Limited-memory BFGS with displacement aggregation

A displacement aggregation strategy is proposed for the curvature pairs stored in a limited-memory BFGS (a.k.a. L-BFGS) method such that the resulting (inverse) Hessian approximations are equal to those that would be derived from a full-memory BFGS method. This means that, if a sufficiently large number of pairs are stored, then an optimization algorithm employing the limited-memory method can achieve the same theoretical convergence properties as when full-memory (inverse) Hessian approximations are stored and employed, such as a local superlinear rate of convergence under assumptions that are common for attaining such guarantees. To the best of our knowledge, this is the first work in which a local superlinear convergence rate guarantee is offered by a quasi-Newton scheme that does not either store all curvature pairs throughout the entire run of the optimization algorithm or store an explicit (inverse) Hessian approximation. Numerical results are presented to show that displacement aggregation within an adaptive L-BFGS scheme can lead to better performance than standard L-BFGS.

     

Extra-Updates Criterion for the Limited Memory BFGS Algorithm for Large Scale Nonlinear Optimizatio

This paper studies recent modifications of the limited memory BFGS (L-BFGS) method for solving large scale unconstrained optimization problems. Each modification technique attempts to improve the quality of the L-BFGS Hessian by employing (extra) updates in a certain sense. Because at some iterations these updates might be redundant or worsen the quality of this Hessian, this paper proposes an updates criterion to measure this quality. Hence, extra updates are employed only to improve the poor approximation of the L-BFGS Hessian. The presented numerical results illustrate the usefulness of this criterion and show that extra updates improve the performance of the L-BFGS method substantially.     

A regularized limited memory BFGS method for nonconvex unconstrained minimization

The limited memory BFGS method (L-BFGS) is an adaptation of the BFGS method for large-scale unconstrained optimization. However, The L-BFGS method need not converge for nonconvex objective functions and it is inefficient on highly ill-conditioned problems. In this paper, we proposed a regularization strategy on the L-BFGS method, where the used regularization parameter may play a compensation role in some sense when the condition number of Hessian approximation tends to become ill-conditioned. Then we proposed a regularized L-BFGS method and established its global convergence even when the objective function is nonconvex. Numerical results show that the proposed method is efficient.     

Enriched Methods for Large-Scale Unconstrained Optimization

This paper describes a class of optimization methods that interlace iterations of the limited memory BFGS method (L-BFGS) and a Hessian-free Newton method (HFN) in such a way that the information collected by one type of iteration improves the performance of the other. Curvature information about the objective function is stored in the form of a limited memory matrix, and plays the dual role of preconditioning the inner conjugate gradient iteration in the HFN method and of providing an initial matrix for L-BFGS iterations. The lengths of the L-BFGS and HFN cycles are adjusted dynamically during the course of the optimization. Numerical experiments indicate that the new algorithms are both effective and not sensitive to the choice of parameters.     

The Global Convergence of Self-Scaling BFGS Algorithm with Nonmonotone Line Search for Unconstrained Nonconvex Optimization Problems

The self-scaling quasi-Newton method solves an unconstrained optimization problem by scaling the Hessian approximation matrix before it is updated at each iteration to avoid the possible large eigenvalues in the Hessian approximation matrices of the objective function. It has been proved in the literature that this method has the global and superlinear convergence when the objective function is convex （or even uniformly convex）. We propose to solve unconstrained nonconvex optimization problems by a self-scaling BFGS algorithm with nonmonotone linear search. Nonmonotone line search has been recognized in numerical practices as a competitive approach for solving large-scale nonlinear problems. We consider two different nonmonotone line search forms and study the global convergence of these nonmonotone self-scale BFGS algorithms. We prove that, under some weaker condition than that in the literature, both forms of the self-scaling BFGS algorithm are globally convergent for unconstrained nonconvex optimization problems.     

