Georgia Institute of Technology College of Sciences

Consider the problem of selling items to a unit-demand buyer. Most work on maximizing seller revenue considers either a setting that is single dimensional, such as where the items are identical, or multi-dimensional, where the items are heterogeneous. With respect to revenue-optimal mechanisms, these settings sit at extreme ends of a spectrum: from simple and fully characterized (single-dimensional) to complex and nebulous (multi-dimensional). In this paper, we identify a setting that sits in between these extremes. We consider a seller who has three services {A,B,C} for sale to a single buyer with a value v and an interest G from {A,B,C}, and there is a known partial ordering over the services. For example, suppose the seller is selling {internet}, {internet, phone}, and {internet, cable tv}. A buyer with interest {internet} would be satisfied by receiving phone or cable tv in addition, but a customer whose interest is {internet, phone} cannot be satisfied by any other option. Thus this corresponds to a partial-ordering where {internet} > {internet, phone} and {internet} > {internet, cable tv}, but {internet, phone} and {internet, cable tv} are not comparable. We show formally that partially-ordered items lie in a space of their own, in between identical and heterogeneous items: there exist distributions over (value, interest) pairs for three partially-ordered items such that the menu complexity of the optimal mechanism is unbounded, yet for all distributions there exists an optimal mechanism of finite menu complexity. So this setting is vastly more complex than identical items (where the menu complexity is one), or even “totally-ordered” items as in the FedEx Problem [FGKK16] (where the menu complexity is at most seven, for three items), yet drastically more structured than heterogeneous items (where the menu complexity can be uncountable [DDT15]). We achieve this result by proving a characterization of the class of best duals and by giving a primal recovery algorithm which obtains the optimal mechanism. In addition, we (1) extend our lower-bound to the Multi-Unit Pricing setting, (2) give a tighter and deterministic characterization of the optimal mechanism when the buyer’s distribution satisfies the declining marginal revenue condition, and (3) prove a master theorem that allows us to reason about duals instead of distributions. Joint work with Nikhil Devanur, Raghuvansh Saxena, Ariel Schvartzman, and Matt Weinberg.

A low-diameter decomposition (LDD) of an undirected graph G is a partitioning of G into components of bounded diameter, such that only a small fraction of original edges are between the components. This decomposition has played instrumental role in the design of low-stretch spanning tree, spanners, distributed algorithms etc. A natural question is whether such a decomposition can be efficiently maintained/updated as G undergoes insertions/deletions of edges. We make the first step towards answering this question by designing a fully-dynamic graph algorithm that maintains an LDD in sub-linear update time. It is known that any undirected graph G admits a spanning tree T with nearly logarithmic average stretch, which can be computed in nearly linear-time. This tree decomposition underlies many recent progress in static algorithms for combinatorial and scientific flows. Using our dynamic LDD algorithm, we present the first non-trivial algorithm that dynamically maintains a low-stretch spanning tree in \tilde{O}(t^2) amortized update time, while achieving (t + \sqrt{n^{1+o(1)}/t}) stretch, for every 1 \leq t \leq n. Joint work with Sebastian Krinninger.

Physical sensors (thermal, light, motion, etc.) are becoming ubiquitous and offer important benefits to society. However, allowing sensors into our private spaces has resulted in considerable privacy concerns. Differential privacy has been developed to help alleviate these privacy concerns. In this talk, we’ll develop and define a framework for releasing physical data that preserves both utility and provides privacy. Our notion of closeness of physical data will be defined via the Earth Mover Distance and we’ll discuss the implications of this choice. Physical data, such as temperature distributions, are often only accessible to us via a linear transformation of the data. We’ll analyse the implications of our privacy definition for linear inverse problems, focusing on those that are traditionally considered to be "ill-conditioned”. We’ll then instantiate our framework with the heat kernel on graphs and discuss how the privacy parameter relates to the connectivity of the graph. Our work indicates that it is possible to produce locally private sensor measurements that both keep the exact locations of the heat sources private and permit recovery of the ``general geographic vicinity'' of the sources. Joint work with Anna C. Gilbert.

Studying random samples drawn from large, complex sets is one way to begin to learn about typical properties and behaviors. However, it is important that the samples examined are random enough: studying samples that are unexpectedly correlated or drawn from the wrong distribution can produce misleading conclusions. Sampling processes using Markov chains have been utilized in physics, chemistry, and computer science, among other fields, but they are often applied without careful analysis of their reliability. Making sure widely-used sampling processes produce reliably representative samples is a main focus of my research, and in this talk I'll touch on two specific applications from statistical physics and combinatorics.I'll also discuss work applying these same Markov chain processes used for sampling in a novel way to address research questions in programmable matter and swarm robotics, where a main goal is to understand how simple computational elements can accomplish complicated system-level goals. In a constrained setting, we've answered this question by showing that groups of abstract particles executing our simple processes (which are derived from Markov chains) can provably accomplish remarkable global objectives. In the long run, one goal is to understand the minimum computational abilities elements need in order to exhibit complex global behavior, with an eye towards developing systems where individual components are as simple as possible.This talk includes joint work with Marta Andrés Arroyo, Joshua J. Daymude, Daniel I. Goldman, David A. Levin, Shengkai Li, Dana Randall, Andréa Richa, William Savoie, Alexandre Stauffer, and Ross Warkentin.