1
$\begingroup$

$y$ is scalar observations and so C will be a 1x2 matrix.

I want to represent the following model as a state space representation so as to estimate the hidden states from the noisy observations $y$ using Kalman filter.

The state space model :

x(t+1) = Ax(t) + w(t)

y(t) = Cx(t) + v(t)

w(t) = N(0,Q)

v(t) = N(0,R)

$w(t)$ is a pseudo-random binary signal that excites/ drives ; and $v(t) = N(0,\sigma^2_v)$ is the measurement noise.

  1. The model is an FIR (MA) filter

$$x(t) = h_1 \epsilon(t-1) + h_2 \epsilon(t-2) + \epsilon(t)$$ $$y(t) = x(t) + v(t)$$ $$ y(t) = h^T \epsilon(t) + v(t)$$

(In vector form)

where $\epsilon(t) = w(t)$.

  1. The other model is an IIR (AR) filter $$x(t) = ax(t-1) + bx(t-2)+ w(t)$$

The state space representation:

$$x(t+1) = a^Tx(t) + w(t)$$

$$y(t) = h^Tx(t) + v(t)$$

How do I represent these as state space so as to apply Kalman Filter?

There are several ways to represent time series models. This is how I proceeded, but unsure because the output of the log-likelihood is a matrix of 2 by 2 with off diagonal elements being infinity and the diagonal elements are same positive values. So, the dimension and the value of log-likelihood is incorrect, I should get negative instead of positive values.

  1. FIR :

Re-writing the above model as:

$$x(t+1) = h_1 \epsilon(t) + h_2 \epsilon(t-1) + \epsilon(t+1)$$ $$y(t) = Cx(t)+v(t)$$

State Space :

$ \left[ \begin{array}{c} x(t+1) \\ x(t)\\ x(t-1) \end{array} \right] $ = $ \left[ \begin{array}{ccc} 1 & h_1 & h_2 \\ 0 & 1 & h_1 \\ 0 & 0 & 1\end{array} \right] $ $\times$ $ \left[ \begin{array}{c} e(t+1)\\ e(t) \\ e(t-1)\end{array} \right] $

$ y(t)$ = $\left[ \begin{array}{c} 1 \hskip 5 pt 0 \hskip 5 pt 0\end{array} \right] $ $\times$ $ \left[ \begin{array}{c} e(t+1)\\ e(t) \\ e(t-1)\end{array} \right] $ + $v(t)$

  1. IIR (AR model)

    $ \left[ \begin{array}{c} x(t+1) \\ x(t) \end{array} \right] $ = $ \left[ \begin{array}{cc} a & b \\ 1 & 0 \end{array} \right] $ $\times$ $ \left[ \begin{array}{c} x(t)\\ x(t-1)\end{array} \right] $ + $\left[ \begin{array}{c} 1\\ 0\end{array} \right] $ $\times$ $\left[ \begin{array}{c} w(t+1)\\ w(t)\end{array} \right] $

    $ y(t)$ = $\left[ \begin{array}{cc} 1 \hskip 5 pt 0\end{array} \right]$ $\times$ $ \left[ \begin{array}{c} x(t)\\ x(t-1)\end{array} \right] $ + $v(t)$

$\endgroup$
  • 1
    $\begingroup$ no one is worth for downvoting $\endgroup$ – dato datuashvili Mar 16 '15 at 18:05
1
$\begingroup$

Your FIR state space representation seems to be doing too much.

The way I would write it is to have the process noise is $\epsilon(t)$ as your input, and your state as two time-delayed copies of it:

$$x(t+1) = \left[ \begin{array}{c} \epsilon(t+1)\\ \epsilon(t) \\ \epsilon(t-1) \end{array} \right] = \left[ \begin{array}{ccc} 0 & 0 & 0\\ 1 & 0 & 0\\ 0 & 1 & 0 \end{array} \right] x(t) + \left[\begin{array}{c} 1\\ 0\\ 0 \end{array} \right] \epsilon(t + 1) $$

then your output equation is just: $$ y(t) = \left[ \begin{array}{ccc} 1 & h_1 & h_2 \end{array} \right] x(t) + v(t) $$

Your IIR representation is too confused for me to make head or tail out of. Can you clarify that you have the right information there?

Other questions:

$\endgroup$
  • $\begingroup$ Thank you for your answer. The IIR has a typo, I meant $u(t)$ and not $w(t)$ where $u(t)$ and $\epsilon(t)$ are both PRBS. A nonlinear signal is quantized using the mean of the nonlinear signal. The sequence of 0/1 is then represented using a mapping equation which converts the 0/1 to real numbers. This becomes the excitation input that drives both the models. I need to estimate the transition matrix for this input as well. I have a question regarding the MA model: if number of lags =2 (as in the Question) then why the state space is becoming three dimensional and not 2 dimensional? $\endgroup$ – SKM Mar 16 '15 at 21:28
  • $\begingroup$ i won't edit it, Peter, but your state-space representation is actually a discrete-time representation, but appears (by use of "$(\cdot)$" and "$t$") to be continuous time. $\endgroup$ – robert bristow-johnson Mar 17 '15 at 18:57
  • $\begingroup$ @robertbristow-johnson: While it's possible to have an FIR continuous-time system, I believe the OP is actually asking about a discrete-time FIR system. I think you're mistaken. $\endgroup$ – Peter K. Mar 18 '15 at 2:13
  • $\begingroup$ @SKM: Good question regarding lags vs order. The state space representation that I've chosen is degenerate, so can probably be simplified to a two-state version. I've left it as it is, though, because the tapped delay line representation is (I think) clearer. $\endgroup$ – Peter K. Mar 18 '15 at 2:43
  • $\begingroup$ @PeterK.: YOu are correct in mentioning that I am asking for discrete system (Will add that tag), but the answer is not clear to me. On expanding by multiplying, I am not getting the $h$ coefficients in the state representation. Also, $x(t)$ on RHS of the first equation (State Eq) $x(t)$ is a scalar but how do I model that? $x(t)$ is present as the output and not in the input of the main process representation. $\endgroup$ – SKM Mar 18 '15 at 18:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.