Unique reconstruction for discretized inverse problems: a random sketching approach

Inverse problem theory is often studied in the ideal infinite-dimensional setting. Through the lens of the PDE-constrained optimization, the well-posedness PDE theory suggests unique reconstruction of the parameter function that attain the zero-loss property of the mismatch function, when infinite amount of data is provided. Unfortunately, this is not the case in practice, when we are limited to finite amount of measurements due to experimental or economical reasons. Consequently, one must compromise the inference goal to a discrete approximation of the unknown smooth function. What is the reconstruction power of a fixed number of data observations? How many parameters can one reconstruct? Here we describe a probabilistic approach, and spell out the interplay of the observation size (r) and the number of parameters to be uniquely identified (m). The technical pillar is the random sketching strategy, in which the matrix concentration inequality and sampling theory are largely employed. By analyzing randomly sub-sampled Hessian matrix, we attain well-conditioned reconstruction problem with high probability. Our main theory is finally validated in numerical experiments. We set tests on both synthetic and the data from an elliptic inverse problem. The empirical performance shows that given suitable sampling quality, the well-conditioning of the sketched Hessian is certified with high probability.