Abstract
dc:descriptionTrend filtering is a nonparametric regression method that penalises divided differences of the underlying signal, yielding locally adaptive, piecewise polynomial estimators. Although its statistical and computational properties are well understood under Gaussian loss, the non-Gaussian setting—central to count and exponential-family observations such as epidemic incidence— remains computationally challenging, since the non-quadratic loss admits no closed-form proximal step. This dissertation develops methodologies and algorithms for trend filtering with Poisson loss. First, we formulate the problem of Poisson trend filtering on general graphs, and develop two generic proximal algorithms, based on linearised alternating direction method of multipliers and proximal Newton methods respectively. Both algorithms are built upon the alternating direction method of multipliers (ADMM) decomposition of the Lagrangian function of the objective and attain per-iteration linear-time complexity. Second, we apply this framework to estimate time-varying reproduction numbers from infectious disease incidence data, demonstrating that locally adaptive, piecewise polynomial trend filtering yields accurate estimates with principled, cross-validated tuning; the method is released as the R package rtestim. Third, we develop an accurate algorithm for the linear systems on divided difference operators that arise as inner ADMM subproblems. By reinterpreting the normal equations as posterior-mean computation in a linear Gaussian state-space model, the solver achieves linear-time complexity for fixed penalty order and provides substantially improved numerical stability when design points are irregularly spaced and the linear system is ill-conditioned. All three contributions are validated through synthetic experiments and real-world applications, including Covid-19 surveillance and spatial temperature analysis. Finally, we present ongoing theoretical work showing that introducing observation-dependent weights into the exponential-family trend filtering penalty on lattices recovers minimax optimal convergence rates, addressing the excess prediction risk caused by heteroscedasticity in the effective noise. Together, these contributions provide a modular computational and theoretical toolkit for locally adaptive nonparametric regression under non-Gaussian loss on structured domains.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Statistics
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Liu, Jiaping
- Contributors dc:contributor
-
- McDonald, Daniel
Rights
dc:rights- Statement dc:rights
-
- Attribution-NonCommercial-NoDerivatives 4.0 International
- Language dc:language
- eng
Identifiers
dc:identifier.*- Handle dc:identifier
- http://hdl.handle.net/2429/95559