Functional modeling: Newton's First Law is a case of functional modeling that doesn't require mathematics to comprehend.
Geometry itself was originally very empirical.
Then a perfect model was made in mathematics: The equation of a circle or a square, for instance. These objects don't exist in reality, but they make excellent approximations to our empirical models that they were indeed derived from. We recognize a class of objects that is 'squarish', or 'circlish'.
So the line between mathematics and physics is fuzzy at times. An excellent example is basic addition/subtraction. If there are four enemies in front of us, and we kill one, there is now three. That's a very observationally based model, and it often makes a lot of sense for us to view things as concrete wholes. Integers. This is so ingrained into mathematics that some would think it existed as a theory before it was observed, but I would think that it was observed first, but was so simple to incorporate: "hey! let's give each situation a different name: this will be one, this will be two, and we can add them such that they form the next situation which we'll call three!"
Once a set of rules is put together into a certain classification, the mathematicians will then formulate the rules and push them to the limits, and find errors in the logic and iron them out and make them more consistent. By pushing them to the limits, they will find consequences of implications of the original logic that may lead to or confirm other logic... If it doesn't, we reformulate the rules. In many cases, the full mathematical range of a set of rules is not observable (so we make new mathematical language: "this only works on x>0", but also in many cases, the mathematics makes predictions (based on the original logic, which most likely came from observed phenomena).
We can make up any function we chose in mathematics. It doesn't have to represent any physical observations to be mathematics.