From: Martin Kronbichler Date: Sun, 17 Dec 2017 11:32:19 +0000 (+0100) Subject: Further improve manifold documentation text. X-Git-Tag: v9.0.0-rc1~637^2 X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=refs%2Fpull%2F5634%2Fhead;p=dealii.git Further improve manifold documentation text. --- diff --git a/doc/doxygen/headers/manifold.h b/doc/doxygen/headers/manifold.h index e4bd55b894..ca48315098 100644 --- a/doc/doxygen/headers/manifold.h +++ b/doc/doxygen/headers/manifold.h @@ -301,10 +301,11 @@ * circular part with a straight-sided description in the center of the domain * where the spherical manifold is not valid. * - * In general, the process of blending in deal.II is achieved by the so-called - * transfinite interpolation. Its formula 2D is, for example, described on - * - * Wikipedia. Given a point $(u,v)$ on a chart, the image of this point + * In general, the process of blending different manifold descriptions in + * deal.II is achieved by the so-called transfinite interpolation. Its formula + * in 2D is, for example, described on + * Wikipedia. Given a point $(u,v)$ on a chart, the image of this point * in real space is given by * @f{align*}{ * \mathbf S(u,v) &= (1-v)\mathbf c_0(u)+v \mathbf c_1(u) + (1-u)\mathbf c_2(v) + u \mathbf c_3(v) \\ @@ -315,36 +316,40 @@ * four curves describing the lines of the cell. * * If we want to find the center of the cell according to the manifold (that - * is also used when the grid is refined), we want to evaluate this formula in - * the point $(u,v) = (0.5, 0.5)$. In that case, $\mathbf c_2(0.5)$ is the - * position of the midpoint of the lower face (indexed by 2 in deal.II's - * ordering) that is derived from its own manifold, $\mathbf c_1(0.5)$ is the - * position of the midpoint of the upper face (indexed by 3 in deal.II), - * $\mathbf c_2(0.5)$ is the midpoint of the face on the left (indexed by 0), - * and $\mathbf c_3(0.5)$ is the midpoint of the right face. In this formula, - * the weights equate to $\frac{\displaystyle 1}{\displaystyle 2}$ for the - * four midpoints in the faces and to $-\frac{\displaystyle 1}{\displaystyle - * 4}$ for the four vertices. These weights look weird at first sight because - * the vertices enter with negative weight but the mechanism does what we - * want: In case of a cell with curved description on two opposite faces but - * straight lines on the other two faces, the negative weights of -1/4 in the - * vertices balance with the center of the two straight lines in radial - * direction that get weight 1/2. Thus, the average is taken over the two - * center points in curved direction, exactly placing the new point in the - * middle. - * - * In three spatial dimensions, the weights are +1/2 for the face midpoints, - * -1/4 for the line mid points, and +1/8 for the vertices, again balancing - * the different entities. In case all the surrounding of a cell is straight, - * the formula reduces to the obvious weight 1/8 on each of the eight - * vertices. + * is also used when the grid is refined), the chart is the unit cell + * $(0,1)^2$ and we want to evaluate this formula in the point $(u,v) = (0.5, + * 0.5)$. In that case, $\mathbf c_0(0.5)$ is the position of the midpoint of + * the lower face (indexed by 2 in deal.II's ordering) that is derived from + * its own manifold, $\mathbf c_1(0.5)$ is the position of the midpoint of the + * upper face (indexed by 3 in deal.II), $\mathbf c_2(0.5)$ is the midpoint of + * the face on the left (indexed by 0), and $\mathbf c_3(0.5)$ is the midpoint + * of the right face. In this formula, the weights equate to + * $\frac{\displaystyle 1}{\displaystyle 2}$ for the four midpoints in the + * faces and to $-\frac{\displaystyle 1}{\displaystyle 4}$ for the four + * vertices. These weights look weird at first sight because the vertices + * enter with negative weight but the mechanism does what we want: In case of + * a cell with curved description on two opposite faces but straight lines on + * the other two faces, the negative weights of $-\frac{\displaystyle + * 1}{\displaystyle 4}$ in the vertices balance with the center of the two + * straight lines in radial direction that get weight $\frac{\displaystyle + * 1}{\displaystyle 2}$. Thus, the average is taken over the two center points + * in curved direction, exactly placing the new point in the middle. + * + * In three spatial dimensions, the weights are $+\frac{\displaystyle + * 1}{\displaystyle 2}$ for the face midpoints, $-\frac{\displaystyle + * 1}{\displaystyle 4}$ for the line mid points, and $\frac{\displaystyle + * 1}{\displaystyle 8}$ for the vertices, again balancing the different + * entities. In case all the surrounding of a cell is straight, the formula + * reduces to the obvious weight $\frac{\displaystyle 1}{\displaystyle 8}$ on + * each of the eight vertices. * * In the MappingQGeneric class, a generalization of this concept to the - * support points of the polynomial grid representation, the nodes of the - * Gauss-Lobatto quadrature, is implemented by evaluating the boundary curves - * in the respective points $(u_i,v_i)$ of the Gauss-Lobatto points. The - * weights have been verified to yield optimal convergence rates $\mathcal - * O(h^{k+1})$ also for very high polynomial degrees, say $k=10$. + * support points of a polynomial representation of curved cells, the nodes of + * the Gauss-Lobatto quadrature, is implemented by evaluating the boundary + * curves in the respective Gauss-Lobatto points $(u_i,v_i)$ and combining + * them with the above formula. The weights have been verified to yield + * optimal convergence rates $\mathcal O(h^{k+1})$ also for very high + * polynomial degrees, say $k=10$. * * In the literature, other boundary descriptions are also used. Before * version 9.0 deal.II used something called Laplace smoothing where the @@ -357,20 +362,21 @@ * If the transition from a curved boundary description to a straight * description in the interior is done wrong, it is typically impossible to * achieve high order convergence rates. For example, the Laplace smoothing - * inside a single layer leads to a singularity in the fourth derivative of the + * inside a single cell leads to a singularity in the fourth derivative of the * mapping from the reference to the real cell, limiting the convergence rate * to 3 in the cells at the boundary (and 3.5 if global L2 errors were * measured in 2D). Other more crude strategies, like completely ignoring the - * manifold for the additional points of a high-order mapping, could lead to - * even worse convergence rates. The current implementation in deal.II, on the - * other hand, has been extensively verified in this respect and should behave - * optimally. + * presence of two different manifolds and simply computing the additional + * points of a high-order mapping in a straight coordinate system, could lead + * to even worse convergence rates. The current implementation in deal.II, on + * the other hand, has been extensively verified in this respect and should + * behave optimally. * * A bad strategy for blending a curved boundary representation with flat * interior representations obviously also reflects mesh quality. For example, * the above case with only 3 circumferential cells leads to the following - * mesh with Laplace smoothing rather than the interpolation from the - * boundary: + * mesh with Laplace manifold smoothing rather than the interpolation from the + * boundary as is implemented in deal.II: * * @image html hypershell-boundary-only-3-old.png "" *