Noise-Adaptive DAG Learning: New Tutorial Boosts Causal Discovery Robustness

Directed acyclic graphs (DAGs) are fundamental for causal reasoning, but inferring them from observational data has long been hampered by combinatorial search over acyclicity constraints and sensitivity to heteroscedastic noise. This tutorial, submitted to IEEE Signal Processing Magazine's special issue on causal inference, reviews recent signal processing and optimization advances that reframe structure learning as a continuous score-based problem over adjacency matrices. By leveraging smooth characterizations of acyclicity (e.g., NOTEARS-type regularizers), the method scales to hundreds of variables.

The core innovation is "concomitant estimation": jointly learning sparse causal graphs and the variances of exogenous noise terms. This makes the estimator noise-adaptive, handling heteroscedasticity and distribution shifts without prior knowledge of noise levels. The tutorial covers structural equation models, historical scoring methods, and emerging directions like online learning, nonlinear DAGs, and neural causal discovery. For practitioners, this means more reliable causal inference in domains like genomics, finance, and climate science where noise heterogeneity is the norm.