The Inflation Technique for Causal Inference with Latent Variables

I. INTRODUCTION

• With latent variables, observed conditional independences are necessary but generally insufficient to establish compatibility, making causal inference more difficult.For finite-cardinality observed variables, compatibility can be framed as nonlinear quantifier elimination, whose exact algorithms are generally impractical beyond simple scenarios.Source: §I · p. 3
• Inflation-derived polynomial inequalities can detect the Triangle scenario's W-type distribution as incompatible when conditional-independence, entropic, and covariance methods cannot.The paper describes systematic and more efficient algorithms for deriving these inequalities, including from marginal-polytope facets and hypergraph transversals.Source: §I · p. 4
• The inflation technique derives necessary compatibility conditions by duplicating nodes while mirroring ancestry and translating constraints from the inflated structure back to the original.The approach uses marginal-problem inequalities and factorization conditions from d-separation, especially for sets with disjoint ancestries.Source: §I · p. 4
• The method is presented as a systematic tool for latent-variable causal inference that needs no bounds on latent-variable cardinalities, but the authors do not establish that its generated inequalities are sufficient.The paper also discusses applications to quantum causal models and generalized probabilistic theories.Source: §I · pp. 4–5

II. BASIC DEFINITIONS OF CAUSAL MODELS AND COMPATIBILITY

• A causal structure is a DAG distinguishing observed and latent nodes, while a causal model adds conditional distributions for each node given its parents.Compatibility means that some choice of those causal parameters yields the given observed distribution.Source: §II · pp. 5–6

III. THE INFLATION TECHNIQUE FOR CAUSAL INFERENCE

• Each observed node in an inflation has an ancestral subgraph equivalent to its original counterpart after copy indices are removed.Equivalently, each copied node has parents corresponding to the original node's parents.Source: A. Inflations of a Causal Model · p. 6
• The Web inflation duplicates each latent node and quadruples each observed node to represent the possible parentage choices.The paper also presents Spiral, Capped, and Cut inflations; Cut does not contain the Triangle scenario as a subgraph.Source: A. Inflations of a Causal Model · p. 6, Figure 2 · p. 7, Figure 5 · p. 7
• The inflated model assigns copied nodes the same causal dependence on their parents as in the original model.The construction determines a unique inflated model for a given original model and inflation structure.Source: A. Inflations of a Causal Model · p. 7
• When a set of copies has an ancestral subgraph matching its original counterpart, its marginal distribution equals that counterpart's distribution.Such sets are called injectable; an injectable set's ancestry cannot contain multiple copies of any original node.Source: A. Inflations of a Causal Model · p. 8

B. Witnessing Incompatibility

• If a distribution is compatible with the original structure, its corresponding marginals on injectable sets must be compatible with the inflation, so an inflation-level contradiction rules out original compatibility.The paper applies this transfer to families of distributions on injectable sets and their images in the original structure.Source: B. Witnessing Incompatibility · pp. 8–9
• The Cut inflation rules out the perfectly three-way-correlated distribution in the Triangle scenario because perfect correlations imply dependence between copies that must be independent.The argument uses the Cut inflation’s ancestral independence of A2 and B1 to derive a contradiction.Source: B. Witnessing Incompatibility · p. 9
• The Spiral inflation proves that the W-type distribution is incompatible with the Triangle scenario, a result the authors say existing causal-inference techniques cannot derive.They identify failures of conditional-independence checks, entropic inequalities, and covariance-matrix methods as examples.Source: B. Witnessing Incompatibility · p. 10
• For the PR-box distribution, duplicated settings and outcomes yield mutually inconsistent output constraints, proving incompatibility with the Bell scenario.The Bell inflation shown in Fig. 8 is sufficient to witness any distribution’s incompatibility with the Bell scenario.Source: B. Witnessing Incompatibility · pp. 11–12
• The Bell-inflation argument parallels standard PR-box incompatibility proofs by showing that no joint assignment can satisfy all four output constraints.The authors describe the inflation argument as the same reasoning in disguise.Source: B. Witnessing Incompatibility · p. 12

