File: node95.html

package info (click to toggle)
lapack 3.0.20000531a-28
links: PTS
area: main
in suites: sarge
size: 61,920 kB
ctags: 46,200
sloc: fortran: 584,835; perl: 8,226; makefile: 2,331; awk: 71; sh: 45
file content (453 lines) | stat: -rw-r--r-- 14,146 bytes
parent folder | download | duplicates (4)
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2 Final//EN">
<!--Converted with LaTeX2HTML 98.2 beta6 (August 14th, 1998)
original version by:  Nikos Drakos, CBLU, University of Leeds
* revised and updated by:  Marcus Hennecke, Ross Moore, Herb Swan
* with significant contributions from:
  Jens Lippmann, Marek Rouchal, Martin Wilck and others -->
<HTML>
<HEAD>
<TITLE>Computing s and </TITLE>
<META NAME="description" CONTENT="Computing s and ">
<META NAME="keywords" CONTENT="lug_l2h">
<META NAME="resource-type" CONTENT="document">
<META NAME="distribution" CONTENT="global">
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">
<LINK REL="STYLESHEET" HREF="lug_l2h.css">
<LINK REL="previous" HREF="node94.html">
<LINK REL="up" HREF="node92.html">
<LINK REL="next" HREF="node96.html">
</HEAD>
<BODY >
<!--Navigation Panel-->
<A NAME="tex2html5523"
 HREF="node96.html">
<IMG WIDTH="37" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="next"
 SRC="next_motif.png"></A> 
<A NAME="tex2html5517"
 HREF="node92.html">
<IMG WIDTH="26" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="up"
 SRC="up_motif.png"></A> 
<A NAME="tex2html5513"
 HREF="node94.html">
<IMG WIDTH="63" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="previous"
 SRC="previous_motif.png"></A> 
<A NAME="tex2html5519"
 HREF="node1.html">
<IMG WIDTH="65" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="contents"
 SRC="contents_motif.png"></A> 
<A NAME="tex2html5521"
 HREF="node152.html">
<IMG WIDTH="43" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="index"
 SRC="index_motif.png"></A> 
<BR>
<B> Next:</B> <A NAME="tex2html5524"
 HREF="node96.html">Error Bounds for the</A>
<B> Up:</B> <A NAME="tex2html5518"
 HREF="node92.html">Further Details: Error Bounds</A>
<B> Previous:</B> <A NAME="tex2html5514"
 HREF="node94.html">Balancing and Conditioning</A>
 &nbsp <B>  <A NAME="tex2html5520"
 HREF="node1.html">Contents</A></B> 
 &nbsp <B>  <A NAME="tex2html5522"
 HREF="node152.html">Index</A></B> 
<BR>
<BR>
<!--End of Navigation Panel-->

<H3><A NAME="SECTION03481300000000000000"></A><A NAME="secspec"></A>
<BR>
Computing <B><I>s</I></B> and <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$">
</H3>

<P>
To explain <B><I>s</I></B> and <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$"><A NAME="11585"></A>, we need to
introduce<A NAME="11586"></A>
the <B>spectral projector</B> <B><I>P</I></B> [<A
 HREF="node151.html#stewart73">94</A>,<A
 HREF="node151.html#kato">76</A>], and the
<B>separation of two matrices</B><A NAME="11590"></A>
<B><I>A</I></B> and <B><I>B</I></B>, 
<!-- MATH
 ${\rm sep}(A,B)$
 -->
<IMG
 WIDTH="77" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img616.png"
 ALT="${\rm sep}(A,B)$">
[<A
 HREF="node151.html#stewart73">94</A>,<A
 HREF="node151.html#varah">98</A>].

<P>
We may assume the matrix <B><I>A</I></B> is in Schur form, because reducing it
to this form does not change the values of <B><I>s</I></B> and <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$">.
Consider a cluster of <IMG
 WIDTH="52" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img617.png"
 ALT="$m \geq 1$">
eigenvalues, counting multiplicities.
Further assume the <B><I>n</I></B>-by-<B><I>n</I></B> matrix <B><I>A</I></B> is
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
A  = \left( \begin{array}{cc} A_{11} & A_{12} \\0 & A_{22} \end{array} \right)
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><A NAME="eq2.1"></A><IMG
 WIDTH="145" HEIGHT="54" BORDER="0"
 SRC="img618.png"
 ALT="\begin{displaymath}
A = \left( \begin{array}{cc} A_{11} &amp; A_{12} \\ 0 &amp; A_{22} \end{array} \right)
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.1)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
where the eigenvalues of the <B><I>m</I></B>-by-<B><I>m</I></B> matrix
<B><I>A</I><SUB>11</SUB></B> are exactly those in which we are
interested. In practice, if the eigenvalues on the diagonal of <B><I>A</I></B>
are in the wrong order, routine xTREXC
<A NAME="11598"></A><A NAME="11599"></A><A NAME="11600"></A><A NAME="11601"></A>
can be used to put the desired ones in the upper left corner
as shown.

<P>
We define the <B>spectral projector</B>, or simply projector <B><I>P</I></B> belonging
to the eigenvalues of <B><I>A</I><SUB>11</SUB></B> as
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
P = \left( \begin{array}{cc} I_m & R \\0 & 0 \end{array} \right)
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><A NAME="eq2.2"></A><IMG
 WIDTH="124" HEIGHT="54" BORDER="0"
 SRC="img619.png"
 ALT="\begin{displaymath}
P = \left( \begin{array}{cc} I_m &amp; R \\ 0 &amp; 0 \end{array} \right)
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.2)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
where <B><I>R</I></B> satisfies the system of linear equations
<BR><P></P>
<DIV ALIGN="CENTER">


<!-- MATH
 \begin{equation}
A_{11}R - RA_{22} = A_{12}.
\end{equation}
 -->
<A NAME="eq2.3"></A>
<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP>
<B><I>A</I><SUB>11</SUB><I>R</I> - <I>RA</I><SUB>22</SUB> = <I>A</I><SUB>12</SUB>.
</B>
</TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.3)</TD></TR>
</TABLE>
</DIV>
<BR CLEAR="ALL"><P></P>
Equation (<A HREF="node95.html#eq2.3">4.3</A>) is called a Sylvester equation<A NAME="11611"></A>.
Given the Schur form (<A HREF="node95.html#eq2.1">4.1</A>), we solve equation
(<A HREF="node95.html#eq2.3">4.3</A>) for <B><I>R</I></B> using the subroutine xTRSYL.
<A NAME="11614"></A><A NAME="11615"></A><A NAME="11616"></A><A NAME="11617"></A>

<P>
We can now define <B><I>s</I></B> for the eigenvalues of <B><I>A</I><SUB>11</SUB></B>:
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
s = \frac{1}{\|P\|_2} = \frac{1}{\sqrt{1+\|R\|_2^2}}.
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><IMG
 WIDTH="189" HEIGHT="57" BORDER="0"
 SRC="img620.png"
 ALT="\begin{displaymath}
s = \frac{1}{\Vert P\Vert _2} = \frac{1}{\sqrt{1+\Vert R\Vert _2^2}}.
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.4)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
In practice we do not use this expression since <B>|R|<SUB>2</SUB></B> is hard to
compute. Instead we use the more easily computed underestimate
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
\frac{1}{\sqrt{1+\|R\|_F^2}}
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><IMG
 WIDTH="88" HEIGHT="57" BORDER="0"
 SRC="img621.png"
 ALT="\begin{displaymath}
\frac{1}{\sqrt{1+\Vert R\Vert _F^2}}
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.5)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
which can underestimate the true value of <B><I>s</I></B> by no more than a factor

<!-- MATH
 $\sqrt { \min ( m,n-m ) }$
 -->
<IMG
 WIDTH="133" HEIGHT="38" ALIGN="MIDDLE" BORDER="0"
 SRC="img622.png"
 ALT="$\sqrt { \min ( m,n-m ) }$">.
This underestimation makes our error bounds more conservative.
This approximation of <B><I>s</I></B> is called <TT>RCONDE</TT> in xGEEVX and xGEESX.
<A NAME="11627"></A>

