---
title: Multi-Group Ex-Ante Agent Game (MEAG)
url: https://www.emergentmind.com/topics/multi-group-ex-ante-agent-game-meag
type: topic
---

# Multi-Group Ex-Ante Agent Game (MEAG)

Searching arXiv for the primary paper and closely related ex-ante/group-game work.
The **Multi-Group Ex-Ante Agent Game (MEAG)** is a complete-information normal-form game constructed from a **multi-group Bayesian game (MBG)** by replacing each **group-type pair** with a distinct agent. In the formulation of "Multi-group Bayesian Games" [2510.02078], the construction is the central device for converting equilibrium search from a space of type-contingent mappings into an ordinary finite action space. The resulting normal-form representation preserves equilibrium meaning through a bijection \(\Gamma\), supports both **multi-group Bayesian Nash equilibrium (MBNE)** and **strongly MBNE**, and provides a tractable entry point for potential-game analysis, semi-tensor-product algebra, and equilibrium computation.

## 1. Origin in multi-group Bayesian games

MEAG is defined only relative to an underlying MBG. In that model, the original player set is partitioned into groups,
\[
G=\mathop{\cup}\limits_{l=1}^{r} G_l,\qquad G_i \bigcap G_j=\emptyset,\ \forall i\neq j,\ i,j\in \mathcal{D}_{r},
\]
where \(G_l\) is the \(l\)-th group and \(|G_l|=m_l\). The type and action spaces are grouped accordingly:
\[
\mathcal{T}_{G_l}={\prod}_{j \in G_l} \mathcal{T}_j,\qquad \mathcal{A}_{G_l}={\prod}_{j \in G_l} \mathcal{A}_j,
\]
so that
\[
\mathcal{T}={\prod}_{l=1}^r \mathcal{T}_{G_l},\qquad \mathcal{A}={\prod}_{l=1}^r \mathcal{A}_{G_l}.
\]

The key informational assumption is that private information is shared **within** groups and incomplete **across** groups. Thus players in group \(l\) observe the whole group type vector \(T_l\in\mathcal{T}_{G_l}\), but do not observe \(T_{-l}\). Beliefs are therefore defined at the group level:
\[
p\left(T_{-l} \mid T_{l}\right)=\frac{p\left(T_{-l},T_{l}\right)}{p\left(T_{l}\right)}
=\frac{p\left(T_{-l}, T_{l}\right)}{\sum_{T_{-l}^{\prime} \in \mathcal{T}_{-G_l}} p\left(T_{-l}^{\prime}, T_{l}\right)}.
\]

Formally, the MBG is the tuple
\[
\mathcal{G}=( G,\mathcal{T},\mathcal{A},p,C),
\]
where \(p:\mathcal{T}\to[0,1]\) is a common-knowledge probability distribution and \(c=\{c_1,\dots,c_m\}\) is the payoff-function set. A strategy is group-contingent:
\[
s_l:\mathcal{T}_{G_l}\rightarrow \mathcal{A}_{G_l}.
\]
Hence an MBG equilibrium is sought over a profile of mappings rather than over a simple finite action set. That mapping-space difficulty is the immediate motivation for the MEAG transformation [2510.02078].

## 2. Construction of the MEAG

The MEAG compiles every possible group type into a separate normal-form decision maker. Its agent set is
\[
\widehat{G}=\{(l,T_{lk})\}_{l\in \mathcal{D}_{r},\,T_{lk}\in \mathcal{T}_{G_l}},
\]
where agent \((l,T_{lk})\) represents the event that group \(l\) has realized type \(T_{lk}\). The corresponding action set is inherited from the original group:
\[
\widehat{\mathcal{A}_{(l,T_{lk})}}=\mathcal{A}_{G_l},
\qquad
\widehat{\mathcal{A}}=\prod_{(l,T_{lk})\in\widehat G}\widehat{\mathcal{A}_{(l,T_{lk})}}.
\]

The transformed payoffs are ex-ante expected payoffs under the prior. For player \(i\in G_l\), the payoff attached to agent \((l,T_{lk})\) is
\[
\widehat{c}_{(l,T_{lk})}^i(\widehat{A})
=
\sum_{T_{-l}\in\mathcal{T}_{-G_l}}
p(T_{-l},T_{lk})\,
c_i\!\left(T_{-l},T_{lk};[\Gamma^{-1}(\widehat A)](T_{-l},T_{lk})\right).
\]
For the cooperative interpretation inside group \(l\), these are averaged:
\[
\widehat{C}_{(l,T_{lk})}(\widehat{A})
=
\frac{\sum_{i\in G_l}\widehat{c}_{(l,T_{lk})}^i(\widehat{A})}{m_l},
\]
mirroring the MBG group payoff
\[
C_l(T;A)=\frac{\sum_{i\in G_l}c_i(T;A)}{m_l}.
\]

