Cosmologists usually summarize a map of the universe by its power spectrum, which measures how much structure exists at each spatial scale. It throws away the Fourier phase, the part of the transform that records where structures sit. Speech engineers faced the same loss decades ago and found a stable way to use phase: take its derivative with respect to frequency. This quantity, the group delay, turns wrapped, noisy-looking phase into a smooth signal that marks where energy is located. We propose porting group delay to 2D maps from the CAMELS simulations, a suite of thousands of simulated universes that vary two cosmological parameters (matter density $\Omega_m$ and clustering amplitude $\sigma_8$) together with four "feedback" parameters. The feedback parameters control how exploding stars and growing black holes blow gas around.
The physical bet is simple. Feedback mostly moves gas relative to dark matter, and cosmology mostly changes how much structure there is. A displacement shows up as a slope in Fourier phase, which is exactly what group delay measures. We therefore expect a translation-invariant group-delay summary, especially the scale-dependent offset between the group delay of the gas map and that of the dark-matter map, to break the degeneracy between cosmology and feedback that power spectra leave open. Because it is a fixed, low-dimensional statistic rather than a learned filter bank, it may also hold up better when a model trained on one simulation code is tested on another.
Core Claim
A windowed, modified group delay computed on CAMELS 2D fields yields a compact summary with two properties. First, together with auto- and cross-power spectra, it recovers a substantial share of the $\Omega_m$ and $\sigma_8$ information that end-to-end convolutional networks (CNNs, image networks that learn their own filters) extract. Second, it shows a smaller cross-code bias when trained on IllustrisTNG and tested on SIMBA. The new ingredient is the cross-field group-delay offset. This is the gradient, across spatial frequency, of the phase difference between gas and dark matter. The mean cross-correlation coefficient $r(k)$ captures only the average cosine of that phase difference, not how the difference changes with scale. If the offset adds nothing beyond cross-spectra, the proposal fails, and we have designed the experiments so that this outcome is visible.
Why this intersection is open
The CAMELS multifield papers (CMD, 2022; CMD; public data release) train real-space CNNs that reach "a few percent" error on $\Omega_m$ and $\sigma_8$ while marginalizing over feedback. They also document that these networks degrade when moved across simulation codes. Those papers motivate this work: the information is there, but nobody knows which part of it is robust.
DeepLSS breaks $\Omega_m$–$\sigma_8$ and intrinsic-alignment degeneracies with a CNN on gravity-only maps and attributes the gain partly to phase information. It never measures phase. We test that attribution directly, and we do so with baryons included.
The beyond-two-point line of work constrains our design. The SimBIG bispectrum, skew spectra, wavelet scattering and learned equivariant marks all break $\Omega_m$–$\sigma_8$ on gravity-only mocks. The bispectrum sums phases over closed triangles. Scattering applies a modulus at each layer and so discards phase. The wavelet-scattering paper also shows that sensitive small-scale statistics become biased under forward-model mismatch and require scale cuts. That last negative result means we must test robustness, not merely information content.
Two papers supply the method. Modified group delay defines the stabilized phase derivative. Joint MODGDF+MFCC shows that group delay is most useful when combined with magnitude features rather than used alone: MODGDF alone scored below MFCC, but the joint features were far better. We inherit that lesson and never evaluate group delay in isolation as the headline. Among the galaxy-level studies, Contardo et al. identify the supernova-feedback parameter $A_{\rm SN1}$ as nearly degenerate with $\Omega_m$ at the galaxy level. That gives us a concrete degeneracy to target in field space.
The statistic
Take a tapered window $w$ over a patch of a field $\delta$ (gas, dark matter, total mass, temperature). Let $X(\mathbf{k})$ be the Fourier transform of $w\,\delta$ and $Y_j(\mathbf{k})$ the transform of $x_j\,w\,\delta$, where $x_j$ is the $j$-th position coordinate. This is the 2D analogue of the "time-weighted signal" used in speech. The exact spatial group delay is then
$$\tau_j(\mathbf{k}) = -\frac{\partial \phi(\mathbf{k})}{\partial k_j} = \frac{\mathrm{Re}\,[X^*(\mathbf{k})\,Y_j(\mathbf{k})]}{|X(\mathbf{k})|^2}.$$
For a single compact object, $\boldsymbol{\tau}$ is simply its position. The formula needs no phase unwrapping.
