27. Deriving the Quasi Unconstrained Estimator (Eq. 16) in Ketz (2019)

27. Deriving the Quasi Unconstrained Estimator (Eq. 16) in Ketz (2019)

Settings

  1. Recall the theoretical GMM quadratic approximation of the paper's Eq. (2), expanded around the true parameter $\theta_T$: $$\mathcal{Q}_T(\theta) = Q_T(\theta_T) + G_T(\theta_T)'\mathcal{W}G_\theta(\theta - \theta_T) + \frac{1}{2}(\theta - \theta_T)'G_\theta'\mathcal{W}G_\theta(\theta - \theta_T) + R_T(\theta) \tag{1}$$where $G_T(\cdot)$ is the $l\times 1$ sample moment vector, $\mathcal{W}$ is the $l\times l$ (population) weight matrix, and $G_\theta$ is the $l\times k$ Jacobian evaluated at $\theta_T$.
  1. Sample analogue: replace population objects with consistent sample estimates $\hat{\mathcal{W}} \xrightarrow{p} \mathcal{W}$ and $\hat{G}_\theta \xrightarrow{p} G_\theta$, and replace the (unobservable) expansion center $\theta_T$ with a feasible consistent initial estimator $\hat{\theta}$ satisfying $\hat{\theta} \xrightarrow{p} \theta_T$. Drop the higher-order remainder $R_T$: $$\hat{\mathcal{Q}}_T(\theta) = Q_T(\hat{\theta}) + G_T(\hat{\theta})'\hat{\mathcal{W}}\hat{G}_\theta(\theta - \hat{\theta}) + \frac{1}{2}(\theta - \hat{\theta})'\hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta(\theta - \hat{\theta}) \tag{2}$$
  1. Objective: solve for the **quasi unconstrained estimator** $\tilde{\theta}_T \equiv \arg\min_{\theta \in \mathbb{R}^k}\hat{\mathcal{Q}}_T(\theta)$ in closed form, thereby filling in the algebra that the paper skips between Eq. (2) and Eq. (16).

Proof

Step 1: First-Order Condition

Take the gradient of $\hat{\mathcal{Q}}_T(\theta)$ with respect to $\theta$. Let $\delta \equiv \theta - \hat{\theta}$; the constant $Q_T(\hat{\theta})$ vanishes. Using the matrix-calculus identities $\nabla_\theta(a'\delta) = a$ and $\nabla_\theta(\delta' A \delta) = (A + A')\delta = 2A\delta$ (with $A = \hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta$ symmetric): $$\Big[\nabla_\theta \hat{\mathcal{Q}}_T\Big]_{k \times 1} = \hat{G}_\theta'\hat{\mathcal{W}}G_T(\hat{\theta}) + \hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta\,\delta \tag{3}$$
Setting the FOC to zero at $\theta = \tilde{\theta}_T$ (i.e., $\delta = \tilde{\theta}_T - \hat{\theta}$): $$\hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta\,(\tilde{\theta}_T - \hat{\theta}) = -\hat{G}_\theta'\hat{\mathcal{W}}G_T(\hat{\theta}) \tag{4}$$

Step 2: Invert the Sample Information Matrix

Define the sample information matrix $\hat{\mathcal{J}} \equiv \hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta$. Under the standard rank condition on $\hat{G}_\theta$ ($\text{rank}(\hat{G}_\theta) = k$) and positive definiteness of $\hat{\mathcal{W}}$, $\hat{\mathcal{J}}$ is positive definite — and therefore invertible — with probability approaching one. Left-multiply (4) by $\hat{\mathcal{J}}^{-1}$: $$\tilde{\theta}_T - \hat{\theta} = -\hat{\mathcal{J}}^{-1}\hat{G}_\theta'\hat{\mathcal{W}}G_T(\hat{\theta}) \tag{5}$$

Step 3: Rearrange for the Closed-Form Estimator

Move $\hat{\theta}$ to the right-hand side: $$\tilde{\theta}_T = \hat{\theta} - (\hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta)^{-1}\hat{G}_\theta'\hat{\mathcal{W}}G_T(\hat{\theta}) \tag{6}$$
which is exactly Eq. (16) in the paper (up to the paper's hat conventions for sample estimates).

Conclusion

The quasi unconstrained estimator $\tilde{\theta}_T$ admits the closed-form Gauss-Newton one-step update: $$\boxed{\;\tilde{\theta}_T = \hat{\theta} - (\hat{G}_\theta'\hat{\mathcal{W}}\hat{G}_\theta)^{-1}\hat{G}_\theta'\hat{\mathcal{W}}G_T(\hat{\theta})\;} \tag{7}$$
starting from any $\sqrt{T}$-consistent initial estimator $\hat{\theta}$. Q.E.D.

