[Paper Review] Variable Total Variation Regularization for Backward Time-Space Fractional Diffusion Problem

Junxiong Jia, Jigen Peng+2 more|arXiv (Cornell University)|Jan 16, 2016

Fractional Differential Equations Solutions20 references3 citations

TL;DR

This paper proposes a variable total variation (TV) regularization method for solving the backward time-space fractional diffusion problem, using Fourier-domain minimization to reconstruct initial data from noisy final-time measurements. The approach combines fractional semigroup theory for well-posedness and a modified Bregman iterative algorithm that adaptively updates the regularization term, achieving edge preservation and reduced staircasing with proven convergence and parameter selection strategy, outperforming standard Tikhonov and TV models in numerical tests.

ABSTRACT

In this paper, we consider a backward problem for a time-space fractional diffusion process. For this problem, we propose to construct the initial data by minimizing data residual error in fourier space domain and variable total variation (TV) regularizing term which can protect the edges as TV regularizing term and reduce staircasing effect. The well-posedness of this optimization problem is studied under a very general setting. Actually, we write the time-space fractional diffusion equation as an abstract fractional differential equation and get our results by using fractional semigroup theory, so our results can be applied to other backward problems for more general fractional differential equations. Then a modified Bregman iterative algorithm is proposed to approximate the minimizer. The new features of this algorithm is that the regularizing term changed in each step and we need not to solve the complexed Euler-Lagrange equations of variable TV regularizing term (just need to solve a simpler Euler-Lagrange equations). The convergence of this algorithm and the strategy of choosing parameters are also obtained. Numerical implementations are given to support our analysis to show the flexibility of our minimization model.

Motivation & Objective

• To address the ill-posed backward problem for time-space fractional diffusion equations with noisy final-time data.
• To develop a flexible regularization model that preserves edges and reduces the staircasing effect common in total variation regularization.
• To establish well-posedness of the optimization problem using fractional semigroup theory for general fractional operators.
• To design an efficient iterative algorithm that avoids solving complex Euler-Lagrange equations by updating the regularization term at each step.
• To validate the method numerically and demonstrate its superiority over Tikhonov and standard TV regularization in recovering discontinuous initial data.

Proposed method

• Formulates the backward problem as a minimization of data residual error in the Fourier domain with a variable total variation (TV) regularization term.
• Rewrites the time-space fractional diffusion equation as an abstract fractional differential equation using fractional semigroup theory.
• Applies the Bregman iterative method with adaptive regularization parameter updates to avoid solving the complex Euler-Lagrange equation of variable TV at each step.
• Uses a stopping criterion based on residual decay to control iteration count and prevent overfitting.
• Employs a modified Bregman algorithm where the regularization term is updated iteratively, simplifying subproblem solutions to simple Euler-Lagrange equations.
• Validates the method numerically on synthetic examples with varying noise levels and fractional parameters.

Experimental results

Research questions

• RQ1 How can variable total variation regularization be effectively applied to the backward time-space fractional diffusion problem to improve edge recovery and reduce staircasing?
• RQ2 What is the theoretical well-posedness of the proposed minimization model under general fractional operator settings?
• RQ3 How does the modified Bregman iterative algorithm ensure convergence while avoiding the need to solve complex Euler-Lagrange equations?
• RQ4 What is the impact of the fractional order α on the degree of ill-posedness, and does it exhibit discontinuous behavior at α=1?
• RQ5 How does the performance of the variable TV model compare quantitatively with Tikhonov and standard TV regularization in recovering discontinuous initial data?

Key findings

• The proposed variable TV model achieves a relative error of 13.0666% for Example 2 with σ=0.0005, outperforming Tikhonov (13.7772%) and standard TV (13.0053%) models.
• For σ=0.005, the variable TV model achieves 22.7810% relative error, slightly better than Tikhonov (25.2101%) and standard TV (22.7222%) models.
• Numerical verification confirms theoretical predictions: when λ=10¹¹, M=10; for λ=1/4×10¹¹ and λ=1/16×10¹¹, M=38 and 167 respectively, closely matching theoretical expectations.
• The degree of ill-posedness decreases continuously as α approaches 1 from below, but jumps significantly at α=1, indicating a discontinuity in ill-posedness behavior.
• The method behaves like Tikhonov regularization for smooth initial data and like TV regularization for piecewise constant data, demonstrating adaptive flexibility.
• Figure 6 shows that even α=0.99 yields significantly weaker ill-posedness than α=1, suggesting that fractional models with α<1 are inherently less ill-posed than integer-order counterparts.