difference between the distribution in figure 1 and the distribution in 
figure 4. The dotted lines in figure 5 are the dashed line multiplied by 
the cost function in figure 2. The area under these dotted lines is calcu- 
lated. That area is the conditional value of sample information for the 
special case of 15 d. f. and the mean equal to 4.5 um. This procedure 
is repeted for all possible means. The resulting areas, or costs, are shown 
in figure 6. Finally these costs are weighted by the probability distribu- 
tion in figure 1, giving the EVSI for 15 d. f. (— 6550 kronor). This 
entire procedure is repeated for all possible degrees of freedom, resulting 
in figure 3. 
Then the Expected Net Gain of sampling will be the bottom curve of 
figure 3, or the difference between the EVSI and the cost curves in the 
same figure. This gives us an expected maximum for 100 degrees of free- 
dom, or 106 measurements, which is our optimum number of observa- 
tions. 
6. Do the measurements 
Here we should preferably take a random sample of our population 
of errors, but because of the cost of the grid and the cost of the measure- 
ment and the calculations, we often make the measurements in certain 
predetermined locations. These locations should then both give us a 
good weight coefficient matrix in the computations and be representative 
for the population. The locations should preferably be determined by an 
international working group, as the previously mentioned "Standard tests 
of photogrammetric instruments”, in cooperation with the manufacturers. 
7. Revise the assigned probabilities and determine the « that gives mi- 
nimum expected loss 
In the light of what has been observed there might exist reasons for 
revision of the distribution of the probability for different variances. Let 
us assume that figure 7 now is valid (instead of figure 1). 
Our problem is now to compare the risks for the two different kinds 
of decision errors a-error and B-error and to find the level of significance, 
a, that minimizes the expected cost of these errors, i. e. of the decision. 
If we try with « = 30 % we get the risks for different errors as the 
fully drawn line of figure 8. (The curves are computed according to 
152