Metric Dimensions of March Madness Brackets

Say you and some friends decide to make brackets for March Madness and are told how each of your brackets scored. The question we ask is: when can you determine how the actual tournament went given your scores? We determine the exact minimum number of brackets needed to do this for any March Madness-style tournament regardless of the scoring system used, and more generally we prove effective bounds for the problem for arbitrary single-elimination tournaments.

💡 Research Summary

The paper “Metric Dimensions of March Madness Brackets” investigates a fundamental reconstruction problem for single‑elimination tournaments such as the NCAA March Madness. Given a collection of participants’ brackets together with the scores each bracket received under an arbitrary scoring rule, the authors ask: how many brackets are required, at a minimum, to uniquely determine the actual outcomes of every match? This question is framed in the language of metric dimension: the set of all possible brackets forms a metric space where the distance between two brackets is the total weight of the matches they predict identically. A “resolving set” (or metric basis) is a subset of brackets whose distance vectors to any other bracket are all distinct; the size of the smallest such set is the metric dimension dim(S,σ) for a tournament S under scoring system σ.

Key definitions

• A single‑elimination tournament is modeled as a directed acyclic graph S with a unique sink (the final) and exactly one outgoing edge from every non‑sink vertex. Vertices that are sources are the initial players; all other vertices are matches.
• For any vertex u, P(u) denotes the set of players that could possibly reach u via a directed walk; thus P(x) for a match x is the set of players that could win that match.
• A bracket B is a function assigning to each match its predicted winner, respecting the tournament’s structure.
• A scoring system σ assigns a positive weight to each match. The σ‑score between two brackets B and B′ is the sum of σ(x) over all matches x where B and B′ agree on the winner.

Main results

1. Standard binary‑tree tournaments (the usual March Madness format, with n = 2^k players). For any scoring system σ, the metric dimension is exactly n/2. Moreover, there exists a universal set of n/2 brackets that resolves the tournament for every possible σ. The lower bound is proved by showing that with fewer than n/2 brackets there will be at least one player on each half of the tree never predicted to win the final, allowing the construction of two indistinguishable brackets. The upper bound is obtained by an inductive construction: resolve the right half (n/2 players) recursively, then pair each left‑half player with a right‑half counterpart to create bracket pairs (B_i, B_i′) that differ only on that left player’s path. The score differences between B_i and B_i′ expose exactly which matches the left player wins, allowing the full tournament to be reconstructed.

2. General single‑elimination tournaments. For any tournament with n players and any σ, dim(S,σ) ≤ n − 1. This bound is tight because a tournament consisting of a single match requires n − 1 brackets to distinguish all possibilities.

3. A universal lower bound. Define for each match x the quantity |P(x)| − max_{u∈N⁻(x)}|P(u)|. Then dim(S,σ) is at least the maximum of this expression over all matches. For the final match z, if its two immediate predecessors each have exactly n/2 possible winners (the usual balanced case), the bound yields dim(S,σ) ≥ n/2. Hence for most “reasonable” tournaments the metric dimension grows linearly with the number of players, regardless of σ. The bound can be as low as 1 for highly unbalanced tournaments, and there exist scoring systems that make the bound tight.

4. Resolving number. The authors introduce res(S,σ), the smallest r such that any set of r brackets is automatically a σ‑resolving set. Let q_max be the maximum, over all players a, of the probability that a uniformly random bracket predicts a as the champion. If N = 2^{n‑1} is the total number of brackets, then (1 − 2q_max)N < res(S,σ) ≤ (1 − q_max)N. For a balanced tournament q_max = 1/n, so the resolving number is essentially the whole bracket space; almost every bracket must be present to guarantee resolution.

Proof techniques

• The lower bound constructions exploit the existence of “unrepresented” players on each side of a match, allowing the creation of indistinguishable bracket pairs.
• The upper bound uses a recursive decomposition of the tournament into sub‑tournaments, together with carefully crafted bracket pairs that differ only on a single player’s path; the score differences between these pairs encode exactly the matches that player wins.
• The resolving number analysis treats the random bracket as a probability distribution over champions, applying simple union‑bound arguments to estimate how many brackets are needed before the chance of two brackets sharing the same distance vector becomes negligible.

Implications
The work provides a precise answer to the “how many brackets do I need to know the whole tournament?” question, showing that for standard March Madness only half the number of players (n/2) of carefully chosen brackets suffice, independent of the scoring rule. For arbitrary tournament shapes the requirement is at most n − 1, and often linear in n. The resolving number result indicates that if one merely collects a large random sample of brackets, one must essentially gather almost the entire bracket space to guarantee uniqueness, highlighting the difficulty of passive reconstruction.

Structure of the paper

• Section 2 sketches the proof of the main theorem for standard tournaments.
• Section 3 develops preliminary lemmas about directed acyclic graphs and player sets P(u).
• Section 4 establishes the lower bounds, including the universal bound of Theorem 1.3.
• Section 5 proves the upper bounds, including the general n − 1 bound and the exact n/2 result for balanced trees.
• An appendix treats additional results on the resolving number.

Overall, the paper bridges combinatorial tournament theory with metric‑dimension concepts, delivering both exact and asymptotic characterizations of the information needed to reconstruct a tournament from bracket scores. This contributes to the theory of information recovery in sports analytics, betting markets, and more broadly to metric‑dimension applications in discrete structures.