How can I prove $\log(\Phi(x))$ is concave, where $\Phi(x)$ is CDF of N(0, 1) ?

Since $\frac{d^2}{dx^2} \log \Phi(x) = \cfrac{\phi(x)[-x \Phi(x) - \phi(x)]}{(\Phi(x))^2}$, it is enough to show that $-x \Phi(x) - \phi(x) < 0$ for all $x \in \mathbb{R}$.

Here's a plot of $\log \Phi$: Graph of log Phi


This may be inaccurate but I consider like below:

  • $-x \Phi(x) - \phi(x)$ is decreasing since

\begin{align} \frac{d}{dx}( -x \Phi(x) - \phi(x)) &= -\Phi(x) - x \phi(x) + x \phi(x) \\ &= -\Phi(x) < 0 \end{align}

  • $\lim_{x \to -\infty} -x \Phi(x) - \phi(x) = 0$, since $\phi(-\infty)=0$ and by l'Hospital's rule

\begin{align} \lim_{x \to -\infty} -x \Phi(x) &= - \lim_{x \to -\infty} \cfrac{\Phi(x)}{\frac{1}{x}} \\ &= \lim_{x \to -\infty} x^2 \phi(x) \\ &= \frac{1}{\sqrt{2\pi}} \lim_{x \to -\infty} \frac{x^2}{\exp(\frac{x^2}{2})}\\ &= \frac{1}{\sqrt{2\pi}} \lim_{x \to -\infty} \frac{2}{\exp(\frac{x^2}{2})} \\ &= 0 \end{align}

Therefore, $-x \Phi(x) - \phi(x) < 0$ for all $x \in \mathbb{R}$

  • $\begingroup$ Your P.S. looks good to me. $\endgroup$
    – A.Γ.
    Dec 20, 2016 at 17:41

2 Answers 2


Instead of using the second derivative, it is easier to verify the log-concavity using the Prekopa-Leindler inequality.

Let $q$ be a log-concave density on $\mathbb{R}$ with corresponding CDF $Q$. For arbitrary $a, b \in \mathbb{R}$ and $\lambda \in [0, 1]$, set $t = \lambda a + (1 - \lambda)b$ and define the following three functions: $$\begin{align*} h(x) &= q(x) I\{x \le t\} \\ f(x) &= q(x) I\{x \le a\} \\ g(x) &= q(x) I\{x \le b\} \end{align*}$$

It is easy to see that these functions satisfy the inequality $$f(x)^\lambda g(y)^{1 - \lambda} \le h(\lambda x + (1 - \lambda) y)$$ for all $x, y \in \mathbb{R}$. This means we may apply the Prekopa-Leindler inequality: $$\left(\int f(x) \, dx\right)^\lambda \left(\int g(x) \, dx\right)^{1 - \lambda} \le \int h(x) \, dx$$

Taking the logarithm on both sides and applying the definition of the CDF $Q$ we now get $$\lambda \log Q(a) + (1 - \lambda) \log Q(b) \le \log Q(t) = \log Q(\lambda a + (1 - \lambda) b),$$ i.e. $Q$ is log-concave.


Alternatively to the (not easy to prove) result on preservation of log-concavity under integration one can take a straightforward approach. Here is the exercise from Boyd, Vandenberghe, Convex optimization with solution steps enter image description here


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .