Forcing monotonicity

[hanslovsky]

Forcing monotonicity is a great idea but will fail at the current state if for any integer number less or equal to the comparison range, all samples violate the monotonicity constraint. In this case no support for the fit is present and, therefore, an exception is thrown during estimation.

To overcome this, we should reconsider monotonicity enforcement:

1. Omit it completely
2. Leave it as is with an adaptive range or filling in values for which no support is present
3. Enforce monotonicity only on the fit, not on the samples that contribute to the fit.

Going with 3 in combination with an adaptive range makes the most sense to me.