Two theoretical traditions in statistics have been producing the same estimators for different reasons.
Note (2026-11-13; read against the paper's full text, version 3 of July 2026): the essay has the asymmetry between the two directions the wrong way round. In the paper, the direction from Neyman orthogonality to pathwise differentiability (its Theorem 1) needs a nondegenerate Jacobian and regularity conditions but 'without requiring any variation independence or product structure', whereas the direction from pathwise differentiability to Neyman orthogonality (its Theorem 2: a mean-zero estimating function whose value at the truth is an influence function is automatically Neyman orthogonal) 'does require local product structure'. So it is the direction the essay calls weaker, from pathwise differentiability to orthogonality, that needs the extra condition. The paper's equivalence is between a Neyman orthogonal estimating function and an influence function of a pathwise differentiable functional; it says an influence function is determined only up to the orthogonal complement of the tangent space, and that the efficient influence function is the one lying in the tangent space, with the two coinciding in the nonparametric model. The essay's phrase that the debiased machine-learning estimator and the semiparametric efficient estimator are formally equivalent is therefore the essay's compression: the paper does not claim that every orthogonal moment gives the efficient estimator. The 'local product structure' is, in the paper, an assumption it says the van der Laan and Robins framework used implicitly; the paper makes it explicit as a requirement that coordinate perturbations form regular submodels, a little weaker than a literal product neighbourhood. The essay's remarks on Le Cam, two decades of development and 'the same geometric object' are the essay's.
Neyman orthogonality is the central device of double/debiased machine learning. It constructs moment conditions that are insensitive to first-stage estimation errors — orthogonal to perturbations of the nuisance parameter. The estimator remains root-n consistent even when the nuisance parameter is estimated at slower-than-parametric rates. The tradition is modern, computational, designed for high-dimensional settings where machine learning estimates the nuisance.
Pathwise differentiability is the central concept of classical semiparametric efficiency theory. It identifies functionals that can be estimated at parametric rates by requiring that the functional's value changes smoothly along smooth paths through the model. The efficient influence function — the optimal estimating equation — falls out of the pathwise derivative. The tradition is older, geometric, rooted in Le Cam's theory of statistical experiments.
Chen, Kennedy, and Balakrishnan prove these are the same thing — under a regularity condition they call “local product structure.” When the nuisance parameter and the target parameter can be locally varied independently (the product structure), Neyman orthogonality and pathwise differentiability impose identical constraints on the estimating equation. The debiased ML estimator and the semiparametric efficient estimator are not merely similar but formally equivalent.
The equivalence is not symmetric in its requirements. Going from pathwise differentiability to Neyman orthogonality requires weaker conditions than the reverse direction. The two traditions ask different structural questions — one about smoothness along paths, the other about robustness to perturbations — that happen to have identical answers when the parameter space has the right geometry.
Two communities, two decades of parallel development, one estimator. The convergence was visible empirically — practitioners noticed the same formulas appearing in both frameworks. The proof explains why: the frameworks were always describing the same geometric object from different angles.