33. Envelope Cancellation of the Linear Block in BLP GMM Gradient

33. Envelope Cancellation of the Linear Block in BLP GMM Gradient

Settings

  1. From [[32. Deriving the Concentrated-Out Linear IV-GMM Estimator|Note 32]]: $\hat b(\theta) = (X'ZWZ'X)^{-1}X'ZWZ'\,Y(\theta)$, structural residual
depends on $\theta$ through both $Y(\theta)$ and $\hat b(\theta)$.
  1. GMM objective and moment (from [[32. Deriving the Concentrated-Out Linear IV-GMM Estimator|Note 32 (3)]]):
Optimizer needs $\nabla_\theta q$. Does the gradient carry $\dfrac{\partial \hat b}{\partial \theta}$?

Proof

  1. Total derivative of $\xi$:
  1. Chain rule on $q$:
  1. Normal equation from [[32. Deriving the Concentrated-Out Linear IV-GMM Estimator|32 (7)]]: $\hat b$ was chosen so that
kills the second term of (3) — the $\dfrac{\partial \hat b}{\partial \theta}$ block never has to be formed.

Conclusion

This is the classical **envelope theorem** — (4) is the inner-optimum FOC $\partial q/\partial b\vert_{\hat b} = 0$ up to the $-\frac{2}{N^2}$ scaling.
What's avoided: differentiating $(X'ZWZ'X)^{-1}$ inside $\hat b$, i.e. the tensor calc $\dfrac{\partial (X'ZWZ'X)^{-1}}{\partial \theta}$.
Q.E.D.
Loading...