Filter-Bank Euclidean Alignment (FBEA)

• Filter-Bank Euclidean Alignment (FBEA) extends traditional Euclidean Alignment by decomposing EEG signals into filter banks before applying covariance whitening to reduce cross-domain variability.
• It can be implemented either independently per sub-band or jointly across channels and filter banks, adapting to different EEG paradigms such as SSVEP and motor imagery.
• Empirical evidence shows FBEA boosts classification accuracy—improving motor imagery performance by up to 27% and accelerating convergence in deep learning models.

Searching arXiv for papers relevant to Filter-Bank Euclidean Alignment and Euclidean Alignment in EEG/SSVEP. Filter-Bank Euclidean Alignment (FBEA) is a filter-bank-aware covariance-whitening strategy for EEG transfer learning and domain adaptation. In the 2026 SSVEP literature, it is explicitly introduced as a preprocessing method for cross-subject classification that computes a reference covariance in the joint filter-bank–channel space and aligns each sample by left-multiplication with the inverse square root of that reference matrix. In the 2025 revisit of Euclidean Alignment (EA), the term itself is not formalized, but the paper provides the mathematical ingredients and pipeline guidance from which a filter-bank extension follows directly: apply temporal filter banks first, then perform EA before downstream spatial filtering or decoding. Taken together, these sources indicate that FBEA denotes a family of filter-bank-aware Euclidean whitening procedures grounded in EA, with the exact implementation depending on whether alignment is performed independently per sub-band or jointly across the concatenated filter-bank representation (Wu, 13 Feb 2025, Wang et al., 29 Jan 2026).

1. Euclidean Alignment as the conceptual basis

FBEA inherits its core logic from Euclidean Alignment, which was proposed to reduce data distribution discrepancies among subjects or sessions in EEG-based brain-computer interfaces. The underlying problem is large intra-subject and inter-subject variability, together with non-stationarity across sessions, which can induce negative transfer when source-domain data are used to assist a new subject. EA addresses this at the level of second-order statistics by whitening trials so that the mean covariance matrix of each domain equals the identity matrix after alignment; the revisit paper states that this makes EEG data distributions from different domains more consistent. For a domain with trials $\{X_n\}_{n=1}^N$, $X_n \in \mathbb{R}^{c \times t}$, EA defines the reference matrix

$\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$

and transforms each trial as

$\tilde{X}_n=\bar{R}^{-1/2}X_n.$

It then follows that

$$\frac{1}{N}\sum_{n=1}^{N}\tilde{X}_n\tilde{X}_n^\top = I.$$

The revisit paper characterizes EA as Flexible, Efficient, and Unsupervised, and emphasizes that different domains are processed identically, each using its own domain-level reference matrix rather than a single global whitening transform (Wu, 13 Feb 2025).

2. From EA to filter-bank formulations

The filter-bank extension arises because many EEG decoding pipelines do not operate on broadband trials alone. The revisit paper explicitly notes that EA should be placed between the temporal filtering block and the spatial filtering block, and its reference list includes “Euclidean alignment for transfer learning in multi-band common spatial pattern,” while its performance table reports a large gain for FBCSP with EA. This supports a direct extension of EA into multi-band or filter-bank pipelines. By contrast, the later SSVEP paper introduces Filter-Bank Euclidean Alignment (FBEA) as a named method and motivates it by the fact that SSVEP signals contain non-negligible harmonic responses, so channel-level alignment may ignore complementary information across frequencies. The paper therefore positions FBEA as a way to exploit frequency information from SSVEP filter banks in order to reduce the marginal distribution shift between source and target domains (Wu, 13 Feb 2025, Wang et al., 29 Jan 2026).

The two papers support two closely related but non-identical formulations.

Source	Status of FBEA	Alignment object
Revisit of EA	Not explicitly named as a formal algorithm	Natural extension: EA applied after filter-bank decomposition, plausibly per sub-band
SSVEP domain adaptation paper	Explicitly introduced as FBEA	Single covariance in the joint filter-bank–channel space

This suggests that FBEA is not yet a universally standardized algorithmic label. Rather, the literature presently supports both a branch-wise reading of filter-bank EA and a joint-space reading specialized to SSVEP.

3. Mathematical formulations

In the EA-based reading, which is explicitly marked in the revisit paper as an inference from the EA framework rather than a dedicated formal subsection, let the $b$-th filter-bank component of trial $n$ be

$$X_n^{(b)} \in \mathbb{R}^{c \times t_b}, \qquad b=1,\ldots,B.$$

A natural filter-bank extension is then to compute a band-specific reference matrix

$$\bar{R}^{(b)}=\frac{1}{N}\sum_{n=1}^{N}X_n^{(b)}\left(X_n^{(b)}\right)^\top,$$

and align each band independently by

$$\tilde{X}_n^{(b)}=\left(\bar{R}^{(b)}\right)^{-1/2}X_n^{(b)}.$$

For each band,

$X_n \in \mathbb{R}^{c \times t}$0

Under this interpretation, FBEA is simply EA repeated branch-wise inside a filter-bank architecture before FBCSP, RCSP, CNNs, or other decoders (Wu, 13 Feb 2025).

The 2026 SSVEP paper defines a different but related formulation. After filter-bank decomposition, each sample is represented as

$X_n \in \mathbb{R}^{c \times t}$1

where $X_n \in \mathbb{R}^{c \times t}$2 is the number of filter banks, $X_n \in \mathbb{R}^{c \times t}$3 the number of EEG channels, and $X_n \in \mathbb{R}^{c \times t}$4 the number of sampling points. The paper then defines

$X_n \in \mathbb{R}^{c \times t}$5

with

$X_n \in \mathbb{R}^{c \times t}$6

and aligns each sample as

$X_n \in \mathbb{R}^{c \times t}$7

The paper states that after alignment, the covariance matrix becomes the identity matrix, thereby achieving signal whitening. Because the stated covariance size is $X_n \in \mathbb{R}^{c \times t}$8, a dimensional reading strongly implies that the filter-bank and channel axes are treated jointly, so FBEA here is not defined as separate alignment for each band. Instead, it is a single whitening transform over the concatenated filter-bank–channel representation (Wang et al., 29 Jan 2026).

A recurrent misconception is therefore that FBEA always means independent bandwise EA. The published record does not support such a universal reading. One paper supports that construction as a natural EA extension, whereas the other explicitly defines a joint covariance over all filter-bank components and channels.

4. Pipeline placement and operational role

The revisit paper gives a strong prescription on placement: EA should be placed between the temporal filtering block and the spatial filtering block. In motor-imagery pipelines, experiments compared no EA, EA between temporal filtering and spatial filtering, and EA after both temporal and spatial filtering. Although EA improved performance in either location, earlier placement worked better. In a filter-bank system, this yields the direct operational template

$X_n \in \mathbb{R}^{c \times t}$9

which is the cleanest EA-grounded interpretation of FBEA. The same paper also reports that average re-referencing after temporal filtering can further help, and it treats domain-wise alignment as the correct usage: source and target domains are aligned separately using their own reference matrices, then pooled or used for transfer learning downstream (Wu, 13 Feb 2025).

In the SSVEP paper, FBEA is a preprocessing front-end to a larger adaptation framework. The specified preprocessing order is: channel selection of 9 occipital-region channels

$\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$0

temporal segmentation into the interval $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$1, filter-bank decomposition into $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$2 sub-bands, FBEA alignment on the filter-bank-decomposed signal, and then feeding the aligned data into the CNN-based model $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$3. This front-end precedes the trainable Cross-Subject Self-Training (CSST) framework, whose first stage is Pre-Training with Adversarial Learning (PTAL), whose second stage is Dual-Ensemble Self-Training (DEST), and which also includes a Time-Frequency Augmented Contrastive Learning (TFA-CL) module. FBEA is therefore not a trainable layer but a signal-level preprocessing step that operates before feature extraction and model training (Wang et al., 29 Jan 2026).

