776 THE BELL SYSTEM TECHNICAL JOURNAL, JULY 1952 



will be much less than the average amplitude of the original signal. 

 Thus, by using the decorrelator alone, we can send a message over a 

 channel with the same bandwidth as before but Avith less average power. 

 At first, this might sound like a worthwhile saving; however this lower 

 average power is accompanied by an even higher peak amplitude which 

 makes any direct saving less attractive. Furthermore, the low frequency 

 attentuation of the decorrelator makes the signal \'iilnerable to low 

 frequency disturbances, since the correlator must restore (emphasize) 

 these low frequency components. 



A proper but not entirely adequate method of evaluating the effective- 

 ness of a predictor is by measuring the ratio of signal power to error 

 power. This is called "Power Reduction" and is generally expressed in 

 db. Power reduction simply provides a scale by which' we can weigh a 

 linear predictor's capabilities. The "not entirely adequate" refers to 

 the fact that minimum error power may not provide simultaneously the 

 lowest amount of redundancy for that given type of prediction. 



As an example. Fig. 13 shows the power reduction for the relative 

 weighting of the previous horizontal signal sample as compared to the 

 present value of the signal, for three pictui-es-later to be described as 

 Scene A, B and C. The top-most curve is for Scene B, which is a 

 simple, soft picture that contains very little detail. For this picture, 

 the minimum error power coincides (^\ithin measurable limits) with the 

 minimum redundancy. For Scene A and particularly Scene C, mini- 

 mum error power is considerably different than that for minimum re- 

 dundancy. This difference between minimum error power and minimum 

 redundancy also apphes to decorrelators using other types of predictors 

 as well. Minimum redundancy may also be a misleading criterion of a 

 predictor's performance, since the value of the prediction must depend 

 on the particular type of encoder used, and some types of encoding will 

 require certain types of redundant information to be retained. 



The follo^\'ing pictures are representations of the error signal as photo- 

 graphed from a 10-inch laboratory monitor. The signals were band 

 limited to 4.3 mc. Fig. 14 represents the "original" for three scenes called 

 A, B and C. These pictures represent, to a first approximation, the 

 gamut of pictures normally expected to be transmitted. They are by 

 no means the best or the worst pictures than can be imagined; how- 

 ever any system should be able to reproduce these pictures Avithout 

 appreciable distortion. For example. Scene C should be capable of 

 being sent continuously without the elastic delay running out, etc. 



Fig. 15 shows how the error signal appears for "pre\dous value" pre- 

 diction. "Previous value" prediction is excellent for jflat white or dark 



