24W Page 134
Here is a list of references to methods of starting the convergence of a rectangle:
Techniques for starting convergence on Colossus R1 p 93.
Necessity of a good start. Suggestion of starting from eleven selected rows, trying all possible signs R2 p 4 (see also R2 pp 18, 22, R3 p 21).
Eye starts R3 p 108.
Random starts, with purging R4 p 3.
Methods of starting and suggestion that the choice amongst certain standard methods should be optional R4 p 23.
Statistics for various methods of starting R4 p 68.
Here are some references to methods of analysis of a rectangle, not connected with methods of starting.
Solving rectangle by linear equations. Crude convergence. Solving a rectangle by minimising a quadratic form R1 pp 40, 56.
Maximum likelihood solution of a rectangle R2 pp 16, 29, 32, 34, 35, 37, 39, 40.
There are other consequences of the knowledge that more than one convergence is possible, besides the importance of a good start. One is that the convergence must be done with care. The standard of acceptance of a character should be lowered gradually and arithmetical mistakes should be avoided. There are several examples of a wrong convergence being reached due to mistakes of various kinds. Another consequence is that a better convergence than the first one can often be obtained by a 'restart' in which the highest scoring characters of the first convergence are taken for the restart of another convergence (see R3 p 98). The validity of this method (apart from the successes attained) (see e.g. R2 p 101) depends on the empirical observation that the high-scoring characters tend to have the right sign even if the rectangle has not reached 'significance' (R3 pp 16, 17, 36). (For the meaning of the term significance see below - significance test IV.)
(d) Flags.
It has been found that crude convergence of a rectangle from a random start is liable to lead to a convergence which is not the best one.
Back to General Report on Tunny. Contents.