Two multivariate classification methods have been applied to discriminant variables measured on a set of 20 Nevada Test Site (NTS) underground nuclear explosions and 27 North American Earthquakes. The classification methods used in this study are cluster analysis, and linear and quadratic multiple discriminant analysis. Discriminant analysis requires that the parameters of the underlying stochastic model be specified a priori, which is usually accomplished with a sample training set of events. Cluster analysis, however, follows a non-parametric approach and requires no prior information. The discriminant variables are the seven energy ratios calculated for these events by Booker and Mitronovas (1964). An additional discriminant variable, M(s)/m(b) was estimated for a reduced set of nine earthquakes and 21 underground explosions. Four M(s)/m(b) estimates were obtained for each event in this subset by randomly generating M(s) estimates within four specified variance levels about the M(s) versus M(b) lines given by Alexander and Lambert (1973). The variances used in generating the M(s) estimates are multiples of the variance found by von Seggern (1972) to be representative of M(s) measurements for NTS explosions. (Author).

Some procedures for discriminating between earthquakes (Q) and explosions (X) set aside a region of the discrimination parameter space (x) in which no decision is made; this may be called the "unidentified" or "undecided" region, (U). Tile existing statistical literature seems not to explicitly provide any such option, although an undecided region may be rigorously supported by the decision theory literature on "action spaces." In this report we show how the concept of U arises naturally from the concept of the costs of classification: positive costs from misclassifications and "no decisions", and negative costs (benefits) from correct classifications. The resulting approach is a generalization of the "Expected Cost of Misclassification" (ECM) approach; we call the generalization the "Expected Cost of Classification" (ECC) approach. We also show how thresholds for detecting X as outliers of a Q population may be derived from cost considerations together with uniform distributions for X, and, together with plausible prior probabilities for Q and X, lead to reasonable thresholds in realistic scenarios.

The subject of this NATO Advanced Study Institute was seismic monitoring under a nuclear test ban - an application of scienti fic knowledge and modern technology for a political purpose. The international political objective of a comprehensive nuclear test ban provided in turn the motivation for our technical and scientific discussions. In order to obtain a historical perspec tive on the progress of the work towards a comprehensive test-ban treaty (CTB), it is necessary to go back to 1958, when a confer ence of scientific experts in Geneva made the first steps toward an international seismic monitoring system. However, agreement on actual capabilities of a monitoring system for verifying compliance with such a treaty was not achieved, and thus the conference did not lead to immediate political results. After the Partial Test Ban Treaty of 1963, which banned nuclear explosions in the atmosphere, outer space and under the seas, renewed interest in the seismological verification of a CTB took place. A number of countries initiated large-scale research efforts toward detecting and identifying underground nuclear explosions, and it was in this context that the large aperture seismic arrays NORSAR and LASA were established. This type of development resulted in excellent seismic data in digital form and was thus of great irnprotance to the seismological com munity.