Abstract
The data drawn from biological, economic, and social systems are oftenconfounded due to the presence of unmeasured variables. Prior work in causaldiscovery has focused on discrete search procedures for selecting acyclicdirected mixed graphs (ADMGs), specifically ancestral ADMGs, that encodeordinary conditional independence constraints among the observed variables ofthe system. However, confounded systems also exhibit more general equalityrestrictions that cannot be represented via these graphs, placing a limit onthe kinds of structures that can be learned using ancestral ADMGs. In thiswork, we derive differentiable algebraic constraints that fully characterizethe space of ancestral ADMGs, as well as more general classes of ADMGs, aridADMGs and bow-free ADMGs, that capture all equality restrictions on theobserved variables. We use these constraints to cast causal discovery as acontinuous optimization problem and design differentiable procedures to findthe best fitting ADMG when the data comes from a confounded linear system ofequations with correlated errors. We demonstrate the efficacy of our methodthrough simulations and application to a protein expression dataset.