diff options
author | enge <enge@211d60ee-9f03-0410-a15a-8952a2c7a4e4> | 2009-06-11 17:17:21 +0000 |
---|---|---|
committer | enge <enge@211d60ee-9f03-0410-a15a-8952a2c7a4e4> | 2009-06-11 17:17:21 +0000 |
commit | 8ba2bc726642350ab369cd4f16ad040bc7af346e (patch) | |
tree | afc6acd89b761f5ce0201a84d536c77d5bd3ae13 | |
parent | e6cc675b57135e599527d37d9de4e63e4acc84ad (diff) | |
download | mpc-8ba2bc726642350ab369cd4f16ad040bc7af346e.tar.gz |
algorithms.tex: added second error bound for norm
git-svn-id: svn://scm.gforge.inria.fr/svn/mpc/trunk@599 211d60ee-9f03-0410-a15a-8952a2c7a4e4
-rw-r--r-- | doc/algorithms.tex | 36 |
1 files changed, 33 insertions, 3 deletions
diff --git a/doc/algorithms.tex b/doc/algorithms.tex index 57e7d48..ac3179b 100644 --- a/doc/algorithms.tex +++ b/doc/algorithms.tex @@ -275,7 +275,7 @@ functions with real arguments and values. Those already contained in \subsubsection {Division} -\label {sssec:realpropdiv} +\label {sssec:proprealdiv} Let \[ @@ -549,7 +549,37 @@ The analogous bound for the second error term yields 2^{\Exp (\appro x) - p} \end {equation} The values $\epsilon_{X, 1}^+$ may be estimated as explained at the end -pf \S\ref {sssec:propmul}. +of \S\ref {sssec:propmul}. + +Alternatively, one might write +$|\corr {x_1}^2 - \appro {x_1}^2| += \appro {x_1}^2 \cdot +\left| 1 - \left( \frac {|\corr {x_1}|}{|\appro {x_1}|} \right)^2 \right|$ +and estimate the both-sided relative error as in \S\ref {sssec:proprealdiv} +to obtain, under the assumption that $\epsilon_{R, 1}^- < 2$, +\[ +| \corr {x_1}^2 - \appro {x_1}^2 | +\leq +\max \big( + \epsilon_{R, 1}^+ (2 + \epsilon_{R, 1}^+), + \epsilon_{R, 1}^- (2 - \epsilon_{R, 1}^-) +\big) +\appro {x_1}^2 +\] +Adding the corresponding bound on the second term, using +Proposition~\ref {prop:relerror} and letting +$k_1 = \max ( k_{R, 1}, k_{I, 1})$ and +$\epsilon_1 = \max ( \epsilon_{R, 1}^-, \epsilon_{R, 1}^+, + \epsilon_{I, 1}^-, \epsilon_{I, 1}^+ )$ +leads to +\begin {equation} +\label {eq:propnormalt} +\error (\appro x) +\leq +\epsilon_1 (2 + \epsilon_1) 2^{\Exp (\appro x)} +\leq +2 k_1 (2 + \epsilon_1) 2^{\Exp (\appro x) - p} +\end {equation} \subsubsection {Division} @@ -563,7 +593,7 @@ Let Then the propagated error may be derived by cumulating the errors obtained for multiplication in \S\ref {sssec:propmul}, the norm in \S\ref {sssec:propnorm} and the division by a real in -\S\ref {sssec:realpropdiv}. +\S\ref {sssec:proprealdiv}. We note \begin{align*} \corr a &= \Re (\corr {z_1} \overline {\corr {z_2}}) & |