Receivers for massive Internet-of-Things uplinks increasingly drop the assumption that every device's packet arrives at the base station in lockstep. Instead, they search over a window of possible arrival delays to find which devices are active and when they started transmitting. These receivers are almost always judged in simulation, using channels drawn from idealized statistical models, and the field is now pushing toward physics-based channel generators that solve the wave equation in a computer model of the site. This note argues that the pairing has a specific, unstudied failure mode. Any wave solver must cut the infinite world down to a finite box, and the artificial edge of that box reflects a little energy back inside. That echo arrives at a clean, deterministic extra delay. A receiver whose job is to find delayed copies of pilot signals can read it as a late-arriving user or as a real multipath tap. The receiver is then scored against channels that contain energy no real building would produce.
The proposal is to measure this effect, predict it, and remove it. We will build a room-scale, full-wave channel generator from a pseudospectral solver, meaning one that computes spatial derivatives exactly with fast Fourier transforms rather than with local finite-difference stencils. The solver will have absorbing layers whose leakage can be dialed up or down, plus a reference version with no reachable boundary at all. Using it, we will derive and test how much boundary leakage an asynchronous detector can tolerate. We will also check whether that leakage silently inflates the inter-user channel correlation that real measurements show is so damaging, and we will validate the surviving channels against measured data. The practical output is a "reflection budget": how well the simulated boundary must absorb, as a function of antenna count, pilot length and signal-to-noise ratio, before simulated results on asynchronous access can be trusted.
Core Claim
Asynchronous activity-and-delay detectors accumulate energy over every receive antenna and every pilot symbol. A spurious echo from an artificial boundary is accumulated in exactly the same way. Let $r$ be the residual amplitude reflection coefficient of the absorbing layer for a given incidence angle, $M$ the number of base-station antennas, $L$ the pilot length in symbols, $\rho$ the per-antenna received signal-to-noise ratio of the true user, and $\gamma$ the detector's decision threshold on its normalized test statistic. A ghost at the echo delay then becomes detectable roughly when
$$ |r|^2 \, M \, L \, \rho \;\gtrsim\; \gamma , $$
which gives a maximum tolerable reflection, in decibels, of about $\gamma_{\mathrm{dB}} - 10\log_{10}(M L \rho)$. The form comes from the ghost tap's energy summing over $M$ antennas and $L$ symbols just as a genuine user's does, compared against unit-variance noise. As an illustration, with $M=64$, $L=100$ and $\rho$ at 0 dB, the factor $ML\rho$ is about 38 dB. An absorbing layer that reflects at −30 dB, which sounds excellent by numerical-analysis standards, would then produce detectable ghosts.
The scaling makes three predictions that can each be falsified. The required absorption tightens by about 3 dB for every doubling of antennas or pilot length. Ghosts concentrate on users near the domain edge and on paths that strike the absorbing layer at grazing angles, where perfectly matched layers are known to absorb worst. And because nearby users see the same artificial boundary, their ghosts arrive from nearly the same direction, which adds spurious correlation between users' channels.
The last prediction connects the problem to sim-to-real evaluation. The Argos campaign of 64-antenna measurements at 2.4 GHz (Massive MIMO Channels With Inter-User Angle Correlation) found intra-cluster correlation above 0.48. It also found that over 30% of far-apart user pairs show at least twice the correlation of independent Rayleigh fading. A full-wave simulator with a leaky boundary could appear to match these statistics for the wrong physical reason.
Research Program
The physical motivation comes from wave-simulation work showing that truncation artifacts are not cosmetic. Reflections from an artificial boundary corrupt interior fields and can even feed back into source physics in jet-noise simulation (A method for removal of reflection artifact in CFD simulation of supersonic jet noise). Periodic boxes trap radiated energy and create non-physical interference (Computing nonlinear Schrödinger equations with Hermite functions beyond harmonic traps).
We reuse the standard tool, the perfectly matched layer: a buffer zone in which the equations are mathematically stretched so that outgoing waves decay instead of reflecting. This is the approach taken for elastic waves in Seismo-VLAB and for leaky optical cavities in the dielectric-cavity PML work. Unlike those papers, we do not treat the residual reflection as something to minimize and then forget. We treat it as the independent variable of the experiment.
The solver will be a Fourier pseudospectral time-domain scheme for the scalar wave equation. It will first run on two-dimensional indoor floor plans at about 900 MHz, where a 30 m room spans roughly 90 wavelengths, and then on a limited three-dimensional room. The spectral choice is deliberate and is supported by A Spectral Split-Step Padé Method for Guided Wave Propagation. That paper replaces finite differences with sine transforms in ocean acoustics, operates near 2 points per wavelength instead of the 10 or more needed by finite differences, and converges with 128 depth points where the finite-difference split-step method failed at 512. For our purpose this matters beyond cost. Low-order stencils introduce numerical dispersion, an error in which waves travel at slightly wrong, frequency-dependent speeds. That error accumulates as delay errors, which would contaminate exactly the delay statistics we are measuring.
The same paper explicitly leaves the integration of matched layers open. A Fourier scheme is periodic by construction, so without absorption the boundary echo becomes a wrap-around echo. That makes the boundary question unavoidable rather than optional.
The program builds four channel generators over identical geometry. The first is the pseudospectral solver with matched layers, with layer thickness and grading swept so that residual reflection spans a wide decibel range. The angle-resolved reflection will be measured directly against the reference below rather than taken from theory. The second is a time-gated reference: the same solver on a domain enlarged until every boundary echo arrives after the end of the recorded impulse response. This is expensive, but it contains no artificial reflection inside the observation window and serves as the ground truth. The third is an equivalent finite-difference time-domain solver with standard matched layers. The fourth is the naive independent-Rayleigh model used by the asynchronous-access papers.
From each generator we extract per-antenna channel impulse responses for a 16- or 64-element virtual array and hundreds of user positions. We then drive two families of asynchronous receivers with these channels. The first is the joint activity–delay–channel estimators built on orthogonal and memory approximate message passing in Joint Activity-Delay Detection and Channel Estimation for Asynchronous Massive Random Access. These are iterative sparse-recovery algorithms designed for the non-independent pilot matrices that arrival delays create. The second is the preamble-free on-off division multiple access receiver with a double sliding window and successive interference cancellation, which decodes users one at a time and subtracts each from the signal. It comes from A Fully Asynchronous Unsourced Random Access Scheme and its fading-channel extension Fully Asynchronous Unsourced Random Access over Fading Channels.
Both receiver families were validated only on statistical channels. Our study does not modify them. It asks what their evaluation would report, and how wrong that report would be, under physically generated channels with and without artificial echoes.
The final stage closes the loop to deployment. Generated correlation statistics are compared against the Argos measurements. A small asynchronous uplink experiment with software-defined radios is planned in an indoor room that we also model in the solver. The COSMOS testbed's USRP N210 radios around 900 MHz (Experimentation with Full-Duplex Wireless in the COSMOS Testbed) are a natural platform. That paper used them for single-link full-duplex, so the multi-user asynchronous setup is new engineering on our side.
Baselines
The naive reference point is the evaluation the field already uses: independent Rayleigh fading with path loss, run exactly as in the joint activity–delay paper's benchmark. That benchmark uses a single cell with 400 potential users within a 1 km radius, 16 base-station antennas, and a maximum timing offset of 4 symbol periods. We will first reproduce its reported runtime ordering, in which memory message passing took 4.31 s against 6.94 s for the orthogonal variant, a 37% reduction at matched accuracy. We will also reproduce the fading-channel unsourced result: about 11 dB of energy per bit over noise to support 50 active users at a per-user error probability of at most 0.1, and an asynchrony penalty under 1 dB at moderate load. Reproducing these numbers confirms that our receiver implementations are faithful before any channel is swapped.
The strongest existing channel generator is site-specific ray tracing with NVIDIA Sionna, as used in How to Bridge the Sim-to-Real Gap in Digital Twin-Aided Telecommunication Networks and PyJama. It is the strongest competitor because it has no artificial boundary at all and is orders of magnitude cheaper than any full-wave solver. It will run on the same floor plans, materials and user positions as our solver. Its maximum reflection order and path count will be tuned with the same compute budget we give to matched-layer tuning, and it will be scored against the same time-gated reference. Lacuna has no reported result on delay-tail or impulse-response fidelity of Sionna against full-wave or measured channels. The only quantitative sim-to-real figure in the file is that prediction-powered calibration reached higher capacity using less than 10% real-world data, which concerns a different objective. Our full-wave pipeline "beats" ray tracing only if its detector metrics are measurably closer to the time-gated reference and to the software-radio measurements. Matching a bulk statistic is not enough.
The finite-difference solver with standard matched layers is the baseline a numerical-methods reviewer would expect. It gets the same wall-clock budget and a grid density chosen by its own convergence test.
Two ablations isolate the key ingredient. In the ghost-injection ablation, we take the clean Rayleigh channels and add only a single synthetic tap whose delay, direction and amplitude are copied from the measured boundary echo. If this alone reproduces the detector degradation seen with the leaky solver, the mechanism is confirmed as "a coherent spurious tap," and the reflection budget can be applied to any simulator. In the matched-tail ablation, we fit a tapped-delay-line model to the power delay profile and inter-user correlation of the time-gated reference. This tests whether a cheap statistical model with the right tail is all a fair evaluation needs.
Why not just…?
Why not just make the absorbing layer thick enough and move on? That is the eventual fix, and we do not argue against it. The open question is how thick is "enough." Numerical practice judges layers by reflection in decibels relative to the incident wave. The scaling above says the relevant comparison is instead against $\gamma/(ML\rho)$, which grows with array size and pilot length. Grazing-angle and corner weaknesses also mean a layer that is excellent on average can still leak badly for particular users. The reflection-budget experiments produce the number practitioners would need.
Why not just use ray tracing, which has no artificial boundary? This is a real threat. If Sionna's impulse responses match the time-gated reference on detector metrics, the full-wave branch becomes unnecessary for this question. The concern is that ray tracing omits diffraction and fine geometry. The sim-to-real digital twin paper reports that resulting phase errors degrade multi-antenna processing and require separate calibration. Ray tracers also truncate reflection order, which shortens precisely the delay tail that asynchronous detectors are sensitive to. The head-to-head comparison against the time-gated reference and against radio measurements settles this directly, and we will report the outcome either way.
Why not just use finite differences, which handle complex geometry more easily? Numerical dispersion in low-order stencils shifts arrival times in a frequency-dependent way. Its cost at acceptable accuracy is far higher, as the spectral split-step result of 128 versus a non-converged 512 points illustrates. The finite-difference baseline at matched compute will show whether its delay errors are large enough to matter for the detectors. If they are not, we will say so and recommend it, because it is simpler.
Why not just evaluate receivers on measured channels like Argos? Measured channels are the gold standard for realism, and we use them for the correlation check. However, the Argos campaign reports channel vectors and correlation statistics, not ground-truth arrival delays. It also offers no knob that can be turned to establish cause and effect. Only a controlled solver lets us change boundary leakage while holding everything else fixed.
Why not just enforce synchronization so delay search is unnecessary? For sporadic, battery-limited devices, closed-loop synchronization costs signaling and energy that the asynchronous-access papers are designed to avoid. Even the preamble alone costs 1.5 dB relative to the preamble-free scheme. Low-cost hardware also drifts, which is why misaligned over-the-air computation designs estimators for offsets rather than assuming them away. The receivers under study exist because synchronization is off the table, so their evaluation must be right.
Proposed Experiments and Decisive Checks
The central experiment sweeps residual boundary reflection across the decibel range, with $M$ set to 16 and 64 and two pilot lengths. It records three kinds of metric. For the message-passing estimators, these are activity false-alarm and miss rates, delay-estimation error and channel estimation error. For the unsourced receiver, they are per-user error probability and the energy per bit required to reach 0.1. For the channel pool, they are the inter-user correlation distributions.
The first decisive check is whether the onset of ghost-induced false alarms tracks the predicted threshold $\gamma_{\mathrm{dB}} - 10\log_{10}(ML\rho)$, including the roughly 6 dB shift between 16 and 64 antennas. A flat response would falsify the claim that boundary leakage is a meaningful confound.
The second check is the correlation audit. If the leaky solver produces Argos-like correlation that disappears in the time-gated reference, then matching measured statistics is not evidence of a faithful simulator. That would be a cautionary result for digital-twin calibration.
The third check is the reversal test. The asynchronous unsourced papers report that staggered arrivals help successive cancellation and that asynchrony costs under 1 dB. We will test whether that conclusion survives on the time-gated reference channels, with their realistic correlation and delay spread, and whether its apparent survival or failure under leaky channels is an artifact of the boundary.
The fourth check is the deployment comparison in the modeled room with software-defined radios. Here the generator whose detector metrics land closest to measurement is declared the fair evaluation channel, whether that turns out to be the spectral solver, ray tracing, or the matched-tail statistical model. The study succeeds if it produces a validated reflection budget and a clear ranking of channel generators for evaluating asynchronous receivers, even if the winner is the cheap one.
References
- Massive MIMO Channels With Inter-User Angle Correlation: Open-Access Dataset, Analysis and Measurement-Based Validation
- A method for removal of reflection artifact in computational fluid dynamic simulation of supersonic jet noise
- Computing nonlinear Schrödinger equations with Hermite functions beyond harmonic traps
- Seismo-VLAB: An Open-Source Software for Soil–Structure Interaction Analyses
- Perfectly-matched-layer method for optical modes in dielectric cavities
- A Spectral Split-Step Padé Method for Guided Wave Propagation
- Joint Activity-Delay Detection and Channel Estimation for Asynchronous Massive Random Access
- A Fully Asynchronous Unsourced Random Access Scheme
- Fully Asynchronous Unsourced Random Access over Fading Channels
- Experimentation with Full-Duplex Wireless in the COSMOS Testbed
- How to Bridge the Sim-to-Real Gap in Digital Twin-Aided Telecommunication Networks
- PyJama: Differentiable Jamming and Anti-Jamming with NVIDIA Sionna
- Federated Edge Learning with Misaligned Over-The-Air Computation