PCA can invent structure that doesn’t exist and miss what’s really there

patrickmineault · x · 2026-07-25

• A reposted commentary argues that dimensionality reduction can invent structure that isn’t really there and also miss structure that does exist.
• The attached PNAS commentary screenshot makes the same point about PCA: when data are Gaussian it can work well, but with non-Gaussian or highly structured data it may hide the true manifold.
• The example emphasizes that the “simplest explanation” is not always the best one, especially when fidelity to the data conflicts with human-friendly interpretability.
• It points viewers to an intro video with strategies for avoiding common mistakes when using dimensionality reduction methods.

Original post →

More from Research

• AI can scan more code than humans, but engineers still own quality — ingliguori · 2026-07-25
• A new AGI essay argues the field is climbing the same mountain from two slopes — op7418 · 2026-07-25
• Paper shows models implicitly optimize for hidden graders in audit-style tasks — Sauers_ · 2026-07-25
• DS-SAC paper compares tree-search robust optimization against RANSAC baselines — ducha_aiki · 2026-07-25
• DS-SAC proposes a deterministic geometric estimator that beats RANSAC on AUC — ducha_aiki · 2026-07-25
• New paper warns that LLM-generated covariates can break statistical inference — PtrPomorski · 2026-07-25

Research channel →