We selected the second option, the reason being that, on the one hand, our method needs only a little modification and the overall increment on the computational cost of the MG algo- rithm is very small, and, on the other hand, constructing a successful adaptive interpolation operator is not an easy task and many different ways to do it need to be tested (see [1, 24] for comments on this respect). Further, these methods can be computationally costly, and their storage requirements are high