Radio spectrum is both scarce and expensive. Traditionally, wireless networks prevent interference by carving up the spectrum into orthogonal slices—giving different frequencies or time slots to different users. However, this approach leaves large amounts of capacity on the table.
Underlay spectrum sharing breaks this paradigm by allowing secondary networks (such as direct device-to-device links, small cells, or unmanned aerial vehicles) to reuse the exact same frequency bands simultaneously occupied by primary cellular users. The catch is that these secondary signals act as immediate co-channel noise to the primary network, and vice versa. As established by Arash Asadi et al. in their foundational survey, A Survey on Device-to-Device Communication in Cellular Networks, the central challenge of this architecture is balancing the massive spectral efficiency gains of underlay communication with the strict requirement to protect the primary network's Quality of Service (QoS).
To make underlay spectrum sharing work, networks must dynamically manage this mutual interference. This is achieved through a combination of spatial geometry, algorithmic coordination, physical-layer engineering, and robust machine learning.
The Geometry of Local Interference
Interference is fundamentally a local, spatial constraint rather than a global one. If two secondary device pairs are sufficiently far apart, they can safely reuse the same frequency band without drowning each other out.
To exploit this mathematically, researchers often map wireless networks to graph structures. In a framework proposed by Yi Li et al. in Joint Mode Selection and Resource Allocation for D2D Communications via Vertex Coloring, the network is modeled as an interference graph where nodes represent communication links and edges connect links that would destructively interfere. This transforms the complex problem of joint channel and mode selection into a vertex-coloring problem, allowing the network to rapidly isolate conflicting links and maximize frequency reuse with low computational overhead.
[D2D Pair A] <--- (Interference Edge) ---> [D2D Pair B]
| |
Color 1 Color 2
(Channel 1) (Channel 2)
In multi-hop or ad-hoc scenarios, these spatial constraints become even more complex because channel qualities are often initially unknown. Yaqin Zhou et al. addressed this in Almost Optimal Channel Access in Multi-Hop Networks With Unknown Channel Variables by converting channel selection into a Maximum Weighted Independent Set (MWIS) problem. By applying a robust polynomial-time approximation scheme, they demonstrated that networks can achieve near-optimal throughput without falling victim to the exponential complexity typically caused by unknown environmental variables.
Furthermore, real-world interference is not static; it is governed by the temporal dynamics of data traffic. Jianfang Xin et al. combined stochastic geometry with queuing theory in Performance Analysis of Cellular Networks with D2D communication Based on Queuing Theory Model. They showed that because real-world data packets arrive randomly and fill finite buffers, the resulting interference environment fluctuates dynamically based on both the spatial density of active devices and their instantaneous queue states.
Cross-Tier Coordination and Game Theory
In heterogeneous networks (HetNets), underlay spectrum sharing creates a multi-tiered interference environment. For example, a macrocell base station must coexist with smaller femtocells or device-to-device (D2D) pairs. This creates two distinct types of conflict:
- Inter-tier interference: Small cells or D2D transmitters disrupting primary macrocell uploads or downloads.
- Intra-tier interference: Nearby secondary devices clashing with each other on the shared channel.
To mitigate cross-tier conflicts without relying on a heavy, slow centralized controller, researchers turn to decentralized coordination mechanisms. One approach is predictive overhearing. In Shadow chasing : A resource allocation scheme for heterogeneous networks, Ahmed R. Elsherif et al. developed a "Shadow Chasing" protocol. By combining outdated network control data with real-time "overhearing" of user feedback, secondary femtocells can probabilistically predict which radio frequencies are currently safe to use, preventing them from stepping on primary macrocell transmissions.
When multiple secondary users must compete for the same pool of shared channels, the interaction can be modeled using game theory. In these scenarios, each device acts self-interestedly to maximize its own data rate, yet the system must guide them toward a socially optimal equilibrium:
- Potential Games: Demia Della Penda et al. demonstrated in Distributed Channel Allocation for D2D-Enabled 5G Networks Using Potential Games that modeling the network as a potential game allows micro-cells and D2D pairs to negotiate channel selection locally via message-passing, converging rapidly to a stable, high-throughput state.
- Coalitional Games: In A coalitional graph game framework for network coding-aided D2D communication, Yulei Zhao et al. utilized coalitional graph games to allow devices to form cooperative transmission groups, optimizing relay selection and spectrum allocation simultaneously.
- Matching Theory with Externalities: A key challenge in underlay networks is that one user's channel selection alters the interference floor for everyone else—a phenomenon known as an externality. To solve this, Tao Yang et al. proposed a two-stage framework in A Cascaded Channel-Power Allocation for D2D Underlaid Cellular Networks Using Matching Theory. They paired matching theory with a virtual pricing mechanism, forcing secondary links to "pay" a penalty based on the amount of interference they cause to primary users, effectively forcing power levels into a safe, cooperative bound.
This multi-tier overlap also introduces security vulnerabilities, as secondary links can inadvertently expose primary transmissions to eavesdropping. Kecheng Zhang et al. addressed this in Secrecy-Optimized Resource Allocation for Device-to-Device Communication Underlaying Heterogeneous Networks by transforming non-convex interference equations into convex sub-problems, maximizing the secure data capacity (secrecy rate) of the underlay network.
Physical Layer Engineering: MIMO and Smart Surfaces
While algorithmic scheduling avoids interference logically, physical-layer engineering manipulates the radio environment itself to nullify or bypass interference entirely.
Cognitive Interference Alignment
If the primary network utilizes Orthogonal Frequency Division Multiplexing (OFDM), secondary users can exploit the inherent mathematical redundancy of the signal. In Cognitive interference alignment for OFDM two-tiered networks, Marco Maso et al. derived a semi-unitary linear precoding strategy called Cognitive Interference Alignment (CIA). This technique allows secondary small cells to pre-code their signals so that they land precisely within the unused dimensions (the null space) of the macro-cell’s receiver, rendering the secondary transmissions completely invisible to the primary network.
Massive MIMO and Edge Vulnerabilities
Deploying large antenna arrays at the base station allows for spatial beamforming, which focuses radio energy into narrow beams to isolate users geographically. Serveh Shalmashi et al. analyzed this interaction in Energy Efficiency and Sum Rate Tradeoffs for Massive MIMO Systems with Underlaid Device-to-Device Communications, proving that Massive MIMO can significantly blunt the impact of underlay interference.
However, as noted by Howard H. Yang et al. in Energy-Efficient Design of MIMO Heterogeneous Networks with Wireless Backhaul, co-channel deployment still demands active inter-tier management. This is especially true for users at the cell edge, who possess weak desired signals and are easily overwhelmed by nearby underlay transmissions (Yashuai Cao et al.).
Reconfigurable Intelligent Surfaces (RIS)
To rescue these vulnerable cell-edge users, modern architectures deploy Reconfigurable Intelligent Surfaces (RIS)—metasurfaces acting as "smart mirrors" for radio waves.
[Secondary Transmitter] ------> (Direct Interference) ------> [Primary Receiver]
\ ^
\---> [ RIS (Optimized Phase Shifts) ] ----------------/
(Destructive Cancellation Beam)
By dynamically adjusting the discrete phase shifts of the elements on an RIS, the network can reflect an incoming signal such that it arrives at the primary receiver exactly $180^\circ$ out of phase with the direct interference path, destructively canceling out the co-channel noise.
- Yali Chen et al. demonstrated in Reconfigurable Intelligent Surface Assisted Device-to-Device Communications that jointly optimizing transmission power and RIS phase shifts yields up to a 64.6% boost in total system data rates.
- Yashuai Cao et al. expanded this to multi-RIS environments using an alternating direction method of multipliers (RM-ADMM) to coordinate user pairing and phase profiles simultaneously.
- Nguyen Thị Thanh Van et al. demonstrated in Intelligence Reflecting Surface-Aided Integrated Data and Energy Networking Coexisting D2D Communications that this spatial control can be extended to simultaneously manage high-speed data streams, wireless power transfer, and D2D links without drowning the primary network.
When the spatial path remains too degraded, physical helpers can step in. Networks can deploy amplify-and-forward helper stations (Jihoon Chung et al.) or leverage cooperative peer-to-peer relaying among moving nodes (Hongliang Zhang et al.) to boost weak signals above the co-channel interference floor.
Adapting to Uncertainty via Machine Learning
Traditional optimization models assume perfect, instantaneous knowledge of the radio environment, known as Channel State Information (CSI). In reality, CSI is often fuzzy, delayed, or incomplete. As emphasized by Safiu A. Gbadamosi et al. in Adaptive Resource Allocation and Mode Switching for D2D Networks With Imperfect CSI in AGV-Based Factory Automation, relying on imperfect CSI makes traditional interference management fragile, particularly in metal-heavy, high-interference environments like smart factories.
To prevent performance collapse without resorting to overly conservative "worst-case" margins that waste spectrum, modern underlay networks embed machine learning directly into their resource allocation loops.
Learning the Noise Topology
Instead of guessing the channel noise, networks can learn its geometric distribution. In Learning-Based Robust Resource allocation for D2D Underlaying Cellular Network, Weihua Wu et al. utilized Support Vector Clustering (SVC) to map out the precise boundaries of channel uncertainty. This machine learning layer allows the power control algorithm to operate right up to the edge of maximum capacity while mathematically guaranteeing that primary users remain protected from unexpected interference spikes.
Deep Reinforcement Learning (DRL) for Dynamic Environments
When nodes are moving—such as in mobile edge computing (MEC) or drone-based networks—the interference environment changes too quickly for mathematical re-derivation. In these settings, Deep Reinforcement Learning (DRL) allows the network to learn by trial and error:
[ Environment State (CSI, Queues) ] ---> [ DRL Agent ]
^ |
| v
[ Reward / Penalty ] <-------- [ Action (Power, Channel) ]
- Mobile Edge Computing: In Resource Allocation in Information-Centric Wireless Networking With D2D-Enabled MEC: A Deep Reinforcement Learning Approach, Dan Wang et al. utilized a Policy Gradient DRL framework to dynamically manage channel selection and continuous power levels, balancing local data offloading with strict signal-to-interference requirements.
- UAV Trajectory Optimization: When the transmitters are airborne drones, their physical movement continuously alters the cross-tier interference mapping. Fanyi Wu et al. developed a multi-agent DRL approach in UAV-to-Device Underlay Communications: Age of Information Minimization by Multi-agent Deep Reinforcement Learning. The DRL agents continuously adjust the flight paths and transmission powers of the UAVs, minimizing data latency (Age of Information) while preventing the airborne signals from blinding ground-based cellular receivers.
Overcoming Latency Constraints
A final frontier in underlay sharing is the decoding delay. When high-speed mobile broadband (eMBB) and ultra-reliable low-latency (URLLC) traffic share the same frequency resources, traditional orthogonal slicing fails. Andrea Matera et al. demonstrated in Non-Orthogonal eMBB-URLLC Radio Access for Cloud Radio Access Networks with Analog Fronthauling that by shifting to Non-Orthogonal Multiple Access (NOMA) combined with edge-based successive interference cancellation, networks can decode overlapping underlay streams instantly, satisfying strict latency constraints without sacrificing raw spectral throughput.
Key Foundational Work
Shadow chasing : A resource allocation scheme for heterogeneous networks (2012) — contributed 1 concept elements to this direction.
A Survey on Device-to-Device Communication in Cellular Networks (2013) — contributed 1 concept elements to this direction.
Cognitive interference alignment for OFDM two-tiered networks (2012) — contributed 1 concept elements to this direction.