Cross Validated
2026-11-04 21:31 UTC
By Asigan
AI-113-20261104-social-media-4a68a2f3
Shao and Tu's The Jackknife and Bootstrap, Theorem 2.1: why do we need jackknife in this case?
In Shao and Tu's The Jackknife and Bootstrap , section 2.1.1, they proved the following theorem (Theorem 2.1), which, they also pointed out, is the first consistency theorem on jackknife: If $X_1, \cdots, X_n$ are i.i.d. samples from an unknown probability measure and $\mu, \Sigma$ is the mean vector and covariance matrix of $X_1$ , $T_n= g(\frac 1n \sum_{i=1}^{n} X_n)$ estimates $g(\mu)$ , then, under certain regularity conditions, $\frac{v_{\text{Jack}}}{\sigma_n^2 }\to 1$ almost surely, where \begin{equation*} v_{\text{Jack}}= \frac{n-1}{n} \sum_{i=1}^n \left( T_{n, i}- \frac{1}{n} \sum_{j=1}^{n} T_{n, j} \right)^2 \text{ where } T_{n, i}= g\left( \frac{1}{n-1}\sum_{j\ne i} X_j \right),\quad \sigma_n^2= \frac{1}{n} \nabla g(\mu) ^T \Sigma \nabla g(\mu) . \end{equation*} I understand the proof, but I have two questions on the theorem itself: (1) Under what circumstances will we use this jackknife variance estimator, rather than plug-in estimator $\hat\sigma_n^2= \frac{1}{n} \nabla g(\hat\mu) ^T \hat\Sigma \nabla g(\hat\mu)$ (where we plug in sample mean and sample variance)? By law of large number this is also consistent provided that $\nabla g$ is continuous at $\mu$ , and in Shao's book, the theorem also required $\nabla g$ is continuous at $\mu$ to make the jackknife variance estimate consistent. (2) In the expression of $v_{\text{Jack}}$ , why do we keep the factor $\frac{n-1}{n}$ ? This factor converges to $1$ as $n\to +\infty$ , so discarding it will still make $v_{\…
In Shao and Tu's The Jackknife and Bootstrap , section 2.1.1, they proved the following theorem (Theorem 2.1), which, they also pointed out, is the first consistency theorem on jackknife: If $X_1, \cdots, X_n$ are i.i.d. samples from an unknown probability measure and $\mu, \Sigma$ is the mean vector and covariance matrix of $X_1$ , $T_n= g(\frac 1n \sum_{i=1}^{n} X_n)$ estimates $g(\mu)$ , then, under certain regularity conditions, $\frac{v_{\text{Jack}}}{\sigma_n^2 }\to 1$ almost surely, where \begin{equation*} v_{\text{Jack}}= \frac{n-1}{n} \sum_{i=1}^n \left( T_{n, i}- \frac{1}{n} \sum_{j=1}^{n} T_{n, j} \right)^2 \text{ where } T_{n, i}= g\left( \frac{1}{n-1}\sum_{j\ne i} X_j \right),\quad \sigma_n^2= \frac{1}{n} \nabla g(\mu) ^T \Sigma \nabla g(\mu) . \end{equation*} I understand the proof, but I have two questions on the theorem itself: (1) Under what circumstances will we use this jackknife variance estimator, rather than plug-in estimator $\hat\sigma_n^2= \frac{1}{n} \nabla g(\hat\mu) ^T \hat\Sigma \nabla g(\hat\mu)$ (where we plug in sample mean and sample variance)? By law of large number this is also consistent provided that $\nabla g$ is continuous at $\mu$ , and in Shao's book, the theorem also required $\nabla g$ is continuous at $\mu$ to make the jackknife variance estimate consistent. (2) In the expression of $v_{\text{Jack}}$ , why do we keep the factor $\frac{n-1}{n}$ ? This factor converges to $1$ as $n\to +\infty$ , so discarding it will still make $v_{\…
Full article content could not be extracted automatically. Read the original below.
Source:
Cross Validated
· stats.stackexchange.com