Structural Causal Bottleneck Models (SCBMs)

• Structural Causal Bottleneck Models (SCBMs) are defined by causal DAGs where high-dimensional effects are transmitted solely via low-dimensional bottlenecks.
• They integrate statistical dimension reduction with formal causal guarantees, leading to improved identifiability, interpretability, and out-of-distribution reliability.
• Empirical studies validate SCBMs by accurately recovering true bottleneck variables and reducing estimation errors, especially in transfer learning and remote sensing applications.

Structural Causal Bottleneck Models (SCBMs) are a principled class of structural causal models in which the causal effects between high-dimensional variables are mediated exclusively through low-dimensional summary statistics, termed “bottlenecks.” SCBMs generalize both concept bottleneck models and causal information bottleneck approaches by combining the expressivity of statistical dimension reduction with formal guarantees rooted in causal inference. Through the explicit structure imposed on the underlying directed acyclic graph (DAG), SCBMs provide both a practical estimation framework and improved interpretability, identifiability, and out-of-distribution reliability in high-dimensional causal systems (Bing et al., 9 Mar 2026).

1. Formal Definition and Structural Equations

Let $G = (V,E)$ be a directed acyclic graph describing the system’s causal structure. Each node $i \in V$ is associated with a high-dimensional observed variable $X_i \in \mathcal{X}_i$ and an exogenous noise $\eta_i \in \mathcal{H}_i$. For every edge $(i \to j) \in E$, a low-dimensional bottleneck space $\mathcal{Z}_{(i,j)}$ is introduced. The SCBM is parameterized by two function families:

• Bottleneck maps $b_{(i,j)}: \mathcal{X}_i \to \mathcal{Z}_{(i,j)}$
• Effect maps $f_{(i,j)}: \mathcal{Z}_{(i,j)} \to \mathcal{X}_j$

The factored additive structural equations for each endogenous node $j$ are:

$$Z_{(i,j)} = b_{(i,j)}(X_i) \quad \text{for each } i \in \mathrm{pa}(j)$$

$$X_j = \sum_{i\in\mathrm{pa}(j)} f_{(i,j)}(Z_{(i,j)}) + \eta_j$$

Alternatively, gathering bottleneck variables $Z_j := (Z_{(i,j)})_{i\in\mathrm{pa}(j)}$,

$$Z_j = b_j((X_i)_{i\in\mathrm{pa}(j)}), \quad X_j = f_j(Z_j, \eta_j)$$

The key assumption is that all causal influence traverses the low-dimensional bottleneck(s), allowing for dimension reduction while preserving the relevant mechanistic dependencies (Bing et al., 9 Mar 2026).

2. Identifiability and Structural Equivalence

SCBMs exhibit attractive identifiability properties under mild assumptions. If each effect map $f_j$ is additive in noise and almost-surely injective in the bottleneck, then the bottleneck maps $\{b_{(i,j)}\}$ are identifiable up to invertible reparameterization—specifically, up to bijections on the bottleneck spaces. Two SCBMs are structurally equivalent if they induce the same observational and interventional distributions; this equivalence class is indexed precisely by invertible transformations of the bottleneck variables (Bing et al., 9 Mar 2026). Failure to specify a sufficiently high bottleneck dimension harms identifiability, while over-specification simply introduces slack variables without affecting the minimal sufficient summary property.

3. Connection to Information Bottleneck and Causal Abstraction

SCBMs generalize the information bottleneck (IB) framework of Tishby & Zaslavsky, replacing the classical $X \to T \to Y$ paradigm with explicit causal DAGs and edgewise compressions. For each node, the desiderata become:

• Minimal compression: $\min I(X_i;Z_i|Z_{\mathrm{pa}(i)})$
• Causal sufficiency: $$I(X_{\mathrm{ch}(i)}; X_i | Z_i, Z_{\mathrm{pa}(i)}) = 0$$

A nodewise Lagrangian trades off compression against the sufficiency for modeling children, allowing for local sequential optimization in a causal order. Unlike purely statistical IB or classical representation learning, SCBMs ensure that the learned abstractions are interventionally faithful and sufficient for causal reasoning. They provide a formal alternative to causal representation learning, which typically operates over unknown graphs and latent variables, and to causal abstraction/feature learning, which relies on global or clusterwise summarization (Simoes et al., 2024, Bing et al., 9 Mar 2026).

