# What does big O mean as a term of an approximation ratio?

I'm trying to understand the approximation ratio for the Kenyon-Remila algorithm for the 2D cutting stock problem.

The ratio in question is $(1 + \varepsilon) \text{Opt}(L) + O(1/\varepsilon^2)$.

The first term is clear, but the second doesn't mean anything to me and I can't seem to figure it out.

-

## migrated from cstheory.stackexchange.comJan 3 '13 at 7:24

This question came from our site for theoretical computer scientists and researchers in related fields.

The expression "$A(L) \le (1 + \varepsilon) \text{Opt}(L) + O(1/\varepsilon^2)$" is, as usual, shorthand for the following:

There exist constants $c>0$ and $\varepsilon_0>0$ such that for all $\varepsilon$ with $0<\varepsilon<\varepsilon_0$, the inequality $A(L) \le (1 + \varepsilon) \text{Opt}(L) + c/\varepsilon^2$ holds.

-
I guess I was just looking into too much. Many thanks (and thanks to Jukka Suomela as well for that link). –  Jacob Fogner Jan 2 '13 at 0:37

This seems a looser variant of polynomial time approximation scheme (PTAS). If \epsilon is not small, you can achieve approximation with ratio very close to 1+\epsilon because O(\epsilon^{-2}) \le c \epsilon^{-2} is small. (c is a fixed positive real number independent of any other variable.) If \epsilon is small, the 2nd term gets larger.

However, OPT(L) is usually much larger than a constant. No matter how large O(\epsilon^{-2}) becomes, it is still a constant (since \epsilon is a given target real number for the approximation ratio). So Kenyon-Remila theorem means: "Constructed \le (1+\ep) OPT + O(1) for any given app ratio 1+\ep, where the O(1) term is a constant depending on \ep. It is actually O(\ep^{-2})."

-

If I remember correctly, this usage you feel confused predated (and actually inspired) the usage of $O$ for asymptotic functions. For example, you can have a short look at http://en.wikipedia.org/wiki/Prime_number_theorem.
In fact, the algorithm in question is a so-called "asymptotic polynomial time approximation scheme" (APTAS), which means that the approximation ratio is measured asymptotically in terms of the optimum; the approximation ratio of $(1+\epsilon)$ is approached as the optimal value gets larger since the additive term $O(\epsilon^{-2})$ is indepentent from both $n$ (the number of items in the instance $L$) and $Opt(L)$, the optimal value of the instance $L$.