Skip to main content
1 of 4
Michael Hardy
  • 11k
  • 1
  • 33
  • 56

Suppose $X_1,\ldots,X_n \sim N(\mu,\sigma^2).$

The family of distributions is $\{ N_n\left(\begin{bmatrix} \mu \\ \vdots \\ \mu \end{bmatrix},\sigma^2 I_n\right) : \mu\in \mathbb R \}$ (in other words, $\sigma$ is fixed, so that the only difference between one distribution and another in this family is a different value of $\mu$). This is the family of $n$-dimensional normal distributions in which the expected value is that column of $\mu$s and the matrix of covariances is $\sigma^2$ times the $n\times n$ identity matrix. This family of distributions of $n$-tuples is not complete since it admits nontrivial unbiased estimators of $0$; for example $X_1-X_2$ is such an estimator.

Now suppose $T = T(X_1,\ldots,X_n) = X_1+\cdots+X_n.$ It follows that $T\sim N(n\mu,n\sigma^2).$ The family of distributions of $T$ is $\{ N(n\mu,n\sigma^2), \mu \in\mathbb R\},$ so again $\sigma$ is fixed, so the difference between two members of this family is a different value of $\mu$. This family is complete since it admits no nontrivial unbiased estimators of $0$. The fact that this family of distributions is complete is also expressed by saying that the statistic $T$ is complete. Any time you define a statistic that is a function of $(X_1,\ldots,X_n),$ having already defined a family of distributions for that $n$-tuple, that definition induces another family of distributions for the statistic you have defined. To say that that statistic is complete merely means that that induced family of distributions is complete.

Michael Hardy
  • 11k
  • 1
  • 33
  • 56