A Proof of Pearson's Chi-Square Goodness-of-Fit Test

Contents

• Theorem
• Proof
  • Part I
  • Part II
  • Part III
  • Part IV
  • Part V
• Another Incomplete Proof
• The textbooks which have no proof
• Note
• References

Theorem

• [Hogg, PSI, p.416] To generalize, we let an experiment have \(k\) (instead of only two) mutually exclusive and exhaustive outcomes, say, \(A_1, A_2, ..., A_k\). Let \(p_i=P(A_i)\), and thus \(\sum_{i=1}^{k}p_i=1\). The experiment is repeated \(n\) independent times, and we let \(Y_i\) represent the number of times the experiment results in \(A_i\), \(i=1, 2, ..., k\). This joint distribution of \(Y_1, Y_2, ..., Y_{k−1}\) is a straightforward generalization of the binomial distribution, as follows.


• [Hogg, PSI, p.416] In considering the joint pmf, we see that \[ f(y_1, y_2, ..., y_{k-1})=P(Y_1=y_1, Y_2=y_2, ..., Y_{k-1}=y_{k-1}), \] where \(y_1, y_2, ..., y_{k-1}\) are nonnegative integers such that \(y_1+y_2+\cdots+y_{k-1}\leq n\). Note that we do not need to consider \(Y_k\), since, once the other \(k-1\) random variables are observed to equal \(y_1, y_2, ..., y_{k-1}\), respectively, we know that \[ Y_k=n-y_1-y_2-\cdots-y_{k-1}=y_k, \text{ say}. \] From the independence of the trials, the probability of each particular arrangement of \(y_1\) \(A_1\)s, \(y_2\) \(A_2\)s, ..., \(y_k\) \(A_k\)s is \[ p_1^{y_1}p_2^{y_2}\cdots p_k^{y_k}. \] The number of such arrangements is the multinomial coefficient \[ {n\choose y_1, y_2, ..., y_k}=\frac{n!}{y_1! y_2!\cdots y_k!}. \] Hence, the product of these two expressions gives the joint pmf of \(Y_1, Y_2, ..., Y_{k-1}\): \[ f(y_1, y_2, ..., y_{k-1})=\frac{n!}{y_1! y_2!\cdots y_k!}p_1^{y_1}p_2^{y_2}\cdots p_k^{y_k}. \] (Recall that \(y_k=n-y_1-y_2-\cdots-y_{k-1}\).)


• [Hogg, PSI, p.416] Pearson then constructed an expression similar to \(Q_1\) (Equation 9.1-1), which involves \(Y_1\) and \(Y_2=n-Y_1\), that we denote by \(Q_{k-1}\), which involves \(Y_1, Y_2, ..., Y_{k-1}\), and \(Y_k=n-Y_1-Y_2-\cdots -Y_{k-1}\), namely, \[ Q_{k-1}=\sum_{i=1}^{k}\frac{(Y_i-np_i)^2}{np_i}. \] He argued that \(Q_{k-1}\) has an approximate chi-square distribution with \(k-1\) degree of freedom in much the same way we argued that \(Q_1\) is approximately \(\chi^2(1)\). We accept Pearson's conclusion, as the proof is beyond the level of this text.

Proof

The main idea is from this lecture note by David Hunter in Penn State Department of Statistics. [Keeping, sec.13.5, 13.6, appendix A.17] has a partial discussion. If you know where I can find a complete proof in a book. Please give me an email.

Part I

• We need some preliminaries.
• [Hogg, IMS, p.140] In Section 2.5 we discussed the covariance between two random variables. In this section we want to extend this discussion to the \(n\)-variate case. Let \(\mathbf{X}=(X_1, ..., X_n)'\) be an \(n\)-dimensional random vector. Recall that we defined \(\text{E}(\mathbf{X})=(\text{E}(X_1), ..., \text{E}(X_n))'\), that is, the expectation of a random vector is just the vector of the expectations of its components.


