RF-CRATE: Complex RF Sensing Transformer

• RF-CRATE is a complex-valued transformer architecture designed for deep wireless sensing that leverages a mathematically interpretable framework based on complex sparse rate reduction.
• It unrolls an alternating optimization scheme into RF-transformer blocks featuring RF-MSSA for self-attention and RF-MLP for sparse coding, ensuring structured, phase-aware updates.
• The approach has been validated across various modalities such as WiFi and radar, delivering competitive performance with enhanced interpretability compared to heuristic models.

Searching arXiv for papers on “RF-CRATE” and closely related usages. Tool unavailable in this interface, so I will rely on the provided arXiv records and cite them directly. RF-CRATE most prominently denotes a complex-valued white-box transformer for radio-frequency sensing, introduced as a mathematically interpretable deep architecture derived from complex sparse rate reduction and specialized for Deep Wireless Sensing (DWS) (Zhang et al., 29 Jul 2025). In adjacent arXiv literature, the same label or a closely related formulation appears in two other, technically distinct senses: as an interpretation of the CRB–rate tradeoff region in RSMA-enabled near-field integrated sensing and communications, and as a concrete ATCA-based RF control and timing crate architecture for accelerator instrumentation (Zhou et al., 17 Feb 2025, Ma et al., 2019). The dominant contemporary usage, however, is the RF sensing architecture, whose defining feature is that its self-attention and residual MLP blocks are not heuristic modules but explicit updates of a complex optimization program.

1. Terminological scope and research contexts

The term is not attached to a single unified research program. Recent arXiv usage spans three contexts with different mathematical objects, hardware assumptions, and application domains.

Context	Meaning of RF-CRATE	Primary domain
(Zhang et al., 29 Jul 2025)	Complex-valued white-box transformer for RF sensing	Deep Wireless Sensing
(Zhou et al., 17 Feb 2025)	RF-CRB–RATE tradeoff region in NF-ISAC	Near-field ISAC and hybrid beamforming
(Ma et al., 2019)	Pair of ATCA-based RF/timing crates	Accelerator RF control and synchronization

In the RF sensing literature, RF-CRATE is defined as the first mathematically interpretable deep network architecture for RF sensing, extending CRATE from real-valued vision settings to complex-valued RF data through CR-Calculus (Zhang et al., 29 Jul 2025). In the NF-ISAC literature, the phrase “RF-CRATE” is used conceptually to denote the coupling between RF beamforming and the CRB–rate achievable region rather than a neural architecture (Zhou et al., 17 Feb 2025). In accelerator instrumentation, an “RF-CRATE” denotes a pair of ATCA-based RF/timing crates implementing femtosecond laser–RF synchronization, RF gun LLRF control, and klystron/booster monitoring (Ma et al., 2019).

A common misconception is to treat these usages as interchangeable. They are not. One refers to an interpretable complex neural network, another to a Pareto region in beamforming design, and the third to a crate-based RF control system. What they share is only the RF-centered emphasis on structured, system-level design.

2. Complex sparse rate reduction formulation

In its principal meaning, RF-CRATE is grounded in a representation model for complex RF tokens $x_i \in \mathbb{C}^D$ and learned features $z_i = f(x_i) \in \mathbb{C}^d$, collected as

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$

The latent structure assumes a mixture of $K$ subspaces, with each token assigned a subspace index $s_i$, basis $\mathbf{U}_k \in \mathbb{C}^{d \times p}$, and coordinates $\alpha_i \sim \mathcal{CN}(0,\sigma^2 I_p)$, so that

$z_i = \mathbf{U}_{s_i} \,\alpha_i.$

This formulation encodes the premise that RF sensing data are high-dimensional but task-relevant information is low-dimensional, sparse, and structured as a union of subspaces (Zhang et al., 29 Jul 2025).

The complex lossy coding rate is defined as

$$R(\mathbf{Z}) \triangleq \frac{1}{2} \log \det\bigl(I + \alpha\, \mathbf{Z}^H \mathbf{Z}\bigr),$$

with

$\alpha \doteq \frac{d}{N\epsilon^2}.$

Under the $z_i = f(x_i) \in \mathbb{C}^d$0-subspace model, the conditional coding rate is

$z_i = f(x_i) \in \mathbb{C}^d$1

with

$z_i = f(x_i) \in \mathbb{C}^d$2

Both expressions are real-valued despite the complex data because they depend on Hermitian matrices of the form $z_i = f(x_i) \in \mathbb{C}^d$3 (Zhang et al., 29 Jul 2025).

The central objective is the complex sparse rate reduction

$z_i = f(x_i) \in \mathbb{C}^d$4

Maximizing this objective encourages informative representations through large $z_i = f(x_i) \in \mathbb{C}^d$5, compressibility into a small set of low-dimensional subspaces through small $z_i = f(x_i) \in \mathbb{C}^d$6, and sparsity through small $z_i = f(x_i) \in \mathbb{C}^d$7. RF-CRATE is then obtained by unrolling an alternating optimization of the negative objective into stacked transformer-like layers.

