Minimal Balanced Collections in Game Theory

• Minimal balanced collections are families of nonempty subsets whose characteristic vectors admit a unique, strictly positive balancing system that sums to the all-ones vector.
• They support efficient recursive generation algorithms like Peleg's, with exact enumeration revealing super-exponential growth as the number of players increases.
• They play a pivotal role in cooperative game theory by indexing Bondareva–Shapley inequalities and enabling finite, combinatorial tests for core nonemptiness, exactness, and stability.

Minimal balanced collections are finite families of nonempty subsets whose characteristic vectors admit a strictly positive balancing system summing to the all-ones vector, and whose balancedness is destroyed by removing any member. They generalize partitions of a finite set and occupy a central position in cooperative game theory because they index the irredundant Bondareva–Shapley inequalities for balanced games, support constructive generation algorithms, and enable finite tests for core nonemptiness, exactness, extendability, and core stability. Recent work has also clarified their large-$n$ asymptotics and their role in the facial structure of the balanced-game cone (Mermoud et al., 8 Jul 2025).

1. Definitions and equivalent formulations

Let $N$ be a finite player set of size $n$, and for each nonempty coalition $S\subseteq N$ let $1^S\in\{0,1\}^N$ denote its characteristic vector. A collection $B\subseteq 2^N\setminus\{\emptyset\}$ is balanced if there exists a system of strictly positive balancing weights $\lambda^B\in\mathbb{R}^{|B|}_{++}$ such that

$\sum_{S\in B}\lambda^B_S\,1^S=1^N.$

Equivalently, $1^N$ lies in the relative interior of the cone generated by $\{1^S:S\in B\}$. A balanced collection is minimal balanced if no proper subcollection is balanced. An equivalent algebraic characterization is that the balancing weights are unique once the normalization is fixed; in matrix form, if $N$0 is the $N$1–$N$2 incidence matrix with columns $N$3, then minimality is equivalent to full column rank together with existence of the unique positive solution of $N$4 (Mermoud et al., 8 Jul 2025).

The same notion admits a geometric convex-hull formulation. If $N$5 and $N$6, then $N$7 is balanced if and only if

$N$8

After projection to the hyperplane $N$9, balancedness is exactly the condition that the barycenter $n$0 lies in the relative interior of the convex hull of the face-centers $n$1. This formulation is useful for the combinatorics of enumeration and for connecting minimal balanced collections to convex-geometric questions (Bludov et al., 24 Nov 2025).

A notational convention used in the asymptotic literature is to write $n$2 for the set of all minimal balanced collections on $n$3, excluding the trivial collection $n$4. This convention matters when comparing counts across papers (García-Segador et al., 16 Jan 2026).

2. Enumeration, conventions, and small cases

Closed-form counts are not known beyond $n$5, but implementation of Peleg’s recursive algorithm yields the exact numbers of minimal balanced collections on $n$6 for $n$7 (Mermoud et al., 8 Jul 2025).

$n$8	Number of minimal balanced collections
1	1
2	2
3	6
4	42
5	1,292
6	200,214
7	132,422,036

These values exhibit very rapid growth, described in the source as apparently super-exponential. The same source notes that the number of maximal unbalanced collections grows much more slowly (Mermoud et al., 8 Jul 2025).

A standard point of confusion is the treatment of the trivial collection $n$9. For $S\subseteq N$0, one source reports exactly five minimal balanced collections other than $S\subseteq N$1, namely

$S\subseteq N$2

so $S\subseteq N$3 under the convention excluding $S\subseteq N$4; the count becomes $S\subseteq N$5 when $S\subseteq N$6 is included (García-Segador et al., 16 Jan 2026).

Small examples already show the two principal structural patterns. One is partition-like, such as $S\subseteq N$7 for $S\subseteq N$8, with weights $S\subseteq N$9. The other is genuinely non-partitional, such as $1^S\in\{0,1\}^N$0, whose balancing weights are $1^S\in\{0,1\}^N$1. This suggests that minimal balanced collections strictly extend the combinatorics of set partitions rather than merely rephrasing them (Mermoud et al., 8 Jul 2025).

