\begin{align*}
& σN+1^2(\x) - σ_N^2(\x)
= & \bphi(\x)^\T \SN+1 \bphi(\x) - \bphi(\x)^\T \S_N \bphi(\x) \
= & \bphi(\x)^\T (\SN+1 - \S_N) \bphi(\x) \
= & - \bphi(\x)^\T \l( \f{β \S_0 ΦN+1^\T ΦN+1 \S_0}
{1 + β ΦN+1 \S_0 ΦN+1^\T}
- \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T} \r) \bphi(\x)
= & - \bphi(\x)^\T \l( \f{ β \S_0 Φ_N^\T Φ_N \S_0
- β \S_0 \bphi(\xN+1)^\T \bphi(\xN+1) \S_0 }
{ 1 + β Φ_N \S_0 Φ_N^\T
- β \bphi(\xN+1) \S_0 \bphi(\xN+1)^\T }
- \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T} \r) \bphi(\x)
\end{align*}
\begin{align*} & (β \S_0 Φ_N^\T Φ_N \S_0 + β \S_0 \bphi(\xN+1)^\T \bphi(\xN+1)) (1 + β Φ_N \S_0 Φ_N^\T)
- (β \S_0 Φ_N^\T Φ_N \S_0)
(1 + β Φ_N \S_0 Φ_N^\T + β \bphi(\xN+1) \S_0 \bphi(\xN+1)^\T)
\end{align*}
\begin{align*} & (x + β \S_0 \bphi(\xN+1)^\T \bphi(\xN+1)) (1 + y)
- x (1 + y + β \bphi(\xN+1) \S_0 \bphi(\xN+1)^\T)
= & (β \S_0 \bphi(\xN+1)^\T \bphi(\xN+1)) (1 + y)
- x (β \bphi(\xN+1) \S_0 \bphi(\xN+1)^\T)
\end{align*}
\begin{align*}
σ_N^2(\x) = \f{1}{β} + \bphi(\x)^\T \S_N \bphi(\x)
\end{align*}
これが\(N\)の単調増加関数であることを示せばよい。
第2項に
\begin{align*}
\S_N-1 = & \S_0-1 + β Φ_N^\T Φ_N
\end{align*}
を代入すると
\begin{align*}
\bphi(\x)^\T (\S_0-1 + β Φ_N^\T Φ_N)-1 \bphi(\x)
\end{align*}
(3.110) (Woodburyの公式(C.7)で\(\A=\M,\B=\v,\C=\v^\T,\D=1\)と置いたもの.)
\begin{align*}
(\M + \v \v^\T)-1 = \M-1 - \f{(\M-1 \v) (\v^\T \M-1)}{1 + \v^\T \M-1 \v}
\end{align*}
を用いて
\begin{align*}
& \bphi(\x)^\T (\S_0-1 + β Φ_N^\T Φ_N)-1 \bphi(\x) \
= & \bphi(\x)^\T \l(\S_0 - \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T}\r) \bphi(\x) \
\end{align*}
第2項の分子
\begin{align*}
& β \S_0 Φ_N^\T Φ_N \S_0 \
= & \bphi(\x)^\T \l(\S_0 - \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T}\r) \bphi(\x) \
\end{align*}
(3.59) \[ σ_N^2(\x) = \f{1}{β} + \bphi(\x)^\T \S_N \bphi(\x) \] (3.110) (Woodburyの公式(C.7)で\(\A=\M,\B=\v,\C=\v^\T,\D=1\)と置いたもの.) \[ (\M + \v \v^\T)-1 = \M-1 - \f{(\M-1 \v) (\v^\T \M-1)}{1 + \v^\T \M-1 \v} \]
\begin{align*} \S_N-1 = & \S_0-1 + β Φ_N^\T Φ_N \end{align*}
\begin{align*}
\S_N = & ( \S_0-1 + β Φ_N^\T Φ_N )-1
