Problem
Researchers and businesses often work with datasets that have hundreds or thousands of variables. Principal component analysis (PCA) is a standard technique for reducing that complexity: it finds a small number of patterns, called principal components (PCs), that together explain most of the variation in the data. But standard PCA produces patterns that use every variable at once, making them hard to interpret. Sparse PCA fixes this by limiting each pattern to a small number of variables. The problem is that existing methods for finding multiple sparse patterns at the same time cannot guarantee that the patterns they return are mathematically valid, or anywhere near the best possible answer.
Science
We developed three algorithms that, for the first time, find multiple sparse and mutually independent patterns in a dataset and prove how close those patterns are to the best possible solution. We tested our methods on eleven real-world datasets from the UCI database, ranging from 13 to 1,300 variables, as well as on synthetic data. On the pitprops dataset, a standard benchmark used across the field, previous methods required around 45 to 60 variables to explain less than 70 per cent of the correlation in the data. We solved the same problem with 24 variables and explained 81 per cent of the correlation.
Outcome
Our three algorithms collectively solve sparse PCA problems with multiple patterns to within 3 per cent of the best possible answer, for datasets with up to 1,000 variables, in minutes. Existing methods routinely return patterns that violate the mathematical requirement that patterns must be independent of each other, even when the data were generated from patterns that are independent. Our methods never do. On the pitprops dataset, we also proved that the solutions previously thought to be unachievable at low sparsity levels are in fact achievable, and can be computed in seconds.
Change
Our research gives analysts in finance, medicine and any other field that relies on large datasets a set of tools that are both faster and more trustworthy than what existed before. In finance, for example, each pattern can be interpreted as a portfolio of stocks in a particular industry sector: getting those patterns right matters. Our code is publicly available at https://cran.r-project.org/web/packages/msPCA/index.html . The research also calls on the field to move away from deflation methods, which compute patterns one at a time and cannot guarantee independence, toward approaches that optimise all patterns simultaneously.
Ryan Cory-Wright, ³Ô¹ÏºÚÁÏ Business School; Jean Pauphilet, London Business School, 'Sparse PCA with multiple components', Operations Research, 25 June 2026