• [Hogg, IMS, p.140] Now suppose \(\mathbf{W}\) is an \(m\times n\) matrix of random variables, say, \(\mathbf{W}=[W_{ij}]\) for the random variables \(W_{ij}\), \(1\leq i\leq m\) and \(1\leq j\leq n\). Note that we can always string out the matrix into an \(mn\times 1\) random vector. Hence, we define the expectation of a random matrix \[ \text{E}[\mathbf{W}]=[\text{E}(W_{ij})]. \tag{2.6.10} \] As the following theorem shows, the linearity of the expectation operator easily follows from this definition:


• [Hogg, IMS, p.141] Let \(\mathbf{X}=(X_1, ..., X_n)'\) be an \(n\)-dimensional random vector, such that \(\sigma_i^2=\text{Var}(X_i)\lt \infty\). The mean of \(\mathbf{X}\) is \(\mathbf{\mu}=\text{E}[\mathbf{X}]\) and we define its variance-covariance matrix as \[ \text{Cov}(\mathbf{X})=\text{E}[(\mathbf{X}-\mathbf{\mu})(\mathbf{X}-\mathbf{\mu})']=[\sigma_{ij}], \tag{2.6.13} \] where \(\sigma_{ii}\) denotes \(\sigma_i^2\). As Exercise 2.6.8 shows, the \(i\)th diagonal entry of \(\text{Cov}(\mathbf{X})\) is \(\sigma_i^2=\text{Var}(X_i)\) and the \((i, j)\)th off diagonal entry is \(\text{Cov}(X_i, X_j)\).


• [Hogg, IMS, p.350] As another simple application, consider the multivariate analog of the sample mean and sample variance. Let \(\{\mathbf{X}_n\}\) be a sequence of iid random vectors with common mean vector \(\mathbf{\mu}\) and variance-covariance matrix \(\Sigma\). Denote the vector of means by \[ \overline{\mathbf{X}}_n=\frac{1}{n}\sum_{i=1}^{n}\mathbf{X}_i. \tag{5.4.5} \] Of course, \(\overline{\mathbf{X}}_n\) is just the vector of sample means, \((\overline{X}_1, ..., \overline{X}_p)'\). By the Weak Law of Large Numbers. Theorem 5.1.1, \(\overline{X}_j\to \mu_j\), in probability, for each \(j\). Hence, by Theorem 5.4.1, \(\overline{\mathbf{X}}_n\to \mathbf{\mu}\), in probability.


• 這張圖可以幫助你理解 \[ \begin{array}{cccc|ccc} \mathbf{X}_1 & \mathbf{X}_2 & \cdots & \mathbf{X}_n & \overline{\mathbf{X}}_n & \to & \mathbf{\mu} \\ || & || & \cdots & || & || & & || \\ \begin{bmatrix}X_{11}\\X_{21}\\\vdots\\X_{p1}\end{bmatrix} & \begin{bmatrix}X_{12}\\X_{22}\\\vdots\\X_{p2}\end{bmatrix} & \cdots & \begin{bmatrix}X_{1n}\\X_{2n}\\\vdots\\X_{pn}\end{bmatrix} & \begin{bmatrix}\overline{X}_1\\\overline{X}_2\\\vdots\\\overline{X}_p\end{bmatrix} & \begin{matrix}\to\\\to\\\vdots\\\to\end{matrix} & \begin{bmatrix}\mu_1\\\mu_2\\\vdots\\\mu_p\end{bmatrix} \end{array} \] 這裡的 \(\overline{X}_k\) 定義成 \(\overline{X}_k=\frac{\sum_{l=1}^{n}X_{kl}}{n}\)，我覺得寫成 \(\overline{X}_{k\Box}\) 會更一目瞭然。 這裡所說的common mean vector \(\mathbf{\mu}\) and variance-covariance matrix \(\Sigma\) 是指 \[ \text{E}(\mathbf{X}_1)=\text{E}(\mathbf{X}_2)=\cdots =\mathbf{\mu} \] 也就是說， \[ \text{E}(X_{i1})=\text{E}(X_{i2})=\cdots=\mu_i \text{ for }i=1, 2, ... \] 及 \[ \text{Cov}(\mathbf{X}_1)=\text{Cov}(\mathbf{X}_2)=\cdots =\Sigma \]

Part II

• Let \(\mathbf{X}_1, \mathbf{X}_2, ...\) be a sequence of independent and identically distributed \(k\)-variate random vectors. For each \(i\), \(\mathbf{X}_i\) has a multinomial distribution \(\text{multinomial}(1, \mathbf{p})\), where \[ \mathbf{p}=\begin{bmatrix} p_1\\ p_2\\ \vdots\\ p_k \end{bmatrix} \]


• Each \(X_{ij}\) is either \(0\) or \(1\). As the following figure shows. Each column has exactly one \(1\) and \(0\) elsewhere. \[ \begin{array}{ccccc} \mathbf{X}_1 & \mathbf{X}_2 & \cdots & \mathbf{X}_n & \\ || & || & \cdots & || & \\ \begin{bmatrix}X_{11}\\X_{21}\\\vdots\\X_{k1}\end{bmatrix} & \begin{bmatrix}X_{12}\\X_{22}\\\vdots\\X_{k2}\end{bmatrix} & \cdots & \begin{bmatrix}X_{1n}\\X_{2n}\\\vdots\\X_{kn}\end{bmatrix} & \begin{matrix} \sum_{j=1}^{n}X_{1j}=n\overline{X}_1=n\overline{X}_{1\Box}=Y_1\\\sum_{j=1}^{n}X_{2j}=n\overline{X}_2=n\overline{X}_{2\Box}=Y_2\\\vdots\\\sum_{j=1}^{n}X_{kj}=n\overline{X}_k=n\overline{X}_{k\Box}=Y_k\end{matrix} \end{array} \] Recall that \(Y_i\) represent the number of times the experiment results in \(A_i\), \(i=1, 2, ..., k\).


• By [Casella, p.182, line 5], \(X_{ij}\sim \text{B}(1, p_i)\). So \[ \text{E}(X_{ij})=p_i\text{ and }\text{Var}(X_{ij})=p_i(1-p_i). \] By [Casella, p.182, line -10], \[ \text{Cov}(X_{kj}, X_{lj})=-p_k p_l. \] Thus, \[ \text{E}(\mathbf{X}_1)=\text{E}(\mathbf{X}_2)=\cdots=\mathbf{p} \] and variance-covariance matrix of \(\mathbf{X}_1, \mathbf{X}_2, ...\) is \[ \text{Cov}(\mathbf{X}_1)=\text{Cov}(\mathbf{X}_2)=\cdots=\Sigma= \begin{bmatrix} p_1(1-p_1) & -p_1 p_2 & \cdots & -p_1 p_k \\ -p_2 p_1 & p_2(1-p_2) & \cdots & -p_2 p_k \\ \vdots & \vdots & \ddots & \vdots \\ -p_k p_1 & -p_k p_2 & \cdots & p_k(1-p_k) \end{bmatrix} \] Note that the sum of every entries in any column or row of \(\Sigma\) is \(0\). Hence, \(\det{\Sigma}=0\) and \(\Sigma\) is not invertible.

Part III

• For each \(j=1, 2, ...\), consider \[ \mathbf{Z}_j= \begin{bmatrix} X_{1j}\\ X_{2j}\\ \vdots\\ X_{k-1, j} \end{bmatrix} \] Check that \[\text{E}(\mathbf{Z}_1)=\text{E}(\mathbf{Z}_2)=\cdots =\mathbf{p}^*= \begin{bmatrix} p_1\\ p_2\\ \vdots\\ p_{k-1} \end{bmatrix} \tag{*} \] and the variance-covariance matrix of \(\mathbf{Z}_1, \mathbf{Z}_2, ...\) is \[ \text{Cov}(\mathbf{Z}_1)=\text{Cov}(\mathbf{Z}_2)=\cdots =\Sigma^*= \begin{bmatrix} p_1(1-p_1) & -p_1 p_2 & \cdots & -p_1 p_{k-1} \\ -p_2 p_1 & p_2(1-p_2) & \cdots & -p_2 p_{k-1} \\ \vdots & \vdots & \ddots & \vdots \\ -p_{k-1} p_1 & -p_{k-1} p_2 & \cdots & p_{k-1}(1-p_{k-1}) \end{bmatrix} \tag{**} \] That is, the upper-left \((k-1)\times (k-1)\) submatrix of \(\Sigma\).


