Year: 2025
Type: dissertation
Abstract: This thesis considers the non-convex methods for low-rank matrix estimation based on Burer-Monteiro factorization, which offers better scalability and lower per-iteration costs compared to convex methods.However, non-convex optimization faces challenges from spurious local minima or critical points, which can lead gradient-based algorithms to converge to suboptimal solutions.Prior research has shown that a sufficiently large training dataset can eliminate spurious minima in non-convex problems, enabling successful low-rank estimation.However, these guarantees are often impractical in real-world applications for two reasons.First, they require a large amount of data to remove spurious local minima across the entire parameter space, which is costly and challenging for large-scale applications.Second, they assume "regular" measurement conditions (e.g., restricted isometry property or incoherence), which are rarely satisfied in practice.To address these challenges, we first prove that a good initial point can reduce the required sample size to avoid spurious minima near the ground truth.This insight reveals that practical low-rank estimation I still remember traveling to Champaign from California for the first time, almost five years ago.On the way to campus from ORD, I saw cornfields.Lots of cornfields.It was a cold day and everything was a bit bleak.I would never have guessed that of all places, Illinois would be the one that feels most like a warm home to me in America.For this, and for many other things, I have to thank all the people that made my PhD possible.First and foremost I would like to thank my advisor Richard Y. Zhang.I am extremely grateful to have Richard as my advisor.When I first started working with Richard, it was immediately obvious to me that Richard was one of the smartest people that I had ever met.He thinks incredibly fast and comes up with deep insights at the speed of light.During our first meetings, it would often take me a week just to process what we talked about in one hour.Especially during my first year, I was constantly astounded and even a bit overwhelmed by how effortlessly he came up with new mathematical ideas.But he always assured me that a lot of this comes with experience, and that I would eventually "get there" as well.During this process, Richard not only helped me improve my technical abilities, but also pushed me to be a much better communicator of my ideas as well.Good writing and presentation is something that he values deeply.One summer, we even paused research for a few weeks, so that I could really learn to communicate a deep mathematical algorithm with as much simplicity and clarity as possible.That process has helped me tremendously since.Before my talk at the SIAM optimization conference in Seattle, we went through my slides again and again.The technical proofs in my paper was just too hard to convey to a broader audience.In the end we came up with a geometric explanation that not only made my talk successful, but also deepened my own understanding of the problem.I still remember how proud and happy he was after the talk.For all of these moments and many others, I will always be thankful.I would also like to thank my committee members, Professors Rayadurgam Srikant, Maxim Raginsky, Zhizhen Zhao.I am very fortunate to have them on both my preliminary exam and my final defense committee.Their valueable feedback and insights have helped me improve my thesis greatly.I also owe a great deal to
Author: Zhang, Jialun
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Keywords: low-rank matrix estimation, Burer-Monteiro factorization, non-convex optimization, spurious local minima, initialization, sample complexity, incoherence, restricted isometry property, gradient-based optimization, convergence guarantees, matrix completion, robust principal component analysis
Open Access status: green
OpenAlex ID: W7220806425
