Self-localizing Radio Mapping (SLoRM)

Constructing radio maps when measurements arrive without reliable location labels.

Why Self-localizing Radio Mapping?

A radio map connects wireless measurements to physical space. Traditional construction assumes that every received signal strength (RSS) or channel state information (CSI) sample comes with an accurate location label. RSS is the measured received power of a wireless signal, while CSI describes the measured or estimated channel response between transmitter and receiver. In practice, the location label may be expensive, unavailable indoors, privacy-sensitive, or unreliable under non-line-of-sight (NLOS) propagation, where the direct signal path is blocked.

Self-localizing radio mapping (SLoRM) asks a harder question: Can the wireless data help locate itself? The reference papers answer this in several settings, from indoor region maps built from unlabeled RSS, to angular power maps in massive multiple-input multiple-output (MIMO), to blind indoor MIMO beam maps from unlabeled MIMO orthogonal frequency-division multiplexing (OFDM) channel sequences.

The unifying idea is that wireless measurements are not isolated points. They have sequential order, spatial continuity, propagation geometry, and radio signatures that can be exploited to infer where the measurements were likely collected.

The Core Idea

A simple formulation is to recover hidden labels \(z_t\) or positions \(x_t\) from a sequence of measurements \(y_t\):

(\widehat{X}_{1:T},\widehat{\theta})= \arg\max_{X_{1:T},\theta} \sum_{t=1}^{T}\log p(y_t\mid x_t,\theta)+ \sum_{t=2}^{T}\log p(x_t\mid x_{t-1}) +\lambda R(X_{1:T},Y_{1:T}). \tag{1}

Here, \(X_{1:T}\) is the unknown trajectory, \(Y_{1:T}\) is the measurement sequence, \(\theta\) denotes propagation or map parameters, and \(p(x_t\mid x_{t-1})\) encodes the fact that a moving device cannot jump arbitrarily between far-away locations. The term \(R(X_{1:T},Y_{1:T})\) is a spatial regularizer that encourages the recovered physical distances to agree with radio-similarity distances extracted from the measurements, and \(\lambda\) controls how strongly this regularization is trusted. A propagation parameter is a quantity that describes how the wireless signal changes with distance, blockage, scattering, or antenna direction.

The model \(p(y_t\mid x_t,\theta)\) can be constructed based on specific applications. For example, in an indoor scenario where only received signal strength (RSS) from several access points (APs) are available, a signal subspace model can be used:

y_t = U_{z_t} a_t + n_t,

where \(U_{z_t}\) is a low-dimensional signal subspace (a tall matrix representing propagation parameters) for region \(z_t\), usually a discretized location, \(a_t\) is a coefficient vector to be estimated, and \(n_t\) is measurement noise. A signal subspace is a low-dimensional set that captures the dominant patterns in high-dimensional measurements. Sequential constraints help segment the unlabeled RSS stream into physically meaningful regions.

Two Self-Localization Examples at a Glance

The first example is coarse indoor localization from unlabeled RSS measurements. The second example is fine-grained MIMO beam-map construction from sparse CSI sequences. Together, they show that SLoRM can work at different spatial resolutions: from "which region am I in?" to "what full beam-space channel should be predicted here?"

Indoor region-based self-localizing radio mapping demo

Figure 1. Region-based self-localizing radio mapping in an indoor office environment [1]. The area is about 30 m by 16 m and is manually divided into 10 application-relevant regions. A mobile device follows a blind trajectory and visits the regions without recorded footprints or timestamps, while 21 ultra-wideband (UWB) sensors at 6 GHz collect RSS measurements. The task is to infer the visited region from unlabeled RSS data, even though the device location, visit order, and region dwell time are unknown.

Generative MIMO beam-map construction framework

Figure 2. Generative MIMO beam-map construction from sparse CSI without explicit location labels [5]. A CSI encoder maps sparse measurements into a latent CSI feature; the radio-map module aligns that latent feature with location labels and radio-map embeddings; and a generative decoder reconstructs full CSI. In this view, self-localization is not a separate preprocessing step: location inference and radio-map reconstruction are learned together through the latent representation.

A Typical Workflow

Step 1: Collect unlabeled measurements

The measurements may be RSS from a device walking through indoor regions, reference signal received power (RSRP) from a commercial fifth-generation (5G) massive MIMO network, or MIMO-OFDM CSI collected along an unknown indoor trajectory. RSRP is a 5G measurement of received synchronization-signal power. The device does not need to report precise coordinates.

Step 2: Infer hidden geometry

Depending on the setting, the algorithm may segment an RSS sequence, infer a hidden Markov trajectory, identify line-of-sight/non-line-of-sight (LOS/NLOS) conditions, or use a CSI-distance metric as a spatial regularizer. Line-of-sight means the direct path is unblocked; non-line-of-sight means the direct path is blocked. This step turns raw radio observations into estimated positions, regions, or ordering constraints.

Step 3: Attach measurements to recovered labels

Once the hidden labels are inferred, the original measurements can be reused as if they were labeled. This produces region-based radio maps, angular power maps, covariance maps, or MIMO beam maps.

Theoretical Foundations: Labels from Continuity and Structure

References [1]-[5] in this section refer to the works listed under Ours in the suggested reading.

1. Identifiability for coarse regions in an indoor scenario

In an indoor scenario where only RSS from several APs are available, it is more feasible to identify the location over coarsely discretized grids, such as region-based localization, i.e., finding which room the mobile locates in. In such scenario, the travelling along grid cells is much slower than the RSS measurements. Consequently, the localization problem can be formulated into a data clustering problem, and more precisely, a sequence segmentation problem.

Our paper [1] establishes identifiability results for recovering region-level location labels, identifying which room or region the RSS measurement belongs to. Several analytical results explain why a segmentation-assisted clustering algorithm is more stable than ordinary clustering: asymptotic hardening and consistency connect the noisy empirical cost to a deterministic proxy cost, while unimodality, flatness, and monotonicity propositions characterize the proxy cost near true segment boundaries. The resulting merge-and-split algorithm is not just a heuristic. Under the special case \(d_k=0\) and equal noise variance, Theorem 1 in [1] proves that the algorithm terminates at the globally optimal segment boundaries. Here, \(d_k\) is the dimension of the signal subspace for region \(k\). In short, there is a global optimality structure that can be stably found by a merge-and-split algorithm, and hence, the region-level labels are guaranteed to be identified.

This gives a theoretical reason why temporal order can replace explicit labels: the sequence structure changes a high-noise clustering problem into a boundary-finding problem with useful cost geometry.

2. Geometry controls whether hidden trajectories are identifiable

For an outdoor case, a hidden Markov model (HMM) connects hidden user locations to CSI evolution [3]. The key theoretical question is whether the trajectory can be recovered from noisy large-timescale measurements at all. The Cramer-Rao lower bound (CRLB) analysis in [3] shows two contrasting regimes: if BSs follow a sufficiently rich large-area Poisson topology, the localization-error lower bound can vanish asymptotically even at low signal-to-noise ratio (SNR); if BSs are confined to a limited region, the lower bound remains nonzero even with infinitely many independent measurements. A BS is a fixed network node that communicates with mobile users.

The lesson is that self-localization is governed by geometry, not only by algorithm choice. Sequential data are useful when the measurement environment provides enough independent spatial views of the hidden trajectory.

3. The spatial continuity theorem justifies NLOS regularization

The spatially regularized blind mapping paper [4] gives a theorem-level explanation of why NLOS measurements can still be useful. Under a quasi-specular multipath model and a large-number-of-paths approximation, its spatial continuity theorem proves that a properly defined CSI distance \(\widehat{u}(h_1,h_2)\) scales with physical distance \(d\), approximately as

\widehat{u}(h_1,h_2)\approx \frac{B}{c}d,

where \(h_1\) and \(h_2\) are two channel observations, \(B\) is the signal bandwidth, and \(c\) is the speed of light. This result supports the regularization term in (1): nearby physical positions should have nearby CSI signatures when they share the same NLOS propagation condition.

This is the main methodological point of SLoRM: the hidden labels are not guessed directly. They are recovered by imposing the right mathematical constraints on how radio measurements can change over space and time.

What Performance Advantage Do We Get?

The bracketed references in this section refer to Ours in the suggested reading.

Scenario	What the method uses	Reported advantage
Indoor region-based RSS map [1]	Signal subspace model with sequential prior and graph-based region matching	In real office-space measurements over a 30 m by 16 m area with 21 sensors, [1] reported 0.49 m and 0.68 m mean region-localization error on two test days, outperforming weighted centroid localization (WCL) and supervised k-nearest neighbor (KNN), support vector machine (SVM), and deep neural network (DNN) baselines trained with labels. In a larger 53 m by 55 m, 24-region real indoor area, [1] reported 0.34 m mean region-localization error.
Blind angular power map [3]	HMM trajectory recovery from unlabeled massive MIMO measurements	In a real commercial 5G massive MIMO dataset, [3] reported mean localization error below 18 m using sparse SSB RSRP measurements, even though neighboring-cell measurements were often missing. SSB means synchronization signal block.
Generative MIMO beam-map learning [5]	Sparse CSI sequence encoder, radio-map embedding, and generative CSI reconstruction	In the measured indoor distributed MIMO dataset from the DICHASUS factory environment, [5] reported 0.49 m mean positioning error and 1.08 m 95th-percentile error. Compared with switching Kalman filtering (SKF), semi-supervised channel charting with noisy labels (Semi-CC-Noisy), and real-world channel charting (Real-world CC) baselines, [5] reported 35.1%-68.3% lower localization error and 32.9%-66.1% lower 95th-percentile error.
Spatially regularized indoor MIMO mapping [4]	CSI-distance regularization, LOS/NLOS assignment, and Bayesian trajectory inference	In the indoor ray-tracing MIMO-OFDM dataset in [4], the method reported 0.68 m average localization error, 2% LOS/NLOS identification error, and 3.3% beam-map relative error.

These values are tied to the experimental settings and baselines in the cited papers. They depend on the measurement environment, mobility pattern, access point (AP) or BS topology, and available channel features.

Why This Research Matters

Self-localizing radio mapping is about turning unlabeled radio traces into usable spatial knowledge. This matters whenever manual labeling is expensive, privacy-sensitive, or technically unreliable.

For theoretical research, self-localizing radio mapping is an inverse problem that combines hidden-state inference, clustering, graph matching, geometry, and identifiability. It asks when unlabeled measurements contain enough sequential or structural information to recover the labels they are missing.

For real industrial applications, self-localizing maps can reduce manual surveying, protect user privacy, and help indoor networks, massive MIMO systems, and future sixth-generation (6G) environments maintain useful channel knowledge over time.

For prospective students who are interested in theoretical research, this topic is a good entry point into signal processing and machine learning because the central question is simple but deep: where did these measurements come from?