An MTP$_2$ property for conditional distributions

What the math gives to ML

The paper introduces conditional multivariate total positivity of order two (cMTP2), a lattice log-supermodularity condition that directly constrains how a conditional distribution changes jointly with its conditioning variables and queried outcomes. Its key transferable asset is that cMTP2 is stronger than stochastic and tail monotonicity while weaker than imposing MTP2 on a full joint density, offering a flexible structural regularizer for conditional probabilistic models. A practical adaptation is to penalize violations of the cMTP2 lattice inequality on pairs of conditioning points and output thresholds, thereby encouraging coherent uncertainty shifts without requiring a fully MTP2 joint density. This is most promising for distributional regression, ordinal prediction, conditional generative models, and probabilistic forecasting where monotone dependence is plausible.