C. Deriving Causal Compatibility Inequalities

• A causal compatibility inequality is a necessary condition: violating it witnesses incompatibility, while satisfying it does not guarantee compatibility.The inequalities are derived as constraints that every compatible distribution must satisfy.Source: C. Deriving Causal Compatibility… · p. 12
• An inequality for an inflation becomes one for the original structure by replacing distributions on copied variables with distributions on their corresponding original-variable sets.The translation applies to injectable sets and their images under dropping copy-indices.Source: C. Deriving Causal Compatibility… · p. 12
• For binary Triangle variables, a correlator inequality yields a tradeoff between AC and BC correlations when A and B are not completely biased.Its symmetry variants express a monogamy property: if two variables with unbiased marginals are perfectly correlated, both are independent of the third.Source: C. Deriving Causal Compatibility… · p. 13
• Inequality (31) robustly witnesses incompatibility between 3-way perfect correlation and the Triangle scenario, excluding distributions close to perfect correlation; polynomial inequality (34) also holds for continuous variables in [−1, +1].Source: C. Deriving Causal Compatibility… · p. 13
• The gap between the paper’s inner and outer constructions remains unresolved and may reflect limited model constructions or insufficiently sensitive small inflations.The authors defer closing this gap to future work.Source: C. Deriving Causal Compatibility… · p. 14
• An entropic constraint produces a Triangle compatibility inequality for arbitrary-cardinality variables, although inflation is not expected to improve computation over standard entropic-inequality algorithms.The authors emphasize instead that inflation provides a narrative for the inequality without reference to unobserved variables.Source: C. Deriving Causal Compatibility… · p. 14
• A joint-distribution inequality that includes PABC(000) captures genuine three-way correlations and rules out the W-type distribution in the Triangle scenario.For that distribution, the inequality’s right-hand side vanishes while its left-hand side does not.Source: C. Deriving Causal Compatibility… · p. 15

IV. SYSTEMATICALLY WITNESSING INCOMPATIBILITY AND DERIVING INEQUALITIES

• For the Triangle example, perfect correlations between A2 and C1 and between B1 and C1 imply perfect correlation between A2 and B1, contradicting the inferred A2B1 marginal.The contradiction witnesses incompatibility with the Triangle scenario.Source: §IV · p. 17
• The general procedure identifies ai-expressible sets, infers their distributions from the original distribution, and tests whether those distributions admit a common joint distribution.Constraints on consistent marginals become causal compatibility inequalities after substituting the factorization relations.Source: §IV · p. 17
• The method assigns each injectable set its corresponding original marginal, then forms ai-expressible-set distributions as products over ancestrally independent injectable components.Constraints on these distributions are rewritten first for the inflation and then for the original structure by removing copy indices.Source: §IV · p. 17
• If distributions on the inflation’s ai-expressible sets cannot be marginals of one joint distribution, the original distribution is incompatible with the original causal structure.The test is necessary but not sufficient for compatibility.Source: §IV · p. 17
• Ai-expressible sets can be computed from injectable distributions using ancestral independences and marginalization, linking distributions on the inflation to the original causal model.The article focuses on ai-expressible sets, leaving more general expressible sets for future work except in Appendix D.Source: §IV · pp. 16–17

A. Identifying the AI-Expressible Sets

• Injectability can be checked on pairs: a set is injectable exactly when the projection removing copy indices is injective on its ancestors.The Spiral inflation example lists five maximal ai-expressible sets.Source: A. Identifying the AI-Expressible Sets · p. 18
• Injectable sets are precisely the cliques of the injection graph, whose nodes are inflated observed variables and whose edges mark injectable pairs.All nonempty cliques are needed, not only maximal ones.Source: A. Identifying the AI-Expressible Sets · p. 18
• Ai-expressible sets correspond to cliques in a graph whose nodes are injectable sets and whose edges connect ancestrally independent sets.The clique’s injectable sets provide the partition used in the factorization relation; enumerating maximal cliques suffices.Source: A. Identifying the AI-Expressible Sets · p. 18, Figure 10 · p. 18

