Paper 2007/435
Irreducibility to the One-More Evaluation Problems: More May Be Less
Daniel R. L. Brown
Abstract
For a random-self-reducible function, the evaluation problem is irreducible to the one-more evaluation problem, in the following sense. An irreduction algorithm exists that, given a reduction algorithm from the evaluation to the one-more evaluation problem, solves a separator problem: the evaluation problem itself. Another irreduction shows that if the computational Diffie-Hellman problem is reduced to the gap Diffie-Hellman problem, then the decision Diffie-Hellman problem is easy. Irreductions are primarily of theoretical interest, because they do not actually prove inequivalence between problems. What these irreductions suggest, though, is that one-more variants of the RSA and discrete logarithm problems may be easier than the standard variants, and that the gap Diffie-Hellman problem may be easier than the standard Diffie-Hellman problem.
Note: Minor changes.
Metadata
- Available format(s)
- PDF PS
- Category
- Foundations
- Publication info
- Published elsewhere. Unknown where it was published
- Keywords
- irreductionone-more evaluationgap DHP
- Contact author(s)
- dbrown @ certicom com
- History
- 2010-06-09: last of 4 revisions
- 2007-11-24: received
- See all versions
- Short URL
- https://ia.cr/2007/435
- License
-
CC BY
BibTeX
@misc{cryptoeprint:2007/435, author = {Daniel R. L. Brown}, title = {Irreducibility to the One-More Evaluation Problems: More May Be Less}, howpublished = {Cryptology {ePrint} Archive, Paper 2007/435}, year = {2007}, url = {https://eprint.iacr.org/2007/435} }