MBTR k=2 Question about dscribe HOT 3 CLOSED

Yes.

from dscribe.

Hi @sirmarcel,

In short: this behaviour you are seeing is intentional and is related to the way the discretization of the Gaussians behaves when using very low number of discretization points (in your case n=3). As soon as you crank up the number of points you will see that this "effect" disappears. See below if interested.

Longer version: To calculate the value of the discretized Gaussians, dscribe uses the cumulative distribution function of a normal distribution: CDF(x). The value at a grid point x is calculated by calculating using a simple finite difference: MBTR(x) = (CDF(x+dx/2) - CDF(x-dx/2) )/dx. By using this approach the normalization of the Gaussians is ensured even with a very low number of discretization points (as seen in your case, where the area under the three points always sums up to 1 as intended). If the implementation would directly use the Gaussian probability distribution, the area under the Gaussians would slightly depend on their location when using small number of discretization points. I wanted to avoid this and thus chose the described option. Both approaches converge to the same result when n is reasonably large.

Good question, hope this clears things up!

from dscribe.

Hi @lauri-codes,

Thanks for the prompt and in-depth reply! I think I understand your approach, and the observed "weirdness" is due to a combination of a) very low sigma and b) low n, right?

from dscribe.