Thread: variance and covariance need to clarify!
View Single Post
  #2  
Old January 11th, 2009, 10:52 AM
Laurent Laurent is offline
Super Member

  
Join Date: Aug 2008
Location: Lyon, France
Posts: 780
Country:
Thanks: 44
Thanked 501 Times in 420 Posts
Laurent is a name known to allLaurent is a name known to allLaurent is a name known to allLaurent is a name known to allLaurent is a name known to allLaurent is a name known to all
Default
Quote:
Originally Posted by ArixII View Post
Hi,
There is an example in an article where the author calculates variance and covariance, but I can't understand how (I'm novice in the topic). The example is about a vehicle that is traveling straight on the road:

The postition is measured with an error of 10 feet (one standard deviation)
The command acceleration is a constant 1 foot/sec^2
The acceleration noise is 0.2 feet/sec^2 (one standard deviation)
The position is measured 10 times per second.

Since the position is proportional to 0.0005 times the acceleration, and the acceleration noise is 0.2 feet/sec^2, the variance of the position noise is (0.005)^2 * (0.2)^2 = 10^(-6)
...

[and what formula is used to calculate this? ]
It seems to be simply a consequence of {\rm Var}(\lambda X)=\lambda^2{\rm Var}(X), which is easily seen from the definition of the variance.

So the variance of X=0.005 A is {\rm Var}(X)=0.005^2 {\rm Var}(A)=0.005^2\cdot 0.2^2.

Quote:
[author continues]
...Similarly, since the velocity is proportional to 0.1 times the acceleration, the variance of the velocity noise is (0.1)^2 * (0.2)^2 = 4 * 10^(-4). Finally, the covariance of the position noise and velocity noise is equal to the standard deviation of the position noise times the standard deviation of the velocity noise, which can be calculated as (0.005 * 0.2) * (0.1 * 0.2) = 2 * 10^(-5).
=================================================

Well I know that:
cov(X,Y) = E[(X - mu_x) * (Y - mu_y)]
but this is not what he used, right? does somebody knows the formulas used in such calculations?

Thank you!
Again, this is in fact quite simple, and very specific to your setting: you have X=\lambda A from above, and V=\mu A where \lambda=0.005 and \mu=0.1 (with appropriate units), so that {\rm Cov}(X,V)=E[(X-E[X])(V-E[V])]=E[\lambda(A-E[A])\mu(A-E[A])]=\lambda\mu E[(A-E[A])^2]=\lambda\mu{\rm Var}(A)=\sqrt{\lambda^2{\rm Var}(A)}\sqrt{\mu^2{\rm Var}(A)}=\sqrt{{\rm Var}(X)}\sqrt{{\rm Var}(V)}=\sigma(X)\sigma(V).

I hope I'm right, I didn't try to understand the physics beneath. The whole thing relies on the way proportional quantities behave for the variance/covariance.
Reply With Quote