Metric Distortion in Peer Selection (2502.21084v1)

Abstract: In the metric distortion problem, a set of voters and candidates lie in a common metric space, and a committee of $k$ candidates is to be elected. The goal is to select a committee with a small social cost, defined as an increasing function of the distances between voters and selected candidates, but a voting rule only has access to voters' ordinal preferences. The distortion of a rule is then defined as the worst-case ratio between the social cost of the selected set and the optimal set, over all possible preferences and consistent distances. We initiate the study of metric distortion when voters and candidates coincide, which arises naturally in peer selection, and provide tight results for various social cost functions on the line metric. We consider both utilitarian and egalitarian social cost, given by the sum and maximum of the individual social costs, respectively. For utilitarian social cost, we show that the voting rule that selects the $k$ middle agents achieves a distortion that varies between $1$ and $2$ as $k$ varies from $1$ to $n$ when the cost of an individual is the sum of their distances to all selected candidates (additive aggregation). When the cost of an individual is their distance to their $q$th closest candidate ($q$-cost), we provide positive results for $q=k=2$ but mostly show that negative results for general elections carry over to our setting: No constant distortion is possible when $q\leq k/2$ and no distortion better than $3/2$ is possible for $q\geq k/2+1$. For egalitarian social cost, selecting extreme agents achieves the best-possible distortion of $2$ for additive cost and $q$-cost with $q> k/3$, whereas no constant distortion is possible for $q\leq k/3$. Overall, having a common set of voters and candidates allows for better constants compared to the general setting, but cases in which no constant is possible in general remain hard in this setting.