Browse the glossary using this index

Special | A | B | C | D | E | F | G | H | I | J | K | L | M | N | O | P | Q | R | S | T | U | V | W | X | Y | Z | ALL

M

Brian Etherton

Minimum Error Variance

by Brian Etherton - Thursday, July 20, 2006, 2:56 PM
 
One of the prime equations for data assimlation is the Kalman Filter Equation. This equation is:

x(a)-x(f) = PH(T){HPH(T)+R}(-1)[d-Hx(f)]

Where x(f) is the forecast, x(a) is the analysis, P is the background error covariance matrix, and R is the observation error covariance matrix. x(a) and x(f) are vectors, containing every model variable at every model gridpoint.

P gives the information regarding errors of the model first guess field
R gives information about the errors of observations (instrument and representativeness errors).

This equation, and the notion of 'covariance matrices', is often a little overwhemling. It can be simpler to consider the case of trying to best know the value of one variable.

For just one variable...
T(a) - T(f) = s(f)**2/(s(o)**2+s(f)**2)*[T(o)-T(f)]

Each term in the equation for one variable has a match with the full blown Kalman Filter equation. T(a) is the analysis value (best estimate), as is x(a), T(f) is the model first guess, as is x(f). And so on.

Consider the temperature of the air at 500mb over Denver Colorado.

To estimate this value, T(a) (a for 'analysis') we have two estimates:

#1 is the value as reported from the DNR sounding.
An observation

I'll call that T(o) (o for 'observed') Using known history, the average error of these observations has an error variance s(o)**2

#2 is the value from the model first guess field
Model First Guess
I'll call that "T(f)" (f for 'first guess') Using known history, the average error of these observations has an error variance s(f)**2

Here is where the graphic comes in!

The graphic I am envisioning is rather similar to this one:

Two become one

This graph would be re-worked such that instead of FCST A and FCST B, we have First Guess (T(f)) and Observation (T(o))

The applet I am envisioning is something that can be changed from the image on the left to the image on the right.

Sliders (as well as text entires) will allow the user to adjust both the values of T(f) and T(o) as well as the error variances s(f)**2 and s(o)**2. Also there would be a box showing the value of T(a).

What if our forecast model was always perfect? In that case, the error variance of the observations, s(f)**2, would be zero. In that case, the equation would collapse to T(a)-T(f) = [T(o)-T(f)]*(0), or T(a)-T(f)=0. At this time, the image would show that T(a) and T(f) are the same value, and that the spread (the tails of the Gaussian curves in the figure above) would be zero.

Consider now the other extreme, where the observations are perfect (but the forecast is not). In that case, the equation becomes T(a)-T(f) = [T(o)-T(f)]*(s(f)**2/s(f)**2), or T(a)-T(f)=T(o)-T(f), and thus, T(a)=T(o).

Users could adjust the values of the error variances. The 'analysis value', T(a) would move towards whichever value (T(a) or T(f)) had a smaller error variance associated with it. As the error variances were increased, the picture would look more like the right side of the above image. If error variances were decreased, the image would look more like the figure on the left of the above image.

Included with the estimate of T(a), the actual value, there would be a Gaussian shape around it with variance equal to [1/(s(o)**2+s(f)**2)](-1), showing that when the error variances of the two estimates are small, the error variance (uncertainty) of the analysis is also small.

This module will show the correlation between error statistics of the two estimates (first guess and observation) and the analysis value.