1
$\begingroup$

Here is the setting:

The likelihood of data is \begin{align} p(\boldsymbol{x} | \mu, \sigma^2) &= (\frac{1}{2\pi \sigma^2})^{\frac{n}{2}} exp\left\{ -\frac{1}{2\sigma^2} \sum\limits_{i=1}^n (x_i - \mu)^2 \right\} \nonumber \\ &= \frac{1}{(2\pi)^{n/2}} (\sigma^2)^{-n/2} exp\left\{ -\frac{1}{2\sigma^2} \left[ \sum\limits_{i=1}^n (x_i - \overline{x})^2 + n(\overline{x} - \mu)^2 \right] \right\}, \nonumber \end{align} and we use the Normal-inverse-Gamma as prior \begin{align} p(\mu , \sigma^2) &= \mathcal{N} (\mu | \mu_0 , \sigma^2 V_0) IG(\sigma^2 | \alpha_0 , b_0 ) \nonumber \\ &= \frac{1}{\sqrt{2\pi V_0}} \frac{b_0^{\alpha_0}}{\Gamma(\alpha_0)}\frac{1}{\sigma} (\sigma^2)^{-\alpha_0 - 1} exp\left( -\frac{1}{2\sigma^2} [V_0^{-1}(\mu - \mu_0 )^2 + 2b_0] \right). \nonumber \end{align} Then, the posterior can be derivated via \begin{align} p(\mu , \sigma^2 | \boldsymbol{x}) &\propto p(\boldsymbol{x} | \mu , \sigma^2 ) p(\mu , \sigma^2) \nonumber \\ &\propto \left[ (\sigma^2)^{-n/2} exp \left( -\frac{1}{2\sigma^2} \big[\sum\limits_{i=1}^b (x_i - \overline{x})^2 + n(\overline{x} - \mu )^2\big] \right) \right] \times \left[ \sigma^{-1} (\sigma^2)^{-\alpha_0 - 1} exp \left( -\frac{1}{2\sigma^2} \left[ V_0^{-1} (\mu - \mu_0 )^2 + 2 b_0 \right] \right) \right] \nonumber \\ &= \sigma^{-1} (\sigma^2)^{-(\alpha_0 + \frac{n}{2}) - 1} exp \left( -\frac{1}{2\sigma^2} \big[ V_0^{-1} (\mu - m_0 )^2 + 2 b_0 + \sum\limits_{i=1}^n (x_i - \overline{x})^2 + n(\overline{x} - \mu)^2 \big] \right) \nonumber \\ &= \sigma^{-1} (\sigma^2)^{-(\alpha_0 + \frac{n}{2}) - 1} exp \Big\{ -\frac{1}{2\sigma^2} \Big[ (V_0^{-1} + n)(\mu - \frac{V_0^{-1} m_0 + n\overline{x}}{V_0^{-1} + n})^2 + \big(b_0 + \frac{1}{2}\sum\limits_{i=1}^n (x_i - \overline{x})^2 + \frac{V_0^{-1} n}{2(V_0^{-1} + n)} (m_0 - \overline{x})^2 \big) \Big] \Big\} \nonumber \end{align} We recognize this is an unnormalized Normal-inverse-Gamma distribution, therefore \begin{align} p(\mu , \sigma^2 | \boldsymbol{x}) = NIG(\mu , \sigma^2 | m_n , V_n , \alpha_n , b_n ), \nonumber \end{align} where \begin{align} m_n &= \frac{V_0^{-1} m_0 + n \overline{x}}{V_0^{-1} + n} \nonumber \\ V_n^{-1} &= V_0^{-1} + n \nonumber \\ \alpha_n &= \alpha_0 + \frac{n}{2} \nonumber \\ b_n &= b_0 + \frac{1}{2}\sum\limits_{i=1}^n (x_i - \overline{x})^2 + \frac{V_0^{-1} n}{2(V_0^{-1} + n)}(m_0 - \overline{x})^2. \nonumber \end{align} As indicated in this paper (see Eq(200)), the last term can be further expressed as \begin{align} b_n &= b_0 + \frac{1}{2} \left[ m_0^2 V_0^{-1} + \sum\limits_{i=1}^n x_i^2 - m_n^2 V_n^{-1} \right]. \nonumber \end{align} But I fail to prove it, i.e., \begin{align} \sum\limits_{i=1}^n (x_i - \overline{x})^2 + \frac{V_0^{-1} n}{(V_0^{-1} + n)}(m_0 - \overline{x})^2 &= \left[ m_0^2 V_0^{-1} + \sum\limits_{i=1}^n x_i^2 - m_n^2 V_n^{-1} \right]. \nonumber \end{align}

$\endgroup$
1
$\begingroup$

This is much simpler to prove compared with your earlier question. \begin{align} \sum\limits_{i=1}^n (x_i - \overline{x})^2 &+ \frac{V_0^{-1} n}{(V_0^{-1} + n)}(m_0 - \overline{x})^2 = \sum\limits_{i=1}^n x_i^2 - n\overline{x}^2\\ &\qquad + \frac{V_0^{-1} n}{(V_0^{-1} + n)}(m_0^2 - 2 m_0\overline{x} + \overline{x}^2)\\ &= \sum\limits_{i=1}^n x_i^2 - n\overline{x}^2 + \frac{V_0^{-1} (n+V_0^{-1}-V_0^{-1})}{(V_0^{-1} + n)}m_0^2\\ &\quad-2\frac{V_0^{-1} nm_0\overline{x}}{(V_0^{-1} + n)} +\frac{(V_0^{-1}+n-n) n}{(V_0^{-1} + n)}\overline{x}^2\\ &= \sum\limits_{i=1}^n x_i^2 + V_0^{-1}m_0^2 - \frac{V_0^{-2}m_0^2}{(V_0^{-1} + n)}\\ &\quad -2\frac{V_0^{-1} m_0n\overline{x}}{(V_0^{-1} + n)} -\frac{n^2\overline{x}^2}{(V_0^{-1} + n)}\\ &= V_0^{-1}m_0^2 + \sum\limits_{i=1}^n x_i^2 - \frac{(n\overline x+V_0^{-1} m_0)^2}{V_n^{-1}}\\ &= V_0^{-1}m_0^2 + \sum\limits_{i=1}^n x_i^2 -V_n^{-1}m_n^2 \end{align}

$\endgroup$

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.