All issues

This paper is devoted to some variants of improving the convergence rate guarantees of the gradient-type algorithms for relatively smooth and relatively Lipschitz-continuous problems in the case of additional information about some analogues of the strong convexity of the objective function. We consider two classes of problems, namely, convex problems with a relative functional growth condition, and problems (generally, non-convex) with an analogue of the Polyak – Lojasiewicz gradient dominance condition with respect to Bregman divergence. For the first type of problems, we propose two restart schemes for the gradient type methods and justify theoretical estimates of the convergence of two algorithms with adaptively chosen parameters corresponding to the relative smoothness or Lipschitz property of the objective function. The first of these algorithms is simpler in terms of the stopping criterion from the iteration, but for this algorithm, the near-optimal computational guarantees are justified only on the class of relatively Lipschitz-continuous problems. The restart procedure of another algorithm, in its turn, allowed us to obtain more universal theoretical results. We proved a near-optimal estimate of the complexity on the class of convex relatively Lipschitz continuous problems with a functional growth condition. We also obtained linear convergence rate guarantees on the class of relatively smooth problems with a functional growth condition. For a class of problems with an analogue of the gradient dominance condition with respect to the Bregman divergence, estimates of the quality of the output solution were obtained using adaptively selected parameters. We also present the results of some computational experiments illustrating the performance of the methods for the second approach at the conclusion of the paper. As examples, we considered a linear inverse Poisson problem (minimizing the Kullback – Leibler divergence), its regularized version which allows guaranteeing a relative strong convexity of the objective function, as well as an example of a relatively smooth and relatively strongly convex problem. In particular, calculations show that a relatively strongly convex function may not satisfy the relative variant of the gradient dominance condition.

The article is devoted to studying the application of convex optimization methods to solve the Cauchy problem for the Helmholtz equation, which is ill-posed since the equation belongs to the elliptic type. The Cauchy problem is formulated as an inverse problem and is reduced to a convex optimization problem in a Hilbert space. The functional to be optimized and its gradient are calculated using the solution of boundary value problems, which, in turn, are well-posed and can be approximately solved by standard numerical methods, such as finite-difference schemes and Fourier series expansions. The convergence of the applied fast gradient method and the quality of the solution obtained in this way are experimentally investigated. The experiment shows that the accelerated gradient method — the Similar Triangle Method — converges faster than the non-accelerated method. Theorems on the computational complexity of the resulting algorithms are formulated and proved. It is found that Fourier’s series expansions are better than finite-difference schemes in terms of the speed of calculations and improve the quality of the solution obtained. An attempt was made to use restarts of the Similar Triangle Method after halving the residual of the functional. In this case, the convergence does not improve, which confirms the absence of strong convexity. The experiments show that the inaccuracy of the calculations is more adequately described by the additive concept of the noise in the first-order oracle. This factor limits the achievable quality of the solution, but the error does not accumulate. According to the results obtained, the use of accelerated gradient optimization methods can be the way to solve inverse problems effectively.

We consider strongly-convex-strongly-concave saddle-point problems with general non-bilinear objective and different condition numbers with respect to the primal and dual variables. First, we consider such problems with smooth composite terms, one of which has finite-sum structure. For this setting we propose a variance reduction algorithm with complexity estimates superior to the existing bounds in the literature. Second, we consider finite-sum saddle-point problems with composite terms and propose several algorithms depending on the properties of the composite terms. When the composite terms are smooth we obtain better complexity bounds than the ones in the literature, including the bounds of a recently proposed nearly-optimal algorithms which do not consider the composite structure of the problem. If the composite terms are prox-friendly, we propose a variance reduction algorithm that, on the one hand, is accelerated compared to existing variance reduction algorithms and, on the other hand, provides in the composite setting similar complexity bounds to the nearly-optimal algorithm which is designed for noncomposite setting. Besides, our algorithms allow one to separate the complexity bounds, i. e. estimate, for each part of the objective separately, the number of oracle calls that is sufficient to achieve a given accuracy. This is important since different parts can have different arithmetic complexity of the oracle, and it is desired to call expensive oracles less often than cheap oracles. The key thing to all these results is our general framework for saddle-point problems, which may be of independent interest. This framework, in turn is based on our proposed Accelerated Meta-Algorithm for composite optimization with probabilistic inexact oracles and probabilistic inexactness in the proximal mapping, which may be of independent interest as well.

The article is devoted to first-order adaptive methods for optimization problems with relatively strongly convex functionals. The concept of relatively strong convexity significantly extends the classical concept of convexity by replacing the Euclidean norm in the definition by the distance in a more general sense (more precisely, by Bregman’s divergence). An important feature of the considered classes of problems is the reduced requirements concerting the level of smoothness of objective functionals. More precisely, we consider relatively smooth and relatively Lipschitz-continuous objective functionals, which allows us to apply the proposed techniques for solving many applied problems, such as the intersection of the ellipsoids problem (IEP), the Support Vector Machine (SVM) for a binary classification problem, etc. If the objective functional is convex, the condition of relatively strong convexity can be satisfied using the problem regularization. In this work, we propose adaptive gradient-type methods for optimization problems with relatively strongly convex and relatively Lipschitzcontinuous functionals for the first time. Further, we propose universal methods for relatively strongly convex optimization problems. This technique is based on introducing an artificial inaccuracy into the optimization model, so the proposed methods can be applied both to the case of relatively smooth and relatively Lipschitz-continuous functionals. Additionally, we demonstrate the optimality of the proposed universal gradient-type methods up to the multiplication by a constant for both classes of relatively strongly convex problems. Also, we show how to apply the technique of restarts of the mirror descent algorithm to solve relatively Lipschitz-continuous optimization problems. Moreover, we prove the optimal estimate of the rate of convergence of such a technique. Also, we present the results of numerical experiments to compare the performance of the proposed methods.

