Polynomial-time algorithms for linear programming based only on primal scaling and projected gradients of a potential function

This paper presents extensions and further analytical properties of algorithms for linear programming based only on primal scaling and projected gradients of a potential function. The paper contains extensions and analysis of two polynomial-time algorithms for linear programming. We first present an extension of Gonzaga's O(nL) iteration algorithm, that computes dual variables and does not assume a known optimal objective function value. This algorithm uses only affine scaling, and is based on computing the projected gradient of the potential function $$q\ln (x^T s) - \sum\limits_{j = 1}^n {\ln (x_j )} $$ wherex is the vector of primal variables ands is the vector of dual slack variables, and q = n + \(\sqrt n \) . The algorithm takes either a primal step or recomputes dual variables at each iteration. We next present an alternate form of Ye's O( \(\sqrt n \) L) iteration algorithm, that is an extension of the first algorithm of the paper, but uses the potential function $$q\ln (x^T s) - \sum\limits_{j = 1}^n {\ln (x_j ) - \sum\limits_{j - 1}^n {\ln (s_j )} } $$ where q = n + \(\sqrt n \) . We use this alternate form of Ye's algorithm to show that Ye's algorithm is optimal with respect to the choice of the parameterq in the following sense. Suppose thatq = n + n ^t wheret?0. Then the algorithm will solve the linear program in O(n ^r L) iterations, wherer = max{t, 1 ? t}. Thus the value oft that minimizes the complexity bound ist = 1/2, yielding Ye's O( \(\sqrt n \) L) iteration bound.     

Limiting behavior of the affine scaling continuous trajectories for linear programming problems

We consider the continuous trajectories of the vector field induced by the primal affine scaling algorithm as applied to linear programming problems in standard form. By characterizing these trajectories as solutions of certain parametrized logarithmic barrier families of problems, we show that these trajectories tend to an optimal solution which in general depends on the starting point. By considering the trajectories that arise from the Lagrangian multipliers of the above mentioned logarithmic barrier families of problems, we show that the trajectories of the dual estimates associated with the affine scaling trajectories converge to the so called centered optimal solution of the dual problem. We also present results related to asymptotic direction of the affine scaling trajectories. We briefly discuss how to apply our results to linear programs formulated in formats different from the standard form. Finally, we extend the results to the primal-dual affine scaling algorithm.     

The primal power affine scaling method

In this paper, we present a variant of the primal affine scaling method, which we call the primal power affine scaling method. This method is defined by choosing a realr>0.5, and is similar to the power barrier variant of the primal-dual homotopy methods considered by den Hertog, Roos and Terlaky and Sheu and Fang. Here, we analyze the methods forr>1. The analysis for 0.50<r<1 is similar, and can be readily carried out with minor modifications. Under the non-degeneracy assumption, we show that the method converges for any choice of the step size . To analyze the convergence without the non-degeneracy assumption, we define a power center of a polytope. We use the connection of the computation of the power center by Newton's method and the steps of the method to generalize the 2/3rd result of Tsuchiya and Muramatsu. We show that with a constant step size such that /(1-)^2r > 2/(2r-1) and with a variable asymptotic step size k uniformly bounded away from 2/(2r+1), the primal sequence converges to the relative interior of the optimal primal face, and the dual sequence converges to the power center of the optimal dual face. We also present an accelerated version of the method. We show that the two-step superlieear convergence rate of the method is 1+r/(r+1), while the three-step convergence rate is 1+ 3r/(r+2). Using the measure of Ostrowski, we note thet the three-step method forr=4 is more efficient than the two-step quadratically convergent method, which is the limit of the two-step method asr approaches infinity.Partially supported by the grant CCR-9321550 from NSF.     

Probability of unique integer solution to a system of linear equations

We consider a system of m linear equations in n variables Ax = d and give necessary and sufficient conditions for the existence of a unique solution to the system that is integer: x ∈ {−1, 1}ⁿ. We achieve this by reformulating the problem as a linear program and deriving necessary and sufficient conditions for the integer solution to be the unique primal optimal solution. We show that as long as m is larger than n/2, then the linear programming reformulation succeeds for most instances, but if m is less than n/2, the reformulation fails on most instances. We also demonstrate that these predictions match the empirical performance of the linear programming formulation to very high accuracy.     

