Fractional Cross-Attention (FCA)

• Fractional Cross-Attention (FCA) is a multi-domain attention mechanism that employs fractional operators like the fractional Fourier transform to blend spatial and spectral information.
• The design utilizes parallel transform branches at 0°, 45°, and 90° along with a log-magnitude branch to capture fine, semi-global, and global features in a unified framework.
• Empirical studies in deformable image registration show FCA improves alignment accuracy and efficiency, making it a promising approach for complex multi-scale tasks.

Searching arXiv for relevant papers on “Fractional Cross-Attention” and closely related FCA usages. Fractional Cross-Attention (FCA) denotes cross-attention mechanisms in which the interaction between query and context is modified by a fractional operator or by a fractional notion of scale, rather than by a single standardized layer definition. In the most explicit arXiv usage, FractMorph defines FCA as a transformer block for deformable image registration that applies parallel fractional Fourier transform (FrFT) branches at $0^\circ$, $45^\circ$, and $90^\circ$, together with a log-magnitude branch, and then fuses the resulting representations through cross-attention between fixed and moving image streams (Kebriti et al., 17 Aug 2025). Related work uses “fractional” in other senses, including fractional shift equivariance and fractional-Laplacian dynamics, while the acronym “FCA” also appears in technically different modules. This suggests that FCA is best understood through its specific mathematical construction in each paper rather than through the acronym alone.

1. Terminological scope and acronym overlap

The acronym FCA appears with different expansions across the cited literature. This suggests a non-uniform nomenclature in which “fractional,” “feature,” “fine/coarse,” and “fully-connected” attention mechanisms may all be abbreviated as FCA, despite having different objectives and operators.

Paper	FCA expansion or related fractional notion	Setting
"Cross Attention Network for Semantic Segmentation" (Liu et al., 2019)	Feature Cross Attention	Semantic segmentation
"Fine- and Coarse-Granularity Hybrid Self-Attention for Efficient BERT" (Zhao et al., 2022)	FCA	Efficient BERT
"LMFCA-Net: A Lightweight Model for Multi-Channel Speech Enhancement with Efficient Narrow-Band and Cross-Band Attention" (Zhang et al., 17 Feb 2025)	T-FCA / F-FCA decoupled fully-connected attention	Multi-channel speech enhancement
"FractMorph: A Fractional Fourier-Based Multi-Domain Transformer for Deformable Image Registration" (Kebriti et al., 17 Aug 2025)	Fractional Cross-Attention	3D deformable image registration
"Alias-Free ViT: Fractional Shift Invariance via Linear Attention" (Michaeli et al., 26 Oct 2025)	Fractional shift-equivariant attention	Vision transformers
"Fractional neural attention for efficient multiscale sequence processing" (Qu et al., 13 Nov 2025)	Fractional Neural Attention with cross-attention implications	Sequence modeling

Within this landscape, the FractMorph usage is the clearest instance in which the phrase “Fractional Cross-Attention” is itself the formal module name. The other papers are nonetheless important for interpreting the broader research space, because they show that “fractional” can refer to fractional Fourier domains, fractional spatial shifts, or fractional diffusion, and that FCA can also denote unrelated attention modules.

2. FractMorph FCA as a multi-domain cross-image attention block

In FractMorph, deformable image registration is framed as the alignment of complex anatomical structures in medical images under both fine-grained local deformation and large-scale global deformation. The stated motivation is that most methods struggle to simultaneously model local, semi-global, and global deformations within a unified, efficient framework, and that standard transformer cross-attention fundamentally operates on a single domain and may not exploit the multi-scale nature of medical image content effectively (Kebriti et al., 17 Aug 2025).

The architecture is a 3D dual-parallel transformer-based model. At every encoding and decoding level, moving and fixed image features are processed in parallel streams. Each FCA block applies parallel FrFT branches at fractional angles of $0^\circ$, $45^\circ$, and $90^\circ$, along with a log-magnitude branch, so that local, semi-global, and global features are extracted at the same time. The enriched features are then fused via cross-attention between the fixed and moving image streams. A lightweight U-Net style network predicts a dense deformation field from the transformer-enriched features (Kebriti et al., 17 Aug 2025).

