[proofplan] We write the centered sum of squares as a matrix quadratic form using the [orthogonal projection](/theorems/437) onto the subspace of $\mathbb{R}^n$ perpendicular to the all-ones vector. An orthogonal change of coordinates diagonalizes this projection and transforms the centered data into $n-1$ independent Gaussian contrast vectors and one sample-mean direction. The contrast vectors have distribution $\mathcal{N}_p(0,\Sigma)$, and the centered cross-product matrix is exactly the sum of their outer products. This is the defining representation of the Wishart distribution with $n-1$ degrees of freedom and scale matrix $\Sigma$. [/proofplan] [step:Represent the centered cross-product matrix by the centering projection] Let $Y: \Omega \to \mathbb{R}^{n \times p}$ be the random matrix whose $i$-th row is $Y_i^\top$. Let $\mathbf{1}_n \in \mathbb{R}^n$ denote the column vector whose entries are all $1$, and define the centering projection $P \in \mathbb{R}^{n \times n}$ by \begin{align*} P &= I_n - \frac{1}{n}\mathbf{1}_n\mathbf{1}_n^\top . \end{align*} For each $\omega \in \Omega$, the $i$-th row of $PY(\omega)$ is \begin{align*} Y_i(\omega)^\top - \bar{Y}(\omega)^\top . \end{align*} Therefore \begin{align*} Y^\top P Y &= (PY)^\top(PY) \\ &= \sum_{i=1}^n (Y_i-\bar{Y})(Y_i-\bar{Y})^\top \\ &= (n-1)S. \end{align*} Here we used $P^\top=P$ and $P^2=P$, which follow directly from the definition of $P$ and the identity $\mathbf{1}_n^\top\mathbf{1}_n=n$. [/step] [step:Diagonalize the centering projection by an orthogonal basis of contrasts] Choose vectors $q_1,\dots,q_{n-1} \in \mathbb{R}^n$ forming an orthonormal basis of the subspace \begin{align*} \mathbf{1}_n^\perp &= \{x \in \mathbb{R}^n : x^\top \mathbf{1}_n = 0\}, \end{align*} and set \begin{align*} q_n &= \frac{1}{\sqrt{n}}\mathbf{1}_n. \end{align*} Let $Q \in \mathbb{R}^{n \times n}$ be the orthogonal matrix whose $a$-th row is $q_a^\top$. Then $Q Q^\top=I_n$, and the projection $P$ has the decomposition \begin{align*} P = Q^\top D Q, \qquad D = \begin{pmatrix} I_{n-1} & 0 \\ 0 & 0 \end{pmatrix}. \end{align*} Indeed, $Pq_a=q_a$ for $a \in \{1,\dots,n-1\}$ because $q_a^\top\mathbf{1}_n=0$, while $Pq_n=0$ because $q_n$ spans $\operatorname{span}\{\mathbf{1}_n\}$. [guided] The matrix $P$ subtracts row averages, so it should preserve exactly the directions in $\mathbb{R}^n$ whose coordinates sum to zero. We make this precise by defining \begin{align*} \mathbf{1}_n^\perp &= \{x \in \mathbb{R}^n : x^\top \mathbf{1}_n = 0\}. \end{align*} This subspace has dimension $n-1$. Choose an orthonormal basis $q_1,\dots,q_{n-1}$ for it, and complete it by setting \begin{align*} q_n &= \frac{1}{\sqrt{n}}\mathbf{1}_n. \end{align*} The vectors $q_1,\dots,q_n$ form an orthonormal basis of $\mathbb{R}^n$, because $q_n$ has Euclidean norm $1$ and is orthogonal to every vector in $\mathbf{1}_n^\perp$. Let $Q \in \mathbb{R}^{n \times n}$ be the orthogonal matrix whose $a$-th row is $q_a^\top$. The projection $P$ acts as the identity on $\mathbf{1}_n^\perp$: if $x \in \mathbf{1}_n^\perp$, then \begin{align*} Px &= x - \frac{1}{n}\mathbf{1}_n\mathbf{1}_n^\top x \\ &= x. \end{align*} It annihilates the mean direction: \begin{align*} Pq_n &= \frac{1}{\sqrt{n}}\mathbf{1}_n - \frac{1}{n}\mathbf{1}_n\mathbf{1}_n^\top\frac{1}{\sqrt{n}}\mathbf{1}_n \\ &= \frac{1}{\sqrt{n}}\mathbf{1}_n - \frac{1}{n\sqrt{n}}\mathbf{1}_n(\mathbf{1}_n^\top\mathbf{1}_n) \\ &= \frac{1}{\sqrt{n}}\mathbf{1}_n - \frac{1}{\sqrt{n}}\mathbf{1}_n \\ &=0. \end{align*} Thus, in the orthonormal basis $q_1,\dots,q_n$, the matrix of $P$ is \begin{align*} D = \begin{pmatrix} I_{n-1} & 0 \\ 0 & 0 \end{pmatrix}, \end{align*} which means \begin{align*} P = Q^\top