Cell DTX/DRX: Energy Saving in 5G Networks

• Cell DTX/DRX is a discontinuous-operation mechanism that selectively mutes transmission and/or reception to reduce energy consumption in cellular systems.
• Studies show that BS-side DTX can reduce base station power consumption by shifting from a high standby power (P0) to lower sleep power (Ps), achieving up to 73% savings at low loads.
• Advanced control strategies using prediction, RL, and wake-up signaling optimize DTX/DRX performance by balancing energy savings with latency and QoS requirements.

Cell Discontinuous Transmission and Reception (Cell DTX/DRX) denotes a family of discontinuous-operation mechanisms in cellular systems in which transmission, reception, or both are deliberately suspended during selected intervals to reduce energy consumption. In the narrow 5G NR Release 18 sense, cell DTX/DRX is a Layer-2, time-domain mechanism at the gNB that aggregates user transmissions into active periods and leaves the remaining time silent so that Advanced Sleep Modes (ASM) can be enabled (Mao et al., 28 Jul 2025). In the broader research literature, the same theme also encompasses BS-side DTX, UE-side DRX, wake-up-radio schemes, and distributed time-slot muting methods, all governed by the same central trade-off: longer inactive intervals reduce power but increase latency, non-reachability, or interference-coupling complexity (Holtkamp et al., 2013).

1. Scope, terminology, and architectural locus

The literature distinguishes sharply between BS-side cell DTX and UE-side DRX, even though both are discontinuous-operation mechanisms. BS-side cell DTX is a network-energy mechanism: the cell transmitter, or more generally the radio/baseband chain associated with the cell, is muted or put into a lower-power state during selected intervals. UE-side DRX is a device-energy mechanism: the UE stops monitoring the PDCCH except during configured on-durations, thereby reducing receiver duty cycle (Azari et al., 2021, Herrería-Alonso et al., 2015). Release 18 cell DTX/DRX belongs to the former category and is explicitly designed to enable RU shutdown and ASM at the gNB (Mao et al., 28 Jul 2025).

The main research lineages can be organized as follows.

Mechanism	Locus	Primary control variables
Cell DTX / cell DTX/DRX	BS / gNB	$T_{\text{cycle}}, T_{\text{on}}, t, \beta$
UE DRX	UE	$$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$
Wake-up radio / wake-up signaling	UE with BTS/gNB control	$t_c, t_{on}, t_{of}, N_w$ or MAC CE timing

In UE DRX, a UE has four radio activity states—data transmission, data reception, active, and sleep—and the dominant inefficiency is often time spent in the active state waiting for possible data rather than actual TX/RX time (Azari et al., 2021). In cell DTX, the analogous inefficiency is the large static BS supply-power component, which makes always-on operation wasteful even when RF transmit power is low (Holtkamp et al., 2013). This parallel explains why the same conceptual vocabulary recurs at both layers.

A recurrent source of confusion is to treat all DRX/DTX work as timer tuning. That description is incomplete. The literature includes timer-based DRX, BS sleep-state modeling, random BS muting in stochastic-geometry networks, predictive DRX selection with LSTM or ARIMA, wake-up radio architectures with dedicated PDWCH, MAC-CE-based RL control, and, more recently, contextual-bandit and DQN methods for Release 18 cell DTX/DRX configuration (Azari et al., 2021, Rostami et al., 2019, Pastore et al., 2024, Mao et al., 28 Jul 2025).

2. Fundamental power–delay structure and early analytical limits

A canonical analytical treatment appears in the two-user downlink model of “Fundamental Limits of Energy-Efficient Resource Sharing, Power Control and Discontinuous Transmission” (Holtkamp et al., 2013). There, a BS serves two downlink users over shared bandwidth $W$, with resource sharing factor $\mu \in [0,1]$, AWGN links, and a minimum guaranteed link spectral efficiency $\varsigma_{\min}$. The key step is the embedding of transmit-power optimization into an affine BS supply-power model with an explicit sleep state: $P_{\text{PM}} = P_0 + m P_{\text{Tx}},$ and, including DTX,

$$P_{\text{supply}} = (1-t)P_s + t(P_0 + mP_{\text{Tx}}),$$

where $t$ is the BS activity factor, $P_s$ is sleep-mode power, $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$0 is standby or zero-load supply power, and $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$1 is the slope relating RF transmit power to supply power (Holtkamp et al., 2013).

Under the simulation setup reported there—2 GHz carrier, 250 m cell radius, 10 MHz bandwidth, $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$2 W, $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$3, $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$4 W, $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$5 dBm, and 10,000 random user drops—the optimal combined resource-sharing, power-control, and DTX scheme yields a 5.5 dB supply-power gain at low load and 2 dB at high load relative to a conventional SOTA BS, corresponding to about 73% and 27% energy saving, respectively (Holtkamp et al., 2013). At low load, ON/OFF DTX and the fully optimized RS-PC-DTX curve show almost no difference, whereas at higher rates RS-PC-DTX gains about 1 dB over ON/OFF DTX. The interpretation given in the paper is explicit: at low load the static term $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$6 dominates, so sleep time is far more valuable than fine-grained power control; at high load, DTX opportunities shrink and power control becomes materially relevant (Holtkamp et al., 2013).

That result established a theme that remains central in later work: cell sleep is the dominant energy-saving lever at low to moderate load, whereas transmit-power control, scheduling, or other refinements mainly matter once the cell is already active most of the time. The same paper is equally explicit about its limits. It assumes exactly one sleep mode, instantaneous entry and exit, no switching overhead, no wake-up delay, no maximum DTX frequency, and no explicit LTE control-channel constraints. The reported savings are therefore upper bounds under idealized assumptions rather than deployment figures (Holtkamp et al., 2013).

A related misconception is that power-control optimality automatically implies supply-power optimality. The affine power model shows why this is false: if $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$7 is large, reducing $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$8 by a few watts has negligible effect relative to suspending operation long enough to move from $$T_\text{I}, T_\text{sc}, T_\text{lc}, T_\text{on}, N_\text{sc}$$9 toward $t_c, t_{on}, t_{of}, N_w$0 (Holtkamp et al., 2013).

3. UE DRX as the reception-side counterpart

In connected-mode DRX, the standard parameterization is given by the inactivity timer $t_c, t_{on}, t_{of}, N_w$1, short cycle $t_c, t_{on}, t_{of}, N_w$2, long cycle $t_c, t_{on}, t_{of}, N_w$3, on-duration $t_c, t_{on}, t_{of}, N_w$4, and number of short cycles $t_c, t_{on}, t_{of}, N_w$5 before transition to long DRX (Azari et al., 2021). After data activity, the UE remains continuously active for $t_c, t_{on}, t_{of}, N_w$6; if no packet arrives before expiry, it enters short DRX, and after $t_c, t_{on}, t_{of}, N_w$7 short cycles it moves to long DRX. This architecture is well known to improve UE battery life, but it inherently creates a delay–energy trade-off because packets arriving while the UE sleeps must be buffered until the next on-duration (Herrería-Alonso et al., 2015).

A mathematically explicit formulation of this trade-off is the coalesced DRX model of “Adaptive DRX Scheme to Improve Energy Efficiency in LTE Networks with Bounded Delay” (Herrería-Alonso et al., 2015). There, the eNB buffers packets for a UE in DRX mode and waits until a configurable downstream queue threshold $t_c, t_{on}, t_{of}, N_w$8 is reached before transmission at the next on-duration. Larger $t_c, t_{on}, t_{of}, N_w$9 means more batching, fewer DRX exits, and more UE sleep time, but it also increases the waiting time of the earliest packets in each batch. The paper derives an average-queueing-delay expression using a vacational queueing model and then designs an adaptive controller that adjusts $W$0 online to keep average queueing delay near a target $W$1 while maximizing low-power time (Herrería-Alonso et al., 2015). In Poisson, self-similar, and real video-streaming scenarios, the adaptive scheme keeps average delay close to target over a wide range of loads and yields significant additional UE sleep time relative to standard DRX (Herrería-Alonso et al., 2015).

The more recent IIoT study “Discontinuous Reception with Adjustable Inactivity Timer for IIoT” shifts attention from queue thresholds to inactivity-timer control (Ruíz-Guirola et al., 2024). Its semi-Markov model uses states for active reception, short-cycle on and sleep, and long-cycle on and sleep, and then compares classical IT restart behavior with a BS-controlled rule in which the BS explicitly indicates whether the inactivity timer should be restarted based on its buffer status (Ruíz-Guirola et al., 2024). The paper reports energy-saving gains of up to 30% regardless of arrival rate and delay constraints, with power consumption reduced by about 12–15% absolute, or approximately 30% relative, compared with standard inactivity-timer behavior across traffic rates and TTIs (Ruíz-Guirola et al., 2024). The accompanying delay increase is limited: instantaneous delay excess is reported as < 2 ms, and with $W$2 ms, 98% of packets remain below 15 ms (Ruíz-Guirola et al., 2024). The same work also shows that exhaustive optimization may require more than $W$3 parameter combinations, whereas the GA-based search operates at about $W$4, or approximately 0.89% of exhaustive complexity (Ruíz-Guirola et al., 2024).

These UE-side studies are not cell DTX in the strict BS-sleep sense, but they are the reception-side analog of the same design problem. They also support a broader systems inference: once UE inactivity can be predicted or controlled more sharply, the cell gains more coherent opportunities for transmission silencing.

4. Cell-side DTX under interference, mobility, and distributed coordination

A different research line studies DTX not through BS supply-power models but through its effect on interference fields, retransmissions, and reliability. In “Enhancing Performance of Random Caching in Large-Scale Heterogeneous Wireless Networks with Random Discontinuous Transmission,” each BS is active in a slot with probability $W$5, independently across slots and BSs (Wen et al., 2018). In high mobility, where the interferer geometry is effectively independent across retransmissions, the successful transmission probability is increasing in $W$6, and the optimal solution is $W$7, meaning no DTX. In the static scenario, however, temporal interference correlation limits retransmission gains, and a non-trivial $W$8 emerges because random DTX decorrelates interference while also risking serving-BS inactivity (Wen et al., 2018).

“On Meta Distribution and Local Delay for Cache-Enabled Networks with Random DTX” deepens that result by separating mean STP from local-delay behavior (Yang et al., 2020). There the conditional STP depends on $W$9, and the mean STP is monotonically increasing in $\mu \in [0,1]$0, so a pure mean-STP objective again prefers no DTX. By contrast, the mean local delay exhibits an intermediate optimum $\mu \in [0,1]$1: very small $\mu \in [0,1]$2 makes the serving BS inactive too often, while large $\mu \in [0,1]$3 raises interference and interference correlation, both of which enlarge local delay (Yang et al., 2020). The same paper also shows that network jitter is high near the $\mu \in [0,1]$4 minimizing mean delay and small as $\mu \in [0,1]$5 or $\mu \in [0,1]$6 (Yang et al., 2020). This directly refutes the idea that one scalar objective, such as average STP, is sufficient to configure cell DTX.

A more deployment-flavored formulation appears in “Distributed DTX Alignment with Memory” (Holtkamp et al., 2014). In a reuse-1 OFDMA macrocell network, each BS chooses which time slots are transmission slots and which are DTX slots under per-UE rate constraints. The paper compares sequential alignment, random alignment, p-persistent ranking, and a memory-based scheme in which each slot keeps a bounded score $\mu \in [0,1]$7 updated according to past usage and current slot quality $\mu \in [0,1]$8 (Holtkamp et al., 2014). All methods converge in at most six OFDMA frames, but their operating points differ sharply. DTX alignment with memory achieves up to 40% savings in power consumption and more than 20% lower retransmission probability than the state of the art random alignment baseline (Holtkamp et al., 2014). The significance is that stable, distributed slot selection can function as a constructive interference-coordination mechanism without explicit inter-BS coordination.

A plausible implication of these stochastic-geometry and distributed-scheduling results is that “cell DTX/DRX” should not be reduced to a single duty-cycle parameter. In multi-cell settings, the temporal structure of silence, the spatial synchronization or staggering of inactivity, and the chosen objective—power, mean STP, local delay, or jitter—can lead to qualitatively different optima.

5. Prediction, wake-up signaling, and RL control of discontinuous reception

Later work increasingly treats DRX/DTX control as a prediction or signaling problem rather than a fixed timer-selection problem. “Energy and Resource Efficiency by User Traffic Prediction and Classification in Cellular Networks” trains per-user LSTM and ARIMA predictors and then maps predicted short- and long-horizon traffic summaries into one of four DRX parameter sets via a decision tree (Azari et al., 2021). On real traffic traces for 10 users, the ML-based scheme achieves average power consumption almost identical to the min-energy configuration, while its delay is significantly better than the min-energy configuration and close to the min-delay configuration (Azari et al., 2021). It also reports that LSTM generally outperforms ARIMA, especially when enough training history is available and the feature set includes UL/DL counts, sizes, ratio, and protocol (Azari et al., 2021). The paper’s explicit conclusion is that per-user adaptation yields much more energy saving at low latency cost than legacy cell-wide DRX parameter adaptation (Azari et al., 2021).

A more control-plane-centric variant is “Optimizing Wireless Discontinuous Reception via MAC Signaling Learning,” where the BTS does not retune DRX timers but instead learns when to send MAC Control Elements that trigger DRX transitions (Pastore et al., 2024). Under XR traffic, the RL policy can nearly halve the active time for a single UE relative to a naïve MAC CE policy, and still achieves near 20% active time reduction for 9 simultaneously served UEs (Pastore et al., 2024). The control logic is therefore shifted from L3 timer configuration to per-TTI MAC signaling.

Wake-up radio work extends this logic further by separating low-power reachability from full-band processing. In “Novel Wake-up Scheme for Energy-Efficient Low-Latency Mobile Devices in 5G Networks,” a wake-up receiver monitors a dedicated PDWCH in short wake-up cycles and activates the BBU only when a wake-up indicator is present (Rostami et al., 2020). For an average buffering delay of 25 ms, the reported power figures are about 140 mW for DRX and about 100 mW for the wake-up scheme, i.e. approximately 30% reduction (Rostami et al., 2020). “Wake-Up Radio based Access in 5G under Delay Constraints” formalizes the same architecture with a semi-Markov process and derives a delay-constrained optimization in closed form, reporting up to 40% lower power than an optimized DRX reference scheme for a given delay requirement (Rostami et al., 2019).

These schemes are still UE-centric, but they matter for cell DTX/DRX because they create an “always-available rather than always-on” control plane (Rostami et al., 2020). This suggests a hierarchical design in which narrowband wake-up or MAC-signaled DRX acts as the fine-grained reachability layer, while BS-level DTX or cell DTX/DRX governs deeper sleep opportunities at the cell.

6. Release 18 cell DTX/DRX and AI-based gNB configuration

The most explicit contemporary formulation of cell DTX/DRX as a standardized network feature appears in “Deep Reinforcement Learning-based Cell DTX/DRX Configuration for Network Energy Saving” (Mao et al., 28 Jul 2025). There, cell DTX/DRX is a 3GPP Release 18, Layer-2, time-domain mechanism in which each cell operates over cycles of length $\mu \in [0,1]$9 with an active on-duration $\varsigma_{\min}$0 and a non-active period $\varsigma_{\min}$1. The duty cycle is

$\varsigma_{\min}$2

During the active period the cell operates normally; during the non-active period there is no data transmission or reception and no associated L1 channels, so ASM can be enabled (Mao et al., 28 Jul 2025).

The paper adopts 3GPP-inspired sleep modes SM1, SM2, and SM3, and uses a simplified downlink power model

$\varsigma_{\min}$3

where $\varsigma_{\min}$4 is the used bandwidth fraction (Mao et al., 28 Jul 2025). Performance is tracked by the delivered-data ratio

$\varsigma_{\min}$5

with $\varsigma_{\min}$6 the amount of data delivered in time and $\varsigma_{\min}$7 the amount not delivered in time, and the normalized average power

$\varsigma_{\min}$8

with $\varsigma_{\min}$9 in the adopted relative power model (Mao et al., 28 Jul 2025).

A central contribution of that work is its reward design. A linear reward,

$P_{\text{PM}} = P_0 + m P_{\text{Tx}},$0

proves workable but brittle across traffic regimes (Mao et al., 28 Jul 2025). A more theoretically appealing QoS-threshold reward is discontinuous and destabilizes training near the QoS boundary. The paper therefore uses a smooth approximation of the thresholded objective and trains a DQN on a contextual-bandit formulation, with state variables including traffic intensity, packet inter-arrival statistics, packet-size statistics, delay requirements, and cell transmission capability (Mao et al., 28 Jul 2025).

In a 21-cell, 210-UE system-level simulator with 3GPP-compliant channel models and FTP Model 3 traffic, the trained agent achieves at least 45% energy saving in light traffic, at least 22% in medium traffic, and about 12% in heavy traffic with a linear reward, all relative to the no-cell-DTX/DRX baseline (Mao et al., 28 Jul 2025). With the approximated QoS reward, the data-rate degradation stays within about 1% in all traffic categories, while energy saving remains large in light and medium traffic; in heavy traffic the same reward can even lead to about 2% higher power than baseline because it backs off from aggressive energy saving to protect QoS (Mao et al., 28 Jul 2025). The paper’s headline summary is that, compared to not using cell DTX/DRX, the agent can achieve up to ~45% energy saving depending on traffic load while maintaining no more than ~1% QoS degradation (Mao et al., 28 Jul 2025).

This result is significant for two reasons. First, it places cell DTX/DRX squarely in the O-RAN, xApp, and near-RT RIC control space rather than treating it as a static RRC parameter set. Second, it confirms, in a standardized Release 18 configuration space, the older analytical insight from BS power models: energy-saving opportunity is concentrated in light and medium load, while heavy-load operation leaves little room for silence without immediate QoS consequences (Holtkamp et al., 2013, Mao et al., 28 Jul 2025).

The remaining controversies are therefore not about whether discontinuous operation saves energy, but about which layer should decide it, which KPI should dominate the objective, and how much of the idealized DTX/DRX gain survives when control channels, wake-up overhead, inter-cell interference, and service-specific latency distributions are modeled explicitly. The literature consistently supports the same broad conclusion: discontinuous cellular operation is most effective when traffic can be temporally concentrated, when static power dominates dynamic RF power, and when control signaling is flexible enough to create genuine sleep windows rather than merely low-throughput active periods (Holtkamp et al., 2013, Azari et al., 2021, Mao et al., 28 Jul 2025).