Source
Balduzzi, Garnelo, Bachrach, Czarnecki, Pérolat, Jaderberg & Graepel, LIT-1509, read in NOTE-1143 from arXiv:1901.08106v2: Definition 3 and Proposition 5 (p. 5, proof p. 13), Proposition 2 (p. 4) and Definitions 6–7 (proof p. 12).
What was actually shown
- Definition. For populations P and Q, v(P, Q) = pᵀA_{P,Q}q at a Nash pair (p, q) of the zero-sum game with payoff φ(v, w), v ∈ P, w ∈ Q.
- Independence of the choice. A zero-sum game has one value whichever equilibrium attains it (Proposition 5(i)).
- Redundancy. If an added agent’s results against every member equal those of a mixture of members, every mixture available after the addition was available before, so the value is unchanged. Proposition 2 shows the gamescape is unchanged up to an isometric embedding. The reader made the step to the value: a game’s value depends only on the sets of payoff vectors each side can mix, and those sets are unchanged.
- Containment (Proposition 5(iii)). If every agent of P is behaviourally a mixture of agents of Q, Q can reproduce any mixture P plays. By antisymmetry the mirrored play is worth zero, so v(P, Q) ≤ 0, and v(P, R) ≤ v(Q, R) for every R. The paper writes hull(P) ⊂ hull(Q); the proof needs it read behaviourally, not in weight space.
What could have come out otherwise: a population score that rewards numbers, such as a mean over members or a sum of wins, rises when members are duplicated. This one does not.
What this does not say
- That it ranks individuals. It compares populations. In a cyclic game the paper holds that an individual’s performance is meaningless.
- That populations are totally ordered. Only hull-nested populations are ordered by it. Singletons in rock-paper-scissors are populations, and their values cycle.
- That it detects strength a population does not exhibit. It is computed from the evaluation matrix of the agents present. The paper’s concentric rock-paper-scissors populations all score 0 against each other, though the outer one exploits a fixed opponent better.
- Anything ordinal. The value uses payoff magnitudes.
Connections
Of the scoring rules in this group it is the one that a padded population cannot move without any further condition. Compare THEORY-946, which needs a unique equilibrium and covers exact copies only. For CLAIM-451 it is the formal shape of relief by mixing: evaluate the portfolio, not the option (NOTE-1143).
Note of 2026-10-30: the second check
The second check at the status review (ADR-065 §5) followed each step.
Independence. By the minimax theorem every Nash pair of a finite zero-sum game attains max_p min_q pᵀMq = min_q max_p pᵀMq, so the value does not depend on which pair is used.
Redundancy. The value is the maximum, over the convex hull of P’s payoff vectors against Q, of the smallest entry, and dually for Q’s columns. An agent x added to P with φ(x, w) = Σ_k λ_k φ(m_k, w) for every w has its payoff vector inside the hull of P’s, so the hull and the value are unchanged. An agent y added to Q equal to a mixture μ of Q’s members has, by antisymmetry, φ(v, y) = −φ(y, v) = −Σ_k μ_k φ(q_k, v) = Σ_k μ_k φ(v, q_k), a mixture of Q’s columns, so again nothing changes.
Containment. Let each v_i in P satisfy φ(v_i, w) = Σ_k λ_ik φ(q_k, w) for every w. For any mixture p over P put ν = λᵀp, a mixture over Q. If Q plays ν, P’s payoff is pᵀA_{P,Q}ν = νᵀA_{Q,Q}ν = 0 by antisymmetry, so v(P, Q) ≤ 0. P’s payoff vector against any R equals ν’s, so v(P, R) ≤ v(Q, R).
A script (numpy, with an exact linear-programming solver, in the session’s scratchpad) checked these on 300 random antisymmetric games of 4 to 8 agents split into two populations. Adding an agent that behaved as a random mixture of one side’s members left the value unchanged to 10⁻⁹, on either side. Three such mixture agents of Q, taken as a population, scored at most 0 against Q and no more than Q against a third population. Two controls moved: a mean-over-members score went from 0.086 to −0.250 when a member was duplicated, and adding an agent that beats every member of Q moved the value from 0.444 to 1.750.