これは、ことが判明しエコノメトリカのケネス・小とハーベイ・ローゼン記事は 1981年にこれを示したが、非常に特殊な状況で結果が経済学のいくつかのトレーニングはもちろんのこと、掘りの多くを必要とせ。私はよりアクセスしやすい方法でそれを証明することにしました。
証明:を選択肢の数とする。ベクトルの値に応じてε = { ε 1、。。。、ε J }、関数マックスI(δ I + ε I)の異なる値をとります。最初の値に焦点εようマックスI(δ I + ε I)= δ 1 + ε 1。つまり、δを統合しますJϵ={ϵ1,...,ϵJ}maxi(δi+ϵi)ϵmaxi(δi+ϵi)=δ1+ϵ1δ1+ϵ1セットオーバーM1≡{ϵ:δ1+ϵ1>δj+ϵj,j≠1}。
Eϵ∈M1[maxi(δi+ϵi)]=∫∞−∞(δ1+ϵ1)f(ϵ1)[∫δ1+ϵ1−δ2−∞...∫δ1+ϵ1−δJ−∞f(ϵ2)...f(ϵJ)dϵ2...dϵJ]dϵ1=∫∞−∞(δ1+ϵ1)f(ϵ1)(∫δ1+ϵ1−δ2−∞f(ϵ2)dϵ2)...(∫δ1+ϵ1−δJ−∞f(ϵJ)dϵJ)dϵ1=∫∞−∞(δ1+ϵ1)f(ϵ1)F(δ1+ϵ1−δ2)...F(δ1+ϵ1−δJ)dϵ1
JE[maxi(δi+ϵi)]
E[maxi(δi+ϵi)]=∑iEϵ∈Mi[maxi(δi+ϵi)].
Now we apply the functional form of the Gumbel distribution. This gives
===Eϵ∈Mi[maxi(δi+ϵi)]=∫∞−∞(δi+ϵi)eμ−ϵie−eμ−ϵi∏j≠ie−eμ−ϵi+δj−δidϵi∫∞−∞(δi+ϵi)eμ−ϵi∏je−eμ−ϵi+δj−δidϵi∫∞−∞(δi+ϵi)eμ−ϵiexp{∑j−eμ−ϵi+δj−δi}dϵi∫∞−∞(δi+ϵi)eμ−ϵiexp{−eμ−ϵi∑jeδj−δi}dϵi
where the second step comes from collecting one of the exponentiated terms into the product, along with the fact that δj−δi=0 if i=j.
Now we define Di≡∑jeδj−δi, and make the substitution x=Dieμ−ϵi, so that dx=−Dieμ−ϵidϵi⇒−dxDi=eμ−ϵidϵi and ϵi=μ−log(xDi). Note that as ϵi approaches infinity, x approaches 0, and as ϵi approaches negative infinity, x approaches infinity.
==Eϵ∈Mi[maxi(δi+ϵi)]=∫0∞(δi+μ−log[xDi])(−1Di)exp{−x}dx1Di∫∞0(δi+μ−log[xDi])e−xdxδi+μDi∫∞0e−xdx−1Di∫∞0log[x]e−xdx+log[Di]Di∫∞0e−xdx
The Gamma Function is defined as Γ(t)=∫∞0xt−1e−xdx. For values of t which are positive integers, this is equivalent to Γ(t)=(t−1)!, so Γ(1)=0!=1. In addition, it is known that the Euler–Mascheroni constant, γ≈0.57722 satisfies
γ=−∫∞0log[x]e−xdx.
Applying these facts gives
Eϵ∈Mi[maxi(δi+ϵi)]=δi+μ+γ+log[Di]Di
Then we sum over i to get
E[maxi(δi+ϵi)]=∑iδi+μ+γ+log[Di]Di
Recall that Di=∑jeδj−δi=∑jeδjeδi. Notice that the familiar logit choice probabilities Pi=eδi∑jδj are inverses of the Di's, or in other words Pi=1/Di. Also note that ∑iPi=1. Then we have
E[maxi(δi+ϵi)]======∑iPi(δi+μ+γ+log[Di])(μ+γ)∑iPi+∑iPiδi+∑iPilog[Di]μ+γ+∑iPiδi+∑iPilog[∑jeδjeδi]μ+γ+∑iPiδi+∑iPilog[∑jeδj]−∑iPilog[eδi]μ+γ+∑iPiδi+log[∑jeδj]∑iPi−∑iPiδiμ+γ+log[∑jexp{δj}].
Q.E.D.