This study examines seismic discrimination between underground nuclear explosions and earthquakes in the Southwestern United States. A thorough review of theoretical and applied research on this problem, especially as it relates to that region, is presented first, followed by a presentation of the seismic discrimination parameters computed for a suite of events in the Southwest and a series of experiments with multi-dimensional discrimination.Review of past work finds some theoretical support for successful discrimination based upon several distinct measurements on seismic recordings.Although negative first motion and presence of pP both indicate a natural earthquake source, they cannot be found confidently in many cases. The rela-* deeper earthquakes could be identified by higher modes and a relative decrease of shorter period amplitude. The region under study extends from California to the southern Rocky Mountains and from roughly 40*N to the Gulf of California. A region of high seismicity, it is also a complex region encompassing several tectonic provinces, including the Basin-Range Province where the Nevada Test Site is located. Source mechanisms are diverse over the entire region and even within small subregions. Focal depths are almost all < 15 km and few depths can be determined reliably from teleseismic recordings. Twelve earthquakes, -3-mostly selected because of low M estimates from other studies, and eleven s explosions comprised the sources for this study. Seven teleseismic LRSM stations supplied the seismic data.In general, seismic discriminant parameters obtained from the recordings reflected the theoretical expectations of earthquake-explosion differences.Path-station effects were large for every parameter, especially for shortperiod data. Love-wave magnitudes were a superior discriminant when plotted versus mb. Shear waves, if measurable, were also excellent. The common Mmb plot shows three earthquakes with anomalously low Ms: the Denver earthquake, a Benham aftershock, and a Baja California earthquake.All the discrimination parameters (10) were applied in various combinations through a stepwise linear discriminant program, which treated the twentythree events as a training set. Using all parameters, the lowest a posteriori probability of correct classification was .987. Multi-dimensional discrimination using network averages for parameters was superior to using single-station parameters or using linear combinations of individual stations' discriminant functions.