Mathematical Foundation

Tikhonov Functional

\[J(\mathbf{x}) = \|\mathbf{Ax} - \mathbf{y}\|^2 + \alpha\|\mathbf{Lx}\|^2\]

Minimizing yields: \(\mathbf{x}_\alpha = (\mathbf{A}^T\mathbf{A} + \alpha\mathbf{L}^T\mathbf{L})^{-1}\mathbf{A}^T\mathbf{y}\)

α = 1e-3
σ = 1.0 mGal

Zero-Order (L = I)

Standard Tikhonov regularization penalizing solution norm. Best for problems where small solution values are expected. Equivalent to ridge regression.

First-Order

Penalizes first derivative, promoting smooth solutions. Good for gravity field recovery where gradients should be continuous.

Second-Order

Penalizes curvature of solution. Produces very smooth fields, appropriate when signal is expected to be slowly varying.

Kaula's Rule

Degree-dependent regularization based on expected gravity field power spectrum: σₙ² ∝ 1/n⁴. Physically motivated for Earth's gravity.

--
Condition Number
--
Effective Rank
--
Regularized Cond.
--
Resolution (km)

L-Curve Parameter Selection

L-Curve Criterion

Plot \(\log\|\mathbf{Ax}_\alpha - \mathbf{y}\|\) vs \(\log\|\mathbf{Lx}_\alpha\|\) for varying α

Optimal α found at corner of L-shaped curve (maximum curvature)

10^-8 to 10^0
50 points
--
Optimal α
--
Residual Norm
--
Solution Norm
--
Curvature

Interactive Gravity Field Inversion

α = 1e-3
N = 60
2.0 mGal
--
RMS Error (mGal)
--
Max Error (mGal)
--
Correlation
--
Resolved Degree

Regularization Method Comparison

Tikhonov (Classical)

• Direct solution, no iterations
• Requires full matrix storage
• Best for moderate-size problems
• Easy error propagation

TSVD (Truncated SVD)

• Discrete regularization
• Sharp spectral cutoff
• Good for rank-deficient problems
• No smoothing between components

Iterative (CG, CGLS)

• Memory efficient
• Early stopping as regularization
• Suitable for large-scale problems
• Harder to quantify uncertainty