search

xzackli / Bolt.jl

differentiable boltzmann code
MIT License
42 stars 5 forks source link

RSA issues #64

Closed jmsull closed 1 year ago

jmsull commented 2 years ago

The RSA trigger is set to the prescription given in the CLASS 2 paper - but I found for k=1h/Mpc, the RSA turns on too late, and the photon monopole has already started exploding. This (and a related bug I fixed in ie branch) probably is causing the residual large scale-dependent disagreements in the radiation transfer functions shown in #44, and (maybe?) could be contributing significantly to the 0.1% scale-dependent deviations in $P_L(k)$.

jmsull commented 2 years ago

This is almost certainly due to a bug in the photon truncation - will close this when this change is pushed and can be referenced.

jmsull commented 2 years ago

May as well use this issue as an opportunity to deal with the RSA photon dipole issues as well. Something is very wrong even when increased rtol is used (the "8" is 1e-8 rtol up from 1e-5). This is using IE branch code after photon truncation is fixed. Orange and gold are using RSA but are nowhere near the correct value (or the CLASS RSA). Likely a code translation bug.

Screen Shot 2022-06-25 at 19 10 19
jmsull commented 1 year ago

Closing this since 1. I no longer see it in conformal time integration, which is what we decided we will always use, and 2. we may not even need RSA