A functional equation related to generalized entropies and the modular group

We solve a functional equation connected to the algebraic characterization of generalized information functions. To prove the symmetry of the solution, we study a related system of functional equations, which involves two homographies. These transformations generate the modular group, and this fact plays a crucial role in solving the system. The method suggests a more general relation between conditional probabilities and arithmetic.


Motivation and results
In this paper, we study the measurable solutions u : [0, 1] → R of the functional equation for all x, y ∈ [0, 1) such that x + y ∈ [0, 1]. The parameter α can take any positive real value. This equation appears in the context of algebraic characterizations of information functions. Given a random variable X whose range is a finite set E X , a measure of its "information content" is supposed to be a function f [X] : ∆(E X ) → R, where ∆(E X ) denotes the set of probabilities on E X , ∆(E X ) = p : E X → [0, 1] x∈EX p(x) = 1 . (1. 2) The most important example of such a function is the Shannon-Gibbs entropy where 0 log 0 equals 0 by convention. Shannon entropy satisfies a remarkable property, called the chain rule, that we now describe. Let X (resp. Y ) be a variable with range E X (resp. E Y ); both E X and E Y are supposed to be finite sets. The couple (X, Y ) takes values in a subset E XY of E X × E Y , and any probability p on E XY induce by marginalization laws X * p on E X and Y * p on E Y . For instance, X * p(x) = y:(x,y)∈EXY p(x, y). (1.4) The chain rule corresponds to the identities where p| X=x denotes the conditional probability y → p(y, x)/X * p(x). These identities reflect the third axiom used by Shannon to characterize an information measure H: "if a choice be broken down into two successive choices, the original H should be the weighted sum of the individual values of H" [7]. There is a deformed version of Shannon entropy, called generalized entropy of degree α [1, Ch. 6]. For any α ∈ (0, ∞) \ {1}, it is defined as (1.7) This function was introduced by Havrda and Charvát [4]. Constantino Tsallis popularized its use in physics, as the fundamental quantity of non-extensive statistical mechanics [8], so S α is also called Tsallis α-entropy. It satisfies a deformed version of the chain rule: Suppose now that, given α > 0, we want to find the most general functions f [X]-for a given collection of finite random variables X-such that A. f [X](δ) = 0 whenever δ is any Dirac measure-a measure concentrated on a singleton-, which means that variables with deterministic outputs do not give (new) information when measured; B. the generalized α-chain rule holds, i.e. for any variables X and Y with finite range 1 (1.10) The simplest non-trivial case corresponds to E X = E Y = {0, 1} and E XY = {(0, 0), (1, 0), (0, 1)}; a probability p on E XY is a triple p(0, 0) = a, p(1, 0) = b, p(0, 1) = c, such that X * p = (a + c, b) and Y * p = (a + b, c). The equality between the right-hand sides of (1.9) and (1.10) reads when α = 1.
The first is proved analytically, by means of standard techniques in the field of functional equations (cf. [1,9,5]), and the second by a novel geometrical argument, relating the equation to the action of the modular group on the projective line. Theorems 1.2 and 1.3 above imply that any measurable solution u of (1.1) must be symmetric, i.e. u(x) = u(1 − x) for all x ∈ [0, 1], and therefore whenever x, y ∈ [0, 1) and x + y ∈ [0, 1]. When α = 1, this equation is called "the fundamental equation of information theory"; it first appeared in the work of Tverberg [9], who deduced it from a characterization of an "information function" that not only supposed a version of the chain rule, but also the invariance of the function under permutations of its arguments. Daróczy introduced the fundamental equation for general α > 0, and showed that it can be deduced from an axiomatic characterization analogue to that of Tverberg, that again supposed invariance under permutations along with a deformed chain rule akin to (1.8), see [3,Thm. 5].
For α = 1, Tverberg [9] showed that, if u : [0, 1] → R is symmetric, Lebesgue integrable and satisfies (1.12), then it must be a multiple of s 1 (x). In [5], Kannappan and Ng weakened the regularity condition, showing that all measurable solutions of (1.12) have the form u(x) = As 1 (x) + Bx (where A and B are arbitrary real constants), which reduces to u(x) = As 1 (x) when u is symmetric. In fact, they solved some generalizations of the fundamental equation, proving among other things that, when α = 1, the only measurable solutions of (1.1) are multiples of s 1 (x).
For α = 1, Daróczy [3] established that any u : [0, 1] → R that satisfies (1.12) and u(0) = u(1) has the form 2 without any hypotheses on the regularity of u. The proof starts by proving that any solution of (1.12) must satisfy u(0) = 0 (setting x = 0), and hence be symmetric (setting y = 1 − x). Since we are able to prove symmetry of the solutions of (1.1) restricted to rational arguments without any regularity hypothesis, we also get the following result.
More details on the characterization of information functions by means of functional equations can be found in the classical reference [1], which gives a detailed historical introduction. Reference [2] summarizes more recent developments in connection with homological algebra.
It is quite remarkable that Theorem 1.1 serves as a fundamental result to prove that, up to a multiplicative constant, {S α [X]} X∈S is the only collection of measurable functionals (not necessarily invariant under permutations) that satisfy the corresponding α-chain rule, for any generic set of random variables S. In order to do this, one introduces an adapted cohomology theory, called information cohomology [2], where the chain rule corresponds to the 1-cocycle condition and thus has an algebro-topological meaning. The details can be found in the dissertation [10].

