What Are Inverse Problems?

Inverse problems are a class of mathematical and computational problems in which the goal is to recover an unknown cause or underlying model from observed effects or measurements. They contrast with forward problems, where a known cause is used to predict an outcome. In a forward problem, a geophysicist might simulate how seismic waves propagate through a known subsurface structure; the corresponding inverse problem uses recorded surface vibrations to infer the structure itself. Inverse problems arise across signal processing, medical imaging, astronomy, and geophysics, wherever the quantity of scientific or engineering interest cannot be measured directly.

The field draws on functional analysis, integral equations, and numerical linear algebra. Andrey Tikhonov's 1943 work on ill-posed problems, and his subsequent development of regularization theory in the 1960s, provided the foundational mathematical framework that remains central to the discipline today.

Ill-Posedness and Regularization

The defining challenge of inverse problems is ill-posedness, a concept formalized by Jacques Hadamard's three criteria: a well-posed problem must have a solution that exists, is unique, and depends continuously on the data. Most inverse problems violate at least the uniqueness or stability conditions. Small perturbations in the measured data, including sensor noise, can produce wildly different reconstructed solutions when no constraint is imposed.

Regularization techniques address this by introducing additional information or constraints that select a stable, physically meaningful solution from among the many that fit the data. Tikhonov regularization, the most widely used approach, adds a penalty term proportional to the squared norm of the solution or its derivatives, penalizing solutions that oscillate rapidly or deviate far from a prior estimate. The balance between data fidelity and the regularization penalty is controlled by a scalar parameter whose selection is itself a research problem, with methods including cross-validation, the L-curve criterion, and generalized cross-validation. A detailed treatment of these methods appears in modern regularization methods for inverse problems in Acta Numerica.

Image Reconstruction

Image reconstruction is the most computationally intensive and widely deployed application domain for inverse problem methods. In computed tomography (CT), the forward model is the Radon transform, which maps a two-dimensional attenuation distribution to a set of line integrals measured by X-ray detectors. Recovering the image from these projections requires inverting the transform in the presence of noise and limited angular sampling. Filtered back-projection, an analytic inversion method, has been used clinically since the 1970s, but iterative reconstruction algorithms that incorporate regularization now dominate in low-dose settings because they suppress noise more effectively at equivalent image quality. A detailed survey of regularization approaches for imaging inverse problems, including methods based on learned image priors and deep neural networks, is provided in an overview of regularization methods for imaging inverse problems on arXiv.

Integral Equations and Functional Analysis

Many physical forward models are expressed as Fredholm integral equations of the first kind, where the unknown function appears under an integral operator. Inverting such operators is intrinsically unstable because the operator's singular values decay to zero, amplifying any noise component in the data. Functional analysis provides the theoretical language for characterizing these operators, the spaces in which solutions exist, and the convergence behavior of iterative solvers. Krylov subspace methods, including LSMR and LSQR, are standard numerical workhorses for large-scale ill-posed systems. A hybrid LSMR approach for large-scale Tikhonov regularization is analyzed in detail by Chung and Saibaba in SIAM Journal on Scientific Computing, demonstrating how iterative regularization can be applied efficiently to problems with millions of unknowns.

Applications

Inverse problems have applications in a wide range of disciplines, including:

• Medical imaging, including CT, MRI, and ultrasound tomography
• Seismic exploration and subsurface geophysical imaging
• Remote sensing and radar signal reconstruction
• Non-destructive evaluation of materials and structures
• Astronomy and radio telescope aperture synthesis