Inflation Technique in Causal Inference

• Inflation Technique is defined as replicating causal model nodes to create inflated DAGs that test if observed distributions satisfy latent-variable constraints.
• It systematically constructs multiple copies of latent and observable variables to derive new, more stringent polynomial inequalities.
• Applications include enhancing quantum nonlocality tests, causal discovery in statistics, and generalizing Bell and instrumental inequalities.

The term "inflation technique" has been developed in distinct domains of theoretical physics and data assimilation, but in the context of causal inference and Bayesian networks, it refers to a powerful framework for determining whether an observed probability distribution can be explained by a given causal structure, particularly when the structure includes latent (unobserved) variables. The inflation technique circumvents the inherent limitations of conditional independence–based tests by systematically constructing extended ("inflated") versions of the original causal model, enabling the derivation of new polynomial constraints that can be far more restrictive than those accessible via conventional methods.

1. Concept and Rationale

The inflation technique addresses the causal compatibility problem: given a set of observed variables and a hypothesized causal structure (typically a directed acyclic graph, DAG, possibly with latent nodes), it asks whether there exists a joint probability distribution consistent with this structure that reproduces the observed statistics. Classical d-separation criteria, which generate observable conditional independence constraints, are generally insufficient in the presence of latent variables. The inflation approach amplifies the logical structure of the original causal model by introducing multiple independent copies of the latent variables and their associated observable descendants while ensuring that local causal relationships are precisely maintained.

Given a DAG $G$ describing the causal relations and an observed distribution $P_{X_1,\ldots,X_n}$, the inflation technique constructs a new causal structure $G'$—the "inflation"—in which each latent variable and its descendants are replicated, preserving their parental relationships. The core idea is that if the original model $G$ could generate the observed statistics, then so must the inflated model $G'$, subject to consistency and symmetry constraints.

2. Construction of the Inflation

The inflation $G'$ is built by creating several copies of the latent and observable nodes of $G$, establishing links so that the ancestry of each copy in $G'$ mirrors that of the original variable in $G$:

• If a node $X_i$ has parents $Pa_G(X_i)$ in $G$, then each copy $X_{i,k}$ of $X_i$ in $G'$ takes as parents the corresponding copies of $Pa_G(X_i)$.
• The local conditional distributions assigned to each node in $G'$ must be identical to those in $G$: $$P(X_{i,k} | Pa_{G'}(X_{i,k})) = P(X_i | Pa_G(X_i)).$$

This maintains the Markovian structure: each replica's probabilistic behavior is governed by the same laws as in the original model. This expanded structure enables the derivation of strong necessary constraints for causal compatibility.

3. Injectable and Expressible Sets

A pivotal step is identifying "injectable" sets of variables in the inflated DAG $G'$. An injectable set $V' \subset$ observed nodes of $G'$ corresponds, after dropping copy indices, to a set $V$ in $G$ whose ancestral subgraphs are isomorphic. The marginal distribution on $V'$ in an inflated distribution must match the corresponding observed marginal $P_V$:

$P_{V'} = P_V \ \ \ \text{if} \ V' \sim V.$

"Expressible" sets are larger sets whose joint distributions can be reconstructed from the marginals of injectable sets, provided their ancestral subgraphs are d–separated (“ancestrally independent” in the graph).

For ancestrally independent (AI) injectable sets $X',Y' \subset G'$, the factorization

$P_{X'\cup Y'} = P_{X'}\,P_{Y'}$

is enforced. This mechanism translates independence relations in the inflated graph into nontrivial polynomial constraints on the original observed distribution.

4. Derivation of Constraints and Inequalities

To test compatibility, one collects the set of marginals $P_{V'}$ over injectable (and AI–expressible) sets from the original observed distribution. The central requirement is that there exists a global joint distribution over all observed nodes in $G'$ that:

• Returns the assigned $P_{V'}$ on each injectable (expressible) set,
• Respects all imposed d-separation–induced conditional independencies of $G'$.

Any failure of such joint extensibility implies the incompatibility of the original observed distribution with the causal structure $G$.

