《计算机视觉》习题解答（二）第3章

习题解答

$答案仅供参考$

3.1

$$E[x]=1\times \lambda +0\times (1-\lambda )=\lambda$$
$由上一章习题2.10可得：$
$$E[(x-\mu {)}^2]=E[x^2]-E[x]E[x]$$
$又E[x]=\lambda =\mu ，所以$
$$\begin{array}{rl}E[(x-E[x]{)}^2]& =E[(x-\mu {)}^2]\\ & =E[x^2]-E[x]E[x]\\ & =1\times \lambda -{\lambda }^2\\ & =\lambda (1-\lambda )\end{array}$$
$得证$

3.2

$已知\alpha ，\beta >1$
$$Pr(\lambda )=\frac{\Gamma [\alpha +\beta ]}{\Gamma [\alpha ]\Gamma [\beta ]}{\lambda }^{\alpha -1}(1-\lambda {)}^{\beta -1}$$
$对\lambda 进行求导$
$$\begin{array}{rl}P{r}^{\prime }(\lambda )& =\frac{\Gamma [\alpha +\beta ]}{\Gamma [\alpha ]\Gamma [\beta ]}[(\alpha -1){\lambda }^{\alpha -2}(1-\lambda {)}^{\beta -1}+{\lambda }^{\alpha -1}(\beta -1)(1-\lambda {)}^{\beta -2}\ast (-1)]\end{array}$$
$令P{r}^{\prime }(\lambda )=0，解得$
$$\lambda =\frac{\alpha -1}{(\alpha -1)+(\beta -1)}$$
$代入可得$
$$\text{mode}=\frac{\Gamma [\alpha +\beta ]}{\Gamma [\alpha ]\Gamma [\beta ]}{\left(\frac{\alpha -1}{(\alpha -1)+(\beta -1)}\right)}^{\alpha -1}{\left(\frac{\beta -1}{(\alpha -1)+(\beta -1)}\right)}^{\beta -1}$$

3.3

$联列方程组$
$$\{\begin{array}{ll}& \mu =\frac{\alpha }{\alpha +\beta }\\ & {\sigma }^2=\frac{\alpha \beta }{(\alpha +\beta {)}^2(\alpha +\beta +1)}\end{array}$$
$从上式得到\beta =\frac{\alpha }{\mu }-\alpha ，代入下式得到$
$$\{\begin{array}{ll}& \alpha =\frac{{\mu }^2(1-\mu )}{{\sigma }^2}-\mu \\ & \beta =\frac{\mu (1-{\mu }^2)}{{\sigma }^2}+\mu -1\end{array}$$

3.4

$已知贝塔分布$
$$\begin{array}{rl}Pr(\lambda )& =\frac{\Gamma [\alpha +\beta ]}{\Gamma [\alpha ]\Gamma [\beta ]}{\lambda }^{\alpha -1}(1-\lambda {)}^{\beta -1}\\ & =\text{exp}[log[\Gamma [\alpha +\beta ]]+(\alpha -1)log[\lambda ]+(\beta -1)log[1-\lambda ]-log[\Gamma [\alpha ]]-log[\Gamma [\beta ]]]\\ & =\text{exp}[(\alpha -1)log[\lambda ]+(\beta -1)log[1-\lambda ]-[log[\Gamma [\alpha ]]+log[\Gamma [\beta ]]-log[\Gamma [\alpha +\beta ]]]]\end{array}$$
$已知指数族的形式为$
$$Pr(x|\mathbf{\theta })=a[x]\text{exp}[\mathbf{b}[\mathbf{\theta }{]}^{\text{T}}\mathbf{c}[x]-d[\mathbf{\theta }]]$$
$对比可得$
$$\{\begin{array}{ll}a[x]& =1\\ \mathbf{b}[\mathbf{\theta }]& =\left[\begin{array}{c}\alpha -1\\ \beta -1\end{array}\right]\\ & \\ \mathbf{c}[x]& =\left[\begin{array}{c}log[x]\\ log[1-x]\end{array}\right]\\ d[\mathbf{\theta }]& =log[\Gamma [\alpha ]]+log[\Gamma [\beta ]]-log[\Gamma [\alpha +\beta ]]\end{array}$$

3.5

$$\begin{array}{rl}\Gamma [z+1]& ={\int }_0^{\infty }{t}^z{\text{e}}^{-t}\text{d}t\\ & =-{\int }_0^{\infty }{t}^z\text{d}({\text{e}}^{-t})\\ & =-{t^z{\text{e}}^{-t}|}_0^{\infty }+{\int }_0^{\infty }{\text{e}}^{-t}\text{d}(t^z)\\ & =0+z\int {\text{e}}^{-t}{t}^{z-1}\text{d}t\\ & =z\Gamma [z]\end{array}$$
$得证$

