My understanding is that the Kalman gain for a regular multivariate filter is computed as

$$ K = \bar P H^T S^{-1} \\ S = H \bar P H^T + R $$

$K$ being the Kalman gain, $\bar P$ being the prior state covariance, $H$ being the measurement function, $S$ being the innovation covariance, and $R$ being the sensor noise.

The state is updated with this equation:

$$ x = \bar x + Ky $$

with $x$ being the posterior state, $\bar x$ being the prior state, and $y$ being the residual.

In the text I am reading, the $H^T$ term in $K$ is explained as being responsible for transforming the value of y into state space, which leaves the $\bar P S^{-1}$ term responsible for scaling the residual as it is added to $\bar x$.

The part that confuses me is that $\bar P$ is obviously in state space, as it is the state covariance. However, $S$ is in measurement space, as it is clearly the sum of $\bar P$ projected into measurement space plus $R$, which is also in measurement space. If state space and measurement space happen to be in different units that have completely different scales (maybe a voltage measurement is being used to estimate a state variable of temperature), I feel like the ratio would not make sense, or unfairly weight the units with a smaller scale.

Is my thinking here correct, or do the current set of equations I have not account for a possible difference in units between state and measurement space?


1 Answer 1


Just to clarify one point to begin -- the components of the state and measurement vectors may themselves be in different units -- there's not a single unit for either the state nor the measurement (observation). Indeed even the dimension of either one could change over time.

Note that:

  • $\bar{P}$ is not in "units of the state". The units of its components are a product of units from the state (e.g. its $i$th diagonal element is in squared units of the $i$th state component, and the $i,j$ element is in $i$th-component-units by $j$th-component-units).

  • Similarly, $S$ is not "in measurement space", but in products of units; (e.g. its $i$th diagonal element is in squared units of the $i$th observation-component, and the $i,j$ element is in $i$th-component-units by $j$th-component-units)

  • $K$ will be a matrix of terms in ratios of units; the $i,j$ element will be in units that are the ratio of units of the $i$th component of state to the $j$th component of observation.

  • $H$ goes the other direction, converting state to observation (the ratios are the other way up)

  • There's no "$\bar{P}S^{-1}$" term. In general they're not even of conforming dimension; that $H$ in there between them matters. You express concern about the ratio (well let's put "ratio" in inverted commas and say we're speaking loosely), but you never take that ratio without the $H$ being there converting between them (as you already stated it did and as your mathematics clearly has it).

  • Note that $H\bar{P}H^T$ then converts from products-of-units of the state to products-of-units of the measurement; both $S$ and $R$ are on that scale.

I see no discrepancy anywhere -- all the units work as they should.

  • $\begingroup$ Ah, I was forgetting that the elements of K would be in ratios of units of the $i$th and $j$th components state and observation, respectively. $\endgroup$
    – npp1993
    Commented Apr 21, 2017 at 3:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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