• Note that \[ \begin{array}{lll} \Sigma^* &=& \begin{bmatrix} p_1(1-p_1) & -p_1 p_2 & \cdots & -p_1 p_{k-1} \\ -p_2 p_1 & p_2(1-p_2) & \cdots & -p_2 p_{k-1} \\ \vdots & \vdots & \ddots & \vdots \\ -p_{k-1} p_1 & -p_{k-1} p_2 & \cdots & p_{k-1}(1-p_{k-1}) \end{bmatrix} \\ &=& \begin{bmatrix} p_1 \\ & p_2 \\ & & \ddots \\ & & & p_{k-1} \end{bmatrix} - \begin{bmatrix} p_1\\ p_2\\ \vdots\\ p_{k-1} \end{bmatrix} \begin{bmatrix} p_1 & p_2 & \cdots & p_{k-1} \end{bmatrix}\\ &=& \text{diag}(p_1, p_2, ..., p_{k-1}) -\mathbf{p^*}(\mathbf{p}^*)^T \end{array}. \] Now, \(\Sigma^*\) is invertible and its inverse is \[ \begin{array}{lll} (\Sigma^*)^{-1} &=& \begin{bmatrix} \frac{1}{p_1}+\frac{1}{p_k} & \frac{1}{p_k} & \cdots & \frac{1}{p_k} \\ \frac{1}{p_k} & \frac{1}{p_2}+\frac{1}{p_k} & \cdots & \frac{1}{p_k} \\ \vdots & \vdots & \ddots & \vdots \\ \frac{1}{p_k} & \frac{1}{p_k} & \cdots & \frac{1}{p_{k-1}}+\frac{1}{p_k} \end{bmatrix} \\ &=& \begin{bmatrix} \frac{1}{p_1} \\ & \frac{1}{p_2} \\ & & \ddots \\ & & & \frac{1}{p_{k-1}} \end{bmatrix} +\frac{1}{p_k} \begin{bmatrix} 1&1&\cdots&1\\ 1&1&\cdots&1\\ \vdots&\vdots&\ddots&\vdots\\ 1&1&\cdots&1 \end{bmatrix}\\ &=& \text{diag}(\frac{1}{p_1}, \frac{1}{p_2}, ..., \frac{1}{p_{k-1}})+\frac{1}{p_k}\mathbf{1}\mathbf{1}^T, \end{array} \tag{***} \] where \[ \mathbf{1}= \begin{bmatrix} 1\\ 1\\ \vdots\\ 1 \end{bmatrix}. \] It is easy to check that it is indeed the inverse by using that \[ \begin{array}{rcl} (\mathbf{p}^*)^T\text{diag}(\frac{1}{p_1}, \frac{1}{p_2}, ..., \frac{1}{p_{k-1}}) &=& \mathbf{1}^T, \\\text{diag}(p_1, p_2, ..., p_{k-1})\mathbf{1}&=& \mathbf{p}^*, \\ (\mathbf{p}^*)^T \mathbf{1} &=& p_1+p_2+\cdots+p_{k-1}=1-p_k. \end{array} \]

Part IV

• [Hogg, IMS, p.351, def.5.4.2] Let \(\{\mathbf{X}_n\}\) be a sequence of random vectors with \(\mathbf{X}_n\) having distribution function \(F_n(\mathbf{x})\) and \(\mathbf{X}\) be a random vector with distribution function \(F(\mathbf{x})\). Then \(\{\mathbf{X}_n\}\) converges in distribution to \(\mathbf{X}\) if \[ \lim_{n\to \infty} F_n(\mathbf{x})=F(\mathbf{x}), \tag{5.4.8} \] for all points \(\mathbf{x}\) at which \(F(\mathbf{x})\) is continuous. We write \(\mathbf{X}_n\stackrel{D}{\to}\mathbf{X}\).


