ArXiv Authors Challenge Core Results in Optimally-Transported GMM Paper

A new arXiv comment challenges several theoretical results in a recently proposed optimally-transported generalized method of moments framework, known as OTGMM. The authors present counterexamples affecting claims about consistency, asymptotic normality, estimator equivalence and hypothesis testing.

The paper is not a report about a new AI model or software release. Its relevance to machine-learning researchers is methodological: statistical assumptions used to justify estimators, uncertainty estimates and tests can materially affect how data-driven systems are evaluated.

The comment, published on arXiv on November 6, 2026, responds to “Optimally-Transported Generalized Method of Moments” by Schennach and Starck. The authors say they also provide sufficient conditions under which the intended results can be recovered.

Key facts

Small-error guarantees questioned

The comment first examines the framework’s small-error analysis. According to the abstract, the assumptions attached to Theorems 2 and 3 are insufficient to establish the claimed consistency and asymptotic normality results.

Consistency concerns whether an estimator converges to the target value as the amount of data grows. Asymptotic normality describes the estimator’s limiting distribution and is commonly used to construct confidence intervals and statistical tests. If the required assumptions do not ensure either property, users cannot automatically rely on the associated large-sample conclusions.

The paper presents counterexamples rather than merely proposing alternative interpretations. It also supplies sufficient conditions intended to support results that follow the original motivation for OTGMM.

Large-error results receive broader corrections

The authors report more substantial problems with the framework’s large-error analysis. The original Theorem 4 describes an equivalence between the OTGMM estimator and a generalized method of moments estimator using modified moments.

The comment says this equivalence does not hold in the counterexamples it studies. It further reports that a component of the Lagrange multiplier has different limits under the two estimators. That difference affects the subsequent consistency and asymptotic normality claims associated with Theorems 5 and 6.

The paper characterizes the corrections to these later results as more substantial than those required for Theorems 2 and 3. The abstract does not establish how the revised results perform across practical datasets or software implementations.

Why this matters for ML-adjacent research

Generalized method of moments is an econometric estimation framework, not an AI model architecture. However, related assumptions about convergence, misspecification and uncertainty appear across statistical learning workflows, including evaluation of predictive models and causal analyses.

The comment’s discussion of misspecification is particularly relevant to researchers combining machine learning with moment-based estimation. It states that the population value selected under misspecification depends on the metric used. This means that an estimator’s behavior may reflect not only the data and moment conditions, but also the geometry chosen to measure discrepancies.

That does not show that machine-learning systems using related techniques are invalid. It does indicate that researchers should inspect the assumptions behind statistical guarantees rather than treating a named estimator as interchangeable with a more familiar method.

What readers should check next

The paper is an arXiv comment and should be read alongside the original OTGMM work. The abstract identifies the disputed theorems, but assessing the practical effect requires reviewing the full proofs, the counterexamples and the revised sufficient conditions.

Researchers considering OTGMM should verify which version of the original paper they are using, determine whether their application falls under the small-error or large-error setting, and examine whether their chosen metric is consequential under misspecification. They should also avoid assuming that the original testing procedure remains valid without checking the corrected assumptions and null distribution.

The available source does not establish whether the authors’ revised conditions have been independently verified or adopted in production econometric or machine-learning software. Those questions remain open until further technical review or implementation evidence is available.

Source: arXiv cs.LG, “Counterexamples and Sufficient Conditions: Comments on ‘Optimally-Transported Generalized Method of Moments’,” https://arxiv.org/abs/2609.20260