arxiv: 2603.15817 · v2 · submitted 2026-03-16 · 📊 stat.ME · math.ST· stat.TH

Recognition: 2 theorem links

· Lean Theorem

On the Equivalence between Neyman Orthogonality and Pathwise Differentiability

Yuxi Chen , Edward H. Kennedy , Sivaraman Balakrishnan

Authors on Pith no claims yet

Pith reviewed 2026-05-15 09:44 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.TH

keywords Neyman orthogonalitypathwise differentiabilitydouble machine learningsemiparametric estimationnuisance parametersaverage treatment effectpartially linear model

0 comments

The pith

A local product structure on target and nuisance parameters makes Neyman orthogonality equivalent to pathwise differentiability.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that in the semiparametric framework, Neyman orthogonality and pathwise differentiability coincide exactly when the target and nuisance parameters satisfy a local product structure. This regularity condition was already implicit in earlier work but not stated explicitly, which explains why the two ideas often produce identical debiased estimators despite originating in separate literatures. A reader cares because both concepts drive double and debiased machine learning, so their formal link clarifies when either tool can be used without loss of generality. The work further demonstrates that the forward and reverse directions of the equivalence rest on different structural requirements.

Core claim

Under the local product structure on the relationship between target and nuisance parameters, Neyman orthogonality holds if and only if the target functional is pathwise differentiable, with each implication relying on distinct additional conditions on the model.

What carries the argument

The local product structure assumption, a regularity condition ensuring local decomposition between the tangent spaces of the target and nuisance parameters.

If this is right

The two debiasing frameworks can be used interchangeably in models that obey the local product structure, such as average treatment effect estimation.
Derivation of influence functions can proceed from either Neyman orthogonality or pathwise differentiability when the structure holds.
The differing structural requirements for each direction identify cases where one concept is easier to verify than the other.
The equivalence extends directly to the slope parameter in partially linear models and to expected density estimation.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Practitioners could first check for local product structure to decide whether to derive estimators via the orthogonality or differentiability route.
The result may extend to other functionals in semiparametric models beyond the three examples, provided the structure can be verified.
It offers a route to simplify proofs of asymptotic linearity by switching between the two equivalent conditions.

Load-bearing premise

The target and nuisance parameters relate through a local product structure in the semiparametric model.

What would settle it

A concrete semiparametric model that satisfies all stated conditions except local product structure yet has Neyman orthogonality without pathwise differentiability, or vice versa.

read the original abstract

It has been frequently observed that Neyman orthogonality, the central device underlying double/debiased machine learning (Chernozhukov et al., 2018), and pathwise differentiability, a cornerstone concept from semiparametric theory, often lead to the same debiased estimators in practice. Despite the widespread adoption of both ideas, the precise nature of this equivalence has remained elusive, with the two concepts having been developed in largely separate traditions. In this work, we revisit the semiparametric framework of van der Laan and Robins (2003) and identify an implicit regularity assumption on the relationship between target and nuisance parameters -- a local product structure -- that allows us to establish a formal equivalence between Neyman orthogonality and pathwise differentiability. We also show that the two directions of this equivalence impose fundamentally different structural requirements. Finally, we illustrate the theory through three detailed examples of estimating the average treatment effect and expected density in a nonparametric model, as well as the slope in a partially linear model. This helps clarify the relationship between these two foundational frameworks and provides a useful reference for practitioners working at their intersection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows Neyman orthogonality equals pathwise differentiability once a local product structure holds between target and nuisance parameters, with the two directions needing different conditions.

read the letter

The punchline is that this paper identifies a local product structure on the target and nuisance parameters as the condition that makes Neyman orthogonality equivalent to pathwise differentiability. The two directions of the equivalence turn out to require different amounts of structure, which is a useful distinction the authors draw out. What is new is the explicit isolation of that local product structure from the van der Laan and Robins semiparametric setup. Earlier papers noticed that the two concepts often produce the same estimators but did not spell out the precise regularity that makes the equivalence hold. The authors also show that the implication is not symmetric. Their three examples walk through the average treatment effect in a nonparametric model, expected density estimation, and the slope parameter in a partially linear model. These are standard problems, and the derivations there help show how the abstract condition translates to concrete influence functions and orthogonal scores. The paper does a good job of connecting the double/debiased machine learning literature with classical semiparametric theory. Anyone who has wondered why their Neyman-orthogonal estimator also satisfies the pathwise differentiability condition will find the explanation here. The argument is internally consistent and avoids circularity by working from the external framework rather than defining things in terms of the estimators themselves. The soft spot is that the local product structure is presented as an implicit regularity rather than something the paper derives or tests extensively. If that structure is violated, the equivalence fails, and the paper does not spend much time on diagnostics or counterexamples. The examples are detailed but remain within familiar territory, so they do not push the theory into more exotic models where the condition might be harder to verify. The full proofs are referenced but would need close checking in review. This paper is for statisticians and econometricians who work at the intersection of semiparametrics and causal inference. A reader who already knows both Neyman orthogonality and influence functions will get the most value. It deserves a serious referee because the clarification is precise and the examples make it practical. I would recommend sending it to peer review.

