search

Chris-Pedersen / LaCE_manager

Cosmological emulator for the 1D flux power spectrum of the Lyman-alpha forest
3 stars 1 forks source link

Avoid interp mock #70

Closed andreufont closed 2 years ago

andreufont commented 2 years ago

Instead of evaluating the polynomial at k_sim, and then interpolate to k_data, we can evaluate directly the polynomial at k_data.

The contours for the small test (Ohio mocks) look significantly better, so I'm testing this now at NERSC for a larger (compressed) run.

@Chris-Pedersen , are you using the master branch for your compression project? If so we should be careful to clearly identify runs pre and post this change. Would you rather have an option in data_MPGadget that enables this new feature, and have it default to the old code for backward-compatibility?

andreufont commented 2 years ago

This is the comparison of contours for the small test.

test_avoid_interp_ohio

Chris-Pedersen commented 2 years ago

Oh wow, thanks a lot for noticing this. Sorry Ohio mocks mean the DESI mock from Naim, correct? Would we be able to run the same comparison for 0.2*Chabanier2019, which I have been using for the compression project? The biases you get for Delta2_star and n_star are significantly worse than what I have been getting using Chabanier. I guess this is down to the fact that the covariance is just lower overall. Also this is the central sim correct? For this sim ln_t0 should be 0, so I guess its a bit suprising that we don't recover this using the improved fit.

Would you rather have an option in data_MPGadget that enables this new feature, and have it default to the old code for backward-compatibility?

I think I would rather test it comprehensively, and then just merge without backwards compatibility (as doing it the other way has lead to an insane proliferation of flags I never use any more..). Once we are confident this procedure is better we won't want to do it the other way again anyway.

andreufont commented 2 years ago

These are longer chains (using example_compressed.config) with 8 IGM parameters and the new polyfit emulator, comparing 5 x eBOSS vs Ohio (DESI mocks from Naim).

new_poly_central_sim

Chris-Pedersen commented 2 years ago

Ok so the bias in Delta^2 star is significantly reduced using this new fit. I think we should switch to this for the compression paper in that case. Also glad to see that tau_0 is recovered better once the full IGM is varied (green contours are the same in as in the previous figure, just with the full IGM varied this time, correct?)

andreufont commented 2 years ago

Yes, I'm only showing 3 parameters out of the 10 parameters I varied.

I'm happy to use this change for the compression paper, probably better since it is the first paper using the polyfit emulator.

andreufont commented 2 years ago

@Chris-Pedersen - Why don't you give it a shot in one of your chains, and if you are happy with it we merge?

Chris-Pedersen commented 2 years ago

Yes sounds good, will set a few of the compression paper chains to run at Flatiron. I'll make a direct comparison using 20% eBOSS, and if everything looks ok I'll merge.

Chris-Pedersen commented 2 years ago

Running another test on this with a tighter initialisation which will hopefully not have islands. Will post results here and merge once they are done.