Random fields have Fourier amplitudes that often come close to zero, so the raw ratio produces spikes. This is the spectral-null problem that MODGDF was built to solve. Following Murthy et al., we replace $|X|^2$ with $S(|\mathbf{k}|)^{2\gamma}$, where $S$ is the ring-averaged amplitude within the window. We then compress the result as $\mathrm{sgn}(\tau)|\tau|^\alpha$. In cosmology the smoothed envelope $S$ has a natural meaning: it is the local power spectrum.
Absolute group delay moves with the patch, so it cannot serve as a summary directly. Two translation-invariant summaries follow. The first is the within-field dispersion $D_a(k)$, the spread of $\boldsymbol\tau^a$ around its window mean, binned in $|\mathbf{k}|$. It measures whether structure at different scales shares a common center, for example how halos are nested. The second is the cross-field offset
$$\Delta_{ab}(k) = \big\langle\, |\boldsymbol\tau^a(\mathbf{k}) - \boldsymbol\tau^b(\mathbf{k})|^2 \,\big\rangle_{|\mathbf{k}|\in k,\ \text{windows}},$$
which measures how far field $a$ (for example gas) is displaced from field $b$ (for example dark matter) at scale $k$.
Research Program
We build these statistics for the CAMELS Multifield Dataset 2D maps ($256\times256$ pixels at $z=0$, $(25\,h^{-1}\mathrm{Mpc})^2$), using the Latin-hypercube sets of IllustrisTNG, SIMBA and, where available, ASTRID. We use the one-parameter-at-a-time ("1P") sets for diagnostics. The window size and $\gamma,\alpha$ are the main knobs. Window size trades scale resolution against the number of independent patches. We tune them on TNG validation data only, because the cross-code test is meaningless if it is tuned on SIMBA.
Every summary, ours and the baselines', feeds the same inference head. That head is a small multilayer network trained for neural posterior estimation, with an identical architecture search budget for each summary. The 1P sets give finite-difference derivatives of each summary with respect to all six parameters, and from them Fisher forecasts. These tell us, before any learning, whether $\Delta_{ab}$ responds to $A_{\rm AGN}$ and $A_{\rm SN}$ in directions that differ from its response to $\Omega_m$ and $\sigma_8$.
Baselines
The naive reference is the auto power spectrum of the total-mass map. This is the statistic whose $\Omega_m$–$\sigma_8$–feedback degeneracy motivates the work. It is run through the same posterior network on the same TNG training split.
The fair "no phase-slope" control is the full set of auto- and cross-power spectra among gas, dark matter, stars and temperature, including the cross-correlation coefficient $r(k)$. This is the most important comparison. $r(k)$ already registers gas–dark-matter phase misalignment on average, so our statistic must beat this set, not merely beat $P(k)$. Beating it means a statistically significant tightening of the marginal posterior widths for $\Omega_m$ and $\sigma_8$, averaged over held-out TNG simulations, plus a reduction in the posterior correlation between $\Omega_m$ and $A_{\rm SN1}$.
The strongest summary-statistic competitor is the power spectra plus the wavelet scattering transform, following the construction in SimBIG WST. The strongest overall competitor is the end-to-end CNN of the CAMELS multifield paper, which reports "a few percent" errors on $\Omega_m$ and $\sigma_8$. The file gives no numeric RMSE for it, so no reported result was found on Lacuna for a CAMELS CNN numeric error. We will retrain it on our exact splits and with the same field channels, giving it a hyperparameter budget at least as large as the summary pipelines receive.
