Merge pull request #25 from yuta-nakahara/develop-hiddenmarkovnormal-resume

update posterior distribution and predictive distribution

Author: yuta-nakahara

bayesml/hiddenmarkovnormal/hiddenmarkovnormal.md

@@ -9,8 +9,9 @@ The stochastic data generative model is as follows:
 * $K \in \mathbb{N}$: number of latent classes
 * $\boldsymbol{z} \in \{ 0, 1 \}^K$: a one-hot vector representing the latent class (latent variable)
 * $\boldsymbol{\pi} \in [0, 1]^K$: a parameter for latent classes, ($\sum_{k=1}^K \pi_k=1$)
-* $a_{jk}$ : transition probability to latent state k under latent state j
-* $\boldsymbol{A}=(a_{jk})_{0\leq j,k\leq K} \in [0, 1]^{K\times K}$: a parameter for latent classes, ($\sum_{k=1}^K a_{jk}=1$)
+* $a_{j,k} \in [0,1]$ : transition probability to latent state k under latent state j
+* $\boldsymbol{a}_j = [a_{j,1}, a_{j,2}, \dots , a_{j,K}]\in [0,1]^K$, a vector of the transition probability ($\sum_{k=1}^K a_{j,k}=1$)
+* $\boldsymbol{A}=(a_{j,k})_{1\leq j,k\leq K} \in [0, 1]^{K\times K}$: a matrix of the transition probability
 * $D \in \mathbb{N}$: a dimension of data
 * $\boldsymbol{x} \in \mathbb{R}^D$: a data point
 * $\boldsymbol{\mu}_k \in \mathbb{R}^D$: a parameter
@@ -22,7 +23,7 @@ The stochastic data generative model is as follows:
 $$
 \begin{align}
     p(\boldsymbol{z}_{1} | \boldsymbol{\pi}) &= \mathrm{Cat}(\boldsymbol{z}_{1}|\boldsymbol{\pi}) = \prod_{k=1}^K \pi_k^{z_{1,k}},\\
-    p(\boldsymbol{z}_{n} |\boldsymbol{z}_{n-1} ,\boldsymbol{A}) &= \prod_{k=1}^K \prod_{j=1}^K a_{jk}^{z_{n-1,j}z_{n,k}},\\
+    p(\boldsymbol{z}_{n} |\boldsymbol{z}_{n-1} ,\boldsymbol{A}) &= \prod_{k=1}^K \prod_{j=1}^K a_{j,k}^{z_{n-1,j}z_{n,k}},\\
     p(\boldsymbol{x}_{n} | \boldsymbol{\mu}, \boldsymbol{\Lambda}, \boldsymbol{z}_{n}) &= \prod_{k=1}^K \mathcal{N}(\boldsymbol{x}|\boldsymbol{\mu}_k,\boldsymbol{\Lambda}_k^{-1})^{z_{n,k}} \\
     &= \prod_{k=1}^K \left( \frac{| \boldsymbol{\Lambda}_{k} |^{1/2}}{(2\pi)^{D/2}} \exp \left\{ -\frac{1}{2}(\boldsymbol{x}-\boldsymbol{\mu}_{k})^\top \boldsymbol{\Lambda}_{k} (\boldsymbol{x}-\boldsymbol{\mu}_{k}) \right\} \right)^{z_{n,k}},
 \end{align}
@@ -44,7 +45,8 @@ $$
     p(\boldsymbol{\mu},\boldsymbol{\Lambda},\boldsymbol{\pi},\boldsymbol{A}) &= \left\{ \prod_{k=1}^K \mathcal{N}(\boldsymbol{\mu}_k|\boldsymbol{m}_0,(\kappa_0 \boldsymbol{\Lambda}_k)^{-1})\mathcal{W}(\boldsymbol{\Lambda}_k|\boldsymbol{W}_0, \nu_0) \right\} \mathrm{Dir}(\boldsymbol{\pi}|\boldsymbol{\eta}_0) \prod_{j=1}^{K}\mathrm{Dir}(\boldsymbol{a}_{j}|\boldsymbol{\zeta}_{0,j}), \\
     &= \Biggl[ \prod_{k=1}^K \left( \frac{\kappa_0}{2\pi} \right)^{D/2} |\boldsymbol{\Lambda}_k|^{1/2} \exp \left\{ -\frac{\kappa_0}{2}(\boldsymbol{\mu}_k -\boldsymbol{m}_0)^\top \boldsymbol{\Lambda}_k (\boldsymbol{\mu}_k - \boldsymbol{m}_0) \right\} \\
     &\qquad \times B(\boldsymbol{W}_0, \nu_0) | \boldsymbol{\Lambda}_k |^{(\nu_0 - D - 1) / 2} \exp \left\{ -\frac{1}{2} \mathrm{Tr} \{ \boldsymbol{W}_0^{-1} \boldsymbol{\Lambda}_k \} \right\}\biggl] \\
-    &\qquad \times \Biggl[ \prod_{k=1}^KC(\boldsymbol{\eta}_0)\pi_k^{\eta_{0,k}-1}\biggl]\times \biggl[\prod_{j=1}^KC(\boldsymbol{\zeta}_{0,j})\prod_{k=1}^K a_{jk}^{\zeta_{0,j,k}-1}\Biggr],\\
+    &\qquad \times \Biggl[ \prod_{k=1}^KC(\boldsymbol{\eta}_0)\pi_k^{\eta_{0,k}-1}\biggl]\\
+    &\qquad \times \biggl[\prod_{j=1}^KC(\boldsymbol{\zeta}_{0,j})\prod_{k=1}^K a_{j,k}^{\zeta_{0,j,k}-1}\Biggr],\\
 \end{align}
 $$
 
@@ -57,68 +59,94 @@ $$
     C(\boldsymbol{\zeta}_{0,j}) &= \frac{\Gamma(\sum_{k=1}^K \zeta_{0,j,k})}{\Gamma(\zeta_{0,j,1})\cdots\Gamma(\zeta_{0,j,K})}. 
 \end{align}
 $$
-<!--
+
 The apporoximate posterior distribution in the $t$-th iteration of a variational Bayesian method is as follows:
 
 * $\boldsymbol{x}^n = (\boldsymbol{x}_1, \boldsymbol{x}_2, \dots , \boldsymbol{x}_n) \in \mathbb{R}^{D \times n}$: given data
 * $\boldsymbol{z}^n = (\boldsymbol{z}_1, \boldsymbol{z}_2, \dots , \boldsymbol{z}_n) \in \{ 0, 1 \}^{K \times n}$: latent classes of given data
-* $\boldsymbol{r}_i^{(t)} = (r_{i,1}^{(t)}, r_{i,2}^{(t)}, \dots , r_{i,K}^{(t)}) \in [0, 1]^K$: a parameter for the $i$-th latent class. ($\sum_{k=1}^K r_{i, k}^{(t)} = 1$)
 * $\boldsymbol{m}_{n,k}^{(t)} \in \mathbb{R}^{D}$: a hyperparameter
 * $\kappa_{n,k}^{(t)} \in \mathbb{R}_{>0}$: a hyperparameter
 * $\nu_{n,k}^{(t)} \in \mathbb{R}$: a hyperparameter $(\nu_n > D-1)$
 * $\boldsymbol{W}_{n,k}^{(t)} \in \mathbb{R}^{D\times D}$: a hyperparameter (a positive definite matrix)
 * $\boldsymbol{\eta}_n^{(t)} \in \mathbb{R}_{> 0}^K$: a hyperparameter
-
+* $\boldsymbol{\zeta}_{n,j}^{(t)} \in \mathbb{R}_{> 0}^K$: a hyperparameter
 $$
 \begin{align}
-    q(\boldsymbol{z}^n, \boldsymbol{\mu},\boldsymbol{\Lambda},\boldsymbol{\pi}) &= \left\{ \prod_{i=1}^n \mathrm{Cat} (\boldsymbol{z}_i | \boldsymbol{r}_i^{(t)}) \right\} \left\{ \prod_{k=1}^K \mathcal{N}(\boldsymbol{\mu}_k|\boldsymbol{m}_{n,k}^{(t)},(\kappa_{n,k}^{(t)} \boldsymbol{\Lambda}_k)^{-1})\mathcal{W}(\boldsymbol{\Lambda}_k|\boldsymbol{W}_{n,k}^{(t)}, \nu_{n,k}^{(t)}) \right\} \mathrm{Dir}(\boldsymbol{\pi}|\boldsymbol{\eta}_n^{(t)}) \\