= & \S_0 - \f{β \S_0 Φ_N^\T Φ_N \S_0}{1 + β Φ_N \S_0 Φ_N^\T}
\end{align*}
\begin{align*}
& σN+1^2(\x) - σ_N^2(\x)
= & \bphi(\x)^\T \SN+1 \bphi(\x) - \bphi(\x)^\T \S_N \bphi(\x) \
= & - \bphi(\x)^\T \l( \f{β \S_0 ΦN+1^\T ΦN+1 \S_0}
{1 + β ΦN+1 \S_0 ΦN+1^\T}
- \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T} \r) \bphi(\x)
= & - \bphi(\x)^\T \l( \f{β \S_0 ΦN+1^\T ΦN+1 \S_0} {1 + β Φ_N \S_0 Φ_N^\T
- β \bphi(\xN+1)^\T \S_0 \bphi(\xN+1)}
- \f{β \S_0 Φ_N^\T Φ_N \S_0}
{1 + β Φ_N \S_0 Φ_N^\T} \r) \bphi(\x)
\end{align*}
\begin{align*}
E(\m_N) = & \f{β}{2} \|\tt - \bPhi \m_N\|^2 + \f{α}{2} \m_N^\T \m_N \tag{3.82}
\end{align*}
再推定方程式
\begin{align*}
α = & \f{γ}{\m_N^\T \m_N} \
\f{1}{β} = & \f{1}{N - γ} ∑n=1^N \{t_n - \m_N^\T \bphi(\x_n)\}^2 \
\end{align*}
を代入
\begin{align*}
E(\m_N) = & \f{1}{2} (N - γ) \l( ∑n=1^N \{t_n - \m_N^\T \bphi(\x_n)\}^2 \r)-1
\|\tt - \bPhi \m_N\|^2
- \f{1}{2} \f{γ}{\m_N^\T \m_N} \m_N^\T \m_N
= & \f{1}{2} (N - γ) + \f{1}{2} γ
= & \f{N}{2} \
\end{align*}
\begin{align*}
p(\x) = & \N(\x|\μ,\Λ-1) \tag{2.113}
p(\y|\x) = & \N(\y|\A \x + \b, \L-1) \tag{2.114} \
\end{align*}
ならば
\begin{align*}
p(\y) = & \N(\y|\A \μ + \b, \L-1 + \A \Λ-1 \A^\T) \tag{2.115} \
\end{align*}
これを
\begin{align*}
p(\w) = & \N(\w|0, α-1) \tag{3.52} \
p(\tt|\w) = & ∏n=1^N \N(t_n|\w^\T \bphi(\x_n), β-1) \tag{3.10} \
∝ & exp\l( \f{β}{2} ∑n=1^N (t_n - \w^\T \bphi(\x_n))^2 \r) \
= & exp\l( \f{β}{2} (\tt - \bPhi \w)^\T (\tt - \bPhi \w) \r) \
∝ & \N(\tt|\bPhi \w, β-1 \I)
\end{align*}
に適用すると
\begin{align*}
\μ = & 0 \
\Λ-1 = & α-1 \
\A = & \bPhi \
\b = & 0 \
\L-1 = & β-1 \I
\end{align*}
エビデンス関数
\begin{align*}
p(\tt) = & \N(\tt|\A \μ + \b, \L-1 + \A \Λ-1 \A^\T)
= & \N(\tt|0, β-1 \I + α-1 \bPhi \bPhi^\T) \
= & \f{1}{(2π)D/2} \f{1}{\l|β-1 \I + α-1 \bPhi \bPhi^\T\r|1/2}
exp\l\{ \tt^\T (β-1 \I + α-1 \bPhi \bPhi^\T)-1 \tt \r\} \
\end{align*}
(3.11)より
\begin{align*}
ln p(\tt|w,β) & = \f{N}{2} ln β - \f{N}{2} ln(2π) - β E_D(\w)
p(\tt|w,β) & = \l(\f{β}{2π}\r)N/2 exp\{ - β E_D(\w)\} \
\end{align*}
(3.52)より
\begin{align*}
p(\w|α) & = \N(\w|0,α-1\I)
& = \l(\f{α}{2π}\r)M/2 exp\{ - \f{α}{2} \w^\T \w \}
