# Maximally entropy preserving irreversible functions. (CS related)

The topic/problem is related to hashing for data structures used in programming, but I seek formal treatment. I hope that by studying the problem I will be enlightened of the fundamental limitations of hashing. So, the purpose is educational.

Let's consider a family of functions $f_i$ which will be applied on a random variable $X$, with unknown weight function $w$, drawn from some finite set $P$.

$f_i: P \rightarrow Q$

$\Vert P\Vert=n>\Vert Q\Vert =m$ (therefore $f$ is irreversible)

The entropy for $X$ is:

$H(X) = q$

What is the function $f_i$ that maximizes

$\min_w \{H(f_i(X))\}\qquad(\max)$.

Clarification: The objective function we are trying to maximize is the minimum entropy for $f_i(X)$, taken over the various possible weight functions $w$ for $X$.

To paraphrase, what is the transformation that promises to preserve the most entropy in its output, regardless of the input's distribution (, as long as the entropy of the input is some fixed $q$.) The question can also be raised with the ratio

$\min_w \left\{\dfrac{H(f_i(X))}{H(X)}\right\}\qquad(\max)$,

or the difference

$\min_w \{H(f_i(X)) - H(X)\}\qquad(\max)$,

and then $q$ can be permitted to vary. I suspect there will be subtleties that I am not trained well enough to see (; such as $q=0$).

The asymptotic behavior is interesting - when $P$ is much larger than $Q$ and $Q$ is sufficiently large. That is, $\frac n m\to+\infty$ and $m\to+\infty$. Particularly, finding a sequence of functions that tends to preserve the entropy completely as $m$ and $n$ grow indefinitely, or at least preserves it up to $\max(q, log_2m)$, would have great significance. If there are no such functions, is there a proof of that being the case?

I will be happy to have such result for the continuous case, if it simplifies things.

Also, I am interested to find literature that performs comparative study of the effect of certain well-known functions (modulo, division, etc.) on the entropy of their input, based on the general properties of the input's distribution (, not bound to particular distributions.)

Post-graduate CS level. Not the most math savvy, but I have passed introductory course to probability and statistics, and to information theory.

Thanks and Regards,

-
Edit: size of P is greater than Q. Wrong direction of the inequality in my post. –  simeonz Jun 11 '12 at 0:09

You won't find any very good solution that maximizes the minimal output entropy over all possible probability distributions for $X$. No matter which $f$ you decide on, the adversary could then choose a probability distribution that concentrates all of the entropy of $X$ in values that map to the same $Q$ value -- for a resulting entropy of zero. This makes your ratio of entropies rather uninteresting.

It works slightly better to ask about the amount of lost entropy. In the worst case where the output entropy is 0, the input entropy cannot be greater than the log of the largest number of $p$ values that go into the same $q$ bucket. So in order to minimize the lost entropy you should choose an $f$ that maps either $\lfloor n/m\rfloor$ or $\lceil n/m\rceil$ different $p$s to each $q$. Any deviation from this will give the adversary a chance to make you lose more than $\log\frac nm$ entropy.

However, that's the most specific advice you're going to get out of the problem as formulated, since you have abstracted away all structure of $P$, $Q$, and $X$ except for their cardinality.

What one does in practice is to make further (often implicit) assumptions about $X$ and its probability distribution. In the pessimistic extreme, if only $X$ is generated by some stochastic process that does not involve inordinate amounts of computation, then one should hope that a good cryptographic hash function would mix up things enough to extract all of the entropy in it that $Q$ has room for. (Because if it didn't, then because of that the hash function wouldn't be "good" after all).

-
You are absolutely right, of course. Great omission on my side. But this leaves me with the "distribute the universe evenly" requirement to good hash functions, which is not the full story in practice. Aren't there fields that deal with the formal semantics of "mixing" the information, potentially involving models of the data, but with greater generality? Most searches for such treatment lead me to cryptographic articles, and I presume those will be focusing on the security perspective. Or are they connected? –  simeonz Jun 11 '12 at 1:16
Well, cryptographers are the ones most in need of such a theory, but that doesn't been you should ignore their work just because you have a different agenda. After all, the great thing (one of the great things) about mathematics is that abstraction can allow insights found in one area to become useful in another. Furthermore, cryptography has a strong tradition of striving to reduce the "fuzzy" security discussion to precisely defined mathematics as much as possible, so even though the motivation sections of their papers will usually mention the security perspective, ... (cont.d) –  Henning Makholm Jun 11 '12 at 11:24
... that doesn't mean that the actual mathematical content is necessarily tied exclusively to that perspective. Even furthermore, the state of the art in cryptography offers clues to the state of the art in general. If a mature theory of entropy preservation along the lines you sketch existed, surely cryptographers would eagerly latch onto it and use it in the construction and evaluation of cryptographic hash functions (which is in a somewhat sorry state, and not for lack of mathematical ability among the practitioners). Since this isn't happening, I suspect it doesn't exist. –  Henning Makholm Jun 11 '12 at 11:30
Then, the best I could do is to study the state of practice more carefully. Digest through related fields later, if time and skill permit me. Well, I have been warned. Thanks for the response. –  simeonz Jun 11 '12 at 12:56