Let [n]={1,2,…,n}. We define a d-dimensional entropy power region as follows:
Γd(n)≜{[N(∑i∈SXi)]:X1,…,Xn∈Rd independent, random vectors ,S⊆[n],S=∅}.
Then we let the entropy power region be
Γ(n)=∪d≥1Γd(n)⊆R+2n−1.
Let G be a collection of nonempty subsets of [n]. Then given G, we call a function β:G→R+ a \emph{fractional partition} if for each i∈[n], ∑s∈G:i∈sβS=1. A set function ν:2[n]→R+ is called \emph{fractionally superadditive} if
ν([n])≥s∈G∑βsν(s).
Using these definitions, we define the class of all fractionally superadditive set functions ν with ν(∅)=0 as follows
ΓFSA(n)≜{[ν(s)]:ν is fractionally superadditive ∀ fractional partitions β,S⊆[n],S=∅}.
We know that Γ(n)⊆ΓFSA(n) due to Ghassemi and Madiman of which the subset sum EPI of Artstein, Ball, Barthe, and Naor is a special case. Further, we define a set function ν to be supermodular if
ν(s∪t)+ν(s∩t)≥ν(s)+ν(t),
for all sets s,t⊂[n]. If we let ΓSM be the class of all supermodular set functions ν with ν(∅)=0, then it is known that ΓSM(n)⊆ΓFSA(n) and that Γ(n)⊈ΓSM(n) for n=3 or more.
The open problems are the following. First, is the closure of Γ(n) convex? It is known that the closure of Γ(n) is a log-convex cone. Second, how does one characterize the closure of Γ(3)? It is known, for example, that the closure of Γ(2) equals ΓFSA(2).