Relaxing Faithfulness with Intervention-Only Causal Discovery

Forum topic · 小凯 · 2026-07-15

Summary

This paper by Bijan Mazaheri, Jiaqi Zhang, and Caroline Uhler (arXiv:2607.11816) revisits the standard causal discovery workflow, which first uses conditional independence tests on observational data and then applies interventions to orient unknown causal edges. The first step typically relies on faithfulness—the assumption that causally linked variables show statistical dependence. However, many natural systems contain buffering and stabilizing pathways whose effects cancel out, producing robustness that violates faithfulness and causes algorithms to wrongly discard real causal dependencies. The authors show that hard interventions carry information about the presence or absence of causal links that is ignored in the structure-discovery phase. Under a mild assumption called interventional instant faithfulness, which permits such cancellations, the causal structure can be non-parametrically identified using only hard interventions. These results reposition interventions as the primary carrier of causal structural information, to be prioritized over conditional independence testing. The paper also characterizes the equivalence classes that arise when identifiability criteria cannot be met due to limited intervention targets, flipping the traditional paradigm of observational-first causal discovery.

Paper Overview

Field: Machine Learning Authors: Bijan Mazaheri, Jiaqi Zhang, Caroline Uhler Published: 2026-07-13 arXiv: 2607.11816

Abstract

Causal discovery algorithms learn a network that describes the causal dependencies among random variables. A common workflow involves first utilizing conditional independence properties on observational data to determine partially directed causal relationships, then applying interventions to orient the unknown causal directions. A critical assumption for the first step is faithfulness: a requirement that causally linked variables exhibit statistical dependence.

Many natural systems include buffering and stabilizing pathways that cancel out to achieve systemic robustness. This cancellation of pathways violates faithfulness, leading causal discovery algorithms to incorrectly remove causal dependencies.

In this paper, the authors argue that hard interventions contain information about the presence/absence of causal links that is overlooked in the first stage of structure discovery. They prove that a mild assumption—interventional instant faithfulness—which permits such cancellations, is sufficient to non-parametrically identify causal structure using hard interventions alone.

Key Contributions

Relaxing faithfulness: Path cancellations from buffering/stabilizing mechanisms no longer break identifiability.
Intervention-first paradigm: Interventions are shown to be the primary carrier of information about causal structure and should be prioritized over conditional independence testing.
Equivalence classes: When identifiability criteria are not satisfied due to limited intervention scope, the paper characterizes the resulting equivalence class of causal structures.