Source-linked AI summary
Symmetries and Causality: Causal Effect Identification Beyond IID Data
Martin Rabel, Jakob Runge
TL;DR
Causal machine-learning formalisms often struggle with non-IID data and complex queries. This paper introduces an abstract language based on symmetries and invariant mechanisms, together with formal identification machinery. It reproduces standard IID causal results while extending the framework toward broader queries and settings.
Problem
Existing causal formalisms often rely on near-IID setups and do-interventions, making complex real-world causal questions difficult to represent and answer.
Method
The paper models causal systems using symmetries and invariant probability-kernel mechanisms, then formalizes observations, queries, and graphical identification operations.
Results
The framework reproduces core IID causal structures and results, including c-components, hedges, mediation formulas, and mz-transport, while identifying the introduction's composed query.
Takeaways & Limitations
The abstraction provides a unified language intended to support causal reasoning for non-IID systems and complex queries beyond conventional do-interventions.
Takeaways & Limitations
The paper leaves completeness of its general identification strategy open and does not investigate discovery of the model structures.
Abstract
from arXiv · showhide
In the natural sciences, symmetries and cause-effect relationships are ubiquitous. Yet for complex machine-learning tasks, like world-modeling in reinforcement learning, they appear difficult to harness. We propose a formal description of statistical systems based on symmetries in data leaving causal mechanisms invariant. The result is an abstract, simple and general mathematical language for causal reasoning. This paper provides formal descriptions of models and queries, setting up this language, and the formal infrastructure and strategies for their mathematically rigorous identification from data within this formalism. This approach reproduces and matches standard theoretical results on IID data and transport of experimental and non-experimental data. But its main purpose is to unify and substantially extend the scope of causal reasoning, in going beyond IID data and in approaching complex causal queries not captured by do- or soft-interventions. This new perspective on causally relevant aspects of data-modeling additionally sheds new light on well-known structures like c-components or hedges but also includes aspects of missing data and is inherently well-suited for the description of transfer and robustness properties.
1 Introduction
The paper seeks a causal formalism that applies to complex, non-IID systems and flexible queries while retaining rigorous identification. It builds this abstraction around invariant mechanisms and aims to recover established causal results as special cases.
- Motivation: Existing causal formalisms capture important aspects of causality but often assume near-IID data and simple do-interventions, limiting complex real-world questions.The paper identifies both applicability and extraction of answers as central challenges.
- Need for Abstraction: Rather than adding definitional structures such as context variables or new intervention types, the approach removes constraints from models and queries to increase expressiveness.This lower-level language preserves ideas that can be compiled into a unified formalism, while making identification technology necessary.
- Need for Abstraction: Complex scientific and AI problems combine data from different environments, experiments, and temporal settings, making IID assumptions restrictive.The paper presents removing IIDness as a foundational abstraction for broader causal modeling.
- Employed Structure: Measure-theoretic modeling without primitive randomness provides a simple language intended to describe causal reasoning beyond IID data and conventional do-interventions.The authors state that the abstraction still reproduces SCM and potential-outcome results in the IID, do-interventional case.
- Employed Structure: The framework treats invariant mechanisms as probability kernels and explicitly represents symmetries, applicability regions, and shared models for observed and intervened worlds.Known kernels represent interventions, while unknown kernels can be extracted from data, including confounded structures.
- Contributions and Scope: The paper supplies formal machinery linking abstract problem statements to identification strategies, while structure discovery remains outside its scope.This separates flexible modeling and query formulation from the unresolved task of discovering the structures that define models.
2 Models and Observations
The paper models statistical systems through invariant probability-kernel mechanisms organized by symmetries and an acyclic dependency graph. It then connects these abstract models to observable random variables and well-defined distributions under finite-past assumptions.
- Preliminaries: The formalism begins with models, their relation to observations, observedness, and asymptotic limits, while assuming standard Borel measurable spaces.Standard Borel spaces include Euclidean spaces and finite discrete sets.
