Logarithmic Modeling

Many real responses do not scale with raw differences, they scale with ratios. Doubling the intensity of a sound or an earthquake adds a fixed amount to what we perceive, it does not double the perception. That is the signature of a log relationship: inputs multiply while outputs add. A logarithmic model, f(x) = a log(x) + c, captures exactly this. Every time x gets multiplied by the same factor, f(x) increases by the same fixed amount, no matter where you start.

To build one from data, find a point where x = 1. Since log(1) = 0, plugging in x = 1 collapses the model to f(1) = c, so that point hands you c immediately. Then use a second point (p, y) to solve for a: y = a log(p) + c, so a = (y - c) / log(p). With a and c both known, the model is complete and you can plug in any x to predict f(x), including values far outside your original data.

This topic sits right next to exponential modeling because the two are mirror images: an exponential function undoes a log model and vice versa. Where exponential models fit data that keeps accelerating, log models fit data that shoots up early and then slows down, leveling off without ever truly stopping. Recognizing that curved, decelerating shape in a scatterplot is the signal to reach for a logarithmic regression instead of a straight line, which would force a constant rate of change onto data that clearly does not have one.