Distributed Carrier Aggregation Overview

• Distributed Carrier Aggregation is a technique that aggregates independent RF carriers across cells and network layers to expand effective bandwidth and mitigate interference.
• It integrates analytical models with dynamic scheduling, load balancing, and distributed power control, ensuring efficient resource utilization and fair load distribution.
• Applications span from HSPA+/LTE-A to 5G-Advanced and satellite systems, demonstrating measurable gains in throughput, energy efficiency, and operational flexibility.

Searching arXiv for recent and foundational papers on distributed carrier aggregation to ground the article in published work. Carrier Aggregation (CA) denotes the association of two or more independent RF carriers to or from a single user terminal, creating a wider aggregate channel. In HSPA+ and LTE-A, CA allows devices to be served simultaneously by several carriers; in 5G-Advanced, the aggregated downlink and uplink component carriers need no longer reside in the same 3GPP band nor in a one-to-one DL–UL pairing; and in satellite systems, the aggregated carriers may traverse different physical satellites and orbits. Distributed Carrier Aggregation refers to CA realizations in which the relevant control variables are distributed across carriers, cells, teams of base stations, beams, O-RUs, or orbital assets, so that aggregation is achieved through coordinated load balancing, power setting, scheduling, transmitter switching, beam illumination, or stream merging rather than through a single static carrier assignment alone (Bénézit et al., 2014, Han et al., 2023, Gonzalez-Rios et al., 26 Aug 2025, Cai et al., 30 Mar 2026).

1. Analytical foundations

A foundational queueing formulation models a cell with two downlink carriers of peak capacities $C_1$ and $C_2$, two user classes, and spatially heterogeneous radio conditions. SC users are served on exactly one carrier, whereas DC users may be served simultaneously on both carriers. SC flows arrive according to a Poisson process of rate $\alpha$, DC flows according to $\mathrm{Poisson}(\beta)$, sessions carry an i.i.d. exponential volume with mean $\sigma$, and the cell is divided into $J$ areas with per-area peak rates $C_{1,j}$ and $C_{2,j}$. Under Processor-Sharing, if there are $n_1$ users on Carrier 1, $n_2$ users on Carrier 2, and $C_2$0 DC sessions in the system, then a user in area $C_2$1 receives

$C_2$2

The state is a $C_2$3-dimensional vector $C_2$4, and throughput is obtained by Little’s law as

$C_2$5

The stability condition uses the offered load

$C_2$6

with $C_2$7 necessary, and under JFQ+VB essentially sufficient, for ergodicity (Bénézit et al., 2014).

A complementary stochastic-geometry formulation treats CA-enabled heterogeneous cellular networks as $C_2$8 tiers of base stations distributed as independent homogeneous PPPs $C_2$9 of density $\alpha$0, with UEs forming an independent PPP of density $\alpha$1. The system has $\alpha$2 disjoint component carriers, band $\alpha$3 has bandwidth $\alpha$4 and path-loss exponent $\alpha$5, and the load-aware activity of interferers is captured by

$\alpha$6

Single-flow CA is modeled by a UE selecting one tier $\alpha$7 through biased association and then aggregating all bands used by that tier. In the interference-limited single-tier case,

$\alpha$8

and the peak data rate is independent of BS density and transmit powers. This formulation also distinguishes orthogonal deployment from universal cochannel deployment and quantifies the trade-off between spectrum addition and small-cell densification (Lin et al., 2012).

Taken together, these models isolate two recurrent structural issues. The first is intra-cell aggregation under shared service disciplines and state-dependent scheduling; the second is multi-tier or multi-band aggregation under spatial interference and load coupling. This suggests that distributed CA is best understood as a family of resource-coupling problems with different state variables but closely related control objectives.

2. Inter-carrier load balancing and carrier pooling

In the two-carrier queueing model, distributed CA is implemented by two inter-carrier load-balancing rules. For SC sessions, Join the Fastest Queue (JFQ) compares the instantaneous throughput if a new SC flow were assigned to Carrier 1 or Carrier 2: $\alpha$9 The new flow joins the carrier with the larger value, and a tie is broken with probability $\mathrm{Poisson}(\beta)$0. For DC sessions, Volume Balancing (VB) splits a remaining volume $\mathrm{Poisson}(\beta)$1 into $\mathrm{Poisson}(\beta)$2 on Carrier 1 and $\mathrm{Poisson}(\beta)$3 on Carrier 2 so as to equalize completion times over intervals where the state is constant: $\mathrm{Poisson}(\beta)$4 Equivalently, a DC session behaves as a single PS server of capacity

$\mathrm{Poisson}(\beta)$5

The transition rates of the underlying Markov process then follow directly from Poisson arrivals and PS departures, and the stationary distribution $\mathrm{Poisson}(\beta)$6 is the unique stationary solution of the global balance equations, although no closed form is explicit (Bénézit et al., 2014).

The key quantitative results are explicit in several regimes. In the equal-capacity case $\mathrm{Poisson}(\beta)$7, pure DC traffic yields $\mathrm{Poisson}(\beta)$8, pure SC traffic under JFQ degenerates to Join-the-Shortest-Queue and yields $\mathrm{Poisson}(\beta)$9 at high load, and the gain of CA over single-carrier PS is a $\sigma$0 gain both at light load and at heavy load. In the unequal-capacity case $\sigma$1, pure DC traffic yields $\sigma$2, while pure SC traffic under JFQ satisfies numerically $\sigma$3, showing that SC flows are steered to the faster carrier. The HSPA+ Dual-Band example $\sigma$4 Mb/s and $\sigma$5 Mb/s gives pure DC throughput $\sigma$6 and pure SC throughput approximately $\sigma$7. Mixed SC/DC traffic is reported as quasi-insensitive to the traffic mix $\sigma$8: both $\sigma$9 and $J$0 remain nearly constant as $J$1 varies, even though the overall cell throughput $J$2 increases almost linearly with the fraction of DC traffic (Bénézit et al., 2014).

The associated design rules are operational rather than purely analytical. Quasi-ideal balancing is achieved if the cell load $J$3, a single joint scheduler is used across carriers to implement JFQ for SC arrivals and VB for DC sessions, and each carrier enforces fair PS-like resource sharing such as round-robin or proportional/max-min. The implementation recommendations are equally concrete: SC user assignment is performed at session setup from current queue sizes and projected per-carrier rates; DC volume splitting is continuously adapted in proportion to instantaneous per-carrier rates; and a sliding-window average of link rates can stabilize VB splitting under fading. A common misconception is that CA performance is highly sensitive to the fraction of CA-capable users; in this model, the performance is practically insensitive to the SC/DC proportion, and dimensioning can therefore be based on pure DC or pure SC worst-cases (Bénézit et al., 2014).

3. Distributed power control in heterogeneous and dense networks

A second major line of work treats distributed CA as a per-carrier power-setting problem in interference-coupled heterogeneous networks. In a two-tier LTE HetNet, the macro layer consists of high-power eNodeBs and the micro layer of low-power small cells underlaid within each macrocell. CA-capable UEs may be simultaneously scheduled on $J$4 component carriers; the cited study fixes $J$5 with central frequencies $J$6 GHz, $J$7 GHz, $J$8 GHz, each of bandwidth $J$9 MHz, and each base station chooses its downlink power on each CC from the discrete set $C_{1,j}$0 of fractions of its maximum. The game is played by “teams,” each consisting of one macro BS plus the micro BSs in its macrocell, with strategy vector $C_{1,j}$1 and inter-team interference

$C_{1,j}$2

For a UE in tile $C_{1,j}$3, served by location $C_{1,j}$4 on carrier $C_{1,j}$5,

$C_{1,j}$6

The payoff is defined as $C_{1,j}$7, where the utility is a normalized sigmoid of SINR and the total cost $C_{1,j}$8 is the sum of a power-cost $C_{1,j}$9 and a coverage-cost $C_{2,j}$0, with $C_{2,j}$1 the fraction of UEs with SINR below $C_{2,j}$2. The resulting best-reply dynamics form a pseudo-potential game, pure-strategy Nash equilibria exist, and iterative best replies are guaranteed to converge in a finite number of steps (Fazliu et al., 2016).

