Computational Identifiability

The Quiet Revolution in Causal AI: Understanding Computational Identifiability

The recent arXiv preprint on "Computational Identifiability" (2606.19361v1) addresses a foundational gap in causal inference that has long troubled AI practitioners: the difference between theoretical identifiability and practical computability. While traditional causal identification theory tells us whether a causal effect can be estimated from observational data, this work asks a more pragmatic question: can we actually compute it given finite data, computational constraints, and real-world algorithmic limitations?

What the Research Actually Proposes

The paper formalizes a distinction that has been implicit in causal machine learning for years. Standard identifiability conditions—such as the back-door criterion or instrumental variables—guarantee that a causal parameter is uniquely determined by the observed distribution. But these guarantees assume infinite data and perfect computation. Computational identifiability introduces a new layer: it characterizes whether a target quantity can be computed by a feasible algorithm within bounded resources.

This is not merely a theoretical exercise. Consider a causal graph with 50 variables where the back-door adjustment requires conditioning on 10 confounders. Theoretically identifiable, yes. But with 10 binary confounders, you need at least 1,024 strata—and if any stratum has zero observations, your estimate collapses. Computational identifiability would flag this as "non-computable" despite theoretical identifiability.

Why This Matters for AI Practitioners

The implications are immediate and practical:

1. Causal discovery pipelines need redesign. Current tools like DoWhy or CausalNex often report "identifiable" without checking computational feasibility. This work provides a framework to add a computational feasibility check before running expensive experiments or A/B tests. 2. It bridges causal inference and deep learning. Neural causal models often struggle with high-dimensional confounders. Computational identifiability offers a principled way to decide when to use proxy variables, when to apply regularization, or when to abandon causal estimation altogether in favor of predictive models. 3. It changes how we evaluate causal claims. Many published causal estimates from observational studies may be theoretically identifiable but computationally non-identifiable—meaning the reported numbers are artifacts of the algorithm, not reflections of true causal effects.

The Broader Context

This research sits at the intersection of causal inference, computational complexity, and statistical learning theory. It echoes a growing recognition in AI that "in principle" guarantees are insufficient for deployment. As causal AI moves into healthcare, economics, and policy, practitioners need tools that warn them when their estimates are fragile due to computational constraints—not just statistical uncertainty.

The work also implicitly critiques the current practice of reporting only standard errors and confidence intervals. A causal estimate with tight confidence intervals may still be computationally non-identifiable if the algorithm used to compute it is unstable under small perturbations to the input data.

Key Takeaways

• Computational identifiability extends traditional causal identification by asking whether a causal parameter can actually be computed with finite data and bounded resources, not just whether it is theoretically determined.
• AI practitioners should audit their causal pipelines for computational feasibility, especially when dealing with high-dimensional confounders or complex graphical models where theoretical identifiability may be misleading.
• This framework provides a new diagnostic tool for causal machine learning: before running experiments, check whether the target quantity is computationally identifiable given your data size and computational budget.
• Expect future causal inference libraries to incorporate computational identifiability checks as a standard preprocessing step, similar to how they currently check for graphical criteria like d-separation.