Commit d8811310 authored by Carsten Gräser's avatar Carsten Gräser

[doc][manual] Describe multi-index concept

parent df7d8c6c
Pipeline #1995 passed with stage
in 69 minutes and 58 seconds
......@@ -49,6 +49,9 @@
%%%%%%%%%%%%%% Define a 'shellenv' environment for shell output
......@@ -60,6 +63,7 @@
......@@ -285,6 +289,113 @@ reuse existing algorithms that expect to operate on those subtrees in more compl
\subsection{Indexing basis functions in a finite element basis tree}
To work with the basis of a finite element space, the basis vectors need to be indexed. Indexing the basis functions
is what allows to address the corresponding vector and matrix coefficients in suitable vector and matrix data structures.
In simple cases, indexing means simply enumerating the basis functions with natural numbers, but for many applications
hierarchically structured matrix and vector data structures are more natural or efficient leading to the need
for hierarchically structured multi-indices.
One way to construct such multi-indices would be to directly use the tree of finite element spaces.
Assuming that a flat indexing with natural numbers is given for each leaf FE space in the tree,
a unique hierarchical index for all basis functions in the tree can be obtained by simply
enumerating all children of any node and prepending the index of the child tree to the index
within the child tree.
We will illustrate this for the Taylor--Hood tree $(P_2 \times P_2 \times P_2) \times P_1$.
We start by assuming that the basis functions for $P_1$ and $P_2$
both have a flat indexing by natural numbers in $\N_0 = \N \cup \{0\}$.
Then we enumerate the three children $P_2$, $P_2$, $P_2$
in the velocity subtree $P_2 \times P_2 \times P_2$
with the numbers $0$, $1$ and $2$.
Finally we enumerate the two children $(P_2 \times P_2 \times P_2)$ (for the velocity)
and $P_1$ (for the pressure) in the overall tree $(P_2 \times P_2 \times P_2) \times P_1$
with the numbers $0$ and $1$.
This construction leads to multi-indices indices of the form $[0,i,j]$ and $[1,k]$.
The component of the multi-index determines if the basis function
belongs to the velocity or pressure degrees of freedom by being either $0$ or $1$.
For the velocity multi-indices $[0,i,j]$ the $i$ determines the component
of the velocity vector field and the $j$ determines the number of the $P_2$ basis
function for scalar $P_2$ function that determines this component.
For the pressure multi-indices $[0,k]$ the $k$ determines the number of the $P_1$ basis
function for the scalar $P_1$ function that determines the pressure.
If the basis functions $\lambda_I$ are indexed by multi-indices of this form
and for a coefficient vector $X$ which has a compatible hierarchical structure
the velocity vector field $u$ and the pressure $p$ are then given by
u &= (\sum_{j} X_{(0,i,j)}\lambda_{(0,i,j)})_{i=1,\dots,3}, &
p &= \sum_{k} X_{(1,k)}\lambda_{(1,j)}.
This example already demonstrates two important requirements:
On the one hand natural multi-indices do not necessarily have
a fixed length.
On the other hand it shows that one may be interested to construct multi-indices
that do not mimic the structure of the FE tree: To increase data locality in assembled
matrices it may be preferable to use group all velocity corresponding to a single
$P_2$ basis function together, i.\,e. to use the index $[0,j,i]$
for the $j$-the $P_2$ basis function for the $i$-th component.
\todo{CG:Add picture to illustrate both indexing schemes for Taylor--Hood}
To reflect this \dunemodule{dune-functions} essentially allows arbitrary
multi-indices as long as they are consistent in the sense that they
can be viewed as the paths to the leafs in an ordered tree.
%Consistency here means that the multi-indices should form an index tree.
That is, the children of each node are enumerated using consecutive zero-based
indices and paths to the leafs (i.e. the multi-indices) are build by concatenating
those indices starting from the root and ending in a leaf.
As motivated above, it is important to note, that this tree does not need to
coincide with the FE tree. A precise definition of such an index tree is given by
the following:
A tuple $I \in \N_0^k$ for some $k \in \N_0$ is called a multi-index of length $k$
and we write $|I|=k$.
The set of all multi-indices is denoted by
$\mathcal{N} = \bigcup_{k \in \N_0^k} \N_0^k$.
If $I \in \mathcal{N}$ takes the form $I = (I',I'')$ for $I',I'' \in \mathcal{N}$
and $|I| = |I'|+|I''|$, then we call $I'$ a prefix of $I$ and write $I = (I',\dots)$.
If additionally $|I''|>0$ we call $I'$ a strict prefix of $I$ and write $I = (I',*,\dots)$.
A set $\mathcal{I} \subset \mathcal{N}$ is called an \emph{index tree}
if for any $(I,i,\dots) \in \mathcal{I}$ there are also $(I,0,\dots),(I,1,\dots),\dots,(I,i,\dots) \in \mathcal{I}$
but $I \notin \mathcal{I}$.
If $\Lambda$ the set of basis functions of a finite element tree,
then \dunemodule{dune-functions} allows any indexing scheme that
is given by an index map $I: \Lambda \to \mathcal{N}$
whose range $I(\Lambda)$ is an index tree.
When operating with the indices of basis functions it is important
to know the maximal appearing index in order to allocate matrices
and vectors.
For a flat consecutive zero-based index this is just $d-1$ where
$d$ is the dimension of the spanned space. For multi-indices from
an index tree $\mathcal{I}$ we need a slightly more complicated
construction. Let $(I,\dots) \in \mathcal{I}$, i.\,e., $I$ is a
prefix of multi-indices in $\mathcal{I}$ then the size relative
to $I$ is given by
|\mathcal{I}|_I = \op{max}\{k \st \exists (I,k,\dots) \in \mathcal{I} \}-1.
In terms of the ordered tree associated with $\mathcal{I}$ this corresponds
to the number of direct children of the node indexed by $I$.
While \dunemodule{dune-functions} allows to use arbitrary index trees
to enumerate the basis functions, there are some important generic constructions
to derive such index trees from the finite element tree that will be described below.
\subsection{Ordering basis functions in a finite element basis tree}
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment