ACO The ACO seminar (2001-2002)

For 2001-2002, the ACO seminar was organized by Giacomo Zambelli and Amitabh Sinha.

The seminar is in Wean 4601 on Monday's from 12:30pm to 1:30pm.

Fall 2001 Speakers:

October 15, 2001: Geoff Atkinson.

October 29, 2001: Nick Wormald.

November 12, 2001: Amitabh Sinha.

November 19, 2001: Cliff Smyth.

November 26, 2001: Konstantin Andreev.

December 3, 2001: Artur Czumaj.

December 10, 2001: Shang-Hua Teng.

Spring 2002 Speakers:

February 11, 2002: Jason Fulman.

February 18, 2002: Thomas A. Bohman.

February 25, 2002: Kent Hoj Andersen.

March 4, 2002: Venkatesh Natarajan.

March 18, 2002: Michael Perregaard.

March 25, 2002: Alan Frieze.

April 15, 2002: Lisa K. Fleischer.

April 22, 2002: Jochen Konemann.

May 6, 2002: Earl Glen Whitehead, Jr.

Geoff Atkinson, October 15, 2001.

Title: Several Results on the b-independence of random graphs.

Abstract: A set of vertices from a graph is independent if, for any two vertices in the set, there are no edges between those vertices. This idea has been widely studied in graph theory, and several results are known about the size of the largest independent set in a random graph (Bollobas & Erdos, Matula, Frieze, Freize & Luczak). A b-independent set is a set of vertices such that for any two vertices in the set, there are no paths of length b or less between those vertices. I will prove two results on the largest b-independent number of a random graph. First, I will consider the case b is a constant in the random graph, Gnp, where each edge is independently included in the graph with probability p. Second, I will consider the case b=2 in a random regular graph.

Nick Wormald, University of Melbourne, October 29, 2001.

Title: Models of random regular graphs.

Abstract: This is a survey of results on properties of random regular graphs, together with an exposition of some of the main methods of obtaining these results. A major feature in this area is the small subgraph conditioning method. This establishes a relationship between random regular graphs with uniform distribution, and other non-uniform models of random regular graphs in which the probability of a graph occurring is weighted according to some random variable. Information can be obtained in this way on the probability of existence of various types of spanning subgraphs, such as Hamilton cycles and decompositions into perfect matchings, as well as other properties such as smallest dominating set. Some results about interesting non-uniform models appear as spin-offs from the same method.

Amitabh Sinha, November 12, 2001.

Title: Approximating k-cuts via network strength.

Abstract: Given an undirected edge-weighted connected graph, the k-cut problem is to partition the vertex set into k non-empty connected components so as to minimize the total weight of edges whose end points are in different components. The problem is NP-hard. In the first half of this talk, we will briefly look at some earlier work on this problem (Goldschmidt & Hochbaum, 1988; Karger & Stein, 1996; Saran & Vazirani, 1995; Naor & Rabani, 2001). In the second half, we will examine a recent approach which uses polynomial time algorithms for network strength (Cunningham, 1985) to approximate an optimal k-cut. This is joint work with R. Ravi.

Clifford Smyth, School of Mathematics, Institute for Advanced Study, Princeton. November 19, 2001.

Title: Reimer's inequality and Rudich's conjecture.

Abstract: We prove Rudich's conjecture, 1984, which arose from his work on reducibility among cryptographic primitives. Roughly the conjecture states that if a family of subcubes of the discrete cube is a near partition, then there is a significant dependence among the family. The subcubes are viewed as events where a point is selected from the cube according to a product distribution. To prove this we use a correlation inequality that is in some sense "dual" to Reimer's inequality (a.k.a. the van den Berg-Kesten conjecture). We also use this inequality to prove an upper bound on the approximate decision tree complexity of a boolean function that is quadratic in its approximate certificate complexity, answering a question of Tardos, 1989. This is joint work with Jeff Kahn and Mike Saks.

Konstantin Andreev, November 26, 2001.

Title: Error correcting codes using expander graphs.

Abstract: Expander graphs have been the subject of much study in combinatorics and computer science. Recently they were used in a series of constructions of asymptotically good error correcting codes. This is a survey of results on error correcting codes that use expander graphs. First we will look at Expander codes. These are linear time decodable codes introduced by D. Spielman and M. Sipser. Then we will look at a recent work by M. Mitzenmacher, M. Luby and D. Spielman which is a hybrid between Expander codes and low density parity check codes which enhances the error correction rate and moves closer to the Shannon capacity bound.

