The classical Shannon-Nyquist sampling theorem dictates that perfect signal reconstruction requires sampling rates at least twice the signal's highest frequency, often imposing significant inefficiencies in high-dimensional data acquisition. Compresse...
The classical Shannon-Nyquist sampling theorem dictates that perfect signal reconstruction requires sampling rates at least twice the signal's highest frequency, often imposing significant inefficiencies in high-dimensional data acquisition. Compressed Sensing (CS) overcomes this by exploiting signal sparsity, typically employing the Least Absolute Shrinkage and Selection Operator (LASSO) for reconstruction. However, standard LASSO (ℓ1-regularization) suffers from inherent limitations, including shrinkage bias that underestimates large coefficients and inconsistent variable selection in correlated feature spaces. To address these shortcomings, this paper introduces the Variational Garrote (VG) into the CS framework. VG is a probabilistic regression model that effectively approximates ℓ0-regularization by explicitly decoupling variable selection from coefficient estimation. Through extensive experiments on 1D synthetic and real-world audio signals, as well as 2D image reconstruction using Radon transforms, we demonstrate that VG significantly outperforms LASSO. Specifically, VG exhibits superior noise robustness and consistently lower generalization errors in data-scarce regimes. Furthermore, in sparse-view CT reconstruction, VG demonstrates remarkable statistical stability with tighter error margins, highlighting its potential as a reliable and precise alternative for advanced signal recovery applications.