Entropy and Complexity

Shah, Devavrat; Sharma, Mayank

HPL-BRIMS-2000-32

Keyword(s): information theory; algorithms

Abstract: Computational complexity of algorithms for solving problems has been at the heart of theoretical computer science. Traditionally, the computational cost of an algorithm is estimated by "counting" operations combinatorially depending on the algorithm. We present a very different method for estimating cost of solving problem, incorporating ideas from information theory. Algorithms can be viewed as "search" procedures on the input (output) space. This naturally makes computational cost of algorithm as function of "entropy" of input (output) distribution. The relation of computational complexity and "entropy" of distribution depends on the particular "operations" used by algorithm to solve problem. This particular mapping between computational scale and "entropy" scale is independent of problem. We demonstrate the use of this method in classical "searching" and "sorting" problems. We work out many different searching and sorting algorithms' complexity using this method. In process, we also come up with new algorithms to solve some cases of searching and sorting that are aware of input distribution.

16 Pages

Back to Index