SVD Singular Value Decomposition and Damped Least Squares Algorithm

The SVD (Singular Value Decomposition) method and damped least squares algorithm are two effective numerical approaches for solving geophysical inverse problems. Both methods can handle ill-conditioned linear systems and are widely applied in geophysical exploration, seismic wave inversion, and related fields.

The SVD method addresses inverse problems by decomposing the coefficient matrix into the product of three matrices: U, Σ, and V^T. This decomposition clearly reveals the matrix rank and ill-conditioning degree. By truncating small singular values, the solution computation becomes stabilized. The key implementation advantage lies in intuitive control over solution smoothness, though it requires careful threshold selection for singular value truncation. In code implementation, the numpy.linalg.svd() function in Python or svd() in MATLAB can perform this decomposition, followed by filtering singular values below a defined tolerance.

The damped least squares algorithm improves problem well-posedness by introducing a regularization term. It adds a solution norm constraint to the objective function, thereby suppressing high-frequency noise and unstable components. The damping coefficient selection is critical - excessively large values cause over-smoothing, while insufficient values fail to effectively control instability. Algorithm implementation typically involves solving (A^TA + λI)x = A^Tb, where λ represents the damping parameter that can be optimized through L-curve analysis or cross-validation techniques.

Both methods require balancing solution fitness and smoothness. In practical applications, SVD is suitable for analyzing mathematical characteristics of problems, while damped least squares facilitates easier incorporation of physical constraints. For beginners, starting with simple linear inverse problems and gradually adjusting parameters to observe solution changes provides an effective approach to understanding these algorithms. Code practice could begin with implementing basic versions using linear algebra libraries before progressing to automated parameter optimization routines.