|
|
|
|
|
10.36 Proposition Suppose . Then . |
|
|
|
|
|
|
|
|
We next show that allowing an extra error in the final inferred program cannot, in general, be compensated by the best Uap-type additional information. |
|
|
|
|
|
|
|
|
(a) For all , . |
|
|
|
|
|
|
|
|
(b) . |
|
|
|
|
|
|
|
|
Proof: For each , define |
|
|
|
|
|
|
|
|
Also, for each , define . Observe that for all , |
|
|
|
|
|
|
|
|
We leave it as exercise to show that the proposition follows from the above observation together with results from Section 6.1.1. |
|
|
|
|
|
|
|
|
The previous three results imply the following corollary, which reveals the complete relationship between UapdExa and ApdExa for various values of a and d. The proof of the corollary is left to the reader. |
|
|
|
|
|
|
|
|
10.38 Corollary Let a, and d1, . |
|
|
|
|
|
|
|
|
(a) . |
|
|
|
|
|
|
|
|
(b) . |
|
|
|
|
|
|
|
|
(c) For all , . |
|
|
|
|
|
|
|
|
In the paradigms considered above, partial explanations do not contradict the function being learned. There is no reason to believe, however, that the state-of-the-art explanation available to a scientist is necessarily free of errors of commission. A natural, and currently open, line of further investigation is to consider approximations that are correct on a set of certain density and either undefined or incorrect off that set. |
|
|
|
|