---
<!-- trackio-cell
{"type": "markdown", "id": "cell_1c53925c595b", "created_at": "2026-07-27T15:30:57+00:00", "title": "Claim 3 \u2014 AIPW preserves martingale, 98.8%/99.1% variance reduction", "pinned": true, "pinned_at": "2026-07-27T15:30:57+00:00"}
-->
## Claim 3 — AIPW augmentation preserves the martingale property

> "The paper characterizes an augmentation (AIPW) class for which per-arm estimation errors preserve the martingale property under staggered entry and delayed outcomes (Section 5, Theorem 5.1)."

Theorem 4.4 characterizes exactly which augmentations mₜ(w) preserve the per-arm martingale: event-time augmentations mᵢₜ(w)=fᵢ(w)·1[tᵢ(w)≤t] where fᵢ(w) is measurable strictly before the arm-w event time. A constant, pre-specified population mean μw trivially satisfies this (it depends on no random information at all). We estimate μ₀, μ₁ = E[Yᵢ(w) | event occurs] from a **separate, independent** N=30,000-unit oracle simulation (μ₀=1.3885, μ₁=0.6843) — NOT from the R=400 test replications below, so the augmentation constant is genuinely pre-specified rather than fit to the data under test.

We form the AIPW estimator r̂ₜᴬᴵᴾᵂ(w) = r̂ₜᴵᴾᵂ(w) + μw·Σᵢ(1−1[wᵢ=w]/π)·1[tᵢ(w)≤t] and re-run the exact same OLS increment-regression test used in Claim 1.

**Arm 0:** 3/4 pairs pass the martingale increment test (the one borderline pair, s=3→t=9, has slope -2.58 but p=0.09 and t_mean=0.27 — consistent with zero-mean, sparse-early-event noise, not a genuine violation). Variance at t=16: IPW **283.8** → AIPW **3.44**, a **98.8%** reduction.

**Arm 1:** 4/4 pairs pass cleanly. Variance at t=16: IPW **159.4** → AIPW **1.42**, a **99.1%** reduction.

Both arms show the AIPW error process retaining the martingale property (mean-zero, unpredictable increments) while achieving a large, measured variance reduction — exactly the paper's efficiency claim for event-time augmentations. As a sanity boundary: setting μw=0 in our AIPW code reduces it to plain IPW bit-for-bit (checked in `unit_tests.py`), confirming the augmentation machinery is implemented correctly relative to the m=0 special case.

**Verdict: VERIFIED** — AIPW preserves the martingale structure (7/8 pairs across both arms cleanly pass, one early-time borderline pair is finite-sample noise) while cutting error variance by 98.8%/99.1%.

````raw
{
 "mu0": 1.388549313697748,
 "mu1": 0.6843135182774553,
 "arm0": {
  "n_pass": 3,
  "n_pairs": 4,
  "var_ipw": 283.8183623615568,
  "var_aipw": 3.4397668931912193,
  "variance_reduction_frac": 0.9878803934158098
 },
 "arm1": {
  "n_pass": 4,
  "n_pairs": 4,
  "var_ipw": 159.40001698326353,
  "var_aipw": 1.4194461345744733,
  "variance_reduction_frac": 0.9910950691133018
 }
}
````