Eq. (4): integral representation of the off-diagonal choice-probability partials
OpenStochFictPlay.DiscreteChoice.choiceProb_partial_integralFormulaLet have a continuous, everywhere strictly positive density on , and let be the choice probability function of the additive random utility model. Assume is continuously differentiable. Then for every payoff vector and every pair of distinct alternatives ,
the -fold integral of the density over the face where and tie for the maximum, with the -th coordinate pinned at (the change of variables of Hofbauer–Sandholm (2002), eq. (4), p. 6). Since everywhere, the integral is strictly positive and the off-diagonal partial is strictly negative; exchanging the roles of and gives the symmetry . This is the deep analytic input behind the symmetry/negativity of the derivative matrix (Theorem 2.1 of Hofbauer–Sandholm 2002), mission-connected as a child of the Thm 2.1 leaf exists_admissible_perturbation (d555a301) and the stated bridge used by dC_symm_offdiag_neg (131ee9c1).
import Mathlib
namespace StochFictPlay.DiscreteChoice
/-- `f : ℝⁿ → ℝ` is a strictly positive probability density of the utility-shock vector
`ε = (ε₁, …, εₙ)`: it is continuous, strictly positive at every point, and integrates to one
against Lebesgue measure. Continuity is how the formula (4) of Hofbauer–Sandholm (2002),
which evaluates `f` on hyperplanes, is read. -/
structure IsStrictlyPositiveDensity {n : ℕ} (f : (Fin n → ℝ) → ℝ) : Prop where
continuous : Continuous f
pos : ∀ x, 0 < f x
lintegral_eq_one :
MeasureTheory.lintegral MeasureTheory.volume (fun x => ENNReal.ofReal (f x)) = 1
/-- The law of the shock vector `ε` with density `f`: Lebesgue measure on `ℝⁿ` weighted by `f`. -/
noncomputable def noiseLaw {n : ℕ} (f : (Fin n → ℝ) → ℝ) : MeasureTheory.Measure (Fin n → ℝ) :=
MeasureTheory.Measure.withDensity MeasureTheory.volume (fun x => ENNReal.ofReal (f x))
/-- The choice probability function (1) of the additive random utility model,
`Cᵢ(π) = P(argmaxⱼ πⱼ + εⱼ = i)`: the probability, under the law of `ε`, that alternative `i`
has strictly the highest total payoff `πᵢ + εᵢ`. Ties have probability zero because `ε` has a
density. Hofbauer–Sandholm (2002), p. 4, eq. (1). -/
noncomputable def choiceProb {n : ℕ} (f : (Fin n → ℝ) → ℝ) (π : Fin n → ℝ) (i : Fin n) : ℝ :=
(noiseLaw f {e | ∀ j, j ≠ i → π j + e j < π i + e i}).toReal
end StochFictPlay.DiscreteChoice
namespace StochFictPlay.DiscreteChoice
/-- Eq. (4) (Hofbauer–Sandholm 2002, p. 6): for `i ≠ j`, the off-diagonal partial
`∂Cᵢ/∂πⱼ(π)` is the negative of the `(n - 1)`-fold integral of the shock density `f`
over the face where alternatives `i` and `j` tie for the maximum. The integral evaluates
`f` at the tie point `x j = π i + x i - π j` — the change of variables
`x̂ⱼ = πᵢ + xᵢ - πⱼ` that moves the moving boundary of `Cᵢ` onto the fixed
hyperplane where `i` and `j` are tied — and integrates over Lebesgue measure restricted
to `{x | π k + x k < π i + x i for all k ≠ i, j}`. Since `f > 0` everywhere, the
integral is strictly positive, so the partial is strictly negative. This is the deep
analytic input behind `dC_symm_offdiag_neg` (symmetry `∂Cᵢ/∂πⱼ = ∂Cⱼ/∂πᵢ` follows by
exchanging the roles of `i` and `j` in the same integral representation). -/
theorem choiceProb_partial_integralFormula {n : ℕ} (f : (Fin n → ℝ) → ℝ)
(hf : IsStrictlyPositiveDensity f)
(hC : ContDiff ℝ 1 (choiceProb f)) (π : Fin n → ℝ) (i j : Fin n) (hij : i ≠ j) :
fderiv ℝ (choiceProb f) π (Pi.single j 1) i =
-MeasureTheory.integral
(MeasureTheory.Measure.restrict MeasureTheory.volume
{x : Fin n → ℝ | ∀ k : Fin n, k ≠ i → k ≠ j → π k + x k < π i + x i})
(fun x : Fin n → ℝ => f (Function.update x j (π i + x i - π j))) := by sorry
end StochFictPlay.DiscreteChoice