1
$\begingroup$

consider the following sorted sample:

$x_1<x_2<\ldots<x_n$

and the kernel function:

$h(j)=\log(x_j-x_1), j>1$.

Now, it turns out that when I draw data from many continuous distributions, sort them and compute $h(j)$, the plot of the $h(j)$ as a function of $j$ is concave in this sense: $h((j+k)/2)>(h(j)+h(k))/2$ for $1<j<j+1<k<n$

My question is the following: will this observation (concavity of $h(j)$ as a function of $j$) be true for all continuous distributions? If not is there a class of distributions for which it will always be true?

Thanks in advance,

At this point, given the useful comments from @Glen_b and @Momo, I think I should give an example.

this is a vector of values of $x_j-x_1,j>1$ (where the $x$'s are drawn from a continuous distribution, but have been rounded here to the third digit so that they don't take too much place.)

My problem, is that when I plot $h(j)$ for this data-set, the result is clearly nor concave (in the sense I outlined above).

x=c(1.403, 1.406, 1.408, 1.416, 1.417, 1.42, 1.427, 1.441, 1.448, 
1.456, 1.458, 1.465, 1.466, 1.472, 1.472, 1.477, 1.477, 1.479, 
1.482, 1.491, 1.5, 1.504, 1.518, 1.52, 1.52, 1.53, 1.544, 1.545, 
1.561, 1.573, 1.595, 1.595, 1.599, 1.603, 1.605, 1.612, 1.617, 
1.618, 1.618, 1.628, 1.64, 1.644, 1.646, 1.653, 1.679, 1.682, 
1.682, 1.693, 1.694, 1.71, 1.71, 1.711, 1.741, 1.75, 1.756, 1.773, 
1.794, 1.799, 1.804, 1.808, 1.86, 1.882, 1.895, 1.955, 1.992, 
1.995, 2.009, 2.009, 2.063, 2.123, 2.329, 2.356, 2.405, 2.535, 
2.632, 2.635, 2.725, 2.763, 2.783)
$\endgroup$
  • 1
    $\begingroup$ I'm curious where the distribution comes in here. Is $x_i$ a random variable and possibky different for each $i$ or a realisation and thus a real number ? $\endgroup$ – Momo Mar 22 '14 at 22:17
  • $\begingroup$ @Momo: thanks for your question but I do not understand it. Can you try to re-phrase please? Thanks. $\endgroup$ – user603 Mar 22 '14 at 22:54
  • $\begingroup$ I'll try: You talk about continuous distributions. I wonder if $x$ is a random variable with an arbitrary distribution or if it is a real number. $\endgroup$ – Momo Mar 22 '14 at 23:41
  • 2
    $\begingroup$ The concavity of the function $h$ doesn't depend on the distribution of the values of the arguments you apply it to. Please clarify your question. $\endgroup$ – Glen_b -Reinstate Monica Mar 23 '14 at 0:01
  • $\begingroup$ @Glen_b: thanks for your comment. Are you stating that so long as the xi are drawn from a continuous distribution, h(j) will always be concave, in the sense made precise in my question? Sorry if the question sounds naive, I'm not a statistician. $\endgroup$ – user603 Mar 24 '14 at 20:35
1
$\begingroup$

UPDATED:

the problem is related to transformation which preserves concavity of the log function. here we transform an index $j$ into real numbers $x(j)$, then transform the real numbers with $\ln(.)$ function, i.e. $h(j) = \ln(x(j))$ , $ln(.)$ is concave, now the transformation $x(j)$ must be at least not more convex than $\ln(.)$ is concave.

e.g. $e^{x^2}$ is too convex because the sequence $\{0,1,2\}$ becomes $h(j)=\{0,1,\ln(4)\}$, which is convex.

in this case $x(j)$ transformation is not a simple function, but a probabilistic one. it seems to me that for a large sample probability to get a concave sequence should tend to be zero for any distribution.

$\endgroup$
  • $\begingroup$ This is incorrect. Take, for instance, the dataset $(0,e,e,e^2)$, for which $h(2)=\log(e-0)=1,h(3)=\log(e-0)=1,h(4)=\log(e^2-0)=2$, whence $1=h(3)\lt(h(2)+h(4))/2=3/2,$ contradicting your assertion. $\endgroup$ – whuber Mar 25 '14 at 18:32
  • $\begingroup$ $x_j-x_1=\{e,2e,2e+e^2\}$ not $x_j-x_1=\{e,e,e^2\}$ $\endgroup$ – Aksakal Mar 25 '14 at 18:35
  • 1
    $\begingroup$ Your are responding to an earlier version of my comment, which I fixed while you were posting yours. But even so, if you do the calculation with $(e, 2e, 2e+e^2)$ you will still find this is a counterexample. The point is that concavity of the log is irrelevant when the subject concerns the values of the finite sequence $(x_j-x_1).$ $\endgroup$ – whuber Mar 25 '14 at 18:38
  • 2
    $\begingroup$ got it. so the problem then is $h(j)=\ln(x(j))$ , $\ln$ is concave, now the transformation $x(j)$ must be at least not more convex than $\ln$ is concave $\endgroup$ – Aksakal Mar 25 '14 at 18:42
  • 1
    $\begingroup$ That is an excellent analysis: you got to the heart of it. $\endgroup$ – whuber Mar 25 '14 at 20:01

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.