| lmrdia46 {lmomco} | R Documentation |
L-moment Ratio Diagram Components of Tau4 and Tau6
Description
This function returns a list of the L-kurtosis (\tau_4 and sixth L-moment ratio \tau_6, respectively) ordinates for construction of L-moment Ratio (L-moment diagrams) that are useful in selecting a distribution to model the data.
Usage
lmrdia46()
Details
The lmrdia46 returns a list of the tables for drawing the trajectories of the distributions by its access of .lmomcohash$t46list created by the inst/doc/SysDataBuilder02.R script for sysdata.rda construction used by the lmomco package itself. The lookup table references below are pointing to the inst/doc/t4t6 subdirectory of the package.
A lookup table for the Exponential Power distribution is provided as PowerExponential.txt (.lmomcohash$tau46list$pwrexp), and this distribution is a special case of the Asymmetric Exponential Power4 (lmomaep4) (.lmomcohash$tau46list$aep4).
A lookup table for the Symmetric Stable distribution is provided as StableDistribution.txt (.lmomcohash$tau46list$symstable).
A lookup table for the Student t distribution is provided as StudentT.txt
(.lmomcohash$tau46list$st2), and this distribution is the same as the Student 3t (lmomst3) (.lmomcohash$tau46list$st3).
A lookup table for the Tukey Lamda distribution is provided as SymTukeyLambda.txt
(.lmomcohash$tau46list$tukeylam), and this distribution is not quite the same as the Generalized Lambda distribution (lmomgld) (.lmomcohash$tau46list$gld).
The normal distribution plots as a point in a Tau4-Tau6 L-moment ratio diagram as.lmomcohash$tau46list$nor for which \tau_4^\mathrm{nor} = 30/\pi\times \mathrm{atan}(\sqrt{2}) - 9 = 0.1226017 and \tau_6^\mathrm{nor} = 0.04365901 (numerical integration).
Finally, the Cauchy and Slade distributions are symmetrical and can be plotted as well on Tau4-Tau6 L-moment ratio diagram if we permit their trim=1 TL-moments to be shown instead. These are inserted into the returned list as part of the operation of lmrdia46().
Tukey Lambda Notes—The Tukey Lambda distribution is a simpler formulation than the Generalized Lambda.
Q(F) = \frac{1}{\lambda} \biggl[F^\lambda - (1-F)^\lambda \biggr]\mbox{,}
for nonexceedance probability F and \lambda \ne 0 and
Q(F) = \mathrm{log}\biggl(\frac{F}{1-F}\biggr)\mbox{,}
for \lambda = 0 using the natural logarithm.
Inspection of the distribution formulae inform us that the variation in the distribution, the scaling factor 1/\lambda to far left in the first definition, for instance, implies that the L-scale (\lambda_2) is not constant and varies with \lambda. The second L-moment of the Tukey Lambda (all odd order L-moments are zero) is
\lambda_2 = \frac{2}{\lambda}\biggl[ -\frac{1}{1+\lambda} + \frac{2}{2+\lambda}\biggr]\mbox{, and}
the fourth and sixth L-moments are
\lambda_4 = \frac{2}{\lambda}\biggl[ -\frac{1}{1+\lambda} + \frac{12}{2+\lambda} - \frac{30}{3+\lambda} + \frac{20}{4+\lambda}\biggr]\mbox{,}
\lambda_6 = \frac{2}{\lambda}\biggl[ -\frac{1}{1+\lambda} + \frac{30}{2+\lambda} - \frac{210}{3+\lambda} + \frac{560}{4+\lambda} - \frac{630}{5+\lambda} + \frac{252}{6+\lambda}\biggr]\mbox{\, and}
\tau_4 = \lambda_4 / \lambda_2 and \tau_6 = \lambda_6 / \lambda_2. The Tukey Lambda is not separately implemented in the lmomco package. It is provided herein for theoretical completeness, but it is possible to implement the Tukey Lambda by the following example:
tukeylam <- .lmomcohash$tau46list$gld_byt6tukeylam
lmr1 <- tukeylam[tukeylam$lambda2 == 1, ] # L-scale equal to one (for instance)
lmr1 <- vec2lmom(c(0, lmr1$lambda2, 0, lmr1$tau4, 0, lmr1$tau6))
tuk1 <- pargld(lmr1, aux="tau6")
print(tuk1$para, 12)
# xi alpha kappa h
# 2.50038766315e-04 -5.82180675380e+03 -1.71745206920e-04 -1.71702273015e-04
lambda <- mean(tuk1$para[3:4]) # remember optimization is used for parameters in
# GLD parlance and so the two shape parameters are not constrained in pargld()
# to be numerically identical. So, here, let us compute a mean of the two and then
# use that as the Lambda in the distribution.
eps <- 1/tuk1$para[2] - lambda
message("EPS should be very close to zero, eps = ", eps, " !!!!!")
tuk2 <- vec2par(c(0, 1/lambda, lambda, lambda), type="gld") # now Tukey Lambda
lmr2 <- lmomgld(tuk2)
"ofunc" <- function(lambda, lambda2=NA) {
tukeyL2 <- ( 2 / lambda ) * ( -1 / (1+lambda) + 2 / (2+lambda) )
return(lambda2 - tukeyL2)
}
lam <- uniroot(ofunc, interval=c(-1, 1), lambda2=1)$root
tuk3 <- vec2par(c(0, 20/lam, lam, lam), type="gld")
lmr3 <- lmomgld(tuk3)
gld5 <- pargld(lmr3, aux="tau5"); gldlmr5 <- theoLmoms(gld5, nmom=6)
gld6 <- pargld(lmr3, aux="tau6"); gldlmr6 <- theoLmoms(gld6, nmom=6)
plotlmrdia46(lmrdia46(), nogld_byt5opt=FALSE)
points(gldlmr5$ratios[4], gldlmr5$ratios[6], pch=16, col="purple")
points(gldlmr6$ratios[4], gldlmr6$ratios[6], pch=21, col="purple", bg="white")
# See how GLD by tau5 optimization, which leaves Tau6 to float plots on the
# "gld_byt5opt" trajectory, but GLD by tau6 optimization, plots on the Tukey
# Lambda line, and gld6$para[2] / (1/gld6$para[3]) is equal to the 20 in the
# parameter setting for tuk3.
The finally differences in the L-moments between the two lmr objects are all are reasonably close to zero with the recognition that optim() has been involved getting us close to the Tukey Lambda that we desire as a GLD with fixed shape parameters and scale factor equal to the inverse of the shape parameter. The demonstration to how to thus acquire a Tukey Lambda from GLD implementation in the lmomco package is thus shown.
Value
An R list is returned.
aep4 |
|
gld_byt5opt |
|
gld_byt6tukeylam |
|
nor |
|
pdq4 |
|
pwrexp |
|
st2 |
|
st3 |
|
symstable |
|
tukeylam |
(reference copy of |
cau |
|
sla |
|
Author(s)
W.H. Asquith
See Also
Examples
lratios <- lmrdia46()