Gaussian Processes
Apr. 26th, 2008 05:57 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslGaussian Processes have been on my mind since Wednesday. Here are some thoughts:
GPs must obey certain constraints:
* correlation function must yield a valid correlation matrix (be symmetric, obey the triangle inequality for correlation)
* no "quantum entanglement", i.e. observing leads to a normal Bayesian update: the function P(X1) = SUM_x2 P(X2=x2)P(X1|X2=x2)
One can encode a linear trend by adding a linear term to the function.
What about:
* circular domains
* periodic domains: can we encode more correlations? What does the correlation between peaks say about the Fourier Transform?
GPs must obey certain constraints:
* correlation function must yield a valid correlation matrix (be symmetric, obey the triangle inequality for correlation)
* no "quantum entanglement", i.e. observing leads to a normal Bayesian update: the function P(X1) = SUM_x2 P(X2=x2)P(X1|X2=x2)
One can encode a linear trend by adding a linear term to the function.
What about:
* circular domains
* periodic domains: can we encode more correlations? What does the correlation between peaks say about the Fourier Transform?