed optimal first-order methods for solving non-smooth convex optimization problems. In 1982 A. Nemirovski proposed optimal first-order methods for solving smooth convex optimization problems, which utilized auxiliary line search. In 1985 A. Nemirovski and Yu. Nesterov proposed a parametric family of optimal first-order methods for solving convex optimization problems with intermediate smoothness. In 2013 Yu. Nesterov proposed a universal gradient method which combined all good properties of the previous methods, except the possibility of using auxiliary line search. One can typically observe that in practice auxiliary line search improves performance for many tasks. In this paper, we propose the apparently first such method of non-smooth convex optimization allowing the use of the line search procedure. Moreover, it is based on the universal gradient method, which does not require any a priori information about the actual degree of smoothness of the problem. Numerical experiments demonstrate that the proposed method is, in some cases, considerably faster than Nesterov's universal gradient method.

This paper considers a stochastic optimization problem over the fixed point sets of quasinonexpansive mappings on Riemannian manifolds. The problem enables us to consider Riemannian hierarchical optimization problems over complicated sets, such as the intersection of many closed convex sets, the set of all minimizers of a nonsmooth convex function, and the intersection of sublevel sets of nonsmooth convex functions. We focus on adaptive learning rate optimization algorithms, which adapt step-sizes (referred to as learning rates in the machine learning field) to find optimal solutions quickly. We then propose a Riemannian stochastic fixed point optimization algorithm, which combines fixed point approximation methods on Riemannian manifolds with the adaptive learning rate optimization algorithms. We also give convergence analyses of the proposed algorithm for nonsmooth convex and smooth nonconvex optimization. The analysis results indicate that, with small constant step-sizes, the proposed algorithm approximates a solution to the problem. Consideration of the case in which step-size sequences are diminishing demonstrates that the proposed algorithm solves the problem with a guaranteed convergence rate. This paper also provides numerical comparisons that demonstrate the effectiveness of the proposed algorithms with formulas based on the adaptive learning rate optimization algorithms, such as Adam and AMSGrad.

Recent decades have witnessed the rise of data deluge generated by heterogeneous sources, e.g., social networks, streaming, marketing services etc., which has naturally created a surge of interests in theory and applications of large-scale convex and non-convex optimization. For example, real-world instances of statistical learning problems such as deep learning, recommendation systems, etc. can generate sheer volumes of spatially/temporally diverse data (up to Petabytes of data in commercial applications) with millions of decision variables to be optimized. Such problems are often referred to as Big-data problems. Solving these problems by standard optimization methods demands intractable amount of centralized storage and computational resources which is infeasible and is the foremost purpose of parallel and decentralized algorithms developed in this thesis.This thesis consists of two parts: (I) Distributed Nonconvex Optimization and (II) Distributed Convex Optimization.In Part (I), we start by studying a winning paradigm in big-data optimization, Block Coordinate Descent (BCD) algorithm, which cease to be effective when problem dimensions grow overwhelmingly. In particular, we considered a general family of constrained non-convex composite large-scale problems defined on multicore computing machines equipped with shared memory. We design a hybrid deterministic/random parallel algorithm to efficiently solve such problems combining synergically Successive Convex Approximation (SCA) with greedy/random dimensionality reduction techniques. We provide theoretical and empirical results showing efficacy of the proposed scheme in face of huge-scale problems. The next step is to broaden the network setting to general mesh networks modeled as directed graphs, and propose a class of gradient-tracking based algorithms with global convergence guarantees to critical points of the problem. We further explore the geometry of the landscape of the non-convex problems to establish second-order guarantees and strengthen our convergence to local optimal solutions results to global optimal solutions for a wide range of Machine Learning problems.In Part (II), we focus on a family of distributed convex optimization problems defined over meshed networks. Relevant state-of-the-art algorithms often consider limited problem settings with pessimistic communication complexities with respect to the complexity of their centralized variants, which raises an important question: can one achieve the rate of centralized first-order methods over networks, and moreover, can one improve upon their communication costs by using higher-order local solvers? To answer these questions, we proposed an algorithm that utilizes surrogate objective functions in local solvers (hence going beyond first-order realms, such as proximal-gradient) coupled with a perturbed (push-sum) consensus mechanism that aims to track locally the gradient of the central objective function. The algorithm is proved to match the convergence rate of its centralized counterparts, up to multiplying network factors. When considering in particular, Empirical Risk Minimization (ERM) problems with statistically homogeneous data across the agents, our algorithm employing high-order surrogates provably achieves faster rates than what is achievable by first-order methods. Such improvements are made without exchanging any Hessian matrices over the network. Finally, we focus on the ill-conditioning issue impacting the efficiency of decentralized first-order methods over networks which rendered them impractical both in terms of computation and communication cost. A natural solution is to develop distributed second-order methods, but their requisite for Hessian information incurs substantial communication overheads on the network. To work around such exorbitant communication costs, we propose a “statistically informed” preconditioned cubic regularized Newton method which provably improves upon the rates of first-order methods. The proposed scheme does not require communication of Hessian information in the network, and yet, achieves the iteration complexity of centralized second-order methods up to the statistical precision. In addition, (second-order) approximate nature of the utilized surrogate functions, improves upon the per-iteration computational cost of our earlier proposed scheme in this setting.

Based on the ideas of arXiv:1710.06612, we consider the problem of minimization of the Holder-continuous non-smooth functional $f$ with non-positive convex (generally, non-smooth) Lipschitz-continuous functional constraint. We propose some novel strategies of step-sizes and adaptive stopping rules in Mirror Descent algorithms for the considered class of problems. It is shown that the methods are applicable to the objective functionals of various levels of smoothness. Applying the restart technique to the Mirror Descent Algorithm there was proposed an optimal method to solve optimization problems with strongly convex objective functionals. Estimates of the rate of convergence of the considered algorithms are obtained depending on the level of smoothness of the objective functional. These estimates indicate the optimality of considered methods from the point of view of the theory of lower oracle bounds. In addition, the case of a quasi-convex objective functional and constraint was considered.

Two optimization algorithms are proposed for solving a stochastic programming problem for which the objective function is given in the form of the expectation of convex functions and the constraint set is defined by the intersection of fixed point sets of nonexpansive mappings in a real Hilbert space. This setting of fixed point constraints enables consideration of the case in which the projection onto each of the constraint sets cannot be computed efficiently. Both algorithms use a convex function and a nonexpansive mapping determined by a certain probabilistic process at each iteration. One algorithm blends a stochastic gradient method with the Halpern fixed point algorithm. The other is based on a stochastic proximal point algorithm and the Halpern fixed point algorithm; it can be applied to nonsmooth convex optimization. Convergence analysis showed that, under certain assumptions, any weak sequential cluster point of the sequence generated by either algorithm almost surely belongs to the solution set of the problem. Convergence rate analysis illustrated their efficiency, and the numerical results of convex optimization over fixed point sets demonstrated their effectiveness.

In this paper, the convergence of the fundamental alternating minimization is established for non-smooth non-strongly convex optimization problems in Banach spaces, and novel rates of convergence are provided. As objective function a composition of a smooth, and a block-separable, non-smooth part is considered, covering a large range of applications. For the former, three different relaxations of strong convexity are considered: (i) quasi-strong convexity; (ii) quadratic functional growth; and (iii) plain convexity. With new and improved rates benefiting from both separate steps of the scheme, linear convergence is proved for (i) and (ii), whereas sublinear convergence is showed for (iii).

A modified Polak–Ribière–Polyak conjugate gradient algorithm for nonsmooth convex programs

Nonsmooth convex optimization problems are solved over fixed point sets of nonexpansive mappings by using a distributed optimization technique. This is done for a networked system with an operator, who manages the system, and a finite number of users, by solving the problem of minimizing the sum of the operator’s and users’ nondifferentiable, convex objective functions over the intersection of the operator’s and users’ convex constraint sets in a real Hilbert space. We assume that each of their constraint sets can be expressed as the fixed point set of an implementable nonexpansive mapping. This setting allows us to discuss nonsmooth convex optimization problems in which the metric projection onto the constraint set cannot be calculated explicitly. We propose a parallel subgradient algorithm for solving the problem by using the operator’s attribution such that it can communicate with all users. The proposed algorithm does not use any proximity operators, in contrast to conventional parallel algorithms for nonsmooth convex optimization. We first study its convergence property for a constant step-size rule. The analysis indicates that the proposed algorithm with a small constant step size approximates a solution to the problem. We next consider the case of a diminishing step-size sequence and prove that there exists a subsequence of the sequence generated by the algorithm which weakly converges to a solution to the problem. We also give numerical examples to support the convergence analyses.

In this paper, we present a conditional gradient type (CGT) method for solving a class of composite optimization problems where the objective function consists of a (weakly) smooth term and a (strongly) convex regularization term. While including a strongly convex term in the subproblems of the classical conditional gradient method improves its rate of convergence, it does not cost per iteration as much as general proximal type algorithms. More specifically, we present a unified analysis for the CGT method in the sense that it achieves the best known rate of convergence when the weakly smooth term is nonconvex and possesses (nearly) optimal complexity if it turns out to be convex. While implementation of the CGT method requires explicitly estimating problem parameters like the level of smoothness of the first term in the objective function, we also present a few variants of this method which relax such estimation. Unlike general proximal type parameter free methods, these variants of the CGT method do not require any additional effort for computing (sub)gradients of the objective function and/or solving extra subproblems at each iteration. We then generalize these methods under stochastic setting and present a few new complexity results. To the best of our knowledge, this is the first time that such complexity results are presented for solving stochastic weakly smooth nonconvex and (strongly) convex optimization problems.

We analyze boosting algorithms [Ann. Statist. 29 (2001) 1189–1232; Ann. Statist. 28 (2000) 337–407; Ann. Statist. 32 (2004) 407–499] in linear regression from a new perspective: that of modern first-order methods in convex optimization. We show that classic boosting algorithms in linear regression, namely the incremental forward stagewise algorithm ($\text{FS}_{\varepsilon}$) and least squares boosting [LS-BOOST$(\varepsilon)$], can be viewed as subgradient descent to minimize the loss function defined as the maximum absolute correlation between the features and residuals. We also propose a minor modification of $\text{FS}_{\varepsilon}$ that yields an algorithm for the LASSO, and that may be easily extended to an algorithm that computes the LASSO path for different values of the regularization parameter. Furthermore, we show that these new algorithms for the LASSO may also be interpreted as the same master algorithm (subgradient descent), applied to a regularized version of the maximum absolute correlation loss function. We derive novel, comprehensive computational guarantees for several boosting algorithms in linear regression (including LS-BOOST$(\varepsilon)$ and $\text{FS}_{\varepsilon}$) by using techniques of first-order methods in convex optimization. Our computational guarantees inform us about the statistical properties of boosting algorithms. In particular, they provide, for the first time, a precise theoretical description of the amount of data-fidelity and regularization imparted by running a boosting algorithm with a prespecified learning rate for a fixed but arbitrary number of iterations, for any dataset.

In many practical applications, the objective function is convex. The use of convex objective functions makes optimization easier, but ubiquity of such objective function is a mystery: many practical optimization problems are not easy to solve, so it is not clear why the objective function—whose main goal is to describe our needs—would always describe easier-to-achieve goals. In this paper, we explain this ubiquity based on the fundamental ideas about human decision making. This explanation also helps us explain why in decision making under uncertainty, people often make pessimistic decisions, i.e., decisions based more on the worst-case scenarios.