data
: Data and data uncertainties/noise¶
data
is a Matlab structure that defines any number of data and the
associated uncertainty/noise model.
data{1}
defines the first data set (which must always be defined),
and any number of additional data sets can be defined in data{2}
,
data{3}
, …
This allows to consider for example seismic data in data{1}
, and
electromagnetic data in data{2}
.
For each set of data, a Gaussian noise model (both correlated and
uncorrelated) can be specified. The noise model for different data types
(e.g. data{1}
and data{2}
are independent).
Once the noise model has been defined, the log-likelihood related to any
model, m
, with the corresponding forward
response, d
, can be computed using
[d,forward,prior,data]=sippi_forward(m,forward,prior,data)
logL=sippi_likelihood(data,d)
where d
is the output of sippi_forward.
The specification of the noise model can be divided into a description of the measurement noise (mandatory) and the modeling error (optional).
Gaussian measurement noise¶
Gaussian modeling error¶
The modeling error refers to errors caused by using for example an imperfect forward model, see HCM14.
A Gaussian model of the modeling error can be specified by the mean,
dt
, and the covariance, Ct
.
For example
data{1}.dt=[0 0 0];
data{1}.Ct=[4 4 4; 4 4 4; 4 4 4];
is equivalent to
data{1}.Ct=4
which implies a zero mean modeling error with a covariance model where all model parameters has a covariance of 4.
sippi_compute_modelization_forward_error can be used to estimate the modeling error related to using an approximate forward model. See the tomography example, for an example of accounting for correlated modeling errors, following HCM14.