[1] Measured sea-ice draft exhibits variations on all scales. We regard draft profiles up to several hundred kilometers in length as being drawn from a stationary stochastic process. We focus on the estimation of the mean draft H of the process. This elementary statistic is typically computed from a profile segment of length L and has some uncertainty, or sampling error, that is quantified by its variance s L 2 . How efficiently can the variance of H be reduced by the use of more data, that is, by increasing L? Three properties of the data indicate the need for a non-standard statistical model: the variance s 2 L of H falls off more slowly than L À1 ; the autocorrelation sequence does not fall rapidly to zero; and the spectrum does not flatten off with decreasing wave number. These indicate that ice draft exhibits, as a fundamental geometric property, 'long-range dependence.' One good model for this dependence is a fractionally differenced process, whose variance s L 2 is proportional to L À1+2d . From submarine ice draft data in the Arctic Ocean, we find d = 0.27. Mean draft estimated from a 50-km sample has a sample standard deviation of 0.29 m; for 200 km, it is 0.21 m. Tabulated values provide the sample standard deviation s L for various values of L for samples both in a straight line and in a rosette or spoke pattern, allowing for the efficient design of observational programs to measure draft to a desired accuracy.