WSJT-X/lib/ldpc/sparse-LU.html
Steven Franke 5ac886855d Add ldpc sandbox folder.
git-svn-id: svn+ssh://svn.code.sf.net/p/wsjt/wsjt/branches/wsjtx@6437 ab8295b8-cf94-4d9e-aec4-7959e3be5d79
2016-01-25 00:04:21 +00:00

91 lines
3.8 KiB
HTML
Executable File

<HTML><HEAD>
<TITLE> Sparse LU Decomposition Methods </TITLE>
</HEAD><BODY>
<H1> Sparse LU Decomposition Methods </H1>
<P>The sparse modulo-2 matrix LU decomposition routine <A
HREF="mod2sparse.html#decomp"><TT>mod2sparse_decomp</TT></A> (which
is used by the <A HREF="encoding.html#make-gen"><TT>make-gen</TT></A>
program when it is asked to create a sparse generator matrix) tries to
find an sub-matrix of a matrix (for <TT>make-gen</TT>, the parity
check matrix), and an ordering of rows and columns for this
sub-matrix, that leads to the lower-triangular matrix <B>L</B> and the
upper-triangular matrix <B>U</B> making up the LU decomposition being
as sparse as possible. Finding an optimal solution is too difficult,
so instead a heuristic strategy is used.
<P>The overall algorithm finds <B>L</B> and <B>U</B> a column at a
time, from left to right (as reordered, in the case of <B>U</B>). As
this is done, a copy, <B>B</B>, of the original matrix is modified.
To create column <I>i</I> of <B>L</B> and <B>U</B>, some element with
value 1 in <B>B</B> whose row and column indexes, after reordering,
are both greater than <I>i</I> is found. The row and column of this
element are considered to come next in the reordering, and the
contents of the column containing this element is copied to <B>L</B>
and <B>U</B> (upper elements going to <B>U</B>, lower to <B>L</B>).
The row containing this element is then added to some later rows so as
to clear the lower part of this column to zeros.
<P>At the first step of this process - selecting an element with value
1 from the later rows and columns - there will often be several
possibilities. Different choices can lead to the final result being
more or less sparse. The possible strategies for picking an element
are identified by the constants <TT>Mod2sparse_first</TT>,
<TT>Mod2sparse_mincol</TT>, and <TT>Mod2sparse_minprod</TT>. These
strategies operate as follows:
<P><TT>Mod2sparse_first</TT>
<BLOCKQUOTE>
Select the first element with value 1 that is encountered in a top
to bottom, left to right search.
</BLOCKQUOTE>
<P><TT>Mod2sparse_mincol</TT>
<BLOCKQUOTE>
Select the first element with value 1 that is contained in a column
of <B>B</B> that has the smallest number of 1s of any column.
</BLOCKQUOTE>
<P><TT>Mod2sparse_minprod</TT>
<BLOCKQUOTE>
Select an element with value 1 for which the product of the number of
1s in that row of <B>B</B> minus one times the number of 1s in that
column of <B>B</B> minus one is as small as possible.
</BLOCKQUOTE>
<P>The <B>abandon_number</B> and <B>abandon_when</B> parameters can
modify the basic strategy. If <B>abandon_number</B> is greater than
zero, then after <B>abandon_when</B> columns have been selected,
<B>abandon_number</B> of the remaining columns are abandoned as
candidates for possible future selection, the abandoned columns being
those with the greatest number of entries. Abandoning such columns
saves space and time, but may make the final result less sparse than
it would otherwise be, and can possibly result in the matrix appearing
to have lower rank than it actually has.
<P>The methods described here are fairly straightforward adaptations
of standard methods for sparse square matrices of reals, as described, for
example, in
<BLOCKQUOTE>
I. S. Duff, A. M. Erisman, J. K. Reid (1986) <I>Direct Methods for
Sparse Matrices</I>, Oxford: Clarendon Press.
</BLOCKQUOTE>
In the coding context, however, we are interested in matrices of
modulo-2 elements, and it is enough to find a sparse LU decomposition
of any square sub-matrix that can be obtained by selecting columns of
the rectangular parity check matrix. I talked about the application
of sparse matrix methods to encoding of LDPC codes at the 1999 IMA
workshop on Codes, Systems and Graphical Models (see the <A
HREF="refs.html">references</A>).
<HR>
<A HREF="index.html">Back to index for LDPC software</A>
</BODY></HTML>