worth noting that even at very high Pu activity levels (> 10° uCi/m2), underestimation using the regression line does not seem to be a problem. The residual variance is 0.0834, and therefore under a Gaussian distribution of the error, the Pu concentration may be predicted within a factor of about 2 at the 68 percent confidence level, and a factor of about 4 at the 96 percent confidence level. 2. For FIDLER < 5,000, the correlation is only p = 0.22 and the regres- sion equation is very different from that for FIDLER > 5,000: log(Pu) = 0.6275 log(FIDLER) + 0.2372 , (49 pairs) with a residual variance of 0.2126, 2.6 times larger than for FIDLER > 5,000. THE MODEL Let P(x) = log Pu(x), (Pu(x) in pCi/m?) and F(x) = log FIDLER(x), (FIDLER(x) in 10° cpm). We model P(x) and F(x) as realizations of random fields on IR? dimensional space) of the form (two- P(x) = POX) + €(x) F(x) = F(x) + n(x), where P_ (x) and F (x) stand for the "true" underlying fields, and ¢(x) and n(x) for measurement errors, of zero means, constant variances, and uncorrelated with P_(x) and F (x). Hence, errors are assumed multiplicative in the arithmetic scale. Constant variance and errors uncorrelated with true values do not strictly hold even in log scale, but the situation is certainly better than in the original scale. Furthermore, we can think of different error variance levels for the ranges FIDLER < 5,000 and FIDLER > 5,000. In the light of the preceding correlation study, it is reasonable to let E[P(x) |F(x) ] =a F(x) +B . 374 (3)