I am trying to understand the proof of the Pearson-Aitken selection formula, widely used in statistical genetics. A proof that the formula is general is given by Aitken (1936). However, I failed to understand the derivation of equation (4) in this paper, given below:

Let $\phi(\boldsymbol{x},\boldsymbol{y})$ denote the probability density function (pdf) a standard zero-mean unit-variance multivariate Normal distribution, with two components $\boldsymbol{x}$ and $\boldsymbol{y}$. The moments generating function (mgf) is given by: \begin{align} G(\boldsymbol{\alpha}, \boldsymbol{\beta}) &= \int \phi(\boldsymbol{x},\boldsymbol{y}) \exp(\boldsymbol{\alpha}^T\boldsymbol{x}+\boldsymbol{\beta}^T\boldsymbol{y}) d\boldsymbol{x} d\boldsymbol{y} \\\\ & = 1 + \frac{1}{2}(\boldsymbol{\alpha}, \boldsymbol{\beta})^T\boldsymbol{V}(\boldsymbol{\alpha}, \boldsymbol{\beta}) + \cdots \end{align}

Denoting the variance and covariance matrices of $\boldsymbol{x}$ and $\boldsymbol{y}$ by $\boldsymbol{V}_x$, $\boldsymbol{V}_y$, and $\boldsymbol{V}_{xy}$, Aitken (1936) then writes in equation (4): \begin{equation} G(\boldsymbol{\alpha}, \boldsymbol{\beta}) = F(\boldsymbol{\alpha}, \boldsymbol{\beta}) \int \phi(\boldsymbol{x}) \exp((\boldsymbol{\alpha} + \boldsymbol{V}_x^{-1} \boldsymbol{V}_{xy}\boldsymbol{\beta})^T\boldsymbol{x}) d \boldsymbol{x} \end{equation} where he calls $F(\boldsymbol{\alpha}, \boldsymbol{\beta})$ the "partial generating function" which does not contain $\boldsymbol{\alpha}$ in its terms of first and second degree.

Can anyone give me a more detailed explanation of this derivation?

Many thanks in advance.


Actually, I found out that the above can be proved using basic multivariate calculus.

Let $\phi(\boldsymbol{x})$ and $\phi(\boldsymbol{y|x})$ denote the marginal and conditional distribution, respectively. We can write: \begin{equation} G(\boldsymbol{\alpha}, \boldsymbol{\beta}) = \int \phi(\boldsymbol{x})\phi(\boldsymbol{y}|\boldsymbol{x}) \exp (\boldsymbol{\alpha}^T\boldsymbol{x})\exp(\boldsymbol{\beta}^T\boldsymbol{y})d\boldsymbol{x} d\boldsymbol{y} \end{equation} Noting that the conditional mean of $\boldsymbol{y}$ given $\boldsymbol{x}$ is given by $ \boldsymbol{V}_{yx} \boldsymbol{V}_x^{-1} \boldsymbol{x}$, we let $\boldsymbol{z} = \boldsymbol{y} - \boldsymbol{V}_{yx}\boldsymbol{V}_x^{-1}\boldsymbol{x}$, and observe that \begin{equation} \left |\frac{d\boldsymbol{z}}{d\boldsymbol{y}}\right | = |\boldsymbol{I}| = 1 \end{equation} Thus, \begin{equation} G(\boldsymbol{\alpha}, \boldsymbol{\beta}) = \int \phi(\boldsymbol{x})\phi_{\boldsymbol{V}_{y|x}}(\boldsymbol{z}) \exp (\boldsymbol{\alpha}^T\boldsymbol{x})\exp(\boldsymbol{\beta}^T(\boldsymbol{z} + \boldsymbol{V}_{yx}\boldsymbol{V}_x^{-1}\boldsymbol{x}))\left |\frac{d\boldsymbol{y}}{d\boldsymbol{z}}\right | d\boldsymbol{z} d\boldsymbol{x} \end{equation} where $\phi_{\boldsymbol{V}_{y|x}}(\boldsymbol{z})$ denotes a zero-centred multivariate Normal with variance $\boldsymbol{V}_{y|x} = \boldsymbol{V}_y - \boldsymbol{V}_{yx}\boldsymbol{V}_x^{-1}\boldsymbol{V}_{xy}$. Thus writing $F(\boldsymbol{\alpha}, \boldsymbol{\beta}) = \int \phi_{\boldsymbol{V}_{y|x}}(\boldsymbol{z}) \exp(\boldsymbol{\beta}^T\boldsymbol{z})d\boldsymbol{z}$, we obtain the required.

  • $\begingroup$ It should be noted that Aitken's proof is more general though, and multivariate Normal in x and y is not required. $\endgroup$ – Tim Mak Jan 10 '19 at 6:24

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.