Technical Summary: Causal Density Functions

Problem Statement

Causal inference traditionally focuses on comparing whole distributions after "graph surgery" (e.g., edge deletion) or estimating global causal strength measures like the Kullback-Leibler (KL) divergence between observational and interventional models. While effective for global comparisons, these approaches often lack a local, differentiable object that describes how probability mass is transported from an observational regime to an interventional one at a pointwise level. Furthermore, existing methods often rely on discrete graph operations or global optimization objectives that may not be compatible with continuous, soft interventions or gradient-based learning in complex, multi-regime settings.

The paper posits that the fundamental analytic question in causal inference is: What is the density-like object that transports observational probability mass into interventional probability mass?

Methodology: Causal Density Functions and Kan–Do Calculus

The core proposal is the definition of Causal Density Functions, which are Radon–Nikodym (RN) derivatives comparing interventional laws to observational laws.

1. Definition and Calibration

For an intervention on a variable $X_i$, the causal density function $\rho_i(z)$ is defined as:
$$\rho_i(z) = \frac{dP_{do(X_i)}}{dP_{obs}}(z)$$
provided the interventional law $P_{do(X_i)}$ is absolutely continuous with respect to the observational law $P_{obs}$.

This definition yields a fundamental calibration identity (the "RN calibration identity"):
$$E_{do(X_i)}[f(Y)] = E_{obs}[f(Y)\rho_i(Z)]$$
This identity serves as both a theoretical definition and an empirical diagnostic. If an estimated density ratio $\hat{\rho}$ is correct, reweighting observational expectations by $\hat{\rho}$ must reproduce interventional expectations.

2. Categorical Semantics (Kan–Do Calculus)

The paper provides a categorical foundation for these concepts using Kan extensions:

• Conditioning is characterized as a Right Kan extension (a conservative, limit-preserving operation).
• Intervention is characterized as a Left Kan extension (a generative, colimit-like pushforward).
• The Radon–Nikodym derivative emerges as the unique natural transformation mediating the adjoint triple of these extensions.

This framework, termed Kan–Do Calculus, unifies conditioning and intervention as compatible universal operations. It generalizes previous "Judo Calculus" by explicitly incorporating measure-theoretic RN derivatives and establishing a duality where the RN density is the canonical differential between observational and interventional measures.

3. Estimation and Scoring

The authors propose a practical pipeline for estimating these densities and using them for causal discovery:

• Density Estimation: Using normalizing flows or density-ratio models to estimate $P_{obs}$ and $P_{do(X_i)}$ on a shared support, computing $\hat{\rho}(z) = \exp(\log \hat{p}_{do}(z) - \log \hat{p}_{obs}(z))$.
• Calibration: Validating estimates by measuring the moment gap $\Delta_f = |E_{obs}[f(Z)\hat{\rho}(Z)] - E_{do}[f(Z)]|$. Large gaps indicate poor regime overlap or estimation failure.
• Do-curves: Estimating response curves $x \mapsto E[Y | do(X_i=x)]$ by reweighting observational samples.
• Directed Edge Scoring: A pairwise score $s_{ij}$ is defined to rank candidate causes $X_i \to X_j$ by the residual variance of $X_j$ after weighting by $\rho_i$:
  $$s_{ij} = E_{obs}\left[ \frac{\hat{\rho}_i(Z)(X_j - \hat{E}[X_j|X_i])^2}{\text{Var}(X_j)} \right]$$
  Lower scores indicate stronger causal support.
• Sheaf Coherence: To handle multiple regimes, the method incorporates a stability penalty based on Maximum Mean Discrepancy (MMD) across environments, ensuring the estimated structure is coherent across different perturbation regimes.

Key Contributions

1. Definition of Causal Density Functions: The paper defines causal effects locally via RN ratios ($dP_{do}/dP_{obs}$), distinguishing them from global causal-strength measures based on graph deletion or whole-distribution divergence.
2. Estimators for Calibration and Scoring: It derives practical estimators for causal-density calibration, do-curves, and directed edge scores, emphasizing that the primary estimand is the density ratio itself, not just a downstream graph.
3. RN–Kan Semantics: It provides a compact categorical semantics explaining the connection between conditioning, intervention, and change of measure via Kan extensions, without making category theory the sole empirical claim.
4. Empirical Evaluation: The framework is evaluated on synthetic systems with known ground truth and real perturbation datasets (Sachs protein signaling, LINCS, PISA), analyzing both successful recovery and failure modes under weak regime overlap.

Experimental Results

The paper evaluates the estimators across several domains:

• Synthetic Gaussian Systems: Causal density estimates are nearly exact, with RN calibration gaps close to zero ($\approx 0.001$), confirming the theoretical compatibility of change-of-measure and Kan extensions.
• Sachs Protein Signaling (S9): The method recovers biologically meaningful directions (e.g., PKC→P38, Raf→MEK) with small calibration errors ($\approx 0.01$). While the pairwise scorer alone yields a dense graph (SHD=30), adding sheaf-coherence penalties and multivariate regression improves the structure to SHD=16.
• LINCS L1000 Panel: In this highly perturbed biological setting, calibration gaps increase ($\approx 0.12$), reflecting strong regime heterogeneity. However, the method still identifies sparse, stable driver genes (e.g., HSPA8→CDC25B) and produces smooth do-curves.
• PISA 2022 Socio-Economic Data: Used as a stress test for cross-country regimes. While plausible socio-economic structures are recovered, the method exhibits large calibration gaps ($\Delta \approx 4.0$) due to extreme cross-country heterogeneity, demonstrating the intended failure mode when regime overlap is poor.
• Multi-Regime Chain: In a synthetic chain $X_1 \to \dots \to X_8$ with soft interventions, the density field correctly identifies directional bands, though parent selection remains a limiting factor for perfect graph recovery (SHD=8, F1=0.33).

Significance and Claims

The paper claims that causal density functions offer a more granular, differentiable, and locally interpretable alternative to global causal strength measures. By framing interventions as Left Kan extensions and conditioning as Right Kan extensions, the work provides a universal mathematical semantics for causal inference that naturally accommodates soft interventions and continuous parameters.

The authors emphasize that the calibration identity is the most critical diagnostic: it allows researchers to verify whether an estimated density ratio actually transports observational expectations to interventional ones. The paper modestly notes that while the pairwise edge scorer is not yet competitive with mature graph-optimization methods like NOTEARS or DCDI for pure structure recovery, the primary contribution is the estimation and calibration of the causal density itself. The framework is presented as a step toward scalable, multi-regime causal discovery where local density ratios are glued together via sheaf coherence to form globally consistent causal models.

The work avoids claiming to solve all causal discovery problems, instead highlighting that the method is most reliable under good regime overlap and degrades predictably as cross-regime divergence grows, a behavior explicitly captured by the RN calibration gaps.