Lemmas 1 and 2 — logit selection probabilities if and only if extreme value shocks
ProvedMcFadden1974.RandomUtility.logit_iff_extreme_valueFix a universe and a surjective representative utility map . Let the taste shocks of the random utility model be independent and identically distributed with a translation complete law on whose distribution function satisfies the normalization . Then the following are equivalent:
- for every finite subset and every alternative ,
- for every real .
This is the characterization announced on p. 111 of the paper: "under mild conditions the distribution (13) characterizes the population choice models whose selection probabilities satisfy Equation (12)". It justifies reading a conditional logit model as a model of utility-maximizing individuals with extreme value taste shocks.
Formalization Note Direction 2 ⇒ 1 is Lemma 1 and needs neither translation completeness nor the normalization; 1 ⇒ 2 is Lemma 2. The hypotheses are satisfiable: the extreme value law is translation complete (footnote 5 of the paper) and has . Condition 1 is required for every finite subset of the stated universe; distinct alternatives may have equal utilities.
Retired 2026-10-27: a Lean-vs-paper fidelity audit found this statement weaker than, or narrower than, the cited result. A corrected statement will be resubmitted.
import Mathlib import Definitions.Def_McFadden1974_RandomUtility_Model open MeasureTheory ProbabilityTheory Finset
namespace McFadden1974.RandomUtility
/-- **Lemmas 1 and 2 — the extreme value law characterizes logit random utility models**
(McFadden, Conditional Logit Analysis of Qualitative Choice Behavior, in Frontiers in Econometrics
(1974), pp. 111–112, Lemmas 1 and 2; PDF pp. 7–8; p. 111: "The next lemma establishes that under
mild conditions the distribution (13) characterizes the population choice models whose selection
probabilities satisfy Equation (12).").
Fix an alternative universe `X` and a representative utility map `u : X → ℝ` that is onto.
Let the shocks be i.i.d. with a translation complete law `μ` normalized by `G(0) = e^{−1}`.
Then the selection probabilities on every finite subset of `X` satisfy (12) if and only if
`G(ε) = e^{−e^{−ε}}` for every `ε`.
Formalization Note: the direction ← is Lemma 1 (which needs neither translation completeness nor
the normalization); → is Lemma 2 with the normalization `G(0) = e^{−1}`. The hypotheses are
satisfiable: the law (13) is translation complete (footnote 5, p. 111) and has `G(0) = e^{−1}`.
The selection probabilities are those of (2), with strict inequalities and no density assumed;
`IsLogitOn` covers finite subsets of `X`, including repeated utility values. -/
theorem logit_iff_extreme_value (μ : Measure ℝ) [IsProbabilityMeasure μ]
{X : Type*} (u : X → ℝ) (hu : Function.Surjective u)
(hTC : TranslationComplete μ) (h0 : cdf μ 0 = Real.exp (-1)) :
IsLogitOn μ u ↔ IsExtremeValue μ := by sorry
end McFadden1974.RandomUtility
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.