\end{align*}
(3.77)にこれらを代入
\begin{align*}
p(\tt|α,β) & = ∫ p(\tt|\w,β) p(\w|α) \d\w
& = \l(\f{β}{2π}\r)N/2 \l(\f{α}{2π}\r)M/2
∫ exp\{ - β E_D(\w) - \f{α}{2} \w^\T \w \} \d\w \
& = \l(\f{β}{2π}\r)N/2 \l(\f{α}{2π}\r)M/2
∫ exp\{ - β E_D(\w) - α E_W(\w) \} \d\w \
& = \l(\f{β}{2π}\r)N/2 \l(\f{α}{2π}\r)M/2
∫ exp\{ - E(\w) \} \d\w \tag{3.78} \
\end{align*}
ただし
\begin{align*}
E(\w) & = β E_D(\w) + α E_W(\w)
E_D(\w) & = \f{1}{2} ∑n=1^N \{t_n - \w^\T φ(\x_n)\}^2
= \f{1}{2} \| \tt - \bPhi \w \|^2 \tag{3.12} \
E_W(\w) & = \f{1}{2} \w^T \w
\end{align*}
平方完成した後の形を以下のように仮定する。
\begin{align*}
& c + \f{1}{2} (\w - \μ)^\T \A (\w - \μ)
= & c + \f{1}{2} (\w^\T \A \w - 2 \μ^\T \A \w + \μ^\T \A \μ) \
\end{align*}
(3.79)より
\begin{align*}
E(\w) & = \f{β}{2} \| \tt - \bPhi \w \|^2 + \f{α}{2} \w^\T \w
& = \f{β}{2} (\tt - \bPhi \w)^\T (\tt - \bPhi \w) + \f{α}{2} \w^\T \w \
& = \f{β}{2} \{ \tt^\T \tt - 2 \tt^\T \bPhi \w + (\bPhi \w)^\T \bPhi \w \}
- \f{α}{2} \w^\T \w
\end{align*}
\(\w\)の2次の項
\begin{align*}
\f{1}{2} \w^\T \A \w = & \f{β}{2} (\bPhi \w)^\T \bPhi \w + \f{α}{2} \w^\T \w
= & \f{β}{2} \w^\T \bPhi^\T \bPhi \w + \f{α}{2} \w^\T \w \
= & \f{1}{2} \w^\T (α \I + β \bPhi^\T \bPhi) \w \
\end{align*}
よって
\begin{align*}
\A = α \I + β \bPhi^\T \bPhi
\end{align*}
これは(3.54)の\(\S_N-1\)と等しい。
\(\w\)の1次の項
\begin{align*}
\μ^\T \A \w = & β \tt^\T \bPhi \w
\μ^\T \A = & β \tt^\T \bPhi \
\A \μ = & β \bPhi^\T \tt \
\μ = & β \A-1 \bPhi^\T \tt \
\end{align*}
これは(3.53)の\(\m_N\)と等しい。
定数項
\begin{align*}
c + \f{1}{2} \μ^\T \A \μ = & \f{β}{2} \tt^\T \tt
c = & \f{β}{2} \tt^\T \tt - \f{1}{2} \m_N^\T \A \m_N \
= & \f{β}{2} \tt^\T \tt - \m_N^\T \A \m_N + \f{1}{2} \m_N^\T \A \m_N \
= & \f{β}{2} \tt^\T \tt - \m_N^\T \A (β \A-1 \bPhi^\T \tt)
- \f{1}{2} \m_N^\T (α \I + β \bPhi^\T \bPhi) \m_N
= & \f{β}{2} \tt^\T \tt - β \m_N^\T \bPhi^\T \tt
- \f{β}{2} \m_N^\T \bPhi^\T \bPhi \m_N + \f{α}{2} \m_N^\T \m_N
= & \f{β}{2} (\tt^\T \tt - 2 \m_N^\T \bPhi^\T \tt + \m_N^\T \bPhi^\T \bPhi \m_N)
- \f{α}{2} \m_N^\T \m_N
= & \f{β}{2} \|\tt - \bPhi \m_N\|^2 + \f{α}{2} \m_N^\T \m_N
\end{align*}
\begin{align*}