3. Recursive generation and the Peleg algorithm

Peleg’s inductive algorithm provides a constructive generation mechanism. Its key insight is that every minimal balanced collection on $1^S\in\{0,1\}^N$2 can be obtained from those on $1^S\in\{0,1\}^N$3 by one of four constructions, depending on how the new player $1^S\in\{0,1\}^N$4 enters the balancing system. Starting from the list $1^S\in\{0,1\}^N$5 of all minimal balanced collections on $1^S\in\{0,1\}^N$6, together with their balancing vectors, the algorithm examines subsets of indices satisfying bounds on partial sums of balancing weights. According to whether the relevant partial sum fits exactly, lies below $1^S\in\{0,1\}^N$7, or crosses $1^S\in\{0,1\}^N$8, one forms a new collection on $1^S\in\{0,1\}^N$9 by adding $B\subseteq 2^N\setminus\{\emptyset\}$0 to selected coalitions, possibly also adding $B\subseteq 2^N\setminus\{\emptyset\}$1, or moving weight from some $B\subseteq 2^N\setminus\{\emptyset\}$2 to $B\subseteq 2^N\setminus\{\emptyset\}$3, and then adjusts the weights accordingly. A further construction handles the case in which two minimal balanced collections $B\subseteq 2^N\setminus\{\emptyset\}$4 on $B\subseteq 2^N\setminus\{\emptyset\}$5 combine to form a collection $B\subseteq 2^N\setminus\{\emptyset\}$6 of rank $B\subseteq 2^N\setminus\{\emptyset\}$7; then a convex combination of their balancing vectors is used to force $B\subseteq 2^N\setminus\{\emptyset\}$8, and $B\subseteq 2^N\setminus\{\emptyset\}$9 is added to the corresponding coalitions (Mermoud et al., 8 Jul 2025).

This recursive scheme was implemented computationally. The implementation runs in minutes for $\lambda^B\in\mathbb{R}^{|B|}_{++}$0 and about $\lambda^B\in\mathbb{R}^{|B|}_{++}$1 hours for $\lambda^B\in\mathbb{R}^{|B|}_{++}$2. It outperformed a generic vertex-enumeration approach on the polytope

$\lambda^B\in\mathbb{R}^{|B|}_{++}$3

The practical significance is that the full list of minimal balanced collections can be precomputed and stored, after which many game-theoretic queries become finite combinatorial checks rather than fresh optimization problems (Mermoud et al., 8 Jul 2025).

An illustrative construction starts from the $\lambda^B\in\mathbb{R}^{|B|}_{++}$4 collection

$\lambda^B\in\mathbb{R}^{|B|}_{++}$5

On $\lambda^B\in\mathbb{R}^{|B|}_{++}$6, adding player $\lambda^B\in\mathbb{R}^{|B|}_{++}$7 to exactly those coalitions whose $\lambda^B\in\mathbb{R}^{|B|}_{++}$8-sum is $\lambda^B\in\mathbb{R}^{|B|}_{++}$9 yields

$\sum_{S\in B}\lambda^B_S\,1^S=1^N.$0

with the same balancing vector. This example exhibits the local nature of the inductive step: the new collection is obtained by modifying incidence patterns while preserving the balancing equations (Mermoud et al., 8 Jul 2025).

4. Polyhedral structure and asymptotics

For a transferable-utility game $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$1 with $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$2, the balanced-game cone is the polyhedron

$\sum_{S\in B}\lambda^B_S\,1^S=1^N.$3

By Bondareva–Shapley, $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$4 if and only if $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$5 is balanced, and the inequalities indexed by all minimal balanced collections give a full polyhedral description of $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$6 (García-Segador et al., 16 Jan 2026).

The asymptotic theory identifies the dominant contribution to the number of minimal balanced collections. Writing $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$7 for the total number of minimal balanced collections on $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$8 excluding $\sum_{S\in B}\lambda^B_S\,1^S=1^N.$9, and $1^N$0 for those of size exactly $1^N$1, one has

$1^N$2

Thus almost all minimal balanced collections have size exactly $1^N$3. The proof proceeds by identifying size-$1^N$4 minimal balanced collections with invertible $1^N$5 $1^N$6–$1^N$7 matrices $1^N$8 such that $1^N$9 componentwise, counting these matrices up to column permutation, and then applying random-matrix arguments: a random Bernoulli$\{1^S:S\in B\}$0 matrix is singular with probability $\{1^S:S\in B\}$1, and conditioned on invertibility, the probability that $\{1^S:S\in B\}$2 is asymptotically $\{1^S:S\in B\}$3 (García-Segador et al., 16 Jan 2026).

The same asymptotic picture has a direct facial interpretation. Each minimal balanced collection $\{1^S:S\in B\}$4 corresponds to a facet $\{1^S:S\in B\}$5 of $\{1^S:S\in B\}$6. If $\{1^S:S\in B\}$7, then every game $\{1^S:S\in B\}$8 on $\{1^S:S\in B\}$9 has a point-core, that is, a singleton core. If $N$00, then the interior of $N$01 consists of games whose core is never a singleton. Consequently,

$N$02

so the proportion of facets whose games have a non-singleton core tends to $N$03 (García-Segador et al., 16 Jan 2026).

