lgamma1p: Accurate `log(gamma(a+1))`

Description

Compute $$l\Gamma_1(a) := \log\Gamma(a+1) = \log(a\cdot \Gamma(a)) = \log a + \log \Gamma(a),$$ which is “in principle” the same as log(gamma(a+1)) or lgamma(a+1), accurately also for (very) small $a$ $(0 < a < 0.5)$.

Usage

lgamma1p (a, tol_logcf = 1e-14)     
lgamma1p.(a, cutoff.a = 1e-6, k = 3)
lgamma1p_series(x, k)

Arguments

a, x

a numeric vector.

tol_logcf

for lgamma1p(): a non-negative number ...

cutoff.a

for lgamma1p.(): a positive number indicating the cutoff to switch from ...

an integer, the number of terms in the series expansion used internally.

Value

a numeric vector with the same attributes as a.

Details

lgamma1p() is an R translation of the function (in Fortran) in Didonato and Morris (1992) which uses a 40-degree polynomial approximation.

lgamma1p_series(x, k) is Taylor series approximation of order k, (derived via Maple), which is $-\gamma x + \pi^2 x^2/ 12 + O(x^3)$, where $\gamma$ is Euler's constant 0.5772156649....

References

Didonato, A. and Morris, A., Jr, (1992) Algorithm 708: Significant digit computation of the incomplete beta function ratios. ACM Transactions on Mathematical Software, 18, 360--373; see also pbeta.

Examples

Run this code

# NOT RUN {
curve(-log(x*gamma(x)), 1e-30, .8, log="xy", col="gray50", lwd = 3,
      axes = FALSE, ylim = c(1e-30,1))
sfsmisc::eaxis(1); sfsmisc::eaxis(2)
at <- 10^(1-4*(0:8))
abline(h = at, v = at, col = "lightgray", lty = "dotted")
curve(-lgamma( 1+x), add=TRUE, col="red2", lwd=1/2)# underflows even earlier
curve(-lgamma1p (x), add=TRUE, col="blue")
curve(-lgamma1p.(x), add=TRUE, col=adjustcolor("forest green",1/4),
      lwd = 5, lty = 2)
for(k in 1:7)
  curve(-lgamma1p_series(x, k=k), add=TRUE, col=paste0("gray",30+k*8), lty = 3)
# }

Run the code above in your browser using DataCamp Workspace