---
title: 'Fair Game: Formal and Applied Perspectives'
url: https://www.emergentmind.com/topics/fair-game
type: topic
---

# Fair Game: Formal and Applied Perspectives

Searching arXiv for the specified "Fair Game" papers to ground the article in current records.
arxiv_search(query="Fair Game", max_results=10)
Searching arXiv for exact records relevant to "Fair Game".
{"query":"all:\"Fair Game\" OR ti:\"Fair Game\"", "max_results": 10}
“Fair game” is not a single formal notion in contemporary research; it is a family of domain-specific constructions in which fairness is encoded as symmetry, admissibility, proportionality, or corrective control. In combinatorial probability, fairness can mean exact equality of winning probabilities produced by a score-exchanging involution [2406.20049]. In bargaining and evolutionary models, it can refer to the emergence or promotion of \(50\text{–}50\) behavior in the Ultimatum Game under occasional fair actions or targeted intervention [2202.06002, 2102.03461]. In machine learning, it can denote either a modified cooperative game with fair-and-stable payoff allocation among data providers [1911.11555] or a closed-loop auditor–debiaser architecture that adapts fairness objectives over time [2508.06443]. In formal methods, it denotes graph games in which one or both players are constrained by strong transition-fairness conditions [2310.13612, 2501.17255].

| Domain | Formal object | Operational meaning of fairness |
|---|---|---|
| Probability and combinatorics | Coin-toss word games; dice-sum distributions | Symmetry of win laws or indistinguishability from a fair distribution |
| Evolutionary and bargaining theory | Ultimatum-game populations | Fair offers, fair responses, or low-cost intervention toward equitable outcomes |
| Multi-party ML and Fair ML | Cooperative allocation; RL auditor–debiaser loop | Proportional reward allocation or dynamic bias control |
| Formal verification and control | \(\omega\)-regular, mean-payoff, energy games | Obligatory recurrence of designated transitions |
| Strategic and quantum game theory | Bayesian games; stochastic Prisoner’s Dilemma | Equalized payoffs, fair equilibria, or fairness-enforcing strategies |
| Applied systems | Autonomous racing; redistricting | Sportsmanship constraints or near-proportional procedural guarantees |

## 1. Probabilistic symmetry and exact fairness

A mathematically sharp use of “fair game” appears in Litt’s coin-tossing game as analyzed by Basdevant, Hénard, Maurel-Ségala, and Singh. A fair coin is flipped \(n\) times, two words \(A,B\in\{H,T\}^{\ell}\) of equal length are fixed, and overlapping occurrences are counted. The paper defines the overlap set \(\mathrm{Corr}_k(A,B)\), the correlation number \(C(A,B)=\sum_{k\in \mathrm{Corr}_k(A,B)}2^k\), and the generating polynomial \(P_{A,B}(z)=\sum_{k\in \mathrm{Corr}_k(A,B)} z^k\). Its main theorem states that if \(C(A,A)=C(B,B)\), then for every \(n\ge 1\) the random vectors \((N_A(X_n),N_B(X_n))\) and \((N_B(X_n),N_A(X_n))\) have the same law, hence \(P_n(\mathrm{Alice\ wins})=P_n(\mathrm{Bob\ wins})\) [2406.20049].

The result is notable because the relevant invariant is the auto-correlation structure alone. The paper gives \(A=\mathrm{HHT}\) and \(B=\mathrm{THH}\) as the smallest nontrivial example: both have \(C(A,A)=C(B,B)=4\), even though their inter-correlations differ, and the game is exactly fair for every \(n\) [2406.20049]. The proof is bijective rather than asymptotic: an explicit involution \(\Phi\) on length-\(n\) bit-strings exchanges the Alice and Bob scores.

A related but distinct notion of fairness appears in the loaded-dice analysis of craps. There the question is whether two non-uniform six-sided dice can reproduce exactly the sum distribution of two fair dice. Writing \(P(x)=p_1x+\cdots+p_6x^6\), \(Q(x)=q_1x+\cdots+q_6x^6\), and \(F(x)=P(x)Q(x)\), the condition becomes
\[
P(x)Q(x)=\frac{\psi_6(x)^2}{36},
\]
where \(\psi_6(x)=x+x^2+\cdots+x^6\). The analysis finds \(252\) complex solutions counted with multiplicity, \(51\) distinct unordered pairs, and exactly one solution with all \(p_i,q_j\ge 0\): the trivial fair pair \(p_i=q_i=1/6\) [1308.6816]. In this setting, a “fair game” of craps is not generated by hidden loading; fairness is identifiable from the full feasible nonnegative factorization structure.

