Hybrid Cellular and Cell-Free Networks

• HCCNs are wireless network architectures that integrate cellular base stations with distributed, user-centric access points to enhance connectivity.
• They employ dynamic cooperation clustering and SDN control to manage interference, optimize resource allocation, and improve network performance.
• Practical considerations such as synchronization, CSI acquisition, and backhaul capacity drive ongoing research in scalable HCCN deployment.

Hybrid Cellular and Cell-Free Networks (HCCNs) are wireless architectures that combine conventional cellular base stations with cell-free access points or user-centric cooperative service, so that legacy cellular infrastructure and cell-free operation coexist within a single network. In the literature, HCCNs are motivated as a practical and feasible solution for advancing cell-free network development, since cell-free access points can be incrementally introduced into existing cellular networks rather than deployed as a network-wide replacement from the outset. The concept also encompasses converged cell-less operation in which user equipment is not persistently attached to a single cell, but is instead served by dynamic cooperative groups of base stations or access points selected for each transmission event (Dai et al., 26 Jul 2025, Han et al., 2016).

1. Conceptual foundations and taxonomy

The cellular and cell-free endpoints of the design space are sharply distinguished in the source literature. Conventional cellular massive MIMO is cell-centric: the network is partitioned into cells, each user equipment is served by one access point or base station except during handover, and inter-cell interference is inherent to the architecture. By contrast, cell-free massive MIMO is a distributed Massive MIMO system in which many geographically separated access points cooperatively and coherently serve user equipments without imposing cell boundaries during active data transmission, typically in a user-centric fashion through dynamic cooperation clusters (Interdonato et al., 2018, Demir et al., 2021).

HCCNs occupy the intermediate region between these endpoints. One formulation treats the hybrid as a network in which the typical user equipment is jointly served by the nearest cellular base station and all cell-free access points in the service area, so that the user benefits simultaneously from the cellular tier and the cell-free tier (Dai et al., 2024, Dai et al., 26 Jul 2025). A second formulation treats the hybrid as clustered cell-free networking, where the network is decomposed into disjoint subnetworks; within each subnetwork, joint processing is adopted for intra-subnetwork interference mitigation, while subnetworks operate independently and inter-subnetwork interference is treated as colored noise (Xia et al., 2024). A third formulation is heterogeneous massive MIMO, which retains a co-located base station array at the center of each cell and augments it with distributed edge access points at cell boundaries or dead spots (Jiang et al., 27 Jun 2025).

The converged cell-less model extends the taxonomy further. There, horizontal convergence dissolves per-cell association among base stations and access points, while vertical convergence orchestrates heterogeneous technologies such as macro, micro, femto, WLAN, mmWave, and IoT tiers through an SDN control plane. The user equipment does not maintain a persistent association to a single base station or access point; association is implicit and event-driven, and mobility is handled by reforming cooperative groups rather than executing horizontal or vertical handovers (Han et al., 2016).

A common misconception is that hybridization requires network-wide participation of all access points for every user. The literature does not support such a single model. It includes nearest-BS-plus-all-AP service, dynamic user-centric subsets, small cooperative groups such as two or three nodes, and disjoint clustered subnetworks, each representing a distinct operating point in the complexity–performance space (Dai et al., 26 Jul 2025, Interdonato et al., 2018, Han et al., 2016).

2. Architectural realizations

In stochastic-geometry HCCN models, the cellular and cell-free layers are superposed. Base stations, access points, and user equipments are modeled as independent homogeneous PPPs with intensities $\lambda_B$, $\lambda_A$, and $\lambda_U$, respectively, over a finite service region. The typical user equipment is associated with its nearest base station through Voronoi association and is simultaneously served by all access points in the cell-free layer. Base stations and access points share the same time-frequency resource, and conjugate beamforming is employed so that the desired signals from the two tiers add coherently at the receiver (Dai et al., 2024, Dai et al., 26 Jul 2025).

In clustered realizations, the network is partitioned into $M$ disjoint subnetworks indexed by $s \in \{1,\dots,M\}$, each with a set of user equipments $\mathcal{U}_s$ and base stations $\mathcal{B}_s$. Inside each subnetwork, all base stations jointly serve or detect all user equipments in that subnetwork; across subnetworks, operations are independent. This architecture is explicitly described as an architectural hybrid between conventional cellular networks and fully cell-free networks, because it yields cellular-like isolation between subnetworks and cell-free-like cooperation within each subnetwork (Xia et al., 2024).

The heterogeneous massive MIMO realization is more cell-centric. Each cell contains a central base station with $N_b$ co-located antennas and $L_c$ distributed edge access points. TDD reciprocity is assumed, the central base station acts as the CPU for the edge access points, and baseband processing is fully centralized within each cell. This preserves a cellular skeleton while adding distributed antennas where edge coverage is weakest (Jiang et al., 27 Jun 2025).

The converged cell-less architecture introduces an SDN cloud composed of SDN controllers plus core routers as the control plane, while routers and instantaneous backhaul links constitute the data plane. The SDN controller configures cooperative base station or access point groups for each transmission event and dynamically instantiates backhaul or midhaul connectivity. Downlink data can be multicast to the cooperative group to reduce backhaul load, and uplink data can be jointly decoded after nearby receivers forward their observations to the cloud (Han et al., 2016).

The control-plane and user-plane can also be split. In one architecture option, a macro-cellular layer handles initial access, synchronization signals, random access, paging, broadcast and system information, and mobility anchoring, while cell-free clusters deliver user-plane data through user-centric joint transmission. This arrangement preserves wide-area compatibility and mobility anchoring while allowing the data plane to exploit cell-free cooperation where it is beneficial (Interdonato et al., 2018).

3. Signal, interference, and analytical models

A representative downlink HCCN SINR model writes the received SINR at the typical user equipment as

$$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$

where $\lambda_A$0 is the aggregate desired signal from the nearest base station and the cell-free access points, $\lambda_A$1 is intra-cell cellular interference, $\lambda_A$2 is inter-cell cellular interference, and $\lambda_A$3 is the mean AP-origin interference under the large-population approximation (Dai et al., 26 Jul 2025, Dai et al., 2024).

Under conjugate beamforming, the desired term is approximated as

$\lambda_A$4

with the aggregate AP desired contribution

$\lambda_A$5

valid for $\lambda_A$6. Since the exact distribution of the shifted-Nakagami square is unwieldy, $\lambda_A$7 is approximated by a Gamma random variable through moment matching, with shape and scale

$\lambda_A$8

Coverage probability is then characterized through Laplace transforms and higher-order derivatives of the interference terms, and the resulting expressions are semi-closed-form (Dai et al., 26 Jul 2025).

The interference model is structurally important because it reveals coupling between signal and interference. Intra-cell interference $\lambda_A$9 and the desired term $\lambda_U$0 are coupled via the serving-base-station channel, so the average achievable rate analysis introduces $\lambda_U$1 and again uses Gamma moment matching. The average achievable rate is written as

$\lambda_U$2

which makes the mutual coupling explicit in the replacement of $\lambda_U$3 by $\lambda_U$4 (Dai et al., 26 Jul 2025).

Clustered HCCNs use a different analytical framework. The uplink subnetwork ergodic capacity with colored inter-subnetwork noise is

$\lambda_U$5

and the network sum ergodic capacity is $\lambda_U$6. In large systems, the capacity is approximated by pathloss-weighted diagonal terms, which motivates graph-based clustering criteria (Xia et al., 2024).

The user-centric cell-free literature contributes the signal-processing backbone for HCCNs. In the uplink, the centralized spectral efficiency is

$\lambda_U$7

with MMSE combining, partial MMSE, MR, or distributed LSFD-based alternatives. In the downlink, user-centric superposition is expressed through

$\lambda_U$8

where $\lambda_U$9 implements dynamic cooperation clustering by turning AP–UE links on or off (Demir et al., 2021).

4. Cooperation, clustering, and resource management

Resource management is the principal mechanism by which HCCNs move between cellular-like and cell-free-like behavior. In converged cell-less networks, the SDN controller selects the $M$0-nearest base stations or access points as candidates for an active user equipment and forms a small cooperative group subject to simplicity, economy, and uniformity criteria. The literature explicitly states that one terminal per base station or access point is preferred when possible, that as few nodes as needed should be selected to meet rate or SINR targets, that uplink and downlink groups should be kept consistent when feasible, and that multicast should be used for downlink delivery to the cooperative group. Pre-grouping caches recent grouping solutions, while mobility prediction and neighbor awareness are used to avoid hot-spot congestion (Han et al., 2016).

In user-centric cell-free massive MIMO, dynamic cooperation clustering is formalized by the set $M$1 of access points serving user $M$2, with

$M$3

The literature emphasizes scalable clustering rules such as received-power-based and channel-quality-based selection, and in one architecture option a 95%-subset rule is used so that only the strongest access points serve a user, limiting fronthaul without major spectral-efficiency loss (Demir et al., 2021, Interdonato et al., 2018).

Clustered HCCNs cast network decomposition as an optimization problem. In one formulation, the objective is to maximize sum ergodic capacity subject to the joint processing constraint $M$4 for every subnetwork. By introducing membership vectors and the Laplacian of the weighted UE–BS bipartite graph, the objective becomes an integer convex program that minimizes $M$5 under affine partition constraints. The branch-and-bound method solves the full problem, while the BC$M$6F-Net algorithm recursively bisects the largest subnetwork and solves a sequence of two-way integer convex programs to reduce complexity (Xia et al., 2024).

Rate-constrained decomposition offers another clustering mechanism. RC-NetDecomp models a mmWave beamspace network as a bipartite graph between users and beams, normalizes edge weights by the strongest beam per user, and maximizes the number of subnetworks subject to a per-user rate lower bound. The rate constraint is converted into an explicit sum-cut constraint,

$M$7

and the largest feasible $M$8 is found by binary search combined with spectral clustering on a reduced meganode graph (Wang et al., 2022).

Power control and fairness are also integral. The user-centric cell-free framework states max-min fairness as

$M$9

subject to linear power constraints, and sum spectral-efficiency maximization as

$s \in \{1,\dots,M\}$0

These formulations are solved through bisection, fixed-point methods, or weighted MMSE reformulations, depending on the utility function and processing model (Demir et al., 2021).

5. Performance characteristics and trade-offs

Across the stochastic-geometry analyses, increasing access-point density generally improves both coverage and average achievable rate because macro-diversity and shorter AP–UE distances strengthen the cell-free contribution. At the same time, AP-origin interference also grows, and this produces a non-monotonic dependence on AP transmit power. For the default parameters $s \in \{1,\dots,M\}$1, $s \in \{1,\dots,M\}$2, $s \in \{1,\dots,M\}$3, $s \in \{1,\dots,M\}$4, $s \in \{1,\dots,M\}$5, $s \in \{1,\dots,M\}$6 dBm, $s \in \{1,\dots,M\}$7, $s \in \{1,\dots,M\}$8, and service-area radius $s \in \{1,\dots,M\}$9 m, coverage increases with $\mathcal{U}_s$0 for any SINR threshold, while average rate first increases and then decreases as $\mathcal{U}_s$1 grows. For $\mathcal{U}_s$2, $\mathcal{U}_s$3–$\mathcal{U}_s$4 dBm yields higher average rate than $\mathcal{U}_s$5 dBm, which identifies an interference-limited regime at high AP power (Dai et al., 26 Jul 2025).

The earlier hybrid downlink model reaches a consistent conclusion. With a finite disk of radius $\mathcal{U}_s$6 m, $\mathcal{U}_s$7, $\mathcal{U}_s$8, $\mathcal{U}_s$9, $\mathcal{B}_s$0, $\mathcal{B}_s$1, $\mathcal{B}_s$2, and $\mathcal{B}_s$3, the analytical coverage closely matches Monte Carlo simulations and shows that the hybrid architecture boosts edge coverage relative to cellular-only operation while preserving higher peak SINR than pure cell-free operation, because the many short-range AP contributions strengthen weak users and the high-power macro base station retains strong center-user performance (Dai et al., 2024).

The converged cell-less model emphasizes handover avoidance and energy savings. In a $\mathcal{B}_s$4 m $\mathcal{B}_s$5 $\mathcal{B}_s$6 m area with $\mathcal{B}_s$7 base stations uniformly distributed, $\mathcal{B}_s$8 base stations configured as active members, the nearest $\mathcal{B}_s$9 candidates considered for grouping, and group size limited to at most $N_b$0, Monte Carlo curves show higher coverage probability than conventional cellular across the entire threshold range from $N_b$1 to $N_b$2 dB. On the energy side, with $N_b$3 active base stations and the rest sleeping, two-base-station groups achieve the largest base-station-side energy saving relative to three- or four-base-station groups, while uplink joint reception yields increasing user-equipment energy saving as the number of cooperative receivers grows (Han et al., 2016).

Cell-free massive MIMO results provide the uniform-service benchmark against which HCCNs are often judged. Channel-gain CDFs show $N_b$4–$N_b$5 dB gains for disadvantaged users when access points are dense, and max-min power control roughly doubles the 95%-likely spectral efficiency compared to channel-dependent full-power transmission, reaching approximately $N_b$6 bit/s/Hz/UE in both indoor and outdoor deployments. The hybrid message is not that all of these gains require a fully cell-free system, but that user-centric cooperation and macro-diversity are the dominant ingredients behind the edge-performance improvements (Interdonato et al., 2018).

Clustered HCCN studies quantify the complexity–performance compromise. BC$N_b$7F-Net is reported to operate within at most $N_b$8 of the full integer convex program, to reduce runtime to as low as $N_b$9 of the full branch-and-bound solution, and to outperform state-of-the-art clustering benchmarks with up to $L_c$0 capacity gain. In a different graph-partitioning formulation, RC-NetDecomp is reported to outperform existing baselines in average per-user rate, fairness among users, and energy efficiency by constructing weakly interfered subnetworks under an explicit rate constraint (Xia et al., 2024, Wang et al., 2022).

Heterogeneous massive MIMO gives a concrete cost-oriented HCCN instance. In a $L_c$1 km $L_c$2 $L_c$3 km area divided into four $L_c$4 m $L_c$5 $L_c$6 m cells, with $L_c$7 service antennas and $L_c$8 user equipments in total, HmMIMO uses per cell $L_c$9 co-located base-station antennas and $$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$0 edge access points with $$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$1 antennas each. Relative to CFmMIMO, the total number of distributed radio heads is halved, and the paper reports roughly a $$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$2 reduction in fronthaul cost while attaining comparable worst-case spectral efficiency and a middle ground between CFmMIMO uniformity and CmMIMO peak-rate advantage (Jiang et al., 27 Jun 2025).

6. Practical considerations, limitations, and research directions

The feasibility of HCCNs is governed by synchronization, CSI acquisition, fronthaul or backhaul capacity, and scalable control. Coherent joint transmission requires tight time, frequency, and phase synchronization across cooperating nodes; non-coherent combining reduces these requirements but also reduces array gain. Dynamic backhaul instantiation and multicast are repeatedly proposed as mechanisms to contain transport overhead, and several works argue that cooperative group sizes should remain small or local unless the fronthaul is sufficiently rich to support more centralized processing (Han et al., 2016, Interdonato et al., 2018).

The analytical models are deliberately simplified. The stochastic-geometry coverage and rate analyses assume perfect local CSI, independent Rayleigh fading, and no pilot contamination, mobility, correlated shadowing, or backhaul fairness constraints. Their AP aggregate approximations require $$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$3 for the desired term and $$\Omega = \frac{S_0}{I_{B0} + I_B + \bar I_A + \sigma^2},$$4 for AP interference. The clustered optimization papers are largely uplink-centric and do not jointly optimize downlink precoding and power allocation. The heterogeneous massive MIMO study does not explicitly model fronthaul capacity limits, quantization, latency, or hardware impairments (Dai et al., 26 Jul 2025, Dai et al., 2024, Xia et al., 2024, Jiang et al., 27 Jun 2025).

Open problems therefore cluster around the interfaces between physical-layer cooperation and network control. The literature explicitly identifies cooperative backhaul design, protocols for cross-technology control and data separation, synchronization and coherent joint processing at city scale, edge/cloud partitioning of decoding and grouping logic, scalable access-point clustering and selection, power control under fronthaul constraints, pilot contamination and imperfect CSI, correlated shadowing and spatial user clustering, uplink analysis and joint uplink–downlink optimization, and multi-tenant or multi-operator HCCNs as unresolved directions (Han et al., 2016, Dai et al., 26 Jul 2025, Interdonato et al., 2018).

Taken together, these works present HCCNs less as a single canonical topology than as a family of intermediate architectures between cell-centric and fully cell-free operation. This suggests that the defining feature of an HCCN is not a specific geometry, but the coexistence of a cellular control or infrastructure substrate with user-centric cooperative service, realized through nearest-BS-plus-AP overlays, clustered subnetworks, heterogeneous antenna splits, or SDN-controlled cell-less convergence.