Unless otherwise noted, seminars are held on Fridays from 02:00 to 03:00 PM in PGH 648 (Philip Guthrie Hoffman Hall).
Yong-Ki Ma
Professor, Department of Applied Mathematics, Kongju National University
Title: Pricing the credit default swap rate for jump diffusion default intensity processes
Abstract: In this talk, I will be discussing methodology for credit risk modeling in financial mathematics.
Since I am concerned with modeling the default time, the intensity-based approach is adopted in my framework of modeling the credit risk. My research reports a theoretical study addressing an important issue, correlated default, in the risk management of credit derivatives. I present a formula for credit default swap (CDS) spreads under counterparty default risk. The default intensities in my model are driven by dependent jump diffusion processes. This dependence structure is studied by using a copula. I obtain a generalized CDS rate formula based upon the calculation of the joint survival probability of the integrated intensities. Also, I give some numerical results and characterizes the CDS rate behavior in terms of the involved parameters.
Hengrui Luo
Assistant Professor, Department of Statistics, Rice University
Title: Modern Statistical Methods for Tensors: Compression, Communication, and Completion
Abstract:
The talk will develop a common statistical perspective on several recent works on tensor-structured problems, with an emphasis on how preserving multiway structure leads to phenomena and methods that are fundamentally different from treating the data as a vector. Modern applications, including compression and distributed learning for large language model weight tensors, will also be discussed.
Time: 4-5 PM
Noel Cressie
Noel Cressie, Distinguished Professor , School of Mathematics and Physics, University of Wollongong, Australia
Title: NASA/ESA's Mars Sample Return (MSR) mission: Bayesian statistics for planetary protection
Abstract: The Mars Sample Return (MSR) mission of NASA and ESA (European Space Agency) involves the collection and return to Earth circa 2040 of samples from the surface and atmosphere of Mars. A “Tiger Team” of scientists, of which I was a member, released a report in 2024 on planetary protection from potential biohazards in the returned samples. This lecture develops the Bayesian statistical design used to determine how many subsamples will be needed for biohazard testing. There are no data available since the samples are still inside sample tubes on Mars. However, there is prior astrobiological knowledge, and a Bayesian statistical approach can direct how that knowledge can be used. The number of subsamples is seen to depend on a prespecified risk-tolerance threshold (e.g., one in a million) and three other quantities that originate from the Bayesian design. Monte Carlo simulation is used in a digital twin to compute the number of subsamples needed for biohazard testing of each sample tube returned from Mars.
Time: 11:00 - 12:00 PM
Debangan Dey
Assistant Professor, Department of Statistics, Texas A&M University
Title: Bigraphical Matérn–Whittle (BMW) Processes for Large Multivariate Spatial Data on General Domains
Abstract: Large spatial datasets now record many correlated variables at thousands of locations, often on domains where Euclidean distance misrepresents proximity. The central difficulty is modeling cross-variable dependence jointly while retaining variable-level interpretation. We introduce the bigraphical Matérn–Whittle (BMW) process, a multivariate Gaussian process that addresses this challenge using two graphs. A spatial graph generates the Matérn structure of each variable through a fractional power of a graph Laplacian, making the process valid on any topology while allowing variable-specific range, smoothness, and amplitude parameters. A directed acyclic variable graph encodes the scientific structure; we prove that each absent edge implies an exact conditional independence between the corresponding fields. We further show that the operator determinant does not involve the cross-dependence coefficients, keeping matrix-free likelihood evaluation and Bayesian learning of the variable graph computationally tractable at scale. Estimation requires only sparse matrix–vector products and scales to tens of millions of space–variable pairs. In simulations, the method accurately recovered parameters and graphs, remained robust under model misspecification, and reduced held-out prediction error by half on a non-convex domain. In an analysis of a spatial transcriptomics section containing 19,809 cells and 1,122 genes, fitted in approximately one hour on a laptop, borrowing information across the learned gene graph reduced held-out prediction error by 50% to 91%. We also explore theoretical challenges, including the achievable efficiency of estimating the nugget variance.
Cindy Zhang
Assistant Professor, Department of Statistics, Rice University
Title: Joint Modeling of Biomarker Dynamics with a Dual Sigmoid Curve in an Unobserved Disease progression: Application to Alzheimer's Disease
Abstract: Modeling the temporal evolution of biomarkers is critical in Alzheimer's disease (AD) research, particularly during the preclinical stages, when subtle biological changes may emerge years before clinical symptoms appear. This period provides a crucial window for early intervention. However, characterizing biomarker trajectories is challenging because disease progression is not directly observable and biomarker dynamics are often heterogeneous, nonlinear, and asymmetric over time. To address these challenges, we propose a latent variable model that reflects the biological asymmetry of biomarker change across the early and late stages of the disease. This formulation jointly models multimodal longitudinal biomarker trajectories while capturing the unobserved disease progression as a latent process, yielding interpretable parameters for biomarker ordering. Estimation proceeds via a quasi-likelihood approach and asymptotic properties of the resulting estimators are established. In addition, we explicitly account for real-world constraints commonly encountered in AD studies, such as limited sample sizes, missing and infrequent measurements, and partially observed trajectories due to the decades-long disease course. We illustrate our method using two longitudinal AD cohorts: BIOCARD and ADNI.
Gao Wang
Assistant Professor of Neurological Sciences and Biostatistics, Columbia University Vagelos College of Physicians and Surgeons
Title: Extending SuSiE for Fine-Mapping Molecular QTLs at Scale with Applications in Alzheimer's Disease Genetics
Abstract:
SuSiE (Sum of Single Effects) is a Bayesian fine-mapping framework that has been broadly adopted and extended by many research groups for polygenic prediction, cross-ancestry GWAS and eQTL analysis, causal gene identification in transcriptome-wide association, and other related applications. In this talk, I present the core ideas of a series of methodological extensions motivated by our own analysis of multi-omic data in aging brains through the FunGen-xQTL project within the Alzheimer's Disease Sequencing Project Functional Genomics Consortium. These include fine-mapping of epigenomic QTL via function-on-scalar regression (fSuSiE), joint modeling of chromatin accessibility and RNA programs (mfSuSiE), multi-gene eQTL fine-mapping (an mvSuSiE extension), and a multi-trait colocalization method inspired by SuSiE (ColocBoost). We also address practical challenges including fine-mapping under highly polygenic architectures (SuSiE-ASH) and diagnosing LD and z-score mismatch in GWAS summary statistics (SuSiE-RSS-QC). Applying these methods to integrate chromatin, gene expression, splicing, protein abundance, and metabolomic data across aging cohorts, we characterize regulatory cascades at AD-associated loci with cell-type resolution and identify mechanisms including oligodendrocyte excitotoxicity, primary cilia dysfunction, and mitophagy suppression. Finally, we discuss intrinsic limitations of fine-mapping that persist beyond statistical methodology and their implications for study design and functional follow-up.