• [Hogg, IMS, p.351, thm.5.4.4] (Multivariate Central Limit Theorem). Let \(\{\mathbf{X}_n\}\) be a sequence of iid random vectors with common mean vector \(\mathbf{\mu}\) and variance-covariance matrix \(\Sigma\) which is positive definite. Assume that the common moment generating function \(M(\mathbf{t})\) exists in an open neighborhood of \(\mathbf{0}\). Let \[ \mathbf{Y}_n=\frac{1}{\sqrt{n}}\sum_{i=1}^{n}(\mathbf{X}_i-\mathbf{\mu})=\sqrt{n}(\overline{\mathbf{X}}-\mathbf{\mu}). \] Then \(\mathbf{Y}_n\) converges in distribution to a \(\text{N}_p(\mathbf{0}, \Sigma)\) distribution.


• Consider that \[ \begin{array}{rcl} \overline{\mathbf{Z}}_n &=& \frac{\mathbf{Z}_1+\mathbf{Z}_2+\cdots+\mathbf{Z}_n}{n} \\ &=& \begin{bmatrix} \frac{X_{11}+X_{12}+\cdots+X_{1n}}{n}\\ \frac{X_{21}+X_{22}+\cdots+X_{2n}}{n}\\ \vdots\\ \frac{X_{k-1, 1}+X_{k-1, 2}+\cdots+X_{k-1, n}}{n} \end{bmatrix} \\ &=& \begin{bmatrix} \overline{X}_{1\Box}\\ \overline{X}_{2\Box}\\ \vdots\\ \overline{X}_{k-1,\Box} \end{bmatrix} \end{array} \] By (*), (**) and the Multivariate Central Limit Theorem, \[ \sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)\stackrel{d}{\to}\text{N}_{k-1}(\mathbf{0}, \Sigma^*). \tag{****} \]


• [Hogg, IMS, p.202, thm.3.5.1] Suppose \(\mathbf{X}\) has a \(\text{N}_n(\mathbf{\mu}, \Sigma)\) distribution, where \(\Sigma\) is positive definite. Then the random variable \(Y=(\mathbf{X}-\mathbf{\mu})'\Sigma^{-1}(\mathbf{X}-\mathbf{\mu})\) has a \(\chi^2(n)\) distribution.


• \((\mathbf{X}-\mathbf{\mu})'\) means the transpose of \(\mathbf{X}-\mathbf{\mu}\). That is, \[ (\mathbf{X}-\mathbf{\mu})'=(\mathbf{X}-\mathbf{\mu})^T \]


• By [Hogg, IMS, p.202, thm.3.5.1], (\(\mathbf{X}=\sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)\) and \(\mathbf{\mu}=\mathbf{0}\)), \[ \sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T(\Sigma^*)^{-1}\sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*) \] has a \(\chi^2(k-1)\) distribution.

