Answers are grounded in this paper and cite their sources.
Identifying unknown source terms of anomalous transport processes described by a one-dimensional time fractional advection diffusion equation from sparse concentration measurements is a typical ill-posed inverse problem. However, it is hard to obtain a credible likelihood of the measurement error model. This paper therefore casts the source identification problem as a likelihood-free Bayesian inference problem and solves it with approximate Bayesian computation and nested sampling (ABC-NS) algorithm. To improve the ABC-NS algorithm, a sliced-Wasserstein discrepancy is introduced as an alternative to the relative root-mean-square error, and the ellipsoidal proposal is replaced by a bounding multi-ellipsoid rule with a fixed enlargement. The main algorithmic additions are the SWD, the two-cluster multi-ellipsoid rule and joint inference of the source strengths and α; the L1/upwind forward solver is classical. The forward problem is discretised with the L1 scheme in time and an implicit finite difference scheme in space, and the unconditional stability and convergence of the finite difference scheme are proved. Several numerical experiments are reported, and the results show that the ABC-NS algorithm recovers the source strengths with quantified uncertainty, that the acceptance rate decay of rejection-based ABC is drastically alleviated, that the ellipsoidal sampler reaches the noise-matched tolerance with an order of magnitude fewer forward solves than ABC-SMC needs merely to approach it, and that the fractional order can be inferred jointly with the source strengths from the same data, which is not possible within the deterministic regularization framework. The sliced-Wasserstein discrepancy is slightly less sharp than the residual norm when used alone, but it yields higher empirical coverage, although still below nominal, once it is combined with the multi-ellipsoid proposal, which raises the coverage of the nominal 95% intervals from 0.62 to 0.75 at a cost of about one fifth more forward solves than the single-ellipsoid version.
No takes yet. Share an insight, caveat, or question.
Yi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: