digital, computer. The actual computation took only a few mi'^mtesj however 

 the clericaJ, ■workijavolTcd kn correcting errors in the da.ta and in. the program 

 consumed almost two hourso Were the problem to be done for a nrew sef of 

 data, only a few minutes of Univac time would be needed.,, as the program al~ 

 ready exists, and the data can be m.ade ready by a card to l^ipe converter. 

 S pectrum Computatio.n. by Me.ans of the Univac 

 The computation of the covariance surface is an extremely lon.g co.mpu" 

 tation^ involving many millions of multiplications. Thus it could, be most 

 speedily h.andled by the .IBM 704, the NORC , or the LARC, 



Programming the covar.iance surfer e on the Univac was especia.iry d,iffi- 

 cult because of its limi*:ed m.emorv„ Since many nunnbers must be s-^aihable 

 almost simultaneouslyj it was deemed inefficient lo store the dat^ on tape., as 

 ta.pe time would add cons id.er ably to tl^e prograna's running time. The way out. 

 of this d.ilemma was to break up the 90 by 60 array info three arrays: 44 .x 60, 

 2 X 60, and 44 x 60. There is considerable overlap., Since we wa^it a lag of 

 20, the middle group must contain 20 rows above and below^ thus h contains 

 42 rows in a.Ii, These data were packed 4 on a. lineo In this wavj ii: was 

 possible to pack an entire section in the memorv, and s+iil have er.ough in- 

 structions for the program. Since no room was left for sign, a, constant 

 was added to all the data, to :ma.ke them positive. The reader may perceive 

 how these factors added materially to the length of the romputatJ.on run. 

 Every two .numbers had to be isolated by a.n ingenious system of shiftsj then. 



the same constaJ3.t .fe*d to be subtr .acted, out. Only then could the n.ufT\bers be 



multiplied and the product a,C€umula,ted in a counter. Then the calculation. 



88 