Part V

• The final step is to show that \[ \sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T(\Sigma^*)^{-1}\sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*) =\sum_{i=1}^{k}\frac{(Y_i-np_i)^2}{np_i} \] Note that \[ \overline{\mathbf{Z}}_n-\mathbf{p}^*= \begin{bmatrix} \overline{X}_{1\Box}-p_1\\ \overline{X}_{2\Box}-p_2\\ \vdots\\ \overline{X}_{k-1, \Box}-p_{k-1} \end{bmatrix} \] Therefore, \[ \begin{array}{cl} & \sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T(\Sigma^*)^{-1}\sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*) \\ \stackrel{\text{(***)}}{=}& n(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T(\Sigma^*)^{-1}(\overline{\mathbf{Z}}_n-\mathbf{p}^*) \\ =& n(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T\left(\text{diag}(\frac{1}{p_1}, \frac{1}{p_2}, ..., \frac{1}{p_{k-1}})+\frac{1}{p_k}\mathbf{1}\mathbf{1}^T\right)(\overline{\mathbf{Z}}_n-\mathbf{p}^*) \\ =& n\left[(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T\text{diag}(\frac{1}{p_1}, \frac{1}{p_2}, ..., \frac{1}{p_{k-1}})(\overline{\mathbf{Z}}_n-\mathbf{p}^*) +\frac{1}{p_k}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)^T\mathbf{1}\mathbf{1}^T(\overline{\mathbf{Z}}_n-\mathbf{p}^*)\right] \\ =& n\left[\sum_{i=1}^{k-1}\frac{(\overline{X}_{i\Box}-p_i)^2}{p_i} +\frac{1}{p_k}\sum_{i, j=1}^{k-1}(\overline{X}_{i\Box}-p_i)(\overline{X}_{j\Box}-p_j)\right] \\ =& n\left\{\sum_{i=1}^{k-1}\frac{(\overline{X}_{i\Box}-p_i)^2}{p_i} +\frac{1}{p_k}\left[\sum_{i=1}^{k-1}(\overline{X}_{i\Box}-p_i)\right]^2\right\} \\ =& \sum_{i=1}^{k-1}\frac{(n\overline{X}_{i\Box}-np_i)^2}{np_i}+\frac{\left[\sum_{i=1}^{k-1}(n\overline{X}_{i\Box}-np_i)\right]^2}{np_k} \\ \stackrel{\sum_{i=1}^{k}n\overline{X}_{i\Box}=\sum_{i=1}^{k}Y_i=n}{=}& \sum_{i=1}^{k-1}\frac{(n\overline{X}_{i\Box}-np_i)^2}{np_i}+\frac{\left[(n-n\overline{X}_{k\Box})-n(1-p_k)\right]^2}{np_k} \\ =& \sum_{i=1}^{k}\frac{(n\overline{X}_{i\Box}-np_i)^2}{np_i} \\ =& \sum_{i=1}^{k}\frac{(Y_i-np_i)^2}{np_i} \end{array} \]

Another Incomplete Proof

下面是另一個不完整的證法，主要是參考這裡，這個證法看起來不需要用到Multivariate Normal Distribution，但最後一步要證明sum of dependent normal random variables is still normal的時候（標記三個問號???的地方），就不可避免地要用到Multivariate Normal，參考這裡。

• By [Casella, p.182, line 5], \(Y_i\sim \text{B}(n, p_i)\). Then by the Central Limit Theorem or De-Moivre Laplace Theorem, \(Y_i\sim \text{N}(np_i, np_i(1-p_i))\).


• Let \(X_i=\frac{Y_i-np_i}{\sqrt{np_i}}\). Then \[ X_i=\frac{Y_i-np_i}{\sqrt{np_i}}=\frac{Y_i-np_i}{\sqrt{np_i(1-p_i)}}\sqrt{1-p_i}\stackrel{\text{Central Limit Theorem}}{\sim}\sqrt{1-p_i}\text{N}(0, 1). \] By [Casella, p.184, cor.4.6.10], \(X_i\sim \text{N}(0, 1-p_i)\). This follows that \[ \text{Var}(X_i)=1-p_i \tag{i} \]


• If \(i\neq j\), then \[ \begin{array}{rcl} \text{Cov}(X_i, X_j) &=& \text{Cov}(\frac{Y_i-np_i}{\sqrt{np_i}}, \frac{Y_j-np_j}{\sqrt{np_j}}) \\ &\stackrel{\text{Cov}(aX+b, cY+d)=ac\text{Cov}(X, Y)}{=}& \frac{1}{\sqrt{np_i}}\cdot \frac{1}{\sqrt{np_j}}\cdot\text{Cov}(Y_i, Y_j) \\ &\stackrel{\text{[Casella, p.182, line -10]}}{=}& \frac{1}{\sqrt{np_i}}\cdot \frac{1}{\sqrt{np_j}}\cdot (-np_i p_j) \\ &=& -\sqrt{p_i p_j} \end{array} \tag{ii} \]


• By (i) and (ii), \[ \text{Cov}(\mathbf{X})= \begin{bmatrix} 1-p_1 & -\sqrt{p_1 p_2} & \cdots & -\sqrt{p_1 p_k} \\ -\sqrt{p_2 p_1} & 1-p_2 & \cdots & -\sqrt{p_2 p_k} \\ \vdots & \vdots & \ddots & \vdots \\ -\sqrt{p_k p_1} & -\sqrt{p_k p_2} & \cdots & 1-p_k \end{bmatrix} =I-\mathbf{p}\mathbf{p}^T, \tag{iii} \] where \[ \mathbf{p}= \begin{bmatrix} \sqrt{p_1}\\ \sqrt{p_2}\\ \vdots\\ \sqrt{p_k} \end{bmatrix}. \]


• We find the eigenvalues of \(\text{Cov}(\mathbf{X})\). Note that \(\mathbf{p}^T\mathbf{p}=p_1+p_2+\cdots+p_k=1\). \[ \begin{array}{rcl} \det{(\text{Cov}(\mathbf{X})-\lambda I)} &=& \det{(I-\mathbf{p}\mathbf{p}^T-\lambda I)} \\ &=& \det{((1-\lambda)I-\mathbf{p}\mathbf{p}^T)} \\ &=& (1-\lambda)^k \det{\left(I-\frac{1}{1-\lambda}\mathbf{p}\mathbf{p}^T\right)} \\ &\stackrel{\text{Sylvester's Theorem}}{=}& (1-\lambda)^k\left(1-\frac{1}{1-\lambda}\mathbf{p}^T\mathbf{p}\right) \\ &=& -\lambda(1-\lambda)^{k-1}. \end{array} \] The eigenvalues of \(\text{Cov}(\mathbf{X})\) are \(0\) and \(1\) with \(k-1\) multiplicity.


• Since \(\text{Cov}(\mathbf{X})\) is symmetric, \(\text{Cov}(\mathbf{X})\) is orthogonally diagonalizable. (See [Friedberg, p.384, thm.6.20].) That is, there exists an orthogonal matrix \(Q\) (\(QQ^T=Q^TQ=I\)) such that \[ Q\text{Cov}(\mathbf{X})Q^T= \left[ \begin{array}{c|c} I_{k-1} & O \\ \hline O & 0 \end{array} \right] \text{ and } Q\text{Cov}(\mathbf{X})= \left[ \begin{array}{c|c} I_{k-1} & O \\ \hline O & 0 \end{array} \right]Q \tag{iv} \]


• Set \[ \mathbf{Z}=Q\mathbf{X} \tag{v} \]


• Note that \[ \text{Cov}(\mathbf{Z}) =\text{Cov}(Q\mathbf{X}) \stackrel{\text{[Hogg, IMS, p.141, thm.2.63]}}{=}Q\text{Cov}(\mathbf{X})Q^T= \left[ \begin{array}{c|c} I_{k-1} & O \\ \hline O & 0 \end{array} \right] \tag{vi} \]

Now, we prove the main four results.

• For each \(i\), by (v), \(Z_i\) is a linear combination of normal random variables \(X_1, X_2, ..., X_k\). By ???, \(Z_i\) is also normal.


• Since \(\text{E}(X_i)=0\), by [Casella, p.57], \(\text{E}(Z_i)=0\). By (vi), \(\text{Var}(Z_i)=1\) for \(i=1, 2, ..., k-1\).


• By (vi), \(\text{Cov}(Z_i, Z_j)=0\) for \(i\neq j\). That is, \(Z_i\) and \(Z_j\) are uncorrelated. By [Roussas, Course, p.466, cor.2], \(Z_1, Z_2, ..., Z_k\) are independent.