3.6

书上关于共轭分布讲的比较简单，较为详细的可以参见https://www.jianshu.com/p/bb7bce40a15a

$需要证明先验分布和后验分布有相同的形式，即证明$
$$Pr(\mu )\ast Pr(x|\mu )\propto Pr(\mu )$$
$$\begin{array}{rl}Pr(x|\mu )\ast Pr(\mu )& =\frac{1}{2\pi {\sigma }_p}\text{exp}\left[-0.5\left[(x-\mu {)}^2+\frac{}{}\frac{(\mu -{\mu }_p{)}^2}{{\sigma }_p^2}\right]\right]\\ & =\frac{1}{2\pi {\sigma }_p}\text{exp}\left[-0.5\left[x^2-2\mu x+{\mu }^2+\frac{{\mu }^2}{{\sigma }_p^2}-\frac{2\mu {\mu }_p}{{\sigma }_p^2}+\frac{{\mu }_p^2}{{\sigma }_p^2}\right]\right]\\ & =\frac{1}{2\pi {\sigma }_p}\text{exp}\left[-0.5\left[x^2+\frac{{\mu }_p^2}{{\sigma }_p^2}\right]-0.5\left[{\mu }^2(1+\frac{1}{{\sigma }_p^2})-2\mu (x+\frac{{\mu }_p}{{\sigma }_p^2})\right]\right]\\ & (因为提取出来的前两项与\mu 无关，可以将其作为常量)\\ & ={\kappa }_1\text{exp}\left[-0.5\left(\frac{{\sigma }_p^2+1}{{\sigma }_p^2}{\mu }^2-2\frac{{\sigma }_p^{2}x+{\mu }_p}{{\sigma }_p^2}\mu \right)\right]\\ & (这里使用配方法)\\ & ={\kappa }_1\text{exp}\left[-0.5\left[\frac{{\left(\mu -\frac{{\sigma }_p^{2}x+{\mu }_p}{{\sigma }_p^2+1}\right)}^2}{\frac{{\sigma }_p^2}{{\sigma }_p^2+1}}-常量\right]\right]\\ & ={\kappa }_2\text{exp}\left[-0.5\frac{{\left(\mu -\frac{{\sigma }_p^{2}x+{\mu }_p}{{\sigma }_p^2+1}\right)}^2}{\frac{{\sigma }_p^2}{{\sigma }_p^2+1}}\right]\\ & \propto Pr(\mu )\end{array}$$
$得证$

3.7

$和3.4相似，已知正态分布$
$$\begin{array}{rl}Pr(x)& =\frac{1}{\sqrt{2\pi {\sigma }^2}}\text{exp}\left[-0.5\frac{(x-\mu {)}^2}{{\sigma }^2}\right]\\ & =\text{exp}\left[-0.5(\frac{x^2}{{\sigma }^2}-\frac{2\mu x}{{\sigma }^2}+\frac{{\mu }^2}{{\sigma }^2})-log\left[\sqrt{2\pi {\sigma }^2}\right]\right]\end{array}$$
$已知指数族的形式为$
$$Pr(x|\mathbf{\theta })=a[x]\text{exp}[\mathbf{b}[\mathbf{\theta }{]}^{\text{T}}\mathbf{c}[x]-d[\mathbf{\theta }]]$$
$对比可得$
$$\{\begin{array}{ll}a[x]& =1\\ \mathbf{b}[\mathbf{\theta }]& =\left[\begin{array}{c}1/{\sigma }^2\\ \mu /{\sigma }^2\end{array}\right]\\ & \\ \mathbf{c}[x]& =\left[\begin{array}{c}-0.5x^2\\ x\end{array}\right]\\ d[\mathbf{\theta }]& =\frac{{\mu }^2}{2{\sigma }^2}+log\left[\sqrt{2\pi {\sigma }^2}\right]\end{array}$$

3.8

$已知$
$$Pr(\mu ,{\sigma }^2)=\frac{\sqrt{\gamma }}{\sigma \sqrt{2\pi }}\frac{{\beta }^{\alpha }}{\Gamma [\alpha ]}{\left(\frac{1}{{\sigma }^2}\right)}^{\alpha +1}\text{exp}\left[-\frac{2\beta +\gamma (\delta -\mu {)}^2}{2{\sigma }^2}\right]$$
$对\mu 进行求导$
$$\begin{array}{r}\frac{\partial Pr(\mu ,{\sigma }^2)}{\partial \mu }=\text{c}onstant\cdot \text{e}[\cdots ]\left(-\frac{1}{2{\sigma }^2}\right)(2\gamma )(\delta -\mu )(-1)\end{array}$$
$令\frac{\partial Pr(\mu ,{\sigma }^2)}{\partial \mu }=0，解得$
$$\mu =\delta$$

$对{\sigma }^2进行求导$
$$\begin{array}{r}\frac{\partial Pr(\mu ,{\sigma }^2)}{\partial {\sigma }^2}=balabala\end{array}$$
$令\frac{\partial Pr(\mu ,{\sigma }^2)}{\partial {\sigma }^2}=0，代入\mu 的值，解得$
$${\sigma }^2=\frac{2\beta }{2\alpha +3}$$
$综上$
$$\{\begin{array}{ll}\mu & =\delta \\ {\sigma }^2& =\frac{2\beta }{2\alpha +3}\end{array}$$
$代入可得$
$$\text{mode}=balabala(自己代进去就可以)$$

3.9

$$\begin{array}{rl}& \prod _{i=1}^I{\text{Bern}}_{x_i}[\lambda ]\cdot {\text{Beta}}_{\lambda }[\alpha ,\beta ]\\ =& \prod _{i=1}^I{\lambda }^{x_i}(1-\lambda {)}^{1-x_i}\frac{\Gamma [\alpha +\beta ]}{\Gamma [\alpha ]\Gamma [\beta ]}{\lambda }^{\alpha -1}(1-\lambda {)}^{\beta -1}\\ =& \kappa \cdot \frac{\Gamma [\alpha +\beta +I]}{\Gamma \left[\alpha +\sum x_i\right]\Gamma [\beta +\sum (1-x_i)]}{\lambda }^{\alpha +\sum x_i-1}(1-\lambda {)}^{\beta +\sum (1-x_i)-1}\\ =& \kappa \cdot {\text{Beta}}_{\lambda }[\tilde{\alpha },\tilde{\beta }]\end{array}$$
$得证$

3.10

$$\begin{array}{rl}& \prod _{i=1}^I{\text{Cat}}_{x_i}[{\lambda }_{1\cdots K}]\cdot {\text{Dir}}_{{\lambda }_{1\cdots K}}[{\alpha }_{1\cdots K}]\\ =& \left[\prod _{i=1}^I\prod _{j=1}^K{\lambda }_j^{x_{ij}}\right]\frac{\Gamma \left[\sum _{j=1}^K{\alpha }_j\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j]}\prod _{j=1}^K{\lambda }_j^{{\alpha }_j-1}\\ & (x_{ij}是x_i取到j号结果的次数)\\ =& \left[\prod _{j=1}^K{\lambda }_j^{{\sum }_{i=1}^I{x}_{ij}}\right]\frac{\Gamma \left[\sum _{j=1}^K{\alpha }_j\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j]}\prod _{j=1}^K{\lambda }_j^{{\alpha }_j-1}\\ =& \frac{\Gamma \left[\sum _{j=1}^K{\alpha }_j\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j]}\prod _{j=1}^K{\lambda }_j^{{\alpha }_j-1+N_j}\\ & (设\sum _{i=1}^I{x}_{ij}=N_j，即取I次总共取到j号结果的总次数)\\ =& \frac{\Gamma \left[\sum _{j=1}^K{\alpha }_j\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j]}\frac{\prod _{j=1}^K\Gamma [{\alpha }_j+N_j]}{\Gamma \left[\sum _{j=1}^K({\alpha }_j+N_j)\right]}\frac{\Gamma \left[\sum _{j=1}^K({\alpha }_j+N_j)\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j+N_j]}\prod _{j=1}^K{\lambda }_j^{{\alpha }_j-1+N_j}\\ & （\because \sum _{j=1}^K{N}_j=I，即取I次总共取到所有结果的总次数，因为取了I次，所以等于I）\\ =& \frac{\Gamma \left[\sum _{j=1}^K{\alpha }_j\right]}{\prod _{j=1}^K\Gamma [{\alpha }_j]}\frac{\prod _{j=1}^K\Gamma [{\alpha }_j+N_j]}{\Gamma \left[I+\sum _{j=1}^K{\alpha }_j\right]}\cdot {\text{Dir}}_{{\lambda }_{1\cdots K}}\left[\widetilde{{\alpha }_{1\cdots K}}\right]\\ & \\ =& \tilde{\kappa }\cdot {\text{Dir}}_{{\lambda }_{1\cdots K}}\left[\widetilde{{\alpha }_{1\cdots K}}\right]\end{array}$$
$得证$