-    &= \Biggl[ \prod_{i=1}^n \prod_{k=1}^K (r_{i,k}^{(t)})^{z_{i,k}} \Biggr] \Biggl[ \prod_{k=1}^K \left( \frac{\kappa_{n,k}^{(t)}}{2\pi} \right)^{D/2} |\boldsymbol{\Lambda}_k|^{1/2} \exp \left\{ -\frac{\kappa_{n,k}^{(t)}}{2}(\boldsymbol{\mu}_k -\boldsymbol{m}_{n,k}^{(t)})^\top \boldsymbol{\Lambda}_k (\boldsymbol{\mu}_k - \boldsymbol{m}_{n,k}^{(t)}) \right\} \\
+    &q(\boldsymbol{z}^n, \boldsymbol{\mu},\boldsymbol{\Lambda},\boldsymbol{\pi},\boldsymbol{A}) \nonumber \\
+     &= q^{(t)}(\boldsymbol{z}^n) \left\{ \prod_{k=1}^K \mathcal{N}(\boldsymbol{\mu}_k|\boldsymbol{m}_{n,k}^{(t)},(\kappa_{n,k}^{(t)} \boldsymbol{\Lambda}_k)^{-1})\mathcal{W}(\boldsymbol{\Lambda}_k|\boldsymbol{W}_{n,k}^{(t)}, \nu_{n,k}^{(t)}) \right\} \mathrm{Dir}(\boldsymbol{\pi}|\boldsymbol{\eta}_n^{(t)})\left\{\prod_{j=1}^K\mathrm{Dir}(\boldsymbol{a}_j|\boldsymbol{\zeta}_{n,j}^{(t)})\right\}, \\
+    &= q^{(t)}(\boldsymbol{z}^n) \Biggl[ \prod_{k=1}^K \left( \frac{\kappa_{n,k}^{(t)}}{2\pi} \right)^{D/2} |\boldsymbol{\Lambda}_k|^{1/2} \exp \left\{ -\frac{\kappa_{n,k}^{(t)}}{2}(\boldsymbol{\mu}_k -\boldsymbol{m}_{n,k}^{(t)})^\top \boldsymbol{\Lambda}_k (\boldsymbol{\mu}_k - \boldsymbol{m}_{n,k}^{(t)}) \right\} \\
     &\qquad \times B(\boldsymbol{W}_{n,k}^{(t)}, \nu_{n,k}^{(t)}) | \boldsymbol{\Lambda}_k |^{(\nu_{n,k}^{(t)} - D - 1) / 2} \exp \left\{ -\frac{1}{2} \mathrm{Tr} \{ ( \boldsymbol{W}_{n,k}^{(t)} )^{-1} \boldsymbol{\Lambda}_k \} \right\} \Biggr] \\
-    &\qquad \times C(\boldsymbol{\eta}_n^{(t)})\prod_{k=1}^K \pi_k^{\eta_{n,k}^{(t)}-1},\\
+    &\qquad \times C(\boldsymbol{\eta}_n^{(t)})\prod_{k=1}^K \pi_k^{\eta_{n,k}^{(t)}-1}\left[\prod_{j=1}^K C(\boldsymbol{\zeta}_{n,j}^{(t)})\prod_{k=1}^K a_{j,k}^{\zeta_{n,j,k}^{(t)}-1}\right],\\
 \end{align}
 $$
 
 where the updating rule of the hyperparameters is as follows.
 
 $$
 \begin{align}
-    N_k^{(t)} &= \sum_{i=1}^n r_{i,k}^{(t)} \\
-    \bar{\boldsymbol{x}}_k^{(t)} &= \frac{1}{N_k^{(t)}} \sum_{i=1}^n r_{i,k}^{(t)} \boldsymbol{x}_i \\
+    N_k^{(t)} &= \sum_{i=1}^n \gamma^{(t)}_{i,k}, \\
+    M_{j,k}^{(t)} &= \sum_{i=2}^n \xi^{(t)}_{i,j,k},\\
+    \bar{\boldsymbol{x}}_k^{(t)} &= \frac{1}{N_k^{(t)}} \sum_{i=1}^n \gamma^{(t)}_{i,k} \boldsymbol{x}_i, \\
+    S_k^{(t)} &= \frac{1}{N_k^{(t)}}\sum_{i=1}^n \gamma^{(t)}_{i,k} (x_i-\bar{\boldsymbol{x}}_k^{(t)})(x_i-\bar{\boldsymbol{x}}_k^{(t)})^{\top},\\
     \boldsymbol{m}_{n,k}^{(t+1)} &= \frac{\kappa_0\boldsymbol{\mu}_0 + N_k^{(t)} \bar{\boldsymbol{x}}_k^{(t)}}{\kappa_0 + N_k^{(t)}}, \\
     \kappa_{n,k}^{(t+1)} &= \kappa_0 + N_k^{(t)}, \\
-    (\boldsymbol{W}_{n,k}^{(t+1)})^{-1} &= \boldsymbol{W}_0^{-1} + \sum_{i=1}^{n} r_{i,k}^{(t)} (\boldsymbol{x}_i-\bar{\boldsymbol{x}}_k^{(t)})(\boldsymbol{x}_i-\bar{\boldsymbol{x}}_k^{(t)})^\top + \frac{\kappa_0 N_k^{(t)}}{\kappa_0 + N_k^{(t)}}(\bar{\boldsymbol{x}}_k^{(t)}-\boldsymbol{\mu}_0)(\bar{\boldsymbol{x}}_k^{(t)}-\boldsymbol{\mu}_0)^\top, \\
+    (\boldsymbol{W}_{n,k}^{(t+1)})^{-1} &= \boldsymbol{W}_0^{-1} + N_k^{(t)}S_k^{(t)} + \frac{\kappa_0 N_k^{(t)}}{\kappa_0 + N_k^{(t)}}(\bar{\boldsymbol{x}}_k^{(t)}-\boldsymbol{\mu}_0)(\bar{\boldsymbol{x}}_k^{(t)}-\boldsymbol{\mu}_0)^\top, \\
     \nu_{n,k}^{(t+1)} &= \nu_0 + N_k^{(t)},\\
-    \eta_{n,k}^{(t+1)} &= \eta_{0,k} + N_k^{(t)} \\
-    \ln \rho_{i,k}^{(t+1)} &= \psi (\eta_{n,k}^{(t+1)}) - \psi ( {\textstyle \sum_{k=1}^K \eta_{n,k}^{(t+1)}} ) \nonumber \\
-    &\qquad + \frac{1}{2} \Biggl[ \sum_{d=1}^D \psi \left( \frac{\nu_{n,k}^{(t+1)} + 1 - d}{2} \right) + D \ln 2 + \ln | \boldsymbol{W}_{n,k}^{(t+1)} | \nonumber \\
-    &\qquad - D \ln (2 \pi ) - \frac{D}{\kappa_{n,k}^{(t+1)}} - \nu_{n,k}^{(t+1)} (\boldsymbol{x}_i - \boldsymbol{m}_{n,k}^{(t+1)})^\top \boldsymbol{W}_{n,k}^{(t+1)} (\boldsymbol{x}_i - \boldsymbol{m}_{n,k}^{(t+1)}) \Biggr] \\
-    r_{i,k}^{(t+1)} &= \frac{\rho_{i,k}^{(t+1)}}{\sum_{k=1}^K \rho_{i,k}^{(t+1)}}
+    \eta_{n,k}^{(t+1)} &= \eta_{0,k} + \gamma^{(t)}_{1,k}, \\
+    \zeta_{n,j,k}^{(t+1)} &= \zeta_{0,j,k}+M_{j,k}^{(t)}.
+\end{align}
+$$
+
