RIS-Aided Cell Breathing Techniques

• RIS-CB is a technique that uses passive, reconfigurable intelligent surfaces to reshape radio coverage by dynamically controlling reflected links.
• It integrates methods like AP sleep-mode control, cell zooming, user association, and advanced algorithms (e.g., PPO, DCCN) to optimize network performance.
• Empirical results demonstrate significant gains in energy efficiency, delay reduction, and improved handover management in diverse network architectures.

RIS-aided Cell Breathing (RIS-CB) denotes the dynamic expansion or shrinkage of effective radio coverage through reconfigurable intelligent surface reconfiguration, typically combined with sleep-mode control, cell zooming, user association, or handover management. In the recent literature, the term is instantiated in three closely related ways: as an energy-efficient CF-mMIMO transmission scheme that selectively deactivates access points (APs) while nearby RISs preserve coverage (Luo et al., 13 Dec 2025), as a multi-cell RAN strategy that combines Advanced Sleep Modes (ASM), three zooming levels, user association, proximal policy optimization (PPO), and a double cascade correlation network (DCCN) under a delay constraint (Sun et al., 2024), and as a mobility-management mechanism in which near-field (NF) focusing and far-field (FF) steering retain connectivity with the serving base station (BS) and reduce unnecessary handovers by using bit error rate (BER) as the primary decision variable (Mondal et al., 29 Jul 2025). Across these formulations, the common principle is that coverage adaptation is performed not only through transmitter activation or power adjustment, but also through passive, low-power wavefront shaping.

1. Conceptual definition and research variants

RIS-CB generalizes classical cell breathing by allowing the network to reshape the propagation environment itself. In conventional cellular systems, breathing primarily relies on BS transmit-power adaptation or BS on/off switching. In RIS-CB, a reconfigurable intelligent surface supplies controllable reflected links, so coverage holes created by sleeping APs or overloaded cells can be compensated through phase reconfiguration rather than only through higher radiated power. This is the central idea in the CF-mMIMO formulation, where the CPU jointly orchestrates AP activation, RIS phases, and power allocation so that the coverage footprint “breathes” with current load (Luo et al., 13 Dec 2025).

In the multi-cell RAN formulation, RIS-CB is explicitly coupled to cell zooming. Zooming in reduces cell radius to shed load; zooming out expands coverage to accept offloaded users. The paper combines this with multi-depth sleep states—Active, Idle, Micro sleep (SM1), Light sleep (SM2), and Deep sleep (SM3)—so that traffic-aware coverage adaptation and sleep-mode control are optimized together (Sun et al., 2024).

In the mobility-oriented formulation, RIS-CB is not primarily an energy-minimization routine but a handover-retention mechanism. As a UE approaches a cell edge or the serving link degrades, the RIS extends the serving BS’s usable coverage by creating a high-gain reflected path, often in the RIS near-field, thereby reducing the need for hard or soft handover and restricting measurement reports and HO requests (Mondal et al., 29 Jul 2025).

Reference	Setting	Primary controls
(Luo et al., 13 Dec 2025)	Multi-RIS-aided CF-mMIMO	AP sleep, user powers, RIS phases
(Sun et al., 2024)	RIS-aided multi-cell RAN	ASM state, zooming, user association, RIS phases
(Mondal et al., 29 Jul 2025)	RIS-assisted mobility management	HHO/SHO/RIS-CB/RIS-PP, NF/FF RIS configuration

2. Network architectures and signal models

The CF-mMIMO formulation considers a downlink network with $M$ single-antenna APs, $K$ single-antenna UEs, and $L$ RISs, each RIS $l$ having $N$ reflecting elements and being coordinated by a central CPU. AP activity is represented by $s_m \in \{0,1\}$, with active set $\mathcal{A}=\{m:s_m=1\}$. For RIS $l$, the phase matrix is diagonal,

$$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$

with element-wise unit-modulus constraint. The effective composite AP–UE channel via multi-RIS is

$$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$

and $K$0. Under imperfect CSI, $K$1, where $K$2 depends on the RIS configuration through the RIS–UE covariance. With AP activity mask $K$3, the ZF precoder is

$K$4

and the SINR under imperfect CSI becomes

$K$5

where $K$6 captures error-induced interference (Luo et al., 13 Dec 2025).

The multi-cell RAN formulation considers $K$7 single-antenna BSs, $K$8 single-antenna users, and one shared RIS with $K$9 reflecting elements. Sleep mode states satisfy

$L$0

while zooming levels satisfy

$L$1

with coverage radii $L$2 m, $L$3 m, and $L$4 m, respectively. User association obeys $L$5 and $L$6. The effective BS–user channel is

$L$7

where $L$8 and $L$9 for $l$0-bit quantized phases (Sun et al., 2024).

The mobility formulation models two BSs, one RIS with $l$1 passive elements, and a moving UE. It explicitly distinguishes NF and FF operation. The RIS reflection matrix is

$l$2

with $l$3 and $l$4. In the NF, the reflected field at UE location $l$5 is represented by the spherical-wave sum

$l$6

whereas in the FF the RIS acts as an array producing uniform plane-wave beams. The paper also uses the cascaded-channel model

$l$7

and studies BER, outage probability, and capacity as functions of NF/FF RIS operation (Mondal et al., 29 Jul 2025).

3. Optimization and control mechanisms

In the CF-mMIMO setting, RIS-CB is formulated as an energy-efficiency maximization problem over binary AP selection, per-user powers, and RIS phases: $l$8 subject to binary activity, per-AP radiated power limits, unit-modulus RIS constraints, QoS constraints $l$9, and per-user power limits. The fractional objective is handled with the Dinkelbach transformation,

$N$0

and each Dinkelbach step is solved by alternating optimization over three blocks: AP selection using hybrid branch-and-bound (BnB) plus greedy turn-off, transmit power allocation using sequential convex approximation (SCA), and RIS phase optimization using gradient projection under unit-modulus constraints. A lower-complexity alternative replaces gradient projection with Whale Optimization Algorithm (WOA) (Luo et al., 13 Dec 2025).

In the multi-cell RAN setting, the control problem is modeled as a Markov decision process. The state is

$N$1

where $N$2 is pending traffic load, $N$3 is the channel-coefficient matrix, $N$4 is the current sleep-mode state vector, and $N$5 is the current time index. The action is

$N$6

namely the next sleep states, zooming levels, and user association matrix. PPO is used to optimize the long-term control policy, with actor $N$7, critic $N$8, generalized advantage estimation

$N$9

and the clipped surrogate

$s_m \in \{0,1\}$0

RIS phases are optimized on a shorter timescale by DCCN: a first network maps channel state $s_m \in \{0,1\}$1 to RIS coefficients $s_m \in \{0,1\}$2, and a second network maps $s_m \in \{0,1\}$3 to capacity $s_m \in \{0,1\}$4 (Sun et al., 2024).

In the mobility-oriented framework, RIS-CB is embedded in a handover algorithm that includes hard handover (HHO), soft handover (SHO), RIS-aided cell breathing (RIS-CB), and RIS-aided ping-pong avoidance (RIS-PP). The decision process is BER-centric. If $s_m \in \{0,1\}$5, direct candidates are evaluated for HHO; if $s_m \in \{0,1\}$6, SHO may be initiated; if the serving-cell load satisfies $s_m \in \{0,1\}$7 and a RIS-aided path satisfies $s_m \in \{0,1\}$8, RIS-CB is activated to retain the serving BS; and if the serving BER lies in an $s_m \in \{0,1\}$9-margin around $\mathcal{A}=\{m:s_m=1\}$0, RIS-PP uses the RIS to avoid ping-pong. The paper defines selection rules such as

$\mathcal{A}=\{m:s_m=1\}$1

with the RIS path chosen to maximize this probability (Mondal et al., 29 Jul 2025).

4. Performance criteria and operating objectives

The CF-mMIMO RIS-CB formulation uses spectral efficiency

$\mathcal{A}=\{m:s_m=1\}$2

and total power

$\mathcal{A}=\{m:s_m=1\}$3

where AP, fronthaul, CPU, and RIS terms are all modeled explicitly. Energy efficiency is then

$\mathcal{A}=\{m:s_m=1\}$4

A distinctive feature is that sleeping APs reduce both fixed and rate-dependent consumption, while passive RIS panels contribute only small static power. The design directly accounts for MMSE-estimation error through the stochastic interference term $\mathcal{A}=\{m:s_m=1\}$5, so robustness is expressed in an average sense rather than by a worst-case uncertainty set (Luo et al., 13 Dec 2025).

The multi-cell PPO formulation optimizes BS energy consumption under a delay constraint. System energy over the horizon is

$\mathcal{A}=\{m:s_m=1\}$6

where each BS energy depends on dwell times in Active, Idle, SM1, SM2, and SM3 as well as deactivation and reactivation times. Delay is defined as the time from arrival of traffic item $\mathcal{A}=\{m:s_m=1\}$7 at a BS to completion of transmission to its user, subject to $\mathcal{A}=\{m:s_m=1\}$8. The paper states explicitly that the method may achieve energy-savings at the cost of increased delay, requiring a trade-off between these two factors, and therefore gives higher reward weights to delay terms so that delay satisfaction is prioritized before energy minimization (Sun et al., 2024).

The mobility-oriented formulation treats BER as the key link-quality variable. In AWGN for BPSK, the average BER is

$\mathcal{A}=\{m:s_m=1\}$9

By restricting measurement reports and HO requests, the algorithm seeks improvements in spectrum efficiency and energy efficiency, especially in crowded cellular networks. In this version of RIS-CB, the principal operational objective is not a direct fractional EE maximization but a reduction in HO rates and signaling load while ensuring seamless connectivity and improved QoS (Mondal et al., 29 Jul 2025).

5. Reported empirical behavior

The reported results span energy efficiency, delay-aware energy consumption, convergence, and mobility KPIs.

Reference	Simulation setting	Reported outcome
(Luo et al., 13 Dec 2025)	$l$0, $l$1	EE gains over CF-mMIMO without RIS are $l$2 at $l$3, $l$4 at $l$5, $l$6 at $l$7, and $l$8 at $l$9
(Sun et al., 2024)	$$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$0, $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$1, $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$2	Energy consumption is reduced by $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$3 versus DSZR; proposed PSZR with SM1+SM2+SM3 achieves $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$4 savings versus AA
(Mondal et al., 29 Jul 2025)	$$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$5, $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$6 m	HO trigger distance increases by about $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$7 at $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$8 dB; the gain is smaller, about $$\boldsymbol{\Phi}_l=\operatorname{diag}\!\big(e^{j\theta_{l,1}},\dots,e^{j\theta_{l,N}}\big),$$9, at $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$0 dB

For CF-mMIMO, Dinkelbach plus alternating optimization converges in approximately $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$1 iterations, and the low-complexity solution based on greedy selection, SCA, and WOA achieves approximately $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$2 of the energy efficiency of the near-optimal BnB, SCA, and gradient-projection solution. The paper also reports that EE first rises and then falls with AP density, that EE increases with RIS count up to $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$3 and then declines due to interference and static RIS overhead, and that sleep-mode yields the largest single-component EE gain. RIS optimization provides approximately $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$4 EE gain in sparse scenarios $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$5 but only approximately $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$6 in denser scenarios $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$7, where strong direct AP–UE links diminish the relative value of RIS paths.

For the multi-cell PPO system, the training reward reaches approximately $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$8 at $$g_{mk}=h_{\mathrm{au},mk}+\sum_{l=1}^{L}\mathbf{h}_{\mathrm{ar},ml}^{T}\boldsymbol{\Phi}_l\mathbf{h}_{\mathrm{ru},lk},$$9 iterations, whereas the DQN benchmark reaches approximately $K$00. Energy savings increase with deeper ASM support: the reported savings versus the always-active baseline are about $K$01 for PZR with SM1, about $K$02 for PZR with SM1+SM2, and about $K$03 for PSZR with SM1+SM2+SM3. Energy increases with packet size and user count, decreases as inter-arrival times become sparser, and PSZR consistently yields the smallest or near-smallest number of active BSs per slot. The paper attributes this to RIS-aided breathing that minimizes active infrastructure while respecting delay constraints.

For NF/FF mobility management, the numerical figures show that BER, outage probability, and capacity improve significantly with the number of RIS elements, especially in NF operation. The NF focal beam has very narrow 3 dB bandwidth and depth, which increases the “retention corridor” in which the UE can remain connected to the serving BS. The paper further reports that HHO probability modestly increases with $K$04, while SHO probability decreases as $K$05 grows because the RIS-aided serving link is already strong. The reported overall effect is significant reductions in HO rates and signaling load together with improved SE and EE.

6. Practical constraints, misconceptions, and open problems

A recurrent misconception is to equate RIS-CB with ordinary power-control breathing. The cited works do not support that reduction. In the CF-mMIMO and mobility formulations, RIS-CB specifically means that coverage is adapted through passive reflections, so the environment itself is reconfigured rather than only the transmitter power budget. This distinction is central to the claim that RIS-CB can maintain rates or BER margins with much smaller energy overhead than traditional breathing (Luo et al., 13 Dec 2025, Mondal et al., 29 Jul 2025).

Another misconception is that RIS-CB is equivalent to simply adding an RIS to an always-on network. The results do not support that either. In the CF-mMIMO study, the decisive gain comes from joint AP sleep, power allocation, and RIS phase control; in the multi-cell study, the decisive gain comes from the joint action space of sleep state, zooming, user association, and RIS coefficients. A plausible implication is that RIS-CB is fundamentally a cross-layer coordination problem rather than a purely electromagnetic one.

The practical constraints are substantial. The CF-mMIMO paper emphasizes imperfect CSI, MMSE estimation error, and the need for centralized computation, while also noting that exact BnB AP selection has exponential worst-case complexity and that ultra-dense networks may require further decomposition or learning-based selection. The multi-cell PPO study assumes model-based channels, orthogonal subcarriers, single-antenna nodes, and perfect channel knowledge in simulations; it also notes that CSI acquisition for cascaded BS–RIS–UE channels introduces overhead, although per-slot DCCN inference is lightweight. The mobility paper stresses phase quantization, mutual coupling, element losses, controller latency, synchronization, and stability issues near the NF/FF boundary, where hysteresis and time-to-trigger are needed to avoid rapid toggling (Luo et al., 13 Dec 2025, Sun et al., 2024, Mondal et al., 29 Jul 2025).

The open problems stated in the literature include discrete RIS phase optimization, multi-bit control, joint uplink–downlink RIS-CB, multi-antenna UE extensions, user scheduling and stream management under latency constraints, and worst-case robust optimization under bounded channel uncertainties and hardware impairments. In the mobility domain, NF focusing also raises fairness and interference-management questions because strong focal support for one UE may reduce resources for others. Taken together, these works suggest that RIS-CB has become a unifying label for a family of techniques in which sleep control, coverage adaptation, and electromagnetic shaping are jointly optimized, but the exact formulation depends strongly on whether the dominant objective is energy efficiency, delay-constrained operation, or mobility robustness.