Optimal Neural Summarisation for Full-Field Weak Lensing Cosmological Implicit Inference

Traditionally, weak lensing cosmological surveys have been analyzed using summary statistics motivated by their analytically tractable likelihoods, or by their ability to access higher-order information, at the cost of requiring Simulation-Based Inference (SBI) approaches. While informative, these statistics are neither designed nor guaranteed to be statistically sufficient. With the rise of deep learning, it becomes possible to create summary statistics optimized to extract the full data information. We compare different neural summarization strategies proposed in the weak lensing literature, to assess which loss functions lead to theoretically optimal summary statistics to perform full-field inference. In doing so, we aim to provide guidelines and insights to the community to help guide future neural-based inference analyses. We design an experimental setup to isolate the impact of the loss function used to train neural networks. We have developed the sbi_lens JAX package, which implements an automatically differentiable lognormal wCDM LSST-Y10 weak lensing simulator. The explicit full-field posterior obtained using the Hamiltonian Monte Carlo sampler gives us a ground truth to which to compare different compression strategies. We provide theoretical insight into the loss functions used in the literature and show that some do not necessarily lead to sufficient statistics (e.g. Mean Square Error (MSE)), while those motivated by information theory (e.g. Variational Mutual Information Maximization (VMIM)) can. Our numerical experiments confirm these insights and show, in our simulated wCDM scenario, that the Figure of Merit (FoM) of an analysis using neural summaries optimized under VMIM achieves 100% of the reference Omega_c - sigma_8 full-field FoM, while an analysis using neural summaries trained under MSE achieves only 81% of the same reference FoM.