Direct write E-beam lithography at 0.25 micron ground rules will soon be a practical reality. Other technologies such as X-Ray lithography are also moving rapidly towards these dimensions. This im-plies overlay tolerances at the level of 1 sigma = 25 nm. Characterization of tool overlay performance (particularly systematic error components) in this regime is a challenging and important metrology problem. A combination of carefully designed test structures, high quality fabrication, and the best available measuring instrumentation will be a minimum requirement for adequately addressing this need. This paper describes a study of self-compensating differential linewidth structures for such overlay measurements. These structures require only relative (not absolute) dimensional measurements, and linewidth comparisons are made only over very short physical distances. Therefore, the results are in principle insensitive to systematic errors and time-dependent drift in the metrology tool, and to systematic variations in processing uniformity across a macroscopic substrate. Design and fabrication of the test structures is discussed. The test structures used were primarily designed for use with an electrical linewidth measurement tool, but several more "traditional" overlay test structures were also built in for comparison purposes. The structures were exposed on an IBM EL-3 direct write E-beam tool with 0.25 - 0.5 micron ground rule capability, and delineated by RIE of 200 nm of doped polysilicon. The processing was optimized to produce lines with sharply defined vertical edges. Such lines give excellent signals in optical and SEM linewidth measurement tools, and so are well suited for comparison of these techniques with electrical measurements. Comparisons of linewidth differences measured electrically, laser optically, and on an SEM show excellent agreement. The 1 sigma measurement error is found to be in the range 7 - 9 nm for all of the three tools investigated. This remarkable performance clearly demonstrates the advantages of self-compensating structures. It is concluded that, given high quality test structure fabrication, all of these techniques are adequate for the purpose. Therefore, considerations of cost, throughput, and the real estate demands of the test structures would govern a practical choice of measurement technique. These factors are discussed in general terms. The measured linewidth differences which include overlay errors are compared with a sample population of linewidth differences, from the same areas of the wafer, which include only processing and measurement errors. For a sample size of about 100, the variance of the overlay sample (sigma = 31 nm) is more than 5 times that of the reference sample (sigma = 13 nm). Thus, the overlay errors of interest account for more than 80% of the variance in the measured results. Overlay vector plots measured by the differential technique are compared with those measured from optical verniers and from "Nestled L's" (a non-self-compensating test structure).The agreement is poor. Reasons for the inadequacy of these alternative structures are discussed. The value of these high-resolution overlay measurements is illustrated by presentation of vector plots of systematic and random overlay errors for a high resolution IBM EL-3 tool.