23. Derivation of the Bivariate Projected Asymptotic Distribution

23. Derivation of the Bivariate Projected Asymptotic Distribution


Article: "[[33+. On asymptotic size distortions in the random coefficients logit model]]"
Part: P6

Settings

  1. Consider the case where there are $K_1$ unconstrained parameters (corresponding to the block $\lambda_1 \in \mathbb{R}^{K_1}$) and $K_2$ parameters constrained by a lower bound (corresponding to the block $\lambda_2 \in [-h, \infty]^{K_2}$).
  1. Let $\lambda = \begin{pmatrix} \lambda_1 \\ \lambda_2 \end{pmatrix}$ and the unconstrained normal random vector be $\mathcal{Z} = \begin{pmatrix} \mathcal{Z}_1 \\ \mathcal{Z}_2 \end{pmatrix}$.
  1. Let the symmetric information matrix be $\mathcal{J} = \begin{pmatrix} \mathcal{J}_1 & \mathcal{J}_{1,2} \\ \mathcal{J}_{1,2}' & \mathcal{J}_2 \end{pmatrix}$, where $\mathcal{J}_1$ is $K_1 \times K_1$, $\mathcal{J}_2$ is $K_2 \times K_2$, and $\mathcal{J}_{1,2}$ is $K_1 \times K_2$.
  1. **Objective:** Find the constrained minimizer $\hat{\lambda}_h$ by projecting $\mathcal{Z}$ onto the constrained parameter space $\Lambda_h = \mathbb{R}^{K_1} \times [-h, \infty]^{K_2}$. This requires minimizing the quadratic distance function $q(\lambda) = (\lambda - \mathcal{Z})'\mathcal{J}(\lambda - \mathcal{Z})$.

Proof

Step 1: Block Expansion of the Quadratic Form

Define the difference vector $X = \lambda - \mathcal{Z} = \begin{pmatrix} X_1 \\ X_2 \end{pmatrix}$ with $X_i = \lambda_i - \mathcal{Z}_i$. The quadratic form can be expanded using block matrix multiplication: $$q(\lambda) = X' \mathcal{J} X = (X_1', X_2') \begin{pmatrix} \mathcal{J}_1 & \mathcal{J}_{1,2} \\ \mathcal{J}_{1,2}' & \mathcal{J}_2 \end{pmatrix} \begin{pmatrix} X_1 \\ X_2 \end{pmatrix} \tag{1}$$
Multiply the block matrix by the column block vector: $$= (X_1', X_2') \begin{pmatrix} \mathcal{J}_1 X_1 + \mathcal{J}_{1,2} X_2 \\ \mathcal{J}_{1,2}' X_1 + \mathcal{J}_2 X_2 \end{pmatrix} \tag{2}$$
Multiply the row block vector by the resulting column vector: $$= X_1' \mathcal{J}_1 X_1 + X_1' \mathcal{J}_{1,2} X_2 + X_2' \mathcal{J}_{1,2}' X_1 + X_2' \mathcal{J}_2 X_2 \tag{3}$$
Since $X_2' \mathcal{J}_{1,2}' X_1 = (X_1' \mathcal{J}_{1,2} X_2)' = X_1' \mathcal{J}_{1,2} X_2$ (a scalar equals its transpose), combine the symmetric cross-terms and substitute $X$ back: $$q(\lambda_1, \lambda_2) = (\lambda_1 - \mathcal{Z}_1)' \mathcal{J}_1 (\lambda_1 - \mathcal{Z}_1) + 2(\lambda_1 - \mathcal{Z}_1)' \mathcal{J}_{1,2} (\lambda_2 - \mathcal{Z}_2) + (\lambda_2 - \mathcal{Z}_2)' \mathcal{J}_2 (\lambda_2 - \mathcal{Z}_2) \tag{4}$$

Step 2: First-Order Condition for the Unconstrained Block

Since $\lambda_1 \in \mathbb{R}^{K_1}$ is unconstrained, take the gradient of $q(\lambda_1, \lambda_2)$ with respect to $\lambda_1$ and set it to zero: $$\nabla_{\lambda_1} q = 2\mathcal{J}_1 (\lambda_1 - \mathcal{Z}_1) + 2\mathcal{J}_{1,2} (\lambda_2 - \mathcal{Z}_2) = 0 \tag{5}$$
Since $\mathcal{J}_1$ is positive definite (hence invertible), solve for $\lambda_1$ to obtain its optimal value conditional on $\lambda_2$: $$\lambda_1 = \mathcal{Z}_1 - \mathcal{J}_1^{-1} \mathcal{J}_{1,2} (\lambda_2 - \mathcal{Z}_2) \tag{6}$$

