The exposure distribution function in electron beam lithography, which is needed to perform proximity correction, is usually simulated by Monte Carlo techniques, assuming a Gaussian distribution of the primary beam. The resulting backscattered part of the exposure distribution is usually also fitted to a Gaussian term. In this paper we demonstrate a technique, using a very high contrast resist, whereby the normalized point exposure distribution can be measured experimentally, both on solid substrates which cause backscattering, and on thin substrates where backscattering is negligible. The data sets so obtained can be applied directly to proximity correction and represent the practical conditions met in pattern writing. Results are presented of the distributions obtained on silicon, gallium arsenide, and thin silicon nitride substrates at different beam energies. Significant deviations from the commonly assumed double Gaussian distributions are apparent. On GaAs substrates the backscatter distribution cannot adequately be described by a Gaussian function. Even on silicon a significant amount of exposure is found in the transition region between the two Gaussian terms. This deviation, which can be due to non-Gaussian tails in the primary beam and to forward scattering in the resist, must be taken into account for accurate proximity correction in most submicron lithography, and certainly on the sub-100 nm scale.