B. The Marginal Problem and its Solution

• A family of marginals must be consistent on overlapping contexts, but this condition is often insufficient for a joint distribution to exist.The necessary and sufficient constraints are given by consistency equations together with marginal-polytope facet inequalities.Source: B. The Marginal Problem and its Solution · pp. 19–20
• The marginal satisfiability problem is a linear program that tests whether a joint distribution reproduces the given context marginals.The marginal polytope is the convex hull of deterministic assignments consistent on context overlaps.Source: B. The Marginal Problem and its Solution · p. 19
• A violated valid marginal inequality certifies infeasibility, and substituting inflation factorization relations turns such inequalities into causal compatibility inequalities.Farkas infeasibility certificates can provide explicit inequalities witnessing a distribution’s incompatibility.Source: B. The Marginal Problem and its Solution · p. 20
• The method uses marginal satisfiability to witness incompatibility and the marginal constraint problem to derive causal compatibility inequalities.The authors report reproducing a prior Triangle-scenario incompatibility result using the Web inflation.Source: B. The Marginal Problem and its Solution · p. 20

C. A List of Causal Compatibility Inequalities for the Triangle scenario

• 4,884 facet inequalities from the Spiral inflation yield 4,884 polynomial causal compatibility inequalities for the binary Triangle scenario.The full list corresponds to facets of the relevant marginal polytope.Source: C. A List of Causal Compatibility… · p. 20
• 52 polynomial inequalities form an irredundant complete system after pruning the 4,884-inequality set and removing eight positivity inequalities.The authors first identified a 1,433-inequality complete subset, then used numerical constrained maximization to remove redundant inequalities.Source: C. A List of Causal Compatibility… · pp. 20–21
• For the Triangle scenario with binary observed variables, the inequalities are grouped into four symmetry classes, with one representative from each; the symmetry group has 48 elements.Source: C. A List of Causal Compatibility… · p. 21
• The reported inequalities have no slack: each can be saturated by a distribution compatible with the Triangle scenario.The paper identifies deterministic distributions saturating the inequalities.Source: C. A List of Causal Compatibility… · p. 21

D. Causal Compatibility Inequalities via Hardy-type Inferences from Logical Tautologies

• The Spiral inflation makes A2, B2, and C2 independent with full-support marginals, contradicting the constraint that their all-ones outcome has zero probability.The joint distribution on {A2B2C2} must factorize, so it has full support.Source: D. Causal Compatibility Inequalities… · p. 22
• The implication on deterministic assignments becomes a marginal inequality when its disjunction is converted to a sum using the union bound.For the Spiral example, one resulting inequality bounds PA2B2C2(111) by four marginal probabilities.Source: D. Causal Compatibility Inequalities… · p. 22
• For binary A, B, and C with pair contexts AB, AC, and BC, the tautology yields PAC(00) ≤ PAB(00) + PBC(10).The passage identifies this inequality as equivalent to Eq. (31).Source: D. Causal Compatibility Inequalities… · p. 23
• Enumerating logical tautologies of the relevant form systematically yields probabilistic inequalities that are necessary but insufficient conditions for the marginal problem.The procedure seeks tight implications whose consequents are minimal.Source: D. Causal Compatibility Inequalities… · pp. 22–23
• Minimal hypergraph transversals generate many useful inequalities at lower computational cost than full facet enumeration, but are generally weaker.All irredundant polynomial inequalities in Eqs. (51–54) are found by this method; facet enumeration is preferred when tractable.Source: D. Causal Compatibility Inequalities… · pp. 23–24

V. FURTHER PROSPECTS FOR THE INFLATION TECHNIQUE

• Constraints from an inflated causal structure can be transferred to the original causal structure.The paper identifies joint-distribution constraints and ancestral independences as weak inflation-level constraints that have already produced significant new results.Source: §V · p. 24
• The paper turns to three further possibilities for strengthening inflation: d-separation relations, copy-index symmetries, and nonlinear constraints.These are introduced as additional constraints that might enhance the technique.Source: §V · p. 24
• Even weak constraints at the inflation level, including ancestral independences, can yield significant results for causal inference on the original structure.Source: §V · p. 24

A. Appealing to d-Separation Relations in the Inflated Causal Structure beyond Ancestral Independance

• Maximal expressible sets can incorporate d-separation relations beyond ancestral independence and yield stronger compatibility inequalities and incompatibility tests.The added sets can be handled as marginal-problem contexts, preserving linear inequalities before translation into polynomial causal constraints.Source: A. Appealing to d-Separation Relations… · p. 25

