Search

The rise of algorithmic pricing raises concerns of algorithmic collusion. We conduct experiments with algorithmic pricing agents based on Large Language Models (LLMs). We find that (1) LLM-based agents are adept at pricing tasks, (2) LLM-based pricing agents autonomously collude in oligopoly settings to the detriment of consumers, and (3) variation in seemingly innocuous phrases in LLM instructions ("prompts") may increase collusion. Novel off-path analysis techniques uncover price-war concerns as contributing to these phenomena. Our results extend to auction settings. Our findings uncover unique challenges to any future regulation of LLM-based pricing agents, and black-box pricing agents more broadly.

We study a matching problem between agents and public goods, in settings without monetary transfers. Since goods are public, they have no capacity constraints. There is no exogenously defined budget of goods to be provided. Rather, each provided good must justify its cost, leading to strong complementarities in the "preferences" of goods. Furthermore, goods that are in high demand given other already-provided goods must also be provided. The question of the existence of a stable solution (a menu of public goods to be provided) exhibits a rich combinatorial structure. We uncover sufficient conditions and necessary conditions for guaranteeing the existence of a stable solution, and derive both positive and negative results for strategyproof stable matching.

Traditionally, social choice theory has only been applicable to choices among a few predetermined alternatives but not to more complex decisions such as collectively selecting a textual statement. We introduce generative social choice, a framework that combines the mathematical rigor of social choice theory with large language models' capability to generate text and extrapolate preferences. This framework divides the design of AI-augmented democratic processes into two components: first, proving that the process satisfies rigorous representation guarantees when given access to oracle queries; second, empirically validating that these queries can be approximately implemented using a large language model. We illustrate this framework by applying it to the problem of generating a slate of statements that is representative of opinions expressed as free-form text, for instance in an online deliberative process.

In a 2019 paper, Barnard and Steinerberger show that for $f\in L^1(\mathbf{R})$, the following autocorrelation inequality holds: \begin{equation*} \min_{0 \leq t \leq 1} \int_\mathbf{R} f(x) f(x+t)\ \mathrm{d}x \ \leq\ 0.411 ||f||_{L^1}^2, \end{equation*} where the constant $0.411$ cannot be replaced by $0.37$. In addition to being interesting and important in their own right, inequalities such as these have applications in additive combinatorics where some problems, such as those of minimal difference basis, can be encapsulated by a convolution inequality similar to the above integral. Barnard and Steinerberger suggest that future research may focus on the existence of functions extremizing the above inequality (which is itself related to Brascamp-Lieb type inequalities). We show that for $f$ to be extremal under the above, we must have \begin{equation*} \max_{x_1 \in \mathbf{R} }\min_{0 \leq t \leq 1} \left[ f(x_1-t)+f(x_1+t) \right] \ \leq\ \min_{x_2 \in \mathbf{R} } \max_{0 \leq t \leq 1} \left[ f(x_2-t)+f(x_2+t) \right] . \end{equation*} Our central technique for deriving this result is local perturbation of $f$ to increase the value of the autocorrelation, while leaving $||f||_{L^1}$ unchanged. These perturbation methods can be extended to examine a more general notion of autocorrelation. Let $d,n \in \mathbb{Z}^+$, $f \in L^1$, $A$ be a $d \times n$ matrix with real entries and columns $a_i$ for $1 \leq i \leq n$, and $C$ be a constant. For a broad class of matrices $A$, we prove necessary conditions for $f$ to extremize autocorrelation inequalities of the form \begin{equation*} \min_{ \mathbf{t} \in [0,1]^d } \int_{\mathbf{R}} \prod_{i=1}^n\ f(x+ \mathbf{t} \cdot a_i)\ \mathrm{d}x\ \leq\ C ||f||_{L^1}^n. \end{equation*}, Comment: 9 pages

We study the problem of crescent configurations, posed by Erd\H{o}s in 1989. A crescent configuration is a set of $n$ points in the plane such that: 1) no three points lie on a common line, 2) no four points lie on a common circle, 3) for each $1 \leq i \leq n - 1$, there exists a distance which occurs exactly $i$ times. Constructions of sizes $n \leq 8$ have been provided by Liu, Pal\'{a}sti, and Pomerance. Erd\H{o}s conjectured that there exists some $N$ for which there do not exist crescent configurations of size $n$ for all $n \geq N$. We extend the problem of crescent configurations to general normed spaces $(\mathbb{R}^2, \| \cdot \|)$ by studying strong crescent configurations in $\| \cdot \|$. In an arbitrary norm $\|\cdot \|$, we construct a strong crescent configuration of size 4. We also construct larger strong crescent configurations in the Euclidean, taxicab, and Chebyshev norms, of sizes $n \leq 6$, $n \leq 8$, and $n \leq 8$ respectively. When defining strong crescent configurations, we introduce the notion of line-like configurations in $\|\cdot \|$. A line-like configuration in $\|\cdot \|$ is a set of points whose distance graph is isomorphic to the distance graph of equally spaced points on a line. In a broad class of norms, we construct line-like configurations of arbitrary size. Our main result is a crescent-type result about line-like configurations in the Chebyshev norm. A line-like crescent configuration is a line-like configuration for which no three points lie on a common line and no four points lie on a common $\|\cdot \|$ circle. We prove that for $n \geq 7$, every line-like crescent configuration of size $n$ in the Chebyshev norm must have a rigid structure. Specifically, it must be a perpendicular perturbation of equally spaced points on a horizontal or vertical line., Comment: 27 pages, 15 figures

We construct finite sets of real numbers that have a small difference set and strong local properties. In particular, we construct a set $A$ of $n$ real numbers such that $|A-A|=n^{\log_2 3}$ and that every subset $A'\subseteq A$ of size $k$ satisfies $|A'-A'|\ge k^{\log_2 3}$. This construction leads to the first non-trivial upper bound for the problem of distinct distances with local properties.

The local properties problem of Erd\H{o}s and Shelah generalizes many Ramsey problems and some distinct distances problems. In this work, we derive a variety of new bounds for the local properties problem and its variants. We do this by continuing to develop the color energy technique --- a variant of the concept of additive energy from Additive Combinatorics. In particular, we generalize the concept of color energy to higher color energies, and combine these with Extremal Graph Theory results about graphs with no cycles or subdivisions of size $k$.