3.11

$$\begin{array}{rl}& \prod _{i=1}^I{\text{Norm}}_{x_i}\left[\mu ,{\sigma }^2\right]\cdot {\text{NormInvGam}}_{\mu ,{\sigma }^2}[\alpha ,\beta ,\gamma ,\delta ]\\ =& \prod _{i=1}^I\frac{1}{\sqrt{2\pi {\sigma }^2}}\text{exp}\left[-0.5\frac{(x_i-\mu {)}^2}{{\sigma }^2}\right]\cdot \frac{\sqrt{\gamma }}{\sigma \sqrt{2\pi }}\frac{{\beta }^{\alpha }}{\Gamma [\alpha ]}{\left(\frac{1}{{\sigma }^2}\right)}^{\alpha +1}\text{exp}\left[-\frac{2\beta +\gamma (\delta -\mu {)}^2}{2{\sigma }^2}\right]\\ =& \kappa \cdot \frac{\sqrt{\tilde{\gamma }}{\tilde{\beta }}^{\tilde{\alpha }}}{\Gamma [\tilde{\alpha }]}{\left(\frac{1}{\sigma }\right)}^{2\alpha +2+I}\frac{1}{\sigma \sqrt{2\pi }}\text{exp}\left[-\frac{\sum _{i=1}^I(x_i-\mu {)}^2+2\beta +\gamma (\delta -\mu {)}^2}{2{\sigma }^2}\right]\\ =& \kappa \cdot \frac{\sqrt{\tilde{\gamma }}}{\sigma \sqrt{2\pi }}\frac{{\tilde{\beta }}^{\tilde{\alpha }}}{\Gamma [\tilde{\alpha }]}{\left(\frac{1}{{\sigma }^2}\right)}^{\tilde{\alpha }+1}\text{exp}\left[-\frac{2\tilde{\beta }+\tilde{\gamma }(\tilde{\delta }-\mu {)}^2}{2{\sigma }^2}\right]\\ =& \kappa \cdot {\text{NormInvGam}}_{\mu ,{\sigma }^2}[\tilde{\alpha },\tilde{\beta },\tilde{\gamma },\tilde{\delta }]\end{array}$$
$得证$

3.12

待定
$$\begin{array}{rl}& \prod _{i=1}^I{\text{Norm}}_{{\mathbf{x}}_{\mathbf{i}}}\left[\mathbf{\mu },\mathbf{\Sigma }\right]\cdot {\text{NorIWis}}_{\mathbf{\mu },\mathbf{\Sigma }}[\alpha ,\mathbf{\Psi },\gamma ,\mathbf{\delta }]\\ =& \prod _{i=1}^I\frac{1}{(2\pi {)}^{D/2}|\mathbf{\Sigma }{|}^{1/2}}\text{exp}\left[-0.5(\mathbf{x}-\mathbf{\mu }{)}^{\text{T}}{\mathbf{\Sigma }}^{-1}(\mathbf{x}-\mathbf{\mu })\right]\cdot \frac{{\gamma }^{D/2}|\mathbf{\Psi }{|}^{\alpha /2}|\mathbf{\Sigma }{|}^{-(\alpha +D+2)/2}}{2^{\alpha D/2}(2\pi {)}^{D/2}{\Gamma }_D[\alpha /2]}\text{exp}\left[-0.5\left(\text{Tr}[\mathbf{\Psi }{\mathbf{\Sigma }}^{-1}]+\gamma (\mathbf{\mu }-\mathbf{\delta }{)}^{\text{T}}{\mathbf{\Sigma }}^{-1}(\mathbf{\mu }-\mathbf{\delta }){\gamma }^{D/2}\right)\right]\\ =& \kappa \cdot \frac{{\tilde{\gamma }}^{D/2}|\tilde{\mathbf{\Psi }}{|}^{\tilde{\alpha }/2}|\mathbf{\Sigma }{|}^{-(\alpha +D+2+I)/2}}{2^{D(\alpha +I)/2}(2\pi {)}^{D/2}{\Gamma }_D[\tilde{\alpha }/2]}\\ =& \kappa \cdot \frac{{\tilde{\gamma }}^{D/2}|\tilde{\mathbf{\Psi }}{|}^{\tilde{\alpha }/2}|\mathbf{\Sigma }{|}^{-(\tilde{\alpha }+D+2)/2}}{2^{D\tilde{\alpha }/2}(2\pi {)}^{D/2}{\Gamma }_D[\tilde{\alpha }/2]}\end{array}$$