The correct migration of seismic data depends on the accuracy of the chosen velocity model. Rocca and Salvador (1982) showed that small errors in the velocity model may be efficiently corrected by applying a residual migration to previously migrated data, rather than remigrating the original data with a corrected velocity field. The effective velocity used in this residual processing is usually small compared to the original migration velocity. This decreases computational cost relative to a full migration, and allows the initial migration to be done with a less accurate but faster algorithm than would otherwise be required. The possible advantages are many. The overall cost of migration may be reduced, a consideration especially important when migrating 3-D data sets. Migration quality may be improved, because the location of mispositioned reflectors can be corrected and because of the freedom to choose initial migration with a high dip, low dispersion method such as Stolt migration. Interactive residual sharpening of the migrated image also becomes feasible. We discuss the theoretical and practical limitations of residual migration and quantify the related reductions of effective dip, velocity, and frequency after initial migration. We determine how accurate the initial migration velocity must be to justify use of this approach and analyze aliasing and numerical artifacts. Field data examples using Kirchhoff summation and finite-difference migration illustrate the features and drawbacks of the method.