On Tightness of Mutual Dependence Upperbound for Secret-key Capacity of Multiple Terminals

Csiszar and Narayan[3] defined the notion of secret key capacity for multiple terminals, characterized it as a linear program with Slepian-Wolf constraints of the related source coding problem of communication for omniscience, and upper bounded it by some information divergence expression from the joint to the product distribution of the private observations. This paper proves that the bound is tight for the important case when all users are active, using the polymatroidal structure[6] underlying the source coding problem. When some users are not active, the bound may not be tight. This paper gives a counter-example in which 3 out of the 6 terminals are active.

💡 Research Summary

The paper revisits the secret‑key generation problem for a set of terminals that observe correlated discrete memoryless sources. Building on the framework of Csiszár and Narayan (2004), the authors model the public discussion required for “communication for omniscience” (CO) as a linear program (LP) whose constraints are the classic Slepian‑Wolf inequalities. For a given set of active users A⊆M, the feasible rate vectors R∈ℝ^m must satisfy ∑{j∈B}R_j ≥ h(B) for every non‑empty subset B⊆M that does not contain A, where h(B)=H(X_B|X{B^c}) is the conditional entropy. The minimal total public‑discussion rate R_CO(A) is the optimum of this LP, and the secret‑key capacity is C_SK(A)=h(M)−R_CO(A).

Csiszár and Narayan derived an upper bound on C_SK(A) expressed as a mutual‑dependence functional I(A). I(A) is defined by minimizing, over all partitions (C_1,…,C_k) of the whole terminal set M that intersect A, the quantity (1/(k−1))∑{i=1}^k H(X{C_i})−H(X_M). This expression can be rewritten as a normalized Kullback‑Leibler divergence from the joint distribution of the sources to the product of the marginal distributions over the partition blocks, and is commonly interpreted as a measure of multivariate mutual dependence.

The central contribution of the paper is a precise characterization of when the bound is tight. Proposition 1 (tightness condition) states that C_SK(A)=I(A) if and only if there exists a partition (C_1,…,C_k)∈P_k(A) and a feasible rate vector R such that each complement C_i^c yields a tight Slepian‑Wolf constraint, i.e., ∑_{j∈C_i^c}R_j = h(C_i^c) for all i. This condition is both necessary and sufficient and provides an operational test for tightness.

To prove sufficiency, the authors exploit the polymatroidal structure of the entropy function. Proposition 2 shows that if two Slepian‑Wolf constraints are tight, then both their union and intersection are also tight. This follows from the supermodularity of the conditional entropy h, a property that makes the set of tight constraints closed under union and intersection.

When every terminal is active (A=M), the dual of the LP (by the standard LP duality theorem) yields a non‑zero dual vector y whose support corresponds to a set of tight constraints. By grouping columns of the incidence matrix that are identical, the authors construct a partition of M whose complement sets are exactly the unions of those tight constraints. The polymatroid property guarantees that these complement sets are themselves tight, satisfying the condition of Proposition 1. Consequently, Theorem 1 asserts that for the fully active case, the mutual‑dependence upper bound is exact: C_SK(M)=I(M).

The paper then demonstrates that this exactness does not extend to scenarios with helpers (users that are not required to recover the source). A concrete counter‑example with six terminals (m=6) and three active users (A={1,2,3}) is presented. The sources are defined as XORs of four independent uniform bits, leading to a simple conditional‑entropy profile: h(B)=0 for |B|=1,2; h(B)=1 for |B|=3,4; h(B)=2 for |B|=5. Solving the LP yields R_CO(A)=9/4 and thus C_SK(A)=1.75. Exhaustive enumeration of all admissible partitions shows that the minimum I(A) equals 2, strictly larger than the secret‑key capacity. Hence the bound is loose in this case.

The authors conclude that the mutual‑dependence functional I(A) precisely captures secret‑key capacity only when all terminals are active. When helpers are present, the bound may be loose, and a more general expression is needed. They note that exhaustive computer checks confirm tightness for three‑terminal systems and, informally, for four terminals, while the five‑terminal case remains intractable.

Overall, the paper contributes a rigorous proof technique that combines linear‑programming duality with polymatroid theory to clarify the role of Slepian‑Wolf constraints in secret‑key generation. It also highlights the limitations of the current mutual‑dependence measure and opens avenues for future research on tighter bounds in the presence of helper terminals.