We study the problem of estimating the parameters of a linear median regression without any assumption on the shape of the error distribution -- including no condition on the existence of moments -- allowing for heterogeneity (or heteroskedasticity) of unknown form, noncontinuous distributions,...