Structural Causal Bottleneck Models: Targeted Dimension Reduction for Causal Inference

High-dimensional causal inference faces a fundamental tension. On one hand, real-world phenomena from neural population dynamics to climate teleconnections naturally involve high-dimensional random vectors interacting across complex networks. On the other hand, estimating causal effects in such spaces rapidly succumbs to the curse of dimensionality, requiring sample sizes that grow exponentially with dimension. In their recent work, Structural Causal Bottleneck Models, Simon Bing, Jonas Wahl, and Jakob Runge propose an elegant resolution: assume that causal effects depend not on the full high-dimensional parents, but on low-dimensional summaries, or bottlenecks. This assumption, far from being merely a computational convenience, reflects a physical reality in many scientific domains where macroscopic properties drive microscopic outcomes.

The Architecture of Compression

An SCBM modifies the standard structural causal model (SCM) framework by inserting bottleneck functions between parents and children. Formally, instead of the standard structural assignment (X_j := m_j(X_{pa(j)}, \eta_j)), an SCBM defines (X_j := f_j(b_j(X_{pa(j)}), \eta_j)), where (b_j) maps the product of parent spaces to a lower-dimensional bottleneck space (Z_j). The bottleneck function (b_j) is assumed surjective, ensuring that the bottleneck variables (Z_j) are well-defined latent summaries that capture all information from the parents relevant to the child.

In the linear case, this structure implies that the mechanism matrices (A^j_i) between parent (X_i) and child (X_j) are low-rank. This is a substantial departure from standard high-dimensional causal inference, which typically relies on sparsity assumptions or additive models. The low-rank assumption captures a distinct physical intuition: a downstream variable depends only on specific emergent properties of its causes, not their full microstate. For instance, rainfall patterns over West Africa may depend only on the phase of the El Niño Southern Oscillation (ENSO), a low-dimensional summary of Pacific temperature distributions, rather than the full oceanic temperature field.

The authors draw a formal connection to the Information Bottleneck method of Tishby and Zaslavsky (2015), though SCBMs differ crucially in their causal interpretation. While the Information Bottleneck optimizes a tradeoff between compression and prediction, SCBMs enforce a structural constraint on the data generating process. The bottleneck is not merely a convenient encoding for statistical efficiency; it is posited as a property of the causal mechanism itself. This distinction matters because it shifts the burden of proof: rather than asking whether we can compress the data without losing predictive power, we ask whether the true causal process admits a low-dimensional sufficient statistic.

Identifiability and Estimation in Low-Sample Regimes

A central theoretical contribution concerns identifiability: to what extent can we recover the bottleneck variables (Z_j) from observational data? The authors establish conditions under which