Problem statement
Given a vector $\mathbf{x} \in \mathbb{R}^{K}$, where $x_{i} \neq x_{j}, \forall i \neq j$. Prove that:
\begin{equation} \sum_{k=1}^{K} \frac{\exp(x_{k})}{\prod_{i \neq k} (x_{k} - x_{i})} \ge 0. \end{equation}
If it is wrong, I would appreciate if a counter example is provided.
Background
I am reading the paper "The continuous categorical: a novel simplex-valued exponential family". In the paper, the authors proposed a continuous version for the categorical distribution. They then provided the normalizing constant. To calculate the log-likelihood, we need the log of that normalizing constant. However, I do not know how to prove that it is always positive, so that we can take a log. Here, I reformulate the expression of the normalizing constant for convenience. If one read the paper, we can analogize $\mathbf{x}$ as $\pmb{\eta}$. Note that although $\eta_{K} = 0$, we could simply consider it as any vector, and just shift $\eta_{i} - \eta_{K}$ to obtain $\pmb{\eta}$ as in the paper. The non-shifting version does not change the positivity nature of the expression of interest.
In addition, I tried the implementation provided by the authors, but it was unclear to me since the authors calculated the denominator $\propto (x_{k} - x_{i})$, not $(x_{i} - x_{k})$ as specified in the paper. Of course, it would lead to an additional factor $(-1)^{K - 1}$ in the denominator. This factor would then neutralizes the $(-1)^{K + 1}$. However, the implementation still have something with $(-1)^{K-1}$ (Note that the implementation denotes a variable $K$ as the value $(K - 1)$ in the paper.)
I did try the implementation by passing a vector as input to the provided function. However, it does not always return a value, but sometimes, not-a-number (NaN).