Abstract:We consider the problem of recovering elements of a low-dimensional model from linear measurements. From signal and image processing to inverse problems in data science, this question has been at the center of many applications. Lately, with the success of models and methods relying on deep neural networks leading to non-convex formulations, traditional convex variational approaches have shown their limits. Furthermore, the multiplication of algorithms and recovery results makes identifying the best methods a complex task. In this article, we study recovery with a class of widely used algorithms without considering any underlying functional. This result leads to a class of projected gradient descent algorithms that recover a given low-dimensional with linear rates. The obtained rates decouple the impact of the quality of the measurements with respect to the model from its intrinsic complexity. As a consequence, we can directly measure the performance of this class of projected gradient descents through a restricted Lipschitz constant of the projection. By optimizing this constant, we define optimal algorithms. Our general approach provides an optimality result in the case of sparse recovery. Moreover, we uncover underlying linear rates of convergence for some ''plug and play'' imaging methods relying on deep priors by interpreting our results in this context, thus linking low-dimensional recovery and recovery with deep priors under a unified theory, validated by experiments on synthetic and real data.

Instances of Computational Optimal Recovery: Dealing with Observation Errors

Near-Optimal Estimation of Linear Functionals with Log-Concave Observation Errors

Full Recovery from Point Values: an Optimal Algorithm for Chebyshev Approximability Prior

Optimal Recovery of a Family of Operators from Inaccurate Measurements on a Compact Set

On Robust Recovery of Signals from Indirect Observations

Approximability models and optimal system identification

Near-Optimal Recovery of Linear and N-Convex Functions on Unions of Convex Sets

Adaptive reconstruction of imperfectly-observed monotone functions, with applications to uncertainty quantification

S-Procedure Relaxation: a Case of Exactness Involving Chebyshev Centers

On the Optimal Recovery of Graph Signals

Towards optimal algorithms for the recovery of low-dimensional models with linear rates

Robust Sparse Recovery Via Weakly Convex Optimization in Impulsive Noise

Radius of Information for Two Intersected Centered Hyperellipsoids and Implications in Optimal Recovery from Inaccurate Data

Optimal recursive estimation with uncertain observation

Partial Extended Observability Certification and Optimal Design of Moving Horizon Estimators Under Uncertainties

Stability and Recovery for Independence Systems

On the Optimal Recovery of One Family of Operators on a Class of Functions from Approximate Information about Its Spectrum

Optimal Recovery for Causal Inference

Cautious optimization via data informativity

Optimal Learning via Moderate Deviations Theory

Greedy Pursuits: Stability of Recovery Performance Against General Perturbations.