Let f: R -> R be the probability density function of X. It fulfills int f(x) dx = 1 and we furthermore assume that f is differentiable.
We can write the identity as
int_a^oo x f(x) dx / int_a^oo f(x) dx = 2 a.
We rewrite this as
int_a^oo x f(x) dx = 2 a int_a^oo f(x) dx.
Taking the derivative with respect to a on both sides yields
- a f(a) = 2 int_a^oo f(x) dx + 2 a (- f(a)).
Taking another derivative yields
- f(a) - a f'(a) = - 2 f(a) - 2 f(a) - 2 a f'(a).
We rewrite this as
f'(a) = - 3 f(a) / a,
and furthermore
f'(a) / f(a) = - 3 / a.
We can write this as
(ln ○ f)'(a) = - 3 / a.
As far as I can tell, we can not find a solution on all of [0, oo).
Let's assume that failure is impossible up to an early time t. This means that we assume f(x) = 0 for all x < t, and that possibly f(t) > 0, meaning we have a jump discontinuity in the probability density at x = t.
So these are the probability densities that fulfill the "modified" Lindy effect, where failure is impossible before time t > 0.
I haven't double checked this so maybe there is a mistake, but I'm quite convinced that this line of reasoning leads to a unique distribution for each t > 0 and impossibility when t = 0.
So the Pareto distribution with alpha=2 produces the Lindy effect. Nice result! It looks like this generalises to E(X|X>a)=ka (with alpha depending on k).
We can write the identity as
We rewrite this as Taking the derivative with respect to a on both sides yields Taking another derivative yields We rewrite this as and furthermore We can write this as As far as I can tell, we can not find a solution on all of [0, oo). Let's assume that failure is impossible up to an early time t. This means that we assume f(x) = 0 for all x < t, and that possibly f(t) > 0, meaning we have a jump discontinuity in the probability density at x = t.We integrate from t to x
This evaluates to In particular Right now f(t) is still a free parameter, but recall that has to hold, so we find that and So these are the probability densities that fulfill the "modified" Lindy effect, where failure is impossible before time t > 0.I haven't double checked this so maybe there is a mistake, but I'm quite convinced that this line of reasoning leads to a unique distribution for each t > 0 and impossibility when t = 0.