Regularization choices that work best
In tomography, the most consistently successful regularizers for recovering sparse signals are L1-norm penalties and Total Variation. Theoretical foundations for L1 recovery come from compressed sensing, where Emmanuel J. Candès at Stanford University and David L. Donoho at Stanford University proved that under appropriate incoherence or restricted isometry conditions sparse vectors can be exactly recovered from far fewer measurements than classical Nyquist limits require. In practical tomographic imaging, L1 on a sparse transform domain such as wavelets or discrete cosine bases enforces coefficient-level sparsity and is robust when the true image admits a sparse representation.
Structure-aware priors: Total Variation and beyond
When the target is piecewise smooth or contains edges, Total Variation regularization often outperforms plain L1 because it promotes sparsity of image gradients rather than of transform coefficients. This reduces staircasing of homogeneous regions while preserving boundaries, which is especially relevant for medical and geophysical tomography where anatomical or geological interfaces are critical. Michael Lustig at Stanford University demonstrated the practical impact of combining L1 and Total Variation ideas in accelerated MRI, showing substantial gains in realistic imaging tasks.
Non-convex penalties such as Lp norms with p less than one and iterative reweighted L1 methods can yield sparser and more accurate reconstructions in low-noise, highly underdetermined regimes. These approaches are more sensitive to initialization and noise and can require careful tuning of optimization algorithms and stopping criteria. In many tomographic setups, hybrid strategies that combine synthesis sparsity, analysis priors, and TV produce the best trade-off between artifact suppression and detail preservation.
Causes, consequences, and contextual nuance
The principal cause of success for these regularizers is their alignment with the true signal model. If a scene is truly sparse in a chosen basis, L1 recovers coefficients stably; if the scene is piecewise constant, TV preserves edges. Consequences of mismatched priors include oversmoothing of fine detail or artifacts such as ringing and staircase effects. Cultural and environmental contexts influence model selection: in clinical imaging preserving diagnostically relevant edges is prioritized, while in materials tomography capturing fine texture may favor learned or transform-domain sparsity.
Advances in learned priors and plug-and-play algorithms are promising but should be evaluated against established theory and real data. For many tomography problems, starting with L1 and TV regularization remains the most reliable path to recover sparse structures while balancing computational tractability and interpretability.