29. Schur Complement Proof of Constrained Variance Reduction in Ketz (2019)

29. Schur Complement Proof of Constrained Variance Reduction in Ketz (2019)

Setup

GMM with the **optimal weighting matrix** $S^{-1}$ and joint parameter $\theta^* = (\mu', \sigma^2)'$. The asymptotic variance $$V(\theta^*_0) = (G'S^{-1}G)^{-1} \equiv H^{-1}$$is the inverse of a **single** matrix (not a sandwich). Block-partition along $(\mu, \sigma^2)$: $$H=\begin{bmatrix}H_{\mu\mu} & H_{\mu\sigma}\\ H_{\sigma\mu} & H_{\sigma\sigma}\end{bmatrix}$$Compare two regimes: unconstrained $\text{Var}(\hat\mu_{\text{unc}}) = [H^{-1}]_{\mu\mu}$; constrained (impose $\sigma^2 \equiv 0$, i.e. the 50% boundary-hit realisation) $\text{Var}(\hat\mu_{\text{con}}) = H_{\mu\mu}^{-1}$.

Schur Complement Gives the PSD Ordering in One Line

The Schur complement identity yields$$[H^{-1}]_{\mu\mu} = \big(H_{\mu\mu} - H_{\mu\sigma}H_{\sigma\sigma}^{-1}H_{\sigma\mu}\big)^{-1} \tag{1}$$Since $H_{\sigma\sigma}^{-1} \succeq 0$, the outer product $H_{\mu\sigma}H_{\sigma\sigma}^{-1}H_{\sigma\mu} \succeq 0$, so$$H_{\mu\mu} - H_{\mu\sigma}H_{\sigma\sigma}^{-1}H_{\sigma\mu} \preceq H_{\mu\mu}$$Inversion reverses the PSD order on positive-definite matrices, giving$$[H^{-1}]_{\mu\mu} \succeq H_{\mu\mu}^{-1} \iff \text{Var}(\hat\mu_{\text{unc}}) \succeq \text{Var}(\hat\mu_{\text{con}}) \tag{2}$$This is a **matrix ordering**: every linear combination $a'\hat\mu$ has weakly larger variance under the unconstrained estimator, not just the coordinate $\hat\mu_{T,1}$.

The Gap Term Is Exactly "Covariance Contagion"

Define $\Delta \equiv H_{\mu\sigma}H_{\sigma\sigma}^{-1}H_{\sigma\mu}$. Then $\Delta = 0$ iff $H_{\mu\sigma}=0$, i.e. iff the optimally-weighted moment conditions for $\mu$ are orthogonal to those for $\sigma^2$. In the RC-Logit both are identified from the **same** choice-share moments, so $H_{\mu\sigma} \neq 0$ generically. **This term is the mathematical identity of the informal "covariance contagion"**, not a metaphor. Using the local scaling (see [[24. The Second Order Derivation in Ketz (2019)|Ketz §2.1 second-order derivation]]): $H_{\mu\sigma} = O(\sigma)$ (the even-function argument suppresses the leading σ-derivative by one order) and $H_{\sigma\sigma}^{-1} = O(\sigma^{-2})$, hence$$\Delta = O(\sigma) \cdot O(\sigma^{-2}) \cdot O(\sigma) = O(1)$$so **the gap does not vanish as $\sigma \to 0$**. Numerical check at $\sigma \in \{1, 0.1, 0.01\}$ gives $\|\Delta\| \approx 2.04, 0.35, 0.53$, and the minimum eigenvalue of $V_{\text{unc}} - V_{\text{con}}$ is $\ge 0$ to machine precision in all three cases.

Why "Optimal Weighting" Is a Precondition, Not Rhetoric

Under a generic weighting $W$, the asymptotic variance becomes the sandwich $V_W = (G'WG)^{-1}G'WSWG(G'WG)^{-1}$, which is **not** the inverse of any single matrix. Eq. (1)'s Schur-complement identity no longer applies, and the PSD ordering (2) is not guaranteed. The paper's clause "given the use of an optimal weighting matrix" is what activates the whole argument — it is a structural precondition, not decoration.

From (2) to 96.8% Over-Coverage

Two ingredients turn the variance inequality into an over-coverage story rather than a bias–variance trade-off:
  1. **The constraint is truth-consistent at the DGP.** The simulation sets $\sigma^2_0 = 0$, so $\hat\mu_c$ is unbiased and the constrained regime is a pure variance gain. If instead $\sigma^2_0 > 0$, $\hat\mu_c$ would be biased and MSE could go either way — (2) alone would not imply over-coverage.
  1. **The realised sampling distribution is a mixture.** With ≈50% of MC draws hitting the boundary, $\hat\mu_{T,1}$'s true distribution is a mixture of the constrained and unconstrained laws, with variance **strictly between** $H_{\mu\mu}^{-1}$ and $[H^{-1}]_{\mu\mu}$. But the reported Wald SE plugs in the pure unconstrained $[H^{-1}]_{\mu\mu}$ — strictly larger than the mixture's true variance. Larger reported SE → wider Wald interval → coverage above nominal, landing at **96.8%**.
This mixture-vs-plug-in mismatch is exactly what the [[28. Feldman-Cousins LR Test and Critical Value in Ketz (2019)|Feldman–Cousins LR construction]] (§5) sidesteps by not using the Wald form in the first place.
Loading...