Phase steps are an important type of wavefront aberration generated by large telescopes with segmented mirrors. In a closed-loop correction cycle these phase steps have to be measured with the highest possible precision using natural reference stars, that is with a small number of photons. In this paper the classical Fisher information of statistics is used for calculating the Cramér–Rao minimum variance bound, which determines the limit to the precision with which the height of the steps can be estimated in an unbiased fashion with a given number of photons and a given measuring device. Four types of such measurement devices are discussed: a Shack–Hartmann sensor with one small cylindrical lenslet covering a subaperture centred on a border, a modified Mach–Zehnder interferometer, a Foucault test, and a curvature sensor. The Cramér–Rao bound is calculated for all sensors under ideal conditions, that is narrowband measurements including photon shot noise, but without other forms of noise or disturbances. This limit is compared with the ultimate quantum statistical limit for the estimate of such a step, which is independent of any measuring device. For one device, the Shack–Hartmann sensor, the effects on the Cramér–Rao bound of broadband measurements, finite sampling, and disturbances such as atmospheric seeing and detector readout noise are also investigated. The methods presented here can be used to compare the precision limits of various devices for measuring differential segment phases, and for optimizing the devices. Under ideal conditions the Shack–Hartmann and the Foucault devices nearly attain the ultimate quantum statistical limit, whereas the Mach–Zehnder and the curvature devices each require approximately twenty times as many photons in order to reach the same precision.