Primordial non-Gaussianity encodes the particle content of inflation — masses, spins, interaction strengths — in subtle statistical correlations of the cosmic density field. Measuring it would distinguish between inflationary models that produce identical power spectra but different higher-point statistics.
Note (2026-11-13; read against the paper's full text): the essay leaves out the paper's qualification of the local result. The paper says that for local non-Gaussianity multi-tracer scale-dependent bias can exceed the sensitivity of conservative higher-point analyses, but that, as expected, the multi-tracer analysis falls short of the optimal field-level constraint when all modes at the scale of the dark matter halos are included. It also says the path to surpassing the CMB sensitivity is more plausible with scale-dependent bias but that the precise sensitivity is uncertain, and that published equilateral forecasts for a single survey vary by up to an order of magnitude, with equilateral highly sensitive to priors and modelling assumptions and able to benefit from theoretical input such as the redshift evolution of the bias. The essay's closing recommendation to exploit the redistribution of information is its own.
The authors (arXiv:2603.22415) use field-level Cramer-Rao bounds to establish the ultimate information content of galaxy surveys for two types of non-Gaussianity. For local non-Gaussianity (f_NL^local), multi-tracer analysis of scale-dependent bias can exceed the sensitivity of conservative higher-point function analyses. The information about f_NL migrates from the non-Gaussian initial conditions into the amplitude of Gaussian fluctuations via biased tracers.
For equilateral non-Gaussianity, the constraints are far more sensitive to modeling assumptions and theoretical priors. Small changes in how nonlinear structure formation is modeled can dramatically shift the forecasted sensitivity.
The through-claim: nonlinear evolution doesn't just destroy primordial information — it redistributes it. For local non-Gaussianity, structure formation moves the signal from higher-point functions (hard to measure) into the power spectrum of biased tracers (easy to measure). The optimal strategy isn't to fight structure formation but to exploit the redistribution.