P.S. Corollary — Interior Case: $\tilde{\theta}_T^* = \hat{\theta}_T^*$

  1. Initialize (7) at the **constrained** GMM minimizer $\hat{\theta}_T^* \equiv \arg\min_{\theta \in \Theta} Q_T(\theta)$ — this is the estimator practitioners actually compute (BLP95, BLP99, RV14 all report $\hat{\sigma}_j$ obtained this way). $\Theta$ encodes the boundary constraints $g_j(\theta) \le 0$; in Ketz, $g_j(\theta) = -\sigma_j$, so $g_j \le 0 \iff \sigma_j \ge 0$. Assume $\hat{\theta}_T^*$ is **interior**: every constraint is strictly slack, $g_j(\hat{\theta}_T^*) < 0\ \forall j$, i.e., $\hat{\sigma}_j > 0$. (*P.S.* Notation clash warning: $g_j$ here is the KKT inequality-constraint function from optimization theory, **not** the GMM moment vector — which is $G_T$ throughout this note.)
  1. **KKT (Karush-Kuhn-Tucker) conditions** — the multi-inequality generalization of Lagrange multipliers. Form the Lagrangian (拉格朗日函数) $\mathcal{L}(\theta, \mu) \equiv Q_T(\theta) + \sum_j \mu_j\, g_j(\theta)$, one dual multiplier $\mu_j \ge 0$ per inequality. Any constrained local minimizer $\hat{\theta}_T^*$ satisfies **four** conditions simultaneously: $$\underbrace{\nabla_\theta Q_T(\hat{\theta}_T^*) + \sum_{j} \mu_j\, \nabla g_j(\hat{\theta}_T^*) = 0}_{\text{stationarity (驻点条件)}},\quad \underbrace{g_j(\hat{\theta}_T^*) \le 0}_{\text{primal feasibility (原始可行)}},\quad \underbrace{\mu_j \ge 0}_{\text{dual feasibility (对偶可行)}},\quad \underbrace{\mu_j\, g_j(\hat{\theta}_T^*) = 0}_{\text{complementary slackness (互补松弛)}} \tag{8}$$**Complementary slackness** is the load-bearing piece: the *product* $\mu_j\, g_j$ must vanish for each $j$, so either $g_j = 0$ (constraint **active/binding**, 边界起作用) or $\mu_j = 0$ (constraint **slack/silent**, 边界不起作用). Interior $\Rightarrow g_j(\hat{\theta}_T^*) < 0$ strictly $\forall j \Rightarrow \mu_j = 0\ \forall j \Rightarrow$ the multiplier sum in stationarity vanishes: $$\nabla_\theta Q_T(\hat{\theta}_T^*) = 0 \tag{9}$$In one sentence: at an interior optimum, the boundary "wall" isn't pushing back, so $\hat{\theta}_T^*$ satisfies the same FOC as if we had ignored $\Theta$ entirely.
  1. **Translate stationarity into moment orthogonality.** The standard GMM objective is $Q_T(\theta) = \tfrac{1}{2}\, G_T(\theta)'\, \mathcal{W}\, G_T(\theta)$; chain rule (with Jacobian $G_\theta \equiv \partial G_T/\partial \theta'$) gives $$\nabla_\theta Q_T(\theta) = G_\theta(\theta)'\, \mathcal{W}\, G_T(\theta)$$(without the $\tfrac{1}{2}$ prefactor this picks up a factor of $2$ — irrelevant for setting the gradient to zero). Combining with (9) and using sample estimates: $$\hat{G}_\theta'\,\hat{\mathcal{W}}\, G_T(\hat{\theta}_T^*) = 0 \tag{10}$$This is exactly the residual vector the Newton step in (7) tries to zero out — and it's already zero.
  1. Substitute (10) into (7) with $\hat{\theta} = \hat{\theta}_T^*$: $${\color{Orange} \tilde{\theta}_T^* = \hat{\theta}_T^* - \hat{\mathcal{J}}^{-1}\underbrace{\hat{G}_\theta'\,\hat{\mathcal{W}}\, G_T(\hat{\theta}_T^*)}_{=\, 0} = \hat{\theta}_T^*} \tag{11}$$The Gauss-Newton correction is the zero vector — the quadratic approximation $\hat{\mathcal{Q}}_T$ built around an interior stationary point of $Q_T$ has its minimizer exactly at the center. Q.E.D.
Loading...