Convergence analysis of the projective scaling algorithm based on a long-step homogeneous affine scaling algorithm

In this paper we give a new convergence analysis of a projective scaling algorithm. We consider a long-step affine scaling algorithm applied to a homogeneous linear programming problem obtained from the original linear programming problem. This algorithm takes a fixed fraction λ≤2/3 of the way towards the boundary of the nonnegative orthant at each iteration. The iteration sequence for the original problem is obtained by pulling back the homogeneous iterates onto the original feasible region with a conical projection, which generates the same search direction as the original projective scaling algorithm at each iterate. The recent convergence results for the long-step affine scaling algorithm by the authors are applied to this algorithm to obtain some convergence results on the projective scaling algorithm. Specifically, we will show (i) polynomiality of the algorithm with complexities of O(nL) and O(n ² L) iterations for λ<2/3 and λ=2/3, respectively; (ii) global covnergence of the algorithm when the optimal face is unbounded; (iii) convergence of the primal iterates to a relative interior point of the optimal face; (iv) convergence of the dual estimates to the analytic center of the dual optimal face; and (v) convergence of the reduction rate of the objective function value to 1−λ.     

The p-median problem under uncertainty

Consider the need to currently locate p facilities but it is possible that up to q additional facilities will have to be located in the future. There are known probabilities that 0 ? r ? q facilities will need to be located. The p-median problem under uncertainty is to find the location of p facilities such that the expected value of the objective function in the future is minimized. The problem is formulated on a graph, properties of it are proven, an integer programming formulation is constructed, and heuristic algorithms are suggested for its solution. The heuristic algorithms are modified to reduce the run time by about two orders of magnitude with minimal effect on the quality of the solution. Optimal solutions for many problems are found effectively by CPLEX. Computational results using the heuristic algorithms are presented.     

Maximizing throughput in queueing networks with limited flexibility

We study a queueing network where customers go through several stages of processing, with the class of a customer used to indicate the stage of processing. The customers are serviced by a set of flexible servers, i.e., a server is capable of serving more than one class of customers and the sets of classes that the servers are capable of serving may overlap. We would like to choose an assignment of servers that achieves the maximal capacity of the given queueing network, where the maximal capacity is λ^∗ if the network can be stabilized for all arrival rates λ < λ^∗ and cannot possibly be stabilized for all λ > λ^∗. We examine the situation where there is a restriction on the number of servers that are able to serve a class, and reduce the maximal capacity objective to a maximum throughput allocation problem of independent interest: the total discrete capacity constrained problem (TDCCP). We prove that solving TDCCP is in general NP-complete, but we also give exact or approximation algorithms for several important special cases and discuss the implications for building limited flexibility into a system.     

Primal cutting plane algorithms revisited

Dual fractional cutting plane algorithms, in which cutting planes are used to iteratively tighten a linear relaxation of an integer program, are well-known and form the basis of the highly successful branch-and-cut method. It is rather less well-known that various primal cutting plane algorithms were developed in the 1960s, for example by Young. In a primal algorithm, the main role of the cutting planes is to enable a feasible solution to the original problem to be improved. Research on these algorithms has been almost non-existent.  In this paper we argue for a re-examination of these primal methods. We describe a new primal algorithm for pure 0-1 problems based on strong valid inequalities and give some encouraging computational results. Possible extensions to the case of general mixed-integer programs are also discussed.     

The s-monotone index selection rules for pivot algorithms of linear programming

In this paper we introduce the concept of s-monotone index selection rule for linear programming problems. We show that several known anti-cycling pivot rules like the minimal index, Last-In–First-Out and the most-often-selected-variable pivot rules are s-monotone index selection rules. Furthermore, we show a possible way to define new s-monotone pivot rules. We prove that several known algorithms like the primal (dual) simplex, MBU-simplex algorithms and criss-cross algorithm with s-monotone pivot rules are finite methods.     

