Cross-Subject Alignment & Transfer Matrices

• Cross-Subject Alignment is a set of algorithmic and statistical strategies that transform individual neural signals into a shared, subject-agnostic space.
• Transfer matrices, typically linear operators, are crucial for aligning data across subjects, enabling improved decoding and reduced calibration in fMRI and EEG applications.
• Methods such as ridge regression, Procrustes analysis, and manifold embedding reveal how careful regularization and optimization yield robust, interpretable mappings.

Cross-subject alignment refers to the set of algorithmic, statistical, and geometric strategies designed to map neural, electrophysiological, or other biological signals from one individual ("subject") to a representation compatible with that of another, or to a common shared space. Transfer matrices are explicit (typically linear) operators emerging from these alignment frameworks, used to transform data from a source domain (one subject or cohort) into the domain or coordinate system of a target subject or group. This is fundamental for pooling data, building subject-agnostic decoders, and reducing calibration requirements in applications such as fMRI-based brain decoding and EEG-based brain–computer interfaces (BCIs).

1. Fundamental Objectives of Cross-Subject Alignment

The central challenge in cross-subject analysis is the presence of substantial inter-individual variation in both anatomical structure and functional organization, leading to a breakdown of population-level statistical assumptions and limiting the generalizability of models trained on one subject to another. The primary objectives of cross-subject alignment and transfer matrices are:

• To construct mappings $T: X_s \mapsto X_t$ (or $X_s \mapsto Z$ where $Z$ is a shared space) such that task-relevant structure is preserved and task-irrelevant (subject-specific) variation is suppressed.
• To enable the aggregation of data across individuals for increased statistical power and generalization, and to facilitate transfer learning (e.g., applying a decoder trained on subject A to subject B).
• To minimize or eliminate the calibration burden for new users in BCI, neuroimaging, and similar domains.

Several distinct frameworks embody these principles, differing in their statistical, geometric, or algorithmic details, as detailed below (Zhang et al., 2019, Lai-Tan et al., 11 Aug 2025, Dai et al., 7 Feb 2025, Ferrante et al., 2023, He et al., 2018).

2. Key Mathematical Formalisms and Transfer Matrix Construction

Transfer matrices formalize how observations from a source subject ($X_s$) are mapped into the feature space of a target subject ($X_t$) or a shared representation ($Z$). The construction of these matrices depends on domain assumptions:

2.1 fMRI: Functional/Hyperalignment and Low-Rank Projections

• Linear Ridge Regression Transfer: Ridge alignment computes a matrix $W$ by minimizing

$\|X_s W^T - X_t\|_F^2 + \lambda\|W\|_F^2$

s.t. $W \in \mathbb{R}^{p_t \times p_s}$ (Ferrante et al., 2023). The closed form is

$W^T = (X_s^T X_s + \lambda I)^{-1} X_s^T X_t$

yielding a transfer matrix for direct pattern-mapping.

• Procrustes (Orthogonal) Hyperalignment: Seeks an orthogonal $X_s \mapsto Z$0 and (possibly) a scale $X_s \mapsto Z$1 for minimal least-squares misfit:

$X_s \mapsto Z$2

$X_s \mapsto Z$3 is given by $X_s \mapsto Z$4 from the SVD of $X_s \mapsto Z$5, $X_s \mapsto Z$6 by a norm ratio. Used for functional alignment (Ferrante et al., 2023, Xu et al., 2018).

• Graph or Kernel Hyperalignment: Models like the Graph-based Decoding Model (GDM) define blockwise subject-specific projection matrices $X_s \mapsto Z$7, mapping individual data into a K-dimensional aligned space, with alignment objective

$X_s \mapsto Z$8

Here, $X_s \mapsto Z$9 is a graph Laplacian encoding sample (dis)similarity across subjects and $Z$0 is a blockwise feature map (may be kernelized). Transfer matrices $Z$1 are constructed from the leading eigenvectors of a blockwise Laplacian (Li et al., 2019).

• Explicit Brain Transfer Matrix (BTM): The MindAligner framework learns a low-rank factored $Z$2 mapping arbitrary subject fMRI to a "reference" subject, with multi-level losses on decoding, spatial congruence, region-level statistics, and latent geometry (Dai et al., 7 Feb 2025).

2.2 EEG/BCI: Euclidean, Riemannian, and Manifold Methods

• Euclidean-Whitening Alignment (EA): Each subject's trial data $Z$3 is aligned by whitened transform $Z$4, where $Z$5 is their mean covariance. For any new trial,

$Z$6

producing identical covariance structure across subjects (He et al., 2018).

• Riemannian Alignment (RA): Covariance matrices are mapped to a joint reference via congruence:

$Z$7

Projecting SPDs into a tangent space (log-Euclidean or affine-invariant metric), then applying further projections or rotations for alignment (Manivannan et al., 24 Nov 2025, Zhang et al., 2019, Lai-Tan et al., 11 Aug 2025).

• Manifold-based Kernels and Projections: Embeddings using the Geodesic Flow Kernel (GFK) or manifolds (Grassmann, SPD) are constructed, followed by RKHS-alignment via kernelized Maximum Mean Discrepancy (MMD) matrices, which can also be dynamically balanced between marginal and conditional (class-wise) distributions (Luo et al., 2024).
• Calibration and Nonlinear Transfer Functions: Calibration strategies may use nonlinear mappings learned by deep networks or iterative optimization (gradient descent in signal space) to transfer resting-state signals into pseudo-task data for subject adaptation (An et al., 2024).
• Least-Squares Transformation (LST) for SSVEP: For transfer in plug-and-play SSVEP BCIs, linear channel-space transforms $Z$8 are estimated via least-squares regression to align source trials to a new subject's calibration templates (Chiang et al., 2018).

3. Algorithmic Procedures and Practical Workflows

The alignment and estimation of transfer matrices follow certain common algorithmic motifs, though implementation varies by domain:

• Subjectwise covariance or kernel computation: Compute per-subject statistical descriptors (covariance, Gram matrices, kernel representations).
• Spectral/hyperalignment solution: Solve SVD, eigenvalue problems, or Procrustes problems to yield linear or orthogonal alignment/transfer matrices.
• Manifold embedding and domain adaptation: For non-Euclidean signal structure, use tangent-space projections or geodesic interpolations (GFK), followed by kernel/graph-based alignment.
• Rotation calibration: Supervised Procrustes alignment using a small set of calibration (labelled or pseudo-labelled) data from the target subject, yielding closed-form orthogonal rotations (Lai-Tan et al., 11 Aug 2025).
• Iterative domain-adaptive classifier training: For manifold and kernel methods, alternate between domain-adapted classifier fitting and label/pseudo-label refinement (e.g., MDDD (Luo et al., 2024)).
• Downstream application: Aligned/pooled data is used for statistical pattern recognition (decoding, classification, image synthesis) using a variety of models (SVMs, linear regression, neural decoders).

The following table summarizes representative alignment strategies and their transfer matrix forms:

Method/Domain	Transfer Matrix (Operator)	Key Reference
Ridge regression (fMRI)	$Z$9	(Ferrante et al., 2023)
Hyperalignment/Procrustes	$X_s$0 from SVD; $X_s$1	(Ferrante et al., 2023, Xu et al., 2018)
Graph-based decoding	$X_s$2	(Li et al., 2019)
Brain Transfer Matrix (BTM)	$X_s$3	(Dai et al., 7 Feb 2025)
Euclidean Alignment (EEG)	$X_s$4	(He et al., 2018)
Riemannian Alignment (RA)	$X_s$5	(Manivannan et al., 24 Nov 2025)
GFK/Grassmannian	$X_s$6 with $X_s$7 via integrals on $X_s$8	(Luo et al., 2024)
Least-Squares Transfer (LST)	$X_s$9	(Chiang et al., 2018)

This table compiles explicit matrix formulations only where provided directly in the cited works.