Operationally, the block proceeds in three stages. First, each stream is transformed through the FrFT branches. Second, all branch outputs, together with an optional skip connection, are concatenated and fused by pointwise convolution and normalization. Third, standard transformer cross-attention is applied to the FCA-enriched features, followed by layer normalization, MLP, and skip connections. The same pattern is used through encoder and decoder stages, with skip connections and progressive up/down-sampling mirroring U-Net architectures.

The core representational claim is that the three fractional domains and the log-magnitude branch expose complementary cues: direct spatial patterns for fine details and edges, mixed spatial-frequency structure for semi-global organization and robustness to spatial misalignment, and spectral context for overall structural correspondence. In the formulation given for FractMorph, these are available within a single block rather than through separate multi-scale pathways or hierarchical models.

3. Fractional Fourier construction and mathematical formulation

The FractMorph FCA block is built on the fractional Fourier transform, parameterized by an order $p$ or equivalently an angle $\alpha = p\pi/2$. The paper summary states the one-dimensional FrFT as

$$X_p(u) = \mathcal{F}_p\{x\}(u) = \int_{-\infty}^{\infty} K_p(t,u)\,x(t)\,dt,$$

with kernel

$$K_p(t,u) = A_\alpha\, e^{j\left(\frac{t^2 \cot\alpha}{2} - u t \csc\alpha + \frac{u^2 \cot\alpha}{2}\right)},$$

where $45^\circ$0 or $45^\circ$1 gives the pure spatial domain, $45^\circ$2 or $45^\circ$3 gives the pure frequency domain, and intermediate values such as $45^\circ$4 produce a semi-global spatial-frequency blend (Kebriti et al., 17 Aug 2025).

For 3D volumes, the transform is computed by sequentially applying 1D FrFT along each axis:

$45^\circ$5

The inverse is performed similarly with $45^\circ$6. Within each branch $45^\circ$7, the per-branch operator is described as

$45^\circ$8

The branch set comprises $45^\circ$9, $90^\circ$0, $90^\circ$1, and the log-magnitude branch

$90^\circ$2

The summary also specifies domain-dependent convolutions: local $90^\circ$3 convolution for the spatial branch and $90^\circ$4 convolution for the spectral branch, attributed to the convolution theorem.

After branch fusion, the attention stage uses standard transformer-style cross-attention on the FCA-enriched features. For moving-to-fixed alignment, the formulation is

$90^\circ$5

with $90^\circ$6 from the moving image and $90^\circ$7 from the fixed image. The same process is applied symmetrically in the opposite direction. In this construction, the “fractional” component lies in the multi-domain FrFT enrichment preceding cross-attention, not in a replacement of the softmax-attention rule itself.

4. Empirical behavior in deformable image registration

On the ACDC cardiac MRI dataset, FractMorph is reported to achieve an overall Dice Similarity Coefficient of $90^\circ$8, an average per-structure DSC of $90^\circ$9, and a $0^\circ$0th-percentile Hausdorff distance of $0^\circ$1 mm on the reported data split (Kebriti et al., 17 Aug 2025). The same abstract introduces FractMorph-Light, a lightweight variant with only $0^\circ$2M parameters, which maintains the superior accuracy of the main model while using approximately half the memory.

The ablation results summarized for the FCA block support the multi-branch formulation. Full FCA with all branches gives the best performance. Disabling the $0^\circ$3 branch causes overall DSC to drop by $0^\circ$4, average DSC by $0^\circ$5, and HD95 to increase. The $0^\circ$6 branch is described as critical for global matching, the $0^\circ$7 branch for semi-global structure, and the local branch for fine detail. The log-magnitude branch adds further improvement, especially when combined with phase features. Removing the skip connection in FCA decreases DSC by $0^\circ$8 and increases HD95, while the ablations are reported to maintain topological or diffeomorphic regularity, with Jacobian determinant statistics remaining stable (Kebriti et al., 17 Aug 2025).

The qualitative interpretation offered in the summary is that FCA enables accurate alignment under subtle local, large-scale global, and intermediate deformations within a single end-to-end network, without the need for scenario-specific tuning or hierarchical multi-scale networks. That interpretation is consistent with the reported branch roles and the ablation outcomes. A plausible implication is that the FrFT decomposition is functioning not merely as a frequency augmentation, but as a structured prior on deformation scale.

