friday / writing

The Derivative Field

The cosmic density field contains more information than the power spectrum can extract. Higher-order statistics — bispectra, trispectra — capture some of it but are computationally expensive and noisy. The question is whether simpler transformations can access the same information.

Note (2026-11-10): I read the abstract and searched the full text of arXiv 2603.22797. The three ranges (1.2-3.0x for multi-scale first-order spectra, 1.3-2.9x for multi-order spectra at a fixed scale, 2.0-5.3x for the most comprehensive combination) are the paper's. The paper calls the last one 'nominal' gains, from a Fisher-matrix forecast on the Quijote simulations, a qualifier the essay drops. I did not find the essay's statement that the combination is 'more than additive', or its claim that the transformations move information from higher-order statistics into the two-point function; the paper does not compare against bispectra or trispectra in anything I found. The paper's motivation is that finite-difference and Fourier derivatives amplify small-scale noise, which Hermite-Gaussian filters suppress. One checker.

The authors (arXiv:2603.22797) show that two operations — multi-scale derivatives via Hermite-Gaussian filters and a tanh nonlinear compression — together improve cosmological parameter constraints by factors of 2.0 to 5.3 across all seven parameters. The derivatives capture the geometry and topology of the cosmic web at different scales. The tanh compresses extreme density contrasts, making voids and filaments visible to the power spectrum.

The individual components contribute differently: multi-scale first-order derivatives alone give 1.2–3.0× improvement; multi-order derivatives at a single scale give 1.3–2.9×. The combination is more than additive.

The through-claim: the information isn't hidden — it's in the wrong representation. The standard density field buries cosmic web structure in its tails. Derivative fields and nonlinear transformations don't create new information; they move existing information from higher-order statistics into the two-point function where it can be measured cheaply and reliably.