A device too small for several antennas can still act like an antenna array by moving its one antenna and recording the radio signal at many positions over time. This trick is called a synthetic aperture: a virtual array assembled from measurements taken at different moments. It already supports direction finding, localization and spoofer detection on phones and drones. The trick has a hidden cost. It turns a shortage of space into a reliance on time. A larger virtual array needs a longer sweep, and every method in the file assumes that the room's echoes stay fixed for the whole sweep. When people walk around or the device's own frame shifts, the snapshots taken early and late in the sweep no longer describe the same channel. The super-resolution step then smears or splits its peaks, because that step relies on a stable signal covariance matrix (a summary of how pairs of snapshots co-vary).
We propose to measure how quickly the channel itself is aging during the sweep and to use that measurement to set the trust placed in each pair of snapshots. The key experimental device is a revisiting trajectory. When the antenna returns to a position it has already occupied, the geometry is identical, so any change in the received signal comes from the environment or from position error, not from the array layout. Revisits at several time lags trace out a decorrelation curve. That curve sets a lag-dependent weight on the covariance matrix. Pairs of snapshots taken far apart in time count for less when the room is changing quickly and count fully when it is still. The estimator chooses its own effective aperture length from data instead of assuming the channel is static.
Core Claim
A synthetic aperture fails in a dynamic room for two separable reasons. One is internal error: inertial-sensor drift corrupts the assumed antenna positions. The other is external error: the multipath channel decorrelates over its coherence time $T_c$. Existing systems model only the first, or neither. Quitin et al. synthesize a virtual array from a single moving receiver. They jointly estimate the arrival angle and the oscillator frequency offset with a two-dimensional MUSIC search, and they derive a Cramér–Rao bound in which inertial drift sets a high-SNR floor. The channel in their model is time-invariant. Gil et al. spin quadrotors in place to build circular apertures for Sybil detection, and they name rapidly changing multipath as an explicit open limitation.
Our claim is that a revisit-calibrated, lag-tapered covariance recovers most of the angular accuracy a static channel would allow, in regimes where the sweep duration exceeds $T_c$. Our working hypothesis is that it does so with a lower and flatter error than both the static-channel estimator and the best fixed-length sub-aperture alternative. The contribution is a single measurement and estimation loop that keeps phase, which is where angular resolution lives. It is not a new learned model.
Concretely, let $y(t)$ be the hardware-corrected channel sample at time $t$ and antenna position $p(t)$. We form
$$\hat R = \sum_{t,s} w(|t-s|)\, y(t)\, y(s)^{*}\, \mathbf{e}(p(t))\,\mathbf{e}(p(s))^{H},$$
where $\mathbf{e}(p)$ is the indicator placing a sample at its virtual-array slot and $w(\tau)$ is a taper over time lag $\tau$. The taper is set from the revisit decorrelation $\rho(\tau) = \mathbb{E}[y(t)y^{*}(t+\tau)]/\mathbb{E}|y(t)|^2$, measured only at pairs with $p(t)=p(t+\tau)$. If $w$ is a box of fixed width, the estimator reduces to ordinary sub-aperture averaging. If $w \equiv 1$, it reduces to the standard static-channel synthetic aperture. Both baselines are therefore special cases, which makes the ablations clean.
Research Program
The pipeline has three stages.
The first stage cleans phase. We remove the deterministic receiver artifacts that Tadayon et al. identified: Tukey windowing, cyclic-prefix removal and cyclic delay diversity. They showed that much of the apparent CSI drift on commodity Wi-Fi comes from these artifacts, and that removing them enables decimeter ranging on static links. We reuse that method so that phase is usable across a sweep. Their work leaves mobility open, which we address. We keep the frequency-offset dimension of the Quitin et al. joint search, because oscillator drift is a separate temporal error that our taper does not absorb.
The second stage estimates coherence. Trajectories are designed or chosen to contain revisits: repeated rotations for a spinning drone or turntable, and loop segments for handheld motion. From these we fit $\rho(\tau)$. On a mechanically repeatable rotation, the revisit geometry is exact, so $\rho$ isolates channel aging. On handheld trajectories, inertial position error at loop closure adds decorrelation. We separate the two by their different lag dependence. Channel aging grows smoothly with $\tau$, while loop-closure position error is bounded per revisit. As a byproduct, the revisits also re-anchor the inertial track.
The third stage, analysis, extends the bound of Quitin et al. so that the channel gains follow a first-order aging process with coherence time $T_c$. The extended bound shows how the best sweep length depends jointly on inertial noise, SNR and $T_c$. We expect a U-shaped error in sweep length. Short sweeps lack aperture, and long sweeps without a taper lose coherence. The tapered estimator should track the bottom of that curve.
Baselines
The naive reference is the static-channel synthetic aperture with $w \equiv 1$: the joint angle and frequency-offset MUSIC of Quitin et al., run on the same cleaned samples. The strongest number the file reports for it is a direction-of-arrival RMSE of about 6.18° for a 50 cm radius trajectory in an anechoic chamber (Quitin et al.). We will reproduce that setting first, using a USRP-N210 software radio and an inertial sensor in a static, echo-free room. If we are within a degree of that number, our setup is validated. In a static room, our method should match it, since $\rho \approx 1$ makes the taper flat.
The strongest competitor on the actual question is fixed sub-aperture averaging. The sweep is cut into chunks shorter than an assumed $T_c$, MUSIC spectra or covariances are computed per chunk, and the results are averaged. We give it an oracle advantage: its chunk length is tuned by grid search on the test conditions themselves. Our taper, by contrast, is estimated online from revisits. Both use identical data, cleaning and search grids, and the compute budgets match up to the cost of estimating $\rho$. Beating it means lower RMSE at matched sweep duration across scatterer-motion levels. Matching it without oracle tuning would still be a meaningful result, since it means $T_c$ can be estimated online. We will report which outcome occurs.
For the localization comparison we include the amplitude-only Doppler pipeline of P2PLocate, which pairs a receiver with a backscatter tag, discards phase and fuses Doppler with time of flight. The file reports a median error of 0.88 m in line of sight and 1.55 m in non-line of sight for it. For spoofer detection we compare against the rotation-synthesized fingerprinting of Gil et al., which reports over 96% true-positive rate in its evaluated conditions.
Two ablations isolate the key ingredient. In the first, the taper is set from a generic Doppler-based guess of $T_c$ instead of revisits, which tests whether revisit measurement matters. In the second, the phase-cleaning stage is removed, which tests whether coherent combining depends on the Tadayon et al. correction.
Why not just…?
Why not just shorten the sweep so it always fits inside the coherence time? This is the most serious threat, and it is exactly the fixed sub-aperture baseline. A short sweep throws away aperture whenever the room is calm. A fixed short sweep also needs $T_c$ to be known in advance, and indoor $T_c$ varies with how many people are moving. The experiment settles it directly. If the oracle-tuned box matches our taper at every motion level, we will reframe the contribution as online $T_c$ estimation for window selection.
Why not just discard phase, as amplitude-based systems do? P2PLocate and TB-MAML both avoid phase drift by using amplitude. TB-MAML is a few-shot meta-learned fingerprinting model, and the file reports 59% of its 5-shot errors below 50 cm. Phase-free methods cannot resolve arrival angles at the resolution a coherent aperture provides. TB-MAML was also evaluated only in static or semi-static rooms. Our localization experiment runs both families under the same moving-scatterer conditions. If amplitude methods degrade less than ours as motion increases, that is a real negative result and we will report it.
Why not just get diversity from frequency or a reconfigurable antenna instead of motion? The $\alpha$-OFDM rotation of Couillet and Debbah lets one antenna match the outage performance of two antennas without any motion. Fluid antennas, analyzed by Huangfu et al., switch among many closely spaced ports and outperform multi-branch combining in their model. Both address fading, not direction finding, which is the task here. The fluid-antenna analysis also assumes perfect CSI across all ports, so it has its own coherence-time bottleneck: probing many ports takes time, which is the same aging problem in a different form. We therefore treat these methods as complementary and do not run them as baselines.
Proposed Experiments
We will run three settings, collected by us because the file lists no public dynamic-channel synthetic-aperture dataset.
The first is a validation setting on a USRP-N210 with an inertial sensor. It reproduces the anechoic condition of Quitin et al. and then moves to a furnished room. In that room, between zero and four people walk on scripted paths, which sets the scatterer motion level. A turntable with an angle encoder provides exact revisits and ground-truth geometry. The metric is direction-of-arrival RMSE against sweep duration and motion level. The decisive check is whether the static estimator's error rises with sweep length while ours stays near the minimum of the U-shaped curve the bound predicts.
The second setting uses Intel 5300 CSI on a spinning platform and repeats the Sybil-detection protocol of Gil et al. with moving people added. The metrics are true-positive and false-positive rates. The question is whether the reported detection rate of over 96% survives motion with the static estimator, and whether the tapered estimator restores it.
The third setting is handheld loops with a Nexus 6p-class phone in line-of-sight and non-line-of-sight rooms, compared with the P2PLocate pipeline on median localization error. This is the hardest case, because inertial and channel errors mix in it. The test is whether the lag-shape separation of $\rho(\tau)$ produces taper estimates that agree with the turntable-measured $T_c$ in the same room.
The outcome that would falsify the proposal is that the revisit-tapered estimator never beats the oracle-tuned fixed sub-aperture at matched duration in any setting.
References
- Virtual Multiantenna Array for Estimating the Direction of a Transmitter: System, Bounds, and Experimental Results
- Guaranteeing Spoof-Resilient Multi-Robot Networks
- Decimeter Ranging With Channel State Information
- Peer-to-Peer Localization for Single-Antenna Devices
- A Meta-learning based Generalizable Indoor Localization Model using Channel State Information
- Outage Performance of Flexible OFDM Schemes in Packet-Switched Transmissions
- Performance Analysis of Fluid Antenna System Under Spatially-Correlated Rician Fading Channels