NOTE: This website is no longer maintained, please visit my current website at http://www.cs.rit.edu/~ib.

## Ivona Bezakova

Ph.D. student
Department of Computer Science
University of Chicago

E-mail: firstname @ cs.uchicago.edu

Papers:

• Sampling Binary Contingency Tables with a Greedy Start,
with Nayantara Bhatnagar and Eric Vigoda.
Submitted to Random Structures & Algorithms. Extended abstract appears in SODA '06.

We study the problem of counting and randomly sampling binary contingency tables. For given row and column sums, we are interested in approximately counting (or sampling) 0/1 n x m matrices with specified row/column sums. We present a simulated annealing algorithm with running time O((nm)^2D^3d_{max}log^5(n+m)) for any row/column sums where D is the number of non-zero entries and d_{max} is the maximum row/column sum. This is the first algorithm to directly solve binary contingency tables for all row/column sums. Previous work reduced the problem to the permanent, or restricted attention to row/column sums that are close to regular. The interesting aspect of our simulated annealing algorithm is that it starts at a non-trivial instance, whose solution relies on the existence of short alternating paths in the graph constructed by a particular Greedy algorithm.

• Accelerating Simulated Annealing Algorithm for the Permanent and Combinatorial Counting Problems,
with Daniel Stefankovic, Vijay V. Vazirani, and Eric Vigoda.
Submitted to SIAM Journal on Computing (SICOMP). Extended abstract appears in SODA '06.
Powerpoint slides are here: [ppt].

We present an improved "cooling schedule" for simulated annealing algorithms for combinatorial counting problems. Under our new schedule the rate of cooling accelerates as the temperature decreases. Thus, fewer intermediate temperatures are needed as the simulated annealing algorithm moves from the high temperature (easy region) to the low temperature (difficult region). We present applications of our technique to colorings and the permanent (perfect matchings of bipartite graphs). Moreover, for the permanent, we improve the analysis of the Markov chain underlying the simulated annealing algorithm. This improved analysis, combined with the faster cooling schedule, results in an O(n^7log^4{n}) time algorithm for approximating the permanent of a 0/1 matrix.

• Graph Model Selection using the Maximum Likelihood Paradigm,
with Adam Kalai, and Rahul Santhanam.
Appears in ICML '06.
A more networks-oriented version is here. Preprint.

In recent years, there has been a proliferation of theoretical graph models, e.g., preferential attachment and small-world models, motivated by real-world graphs such as the Internet topology. To address the natural question of which model is best for a particular data set, we propose a model selection criterion for graph models. Since each model is in fact a probability distribution over graphs (equivalently a compression scheme), we suggest using the Maximum Likelihood (ML) principle to compare graph models and select their parameters. Unfortunately, for the case of graph models, computing this metric is a difficult algorithmic task. However, we design and implement MCMC algorithms for computing the maximum likelihood for four popular models: a power-law random graph model, a preferential attachment model, a small-world model, and a uniform random graph model. We hope that this novel use of ML and our new algorithms will find other applications.

• Analysis of Sequential Importance Sampling for Contingency Tables,
with Alistair Sinclair, Daniel Stefankovic, and Eric Vigoda.
Appears in ESA '06.

The sequential importance sampling (SIS) algorithm has gained considerable popularity for its empirical success. One of its noted applications is to the binary contingency tables problem, an important problem in statistics, where the goal is to estimate the number of \$0/1\$ matrices with prescribed row and column sums. We give a family of examples in which the SIS procedure, if run for any subexponential number of trials, will underestimate the number of tables by an exponential factor. This result holds for any of the usual design choices in the SIS algorithm, namely the ordering of the columns and rows. These are apparently the first theoretical results on the efficiency of the SIS algorithm for binary contingency tables. Finally, we present experimental evidence that the SIS algorithm is efficient for row and column sums that are regular. Our work is a first step in determining the class of inputs for which SIS is effective.

• Allocating Indivisible Goods,
with Varsha Dani.
Preprint. Survey of results invited for publication in ACM SIGecom Exchanges, Vol. 5.3, 2005.

The Max-Min Fairness problem is as follows: Given m indivisible goods and k players, each with a specified valuation function on the subsets of the goods, how should the goods be split between the players so as to maximize the minimum valuation. Viewing the problem from a game theoretic perspective, we show that for two players and additive valuations the expected minimum of the (randomized) cut-and-choose mechanism is a 1/2-approximation of the optimum. To complement this result we show that no truthful mechanism can compute the exact optimum. We also consider the algorithmic perspective when the (true) additive valuation functions are part of the input. We present a simple 1/(m-k+1) approximation algorithm which allocates to every player at least 1/k fraction of the value of all but the k-1 heaviest items and we give an algorithm with additive error against the fractional optimum bounded by the value of the largest item. The two approximation algorithms are incomparable in the sense that there exist instances when one outperforms the other.

• The Poincare Constant of a Random Walk in High-Dimensional Convex Bodies,
Master Thesis at the University of Chicago, 2002.

Estimating the volume of a convex body is an important algorithmic problem. High dimensions pose an especially difficult task for the algorithm designer. To be considered efficient, the running time must be polynomial in the dimension. The first provably efficient approximation algorithm was presented by Dyer, Frieze, and Kannan in 1989, and steadily improved by various authors over the next decade. Fundamental to these works is the analysis of random walks for generating random points from a convex body. Kannan, Lovasz, and Simonovits recently analyzed a natural random walk known as the ball walk. By bounding the so-called conductance, they obtain (via a Cheeger-type inequality) a bound on the Poincare constant of the random walk. Their bound proves the walk converges to a random point in time O*(n^3D^2), where n is the dimension and D is the diameter of the body. We survey and present a self-contained view of their work. In addition, following an outline proposed by Jerrum, we slightly modify the KLS analysis to bound the Poincare constant without using a Cheeger-type inequality.

• Compact Representations of Graphs and Adjacency Testing,
"Magister" Thesis at the Comenius University, Bratislava, Slovakia, 2000.
A section of the thesis appeared at the Student Science Conference, Bratislava, Slovakia, 2000.

A succinct representation of a graph is a widely studied problem. We examine this representation in two aspects: (1) its space complexity, and (2) the time complexity of the basic "graph operations" - the adjacency test and update operations (adding or removing an edge). The adjacency matrix is optimal with respect to the adjacency test but for n vertices its space complexity is always O(n^2) bits, regardless of the number of edges m. The adjacency list representation is space-optimal but the adjacency test takes more time than it necessarily needs. The aim of this thesis was to find a representation as close as possible to the optimal space complexity of O(m log n) bits, and to the optimal time complexity of O(log n) bit-operations for the adjacency test (and the update operations). We present a representation using the technique of hashing: it takes O(m log n) bits and the adjacency test can be performed in O(log n loglog n) bit-operations. The "list-of-all-neighbors" procedure takes O((k + loglog n)log n) bit-operations, where k is the number of neighbors to be listed, and the expected amortized time complexity of the update operations is O(log n loglog n). We also suggest a new representation based on computing the distance from all vertices to some given vertex. We show that this representation is optimal (in both time and space complexities) for the family of planar graphs.

• Planar Finite Automata,
with Martin Pal.
Student Science Conference, Bratislava, Slovakia, 1999.

Several aspects of graph representations of finite automata have been studied in the literature, e.g., the complexity considerations involving the number of states. It is natural to ask whether we indeed need the full generality of possible interconnections among the states provided by the general definition of finite automata. We study the influence of restricting the interconnection network to planar graphs. We introduce planar finite automata as a type of automata whose graph representation is planar. We study families of languages defined by deterministic, nondeterministic, and epsilon-free nondeterministic planar finite automata and compare them to R, the family of regular languages. We show that the deterministic planar automata are weaker than finite automata, except for the case of a one-letter alphabet. We also prove that R is equivalent to the family of languages accepted by (epsilon-free) nondeterministic planar automata. In addition, we study the influence of restricting the graph representation to a particular architecture studied in parallel and distributed computing, the d-dimensional mesh. We show that for any k there is a language which cannot be accepted by a mesh k-epsilon-bounded nondeterministic finite automaton, i.e., an automaton whose graph representation can be embedded into a mesh and its maximal number of epsilon moves is bounded by k.

Course Projects:

Last Update: May 30, 2006