Hello
I am using camb to obtain the ISW cross-correlation spectra of the CMB with a certain mass tracer (use do_lensing=true and a modified tracer).
When I run the code for different values of w around a fiducial model, say: 1%, 2%, 5%, etc., I get a noticeable scatter that hinders me from deriving the slope around the fiducial (see plot). Increasing the accuracy_boost improve things a bit but I have to go higher than 3 and things get very, very slow.
Can anyone suggest how to tackle this problem?
thank you
http://roger.phy.unp.ac.za/~angel/cross.eps
CAMB - numerical precision for different w
-
- Posts: 2
- Joined: February 22 2006
- Affiliation: Univeristy of KwaZulu-Natal - South Africa
CAMB - numerical precision for different w
It looks to me like you're trying to determine changes in [tex]C_l[/tex] which are less than one percent. I don't know as much about CAMB's accuracy, but I'd get very suspicious of any result out of cmbfast that relied on things which changed the [tex]C_l[/tex] by such a small amount. I suspect you'll have to do at least one of the following:
1. Look at much larger changes in your parameter ([tex]w[/tex])
2. Turn the accuracy way up and cope with the slowdown (and hope that there aren't unknown errors in the code at the sub-percent level)
3. Come up with some kind of analytic or semi-analytic technique to quickly calculate [tex]dC_l/dw[/tex] to the precision you need (looks like 0.1% or better)
1. Look at much larger changes in your parameter ([tex]w[/tex])
2. Turn the accuracy way up and cope with the slowdown (and hope that there aren't unknown errors in the code at the sub-percent level)
3. Come up with some kind of analytic or semi-analytic technique to quickly calculate [tex]dC_l/dw[/tex] to the precision you need (looks like 0.1% or better)
-
- Posts: 2
- Joined: February 22 2006
- Affiliation: Univeristy of KwaZulu-Natal - South Africa
CAMB - numerical precision for different w
Thanks Ben
I was thinking about your first solution. I agree, it might be better to look at larger variations of w, then fit a more general curve and then get the derivatives at the point of interest.
Angel
I was thinking about your first solution. I agree, it might be better to look at larger variations of w, then fit a more general curve and then get the derivatives at the point of interest.
Angel