Referee Report

3 major / 2 minor

Summary. The manuscript claims that an implicit local product structure assumption on the relationship between target and nuisance parameters (drawn from the semiparametric framework of van der Laan and Robins 2003) establishes a formal equivalence between Neyman orthogonality and pathwise differentiability. The two directions of the equivalence are shown to impose distinct structural requirements, and the theory is illustrated via three examples: estimation of the average treatment effect, expected density in a nonparametric model, and the slope parameter in a partially linear model.

Significance. If the derivations hold, the result bridges two largely separate traditions in semiparametric statistics and double/debiased machine learning, clarifying why the concepts frequently produce the same estimators in practice. It supplies a precise regularity condition that practitioners can check and offers a reference point for work at the intersection of these frameworks.

major comments (3)

[Theoretical development (around the statement of the main equivalence)] The central equivalence relies on the local product structure being both implicit in van der Laan and Robins (2003) and sufficient to equate the two notions; the manuscript must supply the explicit steps showing how this structure converts the pathwise derivative condition into the Neyman orthogonality condition (and vice versa) without additional regularity.
[Examples (average treatment effect, expected density, partially linear model)] The three examples are presented as detailed illustrations, yet the provided description indicates that full step-by-step verification of the equivalence (including explicit computation of the relevant influence functions and verification that the local product structure holds) is not supplied; this verification is load-bearing for the claim that the theory applies in standard settings.
[Discussion of directional differences] The claim that the two directions impose fundamentally different structural requirements needs concrete counter-examples or minimal conditions showing a setting where one direction holds but the other fails; without such separation the distinction remains asserted rather than demonstrated.

minor comments (2)

[Introduction / Setup] Notation for the local product structure should be introduced with a formal definition (e.g., a displayed equation) before it is used in the equivalence statements.
[References] The abstract and introduction cite Chernozhukov et al. (2018) and van der Laan and Robins (2003); ensure the reference list contains the full bibliographic details for both.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their detailed and constructive comments, which highlight opportunities to strengthen the clarity and rigor of our derivations and examples. We agree that explicit steps and verifications will improve the manuscript and will incorporate them in the revision. We address each major comment below.

read point-by-point responses

Referee: The central equivalence relies on the local product structure being both implicit in van der Laan and Robins (2003) and sufficient to equate the two notions; the manuscript must supply the explicit steps showing how this structure converts the pathwise derivative condition into the Neyman orthogonality condition (and vice versa) without additional regularity.

Authors: We agree that the proof steps should be expanded for transparency. In the revised manuscript we will add a new subsection in Section 3 that walks through the equivalence in both directions, starting from the definition of pathwise differentiability, applying the local product structure to separate the target and nuisance directions, and arriving at the Neyman orthogonality condition (and the converse) using only the assumptions already stated in the paper. revision: yes
Referee: The three examples are presented as detailed illustrations, yet the provided description indicates that full step-by-step verification of the equivalence (including explicit computation of the relevant influence functions and verification that the local product structure holds) is not supplied; this verification is load-bearing for the claim that the theory applies in standard settings.

Authors: We will supply the requested verifications. The revised version will include an expanded appendix (or dedicated subsection) for each of the three examples that (i) explicitly computes the relevant influence functions, (ii) verifies that the local product structure holds, and (iii) confirms that the two notions coincide under the stated conditions. revision: yes
Referee: The claim that the two directions impose fundamentally different structural requirements needs concrete counter-examples or minimal conditions showing a setting where one direction holds but the other fails; without such separation the distinction remains asserted rather than demonstrated.

Authors: We will add concrete counter-examples. In the revision we will construct two minimal parametric models (one where pathwise differentiability holds but Neyman orthogonality fails, and one in the opposite direction) that satisfy the local product structure, thereby demonstrating the distinct structural requirements of each direction. revision: yes

Circularity Check

0 steps flagged

No significant circularity; equivalence derived from external framework

full rationale

The paper's central derivation identifies a local product structure as an implicit regularity condition in the van der Laan and Robins (2003) semiparametric framework, then uses it to prove equivalence between Neyman orthogonality and pathwise differentiability. This is a self-contained theoretical argument resting on external literature rather than any self-definition, fitted input renamed as prediction, or load-bearing self-citation chain. The three illustrative examples (ATE, expected density, partially linear model) serve only to clarify the established equivalence and do not reduce the main claim to the paper's own inputs by construction. No steps satisfy any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the local product structure as a domain assumption within the van der Laan and Robins semiparametric framework; no free parameters or invented entities are introduced.

axioms (1)

domain assumption local product structure between target and nuisance parameters
Identified as the implicit regularity assumption that enables the equivalence between the two concepts.

pith-pipeline@v0.9.0 · 5505 in / 1172 out tokens · 27054 ms · 2026-05-15T09:44:58.288005+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

identify an implicit regularity assumption... local product structure... equivalence between Neyman orthogonality and pathwise differentiability
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Neyman orthogonal estimating function... influence function... pathwise differentiability

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Doubly Robust Instrumented Difference-in-Differences
econ.EM 2026-05 unverdicted novelty 7.0

Derives the efficient influence function and doubly robust estimators for the local average treatment effect on the treated in instrumented DiD designs with staggered exposure and covariates.
Calibeating Prediction-Powered Inference
stat.ML 2026-04 unverdicted novelty 7.0

Post-hoc calibration of miscalibrated black-box predictions on a labeled sample improves efficiency of prediction-powered inference for semisupervised mean estimation.