Map of Tournaments: Metrics & Applications

• Map of tournaments is a geometric representation of finite, complete asymmetric digraphs that captures structural similarity using isomorphism-invariant metrics.
• It employs techniques such as multidimensional scaling on computed distance matrices to convert complex tournament data into planar embeddings for clear visualization.
• The approach integrates combinatorial, algorithmic, and geometric tools, offering insights for applications in social choice, sports analytics, and extremal graph theory.

A map of tournaments is a geometric representation of the space of finite tournaments—complete asymmetric digraphs—equipped with carefully chosen metrics that reflect structural similarity. Such maps enable the visualization, classification, and comparative study of both synthetic and real-life tournaments, and serve as a conceptual foundation for recent advances in tournament theory, including structure theory (Erdős–Hajnal-type dichotomies), quasi-randomness, and extensions incorporating cooperativity as in Coxeter tournaments. The construction and interpretation of maps of tournaments demand rigorous combinatorial, algorithmic, and geometric tools, with applications ranging from social choice to sports analytics and extremal combinatorics (Nikolow et al., 26 Jan 2026, Ghazal et al., 2021, Coregliano, 2015, Kolesnik et al., 2023).

1. Mathematical Foundations and Tournament Structures

A tournament $T$ on $n$ vertices is an oriented complete graph: for every pair $\{u, v\}$, exactly one of $(u, v)$ or $(v, u)$ is present. Subtournaments are induced tournaments on subsets of vertices. Structural classification leverages properties such as transitivity, local acyclicity, and regularity:

• Transitive tournaments contain no directed cycles; every induced subgraph is transitive.
• Locally transitive tournaments avoid $W_4$ (a single winner with a $3$-cycle among the rest) and $L_4$ (a single loser, rest forming a $3$-cycle), so that every in- and out-neighborhood is transitive (Coregliano, 2015).
• Galaxies, stars, and nebulas are families defined via ordered decompositions into star substructures under some vertex order; galaxies require non-overlapping star-centers, nebulas relax this, and super nebulas generalize with further gadgets (Ghazal et al., 2021).

This structural taxonomy informs both enumeration and extremal results, as in the directed Erdős–Hajnal conjecture, where the existence or prohibition of specific subtournaments dictates the lower bound on the size of large transitive subtournaments.

2. Metrics and Distance Measures on Tournaments

Maps of tournaments rely on defining an isomorphism-respecting distance metric $d(T_1,T_2)$ over the space $\mathcal{T}_n$ of $n$-vertex tournaments.

• Graph-edit distance (GED): The minimum number of edge reversals (counted over all labelings) required to transform $T_1$ into $T_2$; formalized as

$$d_{\mathrm{GED}}(T_1, T_2) = \min_{\pi \in \mathrm{Sym}(n)} |\{ \{i,j\} : (i, j) \in E(T_1) \not\leftrightarrow (\pi(i), \pi(j)) \in E(T_2) \}|.$$

Exact computation is NP-hard but feasible for moderate $n$ via integer linear programming (Nikolow et al., 26 Jan 2026).

• Katz-centrality distance: Based on the sorted vector of Katz centralities, which for $T$ are defined recursively as

$$c_T(v) := \beta + \alpha \sum_{(v \to w) \in E(T)} c_T(w), \quad (\alpha \in [0,1), \beta > 0),$$

then

$$d_{\mathrm{Katz}}(T_1, T_2) = \| \mathbf{c}(T_1)^\downarrow - \mathbf{c}(T_2)^\downarrow \|_1.$$

Sorting produces isomorphism invariance. Katz distance is a pseudometric (distinct tournaments may share the same profile) (Nikolow et al., 26 Jan 2026).

These metrics are foundational for embedding, nearest-neighbor analysis, and correlating structural or computational properties.

3. Construction and Interpretation of Tournament Maps

The primary methodology for constructing a map is as follows (Nikolow et al., 26 Jan 2026):

1. Data Collection: Select a set $S \subseteq \mathcal{T}_n$ consisting of tournaments of fixed order $n$, encompassing canonical examples (fully ordered, "rock–paper–scissors" maximal cyclicity), synthetic models (uniform random, Condorcet noise, strength-based), and real-world instances (sports leagues, voting majority tournaments).
2. Distance Matrix Computation: For chosen metric $d$, compute the $|S| \times |S|$ distance matrix $D = (d_{ij})$ for all $T_i, T_j \in S$.
3. Dimensionality Reduction via MDS: Apply classical multidimensional scaling (MDS): derive the double-centered Gram matrix $B = -\frac{1}{2} H D^2 H$ ($H$ is the centering matrix), extract the top two eigenvectors and eigenvalues, and use these to assign planar coordinates to tournaments.
4. Refinement (Optional): Use SMACOF-style iterative majorization to minimize "stress" and optimize the geometric embedding.
5. Visualization and Analysis: Plot tournaments, optionally annotating by model class or outcome (e.g., number of Copeland winners). Landmark tournaments (e.g., the transitive tournament $T_0$, the "rock–paper–scissors" maximally cyclic instance) serve as reference points.