• Now, \[ \begin{array}{cl} \text{check that}& \mathbf{p}^T\mathbf{X}=0 \text{ by their definitions}\\ \Rightarrow& \text{Cov}(\mathbf{X})\mathbf{X}\stackrel{\text{(iii)}}{=}(I-\mathbf{p}\mathbf{p}^T)\mathbf{X}=\mathbf{X}-\mathbf{p}\mathbf{p}^T\mathbf{X}=\mathbf{X} \\ \Rightarrow& \mathbf{Z}\stackrel{\text{(v)}}{=}Q\mathbf{X}=Q(\text{Cov}(\mathbf{X})\mathbf{X})=(Q\text{Cov}(\mathbf{X}))\mathbf{X}\stackrel{\text{(iv)}}{=}\left[ \begin{array}{c|c} I_{k-1} & O \\ \hline O & 0 \end{array} \right] Q\mathbf{X} \\ \Rightarrow & Z_k=0 \end{array} \] This follows that \[\sum_{i=1}^{k-1}Z_i^2=\sum_{i=1}^{k}Z_i^2 \] and \[ \sum_{i=1}^{k-1}Z_i^2 =\sum_{i=1}^{k}Z_i^2 =\mathbf{Z}^T\mathbf{Z} =(Q\mathbf{X})^T(Q\mathbf{X}) =\mathbf{X}^T Q^T Q \mathbf{X} =\mathbf{X}^T \mathbf{X} =\sum_{i=1}^{k}X_i^2. \]

Finall, by the above four results and [Casella, p.219, lem.5.3.2], we have \[ \sum_{i=1}^{k}X_i^2\sim \chi^2(k-1). \]

The textbooks which have no proof

• [Hogg, PSI, p.416] We accept Pearson's conclusion, as the proof is beyond the level of this text.
• [Hogg, IMS, p.284] It is proved in a more advanced course that, as \(n\to \infty\), \(Q_{k−1}\) has an approximate \(\chi^2(n-1)\) distribution.
• [Mood, p.445] We will not prove the above theorem, but we will indicate its proof for \(k=1\).
• [DeGroot, p.626] In 1900, Karl Pearson proved the following result, whose proof will not be given here.

Note

• This theorem can be proved by maximum likelihood estimator. See [Rice, p.341] or [Roussas, Course, p.370] or [Spokoiny, p.205]
• This proof is similar to the approximation of multinomial by multivariate normal. See here
• Some books don't require that \(\Sigma\) is invertible. In that case, you can use [Hogg, IMS, p.202, thm.3.5.2] directly. At the end of Part II, we have \[ \sqrt{n}(\overline{\mathbf{X}}_n-\mathbf{p})\stackrel{d}{\to} \text{N}_k(\mathbf{0}, \Sigma). \] By [Hogg, IMS, p.202, thm.3.5.2], (\(A=\left[\begin{array}{c|c}I_{k-1}&\mathbf{0}\end{array}\right], \mathbf{b}=\mathbf{0}\)), \[ \sqrt{n}(\overline{\mathbf{Z}}_n-\mathbf{p}^*)=A\sqrt{n}(\overline{\mathbf{X}}_n-\mathbf{p})\stackrel{d}{\to} N_{k-1}(\mathbf{0}, \Sigma^*). \] A same result as (****).

References

• [Casella] Casella and Berger's Statistical Inference
• [DeGroot] DeGroot and Schervish's Probability and Statistics
• [Friedberg] Friedberg, Insel and Spence's Linear Algebra
• [Hogg, PSI] Hogg and Tanis's Probability and Statistical Inference
• [Hogg, IMS] Hogg, McKean and Craig's Introduction to Mathematical Statistics
• [Keeping] Keeping's Introduction to Statistical Inference
• [Mood] Mood, Graybill and Boes's Introduction to Theory of Statistics
• [Rice] Rice's Mathematical Statistics and Data Analysis
• [Roussas, Course] Roussas's A Course in Mathematical Statistics
• [Spokoiny] Spokoiny and Dickhaus's Basics of Modern Mathematical Statistics