it turns out that the problem may be solved with only real numbers in 
slightly modified formulae. 
Numerical stability may become poor when deleting observations. Another 
algorithm for observation removal is recommended in /2/. It requires 
slightly more arithmetic operations, but is believed to be more 
reliable. 
A COMPARATIVE TEST OF GT AND TFU 
Aspects of Testing 
Different adjustment methods have been evaluated with respect to the 
computing time, either by counting the number of algebraic operations 
necessary or by running test programs. The theoretical operation count 
may give a good indication of performance. However, there are other 
factors which will influence the final performance in a specific 
application. 
Obviously, the data structure should fit the algorithm, to make it as 
effective as possible. The sparsity pattern of the normals should be 
exploited if possible, both with respect to storage and execution time. 
However, it should be considered whether the sparsity pattern is 
substantially different in different applications. 
On the other hand, the procedural structure of the software should 
reflect the operational procedures of the OLT system. This includes, 
for instance, the number of observations per update and how often 
blunder detection is wanted. At which time object control is to be 
included, and in which form it is included, will also have considerable 
impact on the program structure. 
As a conclusion, the performance of a specific algorithm will depend on 
the software structure, and therefore, performance tests of algorithms 
should preferably be run in a software environment similar to the 
actual application software. 
The Actual Test 
To test the performance of the GT algorithm it was decided to compare 
it with the TFU algorithm, since the latter had already been found 
useful in an earlier test. 
A copy of the program written and used by Wyatt /13/ for testing TFU, 
was made available for the test, and used as reference. 
A new program was written with the same operational specifications, but 
based on the GT algorithm for sequential adjustment. Most of the data 
structures were identical, changes being introduced only where 
necessary for the correct operation of the GT algorithm. 
The test now consists of a simulated aerotriangulation of a block of 
- 362 +