Spatial–Frequency Cooperative Modulation

• Spatial–Frequency Cooperative Modulation (SFCM) is a framework that integrates spatial selectivity with frequency control to enhance performance across diverse engineering fields.
• It increases spectral and energy efficiency by embedding spatial and frequency indices, achieving up to a 33% throughput gain without additional power usage.
• SFCM enables high-fidelity image reconstruction and sub-meter sensor resolution through cooperative neural architectures and adaptive signal processing methods.

Spatial–Frequency Cooperative Modulation (SFCM) refers to a diverse set of architectures and modulation strategies designed to jointly exploit spatial and frequency degrees of freedom for enhanced discrimination, efficiency, or regularization in communication systems, computational imaging, and deep neural models. SFCM appears in three primary domains: physical-layer wireless/optical modulation, inverse-problem neural architectures, and frequency-sensitive image processing networks. In each, spatial and frequency channels or representations interact to achieve capabilities or efficiency unattainable through purely spatial or purely frequency approaches.

1. Principles and Definitions

SFCM, in its broadest sense, integrates spatial selectivity (e.g., antenna index, local attention, or spatial grouping) with frequency control (e.g., frequency offset, spectral gating, or Laplacian pyramids). The strategy is to leverage each domain’s unique discriminativity:

• In communication systems, SFCM simultaneously encodes information in the transmitter’s spatial index and slight frequency offsets, thus increasing both spectral and energy efficiency beyond conventional orthogonal methods (Zou et al., 2016).
• In deep learning for inverse imaging, SFCM blocks coordinate Laplacian-based frequency cues and spatial attention to reinforce the capture of both global spectrum and local structure during reconstruction from sparse data (Meng et al., 2024).
• In image enhancement, SFCM networks dynamically modulate between global frequency representations and local spatial detail modulation to improve quality and speed, frequently via Fourier gating combined with spatial attention (Jiang et al., 2023).

Through cooperation, the spatial and frequency branches of such systems address the typical limitations encountered by using either domain in isolation, such as frequency discrimination limits, noise, or insufficient spatial resolution.

2. SFCM in Physical-Layer Communications

The SFCM scheme for communications, as formulated in "Conceptual Proposal: Frequency Offset Modulation for High-Efficiency Communications" (Zou et al., 2016), unifies spatial modulation (SM) and frequency offset modulation (FOM):

Architecture

• A transmitter array of $N_s$ antennas; during each symbol interval, a single antenna is activated.
• The active antenna’s carrier is offset by one of $N_f$ discrete frequency offsets.
• Each symbol encodes three streams:
  • M-ary QAM/PSK symbol bits ($\log_2 M$),
  • spatial index bits ($\log_2 N_s$),
  • frequency offset bits ($\log_2 N_f$).

Transmitted Signal Model

$$x_{s,\ell}(t) = \sqrt{E_s} \cdot a_m \cdot p(t) \cdot e^{j 2\pi \Delta f_\ell t}$$

where $a_m \in \mathcal{M}$ denotes the QAM symbol and $\Delta f_\ell = \ell \Delta f$, $\ell = 0,...,N_f-1$.

Receiver Processing

The receiver jointly detects the active antenna and frequency offset through maximum-likelihood estimation over all possible $(s,\ell)$ pairs.

Efficiency Advantages

SFCM delivers a throughput per channel use of $N_f$0 bits. Compared to either pure SM or FOM alone, SFCM offers up to a 33% increase in both spectral and energy efficiency, as all the additional spatial and frequency index bits are embedded without increasing radiated power or significant spectral overhead.

Implementation Challenges

Crucial issues include oscillator precision for small frequency offsets, fast reliable antenna selection, and robust channel/frequency estimation. The “cooperative” design implies both spatial and frequency index bits can be mapped adaptively or coded jointly to further improve error rates and capacity.

3. SFCM for Computational Imaging and Inverse Problems

In computational imaging, especially accelerated MRI reconstruction, SFCM has been implemented as a dedicated neural network block engineered to restore information lost under sampling limitations (Meng et al., 2024). The FPS (Frequency–Purification–Scale) block exemplifies this SFCM principle by combining:

Key Components

1. Frequency Modulation Attention (FMA)
   • Builds a Laplacian pyramid from the normalized feature map to extract multi-band (high-frequency) details.
   • Each bandwise component is used to recalibrate self-attention scores, thereby amplifying true high-frequency cues.
   • Equation:

   $N_f$1

   where $N_f$2 is the softmax-based attention map per frequency band.

2. Spatial Purification Attention (SPA)

   • Tokens (patches) with semantically related content are identified with a randomized hash-sorting operation.
   • Grouped localized self-attention is executed within these clusters, excluding irrelevant spatial interactions.
   • Output tokens are unsorted back into their original positions and represent spatially purified features.
3. Fusion and Scale Diversification
   • The outputs from FMA and SPA are concatenated and fused with depth-wise convolution.
   • Followed by a multibranch scale-diversified feed-forward network.

