Refined Inference for Asymptotically Linear Estimators with Non-Negligible Second-Order Remainders

arXiv:2603.16833 · stat.ME, math.ST, stat.TH · Submitted 2026-03-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Refined Inference for Asymptotically Linear Estimators with Non-Negligible Second-Order Remainders".

Paper summary: Tom: So we're getting into the core of this paper, "Refined Inference for Asymptotically Linear Estimators with Non-Negligible Second-Order Remainders". The main thesis is that standard asymptotic linear estimators often rely on a simple expansion, but that expansion breaks down if the second-order remainder term isn't negligible in variance <ref:2603.16833#pg1>.

Jane: They study the specific regime where this remainder actually contributes variance at order n minus one, meaning the total sampling variance is different from what you get by just using the influence function approximation <ref:2603.16833#pg2>.

Lu: What they claim is that they derive a finite-sample variance decomposition that clearly separates the influence-function variance, the remainder variance, and their covariance <ref:2603.16833#pg5>.

Meng: The point is that standard sandwich estimators only focus on the first piece—the influence function part—and they completely ignore what’s happening with the remainder variance or that cross term <ref:2603.16833#pg8>.

Lalam: They then characterize when the sandwich estimator is actually consistent for the total sampling variance by looking at when this scaled remainder variance vanishes, specifically when n times the variance of the remainder goes to zero <ref:2603.16833#pg5>.

Tom: In short, they show that in some regimes where n times the remainder variance stays larger than zero, like when it equals a constant c times R, the sandwich estimator misses a first-order contribution <ref:2603.16833#pg10>.

Jane: This is important because it tells us that what we think is our variance estimate might be systematically off by a term that grows with n in those near-boundary situations <ref:2603.16833#pg7>.

Lu: They then move on to suggesting practical tools like the jackknife and bootstrap, which they show work well under mild additional conditions to recover consistent variance inference <ref:2603.16833#pg4>.

Meng: And for clustered data, they have an analytic expression showing how the intra-cluster correlation really amplifies that gap between the sandwich and the total variance through that remainder term <ref:2603.16833#pg5>.

Lalam: The authors verify these predictions using simulations of a cross-fitted nested-bridge estimator for stepped wedge clustered randomized trials, showing how this amplification is visible <ref:2603.16833#pg5>.

Tom: It’s interesting that they use a surrogate-assisted nested bridge estimator because it cleanly shows this near-boundary regime and the cluster structure makes the effect easy to see in practice <ref:2603.16833#pg4>.

Jane: It seems like they are building up a very concrete framework for when those traditional asymptotic linearity assumptions aren't quite enough to guarantee a good variance estimate <ref:2603.16833#pg7>.

Lu: The structure of the paper is organized to build this up, introducing the general framework first, then developing that variance decomposition in sections three and four <ref:2603.16833#pg4>.

Meng: If we want to deploy these models reliably, we have to know when these estimators are actually reliable under different conditions <ref:2603.16833#pg7>.

Lalam: That’s what this paper is about—refining inference for those tricky situations where the standard tools fall short <ref:2603.16833#pg1>.

Conclusion: Tom: So wrapping up "Refined Inference for Asymptotically Linear Estimators with Non-Negligible Second-Order Remainders". The authors are showing us that the standard sandwich estimator can be misleading if the second-order remainder isn't negligible in variance, especially when n times the remainder variance is a positive constant <ref:2603.16833#pg1>.

Jane: The paper is essentially about moving beyond just first-order inference and giving us a way to account for that missing piece of variance by decomposing the total sampling variance into three distinct components <ref:2603.16833#pg5>.

Lu: They provide specific tools like the jackknife and bootstrap as reliable alternatives when those conditions aren't met, showing they achieve correct variance targets under mild additional requirements <ref:2603.16833#pg4>.

Meng: For someone shipping models, the implication is that you can’t just rely on one formula; you need to check if your data structure—like clustering—is making things worse through that remainder term [two thousand six hundred three point one six eight three

stat.ME, math.ST, stat.TH

Submitted: 2026-03-17

Updated: 2026-11-16

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 90/100

The gist: Asymptotically linear estimators in semiparametric models are usually studied through a von Mises expansion in which first-order inference is based on the influencefunction variance…

Key concepts

Influence Function Variance
This is a key component in the von Mises expansion used to study asymptotically linear estimators. It represents the first-order contribution to the estimator's variance, which is what standard sandwich estimators primarily target.
Remainder Variance
This term captures the second-order error or remainder in the asymptotic expansion. When this term is not negligible, it causes standard first-order inference methods, like simple sandwich estimators, to be inaccurate.
Sandwich Validity
This refers to a condition where the standard sandwich variance estimator accurately targets the true sampling variance. The paper shows this validity depends on whether the scaled remainder variance vanishes, which is not always guaranteed in practice.

Terminology

Summary

Asymptotically linear estimators in semiparametric models are usually studied through a von Mises expansion in which first-order inference is based on the influencefunction variance, this reduction is valid only when the second-order remainder is negligible not only in probability but also in variance, a requirement not implied by the usual product-rate conditions ensuring asymptotic linearity <ref:2603. p. 0

Where the paper says it

Refined Inference for Asymptotically Linear Estimators with Non-Negligible Second-Order Remainders Amanda Lin Li ∗ May 26, 2026 Abstract Asymptotically linear estimators in semiparametric models are usually studied through a von Mises expansion in which first-order inference is based on the influencefunction variance. This reduction is valid only when the second-order remainder is negligible not only in probability but also in variance, a requirement not implied by the usual product-rate conditions ensuring asymptotic linearity.

16833#pg2>. The paper derives a finite-sample variance decomposition separating the influence-function variance, the remainder variance, and their covariance, and characterizes sandwich validity through the vanishing of scaled remainder variance <ref:2603. p. 0
Where the paper says it

We study the regime in which the remainder contributes variance at order n −1 , so that the total sampling variance differs from the standard influence-function approximation by a non-vanishing first-order term. We derive a finite-sample variance decomposition separating the influence-function variance, the remainder variance, and their covariance, and characterize sandwich validity through the vanishing of scaled remainder variance: under a negligible cross term, the sandwich estimator is consistent for the total sampling variance when n Var(Rrem) → 0 and materially underestimates it in the complementary near-boundary regime n Var(Rrem) → cR > 0.

16833#pg5>.

Variance Decomposition

The sampling variance of an asymptotically linear estimator decomposes into three pieces: the influence-function component, a remainder-variance component, and a cross term <ref:2603.16833#pg6>. The sandwich variance estimator targets only the influence-function component and does not by itself account for Var(Rrem) or the cross term <ref:2603. p. 4

Where the paper says it

| {z } cross term . (4) The sandwich variance estimator Var dSand(Ψ) ˆ targets only the influence-function component σ 2 EIF/n and does not by itself account for Var(Rrem) or the cross term.

16833#pg8>. In the near-boundary regime, where n Var(Rrem) → cR > 0, the sandwich misses a first-order contribution <ref:2603. p. 1
Where the paper says it

From this decomposition we obtain a sharp criterion. Under a negligible cross term, the sandwich variance is consistent for the total sampling variance when the remainder variance, scaled by n, tends to zero; in the complementary near-boundary regime, it misses a first-order contribution.

16833#pg10>.

Refined Variance Estimators

When the sharp condition fails, two familiar resampling tools recover consistent variance inference under mild additional conditions: the leave-one-out jackknife and the pairs bootstrap. The leave-one-out jackknife works through a self-normalization mechanism that absorbs the remainder contribution automatically. p. 2

Where the paper says it

The jackknife works through a self-normalization mechanism that absorbs the remainder contribution automatically; the bootstrap works under a Mallows-distance condition that is satisfied in the settings of practical interest [Bickel and Freedman, 1981, Cheng and Huang, 2010]. For clustered data we derive an analytic expression showing how intra-cluster correlation amplifies the gap between the sandwich and the total variance, and we verify the predictions in simulations of a cross-fitted nested-bridge estimator for stepped-wedge clusterrandomized trials.

The bootstrap works under a Mallows–2 condition that is satisfied in the settings of practical interest. p. 2
Where the paper says it

The jackknife works through a self-normalization mechanism that absorbs the remainder contribution automatically; the bootstrap works under a Mallows-distance condition that is satisfied in the settings of practical interest [Bickel and Freedman, 1981, Cheng and Huang, 2010]. For clustered data we derive an analytic expression showing how intra-cluster correlation amplifies the gap between the sandwich and the total variance, and we verify the predictions in simulations of a cross-fitted nested-bridge estimator for stepped-wedge clusterrandomized trials.

Clustered Data Amplification

For clustered data, an analytic expression shows how intra-cluster correlation amplifies the gap between the sandwich and the total variance through the remainder term. The variance of the remainder satisfies Var(R2) = V0 z i. p. 0

Where the paper says it

We study the regime in which the remainder contributes variance at order n −1 , so that the total sampling variance differs from the standard influence-function approximation by a non-vanishing first-order term. We derive a finite-sample variance decomposition separating the influence-function variance, the remainder variance, and their covariance, and characterize sandwich validity through the vanishing of scaled remainder variance: under a negligible cross term, the sandwich estimator is consistent for the total sampling variance when n Var(Rrem) → 0 and materially underestimates it in the complementary near-boundary regime n Var(Rrem) → cR > 0.

i.d. baseline + ρICC · m · VICC z ICC amplification + O(J −2), for positive constants V0, VICC. p. 2
Where the paper says it

The empirical illustrations come from a surrogate-assisted nested-bridge estimator developed in a companion paper [Li et al. We use it because it cleanly exhibits the near-boundary regime and because its cluster structure makes the intracluster amplification visible.

Asymptotic Validity and Coverage

The jackknife and bootstrap procedures target the correct variance (σ 2EIF+cR)/n. The studentized bootstrap pivot satisfies T boot n = √n(Ψˆ − Ψ(P0))/q Var dBoot ⇒ N(0, 1). p. 3

Where the paper says it

2 (First-Order von Mises Expansion). The estimator Ψˆ satisfies (1) with Rrem = OP (n −1/2 ) and D∗ ∈ L 2 0 (P0) having σ 2 EIF = EP0 [(D∗ ) 2 ] ∈ (0, ∞).

Similarly, the jackknife Wald interval achieves asymptotically exact coverage P(Ψˆ±zα/2q Var dJK) ⇒ N(0, 1). p. 7
Where the paper says it

, n, let Ψˆ (−i) denote the ALE computed on the sample with unit i removed. The jackknife variance estimator is: Var dJK(Ψ) = ˆ n − 1 n Xn i=1 Ψˆ (−i) − Ψ¯ (−·) 2 , (6) where Ψ¯ (−·) = n −1 P i Ψˆ (−i) .

Comparison of Estimators

The three estimators are first-order equivalent under Assumptions 2.2–4.2, with Var dJK = Var dHC + oP (n −1). p. 7

Where the paper says it

, n, let Ψˆ (−i) denote the ALE computed on the sample with unit i removed. The jackknife variance estimator is: Var dJK(Ψ) = ˆ n − 1 n Xn i=1 Ψˆ (−i) − Ψ¯ (−·) 2 , (6) where Ψ¯ (−·) = n −1 P i Ψˆ (−i) .

The jackknife requires J refits and tJ−1 critical values, while the bootstrap yields BCa intervals and second-order accuracy. p. 0
Where the paper says it

Simulations illustrate the regime and confirm the predicted coverage behaviour of the competing variance estimators. Keywords: asymptotic linearity; semiparametric inference; second-order remainder; jackknife; bootstrap; variance decomposition; clustered data.

The HC sandwich is numerically identical to the jackknife, expressing the correction as a multiplicative inflation of the sandwich. p. 0
Where the paper says it

We derive a finite-sample variance decomposition separating the influence-function variance, the remainder variance, and their covariance, and characterize sandwich validity through the vanishing of scaled remainder variance: under a negligible cross term, the sandwich estimator is consistent for the total sampling variance when n Var(Rrem) → 0 and materially underestimates it in the complementary near-boundary regime n Var(Rrem) → cR > 0. We then establish asymptotic validity of two refined procedures in the nearboundary regime: the leave-one-out jackknife and the pairs bootstrap.

The results confirm structural near-boundary operation when ˆρ remains above 1. p. 5
Where the paper says it

In the running example, the ratio is ≈ 0. 63 (equivalently, Var dSand is ≈ 63% of the empirical variance) and is constant across J ∈ {10, 100}, confirming near-boundary operation.

14 across all (J,ICC) cells. p. 5
Where the paper says it

Proposition 3. 1(c) has a simple consequence: if the empirical variance ratio Var dSand/Var demp remains approximately constant across increasing sample sizes, the estimator is operating at or near the product-rate boundary.

Strong Decay Regime

In the strong-decay regime where n Var(Rrem) → 0, the variance ratio ˆρ decreases monotonically to 1.0 as n grows when nuisance estimation operates in the strong-decay regime. p. 1

Where the paper says it

1 Introduction The bookkeeping step is doing more work than it appears. In the strong-decay regime, where n Var(Rrem) → 0, the first-order linearization is sufficient both for centering and for first-order variance estimation.

This contrasts with the persistent near-boundary behaviour of Table 1 in the main paper. p. 1
Where the paper says it

We call this the near-boundary regime. Empirical traces of this behaviour have appeared in several specific contexts.

