Running a Full-Duplex Receiver Hot: Treating ADC Clipping as a Mask

A full-duplex radio transmits and receives on the same frequency at the same time. Its own transmission leaks into its receiver 50 to 110 dB stronger than the signal it actually wants to hear. Because of that leakage, the receiver's analog-to-digital converter (ADC) has to be set conservatively. Its gain is backed off so that the loud leakage never hits the converter's maximum code. The weak wanted signal then sits in the bottom few bits and is buried in quantization noise. This note proposes the opposite operating choice: turn the gain up, let the rare leakage peaks clip, and treat every clipped sample as a masked position whose location and sign are known. A masked-reconstruction model is a network trained to fill in deliberately hidden parts of a signal from the visible context. Here it is trained with masks shaped like real saturation events and conditioned on the transmit waveform, which the radio knows exactly. Its job is to estimate the leakage from the unclipped samples so it can be subtracted, while the receiver decodes the wanted signal with the clipped samples treated as erasures.

The opportunity matters most for low-resolution converters (4 to 6 bits), which wideband millimeter-wave arrays use to save power. At those resolutions, avoiding saturation forces the transmitter to throttle power or steer its beam away from the user. If deliberate clipping plus erasure-aware processing recovers even one effective bit for the wanted signal, it relaxes a constraint that currently shapes the whole transceiver.

Core Claim

Our claim is narrower and more physically honest than "reconstruct the weak signal from saturated samples." At a clipped sample the wanted signal is genuinely lost, because it is tens of dB below the leakage and the converter output is pinned at the rail. Clipping therefore acts as an erasure channel for the wanted signal, with known erasure positions. The wanted signal survives only through its own redundancy: it is band-limited and oversampled, and it carries pilots and coding.

The difficult and underexplored part is the leakage. In full-duplex work the leakage is called self-interference (SI), and it must be estimated, including the power amplifier's nonlinear distortion, from data in which clipping has censored exactly the high-amplitude samples. Those peaks are where amplifier compression lives. Dropping clipped samples therefore discards the most informative data for nonlinear SI estimation. The clipped samples still carry an inequality: the true value exceeded the rail, with known sign.

We claim two things. First, an erasure-aware receiver run at higher gain beats the conventional backed-off receiver in achievable rate for the wanted signal when the ADC has 4 to 6 bits. Second, a bidirectional masked model conditioned on the transmit sequence and on the rail-sign of clipped samples estimates nonlinear, multi-tap SI from censored data better than polynomial cancellers that simply drop clipped rows. It also needs fewer calibration samples than a censored-likelihood polynomial fit while approaching it in accuracy.

The size of the possible gain follows from two textbook facts. Each 6.02 dB of extra gain buys roughly one bit of effective resolution for the wanted signal. For a complex-Gaussian-like OFDM leakage with RMS amplitude $\sigma$, the fraction of samples whose envelope exceeds a clip level $A$ is

$$ P_{\text{clip}} = \exp\!\left(-A^2/\sigma^2\right). $$

This formula comes from the Rayleigh distribution of the envelope, and it is approximate for per-rail I/Q clipping. Lowering the clip margin from 9 dB to about 4.8 dB above RMS raises the clipped fraction from roughly 0.04% to roughly 5%. That buys about 4 dB, or two-thirds of a bit. Whether this trade pays off depends entirely on how well the SI can still be estimated and the erasures absorbed. That is the empirical question.

Where this sits in prior work

Several papers motivate the regime. Full-Duplex MIMO Relaying: Achievable Rates under Limited Dynamic Range models hardware distortion as noise proportional to signal power and finds a critical interference-to-noise ratio beyond which full duplex should fall back toward half duplex. It treats saturation as something to avoid, never as something to process. Hybrid Beamforming for Millimeter Wave Full-Duplex under Limited Receive Dynamic Range builds per-LNA and per-ADC power constraints into a convex beamformer. It reports that 12-bit ADCs are nearly optimal while 4-bit ADCs force throttled power or redirected beams. That paper defines our target regime, and its design is the "avoid saturation spatially" alternative we compare against. Millimeter-Wave Full Duplex Radios: New Challenges and Techniques gives a link budget needing about 85 dB of total isolation, of which only about 15 dB falls to the digital stage. Our work asks whether the analog stages can be allowed to leave more residual leakage.

Two papers supply methods. Sample-Efficient Self-Interference Cancellation for In-Band Full Duplex Radios via In-Context Learning recasts digital SI calibration as prompting a GPT-2-style model with pilot I/Q tokens. It uses no positional encoding, bidirectional attention over the prompt, per-context gain normalization, and context decimation to avoid a copy shortcut. It reaches its floor with 64× fewer calibration samples than the best model-based baseline on USRP N210 hardware. It assumes no clipping, single-tap channels, and dedicated pilots. We reuse its tokenization, gain normalization and decimation. We change three things: the objective becomes masked reconstruction, masks follow saturation, and calibration may use payload periods with the wanted signal present. Masked pre-training on physical signals, as in Brant: Foundation Model for Intracranial Neural Signal and BrainLM: A foundation model for brain activity recordings, shows that bidirectional models trained to fill 40 to 90% hidden spans learn usable structure. Brant reports an imputation MSE of 0.0814 at 40% missingness. Those models treat masked positions as pure erasures and visible ones as clean. We add a third token state, "clipped with known sign," which carries an inequality rather than a value.

Two papers constrain the design. Direct Reconstruction of Saturated Samples in Band-Limited OFDM Signals restores clipped OFDM samples by sinc-basis interpolation from nearby unclipped samples. It reduces BER from 0.14 to near zero, but it fails at high clip density and knows nothing about a known, overlaid interferer. It is the natural classical erasure handler for the wanted signal and a baseline. Learning to reconstruct from saturated data: audio declipping and high-dynamic range imaging learns declipping without ground truth, but its identifiability argument requires scale invariance. Amplifier nonlinearity breaks that assumption, because scaling the drive changes the distortion. We therefore do not use its equivariance loss. We get supervision instead by capturing at the radio's native resolution and emulating clipping, low-resolution converters in software, so targets are real samples. Are EEG Foundation Models Worth It? Comparative Evaluation with Traditional Decoders in Diverse BCI Tasks found that large pretrained models do not uniformly beat tuned classical decoders. We take that as a direct warning, which is why strong classical clipping-aware pipelines are the primary baselines rather than straw men.

Research Program

The input to the model consists of short windows of received baseband samples. Each window is stacked as real and imaginary parts and paired with the transmit samples spanning the SI channel's delay spread, so that multi-tap channels are expressible. Each received sample carries a three-way flag: unclipped, clipped at the positive rail, or clipped at the negative rail, assigned separately on the I and Q components.

The encoder is bidirectional. Instead of the absolute position-free attention of the in-context paper, it uses relative position only within the delay-spread window, because multi-tap channels make ordering matter. Pre-training masks a mixture of random spans and saturation-shaped masks. The saturation-shaped masks are produced by actually lowering the emulated rail on the transmit-driven leakage, so masks cluster on amplifier compression peaks the way real clipping does.

The training target is the SI component at every sample. At clipped samples, the loss combines regression against the native-resolution reference with a hinge penalty when the estimate violates the known rail inequality. At deployment that hinge is the only signal available, so we train with it.

Downstream, the estimated SI is subtracted at unclipped samples. The wanted OFDM signal is then demodulated by erasure-aware least squares over its occupied subcarriers, with the clipped positions removed. Soft outputs go to a decoder with erasure flags. We deliberately keep the wanted-signal demodulator classical. This makes any gain attributable to SI estimation from censored data and to the operating-point shift, not to an opaque end-to-end decoder.

Data comes from three sources. The first is OpenDPD's real power-amplifier measurements at 200 MHz, for nonlinearity at wider bandwidth than the 20 MHz the in-context work tested. The second is over-the-air USRP N210 captures in the style of the in-context paper; we use theirs if released, and otherwise collect from several nominally identical units to test cross-device transfer. The third is SI channels with multiple taps, built by convolving these with tapped-delay profiles of increasing spread. The wanted signal is added at controlled power, and the ADC is emulated at 4, 6, 8 and 12 bits with a swept gain setting. The known limit of software emulation is that it captures hard ADC saturation but not analog LNA compression. We state this as a scope boundary and test one hardware gain sweep on real radios to check it.

Baselines

The naive reference is the conventional backed-off receiver. The ADC gain is set so that clipping occurs with probability below about $10^{-4}$, and SI is cancelled by a memory polynomial with conjugate terms, in the spirit of An Augmented Nonlinear LMS for Digital Self-Interference Cancellation in Full-Duplex Direct-Conversion Transceivers. That paper reports up to 70 dB of digital attenuation and stable SINR up to 25 dBm transmit power in 802.11-style simulation. This baseline defines what "run hot" must beat. It is run on the identical captures, emulated converter and calibration budget.

The strongest existing learned method is the in-context canceller of Sample-Efficient Self-Interference Cancellation for In-Band Full Duplex Radios via In-Context Learning. Its reported results are 4.4× better sample efficiency than alternating least squares in simulation and a 64× reduction in calibration samples on USRP hardware. It has no reported result under ADC clipping or multi-tap channels; we found none on Lacuna for clipping-aware learned SI cancellation. We run it two ways, at backed-off gain and at high gain with clipped samples dropped from its prompt. Model size, training data, and calibration samples are matched to ours.

The strongest classical competitor on our specific question is a clipping-aware polynomial pipeline. The polynomial is fit by censored maximum likelihood, a Tobit-style objective using the rail inequality at clipped samples. It is paired with the same erasure-aware OFDM demodulator, and with sinc interpolation from Direct Reconstruction of Saturated Samples in Band-Limited OFDM Signals as an alternative erasure handler. Its polynomial order and memory depth are tuned on a held-out split with the same tuning budget as our model's hyperparameters.

Beating these baselines means two things. First, a higher achievable rate and lower error-vector magnitude for the wanted signal at 4 and 6 bits, at equal calibration samples, across at least two interference-to-noise ratios above the critical regime. Second, matching censored-likelihood accuracy with substantially fewer calibration samples, or exceeding it under drift and long delay spreads.

Two ablations isolate the key ingredient. The first trains the same network with random masks only, and the second removes the rail-sign flag and hinge loss. If neither ablation hurts, saturation-shaped masking is not what drives the result. A native-resolution, unclipped oracle bounds everything from above.

Why not just…?

Why not just buy more ADC bits or more analog cancellation? At 12 bits the problem largely disappears, as Hybrid Beamforming for Millimeter Wave Full-Duplex under Limited Receive Dynamic Range shows. Analog cancellers such as the frequency-domain equalization design in Demo abstract: Full-duplex with a compact frequency domain equalization-based RF canceller reach 95 dB over 20 MHz. Our proposal does not argue against either. It targets the regime where converter power forces 4 to 6 bits, or where analog isolation falls short of the 85 dB budget. The bit sweep reports where the gain vanishes. If it vanishes already at 6 bits, the idea has a narrow niche and we will say so.

Why not just fit a polynomial canceller with a censored-likelihood objective and decode with erasures? This is the real threat. It may capture most of the benefit, because the physics of the problem is mostly low-order. Our experiments treat it as the main baseline. The learned model's case rests on three regimes where polynomial fits are known to struggle: few calibration samples, which is the in-context paper's 64× finding; long multi-tap delay spreads, where the parameter count grows; and drift across devices. If the censored polynomial matches us everywhere, the operating-point result still stands as a contribution with classical tools, and we will report that the learned model is unnecessary.

Why not just reduce peak-to-average power at the transmitter so peaks never clip? Clipping or tone reservation on transmit lowers the leakage peaks, but it spends transmit power, spectral purity or subcarriers on every symbol. Receiver-side erasure handling pays only when clipping actually occurs. We include a transmit peak-reduction baseline tuned to equal clip probability and compare end-to-end rate for both directions of the link.

Why not just drop clipped samples from the existing in-context canceller's prompt? That is one of our runs. Our hypothesis is that dropping is biased, because clipping removes the compression peaks that reveal the amplifier's nonlinearity. The rail-sign ablation tests directly whether the inequality information is worth anything beyond the dropped-row version.

Decisive Checks

The first check is a curve of achievable rate against emulated ADC gain, for each resolution, on OFDM leakage with OpenDPD amplifier models and multi-tap channels. The proposal succeeds only if, at 4 and 6 bits, the rate peaks at a gain where clipping is in the percent range, and that peak exceeds the backed-off optimum of the conventional receiver. The second check compares SI residual power against calibration-sample count for our model and the censored polynomial at high gain, as delay spread grows from one tap to many. The third check is cross-device transfer on USRP captures under high-gain operation. It tests whether gain normalization still prevents the roughly 10 dB collapse the in-context paper saw without it. The fourth check is a real-radio gain sweep, to verify that conclusions drawn from software-emulated clipping survive analog compression. If the rate peak moves back to backed-off gain on hardware, the emulation was misleading and the claim fails.