Inverse problems have gained popularity in statistical research recently. This dissertation consists of two statistical inverse problems: a Bayesian approach to detection of small low emission sources on a large random background, and parameter estimation methods for partial differential equation (PDE) models. Source detection problem arises, for instance, in some homeland security applications. We address the problem of detecting presence and location of a small low emission source inside an object, when the background noise dominates. The goal is to reach the signal-to-noise ratio levels on the order of 10^-3. We develop a Bayesian approach to this problem in two-dimension. The method allows inference not only about the existence of the source, but also about its location. We derive Bayes factors for model selection and estimation of location based on Markov chain Monte Carlo simulation. A simulation study shows that with sufficiently high total emission level, our method can effectively locate the source. Differential equation (DE) models are widely used to model dynamic processes in many fields. The forward problem of solving equations for given parameters that define the DEs has been extensively studied in the past. However, the inverse problem of estimating parameters based on observed state variables is relatively sparse in the statistical literature, and this is especially the case for PDE models. We propose two joint modeling schemes to solve for constant parameters in PDEs: a parameter cascading method and a Bayesian treatment. In both methods, the unknown functions are expressed via basis function expansion. For the parameter cascading method, we develop the algorithm to estimate the parameters and derive a sandwich estimator of the covariance matrix. For the Bayesian method, we develop the joint model for data and the PDE, and describe how the Markov chain Monte Carlo technique is employed to make posterior inference. A straightforward two-stage method is to first fit the data and then to estimate parameters by the least square principle. The three approaches are illustrated using simulated examples and compared via simulation studies. Simulation results show that the proposed methods outperform the two-stage method.

I propose a computationally efficient method to approximate the inverse of the normal operator arising in the multi-parameter linearized inverse problem for reflection seismology in two and three spatial dimensions. Solving the inverse problem using direct matrix methods like Gaussian elimination is computationally infeasible. In fact, the application of the normal operator requires solving large scale PDE problems. However, under certain conditions, the normal operator is a matrix of pseudodifferential operators. This manuscript shows how to generalize Cramer's rule for matrices to approximate the inverse of a matrix of pseudodifferential operators. Approximating the solution to the normal equations proceeds in two steps: (1) First, a series of applications of the normal operator to specific permutations of the right hand side. This step yields a phase-space scaling of the solution. Phase space scalings are scalings in both physical space and Fourier space. Second, a correction for the phase space scaling. This step requires applying the normal operator once more. The cost of approximating the inverse is a few applications of the normal operator (one for one parameter, two for two parameters, six for three parameters). The approximate inverse is an adequately accurate solution to the linearized inverse problem when it is capable of fitting the data to a prescribed precision. Otherwise, the approximate inverse of the normal operator might be used to precondition Krylov subspace methods in order to refine the data fit. I validate the method on a linearized version of the Marmousi model for constant density acoustics for the one-parameter problem. For the two parameter problem, the inversion of a variable density acoustics layered model corroborates the success of the proposed method. Furthermore, this example details the various steps of the method. I also apply the method to a 1D section of the Marmousi model to test the behavior of the method on complex two-parameter layered models.

