Can Riemannian alignment on the SPD manifold correct EEG forward modeling shifts?

The shift comes from each subject's forward matrix

EEG records scalp voltages that mix latent brain sources through a subject-specific forward process. The paper models an epoch as x = A_j s, where s holds uncorrelated latent sources and A_j is a domain-specific, invertible forward matrix tied to factors such as head geometry and sensor placement. Because A_j differs across subjects, the observed sensor covariance C = A_j E A_j^T differs too, even when the latent source covariance E is stable.

This is a latent model shift, not a change in the source-to-label relationship. The paper assumes the label depends on the log of the latent source covariance through a shared linear model, so once the A_j-induced discrepancy is removed, one shared regression model should transfer across domains. That assumption is what makes the alignment target well defined.

Polar decomposition splits the shift into mean and rotation

The paper decomposes each forward matrix as A_j = Q_j exp(P_j), using the uniqueness of the polar decomposition for square invertible matrices. The exp(P_j) term is a domain-specific transition, and Q_j is an orthogonal matrix. On the SPD manifold this maps to two geometric operations: mean shifts, which correspond to global power changes affecting all channels similarly, and rotation shifts, which correspond to scale changes that differ across classes.

That decomposition is the paper's central theoretical claim: the forward-modeling shift can be recovered solely through linear transformations on the SPD manifold, specifically mean matching followed by a global rotation. It also explains why earlier Riemannian alignment methods worked without a stated reason, and it tells you which two operations an alignment pipeline needs.

SPDAlign aligns means, then estimates one rotation

SPDAlign is an end-to-end Riemannian network with three parts: a feature extractor mapping raw EEG onto the SPD manifold, an alignment module, and a shared linear classifier. The alignment module recenters each domain's Fréchet mean to the identity matrix, then applies an orthogonal transformation. Training runs in two stages: the first 70% of iterations trains the feature extractor without rotation alignment, then the extractor is frozen, the rotation is estimated and applied, and the classifier is fine-tuned on aligned representations.

The rotation is estimated differently by setting. During source training, orthogonal Procrustes uses labels to build class-constrained correspondences, with a randomly chosen reference domain and downsampling to equalize class counts. During target inference, no target labels are used: Wasserstein Procrustes jointly estimates the orthogonal matrix and the correspondences between the reference source domain and the target domain, initialized at the identity and optimized for 50 epochs at a learning rate of 0.01. If label shifts are present, the Wasserstein term can be replaced by unbalanced optimal transport, which relaxes the marginal constraints.

The oracle gap isolates rotation estimation as the bottleneck

To test whether forward modeling is really the dominant source of cross-subject shift, the authors evaluated an oracle variant that uses target labels only to estimate the orthogonal transformation by empirical risk minimization, keeping the classifier fixed and shared. The oracle's high performance across datasets supports the claim that a substantial portion of the cross-subject shift can be addressed by conditional rotation alignment, and that the remaining unsupervised gap is mainly suboptimal rotation estimation rather than model mismatch.

That distinction matters for reading the headline numbers. The unsupervised SPDAlign result and the oracle result answer different questions: one is what the method achieves without target labels, the other is what the assumed shift model could achieve if the rotation were known. The paper reports state-of-the-art performance in three of four downstream tasks, but the gap to the oracle is the more informative signal about where future work should focus.

Rotation alignment carries more weight than mean alignment

The ablation reports grand-average change in balanced accuracy across all test sets relative to the full configuration. Removing rotation alignment while keeping moment alignment drops accuracy by 5.6 percentage points on average (t = 2.37, p = 0.02), a statistically significant loss. Removing moment alignment while keeping rotation drops it by 2.5 points on average (t = 0.82, p = 0.4), which is not significant. Removing both drops accuracy by 11.3 points on average (t = 3.58, p = 0.0006).

The authors note a caveat that shapes how to read the moment-alignment row: their implementation uses a global reference mean across domains, and without it performance degrades to chance level, so partial moment alignment is already implicitly enforced. That implicit alignment reduces the apparent contribution of the explicit moment step, which is why the rotation component looks comparatively more critical in this table.

Global linearity keeps classifier weights mappable to sources

Because SPDAlign composes a linear feature extractor, linear manifold transformations, and a linear classifier, it stays globally linear and is equivalent to the tangent space mapping framework. That means classifier weights can be mapped back to sensor space, and the resulting spatial patterns reflect source-sensor relationships and source contributions. The paper argues this matters for identifying which measurement channels and frequencies carry a cognitive process, and for trustworthy clinical BCI systems.

On the mental workload dataset, the extracted patterns peak in the theta band, and discriminative source activity concentrates in frontal and parietal areas. The authors present this as consistent with neuroscience work identifying those regions and that band as markers of mental workload. This is an interpretation anchored to one dataset and one class contrast, so it illustrates the framework's readout rather than establishing a general neurophysiological claim.

Where the evidence stops and what to watch

The evaluation covers four public benchmark datasets — motor imagery, workload estimation, mental stress, and driving fatigue — under a cross-subject decoding scheme within each dataset, compared against deep learning models, EEG foundation models, geometric frameworks, and other unsupervised domain adaptation methods. The paper's own limitation is computational cost: eigendecomposition-based operations restrict the framework to moderate-dimensional SPD features. The study is also limited to offline evaluation, though the authors describe extending to online adaptation by tracking running statistics and using online Wasserstein estimation.

Simulations show unsupervised rotation estimation degrades under extreme rotation shifts or limited class separability, because distribution-level optimal transport can pair semantically mismatched points when class-conditional structures overlap, biasing the estimated global rotation. The practical priority that follows is specific to this framework: treat rotation estimation quality, not the mean-alignment step, as the binding constraint on unsupervised cross-subject performance, and check class separability and the size of the rotation shift before expecting the label-free path to approach the oracle.

Sources used in this answer