Appendix A. Some basic computations between reference and real elements

Volume integral

One has

\int_T f(x)\ dx = \int_{\widehat{T}} \widehat{f}(\widehat{x})
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_0};
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_1};
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_{P-1}}
\right)|\ d\widehat{x}.

Denoting J_{\tau}(\widehat{x}) the jacobian

\fbox{$ J_{\tau}(\widehat{x}) :=
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_0};
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_1};
\frac{\partial\tau(\widehat{x})}{\partial \widehat{x}_{P-1}}
\right)| =
(\mbox{det}(K(\widehat{x})^T K(\widehat{x})))^{1/2}$,}

one finally has

\fbox{$\int_T f(x)\ dx = \int_{\widehat{T}} \widehat{f}(\widehat{x}) J_{\tau}(\widehat{x})\ d\widehat{x}$.}

When P = N, the expression of the jacobian reduces to J_{\tau}(\widehat{x})
= |\mbox{det}(K(\widehat{x}))|.

Surface integral

With \Gamma a part of the boundary of T a real element and \widehat{\Gamma} the corresponding boundary on the reference element \widehat{T}, one has

\fbox{$\int_{\Gamma} f(x)\ d\sigma =
\int_{\widehat{\Gamma}}\widehat{f}(\widehat{x}) \|B(\widehat{x})\widehat{n}\| J_{\tau}(\widehat{x})\ d\widehat{\sigma}$,}

where \widehat{n} is the unit normal to \widehat{T} on \widehat{\Gamma}. In a same way

\fbox{$\int_{\Gamma} F(x)\cdot n\ d\sigma =
\int_{\widehat{\Gamma}} \widehat{F}(\widehat{x})\cdot(B(\widehat{x})\cdot\widehat{n}) J_{\tau}(\widehat{x})\ d\widehat{\sigma}$,}

For n the unit normal to T on \Gamma.

Derivative computation

One has

\nabla f(x) = B(\widehat{x})\widehat{\nabla} \widehat{f}(\widehat{x}).

Second derivative computation


\nabla^2 f =
\left[\frac{\partial^2 f}{\partial x_i \partial x_j}\right]_{ij},

the N \times N matrix and

\widehat{X}(\widehat{x}) =
\sum_{k = 0}^{N-1}\widehat{\nabla}^2\tau_k(\widehat{x})\frac{\partial f}{\partial x_k}(x) =
\sum_{k = 0}^{N-1}\sum_{i = 0}^{P-1}
\widehat{\nabla}^2\tau_k(\widehat{x})B_{ki}\frac{\partial \widehat{f}}{\partial \widehat{x}_i}(\widehat{x}),

the P \times P matrix, then

\widehat{\nabla}^2 \widehat{f}(\widehat{x}) = \widehat{X}(\widehat{x}) + K(\widehat{x})^T \nabla^2 f(x) K(\widehat{x}),

and thus

\nabla^2 f(x) = B(\widehat{x})(\widehat{\nabla}^2 \widehat{f}(\widehat{x}) - \widehat{X}(\widehat{x})) B(\widehat{x})^T.

In order to have uniform methods for the computation of elementary matrices, the Hessian is computed as a column vector H f whose components are \frac{\partial^2 f}{\partial x^2_0}, \frac{\partial^2 f}{\partial
x_1\partial x_0},\ldots, \frac{\partial^2 f}{\partial x^2_{N-1}}. Then, with B_2 the P^2 \times P matrix defined as

\left[B_2(\widehat{x})\right]_{ij} =
\sum_{k = 0}^{N-1}
\frac{\partial^2 \tau_k(\widehat{x})}{\partial \widehat{x}_{i / P} \partial \widehat{x}_{i\mbox{ mod }P}}

and B_3 the N^2 \times P^2 matrix defined as

\left[B_3(\widehat{x})\right]_{ij} =
B_{i / N, j / P}(\widehat{x}) B_{i\mbox{ mod }N, j\mbox{ mod }P}(\widehat{x}),

one has

\fbox{$H f(x) = B_3(\widehat{x})
\left(\widehat{H}\ \widehat{f}(\widehat{x}) - B_2(\widehat{x})\widehat{\nabla} \widehat{f}(\widehat{x})\right)$.}

Example of elementary matrix

Assume one needs to compute the elementary “matrix”:

t(i_0, i_1, \ldots, i_7) =
\partial^2_{i_7/ P, i_7\mbox{ mod } P}\varphi_{i_6}^{i_5}\ dx,

The computations to be made on the reference elements are

\widehat{t}_0(i_0, i_1, \ldots,i_7) =
\partial^2_{i_7 / P, i_7\mbox{ mod } P}(\widehat{\varphi})_{i_6}^{i_5} J(\widehat{x})\ d\widehat{x},


\widehat{t}_1(i_0, i_1, \ldots, i_7) =
\partial_{i_7}(\widehat{\varphi})_{i_6}^{i_5} J(\widehat{x})\ d\widehat{x},

Those two tensor can be computed once on the whole reference element if the geometric transformation is linear (because J(\widehat{x}) is constant). If the geometric transformation is non-linear, what has to be stored is the value on each integration point. To compute the integral on the real element a certain number of reductions have to be made:

  • Concerning the first term (\varphi_{i_1}^{i_0}) nothing.
  • Concerning the second term (\partial_{i_4}\varphi_{i_3}^{i_2}) a reduction with respect to i_4 with the matrix B.
  • Concerning the third term (\partial^2_{i_7 / P, i_7\mbox{ mod }P}
\varphi_{i_6}^{i_5}) a reduction of \widehat{t}_0 with respect to i_7 with the matrix B_3 and a reduction of \widehat{t}_1 with respect also to i_7 with the matrix B_3 B_2

The reductions are to be made on each integration point if the geometric transformation is non-linear. Once those reductions are done, an addition of all the tensor resulting of those reductions is made (with a factor equal to the load of each integration point if the geometric transformation is non-linear).

If the finite element is non-\tau-equivalent, a supplementary reduction of the resulting tensor with the matrix M has to be made.