A Batched Discrepancy Principle for Stochastic Landweber-Kaczmarz Iteration and its Application to EIT
This paper proposes a stochastic mini-batch Landweber-Kaczmarz method with a novel batched discrepancy principle as a stopping rule for solving ill-posed inverse problems, establishing its theoretical convergence under the tangential cone condition and validating its performance through numerical experiments on electrical impedance tomography.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine trying to see inside a dark room by only feeling the temperature on the walls. You know that heat travels differently through wood, metal, or air, so by measuring how the temperature changes at the surface, you could theoretically figure out what objects are hidden inside. This is the essence of a medical imaging technique called electrical impedance tomography, which aims to map the electrical conductivity inside the human body using only boundary measurements. However, this is a notoriously difficult puzzle. The data collected is often noisy, and the mathematical equations used to reverse-engineer the internal picture are unstable; tiny errors in the measurements can lead to wildly incorrect images. To solve this, scientists use a process called regularization, which acts like a filter to smooth out the noise and find a plausible solution. But when the number of measurements is huge, the standard way of solving these equations becomes painfully slow, like trying to count every single grain of sand on a beach one by one.
Researchers have long sought a faster way to tackle these massive systems of equations. One popular approach, known as stochastic gradient descent, speeds things up by looking at just one equation at a time, randomly picking a piece of the puzzle to solve before moving to the next. While this is much faster, it can be jittery and unstable, often wandering off course because it relies on such a small sample of information. Another method looks at all the equations at once, which is stable but incredibly slow. The challenge has been finding a middle ground that is both fast and reliable. In a new study, a team of mathematicians has developed a technique that sits right between these two extremes. They call it a "mini-batch" approach, where instead of looking at just one equation or the entire set, the computer randomly selects a small group of equations to solve in each step. This reduces the noise and instability of the single-equation method without incurring the heavy cost of checking everything at once.
The core of this new work is not just the method of solving the equations, but a new rule for knowing when to stop. In these iterative processes, you cannot simply run the calculation forever; if you go too long, the noise in the data starts to corrupt the image, making it worse rather than better. The standard rule for stopping is to check the "discrepancy," or the difference between the calculated result and the actual noisy data, and stop when this difference matches the expected level of noise. However, checking this difference for every single equation in a massive system is just as slow as the original problem. The researchers introduced a clever shortcut: a "batched discrepancy principle." Instead of checking the error for the whole system, the computer only checks the error for the small, random group of equations it just solved. By averaging these small checks over a sequence of recent steps, the method can accurately estimate when the solution is good enough to stop, without ever having to look at the entire dataset at once.
To test this idea, the team applied it to a simulated version of the electrical impedance tomography problem, using a mathematical model of a two-dimensional domain with a complex, wavy pattern of conductivity. They compared their new batched method against the traditional, slower approach. The results were striking. By carefully choosing the size of the random group of equations and how many recent steps to average together, the new method produced images of the same high quality as the traditional method but required significantly less computing time. In their simulations, the most efficient setup used a very small group of equations—just one at a time—but averaged the stopping decision over a long history of about 150 previous steps. This allowed the algorithm to move quickly while still making a reliable decision about when to stop. The study proves mathematically that this approach works, showing that the images converge to the correct solution as the noise in the data gets smaller.
The researchers also explored how different settings affected the outcome. They found that if the group of equations was too small and the history of steps too short, the stopping rule became unstable, causing the algorithm to quit too early and produce a blurry, incomplete image. Conversely, if the group was too large, the speed advantage disappeared. The sweet spot was a balance where the randomness of the small group was tamed by looking back at a sufficient number of past steps. This work demonstrates that for large-scale imaging problems, you do not need to check every single piece of data to know when you have found the answer. By using a smart, randomized sampling strategy, it is possible to solve complex medical imaging problems much faster, potentially making these technologies more practical for real-world use where time and computing power are limited. The study confirms that this specific combination of random sampling and a new stopping rule is a robust and efficient way to handle the difficult mathematics behind seeing inside the human body.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.