Convex Combination of Simplex Points
The basic vocabulary for blending two mixed strategies. Given $x, y \in \Delta(I)$ and $\alpha \in [0, 1]$, the mixed point
$$ \mathrm{mix}(\alpha; x, y) \;=\; \alpha \cdot x + (1 - \alpha) \cdot y $$
is again a simplex point. This is the substrate for compound lotteries (MSZ Axiom 2.16), the inductive perturbation step in Loomis-style minimax proofs, and any algorithm that interpolates between two distributions (fictitious play, no-regret dynamics, Sion-style continuity arguments).
Statement
For a finite index I, a strictly ordered field 𝕜, and α ∈ [0, 1]:
stdSimplex.mix α hα₀ hα₁ x yis the element ofstdSimplex 𝕜 Iwhosei-th coordinate is $\alpha \cdot x_i + (1 - \alpha) \cdot y_i$.- Bilinearity of
wsum(wsum_mix): for anyf : I → 𝕜, $$ \operatorname{wsum}(\mathrm{mix}(\alpha; x, y), f) \;=\; \alpha \cdot \operatorname{wsum}(x, f) + (1 - \alpha) \cdot \operatorname{wsum}(y, f). $$ - Strict-monotonicity over a convex combination (
linear_comb_gt_of_ge_gt,linear_comb_lt_of_le_lt, and theirwsumversionswsum_mix_gt_of_ge_gtandwsum_mix_lt_of_le_lt): if one endpoint of the combination satisfies a non-strict bound and the other a strict bound, the combination remains strict for anyα < 1. - Neighborhood-of-1 existential (
mix_gt_of_gt_nbhand dualmix_lt_of_lt_nbh): ifc < x, there is an interiort ∈ (0, 1)withc < t · x + (1 - t) · y. The proof appeals to continuity of $s \mapsto s \cdot x + (1 - s) \cdot y$ at $s = 1$.
Hypothesis convention
Hypotheses for α are passed as plain (α : 𝕜) (hα₀ : 0 ≤ α) (hα₁ : α ≤ 1)
rather than via a unit-interval subtype, matching Mathlib idioms and avoiding
a Core-level subtype declaration that would not pay for itself. The legacy
Interval := { t : ℝ // 0 ≤ t ∧ t ≤ 1 } subtype, previously defined in
StrategicGame/MinimaxLoomis.lean, has been retired in favor of this
convention.
Use
EconCSLib.Utility.Lottery:Lottery.mixis a definitional alias ofstdSimplex.mix;Lottery.expectedValue_mixis a one-line wrapper aroundwsum_mix.EconCSLib.StrategicGame.MinimaxLoomis(simplified Loomis induction) andEconCSLib.StrategicGame.Loomis(general-BLoomis): the inductive step "blend the existing optimiser with the optimiser of the smaller restricted game" is the canonical consumer ofstdSimplex.mix,wsum_mix,wsum_mix_*_of_*, andmix_*_nbh.
References
- [MSZ] Maschler, Solan, Zamir, Game Theory, Axiom 2.16 and the surrounding discussion of compound lotteries.
- [MFoGT] Laraki, Renault, Sorin, Mathematical Foundations of Game Theory, Section 2.5 (Loomis induction).