1: \begin{abstract}

2: It is natural to expect the following loosely stated approximation principle to hold: a numerical

3: approximation solution should be in some sense as smooth as its target exact

4: solution in order to have optimal convergence. For piecewise polynomials,

5:  that means we have to at least maintain

6:  numerical smoothness in the

7: interiors as well as across the interfaces of cells or elements.

8: % Recently, Sun

9: %\cite{SUN} validated this principle in one dimensional case by

10: %rigorously  defining numerical smoothness as the boundedness of the scaled

11: %jumps of derivatives across mesh points and proving that numerical

12: %smoothness is a necessary condition under uniform meshes. However, in order to

13: %broaden the application of this useful principle, one must validate

14: %the case for nonuniform meshes and for two and three dimensions.

15: In this paper we give clear definitions of numerical smoothness that address the across-interface smoothness in terms of {\it scaled} jumps in derivatives \cite{SUN} and the interior numerical smoothness in terms of differences in derivative values. Furthermore, we prove rigorously that the principle can be simply stated as {\it numerical smoothness is necessary for optimal order convergence.}

16: It is valid on quasi-uniform meshes by triangles and quadrilaterals in two dimensions and by tetrahedrons and hexahedrons

17: in three dimensions. With

18: this validation we can justify, among other things, incorporation of this

19: principle in creating adaptive numerical approximation for the solution of

20: PDEs or ODEs, especially in designing proper smoothness indicators or detecting potential non-convergence and instability.

21: \end{abstract}

22: