Functional Causal Models Overview

• Functional causal models are a deterministic framework where each variable is defined as a function of its direct causes and an independent noise term.
• They enable identifiability analysis through methods like sink elimination and nonparametric estimators, thereby allowing full causal graph recovery under certain conditions.
• Practical implementations using additive noise, Bayesian networks, and deep generative models highlight FCMs’ applicability in fields such as neuroimaging, climate science, and dynamic systems.

A functional causal model (FCM) is a structural approach to modeling causal systems in which each variable is a deterministic function of its direct causes and an independent exogenous noise term, often within a directed (possibly cyclic) graphical structure. FCMs unify frameworks including structural equation models, additive noise models, functional data models, and recent operator- and network-based generalizations. The FCM paradigm enables precise representation, identifiability analysis, and algorithmic discovery of causal relations in both finite and infinite-dimensional (functional) domains.

1. Formal Structure and Mathematical Foundations

Let $V = \{1, \dots, d\}$ index observed variables $X_1, \dots, X_d$. An FCM comprises

• a graph $G = (V, E)$ encoding direct functional dependencies,
• a set of mutually independent exogenous noises $\{E_i\}$,
• deterministic functions $f_i$ mapping parental variables and their noise to each $X_i$.

Abstractly, the model specification is

$$X_i = f_i( X_{\mathrm{Pa}_i}, E_i ) \qquad i=1, ..., d$$

where $\mathrm{Pa}_i$ denotes the parents of node $i$. This structure generates a unique joint distribution: $$P(X_1, ..., X_d) = \int \prod_{i=1}^d \delta\big( x_i - f_i(x_{\mathrm{Pa}_i}, e_i) \big) \prod_{i=1}^d dP_{E_i}(e_i)$$ which factorizes as

$$P(X) = \prod_{i=1}^d P( X_i \mid X_{\mathrm{Pa}_i} )$$

if the graph is acyclic and the $E_i$ are independent (Goudet et al., 2017).

This formalism extends naturally to random functions. For example, each $X_i$ may be a function $f_i \in \mathcal{H}_i$, where functional dependencies and exogenous processes are suitably defined over Hilbert or Banach spaces (Yang et al., 2024, Zhou et al., 2022, Roy et al., 2023).

2. Identifiability and Functional Model Classes

Identifiability in the FCM context centers on conditions under which the full causal graph (not only the equivalence class) can be recovered from observed data. Identifiable Functional Model Classes (IFMOCs) provide precise criteria (Peters et al., 2012):

• Bivariate identifiability: In the canonical two-variable functional model $Y = f(X, N)$ with $N \perp X$, reverse functional representations in the model class are generically impossible except in degenerate cases (e.g., joint Gaussianity in linear models).
• Multivariate lift: For a DAG-aligned FCM, each function $f_i$ must retain bivariate identifiability under arbitrary fixing of non-descendant variables. This allows for full-graph identifiability if the model class is “rich enough”—e.g., nonlinear additive noise models or post-nonlinear models with non-Gaussian noise (Peters et al., 2012, Yang et al., 2024, Zhou et al., 2022).

Causal discovery under IFMOC proceeds by sink elimination: recursively identify nodes whose residuals (from regressing on putative parents) are independent of all others, iterating to uniquely uncover the graph. When the true data-generating process lies outside any IFMOC, such methods terminate safely (no output or “I do not know”) rather than yielding incoherent conclusions (Peters et al., 2012).

3. Algorithmic Discovery and Representational Approaches

Standard FCM and Specializations

• Additive noise models (ANM): $Y = g(X) + E$, $E \perp X$—central for causal directionality as identifiability results hold unless $g$ is linear and $E$ Gaussian (Tu et al., 2022).
• Post-nonlinear models (PNL): $Y = h( g(X) + E )$ with invertible $h$.
• Linear non-Gaussian models (LiNGAM): $X = BX + E$ where $B$ is lower triangular, $E$ non-Gaussian—full identifiability by the Darmois–Skitovich theorem (Yang et al., 2024).

Operator and Functional Data Models

• Functional LiNGAM for random functions: $f_i = \sum_{j<i} T_{ij} f_j + h_i$, with $T_{ij}$ bounded operators, $h_i$ mutually independent non-Gaussian random elements in infinite-dimensional Hilbert spaces (Yang et al., 2024). Identifiability holds for generic operator coefficients and non-Gaussian noise.

Embedding and Expansion

• Basis expansions: Random functions are represented as $f_i(t) \approx \sum_{k=1}^M \xi_{ik} \phi_{ik}(t)$; model estimation is performed on the principal component scores, enabling practical implementation and straightforward incorporation of penalization for sparsity (Yang et al., 2024, Roy et al., 2023, Zhou et al., 2022).

Bayesian and Score-Based Learning

• Functional Bayesian networks: Causal graphs on $p$ functional variables, structural equations on basis coefficients with mixtures of Gaussians noise, recovered via MCMC and spike-and-slab priors; identifiability from non-Gaussian error mixtures (Zhou et al., 2022).
• Expectation-Maximization for multivariate functional DAGs: Block bilinear operators describe func2func regression and DAG structure enforced by algebraic acyclicity, solved by penalized EM with group-lasso and Lagrange multipliers (Lan et al., 2024).
• Dynamical and optimal-transport methods: Bivariate causal direction identified via zero-divergence flows in the OT formulation of FCM under ANM/PNL constraints (Tu et al., 2022).

Deep Generative Approaches

• Causal Generative Neural Networks: Parameterize each $f_i$ as a neural network; train via Maximum Mean Discrepancy (MMD) to minimize distance between observed and generated distributions; under infinite data and sufficient model richness, approach the true FCM distribution (Goudet et al., 2017).

4. Extensions: Functional Data, Cyclicity, Laws, and Abstraction

Functional Data Causal Modeling

FCMs generalize naturally to settings with functional (curve-valued) treatments, mediators, or outcomes. Models include functional linear structural equations and bilinear regression operators, with applicability to fMRI, EEG, and other domains (Yang et al., 2024, Zhou et al., 2022, Roy et al., 2023, Zhao et al., 2018, Gao et al., 2023).

Cyclic and Equilibrium Models

For directed graphs with cycles, unique probabilistic assignment requires more sophisticated constructions. Averagely uniquely solvable cyclic FCMs admit Markov factorization, and the $p$-separation generalizes $d$-separation to cyclic fCMs; these advances resolve prior obstacles to assigning distributions and conditional independencies in feedback models (Ferradini et al., 6 Feb 2025).

Causal Constraints Models

SCMs and FCMs are special cases of causal constraints models (CCMs), which encode algebraic relations (e.g., conservation laws, equilibrium equations) invariant under specified sets of interventions. CCMs naturally model dynamical systems at equilibrium and encode functional laws (e.g., PV = Nk_BT) unavailable to any SCM (Blom et al., 2018).

Abstraction and Multi-Level Models

Functional and graphical abstractions (such as $\alpha$-abstractions, $\tau$-abstractions, Cluster DAGs, and Partial Cluster DAGs) formally relate fine-grained FCMs to coarse-grained models, providing a rigorous bridge for transfer of theoretical and algorithmic results between levels of granularity (Schooltink et al., 2024).

Category-Theoretic and Diagrammatic Reformulation

String diagrams in symmetric monoidal categories provide a formal, compositional framework unifying Bayesian networks, SCMs, and general functional causal models, with diagrammatic representations for intervention, conditioning, counterfactuals, and identifiability (Lorenz et al., 2023).

5. Practical Estimation and Robustness

Kernel and Nonparametric Estimation

• Kernel-based functional causal estimators: Use operator-valued kernels for functional potential outcome embedding; Fréchet means and kernel ridge regression enable nonparametric estimation of average treatment effects and dynamic effects for function-valued outcomes under minimal model assumptions (Raykov et al., 6 Mar 2025).
• Handling time-to-event and censored data: Functional accelerated failure time models with regression adjustment, inverse probability weighting, and doubly robust estimators have been developed for causal inference with functional predictors and censored outcomes (Gao et al., 2023).

Handling Unmeasured Confounding and Spatio-Temporal Dynamics

• Partial Functional Dynamic Backdoor Diffusion-based Models (PFD-BDCM): Integrate basis expansion for functional data, valid backdoor sets, and diffusion-based generative modeling to support causal inference under unmeasured confounders with spatial and temporal dependency, establishing error bounds on counterfactual queries (Liu et al., 30 Aug 2025).

6. Empirical and Application Highlights

Empirical studies have established the superiority of properly specified FCM-based discovery methods over faithfulness-based CI approaches for full graph recovery when the underlying data-generating process meets the identifiability assumptions (Peters et al., 2012, Zhou et al., 2022, Yang et al., 2024, Lan et al., 2024). Applications span brain connectivity (fMRI, EEG), clinical time-to-event analysis, climate and traffic networks, and air pollution counterfactual estimation (Roy et al., 2023, Zhou et al., 2022, Gao et al., 2023, Liu et al., 30 Aug 2025, Lan et al., 2024). Doubly robust and kernel-based estimators show enhanced statistical properties and interpretability for dynamic and non-linear causal effects in functional data (Raykov et al., 6 Mar 2025).

7. Limitations and Directions for Future Research

• Model class misspecification: FCM-based discovery is reliable only if the correct functional form is present; otherwise, procedures may be undecided or mislead (Peters et al., 2012).
• Cyclic and equilibrium phenomena: Graphical semantics become complex in feedback or equilibrium settings, requiring richer language via CCMs or generalizations of $d$-separation (Ferradini et al., 6 Feb 2025, Blom et al., 2018).
• Algorithmic scalability: Gröbner-basis methods and high-dimensional nonparametrics can be computationally intensive; low-rank and sparse methods show promise (Lee et al., 2015, Raykov et al., 6 Mar 2025, Lan et al., 2024).
• Unmeasured confounding: Advanced methods employing backdoor adjustment, operator-valued kernels, and spatio-temporal models are active research frontiers (Liu et al., 30 Aug 2025, Raykov et al., 6 Mar 2025).
• Multi-resolution and abstraction: Formal alignment between functional and graphical abstraction supports cross-resolution modeling, but practical synthesis and identifiability analysis remain challenging (Schooltink et al., 2024).
• Diagrammatic reasoning: String-diagram and categorical formulations offer foundational clarity and compositionality, particularly for counterfactuals and multi-world models (Lorenz et al., 2023).

Functional causal models constitute a mathematically rigorous and algorithmically fertile framework for causal representation, discovery, and effect estimation across domains ranging from finite discrete systems to infinite-dimensional functional data, with powerful extensions for dynamic, cyclic, and abstracted systems.