Bernstein Functions at Work: Coalescents, Copulas, and Subordination

What the math gives to ML

The paper supplies constructive Laplace-transform certificates for positivity and monotonicity: Bernstein functions are positive mixtures of exponential increments, while special Bernstein functions have reciprocal transforms whose potential density is nonincreasing. This suggests replacing unconstrained distance- or lag-dependent neural kernels with finite completely monotone mixtures, giving a learnable multiscale component with guaranteed nonnegative weights and monotone decay. The most practical first transfer is a stable attention or state-space decay kernel whose parameters are learned in log-space and whose positivity and monotonicity are guaranteed by construction rather than encouraged by a penalty.