error associated with the model. Only by comparing estimated values 

 with values measured In high frequencies over many data sets and geo- 

 logic environments, can a reasonable statistical base be assembled to 

 assess the prediction capabilities of the model quantitatively. 



The effect of Instrument noise (Illustrated In Figure 4-5) when 

 encountered by the signal spectrum, has the effect of reducing the 

 spatial frequency range of amplitude estimates available to the regres- 

 sion analysis. Certain spectra examined In the course of this study 

 varied from complete whlte-nolse spectra to spectra showing only two or 

 three amplitude estimates above the noise. Such spectra are of no use 

 In model generation. The length of the spectrum available for regres- 

 sion analysis depends as well on the length of data available to the 

 FFT. Long profiles from large quasi-stationary provinces produce cor- 

 respondingly long amplitude spectra and therefore more reliable model 

 estimates. As mentioned In Appendix B, a minimum of one hundred points 

 Is allowed in a profile for analysis. The use of a higher minimum pro- 

 file length, while Improving recession estimates, would allow more non- 

 statlonarlty In the profile and reduce the ability to resolve small 

 provinces . 



No attempt has been made to quantify the level of non-statlonarlty 

 (as defined for this study) by examining the mean variabilis of the 

 spatial domain estimates used in province picking (see Appendix B). IXie 

 to the nature of the sea floor, certain provinces appear over thousands 

 of square miles, while others fall smaller than the hundred point mini- 

 mum required for analysis and must be combined. Such variability prob- 

 ably precludes estimating the degree of non-stationarity with any accu- 

 racy; we can only attempt to constrain the effect with the province 



108 