- Preliminaries: Symmetries are expressed through group actions, with invariance meaning a mapping is unchanged under the action and equivariance meaning it transforms compatibly.The paper represents a symmetry as an induced action of a subgroup.
- Models: A mechanism is a probability kernel from parent spaces to an output space, together with an applicability region, a symmetry acting freely and transitively there, and an equivariant mapping.Each model partitions the index set into disjoint applicability regions whose mechanisms are either observed-unknown or intervention-known.
- Examples: IID data, time-series, and experimental settings can be represented within the same model language through suitable symmetries and mechanism specifications.The examples include permutation symmetries for IID samples and time-translation symmetries for time-series.
- Models: Models use an acyclic I-graph and finite modeled past to compose mechanisms in causal order and define a joint distribution over indexed variables.Finite-past boundary constructions can represent settings such as initial states in time-series models.
- Observable World: Although the model itself is non-probabilistic, its observable world consists of noisy random variables generated by measurable mechanisms and has a uniquely determined distribution in law.The resulting observable world satisfies a local Markov property under the stated assumptions.
3 Structured Kernels
Structured kernels combine probability kernels in causal order and attach their dependencies to model-aligned structural graphs. Graphical operations such as simplification, c-component computation, gluing, and revealing support systematic regular computations and missing-value imputation.
- Preliminaries: The framework defines formal objects for knowledge obtained from observations and a rigorous calculus for transforming and computing with them.Its basic kernel operations include products, marginalizations, disintegrations, and reordering.
- Preliminaries: Kernel products are asymmetric because causal ordering determines how shared arguments and outputs are passed between factors.Marginalization removes terminal terms, while disintegration turns initial terms into arguments.
- Regular Computation: Regular computations are finite combinations of these operations, which yield suitably unique results for generic kernels.The paper conjectures that every finitely computable functional of generic kernels is regular.
- Structural Models: A structural graph is a directed acyclic dependency object whose inner nodes carry kernels and whose outer nodes represent shared arguments.Model alignment assigns kernel arguments to parent nodes and wires them into a structured kernel.
- Graphical Operations: Graphical operations represent important regular computations: simplification removes irrelevant descendants, while c-component computation and gluing decompose and reconstruct structured kernels.These operations generalize familiar IID causal structures within the abstract framework.
- Graphical Operations: The revealing operation can compute a model with altered latent structure and thereby yields a non-parametric causal missing-value imputation technique.For a structural c-component, it computes a regular functional after replacing part of the latent set.
- Graphical Operations: The paper conjectures that every regularly computable structured kernel can be obtained through a finite sequence of graphical operations, but this remains unresolved.The authors identify more general revealing operations as a principal concern.
4 Extraction from Data
This section defines local graph embeddings and repeated embedding families, then augments them with latent and ancestral structures to support identification. A backdoor-free family makes μ(G, L) identifiable.
- 4.1 Families of Embeddings: Local graph embeddings injectively map model-aligned local graphs into the I-graph while including proper parents, preserving proper edges, and aligning mechanisms.These conditions are illustrated in Figure 4.
- 4.1 Families of Embeddings: Families of embeddings collect repeated occurrences of one model-aligned local graph around an anchor node, with a symmetry group and applicability range.They provide repeated observations needed for statistical reasoning.
- 4.2 Decorated Families: Latent subsets and ancestral structures capture unobserved nodes and relevant directed ancestral relations, but minimal structures need not be unique.For X1 → Y ← X2, either X1 or X2 can form a minimal latent subset when exactly one is observed per embedding.
- 4.2 Decorated Families: A decorated family is backdoor-free when no latent node is external and no ancestral edge starts at a latent node.Identification depends on the interaction of latent nodes and backdoor paths.
- 4.3 Extracting Backdoor-Free Families: An algorithm grows direct mechanism embeddings by absorbing latent external nodes and children of latent nodes until the family becomes backdoor-free.The same mechanism may occur with different neighborhoods, so absorbing parents or children is not unique.
- 4.4 Identifiability from Embeddings: For a backdoor-free decorated family, μ(G, L) is identifiable.Theorem 1 supplies the identification result for the extracted structure.
