The Lensing Posterior

At its heart, CMBLensing.jl is centered around the "CMB lensing posterior". We work with both the "joint posterior", which is joint over all variables,

\[ \mathcal{P}(f,\phi,\theta\,|\,d), \]

or the "marginal posterior",

\[ \mathcal{P}(\phi,\theta\,|\,d) \equiv \int \! \mathcal{D}f \; \mathcal{P}(f,\phi,\theta\,|\,d), \]

which is simply the joint posterior analytically marginalized over $f$. Here,

  • $f$ are the CMB fields (T/Q/U),
  • $\phi$ is the lensing potential,
  • $\theta$ are any cosmological parameters,
  • $d$ is the data.

The default data model which is assumed, which is generally flexible enough to handle real experiments (but can be customized), is:

\[ d = \mathbb{A} \, \mathbb{L}(\phi) \, f + n, \]


\[ \mathbb{A} = \mathbb{M} \, \mathbb{B} \]


  • $\mathbb{L}(\phi)$ is the lensing operation
  • $\mathbb{B}$ is an instrumental transfer function or "beam"
  • $\mathbb{M}$ is a user-chosen mask
  • $n$ is the instrumental noise.

Given this model, the joint posterior (up to an unimportant normalization constant) is:

\[ -2\ln\mathcal{P}(f,\phi,\theta\,|\,d) = \frac{\big(d - \mathbb{A} \, \mathbb{L}(\phi) \, f\big)^2}{\mathbb{C}_n} + \frac{f^2}{\mathbb{C}_f(\theta)} + \frac{\phi^2}{C_\phi(\theta)} + \log\det \mathbb{C}_f(\theta) + \log\det C_\phi(\theta), \]

and the marginal posterior is:

\[ -2\ln\mathcal{P}(\phi,\theta\,|\,d) = \frac{d^2}{\mathbb{\Sigma}_d(\phi,\theta)} + \frac{\phi^2}{C_\phi(\theta)} + \log\det \mathbb{\Sigma}_d(\phi,\theta) + \log\det C_\phi(\theta), \]


\[ \Sigma_d = \mathbb{A} \, \mathbb{L}(\phi) \, \mathbb{C}_f(\theta) \, \mathbb{L}(\phi)^\dagger \mathbb{A}^\dagger + \mathbb{C}_n \]


  • $\mathbb{C}_n$ is the noise covariance
  • $\mathbb{C}_f$ is the CMB covariance (i.e. the CMB T, E, and B $C_\ell$'s)
  • $\mathbb{C}_\phi$ is the lensing potential covariance (i.e. $C_\ell^{\phi\phi}$)

and we have used the slighly sloppy notation $x^2/\mathbb{C}$ to mean $x^\dagger \mathbb{C}^{-1} x$.

Posterior basics

using CMBLensing, PythonPlot

CMBLensing uses the function logpdf to compute the log of the joint posterior probability.

To evaluate this posterior, we need the arguments of the probability distribution, $f$, $\phi$, and $\theta$. We also need the data $d$ and host of other operators and covariances which enter the expressions above, which CMBLensing stores in a DataSet object.

First lets load up some simulated data. The function load_sim handles constructing a DataSet and is the recommended way to create the various fields and covariances needed. In this case, let's use 1$\mu$K-arcmin noise and a border mask:

(;f, f̃, ϕ, ds) = load_sim(
    θpix      = 2,
    Nside     = 256,
    T         = Float64,
    pol       = :P,
    μKarcminT = 1,
    L         = LenseFlow(10),
    seed      = 0,
    pixel_mask_kwargs = (edge_padding_deg=1, apodization_deg=0, num_ptsrcs=0),
    bandpass_mask     = LowPass(5000)

The DataSet object, by convention called ds, stores all the aforementioned quantities:

(:d, :Cf, :Cn, :Cn̂, :M, :M̂, :B, :B̂, :logprior, :Cϕ, :Cf̃, :D, :G, :L, :Nϕ)

For example, the data is:



Or the diagonal of the $\mathbb{C}_f$ operator:

plot(diag(ds.Cf), which=[:El :Bl])


We can now evaluate the posterior, for example at the true $f$ and $\phi$:

logpdf(ds; f, ϕ)

There is a particular change-of-variables called the "mixed parameterization" (see our paper for details) which helps make the the posterior a little less correlated and more amenable to sampling and maximization. We can apply the change-of-variables to some point in parameter space as:

f°, ϕ° = mix(ds; f, ϕ);

Then we can evaluate the posterior in the mixed parameterization:

logpdf(Mixed(ds); f°, ϕ°)

Note that this is the same value as above since we are just evaluating the same point in parameter space, just parameterized in terms of different variables.

