---
title: Power-Invariant Radio Map
url: https://www.emergentmind.com/topics/power-invariant-radio-map-rm
type: topic
---

# Power-Invariant Radio Map

A power-invariant Radio Map (RM) is a radio-field representation whose spatial pattern is constructed so that unknown or changing transmit power does not alter the map structure. In the cited literature, this is realized in two closely related ways: by normalizing sparse received signal strength (RSS) samples by the maximum observed sample to build a relative RM for non-cooperative NLoS localization, and by normalizing large-scale channel gain with respect to transmit power so that the shape of the gain map does not depend on the absolute base-station power [2509.01875, 2604.05788]. Power-invariant RMs therefore occupy a junction between localization, radio cartography, sparse reconstruction, geometry-aware learning, and active sensing.

## 1. Definitions and invariant formulations

Two formulations appear in the recent literature.

| Setting | Power-invariant quantity | Invariance statement |
|---|---|---|
| Sparse RSS in NLoS localization | $\tilde y_i=\dfrac{y_i}{\max_{j=1,\dots,M}y_j}$ | Unknown transmit power is removed by normalization |
| Large-scale gain cartography | $G'(\mathbf q)=G(\mathbf q)-P_{\rm tx}$ | The shape of the gain map does not depend on the absolute transmit power |

In the NLoS localization setting, the measurements are raw RSS values $y_i$ at sensor locations $\{\mathbf r_i\}\subset\mathcal S_s$. RadioDiff-Loc constructs power-invariant measurements $\{\tilde y_i\}\in[0,1]$ by normalizing every sampled RSS by the maximum observed sample, and then defines a relative RM through
\[
\mathcal R(\mathbf x)\;\approx\;\mathbb E\bigl[\tfrac{r(\mathbf x)}{\max_j r(\mathbf r_j)}\mid\mathrm{layout},\{\tilde y_i\}\bigr].
\]
This is explicitly a relative, power-invariant RM rather than an absolute field estimate [2509.01875].

In the gain-cartography setting, the primary quantity of interest is the large-scale channel gain
\[
G(\mathbf q)\;=\;P_{\rm tx}\;-\;L(\mathbf q)\;-\;S(\mathbf q)\;+\;\varepsilon(\mathbf q),
\]
or, in linear units,
\[
G_{\rm lin}(\mathbf q)\;=\;\frac{P_{\mathrm{rx}(\mathbf q)}}{P_{\mathrm{tx}}}.
\]
If $P_{\rm tx}$ is changed by a constant $\Delta P$, then $G(\mathbf q)$ shifts by $\Delta P$ everywhere, whereas the normalized gain
\[
G'(\mathbf q)=G(\mathbf q)-P_{\rm tx}
\]
does not change its spatial pattern. The same normalization is written in dB as
\[
G(\mathbf q)\;=\;10\log_{10}\!\Bigl(\frac{P_{\mathrm{rx}(\mathbf q)}}{P_{\mathrm{tx}}}\Bigr)
\quad\Longleftrightarrow\quad
G(\mathbf q)=P_{\rm rx}(\mathbf q)-P_{\rm tx}.
\]
This establishes power invariance as invariance of map shape, not invariance of absolute received power [2604.05788].

## 2. Construction from sparse unknown-power RSS

RadioDiff-Loc addresses non-cooperative scenarios in which the transmit power $P_t$ is unknown, so all RSS readings are only proportional to the true field. Its normalization step is
\[
\tilde y_i \;=\;\frac{y_i}{\max_{j=1,\dots,M}y_j},\quad i=1,\dots,M,
\]
with $\{\mathbf r_i\}$ the sparse sampling positions and $\{y_i\}$ the measured RSS. The normalized values $\{\tilde y_i\}$ become the conditioning input for constructing a relative RM [2509.01875].

The conditioning representation is spatially embedded. The model uses a binary layout map $H\in\{0,1\}^{N\times N}$, a sparse-measurement tensor $Y\in\mathbb R^{N\times N}$,
\[
Y_{m,n} =
\begin{cases}
\tilde y_i, & (m,n)=\mathbf r_i,\\
0, & \text{otherwise,}
\end{cases}
\]
and a binary mask
\[
M_{m,n}=\mathbf1\{(m,n)\in\{\mathbf r_i\}\}.
\]
These are stacked into the three-channel input
\[
X = [\,H\,,\,M\odot Y\,,\,M\odot Y\,]\in\mathbb R^{N\times N\times 3}.
\]

This formulation is specific to unknown-power RSS localization in obstructed environments. A plausible implication is that the normalization step is not merely a preprocessing convenience; it is the mechanism that makes the learned RM transferable across different unknown source powers while preserving the geometry-dependent intensity pattern.

## 3. Information-optimal sparse sampling at obstacle vertices

RadioDiff-Loc does not treat sparse measurement placement as arbitrary. Its sampling strategy is derived from the physical insight that diffracted electromagnetic energy concentrates near building edges, and, more specifically, that most of the diffracted energy in a NLoS scenario concentrates near sharp edges and corners through the “knife-edge” effect [2509.01875].

The sampled RSS vector is modeled as
\[
\mathbf y=K\,\mathbf u +\mathbf n,
\]
where $\mathbf u$ are unknown boundary fields, $K$ is a Green’s-function matrix, and $\mathbf n$ is noise. The Fisher information matrix is
\[
J = \sigma^{-2}\,K^H K.
\]
Under far-field diffraction each column $k_j$ of $K$ has norm $\|k_j\|\propto1/s_j$, where $s_j$ is the distance from the corner. Hence
\[
J_{jj}\;\propto\;\frac1{s_j^2},
\]
which is maximized at $s_j=0$, namely at the obstacle vertex.

The same conclusion is obtained from mutual information. The mutual information between $\mathbf u$ and $\mathbf y$ is
\[
\tfrac12\log\det\bigl(I+\sigma^{-2}KCK^H\bigr),
\]
and it grows fastest where $\|k_j\|$ is largest, again at vertices. The stated consequence is direct: placing sparse sensors at obstacle corners yields maximal information for RM reconstruction and localization, even with very few samples. This makes vertex sampling a structural component of the power-invariant RM pipeline rather than a heuristic measurement policy.

## 4. Conditional diffusion completion and RM-based localization

The completed power-invariant RM in RadioDiff-Loc is generated by a decoupled denoising diffusion model conditioned on the layout map and the sparse normalized RSS map [2509.01875]. Let $\mathcal R_0\in\mathbb R^{N\times N}$ denote a ground-truth power-invariant RM. The forward kernel is
\[
q(\mathcal R_t\mid \mathcal R_0)
=\mathcal N\bigl(\gamma_t\,\mathcal R_0\,,\,\delta_t^2\,I\bigr),
\]
with continuous-time form
\[
d\mathcal R_t = f_t\,\mathcal R_t\,dt + g_t\,dW_t,
\]
where $f_t=\tfrac{d\ln\gamma_t}{dt}$ and $g_t^2=\tfrac{d\delta_t^2}{dt}-2f_t\delta_t^2$.

A U-Net or similar CNN $s_\theta(\mathcal R_t,t\,;\,X)$ is trained to estimate the score $\nabla_{\mathcal R_t}\log q(\mathcal R_t\mid X)$. In practice the simplified denoising–noise prediction loss is
\[
\mathcal L(\theta)
=\mathbb E_{t,\mathcal R_0,\epsilon}\;
\Bigl\|\epsilon \;-\;\epsilon_\theta\bigl(\mathcal R_t,\,t\,;\,X\bigr)\Bigr\|_2^2,
\]
where $\mathcal R_t=\gamma_t\mathcal R_0+\delta_t\epsilon$ and $\epsilon\sim\mathcal N(0,I)$.

At inference, the model simulates the reverse DDPM updates
\[
\widehat{\mathcal R}_{t-1}
= \frac{1}{\sqrt{\alpha_t}}
\Bigl(\widehat{\mathcal R}_t
-\frac{\beta_t}{\sqrt{1-\bar\alpha_t}}
\,\epsilon_\theta(\widehat{\mathcal R}_t,t\,;\,X)\Bigr)
+\sqrt{\beta_t}\,z_t,
\quad z_t\sim\mathcal N(0,I),
\]
where $\alpha_t=1-\beta_t$ and $\bar\alpha_t=\prod_{s=1}^t\alpha_s$. The sample at $t=0$ is the completed, power-invariant RM $\widehat{\mathcal R}_0$.

Localization is then performed by identifying the brightest point on the generated RM:
\[
\hat{\mathbf d}
= \arg\max_{\mathbf x\in\mathcal S_r}\widehat{\mathcal R}(\mathbf x).
\]
The paper states that this approximates the MAP estimate under a uniform prior. Repeating the generative sampling yields an ensemble $\{\hat{\mathbf d}^{(k)}\}$ for uncertainty quantification. The same completed RM can also be fused with classical RSS-based localization methods in a “dual-driven” paradigm: trilateration or nonlinear least squares, fingerprinting/MLE, and centroid methods such as Top-$k$ weighted centroid or threshold-region centroid.

