Google Scholar · arXiv · ORCID · CV
Much of my work starts from one observation: algorithms like Gaussian elimination make choices as they run, and with random inputs those choices become random objects worth studying. Sometimes this tells us something about the algorithm, such as how stable it is or how bad growth can get. Sometimes the permutations, trees, and fractals that fall out turn out to be interesting on their own. I’m also drawn to universality, including finding random matrix statistics in places you wouldn’t expect them.
Students: many of these questions have a computational way in. If you’re interested in working on one, get in touch.
Every time you solve a linear system with partial pivoting (GEPP), you produce a permutation: the factor P in PA = LU. If A is random, what does P look like? Is it uniform? What are its cycles, its longest increasing subsequence, its shape?
Random butterfly matrices make this question tractable. GEPP on butterfly matrices produces butterfly permutations, a family of separable permutations that sit inside Sylow p-subgroups of the symmetric group, which are iterated wreath products. Their recursive structure leads to:
cycle counts driven by branching processes
longest increasing subsequences that grow like a power of n other than √n
tree statistics (heights, Horton–Strahler numbers) with their own scaling laws
in the limit, random fractal measures whose dimension you can tune
Outside the butterfly setting, GEPP on Gaussian, GOE/GUE, and Bernoulli matrices never gives uniform permutations. I’m now extending this approach to column-pivoted QR.
Distribution of the number of pivots needed using Gaussian elimination with partial pivoting on random matrices. Ann. Appl. Probab. 34(2), 2294–2325 (2024). arXiv · journal · poster
On the longest increasing subsequence and number of cycles of butterfly permutations (with Chenyang Zhong). (2024) arXiv · poster
Heights of butterfly trees (with Chenyang Zhong). Random Structures & Algorithms 68(3), e70068 (2026). arXiv · journal
The Horton–Strahler number of butterfly trees. (2025) arXiv
Random permutations using GEPP (with Kenji Gunawan and Chenyang Zhong). (2026) arXiv
Pivoting exists to control growth, and growth controls error. Worst-case growth for GEPP is exponential, yet in practice it almost never shows up. I want to understand that gap, and what you give up by changing the pivoting strategy.
One way to skip pivoting is to precondition with random butterfly matrices (Parker, 1995). This is essentially free, since butterfly products cost O(n² log n), and it saves the data movement that pivoting costs in practice. With Tom Trogdon, I gave the full distribution of the growth factor for Haar-butterfly matrices, where earlier work on random growth factors only had empirical moments. Follow-up work treats complete pivoting, butterfly Hadamard matrices, and worst-case growth over orthogonal matrices, including how GEPP and GECP compare on the same system.
I also returned to Wilkinson’s classical worst-case matrix, which turns out to be a discretized Volterra operator. This gives its sharp 2-norm condition number and explains the lattice of large forward errors it produces in floating point.
Recent computer-assisted searches have overturned my conjectured extremal orthogonal matrix, and I’m now working with an undergraduate researcher on the right asymptotic constant.
Growth factors of random butterfly matrices and the stability of avoiding pivoting (with Tom Trogdon). SIAM J. Matrix Anal. Appl. 44(3), 945–970 (2023). arXiv · journal · poster
Growth factors of orthogonal matrices and local behavior of Gaussian elimination with partial and complete pivoting. SIAM J. Matrix Anal. Appl. 45(3) (2024). arXiv · journal
Complete pivoting growth of butterfly matrices and butterfly Hadamard matrices. Linear Multilinear Algebra 74(7), 934–958 (2026). arXiv · journal
Maximal growth matrices are shifted Volterra operators. Submitted (2026). pdf
My dissertation, Numerical, spectral and group properties of random butterfly matrices (UC Irvine, 2021), is where this all started.
Universality predicts that eigenvalue spacing laws should appear well outside of matrices, and they do: bus arrivals, parked cars, subway trains. I like adding new examples, and occasionally finding systems where the prediction fails.
Ocean waves (with Tom Trogdon and Jim Thomson, APL–UW): spacings between wave peaks in SWIFT buoy data follow the Wigner surmise.
Interstate rest areas (with Arizona undergrads Jia Cai and Yunke Wan): Wigner surmise statistics nationally, Poisson statistics in the Rocky Mountain region. Unlike the traffic examples, this one is built from fixed infrastructure. Random matrix statistics and safety rest areas on interstates in the United States. J. Stat. Mech. (2024) 063404. arXiv · journal
Earthquakes and solar flares: honors thesis by Anupama Hazarika, which I supervised. Thesis
Marathon runners (with Michael Conroy): a case where the prediction fails, with power-law spacings in place of random matrix statistics.
Totally positive Hessenberg Toeplitz matrices (with Nick Ercolani and Jon Ramalheira-Tsu). Combining total positivity with Toeplitz structure brings together two sets of spectral tools. Total positivity and spectral theory for Toeplitz Hessenberg matrix ensembles. (2023) arXiv
Burning number of graphs. The burning number models how fast information spreads on a network, and the burning number conjecture says b(G) ≤ ⌈√n⌉ for connected graphs on n vertices. I prove it for comb graphs, with an exact formula in the spine-dominant regime and a sharp asymptotic bound in the tooth-dominant regime. Burning rooted graph products. (2026) arXiv
I’m also interested in random matrices over the integers, finite fields, and p-adics, and their connections to Cohen–Lenstra heuristics.
Attitudes toward disability and accommodations in university mathematics (with Alison Mirin, Paulo Tan, and Rachel VanDaalen). Recognizing disability as diversity: The importance of and plans for exploring attitudes and ableism in tertiary maths. Delta 2025 Proceedings, 41–51. journal · video
Universal design in an online linear algebra course (with Melinda Lanius, 2023).
UCI Math CEO (2019–2021, with Alessandra Pantano): I built the reporting structure used to measure the program’s effect on student math and reading proficiency. Supported by a UCI Inclusive Excellence Spirit Award and a Summer Inclusive Excellence Grant.