You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
For a deterministic forward map G, the current EKP implementation works on the mean \bar{y} and covariance \Gamma of the data. Internally a regularization technique in the form drawing new samples of data for each iteration and ensemble member.
Effectively this solves the problem y = G(\theta) + N(0,\Gamma) by generating samples of \gamma_i ~ N(0,\Gamma) and using samples y_i = \bar{y} + \gamma_i in the algorithm. As we provide \bar{y} we achieve an unbiased result.
Also note - There should be no issue in this formulation working where \bar{y} is replaced by a sample y (and the algorithm still perturbs this internally). In this case one will acheive a (naturally) sample-biased result.
Nondeterministic setup
For nondeterministic forward map G (with mean \bar{G}) we would like to use a sample y and covariance \Gamma here. We do not regularize (redraw the data at every iteration) because the internal variability is sufficient.
Effectively this solves the problem y = G(\theta) = \bar{G}(\theta) + N(0,\Gamma) , where we don't have access to \bar{G}, thus a sample of G already contains the correct variability and no regularization should be performed. As we provide only a sample y this will acheive a (naturally) biased result
To obtain consistency with the current CES framework, the sample of data for EKI and MCMC should be the same.
Solution?
Perhaps the simplest way to do this would be to have a flag deterministic_forward_map=true.
If true, the sample is perturbed at every iteration and ensemble member (as is traditional)
If false, the sample is unperturbed.
The text was updated successfully, but these errors were encountered:
Deterministic setup
For a deterministic forward map G, the current EKP implementation works on the mean
\bar{y}
and covariance\Gamma
of the data. Internally a regularization technique in the form drawing new samples of data for each iteration and ensemble member.Effectively this solves the problem
y = G(\theta) + N(0,\Gamma)
by generating samples of\gamma_i ~ N(0,\Gamma)
and using samplesy_i = \bar{y} + \gamma_i
in the algorithm. As we provide\bar{y}
we achieve an unbiased result.Also note - There should be no issue in this formulation working where
\bar{y}
is replaced by a sampley
(and the algorithm still perturbs this internally). In this case one will acheive a (naturally) sample-biased result.Nondeterministic setup
For nondeterministic forward map G (with mean
\bar{G}
) we would like to use a sampley
and covariance\Gamma
here. We do not regularize (redraw the data at every iteration) because the internal variability is sufficient.Effectively this solves the problem
y = G(\theta) = \bar{G}(\theta) + N(0,\Gamma)
, where we don't have access to\bar{G}
, thus a sample of G already contains the correct variability and no regularization should be performed. As we provide only a sampley
this will acheive a (naturally) biased resultTo obtain consistency with the current CES framework, the sample of data for EKI and MCMC should be the same.
Solution?
Perhaps the simplest way to do this would be to have a flag
deterministic_forward_map=true
.The text was updated successfully, but these errors were encountered: