When two cellular networks share one band without talking to each other, they take turns. One network goes silent in agreed time slots so the other can transmit, and any slot where both transmit anyway is written off as a collision. We propose to stop writing those slots off. The receiver separates the overlapping signals, using a separator trained only on recordings that the turn-taking scheme already produces for free. While network A is muted, a receiver in network B hears only B. While B is muted, it hears only A. Both recordings arrive through the receiver's own real radio channels. Adding these clean recordings together, with random gains and offsets, gives realistic training mixtures. The separator never needs to know the other operator's configuration, and it never needs paired ground truth.
This matters because measured spectrum is mostly idle. Surveys summarized in a review of cognitive radio networks report that over 60% of licensed spectrum below 6 GHz is unused at any given place and time. The best uncoordinated sharing schemes still treat every overlap as lost capacity. Girmay et al. share a band between LTE and 5G New Radio (NR) by muting subframes and use a convolutional classifier to label each window as LTE, NR, or "Overlap." They then back off from overlaps. If overlaps become recoverable, a scheduler can mute less and tolerate some deliberate overlap. That converts the remaining idle time into throughput.
Core Claim
In muting-based uncoordinated sharing, the muting schedule itself is a source of clean single-network signal captures. A single-antenna separator trained only on synthetic sums of those captures can recover a large fraction of overlapped subframes. A scheduler that knows how reliable this recovery is can then reduce muting and exceed the efficiency of pure avoidance. The new element is this loop: sharing produces the separator's training data, and the separator changes how much sharing must rely on avoidance.
The training mixture is built as
$$\tilde{y}(t) = r_B(t) + a\, r_A(t-\delta)\, e^{j2\pi \nu t} + n(t).$$
Here $r_B$ is a capture of network B alone, recorded while A was muted. $r_A$ is a capture of A alone, recorded while B was muted. $a$ is a random complex gain that sets the power ratio and phase between the two. $\delta$ is a random timing offset and $\nu$ a random residual frequency offset, covering the fact that the two operators are not perfectly aligned. $n$ is added noise. The separator is trained to output both received components, $r_B$ and the shifted copy of $r_A$, from $\tilde{y}$. It outputs signals as they arrive at the antenna, not as they were transmitted. The standard demodulator that follows already equalizes the channel using the network's own reference signals, so the separator never has to estimate the cross-operator channel.
Positioning Against Prior Work
The closest separation work is the RF Challenge. It trains one-dimensional UNet and WaveNet separators on single-antenna co-channel mixtures. It reports bit error rates (BER) up to 100 times lower than a linear minimum-mean-square-error filter, and BER near $10^{-3}$ at -18 dB signal-to-interference-plus-noise ratio (SINR), where the linear filter is at 0.5. That paper supplies our architecture and our sanity benchmark. It differs from this proposal in three ways. Its training data is simulator-generated, it requires knowing the signal-of-interest class in advance, and it is not connected to any spectrum-access decision.
ArrayDPS motivates the idea that priors learned from clean single-source data can drive separation without paired mixtures. It uses a diffusion model, a generator that learns by gradually denoising random noise into samples, trained on clean speech. On 3-channel SMS-WSJ it reaches 17.2 dB signal-to-distortion ratio against 13.4 dB for classical Independent Vector Analysis (IVA). It also constrains the design: it needs hundreds of sampling steps, which does not fit a 1 ms subframe. We therefore keep the "train only on clean single sources" principle but use a fast feed-forward separator rather than iterative posterior sampling.
Torkkola showed that building modulation knowledge into blind separation cuts convergence from 800–1,200 samples to 80 in fading channels. His method assumes a multi-antenna, narrowband mixture. That rules it out for the wideband, single-antenna regime we target, though it reappears below as a baseline in the multi-antenna setting.
Vargas et al. treat radar–communication coexistence as convex dual-blind deconvolution. Their solver is far too slow for cellular rates.
Girmay et al. supply the system we modify: muting, the three-class overlap detector (above 98% accuracy at SNR of 5 dB or more), and the lookup-table scheduler.
Research Program
The first component is capture harvesting. We reproduce Girmay et al.'s LTE–NR setup with the Vienna LTE and MATLAB 5G simulators. Each receiver logs baseband in-phase/quadrature samples from windows its classifier labels "LTE only" or "NR only," and keeps a rolling buffer of these captures. Because the captures come from real scheduling, they reflect actual channel draws, resource-block occupancy, and impairments at that receiver.
The second component is the separator: an RF-Challenge-style UNet, sized so that inference on one subframe completes in well under 1 ms on a target accelerator. It is trained on remixes built from the buffer. When the classifier flags a window as overlap, the separator runs before the standard LTE and NR decoders. A light online fine-tuning step refreshes the model as new captures arrive, so it tracks channel drift.
The third component is overlap-aware scheduling. We measure the separator's block error rate as a function of power ratio and SINR, and fold that into the capacity lookup table. The binary search then chooses both a muting fraction and a tolerated-overlap fraction, rather than muting alone.
The research questions follow from these components. First, how much of the recoverable gain comes from training on captures from the receiver's own location rather than on simulated waveforms? Second, how stale can captures be before channel drift erodes separation? Third, does the scheduling change turn per-subframe recovery into end-to-end spectrum efficiency, or does it only reshuffle errors between the two networks?
Baselines
The do-nothing reference is the Girmay et al. scheme unchanged. Overlaps are discarded, and the reported operating point is a 20.7% spectrum-efficiency gain on NR and 11.8% on LTE over static allocation, 13.5% and 8.3% throughput gains, and buffer overflow falling from 14.6% to about 9% (Girmay et al.). We rerun it in our reproduction with identical traffic, channels, and synchronization. Beating it means higher spectrum efficiency on both bands at equal or lower buffer overflow, with no regression for the network that previously "won" the slots.
The strongest separation competitor is a supervised RF-Challenge UNet trained on simulator-generated LTE and NR mixtures (RF Challenge). It has the same architecture and parameter count, the same number of training mixtures, and the same tuning budget as our model. The only difference is whether training components come from the simulator or from in-situ captures. On the RF Challenge dataset itself, the reported figure is BER of about $10^{-3}$ at -18 dB SINR. We first confirm that our implementation reaches this before any comparison. On LTE/NR overlaps, no reported result was found on Lacuna, so both models are measured here. Beating this baseline means lower block error rate on overlapped subframes under test channels not seen in the simulator's training distribution.
The classical baseline is successive interference cancellation. The standard receiver decodes the stronger signal, re-modulates it, subtracts it, and decodes the residual. It requires no learning, and no reported result was found on Lacuna for it in this setting. In the two-antenna variant we add frequency-domain IVA, which separates each subcarrier independently and links them to resolve the ordering of outputs. This is the modern form of the multi-antenna blind separation that Torkkola and ArrayDPS build on.
The key ablation removes the receiver's own captures, which recovers the simulator-trained baseline. A second ablation keeps the separator but leaves Girmay's scheduler unchanged, isolating the contribution of overlap-aware scheduling.
Why not just…?
Why not just train on simulated LTE and NR, since both are fully specified standards? This is the most serious threat. Modulation formats are not really unknown here. What the receiver does not know is the cross-operator channel, timing and frequency offsets, the other operator's resource allocation, and hardware non-idealities. We expect in-situ captures to win when the test channel departs from the simulator's channel models. The simulator-versus-capture ablation, at equal data size, settles this directly. If simulated training matches capture training, we will report it and the contribution reduces to the scheduling loop.
Why not just use successive interference cancellation? At large power imbalance it may match a learned separator at much lower cost. We therefore sweep the power ratio from 0 to 20 dB. We expect the learned separator to matter mainly near equal power, where cancellation fails because neither signal decodes first. If cancellation dominates everywhere the scheduler would actually operate, the scheduler should use cancellation instead, and we will say so.
Why not just mute more aggressively or schedule better? Muting is strictly orthogonal, depends on GPS-level synchronization, and cannot recover unintended overlaps from bursty traffic (Girmay et al.). The scheduling ablation shows whether recovered overlaps add capacity beyond what any muting fraction in the original table achieves.
Why not use multi-antenna blind separation? With as many antennas as sources, per-subcarrier IVA may work well, and we test it in the two-antenna variant. Our main regime is single-antenna user equipment and IoT receivers, where such methods cannot separate two sources at all.
Proposed Experiments
We begin by checking that our UNet matches the RF Challenge figure on its own data. We then run the full LTE–NR reproduction, reporting block error rate on overlapped subframes against SINR and power ratio. At the system level we report spectrum efficiency, throughput, and buffer overflow for both bands. We also record separator latency per subframe.
The drift experiment trains on captures that are 10 ms to several seconds old relative to the overlap being separated, at several UE speeds. It measures how quickly accuracy decays and whether online fine-tuning restores it.
The decisive check has two parts. First, capture-trained separation must beat the simulator-trained UNet on held-out channel models. Second, overlap-aware scheduling must raise spectrum efficiency above the 20.7% (NR) and 11.8% (LTE) avoidance baseline at no worse buffer overflow. If the first part fails, the in-situ-capture claim fails. If the second fails, per-subframe separation does not translate into usable spectrum.
References
- Intelligent and efficient development of wireless networks: A review of cognitive radio networks
- Enabling Uncoordinated Dynamic Spectrum Sharing Between LTE and NR Networks
- RF Challenge: The Data-Driven Radio Frequency Signal Separation Challenge
- ArrayDPS: Unsupervised Blind Speech Separation with a Diffusion Prior
- Blind Separation of Radio Signals in Fading Channels
- Joint Radar-Communications Processing from a Dual-Blind Deconvolution Perspective