## 2. Fairness in bargaining populations and intervention design

In evolutionary Ultimatum-Game models, fairness is often operationalized by offer and acceptance levels rather than by symmetry of win probabilities. In the “good Samaritan” model, each player carries a strategy \(S_i=(p_i,q_i)\), with \(p_i\in[0,1]\) the fraction offered as proposer and \(q_i\in[0,1]\) the minimal accepted offer as responder. In each round, with probability \(\rho\), a player temporarily abandons \((p_i,q_i)\) and uses the fixed fair strategy \(S_h=(C_h,C_h)\) with \(C_h=0.5\). A mean-field replicator-style ODE is derived:
\[
\dot x=(1-x)\bigl[(1-\rho)x^2+(\rho-C_h)x+\rho\bigr],
\]
with a saddle-node bifurcation at
\[
\rho_c=\frac12-\frac{1}{\sqrt 5}\simeq 0.0528.
\]
For \(0<\rho<\rho_c\) there are two attractors, while for \(\rho>\rho_c\) only full fairness remains [2202.06002].

The numerical phenomenology depends strongly on network structure. In well-mixed populations, fairness exhibits a first-order transition and hysteresis: for \(\rho\lesssim 1\%\), the population remains in a low-fairness state with \(\langle p\rangle\approx 0.1\); between roughly \(1\%\) and \(10\%\), runs are bistable; and small fair frequencies around \(5\%\) are sufficient to tip a homogeneous society into universal \(50\text{–}50\) sharing [2202.06002]. On Barabási–Albert graphs the transition becomes continuous, and even a single highest-degree node acting as a good Samaritan can drive the population to full fairness once \(\rho\approx 20\%\) [2202.06002].

A more interventionist formulation appears in the spatial Ultimatum Game with an external investor. Players occupy a \(2\)-D lattice, use one of four discrete strategies \(HH,HL,LH,LL\), update by the Fermi rule with noise \(K=0.1\), and may mutate with probability \(\mu\). The investor observes either global frequencies \(I_{\mathrm{macro}}\) or local neighborhood frequencies \(I_{\mathrm{local}}(i)\), gives supplementary endowments \(\Delta p_i\) or \(\Delta q_i\), and incurs long-run cost
\[
C_{\mathrm{total}}=\sum_{t=1}^{T}\sum_{i=1}^{Z} c_t(i).
\]
The main design finding is asymmetrical: globally, the cheapest scheme targets only \(HH\)-players with \(p_f\approx 0.3\); locally, cost efficiency improves when support is restricted to neighborhoods with \(n_f\approx 0.25\). For \(\mu=10^{-2}\), achieving \(F\ge 90\%\) is summarized as \(C_{\mathrm{tot}}\sim 0.7\times T\times Z\) for the global \(HH\)-only rule and \(C_{\mathrm{tot}}\sim 0.1\times T\times Z\) for the local rare-proposer rule [2102.03461].

## 3. Fairness as stability, proportionality, and strategic advantage

In multi-party machine learning, “fair game” is defined through a modified cooperative-game model for non-rivalrous goods. A coalition-value map \(v:2^N\to\mathbb R_{\ge 0}\) is assumed monotonic, and an outcome is a payoff vector \(x=(x_1,\dots,x_n)\). Stability is modified from the classical core: \(x\) is stable iff for every coalition \(S\subseteq N\), there exists \(k\in S\) with \(x_k\ge v(S)\). Fairness is proportionality to a baseline contribution vector \(\phi\), taken to be the Shapley value, so that \(x_i=a\phi_i\) for some \(a>0\). Proposition 5 reduces stability checking to the \(n\) prefix inequalities \(x_i\ge v(\{1,\dots,i\})\), and Theorem 6 gives the unique optimal fair-and-stable allocation when it exists:
\[
x_i^*=\frac{v(N)}{\phi_n}\phi_i,
\]
provided the corresponding prefix constraints hold [1911.11555]. In experiments, Fisher-information valuation satisfies these constraints, whereas mutual-information valuation does not, forcing a stable but non-proportional compromise [1911.11555].