Clusters and gradients in this geometric representation reflect key combinatorial and algorithmic properties, such as top-cycle size, computation complexity of winner identification, and correspondence with statistical or real-world models.

4. Data Models, Real-life Examples, and Experimental Results

Synthetic and empirical data sampled for tournament maps include:

• Uniform model: Each unordered pair's edge is oriented independently with probability $1/2$.
• Condorcet noise model: Start with the transitive tournament $T_0$; each edge is flipped (reversing the winner) independently with a fixed probability.
• Strength models: Assign weights $w(i)$ to each player; the edge $\{i, j\}$ is oriented $i \to j$ with probability $w(i)/(w(i) + w(j))$. Choices of $w(i)$ include exponential, polynomial, and randomized families.
• Real-life datasets: Aggregated season data from NBA (20 seasons, 29–30 teams), Polish Sports Bridge Association (17 round-robins, 16 players), and voting-derived majority tournaments (impartial culture, 8 candidates, 96 voters) (Nikolow et al., 26 Jan 2026).

Typical findings:

Model/Class	Typical Map Location	Characteristic Structure
Transitive ($T_0$)	Corner/extreme	Hierarchies
Uniform/random	Middle band/arc	Balanced, cyclic
Strength ($\alpha \gg 1$)	Near $T_0$	Slightly noisy hierarchy
NBA tournaments	Cluster near $T_0$	Hierarchies
Bridge tournaments	Farther arc/center	Large cycles/symmetry
Election majority (IC)	Cluster near $T_0$	Surprising order

These positional patterns reliably predict computational hardness for social choice functions and sensitivity to manipulation in competitive formats.

5. Theoretical Extensions and Taxonomies

The map framework integrates naturally with recent advances:

• Erdős–Hajnal Theory: Certain infinite classes of forbidden substructure (e.g., pairs such as super nebula and $\Delta$-galaxy) force large transitive subtournaments, even where single-member forbidden cases remain open (Ghazal et al., 2021).
• Quasi-carousel and quasi-random tournaments: Quasi-carousel tournaments occupy an intermediate position between locally transitive and quasi-random tournaments, characterized by almost-balanced degree sequences and vanishing densities for forbidden $W_4$ and $L_4$ substructures (Coregliano, 2015).

Complexity and statistical properties often align with region in the map:

• Distance from $T_0$ correlates with top-cycle size and computation time for Slater winners.
• Real-world tournaments gravitate toward structured–hierarchical regions, while synthetic models densely populate extremal and intermediate zones.

6. Generalizations: Coxeter Tournaments

Coxeter tournaments generalize the classical concept via root systems $\Phi$ of types $A_{n-1}, B_n, C_n, D_n$, allowing not only competitive (negative) but also cooperative (positive), half-edge (solitaire competitive), and loop (solitaire cooperative) interactions (Kolesnik et al., 2023). The mean score vector resides in the corresponding Coxeter permutahedron or zonotope, and fundamental combinatorial theorems (Moon, Landau) generalize as geometric, probabilistic, and algorithmic statements within this setting.

Taxonomically:

• Classical tournaments correspond to type $A_{n-1}$, whereas types $B_n$, $C_n$, $D_n$ enable additional symmetries and interaction modalities.
• Statistical models (e.g., Bradley–Terry) and structural conditions (majorization, submodularity) are realized uniformly in this broader context.

The use of generalized tournament maps in the Coxeter setting is a plausible direction, as the geometric and metric toolkit supports multidimensional comparative analyses for complex interactions.

7. Applications and Future Directions

Maps of tournaments facilitate a variety of applications:

• Visualization: Summarize large collections of tournaments, exposing hidden gradients, clusters, and atypical structures.
• Benchmarking and Model Selection: Enable rigorous comparison of tournament generation models and empirical data, detecting which models account for observed data via positional coincidence.
• Algorithmic Hardness: Distance from ordered or cyclic "landmarks" predicts computational complexity for decision and optimization problems on tournaments.
• Extremal Graph Theory: Inform conjectures and proofs by situating specific families (e.g., galaxies, nebulas) within the global landscape and tracking structural progress toward resolving open cases (e.g., Erdős–Hajnal for $K_6$).
• Social Choice and Statistical Ranking: Structural position predicts manipulability, winner diversity, and robustness under noise or model misspecification.

Given the synergy with combinatorics, geometry, algorithmics, and statistics, maps of tournaments are poised to shape the ongoing synthesis and extension of tournament theory (Nikolow et al., 26 Jan 2026, Ghazal et al., 2021, Kolesnik et al., 2023, Coregliano, 2015).