I'm trying to test if $ f : \mathbb{R}^n\rightarrow \mathbb{R} $ given by :

$$ f(x) = \ln \left(\sum\limits_{i=1}^n e^{x_i}\right) $$ is convex.

I tried computing positivity of the Hessian $H$ matrix by definition, multiplying by $a \in \mathbb{R}^n $. I got the expression

$$ a^T H a = \displaystyle\sum_{i=1}^n a_i^2 e^{x_i} -(a_1 + ... + a_n)\frac{a_1 e^{x_1} + ... + a_n e^{x_n}}{e^{x_1} + ... + e^{x_n} } $$

Where $ a_i $ is the $i$-th coordinate of $a$.

Is there a way to check if this expression is higher than zero? I can't see if an averages inequality would help.


Let $\lambda, \mu \in \langle 0,1\rangle$ be such that $\lambda + \mu = 1$ and $x = (x_1, \ldots, x_n), y= (y_1, \ldots, y_n) \in \mathbb{R}^n$. We wish to show that $$f(\lambda x + \mu y) \le \lambda f(x) + \mu f(y)$$

We shall use Hölder's inequality for conjugate exponents $\frac1{\lambda}$ and $\frac{1}\mu$. We have:

\begin{align} f(\lambda x + \mu y) &= \ln \sum_{i=1}^n e^{\lambda x_i + \mu y_i} \\ &= \ln \sum_{i=1}^n e^{\lambda x_i}e^{\mu y_i}\\ &\le \ln\, \left[\left(\sum_{i=1}^n e^{\lambda x_i\cdot \frac1\lambda}\right)^{\lambda}\left(\sum_{i=1}^n e^{\mu y_i\cdot \frac1\mu}\right)^{\mu}\right]\\ &= \lambda \ln \sum_{i=1}^n e^{x_i} + \mu \ln \sum_{i=1}^n e^{y_i}\\ &= \lambda f(x) + \mu f(y) \end{align}

Hence, $f$ is convex on $\mathbb{R}^2$.

  • $\begingroup$ Awesome! Thank you so much! $\endgroup$ – Mateus Nov 16 '17 at 11:23

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.