Extend-and-forward relaying through wireless frequency extension for sub-terahertz communication Authors: Eray Güven, Louis Thivierge, Jean-François Frigon, Güneş Karabulut Kurt Published: 2026 ABSTRACT AbstractThis study introduces extended-and-forward relaying for sub-terahertz (sub-THz) extended fronthaul link (EFL) via jointly implemented frequency division multiplexing at the fronthaul and wireless frequency extension for sub-THz relaying. EFL transmission and reception, phase noise (PN) induced symbol error probability, and operable frequency regions are modeled. Aggregated PN significance over distributedly synchronized (DS) and non-synchronized (NS) EFLs is compared. 180 GHz EFL on cascaded software-defined radio (SDR)-THz testbed is demonstrated with off-the-shelf SDRs, frequency extenders, and various radio frequency measurement tools. DS demonstration achieves 0 bit error rate over transmitted bits per user at measured signal-to-noise ratios between 11 and 13 dB, whereas NS shows 0.01 bit error floor with software controllable SDR clock aggregated 0.384 radian PN deviation at most. A digital post correction is proposed to suppress PN-related clock offset by dynamic phase correction, which achieves up to 0.088 BER decrement in NS, with the improvement varying by user and link distances.Keywords: Sub-terahertz communication, Frequency extension, Frequency division multiplexing, Software-defined radiosSubject terms: Engineering, Optics and photonics, Physics Introduction In wireless communication, there is a growing demand for less congested fronthaul and wideband wireless backhaul links. To address these versatile demands, sub-terahertz (sub-THz) bands not only enable faster communication with wide spectrum availability, but also present segregable sub-bands that minimize the interference in fronthaul. An overview of the recent advancements in THz communication, use cases, challenges, and advancements is given in1. Other than macro-scale deployment challenges such as power inefficiency, high attenuation, and mobility, it has recently been shown that sub-THz communication faces design and link-level challenges, such as data rate scarcity, frequency sparsity, and hardware impairments2. Several sub-THz and THz applications have been investigated in the literature, including long-distance backhaul links3 and indoor radio systems4–6. In7 and8, a phase noise (PN) resilient photonic-based 300 GHz wireless backhaul and fronthaul link is introduced. Multi-channel sub-THz and THz backhaul demonstrations in9,10 at 130 GHz and 220 GHz over distances exceeding 1 km. Recently, PN and peak-to-average-power ratio in sub-THz with orthogonal frequency division multiplexing (OFDM)11, where the modeling of PN is proposed with a Gaussian distribution in12. In the same monolithic scenario, the potential use of sub-THz targeting multi-users with leaky wave antennas in13 for downlink. Similarly, several mm-wave and sub-THz software-defined radio (SDR) demonstrations are gaining attention, such as14,]15. SDRs provide low-cost and flexible real-time DSP capability while also offering reprogrammable FPGA-based RF front-end architectures. To overcome the data rate sufficiency challenge, two recent studies have deployed RF-SoC-based SDR implementations5,6. Another study develops the high-throughput SDR for mmWave beam management16. In6,17, a frequency multiplexing technique for ultrabroadband signal generation via SDR is demonstrated. As a low profile SDR unit, USRP-B200 provides futuristic applications based on SDR, such as CubeSats design, mobile communication, spectroscopy, and scientific observations18. In contrast to the existing monolithic sub-THz use cases, we take advantage of a wireless frequency multiplexing technique19 with RF upconversion in the fronthaul, without using dedicated multiplexing and signal-separation hardware, such as multiplexers, demultiplexers, or diplexers. We refer to this architecture as extend-and-forward (EF) relaying. In this approach, a frequency extender (FEX) captures the sub-6 GHz multi-user frequency division multiplexing (FDM) signal over the air, upconverts it to the sub-THz band, and forwards it toward the BS, where downconversion is performed. This approach allows existing sub-6 GHz infrastructure to leverage the sub-THz spectrum without requiring additional resources. As a solution for ultrabroadband data generation from a monolithic source, EF relaying exploits the integration of independent users to aggregate sub-bands and achieve ultrabroadband transmission in sub-THz. An extended fronthaul link (EFL) between EF relay and BS, not only eliminates the self-interference, but also enables mobile users to benefit from sub-THz communication without mobility challenges20. This potential benefit serves at the main motivation of this work where we aimed its experimental demonstration. Our work differs from recent single-link and single-transmitter sub-THz demonstrations, such as the D-band SDR link in21, the multi-subchannel digital subcarrier multiplexing scheme in22, and the single-source wideband OFDM demonstration in23, none of which perform wireless relaying of independently generated, frequency-multiplexed transmitters. EF relaying offers advantages over traditional THz relaying methods. Unlike amplify-and-forward techniques24, it inherently avoids amplifying interference to co-channel bands. Additionally, it achieves signal privacy benefits without the computational overhead associated with decode-and-forward relaying approaches25. Furthermore, unlike existing SDR-based sub-THz frequency extension studies that primarily focus on single-user link-level transmission, the proposed architecture performs real-time wireless frequency extension of aggregated frequency-division-multiplexed users. This enables multi-user sub-THz relaying without signal regeneration, extending frequency-extension techniques beyond conventional point-to-point operation. Two example use cases of EFL and wireless sub-THz backhaul are illustrated in Fig. 1a. In EFL, sub-6 GHz multiplexed users relayed to a BS. In Fig. 1b, a multi-user multiplexed uplink signal is upconverted on the BS to be directed to the core network (CN). A use case of wireless frequency extension scenario. Extend-and-forward (EF) repeater upconverts sub-6 GHz multiplexed users to sub-THz and forwards the ultrabroadband signal to the BS for eliminating the self-interference. A comparative overview of the proposed architecture relative to recent SDR-based and multi-channel sub-THz demonstrations is given in Table 1. Table 1Comparison with experimental sub-THz studies.ReferencesFrequencyUsers/channelsRelay type5,6145–147 GHz1None14mmWave1None16mmWave1None17sub-THz1 (FDM)None21D-band1None22220 GHzMulti-subchannel (single TX)None23180 GHz1, wideband OFDMNoneThis work180 GHz3, FDM, independent TXEF Comparison with experimental sub-THz studies. The main contributions of this study can be summarized as follows EF relaying is introduced for multiple access in fronthaul links through wireless multiplexing, enabling relay operation without self-interference or computational overhead, indicating potential compatibility with existing sub-6 GHz infrastructure. EF relaying is introduced for multiple access in fronthaul links through wireless multiplexing, enabling relay operation without self-interference or computational overhead, indicating potential compatibility with existing sub-6 GHz infrastructure. Aggregated PN limitation and probability of symbol error boundary on sub-THz EFL are derived. Distributed synchronization (DS) improvement on the PN mitigation in sub-THz EFL is presented. Aggregated PN limitation and probability of symbol error boundary on sub-THz EFL are derived. Distributed synchronization (DS) improvement on the PN mitigation in sub-THz EFL is presented. First real-time demonstration of EF relaying in a sub-THz SDR-based frequency extension testbed is achieved. The impact of aggregated phase noise (PN) is evaluated in comparison to distributedly synchronized (DS) and non-synchronized (NS) EFLs. First real-time demonstration of EF relaying in a sub-THz SDR-based frequency extension testbed is achieved. The impact of aggregated phase noise (PN) is evaluated in comparison to distributedly synchronized (DS) and non-synchronized (NS) EFLs. A dynamic phase compensation method with digital post correction (DPC), mitigating the PN-related clock offset in the NS case and reducing the measured bit error rate by up to 0.088 across tested users and distances in real time, is proposed. A dynamic phase compensation method with digital post correction (DPC), mitigating the PN-related clock offset in the NS case and reducing the measured bit error rate by up to 0.088 across tested users and distances in real time, is proposed. The rest of the paper is organized as follows. In Section “Wireless sub-THz multiplexing analysis”, wireless sub-THz multiplexing steps for the cascaded SDR-THz signal model are presented. In Section “Experimental design”, we provide the experimental evaluation of the proposed model. Next, Section “Experimental results” discusses the results and the model’s applicability in real-time, accompanied by numerical results. Finally, the conclusion of this study is presented in Section “Conclusion”. Wireless sub-THz multiplexing analysis In this section, the electronics-based sub-THz up and downconversion is analyzed for 2 stages of cascaded EFL. In the first stage S(1), N users, each equipped with a single antenna and operating over a narrowband channel, are assigned sub-6 GHz uplink frequencies. Uplink user signals are multiplexed in frequency by a single antenna-equipped BS. In the second stage of S(2), the multiplexed signal is extended into the sub-THz band to be transmitted to a BS over EFL. Each user is represented by a radio (by an SDR in experimental validation in Section “Experimental design”). BS is represented by a FEX-1 to perform joint multiplexing and frequency upconversion, and the CN is represented by the FEX-2 integrated with an SDR to perform sub-THz downconversion and message decoding. For further information, the two-stage SDR-THz testbed is detailed in2. In the following, we present the received signal model in the EFL transmission and EFL reception steps. EFL transmission The baseband signal of a generic homodyne user is defined as follows where |s(t)| is the envelope of the signal, is the nominal transmitted signal phase. The intermediate frequency (IF) passband output of the n-th user with PN (denoted with ) on it’s local oscillator (LO) is the following where in sub-6 GHz is the IF frequency on S(1) of the n-th user, n ∈ N. Without loss in generality, a multi-users scenario follows with the following superposed signal . For simplicity, we continue our analysis with single user N=1 case. Through sub-THz upconversion, x(t) is mixed with an LO by an external sub-harmonic mixer with as follows where the sub-THz LO source frequency is upconverted with M factor as following Using the binomial expansion of the generalized power reduction identity26,Section 1.32, page 31], M=6 upconverted LO is following where we denote the followings for the convenience Here, the M-th order non-linearity is denoted with the coefficient 0<α <1. In short, Eq. (3) results with upconverted EFL mixed signal of contains comb of harmonics with components as following where we denote each double-sided harmonic component as follows This expression presents the sub-THz upconverted signal model at the transceiver. In practice, the terms A, B, C, D, E are effectively suppressed at front-end after the mixer, by passive components such as the waveguide, antenna, and high-pass filters. Using the harmonic components from A to G are expressed as where . Assuming that the highest frequency components are the desired sub-THz bands, F and G components are desired for wireless transmission. The desired M-th order upconverted components F and G are spectrally isolated from the lower-order components A–E with following where coefficients are illustrated in Eq. (8), and is following Therefore, all lower-order harmonic terms vanish and only the desired M-th order components remain. Non-synchronized EFL reception Uplink IF of sub-6 GHz signal recovery at BS in NS manner is presented in this section. First, a sub-THz to IF downconversion with an LO mixing as follows where Similar to Eq. (3), the expression for non-synchronized EFL with an existing is given below Using the power reduction identity in Eq. (5), results in Eq. (14) at next page with highlighted sub-THz components using . This derivation shows the every harmonics including the sub-THz. In coherent communication systems to minimize phase drifts, synchronization related errors, and keep the architecture simple, and are selected as equal. In this case, sub-THz components in Eq. (14) corresponds to . For example, and base terms are already at and . Also note the cross mixing terms of and adds onto same sub-THz component. It can be seen that cross mixing term coefficients are 45 while the base sub-THz terms are 10. This shows that method of this sub-THz generation relies on more the synchronization of and than base M-th sub-THz harmonic. In contrast to the NS architecture, a more robust EFL reception can be achieved through distributed synchronization (DS), where . In this case, a common oscillator source is utilized for both upconversion and downconversion processes. The DS-based EFL analysis is presented in the following subsection Distributedly synchronized EFL reception In order to mitigate , the expression for DS-EFL is obtained by using Eqs. (8) and (11) as following which repeats the previous steps for as given in Eq. (16), with highlighted sub-THz components using . It can be observed that the scaling factor in sub-THz components of F and G results in 1/2048 amplitude (33.11 dB loss) in both NS and DS. Following the and , IF downconversion and baseband processes follow respectively. The expression of obtained in clearly shows that the DS case preserves the total harmonic generation power, and doesn’t include any terms. To fully understand the impact of PN terms in NS systems and to emphasize the importance of synchronization, the symbol error tolerance under PN-affected ELF systems is presented in the next subsection. Symbol error in EFL with PN Previously, PN distribution with zero mean Gaussian is mentioned in several studies27–29. Here, we present the probability of symbol error caused by PN in EFL. We denote the aggregated PN with L sources as follows There exists a symmetric decision bound () for symbol phase angle that minimizes the symbol error probability of , expressed as Let . For M-ary PSK systems, the angular distance between adjacent symbols is 2 π /M; therefore, the symbol decision boundary angle is . The probability of symbol error due to the PN being larger than the decision boundary is as follows This can be expressed within the Q function form as follows where . Only PN-related symbol error bound for quadrature phase shift keying (QPSK) is given in Fig. 2, which shows that received signal is PN-resilient by the decision boundary angle, as also derived in Eq. (19). Symbol error is monotonicity increases with the increased PN deviation. QPSK symbol error probability versus phase noise (PN) standard deviation (top), and corresponding constellation diagrams under unbiased Gaussian PN with standard deviations: (a) , (b) , (c) , and (d) radians (bottom). Although PN remains a fundamental impairment in THz communication systems, another critical limitation in cascaded sub-THz architectures is frequency sparsity resulting from hardware constraints. Therefore, the following subsection analyzes the operable frequency regions of the proposed architecture. EFL operation region In2, it is explained that frequency extended SDR-THz testbeds can only operate in discrete frequency points, depending on the LO resolution of each stage of S(1) and S(2). For example, S(2) constrained frequency operable region g is approximated in2. Here, we provide EFL operation region analysis with S(1) included, as wireless multiplexing introduces a non-negligible sparsity region. A generalized two-stage SDR-THz operation region is as follows where is the target nominal sub-THz frequency and is the smallest frequency step. Upconversion in Eq. (7) denoted as where and in2 After the cancellation of the nominal terms, the smallest frequency step becomes the following An image rejection filter , suppressing the lower sideband into the upper single band as follows A broadband constraint is introduced, establishing a linear relationship between and , expressed as . In this case, step size () is following Therefore, an EFL is not able to operate within interval, constituting the minimum step size between two operable frequency points. Experimental design Below, we support the analysis presented above with measurement results. First, we show that the DS receiver preserves the total upconverted signal power and removes any dependence on , whereas the NS receiver retains cross-mixing terms that depend on both and (Sections “Non-synchronized EFL reception” and “Distributedly synchronized EFL reception”). Second, the symbol-error analysis, expressed through and the decision boundary π /M, is shown to provide a PN-resilient region below which symbol errors remain negligible (Section “Symbol error in EFL with PN”). Finally, the operable-frequency-region analysis is shown to bound the minimum usable spacing between simultaneously multiplexed users (Section “EFL operation region”). Instrumentation The EF relaying experiment was conducted in an indoor lab environment with the SDR-THz testbed in the Poly-Grames Research Center30. Three independent Universal Serial Radio Peripheral (USRP) B200 (SDR-1, SDR-2, and SDR-3) were used to emit signals. They were each emitting over the air a predetermined message at different assigned carrier frequencies. Transmitter SDRs emit with an omnidirectional swivel dipole antenna of maximum 5 dBi gain. Sub-THz up and downconversion is performed by Virginia Diodes Inc. (VDI) frequency extension modules (FEX) with WR-5.1 input port flange. FEXs are referred to as FEX-1 and FEX-2 in upconversion and downconversion, respectively. FEX-1 and FEX-2 are used to perform the upconversion and the down conversion respectively with an integrated M=6 multiplier (a wideband doubler and trippler) which ends up with M+1 harmonics as explained in Section “Wireless sub-THz multiplexing analysis”. The selection of M=6 is determined by the architecture of the employed FEX modules. While the same carrier frequency may be generated using higher multiplication factors and lower source frequencies, larger multiplication factors increase harmonic generation, insertion loss, and phase-noise amplification. Among the available operating frequencies which are compatibale with system, 180 GHz was selected as a mid-band operating point sufficiently far from the cutoff regions of the front-end hardware. Both FEX-1 and FEX-2 are equipped with directive horn antennas, with a gain of 11 dBi and a 3 dB beamwidth of on the E-plane and on the H-plane. A detailed radiation analysis for the front-end in31. As an uplink scenario depicted in Fig. 3, SDR-1, SDR-2, and SDR-3 represent independent users, FEX-1 represents the BS, and SDR-4, integrated with FEX-2, represents the CN. Each SDR is hosted by a computer. Additionally, the misalignment on S(2) is non-measured; therefore, the RSS level on S(2) is suboptimal. Wireless EF relaying with multi-user SDR-THz testbed block diagram. In NS, LO upconversion with is generated by while LO source is referred as . In DS, source is utilized for both FEX-1 and FEX-2. In the synchronized case, a pulse signal generator (E8267D) provides a 30 GHz LO source with 5 dBm to both FEX-1 and FEX-2 over a wideband (6-40 GHz) two-way splitter. The down-converted signal from FEX-2 is forwarded to another USRP B200 (SDR-4) via wire. Therefore, it corresponds to synchronized SDR-THz receiver model with in the Section “Wireless sub-THz multiplexing analysis”. In the non-synchronized case, two independent yet identical PSGs are used. Due to the direct channel from S(1) to SDR-4, the experiment is placed in the anechoic chamber to mitigate the IF reflections in uplink. Moreover, the FDM band power and quality are observed over the Agilent N9030A spectrum analyzer. Each SDR hosts a Spartan-6 FPGA and an AFD-9364 transceiver with 50-6000 MHz, which uses an external LO on board 40 MHz TCXO with ± 2 ppm and . USRP-B200 PN at 4 GHz is noted around 1.0 deg RMS32. A maximum of instantaneous 56 MHz bandwidth is reachable, whereas the USB 2.0 interface bottlenecks it by underflows from ∼ 4 MS/s up to ∼ 8 MS/s. To avoid, the host computer of the 3 SDRs avoided with 1 MS/s. Maximum achievable gain is used in each transmitter SDR with less than 5 dBm output power for each, while receiver SDR-4 utilizes automatic gain control (AGC). Each SDR is calibrated in output power by equalizing the software controllable amplifier gain (During the linear region of the SDR amplifier, it is possible to apply gain calibration to correspond the SDR output power in dBm, whereas we skip this step.). Due to the missing gain calibration, the SDR spectrum gain is only relative. Nevertheless, the SA output illustrates the input power of the IF downconverted signal (Fig. 4). (a) Real-time SDR-1 analysis on SDR-4 and SA, (b) NS-17B measurement with 3 USRP-B200. Measurement