## 5. Geometry-aware gain RM reconstruction and uncertainty modeling

GeoUQ-GFNet studies sparse gain radio map reconstruction from a geometry-aware and active sensing perspective. It defines the true gain RM after discretizing $\mathcal Q\subset\mathbb R^2$ into an $H\times W$ grid with cell centers $\mathbf q_{i,j}$ as
\[
\mathbf G\in\mathbb R^{H\times W},\qquad G_{i,j}=G(\mathbf q_{i,j}),
\]
and performs reconstruction on the normalized gain so that the learned map is independent of the actual base-station transmit power [2604.05788].

The geometry priors are encoded as
\[
\mathbf P = [\,\mathbf O,\,\mathbf H,\,\mathbf R_x,\,\mathbf R_y,\,\mathbf D,\,\mathbf L\,],
\]
where $\mathbf O$ is obstacle occupancy, $\mathbf H$ is building height, $\mathbf R_x,\mathbf R_y$ are horizontal offset maps from the transmitter, $\mathbf D$ is transmitter–grid distance, and $\mathbf L$ is a line-of-sight–blocked indicator. Sparse observations are
\[
\mathbf G_s = \mathbf M_s\odot \mathbf G,
\]
with sampling mask $\mathbf M_s\in\{0,1\}^{H\times W}$ and accessibility mask $\mathbf M_a$. A dense initialization $\mathbf G_{\rm init}$ is obtained by a fast neighborhood-filling of $\{\mathbf G_s,\mathbf M_s\}$. The full input is
\[
\mathbf X  = \bigl[\mathbf P,\;\mathbf G_s,\;\mathbf M_s,\;\mathbf M_a,\;\mathbf G_{\rm init}\bigr].
\]

The architecture first separates
\[
\mathbf X_{\rm str} = [\mathbf O,\mathbf H,\mathbf M_a,\mathbf L,\text{edge cues}], \quad
\mathbf X_{\rm rel} = [\mathbf R_x,\mathbf R_y,\mathbf D], \quad
\mathbf X_{\rm obs} = [\mathbf G_s,\mathbf M_s,\mathbf G_{\rm init}],
\]
and encodes them into $\mathbf F_{\rm str},\mathbf F_{\rm rel},\mathbf F_{\rm obs}$. The observation branch is geometry-gated:
\[
\mathbf G_{\rm geo}=g_{\rm geo}(\mathbf X_{\rm str}),\quad
\mathbf G_{\rm mask}=g_{\rm mask}([\mathbf M_s,\mathbf M_a]),\quad
\tilde{\mathbf F}_{\rm obs}
=\mathbf F_{\rm obs}\odot\bigl(1+\mathbf G_{\rm geo}+\mathbf G_{\rm mask}\bigr).
\]
The paper states that this explicitly blocks propagation of spurious information across walls or into inaccessible areas.

The encoder uses three cascaded stages with Ghost residual blocks, while the deeper two stages insert Grid-KAN nonlinear enhancers,
\[
\mathcal K(\mathbf Z) =\psi\Bigl[\exp\!\bigl(-\tfrac{(\mathbf H-c_m)^2}{2\sigma_m^2}\bigr)_{m=1}^M\Bigr],
\]
and the bottleneck employs a large-kernel depthwise convolutional context module. The decoder is FPN-style, producing a shared representation $\mathbf F_{\rm ref}$ that feeds a residual gain head and an uncertainty head. Reconstruction is
\[
\Delta\mathbf G =f_{\rm mean}(\mathbf F_{\rm ref}),\quad
\hat{\mathbf G}_u=\mathbf G_{\rm init}+\Delta\mathbf G,\quad
\hat{\mathbf G}
=\mathbf M_s\odot\mathbf G_s +(1-\mathbf M_s)\odot\hat{\mathbf G}_u,
\]
and the uncertainty output is
\[
\mathbf S=f_{\rm var}\!\bigl(\mathcal K([\mathbf F_{\rm ref},\mathbf M_s,\mathbf M_a,\mathbf L])\bigr), \quad
\hat{\mathbf U} = \exp\Bigl(\tfrac12\,\mathbf S\Bigr), \quad
s_{i,j}=\log\hat\sigma_{i,j}^2.
\]

Training uses a weighted sum of four losses over the unobserved accessible set $\Omega_{\rm unobs}=\{(i,j)\mid M_{a,i,j}=1,\,M_{s,i,j}=0\}$:
\[
\mathcal L =\lambda_1\,\mathcal L_1 +\lambda_{\rm grad}\,\mathcal L_{\rm grad} +\lambda_{\rm nll}\,\mathcal L_{\rm NLL} +\lambda_{\rm var}\,\mathcal L_{\rm reg},
\]
with default weights
\[
\lambda_1=1.0,\;\lambda_{\rm grad}=0.05,\;\lambda_{\rm nll}=0.2,\;\lambda_{\rm var}=0.001.
\]
The stated losses are masked $\ell_1$, gradient consistency, heteroscedastic Gaussian negative log-likelihood, and a small regularizer on $\mathbf S$.

