%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % %A efficiency.tex DESIGN documentation Leonard Soicher % % % \def\DESIGN{\sf DESIGN} \def\GRAPE{\sf GRAPE} \def\nauty{\it nauty} \def\Aut{{\rm Aut}\,} \def\x{\times} \Chapter{Matrices and efficiency measures for block designs}
In this chapter we describe functions to calculate certain matrices
associated with a block design, and the function `BlockDesignEfficiency'
which determines certain statistical efficiency measures of a 1-design.
We also document the utility function `DESIGN_IntervalForLeastRealZero',
which is used in the calculation of E-efficiency measures, but has much
wider application.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \Section{Matrices associated with a block design}
\>PointBlockIncidenceMatrix( <D> )
This function returns the point-block incidence matrix <N> of the
block design <D>. This matrix has rows indexed by the points of <D>
and columns by the blocks of <D>, with the $(i,j)$-entry of <N> being
the number of times point $i$ occurs in `<D>.blocks[<j>]'.
This function returns the concurrence matrix <L> of the block design <D>.
This matrix is equal to $NN^{\rm T}$, where $N$ is the point-block
incidence matrix of <D> (see "PointBlockIncidenceMatrix") and
$N^{\rm T}$ is the transpose of $N$. If <D> is a binary block design
then the $(i,j)$-entry of its concurrence matrix is the number of blocks
containing points $i$ and $j$.
This function returns the information matrix <C> of the block design <D>.
This matrix is defined as follows. Suppose <D> has $v$ points and $b$
blocks, let $R$ be the $v\times v$ diagonal matrix whose $(i,i)$-entry
is the replication number of the point $i$, let $N$ be the point-block
incidence matrix of <D> (see "PointBlockIncidenceMatrix"), and let $K$
be the $b\times b$ diagonal matrix whose $(j,j)$-entry is the length of
`<D>.blocks[<j>]'. Then the *information matrix* of is
$C:=R-NK^{-1}N^{\rm T}$. If <D> is a 1-$(v,k,r)$ design then this expression
for <C> simplifies to $rI-k^{-1}L$, where $I$ is the $v\times v$ identity
matrix and $L$ is the concurrence matrix of <D> (see "ConcurrenceMatrix").
Let <D> be a 1-$(v,k,r)$ design with $v>1$, let <eps> be a positive
rational number (default: $10^{-6}$), and let <includeMV> be a boolean
(default: `false'). Then this function returns a record containing
information on statistical efficiency measures of <D>. These measures
are defined below. See \cite{Extrep}, \cite{BaCa} and \cite{BaRo}
for further details. All returned results are computed using exact
algebraic computation.
The component `<eff>.A' contains the A-efficiency measure for ,
`<eff>.Dpowered' contains the D-efficiency measure of raised to the
power $v-1$, and `<eff>.Einterval' is a list $[a,b]$ of non-negative
rational numbers such that if $x$ is the E-efficiency measure of <D>
then $a\le x\le b$, $b-a\le$<eps>, and if $x$ is rational then $a=x=b$.
Moreover `<eff>.CEFpolynomial' contains the monic polynomial over the
rationals whose zeros (counting multiplicities) are the canonical
efficiency factors of the design <D>. If `<includeMV>=true' then
additional work is done to compute the MV- (also called E$'$-) efficiency
measure, and then `<eff>.MV' contains the value of this measure. (This
component may be set even if `<includeMV>=false', as a byproduct of
other computation.)
We now define the canonical efficiency factors and the A-, D-, E-,
and MV-efficiency measures of a 1-design.
Let $D$ be a 1-$(v,k,r)$ design with $v\ge 2$, let $C$ be the information
matrix of $D$ (see "InformationMatrix"), and let $F:=r^{-1}C$.
The eigenvalues of $F$ are all real and lie in the interval $[0,1]$.
At least one of these eigenvalues is zero: an associated eigenvector is
the all-$1$ vector. The remaining eigenvalues $\delta_1\le\delta_2\le \cdots\le\delta_{v-1}$ of $F$ are called the *canonical efficiency
factors* of $D$. These are all non-zero if and only if $D$ is connected
(that is, the point-block incidence graph of $D$ is a connected graph).
If $D$ is not connected, then the A-, D-, E-, and MV-efficiency measures
of $D$ are all defined to be zero. Otherwise, the *A-efficiency
measure* is $(v-1)/\sum_{i=1}^{v-1}1/\delta_i$ (the harmonic mean
of the canonical efficiency factors), the *D-efficiency measure*
is $(\prod_{i=1}^{v-1}\delta_i)^{1/(v-1)}$ (the geometric mean of
the canonical efficiency factors), and the *E-efficiency measure* is
$\delta_1$ (the minimum of the canonical efficiency factors).
If $D$ is connected, and the MV-efficiency measure is required,
then it is computed as follows. Let $F:=r^{-1}C$ be as before,
and let $P:=v^{-1}J$, where $J$ is the $v\times v$ all-1 matrix. Set
$M:=(F+P)^{-1}-P$, making $M$ the ``Moore-Penrose inverse'' of $F$ (see \cite{BaCa}). Then the *MV-efficiency measure* of $D$ is the minimum
value (over all $i,j\in\{1,\ldots,v\}$, $i\not=j$) of
$2/(M_{ii}+M_{jj}-M_{ij}-M_{ji})$.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% \Section{Computing an interval for a certain real zero of a rational polynomial}
We document a {\DESIGN} package utility function used in the calculation
of the `Einterval' component above, but is more widely applicable.
Suppose that <f> is a univariate polynomial over the rationals, <a>,
<b> are rational numbers with $<a>\le <b>$, and <eps> is a positive
rational number.
If <f> has no real zero in the closed interval $[<a>,<b>]$, then this
function returns the empty list. Otherwise, let $\alpha$ be the least
real zero of <f> such that $<a>\le\alpha\le <b>$. Then this function
returns a list $[c,d]$ of rational numbers, with $c\le\alpha\le d$
and $d-c\le <eps>$. Moreover, $c=d$ if and only if $\alpha$ is rational
(in which case $\alpha=c=d$).
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung ist noch experimentell.
