The long term accuracy (i.e., scale factor) of a Doppler velocity log (DVL) for speed over seabed estimation is studied theoretically, using a simplified 2-D physical model which accurately describes the absorption and terrain biases, for the narrow beamwidths of existing DVLs. The model is used to assess performance limits of existing DVL designs and motivate new designs. A classical piston DVL, which averages the Doppler phase over the entire beam footprint, is compared with the piston of same diameter and operating frequency operated in a new range-gated mode, to reduce the spread in echo strength which biases the speed estimation. The range gating is shown to lead to a very significant reduction of both biases. As an example, the bias resulting from both effects is lower than 0.01% for a range-gated 150-kHz piston of 14-cm diameter operated at 500-m altitude, whereas the corresponding value is 0.8% for the same piston operated conventionally, and 0.15% for a conventional 600-kHz piston of 7-cm diameter. Thus the 150-kHz range-gated DVL has better accuracy than the conventional 600-kHz DVL but much greater operating range. To exploit this benefit of range gating, the direction of arrival of the seabed echoes from the gate must be accurately estimated in the piston frame, e.g., the gate must be centered at boresight to the piston with low bias. It is shown how this can be achieved by an interferometric DVL (InDVL) design which splits the piston into at least two phase centers, and centers the gate using the zero crossing of the differential Doppler phase. It is also shown that, over locally planar seabeds, the residual spread in Doppler phase over the range gate can be removed by replacing the autocorrelation currently used in wideband Doppler processing by the correlation of the seabed echo with a suitably time scaled copy of itself. Finally, a new source of DVL bias, known as diffraction bias, is shown to be the main bias limiting InDVL performance. This is due to a second-order 3-D effect related to the curvature of the range bins not accounted for by the simplified 2-D physical model. For the 150-kHz InDVL, this bias is close to 0.1%, making it the main factor limiting theoretical performance.