Artur Czumaj, NJIT, December 3, 2001.

Title: Selfish Traffic Allocation for Server Farms.

Abstract: We study the price of selfish routing in non-cooperative networks, like the Internet. We investigate the notion of the worst-case coordination ratio in networks as recently introduced by Koutsoupias and Papadimitriou. The analysis of the coordination ratio seeks the price of uncoordinated individual utility-maximizing decisions (``the price of anarchy''). Koutsoupias and Papadimitriou proposed that model to study routing problems in which a set of several agents want to send traffic from the sources to destinations along parallel links with linear cost functions.

In this talk, we will first consider linear cost functions, and then we will discuss general, monotone cost functions as well as families of cost functions motivated by queueing theory. Our special focus lies on cost functions that describe the behavior of Web servers that can open only a limited number of TCP connections.

For the worst-case coordination ratio in the simplest system consisting of m parallel links with linear cost functions and show that it is Theta ( log(m) / logloglog(m) ). Our bound is asymptotically tight and it resolves an open question posed recently by Koutsoupias and Papadimitriou . We also obtain a tight bound in the special case with identical links (i.e., all of the same speed), and show that it is (1+o(1)) log(m) / loglog(m) .

Then, we consider arbitrary cost functions. Our results can be summarized as follows.

* We give an exact characterization of those cost functions that have bounded/unbounded coordination ratio. For example, we show that polynomial functions have bounded ratio whereas exponential functions have unbounded ratio.

* We present the first models and results that take into account that Internet traffic is not homogeneous and session lengths are not exponentially distributed. In particular, we assume that the lengths of TCP sessions follow general probability distributions and different data streams may use different distributions.

* We show that monotone queueing theoretical cost functions behave extremely bad under game theoretic measures. In fact, the increase in cost due to selfish behavior is unbounded. Moreover, under bicriteria measures selfish behavior can lead to a bandwidth degradation by a factor m, where m is the total number of servers (or parallel links).

* Finally, we consider server systems that reject packets in case of overload. Again we obtain an unbounded coordination ratio on the fraction of rejected requests. On the positive side, however, we show that the fraction of rejections can be bounded above by a small term that is polynomial in the number of servers and even exponential in the number of TCP connections that can be opened simultaneously.

Joint work with Piotr Krysta and Berthold Voecking (MPI Saarbruecken, Germany)

Shang-Hua Teng, UIUC / BU / Akamai, December 10, 2001.

Title: Why the Simplex Algorithm Usually takes Polynomial Time.

Abstract: Theorists have been challenged by the existence of remarkable algorithms that are known by scientists and engineers to work well in practice, but whose theoretical analyses are negative or inconclusive. The root of the problem is that algorithms are usually analyzed in one of two ways: by worst-case or average-case analysis. The former can improperly suggest that an algorithm will perform poorly, while the latter can be overly optimistic because the random inputs it considers may bear little resemblance to those encountered in practice.

We propose an analysis that we call smoothed analysis that can help explain the success of many algorithms that both worst-case and average-case cannot. In smoothed analysis, we measure the performance of an algorithm under slight random perturbations of arbitrary inputs. In particular, we consider Gaussian perturbations of inputs to algorithms that take real and complex inputs, and we measure the running time of algorithms in terms of the input size and the variance of the perturbations.

We show that the simplex algorithm has polynomial smoothed complexity. The simplex algorithm is the classic example of an algorithm that performs well in practice but takes exponential time in the worst case. In particular, we prove for every matrix $A$ with entries of absolute value at most $1$, every vector $c$, and every vector $b$ whose entries are $1$ or $-1$, we show that the simplex algorithm using the shadow-vertex pivot rule takes expected time polynomial in $1/sigma$ and the sizes of $A$ and $c$ to solve minimize c'x subject to (A+ \sigma G) x <=b, where $\sigma G$ is a matrix of Gaussian random variables of mean 0 and variance $\sigma^2$.

This is joint work with Dan Spielman (MIT).

Jason Fulman, University of Pittsburgh, February 11, 2002.

Title: GL(n,q) and increasing subsequences in nonuniform random permutations.

Abstract: Motivated by random matrices over finite fields, we study random integer partitions. Exploiting connections with identities from q-series, we derive bounds on the length of the longest increasing subsequence in nonuniform random permutations. This is different from the traditional methods (which we shall survey) for studying the longest increasing subsequence.

Thomas A. Bohman, Carnegie Mellon University, February 18, 2002.

Title: Linear versus Hereditary Discrepancy.

Abstract: The concept of hypergraph discrepancy provides a unified combinatorial framework for handling classical discrepancy problems from geometry and number theory. Since the discrepancy of a hypergraph can be small for somewhat arbitrary reasons, variants of hypergraph discrepancy have been introduced in the hopes of getting a better measure of the `intrinsic balance' of a hypergraph. Two such variants are linear discrepancy and hereditary discrepancy. Lovasz, Spencer and Vesztergombi proved that the linear discrepancy of a hypergraph H is bounded above by the hereditary discrepancy of H, and conjectured a sharper bound that involves the number of vertices in H. In this talk we give a short proof of this conjecture for hypergraphs of hereditary discrepancy 1. For hypergraphs of higher hereditary discrepancy we give some partial results and propose a sharpening of the conjecture.

This is joint work with Ron Holzman.

Kent Hoj Andersen, February 25, 2002.

Title: Split Closure and Intersection Cuts.

Abstract: In the seventies, Balas introduced intersection cuts for a Mixed Integer Program (MILP), and showed that these cuts can be obtained by a closed form formula from a basis of the standard linear programming relaxation. In the early nineties, Cook, Kannan and Schrijver introduced the split closure of an MILP, and showed that the split closure is a polyhedron. In this paper, we show that the split closure can be obtained using only intersection cuts. We give two different proofs this result, one geometric and one algebraic. Furthermore, the result is used to provide a new proof of the fact that the split closure is a polyhedron. Finally, we extend the result to more general two-term disjunctions.

This is joint work with Gerard Cournuejols and Yanjun Li

Venkatesh Natarjan, March 4, 2002.

Title: Edge Colorings for a Class of Planar Multigraphs.

Abstract: In this talk, we present a simple lower bound for the edge-chromatic number of any multigraph, which is conjectured to be tight for any planar graph. This talk focusses on the work of Odile Marcotte, who showed this bound to be tight for all graphs excluding K_{3,3} minors and K_5 - e minors.

Michael Perregaard, March 18, 2002.

Title: Generating Disjunctive Cuts for Mixed Integer Programs

Abstract: Disjunctive cuts is a very broad class of cuts for mixed integer programming. In general, any cut that can be derived from a disjunctive argument can be considered a disjunctive cut. Here we consider specifically cuts that are valid inequalities for some simple disjunctive relaxation of the mixed integer program. Such a relaxation can e.g. be obtained by relaxing the integrality condition on all but a single variable. The lift-and-project procedure developed in the early nineties is a systematic way to generate an optimal (in a specific sense) disjunctive cut for a given disjunctive relaxation. It involves solving a higher dimensional cut generating linear program (CGLP) and has been developed for the simplest possible disjunctions; those requiring that a simple variable be either zero or one. In our work we consider the problem of efficiently generating disjunctive cuts for any given disjunction. That is, once we are presented with a disjunctive relaxation of a mixed integer program, how can we efficiently generate one or more cuts that cuts off an optimal solution to the LP relaxation? This problem naturally falls into two cases: Two-term disjunctions, as those the original lift-and-project procedure was designed to solve, and more general multiple-term disjunctions. For the two-term disjunctions we show how one can effectively reduced the CGLP, but the main result is that we show a precise correspondence between the lift-and-project cuts obtained from the CGLP and simple disjunctive cuts from rows of the LP relaxation simplex tableau. The implication is that lift-and-project cuts from the high dimensional CGLP can be obtained directly from the LP relaxation. Furthermore, if integrality on all variables are considered then this becomes a correspondence between strengthened lift-and-project cuts and Gomory's mixed integer cuts. Using this correspondence we present a procedure to efficiently generate an optimal mixed integer Gomory cut (optimal in the sense of the CGLP) through pivots in the simplex tableau of the LP relaxation. In the case of multiple-term disjunctions we present procedures that provide an optimal solution to the high dimensional CGLP, by solving the cut problem in the original space without recourse to the many auxiliary variables present in the CGLP. Finally, we propose a procedure that generates a set of disjunctive cuts for a given disjunction.