details The setup configuration is given in Fig. 3. In S(1), the antenna to antenna distances between the FEX-1 to SDR-1, SDR-2, SDR-3 are 40 cm. Each SDR is assigned a pre-determined binary message to emit perpetually. Message-1 is a text script with ASCII encoding in binary. Message-1 is a text script with ASCII encoding in binary. Message-2 and Message-3 are pseudorandomly generated numbers within (0, 255]. Message-2 and Message-3 are pseudorandomly generated numbers within (0, 255]. Each symbols are oversampled with 4 and differentially encoded quadrature key phase shifting (QPSK) modulated, which has normalized average symbol energy with unit power. MHz, Hz33 and Δ M = 1 limit the operation point, which is used for a guard band, where the dominant term is the . This spacing is based on the the operable-frequency-region bound MHz in Eq. (25), which sets the minimum usable separation between simultaneously multiplexed users (Using MHz and the SDR operating range of 50 MHz-6 GHz, an upper bound of 29 frequency-multiplexed users can in principle be accommodated in S(1) before extension. Note that this is an upper bound as uniform per-user IF bandwidth allocation is assumed and guard-band and front-end linearity constraints are not considered. This upper bound does not account for the aggregated phase-noise growth with the number of sources L (Eq. 19), which is expected to reduce the number of users.). The SDR transceivers employed in the experiments support an operating range between 50 MHz and 6 GHz, while the PolyGrames anechoic chamber supports measurements from 4 GHz to 400 GHz due to its absorber design and chamber calibration range. In addition, the resonant frequency range of the probe antennas resides approximately between 3.8 GHz and 4.6 GHz. In this case, SDR-IF carrier frequencies are GHz for SDR-1, GHz for SDR-2 and GHz for SDR-3. A direct channel from user SDRs to CN is mitigated by the anechoic chamber to reveal the sub-THz performance (Authors note that a direct channel is not an undesired link34. It is only mitigated to observe the sub-THz performance). The following tests for non-synchronized cases, resulting in , are evaluated: A NS EFL with 6 cm (NS-6B), a NS EFL with 17 cm (NS-17B). Also, distributedly synchronized tests for 6 cm ELF (DS-6B) and distributedly synchronized 17 cm EFL (DS-17B) cases, which present the upper limit as a performance benchmark with , are compared with NS-6B and NS-17B. The selected transmission distances were chosen to investigate synchronization robustness under different propagation conditions while remaining within the operational limits of the experimental setup. The measurements at 6 cm and 17 cm demonstrate that synchronization performance is not solely determined by attenuation effects. At , the wavelength is given by The free-space path loss is expressed as , or in dB form, For For . Thus, increasing the separation from 6 cm to 17 cm introduces an additional propagation loss of approximately . Importantly, this controlled difference in path loss is intentionally leveraged in the experimental design. Since the FEX-to-user distances are kept fixed throughout all measurements, the relative attenuation is dominated by predictable free-space propagation effects rather than uncontrolled channel variations. This helps isolate the impact of THz attenuation and system impairments while reducing sensitivity to unknown factors such as slight antenna misalignment or uncharacteristic insertion losses. In real-life scenarios, a clock distribution method such as Octoclock with a global reference signal35,36 or White Rabbit37, as we bypass this with a single LO with in Fig. 3. The DS and NS test cases defined in Sections “Non-synchronized EFL reception” and II-C are designed to reveal the different synchronization behavior. DS is expected to exhibit converged phase-locked-loop tracking and error-free or near error-free BER (Fig. 5), while NS is expected to exhibit irregular phase wrapping and a persistent BER floor (Fig. 7), as will be confirmed by the measurements presented in Section “Experimental results”. DS-6B and DS-17B measurements (synchronized cases) on SDR-4 from SDR-1 (orange), SDR-2 (blue), and SDR-3 (green) where (a) and (d) are bit error rate and SNR evaluation, (b) and (e) are constellation and spectrum, and (c) and (f) are locked phase with normalized CFO. The x- and y-axis titles are expressed only once per subfigure and correspond to all measurements. NS-6B and NS-17B measurements (non-synchronized cases) on SDR-4 from SDR-1 (orange), SDR-2 (blue), and SDR-3 (green) where (a) and (d) are bit error rate and SNR evaluation, (b) and (e) are constellation and spectrum, and (c) and (f) are locked phase with normalized CFO. The x- and y-axis titles are consistent and correspond to all measurements. Digital design details All SDRs data and power connections are done over USB 3.0 via UHD 4.7. IF bandwidth in S(1) is proportional with 2 baseband sampling rate and the oversampling ratio of 4. Targeting the 400 kHz IF bandwidth for each user SDR (with usable bandwidth due to root raised cosine (RRC) pulse shaping and built-in half band filter), 32 MHz master clock rate with 32 interpolation factor and decimation. In the reception on SDR-4, a matched filter to the transceiver’s RRC taps mitigates inter-symbol interference (ISI) and downsamples afterwards. Adaptive filter algorithm of decision-directed (DD) least mean square (LMS) equalizer blindly equalizes the channel while CFO is estimated and compensated by digital phase-locked loop (DPLL)38. Timing synchronization39 tunes the sampling time to maximize the correlation of transmitted and detected receieved symbols, while algorithm40 estimates the SNR by the kurtosis of the signal and the noise for 100 sample interval. IQ symbols are collected and analyzed over SDR-4 and SA. z(t) on SDR-4 is windowed with Blackman-Harris before the discrete Fourier transform (DFT), and the power is normalized with maximum power spectrum density (PSD). SDR-4 utilizes built-in bandpass filters to eliminate undesired users. Lastly, each 100 symbols are packed as 1 frame, where the instantaneous BER for each frame is calculated and illustrated in Figs. 5 and 7. The digital design is further enhanced through the post-processing DPC algorithm, which is detailed in the following subsection. The decision boundary π /M and the PN-resilience criterion supports the CFO and SNR estimation blocks. Digital post correction (DPC) Based on the knowledge of aggregated SCO due to the phase jitter, several steps of the DPC algorithm aim to break down the phase reliance on each DSP step at the receiver. The proposed DPC chain is applied exclusively to the NS receiver architecture. It consists of two main stages. First, an AGC-based compensation is applied to mitigate amplitude fluctuations and restore a nearly constant-envelope received signal. Second, several conventional synchronization blocks are replaced with alternative synchronization methods that are less dependent on carrier phase information, at the expense of reduced PLL tracking capability and slower convergence. After digital downconversion from IF to baseband, DPC decimates each sample and normalize the fast fluctuations with digital AGC41 with attack and decay rate of 1. Following this, the DD-based Mueller & Müller algorithm39 is replaced with Gardner42, which estimates timing error from midpoint samples. Figure 6 shows IQ symbols with DPC. Similarly, the DD adaptive filter is replaced with the constant modulus algorithm (CMA)43 with 0.001 convergence rate, where the linear equalization is placed based on the constant amplitude instead of the phase. Since CMA adaptation relies exclusively on signal magnitude, the proposed DPC constitutes a suboptimal solution relative to fully phase-aware synchronization schemes. IQ recovery example with DPC on User-1 measurement in NS-6B. Experimental results During the experiments, inter-modulation interference occurs on S(2) due to the generated RF combs, while self-interference on BS does not exist thanks to the separated sub-6 and sub-THz spectrum use. results in nearly 200 MHz band gap among users, which is wide enough to avoid using guard bands in the uplink of S(1). Therefore, the operation region can only be tunable with integer coefficients of MHz. This measured band gap matches the operable-frequency-region bound in Eq. (25), confirming that the 200 MHz spacing adopted for SDR-1, SDR-2, and SDR-3 is the minimum separation permitted by the cascaded upconversion architecture. Despite the calibration in output power on each SDR, the attenuation on each band is ± 5 dB differs on Figs. 5 and 7. On both DS and NS cases, BS aims coherent QPSK demodulation for all users, synchronizing in time and frequency. In the NS case, PN-related SCO drops the accuracy of synchronization steps, leading to periodic bit errors with respect to the sampling interval. Figure 6 reveals the severity of CFO during NS-6B measurement. As a result of the phase jitter and non-synchronized sampling process, NS-6B and NS-17B measurements constitute 0.01 BER floor due to the SCO and PN. Following this, a constant CFO deviation leaves a residual offset. Figure 5a,d present the instantaneous BER and SNR variations over time within a frame. As the backhaul distance increases from 6 cm to 17 cm, the impact of sub-THz propagation attenuation is directly observed through the reduction in SNR. Nevertheless, the maintained synchronization stability prevents the emergence of a BER error floor. The resulting attenuation trend is also clearly observable in the measured spectra shown in Fig. 5b,e, where the RSS consistently reflects the expected distance-dependent FSPL behavior. The experiments show that increased distance from 6 to 17 cm reflects in SNR around 8 dB loss across all users, whereas the ideal FSPL shows 9.05 dB. We interpret this difference with SNR estimation error and instrumentation. In Fig. 5c,f, coarse term of normalized CFO strictly with at most ± 0.2 depends on hardware, whereas the fluctuations change with synchronization status where deviation of SDR-3 in DS-17B increased to 0.025 in SDR-3 NS-17B case. On the other hand, as a significant tradeoff, the IQ recovery with DPC costs loosen PLL performance. Figure 5e illustrates that the increased transmission distance exposes the system to stronger additive Gaussian noise effects. However, Fig. 5f indicates that no dominant phase noise impairment is observed. Therefore, the performance degradation is mainly attributed to power loss rather than severe oscillator instability, implying that increasing the received signal power can effectively compensate for the observed impairments. Similarly, the reduction in received power is also visible in the spectral analysis results. An additional observation is that, among the three SDR units, SDR-1 exhibits approximately 1 dB higher average SNR compared to the others, although the relative spectral gain is not necessarily higher. This suggests that variations originating from internal hardware components, such as amplifiers and filtering stages, can introduce up to 1 dB performance differences even among nominally identical SDR platforms. Lack of Doppler spread by the stationary environment, estimated coarse component of the normalized CFO for each user is same for both NS-6B and NS-17B. Locked phases in Figs. 5c and 7c show that the users are not coherent in phase, and existing PN drops the CFO estimation accuracy. While the coarse CFO is sourced by SDR impairments44, the fluctuations are reasoned by both impairment and two stage channel, which subject to fine CFO estimation. As an example, average normalized CFO on SDR-2 fluctuates around 0.01 while SDR-3 does -0.09. In result, RSS in NS cases are heavily exposed to ELF distance, ∼ 10 dB decrement is observed by increasing the distance. Especially in the NS case, the average SNR of SDR-3 drops to -12.12 dB with <-8 dB deep fading, and the highest BER of 0.43 among all the tests is observed, even though the spectrum relative gain is nearly 20 dB higher than the noise floor. With the use of in DS, the two-way splitter removes the sampling offset at the expense of increased insertion loss of nearly 8 dB. In the case of DS-6B, BER is constant 0 and >10 dB SNR is achieved. Smooth phase wrapping proves the converged DPPL and minimized CFO deviation in Fig. 5c,f. This measured contrast confirms the analytical results in (14) and (16), where the DS receiver was shown to eliminate any dependence on while the NS receiver retains cross-mixing terms that degrade synchronization. In this regard, small-scale (fine) CFO produce noise as depicted in Fig. 5b,e. With an assumption of all noise as PN, is measured, where Fig. 2 shows that such PN is below the symbol error boundary. This measured value falls within the PN-resilient region as given by Eqs. (19) and (20), confirming that the residual phase noise in DS-6B remains below the decision boundary π /M derived analytically. With the increased EFL distance from 6 to 17 cm, bit errors occur without an error floor in DS-17B. Despite having a low SNR of <5.5 dB, full synchronization achieves errorless frames in DS-17B. “On the other hand, while providing an ideal performance benchmark under shared LO conditions, the DS configuration, is not practically scalable to distributed high-frequency EF relaying systems due to LO distribution loss, phase instability, and synchronization overhead. Therefore, it is used solely as a laboratory upper bound, whereas the proposed DPC method targets realistic NS deployments.” Derivations in Section “Wireless sub-THz multiplexing analysis” shows that is the most dominant term of in Eq. (14), carrying out only the SDR PN. It is observed that the and terms are carried out over 180 GHz with multiplicatively on . Product results of the last term in highlight that, even if the emitter SDRs are PN-free, PN over FEX-1 and FEX-2 still appear in the IF downconverted signal. Nevertheless, the higher out-of-band components are filtered out by the IF filter of the SDR-4 before the real-time processing. Both Section “Wireless sub-THz multiplexing analysis” derivations and the demonstration prove that a synchronized BS-Core of SDR-THz EFL is not only PN resilient but also poses higher signal quality and is essential for error-free communication. Significance of is noted32 as GHz for each USRP B200. In contrast to and , Eqs. (8) and (14) show that is not aggregated by FEX, but carried out through all harmonics. Figure 7 with identical configurations in NS-6B and NS-17B shows the direct impacts of non-synchronized clocks. It is observed that the initial PN impact is different for sampling times of FEX-1 and FEX-2, leading to sampling clock offset (SCO) estimation, resulting in periodical bit errors after the demodulation shown in Fig. 7a. Compared to the DS case, Fig. 7b shows that the PSD remains the same, yet Fig. 7c shows that the low-frequency PN occurs as the phase wrapping reveals irregularities espeically in SDR-1 and SDR-3. Phase of SDR-1 wraps successfully with 2 π periods; however, the slope is not smooth. In SDR-2, PN is more severe where phase wraps with fluctuations. In SDR-3, the rate of the phase is high due to the coarse CFO. Nevertheless, PN is not as significant as SDR-2, where the CFO changes in each symbol. BER and constellation diagram with DPC is depicted in Fig. 8a,c, where SCO correction dropped the BER across all users. BER measurements are summarized in Table 1. According to these results, the proposed DPC framework reduces the BER in both the 6B and 17B measurement scenarios. However, the magnitude of the improvement varies across users due to hardware and channel dependent synchronization characteristics (Table 2). Digital post correction (DPC) impact on NS-6B in (a) and (b) and NS-17B in (c) and (d) measurements on SDR-4 from SDR-1 (orange), SDR-2 (blue), and SDR-3 (green) for (a) and (c) bit error rate and constellation diagrams, (b) and (d) locked phase with normalized CFO. The x- and y-axis titles are consistent and correspond to all measurements. Table 2Measured average BER comparison between conventional NS and NS w/DPC for all users.ScenarioUser 1User 2User 3NS-6B w/DPC0.0520.0610.068NS-17B w/DPC0.0440.0570.110NS-6B0.1400.1390.111NS-17B0.1250.0660.153 Measured average BER comparison between conventional NS and NS w/DPC for all users. It shows that DPC suppresses SCO where the residual PN is visible by the constellation diagram, where symbol phase error is predominantly below the decision boundary. However, residual SCO still exists as it resolves in 0.01 bit error floor. Nevertheless, dynamic gain adjustment leads to distortion in the phase wrapping where the normalized CFO is higher than in the NS and DS cases without DPC use. As expected, PN on users in NS-17B is more visible as SNR in this EFL is lower than in NS-6B. Therefore, in addition to the residual SCO errors, PN-related errors are further accumulated. Overall, the six tested configurations demonstrate that synchronization architecture, rather than link distance alone, is the primary determinant of EFL performance, with DPC offering a practical middle ground between the non-synchronized and distributedly synchronized extremes. NS-6B and NS-17B are dominated by SCO-induced degradation, with average BER of 0.130 and 0.115, respectively, and SNR as low as -12.12 dB for the worst-case user. DS-6B is the strongest configuration, achieving zero measured bit errors at 11–13 dB SNR, while DS-17B trades SNR for distance (falling below 5.5 dB) yet avoids the persistent bit-error floor seen in the NS cases. Applying DPC to the NS cases lowers the average BER to 0.060 (NS-6B) and 0.070 (NS-17B), up to 0.088 for the best-case user, at the cost of a higher residual CFO. These results confirm that DS remains the strongest configuration, DPC offers a practical partial correction for NS deployments, and both BER and CFO are governed by the same PN/SCO mechanisms. Conclusion Extend-and-Forward (EF) relaying is introduced for sub-THz extended fronthaul links (EFL), toward integration with existing sub-6 GHz infrastructure. In this framework, frequency-division multiplexed (FDM) multi-user signals operating in the sub-6 GHz band are wirelessly extended to the sub-THz domain. The backbone of EFL signal modeling is established, addressing the primary challenge of phase noise (PN) induced symbol error probability and identifying operable frequency regions. The impact of clock synchronization on sampling offset is demonstrated with real-time both in the signal and symbol domains using a 180 GHz EFL setup on an SDR-THz testbed. Compared to the non-synchronized case, distributed clock synchronization across sub-THz up and down-converters achieves lower bit error rates and higher SNR, even under low spectral power conditions. Our proposed DPC algorithm further reduces the bit error rate by up to 0.088, despite the presence of a residual clock offset that introduces a bit error floor, enhancing performance over non-synchronized cases. Prospective use cases are presented, including 1) EF repeaters for EFL to enable interference-free relaying, and 2) an ultrabroadband sub-THz backhaul link for multiple access. We note that the present validation is a proof-of-concept and has limitations regarding the tested link distances, the number of simultaneous users, and system-level applicability toward expansion to a full multi-cellular or multi-hop deployment. Furthermore, the DS configuration, which achieves the best measured performance, relies on a shared local oscillator across FEX-1 and FEX-2, and therefore imposes precise synchronization requirements that are not readily scalable to distributed, geographically separated relay deployments. The future work directions below are aimed directly at addressing these four limitations: distance, user scale, system-level applicability, and synchronization. As future work, we plan to investigate the impact of misalignment using a heterogeneous SDR-THz testbed that integrates SDR technologies of varying quality and data rates, aiming to explore the performance limits beyond the tested distances. From the implementation side, we aim to extend the current set-up by active RF components, transitioning from a conceptual study toward real-time applications, moving toward the system-level applicability noted above. This will provide a means to quantify the impact of the impairments, including the synchronization overhead inherent to the DS configuration. Other open issues include exploring the EFL setup in a multi-cellular scenario where a higher number of users are considered, addressing the multi-user scalability limitation discussed above. Author contributions E.G. wrote the main manuscript text. E.G. and L.T. performed the experiments. J.F. and G.K. reviewed and refined the experimental procedures. All authors reviewed the manuscript. Funding This work was supported in part by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant program and also by John R. Evans Leader Fund (JELF) of Canada Foundation of Innovation (CFI). A public demonstration of this study was presented at ICC 2025 in Montreal, Canada45. Data availability The datasets and collected measurements used and/or analysed during the current study available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets and collected measurements used and/or analysed during the current study available from the corresponding author on reasonable request.