Andrea Montanari
Large matrices beyond singular value decomposition:
A number of data sets are naturally described in matrix
form. Examples range from microarrays to collaborative
filtering
data. In many of these examples, singular value
decomposition
(SVD)
provides an efficient way to construct a low-rank
approximation
thus achievieng a large dimensionality reduction.
SVD is also an important tool in the design of approximate
linear
algebra algorithms for massive data sets.
It is a recent discovery that --for `generic' matrices--
SVD is sub-optimal, and can be significantly improved upon.
There has been considerable progress on this topic over the
last
year, partly spurred by interest in the Netflic
challenge. I
will overview this progress.
Yun Song
Combinatorial stochastic processes in mathematical population genetics:
Random combinatorial structures arise naturally in mathematical
population genetics. For example, under certain population genetics
models, the probability of observing a sample of DNA sequences is
closely related to the distribution of exchangeable random partitions.
This so-called sampling probability has a number of important
applications in population genetics, but unfortunately closed-form
formulas are rarely known. Because obtaining exact analytic results
is a formidable problem, recent research has focused on developing
sophisticated and computationally-intensive Monte Carlo techniques,
such as MCMC or importance sampling. In this talk, I will review the
aforementioned connection between combinatorial stochastic processes
and population genetics, and describe a technique based on
asymptotic analysis for obtaining novel closed-form sampling formulas.
Shaddin Dughmi
Randomization in Algorithmic Mechanism Design:
Algorithmic mechanism design is concerned with
problems where the inputs are a-priori unknown to the algorithm
designer, and must be elicited from self-interested agents. Designing
algorithms for resource allocation problems that achieve a good
approximation ratio in polynomial time, despite the selfish behavior
of agents holding the inputs, is a central research direction at the
intersection of computer science and economics. Until recently, most
of the known positive and negative results pertain to deterministic
mechanisms. In particular, recent results have indicated that the
power of deterministic truthful mechanisms can be severely bounded.
However, we have recently shown that employing randomization may
bridge the gap between truthful mechanisms and algorithms with full
access to the inputs. I will discuss recent results on the power and
limitations of randomization in algorithmic mechanism design, using
examples such as multi-unit and combinatorial auctions.
Jan Vondrak
Randomized rounding in the matroid polytope:
The question how to round a fractional solution in a matroid polytope
has emerged recently in various settings, ranging from the submodular
welfare maximization and submodular maximization with linear
constraints, to problems like congestion minimization in network
routing, the degree-constrained spanning tree and the asymmetric
traveling salesman problem. In all these applications, it is useful
to have a randomized rounding procedure which preserves the marginal
probabilities on the elements and satisfies strong concentration
bounds for certain functions of the rounded solution. Various
approaches have been proposed in different settings, e.g. the
dependent rounding of Gandhi et al. for bipartite matchings, pipage
rounding of Calinescu et al. for matroids, and maximum entropy
sampling of Asadpour et al. for spanning trees.
I will describe a simple rounding procedure which has the above
properties for any matroid and satisfies Chernoff-type concentration
bounds for linear functions as well as monotone submodular functions.
I will illustrate the usefulness of this technique on various
examples.
Joint work with Chandra Chekuri (UI Urbana-Champaign) and Rico
Zenklusen (ETH Zurich).
Grant Schoenebeck
Unconditional hardness results for semi-definite programs:
Integrality gaps for the Lasserre hierarchy:
Semidefinite programs provide the best approximation algorithms for
many NP-hard combinatorial optimization problems. Recently,
techniques were developed to give unconditional lower bounds for
algorithms based on SDPs.
I will define and give background about semidefinite programming (and
linear programming) hierarchies, which include a large class of SDPs
(and
LPs) including all the most famous SDP algorithms (which actually
occur very low in the hierarchy).
I will then show a lower bound
that
implies no SDP in this large class can refute a random 3-XOR instance
(even though such an instance is very far from satisfiable).
Finally,
I will discuss some corollaries, implications, and open problems.
This result is the first construction of a Lasserre integrality gap,
and simplifies, strengthens, and helps to explain several previous
results.
Parikshit Gopalan
List-decoding using deletions:
A list-decoder for an error-correcting code is an algorithm
that recovers a list of all codewords within a given radius of any
received word. In designing list-decoders, one has to ask: what is the larg=
est radius up to which the list is
guaranteed to be small? A "black-box" approach to this problem is to
apply the classical Johnson bound, which guarantees a small list up to
the Johnson radius. This radius only depends on the minimum distance
and is oblivious to all other information about the code. Thus for
many natural codes, one could hope to do better.
We present a "white-box" approach which for some codes allows
decoding well beyond the Johnson radius. We identify a few "bad" codewords =
and then "delete" them
from the code, thereby improving the minimum distance and now invoke the Jo=
hnson bound. We will apply this template to decoding linear transformations=
, tensor-product codes and
low-degree polynomials.
This talk covers joint work with Venkatesan Guruswami, Adam Klivans,
Prasad Raghavendra and David Zuckerman.