Neyrinck, Szapudi & Szalay (2009)

The big idea¶

As structure grows non-linear, the matter power spectrum $P_\delta(k)$ becomes a bad summary statistic: gravitational evolution couples Fourier modes, the late-time density is dominated by a few sharp peaks (haloes), and the cosmological information that was cleanly held in the linear power spectrum leaks into higher-order statistics and into a large, “translinear” covariance. Neyrinck, Szapudi & Szalay show that a single, simple operation undoes much of this damage: replace $\delta$ with the log-density $\log(1+\delta)$ before measuring the power spectrum.

The log-mapped power spectrum $P_{\log(1+\delta)}(k)$ has a shape “hardly departing from the linear power spectrum for $k \lesssim 1\,h\,\mathrm{Mpc}^{-1}$ at all redshifts,” and — the headline — recovers pristine Fisher information, yielding about 10× more cumulative signal-to-noise at $z=0$ than the standard power spectrum over a range of scales.

Why it works¶

The density field is statistically invariant under translations and rotations; all the cosmological information of the Gaussian initial conditions lives in the power spectrum, with every higher moment zero. Non-linear growth breaks this by making the one-point PDF non-Gaussian ($\delta$ develops a long tail and a hard floor at $\delta=-1$). Because the late-time PDF is approximately lognormal (Coles & Jones (1991)), restoring Gaussianity to the one-point distribution — by taking the log — pulls the strayed information back into the two-point function. Phase correlations (which build genuine cosmic-web filaments) affect higher-order statistics but not the power spectrum, so a monotone one-point remap recovers what it can without needing the phases.

Use in progenax¶

• Why the 2-point carrier is the log-density $\xi_s$, not the linear $\langle\rho\rho\rangle$. This is the cosmology precedent for the choice made throughout Differentiable inference — natal cloud parameters from cluster substructure: the fat power-law tail makes linear-density 2-point statistics divergent / information-poor, while the log-density two-point is well-behaved and information-rich. The gaussianized_xi series predicts exactly this $\xi_s(r)$.

• The “predict the statistic, restore the information” philosophy of the cosmology playbook that the differentiable-inference layer adopts.

Notes¶

• Paper I of a pair; the discrete-field / galaxy extension and the Fisher-vs-resolution behaviour are in Neyrinck, Szapudi & Szalay (2011). The information-theoretic optimality of the transform is formalised by Carron & Szapudi (2013).

• The log transform is exactly reversible and preserves cell-by-cell ranking — the same property the FDF rank copula exploits.