Answers are grounded in this paper and cite their sources.
High-dimensional image inverse problems are inherently ill-posed and computationally challenging. Within a Bayesian framework employing a hierarchical model with nonsmooth heavy-tailed priors, we propose a subspace splitting Gibbs sampling (SSGS) algorithm to efficiently simulate the resulting complex posterior distributions. Specifically, a split-augmented strategy is established to decouple the restoration and denoising steps into distinct computational domains. While the restoration step can be efficiently handled via the fast Fourier transform, the denoising step constitutes the dominant computational bottleneck, as MCMC sampling of high-dimensional hyperparameters under a nonsmooth heavy-tailed prior is often intractable. To overcome this limitation, we propose a subspace strategy that substantially reduces the number of hyperparameters updated at each iteration, leading to significant computational savings. Moreover, we establish a rigorous theoretical guarantee showing that, under the proposed subspace strategy, the Markov chain induced by the nonsmooth heavy-tailed prior remains geometrically ergodic, thereby ensuring convergence of the algorithm. The effectiveness and efficiency of the proposed method are demonstrated through extensive image inverse numerical experiments.Journal vv (yyyy) aaaaaaAuthor et al infer it within the Gibbs sampling scheme. This allows the effective noise level to adapt dynamically during the reconstruction process, yielding a more principled and robust inference procedure without requiring an accurate initial estimate of σ 2 .2. Motivated by the observation that the posterior distribution of transform coefficients β and associated hyperparameters in the hyperprior of the heavy-tailed prior concentrates in a low-dimensional subspace, we propose a novel subspace strategy that restricts hyperparameter updates to dominant components only. This substantially reduces the computational burden associated with high-dimensional hyperparameter sampling while preserving the statistical fidelity of the model.3. We prove that the proposed SSGS algorithm is geometrically ergodic under the condition that the employed sparsity-inducing prior admits a Gaussian scale mixture representation and induces a geometrically ergodic Markov chain. This condition is satisfied by a broad class of commonly used nonsmooth heavy-tailed priors. Additionally, although not central to the main algorithmic development, we provide a general proof of geometric ergodicity for Gibbs samplers associated with horseshoe priors.4. We demonstrate the effectiveness of the proposed SSGS algorithm on image deblurring problems. Numerical results indicate that the method achieves superior reconstruction quality compared to existing approaches, while significantly reducing per-iteration computational cost due to the proposed subspace hyperparameter sampling strategy.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: