EFFICIENT CODING 749 



The encoding of the reduced signal into binary digits presents no 

 theoretical difficulties. A PCM type coder tube^ with the appropriate 

 Shannon-Fano groups built into it is all that is needed. The biggest 

 practical complication arises out of the fact that the code groups arc of 

 different length. Some messages, such as written text, can be fed into 

 the system as fast as it can handle them. The transmission time will 

 then ^'ary with the message complexity. Others, such as television are 

 generated and must be accepted and delivered at a constant rate. One 

 solution is then to take the binary digits in big and little batches as they 

 come from the coder and store the surplus in a sort of pulse "surge tank" 

 before they are sent over the channel at a regular rate. At the receiver, 

 a similar sort of storage register is necessary as the pulses arrive over 

 the channel at a regular rate and are used by the decoder at a varying 

 rate. Devices which will perform this variable delay function satisfacto- 

 rily for signals vvdth relatively slow samphng frequency are available, 

 and as the art progresses there is every reason to believe that high speed 

 sampled signals like television can be handled also. 



It vnll be noticed that the digramming or prediction operation, while 

 it involves memory, does not introduce appreciable transmission delay. 

 Each symbol of the reduced signal appears the moment the correspond- 

 ing message sample is applied. The total transmission delay required 

 for statistical coding thus depends upon how much variation is required 

 in the variable delay units. This in turn depends upon the degree of 

 stationarity in the "local information rate" of the message. For example, 

 in television, if each line could be described (by the n-grammer and 

 subsequent coder) in the same total number of binary digits, then the 

 total delay variation and total delay would be less than one line time. 

 Since this is not true, we either must have enough channel capacity to 

 send in one line time the niunber of digits corresponding to the "worst" 

 line, or enough variable delay to average the existing rate over many 

 lines. 



Probably the most practical solution is to provide sufficient channel 

 capacity and variable delay to take care of all but a small fraction of 

 the possible message sequences. Then when an unusual stretch of message 

 continues long enough for the variable delay to be nearly all used up, the 

 system should fail in some relatively harmless way. In television, the 

 sampling rate could be momentarily reduced, for example. This would 

 degrade the resolution in rare situations, but a small amount of this 

 could be tolerated in return for transmission savings. 



If long blocks of the message are efficiently encoded as a group, then 

 an error in transmission may cause the whole block to be reproduced 



