0
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 3 years have passed since last update.

PRML 演習問題3.20 解答

Last updated at Posted at 2020-06-07

##問題

(3.86)の$\mathbf{\alpha}$に関する最大化が再推定方程式(3.92)に帰着されることを示すのに必要なすべての段階を(3.86)から始めて確かめよ

\begin{align*}
\mathrm{lnp(t|\alpha,\beta)=\frac{M}{2}ln{\alpha}+\frac{N}{2}ln{\beta}-E(m_N)-\frac{1}{2}ln|A|-\frac{N}{2}ln(2\pi)}
\tag{3.86}\\
\mathrm{{\alpha}=\frac{\gamma}{m_N^Tm_N}}
\tag{3.92}
\end{align*}

##方針
PRML本文の3.5.2にしたがって式変形を行う。まずは(3.86)の項の内$\mathbf{\alpha}$に関する項に着目し取り組んでいく。

##解答

\begin{align*}
\mathrm{A={\alpha}I+{\beta\Phi\Phi^T}}
\tag{3.81}\\
\mathrm{E(m_N)=\frac{\beta}{2}{\parallel}t-{\Phi}^Tm_N{\parallel}^2+\frac{\alpha}{2}m_N^Tm_N}
\tag{3.82}
\end{align*}

はじめに(3.87)の固有ベクトル方程式を考える

\begin{align*}
\mathrm{\Big({\beta\Phi\Phi^T}\big)u_i=\lambda_iu_i}
\tag{3.87}
\end{align*}

(3.81)よりAは固有値$\mathrm{{\alpha}+\lambda_i}$をもつ
したがって(3.86)の第4項の$\mathbf{\alpha}$に関する導関数を考えると以下のようになる

\begin{align*}
\mathrm{\frac{d}{d{\alpha}}ln|A|=\frac{d}{d{\alpha}}ln{\prod_i}(\lambda_i+\alpha)=\frac{d}{d{\alpha}}{\sum_i}ln(\lambda_i+\alpha)}
\tag{3.88}
\end{align*}

よって(3.86)の$\mathbf{\alpha}$に関する導関数は以下で与えられる

\begin{align*}
\mathrm{\frac{d}{d\alpha}lnp(t|{\alpha},{\beta})}=\mathrm{\frac{1}{2}\Big(\frac{M}{\alpha}-m_N^Tm_N-\sum_i\frac{1}{\lambda_i+\alpha}\Big)}
\end{align*}

導関数=0とおくと(3.86)の$\mathbf{\alpha}$に関する停留点は以下の(3.89.a)を満たす

\begin{align*}
\mathrm{\frac{1}{2}\Big(\frac{M}{\alpha}-m_N^Tm_N-\sum_i\frac{1}{\lambda_i+\alpha}\Big)}=0
\tag{3.89.a}
\end{align*}

整理すると

\begin{align*}
\mathrm{{\alpha}m_N^Tm_N=M-\alpha\sum_i\frac{1}{\lambda_i+\alpha}}
\tag{3.90.a}
\end{align*}

Mは以下で示すことができるので

\begin{align*}
\mathrm{M=\sum_i^M\frac{\lambda_i+\alpha}{\lambda_i+\alpha}}
\end{align*}

(3.90.a)を変形し$\mathbf{\gamma}$とおく

\begin{align*}
\mathrm{{\alpha}m_N^Tm_N=\sum_i\frac{\lambda_i}{\lambda_i+\alpha}=\gamma}
\end{align*}

したがって周辺尤度を最大化する$\mathbf{\alpha}$の値は以下の(3.92)を満たす

\begin{align*}
\mathrm{{\alpha}=\frac{\gamma}{m_N^Tm_N}}
\tag{3.92}
\end{align*}

よって題意は示された

0
0
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
0
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?