5 Queries and Prediction
The paper formulates causal questions as basic and structured queries, connects them to identifiable structured kernels, and provides an extraction–decomposition–assembly strategy. In IID settings, this framework matches established identification results while also covering transport and mediation queries.
- 5 Queries and Prediction: The framework describes models, extracts structured kernels from observations, and uses graphical transformations to formulate and identify causal queries.Query-identifiability is stated directly relative to a realized world rather than through embeddings or structured kernels in its hypothesis.
- 5.1 Basic Queries: A basic query specifies target variables, conditioning variables, and intervention parameters, and is identifiable when its associated probability kernel is identifiable from the dataset.Interventional distributions are a special case of basic queries.
- 5.2 Structured Queries: Structured queries embed a local graph with target inner nodes, while the underlying basic query uses the embedded targets and external nodes.Canonical structured queries can be built by repeatedly adding I-graph parents until reaching queried external variables or no further parents.
- 5.3 Identification: Theorem 2 establishes query identifiability when μ(G, L) has a regular functional whose component kernels are all identifiable.The EDAIdentify algorithm extracts candidate structures, decomposes them, and assembles identification strategies.
- 5.4 Relation to the IID-Case: Under IID data, do-interventions, and uniform missingness, EDAIdentify matches the completeness of the standard ID algorithm for unconditional interventions.Conditional queries may require meta-queries because do-calculus Rule 2 uses internal intervention structure excluded by the framework.
- 5.4 Relation to the IID-Case: The framework also reproduces the mz-transportability setup and the mediation formula for natural direct effects without introducing counterfactuals.Its query formulation captures the question motivating natural direct effects directly.
6 Conclusion
The paper introduces an abstract language for causal reasoning and uses it to identify causal queries, while recovering established structures and results. It also notes open questions about learning the formalism's models from data.
- The paper introduces a simple, widely applicable abstract language for describing causal reasoning.
- Its weak structural assumptions support learning and reasoning about models and identifying causal queries, including conventional questions as special cases.
- The formalism recovers c-components, c-trees, c-forests, and hedges, validating its results against known IID special cases.
- The approach correctly reproduces the mediation formula for natural direct effects and mz-transport results.
- How and to what degree the formalism's models can be learned from data remains an open question.
Reading Guide
The reading guide organizes the paper's technical development from structural graphs and kernels to graphical operations, model extraction, and multi-level queries. It also highlights the appendix proofs and the separation of randomness from measure-theoretic models.
- Reading Guide: The appendix explains that graphical operations and Theorem 1's extraction from backdoor-free families are its two most technically relevant results.
- Reading Guide: Technical details are distributed across appendices covering graphical operations, model-data relations, multi-level setups, and the connection to textbook probability results.
- Structural Graphs: Structural graphs consist of inner and outer nodes connected by directed edges into inner nodes, with acyclic graphs admitting causal orders.
- Structural Graphs: A diagonal map can model one mechanism receiving the same variable as multiple parents, represented as an allowed soft-intervention.
- Structural Kernels: Structured kernels align probability kernels with graph nodes and combine them in causal order while preserving graph-based sparsity and latent structure.
A.2 Regularity
This section formalizes computation and regularity for probability kernels while emphasizing that identifiability depends on support. It uses singular interventions to show why observational conditionals may not determine causal queries.
- Regularity: Identifiability asks whether observations uniquely determine answers, whereas estimation concerns approximating those answers from finite data.
- Support: Disintegrations are generally unique only almost everywhere, so learned conditionals are supported only on values encountered with non-zero probability.
- Singular Interventions: Two IID models can share the same observational joint distribution while disagreeing on a point intervention, because the intervention targets a probability-zero value.
- Singular Interventions: Replacing an observational measure by a query measure yields a well-defined transferred kernel only when the query measure is dominated by the observational one.
- Regularity: Regular functionals are built from finite combinations of kernel products, disintegrations, marginalizations, and transpositions.
A.3 Graphical Operations
The section develops graphical operations for transforming and combining structured kernels. Its core construction decomposes kernels into atom and c-component representations, enabling simplification, extraction, gluing, and reconstruction through regular functionals.
- Gluing: Two subgraph kernels can be glued into a larger kernel when every structural c-component lies wholly within at least one of the two subgraphs.The gluing operation uses kernels from the two smaller graphs as inputs.
- Simplification: A structurally simplified subgraph kernel can be regularly computed by removing suitable inner nodes and marginalizing their associated structure.The removable set must be descendant-closed, and the corresponding simplified graph is obtained by a direct simplification.
- Atoms: Atoms decompose a structured kernel into per-observable-node factors that are regularly computable and contain enough information to reconstruct the original kernel.The kernel is a regular functional of its atoms.
- C-components: The full kernel can be reconstructed as a regular functional of all non-trivial c-component kernels jointly.This generalizes the atom-based reconstruction to c-component structure.
- Revealing: Revealing operations compute a kernel with selected latent variables revealed from the full kernel and the corresponding c-component kernel.The result is a regular functional of µ(G, L) and µ(Gc, L′).
B.1 Results in the Main Text
The paper constructs a shallow distribution and observable world under acyclicity and finite-past assumptions, then characterizes their probabilistic and graphical properties. It also argues that stationarity is not systematically required for identifiability through symmetry.
- Probabilistic foundations: Under acyclicity and finite past, a probability measure Pθ exists on the product space and is characterized by its finite-dimensional marginalizations.This measure is called the shallow distribution and is parametrized by the intervention mechanisms.
- Observable worlds: An observable world exists under the same assumptions, is unique up to equality in distribution, and satisfies the local Markov property.Its variables are generated from parent variables and jointly independent uniform noises.
- Observable worlds: The observable-world construction proceeds inductively in causal order, matching each finite-dimensional marginal to the model distribution.At each step, a measurable mechanism and an independent uniform noise generate the next variable.
- Graphical properties: The framework provides structural-graph properties for single-level and multi-level settings, including preservation of mechanisms and parent sets outside designated ancestral regions.These properties support later identification arguments for structured kernels.
- Identifiability scope: Stationarity is unnecessary for a systematic symmetry-based approach to identifiability, although it can help as an a-priori assumption about a specific model realization.Using stationarity merely to structure the problem may relocate or obscure the difficulty rather than solve it.
C.4 Identification from Families of Embeddings
This section shows that backdoor-free families of embeddings support identification of structured kernels and enable algorithmic construction of useful embedding families. It also extends the query language while identifying observational-support and conjectural limitations.
- Kernel identification: The main technical contribution is identification of structured kernels from suitably repeated, backdoor-free embeddings.Theorem 1 states that µ(Gobs(G), L) is δi-identifiable for a backdoor-free family.
- Limitations: The framework's validity guarantees do not ensure non-degeneracy of embedding variables, and degenerate observational support can severely restrict transfer conclusions.If all embeddings yield the same value, there may be no observational support elsewhere; related support problems can also occur without X-degeneracy.
- Kernel identification: Backdoor-free embeddings let the observed graph behave as a union of structural c-components, allowing data attached to embeddings to identify the associated kernel.The proof transfers c-component structure between each embedding graph and the observed graph before applying identification results.
- Algorithmic construction: Algorithm ExtractCS-R extends a decorated family into non-unique c-connected backdoor-free embedding families.By Theorem 1, kernels in Kbasic are identifiable from such c-connected families.
- Identification scope: Smaller observed c-structures can yield better identification than the largest local c-structure when the latter has worse symmetry.The relevant comparison is between observationally available subgraphs and the largest structure.
- Query extensions: A more expressive structured query formulation captures additional basic queries by weakening the hypothesis while still excluding queries that cause observations.Under the stated identifiability conditions, the underlying query is identifiable when its component kernels are δi-identifiable.
D.3 Identification of Queries
The paper identifies structured queries by linking their kernels to realized-world distributions and composing identifiable subkernels through regular functionals. Basic queries include do- and soft-interventions while supporting more flexible and extensible constructions.
- Query identification: The technical pipeline associates a structured query graph with the realized-world distribution before connecting its structured kernel to observational data.This association is described as the main technical step of the query formalism.
- Query identification: Identification is preserved when identifiable representatives are substituted into regular functionals, with equality interpreted almost everywhere where appropriate.The supporting lemmas establish that ν = F[ν1, . . . , νn] represents the target kernel almost surely.
- Query identification: A regular functional of individually identifiable kernels yields an identifiable structured kernel and underlying query.Theorem 2 applies when µ(Gquery, L) = F[µ1, . . . , µn] and every µi is identifiable.
- Scope: Basic queries contain conventional do- and soft-interventions as special cases but are more flexible and readily extensible.The formalism also supports multi-level, nested, meta-, stationary-distribution, and post-intervention fixed-point questions.
D.5 Examples for Multi-Level Queries
The multi-level river example distinguishes queries by how observational and query structures share latent contexts. More local queries can be more precise, but require different data for identification and may induce exchangeability rather than IID sampling.
- River example: Three river queries ask how water level changes after enforcing throughput at an observed site, a nearby unobserved site, or a site sharing geological context.The queries differ in where the intervention is applied and which contextual properties are shared.
- Identification conditions: The three queries require progressively different identification resources, ranging from many time-points at one site to many sites and shared sediment contexts.The third query requires learning a composition involving Y, Cgeom, and Csediment across many sites.
- Precision and variance: More local queries can be more precise because Y1 mainly reflects local noise, whereas Y2 and Y3 also include contextual variation.Finite-sample estimator variance additionally depends on estimation error and the number of available observations.
- Dependence structure: At a new site, repeated Y1 observations are IID, while repeated Y2 and Y3 observations are conditionally IID given shared geometry and therefore exchangeable.The formalism represents this distinction through the query graph's shared latent nodes.
- Formal distinction: The formalism distinguishes meaningful queries by how their query graphs attach to observational graphs, a distinction unavailable in the single-level case.Different attachments can change which joint distributions are identifiable through gluing.
E.1 Standard Results for the Single-Context Case
The paper reviews the standard single-context IID framework, including SCM definitions, interventions, causal-effect identifiability, and the hedge-based completeness of the ID algorithm.
- SCMs and graphs: An SCM specifies IID endogenous variables, independent exogenous noises, parent sets, and measurable mechanisms satisfying structural equations.The paper assumes unique solvability and suitable parent-set minimality.
- Interventions: Interventions preserve exogenous noises and mechanisms outside the intervened variables, while hard interventions remove their parents; do-interventions assign constants.Do-interventions are represented by Pearl's do-operator and graphically by removing incoming edges.
- Identifiability: A causal effect is identifiable when its interventional distribution is uniquely computable from the observed joint distribution across all SCMs inducing the graph.This generic definition does not rely on internal assumptions such as linearity.
- Completeness: An interventional distribution is identifiable if and only if no hedge exists for the intervention and outcome.A hedge is the graph-theoretic obstruction used by the completeness result.
- Completeness: The ID algorithm is complete because failure produces a hedge, and the hedge criterion establishes that no further identification is possible.The paper states both the failure certificate and the resulting completeness corollary.
E.2 Translating the Standard IID-Setup
The paper translates standard IID causal models into its symmetry-based formalism by encoding variables, mechanisms, observations, and interventions as structured data. In this representation, the formalism recovers standard causal-effect identification results.
- IID representation: IID models are represented by dataset and sample permutation symmetries acting on finitely many variable indices, yielding exchangeable samples conditional on dataset identity.The viewport asymptotically observes selected variables while excluding hidden variables.
- Mechanism translation: Structural mechanisms are translated into probability kernels by pushing noise distributions through measurable maps.These kernels reproduce conditional distributions and assemble into the observed joint distribution.
- Intervention translation: The translated model uses separate observational and query datasets, with intervention mechanisms encoded as known kernels in the query component.For do-interventions, these kernels are singular at the intervened values and have no parents.
- Recovery of standard results: In the IID case, knowledge of the observational distribution and conditional distributions supports direct identification, and regular-functional identifiability implies standard causal-effect identifiability.The paper further proves completeness for single-context do-interventions with an empty conditioning set.
- Algorithmic validation: The extraction phase validates data availability and can enable a frontdoor-like strategy in a multi-site, multi-time example.It avoids dubious claims when local graphs lack sufficient data.
E.4 Single-Context with Non-Empty Conditioning Set
The section examines conditional queries involving interventions, showing when conditioning on an intervention is identifiable and how internal intervention structure affects that conclusion.
- Selection bias: Conditioning can introduce selection bias when an intervention is distributed, such as Xdo ∼N(0, 1), requiring the joint Pθ(Y, M) for estimation.That joint distribution is not identifiable from the available observations in the described setting.
- Conditional effects: The formalism distinguishes conditional effects that are generally non-identifiable from cases where the intervention takes only a single value.Both assessments can be correct because the approach does not assume exploitable internal structure on intervened mechanisms.
- Conditional effects: Conditioning on an intervened variable recovers the do-calculus interpretation when the intervention is not singular.The query Pθ(Y|M = m, Xdo = x) is identifiable because Yx,m is identifiable.
- Alternative regularity assumptions: A second route classifies certain kernels as regular, making selection bias trivial but requiring a custom notion of regularity for complex variables.The paper identifies multidimensional or categorical variables as settings where this approach may provide substantial insights.
E.5 Standard Results for the Multi-Context, Experimental Case
The multi-context experimental framework formalizes transportability across related systems and recovers completeness results through c-forests, hedges, and the EDAIdentify and TRmz procedures.
- Multi-context setup: Selection diagrams represent related domains with shared causal graphs, selection variables for changed mechanisms, and experiments available on specified variable subsets.Each changing observed mechanism receives a selection variable connected to its corresponding node.
- Multi-context setup: mz-transportability asks whether a target-domain intervention effect is uniquely computable from observational and interventional distributions collected across source domains.The target effect is P∗(Y | do(X = x)), and experiments may differ by domain.
- Obstructions: The framework extends hedge-based reasoning to mz-shedges and mz*-shedges that organize c-forest obstructions across domains and experiments.The mz*-shedge additionally requires a directed path between specified rooted regions that avoids X.
- Obstructions: In a standard example, P(Y | do(X = x)) is identifiable whereas P(Y, Z | do(X = x)) is not, because the hedge applies to the joint target but not to Y alone.The example motivates the distinction between target ancestry and the stronger path condition in the shedge definition.
- Caveats: Several transportability statements require qualification because the cited shedge definition has inducedness and path-condition problems in extended examples.The text says the path requirement would make theorems 2 and 3 fail for a simple extended example.
- Completeness: The stated TRmz and EDAIdentify procedures are complete, with failures producing graph structures containing mz*-shedges or with no mz-shedge, respectively.The EDAIdentify result states that it returns a non-empty set exactly when no mz-shedge exists.
E.8 Mediation and Counterfactuals
The paper shows that its query language identifies the mediation formula for natural direct effects without introducing counterfactuals. It situates this result within a broader framework for extending causal identification beyond standard IID and intervention settings.
- Mediation formula: The query graph in Figure 9 yields the mediation formula for a structured query involving Y.
- Mediation formula: The algorithm identifies the mediation formula for natural direct effects from extracted kernels.Taking the expectation of Y computes both terms of the usual difference immediately.
- Counterfactuals: Natural direct effects can be expressed in the paper’s query language without introducing counterfactuals.The paper argues that counterfactuals are not needed to phrase these questions within its more flexible formalism.
- Mediation formula: With observed confounders, the same query includes them as additional kernel arguments and integrates them out.
- Scope: The framework aims to expand causal identification to data structures and queries beyond the IID and do-intervention setups used in much prior work.The paper compares its approach with IID and do-intervention results while emphasizing a broader accessible query language.