B. Imposing Symmetries from Copy-Index-Equivalent Subgraphs of the Inflated Causal Structure

• Copy-index-equivalent variables have equal parent-response mechanisms, which can imply equality of marginal distributions even for non-injectable contexts.Such equalities can be added to compatibility or quantifier-elimination problems and may strengthen the resulting constraints.Source: B. Imposing Symmetries from… · p. 25

C. Incorporating Nonlinear Constraints

• Additional d-separation relations can be imposed as nonlinear constraints during quantifier elimination when the relevant joint set is not expressible.For compatibility testing, substituting the observed marginals into the polynomial system reduces the task to checking whether a solution exists.Source: C. Incorporating Nonlinear Constraints · p. 26
• Nonlinear quantifier elimination is generally impractical for large polynomial systems with many unknowns to eliminate.The paper notes that algebraic-geometric structure may help; mixed linear and nonlinear elimination is also suggested as a possible route to further inequalities.Source: C. Incorporating Nonlinear Constraints · p. 26

D. Implications of the Inflation Technique for Quantum Physics and Generalized Probabilistic Theories

• An inflationary fan-out is a latent node with at least two copy-index-equivalent children.Source: D. Implications of the Inflation… · p. 28
• Without an inflationary fan-out, inflation-derived inequalities are GPT-valid; with one, quantum and GPT violations may occur.The distinction is linked to the impossibility of broadcasting in quantum theory and analogous GPTs.Source: D. Implications of the Inflation… · p. 28
• A quantum model can be built for an inflation without fan-outs by reusing channels or measurements, or tracing out or marginalizing pruned outputs.This construction lets constraints on the inflated structure translate to the original; it preserves quantum incompatibility of perfect correlation with the Triangle scenario and leaves Eqs. (34,39) without quantum violations.Source: D. Implications of the Inflation… · p. 29

VI. CONCLUSIONS

• The inflation technique enhances existing methods using entropies, correlators, or probabilities; its computational cost depends on the seed technique.Source: §VI · p. 29
• Polynomial inflation inequalities can rule out the Triangle W-type distribution where entropic techniques cannot, and are stronger in at least some cases.Source: §VI · p. 29
• Some inflation-derived inequalities are necessary conditions for compatibility in any GPT, extending beyond classical models to include quantum causal models.Source: §VI · p. 29, D. Implications of the Inflation… · p. 29
• The authors lack a systematic way to find quantum violations of their polynomial inequalities beyond Bell scenarios, and do not know how to bound the useful inflation size.They note that semialgebraicity gives a finite upper bound on the maximum degree of tight characterizing polynomials, but not a known bound on that degree.Source: §VI · p. 30
• A later result proves that a suitably large inflation witnesses every incompatible distribution, making inflation necessary and sufficient for causal compatibility.The result is attributed to Ref. [108].Source: §VI · p. 30

Appendix A: Algorithms for Solving the Marginal Constraint Problem

• The marginal constraint problem is facet enumeration for a polytope whose vertices are deterministic assignments consistent across overlapping contexts.The marginal description matrix supplies the vertices to the enumeration algorithm.Source: Appendix A · p. 31
• Fourier–Motzkin elimination projects the system b = Vx, x ≥ 0, with normalized weights, into b-space by eliminating x.The paper reports custom-coding a linear quantifier-elimination routine in Mathematica.Source: Appendix A · p. 31

Appendix B: Explicit Marginal Description Matrix of the Cut Inflation with Binary Observed Variables

• For the binary Cut inflation, the three maximal ai-expressible pairs yield a 12 × 8 marginal description matrix.The joint distribution vector has 8 elements and the marginal distribution vector has 12.Source: Appendix B · p. 32

Appendix C: Constraints on Marginal Distributions from Copy-Index Equivalence Relations

