Network-Weighted Proportional Fairness

Updated 6 July 2026

Network-weighted proportional fairness is defined as a logarithmic utility framework that incorporates explicit utility weights or endogenous network factors (e.g., channel prices, congestion) to guide fair resource distribution.
It applies to diverse scenarios, including multi-channel wireless and HetNets, ensuring Pareto efficient allocations by balancing equivalent airtime and congestion costs across networked resources.
Methodologies involve decomposing non-convex problems into convex subproblems, deploying distributed algorithms, and analyzing heavy-traffic limits to achieve scalable and robust network performance.

Searching arXiv for relevant papers on weighted/network-aware proportional fairness and closely related formulations. Network-weighted proportional fairness is the use of proportional-fair utility in network resource allocation when the fairness criterion is shaped either by explicit weights or by network-coupled quantities such as channel prices, interference-neighborhood aggregates, or path congestion costs. In the most standard utility-theoretic sense, weighted proportional fairness is the $\alpha=1$ member of the weighted $\alpha$ -fair family; in multi-channel wireless systems, proportional fairness can also be interpreted as equalizing an “equivalent airtime” that is a weighted sum of per-channel airtimes, with the weight of a channel being its price or value [(Borst et al., 2012), 0611075]. Across bandwidth-sharing networks, multi-band wireless systems, multi-RAT HetNets, and heavy-traffic flow models, the recurring structure is a sum of logarithms over network-coupled throughputs, with the relevant “weights” arising either exogenously from utility coefficients or endogenously from topology, congestion, and shadow prices (Aryafar et al., 2019, Wang et al., 2018).

1. Utility-theoretic formulation

In the canonical communication-network model, the network has a finite set of links/resources $J$ and a finite set of routes/users $R$ . The topology is encoded by an incidence matrix

$A=(A_{jr})_{j\in J,r\in R},$

with feasible aggregate route allocations

$\mathcal C=\{\Lambda\in \mathbb R_+^{R}: A\Lambda\le C\}.$

If $y_r$ is the number of flows on route $r$ , then the per-flow rate is

$x_r=\frac{\Lambda_r}{y_r},$

and the network utility function is

$U(y,\Lambda)=\frac{1}{\sum_{i\in R} y_i}\sum_{r\in R} y_r\,U_r\!\left(\frac{\Lambda_r}{y_r}\right).$

The weighted $\alpha$ 0-fair family is

$\alpha$ 1

At $\alpha$ 2, this is weighted proportional fairness; if all $\alpha$ 3, it is ordinary proportional fairness in the Kelly sense (Borst et al., 2012).

The same paper gives a structural characterization that is unusually strong. A utility is network iso-elastic if multiplying all allocated bandwidths by a common factor does not change the preference ordering among network states: $\alpha$ 4 Its main theorem states that, for the separable average-utility form above, network iso-elasticity is equivalent to weighted $\alpha$ 5-fairness up to an additive constant. Weighted proportional fairness therefore appears not merely as one convenient choice, but as the $\alpha$ 6 member of the uniquely characterized family of network utility functions with this scaling property (Borst et al., 2012).

A plausible implication is that “network-weighted proportional fairness” has two legitimate readings. In one reading, the weights are explicit coefficients $\alpha$ 7 multiplying $\alpha$ 8. In the other, the weighting is induced by how the network embeds throughput into the feasible region or into the effective marginal cost of service.

2. Prices, equivalent airtime, and endogenous weighting

A central network-theoretic interpretation of proportional fairness in wireless systems is given by the deterministic multi-channel study that shows: PF leads to equal airtime allocation to users for the single-channel case, and equal equivalent airtime allocation to users for the multi-channel case, where the equivalent airtime enjoyed by a user is a weighted sum of the airtimes enjoyed by the user on all channels, with the weight of a channel being the price or value of that channel. The same work also establishes the Pareto efficiency of PF solutions [0611075]. This moves the notion of fairness away from raw airtime and toward airtime weighted by endogenous scarcity.

The same theme reappears in interference-coupled access systems. In distributed resource allocation for multi-band wireless systems, the client-weight aggregate at AP $\alpha$ 9 is

$J$ 0

while the interference-neighborhood aggregate is

$J$ 1

The resulting optimal access rule is

$J$ 2

This is an explicit network-coupled weighting law: an AP’s aggressiveness is determined by its own aggregate client weight relative to the total weight in its contention neighborhood (Hou et al., 2011).

In multi-hop deadline-constrained coding, the same role is played by shadow prices. With per-cell dual variables $J$ 3, the optimality condition contains the path-price term

$J$ 4

Here the left-hand side is a route-dependent congestion-weighted airtime cost accumulated over every cell on the flow’s path, while the right-hand side is a coding-side marginal tradeoff determined by channel loss and deadline (Karumbu et al., 2012). This suggests that network-weighted PF is often mediated by dual prices even when no explicit utility weights are introduced.

A related multi-RAT result makes the same distinction explicitly: in proportional-fair RAT aggregation, the only explicit weights are per-user priority weights $J$ 5, while most of the “weighting” induced by the network comes implicitly through heterogeneous PHY rates, BS resource constraints, and equilibrium water-fill levels (Aryafar et al., 2019).

3. Canonical network realizations

The topic spans several concrete network models in which the PF objective remains logarithmic but the throughput variable is network-coupled in different ways.

Setting	PF form	Network-coupled term
Multi-band self-organized wireless systems (Hou et al., 2011)	$J$ 6	$J$ 7, $J$ 8, $J$ 9
Multi-RAT traffic aggregation in HetNets (Aryafar et al., 2019)	$R$ 0	Per-BS resource constraints, water-fill levels $R$ 1, dual prices $R$ 2
Dual connectivity in HetNets (Taksande et al., 2020)	$R$ 3	PF metric uses the sum of the average throughput received by the MT from the two BSs to which it is connected
Multi-hop coding with deadlines (Karumbu et al., 2012)	$R$ 4	Route set $R$ 5, per-cell airtime prices $R$ 6, end-to-end symbol error probability $R$ 7

In multi-RAT aggregation, the aggregate throughput of client $R$ 8 is

$R$ 9

and the global weighted PF problem is

$A=(A_{jr})_{j\in J,r\in R},$ 0

subject to per-BS resource constraints $A=(A_{jr})_{j\in J,r\in R},$ 1 (Aryafar et al., 2019). In dual connectivity, by contrast, the utility remains unweighted,

$A=(A_{jr})_{j\in J,r\in R},$ 2

but the scheduler becomes network-aware because the PF denominator is the sum of the average throughputs from the two connected BSs rather than a purely local average (Taksande et al., 2020).

These examples show that network-weighted PF need not mean the same thing in every model. It may mean explicit utility weights, network-dependent denominators, path-price accumulation, or interference-neighborhood aggregates.

4. Algorithms, decomposition, and control

The algorithmic literature is correspondingly diverse. In deterministic multi-channel wireless networks, several PF-optimization algorithms are presented, including a fast algorithm that is amenable to parallel implementation [0611075]. In multi-band self-organized wireless systems, the problem naturally dissociates into a convex subproblem for channel access and client scheduling and a second subproblem for channel selection and client association, for which a Gibbs-sampler based approach and fast greedy algorithms are derived (Hou et al., 2011). In multi-RAT aggregation, AFRA is a distributed algorithm executed autonomously by each BS, and the paper shows that it converges to the desired PF solution and provides tight bounds on its convergence speed (Aryafar et al., 2019).

Cross-layer models often require more elaborate decomposition. In proportional fair coding for wireless mesh networks, the joint airtime/coding-rate utility problem is non-convex. One of the technical contributions is to show that the proportional fair utility optimisation can nevertheless be decomposed into a sequence of convex optimisation problems, and that the solution to this sequence of convex problems is the unique solution to the original non-convex optimisation (Karumbu et al., 2012). The dual update takes the familiar NUM form

$A=(A_{jr})_{j\in J,r\in R},$ 3

so per-cell airtime prices become the mechanism by which network contention is communicated to sources (Karumbu et al., 2012).

PF has also been realized through online adaptation and control. In coexisting WiFi networks, a simple algorithm allows WiFi nodes to independently tune their contention window to achieve proportional fairness, using asynchronous distributed optimization and only passive observations in a single collision domain (Gawłowicz et al., 2021). In 802.11e EDCA WLANs, a centralized adaptive control approach uses a Linear Quadratic Integral controller to periodically update $A=(A_{jr})_{j\in J,r\in R},$ 4 until the optimal fair point of operation is reached; the underlying utility is

$A=(A_{jr})_{j\in J,r\in R},$ 5

subject to per-AC delay constraints (Chen et al., 2015, Chen et al., 2015). These implementations are not all “weighted” in the explicit utility sense, but they demonstrate that network-aware PF is compatible with decentralized, dual, and control-theoretic realizations.

5. Stochastic and asymptotic theory

The asymptotic theory of weighted PF is unusually developed in bandwidth-sharing networks. For weighted proportionally fair sharing policies, heavy-traffic bounds are established for the weighted sum of the expected numbers of flows on different routes. If $A=(A_{jr})_{j\in J,r\in R},$ 6 is the number of critically loaded links and $A=(A_{jr})_{j\in J,r\in R},$ 7 is the positive route weight, then

$A=(A_{jr})_{j\in J,r\in R},$ 8

These bounds are heavy-traffic insensitive to a class of phase-type file-size distributions, and in the equal-weight case they imply the interchange of limits for the diffusion approximation under phase-type file-size distributions (Wang et al., 2018).

For unweighted PF, the stationary-measure theory reaches a complementary conclusion. In Markovian networks, PF and an associated reversible allocation are geometrically ergodic and have the same large deviations characteristics; in monotone networks, the same result is proved using stochastic comparisons that hold for general service requirement distribution (Jonckheere et al., 2012). The paper’s central message is that PF is “close” to allocations of service being insensitive to the service time requirement. A plausible implication is that network-weighted PF should be studied not only through pointwise utility optimization, but also through the geometry of its stationary tails and its relation to reversible proxy allocations.

These asymptotic results matter because they show that weighting is not merely a policy knob. In heavy traffic, route weights $A=(A_{jr})_{j\in J,r\in R},$ 9, the number of critically loaded links, and the route/link incidence structure all enter the first-order congestion law (Wang et al., 2018).

A recurring source of confusion is that not every PF scheduler on a network is an instance of explicit network-weighted proportional fairness. In wavelength-multiplexed quantum networks, the PF pumping strategy is standard PF over a network-constrained scheduling problem, not a fully general explicitly network-weighted PF with exogenous topology-dependent weights $\mathcal C=\{\Lambda\in \mathbb R_+^{R}: A\Lambda\le C\}.$ 0; the “network” enters through feasible topologies and through how simultaneous pumping changes achievable secret key rates (Bhambay et al., 18 Jul 2025). In dual connectivity HetNets, the scheduler is similarly network-aware because it uses aggregate average throughput across two BSs, but again without explicit utility weights (Taksande et al., 2020). In coexisting WiFi networks, the fairness notion is station-level proportional fairness, not network-level weighted proportional fairness (Gawłowicz et al., 2021).

This distinction also appears outside classical wireless bandwidth allocation. In private shared networks, the objective is unweighted proportional fairness over useful rates, but the feasible region is reshaped by dummy traffic and delay constraints, so privacy induces effective per-flow network weights and penalties inside a proportional-fair utility maximization framework (Feghhi et al., 2016). In RF energy-harvesting ALOHA networks, proportional fairness remains the unweighted objective

$\mathcal C=\{\Lambda\in \mathbb R_+^{R}: A\Lambda\le C\}.$ 1

but heterogeneity enters through the parameter

$\mathcal C=\{\Lambda\in \mathbb R_+^{R}: A\Lambda\le C\}.$ 2

which captures the double near-far effect (Hadzi-Velkov et al., 2018). In cluster computing, PF is again the sum-log allocation rule over multiple resources, and the paper explicitly presents it as analogous to sharing bandwidth in a network where each link is a distinct resource (Bonald et al., 2014).

Taken together, these results support a broad but technically precise usage. Network-weighted proportional fairness may refer to explicit weighted log-utility, to endogenous weighting by channel prices or path prices, or to PF scheduling rules whose normalizers are network-wide rather than local. What remains common is the logarithmic utility structure and the fact that fairness is assessed over a throughput variable that is inseparable from network topology, coupling, and congestion.