<P>
The <B>separation</B> 
<!-- MATH
 ${\rm sep}(A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
of the matrices <B><I>A</I><SUB>11</SUB></B> and
<B><I>A</I><SUB>22</SUB></B> is defined as the smallest singular value of the linear
map in (<A HREF="node95.html#eq2.3">4.3</A>) which takes <B><I>X</I></B> to 
<!-- MATH
 $A_{11}X - XA_{22}$
 -->
<B><I>A</I><SUB>11</SUB><I>X</I> - <I>XA</I><SUB>22</SUB></B>, i.e.,
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
{\rm sep}(A_{11},A_{22}) = \min_{X \neq 0} \frac{\|A_{11}X - XA_{22}\|_F}
{\| X \|_F}.
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><A NAME="eq2.4"></A><IMG
 WIDTH="300" HEIGHT="48" BORDER="0"
 SRC="img624.png"
 ALT="\begin{displaymath}
{\rm sep}(A_{11},A_{22}) = \min_{X \neq 0} \frac{\Vert A_{11}X - XA_{22}\Vert _F}
{\Vert X \Vert _F}.
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.6)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
This formulation lets us estimate 
<!-- MATH
 ${\rm sep}(A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
using the condition estimator
<A NAME="11645"></A>
xLACON [<A
 HREF="node151.html#hager84">59</A>,<A
 HREF="node151.html#higham1">62</A>,<A
 HREF="node151.html#nick2">63</A>], which estimates the norm of
a linear operator <IMG
 WIDTH="42" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img625.png"
 ALT="$\Vert T \Vert _1$">
given the ability to compute <B><I>Tx</I></B> and
<B><I>T</I><SUP><I>T</I></SUP><I>x</I></B> quickly for arbitrary <B><I>x</I></B>.
In our case, multiplying an
arbitrary vector by <B><I>T</I></B>
means solving the Sylvester equation (<A HREF="node95.html#eq2.3">4.3</A>)<A NAME="11648"></A>
with an arbitrary right hand side using xTRSYL, and multiplying by
<B><I>T</I><SUP><I>T</I></SUP></B> means solving the same equation with <B><I>A</I><SUB>11</SUB></B> replaced by
<B><I>A</I><SUB>11</SUB><SUP><I>T</I></SUP></B> and <B><I>A</I><SUB>22</SUB></B> replaced by <B><I>A</I><SUB>22</SUB><SUP><I>T</I></SUP></B>. Solving either equation
costs at most <B><I>O</I>(<I>n</I><SUP>3</SUP>)</B> operations, or as few as <B><I>O</I>(<I>n</I><SUP>2</SUP>)</B> if <IMG
 WIDTH="57" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img167.png"
 ALT="$m \ll n$">.
Since the true value of <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$">
is <B>|T|<SUB>2</SUB></B> but we use <B>|T|<SUB>1</SUB></B>,
our estimate of <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$">
may differ from the true value by as much as
<IMG
 WIDTH="96" HEIGHT="38" ALIGN="MIDDLE" BORDER="0"
 SRC="img626.png"
 ALT="$\sqrt{m(n-m)}$">.
This approximation to <IMG
 WIDTH="29" HEIGHT="30" ALIGN="MIDDLE" BORDER="0"
 SRC="img2.png"
 ALT="${\rm sep}$">
is called
<TT>RCONDV</TT> by xGEEVX and xGEESX.
<A NAME="11655"></A>

<P>
Another formulation which in principle permits an exact evaluation of

<!-- MATH
 ${\rm sep}( A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
is
<BR>
<DIV ALIGN="RIGHT">


<!-- MATH
 \begin{equation}
{\rm sep}(A_{11},A_{22}) = \sigma_{\min} ( I_{n-m} \otimes A_{11} -
A_{22}^T \otimes I_m )
\end{equation}
 -->