Haplotyping populations by pure parsimony based on compatible genotypes and greedy heuristics

The population haplotype inference problem based on the pure parsimony criterion (HIPP) infers an m × n genotype matrix for a population by a 2m × n haplotype matrix with the minimum number of distinct haplotypes. Previous integer programming based HIPP solution methods are time-consuming, and their practical effectiveness remains unevaluated. On the other hand, previous heuristic HIPP algorithms are efficient, but their theoretical effectiveness in terms of optimality gaps has not been evaluated, either. We propose two new heuristic HIPP algorithms (MGP and GHI) and conduct more complete computational experiments. In particular, MGP exploits the compatible relations among genotypes to solve a reduced integer linear programming problem so that a solution of good quality can be obtained very quickly; GHI exploits a weight mechanism to selects better candidate haplotypes in a greedy fashion. The computational results show that our proposed algorithms are efficient and effective, especially for solving cases with larger recombination rates.     

Competitive online scheduling of perfectly malleable jobs with setup times

We study how to efficiently schedule online perfectly malleable parallel jobs with arbitrary arrival times on m ? 2 processors. We take into account both the linear speedup of such jobs and their setup time, i.e., the time to create, dispatch, and destroy multiple processes. Specifically, we define the execution time of a job with length p_j running on k_j processors to be p_j/k_j + (k_j − 1)c, where c > 0 is a constant setup time associated with each processor that is used to parallelize the computation. This formulation accurately models data parallelism in scientific computations and realistically asserts a relationship between job length and the maximum useful degree of parallelism. When the goal is to minimize makespan, we show that the online algorithm that simply assigns k_j so that the execution time of each job is minimized and starts jobs as early as possible has competitive ratio 4(m − 1)/m for even m ? 2 and 4m/(m + 1) for odd m ? 3. This algorithm is much simpler than previous offline algorithms for scheduling malleable jobs that require more than a constant number of passes through the job list.     

Average behavior of greedy algorithms for the minimization knapsack problem: General coefficient distributions

For the minimization knapsack problem with Boolean variables, primal and dual greedy algorithms are formally described. Their relations to the corresponding algorithms for the maximization knapsack problem are shown. The average behavior of primal and dual algorithms for the minimization problem is analyzed. It is assumed that the coefficients of the objective function and the constraint are independent identically distributed random variables on [0, 1] with an arbitrary distribution having a density and that the right-hand side d is deterministic and proportional to the number of variables (i.e., d = μn). A condition on μ is found under which the primal and dual greedy algorithms have an asymptotic error of t.     

Primal interior-point decomposition algorithms for two-stage stochastic extended second-order cone programming

ABSTRACT

We study and solve the two-stage stochastic extended second-order cone programming problem. We show that the barrier recourse functions and the composite barrier functions for this optimization problem are self-concordant families with respect to barrier parameters. These results are used to develop primal decomposition-based interior-point algorithms. The worst case iteration complexity of the developed algorithms is shown to be the same as that for the short- and long-step primal interior algorithms applied to the extensive formulation of our problem.     

Convergence of the Dual Variables for the Primal Affine Scaling Method with Unit Steps in the Homogeneous Case

In this paper, we investigate the behavior of the primal affine scaling method with unit steps when applied to the case where b=0 and c>0. We prove that the method is globally convergent and that the dual iterates converge to the analytic center of the dual feasible region.     

