Predictive Cross-Layer Resource Allocation

• Predictive cross-layer resource allocation is a method that jointly optimizes control variables from multiple layers using forecasted channel, interference, and QoS data.
• It addresses inefficiencies from traditional layer separation by integrating scheduling, power, routing, and resource allocation in unified optimization frameworks.
• Emerging techniques include feedback-based inference, predictive convex optimization, Lyapunov control, and machine learning approaches to enhance overall system performance.

Predictive cross-layer resource allocation is the joint optimization of resource-control variables drawn from multiple layers—such as routing, timing, network selection, packet scheduling, subcarrier assignment, beamforming, rate, and power—using predictive information about channel conditions, interference, traffic arrivals, mobility, queue evolution, or local quality-of-service (QoS) constraints. In the wireless literature, the cross-layer premise emerged from the observation that the conventional separate design of MAC packet scheduling and PHY resource allocation yields inefficient resource utilization, motivating integrated MAC-PHY formulations for OFDM, UWB, and OFDMA systems [0611080], (0907.3793, Ukil et al., 2011). The predictive component later expanded from feedback-driven inference based on ACK/NAK histories to future large-scale channel information, predictive Lyapunov control over multi-frame windows, time-varying distributed convex optimization with memory constraints, interference forecasting for URLLC, and radio-map-assisted route, timing, and power control in dynamic low-altitude networks (0905.4700, Yao et al., 2016, Yu et al., 2015, Coffman et al., 2020, Jayawardhana et al., 2023, Li et al., 25 Jul 2025).

1. Cross-layer foundation and early problem statements

Early cross-layer resource allocation papers focused on the inadequacy of layer separation in packet-switched and multi-carrier systems. In packet-switched OFDM networks, a cross-layer resource allocation algorithm was proposed on top of a novel multi-server scheduling framework in order to achieve overall high system power efficiency, with four explicit elements: MPGPS, a MPGPS-based joint MAC-PHY resource allocation scheme, A-MPGPS, and O-MPGPS [0611080]. In multiband UWB systems, the cross-layer allocation problem was framed around combining PHY-layer channel quality and MAC-layer QoS information so that sub-bands can be assigned in a way that both improves spectrum usage and differentiates service quality among users (0907.3793). Related UWB work under heterogeneous constraints used PHY-layer effective SINR and interference power together with MAC-layer two-class service differentiation in order to satisfy hard QoS requirements and limit interference to licensed users (Khalil et al., 2010).

A second early line concerned delay-tolerant traffic rather than strictly delay-bounded traffic. In heterogeneous multiuser OFDMA downlink, the Aggregate Throughput Optimization scheme targeted NRT and BE traffic by maximizing aggregate delivered data over a long-term window and exploiting the inherent time-diversity gain in mobile wireless environment for delay-tolerant applications (Ukil et al., 2011). In integrated cellular and Wi-Fi networks, the problem was posed as joint network selection, subchannel allocation, and power allocation with a power-delay tradeoff, where the two-timescale structure couples large-timescale association and small-timescale PHY/MAC decisions (Yu et al., 2015).

These formulations established the central architectural idea: cross-layer allocation is not merely “PHY adaptation with queue weights,” but a control problem in which layer-specific states and constraints are explicitly coupled. This suggests why predictive extensions became natural: once routing, scheduling, and power are optimized jointly, future information at any one layer can alter the optimal decisions at the others.

2. Predictive information structures

The predictive variable in predictive cross-layer resource allocation is not uniform across the literature. Different systems use different sources of foresight, and the resource allocator is shaped accordingly.