Independent combinatorial bounds corroborate the same scale of growth. If $N$04 denotes the total number of minimal balanced collections on $N$05, then

$N$06

For fixed $N$07, one also has

$N$08

which supports the conclusion that the $N$09 term eventually dominates. The derivation uses a $N$10-action by column complementation, an orbit lemma relating strictly positive weight matrices to matrices whose unique weight vector has mixed signs, and counting over $N$11 for the square case (Bludov et al., 24 Nov 2025).

5. Applications to cooperative game theory

Once the full list $N$12 is precomputed and stored, many standard decision problems in cooperative game theory reduce to explicit finite summations over minimal balanced collections rather than solving a linear program for each query. The basic example is core nonemptiness. By Bondareva–Shapley,

$N$13

The reported implementation checked $N$14 random $N$15-player games in approximately $N$16 second by this test, versus approximately $N$17 seconds using a simplex linear program; a separate illustrative example reports $N$18 seconds versus $N$19 seconds for a random $N$20-player game (Mermoud et al., 8 Jul 2025).

The same precomputed data support more refined properties. Exactness of a coalition $N$21, meaning that there exists $N$22 with $N$23, reduces to checking that the modified game $N$24, where $N$25, still satisfies the Bondareva–Shapley inequalities. Strict vital-exactness, extendability, feasible collections, and blocking pairs are likewise tested by iterating over $N$26 and performing simple linear or inequality checks. The stated complexity is $N$27, in contrast to repeated linear-programming-based methods (Mermoud et al., 8 Jul 2025).

A more elaborate application concerns stable sets in the sense of von Neumann and Morgenstern. If $N$28 are imputations, then $N$29 is dominated by $N$30 if there is a coalition $N$31 with $N$32 and $N$33 for all $N$34. Grabisch–Sudhölter’s nested balancedness criterion yields a finite constructive test for when the core is stable. At the first level, for each feasible collection of exact coalitions $N$35, one associates to every $N$36 a further minimal balanced collection $N$37 in the strictly vital-exact coalitions. At the second level, one forms a finite set $N$38 and checks minimal balanced subsets $N$39, excluding certain degenerate $N$40. The operative theorem states that a balanced game $N$41 has a stable core if and only if, for every feasible collection $N$42 and every admissible choice of $N$43, there exists such a minimal balanced subset $N$44 with

$N$45

or with $N$46 in the boundary case. The second level requires generalizing the notion of balanced collection to balanced sets. The resulting test avoids multiparametric linear programs, although it does require many small vertex-enumeration subproblems on $N$47, whose size is at most $N$48. A $N$49-player game of Biswas et al. (1999) is used as a counterexample by exhibiting a blocking feasible pair and then a balanced subset $N$50 that fails the nested condition (Mermoud et al., 8 Jul 2025).

A plausible implication is that minimal balanced collections serve not merely as certificates of balancedness, but as the basic finite combinatorial objects from which several otherwise unrelated solution-theoretic tests can be organized.

6. Terminological scope and adjacent uses of “balanced”

The expression “balanced” is not uniform across the literature, and conflating the underlying objects can obscure the coalition-theoretic meaning of minimal balanced collections. In the theory of root systems, Moroianu and Schwahn call a subset $N$51 balanced if there exists a choice of signs $N$52 such that

$N$53

and well-balanced if, in addition, the complement $N$54 is strongly orthogonal. They compute the maximal and minimal size of well-balanced subsets in all simple root systems and show that

$N$55

This is a different notion from balanced collections of coalitions, even though both involve finite families satisfying a balancing condition (Moroianu et al., 11 May 2026).

In topological combinatorics and rigidity theory, Oba studies balanced simplicial complexes, where “balanced” means that the $N$56-skeleton is $N$57-colorable. A minimal $N$58-cycle complex is then the support complex of a minimal $N$59-cycle, and the main theorem proves that for any $N$60-subset $N$61 of colors, the rank-selected subgraph is infinitesimally rigid in $N$62. This yields the balanced lower-bound theorem for balanced minimal $N$63-cycle complexes. Again, the terminology overlaps, but the object is a colored simplicial complex rather than a family of coalitions with positive balancing weights (Oba, 2023).

Two recurring misconceptions are therefore easy to isolate. First, numerical counts depend on whether the trivial collection $N$64 is included. Second, the adjective “balanced” is domain-dependent: balanced collections in cooperative game theory, balanced subsets in root systems, and balanced simplicial complexes are mathematically distinct constructions. The coalition-theoretic notion is characterized by positive solutions to $N$65, minimality under deletion, and its role in the polyhedral and algorithmic structure of balanced games.