Short, plain-language notes on each of my papers — what problem it solves and why it matters.
May 2025 · Multi-source learning
Finding structure every data source agrees on: StablePCAPooling datasets and running PCA lets the biggest, noisiest batch write the principal components. StablePCA maximizes the worst-case explained variance across every mixture of sources — and a convex relaxation solves it about 39× faster than semidefinite programming.
Jul 2025 · Domain adaptation
Classifying reliably in a domain you've never labelled: CG-DROWith labels from several source domains and none from the target, the worst-case classifier is not asymptotically normal — so both the bootstrap and normal intervals break. A perturbation scheme restores uniformly valid confidence intervals in about twenty seconds.
Dec 2024 · Causal inference
Making invariance-based causal discovery actually scalableInvariance is a beautiful idea for finding causal predictors, but testing every subset of variables blows up exponentially. Letting the environment weights go negative turns the search into one continuous problem — nonconvex, yet every stationary point lands near the truth.
2026 · The Annals of Statistics
Learning models that hold up on a population you never labelledA model fit to pooled multi-source data quietly optimizes for the majority. Maximizing the worst-case reward over every mixture of the sources collapses to a weighted average of per-source models, with weights from a tiny quadratic program.
The R Journal · Software
Confidence intervals in high dimensions, in one R package: SIHRLasso-type estimators are biased, so their outputs don't come with valid confidence intervals. SIHR turns a decade of debiasing methodology into five functions, for linear and logistic models alike — debiasing once per target rather than once per coefficient.
© 2026 Zhenyu (Zach) Wang 王振宇