4. Estimation Procedure and Algorithmic Methods

Estimation in SCBMs proceeds edgewise, using only observational data and a known DAG $G$. For each edge $(i \to j)$:

• Fit a regressor $m_{(i,j)}$ from $X_i$ to $X_j$, conditioning on a valid adjustment set (e.g., $\mathrm{pa}(i) \cup \mathrm{pa}(j)$).
• Linear case: Factor the regression weight matrix $M$ (of rank equal to $\mathrm{dim}\,\mathcal{Z}_{(i,j)}$) using SVD or independent column selection, yielding a map $B$ for bottleneck extraction and $F$ as a mapping to the target.
• Nonlinear case: Jointly train encoder–decoder networks $\hat{b}_{(i,j)},\hat{f}_{(i,j)}$ to minimize predictive loss, typically using backpropagation and reconstruction loss.

Sequentially traversing a topological order, one recursively fits and factorizes all edgewise mechanisms, blocking confounding via adjusted conditioning. Care must be taken to prevent rank collapse and to select model classes appropriate to the presumed causal mechanism (linear vs. nonlinear) (Bing et al., 9 Mar 2026).

5. Empirical Validation and Use Cases

Experimental evidence demonstrates that SCBMs accurately recover true low-dimensional bottleneck variables (up to bijection) in both linear (up to $|V|=100$, $d_X=100$, $d_Z=10$) and nonlinear settings, with mean $R^2 \approx 1$ given sufficient dimensionality and sample size. Underestimating bottleneck dimension sharply reduces recoverability metrics, while overestimation is benign. In transfer learning scenarios with limited labeled pairs but abundant partially observed data, conditioning on learned bottlenecks (rather than high-dimensional original covariates) substantially reduces effect-estimation error (Bing et al., 9 Mar 2026).

The Process-Guided Concept Bottleneck Model (PG-CBM) is a prominent SCBM instance developed for above-ground biomass prediction using remote sensing. Its architecture factorizes $p(X,C,Y) = p(X) p(C|X) p(Y|C)$, enforcing causal mediation and embedding domain constraints (monotonicity, non-negativity, spatial smoothness, bias mitigation) through dedicated penalty terms. Empirical comparisons show PG-CBM achieves lower RMSD, reduced structure-dependent bias, and improved out-of-distribution robustness relative to both vanilla CBMs and black-box deep learning models (Asiyabi et al., 15 Jan 2026).

6. Comparison to Related Methods

Framework	Bottleneck Summary	Graph Knowledge	Intervention Handling
Structural Causal Bottleneck	Edgewise, continuous/sur.	Known	Observational + DAG
Causal Representation Learning	Fully latent, invertible	Unknown	Not explicit
Causal Abstraction/Feature Learn.	Global/clusterwise	Varies	Indirect
Causal Information Bottleneck	Single pair, discrete	Known	Explicit, uses $p(Y|do(X))$

SCBMs occupy a unique position by providing local edgewise interventions, accommodating both continuous and discrete bottlenecks, requiring only observational data and DAG knowledge (but not interventional distributions), and permitting surjective (non-invertible) compression (Bing et al., 9 Mar 2026, Simoes et al., 2024).

7. Limitations and Open Research Questions

Current limitations of SCBMs include reliance on a known or pre-learned causal DAG, sensitivity to bottleneck dimension specification, and the potential for dependent noise terms in the induced SCM over bottlenecks, which may complicate identification via front-door adjustment. Error propagation over compositional edgewise estimation and scalable methods for joint graph and bottleneck discovery are open areas of investigation. Extending SCBMs to partially unknown graph settings and integrating domain process constraints remain active directions for research (Bing et al., 9 Mar 2026, Asiyabi et al., 15 Jan 2026).

• "Structural Causal Bottleneck Models" (Bing et al., 9 Mar 2026)
• "Process-Guided Concept Bottleneck Model" (Asiyabi et al., 15 Jan 2026)
• "The Causal Information Bottleneck and Optimal Causal Variable Abstractions" (Simoes et al., 2024)
• "Causally Reliable Concept Bottleneck Models" (Felice et al., 6 Mar 2025)