Given a prior or bounding box in parameter space, one could propagate that set of model params through the model to construct a matrix of simulator outputs. Then, doing SVD or some other rank-reducing projection, provides a transformation into a latent space. This reduces time in calculating the likelihood for large output spaces, and removes noise. Then, the model discrepancy can directly act on the coefficient space.
This SVD can be done directly on the matrix of model simulator outputs, or on those weighted by the experimental covariance.
This should be done at the level of the Constraint object. It is probably most cleanly done as a constructor option.
Given a prior or bounding box in parameter space, one could propagate that set of model params through the model to construct a matrix of simulator outputs. Then, doing SVD or some other rank-reducing projection, provides a transformation into a latent space. This reduces time in calculating the likelihood for large output spaces, and removes noise. Then, the model discrepancy can directly act on the coefficient space.
This SVD can be done directly on the matrix of model simulator outputs, or on those weighted by the experimental covariance.
This should be done at the level of the
Constraintobject. It is probably most cleanly done as a constructor option.