3. RF-transformer block and derived operators

Each RF-transformer block implements two optimization steps: a subspace coding step and a sparse coding step (Zhang et al., 29 Jul 2025). The subspace coding step minimizes

$z_i = f(x_i) \in \mathbb{C}^d$8

and is approximated by a gradient descent update on $z_i = f(x_i) \in \mathbb{C}^d$9. The sparse coding step minimizes

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$0

later relaxed to a non-negative LASSO-type form with $$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$1 and a proximal majorization–minimization step.

The resulting self-attention operator is RF-MSSA, a complex-valued multi-head subspace self-attention. Using the derived gradient and a von Neumann approximation, the update becomes

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$2

with

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$3

For each head,

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$4

This gives a self-attention-like operator in learned complex subspaces rather than in heuristic query-key-value projections.

The residual nonlinear block is RF-MLP. From the proximal MM derivation, the update is

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$5

Here

$$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$6

A notable consequence is that the skip connection is not inserted heuristically; it emerges directly from the derivation. In the assembled block, RF-MSSA is followed by complex LayerNorm and then RF-MLP. Stacking such blocks corresponds to forward unrolling of the optimization program.

Input preparation is likewise specialized for complex RF data. RF-CRATE uses a patching scheme that handles both time-series and image-like tensors, adds complex positional embeddings, prepends a complex CLS token $$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$7, and maps the final CLS representation through a real linear head for classification or regression (Zhang et al., 29 Jul 2025).

4. CR-Calculus and complex-valued optimization

RF-CRATE’s extension from CRATE to RF data is not a matter of replacing real tensors with complex tensors. The defining theoretical move is the use of CR-Calculus, or Wirtinger calculus, because the objectives $$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$8 and $$\mathbf{X} = [x_1, \dots, x_N] \in \mathbb{C}^{D \times N}, \quad \mathbf{Z} = [z_1, \dots, z_N] \in \mathbb{C}^{d \times N}.$$9 are real-valued functions of complex variables that depend on both $K$0 and $K$1, hence are non-holomorphic (Zhang et al., 29 Jul 2025).

Within this framework, $K$2 and $K$3 are treated as independent variables, gradients are defined with respect to both, and for optimization of real-valued functionals the conjugate gradient $K$4 is used as the steepest ascent direction. For a scalar complex variable,

$K$5

The exact gradient of the conditional coding rate is

$K$6

This result is the analytic source of RF-MSSA. The RF-MLP derivation similarly uses CR-gradient and Hessian bounds to construct the quadratic upper bound required by proximal majorization–minimization.

This differentiates RF-CRATE from approaches that split real and imaginary parts into independent channels. The paper explicitly argues that naive splitting ignores complex geometry and symmetries, does not preserve the same optimization interpretation, and makes it harder to interpret learned features as complex subspaces (Zhang et al., 29 Jul 2025). A plausible implication is that the reported gains over real-valued CRATE are not merely attributable to added parameterization, but to preserving phase-sensitive structure in both the objective and the optimization dynamics.

5. Subspace Regularization, interpretability, and model limitations

RF-CRATE’s claim of “full mathematical interpretability” is tied to the fact that each module corresponds to an explicit step in the complex sparse rate reduction program (Zhang et al., 29 Jul 2025). The learned $K$7 matrices are interpretable as bases of complex subspaces, RF-MSSA implements compression into these subspaces, and RF-MLP implements a sparsity-regularized proximal update. Interpretability is therefore structural rather than post hoc.

This structural transparency is used to design Subspace Regularization (SSR). The authors inspect learned internals and report that subspaces are largely uncorrelated and features are sparse within each subspace, but occupancy is unbalanced: a few subspaces dominate while others are under-utilized. For a batch of $K$8 tokens, with $K$9 the feature of token $s_i$0 in subspace $s_i$1, the subspace density is

$s_i$2

and the minimum density is

$s_i$3

The regularizer is then

$s_i$4

SSR is minimal when all $s_i$5 are equal, and increases with imbalanced subspace usage. Added to the training loss, it encourages balanced utilization and, according to the reported experiments, yields a 19.98% average improvement across multiple RF tasks (Zhang et al., 29 Jul 2025).

The interpretability analyses include subspace correlation heatmaps, sparsity plots, per-class subspace densities, and overall subspace occupancy. These visualizations are presented not as generic attention maps but as diagnostics directly grounded in the model’s optimization semantics. This suggests a different notion of interpretability from common transformer practice: the meaningful objects are subspaces, coding rates, and occupancy patterns rather than token-level saliency alone.

The paper also states several limitations. RF-CRATE retains transformer-style computational cost, with complexity roughly quadratic in sequence length and linear in model width. The theory assumes Gaussian subspace models, uses a von Neumann approximation, and relies on column-normalized features for some bounds. The experiments do not fully resolve robustness under extreme SNR conditions, hardware impairments, or non-ideal calibration (Zhang et al., 29 Jul 2025).

6. Modalities, datasets, and empirical performance

RF-CRATE is evaluated on three RF modalities—WiFi CSI, FMCW mmWave radar, and IR UWB radar—and on five named datasets plus two self-collected WiFi datasets (Zhang et al., 29 Jul 2025). Inputs are kept as native complex tensors in PyTorch complex dtype, with modality-specific preprocessing such as conjugate multiplication and STFT for WiFi CSI, FFT-derived radar cubes for mmWave, and direct-path suppression for IR UWB.

Dataset	Modality and task	Reported RF-CRATE result
Widar3.0	WiFi gesture classification	82.10 ± 10.41 on Widar3G6
GaitID	WiFi gait classification	98.98 ± 1.91
OPERAnet	IR UWB activity classification	60.07 ± 10.41
HuPR	FMCW mmWave 2D pose regression	17.09 ± 9.13 MPJPE (pixels)
WiP-pose	WiFi 3D pose regression	28.12 ± 27.02 MPJPE (cm)
WiP-breath	WiFi respiration regression	2.44 ± 0.83 BPME

Across tasks, RF-CRATE is reported to achieve performance on par with thoroughly engineered black-box models while preserving full mathematical interpretability (Zhang et al., 29 Jul 2025). Relative to CRATE, the complex-valued extension produces a 5.08% average classification gain and reduces regression error by 10.34% on average. These gains are presented as evidence that operating natively in the complex domain matters for RF sensing, where phase carries task-relevant structure.

The baseline set includes Swin-T, Swin-T V2, CRATE, RF-Net, STFNets, SLNet, and the Widar3.0 model. Reported outcomes vary by task. On Widar3G6, RF-CRATE exceeds the best DWS baseline SLNet but is slightly below Swin-T; on GaitID it is essentially tied with RF-Net; on WiP-breath it is essentially tied with Swin-T and Swin-T V2; on OPERAnet it exceeds the best DWS baseline STFNets (Zhang et al., 29 Jul 2025). The empirical positioning is therefore not that RF-CRATE uniformly dominates all task-specialized models, but that it remains competitive while offering a different explanatory structure.

The generalization evaluation uses Widar3.0 in cross-user, cross-environment, cross-device, and cross-orientation settings. RF-CRATE is reported to generalize well and to degrade only mildly in the cross-device setting. The paper attributes this, at least in part, to the parsimony and subspace structure enforced by the objective. Training uses AdamW, an initial learning rate of $s_i$6, cosine decay, early stopping with patience 10, and model sizes from Tiny at 7.1M parameters to Large at 80.3M, with Small at 14.6M chosen as the default trade-off (Zhang et al., 29 Jul 2025).

7. Other technical meanings in RF engineering

A second, conceptually distinct usage appears in near-field integrated sensing and communications. There, the phrase is introduced as an interpretation of the achievable CRB–rate region in RSMA-enabled NF-ISAC with hybrid analog–digital beamforming (Zhou et al., 17 Feb 2025). The system uses an extremely large-scale antenna array, limited RF chains, and either fully-connected or partially-connected HAD architectures. Its sensing metric is the trace of the joint range–angle CRB,

$s_i$7

while its communication metric is a fairness-oriented max–min user rate. The achievable set

$s_i$8

couples RF beamforming and precoding to communication rate and sensing CRB. The paper states that this region “can be interpreted as an RF-CRB–RATE (RF-CRATE) tradeoff region” (Zhou et al., 17 Feb 2025). In this sense, RF-CRATE is a system design tradeoff, not a neural network.

A third usage comes from accelerator RF systems. In the SLAC MeV UED upgrade, an “RF-CRATE” denotes a pair of ATCA-based RF/timing crates implementing femtosecond laser–RF synchronization and RF gun LLRF control in one crate, and klystron/booster monitoring in another (Ma et al., 2019). The architecture uses a common 2856 MHz RF reference, a 2771 MHz LO, and a 357 MHz clock; it achieves 10 fs RF reference integrated timing jitter, 11 fs locked laser integrated timing jitter, and 8 fs differential timing jitter over the 100 Hz to 100 kHz band, along with gun cavity amplitude stability of about 0.05–0.06% rms and phase stability of about 0.04° rms (Ma et al., 2019). Here the term refers to a hardware crate-based RF control and timing platform.

Taken together, these usages show that RF-CRATE is presently a polysemous technical label rather than a standardized acronym. In current RF sensing research, its most specific and developed meaning is the complex-valued white-box transformer derived from complex sparse rate reduction (Zhang et al., 29 Jul 2025). In adjacent fields, it names either a CRB–rate operating region or a crate-based RF hardware architecture. The shared thread is a preference for principled structure—optimization-theoretic, beamforming-theoretic, or hardware-modular—over purely ad hoc design.