The secondary simulation verifies that under OP (n −1/2) parametric nuisance convergence, n Var(Rrem) → 0. p. 1
Where the paper says it

When nuisance estimators are flexible, such as adaptive nonparametric regressions, crossfitted machine learners, or sieve estimators [Chernozhukov et al. , 2018], their convergence rate often hovers near the n −1/4 threshold that the product-rate condition allows.

Secondary Simulation Results

The secondary AIPW simulation validates Theorem 4.2, showing that ρˆ decreases monotonically from 1. p. 0

Where the paper says it

Jackknife validity is obtained through a self-normalization argument; bootstrap validity is established directly under a Mallows–2 condition. We also extend the theory to clustered data and derive an analytic expression showing how intra-cluster correlation amplifies the sandwich gap through the remainder term.

040 to 1.003, consistent with the sharp sandwich characterization. p. 5
Where the paper says it

Without cross-fitting, the cross term can be substantial and of either sign. 4 Sandwich Consistency: A Sharp Characterization We now give a precise statement of when the sandwich is and is not consistent for the total variance [Newey, 1994].

This confirms that under OP (n −1/2) nuisance convergence, nVar dSand P → σ 2EIF and nVar dJK P → σ 2EIF, so no correction beyond the sandwich is needed. p. 9
Where the paper says it

The jackknife tracks the total sampling variance; the sandwich stays at σ 2 EIF. In both regimes, E[Var dJK(Ψ)] ˆ −Var(Ψˆ −Ψ(P0)) = O(n −3/2 ), whereas E[Var dSand(Ψ)] ˆ −Var(Ψˆ − Ψ(P0)) = −cR/n + o(n −1 ).

The coverage probabilities lie within ±2 Monte Carlo standard errors of 0. p. 0
Where the paper says it

We also extend the theory to clustered data and derive an analytic expression showing how intra-cluster correlation amplifies the sandwich gap through the remainder term. Simulations illustrate the regime and confirm the predicted coverage behaviour of the competing variance estimators.

95.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI systems can now reliably perform first-order variance inference for asymptotically linear estimators in semiparametric models when second-order remainders contribute variance at order n minus 1, as specified by the near-boundary regime.

  2. The improved system can distinguish between asymptotic linearity and valid first-order variance inference, allowing it to identify when the standard product-rate condition is insufficient for robust inference in finite samples.

  3. It can utilize the leave one out jackknife or pairs bootstrap procedures to recover consistent variance estimation under the near boundary regime where n Var(Rrem) goes to cR greater than zero.

  4. The system can provide a finite sample variance decomposition separating influence function variance, remainder variance, and covariance term, allowing for a sharp criterion for sandwich validity based on the vanishing of scaled remainder variance.

  5. For clustered data settings, the improved system can quantify how intra-cluster correlation amplifies the variance gap between the sandwich estimator and total sampling variance through an analytic expression showing this amplification.

  6. It can employ a method called HC corrected sandwich which inflates the sandwich by an estimated variance ratio, providing a measure of how far an estimator is from strong decay regime, as defined by Var dHC(Psi) = hat rho · Var dSand(Psi).

  7. The system can perform bias corrected and accelerated confidence intervals using the pairs bootstrap, enabling second order coverage accuracy that corrects for skewness in the remainder distribution.

  8. It can establish asymptotic normality conditions for studentized jackknife and bootstrap pivots under Assumption 4.3, ensuring that studentized pivots converge to N(0, 1) with asymptotically exact coverage when Mallows-2 consistency is met.

Abstract

First-order semiparametric inference typically uses the same influence-function linearization to justify both centering and sandwich variance estimation. We show that these are distinct questions. In a near-boundary regime, the second-order remainder can be nearly mean zero yet large enough to contribute O(n-1) sampling variation. When plug-in influence-function error and the cross term are negligible, the usual sandwich estimator captures only the influence-function component, and Wald intervals can undercover even when the point estimator is nearly unbiased. We derive a finite-sample variance decomposition and show that, when the cross term is asymptotically negligible, sandwich consistency is characterized by the condition nVar(Rrem) --> 0: the sandwich consistently estimates total variance in the strong-decay regime where this holds, and materially underestimates it in the near-boundary regime where nVar(Rrem) -->cR > 0. We then study two corrections. Under a leave-one-out remainder linearization condition, the jackknife consistently estimates the total variance; under a conditional Mallows-2 bootstrap condition, the pairs bootstrap does likewise. For clustered data we distinguish within-cluster dependence from first-order influence-function mismatch. Simulations for a surrogate-assisted targeted learning estimator in stepped-wedge cluster-randomized trials illustrate variance discrepancies and the limits of jackknife correction.

Sources

Related papers