Estimating Entropy¶
We are interested in the amount of entropy (or information) that is transferred to humans in a visual hash. It is worth taking a moment to define terms in a human-centric manner.
Self-information¶
The self-information of a given image is given by
where \(P(x)\) is the probability of observing state \(x\).
Information¶
The information (or entropy) of a system is actually the weighted average of the self-information all possible states.
Estimating the entropy take II¶
Let us consider a system consisting of \(N\) parts, each of which has \(n\) distinct configurations. If I randomly modify one part, then the chance of changing the system is \(q = 1 - \frac{1}{n}\). The entropy of such a system is given by
If I change a fraction \(f\) of the parts of the system, the probability of changing the system is given by
where \(q\) is the probability of a given thing not changing the system when it is modified, and math: ‘A’ is the variable accounting for user error. Let us now consider a fraction \(f_\gamma\) that has a probability \(\gamma\) of not changing the system.
We will assume that we can measure \(f_\gamma\), and also that we can (and do) measure \(f_{\gamma^2}\), the fraction that leads to the square of probability \(\gamma\).
We can solve for the following:
Taking these together, we can show that the entropy is given by
A nice option seems to be
which balances the two at an average of 50%, thus avoiding a bias in one direction or the other. This is the golden ratio minus one.