EDIT: It turns out I just made a dumb mistake:


A portion of my "factor" is actually inside the derivative. Thanks to everyone who helped!

I'm using Mathematica to try to solve: https://math.stackexchange.com/questions/1700486

The probability distribution function (PDF) of the maximum of n standard normally distributed variables is:

(2^(1/2 - n)*n*(1 + Erf[x/Sqrt[2]])^(-1 + n))/(E^(x^2/2)*Sqrt[Pi]) 

To find the mode, I take the derivative and set equal to 0. The "raw" derivative is:

(2^(1 - n)*(-1 + n)*n*(1 + Erf[x/Sqrt[2]])^(-2 + n))/(E^x^2*Pi) -  
 (2^(1/2 - n)*n*x*(1 + Erf[x/Sqrt[2]])^(-1 + n))/(E^(x^2/2)*Sqrt[Pi]) 

and Simplify will reduce it to:

-((n*(1 + Erf[x/Sqrt[2]])^(-2 + n)*(2 - 2*n + E^(x^2/2)*Sqrt[2*Pi]*x +  

Since I only need to see when this value is $0$, I can multiply it by anything that's not $0$ (or undefined), in particular:

Exp[x^2] 2^n Pi/n (1 + Erf[x/Sqrt[2]])^(2 - n)

Multiplying the simplified derivative by that quantity yields:

2^n E^x^2 Pi (2^(1 - n)/(E^x^2 Pi) - (2^(1/2 - n) x (1 + Erf[x/Sqrt[2]])) /
 (E^(x^2/2) Sqrt[Pi])) 

Simplifying again, we have:

2 - E^(x^2/2) Sqrt[2 Pi] x - E^(x^2/2) Sqrt[2 Pi] x Erf[x/Sqrt[2]]

The problem? This value no longer depends on $n$.

However, graphing the PDF for various values of $n$ shows that the mode changes:

enter image description here


I wonder if your calculations have gone astray somewhere. I seem to get a different result.

Below pdf is your expression for the PDF you are working with, factor is the non-zero expression by which you want to multiply your derivative:

pdf = (2^(1/2 - n)*n*(1 + Erf[x/Sqrt[2]])^(-1 + n))/(E^(x^2/2)*Sqrt[Pi]);
factor = Exp[x^2]*2^n*Pi/n*(1 + Erf[x/Sqrt[2]])^(2 - n);

Calculating D[pdf, x] // Simplify agrees with your raw and simplified results in the question. However, multiplying by your factor and simplifying did not remove the dependence on n:

D[pdf, x] factor // Simplify

Mathematica graphics

Solve[D[pdf, x] factor == 0, n]

Mathematica graphics

  • $\begingroup$ I was about to write an answer, but I'll leave this comment instead: have a look at FullSimplify[D[PDF[OrderDistribution[{NormalDistribution[], n}, n], x], x]]. $\endgroup$ – J. M. will be back soon Mar 20 '16 at 16:25
  • $\begingroup$ Thanks, @MarcoB! As a minor note, finding the value of n for where the derivative is 0 is quite interesting, but I think I need the solution for x to find the mode :) $\endgroup$ – barrycarter Mar 20 '16 at 17:27
  • $\begingroup$ Something like With[{n = 25}, x /. FindRoot[-2 + 2 n + Exp[x^2/2] Sqrt[2 π] x (-2 + Erfc[x/Sqrt[2]]), {x, -Sqrt[2] InverseErfc[2^(1 - 1/n)]}]] then, @barry? $\endgroup$ – J. M. will be back soon Mar 20 '16 at 17:47
  • $\begingroup$ Yes, that's what I had been doing, and was hoping to find a closed form solution, but no luck, alas. github.com/barrycarter/bcapps/blob/master/STACK/bc-prob-max.m for the morbidly curious :) $\endgroup$ – barrycarter Mar 20 '16 at 17:53

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.