On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models

arXiv:2606.22391 · math.ST, econ.EM, stat.ML, stat.TH · Submitted 2026-06-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models".

Jane: The paper was written by Lin Liu, Rajarshi Mukherjee and James M. Robins from Institute of Natural Sciences, Ministry of Education Local Science Center, School of Mathematical Sciences, Shanghai Jiao Tong University-Yale and Department of Biostatistics, Harvard T. H. Chan School of Public Health and Department of Epidemiology and Department of Biostatistics, Harvard T. H. Chan School of Public Health.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Paper discussion segment 1: Tom: So, let's take a closer look at how "On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models" summarizes its findings and what those implications mean for real-world data analysis.

Jane: The paper is essentially showing that when we look at various functionals—like the quadratic functional or even expected conditional covariance—the first-order DML estimator isn't always the most efficient choice.

Lu: They use rigorous math to show that for certain types of parameters, this minimax estimator is asymptotically inadmissible, which means there are demonstrably better ways to estimate them in the long run.

Meng: The practical implication here is that if we blindly trust DML for functions falling into the monotone bias class, we might be accepting a measurable level of suboptimal performance that needs to be accounted for in our error bars.

Lalam: It’s a call for statistical honesty, recognizing that even when using state-of-the-art AI methods, there are theoretical limits to the accuracy of basic estimators.

Tom: The authors explicitly show that two out of the three functionals they studied—the quadratic functional and the quadratic density integral functional— fall into this critical monotone bias class.

Jane: That classification is key because it tells us precisely where our current standard methods fail; we can't just treat all complex parameters as being handled equally well by DML.

Lu: This structural categorization helps us understand that these functions possess a specific type of systematic error that simple first-order techniques cannot correct for.

Meng: For us, this means we need to build conditional logic into our systems: if the functional falls into this class, DML shouldn's be the default choice.

Lalam: Understanding this failure mode fosters a culture of responsibility in data science, driving us toward more nuanced and adaptive solutions instead of simple black-box reliance.

Tom: That really lays the groundwork for discussing how we move from identifying a problem to finding the specific solutions proposed by looking at next segment.

Paper discussion segment 2: Tom: Moving past just identifying the issues, let's look at how "On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models" proposes ways to fix these deficiencies by exploring the concept of bias reduction.

Jane: They are focusing on a specific group called the monotone bias class where we can systematically reduce that inherent error, or bias, using techniques much more advanced than standard DML.

Lu: The solution involves introducing second-order estimators, which is a major leap from first-order methods and represents a significant conceptual advancement in statistical design.

Meng: From an engineering standpoint, this means we are designing algorithms that actively calculate the systematic error at a higher order of approximation and then correct it automatically.

Lalam: This shift, as I see it, moves our focus from merely accepting the existing bias to actively engineering a statistical method designed to achieve near-perfect accuracy by correcting known systemic errors.

Tom: So, if DML is like using a reliable first-draft sketch of a building, these second-order methods are like having the detailed blueprints that account for structural load variations and corner tolerances.

Jane: It’s about injecting intelligence into the estimation process itself, not just relying on the inherent robustness that comes with machine learning alone.

Lu: And this capability to automatically estimate and correct bias based on higher-order influence functions is what gives these methods their power—it's adaptive correction built into the core mathematics.

Meng: This complexity means implementation isn't trivial; it requires building robust systems that can handle these layered corrections without introducing new sources of instability or computational errors.

Lalam: The fact that this correction is systematic, tied to the structure of the bias itself, rather than just an arbitrary tuning parameter, gives us confidence in its theoretical backing.

Tom: That's a clear progression: we found the problem areas with DML's first-order nature, and now we have a mathematically defined path toward solving those shortcomings through second-order methods.

Paper discussion segment 3: Tom: Now that we understand the technical upgrades—the second-order estimators—let's discuss the specific improvements proposed by "On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models" and what they offer in practice.

Jane: The paper introduces these sophisticated second-order estimators, often referred to as empirical higher-order influence function or HOIF estimators, which are designed to directly improve upon DML performance.

Lu: These HOIF estimators are capable of correcting the systematic bias of the first-order estimator by estimating a lower bound of that bias and then neutralizing it out.

Meng: The practical challenge here is implementing these complex adjustments; we need systems that can reliably calculate and apply these second-order corrections without introducing new computational instabilities or failure modes.

Lalam: This shift, as I see it, moves our focus from merely accepting an approximation to actively engineering a statistical method designed to achieve near-true accuracy by correcting known systemic errors.

Tom: So, we are moving beyond simple linear fixes; these second-order methods are like having the detailed blueprints that account for structural load variations and corner tolerances.

Jane: It’s about injecting intelligence into the estimation process itself, not just relying on the inherent robustness of machine learning alone to get a better result.

Lu: This capability to automatically estimate and correct a bias based on higher-order influence functions is what gives these methods their power—it's adaptive correction built into the core mathematics.

Meng: Implementing this requires building robust systems that can handle these layered corrections without introducing new sources of instability or computational errors, ensuring the model doesn't diverge under real data pressure.

Lalam: The fact that this correction is systematic, tied to the structure of the bias itself rather than an arbitrary tuning parameter, gives us confidence in its theoretical backing and its potential for a robust statistical future.

Tom: It’s a clear progression: we found the problem areas with DML's first-order nature, and now we have a mathematically defined path toward solving those shortcomings through second-order methods.

Conclusion: Tom: So, as we wrap up our discussion of "On the Asymptotic Inadmissibility of Double Machine Learning Estimators Under Structure-Agnostic Models," the main message is that while DML is a powerful tool, it's not perfect for all functions we might want to measure.

Jane: Exactly, Tom; it’s important for our listeners to know that even when the model is structure-agnostic, there are specific parameters where the standard first-order approach falls short and the DML estimator becomes suboptimal.

Lu: The theory shows that by embracing second-order methods like HOIF estimators, we can achieve a level of precision in AI that was previously inaccessible in these particular classes of data.

Meng: And I think it’s great news for our implementation teams because it tells us exactly when DML needs to be replaced with an advanced, biased-correct estimator, allowing us to build truly robust statistical pipelines.

Lalam: This paper inspires a culture of statistical responsibility, moving away from blind trust in single methods toward actively seeking out and correcting the systematic biases that exist within any complex estimation process.

Tom: It’s certainly a lot of nuance to take home today, but that’s what makes this field so exciting—knowing the boundaries is as valuable as knowing the limits.

Jane: We hope this has given our listeners a clear picture of how to use these insights when they encounter similar problems in their own work.

Lu: I think the theoretical implications for future research are truly vast, especially when we consider applying these second-order corrections to even more complex data structures.

Meng: It’s practical guidance that leads to better engineering design and Lalam is right, it helps us make informed decisions about reliability in the real world.

Lalam: Ultimately, it fosters a sense of transparency in how we use AI and statistics to find truth, knowing when our tools are less than perfect.

Tom: We’ve covered the core ideas beautifully today; I think we’ve given a really solid deep dive into this paper's findings before we move on to look at some more practical applications in the next segment.

Institute of Natural Sciences, Ministry of Education Local Science Center, School of Mathematical Sciences, Shanghai Jiao Tong University-Yale · Department of Biostatistics, Harvard T. H. Chan School of Public Health · Department of Epidemiology and Department of Biostatistics, Harvard T. H. Chan School of Public Health

math.ST, econ.EM, stat.ML, stat.TH

Submitted: 2026-06-21

Updated: 2026-09-04

Comments: 27 pages

License: http://creativecommons.org/publicdomain/zero/1.0/

Importance score: 86/100

The gist: " * Background and Motivation In statistical inference, a common goal is to infer a low-dimensional, smooth functional psi(theta) of the underlying data-generating law P theta.

Key concepts

Double Machine Learning (DML)
A statistical method discussed in the paper. While powerful for data analysis, the hosts note that DML is not always optimal; it may be suboptimal for certain parameters because it is a first-order estimator.
Asymptotically Inadmissible
A theoretical finding showing that an estimator, like DML, is not the most efficient choice in the long run. This means there are demonstrably better ways to estimate certain parameters than the standard DML approach.
Monotone Bias Class
A specific classification of functions where standard first-order techniques, including DML, fail. Identifying if a parameter falls into this class is key because it signals a systematic error that needs correction.
Second-Order Estimators (HOIF)
Advanced statistical methods proposed to fix DML's shortcomings. These estimators actively calculate and correct systematic bias by estimating higher-order influence functions, leading to greater precision.