The numerical results quantify the role of CA as an interference-mitigation primitive rather than merely a bandwidth-expansion mechanism. In a 57-team scenario, the Best-Reply Power Setting (BPS) scheme outperforms fixed-power and eICIC/LP-ABS baselines, with up to $C_{2,j}$3 higher sum-utility versus eICIC/LP-ABS, total network transmit power approximately that of the min-power strategy, and only $C_{2,j}$4 unserved UEs. The convergence speed is reported as an average of 6–10 best-reply iterations per CC per team, independent of network size. The same framework is explicitly stated to generalize to other multi-carrier and 5G scenarios, including sub-6 GHz plus mmWave, cloud-RAN clusters, and gNB-CU/DU splits (Fazliu et al., 2016).

A related dense-network formulation considers cells rather than teams as the players. Each cell $C_{2,j}$5 chooses a power vector $C_{2,j}$6, where $C_{2,j}$7, and the downlink SINR of user $C_{2,j}$8 on carrier $C_{2,j}$9 is

$n_1$0

The payoff combines a sigmoid-like utility, a per-carrier power-cost $n_1$1, and a coverage penalty $n_1$2. With $n_1$3, the game is one of strategic substitutes with aggregation and admits a pseudo-potential, so best-response dynamics converge to a pure-strategy NE. In the reported simulations, the distributed BPS solution yields $n_1$4–$n_1$5 higher total downloaded data and demand met in peak periods, $n_1$6 fewer failed downloads, macro-cell energy efficiency $n_1$7, micro-cell energy efficiency $n_1$8–$n_1$9, average user throughput increase of $n_2$0–$n_2$1, and RB-usage efficiency gains of $n_2$2–$n_2$3 at macros and $n_2$4–$n_2$5 at micros, all relative to a full-power + eICIC baseline (Fazliu et al., 2017).

A recurring misconception is that CA by itself suppresses interference. These power-control results show the opposite: when carriers have distinct budgets and propagation profiles, CA becomes a control surface on which interference management is performed, and the gain depends on explicit optimization of per-carrier power, coverage, and utility rather than on aggregation alone.

4. Flexible spectrum orchestration in 5G-Advanced and O-RAN

In 5G-Advanced Release 18, flexible spectrum orchestration generalizes CA by decoupling the historical assumption that DL and UL component carriers belong to the same band or to a fixed pairing. A serving cell is identified by a cell index $n_2$6 and may be DL-only, UL-only, or bi-directional. The PCell in the MCG always carries SSB/PRACH/PDCCH, while SCells in either MCG or SCG may be activated or deactivated dynamically and may carry only UL or only DL. Spectrum pooling allows multiple narrow-band carriers, possibly in different bands, to be pooled into a single logical cell for DL or UL. The architecture imposes concrete synchronization and control constraints: when enabling an SCell without SSB, the residual timing difference relative to a reference cell must be at most $n_2$7 ns; the received power difference between carriers must be at most $n_2$8 dB; and QCL relations are specified between an SCell’s TRS, the SCell’s RS, and an SSB of a reference cell. Each DL or UL carrier may be split into up to four BWPs, and different BWPs may be chosen per cell to match traffic load (Han et al., 2023).

The formal scheduling layer is equally explicit. In the multi-cell scheduling problem, $n_2$9 indicates whether UE $C_2$00 is scheduled on RB $C_2$01 of cell $C_2$02, and the sum-rate objective is

$C_2$03

subject to per-RB exclusivity, per-cell power budgets, fairness or minimum-throughput constraints, and a DCI-size limit

$C_2$04

This yields a practical maximum of $C_2$05 cells in one DCI. The same framework also formulates transmitter switching through binary activity indicators $C_2$06 and an energy–latency minimization over time, with the per-interval switching rule

$C_2$07

Reported Release 18 performance figures include mean UE throughput gains of $C_2$08 and $C_2$09 for two UL Tx switching frameworks with 5 UEs, a small difference of $C_2$10 between the frameworks, and BS energy savings of $C_2$11–$C_2$12 from the SSB-less SCell scheme over resource utilization from $C_2$13 to $C_2$14 (Han et al., 2023).

O-RAN introduces a further distribution layer in which multi-CC scheduling must be coordinated across cooperating O-RUs, especially for JT users. In the cited MU-MIMO O-RAN formulation, $C_2$15 indicates whether UE $C_2$16 in cell $C_2$17 is scheduled on RBG $C_2$18 of component carrier $C_2$19, and JT consistency requires

$C_2$20

After extending eigen-based zero-forcing transceiver design to JT users and deriving a tractable separable rate approximation using massive-MIMO asymptotics, the problem is solved by a centralized BCD benchmark and a distributed scheduler aligned with the O-RAN architecture. The distributed method uses three stages: local per-core optimization, PU$C_2$21 coordination for JT-UE consistency and QoS, and a final NJT-UE refinement. Only one round of inter-PU exchange is required. In the reported results, the distributed scheduler achieves $C_2$22 of the centralized ESR while preserving $C_2$23, whereas a no-coordination variant drops below $C_2$24 satisfaction under higher QoS load. At $C_2$25, the centralized PCS requires $C_2$26 s, $C_2$27 s, and $C_2$28 s of CPU time, while the distributed PDS requires $C_2$29 s, $C_2$30 s, and $C_2$31 s, respectively (Cai et al., 30 Mar 2026).

These developments show that distributed CA in 5G-Advanced is no longer confined to “aggregating carriers” in a narrow frequency-domain sense. It now includes flexible DL/UL cell formation, multi-cell DCI design, UL transmitter switching, energy-aware SCell activation, and cross-cell consistency constraints for JT.

5. Satellite, multi-beam, and multi-orbit realizations

In high-throughput satellite systems, distributed CA is coupled to beam hopping and to the spatial overlap of adjacent beams. A joint BH-CA formulation uses beams $C_2$32, users $C_2$33, carriers $C_2$34, and time-slots $C_2$35. The principal decision variables are $C_2$36 for user-carrier assignment, $C_2$37 for fill-rate allocation, $C_2$38 for beam illumination, and $C_2$39 as the linearization helper for $C_2$40. The supplied capacity to user $C_2$41 is

$C_2$42

and the optimization simultaneously maximizes the minimum per-user supply/demand ratio and the minimum per-beam supply/demand ratio, subject to aggregation limits, fill-rate constraints, illumination limits, adjacency interference constraints, and assignment–allocation coupling. By introducing $C_2$43, replacing the max–min by

$C_2$44

and linearizing $C_2$45 with the $C_2$46-variables, the original MINLP is converted into an MILP solvable by Gurobi or CPLEX (Kibria et al., 2022).

The reported BH-CA results emphasize rate matching and fairness. In the reference setup, a GEO system has $C_2$47 beams, $C_2$48 users per beam, $C_2$49 carriers per beam of $C_2$50 MHz each, $C_2$51, $C_2$52 transponders, $C_2$53 ms, and $C_2$54. For a total demand of $C_2$55 Mbps, conventional BH supplies $C_2$56 Mbps, while joint BH-CA supplies $C_2$57 Mbps; unused capacity is approximately $C_2$58 Mbps under BH-CA versus $C_2$59 Mbps under BH; and user fairness $C_2$60 is approximately $C_2$61–$C_2$62 in every beam under BH-CA versus $C_2$63–$C_2$64 under BH alone. The summary explicitly states that distributed CA naturally emerges when regions optimize locally which carriers to share and a top-level scheduler enforces $C_2$65 (Kibria et al., 2022).

A separate line of work experimentally validates CA above the PHY layer in a multi-orbit satellite testbed. The CADSAT in-lab demonstrator consists of a Gateway module, a multi-orbit channel emulator, and a User Terminal module. The GW includes a Traffic Generator, a Load Balancing & PDU Scheduler, GSE Encapsulation & DVB-S2 PHY Framing, and Adaptive Buffering. The UT contains two parallel SDR chains, GSE de-encapsulation, and a FIFO Merge Buffer. The scheduler uses the load-balancing factor

$C_2$66

with Carrier 1 chosen as the dominant link, and in GEO+MEO scenarios prepends an initial prefix on the faster MEO link to compensate for the differential propagation delay $C_2$67, where

$C_2$68

In the example $C_2$69 dB, $C_2$70PSK $C_2$71, $C_2$72, $C_2$73 MHz, and $C_2$74 ms, one obtains $C_2$75 and $C_2$76 PDUs. Across GEO, MEO, and mixed GEO/MEO test campaigns, round-robin scheduling yields mean misplacement $C_2$77 PDUs and max misplacement $C_2$78 PDUs in the GEO-RR case, whereas the proposed load-balancing scheduler limits mean misplacement to $C_2$79, $C_2$80, $C_2$81, and $C_2$82 PDUs in GEO, MEO, MEO$C_2$83GEO, and GEO$C_2$84MEO cases, respectively, with aggregate throughput approximately $C_2$85 Mbps in all cases (Gonzalez-Rios et al., 26 Aug 2025).

These satellite results also correct a common misunderstanding. Distributed CA need not imply PHY-level signal combining: the testbed uses no MRC and no cross-carrier LO sharing, and the UT complexity is limited to per-carrier decoding plus FIFO-based stream merging.

6. Deployment trade-offs, performance regimes, and recurring misconceptions

Several recurrent trade-offs govern distributed CA across the terrestrial and satellite literature. In load-aware HetNets, universal cochannel deployment typically yields the largest rate, while orthogonal deployment incurs a capacity loss that can be reduced by appropriately tuning biasing factors $C_2$86. In the comparison between 1-band/$C_2$87-tier spatial reuse and $C_2$88-band/1-tier spectrum addition, the proposition states that if $C_2$89, then small-cell densification outperforms pure spectrum addition. The same analysis reports that CA provides super-linear peak gains if bands differ in path-loss exponents and recommends aggregating bands with higher $C_2$90 last (Lin et al., 2012).

In queueing-based terrestrial CA, the dominant operating condition is $C_2$91, together with a single joint scheduler and PS-like sharing. Under those assumptions, DC sessions achieve full aggregated capacity, SC sessions are steered to the currently fastest carrier, and mixing CA and non-CA traffic has negligible mutual impact. A misconception that unequal carriers invalidate effective aggregation is not supported by the model: when $C_2$92, JFQ steers SC traffic toward the larger-capacity carrier, while VB still lets DC users exploit $C_2$93 (Bénézit et al., 2014).

In interference-limited HetNets, another common misconception is that peak-rate scaling can be obtained simply by densifying the network or raising transmit power. The stochastic-geometry analysis shows the opposite in the single-tier interference-limited case: the peak data rate is independent of BS density and transmit powers, which strongly motivates other approaches such as CA to increase the peak data rate (Lin et al., 2012).

In 5G-Advanced and dense networks, distributed CA is also tightly linked to energy management and control signaling. Multi-cell scheduling gain saturates as DCI size approaches the polar-code limit of $C_2$94 bits, UL Tx switching is bounded by the maximum number of monitored or simultaneously transmitted bands, and SSB-less SCells rely on wake-up signaling as an energy-saving fallback. These considerations indicate that the effective unit of aggregation is no longer only the RF carrier but the jointly scheduled, jointly controlled, and sometimes jointly deactivated cell-carrier pair (Han et al., 2023).

In satellite systems, the main trade-off is between flexible resource pooling and delay-induced packet reordering. The in-lab results show that a carefully designed gateway scheduler and modest buffering are sufficient to keep reordering within a few PDUs, but also that differential propagation delays in multi-orbit CA require explicit prefix scheduling and buffer dimensioning (Gonzalez-Rios et al., 26 Aug 2025). This suggests that the term “distributed” changes its technical content across domains: in terrestrial RANs it is often interference- and scheduler-centric, whereas in SatCom it is often topology- and delay-centric.

Across these domains, the consistent pattern is that distributed CA is not a single standardized algorithm but a design space defined by how aggregation is coupled to load, interference, coordination latency, and control-plane structure. The cited work supports three broad conclusions: carrier pooling can be made close to ideal under explicit stability and scheduling conditions; per-carrier control is a principal mechanism for interference and energy management; and the scope of CA has expanded from intra-cell bandwidth aggregation to multi-cell, multi-beam, and multi-orbit resource orchestration (Bénézit et al., 2014, Fazliu et al., 2016, Fazliu et al., 2017, Kibria et al., 2022, Han et al., 2023, Gonzalez-Rios et al., 26 Aug 2025, Cai et al., 30 Mar 2026).