Characterisation of distributions through $\delta$-records and martingales

Given parameters $c>0, \delta\ne0$ and a sequence $(X_n)$ of real-valued, integrable, independent and identically $F$-distributed random variables, we characterise distributions $F$ such that $(N_n-cM_n)$ is a martingale, where $N_n$ denotes the number of observations $X_k$ among $X_1,\ldots,X_n$ such that $X_k>M_{k-1}+\delta$, called $\delta$-records, and $M_k=\max\{X_1,\ldots, X_k\}$. The problem is recast as $1-F(x+\delta)=c\int_{x}^{\infty}(1-F)(t)dt$, for $x\in T$, with $F(T)=1$. Unlike standard functional equations, where the equality must hold for all $x$ in a fixed set, our problem involves a domain that depends on $F$ itself, introducing complexity but allowing for more possibilities of solutions. We find the explicit expressions of all solutions when $\delta < 0$ and, when $\delta > 0$, for distributions with bounded support. In the unbounded support case, we focus attention on continuous and lattice distributions. In the continuous setting, with support $\mathbb{R}_+$, we reduce the problem to a delay differential equation, showing that, besides particular cases of the exponential distribution, mixtures of exponential and gamma distributions and many others are solutions as well. The lattice case, with support $\mathbb{Z}_+$ is treated analogously and reduced to the study of a difference equation. Analogous results are obtained; in particular, mixtures of geometric and negative binomial distributions are found to solve the problem.