Predictive input	Resource decision	Representative papers
ACK/NAK feedback history	User selection, power allocation, rate allocation	(0905.4700)
Future large-scale channel information	Scheduling threshold, water-filling level, energy-saving allocation	(Yao et al., 2016)
Predicted mobility, channel states, and traffic arrivals over a window	Network selection, subchannel allocation, power allocation	(Yu et al., 2015)
Time-varying reference signal with local QoS constraints	Distributed projected-gradient-style predictive allocation	(Coffman et al., 2020)
Past interference sequence decomposed by EMD	Finite-blocklength URLLC channel-use allocation	(Jayawardhana et al., 2023)
Predetermined 4D trajectories and radio maps	Strategic routing, tactical timing, operational power	(Chen et al., 1 Sep 2025, Li et al., 25 Jul 2025)

In the ACK/NAK-driven FDD-OFDM design, the transmitter does not know the scalar channel state $X_k$ directly. Instead, it maintains the feasible set of channel values consistent with past feedback,

$$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$

and uses lower and upper bounds induced by previous thresholds to choose the next user, power, and rate (0905.4700). In this formulation, prediction is feedback-driven channel inference rather than explicit CSIT.

In delay-tolerant single-user downlink, the predictive input is future large-scale channel information obtained from predicted trajectory and radio map. The main analytical claim is that future large-scale channel information can capture almost all the performance gain from knowing the full future channel, because the optimal policy depends on the threshold $g_{\rm th}^*$ and water-filling level $\nu^*$, and both can be inferred from the distribution induced by framewise large-scale gains (Yao et al., 2016). In low-altitude networks, the same large-scale perspective is generalized into a hierarchy: 4D mission trajectories, a 6D radio environment model, and a 7D state descriptor are fused to generate link-level and network-level channel prediction, which are then mapped to strategic, tactical, and operational decisions (Chen et al., 1 Sep 2025).

A common source of confusion is that some systems are called predictive even though the foresight horizon is very short. In distributed WiMedia UWB, the allocation is updated at the beginning of each superframe using channel quality information estimated from the current channel and exchanged during the beacon period; the paper explicitly states that this is not a truly predictive channel forecast in the sense of estimating future time evolution over multiple frames (0907.3793). By contrast, predictive Lyapunov scheduling assumes perfect prediction of users’ locations, cellular channel states, and traffic arrivals over a prediction window of $W$ frames (Yu et al., 2015), while URLLC interference forecasting predicts the next interference value one step ahead and immediately maps it into finite-blocklength channel uses (Jayawardhana et al., 2023).

3. Canonical optimization formulations and algorithmic motifs

A defining feature of predictive cross-layer resource allocation is that the predictive signal is embedded into a formal optimization or control law. Several recurring templates appear.

The first is the finite-horizon MDP with belief-state evolution. In FDD-OFDM without explicit CSIT, the objective is to maximize average goodput under a total power constraint and a target PER. The state contains the current lower and upper channel bounds, the threshold estimate, remaining power, and pointers to ACK/NAK successor states, while the action is the triple $\alpha_m=(p_m,r_m,a_m)$ (0905.4700). Under sufficiently small target packet error rate, the paper derives a closed-form asymptotically optimal policy: select

$a_m=\arg\max_k L(\Theta_{m-1},\bar{v}_k^{m-1}),$

allocate

$$p_m= \frac{\epsilon \bar{P}_m}{1-(1-\epsilon)^{M-m+1}},$$

and choose

$$r_m=\frac{NT}{DM}\log_2\left(\left(\frac{p_m}{N}\right)^D\theta_m\right).$$

The second is predictive convex optimization with state augmentation. For flexible loads with local QoS, the difficulty is that rate and energy constraints have memory and the feasible set depends on the previous state. The main technical device is to introduce a fictitious memory state,

$z^i_t \triangleq [d^i_{t-1|t}, (x^i_t)^T]^T,$

so that the time-varying aspects move into the objective while the constraints become fixed (Coffman et al., 2020). The resulting update is a projected-gradient-style distributed iteration,

$$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$0

and the analysis yields a global ISS guarantee under the explicit step-size condition

$$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$1

The third is predictive Lyapunov optimization over a multi-frame window. In integrated cellular and Wi-Fi networks, ENSRA solves a one-frame drift-plus-penalty problem, while P-ENSRA uses perfect prediction over $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$2 frames and adds the predictive control parameter $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$3 to value earlier service within the window (Yu et al., 2015). The central predictive drift relation includes the negative queue-pressure term

$$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$4

This formulation makes prediction part of queue shaping, not merely of PHY adaptation.

The fourth is finite-blocklength mapping from predicted SINR to radio resources. In URLLC, predicted interference $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$5 yields predicted SINR $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$6, which is inserted into

$$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$7

with $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$8, so that the scheduler computes the minimal number of channel uses $$\mathbb{X}_{k,m+1}= \begin{cases} \mathbb{X}_{k,m}\cap \{X_k\ge \theta_m\}, & v_{k,m}=1\ \mathbb{X}_{k,m}\cap \{X_k< \theta_m\}, & v_{k,m}=0 , \end{cases}$$9 needed to satisfy the target error probability (Jayawardhana et al., 2023). The predictive component is therefore directly operational: interference prediction changes the PHY/MAC allocation before transmission.

A fifth motif appears in radio-map-assisted aerial networks. In the low-altitude single-commodity problem, the inner resource-allocation subproblem minimizes the worst-case interference leakage while guaranteeing delivery of a data package. The optimal power policy has the structure

$g_{\rm th}^*$0

and the paper proves that the throughput function $g_{\rm th}^*$1 is strictly increasing in $g_{\rm th}^*$2, which enables bisection-based global optimization of the predictive resource allocation subproblem (Li et al., 25 Jul 2025).

4. System families and representative realizations

In OFDM and UWB systems, cross-layer allocation typically couples channel abstraction at PHY with service differentiation at MAC. WiMedia UWB uses an effective SINR mapping,

$g_{\rm th}^*$3

to compress per-subcarrier SINRs into a scalar channel-state metric, then combines this with service weights $g_{\rm th}^*$4 in an allocation-level function $g_{\rm th}^*$5 for distributed sub-band assignment (0907.3793). A related UWB formulation under heterogeneous constraints uses the cross-layer metric $g_{\rm th}^*$6, where $g_{\rm th}^*$7 is a MAC-layer weight and $g_{\rm th}^*$8 is PHY-layer effective SINR, followed by class-dependent interference control for HQoS and SQoS users (Khalil et al., 2010). In OFDMA NRT/BE traffic, the ATO approach uses current CSI and long-term throughput requirements $g_{\rm th}^*$9 to defer service within a window $\nu^*$0 and exploit time diversity (Ukil et al., 2011).

In systems without explicit CSIT, prediction can arise purely from protocol feedback. The FDD-OFDM ACK/NAK design adapts power, rate, and user allocation causally from feedback flows already present in the link layer, and the scheduled rate converges asymptotically to the true capacity that would have been selected with perfect CSIT as more feedback arrives (0905.4700). This is a cross-layer scheduler whose predictive state is the feedback-induced uncertainty interval.

In energy-aware wireless delivery, future large-scale channel information supports predictive scheduling even when future small-scale fading is unavailable. The delay-tolerant download problem is cast as joint slot scheduling and power allocation with sleep/active mode decisions, and the main conclusion is that future large-scale channel knowledge is almost sufficient to realize the main energy-saving gain of full future-channel knowledge (Yao et al., 2016). In integrated cellular and Wi-Fi networks, the same delay-tolerant logic appears at a larger architectural scale: future Wi-Fi coverage, channels, and traffic demand allow deferral of traffic away from expensive cellular transmission (Yu et al., 2015).

For URLLC, the predictive cross-layer coupling is between interference estimation and finite-blocklength scheduling. The EMD-based hybrid predictor decomposes the observed interference sequence into intrinsic mode functions and a residual, forecasts each component with LSTM or ARIMA, reconstructs the predicted interference, and then computes channel uses from the finite-blocklength expression (Jayawardhana et al., 2023). For THz mesh backhaul, DEFLECT separates routing and resource allocation: a heuristic distance-aware routing metric is followed by a DRL-based long-term resource allocator with a multi-task structure for joint power and sub-array allocation and a hierarchical architecture for node-specific control and recovery from broken links (Hu et al., 2023).

