Scenario 1: Let $X_i,i\in I$ where $I$ is the index set. For simplicity let $I$ be the finite set $I=\{1,2,\ldots,n\}$. Let each $X_i$ be independent and distributed as $$ X_i\sim N(\delta,1), $$ and let $\mathbf{X}=(X_1,X_2,\ldots,X_n)$. Define the index estimator $$ \gamma(\mathbf{X})=\text{argmax}_{i\in I}{\mathbf{(X)}} $$ The distribution of $\gamma(\mathbf{X})$ is uniform. This is easy to see because each $X_i$ is equally likely to be the maximum since they are all independent and identically distributed.

Scenario 2: In this scenario, $X_i$ are still independent, however, not identically distributed. Each $X_i$ has the distribution $$ X_i\sim N(\delta_i,1), $$ and $\gamma(\mathbf X)$ has the same definition, i.e., $$ \gamma(\mathbf{X})=\text{argmax}_{i\in I}{\mathbf{(X)}}. $$ My question is what is the distribution of $\gamma(\mathbf{X})$ in this scenario? I started with \begin{eqnarray} \Pr(\gamma(\mathbf X)=i)&=&\Pr(X_i=\max\{X_1,X_2,\ldots,X_n\})\\ &=& \Pr(X_i\geq\max\{X_k:k\neq i,k=1,2,\ldots,n\})\\ \end{eqnarray} And define $Y=\max\{X_k:k\neq i,k=1,2,\ldots,n\}$ which implies $$ \Pr(Y\leq y)=\prod_{k \neq i}\Pr(X_k\leq y)=\prod_{k \neq i}\Phi(y) $$ Then my problem simplifies to $$ \Pr(\gamma(\mathbf X)=i)=\Pr(X_i-Y\geq 0). $$ It is the distribution of $X_i-Y$ that makes life complicated for me. However, if you can proceed in another way it might be easier.


For $n = 2$, $p(\gamma(x) = 1) = p(x_{1} - x_{2} > 0)$. Let $z_{1,2} = x_{1} - x_{2}$. Obviously, $z_{1,2} \sim N(\delta_{1} - \delta_{2}, 2)$. Hence, $p(\gamma(x) = 1) = 1 - \Phi((\delta_{1} - \delta_{2}) / \sqrt{2})$.

For $n > 2$, we use multivariate normal distribution. Take $n = 3$ for example. $p(\gamma(x) = 3) = p(x_{3} - x_{2} > 0, x_{3} - x_{1} > 0)$. Let $z_{3,2} = x_{3} - x_{2}$ and $z_{3,1} = x_{3} - x_{1}$. Of course, the vector $(z_{3,2}, z_{3,1})$ follows bivariate normal: \begin{equation*} (z_{3,2}, z_{3,1})' \sim N \left( (\delta_{3} - \delta_{2}, \delta_{3} - \delta_{1})', \begin{pmatrix} 2 & 1 \\ 1 & 2 \end{pmatrix} \right) \end{equation*} Then the probability of $p(\gamma(x) = 3) = p(z_{3,2} > 0, z_{3,1} > 0)$ can be calculated from the CDF of the above bivariate normal distribution. The general case of $n > 2$ is similar. You will define a $n-1$-dimensional multivariate normal, whose covariance matrix has diagonal elements of $2$ and off-diagonal elements of $1$.

  • $\begingroup$ Something counter-intuitive is going, however, consider $n=2$ and if $\delta_1>\delta_{2}$, I will expect $p(\gamma(x)=1)>p(\gamma(x)=2)$, somehow the reverse is true. What are we missing? $\endgroup$ – Chamberlain Foncha Mar 8 '18 at 17:31
  • $\begingroup$ Ok I see the error you omitted a negative sign. $\endgroup$ – Chamberlain Foncha Mar 8 '18 at 18:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.