Lemma 3 — under Axioms 1–5 the log-likelihood has a maximizer iff Axiom 6 holds
ProvedMcFadden1974.MLE.exists_maximizer_iff_axiom6Consider a conditional logit choice experiment: trials, trial with alternatives of attribute vectors , repeated times with alternative chosen times, and log-likelihood (18)
Suppose the full-rank Axiom 5 holds. Then has a global maximizer on if and only if Axiom 6 holds, that is, if and only if there is no nonzero with
Lemma 3 decides when the conditional logit maximum likelihood estimator exists in a finite sample: it fails exactly when some direction in parameter space makes every observed choice weakly best in its alternative set.
Formalization Note Axioms 1–4 are built into the model through the logit form (16) and the likelihood (18), so "Suppose Axioms 1–5 hold" becomes the model plus Axiom 5. Every trial is assumed observed (), as implicit in the paper; without it necessity fails. Axiom 5 is read at every ; its row space does not depend on .
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_MLE_Model import Definitions.Def_McFadden1974_MLE_Axioms
namespace McFadden1974.MLE
open scoped RealInnerProductSpace
/-- **Lemma 3** (McFadden, *Conditional Logit Analysis of Qualitative Choice Behavior*, in P. Zarembka (ed.),
*Frontiers in Econometrics*, Academic Press (1974), p. 116, PDF p. 12): "Suppose Axioms 1–5 hold. Then Axiom 6 is
necessary and sufficient for the existence of a vector θ maximizing L."
For conditional logit data with every trial observed (`R_n ≥ 1`) and satisfying the full-rank
Axiom 5, the log-likelihood (18) has a global maximizer on `ℝ^K` if and only if Axiom 6 holds.
**Formalization Note.** Axioms 1–3 (which give the logit form (12)) and Axiom 4 (linearity of
`v` in `θ`) are built into the model `Data`: the selection probabilities are (16) and `L` is (18).
"Suppose Axioms 1–5 hold" therefore becomes: the model `Data` plus `Axiom5`. `Data.observed` is
the standing assumption `R_n ≥ 1` for every trial; without it necessity fails. Axiom 5 is read at
every `θ`; its row space does not depend on `θ`. -/
theorem exists_maximizer_iff_axiom6
{K : ℕ} (d : Data K) (h5 : d.Axiom5) :
(∃ θhat : EuclideanSpace ℝ (Fin K), ∀ θ, d.L θ ≤ d.L θhat) ↔ d.Axiom6 := by sorry
end McFadden1974.MLE
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.