5. Other fractionalizations of cross-attention

A distinct 2025 line of work treats “fractional” not as fractional Fourier analysis but as sub-pixel translation equivariance. "Alias-Free ViT" replaces softmax attention with shift-equivariant attention mechanisms, including linear attention and cross-covariance attention, and proves equivariance to arbitrary, including fractional, spatial shifts (Michaeli et al., 26 Oct 2025). The attention family is written as

$0^\circ$9

and the XCA variant is

$45^\circ$0

The theoretical argument relies on the shift invariance of the cross-covariance $45^\circ$1, derived through Parseval’s theorem for band-limited continuous signals. Under the alias-free architecture described in the summary, the model attains $45^\circ$2 integer shift consistency and $45^\circ$3 half-pixel shift consistency for XCiT-Nano-AF, whereas APS retains near-$45^\circ$4 integer consistency but $45^\circ$5 half-pixel consistency. In this setting, “fractional cross-attention” is a useful interpretive label for attention that remains equivariant under sub-pixel translations, rather than an FrFT-based block.

A further generalization appears in "Fractional neural attention for efficient multiscale sequence processing" (Qu et al., 13 Nov 2025). There, the fractional operator is the fractional Laplacian, and token interactions are modeled through Lévy diffusion rather than Brownian diffusion. The attention score is defined by a power-law kernel,

$45^\circ$6

with row-normalized weights $45^\circ$7, and the same mechanism is stated to apply to cross-contexts for Fractional Cross-Attention. The theoretical claims are that the resulting attention networks exhibit larger spectral gaps and shorter path lengths, while the empirical claims include competitive text-classification performance even with a single layer and a single head, improvements in image processing and neural machine translation, and stronger BLEU on Multi30K En-De for encoder-decoder models with FCA. Here the “fractional” aspect is neither spectral-domain branching nor shift equivariance, but anomalous diffusion governed by the fractional diffusion equation

$45^\circ$8

Taken together, these papers show that the phrase “fractional cross-attention” can refer to at least three mathematically distinct constructions: FrFT-domain enrichment, fractional-shift-equivariant cross-covariance attention, and fractional-Laplacian attention kernels.

6. Distinctions from earlier FCA modules and recurring misconceptions

A common misconception is to treat FCA as a single architectural primitive across the literature. The earlier semantic-segmentation FCA in "Cross Attention Network for Semantic Segmentation" is not fractional in the FrFT or fractional-calculus sense. It is a Feature Cross Attention module that fuses a shallow spatial branch and a deep context branch, computing spatial attention from the spatial branch and channel attention from the context branch, then applying them sequentially (Liu et al., 2019). The reported Cityscapes validation ablation moves from $45^\circ$9 mIoU for a baseline concat+$90^\circ$0 conv fusion to $90^\circ$1 for the spatial-then-channel FCA order, and the full Cityscapes test result with ResNet101 reaches $90^\circ$2 mIoU. This is cross-branch attention, but not Fractional Cross-Attention.

The efficient-BERT FCA of "Fine- and Coarse-Granularity Hybrid Self-Attention for Efficient BERT" is different again. There, FCA progressively shortens the computational sequence length by retaining informative tokens as fine-granularity units and replacing uninformative tokens with one or several clusters as coarse-granularity units (Zhao et al., 2022). The paper reports a $90^\circ$3 reduction in FLOPs over original BERT with $90^\circ$4 loss in accuracy on GLUE and RACE. Although the summary describes this as “progressively fractional attention,” it is a hybrid self-attention compression strategy rather than a cross-attention mechanism.

LMFCA-Net introduces T-FCA and F-FCA as time-axis and frequency-axis decoupled fully-connected attention for multi-channel speech enhancement (Zhang et al., 17 Feb 2025). Its F-FCA models frequency-axis dependencies for tensors shaped $90^\circ$5, with an efficient implementation based on 1D depth-wise convolutions along the frequency axis and reduced complexity $90^\circ$6 rather than $90^\circ$7. The summary states that this can be conceptually viewed as a fractional or factorized implementation of cross-attention, but the paper’s own terminology is “decoupled fully-connected attention,” not Fractional Cross-Attention.

The consistent lesson is therefore terminological rather than merely historical. “FCA” may denote Feature Cross Attention, fine- and coarse-granularity hybrid self-attention, decoupled fully-connected attention, or Fractional Cross-Attention, and “fractional” itself may refer to FrFT angles, sub-pixel translations, or fractional diffusion. This suggests that precise interpretation requires checking the paper’s operator class, attention equation, and target domain rather than relying on the acronym.