## 6. Active sensing, benchmark evidence, and reconstruction performance

GeoUQ-GFNet couples power-invariant gain RM reconstruction with uncertainty-guided measurement selection [2604.05788]. After each reconstruction step, the candidate region is
\[
\Omega^{(t)}=\{(i,j)\mid M_{a,i,j}=1,\,M^{(t)}_{i,j}=0\},
\]
and the next measurements are chosen by
\[
\mathcal Q^{(t)} =\operatorname{TopK}\Bigl(\hat{\mathbf U}^{(t)}\odot\mathbf M_a\odot(1-\mathbf M^{(t)})\Bigr).
\]
The mask and observed sparse map are then updated:
\[
\mathbf M^{(t+1)}=\mathbf M^{(t)}\cup\mathcal Q^{(t)}, \quad
\mathbf G_s^{(t+1)}=\mathbf M^{(t+1)}\odot\mathbf G.
\]
The paper states that this policy yields faster reduction in reconstruction error than random sampling under the same budget because the model flags geometrically difficult or sparsely supported regions with high predicted uncertainty.

All experiments use the UrbanRT-RM ray-tracing benchmark, described as seven scene layouts × eight BS deployments, 128×128 patches, 3.5 GHz, Sionna RT. The reported metrics over the accessible region are $\mathrm{RMSE}_{\mathrm dB}$ and $\mathrm{MAE}_{\mathrm dB}$. Under overall reconstruction with random 10% sampling, the reported results are as follows.

| Model | $\mathrm{RMSE}_{\mathrm dB}$ | $\mathrm{MAE}_{\mathrm dB}$ |
|---|---:|---:|
| GeoUQ-GFNet | **7.1025** | **1.6191** |
| Eff-UNet | 7.7557 | 2.0449 |
| ViT-UNet | 7.7895 | 2.0978 |
| ResNet-UNet | 7.8529 | 2.1139 |
| Nearest | 11.9412 | 3.7052 |

For active sensing with initial 10% plus 4 rounds of +1% each, at total additional budget 4%, the reported results are:

| Model | UQ-guided RMSE | Random RMSE |
|---|---:|---:|
| GeoUQ-GFNet | **3.1002** | 5.9144 |
| ViT-UNet | 3.4558 | 6.2599 |
| Eff-UNet | 3.5101 | 6.2742 |
| ResNet-UNet | 3.5543 | 6.3106 |
| Nearest | 9.8791 | 9.8791 |

The active sensing curves are summarized in the source as showing that GeoUQ-GFNet’s uncertainty policy steadily drives down RMSE from approximately $6.8$ dB at 10% to approximately $3.1$ dB at +4%, whereas random sampling yields only modest improvement. This places power-invariant RM reconstruction within a broader closed-loop sensing framework rather than a one-shot completion problem.

## 7. Applications, fusion strategies, and conceptual boundaries

Power-invariant RMs serve two application classes in the cited works. In RadioDiff-Loc, the completed relative RM is used for emitter localization in non-line-of-sight environments and is also compatible with existing RSS-based localization algorithms, enabling a “dual-driven” paradigm that fuses physical knowledge and data-driven inference [2509.01875]. In GeoUQ-GFNet, the normalized gain map is motivated by environment-aware wireless communication, network planning, and radio resource optimization, and the uncertainty map is used to guide active measurement selection under limited sensing budgets [2604.05788].

A key conceptual boundary is that power invariance does not recover absolute transmit power. In the gain-map formulation, the only difference between two deployments with powers $P_{\rm tx}$ and $P_{\rm tx}'$ is an additive shift in the entire map; no architectural or data-collection changes are required. In the RSS-based formulation, normalization by the maximum observed sample produces a relative map,
\[
\mathcal R(\mathbf x)\approx\mathbb E\bigl[\tfrac{r(\mathbf x)}{\max_j r(\mathbf r_j)}\mid\mathrm{layout},\{\tilde y_i\}\bigr],
\]
so the output is expressly power-invariant rather than absolutely calibrated.

The two works also delimit two methodological regimes. One regime reconstructs a relative RSS field from sparse, strategically placed measurements using a conditional diffusion model and then localizes by an argmax rule. The other reconstructs a dense gain map and a spatial uncertainty map from sparse measurements and structured scene priors, then uses uncertainty for iterative measurement selection. This suggests that “power-invariant RM” is best understood not as a single algorithm, but as a representation principle: remove the transmit-power ambiguity, preserve the spatial pattern, and then apply a reconstruction or inference mechanism suited to the downstream task.

Source: https://www.emergentmind.com/topics/power-invariant-radio-map-rm