Joint distribution of hidden markov model

1.5k Views Asked by At

Does anyone know how the joint distribution of the HMM in page 281 of these notes is derived?

Specifically the formula is: $$p(\textbf{y},\textbf{x})=\prod_{t=1}^{T}p(y_t|y_{t-1})p(x_t|y_t)$$

Where $t$ is the $t^{th}$ sequence, $\textbf{x}$ is the observation sequence and $\textbf{y}$ is the state sequence.

I know that: $$f_{X,Y}(x,y)=f_{Y|X}(y|x)f_{X}(x)$$

So I'm not sure why both $p$'s are conditional in the HMM formula. Can someone explain this or link me to an explanation?

1

There are 1 best solutions below

0
On

With the understanding that $p(y_1\mid y_0):=p(y_1)$, what they appear to be saying is:

$\begin{align}p(\mathbf y,\mathbf x) ~=~& p(\mathbf y)~p(\mathbf x\mid \mathbf y) \\[1ex] ~=~& p(y_1,..,y_T)~p(\mathbf x\mid \mathbf y) \\[1ex] ~=~& \big(~p(y_1)~p(y_2\mid y_1)\cdots p(y_T\mid y_{T-1})~\big)~ p(\mathbf x\mid \mathbf y) \tag 1 \\[1ex] ~=~& \Big(\prod_{t=1}^T p(y_t\mid y_{t-1})\Big)~ p(x_1,..,x_T\mid \mathbf y) \\[1ex] ~=~& \Big(\prod_{t=1}^T p(y_t\mid y_{t-1})\Big)~ \big(p(x_1\mid \mathbf y)\cdots p(x_T\mid \mathbf y)\big) \tag 2 \\[1ex] ~=~& \Big(\prod_{t=1}^T p(y_t\mid y_{t-1})\Big)~ \big(p(x_1\mid y_1)\cdots p(x_T\mid y_T)\big) \tag 3 \\[1ex] ~=~& \Big(\prod_{t=1}^T p(y_t\mid y_{t-1})\Big)~\Big(\prod_{t=1}^T p(x_t\mid y_t) \Big) \\[1ex] ~=~& \Big(\prod_{t=1}^T p(y_t\mid y_{t-1})~p(x_t\mid y_t) \Big) \end{align}$

$(1)$ is true only when $p(y_t\mid y_{t-1},..,y_1)=p(y_t\mid y_{t-1})$ , every member of $\mathbf y$ depends only on the value of the immediately preceding member rather than the preceding list.

$(2)$ is true only when the members of $\mathbf x$ are mutually independent on condition of any $\mathbf y$

$(3)$ is true only when $p(x_t\mid y_1, .. , y_T) = p(x_t\mid y_t)$ , that is every member of $\mathbf x$ depends only on the corresponding member of $\mathbf y$ rather than the entire list.   (Which would also make (2) okay).

$\Box$

These do appear to match the two assumption of independence made on page 281, right in the paragraph preceding equation (210).

$\blacksquare$