Step 3: Boundary Truncation for the Constrained Block

Substitute equation (6) into the objective to concentrate out $\lambda_1$.
From (6), $\lambda_1 - \mathcal{Z}_1 = -\mathcal{J}_1^{-1} \mathcal{J}_{1,2} (\lambda_2 - \mathcal{Z}_2)$. Let $u = \lambda_2 - \mathcal{Z}_2$ and $A = \mathcal{J}_1^{-1} \mathcal{J}_{1,2}$, so $\lambda_1 - \mathcal{Z}_1 = -Au$. Substitute each term of equation (4):
  • Term 1: $(\lambda_1 - \mathcal{Z}_1)' \mathcal{J}_1 (\lambda_1 - \mathcal{Z}_1) = (-Au)' \mathcal{J}_1 (-Au) = u' A' \mathcal{J}_1 A u = u' \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_1 \mathcal{J}_1^{-1} \mathcal{J}_{1,2} u = u' \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} u$ (using $A' = \mathcal{J}_{1,2}' (\mathcal{J}_1^{-1})' = \mathcal{J}_{1,2}' \mathcal{J}_1^{-1}$ by symmetry of $\mathcal{J}_1$).
  • Term 2: $2(\lambda_1 - \mathcal{Z}_1)' \mathcal{J}_{1,2} (\lambda_2 - \mathcal{Z}_2) = 2(-Au)' \mathcal{J}_{1,2} u = -2 u' \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} u$.
  • Term 3: $(\lambda_2 - \mathcal{Z}_2)' \mathcal{J}_2 (\lambda_2 - \mathcal{Z}_2) = u' \mathcal{J}_2 u$.
Summing the three terms cancels one copy of the cross-quadratic and leaves: $$q^*(\lambda_2) = u' \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} u - 2 u' \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} u + u' \mathcal{J}_2 u = u' \big( \mathcal{J}_2 - \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} \big) u \tag{7}$$
Restoring $u = \lambda_2 - \mathcal{Z}_2$ yields the concentrated quadratic in $\lambda_2$: $$q^*(\lambda_2) = (\lambda_2 - \mathcal{Z}_2)' \underbrace{\big( \mathcal{J}_2 - \mathcal{J}_{1,2}' \mathcal{J}_1^{-1} \mathcal{J}_{1,2} \big)}_{\text{Schur complement}} (\lambda_2 - \mathcal{Z}_2) \tag{8}$$
The bracketed matrix is the Schur complement (舒尔补) of $\mathcal{J}_1$ in $\mathcal{J}$, and inherits positive definiteness from $\mathcal{J} \succ 0$. Because of the componentwise hard boundary $\lambda_2 \ge -h$, the optimal solution is the unconstrained minimizer $\mathcal{Z}_2$ truncated at the lower bound $-h$: $$\hat{\lambda}_2 = \max\{-h, \mathcal{Z}_2\} \tag{9}$$
where the max is applied component-wise.

Step 4: Substitution and Simplification

Substitute the truncated solution $\hat{\lambda}_2$ back into equation (6) to find the final estimator for $\lambda_1$: $$\hat{\lambda}_1 = \mathcal{Z}_1 - \mathcal{J}_1^{-1} \mathcal{J}_{1,2} \big( \max\{-h, \mathcal{Z}_2\} - \mathcal{Z}_2 \big) \tag{10}$$
Apply the componentwise identity $\max\{-h, x\} - x = -\min\{0, x + h\}$ (letting $x = \mathcal{Z}_2$): $$\max\{-h, \mathcal{Z}_2\} - \mathcal{Z}_2 = -\min\{0, \mathcal{Z}_2 + h\} \tag{11}$$
Substitute this identity back into equation (10): $$\hat{\lambda}_1 = \mathcal{Z}_1 - \mathcal{J}_1^{-1} \mathcal{J}_{1,2} \big( -\min\{0, \mathcal{Z}_2 + h\} \big) = \mathcal{Z}_1 + \mathcal{J}_1^{-1} \mathcal{J}_{1,2} \min\{0, \mathcal{Z}_2 + h\} \tag{12}$$

Conclusion

By minimizing the generalized quadratic distance subject to the parameter space constraints, the asymptotic distributions of the estimators (parameterized by the true values $\gamma_0$) are precisely given by: $$\sqrt{T}(\hat{\theta}_{1,T} - \theta_{1,T}) \xrightarrow{d} \mathcal{Z}_1 + \mathcal{J}_1^{-1}(\gamma_0) \mathcal{J}_{1,2}(\gamma_0) \min\{0, \mathcal{Z}_2 + h\} \tag{13}$$
Q.E.D.
Loading...