Theoretical higher order predictions for wavelet l1-norm

Vilasini Tinnaneri Sreekanth, Sandrine Codis, Alexander Barthlemey, Jean-Luc Starck












slides at vilasinits.github.io/Talks/ActionDarkEnergy23-Annecy/

Colloque national Action Dark Energy- 07 Nov 2023


Weak Lensing is one of the crucial probe in modern cosmology.


  • Direct observations of the mass distribution
  • Can help us in better understanding the universe
Source: UNIONS

The usual method of analysing the WL map is by using the 2 point statistics.

Data
2 point statistics
Theory
2 point statistics

Likelihood/MCMC





2 point statistics is not sufficient. Why?

$\Longrightarrow$ The power spectrum shows power as the mean squared amplitude at each frequency line but includes no phase information.

Higher-order statistics

Higher-order statistics

Data
Higher order statistics
Simulation
Higher order statistics

Likelihood/MCMC





Data
Higher order statistics
Simulation

Multiple realisations

Higher order statistics

Likelihood/MCMC

  • Need huge amount of storage
  • Huge amount of computational resourses

This work





Data
Higher order statistics
Simulation

Multiple realisations

Higher order summary statistics

Likelihood/MCMC

Theory for HOS
wavelet $\ell_1$-norm PDF

One-point Probability Distribution Functions

  • Contain a significant amount of cosmological information
  • potential to break cosmological parameter degeneracies
  • Has the advantage of being straightforward to measure.
Source: Codis et al. (2021)

wavelet $\ell_1$-norm

  • $\ell_1 = \sum_k|x_k|$
  • the $\ell_1$-norm carries the information encoded in all pixels of the map.
  • It is shown in Ajani et al. (2021) that it remarkably outperforms commonly used summary statistics, such as the power spectrum or the combination of peak and void counts.
Description of the image
Source: Ajani et al. (2021)

Wavelet decomposition of a convergence map: example

||

+

My work: Theoretical prediction of wavelet $\ell_1$-norm based on Large Deviation Theory

  • Part 1: What is LDT?
  • Part 2: Prediction of one-Point PDF
  • Part 3: Derivation of wavelet $\ell_1$-norm

Part 1: Intuition to LDT

  • Central Limit Theorem gives accurate information around the mean.

  • But what about the tail?
  • $\hookrightarrow$Large Deviations Theory concerns the asymptotic behavior of remote tails of sequences of probability distributions.

  • If we know the behaviour of the tail $\Longrightarrow$ we can get the entire probability distribution.

LDT in Cosmology and prediction of the one-point PDF

Rate Function


Cumulant Generating Function


PDF
Rate Function
From the first principles of Cosmology $\downarrow$ Spherical Collapse Model (one-to-one mapping between initial and final densities) $\zeta(\bar{\tau_k}) = \rho_k = \Bigl( 1 - \frac{\bar{\tau_k}}{\nu}\Bigr)^{-\nu}$ $\psi (\{ \rho_i\}) = \frac{\sigma^2_{R1}}{2} \sum_{k,j}\Xi_{kj}(\{\tau_i \}) \bar{\tau_k} \bar{\tau_j}$


Cumulant Generating Function


PDF
Rate Function
From the first principles of Cosmology $\downarrow$ Spherical Collapse Model $\zeta(\bar{\tau_k}) = \rho_k = \Bigl( 1 - \frac{\bar{\tau_k}}{\nu}\Bigr)^{-\nu}$ $\psi (\{ \rho_i\}) = \frac{\sigma^2_{R1}}{2} \sum_{k,j}\Xi_{kj}(\{\tau_i \}) \bar{\tau_k} \bar{\tau_j}$


Cumulant Generating Function
Legendre-Fenschel transformation $\varphi_{\{ \rho_i\}}(\{ \lambda_i\}) = \sup_{\{ \rho_i\}} \Bigl[ \sum_{i}\lambda_i\rho_i - \psi_{\{ \rho_i\}} (\{ \rho_i\})\Bigr]$


PDF
Rate Function
From the first principles of Cosmology $\downarrow$ Spherical Collapse Model $\zeta(\bar{\tau_k}) = \rho_k = \Bigl( 1 - \frac{\bar{\tau_k}}{\nu}\Bigr)^{-\nu}$ $\psi (\{ \rho_i\}) = \frac{\sigma^2_{R1}}{2} \sum_{k,j}\Xi_{kj}(\{\tau_i \}) \bar{\tau_k} \bar{\tau_j}$

Cumulant Generating Function
Legendre-Fenschel transformation
$\varphi_{\{ \rho_i\}}(\{ \lambda_i\}) = \sup_{\{ \rho_i\}} \Bigl[ \sum_{i}\lambda_i\rho_i - \psi_{\{ \rho_i\}} (\{ \rho_i\})\Bigr]$

PDF
Inverse laplace transformation
$P(\kappa) = \int_{-i\infty}^{+i\infty} \frac{dy}{2\pi i} exp(-\lambda \kappa + \phi_{\kappa,\theta}(\lambda))$
Filter Choice

In this work: : use a function of concentric disks
$\hookrightarrow$ $M_{ap}(\nu) = \kappa_{<\theta_2}(\nu) - \kappa_{<\theta_1}(\nu)$
Deriving wavelet $\ell_1$-norm from PDF
$l_1^{j,i} = \sum_{u=1}^{coef(Sj,i)} |S_{j,i}[u]|$

$S_{j,i} = {w_{j,k}/B_i < w_{j,k} < B_{i+1}}$ $w_{j,k}$ is a wavelet coefficient

  • information encoded in all pixels
  • automatically includes peaks and voids
  • multi-scale approach
  • avoids the problem of defining peaks and voids

This is equivalent to obtaining the $\ell_1$-norm from the PDF using counts x bin method

Results

Preliminary

Preliminary

Preliminary

Conclusion

  • We need different analytical methods to extract non-Gaussianities
  • $\hookrightarrow$ Using Higher-Order statistics
    $\hookrightarrow$ Wavelet $\ell_1$-norm is shown to be a better estimator in comparision to powerspectrum

  • Current methods use simulations based approach $\rightarrow$ Highly resource intensive
  • $\hookrightarrow$ Need theoretical modelling

  • Use LDT based approach to obtain the PDF for aperture mass maps
  • $\hookrightarrow$ Derieved wavelet $\ell_1$-norm from PDF
  • Future work: Extending to any wavelet filter

Thank you!

Questions?

Extra slides

Wavelets

A wavelet is a waveform of effectively limited duration that has an average value of zero and nonzero norm.

This is an example of a DB5 wavelet
The coefficients from Fourier transform and wavelet trandform respectively
Especially useful in cases where we have discontinuties, trends etc that other techniques often tend to miss

Wavelet transform vs fourier transform

LSS example using discrete wavelet transform