Terminology

Summary

"

Background and Motivation

In statistical inference, a common goal is to infer a low-dimensional, smooth functional psi(theta) of the underlying data-generating law P theta. While the Double Machine Learning (DML) estimator psi b1,n has become standard practice for estimating such functionals (e.g., average treatment effect), traditional structural assumptions—such as smoothness or sparsity—are often required to ensure uniformly consistent estimators converge to psi at parametric rates.

The Structure-Agnostic (SA) Model

To circumvent these restrictive assumptions, Balakrishnan et al. (2026) introduced the structure-agnostic (SA) model, PSA(theta b, r n.), which does not impose explicit complexity reducing structural assumptions such as smoothness or sparsity. This model is parameterized by an initial estimator theta b treated as fixed and a set of convergence rates r n, where the subset of all possible data-generating distributions P' is defined such that theta bj - theta j2 r n,j.

Initial Findings and Contribution

Balakrishnan et al. (2026) established that under the SA model, the first-order DML estimator psi b1,n is minimax for several functionals. This paper builds upon this by focusing on the decision-theoretic properties of these estimators: asymptotic (in)admissibility.

The primary technical contribution of this paper is to show that psi b1,n is asymptotically inadmissible for two out of the three functionals studied in Balakrishnan et al. (2026). These two functionals are shown to fall into a class referred to as the monotone bias class. For this class, the paper exhibits second-order (U-statistic) estimators, psi b2,n, which asymptotically dominate DML estimators, and these estimators are identified as empirical higher-order influence function (HOIF) estimators.

Defining Inadmissibility and the Monotone Bias Class

To formalize the analysis, two definitions are provided:

  1. Asymptotic (in)admissibility: An estimator sequence psi n is asymptotically admissible if there exists no other sequence psi'n such that " n to infinity mse(psi'n) - mse(psi n) 0 and n to infinity < 0."

  2. Monotone Bias Class: A a functional psi belongs to this class if, for any r n, there exists another estimator psi b2,n such that bias(psi b2,n) / bias(psi b1,n) - 1 0 and the inequality is strict. Furthermore, the variances must satisfy specific conditions relating to o(1) bias reduction.

Key Results (The Four Functionals) The paper analyzes four examples of psi:

1. Quadratic Functional in the Gaussian Sequence Model (Section 2):

  • psi b1,n is found to be asymptotically inadmissible when r n squared n-1/2.

  • The second-order estimator psi b2,n(k) dominates psi b1,n in the scaled MSE.

2. Quadratic Density Integral Functional (Section 3):

  • psi b1,n is found to be asymptotically inadmissible when r n squared n-1/4.5.

  • The second-order estimator psi b2,n(phi k) corrects the bias of psi b1,n by estimating a lower bound of bias(psi b1,n) is r n 2.

3. Expected Conditional Covariance (Section 4):

  • This functional does not belong to the monotone bias class.

  • Both psi b1,n and psi b2,n are found to be asymptotically minimax, and neither estimator dominates the other.

4. Expected Conditional Variance (ECV) Functional (Section 15):

  • This functional belongs to the monotone bias class.

  • psi b1,n is found to be asymptotically inadmissible when r n squared n-1/2. The second-order estimator psi b2,n(phi k;) dominates psi b1,n.

The Assumption-Lean Model and Inference

The paper then contrasts the SA model with the assumption-lean model (PAL(theta)), which imposes no assumptions beyond the trivial hypothesis that the bias of any estimator... may be of order 1.

  • Under this assumption-lean model, for parameters in the monotone bias class, psi b2,n dominates psi b1,n asymptotically.

  • The paper notes that a Wald confidence interval centered at a DML estimator may under-cover under the assumption-lean model.

  • While Liu et al. (2024) developed valid tests for the monotone bias class, these tests are not consistent under the assumption-lean model, as no consistent tests exist (Robins and Ritov, 1997).

Conclusion

In conclusion, while the DML estimator psi b1,n is minimax under the SA model for certain functionals, it is asymptotically inadmissible for those in the monotone bias class. The paper demonstrates that second-order HOIF estimators (psi b2,n) provide a superior alternative to psi b1,n in this class. Furthermore, it highlights significant challenges regarding uncertainty quantification and inference when moving from the SA model to an assumption-lean model.

