An efficient method for reconstructing the potential in the fractional Calder\'on problem with a single measurement

1. Introduction

• The paper addresses the scarcity of numerical methods for the fractional Calderón problem by developing POD-RBF surrogate-based inversion, a new application of POD-RBF to this problem.POD-RBF had been applied to inverse heat conduction problems, but its application to fractional Calderón problems is described as new.Source: §1 · p. 3
• The inverse task is to reconstruct the potential from a single noisy measurement of the fractional Laplacian applied to the solution on an exterior set.The paper reports applying the method in one- and two-dimensional settings.Source: §1 · pp. 3–4
• The offline phase generates forward-solution observation snapshots, extracts POD basis vectors, and trains an RBF network to map parameters to POD amplitude coefficients.The online phase replaces repeated full-order forward solves with the surrogate’s predictions and exact gradients during Tikhonov-regularized conjugate-gradient optimization.Source: §1 · pp. 4–5
• POD-RBF combines data-driven model reduction with a smooth nonlinear surrogate to address the severely ill-posed inverse problem with accuracy and computational efficiency.The paper also describes the method as accelerating inversion compared with conventional numerical approaches such as finite differences.Source: §1 · pp. 4–5

2.1. Overview of the well-posedness

• Under the stated assumption, the weak forward problem has a unique solution, and the section provides a norm estimate.The observation is also shown to belong to L2(W2) under the stated conditions on the exterior sets and support of f.Source: §2 · p. 6

2.2. Algorithms of the forward problem

• In one dimension, the truncated fractional Laplacian is discretized by finite differences, yielding a symmetric positive-definite Toeplitz system with diagonal potential matrix.The resulting linear system can be solved using Krylov subspace methods with the fast Fourier transform.Source: §2 · p. 7

2.3. Two-dimensional finite difference discretization

• In two dimensions, finite differences on a Cartesian grid yield a Toeplitz-block-Toeplitz system with a diagonal potential matrix.The discretization uses quadrature weights for the fractional Laplacian and accounts for known exterior data.Source: §2 · pp. 7–8

2.4. Discrete observation operator

• The discrete observation operator restricts the full discrete fractional Laplacian to measurement points in W and separates contributions from internal and known external values.The exterior field contributes a constant vector to the observations.Source: §2 · p. 8

3.1. Proper orthogonal decomposition basis

• Training snapshots are generated by solving the forward problem at sampled normalized potential parameters and collecting their observation vectors.The snapshots are centered by subtracting their average before POD is applied.Source: §3 · pp. 8–9
• POD retains the leading modes of the centered observation snapshots, choosing the number of modes K using an energy-decay tolerance.The resulting basis provides a low-dimensional representation of the training observations.Source: §3 · p. 9

3.2. RBF network and surrogate model

• Combining the retained POD modes with the learned amplitude map gives a surrogate that predicts observations for a potential parameter vector.The amplitude map is fitted from the known training snapshots.Source: §3 · pp. 9–10
• The RBF network approximates POD amplitude vectors as a function of normalized potential parameters using radial basis functions centered at training points.The paper uses an inverse multiquadric kernel.Source: §3 · p. 10

4. The inversion algorithm

• The inversion parameterizes the potential with a finite-dimensional vector, approximates the observation map with a POD-RBF surrogate, and regularizes recovery of the parameters.The error analysis includes parameterization, surrogate-model, and measurement-noise contributions.Source: §4 · p. 11
• The Tikhonov functional balances consistency with noisy measurements against stabilization of the parameter vector’s discrete derivatives, with regularization scaled to the noise level.The stated bounds are c1δ^2 ≤ α ≤ c2δ^2.Source: §4 · p. 12
• The POD-RBF surrogate preserves the inversion’s logarithmic convergence order when parameterization and reduced-order errors remain below the noise level, while reducing online computational cost.The analysis accounts for parameterization and surrogate errors alongside measurement noise.Source: §4 · pp. 11, 15

4.2. Algorithm.

• Offline, the algorithm solves forward problems for training samples, forms observation snapshots, computes a POD basis, and trains an RBF network; online, it minimizes the objective by conjugate gradients.The explicit analytic surrogate provides an exact gradient for the optimization.Source: §4.2 · pp. 15–17

Example 2

• The Example 2 reconstructions are shown against the true potential for noise levels δ=1E-7, 1E-6, and 1E-5, with α=δ2.The supplied passage gives the setup but does not state a qualitative reconstruction outcome.Source: §5 · p. 18, Figure 5.2 · p. 20

Example 1

• Higher noise rapidly degrades Example 1 reconstruction, while the figure also tracks POD eigenvalue decay, cumulative energy, and CG residuals.The tested noise levels are δ=1E-7, 1E-5, and 1E-3.Source: §5 · p. 18, Figure 5.1 · p. 19

Example 4

• The two-dimensional example presents the true and reconstructed potentials for δ=1E-6.The supplied passage does not state the reconstruction's accuracy.Source: §5 · p. 20

• The method uses observation snapshots to build a reduced-order surrogate, avoiding repeated expensive nonlocal forward solves during inversion.POD captures dominant measurement modes, and RBF interpolation maps potential parameters to modal amplitudes.Source: §6 · p. 21
• Finite differences handle one- and two-dimensional cases, including extended boundary conditions and an affine observation operator.Source: §6 · p. 21
• Tikhonov regularization and gradient-based optimization accompany the surrogate; the paper reports high computational efficiency and robustness to noise.Source: §6 · p. 21
• Three-dimensional problems and more complex nonlinear fractional equations are proposed as future extensions.Source: §6 · p. 21

• The authors state that they have no conflicts of interest.Source: CONFLICT OF INTEREST STATEMENT · p. 21

• The statement acknowledges NSF of China support and lists additional Gansu and Northwest Normal University project support.Source: FUNDING · p. 22