Read on Twitter
Happy to present new @SPTelescope results today analyzing the gravitational lensing of the CMB with the type of fully optimal/Bayesian method thats going to become critical in the next decade.
A cool aspect of this analysis are the maps of dark matter + baryons we can make:
A cool aspect of this analysis are the maps of dark matter + baryons we can make:
This is possible because the gravitational field from this matter distorts ("lenses") the CMB. From this we can figure out where the gravity is and hence the matter ( brighter color = more matter). And the analysis doesnt give us just a single map, but a set of possible maps.
If you look closely at these maps (shown in the movie), you'll see structures which are pretty much there in every frame, meaning we're sure they're there. Near the edges (where we didn't have much data), it looks like random noise, so we're not that sure what's there at all.
This type of analysis turns out to be not just pretty animations, but lets us get tighter constraints than traditional methods. Here, we beat the existing way (the "quadratic estimate") by 17% in terms of constraining the amplitude of the structures (Aϕ) from this data.
It also gives a way to undo the distortion, ie "delense" the CMB. Eg here are maps of the delensed E-mode polarization from the same patch of sky (if you zoom in, you can see the warping due to the different possible solutions for the undistortion)
And then compute constraints on Aϕ and parameters inferred from the delensed CMB (like one called ΔAL below), while correctly accounting for degeneracies. This (I think) would be *really* hard done in any other way, or, at the very least, it hasn't been done yet.
People have know for a long time this type of thing was possible, but this is the first time figuring out exactly how to do it & actually executing it in real life. A TLDR: @JuliaLanguage + GPU + automatic differentiation + flows + Hamiltonian Monte Carlo https://github.com/marius311/CMBLensing.jl
I suspect we'll be improving the methodology in the future, maybe with ideas from machine learning, likelihood-free inference, variational inference, or others... But for now, we have a solution, its exact and it works, and we can build from it.
Also in the future, the importance of stuff like this will only become more. Eg for the next-gen CMB-S4 @DOE_Stage_4_CMB, that 17% improvement-over-traditional-methods for Aϕ turns into 50%, or as much as 700% for other quantities.
Here eg is what the movie from the first tweet might look like if we had something as sensitive as CMB-S4 today (the right is a simulation, taking one of the matter maps from our real analyis as "truth"). We'd resolve these structure with drastically more certainty.
Anway, check out the paper, it was a big work by many in the collaboration, and don't hesitate to get in touch if you have any questions or comments!
https://arxiv.org/abs/2012.01709
https://arxiv.org/abs/2012.01709
