This paper investigates time of arrival estimation via impulse radio ultra-wideband technology. A dense multipath channel is assumed and frequency-selective propagation effects are taken into account that make the individual received pulses different in shape from the transmitted monocycles. In these conditions a conventional correlation-based estimator cannot be used and other schemes must be resorted to. Assuming only an approximate knowledge of the received pulses' duration, a TOA estimator is derived making use of least mean square techniques. The algorithm operates on energy measurements made on a folded version of the received signal. Its performance is investigated by simulation under implementation conditions of different complexity. In particular, sub-Nyquist sampling rates are considered and analog-to-digital conversion with limited resolution is explored. Comparisons are made with other available schemes.