In this thesis, the two-dimensional, unsteady, laminar and incompressible fluid flow problems governed by partial differential equations are solved by using dual reciprocity boundary element method (DRBEM). First, the governing equations are transformed to the inhomogeneous modified Helmholtz equations, and then the fundamental solution of modified Helmholtz equation is used for obtaining boundary element method (BEM) formulation. Thus, all the terms in the equation except the modified Helmholtz operator are considered as inhomogeneity. All the inhomogeneity terms are approximated by using suitable radial basis functions, and corresponding particular solutions are derived by using the annihilator method. Transforming time dependent partial differential equations to the form of inhomogeneous modified Helmholtz equations in DRBEM application enables us to use more information from the original governing equation. These are the main original parts of the thesis. In order to obtain modified Helmholtz equation for the time dependent partial differential equations, the time derivatives are approximated at two time levels by using forward finite difference method. This also eliminates the need of another time integration scheme, and diminishes stability problems. Stream function-vorticity formulations are adopted in physical fluid dynamics problems in DRBEM by using constant elements. First, the procedure is applied to the lid-driven cavity flow and results are obtained for Reynolds number values up to $2000.$ The natural convection flow is solved for Rayleigh numbers between $10^3$ to $10^6$ when the energy equation is added to the Navier-Stokes equations. Then, double diffusive mixed convection flow problem defined in three different physical domains is solved by using the same procedure. Results are obtained for various values of Richardson and Reynolds numbers, and buoyancy ratios. Behind these, DRBEM is used for the solution of natural convection flow under a magnetic field by using two different radial basis functions for both vorticity transport and energy equations. The same problem is also solved with differential quadrature method using the form of Poisson type stream function and modified Helmholtz type vorticity and energy equations. DRBEM and DQM results are obtained for the values of Rayleigh and Hartmann numbers up to $10^6$ and $300,$ respectively, and are compared in terms of accuracy and computational cost. Finally, DRBEM is used for the solution of inverse natural convection flow under a magnetic field using the results of direct problem for the missing boundary conditions.

In this thesis, the two-dimensional, unsteady, laminar and incompressible fluid flow problems governed by partial differential equations are solved by using dual reciprocity boundary element method (DRBEM). First, the governing equations are transformed to the inhomogeneous modified Helmholtz equations, and then the fundamental solution of modified Helmholtz equation is used for obtaining boundary element method (BEM) formulation. Thus, all the terms in the equation except the modified Helmholtz operator are considered as inhomogeneity. All the inhomogeneity terms are approximated by using suitable radial basis functions, and corresponding particular solutions are derived by using the annihilator method. Transforming time dependent partial differential equations to the form of inhomogeneous modified Helmholtz equations in DRBEM application enables us to use more information from the original governing equation. These are the main original parts of the thesis. In order to obtain modified Helmholtz equation for the time dependent partial differential equations, the time derivatives are approximated at two time levels by using forward finite difference method. This also eliminates the need of another time integration scheme, and diminishes stability problems. Stream function-vorticity formulations are adopted in physical fluid dynamics problems in DRBEM by using constant elements. First, the procedure is applied to the lid-driven cavity flow and results are obtained for Reynolds number values up to 2000. The natural convection flow is solved for Rayleigh numbers between 10^3 to 10^6 when the energy equation is added to the Navier-Stokes equations. Then, double diffusive mixed convection flow problem defined in three different physical domains is solved by using the same procedure. Results are obtained for various values of Richardson and Reynolds numbers, and buoyancy ratios. Behind these, DRBEM is used for the solution of natural convection flow under a magnetic field by using two different radial basis functions for both vorticity transport and energy equations. The same problem is also solved with differential quadrature method using the form of Poisson type stream function and modified Helmholtz type vorticity and energy equations. DRBEM and DQM results are obtained for the values of Rayleigh and Hartmann numbers up to 10^6 and 300, respectively, and are compared in terms of accuracy and computational cost. Finally, DRBEM is used for the solution of inverse natural convection flow under a magnetic field using the results of direct problem for the missing boundary conditions.

We are examining iteratively regularized methods for solving nonlinear inverse problems. Of particular interest for these types of methods are application problems which are unstable. For these application problems, special methods of numerical analysis are necessary, since classical algorithms tend to be divergent.

Synthetic Aperture Radar (SAR) imaging is a well-known technique in the domain of remote sensing, aerospace surveillance, geography and mapping. To obtain images of high resolution under noise, taking into account of the characteristics of targets in the observed scene, the different uncertainties of measure and the modeling errors becomes very important.Conventional imaging methods are based on i) over-simplified scene models, ii) a simplified linear forward modeling (mathematical relations between the transmitted signals, the received signals and the targets) and iii) using a very simplified Inverse Fast Fourier Transform (IFFT) to do the inversion, resulting in low resolution and noisy images with unsuppressed speckles and high side lobe artifacts.In this thesis, we propose to use a Bayesian approach to SAR imaging, which overcomes many drawbacks of classical methods and brings high resolution, more stable images and more accurate parameter estimation for target recognition.The proposed unifying approach is used for inverse problems in Mono-, Bi- and Multi-static SAR imaging, as well as for micromotion target imaging. Appropriate priors for modeling different target scenes in terms of target features enhancement during imaging are proposed. Fast and effective estimation methods with simple and hierarchical priors are developed. The problem of hyperparameter estimation is also handled in this Bayesian approach framework. Results on synthetic, experimental and real data demonstrate the effectiveness of the proposed approach.

The problem of inferring or reconstructing the material properties (cross sections) of a domain through noninvasive techniques, methods using only input and output at the domain boundary, is attempted using the governing laws of neutron diffusion theory as an optimization constraint. A standard Lagrangian was formed consisting of the objective function and the constraints to satisfy, which was minimized through optimization using a line search method. The chosen line search method was Newton's method with the Armijo algorithm applied for step length control. A Gaussian elimination procedure was applied to form the Schur complement of the system, which resulted in greater computational efficiency. In the one energy group and multi-group models, the limits of parameter reconstruction with respect to maximum reconstruction depth, resolution, and number of experiments were established. The maximum reconstruction depth for one-group absorption cross section or multi-group removal cross section were only approximately 6-7 characteristic lengths deep. After this reconstruction depth limit, features in the center of a domain begin to diminish independent of the number of experiments. When a small domain was considered and size held constant, the maximum reconstruction resolution for one group absorption or multi-group removal cross section is approximately one fourth of a characteristic length. When finer resolution then this is considered, there is simply not enough information to recover that many region's cross sections independent of number of experiments or flux to cross-section mesh refinement. When reconstructing fission cross sections, the one group case is identical to absorption so only the multi-group is considered, then the problem at hand becomes more ill-posed. A corresponding change in fission cross section from a change in boundary flux is much greater then change in removal cross section pushing convergence criteria to its limits. Due to a more ill-posed problem, the maximum reconstruction depth for multi-group fission cross sections is 5 characteristic lengths, which is significantly shorter than the removal limit. To better simulate actual detector readings, random signal noise and biased noise were added to the synthetic measured solutions produced by the forward models. The magnitude of this noise and biased noise is modified and a dependency of the maximum magnitude of this noise versus the size of a domain was established. As expected, the results showed that as a domain becomes larger its reconstruction ability is lowered which worsens upon the addition of noise and biased noise.

In this dissertation, two recent problems from tomographic imaging are studied, and results from numerical simulations with synthetic data are presented. The first part deals with ultrasound modulated optical tomography, a method for imaging interior optical properties of partially translucent media that combines optical contrast with ultrasound resolution. The primary application is the optical imaging of soft tissue, for which scattering and absorption rates contain important functional and structural information about the physiological state of tissue cells. We developed a mathematical model based on the diffusion approximation for photon propagation in highly scattering media. Simple reconstruction schemes for recovering optical absorption rates from boundary measurements with focused ultrasound are presented. We show numerical reconstructions from synthetic data generated for mathematical absorption phantoms. The results indicate that high resolution imaging with quantitatively correct values of absorption is possible. Synthetic focusing techniques are suggested that allow reconstruction from measurements with certain types of non-focused ultrasound signals. A preliminary stability analysis for a linearized model is given that provides an initial explanation for the observed stability of reconstruction. In the second part, backprojection schemes are proposed for the detection of small amounts of highly enriched nuclear material inside 3D volumes. These schemes rely on the geometrically singular structure that small radioactive sources represent, compared to natural background radiation. The details of the detection problem are explained, and two types of measurements, collimated and Compton-type measurements, are discussed. Computationally, we implemented backprojection by counting the number of particle trajectories intersecting each voxel of a regular rectangular grid covering the domain of detection. For collimated measurements, we derived confidence estimates indicating when voxel trajectory counts are deviating significantly from what is expected from background radiation. Monte Carlo simulations of random background radiation confirm the estimated confidence values. Numerical results for backprojection applied to synthetic measurements are shown that indicate that small sources can be detected for signal-to-noise ratios as low as 0.1%.

