Rise crest. Each derived E-W profile was 17 km In length and spaced 

 approximately 100 m apart. Amplitude spectra were generated for each 

 profile, and rather than averaging each amplitude estimate, the derived 

 regression parameters a and b were assembled for adjacent beams In 

 groups of 2, 4, 8, and 16. Figure 7-1 plots for both a and b, the 

 estimated parameters for each of the sixteen profiles and the various 

 results of ensembllng. The standard error of each set Is plotted as a 

 solid line, with the theoretical 1/VN relationship shown as a dashed 

 line. The derived standard error for the final average of sixteen 

 points Is necessarily zero. Similar techniques could be applied In the 

 province picking algorithm to Improve reliability. 



Propagation of Error to High Frequency Estimates 



Although the errors associated with the model can not be determined 

 with any accuracy, It Is still Instructive to examine how the errors 

 affect the ability to predict amplitude In frequencies beyond the range 

 of analysis. Due to the power law form of the model, 



A ■• a * s" where A > amplitude 



s *• spatial frequency 

 a, b > regression parameters 



the errors of estimate (error bars) are not linear. It Is therefore 

 somewhat easier to visualize the function In log A-log s space where the 

 error bars are linear. 



Ill 



