University of British Columbia
Causal discovery and density estimation : from algebra to stochastics
Abstract
dc:descriptionThis dissertation employs algebraic and stochastic methods to develop a mathematical understanding of complex statistical models. Chapter 2 proposes a two-step hybrid method for learning a possibly cyclic causal graph given the conditional independence (CI) relations satisfied by a Markov and faithful observational distribution. We first optimize a score over partially ordered partitions of variables to identify the right Markov equivalence class, and then construct a graph within this class. Directed graphs are covariance equivalent if their associated linear structural equation models induce the same collection of covariance matrices, up to a Lebesgue-measure-zero set. Chapter 3 considers a linear map whose fibers correspond to graphs with identical characteristic imset vectors, and studies its associated toric ideal to prove such graphs are covariance equivalent. Chapter 4 develops a parameter estimation method for signature SDEs, a rich class of path-dependent SDEs capturing non-Markovian phenomena. We establish existence and uniqueness of solutions to signature SDEs, and introduce the Expected Signature Matching Method for parameter estimation in linear signature SDEs. We prove consistency and asymptotic normality of the estimator, and demonstrate its performance in simulations. Chapter 5 investigates linear concentration models of maximum likelihood (ML) degree one via homaloidal polynomials. For models associated to spanning tree polynomials of graphs, we show ML degree equals one for chordal graphs and 2^(n-1)-n for n-cycles. We bound matrix sizes in symmetric determinantal representations of polynomials; applied to homaloidal polynomials, such representations yield novel ML-degree-one models. Chapter 6 studies Independent Component Analysis mixture models, non-parametric extensions of Gaussian mixtures. We establish when the mixing coefficients, linear transformations, and source moments are recoverable, up to a finite class, from observed moments up to order d. For n-variate distributions, d ⩾ 3, and a positive constant c_d, we prove algebraic identifiability for up to c_dn^(d-2) mixture components. Chapter 7 explores determinantal ideals encoding certain CI statements in discrete distributions with hidden variables. We provide decompositions, prove radicality and primeness of components under certain conditions, compute component dimensions, and present a combinatorial classification and counting formula. Our results establish a generalized intersection property for discrete CI models with hidden variables.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy - PhD
- Level thesis:degree_level
- doctoral
- Discipline thesis:degree_discipline
- Mathematics
- Grantor dc:publisher
- University of British Columbia
- Year dc:date
- 2026
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Semnani, Pardis
- Contributors dc:contributor
-
- Robeva, Elina
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/2429/94968