A thought process in the verification of a hypothesis is: "If hypothesis H infers a novel prediction, P and at a later time the prediction P is confirmed, then the probability that H is true will increase".
Typically, the conclusion that P is true is based on a single, though occasionally sophisticated, experiment.
As A. F. Chalmers in his book "What is this thing called Science?" points out what is the point of repeating the same tests over again?
My view on this matter has arisen from my much-anticipated continuing series on models.
I would consider the hypothesis H to be a model, and perhaps more interestingly the prediction P, generally, to be another model. Simply because the prediction P appears to be a more nearly simple construct, does not imply that it lacks the same fundamental properties/problems associated with any model. Furthermore, the specific hypothesis referred to Chalmers book is that of Einstein’™s General Relativity and the prediction that a massive object would bend light. Eddington’s eclipse experiment was taken as the validation of the prediction.
The actual test data showed:
- The usable photos from Principe, a tiny island nestled in the crook of Africa’s Gulf of Guinea, showed an average difference of 1.61 plus or minus 0.30 arc seconds.
- The astrograph in Brazil indicated a deflection of about 0.93 arc seconds (depending on how one weighted the individual spoiled photos).
- The little 10-centimeter telescope gave a result of 1.98 plus or minus 0.12 arc seconds.
Only the first of the test data actually supports Einstein’s prediction of a value of 1.75, and that result has a very high margin of error. The second example is actually closer to the result expected if light was considered to have mass and the calculations were done on the basis of straight Newtonian physics. The second and third test would actually falsify General Relativity.
An additional later test has supported Einstein’s model, but it requires a different technology.
(Campbell in Australia does this in 1922 with a test result of1.72 plus or minus 0.11)
Note: The tolerances recorded here are (I believe) instrumental tolerances as are generally ascribed by engineering’s specifications and NOT rigorous statistical specifications.
In any case I believe that the first three paragraphs of this post are commonly held beliefs, which are complete nonsense.
A model is simply a model it cannot be true or false. (It can, however, given certain assumptions, make statistically reliable statements over a limited domain). No single test can possibly prove anything about a model. (This fails on a number of accounts not the least of which is the logical problem that if a implies b then b ----well— does not imply anything)