Search papers, labs, and topics across Lattice.
This paper introduces a theoretical framework for generative compressed sensing in infinite-dimensional Hilbert spaces, addressing the limitations of existing theory confined to finite-dimensional spaces. They extend the concept of local coherence to infinite dimensions and derive resolution-independent sampling distributions. The authors prove stable recovery with a number of measurements proportional to the prior's intrinsic dimension, independent of the ambient dimension, and validate their theory with Darcy flow equation experiments.
Generative models can now be rigorously applied to compressed sensing in infinite-dimensional spaces, enabling stable signal recovery with resolution-independent sampling.
Deep generative models have become a standard for modeling priors for inverse problems, going beyond classical sparsity-based methods. However, existing theoretical guarantees are mostly confined to finite-dimensional vector spaces, creating a gap when the physical signals are modeled as functions in Hilbert spaces. This work presents a rigorous framework for generative compressed sensing in Hilbert spaces. We extend the notion of local coherence in an infinite-dimensional setting, to derive optimal, resolution-independent sampling distributions. Thanks to a generalization of the Restricted Isometry Property, we show that stable recovery holds when the number of measurements is proportional to the prior's intrinsic dimension (up to logarithmic factors), independent of the ambient dimension. Finally, numerical experiments on the Darcy flow equation validate our theoretical findings and demonstrate that in severely undersampled regimes, employing lower-resolution generators acts as an implicit regularizer, improving reconstruction stability.
