1: \begin{abstract}

2: %     The expected signature of an $\R^d$-valued stochastic process is at the core of a number of recently introduced metrics between probability measures on path space. A variation on this popular theme is to consider the expected value of random Cartan developments of the stochastic process into some matrix Lie group $G_N$ of dimension $N$. Such developments are solutions to controlled differential equations, whose vector fields are drawn from some probability measure on Hom$(\R^d,\mathfrak{g}_N)$, $\mathfrak{g}_N$ the Lie algebra associated to $G_N$. In particular, \cite{PCFGAN} construct an MMD metric in the case $G_N=U(N;\C)$, the unitary group. The cost of computing this metric imposes limitations on both $N$ and the class of measures on Hom$(\R^d,\mathfrak{g}_N)$, so that in \cite{PCFGAN} only (adversarially learnt) finitely supported measures are considered. We propose an alternative approach, where we consider the asymptotic regime as $N\to\infty$. Leveraging tools from random matrix theory and free probability theory, we show that a deterministic normalised limit exists and is universal for a broad class of probability measures. Moreover, we prove that this limit solves a quadratic functional equation that appears not to have been previously been studied. We prove uniqueness of the solution to this equation and propose a convergent numerical scheme (with rates) for its computation. We also show how these techniques can be used to dispense with certain Gaussian and independence assumptions in a recent result in \cite{SKlimit}, where $G_N=\text{GL}(N,\R)$ the general linear group is considered.

3: % \end{abstract}

4: