Matchings in bipartite graphs play a foundational role in a variety of mathematical, scientific, and engineering disciplines—from Frobenius’ work on the determinants of matrices, Gale and Shapley’s influential paper on stable marriages and college admissions, Tolstoi and Kantorovich’s work on the optimal transportation problem, to the online world where advertisements are being matched to users billions of times a day. As a consequence, algorithms for finding matchings in bipartite graphs also have a century-old history and the pursuit of efficient algorithms for this problem has led to major achievements in computer science and optimization.
In the 1980s, with the growing availability of parallelizable computer architectures, the study of whether one can parallelize algorithms for fundamental problems gained significant momentum. An efficient parallel algorithm would distribute the work on several processors in a manner that keeps the longest sequence of dependent tasks among processors small—ideally, logarithmic in the size of the problem. Several basic problems such as multiplying matrices, solving a system of equations and computing shortest paths in graphs already had such parallel algorithms.
For the bipartite matching problem, however, it turned out that all algorithms developed so far were inherently sequential in nature and, as such, were not amenable to parallelization. In 1985, Karp, Upfal, and Wigderson3 presented an efficient parallel algorithm for the problem. However, there was a caveat: their algorithm was randomized, that is, it needed access to independent coin tosses. This result was soon followed by a more efficient algorithm by Mulmuley, Vazirani, and Vazirani in 19874 who, using an old algebraic characterization of matchings by Tutte, reduced the problem of computing matchings in graphs to computing determinants of matrices—the latter problem is known to have an efficient parallel algorithm. However, the MVV algorithm, while very different from that of Karp et al., also made crucial use of randomness in its reduction to computing determinants.
This was not the first instance of a problem in which randomness seemed to help—checking whether a number is prime or not was already a notorious example. In 1977, Solovay and Strassen5 discovered a randomized algorithm for primality testing but no efficient algorithm that did not use randomness (called a deterministic algorithm) was discovered until 30 years later. In 1982, Valiant6 showed that a natural routing problem on a network had an efficient randomized algorithm yet every deterministic algorithm for the problem was necessarily inefficient. Research on the power of randomness culminated in a surprising result by Kabanets and Impagliazzo in 20032—removing randomness from certain efficient algorithms can show the non-existence of efficient algorithms themselves—a question that is a whisker away from the P versus NP question.
Thus, understanding the power of randomness in computation very quickly evolved from being a curiosity to being of profound interest in theoretical computer science.
Whether there exists a deterministic algorithm for primality testing or a deterministic parallel algorithm for bipartite matching remained two outstanding questions at the frontiers of our understanding of the role of randomness in computation. The former was solved in 2001 in a remarkable paper by Agrawal, Kayal and Saxena.1 And the latter has been (nearly) recently resolved in the following paper by Fenner, Gurjar and Thierauf.
The authors show the randomized parallel algorithm of MVV can be converted into a deterministic parallel algorithm. At the heart of the MVV approach was an extremely powerful use of randomization—the Isolation Lemma. This lemma asserts there is a randomized algorithm to assign weights to the edges of a bipartite graph such that the minimum-weight perfect matching in the graph is unique—just assign to each edge a weight independently and randomly from a set of integers that is twice the number of edges in the graph. Amazingly, this randomized algorithm does not get to see the graph. Derandomizing the isolation lemma is tantamount to asking the following question—could we also find such a weight assignment when we can, in addition, no longer toss coins?
The main result in this paper is the construction of a list of weight assignments via an almost-polynomial, parallel and deterministic algorithm (which also does not look at the graph) that has the property that for any bipartite graph of a given size, at least one of the weight assignments isolates the minimum weight perfect matching in it. The end result is a gem that comes very close to solving an important open problem, makes an elegant connection between graph theory and derandomization, has been used to make progress on a few other important questions and, as a bonus, the proof is from "The Book."
Join the Discussion (0)
Become a Member or Sign In to Post a Comment