A full-duplex radio transmits and receives on the same frequency at the same time. Its own transmission leaks into its receiver at up to a hundred billion times the power of the distant signal it is trying to hear. Working systems remove this leakage in stages. First, an analog canceller built from a bank of tunable bandpass filters (a "frequency-domain equalization" canceller) removes about 50 dB before digitization. Then a digital model of the leftover interference removes another 40–45 dB. The digital stage is fragile because the leftover interference drifts: amplifier behavior changes with temperature, and capacitor values in the analog canceller shift. Today the digital stage re-learns by pausing data to send calibration pilots. Between pilots it adapts on a received signal that also contains the desired incoming signal, which corrupts the estimate.
Particle physicists handle the same logical problem when they estimate backgrounds. They never try to measure the background inside the region where the signal lives. They simulate a small family of possible background shapes, measure the background in "control regions" where signal cannot occur, such as mass sidebands, and fit the few free normalizations jointly with the signal region. We propose to carry this recipe over to full-duplex radio. The unoccupied frequencies at the edges of an OFDM transmission can never contain the desired signal, but they do contain the transmitter's nonlinear distortion. They are therefore a natural sideband. We fit a low-dimensional, simulation-derived family of residual-interference shapes jointly to these guard frequencies and to the occupied band. The goal is to track drift continuously, with no pilots and no contamination from the desired signal.
Core Claim
Behind a frequency-domain analog canceller, the leftover interference varies over time along only a few physical directions. These are drift of each filter tap's center frequency and quality factor, amplifier compression, and I/Q imbalance, which is the gain and phase mismatch between the receiver's two signal paths. A Monte Carlo library of circuit and amplifier variations can compress these into a handful of basis shapes. Because so few coefficients need estimating, two things follow. The guard-band observations pin down the nonlinear coefficients without in-band pilots. The linear coefficients can be fit in-band while treating the desired signal as a modeled nuisance. We claim this combination holds digital cancellation close to a freshly pilot-calibrated canceller over long drift episodes, at zero pilot overhead. We also claim it degrades much less than least-squares estimation when the digital interference-to-noise ratio drops below the roughly 20 dB threshold identified by Li et al..
What is new is the pairing of a physically defined signal-free observation region with a small set of simulated template coefficients. Neither ingredient exists in current self-interference cancellation. The closest work is Balaji et al.. They pre-train a Transformer on simulated amplifier and channel distributions and adapt it in-context from calibration pilots. That shares our "simulation prior plus online grounding" structure, but it grounds on in-band pilots and models the analog front end as a single direct leakage tap. We ground on signal-free spectrum and explicitly model the analog canceller's drift.
Research Program
Template library. The first component is a simulated library of residual-interference shapes. We start from the equivalent-circuit models of the two-tap canceller in Chen et al. and Kohli et al.. These papers configure the filter taps' center frequency, quality factor, amplitude and phase by offline non-convex optimization against a measured antenna response. We take their optimized operating point and draw Monte Carlo perturbations of it. The perturbations include component drift, reflecting the capacitor quantization steps they identify as the limit on cancellation depth, and antenna-impedance changes. For each draw we compute the linear residual across the band. We combine these with amplifier nonlinearities drawn around the measured Doherty amplifier behavior in the OpenDPD benchmark used by Balaji et al., and with I/Q imbalance of the widely-linear form in Li et al.. Principal components of the simulated residuals, taken separately for linear and nonlinear parts, give the templates. The number kept is chosen by held-out reconstruction error.
Joint fit. The second component is a joint likelihood fit that runs on every received frame. On guard and edge subcarriers, and on any sampled bandwidth beyond the occupied band, the received value is modeled as template-weighted nonlinear residual plus thermal noise. This is the control region. On occupied subcarriers, the model is linear plus nonlinear residual, plus the desired signal with its constellation and channel treated as nuisance parameters, plus noise. The template coefficients get priors from the simulation spread. They are updated recursively across frames as a Kalman-style tracker.
The physics analyses that supply this method are Acero et al., who float Monte Carlo background normalizations against a data-driven control sample, and Sahoo et al. and Ablikim et al., who anchor background shapes in sidebands. Those analyses fit Poisson event counts. We fit complex Gaussian baseband samples whose interference is a known function of our own transmitted data. The input is known exactly and only the coefficients drift, so the physics recipe is better posed here than in its home field.
Main risk: extrapolation. The central risk is the one physicists call control-to-signal extrapolation. Linear leakage has no energy on unused subcarriers, so the guard bands carry no information about it. Only the spectral regrowth from amplifier distortion, plus I/Q-image energy mirrored from edge subcarriers, reaches the guard bands. With 20 MHz Wi-Fi-like framing, only a minority of subcarriers are unused. The design therefore deliberately splits the job. The control region constrains the nonlinear coefficients. The in-band fit constrains the few linear drift coefficients. We will explicitly measure how well guard-band estimates predict the in-band nonlinear residual. Where the digitizer's sampling bandwidth exceeds the occupied bandwidth, the extra spectrum enlarges the control region, and we will report results both with and without it.
Baselines
Naive reference. The digital canceller is calibrated once with pilots at the start of a session and then frozen. This shows how much drift actually costs. A second reference is the same canceller recalibrated with pilots at a fixed interval, with the airtime spent on pilots charged against throughput.
Hardware-validated pipeline. This is the multi-tap least-squares digital canceller behind the frequency-domain analog canceller in Kohli et al.. They report 95 dB total cancellation, 52 dB of it analog and 43 dB digital, over 20 MHz on the COSMOS testbed. We rerun it on the same captures with the same pilot budget. Beating it means matching its 43 dB digital cancellation at zero pilots over a drift episode where the frozen version loses several dB.
Strongest adaptive competitor. This is the in-context Transformer of Balaji et al.. On USRP N210 captures it reached its error floor with about 64× fewer calibration samples than the strongest model-based baselines, and in simulation it was about 4.4× more sample-efficient than widely-linear alternating least squares. The paper reports no digital cancellation figure in dB on COSMOS data, so for that setting there is no reported result found on Lacuna. We will retrain it on the same simulation library we use for templates, so both methods see identical prior information, and give it its preferred pilot context. Beating it means lower residual interference per unit of airtime overhead, and in particular being within 1 dB of it when it is allowed pilots and we are not.
Adaptive filter. The augmented nonlinear LMS of Li et al. reports up to 70 dB digital attenuation and stable signal quality at 25 dBm transmit power in simulation. We run it adapting continuously during reception, with pre-whitening and a tuned step size given the same hyperparameter search budget as our method.
Ablations. Three ablations isolate the key ingredients. The first uses guard-band fitting with an unconstrained memory polynomial instead of templates. The second uses templates fit in-band only, with no control region. The third uses templates drawn from a deliberately mismatched simulation, which measures how much the method depends on template fidelity.
Why not just…?
Why not just adapt on the full received signal and subtract the decoded desired signal (decision-directed estimation)? This is the most serious alternative, because a correctly decoded incoming signal can be removed and the leftover interference re-estimated. It fails exactly when it is needed. Once drift makes decoding errors frequent, errors feed back into the estimate. And below 20 dB interference-to-noise ratio, Li et al. show least-squares estimation degrades even with clean references. We include it as a baseline and sweep the incoming signal-to-noise ratio. If it matches our method at every operating point, the control-region idea adds nothing, and we will say so.
Why not just use the in-context Transformer with more frequent short pilots? Balaji et al. already make pilots cheap. Fast enough drift, though, still forces periodic interruptions. Their evaluation also used single-tap leakage and needed a gain-normalization step, without which performance dropped about 10 dB across devices. Our airtime-normalized comparison under drift settles this. A Transformer could also take guard-band samples as context, so we will run that variant too. If it wins, the contribution becomes the guard-band control region rather than the likelihood fit.
Why not just fit a standard memory polynomial on guard bands, with no simulation? This is a real threat to the template half of the claim. A memory polynomial's many coefficients are poorly determined by a few low-power guard subcarriers, and it cannot say anything about linear drift. The first ablation measures this directly. If unconstrained guard-band fitting performs as well, we will drop the templates and report the simpler method.
Why not just push more cancellation into the analog stage? Two-tap prototypes deliver about 50–52 dB (Chen et al.). Deeper analog cancellation lowers the digital interference-to-noise ratio, which is precisely the regime where Li et al. show digital estimation weakens. So a better analog canceller increases, rather than removes, the need for a low-noise digital estimator.
Proposed Experiments
Data. The primary data are the COSMOS full-duplex captures from Chen et al. and Kohli et al.. These contain transmitted baseband and post-analog-cancellation residual for 20 MHz OFDM frames. Because captures are made with the incoming signal absent, we add separately recorded incoming signals at controlled strength, which gives exact ground truth. The USRP N210 multi-device captures of Balaji et al. test cross-device transfer. If hardware access allows, we will induce live drift on the canceller PCB with thermal cycling and a hand near the antenna.
Metrics. We measure digital cancellation in dB and residual power above the noise floor. We measure the incoming signal's error-vector magnitude and bit error rate. We also measure effective throughput with pilot airtime deducted, tracked over time within each drift episode.
Decisive checks. Three checks decide the question. The first is extrapolation: with the incoming signal off, nonlinear coefficients estimated only from guard bands should predict the in-band nonlinear residual to within 1–2 dB of a pilot-trained oracle. If they miss by much more, the control-region premise fails. The second is low interference-to-noise ratio: we vary the analog attenuation so the digital ratio spans 10–30 dB. Our method's cancellation loss relative to an oracle should grow more slowly below 20 dB than least squares, since it estimates a handful of coefficients instead of a full filter. The third is tracking: over drift episodes, our zero-pilot method should stay within 1 dB of the periodically recalibrated baseline while beating it on effective throughput. The template-mismatch ablation bounds how much of any gain comes from simulation fidelity rather than from the control region itself.
References
- Impact of the Amount of RF Self-interference Cancellation on Digital Self-interference Cancellation in Full Duplex Communications
- Sample-Efficient Self-Interference Cancellation for In-Band Full Duplex Radios via In-Context Learning
- Wideband Full-Duplex Wireless via Frequency-Domain Equalization: Design and Experimentation
- Design and Testbed Deployment of Frequency-Domain Equalization Full Duplex Radios
- An Augmented Nonlinear LMS for Digital Self-Interference Cancellation in Full-Duplex Direct-Conversion Transceivers
- Measurement of neutrino-induced neutral-current coherent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math> production in the NOvA near detector
- Search for lepton-number- and baryon-number-violating tau decays at Belle
- Search for the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>e</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>e</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:mi>ϕ</mml:mi><mml:msub><mml:mi>χ</mml:mi><mml:mrow><mml:mi>c</mml:mi><mml:mn>1</mml:mn></mml:mrow></mml:msub><mml:mo stretchy="false">(</mml:mo><mml:mn>3872</mml:mn><mml:mo stretchy="false">)</mml:mo></mml:math> process at BESIII
- Experimentation with Full-Duplex Wireless in the COSMOS Testbed