& ∫ exp\{ -E(\w) \} \d\w
= & ∫ exp\{ - E(\m_N) - \f{1}{2} (\w - \m_N)^\T \A (\w - \m_N) \} \d\w \
= & ∫ exp\{ - E(\m_N) \} exp\{ - \f{1}{2} (\w - \m_N)^\T \A (\w - \m_N) \} \d\w \
= & exp\{ - E(\m_N) \} ∫ exp\{ - \f{1}{2} (\w - \m_N)^\T \A (\w - \m_N) \} \d\w \
= & exp\{ - E(\m_N) \} (2π)M/2 |\A|1/2 \tag{3.85} \
\end{align*}
積分は多次元ガウス分布の正規化条件より求まる。
\begin{align*}
∫ \N(\w|\m_N, \A) \d\w = & 1
\f{1}{(2π)M/2} \f{1}{|\A|1/2}
∫ exp\l\{ - \f{1}{2} (\w - \m_N)^\T \A (\w - \m_N) \r\} \d\w = & 1 \
∫ exp\l\{ - \f{1}{2} (\w - \m_N)^\T \A (\w - \m_N) \r\} \d\w
= & (2π)M/2 |\A|1/2 \
\end{align*}
対数周辺尤度関数
\begin{align*}
p(\tt|α,β) = & \l(\f{β}{2π}\r)N/2 \l(\f{α}{2π}\r)M/2
∫ exp\{ -E(\w) \} \d\w
\tag{3.78}
= & \l(\f{β}{2π}\r)N/2 \l(\f{α}{2π}\r)M/2
exp\{ - E(\m_N) \} (2π)M/2 |\A|1/2 \
ln p(\tt|α,β) = & \f{M}{2} ln α + \f{N}{2} ln β
- E(\m_N) - \f{1}{2} ln |\A| - \f{N}{2} ln (2π)
\tag{3.86}
\end{align*}
対数周辺尤度関数(3.86) \begin{align*} ln p(\tt|α, β) = & \f{M}{2} ln α + \f{N}{2} ln β - E(\m_N)
- \f{1}{2} ln |\A| - \f{N}{2} ln(2π)
\end{align*}
\begin{align*}
& \p{}{α} ln p(\tt|α, β)
= & \f{M}{2} \p{}{α} ln α - \p{}{α} E(\m_N) - \f{1}{2} \p{}{α} ln |\A| \
= & \f{M}{2α} - \f{1}{2} \m_N^\T \m_N - \f{1}{2} \p{}{α} ln |\A| \
\end{align*}
次の固有ベクトル方程式を考える。
\begin{align*}
(β \Φ^\T \Φ) \u_i = λ_i \u_i \
\end{align*}
ここで
\begin{align*}
\A = α \I + β \Φ^\T \Φ \
\end{align*}
より
\begin{align*}
\A \u_i = (λ_i + α) \u_i \
\end{align*}
が成り立つ。
よって
\begin{align*}
\p{}{α} ln |\A|
= \p{}{α} ln ∏_i (λ_i + α)
= \p{}{α} ∑_i ln (λ_i + α)
= ∑_i \p{}{α} ln (λ_i + α)
= ∑_i \f{1}{λ_i + α}
\end{align*}
\begin{align*}
& \p{}{α} ln p(\tt|α, β)
= & \f{M}{2α} - \f{1}{2} \m_N^\T \m_N - \f{1}{2} ∑_i \f{1}{λ_i + α} \
\end{align*}
\begin{align*}
0 = & \f{M}{2α} - \f{1}{2} \m_N^\T \m_N - \f{1}{2} ∑_i \f{1}{λ_i + α}
0 = & M - α \m_N^\T \m_N - ∑_i \f{α}{λ_i + α} \
α \m_N^\T \m_N = & M - ∑_i \f{α}{λ_i + α} \
α \m_N^\T \m_N = & ∑_i (1 - \f{α}{λ_i + α}) \
α \m_N^\T \m_N = & ∑_i \f{λ_i}{λ_i + α} \
α \m_N^\T \m_N = & γ \
α = & \f{γ}{\m_N^\T \m_N} \
\end{align*}
ここで
\begin{align*}
γ = & ∑_i \f{λ_i}{λ_i + α} \
\end{align*}