A cosmic lightcone treated as video: ViViViT reconstructs the matter field with split space-time attention, holding bispectrum error near 5%
Synopsis
Moving along a cosmological lightcone's line of sight also moves through cosmic time, so the lightcone can be read as a video of structure forming. ViViViT recasts reconstruction of the matter density field from 21 cm brightness-temperature lightcones as video-to-video translation, using attention factorized into transverse frames and radial evolution. On held-out 21cmFAST lightcones it keeps spectral power and phase alignment within a few percent of the ground truth and outperforms the tested 3D U-Net baseline across all measured statistics.
Interpretation
Reconstruction of the matter density lightcone from a 21 cm brightness-temperature lightcone is handled as video-to-video translation, converting the whole lightcone in a single forward pass. Earlier 21 cm reconstruction approaches largely inherited the inductive biases of convolutional networks, with time dependence supplied only as conditioning or used for parameter inference. Trained on 3000 21cmFAST lightcones and evaluated on 200 held-out lightcones, ViViViT outperforms the tested four-level 3D U-Net across all measured two- and three-point statistics.
The model splits attention into transverse within-frame and radial-evolution components, so it keeps a global radial view across the whole lightcone. Full attention over all space-time tokens would grow quadratically with token count, making larger volumes impractical. On a synthetic video benchmark without periodicity, U-Nets at every tested depth lose accuracy as the target correlation length grows, while ViViViT stays at low error for every tested correlation length.
The reconstructed density fields stay close to the ground truth in spectral power, phase alignment and three-point correlations. The tested U-Net loses small-scale power earliest in the lowest-redshift bin and struggles to recover surrounding density once local brightness temperature is suppressed inside ionized regions. ViViViT keeps the median power ratio near unity over most measured modes, holds a higher cross-correlation coefficient than the tested U-Net in the central redshift bin, and reconstructs the bispectrum to within about 5% for most retained triangle configurations.
Perspective
The demonstrated scope is limited to simulated lightcones generated with 21cmFAST, spanning morphologically rich structure across absorption, heating and patchy reionization, with 3000 training lightcones and 200 held-out lightcones compared against a four-level 3D U-Net; the synthetic benchmark separately tests long-range radial information.
The 21cmFAST forward model imprints the same comoving structure periodically at different times, unlike a real observed lightcone, and this may supply extra reconstruction information; the synthetic benchmark establishes long-context capability but does not isolate its contribution to the cosmological task. The reconstruction is deterministic, producing the whole volume in one forward pass without preserving posterior uncertainty; real-data use would also need observational and astrophysical systematics and projection of the physical field to an instrumental measurement, where galactic and extragalactic foregrounds confine reliable spectral information outside a wedge-shaped region in Fourier space.