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) addresses a key limitation of classical causal discovery: the faithfulness assumption. Standard pipelines first use conditional independence tests on observational data to infer partial causal structure, then apply interventions to orient remaining edges. Faithfulness requires causally linked variables to exhibit statistical dependence, but many natural systems contain buffering and stabilizing pathways that cancel out, violating faithfulness and causing algorithms to incorrectly discard real causal dependencies. The authors argue that hard interventions carry information about the presence or absence of causal links that is ignored in the first stage of structure discovery. They prove that a mild assumption—interventional instantaneous faithfulness—permits such cancellations while still allowing nonparametric identification of the causal structure using hard interventions alone. These results reposition interventions as the primary carrier of causal information, to be prioritized over conditional independence testing. The paper also characterizes an equivalence class of structures for cases where the identification criterion is not met due to limited interventional scope.

Paper Overview

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

Abstract (translation)

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 ignored in the first stage of structure discovery. They prove that a mild assumption—interventional instantaneous faithfulness—which permits cancellations, is sufficient to nonparametrically identify causal structure using hard interventions alone.

These results position interventions as a primary carrier of information about causal structure, to be prioritized over conditional independence testing. To complete the paradigm shift, the paper also specifies an equivalence class of structures for when the identification criterion is unmet due to limitations on the scope of interventions.

Key Points

Classical causal discovery relies on faithfulness, which fails in systems with pathway cancellation (buffering, robustness mechanisms).
Hard interventions alone, under the assumption of interventional instantaneous faithfulness, suffice for nonparametric causal structure identification.
The paradigm is flipped: interventions should be treated as the primary source of causal information rather than a post-hoc tool for orienting edges.
When interventional scope is limited and identification fails, the paper characterizes the resulting equivalence class of causal structures.

---

*Source: arXiv:2607.11816*