What should lower limit of sigma be?

[mars0i]

In a scalar model, sigma is variance, so it's nonsensical for it to be < 0, and it can't be 0, since we divide by it. However, Bogacz, following Friston, suggests setting a static sigma to 1. Should the dynamic sigmas also be kept >= 1?

What should the corresponding rule be for matrix sigmas, which are covariance matrices? At the very least, their determinants must be nonzero, since we divide by that number.

Also, if next-sigma would produce an inappropriate sigma, what value should it be given instead of what the default calculation would give it? The minimum value (which won't really make sense for matrices, probably)? Or whatever value it had last time--i.e. it will never change after that until eps changes?

I'm also not sure where the best place is to test for this.

[mars0i]

Currently restricting sigma to >= 1 for scalars. Still have to do further research for matrices. This answer might be right. Although then I have to figure out what the equivalent of >= 1 is.