It is well known that in situ stress measurements at a given site will generally display significant variability. However, it appears that a robust analysis of such measurements – i.e., one that is faithful to the tensorial nature of the data – is not available. Despite the fact that stress is a tensor, stress magnitude and orientation have long been analyzed separately using classical statistics. In earlier work we demonstrated the applicability of tensor statistics to the analysis of stress variability. Here, we investigate further the difference between tensor and classical statistics with respect to the stress tensor correlation matrix. Our analysis commences with the selection of a series of valid correlation matrices. Then, assuming that standardized tensor components follow the multivariate standard normal distribution, these correlation matrices are used to generate marginal probability distributions of the eigen-parameters (i.e., principal stress magnitude and rotation angle). From these, random stress tensors are generated, their eigen-parameters determined, and the probability density distributions of these obtained using both classical and tensor statistics. Our results indicate that, for all correlation matrices, both methods give practically identical distributions for principal stress magnitude. However, the discrepancy between the two methods with regard to probability density distribution of rotation angle is highly dependent on stress component correlation. This discrepancy is usefully quantified in terms of the integral, with respect to orientation, of the difference between the two probability density distributions. This parameter, which we term Area Difference, is closely linked to the magnitude of the correlation matrix determinant, which itself is related to the normal stress correlation coefficient (i.e., ρ σxx σy). Assuming that Area Difference should not exceed 0.1, we identify the critical determinant for any particular value of normal stress component correlation coefficient. Thus, using published in situ stress data and the obtained critical determinants, we show that for a specific normal stress component coefficient, when the correlation matrix determinant is less than its critical determinant, the more precise tensor statistics should be used to calculate the distribution of rotation angle. These results demonstrate that the customary analysis of stress measurement data may lead to erroneous results, and that tensor statistics appears to be preferred approach.

This content is only available via PDF.
You can access this article if you purchase or spend a download.