A different sense of fairness appears in Bayesian games with classical, quantum, and no-signaling advice. In the two-player family \(\mathcal G(\kappa,\tau)\), a fair equilibrium has \(F_A=F_B\), while an unfair equilibrium has \(F_A\ne F_B\). The analysis shows that nonlocal correlations can outperform not only fair classical equilibria but also unfair ones. When the fair-equilibrium benchmark is \(F_{\mathrm{fair}}=9/16\), the outperforming condition reduces to \(\mathcal B>2\), where \(\mathcal B\) is the CHSH expression. For unfair equilibria, the advantageous region is characterized by explicit linear inequalities in \(\mathcal B\), \(m_0+n_0\), and \(m_1+n_1\) [1601.02349]. This directly refutes the misconception that nonlocal advantage is relevant only against fair classical baselines.

Fairness in payoff control becomes more restrictive in stochastic repeated games. In the periodic Prisoner’s Dilemma, a fair zero-determinant strategy for player \(1\) enforces \(\mathcal S_1=\mathcal S_2\). Nakamura and Ueda derive necessary and sufficient existence conditions in terms of \(D_1,D_2,\delta_R,\delta_P\), and show that fair ZD strategies do not necessarily exist in the periodic game, unlike in the standard repeated Prisoner’s Dilemma [2603.19641]. Tit-for-Tat is fair only in the special case
\[
\delta_R+\delta_P=\delta_S+\delta_T,
\]
so equal-payoff enforcement is no longer generic once state dynamics are introduced [2603.19641].

## 4. Dynamic fairness in machine learning and sequential decision-making

In Fair ML, “Fair Game” names a closed-loop framework in which an auditor and a debiaser surround an ML model and adapt over time. The joint auditor–debiaser is modeled as an RL agent interacting with an environment \(E\). The state is \(s_t=(\theta_t,\mu_t^-)\), where \(\theta_t\) are current model parameters and \(\mu_t^-\in\mathbb R^k\) is the vector of last-observed fairness metrics. Actions include threshold choices, reweighting factors, or Lagrange multipliers, while the reward penalizes both predictive loss and fairness violation:
\[
r_t=-\bigl[\mathrm{Loss}(f_t,D_t)+\alpha\cdot \mathrm{FairnessViolation}(\mu_t)\bigr].
\]
The auditor estimates metrics such as Statistical Parity, Demographic Parity ratio, Equalized Odds gap, and Predictive Value Parity gap, and is assumed “anytime-accurate PAC” [2508.06443].

The theoretical guarantee combines auditing accuracy and RL regret. Under an anytime-PAC auditor and an \(O(\sqrt T)\)-regret debiaser, the time-averaged bias satisfies
\[
\frac1T\sum_{t=1}^{T}\mu_t(f_{\theta_t})
\le
\min_{\pi}\frac1T\sum_{t=1}^{T}\mu_t(f_t^\pi)+O(\epsilon+T^{-1/2})
\]
with probability at least \(1-\delta\) [2508.06443]. On UCI-Adult, COMPAS, and German Credit, the reported excerpt gives Fair Game an accuracy of \(0.76\), with SP gap \(0.07\), EO gap \(0.08\), and PVP gap \(0.09\); over a \(50\)-step horizon it reduces SP and EO gaps by about \(30\%\) beyond static-RL at an accuracy cost of about \(5\%\) [2508.06443].

A related but distinct sequential formalism replaces utilitarian welfare by Proportional Fairness in mixed-motive Markov games. The fair altruistic utility is
\[
U_i(s;\alpha)=(1-\alpha)\log p_i(s)+\alpha\sum_{j=1}^{n}\log p_j(s)
=\log p_i(s)+\alpha\sum_{j\ne i}\log p_j(s),
\]
and in two-player social dilemmas cooperation at \((C,C)\) is stabilized when
\[
\alpha>\frac{\log T-\log R}{\log R-\log S}
\]
for Prisoner’s Dilemma and Chicken [2602.08389]. The sequential extension defines a Fair Markov Game and derives Fair MAA2C and Fair MAPPO. In CleanUp, under \(\alpha=1\), PF-MAPPO achieves approximately \(120\) apples with Gini \(\approx 0.1\), whereas utilitarian MAPPO achieves approximately \(40\) apples with Gini \(\approx 0.8\) [2602.08389]. Here fairness is not merely a constraint; it is embedded directly into the objective.

