Forum: help

RE: mixed exponential created by distr not working correctly [ Reply ]
By: Peter Ruckdeschel on 2017-11-19 17:51

I have no time to look at this in detail (will do so later ~ mid Dec),but on first glance this should be the default minimal grid width that is byting you here -- have you looked at changing this in distroptions() ?

Later reply (Dec 1 2017): Indeed, you can fix this (at the cost of large object sizes, and slower computations, though) by using a finer resolution:

distroptions(DefaultNrGridPoints=2^21)
## uses 2e6 gridpoints now instead of 4e3

mixobj <- UnivarMixingDistribution(
Exp(0.000281587060464458),
Exp(4.60503874426382e-05),
Exp(8.66633661057179e-06),
Exp(2.06028276672794e-06),
Exp(3.35389385119948e-07),
Exp(7.42313835767416e-08),
mixCoeff = ws)

qtst <- q(mixobj)
ptst <- p(mixobj)

qtst(ptst(10000))
[1] 10000.42

## so I guess you need to seek a compromise between object size+computation time on one hand and accuracy on the other.

I hope this helps -- best, Peter

mixed exponential created by distr not working correctly [ Reply ]
By: Allen Tan on 2017-11-17 02:40

[forum:45467]

Thank you for this great package, it is very helpful in working with mixed distributions.

However, when I am working with mix exponential distributions with rate very close to 0 (very large means). The q function seems not to work correctly.

example codes:
ws <- c(0.467605678, 0.214252136, 0.177945838, 0.116423842, 0.022384060, 0.001388445)

mixobj <- UnivarMixingDistribution(
Exp(0.000281587060464458),
Exp(4.60503874426382e-05),
Exp(8.66633661057179e-06),
Exp(2.06028276672794e-06),
Exp(3.35389385119948e-07),
Exp(7.42313835767416e-08),
mixCoeff = ws)

qtst <- q(mixobj)
ptst <- p(mixobj)

qtst(ptst(10000))

This should return 10000, if the q function is working correctly, but not in this case. I think it generally works, so I suspect this might be an issue related to machine number precision? It would be great if it can be fixed.

Thanks!
Allen