The transformation is organized by a bijection
\[
\Gamma:\mathcal S\to \widehat{\mathcal A},
\]
defined by
\[
\Gamma(s)=\left(s_l(T_{lk})\right)_{l\in\mathcal D_r,\;T_{lk}\in\mathcal T_{G_l}}.
\]
Conversely, for any \(\widehat A\in\widehat{\mathcal A}\),
\[
[\Gamma^{-1}(\widehat A)]_l(T_l)=\widehat A_{(l,T_l)}\in\mathcal A_{G_l}.
\]

Conceptually, the MEAG turns each value of a strategy function \(s_l(\cdot)\) into an ordinary action coordinate. This suggests a useful way to read the construction: the Bayesian dependence on type is not removed, but rather compiled into a normal-form player set indexed by type realizations. The paper explicitly positions MEAG as the group-structured analogue of the ex-ante agent transformation for standard Bayesian games [2510.02078].

## 3. Equilibrium notions and the correspondence theorem

The MBG distinguishes two within-group behavioral regimes. The first is **MBNE**, corresponding to within-group cooperative play. Group \(l\) chooses, for each realized type \(T_l\), an action maximizing expected average group payoff:
\[
\max_{A_l\in\mathcal A_{G_l}}
\sum_{T_{-l}\in\mathcal T_{-G_l}}
p(T_{-l}\mid T_l)\,
C_l\!\left(T_{-l},T_l; s_{-l}^*(T_{-l}),A_l\right).
\]

The second is **strongly MBNE**, corresponding to within-group noncooperative play with complete information. The chosen group action must simultaneously maximize every member’s expected payoff:
\[
\max_{A_l\in\mathcal A_{G_l}}
\sum_{T_{-l}\in\mathcal T_{-G_l}}
p(T_{-l}\mid T_l)\,
c_i\!\left(T_{-l},T_l; s_{-l}^*(T_{-l}),A_l\right),
\qquad \forall i\in G_l.
\]

The MEAG mirrors this distinction. A Nash equilibrium of the MEAG is defined using \(\widehat C_{(l,T_{lk})}\):
\[
\max_{\widehat A_{(l,T_{lk})}\in \widehat{\mathcal A}_{(l,T_{lk})}}
\widehat C_{(l,T_{lk})}\!\left(\widehat A_{(l,T_{lk})},\widehat A_{-(l,T_{lk})}^*\right),
\]
for every \((l,T_{lk})\). A **strongly Nash equilibrium** is defined agentwise through \(\widehat c_{(l,T_{lk})}^i\):
\[
\max_{\widehat A_{(l,T_{lk})}\in \widehat{\mathcal A}_{(l,T_{lk})}}
\widehat c_{(l,T_{lk})}^i\!\left(\widehat A_{(l,T_{lk})},\widehat A_{-(l,T_{lk})}^*\right),
\qquad \forall i\in G_l.
\]

The central correspondence theorem states that
\[
s^* \text{ is a (strongly) MBNE of } \mathcal G
\iff
\widehat A^*=\Gamma(s^*) \text{ is a (strongly) Nash equilibrium of } \widehat{\mathcal G}.
\]
The proof rewrites the MEAG best-response condition through \(\Gamma^{-1}\), then uses
\[
p(T_{-l},T_{lk})=p(T_{-l}\mid T_{lk})p(T_{lk}),
\]
with \(p(T_{lk})\) constant with respect to the deviating action. The resulting optimization criterion is exactly the MBNE or strongly MBNE condition [2510.02078].

A frequent source of confusion is the paper’s use of the term **strongly Nash equilibrium**. Here it does not denote the standard coalition-proof strong Nash concept. It denotes the equilibrium notion induced by requiring, for each group-type agent, simultaneous individual optimality for all members of the corresponding original group. The terminology is internal to the MBG-MEAG correspondence.

## 4. Potential structure and semi-tensor-product representation

MEAG is not merely an equilibrium-preserving reformulation; it is also the vehicle through which the paper develops a potential-game characterization. In the MBG, a potential function \(F:\mathcal T\times\mathcal A\to\mathbb R\) satisfies, for the cooperative notion,
\[
C_l(T;A_{-l},A_l)-C_l(T;A_{-l},A_l')
=
F(T;A_{-l},A_l)-F(T;A_{-l},A_l'),
\]
and, for the strong notion,
\[
c_i(T;A_{-l},A_l)-c_i(T;A_{-l},A_l')
=
F(T;A_{-l},A_l)-F(T;A_{-l},A_l'),
\qquad \forall i\in G_l.
\]

If the MBG is potential or strongly potential, then its MEAG is potential or strongly potential, with induced potential
\[
\widehat F(\widehat A)=
\sum_{T\in\mathcal T}
p(T)\,
F\!\left(T;\Gamma^{-1}(\widehat A)(T)\right).
\]
This induced potential converts equilibrium search into potential maximization.

To operationalize the construction, the paper uses the **semi-tensor product (STP)**. The prior is encoded by
\

Source: https://www.emergentmind.com/topics/multi-group-ex-ante-agent-game-meag