A cautious BFGS update for reduced Hessian SQP

In this paper, we introduce a cautious BFGS (CBFGS) update criterion in the reduced Hessian sequential quadratic programming (SQP) method. An attractive property of this update criterion is that the generated iterative matrices are always positive definite. Under mild conditions, we get the global convergence of the reduced Hessian SQP method. In particular, the second order sufficient condition is not necessary for the global convergence of the method. Furthermore, we show that if the second order sufficient condition holds at an accumulation point, then the reduced Hessian SQP method with CBFGS update reduces to the reduced Hessian SQP method with ordinary BFGS update. Consequently, the local behavior of the proposed method is the same as the reduced Hessian SQP method with BFGS update. The presented preliminary numerical experiments show the good performance of the method. This work was supported by the National Natural Science Foundation of China via grant 10671060 and 10471060.     

Computing the matrix geometric mean: Riemannian versus Euclidean conditioning,implementation techniques,and a Riemannian BFGS method

This paper addresses the problem of computing the Riemannian center of mass of a collection of symmetric positive definite matrices. We show in detail that the condition number of the Riemannian Hessian of the underlying optimization problem is never very ill conditioned in practice, which explains why the Riemannian steepest descent approach has been observed to perform well. We also show theoretically and empirically that this property is not shared by the Euclidean Hessian. We then present a limited‐memory Riemannian BFGS method to handle this computational task. We also provide methods to produce efficient numerical representations of geometric objects that are required for Riemannian optimization methods on the manifold of symmetric positive definite matrices. Through empirical results and a computational complexity analysis, we demonstrate the robust behavior of the limited‐memory Riemannian BFGS method and the efficiency of our implementation when compared to state‐of‐the‐art algorithms.     

A subspace implementation of quasi-Newton trust region methods for unconstrained optimization

This paper studies subspace properties of trust region methods for unconstrained optimization, assuming the approximate Hessian is updated by quasi- Newton formulae and the initial Hessian approximation is appropriately chosen. It is shown that the trial step obtained by solving the trust region subproblem is in the subspace spanned by all the gradient vectors computed. Thus, the trial step can be defined by minimizing the quasi-Newton quadratic model in the subspace. Based on this observation, some subspace trust region algorithms are proposed and numerical results are also reported.     

On efficiently combining limited-memory and trust-region techniques

Limited-memory quasi-Newton methods and trust-region methods represent two efficient approaches used for solving unconstrained optimization problems. A straightforward combination of them deteriorates the efficiency of the former approach, especially in the case of large-scale problems. For this reason, the limited-memory methods are usually combined with a line search. We show how to efficiently combine limited-memory and trust-region techniques. One of our approaches is based on the eigenvalue decomposition of the limited-memory quasi-Newton approximation of the Hessian matrix. The decomposition allows for finding a nearly-exact solution to the trust-region subproblem defined by the Euclidean norm with an insignificant computational overhead as compared with the cost of computing the quasi-Newton direction in line-search limited-memory methods. The other approach is based on two new eigenvalue-based norms. The advantage of the new norms is that the trust-region subproblem is separable and each of the smaller subproblems is easy to solve. We show that our eigenvalue-based limited-memory trust-region methods are globally convergent. Moreover, we propose improved versions of the existing limited-memory trust-region algorithms. The presented results of numerical experiments demonstrate the efficiency of our approach which is competitive with line-search versions of the L-BFGS method.     

On the limited memory BFGS method for large scale optimization

We study the numerical performance of a limited memory quasi-Newton method for large scale optimization, which we call the L-BFGS method. We compare its performance with that of the method developed by Buckley and LeNir (1985), which combines cycles of BFGS steps and conjugate direction steps. Our numerical tests indicate that the L-BFGS method is faster than the method of Buckley and LeNir, and is better able to use additional storage to accelerate convergence. We show that the L-BFGS method can be greatly accelerated by means of a simple scaling. We then compare the L-BFGS method with the partitioned quasi-Newton method of Griewank and Toint (1982a). The results show that, for some problems, the partitioned quasi-Newton method is clearly superior to the L-BFGS method. However we find that for other problems the L-BFGS method is very competitive due to its low iteration cost. We also study the convergence properties of the L-BFGS method, and prove global convergence on uniformly convex problems.This work was supported by the Applied Mathematical Sciences subprogram of the Office of Energy Research, U.S. Department of Energy, under contract DE-FG02-87ER25047, and by National Science Foundation Grant No. DCR-86-02071.     

成组BFGS修正的紧凑形式与成组记忆修正算法

１．引言实际问题中经常要遇到一族函数极小值问题的求解,即ｍｉｎｆｉ（ｘ）,ｉ＝１,...,Ｐ;（１．１）其中人：Ｒ"、Ｒ具有公共的Ｈｅｓｓｉａｎ矩阵Ｇ（ｘ）。７'ｆｉ（ｘ）,ｒ是适中的数值．如在各种负载下的弹性体研究中,即要遇到问题（ｌ．Ｉ）的求解,其中人（Ｃ）一人Ｃ）＋ｑＣ十Ｃ;（＝１,...,...．对于不同的比则人（Ｘ）具有不同的极小点和不同的梯度Ｄ人（Ｘ）,但具有相同的Ｈｅｓｓｉａｎ矩阵Ｇ（Ｘ）．１９９４年,Ｏ'Ｌｅａｒｙ等【'］把拟一Ｎｅｗｔｏｎ算法推广至成组形式（ｍｕｌｔｉＰｌｅｖｅｒｓｉｏ...,…     

An example of numerical nonconvergence of a variable-metric method

An example is given where truncation error, caused by finite computer arithmetic, leads to the BFGS variable-metric method becoming stuck, despite the approximated Hessian matrix, the gradient vector, and the search vector satisfying analytical conditions for convergence. A restart criterion to eliminate the problem is suggested.     

On quasi-Newton and pseudo-Newton algorithms

This paper surveys some of the existing approaches to quasi-Newton methods and introduces a new way for constructing inverse Hessian approximations for such algorithms. This new approach is based on restricting Newton's method to subspaces over which the inverse Hessian is assumed to be known, while expanding this subspace using gradient information. It is shown that this approach can lead to some well-known formulas for updating the inverse Hessian approximation. Deriving such updates through this approach provides new understanding of these formulas and their relation to the pseudo-Newton-Raphson algorithm.     

A Quasi-Newton Quadratic Penalty Method for Minimization Subject to Nonlinear Equality Constraints

We present a modified quadratic penalty function method for equality constrained optimization problems. The pivotal feature of our algorithm is that at every iterate we invoke a special change of variables to improve the ability of the algorithm to follow the constraint level sets. This change of variables gives rise to a suitable block diagonal approximation to the Hessian which is then used to construct a quasi-Newton method. We show that the complete algorithm is globally convergent. Preliminary computational results are reported.     

A second-order globally convergent direct-search method and its worst-case complexity

Direct-search algorithms form one of the main classes of algorithms for smooth unconstrained derivative-free optimization, due to their simplicity and their well-established convergence results. They proceed by iteratively looking for improvement along some vectors or directions. In the presence of smoothness, first-order global convergence comes from the ability of the vectors to approximate the steepest descent direction, which can be quantified by a first-order criticality (cosine) measure. The use of a set of vectors with a positive cosine measure together with the imposition of a sufficient decrease condition to accept new iterates leads to a convergence result as well as a worst-case complexity bound. In this paper, we present a second-order study of a general class of direct-search methods. We start by proving a weak second-order convergence result related to a criticality measure defined along the directions used throughout the iterations. Extensions of this result to obtain a true second-order optimality one are discussed, one possibility being a method using approximate Hessian eigenvectors as directions (which is proved to be truly second-order globally convergent). Numerically guaranteeing such a convergence can be rather expensive to ensure, as it is indicated by the worst-case complexity analysis provided in this paper, but turns out to be appropriate for some pathological examples.     

A New Diagonal Quasi-Newton Updating Method With Scaled Forward Finite Differences Directional Derivative for Unconstrained Optimization

A new diagonal quasi-Newton updating algorithm for unconstrained optimization is presented. The elements of the diagonal matrix approximating the Hessian are determined as scaled forward finite differences directional derivatives of the components of the gradient. Under mild classical assumptions, the convergence of the algorithm is proved to be linear. Numerical experiments with 80 unconstrained optimization test problems, of different structures and complexities, as well as five applications from MINPACK-2 collection, prove that the suggested algorithm is more efficient and more robust than the quasi-Newton diagonal algorithm retaining only the diagonal elements of the BFGS update, than the weak quasi-Newton diagonal algorithm, than the quasi-Cauchy diagonal algorithm, than the diagonal approximation of the Hessian by the least-change secant updating strategy and minimizing the trace of the matrix, than the Cauchy with Oren and Luenberger scaling algorithm in its complementary form (i.e. the Barzilai-Borwein algorithm), than the steepest descent algorithm, and than the classical BFGS algorithm. However, our algorithm is inferior to the limited memory BFGS algorithm (L-BFGS).     

Constrained definite Hessians tend to be well conditioned

The testing of optimization algorithms requires the running of problems with ill-conditioned Hessians. For constrained problems, it is the projection of the Hessian onto the space determined by the active constraints that must be ill conditioned. In this note it is argued that unless the Hessian and the constraints are constructed together, the constrained Hessian is likely to be well conditioned. The approach is to examine the effects of random constraints on a singular Hessian.This work was supported in part by the Office of Naval Research under Contract No. N00014-76-C-0391.     

Structure of the Hessian matrix and an economical implementation of Newton’s method in the problem of canonical approximation of tensors

A tensor given by its canonical decomposition is approximated by another tensor (again, in the canonical decomposition) of fixed lower rank. For this problem, the structure of the Hessian matrix of the objective function is analyzed. It is shown that all the auxiliary matrices needed for constructing the quadratic model can be calculated so that the computational effort is a quadratic function of the tensor dimensionality (rather than a cubic function as in earlier publications). An economical version of the trust region Newton method is proposed in which the structure of the Hessian matrix is efficiently used for multiplying this matrix by vectors and for scaling the trust region. At each step, the subproblem of minimizing the quadratic model in the trust region is solved using the preconditioned conjugate gradient method, which is terminated if a negative curvature direction is detected for the Hessian matrix.     

A parameterized hessian quadratic programming problem

We present a general active set algorithm for the solution of a convex quadratic programming problem having a parametrized Hessian matrix. The parametric Hessian matrix is a positive semidefinite Hessian matrix plus a real parameter multiplying a symmetric matrix of rank one or two. The algorithm solves the problem for all parameter values in the open interval upon which the parametric Hessian is positive semidefinite. The algorithm is general in that any of several existing quadratic programming algorithms can be extended in a straightforward manner for the solution of the parametric Hessian problem.This research was supported by the Natural Sciences and Engineering Research Council under Grant No. A8189 and under a Postgraduate Scholarship, by an Ontario Graduate Scholarship, and by the University of Windsor Research Board under Grant No. 9432.     

A parameterized hessian quadratic programming problem

We present a general active set algorithm for the solution of a convex quadratic programming problem having a parametrized Hessian matrix. The parametric Hessian matrix is a positive semidefinite Hessian matrix plus a real parameter multiplying a symmetric matrix of rank one or two. The algorithm solves the problem for all parameter values in the open interval upon which the parametric Hessian is positive semidefinite. The algorithm is general in that any of several existing quadratic programming algorithms can be extended in a straightforward manner for the solution of the parametric Hessian problem. This research was supported by the Natural Sciences and Engineering Research Council under Grant No. A8189 and under a Postgraduate Scholarship, by an Ontario Graduate Scholarship, and by the University of Windsor Research Board under Grant No. 9432.