Cooperative Mechanism

• FMA accentuates missing high-frequency anatomical structure (edges, fine texture), counteracting the frequency bias of standard ViT models.
• SPA suppresses attention-induced noise from unrelated or artifact-prone tokens, reducing hallucinations and enhancing locality.
• Their cooperation in under-sampled k-space regimes ensures recovery of high-fidelity structure without explicit regularization.

Empirical Results

FPS-Former demonstrates NMSE reduction (0.0103 vs. 0.0108), SSIM improvement (0.9321 vs. 0.9297), and PSNR gains over state-of-the-art baselines with lower compute cost on CC359, fastMRI, and SKM-TEA datasets.

4. SFCM in Real-Time Image Enhancement Networks

The RSFDM-Net architecture introduces SFCM as a core fusion principle for underwater image enhancement (Jiang et al., 2023). Two central modules operate over spatial and frequency domains:

Adaptive Fourier Gating Mechanism (AFGM)

• Applies a 2D DFT channel-wise, then generates a soft, learnable frequency gating map via a spectral-domain convolutional sub-network.
• Spatial features are split into low- and high-frequency bands, which are selectively enhanced or suppressed before inverse DFT reconstruction.

Multiscale Convolutional Attention Module (MCAM)

• Processes split channel groups with depthwise and dilated convolutions at multiple scales, followed by point-wise projections and sigmoid fusion for local texture sensitivity.
• Outputs are concatenated and modulated with the frequency branch’s output via a global style vector broadcast to all primary net layers.

Cooperation and Training

This SFCM synergy allows robust color and detail restoration in highly degraded underwater images. The network is trained end-to-end with Charbonnier and perceptual losses. Experimental comparison on UIEB T90 yields PSNR = 23.325 dB and SSIM = 0.912, outperforming prior methods with orders of magnitude fewer parameters and FLOPs.

5. SFCM in Brillouin Optical Sensing and Spatial Resolution Optimization

Spatial–Frequency Cooperative Modulation principles also apply in direct-modulation Brillouin optical correlation-domain reflectometry (BOCDR), where spatial resolution is governed by both laser modulation amplitude and frequency (Ochi et al., 2023).

Key Theoretical Results

• Correlation peak width (spatial resolution):

$N_f$3

where $N_f$4 is the speed of light, $N_f$5 is fiber core index, $N_f$6 is Brillouin linewidth, $N_f$7 is frequency excursion, $N_f$8 is modulation frequency.

• Range between peaks:

$N_f$9

Cooperative Mechanism

$\log_2 M$0 and $\log_2 M$1 both appear symmetrically and multiplicatively in the denominator for $\log_2 M$2, indicating that improvements can be achieved by increasing either parameter, but practical limits on $\log_2 M$3 (typically $\log_2 M$4 half Brillouin shift) and range reduction induced by $\log_2 M$5 must be considered. The optimal protocol is to maximize $\log_2 M$6 first, then increase $\log_2 M$7 only when needed, preserving measurement range.

Experimental Confirmation

Tested from $\log_2 M$8 to $\log_2 M$9 GHz and $\log_2 N_s$0 increasing from the 5th to 20th harmonic, measured $\log_2 N_s$1 scales as predicted—down to sub-meter resolution—validating the cooperative effect.

6. Cross-Domain Patterns and Generalizations

SFCM provides a unified conceptual toolset wherever information is distributed across spatial and frequency resources:

• In communication, the concept enables multiplexing gains even under strict energy or bandwidth constraints.
• In neural architectures, it enables robust recovery and enhancement in ill-posed imaging problems, and can be ported to denoising, super-resolution, remote sensing, and astronomy (Meng et al., 2024).
• In sensor systems, SFCM models lead to practical recipes for optimizing physical spatial resolution.

An overarching implication is that SFCM architectures transfer successfully wherever the task requires selective amplification of weak or ambiguous frequency or spatial cues while maintaining fidelity and discrimination.

7. Common Methodological and Practical Considerations

Across domains, the following methodological elements and practical guidelines consistently emerge:

• SFCM exploits the symmetric and cooperative mathematical roles of spatial and frequency indices or modulations.
• When optimizing systems, increase frequency modulation depth/amplitude to the practical hardware limit prior to sacrificing spatial coverage or channel separation (e.g., by increasing modulation frequency or spatial index density).
• Neural and signal-processing implementations rely on differentiable, learnable gates or attention to couple spatial and frequency channels.
• In all cases, the cooperative mechanism does not generally require explicit spatial–frequency regularization terms; the interaction is structural within the architecture or protocol design.

The robust empirical and theoretical foundation for SFCM in physical-layer, neural, and sensor domains suggests it will remain a central strategy for future advances in data-efficient, high-fidelity sensing, communication, and image processing (Zou et al., 2016, Meng et al., 2024, Jiang et al., 2023, Ochi et al., 2023).