DQ. \end{align*} [/guided] [/step] [step:Transform the observations into independent Gaussian contrasts] Define the transformed data matrix $Z: \Omega \to \mathbb{R}^{n \times p}$ by \begin{align*} Z &= QY. \end{align*} Let $Z_a: \Omega \to \mathbb{R}^p$ denote the column vector whose transpose is the $a$-th row of $Z$. Thus \begin{align*} Z_a &= \sum_{i=1}^n q_{a i}Y_i, \end{align*} where $q_{a i}$ is the $i$-th component of $q_a$. For $a \in \{1,\dots,n-1\}$, since $\sum_{i=1}^n q_{a i}=q_a^\top\mathbf{1}_n=0$, the mean of $Z_a$ is \begin{align*} \mathbb{E}[Z_a] &= \sum_{i=1}^n q_{a i}\mu \\ &= 0. \end{align*} Its covariance matrix is \begin{align*} \operatorname{Cov}(Z_a) &= \sum_{i=1}^n q_{a i}^2\Sigma \\ &= \Sigma, \end{align*} because $|q_a|^2=1$. If $a,b \in \{1,\dots,n-1\}$ with $a \neq b$, then \begin{align*} \operatorname{Cov}(Z_a,Z_b) &= \sum_{i=1}^n q_{a i}q_{b i}\Sigma \\ &= (q_a^\top q_b)\Sigma \\ &=0. \end{align*} Since $(Z_1,\dots,Z_n)$ is obtained from the jointly Gaussian vector $(Y_1,\dots,Y_n)$ by a linear transformation, it is jointly Gaussian. Therefore the zero cross-covariances imply that $Z_1,\dots,Z_{n-1}$ are independent, and we have \begin{align*} Z_1,\dots,Z_{n-1} \overset{\mathrm{ind}}{\sim} \mathcal{N}_p(0,\Sigma). \end{align*} [guided] The purpose of multiplying by $Q$ is to replace the original observations by orthogonal linear combinations: $n-1$ contrast directions and one mean direction. Define \begin{align*} Z &= QY, \end{align*} and let $Z_a: \Omega \to \mathbb{R}^p$ be the column vector whose transpose is row $a$ of $Z$. Because row $a$ of $Q$ is $q_a^\top$, we have \begin{align*} Z_a = \sum_{i=1}^n q_{a i}Y_i. \end{align*} For $a \leq n-1$, the vector $q_a$ lies in $\mathbf{1}_n^\perp$, so its coordinates sum to zero. Therefore the common mean $\mu$ cancels: \begin{align*} \mathbb{E}[Z_a] &= \mathbb{E}\left[\sum_{i=1}^n q_{a i}Y_i\right] \\ &= \sum_{i=1}^n q_{a i}\mathbb{E}[Y_i] \\ &= \sum_{i=1}^n q_{a i}\mu \\ &= \left(\sum_{i=1}^n q_{a i}\right)\mu \\ &=0. \end{align*} The covariance calculation uses independence of the original observations. Since $\operatorname{Cov}(Y_i,Y_j)=0$ for $i \neq j$ and $\operatorname{Cov}(Y_i)=\Sigma$, we get \begin{align*} \operatorname{Cov}(Z_a) &= \operatorname{Cov}\left(\sum_{i=1}^n q_{a i}Y_i\right) \\ &= \sum_{i=1}^n q_{a i}^2\operatorname{Cov}(Y_i) \\ &= \left(\sum_{i=1}^n q_{a i}^2\right)\Sigma \\ &= \Sigma. \end{align*} For two distinct contrast rows $a,b \leq n-1$, the same independence calculation gives \begin{align*} \operatorname{Cov}(Z_a,Z_b) &= \sum_{i=1}^n q_{a i}q_{b i}\operatorname{Cov}(Y_i) \\ &= \left(\sum_{i=1}^n q_{a i}q_{b i}\right)\Sigma \\ &= (q_a^\top q_b)\Sigma \\ &=0. \end{align*} The vector $(Z_1,\dots,Z_n)$ is a linear transformation of the jointly Gaussian vector $(Y_1,\dots,Y_n)$, hence is jointly Gaussian. For jointly Gaussian random vectors, vanishing cross-covariance is equivalent to independence. Thus \begin{align*} Z_1,\dots,Z_{n-1} \overset{\mathrm{ind}}{\sim} \mathcal{N}_p(0,\Sigma). \end{align*} [/guided] [/step] [step:Identify the centered cross-product matrix as a Wishart sum] Using the diagonalization $P=Q^\top DQ$ and the definition $Z=QY$, we obtain \begin{align*} (n-1)S &= Y^\top P Y \\ &= Y^\top Q^\top DQY \\ &= Z^\top D Z \\ &= \sum_{a=1}^{n-1} Z_a Z_a^\top. \end{align*} By the defining representation of the Wishart distribution, the sum of $m$ independent outer products $X_aX_a^\top$, where \begin{align*} X_1,\dots,X_m \overset{\mathrm{ind}}{\sim} \mathcal{N}_p(0,\Sigma), \end{align*} has distribution $W_p(m,\Sigma)$. Applying this definition with $m=n-1$ and $X_a=Z_a$ gives \begin{align*} (n-1)S \sim W_p(n-1,\Sigma). \end{align*} This is the desired conclusion. [/step]