LoRa-FHSS: Frequency Hopping for IoT Networks

• LoRa-FHSS is a physical layer option for LoRaWAN that divides packets into 488 Hz fragments and applies pseudo-random frequency hopping to enhance uplink performance.
• It improves network capacity and packet reliability in dense terrestrial and satellite IoT scenarios through header repetition, convolutional coding, and fragment recovery.
• Design trade-offs include managing fragment collisions, Doppler tolerance, and multi-gateway cooperation, with regional adaptations for FCC and EU operating environments.

LoRa-frequency Hopping Spread Spectrum (LR-FHSS) is an uplink-only physical-layer option in LoRaWAN that replaces a single wide-band chirp transmission with a packet split into narrow 488 Hz blocks that hop pseudo-randomly across a wider operating channel width, while relying on header repetition, convolutional coding, and fragment recovery to improve scalability and robustness in dense terrestrial and direct-to-satellite IoT networks (Boquet et al., 2020). Across analytical, implementation, and measurement studies, LR-FHSS is associated with higher network capacity than classical LoRa, the same radio link budget as LoRa in some operating points, and a design space centered on fragment collisions, header detectability, Doppler tolerance, and multi-gateway or satellite reception (Lahoud et al., 3 May 2025).

1. Position within LoRaWAN and intended operating regime

Semtech introduced LR-FHSS in November 2020 as a new PHY mode in the LoRa family, and the protocol literature consistently treats it as a LoRaWAN extension for extremely long-range and large-scale communication scenarios such as satellite IoT (Bukhari et al., 2023). The baseline mode is fast pure ALOHA uplink only; downlink remains LoRa-CSS. This asymmetry is operationally important, because ACKs, receive windows, and control procedures still inherit LoRaWAN Class A behavior even when the uplink PHY is LR-FHSS (Boquet et al., 2020).

Comparative analyses report the same 155 dB link margin as LoRa SF12 for the 162 bps, CR $1/3$ operating point, while emphasizing that LR-FHSS trades a single 125–500 kHz chirp channel for many narrow 488 Hz slices and statistical multiplexing in time and frequency (Boquet et al., 2020). It should not be reduced to LoRa-CSS with hopping: the transceiver literature defines a distinct narrowband PHY with GMSK-modulated hopping blocks, repeated headers, and fragment-wise coding and recovery (Jung et al., 2023).

2. Packetization, hopping mechanics, and regional parametrization

An LR-FHSS uplink packet comprises a PHY-header block, a sequence of $N_H$ header replicas, and a sequence of $N_F$ payload fragments. Header blocks last $T_H=233.472$ ms, payload fragments last $T_F=102.4$ ms, and the header carries the hopping-sequence seed, data rate, payload length, and code rate so that the receiver can reconstruct the fragment schedule (Jung et al., 2023). In one protocol overview, the packet structure is summarized as SyncWord plus $R$ header replicas plus $L$ payload fragments; the payload is convolutionally encoded at rate $\mu$, partitioned into $L$ fragments, and successfully recovered if at least $\lceil \mu L\rceil$ fragments are received (Lahoud et al., 3 May 2025).

Regional instantiations differ materially. In the FCC band, one analytical study uses an Operating Channel Width of 1.523 MHz, an Occupied Bandwidth of 488 Hz, and a hopping organization of $N_H$0 grids $N_H$1 $N_H$2 OBWs with 25.4 kHz spacing; each payload fragment hops pseudo-randomly within one grid (Lahoud et al., 3 May 2025). In the EU 868 MHz band, the reference overview reports $N_H$3 kHz for DR8/9, $N_H$4 Hz, minimum hop spacing $N_H$5 kHz, and about $N_H$6 usable subcarriers per OCW (Boquet et al., 2020).

One published hop model writes the $N_H$7th fragment frequency as

$N_H$8

where $N_H$9 is a 32-bit hash, $N_F$0 is a 9-bit seed from the header, and $N_F$1 is the number of usable subcarriers per device (Boquet et al., 2020). Header repetition and payload coding are DR-dependent: in the FCC analysis DR5 uses $N_F$2 and $N_F$3 at 162 bps, whereas DR6 uses $N_F$4 and $N_F$5 at 325 bps (Lahoud et al., 3 May 2025); the EU overview uses DR8 and DR9 labels for 162 bps and 325 bps modes (Boquet et al., 2020). This suggests that LR-FHSS data-rate nomenclature is regional-parameter dependent rather than globally uniform.

3. Reception architectures, decoding constraints, and fragment recovery

Baseline LoRaWAN reception assumes that one gateway receives at least one header replica and a sufficient fraction of payload fragments. In the reference overview, the gateway listens to the entire OCW and reassembles a packet if it decodes at least one header and at least $N_F$6 of the fragments (Boquet et al., 2020). Direct-to-satellite receiver designs implement this by wideband channelization, header detection, CFO estimation, buffering per-channel time series, payload extraction at predicted hop times, and Viterbi/CRC decoding; one design reports that GMSK outperforms QPSK under Doppler up to 400 Hz and reaches $N_F$7 at SNR $N_F$8 dB for a 48-symbol search window (Jung et al., 2023). Jung et al. extend this architecture with FFT-based channelization up to 3120 carriers, Doppler-rate tracking, and bidirectional SOVA; laboratory tests consider Doppler rates up to 400 Hz/s and co-channel interference up to 40% overlapping hops among devices (Jung et al., 2024).

Several studies identify header acquisition as the dominant bottleneck under high offered load. In a slotted, no-capture analytical model, the header reception probability is

$N_F$9

and the authors explicitly show that in dense traffic $T_H=233.472$0 falls much faster than the payload-reception term, motivating recovery of frames whose payload fragments are observable but whose headers are lost (Fraire et al., 2023). Their linear-programming formulation and sliding-window heuristic exploit the known set of 512 pseudo-random sequences; in simulation both approaches achieve TP$T_H=233.472$1 for all tested loads, and in fast mode with 2000 frames legacy extraction is about $T_H=233.472$2 versus about $T_H=233.472$3 for headerless extraction (Fraire et al., 2023).

Fragment-level macro-diversity removes a different bottleneck: the requirement that packet reconstruction occur at a single gateway. In the multi-gateway scheme, any gateway that decodes one header replica forwards it to a controller, the controller reconstructs the hopping sequence, instructs all gateways to monitor the packet, and then de-duplicates forwarded fragments by $T_H=233.472$4 until $T_H=233.472$5 (Lahoud et al., 3 May 2025). Numerical evaluation shows that, relative to nearest-gateway reception, macro-diversity extends the reliable load region by about $T_H=233.472$6–$T_H=233.472$7; the reported reliable-load gains are from about $T_H=233.472$8 to $T_H=233.472$9 Mbps for DR5 and from about $T_F=102.4$0 to $T_F=102.4$1 Mbps for DR6 (Lahoud et al., 3 May 2025).

4. Analytical models of reliability, throughput, and optimization

A substantial part of the LR-FHSS literature models collisions at fragment granularity. For fragment-level macro-diversity, gateways are represented by a homogeneous Poisson point process $T_F=102.4$2 of density $T_F=102.4$3, devices by an independent Poisson point process $T_F=102.4$4 of density $T_F=102.4$5, and the effective interferer density is thinned by both time and frequency as

$T_F=102.4$6

With single-fragment success $T_F=102.4$7, the number of correctly received fragments obeys $T_F=102.4$8, the payload succeeds when $T_F=102.4$9, and overall packet success is $R$0 (Lahoud et al., 3 May 2025).

At network level, the reference performance analysis casts LR-FHSS as multi-channel ALOHA. With offered load $R$1, ideal LoRa pure ALOHA yields $R$2 per channel, whereas LR-FHSS is approximated by

$R$3

when $R$4 parallel subcarriers are available (Boquet et al., 2020). This model captures why LR-FHSS can maintain useful goodput under loads for which single-channel ALOHA collapses.

Optimization problems built on the collision model go beyond fixed standardized data rates. In a device-level probabilistic allocation strategy, the network server distributes a probability vector $R$5 over setups $R$6, yielding packet success

$R$7

goodput $R$8, and energy efficiency $R$9; exhaustive-search results show that the optimal distribution rarely includes DR9, while DR8 contributes significantly to goodput and energy-efficiency optimizations (Farhat et al., 2024). A different reliability model studies message replication without ACKs and derives $L$0 for frame replication and $L$1 for fragment replication, concluding that DR8 with frame replication is preferable at low traffic whereas DR9 with fragment replication becomes preferable at moderate to heavy traffic (Rathi et al., 21 Jan 2025).

5. Reported performance in simulation, traces, and measurements

Packet-level simulation in EU 868 MHz reports up to $L$2 capacity gain over standard LoRa for LR-FHSS DR8 and $L$3 for DR9 on a one-channel basis, with simulated maximum goodput of $L$4 pkt/h for DR8 and $L$5 pkt/h for DR9 (Boquet et al., 2020). The same study reports PER values of $L$6, $L$7, and $L$8 for LR-FHSS DR8 at 100, 1\,000, and 10\,000 devices, versus $L$9, $\mu$0, and about $\mu$1 for LoRa DR0 in the same scenarios (Boquet et al., 2020).

Measurement-driven studies also report tangible gains in urban operation. In an urban Halifax campaign in the US915 band, LR-FHSS achieved up to a $\mu$2 improvement in Packet Reception Rate over traditional LoRa in dense urban areas; at $\mu$3 km the study reports PRR$\mu$4(DR0)$\mu$5, PRR$\mu$6(DR5)$\mu$7, and a bootstrap 95% confidence interval for PRR$\mu$8-PRR$\mu$9 of $L$0 (Delplace et al., 27 Oct 2025). The same campaign reports minimum RSSI values of about $L$1 dBm for DR5, about $L$2 dBm for DR6, and about $L$3 dBm for DR0, but also notes that part of the apparent sensitivity advantage stems from how the gateway computes RSSI for hopping versus continuous chirp signals (Delplace et al., 27 Oct 2025).

Software-receiver processing of packet traces from an actual SX1261 transmitter gives a complementary view of LR-FHSS under practical impairments. Using real traces with frequency error, one study reports one-to-one SNR thresholds for PRR $L$4 of about $L$5 dB for DR8 and about $L$6 dB for DR9, together with trace-driven network capacities of about $L$7 kbps for DR8 and about $L$8 kbps for DR9 in both AWGN and NTN settings (Bukhari et al., 2023). Its customized receiver raises DR8 AWGN capacity at 3.5 kbps load from $L$9 to $\lceil \mu L\rceil$0 kbps with collision-aware erasure decoding and to $\lceil \mu L\rceil$1 kbps with additional packet-level successive interference cancellation (Bukhari et al., 2023).

Energy evaluations show that LR-FHSS trades airtime for collision resilience and coverage. Measurements on real hardware in a fully operational LR-FHSS network indicate that, with optimal configuration, LR-FHSS end-device battery lifetime can reach 2.5 years for a 50 min notification period, and up to 16 years with a one-day notification interval using a coin-cell battery (Sanchez-Vital et al., 2024). Experiment-based ToA and current-consumption models based on 460+ transmissions per DR report relative ToA error below 0.3%, observe that DR9 reduces ToA by about 50% relative to DR8, and state that battery life can reach about 3.8 years on 2400 mAh for a 10-byte packet at $\lceil \mu L\rceil$2 dBm every 15 min (Ullah et al., 2024).

6. Extensions, caveats, and active research directions

Sequence design and receiver resource allocation remain active lines of work. One study compares Semtech’s default driver family with Li–Fan wide-gap frequency-hopping sequences and reports average cross-correlation $\lceil \mu L\rceil$3 versus $\lceil \mu L\rceil$4; when combined with Early-Decode and Early-Drop demodulator allocation, the reported packet success rate improves from $\lceil \mu L\rceil$5 to $\lceil \mu L\rceil$6 at $\lceil \mu L\rceil$7, and decoded payload throughput can double for $\lceil \mu L\rceil$8, CR1, and $\lceil \mu L\rceil$9 (Maldonado et al., 2024).

Other extensions move beyond a single grant-free uplink. D2D-aided LR-FHSS with network coding derives closed-form outage expressions under a shadowed-Rice satellite channel and reports capacity increases of $N_H$00 for DR6 and $N_H$01 for DR5 at outage $N_H$02, at the cost of a minimum of one and a maximum of two additional transmissions per end device in each time-slot (Maleki et al., 2022). The open-source LR-FHSS-Sim models fragments as SimPy events, reproduces the requirement of at least one collision-free header and a sufficient fraction of payload fragments, and is explicitly intended to host extensions such as SIC, alternative traffic models, Doppler, multi-gateway topologies, and custom hopping schemes (Ana et al., 2024).

Several recurrent comparisons require qualification. First, LR-FHSS is not a downlink PHY in the baseline system; downlinks, including ACKs, remain LoRa-CSS. Second, the literature continues to treat optimal or adaptive hopping sequences, dynamic hop sets, coexistence with legacy LoRa, spectrum planning, channel partitioning, and ADR extensions as open issues (Boquet et al., 2020). Third, spectral-efficiency figures are reported differently across bands: the EU overview gives $N_H$03 bps/Hz for DR8, whereas the US915 measurement paper gives $N_H$04 bit/s $N_H$05 1.523 MHz $N_H$06 bit/s/Hz for DR5 and explicitly states that LR-FHSS strength lies in link budget and interference resilience, not pure throughput (Delplace et al., 27 Oct 2025). This suggests that cross-paper efficiency comparisons should be read in the context of the specific OCW and regional parameter set. Finally, because current deployments may require packet reconstruction at a single gateway, fragment-level macro-diversity suggests that gateway architecture—not only the hopping PHY—can limit realized scalability (Lahoud et al., 3 May 2025).