I read that a fully polynomial time approximation scheme (FPTAS) has time complexity polynomial in the input size and also polynomial in $1/\epsilon$. However, this leaves some ambiguity.

For example, $T(n,1/\epsilon)=\min(n^{1/\epsilon},(1/\epsilon)^n)$ is both polynomial in $n$ (when holding $1/\epsilon$ constant) and polynomial in $1/\epsilon$ (when holding $n$ constant). But $T$ is not polynomial in $n/\epsilon$ or $(n+1/\epsilon)$ because $\sqrt x^\sqrt x$ and $(x/2)^{x/2}$ aren't bounded by polynomials.

Does this mean an algorithm with running time $T$ isn't an FTPAS, and perhaps a better definition would be that it has to be polynomial in $n+1/\epsilon$?


The standard definition is incredibly ambiguous, since it is not clear what "polynomial in $n$ and in $1/\epsilon$" means. Judging from examples, it means $O(n^C(1/\epsilon)^D)$ for some constants $C,D$. Without loss of generality, $C=D$, and then we get the simpler definition $O((n/\epsilon)^C)$.

| cite | improve this answer | |
  • $\begingroup$ Thanks for clarifying. Although might $O((n+1/\epsilon)^C)$ be more precise because when $\epsilon=n$ is very large, $n/\epsilon=1$ but it should still take at least $O(n)$ to find a feasible solution? $\endgroup$ – Akababa Aug 5 '18 at 23:22
  • $\begingroup$ The big O holds in the limit $n\to\infty$ and $\epsilon\to0$. We can assume that $\epsilon < 1 < n$, say. $\endgroup$ – Yuval Filmus Aug 5 '18 at 23:28

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.