Suppose the time-series $X$ is hidden state Gaussian random walk and we observe $Y = X + e$, where $e$ is gaussian white noise independent of $X$.

The Kalman estimator of $X$ in this case has a steady state closed form solution and corresponds to an exponential moving average smoother with constant smoothing parameter. The optimal smoothing parameter looks like $\lambda = \frac{p}{1-p}$ where $p$ is a quadratic formula of signal to noise ratio between $e$ and $\Delta{x}$. See formula for closed form (by searching 'kalman solu‌​tion random walk noise').

If we have instead multiple independent observations $Y_1, Y_2,...$, i.e. independent $e_1, e_2,...$ is there a closed form solution for the optimal estimator of $X$? What would it look like?

If the $e_1, e_2$ were independent and identically distributed (same standard deviation), I can imagine the best estimator for $X$ is to simply average the estimators obtained treating the problem as univariate for each $Y$.

If the $e_1,e_2,...$ have different variance, then some $Y$ series should have low weightings in the overall estimator as their signal to noise ratio is poorer. Perhaps the closed form solution is linear with coefficients proportional to each $Y$ signal to noise ratio.

Is there a known closed form solution? Google seems not to be too helpful on this problem.

  • $\begingroup$ formula for closed form: books.google.com/books?id=Kc6tnRHBwLcC&lpg=PA175&ots=I2VPQwXUGA&dq=kalman%20solution%20random%20walk%20noise&pg=PA175#v=onepage&q=kalman%20solution%20random%20walk%20noise&f=false $\endgroup$ – Matthew Pollard Feb 19 '13 at 6:55
  • $\begingroup$ First of all the constant smoothing parameter is achieved when the filter converges to the steady state. So in that respect exponential smoothing and kalman differ. Secondly I think the question requires clarification in the sense that when you mean Y1 and Y2 are they different time indicies or position in the vector Y for the same time? $\endgroup$ – Cagdas Ozgenc Nov 14 '13 at 12:51
  • $\begingroup$ Your observation noises $e_i$ are independent, but are your random walk $X$ increments independent? If they are, then you just have a bunch of 1-d independent filters and closed-form formula for 1-d case applies to each. $\endgroup$ – Kochede Dec 18 '13 at 2:04

You haven't specified but if the observation noise is independent of the model noise then I think you should first combine the Y estimates with simple observation variance weighted mean as specified here:


after that apply kalman filter as if it is a univariate model. As a result you will have a lower observation noise hence a better estimate of x.

| cite | improve this answer | |

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.