We do not expect to beat the CNN in-distribution. Success against it means closing most of its information gap, which we will define before running as reaching at least 70% of its in-distribution Fisher figure of merit, while showing a smaller TNG→SIMBA bias. We also run a dual-stream CNN that receives the group-delay maps as extra channels, following the magnitude-plus-phase designs of Janbakhshi & Kodrasi and Loweimi et al.. This tests whether explicit phase derivatives help even a flexible model.
The ablations isolate the key ingredient. We replace the modified group delay with the raw phase, with the one-bit sign spectrum, and with the unmodified group delay using the $|X|^2$ denominator. We also drop the cross-field offset and keep only the within-field dispersion.
Why not just…?
Why not just train the CNN on pixels, since it will learn any useful phase structure? It probably does learn it in-distribution. The CAMELS papers show that the same networks lose accuracy across codes, which suggests they also learn code-specific texture. Our claim is not that the CNN lacks the information. It is that a fixed phase-slope statistic keeps the robust part. The TNG→SIMBA test settles this. If the CNN's cross-code bias is no worse than ours, the robustness argument fails and only interpretability remains.
Why not just train on several codes at once? Multi-code training is a real threat to the robustness claim, and we include it. We train every method on TNG plus ASTRID and test on SIMBA. If multi-code training erases the CNN's bias, our contribution shrinks to explaining which physical signal the networks rely on. That explanation still matters, but it is a smaller result, and we will report it as such.
Why not just use cross-power spectra or the bispectrum, which already contain phase information? The cross-spectra baseline is exactly this test for $r(k)$. The bispectrum sums phases over closed triangles, as Hahn et al. use it, and measures coupling between scales rather than how location changes with scale. On gravity-only mocks it gains 2.4× on $\sigma_8$ over the power spectrum. We add a binned bispectrum of the total-mass map to the power-spectrum set. If group delay adds nothing beyond power spectra plus bispectrum, we will report that.
Why not just feed raw phase, as speech work sometimes does? Raw phase was the weakest standalone representation in the dysarthria study (62.76% accuracy), and the ablation checks whether that ordering carries over to cosmic fields.
Decisive Checks
The first check comes before any learning. We take maps whose Fourier phases are randomized within each window while their amplitudes are kept. On these maps $D_a$ and $\Delta_{ab}$ must lose their parameter sensitivity while the power spectra do not. This confirms the statistic reads phase and not amplitude leakage.
The second check uses the 1P Fisher derivatives. $\Delta_{\rm gas,DM}$ should respond strongly to the AGN and supernova parameters and weakly to $\sigma_8$. If it responds the same way to both, the physical premise is wrong and we stop there.
The main test compares, on held-out TNG simulations, the relative RMSE, the posterior width, the $\Omega_m$–$A_{\rm SN1}$ posterior correlation and calibration coverage for every method. The cross-code test is then scored mainly by the normalized bias on SIMBA, meaning the mean error divided by the posterior width. The proposal succeeds if power spectra plus group delay beats the cross-spectra baseline in-distribution and shows lower SIMBA bias than both the scattering pipeline and the CNN under matched training data. It fails cleanly if the cross-spectra baseline matches it.
References
- The CAMELS Multifield Data Set: Learning the Universe’s Fundamental Parameters with Artificial Intelligence
- The CAMELS project: public data release
- DeepLSS: Breaking Parameter Degeneracies in Large-Scale Structure with Deep-Learning Analysis of Combined Probes
- ${\rm S{\scriptsize IM}BIG}$: The First Cosmological Constraints from the Non-Linear Galaxy Bispectrum
- Cosmological constraints from the redshift-space galaxy skew spectra
- Galaxy clustering analysis with SimBIG and the wavelet scattering transform
- Interpretable Equivariant Marks for Contrastive Cosmological Inference
- The modified group delay feature: a new spectral representation of speech
- Continuous speech recognition using joint features derived from the modified group delay function and MFCC
- On the Effects of Parameters on Galaxy Properties in CAMELS and the Predictability of Ωm
- Experimental Investigation on STFT Phase Representations for Deep Learning-Based Dysarthric Speech Detection
- Raw Sign and Magnitude Spectra for Multi-Head Acoustic Modelling