5. Empirical evidence and application domains

The broadest empirical support concerns EA rather than the formalized FBEA label. The revisit paper summarizes EA across multiple BCI paradigms, including motor imagery, mental imagery, event-related potential / P300, RSVP, error-related negativity, sleep stage classification, stress detection, emotion recognition, SSVEP, motor execution, and seizure detection. For filter-bank applications, motor imagery is presented as the clearest use case, because filter-bank methods such as FBCSP are standard and the recommended placement of EA after temporal filtering and before spatial filtering maps directly onto such architectures. In Table 1, the paper reports that FBCSP improved from 52.62% to 79.98%, an improvement of 27.36%, citing Amorim et al. 2024. The same paper also cites deep-learning evidence: EA gave nearly universal performance benefit for TIDNet and EEGNet, and another study reported improved mean accuracy for all evaluated cross-subject deep models together with a 70% acceleration in convergence for shared models. These results concern EA broadly, but they strongly suggest that filter-bank-aware alignment is compatible with both classical spatial filtering and deep networks (Wu, 13 Feb 2025).

The most direct evidence for explicitly named FBEA comes from the SSVEP domain adaptation paper. On the Benchmark dataset with 1 s signal length, the component ablation reports: baseline pure self-training 85.13%, baseline + FBEA 86.08%, CSST without FBEA 93.42%, CSST + FBEA 94.36%, and CSST + FBEA + TFA-CL 94.80%. The corresponding ITR values are 166.05 bits/min, 169.77 bits/min, 188.97 bits/min, 191.87 bits/min, and 193.12 bits/min. In the preprocessing comparison on the same setting, the reported accuracies are: no preprocessing 93.96%, channel normalization 93.87%, trial normalization 93.58%, channel Euclidean alignment 94.25%, and FBEA 94.80%. The paper interprets this pattern as evidence that normalization methods fail to align the data distributions across subjects, that Euclidean alignment helps by aligning geometric centers, and that FBEA performs best because it incorporates filter-bank frequency information (Wang et al., 29 Jan 2026).

6. Related methods, limitations, and open issues

The revisit paper situates EA relative to Riemannian Alignment (RA) and Label Alignment (LA). RA aligns covariance matrices on the Riemannian manifold of symmetric positive definite matrices, whereas EA aligns raw EEG trials in Euclidean space by linear whitening. LA is a supervised extension that performs class-wise EA; the revisit paper explicitly warns that class-wise alignment is no longer EA but LA. This distinction matters for FBEA because filter-bank variants inherit the same conceptual split: an unsupervised whitening step based on second-order statistics remains in the EA family, while label-conditioned alignment belongs elsewhere. The revisit paper also emphasizes that EA’s Euclidean formulation preserves compatibility with any Euclidean space signal processing, feature extraction, and machine learning algorithms, which explains its straightforward use with CSP, RCSP, FBCSP, CNNs, and shallow ConvNets (Wu, 13 Feb 2025).

The same source identifies several limitations that transfer directly to FBEA. First, computing $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$4 may be unstable when the number of channels is large. In filter-bank settings this issue can recur in every band, or in the larger joint filter-bank–channel covariance used by the SSVEP formulation. Second, alignment can alter channel semantics through “channel location shifting,” because aligned channels are linear mixtures of the original channels. Third, EA matches mean covariance but does not guarantee that other forms of mismatch are removed. The SSVEP paper adds further underspecification: it does not state the number of filter banks $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$5, the filter-bank definitions, the exact scope over which $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$6 is computed, whether covariance normalization or time mean-centering is used, or how $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$7 is regularized numerically. It also does not compare the published joint-space FBEA against a per-band alternative. The papers therefore support several open questions: whether joint and per-band formulations behave differently across paradigms, how to stabilize inverse-square-root computation, how to preserve sensor interpretability, and how filter-bank-aware alignment should be adapted to online drift, source-free adaptation, or regression settings. Some of these directions are explicitly proposed in the EA revisit, while others are plausible implications of the current SSVEP formulation (Wang et al., 29 Jan 2026).

Taken in aggregate, the literature supports a precise but non-monolithic definition. FBEA denotes Euclidean alignment performed on filter-bank-expanded EEG representations so that post-alignment second-order statistics are whitened. In the EA revisit, this is most naturally realized by applying EA independently in each sub-band before downstream spatial filtering. In the SSVEP adaptation paper, it is realized by computing a single covariance over the joint filter-bank–channel space and whitening with $\bar{R}=\frac{1}{N}\sum_{n=1}^{N}X_nX_n^\top,$8. The common principle is unchanged: reduce cross-domain distribution mismatch through filter-bank-aware second-order alignment before transfer learning or domain adaptation.