I am learning about Kalman filters/dynamic linear models/state-space models and I am interested in whether the following scheme is possible, in which I try to estimate distribution parameters simultaneously with a sequence.

Suppose that there is some sequence $\{x_0,x_1,\dotsc,x_t,\dotsc\}$ which we would like to estimate (in a sense I will make precise asap). All we are given is the observations $Y_{\leq t} = \{y_0,y_1,\dotsc,y_t\}$, which include some error, so $y_t = x_t + v_t$ and $v_t\sim[0,V_t]$. Let's write the prediction error as $\varepsilon_t = \mathbb E(y_t|Y_{\leq t-1})-y_t$. I would like to write an algorithm that minimises the RMS of prediction errors, that is, minimises $\lVert\{ \varepsilon_1,\varepsilon_2,\dotsc,\varepsilon_t,\dotsc\}\rVert_2$.

I do not want to specify any parameters. I do not know if $\{x_t\}$ is constant, linear, has 'jumps' or anything; let's only write $x_{t+1}\sim [x_t + \mu_t,W_t]$. I am willing to make some quality of life assumptions (such as the signal error $v_t$ having $0$ mean), but would e.g. prefer not to assume that $V_t,\mu_t,W_t$ is a constant or known value. Hence I need to adapt my Kalman gain as I go along.

Finally, I want to output $\mathbb E(y_{t+1}|Y_{\leq t})$. Of course, for this I must first find $\mathbb E(x_t|Y_{\leq t})$, using a Kalman gain/adaption coefficient $K_t$ to determine how much weight to give to my previous estimate versus the new observation: $$\mathbb E(x_t|Y_{\leq t}) = K_t\mathbb E(x_t|Y_{\leq t-1}) + (1-K_t)y_t.$$

The parameter $K_t$ is usually the signal variance divided by the total variance. I want to update my estimate of signal variance $V_t$, and total variance $V_t + W_t$, with each new piece of data. Of course, to update the $V_t,W_t$-estimates I also need to use a Kalman gain.

This is a chicken-and-egg problem. If $ K_t = \frac{ \mathbb E(V_t|Y_{\leq t}) }{ \mathbb E(V_t|Y_{\leq t}) + \mathbb E(W_t|Y_{\leq t}) }$, but $\mathbb E(V_t|Y_{\leq t}) = K_t\mathbb E(V_t|Y_{\leq t-1}) + (1-K_t)(y_t - \mathbb E(x_t|Y_{\leq t}))^2$ (and similarly $\mathbb E(W_t|Y_{\leq t})$), how can I resolve this issue ? Is there something like a prior Kalman gain + a posterior Kalman gain, or some Bayesian tool, that can be brought to bear ? I appreciate any pointers to the literature.


Your Answer

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

Browse other questions tagged or ask your own question.