Further reading

The classical causal-discovery background is given by Spirtes et al. [ 2000 ] . Greedy equivalence search [ Chickering , 2002 ] , FCI-style discovery with latent and selection variables [ Colombo et al. , 2012 ] , and interventional equivalence classes [ Hauser and Bühlmann , 2012 ] provide important reference points for what can be identified under different regimes. Soft interventions with unknown targets are treated by Jaber et al. [ 2020 ] .

Differentiable structure learning includes NOTEARS [ Zheng et al. , 2018 ] , gradient-based neural DAG learning [ Lachapelle et al. , 2020 ] , and interventional extensions [ Brouillard et al. , 2020 ] . For latent confounding, compare non-Gaussian and higher-order approaches [ Cai et al. , 2023 , Morinishi and Shimizu , 2025 , Tramontano et al. , 2025 ] with the recent RelFCI treatment of relational data [ Negro et al. , 2025 ] . The Sachs protein-signaling study [ Sachs et al. , 2005 ] remains a useful empirical touchstone, but its interventions, preprocessing, and target graph must be read carefully. BRIDGE/SKFM belongs beside these methods as a geometric screen and field-estimation layer, not as a replacement for their identification theory.