26. Limit Distribution and Scale-Free Reparameterization

26. Limit Distribution and Scale-Free Reparameterization

Settings

  1. Background: In econometric theory proofs, eliminating a nuisance macroscopic parameter (like $\gamma_0$) can sometimes be achieved through a mathematical trick called **"Scale-Free"** or **"Reparameterization"**.
  1. Condition: This trick works perfectly when $K_2 = 1$ (single variable, in [Ketz(2019)]), which allows us to utilize the ratio property of the $t$-statistic.
  1. Objective: Explain why and how the nuisance parameter $\gamma_0$ can be completely absorbed (normalized to 1) under the $K_2=1$ condition, specifically in the context of the limit distribution involving a truncated maximum function.

Derivation

  1. The Physical Essence of the $t$-Statistic as a Ratio:
The basic structure of a $t$-statistic (used for confidence intervals or hypothesis testing) is always: $$t = \frac{\text{Estimated Error}}{\text{Standard Error}}$$
In the limit state (e.g., based on Equation 14):
* Numerator (Estimated Error): Follows a truncated distribution $\max\{-h, \mathcal{Z}_{K_1+1}\}$. Here, $\mathcal{Z}_{K_1+1}$ is a normal random variable with mean 0 and variance $V_{K_1+1}(\gamma_0)$.
* Denominator (Standard Error): In the limit, this is the square root of the variance, $\sqrt{V_{K_1+1}(\gamma_0)}$.
  1. Algebraic Transformation (Absorbing the Denominator):
Dividing the limit state numerator and denominator yields the $t$-statistic in the limit: $$t_{\text{limit}} = \frac{\max\{-h, \mathcal{Z}_{K_1+1}\}}{\sqrt{V_{K_1+1}(\gamma_0)}}$$
Since the denominator is a positive number, we can use the property of the maximum function ($\frac{\max\{a, b\}}{c} = \max\left\{\frac{a}{c}, \frac{b}{c}\right\}$) and directly pull it inside the bracket: $$t_{\text{limit}} = \max\left\{ \frac{-h}{\sqrt{V_{K_1+1}(\gamma_0)}}, \frac{\mathcal{Z}_{K_1+1}}{\sqrt{V_{K_1+1}(\gamma_0)}} \right\}$$
  1. Perfect Absorption of $\gamma_0$ (The "Scale-Free" Magic):
* Right Term $\frac{\mathcal{Z}_{K_1+1}}{\sqrt{V_{K_1+1}(\gamma_0)}}$: A normal variable with variance $V_{K_1+1}(\gamma_0)$ divided by its standard deviation. This is standard **Z-score normalization**. After division, it becomes a variable that always follows the standard normal distribution $\mathcal{N}(0, 1)$, which we call $Z^*$. The $\gamma_0$ parameter has vanished, and the variance is normalized to 1.
* Left Term $\frac{-h}{\sqrt{V_{K_1+1}(\gamma_0)}}$: This still contains $\gamma_0$. However, our ultimate goal is to find the infimum for the worst-case coverage probability (AsySz), which requires $h$ to traverse from $0$ to $\infty$.
Since $h$ is a variable that takes any value in $[0, \infty)$, regardless of whether the constant denominator $\sqrt{V_{K_1+1}(\gamma_0)}$ is 2 or 200, the entire term $\frac{h}{\sqrt{V_{K_1+1}(\gamma_0)}}$ remains **a new variable that takes any value in $[0, \infty)$**.
We can reparameterize this entire term as a new relative scaling drift parameter, $h^*$.
The final limit statistic becomes: $$t_{\text{limit}} = \max\{-h^*, Z^*\}$$

Conclusion

  1. Why $\gamma_0$ is Eliminated:
In the new formula, $Z^*$ is always a standard normal distribution (variance 1, independent of $\gamma_0$). To find the worst-case coverage, we simply let $h^*$ run from $0$ to $\infty$. The parameter $\gamma_0$ has not truly disappeared; it has been **completely absorbed** by $h$. In the univariate case, changing $\gamma_0$ only scales the coordinate axis, which does not affect the result of finding the infimum (the worst-case scenario).
  1. Contrast: Why This Fails for $K_2 > 1$:
If $K_2 = 2$ (two variance parameters), the underlying $V(\gamma_0)$ is a $2 \times 2$ covariance matrix. Although matrix multiplication can normalize the two diagonal elements to 1, the correlation coefficients (off-diagonal elements) will still remain and be governed by $\gamma_0$. We cannot use a simple scalar $h^*$ to absorb multi-dimensional correlation. This is why the perfect dimensionality reduction (normalizing the variance-covariance structure to completely rid of $\gamma_0$) is a special property unique to the $K_2 = 1$ case.
Loading...