Guiding the modelling of data-driven techniques with some kind of prior information is a pretty old idea. This idea is maybe also what inspired physics-informed neural networks (PINNs). In PINNs, a term is added to the loss function of the network which penalizes certain actions, such that the expressions of the network do not violate known physical laws. This is very much informed by the known physical properties of the problem which is supposed to be modelled. The probably simplest expression for this is a known monotonic relationship between a known variable in the input, and the variable of interest (the output).

But what was monotonicity again?

Monotone functions are a set of functions that develop in only one direction: increasing or decreasing. There is also a distinction between weakly monotone and strictly monotone. Strictly monotone requires the function to always be either increasing or decreasing, while weakly monotone relaxes this requirement and allows the function to stay constant too.

In this context, we will mostly stick to weakly increasing monotonicity, as in reality, most relationships which are monotone are so weakly. Meanwhile the distinction between increasing and decreasing is simply a sign change.

We can define this monotonicity mathematically as follows: Let $f: X \to \mathbb{R}$ be a function on an ordered domain $X \subseteq \mathbb{R}$. We say $f$ is weakly increasing if

That is, moving to a larger input never decreases the output. Replacing the second $\leq$ with a strict $<$ gives strictly increasing monotonicity, and reversing its direction ($\geq$) gives the weakly decreasing case. Note that we can refer to the weakly increasing case also as non-decreasing, which means the same.

Below are examples of weakly increasing monotone functions:

We can see that those functions can still have a lot of different shapes!