## 5. Fairness constraints in graph games and quantitative verification

In formal verification, fair games are games on finite graphs whose plays must satisfy fairness obligations on designated moves. In fair \(\omega\)-regular games, a game graph \(G=(V_\exists,V_\forall,E)\) is augmented by a set of fair edges \(E_f\subseteq E\). A play is \(i\)-fair if, whenever a source vertex of a fair edge owned by player \(i\) is visited infinitely often, that fair edge is also taken infinitely often. Two winning conditions \(\alpha\) and \(\beta\) determine the winner of mutually fair and mutually unfair plays; if exactly one player is fair, the fairly playing player wins [2310.13612]. These fair \(\alpha/\beta\) games are determined, and the parity/parity case admits both a polynomial reduction to ordinary parity games and a direct symbolic \(\mu\)-calculus fixpoint algorithm [2310.13612].

The quantitative extension replaces parity acceptance with mean-payoff or energy objectives while keeping strong transition fairness. For a set \(F\subseteq E\) of fair edges,
\[
\mathrm{Fair}(F)=
\{\pi\mid \forall (u,v)\in F:\ u\in \mathrm{Inf}(\pi)\Rightarrow (u,v)\in \mathrm{Inf}_E(\pi)\}.
\]
When fairness is imposed on player \(1\), the winning condition becomes \(\mathrm{MP}_0\cap \mathrm{Fair}(F)\) or \(\mathrm{En}_c\cap \mathrm{Fair}(F)\); when it is imposed on player \(2\), player \(1\) wins any unfair play by player \(2\) [2501.17255]. The main algorithmic result is gadget-based reduction to ordinary mean-payoff or energy games. Both \(1\)-fair and \(2\)-fair mean-payoff games are determined, with pseudo-polynomial complexity \(O(n^4mW)\) for optimal-value computation or \(O(n^3mW)\) for threshold solving [2501.17255]. By contrast, \(2\)-fair energy games are not determined, although both winning regions remain computable [2501.17255].

These results make clear that, in verification, fairness is not primarily distributive. It is a liveness-style recurrence obligation on transitions, and its main technical role is to constrain admissible infinite behavior.

## 6. Procedural and applied notions of fair play

In autonomous racing, fair play is formalized as sportsmanship constraints. Huang et al. define a blocking predicate \(\mathrm{Block}(x_\tau^D,x_\tau^A)\), then impose a One-Motion Rule forbidding double-blocking over a four-step window and an Enough-Space Rule forbidding a later block when a faster attacker is already near the track boundary. High-level intentions are selected by a Stackelberg game solved by MCTS, while low-level trajectories are computed as a Generalized Nash Equilibrium Problem with shared dynamics, collision-avoidance, and velocity constraints [2503.03774]. In simulations with planning horizon \(T=15\) and \(\Delta t=0.4\,\mathrm s\), the “both aware” setting yields \(\Delta\bar x>0\) and zero violation, whereas an unaware defender blocks aggressively, produces high violation rates, and prevents successful overtaking [2503.03774]. Fairness here is a codified behavioral rule set, not an ex post outcome metric.

Procedural fairness also appears in redistricting. In Redistricting Ghost, two parties alternately place voters into districts of size \(2m+1\), with the minority moving first. If the state has \(j\) districts, \(v=j(2m+1)\) voters, and \(n\) minority voters, the proportional share is defined as
\[
p=\mathrm{round}(jn/v).
\]
The central guarantee is that the minority can always secure at least \(p-1\) districts, while the majority has a complementary cracking upper bound expressed through
\[
f(q)=2q\Bigl(1-\frac{q}{j+q}\Bigr)(m+1)-1
\]
[2401.07440]. In this setting, “fair game” refers to a protocol-level near-proportional guarantee under perfect information and no geographic constraints.

Finally, the fairness of AI-versus-human game benchmarks has been treated as a taxonomy rather than a theorem. The relevant dimensions are input, output, compute, knowledge, experience, psychology, and common sense. On this view, a contest is fair only when no non-game-relevant circumstance advantages either side [1903.07008]. This perspective is useful because it exposes a common ambiguity across the literature: “fair” may refer to equal payoffs, symmetric laws, rule compliance, equitable allocation, or parity of competitive conditions, and these are not interchangeable notions [1903.07008].

In that sense, the modern literature on fair games is best understood as a collection of formally precise but heterogeneous frameworks. Their common feature is not a universal fairness axiom, but the conversion of an intuitive fairness demand into a mathematically checkable object: an involution, a bifurcation threshold, a Shapley-proportional allocation, an auditor feedback loop, a Streett-like recurrence condition, or an explicit sportsmanship rule.

Source: https://www.emergentmind.com/topics/fair-game