+The approximate posterior distribution of the latent variable $q^{(t+1)}(z^n)$ is calculated by the forward-backward algorithm as follows.
+
+$$
+\begin{align}
+    \ln \rho_{i,k}^{(t+1)} &= \frac{1}{2} \Biggl[\, \sum_{d=1}^D \psi \left( \frac{\nu_{n,k}^{(t+1)} + 1 - d}{2} \right) + D \ln 2 + \ln | \boldsymbol{W}_{n,k}^{(t+1)} | \notag \\
+    &\qquad - D \ln (2 \pi ) - \frac{D}{\kappa_{n,k}^{(t+1)}} - \nu_{n,k}^{(t+1)} (\boldsymbol{x}_i - \boldsymbol{m}_{n,k}^{(t+1)})^\top \boldsymbol{W}_{n,k}^{(t+1)} (\boldsymbol{x}_i - \boldsymbol{m}_{n,k}^{(t+1)}) \Biggr], \\
+    \ln \tilde{\pi}_k^{(t+1)} &= \psi (\eta_{n,k}^{(t+1)}) - \psi \left( \textstyle \sum_{k=1}^K \eta_{n,k}^{(t+1)} \right) \\
+    \ln \tilde{a}_{j,k}^{(t+1)} &= \psi (\zeta_{n,j,k}^{(t+1)}) - \psi \left( \textstyle \sum_{k=1}^K \zeta_{n,j,k}^{(t+1)} \right) \\
+    \alpha^{(t+1)} (\boldsymbol{z}_i) &\propto
+    \begin{cases}
+    \prod_{k=1}^{K} \left( \rho_{i,k}^{(t+1)}\right)^{z_{i,k}} \sum_{\boldsymbol{z}_{i-1}} \left[\prod_{k=1}^{K}\prod_{j=1}^{K}\left(\tilde{a}^{(t+1)}_{j,k}\right)^{z_{i-1,j}z_{i,k}}\alpha^{(t+1)}(\boldsymbol{z}_{i-1})\right] & (i>1)\\
+    \prod_{k=1}^{K}\left( \rho_{1,k}^{(t+1)} \tilde{\pi}_k^{(t+1)} \right)^{z_{1,k}} & (i=1)
+    \end{cases} \\
+    \beta^{(t+1)} (\boldsymbol{z}_i) &\propto
+    \begin{cases}
+    \sum_{\boldsymbol{z}_{i+1}} \left[ \prod_{k=1}^{K} \left( \rho_{i+1,k}^{(t+1)}\right)^{z_{i,k}} \prod_{k=1}^{K}\prod_{j=1}^{K}\left(\tilde{a}^{(t+1)}_{j,k}\right)^{z_{i,j}z_{i+1,k}}\beta^{(t+1)}(\boldsymbol{z}_{i+1})\right] & (i<n)\\
+    1 & (i=n)
+    \end{cases} \\
+    q^{(t+1)}(\boldsymbol{z}_i) &\propto \alpha^{(t+1)}(\boldsymbol{z}_i)\beta^{(t+1)}(\boldsymbol{z}_i) \\
+    \gamma^{(t+1)}_{i,k} &= \sum_{\boldsymbol{z}_i} q^{(t+1)}(\boldsymbol{z}_i) z_{i,k}\\
+    q^{(t+1)}(\boldsymbol{z}_{i-1}, \boldsymbol{z}_{i}) &\propto \alpha^{(t+1)}(\boldsymbol{z}_{i-1}) \prod_{k=1}^{K} \left( \rho_{i,k}^{(t+1)}\right)^{z_{i,k}} \prod_{k=1}^{K}\prod_{j=1}^{K}\left(\tilde{a}^{(t+1)}_{j,k}\right)^{z_{i-1,j}z_{i,k}} \beta^{(t+1)}(\boldsymbol{z}_i) \\
+    \xi^{(t+1)}_{i,j,k} &= \sum_{\boldsymbol{z}_{i-1}} \sum_{\boldsymbol{z}_i} q^{(t+1)}(\boldsymbol{z}_{i-1}, \boldsymbol{z}_{i}) z_{i-1,j} z_{i,k}
 \end{align}
 $$
 
