A general wireless networking problem is formulated whereby end-to-end user rates, routes, link capacities, transmit power, frequency and power resources are jointly optimized across fading states. Even though the resultant optimization problem is generally non-convex, it is proved that the gap with its Lagrange dual problem is zero, so long as the underlying fading distribution function is continuous. The implication is that separating the design of wireless networks in layers and per-fading state subproblems can be optimal. Subgradient descent algorithms are further developed to effect the optimal separation in layers and layer interfaces.