• An inflationary isomorphism between observed-node contexts guarantees equal marginal distributions under its specified variable correspondence.When a context maps to itself, the equality expresses invariance under the corresponding permutation.Source: Appendix C · p. 33
• Matching ancestral subgraphs alone does not guarantee equal marginals; the isomorphism must also extend the copy-index correspondence between the contexts.In the instrumental-scenario example, the ancestral subgraphs coincide although the context subgraphs are not isomorphic.Source: Appendix C · pp. 33–34, Figure 13 · p. 34
• Adding equal-marginal constraints can speed elimination, but the authors’ computations found no example where they tightened the marginal constraints.When the isomorphism extends to an automorphism of the whole inflation, a symmetric mixture makes the added equality redundant.Source: Appendix C · p. 34
• Whether these equalities tighten inequalities remains unclear when the isomorphism does not extend to a full automorphism, or when extra d-separation constraints are imposed.Source: Appendix C · p. 34

Appendix D: Using the Inflation Technique to Certify a Causal Structure as “Interesting”

• HLP found 21 potentially interesting structure classes with up to six nodes; they proved 18 interesting using explicit distributions and entropic inequalities.Source: Appendix D · p. 35
• For the three remaining classes, larger Shannon-inequality sets were suggestive but not conclusive because the additional inequalities might follow from observed-variable independences.Source: Appendix D · p. 35

1. Certifying that Henson-Lal-Pusey’s Causal Structure #16 is “Interesting”

• The Pienaar distribution is incompatible with HLP causal structure #16 despite satisfying its sole observed conditional independence, C ⊥⊥ Y.Source: §1 · pp. 35, 37
• The Russian dolls inflation uses injectable sets and an expressible set to derive the contradiction for structure #16.The construction uses {A1C1Y1}, {B2C2Y2}, and {B2C1Y2} as injectable sets, and {A1C2Y1} as expressible.Source: §1 · p. 36
• The contradiction requires P(C2 = 0, A1 = 0) ≥ 1/2 while the distribution gives that probability as 1/4.Source: §1 · p. 37

2. Deriving a Causal Compatibility Inequality for HLP’s Causal Structure #16

• An inequality derived from a joint distribution on {A1B2C1Y1Y2} is strong enough to witness the Pienaar distribution’s incompatibility with structure #16.Source: §2 · pp. 37–38
• Four ancestral independences in the inflation factorize the marginals in the initial inequality.Source: §2 · p. 37
• Because one marginal is only expressible, it must be substituted by its expression before the inequality can be translated to the original structure.The resulting inequality refers exclusively to distributions on injectable sets, enabling the translation.Source: §2 · pp. 37–38

3. Certifying that Henson-Lal-Pusey’s Causal Structures #15 and #20 are “Interesting”

• A distribution with C ⊥⊥ Y, A ⊥⊥ B | Y, and A ⊥⊥ Y | B that violates Eq. (D.22) establishes that HLP structures #15, #16, and #20 are interesting.The inequality also applies to #15 and #20 because their causal models are included among those of #16.Source: §3 · p. 38

Appendix E: The Copy Lemma and Non-Shannon type Entropic Inequalities

• Inflation constructs A′ with matching PAB marginals and A′ ⊥⊥ AC | B, linking the Copy Lemma to Shannon-type derivations of non-Shannon inequalities [121].The paper interprets those inequalities as Shannon-type inequalities applied to an inflated causal structure.Source: Appendix E · p. 39

Appendix F: Causal Compatibility Inequalities for the Triangle Scenario in Machine-Readable Format

• The Spiral inflation yields 52 numerically irredundant polynomial inequalities from consistent marginals; larger inflations can yield stronger inequalities, and these also follow from hypergraph transversals.Source: Appendix F · p. 40

• In Fig. 8, each maximal ai-expressible set contains all setting copies, and its conditional marginal reproduces PAB|XY for the corresponding setting pair.Source: Appendix G · p. 41
• The Fig. 8 inflation gives necessary and sufficient compatibility conditions for the Bell causal structure, and the inequality-generation method recovers all Bell inequalities.Source: Appendix G · pp. 41–42
• The marginal polytope factors into a Bell polytope and a simplex; its facets pair Bell-polytope facets with nonnegativity inequalities, including a CHSH version.Source: Appendix G · p. 42
• An inflationary precursor yields inequality (G.7) after division by setting probabilities and dropping copy indices; the precursor corresponds to the Hardy paradox.Source: Appendix G · p. 42