Spherical graph network reads multifrequency B-mode maps directly, cutting tensor-to-scalar error about a quarter below the strongest bandpower likelihood
Synopsis
A spherical graph convolutional network infers the tensor-to-scalar ratio directly from cut-sky multifrequency cosmic microwave background B-mode maps instead of first compressing them into bandpowers. On 20,000 simulations including lensing, inhomogeneous noise and stochastic Galactic dust and synchrotron, it is compared map by map with a multifrequency bandpower likelihood calibrated on the same simulations, reducing the root-mean-square error by a factor of about 1.3 and giving 68% intervals about 1.2 times narrower. The gain appears with foregrounds present; without them the two are comparable, and both estimators are tied to the foreground model they are built on.
Interpretation
Multifrequency B-mode polarization maps can go straight into a spherical graph convolutional network to yield an estimate of the tensor-to-scalar ratio and its uncertainty, with no power-spectrum estimation or component separation. Earlier map-level inference of this quantity used noiseless, foreground-free flat-sky patches, not the curved sky with Galactic foregrounds and inhomogeneous noise. On foreground simulations with stochastic dust and synchrotron, the network reaches a root-mean-square error of 2.08 on the tensor-to-scalar ratio against 2.79 for the strongest bandpower likelihood, with mean 68% half-widths of 2.15 and 2.59.
The network takes the six frequency bands as input channels and performs spectral convolutions on a graph of HEALPix pixels, learning the frequency combination and spatial weighting itself. The bandpower likelihood first compresses the maps into a 189-element data vector of six auto- and 15 cross-frequency spectra in nine multipole bins, then fits a Gaussian posterior. Training uses 20,000 simulations and testing an independent grid of 100 realizations at each of 11 tensor-to-scalar values; saliency maps show a Spearman rank correlation of 0.80 to 0.85 with the hits map and about 77% of the sensitivity share in the 93 GHz channel.
Perspective
The results rest on a simulated sky for a generic ground-based small-aperture telescope with six bands at 27, 39, 93, 145, 225 and 280 GHz, and training and test sets share fixed large-scale foreground templates taken from observations of the real sky; the benchmark omits delensing, instrumental systematics and variations of the observing strategy.
Both the network and the likelihood are tied to the foreground family they are calibrated on: across seven out-of-distribution foreground tests the network has the lower root-mean-square error in six, but on a three-dimensional dust model with line-of-sight variation it overestimates by about 12 while its reported uncertainty rises only from 2.2 to 2.4; near zero tensor-to-scalar ratio the non-negative prior pulls both estimators high, with the network's bias about one-half to two-thirds that of the likelihood.
