Variance Propagation

Last installment, we observed that the variance in the state values changes over time, starting with a flaky covariance that reflects a lack of knowledge about what the initial system state is. As system activity continues, more and more information from observation of system outputs is incorporated. In time, the variations in the state variables become dependent on the driving internal noise, rather than the peculiar representation of initial uncertainty.

In this installment, we will examine this in more detail, watching the changes in the state covariance analytically.

 

Variance Propagation in an Observer system

Suppose the system starts in a completely unknown initial state. We can represent this by picking an arbitrary state vector of initial values randomly from within the normal operating range, obtaining initial state x0. For purposes of implementation, a zero state is as representative of the input range as any. So a plausible choice for initializing in the presence of no initial state information is simply to start with a zero state vector. For normalized variables, the covariance matrix P0 assumed for this state vector has values 1/3 on the main diagonal by construction.

The state transition equations for the first update has terms for the state transition response, coupled inputs, and random state noise.

   x1 = A x0 + B u0 + w0 

However, we know that from the separability property of linear systems, we can consider the effects of the input terms B u separately from the effects of the noise. Consequently, we can simplify this by considering only the disturbance offset terms e.

   e1 = A e0 +  w0 

When we add in some additional observer terms, the feedback term K will couple some of the original noise back from output to input, and in addition to that, couple some amount of the output observation noise. This has the effect of modifying the noise propagation terms and adding a new one.

   e1 = (I - K C) A e0  +  (I - K C) w0  K v0 

We have previously defined notations for covariances of the noise terms e, w, and v.

1. New internal random noise with covariance Q. This comes from unknown sources and directly disturbs the variables in the next state.
2. Observation random noise with covariance V. This is coupled into the state error through the observer action.
3. State uncertainty propagating in from the initial state vector x0. This is represented by the constructed P0 covariance matrix.

We know that variances are additive when random variables are added. We also know that for a transformed variable M x, where x has covariance P, that the covariance of the result is going to be M P MT. Consequently, we can determine that the new state covariance subsequent to the update will be as follows.

   P1 =  (I - K C) A P0 AT (I - K C)T AT  +  (I - K C) Q (I - K C)T  +  K V K

The output observation equations apply transformation C to the vector of state variables, so this will also show variance effects that can be calculated. The two contributing terms are the state vector and a vector of random observation noise described by covariance V.

y1 = C x1 + v0

Y1  =  C P1 CT + V  

Given the P0 variance of the state, we can repeat this processing to determine the state variance at each instant.

As always, we are not going to fully believe any of this until we see it put into practice. But before we can generate a suitable simulation, we need to know how to produce random vectors having a specified covariance; that's the topic for next time.

 

---