16
$\begingroup$

Worst-case and average case analysis are well-known measures for the complexity of an algorithm. Recently smoothed analysis has emerged as another paradigm to explain why some algorithms which are exponential in the worst-case work so well in practice, for instance the simplex algorithm.

My question is - are there any other paradigms to measure the complexity of an algorithm? I'm particularly interested in ones that attempt to explain why some algorithms which have bad worst-case complexity work well in practice.

$\endgroup$
21
$\begingroup$

There are natural variants of worst-case analysis that are also useful. Perhaps the most famous one is parametrized complexity. Here, we consider a "two-dimensional" measure: the usual input length $n$ and some additional non-negative integer $k$, the parameter. Even though an algorithm may run horribly in the worst case (for all values of $n$ and $k$), it could be that all the cases in one's application that needs to be solved, this parameter $k$ happens to be low, so the algorithm runs well on those instances.

For example, suppose you want to solve Maximum Independent Set on some class of graphs, and develop an interesting algorithm that is surprisingly fast. Investigating further into the class of graphs themselves, you find that all the graphs you examine just happen to have treewidth at most $10$. Well, Bodlaender (cf. Neidermeier [1]) showed that when the treewidth is k, Max Independent Set is fixed parameter tractable: it can be solved in $O(2^k (|E|+|V|))$ time. This gives some explanation as to why your algorithm works well.

[1] R. Niedermeier, Invitation to fixed-parameter algorithms. Oxford Lecture Series in Mathematics and its Applications, Oxford University Press, Oxford, 2006.

$\endgroup$
15
$\begingroup$

There is amortized complexity -- why some operations can be costly in the worst-case, but if you consider many operations, the average cost per operation is good.

A classic example is a data structure that empties itself when it is full by copying all its elements to some storage. The copying operation can be expensive, but it doesn't happen often -- you have to insert enough elements to the data structure to provoke it.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.