Let $f(x)$ be a log-concave probability distribution with support on $\mathbf{R}$ and that is symmetric around zero (for example a zero-mean Gaussian distribution). Denote by $F$ its cumulative density function. We have

  • $F(x)$ is log-concave
  • $F(-x) = 1 - F(x)$ (by symmetry around $0$)

I am trying to prove (or disprove) the following statement for all $x, y \in \mathbf{R}$ such that $x+y \le 0$:

$$ \frac{F(x + y)}{F(-(x +y))} \stackrel{?}{\ge} \frac{F(x)}{F(-x)} \cdot \frac{F(y)}{F(-y)} $$

My feeling is that the above inequality is correct for $x+y \le 0$, and reversed for $x+y \ge 0$. Here is my (unsuccessful, so far) attempt at trying to show this:

  1. Try to show something about the log-concavity of $$ \frac{1}{F(-x)} = \frac{1}{1 - F(x)}. $$ Letting $f'(x)$ be the derivative of the density function, I get $$ \left[ \log \left( \frac{1}{1 - F(x)} \right) \right]'' = f'(x) \frac{1}{1 - F(x)} - f(x)^2 \frac{1}{(1 - F(x))^2} < 0 $$ for $x \in [0, \infty)$ because $f'(x)$ is negative on that interval. This means the the function $1 / (1 - F(x))$ is log-concave on that interval.
  2. Use the the fact that the product of two log-concave functions remains log-concave, hence $$ \frac{F(x)}{F(-x)} $$ is log-concave on $[0, \infty)$
  3. Then... I don't really know. I was hoping to use log-concavity somehow. But I don't see how.

Does someone see how I could make progress in solving this question?

(Source of the question: I am trying to understand how a result developed in this paper (Eq. 2) can be generalized to non-logistic distributions.)


I realized that the inequality is incorrect. (Furthermore, the second derivative of $\log \left( \frac{1}{1 - F(X)} \right)$ derived above contains mistakes).

First of all, I should have probably looked at the log-concavity of $$ \frac{F(x)}{1 - F(x)} $$ directly. Looking at the second derivative of the $\log$, we find $$ \left[ \log \left( \frac{F(x)}{1 - F(x)} \right) \right]'' = \frac{f'(x)F(x)[1-F(x)] - f^2(x)[1 - 2F(x)]}{F^2(x) (1 - F(x))^2} $$ There is no reason to believe that this expression will always be negative, so a priori no way to claim that the function is log-concave.

Second, some simulations with $F(x) = \Phi(x)$ (the Gaussian CDF) show that the inequality does not hold.

Here is some Python code:

import matplotlib.pyplot as plt
import numpy as np
from scipy.stats import norm

x = 3
ys = np.linspace(-5, 5, num=200)

lhs = np.log(norm.cdf(x + ys) / (1 - norm.cdf(x + ys)))
rhs = ((np.log(norm.cdf(x) / (1 - norm.cdf(x))))
       + (np.log(norm.cdf(ys) / (1 - norm.cdf(ys)))))

plt.plot(ys, lhs, label="log(left-hand side)")
plt.plot(ys, rhs, label="log(right-hand side)")
plt.title("both sides of the ineq. as a fct of $y$ for $x = 3$")
plt.xlabel("value of $y$")
plt.legend(loc="upper left")

Matplotlib plot

Note to self: when possible, always run simulations to develop some intuition about the problem.


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.