Distributed saddle point problems (SPPs) have numerous applications in optimization, matrix games and machine learning. For example, the training of generated adversarial networks is represented as a min-max optimization problem, and training regularized linear models can be reformulated as an SPP as well. This paper studies distributed nonsmooth SPPs with Lipschitz-continuous objective functions. The objective function is represented as a sum of several components that are distributed between groups of computational nodes. The nodes, or agents, exchange information through some communication network that may be centralized or decentralized. A centralized network has a universal information aggregator (a server, or master node) that directly communicates to each of the agents and therefore can coordinate the optimization process. In a decentralized network, all the nodes are equal, the server node is not present, and each agent only communicates to its immediate neighbors.

We assume that each of the nodes locally holds its objective and can compute its value at given points, i. e. has access to zero-order oracle. Zero-order information is used when the gradient of the function is costly, not possible to compute or when the function is not differentiable. For example, in reinforcement learning one needs to generate a trajectory to evaluate the current policy. This policy evaluation process can be interpreted as the computation of the function value. We propose an approach that uses a smoothing technique, i. e., applies a first-order method to the smoothed version of the initial function. It can be shown that the stochastic gradient of the smoothed function can be viewed as a random two-point gradient approximation of the initial function. Smoothing approaches have been studied for distributed zero-order minimization, and our paper generalizes the smoothing technique on SPPs.

Non-smooth optimization often arises in many applied problems. The issues of developing efficient computational procedures for such problems in high-dimensional spaces are very topical. First-order methods (subgradient methods) are well applicable here, but in fairly general situations they lead to low speed guarantees for large-scale problems. One of the approaches to this type of problem can be to identify a subclass of non-smooth problems that allow relatively optimistic results on the rate of convergence. For example, one of the options for additional assumptions can be the condition of a sharp minimum, proposed in the late 1960s by B. T. Polyak. In the case of the availability of information about the minimal value of the function for Lipschitz-continuous problems with a sharp minimum, it turned out to be possible to propose a subgradient method with a Polyak step-size, which guarantees a linear rate of convergence in the argument. This approach made it possible to cover a number of important applied problems (for example, the problem of projecting onto a convex compact set). However, both the condition of the availability of the minimal value of the function and the condition of a sharp minimum itself look rather restrictive. In this regard, in this paper, we propose a generalized condition for a sharp minimum, somewhat similar to the inexact oracle proposed recently by Devolder – Glineur – Nesterov. The proposed approach makes it possible to extend the class of applicability of subgradient methods with the Polyak step-size, to the situation of inexact information about the value of the minimum, as well as the unknown Lipschitz constant of the objective function. Moreover, the use of local analogs of the global characteristics of the objective function makes it possible to apply the results of this type to wider classes of problems. We show the possibility of applying the proposed approach to strongly convex nonsmooth problems, also, we make an experimental comparison with the known optimal subgradient method for such a class of problems. Moreover, there were obtained some results connected to the applicability of the proposed technique to some types of problems with convexity relaxations: the recently proposed notion of weak $\beta$-quasi-convexity and ordinary quasiconvexity. Also in the paper, we study a generalization of the described technique to the situation with the assumption that the $\delta$-subgradient of the objective function is available instead of the usual subgradient. For one of the considered methods, conditions are found under which, in practice, it is possible to escape the projection of the considered iterative sequence onto the feasible set of the problem.

In the first part of the paper we present convergence analysis of AGMsDR method on a new class of functions — in general non-convex with $M$-Lipschitz-continuous gradients that satisfy Polyak – Lojasiewicz condition. Method does not need the value of $\mu^{PL}>0$ in the condition and converges linearly with a scale factor $\left(1 - \frac{\mu^{PL}}{M}\right)$. It was previously proved that method converges as $O\left(\frac1{k^2}\right)$ if a function is convex and has $M$-Lipschitz-continuous gradient and converges linearly with a~scale factor $\left(1 - \sqrt{\frac{\mu^{SC}}{M}}\right)$ if the value of strong convexity parameter $\mu^{SC}>0$ is known. The novelty is that one can save linear convergence if $\frac{\mu^{PL}}{\mu^{SC}}$ is not known, but without square root in the scale factor.

The second part presents modification of AGMsDR method for solving problems that allow alternating minimization (Alternating AGMsDR). The similar results are proved.

As the result, we present adaptive accelerated methods that converge as $O\left(\min\left\lbrace\frac{M}{k^2},\,\left(1-{\frac{\mu^{PL}}{M}}\right)^{(k-1)}\right\rbrace\right)$ on a class of convex functions with $M$-Lipschitz-continuous gradient that satisfy Polyak – Lojasiewicz condition. Algorithms do not need values of $M$ and $\mu^{PL}$. If Polyak – Lojasiewicz condition does not hold, the convergence is $O\left(\frac1{k^2}\right)$, but no tuning needed.

We also consider the adaptive catalyst envelope of non-accelerated gradient methods. The envelope allows acceleration up to $O\left(\frac1{k^2}\right)$. We present numerical comparison of non-accelerated adaptive gradient descent which is accelerated using adaptive catalyst envelope with AGMsDR, Alternating AGMsDR, APDAGD (Adaptive Primal-Dual Accelerated Gradient Descent) and Sinkhorn's algorithm on the problem dual to the optimal transport problem.

Conducted experiments show faster convergence of alternating AGMsDR in comparison with described catalyst approach and AGMsDR, despite the same asymptotic rate $O\left(\frac1{k^2}\right)$. Such behavior can be explained by linear convergence of AGMsDR method and was tested on quadratic functions. Alternating AGMsDR demonstrated better performance in comparison with AGMsDR.