(I got interested in this when a funny entropy measure, the collision entropy, popped up in a paper I was reading.)

The Rényi entropies are a family of entropy measures that includes the well-known Shannon entropy along with a bunch of other more obscure ones that sometimes crop up. For simplicity I’m going to use an example of a system with four states. The probabilities of being in each state are labelled, for example, .

Entropy is connected to how ‘spread out’ these probabilities are. For example, is maximally spread out over the four states and has the maximum entropy. At the other end of the scale, is fully concentrated on one state and will have a low entropy. will be somewhere in between.

The Rényi entropies all use a measure of concentration called majorisation, which works in the following way:

First, reorder the elements of the vector in descending order. For the three example vectors so far this gives:

Second, compute partial sums:

If all the partial sums of one vector are bigger than another vector , we say that majorises . majorises the others because each partial sum is higher than for the other vectors - it gets to 1 immediately and stays there. Whereas is majorised by everything - it’s the slowest way to get to 1.

If vector majorises vector , it’s more concentrated and should have lower entropy - writing the entropy as , you want . Functions of this sort are called Schur concave.

The Rényi entropies all have this same Schur concave behaviour. They depend on a free parameter , and roughly speaking the larger is the more weight it gives to more probable states.

More precisely, the Rényi entropy is related to the -norm (actually let’s call it the -norm as we already have for probability knocking around). For a vector of probabilities the Rényi entropy can be written

or in terms of the -norm ,

.

Either way the formula messes up for . The limiting case for turns out to be the Shannon entropy.

As ever, people mostly just care about , , and . The max-entropy case weights everything nonzero equally. The min-entropy only cares about the highest probability event and ignores the rest.

The case is an interesting one. It’s called the collision entropy, and because it squares the probabilities it ends up weighting the most probable events more highly than the Shannon entropy would. I can see some vague link to collisions, as you’re taking the square of each probability, so they’re something like interaction terms.

Intuitively you would expect the 2-norm to be important because it always is, and does indeed seem to have some use in quantum mechanics (the van Enk paper quoted at the start refers back to a paper by Brukner and Zeilinger), but I’m still hazy on how or why.


To add:

  • examples
  • exactly how van Enk’s measure of ‘predictability’ relates to . Already done this, just have to write up the notes.

To read: