If the error is trivial and you think the referee will overlook it (missing factor of 2, etc), you could wait.  If the error is obvious and you think the referee will see it at once, (you forgot to say the manifold is assumed to be compact, and the theorem obviously fails for $\mathbb{R}^n$), you could wait.  Otherwise, as a referee, I'd rather know about it.  I tend to read papers very carefully and sometimes spend a lot of time deciding whether some obscurity is due to an error or my own ignorance (I'm a younger mathematician and will probably outgrow this eventually :)  A constant stream of corrections would be annoying, but one significant one can be appreciated.

If the error will require major revisions or seriously weaken the result, I'd say notify for sure.  If you aren't yet sure of the fix and think you might need to withdraw the paper, notify for sure, so the referee needn't waste time reading it in the meantime.