Title: General truncated linear statistics for the top eigenvalues of random matrices

Abstract: Invariant ensemble, which are characterised by the joint distribution of eigenvalues $P(\lambda_1,\ldots,\lambda_N)$, play a central role in random matrix theory. We consider the truncated linear statistics $L_K = \sum_{n=1}^K f(\lambda_n)$ with $1 \leq K \leq N$, $\lambda_1 > \lambda_2 > \cdots > \lambda_N$ and $f$ a given function. This quantity has been studied recently in the case where the function $f$ is monotonous. Here, we consider the general case, where this function can be non-monotonous. Motivated by the physics of cold atoms, we study the example $f(\lambda)=\lambda^2$ in the Gaussian ensembles of random matrix theory. Using the Coulomb gas method, we obtain the distribution of the truncated linear statistics, in the limit $N \to \infty$ and $K \to \infty$, with $\kappa = K/N$ fixed. We show that the distribution presents two essential singularities, which arise from two infinite order phase transitions for the underlying Coulomb gas. We further argue that this mechanism is universal, as it depends neither on the choice of the ensemble, nor on the function $f$.