The modular group
The group G = SL 2 (Z)/{±I} is called the modular group; it is the image of SL 2 (Z) in P GL 2 (R). We keep using the matrix notation for the images in this quotient. We make G act on P 1 (R) as follows: an element acting on [x : y] ∈ P 1 (R) (homogeneous coordinates) gives Let S and T be the elements of G defined by the matrices The group G is generated by S and T [6, Ch. VII, Th. 2]; in fact, one can prove that S, T ; S 2 , (ST ) 3 is a presentation of G.

Regularity: proof of Theorem 1.2
Lemma 3 in [5] implies that u is locally bounded on (0, 1) and hence locally integrable. Their proof is for α = 1, but the argument applies to the general case with almost no modification, just replacing which is evidently valid too whenever x, y ∈ (0, 1).
To prove the differentiability, we also follow the method in [5]-already present in [9]. Let us fix an arbitrary y 0 ∈ (0, 1); then, it is possible to chose s, t ∈ (0, 1), s < t, such that for all y in certain neighborhood of y 0 . We integrate (1.1) with respect to x, between s and t, to obtain (s−t)u(y) = 1−s The continuity of the right-hand side of (3.1) as a function of y at y 0 , implies that u is continuous at y 0 and therefore on (0, 1). The continuity of u in the right-hand side of (3.1) implies that u is differentiable at y 0 . An iterated application of this argument shows that u is infinitely differentiable on (0, 1).

Symmetry: proof of Theorem 1.3
Define the function h : Observe that h is anti-symmetric around 1/2, that is, we have Let now z ∈ 1 2 , 1 be arbitrary and use the substitutions x = 1 − z and y = 1 − z in (1.1) to derive the identity Using the anti-symmetry of h to modify the right-hand side of the previous equation, we also deduce that Setting x = 0 (respectively y = 0) in (1.1), we to conclude that u(1) = 0 (resp. u(0) = 0). Hence, the function h is subject to the boundary conditions h(0) = h(1) = 0. From (4.3), it follows that h(1/2) = h(0)/2 α = 0. If the domain of h is extended to the whole real line imposing 1-periodicity: We establish first the anti-symmetry around 1/2 of the extended h (Lemma 4.2), which implies that (4.7) follows from (4.6); the latter is a consequence of Lemmas 4.3-4.7. 9) and the change of variables u = x − 1 gives This establishes (4.10).
Proof. Immediately deduced from the previous lemma using the anti-symmetric property in Lemma 4.2.
The transformations x → 2x−1 x and x → 1−x x in Equations (4.6) and (4.7) are homographies of the real projective line P 1 (R), that we denote respectively by α and β. They correspond to elements in G, that satisfy This last matrix corresponds to x → 1 − x. Proof. Let
(4.22) Let X be an arbitrary element of G. Since Y = P XP −1 ∈ G and G is generated by S and T , the element Y is a word in S and T . In consequence, X is a word in P −1 SP and P −1 T P , which in turn are words A and B 2 .
It is possible to find explicit formulas for S and T in terms of A and B 2 . Since P = S −1 T −1 , we deduce that P SP −1 = S −1 T −1 ST S and P T P −1 = S −1 T −1 T T S = S −1 T S. Hence, in virtue of (4.22), To finish our proof of Proposition 1.3, we remark that the orbit of 0 by the action of G on P 1 (R) is Q ∪ {∞}, where Q ∪ {∞} has been identified with {[p : q] ∈ P 1 (R) | p, q ∈ Z} ⊂ P 1 (R). This is a consequence of Bezout's identity: for every point [p : q] ∈ P 1 (R) representing a reduced fraction p q = 0 (p, q ∈ Z\ {0} and coprime), there are two integers x, y such that xq − yp = 1. Therefore x p y q is an element of G and g ′ Since h(1/2) = 0, the following lemma is the missing piece to establish that the extended h vanishes on Q (and hence the original h necessarily vanishes on [0, 1] ∩ Q). Lemma 4.9. For any r ∈ Q \ {0}, there exists a finite sequence w = (w i ) n i=1 ∈ {α, β, α −1 , β −1 } n such that r = w n • · · · • w 1 (1/2) and, for all i ∈ {1, ..., n}, the iterate x i := w i • · · · • w 1 (1/2) does not equal 0 or ∞.
Proof. Since the orbit in P 1 (R) of 1/2 by the group of homographies generated by A and B 2 (i.e. G itself) contains the whole set of rational numbers Q, there exists a w such that r = w n • · · · • w 1 (1/2), where each w i equals α, β or one of their inverses.
If some iterate equals 0 or ∞, the sequence w can be modified to avoid this. Let i ∈ {0, ..., n} be the largest index such that x i ∈ {0, ∞}; in fact, i < n because r = 0, ∞.