This leads to the derivation of new, potentially high-degree polynomial inequalities that observed distributions must satisfy. The process involves constructing a so-called "marginal description matrix" $\mathbf{M}$ relating the joint distribution vector $\mathbf{v}$ of all possible assignments to the injectable marginals $\mathbf{b}$:

$\mathbf{M}\mathbf{v} = \mathbf{b}.$

The existence of a nonnegative solution $\mathbf{v}$ provides necessary conditions; the problem is a linear (or semialgebraic) constraint satisfaction instance and can be solved via polyhedral or hypergraph transversal algorithms.

These inequalities can:

• Generalize (and strengthen) Bell-type inequalities (e.g., CHSH in the quantum context),
• Recover the instrumental inequalities (e.g., Pearl's instrumental scenario).

5. Algorithmic Strategies

Three main classes of algorithms are used for extracting constraints in the inflation framework:

1. Linear quantifier elimination / facet enumeration: Algorithms such as Fourier–Motzkin elimination operate on the linear constraints to derive all tight inequalities (facets) satisfied by the compatible set. This is computationally intensive and limited to relatively small inflations.
2. Hypergraph transversals: For possibilistic (support-based, rather than probability-based) consistency, hypergraph transversal algorithms can efficiently generate Hardy-type inequalities that witness incompatibility even without enumerating all facets.
3. Symmetry-based constraints: Exploiting the symmetries among copies (permutation of indices), equivalence relations among marginals (e.g., $P_{A_1B_1} = P_{A_1B_2}$) can be imposed as additional linear constraints, enhancing the restrictiveness of the resulting system.

Efficiency depends on the complexity of $G'$; facet enumeration is feasible for small inflations, while transversals are applicable in higher-dimensional cases.

6. Comparison to Other Approaches

The inflation technique offers significant advantages over traditional conditional-independence analysis, which may miss restrictions arising from latent variables. Inflation-induced inequalities can be strictly stronger than those derived from entropic or linear constraints. For instance, in the Triangle causal scenario (three observables connected by three pairwise latents), inflation can rule out distributions (e.g., the "W-type" distribution) that satisfy all entropic and observable CI constraints but still cannot be realized by classical causal networks.

It also unifies the derivation of classic results:

• Bell inequalities (e.g., CHSH) in nonlocality tests are recovered.
• Instrumental inequalities (as in Pearl's scenario) are generalized to higher-degree forms.

7. Extensions and Applications

The inflation technique can be "lifted" to handle quantum (and GPT) models, provided the inflation avoids "inflationary fan-outs" that are forbidden in quantum theory by the no-broadcasting theorem. Thus, inflation-induced inequalities are valid in both classical and quantum theories for certain inflations but may be violated by quantum or post-quantum correlations in other inflations. This enables the systematic investigation of the boundary between classical, quantum, and generalized probabilistic theories.

Applications include but are not limited to:

• Quantum information, especially for characterizing nonlocality and ruling out classical hidden-variable models,
• Causal discovery and constraint-based structure learning in statistics and machine learning,
• Biomedical studies involving latent confounding variables.

Summary Table: Key Elements of the Inflation Technique

Element	Description	Example/Role
Inflated DAG	Replicates variables with preserved ancestry	Tests more stringent constraints
Injectable Set	Subset of observables isomorphic (ancestrally) to the original	Matches marginal distributions
Expressible Set	AI combinations of injectable sets	Factorizes to product of marginals
Inequality Type	Polynomial, possibly high-degree	CHSH, instrumental, Hardy-type
Symmetry	Copy index permutation constraints	$P_{A_1B_1} = P_{A_1B_2}$, etc.

The inflation technique provides a principled, algorithmic mechanism by which compatibility of observed distributions with complex latent-variable causal structures can be thoroughly interrogated, surpassing the capabilities of classic d-separation or entropic methods, and yielding a spectrum of new constraints of direct relevance to both classical and quantum domains (Wolfe et al., 2016, Navascues et al., 2017, Boghiu et al., 2022).