Covariance matrix interpolation in PyTorch implemented as a torchtorch.nn.Module calledInterpolateCovariance with the following features:
- Interpolated covariance matrix maintains positive definiteness.
- Convergence to the original covariance matrix near the input coordinates.
The main use case is extending a multivariate gaussian distribution$Z(x_i)$ defined at$N$ descrete strictly increasing series of points$x_1,...,x_N$ to the continuous range$[x_1,x_N]$. Given the covariance of$Z(x_i)$ defined at$x_1,...,x_N$ theInterpolateCovariance module outputs the covariance at an arbitrary set of points.
python -m torch_interpcov