Inverse problems, spring 2015
The course is lectured in English
Figure: electrical conductivity distribution (left), nonlinear reconstruction from voltage-to-current boundary measurements (middle), edge-enhanced nonlinear reconstruction (right).
Inverse problems are about interpreting indirect measurements. The scientific study of inverse problems is an interdisciplinary field combining mathematics, physics, signal processing, and engineering. Examples of inverse problems include
- Three-dimensional X-ray imaging (more information, also see this video and this video)
- Recovering the inner structure of the Earth based on earthquake measurements
- Sharpening a misfocused photograph (more information )
- Reconstructing electric conductivity from current-to-voltage boundary measurements (see this page and this page)
- Finding cracks inside solid structures
- Prospecting for oil and minerals
- Monitoring underground contaminants
- Finding the shape of asteroids based on light-curve data (see this page)
The common features of all this problems are the need to understand indirect measurements and to overcome extreme sensitivity to noise and modelling inaccuracies.
Figure: sharp image (left), misfocused image (middle), sharpened image by deconvolution (right).
Inverse problems research is an active area of mathematics.
At the Department of Mathematics and Statistics the field is represented by three research groups belonging to
a Centre of Excellence of Academy of Finland:
What does the course contain?
The goals of the course are
- introduce discrete matrix models of some widely used measurements, such as tomography and convolution
- show how to detect ill-posedness (sensitivity to measurement noise) in matrix models using Singular Value Decomposition
- compute noise-robust reconstructions using regularization
- write Matlab algorithms for sharpening photographs and computing tomographic reconstructions
- discussion of nonlinear inverse problems, with Electrical Impedance Tomography as an example
The lectures make up 10 credit units. In addition to lectures the course involves a project work. It is done in teams of two and gives 5 credit units to each student.
The course is in total 15 credit units.
Figure: high-resolution X-ray tomographic slice through a walnut (left), reconstruction from few data using an old method (middle), and reconstruction from few data using a modern method (right).
Recommended courses to take before this course: Linear algebra 1 and 2, Applications of matrix computations.
Some previous experience with Matlab programming is very helpful.
Structure of the course
Period III: Lectures as follows:
Tuesday 10-12 in room D123
Wednesday 12-14 in room D123
Friday 12-14 in room C123.
Two hours of exercise classes per week.
Period IV: Project work, which is reported as a poster in a poster session.
Lecture 1 (January 13, 2015)
Introduction to inverse problems, indirect measurements and ill-posedness.
The (huge) files of the presentations are available in this Dropbox folder.
The practical information slides are available in the file PracticalInfo.pdf.
Lecture 2 (January 14, 2015)
One-dimensional convolution of real-valued functions. Book section 2.1.1.
Matlab resources: target1.m, targets_plot.m, PSF.m, deconv1_cont_comp.m, deconv1_cont_plot.m.
Lecture 3 (January 16, 2015)
One-dimensional convolution in both continuous and discrete form. First example of the failure of naive inversion in the case of an ill-posed problem. Book sections 2.1.2 and 2.1.3.
Matlab resources: DC_convmtx.m, deconv1_cont_comp.m, deconv1_cont_plot.m,
deconv2_discretedata_comp.m, deconv2_discretedata_plot.m, deconv3_naive_comp.m, deconv3_naive_plot.m
Lecture 4 (January 20, 2015)
Computational studies of the one-dimensional deconvolution. Simulation of crime-free data. Book sections 2.1.3 and 2.1.4.
Lecture 5 (January 21, 2015)
Singular value decomposition of matrices and how it explains the instability of naive inversion. Book chapter 3.5.
Matlab resources: deconv4_SVD_comp.m, deconv4_SVD_plot.m, deconv5_truncSVD_comp.m, deconv5_truncSVD_plot.m
Lecture 6 (January 23, 2015)
Well-posedness and ill-posedness in the sense of Hadamard. Quick introduction to the concept of regularization. Minimum norm solution and the Moore-Penrose pseudoinverse. Book sections 3.1 and 3.4. and 4.1. (No computational stuff this time.)
There will be an exam after the lecture part of the course.
Mueller J L and Siltanen S: Linear and Nonlinear Inverse Problems with Practical Applications. SIAM 2012.
Course assistant: Andreas Hauptmann
The exercise sessions will be held in the computer room C128. There are two sessions:
The theoretical exercises are to be presented by the students. In each session the students mark the exercises they have completed and according to this list one student will be chosen per exercise to present.
The Matlab exercises will be checked individually at the computers present. The Matlab exercises can be completed in pairs of two. Both students have to be present and able to explain the code.
Points for all completed exercises will be accumulated and part of the final grade.
Exercise 1 (January 20-22, 2015) - Solutions 1
Exercise 2 (January 27-29, 2015)
The idea of the project work is to study an inverse problem both theoretically and computationally in teams of two students. The end product is a scientific poster that the team will present in a poster session in the Industrial Mathematics Laboratory. The poster can be printed using the laboratory's large scale printer. The classical table of contents is recommended for structuring the poster:
2 Materials and methods
Section 2 is for describing the data and the inversion methods used. In section 3 those methods are applied to the data and the results are reported with no interpretation; just facts and outcomes of computations are described. Section 4 is the place for discussing the results and drawing conclusions.
The project is about X-ray tomography. You can measure a dataset yourself in the X-ray facility of the Industrial Mathematics Laboratory:
In the project you are supposed to take a subset of the data with only few projections, such as 20 projection directions. Use one of these methods to recover the walnut slice from sparse data:
- Tikhonov regularization based on conjugate gradient method,
- approximate total variation regularization implemented iteratively with the Barzilai-Borwein method,
- some other suitable method.
Optimally, you should have an automatic method for choosing the regularization parameter and an automatic stopping criteria for the iteration. These are both difficult requirements, so have a simple approach as plan B if a more complicated approach does not work.
First goal consists of two things: (a) two first sections should be preliminary written in LaTeX (not necessarily in poster format yet) and (b) the Matlab codes at the following webpage should be run and studied:
Two things will be graded in the meeting about the first goal: (a) the draft of project work and (b) your understanding of the Matlab codes at the above webpage relevant to your topic. The grade represents 30% of the final grade of the project work. Please agree on a meeting time (in the period April 1-4) with the lecturer for reviewing and grading the first goal.
Second and final goal: poster is presented in the poster session.