4. Empirical Performance and Application Impact

Empirical assessment of cross-subject alignment relies on both task-oriented (decoding, classification, reconstruction) and alignment-oriented (correlation, similarity) metrics:

• fMRI: Ridge regression transfer matrices enable cross-subject natural image reconstruction with accuracy nearly matching within-subject decoders, while anatomical or classical hyperalignment methods perform at chance (Ferrante et al., 2023). MindAligner achieves substantial improvements—up to +17.9% in brain-image retrieval—with only 2.5% of per-subject data, and explicit transfer matrices yield interpretable cortical mappings (Dai et al., 7 Feb 2025).
• EEG/BCI: Euclidean alignment offers 3–20× speed improvements over Riemannian alignment, is simpler to compute, and performs competitively, especially as preprocessing for classical pipelines (He et al., 2018). Geometry-aware congruence transforms (RiFU, DCR) yield 3–4% accuracy improvements in zero-shot leave-one-subject-out setups (Manivannan et al., 24 Nov 2025). Least-squares channel-space transfer can halve calibration trials in SSVEP BCIs without loss of speed or accuracy (Chiang et al., 2018).
• Manifold/Kernels: MDDD with dynamic MMD alignment achieves non-deep performance (up to 84.57% accuracy on SEED) rivaling deep learning (Luo et al., 2024). Manifold-embedded transfer (MEKT) and dynamic alignment matrices consistently outperform classical domain adaptation, especially with ensemble aggregation (Zhang et al., 2019).

A plausible implication is that matrix-based alignment (linear, orthogonal, or manifold) is indispensable for robust cross-subject generalization and calibration efficiency in contemporary neural decoding.

5. Theoretical Guarantees and Geometric/Statistical Principles

Several frameworks provide identifiability or optimality guarantees:

• Graph-based and Kernel-alignment Models: The subspace learned by Laplacian-based methods is unique (up to rotation) when the eigenstructure of the graph Laplacian is well-separated, supporting meaningful sharing across unaligned datasets (Li et al., 2019).
• Manifold and RKHS Alignment: Kernel MMD-based approaches ensure that both marginal and conditional distributions in the aligned feature space are as close as possible, supporting theoretical minimization of cross-domain misclassification under covariate-shift assumptions (Zhang et al., 2019, Luo et al., 2024).
• Statistical Regularization: Regularization (whitening, low-rank projections, norm penalties) is essential to prevent trivial or degenerate solutions, particularly in high-dimension, low-sample regimes typical of neuroimaging and BCI (Ferrante et al., 2023, He et al., 2018).
• Interpretability: Explicit transfer matrices such as MindAligner’s BTM can be analyzed to uncover specific brain areas exhibiting high inter-individual variability, offering neuroscientific as well as methodological insight (Dai et al., 7 Feb 2025).

6. Extensions, Limitations, and Future Directions

Several research directions and open problems are suggested by current practice:

• Nonlinear and Deep Transfer Matrices: There is increasing effort to replace strictly linear mappings (e.g., $X_t$0, $X_t$1) with neural-network-based transfer functions, or with kernelized variants capable of capturing more complex cross-subject differences (Lai-Tan et al., 11 Aug 2025, Chiang et al., 2018, An et al., 2024).
• Plug-and-Play Calibration: Matrix-based transfer enables BCI or neuroimaging systems that require minimal calibration data, advancing practical usability (Chiang et al., 2018, Ferrante et al., 2023).
• Joint/End-to-End Learning: Some frameworks propose simultaneous learning of the decoder and the alignment mapping, potentially increasing robustness at the expense of interpretability (Ferrante et al., 2023).
• Cross-Modality and Multimodal Alignment: Methods such as HADUA handle multimodal inputs (EEG + eye movement), requiring transfer matrices that align not just between subjects but between measurement types (Tang et al., 29 Jan 2026).
• Residual Variability: Purely linear or orthogonally-constrained alignments fail to capture all individual differences, especially nonlinearities and idiosyncratic patterns. Extensions to nonlinear mappings or iterative manifold adaptation are ongoing.

A plausible implication is that although matrix-based cross-subject alignment is now a mature, theoretically-principled field, further gains—particularly for high-dimensional, multimodal, or highly heterogeneous data—will likely depend on nonlinear, adaptive, or multimodal extensions of the alignment and transfer matrix paradigm.