Aerial and low-altitude networks introduce the most explicit cross-layer hierarchy. One formulation treats predictive communication as a layered control framework with strategic routing, tactical timing and local rerouting, and operational power control, driven by 4D mission trajectories, radio maps, and instantaneous local measurements (Chen et al., 1 Sep 2025). A second formulation converts joint routing and predictive resource allocation into a joint bottleneck path planning and resource allocation problem on a dynamic space-time graph, then extends it to multi-commodity transmission tasks through time-frequency allocation (Li et al., 25 Jul 2025).

Learning-based implementations have also entered the area. In MEC-aided cell-free MIMO-OFDMA downlink, joint subcarrier allocation and beamforming are reformulated as a multi-task self-supervised learning problem, leading to CMTSSL, MEC-enabled DMTSSL, and the distance-aware transfer learning algorithm DATL (Zheng et al., 2024). Although this formulation is not a forecasting model in the temporal sense, it predicts cross-layer decisions from channel state while enforcing objective and constraint structure through self-supervised losses.

5. Empirical outcomes, gains, and tradeoffs

Reported gains depend strongly on the source of predictive information and on the traffic regime. In WiMedia UWB, the proposed cross-layer scheme gives the hard-QoS user about a $\nu^*$1 dB gain at BER $\nu^*$2 relative to WiMedia with TFC in the $\nu^*$3 Mbit/s comparison, while the soft-QoS users perform close to the baseline; in the four-user case, the remaining sub-band is shared between the two soft-QoS users and their performance is slightly degraded compared with the single-user baseline (0907.3793). In CSIT-free FDD-OFDM, the ACK/NAK-driven scheduler achieves about $\nu^*$4 of the perfect-CSIT upper bound and gives $\nu^*$5 throughput gain over round-robin scheduling that ignores feedback (0905.4700). In flexible-load resource allocation, the predictive distributed method tracks a BPA balancing reserve deployed signal with prediction horizon $\nu^*$6 and reports a $\nu^*$7-norm tracking error of $\nu^*$8 (Coffman et al., 2020). In integrated cellular and Wi-Fi networks, GP-ENSRA reduces the traffic delay over ENSRA by $\nu^*$9 under the same power consumption, and a representative point reports $W$0 W versus $W$1 W at average delay $W$2 s (Yu et al., 2015).

The dependence on delay tolerance is consistent across multiple systems. The ATO scheduler benefits when the service window length grows; for RT-like hard delay-bound traffic, $W$3, it does not help much (Ukil et al., 2011). The predictive cellular/Wi-Fi gains are weaker when the system must operate with very low delay, because traffic must be served immediately regardless of future favorable conditions (Yu et al., 2015). The large-scale-channel analysis explicitly assumes long deadlines, and the main benefit arises from choosing when to transmit and how much power to use over the horizon (Yao et al., 2016).

Prediction quality also matters differently across domains. In URLLC, EMD-based prediction reduces RMSE by about $W$4, with ARIMA improving from $W$5 to $W$6 and LSTM improving from $W$7 to $W$8; the resulting resource allocation gives $W$9-$\alpha_m=(p_m,r_m,a_m)$0 orders of magnitude lower outage than the baseline prediction-based method (Jayawardhana et al., 2023). In THz backhaul, the routing metric gives $\alpha_m=(p_m,r_m,a_m)$1 lower resource usage than hop-count routing at $\alpha_m=(p_m,r_m,a_m)$2 dB, while DEFLECT DRL reaches long-term resource-efficiency maximization with no packet loss, millisecond-level latency, and recovery from broken links within $\alpha_m=(p_m,r_m,a_m)$3 s (Hu et al., 2023). In low-altitude UAV relay networks, the single-commodity algorithm approaches global optimality and reports more than $\alpha_m=(p_m,r_m,a_m)$4 dB performance gain over classical graph-based methods for delay-sensitive and large data transportation; the multi-commodity method reports $\alpha_m=(p_m,r_m,a_m)$5X improvements in dense service scenarios and an additional $\alpha_m=(p_m,r_m,a_m)$6 dB performance gain by data segmenting (Li et al., 25 Jul 2025). In the conceptual low-altitude predictive-communication framework, the case study claims more than $\alpha_m=(p_m,r_m,a_m)$7 dB reduction in cross-tier interference (Chen et al., 1 Sep 2025).

These results come with explicit limitations. The WiMedia UWB cross-layer scheme is designed for indoor UWB channels with relatively slow variations, so per-superframe updates may be less suitable for fast-varying environments (0907.3793). The flexible-load formulation provides ISS rather than exact asymptotic tracking, and its error is attributed mainly to infeasibility of the reference with respect to all QoS constraints (Coffman et al., 2020). The predictive cellular/Wi-Fi framework assumes perfect prediction within the window (Yu et al., 2015). The large-scale-channel result relies on frame-level ergodicity and delay-tolerant transmission (Yao et al., 2016). The THz DRL approach requires safe initialization and safe exploration in order to avoid packet loss during learning (Hu et al., 2023). The low-altitude graph model depends on predetermined trajectories and radio-map quality (Li et al., 25 Jul 2025).

6. Misconceptions, boundaries, and unresolved issues

A common misconception is that predictive cross-layer resource allocation necessarily requires long-horizon prediction of instantaneous small-scale CSI. Several strands of work contradict that view. The ACK/NAK-based FDD-OFDM scheme operates without explicit CSIT and predicts only a feasible channel region from feedback history (0905.4700). The large-scale-channel analysis argues that future large-scale channel information captures almost all the performance gain of full future-channel knowledge in long-deadline energy-saving allocation (Yao et al., 2016). The WiMedia UWB scheme updates allocation at each superframe from current effective SINR and current QoS weights rather than from multi-step channel forecasting (0907.3793).

A second misconception is that all cross-layer schemes are predictive. Some influential systems are explicitly cross-layer but not forecast-based. The MPGPS/A-MPGPS/O-MPGPS framework integrates packet scheduling, subcarrier allocation, and power allocation in packet-switched OFDM networks [0611080]. The heterogeneous-constraint UWB design states that it is not predictive in the forecasting sense and instead uses current PHY measurements and current MAC classification (Khalil et al., 2010). This distinction matters because the predictive component changes the mathematical structure: belief-state updates, horizon shifting, ISS analysis, or monotone feasibility are not needed in purely reactive cross-layer designs.

A third misconception is that better prediction alone solves the resource-allocation problem. Many successful papers depend as much on structural reformulation as on predictive input: fixed feasible sets through state augmentation (Coffman et al., 2020), negative queue-pressure terms in predictive Lyapunov drift (Yu et al., 2015), finite-blocklength inversion for URLLC (Jayawardhana et al., 2023), hierarchical decomposition into strategic, tactical, and operational layers (Chen et al., 1 Sep 2025), and explicit deterministic bounds plus monotonicity for globally optimal bisection (Li et al., 25 Jul 2025). This suggests that the practical value of prediction often depends on whether the predictive signal can be translated into a tractable decision variable or a provable structural property.

Several unresolved issues recur across the literature. Distributed operation is desirable for privacy, scalability, or decentralization, but it introduces strong assumptions about exchanged information: WiMedia devices share allocation preferences through beacon-period information elements, flexible loads rely on local projections over fixed sets, and hierarchical low-altitude control relies on increasingly accurate information across central, local, and individual echelons (0907.3793, Coffman et al., 2020, Chen et al., 1 Sep 2025). Learning-based implementations add further questions about dimensional disaster, transfer to dynamic scenarios, and the choice between centralized and distributed training; DMTSSL and DATL are explicit responses to those issues in MEC-aided cell-free networks (Zheng et al., 2024). A plausible implication is that future work will continue to move away from “prediction as a scalar forecast” and toward predictive cross-layer control in which uncertainty, timescale separation, and feasibility geometry are themselves first-class optimization objects.