Above, we didn't specify any cosmological parameters, $\theta$. Because of that, they were fixed at their fiducial values (i.e. the fiducial values which generated the simulated data in the call to load_sim earlier). Current only two parameters can be varied, $r$ (tensor-to-scalar ratio), and $A_\phi$ (the amplitude of $C_\ell^{\phi \phi}$). They can be specified as follows, with non-specified parameters left at their fiducial:

logpdf(ds; f, ϕ, θ=(Aϕ=1.1,))

You can see the slight change compared to what we got above. We can even compute a whole slice through the posterior along $A_\phi$:

Aϕs = range(0.5,1.5,length=50)
plot(Aϕs, [logpdf(ds; f, ϕ, θ=(;Aϕ)) for Aϕ in Aϕs])
ylabel(raw"$\log\mathcal{P}(f_{\rm true}, \phi_{\rm true}, A_\phi\,|\,d)$");


Wiener filtering

If we fix $\phi$ and $\theta$ then maximize the joint posterior, $\mathcal{P}(f,\phi,\theta\,|\,d)$, over $f$, we get

\[ \hat f_{\rm wf} = \big[ \mathbb{C}_f(\theta)^{-1} + \mathbb{L}(\phi)^\dagger \mathbb{A}^\dagger\mathbb{C}_n^{-1}\mathbb{A} \, \mathbb{L}(\phi) \big]^{-1} \mathbb{L}(\phi)^\dagger \mathbb{A}^\dagger\mathbb{C}_n^{-1}d \]

This is simply the Wiener filter of the data given a signal covariance which includes correlations induced by $\phi$, and we can compute it by inverting the operator in brackets above with e.g. conjugate gradient.

In CMBLensing.jl, the argmaxf_logpdf function performs this task. Its arguments are similar to logpdf,

f_wf, = argmaxf_logpdf(ds, (;ϕ); conjgrad_kwargs=(tol=1e-1,progress=true));
Conjugate Gradient: 100%|███████████████████████████████| Time: 0:00:11

Wiener filtering effectively delenses the data by a given $\phi$. Here we are Wiener filtering at the true $\phi$, so this is perfect delensing. Note below the large amount of B mode power in the data, as well as the aliasing near the border mask, and how this is removed in the Wiener filtered B, which visually is tracing the true B map up to the level of the noise.

plot([ds.d, f_wf, f], title=["data ", "Wiener filtered ", "true "] .* ["E" "B"]);


Posterior gradients

Fundamental to maximization and sampling algorithms in high dimensions are gradients of the posterior. CMBLensing.jl uses the interface provided by the automatic differentiation library Zygote to compute these gradients.

using Zygote

Zygote provides the gradient function, which takes a function for which to evaluate the gradient as the first argument, and the value at which to evaluate the gradient as the second argument. For example:

gradient(x->3x^2+2x, 1)

Gradients of the CMB lensing posterior work in exactly the same way. For example, to take the gradient with respect to $\phi$ in the mixed parametrization, where we fix $f$ to the Wiener filter computed above and evaluate the gradient at $\phi=0$, we can do:

gϕ = gradient(ϕ° -> logpdf(Mixed(ds); f°=f_wf, ϕ°), 0ϕ)[1];

Here's what this gradient looks like multiplied by $\mathbb{C}_\phi$. This is actually the first gradient step in the iterative joint maximum a posteriori estimation. You can see below this very broadly matches the truth:

plot([ds.Cϕ*gϕ ϕ], title=["gradient step" "true ϕ"])


You are free to manipulate the fields inside of the function whose gradient is being taken, and Zygote will automatically propagate the chain rule for you:

gradient(ϕ -> -2logpdf(ds; f=f_wf, ϕ=3ϕ), 0ϕ)[1] ≈ -6 * gradient(ϕ -> logpdf(ds; f=f_wf, ϕ), 0ϕ)[1]

Automatic differentiation with respect to any and all arguments of logpdf works, including the cosmological parameters:

gradient(Aϕ -> logpdf(ds; f, ϕ, θ=(;Aϕ)), 1)

Currently Zygote is working with many (but not all) operations that you can do to fields. If you run into things which Zygote is not able to differentiate successfully (generally you'll get some error), please feel free to file an Issue with CMBLensing.jl.