 The approximate predictive distribution is as follows:
 
 * $\boldsymbol{x}_{n+1} \in \mathbb{R}^D$: a new data point
+* $(a_{\mathrm{p},j,k})_{1\leq j,k\leq K} \in [0, 1]^{K\times K}$: the parameters of the predictive transition probability of latent classes, ($\sum_{k=1}^K a_{\mathrm{p},j,k}=1$)
 * $\boldsymbol{\mu}_{\mathrm{p},k} \in \mathbb{R}^D$: the parameter of the predictive distribution
 * $\boldsymbol{\Lambda}_{\mathrm{p},k} \in \mathbb{R}^{D \times D}$: the parameter of the predictive distribution (a positive definite matrix)
 * $\nu_{\mathrm{p},k} \in \mathbb{R}_{>0}$: the parameter of the predictive distribution
 
 $$
 \begin{align}
     &p(x_{n+1}|x^n) \\
-    &= \frac{1}{\sum_{k=1}^K \eta_{n,k}^{(t)}} \sum_{k=1}^K \eta_{n,k}^{(t)} \mathrm{St}(x_{n+1}|\boldsymbol{\mu}_{\mathrm{p},k},\boldsymbol{\Lambda}_{\mathrm{p},k}, \nu_{\mathrm{p},k}) \\
-    &= \frac{1}{\sum_{k=1}^K \eta_{n,k}^{(t)}} \sum_{k=1}^K \eta_{n,k}^{(t)} \Biggl[ \frac{\Gamma (\nu_{\mathrm{p},k} / 2 + D / 2)}{\Gamma (\nu_{\mathrm{p},k} / 2)} \frac{|\boldsymbol{\Lambda}_{\mathrm{p},k}|^{1/2}}{(\nu_{\mathrm{p},k} \pi)^{D/2}} \nonumber \\
+    &\approx \sum_{k=1}^K \left( \sum_{j=1}^K \gamma_{n,j}^{(t)} a_{\mathrm{p},j,k} \right) \mathrm{St}(x_{n+1}|\boldsymbol{\mu}_{\mathrm{p},k},\boldsymbol{\Lambda}_{\mathrm{p},k}, \nu_{\mathrm{p},k}) \\
+    &= \sum_{k=1}^K \left( \sum_{j=1}^K \gamma_{n,j}^{(t)} a_{\mathrm{p},j,k} \right)\Biggl[ \frac{\Gamma (\nu_{\mathrm{p},k} / 2 + D / 2)}{\Gamma (\nu_{\mathrm{p},k} / 2)} \frac{|\boldsymbol{\Lambda}_{\mathrm{p},k}|^{1/2}}{(\nu_{\mathrm{p},k} \pi)^{D/2}} \nonumber \\
     &\qquad \qquad \qquad \qquad \qquad \times \left( 1 + \frac{1}{\nu_{\mathrm{p},k}} (\boldsymbol{x}_{n+1} - \boldsymbol{\mu}_{\mathrm{p},k})^\top \boldsymbol{\Lambda}_{\mathrm{p},k} (\boldsymbol{x}_{n+1} - \boldsymbol{\mu}_{\mathrm{p},k}) \right)^{-\nu_{\mathrm{p},k}/2 - D/2} \Biggr],
 \end{align}
 $$
 
-where the parameters are obtained from the hyperparameters of the posterior distribution as follows:
+where the parameters are obtained from the hyperparameters of the predictive distribution as follows:
 
 $$
 \begin{align}
-    \boldsymbol{\mu}_{\mathrm{p},k} &= \boldsymbol{m}_{n,k}^{(t)} \\
+    a_{\mathrm{p},j,k} &= \frac{\zeta_{n,j,k}^{(t)}}{\sum_{k=1}^K \zeta_{n,j,k}^{(t)}}, \\
+    \boldsymbol{\mu}_{\mathrm{p},k} &= \boldsymbol{m}_{n,k}^{(t)}, \\
     \boldsymbol{\Lambda}_{\mathrm{p},k} &= \frac{\kappa_{n,k}^{(t)} (\nu_{n,k}^{(t)} - D + 1)}{\kappa_{n,k}^{(t)} + 1} \boldsymbol{W}_{n,k}^{(t)}, \\
     \nu_{\mathrm{p},k} &= \nu_{n,k}^{(t)} - D + 1.
 \end{align}
 $$
--->
+