Scaling limits of loop-erased random walks and uniform spanning trees

The uniform spanning tree (UST) and the loop-erased random walk (LERW) are strongly related probabilistic processes. We consider the limits of these models on a fine grid in the plane, as the mesh goes to zero. Although the existence of scaling limits is still unproven, subsequential scaling limits can be defined in various ways, and do exist. We establish some basic a.s. properties of these subsequential scaling limits in the plane. It is proved that any LERW subsequential scaling limit is a simple path, and that the trunk of any UST subsequential scaling limit is a topological tree, which is dense in the plane. The scaling limits of these processes are conjectured to be conformally invariant in dimension 2. We make a precise statement of the conformal invariance conjecture for the LERW, and show that this conjecture implies an explicit construction of the scaling limit, as follows. Consider the Löwner differential equation 1 $\frac{{\partial f}}{{\partial t}} = z\frac{{\zeta (t) + z}}{{\zeta (t) - z}}\frac{{\partial f}}{{\partial z}}$ , with boundary valuesf(z,0)=z, in the rangez ∈U= {w ∈ ? : ?w? < 1},t≤0. We choose ζ(t):=B(?2t), where B(t) is Brownian motion on ? $ \mathbb{U} The uniform spanning tree (UST) and the loop-erased random walk (LERW) are strongly related probabilistic processes. We consider the limits of these models on a fine grid in the plane, as the mesh goes to zero. Although the existence of scaling limits is still unproven, subsequential scaling limits can be defined in various ways, and do exist. We establish some basic a.s. properties of these subsequential scaling limits in the plane. It is proved that any LERW subsequential scaling limit is a simple path, and that the trunk of any UST subsequential scaling limit is a topological tree, which is dense in the plane. The scaling limits of these processes are conjectured to be conformally invariant in dimension 2. We make a precise statement of the conformal invariance conjecture for the LERW, and show that this conjecture implies an explicit construction of the scaling limit, as follows. Consider the L?wner differential equation

(1)

, with boundary valuesf(z,0)=z, in the rangez ∈U= {w ∈ ℂ : •w• < 1},t≤0. We choose ζ(t):=B(−2t), where B(t) is Brownian motion on ∂ starting at a random-uniform point in ∂ . Assuming the conformal invariance of the LERW scaling limit in the plane, we prove that the scaling limit of LERW from 0 to ∂ has the same law as that of the pathf(ζ(t),t) (wheref(z,t) is extended continuously to ∂ ) ×(−∞,0]). We believe that a variation of this process gives the scaling limit of the boundary of macroscopic critical percolation clusters. Research supported by the Sam and Ayala Zacks Professorial Chair.     

Multi-degree cyclic scheduling of a no-wait robotic cell with multiple robots

This paper addresses cyclic scheduling of a no-wait robotic cell with multiple robots. In contrast to many previous studies, we consider r-degree cyclic (r > 1) schedules, in which r identical parts with constant processing times enter and leave the cell in each cycle. We propose an algorithm to find the minimal number of robots for all feasible r-degree cycle times for a given r (r > 1). Consequently, the optimal r-degree cycle time for any given number of robots for this given r can be obtained with the algorithm. To develop the algorithm, we first show that if the entering times of r parts, relative to the start of a cycle, and the cycle time are fixed, minimizing the number of robots for the corresponding r-degree schedule can be transformed into an assignment problem. We then demonstrate that the cost matrix for the assignment problem changes only at some special values of the cycle time and the part entering times, and identify all special values for them. We solve our problem by enumerating all possible cost matrices for the assignment problem, which is subsequently accomplished by enumerating intervals for the cycle time and linear functions of the part entering times due to the identification of the special values. The algorithm developed is shown to be polynomial in the number of machines for a fixed r, but exponential if r is arbitrary.     

Nonsmooth minimax programming under locally Lipschitz (Φ, ρ)-invexity

Minimax programming problems involving locally Lipschitz (Φ, ρ)-invex functions are considered. The parametric and non-parametric necessary and sufficient optimality conditions for a class of nonsmooth minimax programming problems are obtained under nondifferentiable (Φ, ρ)-invexity assumption imposed on objective and constraint functions. When the sufficient conditions are utilized, parametric and non-parametric dual problems in the sense of Mond-Weir and Wolfe may be formulated and duality results are derived for the considered nonsmooth minimax programming problem. With the reference to the said functions we extend some results of optimality and duality for a larger class of nonsmooth minimax programming problems.     

Comparative analysis of affine scaling algorithms based on simplifying assumptions

We analyze several affine potential reduction algorithms for linear programming based on simplifying assumptions. We show that, under a strong probabilistic assumption regarding the distribution of the data in an iteration, the decrease in the primal potential function will be with high probability, compared to the guaranteed(1). ( 2n is a parameter in the potential function andn is the number of variables.) Under the same assumption, we further show that the objective reduction rate of Dikin's affine scaling algorithm is with high probability, compared to no guaranteed convergence rate.Research supported in part by NSF Grant DDM-8922636.