<TABLE WIDTH="100%" ALIGN="CENTER">
<TR VALIGN="MIDDLE"><TD ALIGN="CENTER" NOWRAP><A NAME="eq2.5"></A><IMG
 WIDTH="347" HEIGHT="31" BORDER="0"
 SRC="img627.png"
 ALT="\begin{displaymath}
{\rm sep}(A_{11},A_{22}) = \sigma_{\min} ( I_{n-m} \otimes A_{11} -
A_{22}^T \otimes I_m )
\end{displaymath}"></TD>
<TD WIDTH=10 ALIGN="RIGHT">
(4.7)</TD></TR>
</TABLE>
<BR CLEAR="ALL"></DIV><P></P>
where 
<!-- MATH
 $X \otimes Y \equiv [ x_{ij} Y]$
 -->
<IMG
 WIDTH="124" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img628.png"
 ALT="$X \otimes Y \equiv [ x_{ij} Y]$">
is the Kronecker product of <B><I>X</I></B> and <B><I>Y</I></B>.
This method is
generally impractical, however, because the matrix whose smallest singular
value we need is <B><I>m</I>(<I>n</I>-<I>m</I>)</B> dimensional, which can be as large as
<B><I>n</I><SUP>2</SUP>/4</B>. Thus we would require as much as <B><I>O</I>( <I>n</I><SUP>4</SUP> )</B> extra workspace and
<B><I>O</I>(<I>n</I><SUP>6</SUP>)</B> operations, much more than the estimation method of the last
paragraph.

<P>
The expression 
<!-- MATH
 ${\rm sep}( A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
measures the ``separation'' of
the spectra
of <B><I>A</I><SUB>11</SUB></B> and <B><I>A</I><SUB>22</SUB></B> in the following sense. It is zero if and only if
<B><I>A</I><SUB>11</SUB></B> and <B><I>A</I><SUB>22</SUB></B> have a common eigenvalue, and small if there is a small
perturbation of either one that makes them have a common eigenvalue. If
<B><I>A</I><SUB>11</SUB></B> and <B><I>A</I><SUB>22</SUB></B> are both Hermitian matrices, then 
<!-- MATH
 ${\rm sep}( A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
is just the gap, or minimum distance between an eigenvalue of <B><I>A</I><SUB>11</SUB></B> and an
eigenvalue of <B><I>A</I><SUB>22</SUB></B>. On the other hand, if <B><I>A</I><SUB>11</SUB></B> and <B><I>A</I><SUB>22</SUB></B> are
non-Hermitian, 
<!-- MATH
 ${\rm sep}( A_{11},A_{22})$
 -->
<IMG
 WIDTH="105" HEIGHT="34" ALIGN="MIDDLE" BORDER="0"
 SRC="img623.png"
 ALT="${\rm sep}(A_{11},A_{22})$">
may be much smaller than
this gap.

<P>
<HR>
<!--Navigation Panel-->
<A NAME="tex2html5523"
 HREF="node96.html">
<IMG WIDTH="37" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="next"
 SRC="next_motif.png"></A> 
<A NAME="tex2html5517"
 HREF="node92.html">
<IMG WIDTH="26" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="up"
 SRC="up_motif.png"></A> 
<A NAME="tex2html5513"
 HREF="node94.html">
<IMG WIDTH="63" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="previous"
 SRC="previous_motif.png"></A> 
<A NAME="tex2html5519"
 HREF="node1.html">
<IMG WIDTH="65" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="contents"
 SRC="contents_motif.png"></A> 
<A NAME="tex2html5521"
 HREF="node152.html">
<IMG WIDTH="43" HEIGHT="24" ALIGN="BOTTOM" BORDER="0" ALT="index"
 SRC="index_motif.png"></A> 
<BR>
<B> Next:</B> <A NAME="tex2html5524"
 HREF="node96.html">Error Bounds for the</A>
<B> Up:</B> <A NAME="tex2html5518"
 HREF="node92.html">Further Details: Error Bounds</A>
<B> Previous:</B> <A NAME="tex2html5514"
 HREF="node94.html">Balancing and Conditioning</A>
 &nbsp <B>  <A NAME="tex2html5520"
 HREF="node1.html">Contents</A></B> 
 &nbsp <B>  <A NAME="tex2html5522"
 HREF="node152.html">Index</A></B> 
<!--End of Navigation Panel-->
<ADDRESS>
<I>Susan Blackford</I>
<BR><I>1999-10-01</I>
</ADDRESS>
</BODY>
</HTML>