Alan Frieze, March 25, 2002.

Title: On graph irregularity strength.

Abstract: An assignment of positive integer weights to the edges of a simple graph $G$ is called irregular if the weighted degrees of the vertices are all different. The irregularity strength, $s(G)$, is the maximal weight, minimized over all irregular assignments, and is set to $\infty$ if no such assignment is possible. In this paper we show, that $s(G)\leq c_1 n/\d$, for graphs with maximum degree $\D\le n^{1/2}$, and $s(G)\leq c_2 (\log n) n/\d$, for graphs with $\D >n^{1/2}$, where $c_1$ and $c_2$ are explicit constants. These bounds substantially improve previously known results. To prove the results, we are using a combination of deterministic and probabilistic techniques.

Lisa K. Fleisher, April 15, 2002.

Title: The Quickest Multicommodity Flow Problem (or Techniques for Flows Over Time).

Abstract: Standard network flows capture many aspects of routing problems but miss one crucial element: flows occur over time. Ford and Fulkerson recognized this omission when they introduced flows over time. In this model, there is a transit time, or delay, associated with each arc to address the fact that flow does not travel instantaneously across the arc. Now, given a time bound, it is possible to consider the flow-over-time equivalents of standard network flow problems. This allows for more accurate modelling of problems in telecommunications, transportation, financial planning, manufacturing, and logistics. Ford and Fulkerson showed how to solve the maximum flow over time problem by reducing it to a minimum cost flow problem. However, the minimum cost flow over time problem is already NP-hard. And, the obvious linear program formulation of the multicommodity flow over time problem depends linearly on the time bound, so is in general not of polynomial size. Thus, problems in flows over time are inherently more difficult than the standard flow problems. We will introduce flows over time, review some of the classical results, and describe a newly devised FPAS for both the multicommodity flow over time problem and the same problem with costs that is joint work with Martin Skutella. Time permitting we may touch upon further work.

Jochen Konemann, April 22, 2002.

Title: Non-clairvoyant Scheduling for Minimizing Mean Slowdown.

Abstract: We consider the problem of scheduling jobs online non-clairvoyantly, that is, when job sizes are not known. Our focus is on minimizing mean {\em slowdown}, defined as the ratio of response time to the size of the job. We use {\em resource augmentation} in terms of allowing a faster processor to the online algorithm to make up for its lack of knowledge of job sizes. We provide several upper and lower bounds for the problem. For $\epsilon > 0$, our main result is an $O((1+\epsilon)\log_{1+\epsilon} B)$-speed $O(\log^2_{1+\epsilon} B)$-competitive algorithm for minimizing mean slowdown non-clairvoyantly, when $B$ is the ratio between the largest and smallest job sizes. We also prove that any $k$-speed algorithm incurs a slowdown of at least $\Omega(\log B/k)$, thus providing evidence that our upper bound is non-trivial. We show that in the static case (that is, when all jobs have the same arrival time), the Round Robin algorithm is able to match this lower bound. A further set of lower bounds motivates the need for bounded job sizes as well as resource augmentation for any algorithm to minimize slowdown non-clairvoyantly.

This is joint work with N. Bansal, K. Dhamdhere, and A. Sinha.

Earl Glen Whitehead, Jr., University of Pittsburgh, May 6, 2002.

Title: Chromatic polynomials and Tutte polynomials of homeomorphism classes of graphs.

Abstract: First we will discuss a polynomial which contains, as special cases, the chromatic polynomial of all members of a given homeomorphism class of graphs. Then we will discuss a related polynomial which contains, as special cases, the Tutte polynomial of all members of a given homeomorphism class of graphs. As a historical note, we will discuss the earlier work of John Nagle, Department of Physics, CMU.

The is joint work with Ronald C. Read, University of Waterloo. Professor Read and the speaker published the following papers on this topic. (1) Discrete Math. 204 (1999), 337-356. (2) Discrete Math 243 (2002), 267-272.

Past Seminars:

Back to the ACO home page