Optimal non-linear transformations for large scale structure statistics

TL;DR

This paper analytically investigates non-linear transformations, like the power and logarithmic transforms, to efficiently recover Fisher information in large-scale structure statistics, connecting them to optimal observables in cosmology.

Contribution

It derives the optimal non-linear transformations for extracting Fisher information in large-scale structure, linking them to simple power transforms based on the power spectrum slope.

Findings

Power transform with specific exponent recovers linear density contrast.

Transform Gaussianizes the distribution and captures maximum Fisher information.

Transforms remain near optimal even in deeply non-linear regimes.

Abstract

Recently, several studies proposed non-linear transformations, such as a logarithmic or Gaussianization transformation, as efficient tools to recapture information about the (Gaussian) initial conditions. During non-linear evolution, part of the cosmologically relevant information leaks out from the second moment of the distribution. This information is accessible only through complex higher order moments or, in the worst case, becomes inaccessible to the hierarchy. The focus of this work is to investigate these transformations in the framework of Fisher information using cosmological perturbation theory of the matter field with Gaussian initial conditions. We show that at each order in perturbation theory, there is a polynomial of corresponding order exhausting the information on a given parameter. This polynomial can be interpreted as the Taylor expansion of the maximally efficient "sufficient" observable in the non-linear regime. We determine explicitly this maximally efficient observable for local transformations. Remarkably, this optimal transform is essentially the simple power transform with an exponent related to the slope of the power spectrum; when this is -1, it is indistinguishable from the logarithmic transform. This transform Gaussianizes the distribution, and recovers the linear density contrast. Thus a direct connection is revealed between undoing of the non-linear dynamics and the efficient capture of Fisher information. Our analytical results were compared with measurements from the Millennium Simulation density field. We found that our transforms remain very close to optimal even in the deeply non-linear regime with \sigma^2 \sim 10.