Improvements for AI systems

Based on a rigorous analysis of this paper, the improvements required are not merely algorithmic but fundamentally change how we define and validate statistical inference in complex, real-world data environments where structural assumptions fail.

As a fastidious AI researcher, I have identified three specific areas for improvement that elevate current AI systems from simple minimax estimators to robust, statistically rigorous solutions.

The Problem: Standard Double Machine Learning (DML) estimators (psi b1,n) are often minimax—meaning they achieve optimal theoretical rates—but the paper proves that for functionals in the monotone bias class, they are asymptotically inadmissible. This means a better estimator exists.

The Improvement: We must integrate and utilize second-order estimators (psi b2,n), which are constructed using High-Order Influence Functions (HOIFs).

  • Actionable Change: The AI model must be refactored to replace or augment the standard DML loss function with psi b2,n when estimating functionals like the quadratic functional in Gaussian sequence models or the quadratic density integral functional.

  • What the Improved System Can Do: It will achieve a lower Mean Squared Error (MSE) than standard DML in complex scenarios. This allows for more precise and reliable parameter estimation, moving beyond theoretical minimax bounds to practical optimal performance.

The Problem: Current statistical practice often relies on Wald confidence intervals (CI) centered on DML estimators (psi b1,n). The paper demonstrates that under the assumption-lean model (PAL), these CIs can under-cover the true parameter—meaning they are not reliable.

The Improvement: We must implement and standardize robust falsification tests based on the difference between psi b1,n and psi b2,n.

  • Actionable Change: Develop a suite of tests that specifically check the null hypothesis (H 0: bias(psi b1,n) is sufficiently small). If these tests reject H 0, they provide empirical evidence that the DML estimator is significantly underperforming or biased.

  • What the Improved System Can Do: It will provide honesty in uncertainty quantification. Instead of providing a confidence interval that might be falsely inflated due to missing structural assumptions, it will actively signal when statistical validity is compromised, allowing decision-makers to adjust their risk tolerance accordingly.

The Problem: The reliance on structural assumptions (e.g., smoothness or sparsity) is often an unwarranted convenience that leads to flawed inference when those assumptions are violated.

The Improvement: We must adopt the Assumption-Lean (PAL) Model as a baseline for all critical inference stages in a semiparametric estimation pipeline.

  • Actionable Change: Structure the model such to test if results hold even when no explicit structural assumptions are made (i.e, using the PAL model). This requires designing estimators and tests that are robust against this lack of prior knowledge.

  • What the Improved System Can Do: It will eliminate false confidence. The system will be inherently more resilient to model misspecification, ensuring that its conclusions about parameter uncertainty are derived from data-driven evidence rather than from pre-imposed constraints.

By implementing these improvements, the resulting AI system transitions from a Minimax Estimator (which is theoretically optimal but potentially practically suboptimal) to an Asymptotically Optimal and Statistically Valid System. It not only finds the best estimate (psi b2,n) but also provides a scientifically rigorous way to verify that its confidence intervals are trustworthy.

Abstract

Structure-agnostic (SA) models introduced by Balakrishnan et al. (2026) aim to reflect the general lack of knowledge of structural assumptions on data-generating laws such as smoothness or sparsity in practice. Roughly speaking, SA models restrict the observed-data generating law to be in some rn-neighborhood of (black-box machine learning) estimates, treated as given and fixed, where rn encodes the convergence rates of the estimates to the truth. Under SA models, Balakrishnan et al. (2026) show that the popular Double Machine Learning (DML) estimators for three functionals, the quadratic functional in the Gaussian sequence model, the quadratic density integral functional and the expected conditional covariance, are minimax. However, minimax estimators may be inadmissible. In this paper, we show that, for the first two of the three functionals, the DML estimator is asymptotically inadmissible under the SA model. In particular, we show that these two functionals fall into a class of functionals, which we refer to as the monotone bias class. For this class, we exhibit second-order (U-statistic) estimators, which asymptotically dominate DML estimators, under the SA model. These second-order estimators are empirical higher-order influence function (HOIF) estimators introduced in Liu et al. (2017). Furthermore, the empirical HOIF estimator, like the DML estimator, is minimax for the third functional (the expected conditional covariance), although neither asymptotically dominates the other.

Sources

Related papers