Goldstein, D.; Sorkin, V.; Menahem, Y.; Patashov, D.; