Label-Free Low-Rank Adaptation of Neural Direction Finders

Neural networks now estimate the directions of radio, acoustic, or seismic sources more accurately than classical methods when echoes from the ground or sea corrupt the signal. Their weakness is deployment. They are trained on idealized simulated antenna arrays, and real arrays have miscalibrated sensors and propagation environments that differ from simulation. The usual fix is to retrain on many measurements labeled with the true source angle, and in the field such labels are rarely available. This note proposes adapting a pretrained direction finder to a new site using only raw, unlabeled recordings. The training signal is a physics equation that predicts what the array should have measured. Two small sets of parameters are updated. The first is a handful of physical calibration parameters inside that equation. The second is a low-rank adapter inside the network, meaning a small add-on that changes the frozen weights only along a few directions.

Some background on why echoes are hard. An echo is a delayed, scaled copy of the direct signal, so the two arrive coherently: they keep a fixed phase relationship. Coherent arrivals collapse the statistics that classical subspace methods depend on. Two paths look like one, and methods such as MUSIC fail (Merkofer et al.; Zhao et al.). The proposal splits the problem of moving to a new site into two parts. The physics model absorbs what is wrong with the hardware. The adapter absorbs what is new about the incoming signals, such as different reflection strengths, correlation levels, or surface geometry. The central bet is that the low-rank constraint is what keeps label-free adaptation stable when only a few hundred recordings are available. The experiments are designed so that this bet can fail visibly.

Core Claim

The proposal builds on the unsupervised direction finder of Weißer et al.. In that system, a convolutional encoder reads a sample covariance matrix and predicts three things: the angles $\theta$, a Cholesky factor $L$ of the source covariance, and the noise power $\sigma^2$. A fixed decoder then rebuilds the expected covariance from the array equation, and training minimizes a stochastic maximum-likelihood loss. Because the source covariance is estimated explicitly, the system works even at full coherence ($\rho = 1$). The proposal extends the decoder with physical calibration parameters $\phi$: per-sensor complex gains and per-sensor position offsets. The predicted covariance becomes

$$C_y(\theta, L, \sigma^2; \phi) = \Gamma_\phi A_\phi(\theta)\, L L^H A_\phi(\theta)^H \Gamma_\phi^H + \sigma^2 I,$$

where $A_\phi(\theta)$ is the array steering matrix evaluated at the perturbed sensor positions and $\Gamma_\phi$ is the diagonal gain-phase matrix. Adaptation on unlabeled target-site covariances $\hat S$ minimizes

$$\mathcal{L}(\Delta W, \phi) = \mathbb{E}_{\hat S}\left[\log\det C_y + \mathrm{tr}\left(C_y^{-1}\hat S\right)\right].$$

Here the encoder weights are $W_0 + BA$. The pretrained weights $W_0$ stay frozen, and $B, A$ are rank-$r$ factors in the style of LoRA.

The claim is that this split is necessary in both directions. Calibration alone fixes hardware error but leaves the encoder mismatched to new multipath conditions. An encoder adapter alone, with the decoder left uncalibrated, will distort its angle estimates to fit the wrong physics. Full fine-tuning under the same unlabeled loss will overfit the small target sample. The novelty is not "LoRA on a direction-of-arrival network" in isolation. It is the combination of three choices: adaptation without labels, a likelihood with explicit physical parameters, and a low-rank weight update. Each is matched to a different source of shift.

Relation to Prior Work

Three lines of work motivate the proposal. SubspaceNet trains a convolutional autoencoder that turns raw correlations into a full-rank surrogate covariance, with a differentiable Root-MUSIC layer in the training loop. It reports 0.20° error for two coherent sources, against 12.47° for classical Root-MUSIC. DA-MUSIC feeds a recurrent network's surrogate covariance through a differentiable eigendecomposition, and its extended version reports a stark gap between synthetic and real data. Error is about 0.014 rad on synthetic coherent sources and 0.6269 rad on the real GERES seismic array. CAE-CNN handles low-elevation radar multipath with separately trained networks for different elevation sectors. It relies on 200,000 simulated samples and notes the risk of over- or underfitting when data is limited. All of these are supervised and retrain every weight. That synthetic-to-real gap is the evidence this proposal targets.

Weißer et al. supplies the method, since it shows label-free training through a physics decoder is feasible. That paper assumes an ideal 9-element circular array and studies neither adaptation nor calibration.

The low-rank adaptation literature supplies the second ingredient. Arguments that low-rank updates resist overfitting and preserve calibration (Kong et al.; Xu) come entirely from language and vision models. Whether they transfer to small complex-valued signal networks is therefore an open empirical question, not a given. Spectral initializations such as PiSSA and data-driven ones such as EVA will be tested as secondary variants. EVA aligns the adapter with the directions of greatest variance in the target-site activations, which makes it the natural fit for label-free adaptation.

The research file also floats a "duality" between the physical rank collapse of the covariance and the rank of the adapter. I deliberately do not build on this. Nothing in the material suggests the two ranks are causally linked, and treating them as linked would make the idea hard to falsify.

Research Program

The program has four stages.

Pretraining. Encoders are pretrained on ideal synthetic data in two settings, using the losses of the original papers. The first is the 8-element uniform linear array, with datasets loaded in SubspaceNet format through the HYPERDOA codebase. The second is the 9-element circular array with three sources from Weißer et al. Correlation, signal-to-noise ratio, and snapshot count are varied during pretraining. Pretraining is biased toward weak correlation, so that strong-coherence deployment is a genuine shift.

Target sites. Target sites are built from five families of shift. Two are hardware shifts: gain-phase errors and sensor position errors. These are representable by $\phi$. Two are environment shifts: a changed surface reflection coefficient with low-elevation geometry, following the simulator of Zhao et al., and a move to fully coherent sources. These leave the decoder correct and change only what the encoder sees. The fifth is mutual coupling between sensors, which deliberately lies outside $\phi$. It tests what happens when the physics model is incomplete.

Adaptation. For each site, between 50 and 5,000 unlabeled covariance matrices are given to each adaptation method. Angle labels are used only for evaluation.

Analysis. Beyond accuracy, the program asks three questions. Which encoder layers absorb the shift, as seen in the singular spectrum of the learned $BA$? Does the calibrated decoder recover the true $\phi$? Can the adapter rank be chosen without labels, using the likelihood on held-out unlabeled data?

Baselines

The naive reference point is the frozen pretrained encoder applied to the target site with the ideal decoder. Every adaptation method must beat it, and its error measures the size of each shift. Alongside it are classical estimators that need no training. These are MUSIC and Root-MUSIC, and the Toeplitz-reconstruction plus Khatri–Rao method of Qi et al., which reports about 7° error and 74% probability of resolution at 0 dB SNR on its own simulated acoustic data. Where array size permits, the gridless semidefinite approach SPA of Yang et al. is also included. All of these run on exactly the same target snapshots.

The strongest competitor in the unlabeled setting is full fine-tuning of the encoder under the same calibrated likelihood loss. It is the obvious move, and these networks are small enough that it costs little. It receives the same target data, the same number of optimizer steps, and the same learning-rate search budget as the adapter, with model selection by held-out likelihood for both.

A calibration-only baseline updates $\phi$ and nothing else. Two ablations isolate the key ingredients. The first is low-rank adaptation of the encoder with the decoder left uncalibrated, which tests whether calibration is needed. The second adapts only normalization and output-head parameters, which tests whether low rank specifically matters or any small parameter budget would do.

Two supervised references are reported as upper bounds rather than competitors. One is supervised low-rank adaptation with the true labels. The other is SubspaceNet retrained on labeled target data, whose reported synthetic coherent-source result is 0.20°.

On GERES, the strongest reported figure is DA-MUSIC's RMSPE of 0.6269 rad, against 0.9383 for a classical beamformer and 1.2475 for MUSIC (Merkofer et al.). DA-MUSIC used labels to obtain that figure. On the circular-array benchmark, the reference is Weißer et al.'s report that over 95% of fully correlated test samples at 20 dB fall below 10° error. For low-elevation multipath, the only reported figures are classification precisions from Zhao et al.: 96.85% for the 0°–3.2° sector and 97.56% for the 5.8°–10° sector. No reported angle-error result for label-free adaptation was found on Lacuna for any of these benchmarks.

The proposal counts as beating the field under two conditions. First, the hybrid must achieve lower test RMSPE than full fine-tuning and calibration-only at every target-data size up to a few hundred samples, consistently across seeds. Second, on GERES it must approach or beat the labeled 0.6269 rad without using labels for adaptation.

Why not just…?

Why not just fine-tune all the weights, since the networks are small? This is a real threat. Parameter count is not the argument here, because memory is not the bottleneck. The question is whether the low-rank constraint regularizes better than full updates when the only training signal is a likelihood that can be satisfied in many wrong ways. The evidence that low rank regularizes (Kong et al.) comes from other domains. The target-data sweep settles the question. If full fine-tuning matches the adapter at 50–500 samples, the low-rank part of the claim is wrong, and the contribution reduces to calibrated label-free adaptation. That result would still be useful, and it would be reported as such.

Why not just calibrate the array and keep the network frozen? For pure gain-phase and position errors, this may well be enough, because the frozen encoder only sees inputs distorted by a known structured transformation. The program expects calibration-only to nearly match the hybrid on those two families. It is designed to show that calibration-only fails on reflection-coefficient and coherence shifts, where the array is fine but the input distribution is new. If calibration-only also succeeds there, the encoder adapter is unnecessary, and the proposal says so.

Why not just use classical decoherence methods that need no training? Toeplitz reconstruction with Khatri–Rao processing, and SPA, restore rank without data. Qi et al. acknowledge, however, that classical decoherence carries aperture and computation costs. These methods also assume an ideal, uniformly spaced manifold, which is exactly what fails at a real site. SubspaceNet's results in the two-snapshot regime suggest learned estimators are much stronger when snapshots are scarce. The classical baselines run on every target site, so this comparison is measured rather than assumed.

Why not just collect a few labeled samples and do supervised adaptation? If labels exist, use them, and the supervised upper bound shows how much they buy. The proposal is aimed at deployments where target angles are unknown, such as non-cooperative emitters or uncalibrated surface reflections. Its value is measured by how much of the gap between unlabeled and labeled adaptation it closes.

Decisive Checks

Three outcomes decide the idea.

The first is a two-by-two result on the hardware and environment shift families. It is decisive if calibration-only wins on hardware shifts, encoder adapters win on environment shifts, and only the hybrid wins on both. It is refuted if one component dominates across all four.

The second concerns mutual coupling, the out-of-model shift. Here the encoder adapter must absorb the residual error without the calibration parameters drifting to nonsensical values. Recovery of the true $\phi$ is reported on the shift families that $\phi$ can represent, to rule out the encoder and decoder trading errors.

The third is GERES, which is broadband and seismic, whereas the decoder assumes narrowband signals. Failure there would mark the limits of the physics model rather than of adaptation, so the result will be read with that caveat. On the synthetic sites, rank chosen by held-out likelihood must track rank chosen by labels. If it does not, the method cannot be tuned in the field, and that would count as a negative result.