The study of neurons is of fundamental importance in biology and medicine. Neurons are the most basic unit of information processing in the nervous system of humans and all other vertebrates and in complex invertebrates. In addition, networks of neurons (the human brain) are the most sophisticated computational devices known, and the study of neurons individually and working in concert is seen as a step toward understanding consciousness and cognition. In the 1950's Hodgkin and Huxley developed a system of nonlinear ordinary differential equations to describe the behavior of a neuron found in the squid. Equations of this form have since been used to model the behavior of a multitude of neurons across a broad spectrum of species. Hodgkin-Huxley type neuron models helped lay the foundation for computational neuroscience, and they remain widely used in the study of neuron behavior almost sixty years after their development. Hodgkin-Huxley type models accept a set of parameters as input and generate data describing the electrical activity of the neuron as a function of time. We develop inversion algorithms to predict a set of input parameter values from the voltage trace data generated by the model. We test our algorithm on data from the Hodgkin-Huxley equations, and we extend the algorithm to solve the inverse problem associated with a more complex Hodgkin-Huxley type model for a lobster stomatogastric neuron. We find strong empirical evidence that the algorithms produce parameter values that generate a good fit to the target voltage trace, and we prove that under certain conditions the inversion algorithm for the Hodgkin-Huxley equations converges to a perfect match. To our knowledge this is the first parameter optimization procedure for which convergence has been shown theoretically. Understanding the relationship between the parameters of a neuron model and its output has implications for designing effective neuron models and for explaining the mechanisms by which neurons regulate their behavior. Inversion algorithms for Hodgkin-Huxley type neuron models are an important theoretical and practical step toward understanding the relationship between model parameters and model behavior, and toward the larger problem of inferring neuronal parameters from behavior observed experimentally.

<p>Axicons are optical elements that produce Bessel beams,i.e., long and narrow focal lines along the optical axis. Thenarrow focus makes them useful ine.g. alignment, harmonicgeneration, and atom trapping, and they are also used toincrease the longitudinal range of applications such astriangulation, light sectioning, and optical coherencetomography. In this thesis, axicons are designed andcharacterized for different kinds of illumination, using thestationary-phase and the communication-modes methods.</p><p>The inverse problem of axicon design for partially coherentlight is addressed. A design relation, applicable toSchell-model sources, is derived from the Fresnel diffractionintegral, simplified by the method of stationary phase. Thisapproach both clarifies the old design method for coherentlight, which was derived using energy conservation in raybundles, and extends it to the domain of partial coherence. Thedesign rule applies to light from such multimode emitters aslight-emitting diodes, excimer lasers and some laser diodes,which can be represented as Gaussian Schell-model sources.</p><p>Characterization of axicons in coherent, obliqueillumination is performed using the method of stationary phase.It is shown that in inclined illumination the focal shapechanges from the narrow Bessel distribution to a broadasteroid-shaped focus. It is proven that an axicon ofelliptical shape will compensate for this deformation. Theseresults, which are all confirmed both numerically andexperimentally, open possibilities for using axicons inscanning optical systems to increase resolution and depthrange.</p><p>Axicons are normally manufactured as refractive cones or ascircular diffractive gratings. They can also be constructedfrom ordinary spherical surfaces, using the sphericalaberration to create the long focal line. In this dissertation,a simple lens axicon consisting of a cemented doublet isdesigned, manufactured, and tested. The advantage of the lensaxicon is that it is easily manufactured.</p><p>The longitudinal resolution of the axicon varies. The methodof communication modes, earlier used for analysis ofinformation content for e.g. line or square apertures, isapplied to the axicon geometry and yields an expression for thelongitudinal resolution. The method, which is based on abi-orthogonal expansion of the Green function in the Fresneldiffraction integral, also gives the number of degrees offreedom, or the number of information channels available, forthe axicon geometry.</p><p><b>Keywords:</b>axicons, diffractive optics, coherence,asymptotic methods, communication modes, information content,inverse problems</p>