BallotRank: A Condorcet Completion Method for Graphs

Jason Douglas Todd, Ismar Volić · 2026

Read BallotRank: A Condorcet Completion Method for Graphs by Jason Douglas Todd, Ismar Volić (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-14015.txt.

Introduction

Jason Douglas Todd and Ismar Volić build a voting method that borrows from web search ranking in 2026. BallotRank is a ranked preference aggregation method derived from PageRank. PageRank ranks web pages by the links that point to them. BallotRank adapts that idea to ballots, without a damping factor.

The method is Condorcet-consistent, meaning it honors a candidate who beats each rival head-to-head. They tested it on nearly two thousand ranked choice elections. Over twenty thousand internet polls were also examined. At conventional damping values, BallotRank always identified the Condorcet winner.

Catalog record

CREATORJason Douglas Todd, Ismar Volić
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE69 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-14015

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

BallotRank: A Condorcet Completion Method for Graphs Jason Douglas Todd, Ismar Volić Abstract . We introduce BallotRank, a ranked preference aggregation method derived from a modified PageRank algorithm. It is a Condorcet-consistent method without damping, and empirical examination of nearly 2,000 ranked choice elections and over 20,000 internet polls confirms that BallotRank always identifies the Condorcet winner at conventional values of the damping parameter. We also prove that the method satisfies many of the same social choice criteria as other well-known Condorcet completion methods, but it has the advantage of being a natural social welfare function that provides a full ranking of the candidates. 2020 Mathematics Subject Classification Primary 91B12; Secondary 91B14 1. Introduction Ranked choice voting is having a moment. In addition to the intense media coverage of the recent New York City mayoral election, where primaries were held under ranked choice rules, there are also new electoral reforms in Hawaii, the District of Columbia, and elsewhere. Although to the lay public instant runoff (IRV) and its multi-winner variant, single-transferable vote (STV), are synonymous with ranked choice voting, these methods represent but the tip of the iceberg to social choice theorists. Here, older systems like the Borda count and Copeland/Llull method compete for attention with more modern inventions such as minimax and ranked pairs. We add to this list by introducing a new ranked voting method, which we call BallotRank . BallotRank builds upon the equivalence between a preference profile and a graph first noted in [ McG53 ] and extended by [ BJNR21 ] to produce the so-called ‘‘Condorcet graph.’’ 1 1 1 See [ Vol24 ] for an interesting discussion of the equivalence. By modifying this graph and then applying a variant of the PageRank algorithm [ BP98 ] , BallotRank becomes a natural social welfare function. If one elects to eliminate the PageRank damping parameter by setting it to 1, BallotRank is a Condorcet-consistent method satisfying many well-known social choice criteria. If instead the damping parameter is set to conventional values such as 0.85, it still correctly identifies the Condorcet winner in every single one of the more than 20,000 elections we examine. BallotRank is also a more natural social welfare function than some of the other Condorcet completion methods like minimax, ranked pairs, and Schulze/beatpath. The calculation of the rankings, including identifying the winner, requires input from the entire system. The power method of calculating BallotRank iteratively updates the rankings, but it needs the full information from the previous step to do so. Other methods are primarily designed to find the Condorcet winner or break Condorcet cycles, and the fact that they can be used to produce a ranking is an afterthought. BallotRank’s final ranking, on the other hand, is canonical and intrinsic to the method. Several recent works have proposed PageRank- or Markov chain–based methods for aggregating pairwise comparisons. Negahban et al. [ NOS12 , NOS17 ] propose rank centrality , an iterative spectral algorithm that recovers latent item scores from noisy pairwise comparisons under the Bradley–Terry–Luce model. These methods interpret rankings as estimators of qualities and focus on consistency, sample complexity, and optimality guarantees. From a social choice theory viewpoint, Altman and Tennenholtz [ AT05 ] provide an axiomatic characterization of PageRank as a ranking system on directed graphs, establishing a representation theorem based on graph-theoretic axioms tailored to web-ranking contexts. Brandt and Fischer [ BF07 ] build on this and analyze PageRank as a weak tournament solution. The above papers are far apart in intent and spirit from the constructions in this paper. Research closer to ours was done by Bana et al. [ BJNR21 ] , who introduce convergence voting , a PageRank-inspired social choice function defined on pairwise comparison graphs and interpreted as the outcome of an iterative negotiation process converging to a stationary distribution. Their method is explicitly framed as a consensus rule lying between Condorcet and Borda methods and does not aim to satisfy Condorcet consistency. We will provide more details about the difference between convergence voting and BallotRank in Section 5 . The paper is structured as follows. In Section 2 , we first present the intuition and mathematics behind the PageRank algorithm before modifying it to account for the possibility of cycles. In Section 3 , we work through a toy example to fix ideas. Section 4 then examines BallotRank empirically, applying it to online polling and real-world legislative elections and paying particular attention to the problem of Condorcet winners, which BallotRank routinely identifies. Section 5 compares the empirical performance of BallotRank to minimax, ranked pairs, and Schulze methods – standard and familiar methods – as well as the newer convergence voting mentioned above. As we will demonstrate, BallotRank can behave differently from these methods, and we regard this as an advantage owing to the fact that BallotRank uses more information from the profile to determine the ranking than these other methods. We switch gears in Section 6 to consider BallotRank against a standard set of social choice criteria including monotonicity, Pareto, and independence of irrelevant alternatives. We conclude that its behavior is largely in line with the standard Condorcet completion methods mentioned above, but there are also appreciable differences. Finally, Section 7 offers a summary of the results and possible directions for further investigation. 2. From PageRank to BallotRank This section presents PageRank and BallotRank intuitively, formalizes both, and translates them into the language of linear algebra. Along the way, we prove that under certain conditions, BallotRank is a Condorcet-consistent method. 2.1. Social choice under PageRank and BallotRank The intuition behind PageRank begins with the common topological interpretation of a random walk along a network, but first we must be clear about the network itself. We first introduce an unweighted graph and add weights, then sketch out the PageRank algorithm, and finally discuss the necessary addition of self-loops. 2.1.1. Basic network In our setting, n nodes represent candidates (more generally, choices) and edges run from losers to winners. 2 2 2 This differs from the graph theoretic depiction of [ McG53 ] , in which edges run from winners to losers. We will come back to the issue of determining the losers and winners of head-to-head contests in Section 2.2.1 . We assume the network is fully connected in the sense that it is complete if regarded as an undirected graph. With edge orientations, the in-degree and out-degree vary depending upon candidate popularity. A Condorcet winner, by definition, beats all other candidates and hence has an out-degree of zero and an in-degree equal to n-1 ; a Condorcet loser, conversely, has an out-degree of n-1 and an in-degree of zero. 2.1.2. Adding weights The first complication involves moving from this unweighted network, in which edges define a binary social preference, to a weighted network, with weights defined by pairwise margins of victory. Thus if candidate a is preferred to candidate b by 363 voters and b is preferred to a by only 300, the network would feature an edge from node b to node a with weight 63. In matrix form, this network graph is an n\times n adjacency matrix whose A_{ij} entry equals j ’s margin of victory if j is relatively more preferred

Read the complete text