cosmomc 2D contour
-
- Posts: 12
- Joined: April 22 2006
- Affiliation: Sun Yat-Sen University
- Contact:
cosmomc 2D contour
I use getdist to plot 2D contours. I finid the 2D contours do not match the color map of marginalized likelihood. I know both the marginalized likelihood and 2D contours are smoothed, but are they smoothed in different manner?
Here is an example of 2D plot http://www.cita.utoronto.ca/~zqhuang/ALLw0wa_2D.ps
Here is an example of 2D plot http://www.cita.utoronto.ca/~zqhuang/ALLw0wa_2D.ps
-
- Posts: 1943
- Joined: September 23 2004
- Affiliation: University of Sussex
- Contact:
Re: cosmomc 2D contour
By default plot colours are not marginalized likelihoods but mean likelihoods. They only agree for Gaussians (see CosmoMC paper appendix).
-
- Posts: 12
- Joined: April 22 2006
- Affiliation: Sun Yat-Sen University
- Contact:
cosmomc 2D contour
Thanks!
Actually I calculated marginalized likelihood (i.e. the sum of numbers of samples per grids) without any smoothing. I sort the grids according to the marginalized likelihood. Then I pick out the grids sequentially until I get enough grids containing 68.3/95.4 percent of total number of samples.
All the grids I picked out should be the interior of unsmoothed 68.3/95.4 CL contours. However, they are very very unsmoothly distributed (see http://www.astro.utoronto.ca/~zqhuang/u ... 695458.eps). This should not be surprising, since the total number of samples used for plotting is only~10^4.
Nevertheless, I decided to smooth the contours on my own. When doing the smoothing, I carefully kept the total number of samples inside the contour to be 68.3/95.4 percent. Here are the smoothed contours (http://www.astro.utoronto.ca/~zqhuang/u ... 696316.eps).
Not surprising, the contours I found are different from getdist output. Because typically cosmomc doesn't produce enough samples to decide smooth 2D contours, it is hard to say who is wrong. But it is very interesint to do a comparism.
Actually I calculated marginalized likelihood (i.e. the sum of numbers of samples per grids) without any smoothing. I sort the grids according to the marginalized likelihood. Then I pick out the grids sequentially until I get enough grids containing 68.3/95.4 percent of total number of samples.
All the grids I picked out should be the interior of unsmoothed 68.3/95.4 CL contours. However, they are very very unsmoothly distributed (see http://www.astro.utoronto.ca/~zqhuang/u ... 695458.eps). This should not be surprising, since the total number of samples used for plotting is only~10^4.
Nevertheless, I decided to smooth the contours on my own. When doing the smoothing, I carefully kept the total number of samples inside the contour to be 68.3/95.4 percent. Here are the smoothed contours (http://www.astro.utoronto.ca/~zqhuang/u ... 696316.eps).
Not surprising, the contours I found are different from getdist output. Because typically cosmomc doesn't produce enough samples to decide smooth 2D contours, it is hard to say who is wrong. But it is very interesint to do a comparism.
-
- Posts: 54
- Joined: September 28 2004
- Affiliation: University College London
cosmomc 2D contour
I have in the past not found the getdist smoothing to work well for long thin diagonal distributions like the one you are attempting to plot. Following advice from Antony I do numerous tests on them now where I test smoothing vs non-smoothed, as well as sensitivity to increasing the number of bins. Recently I have become even more conservative and compare the resulting plot against my own contouring code before accepting it.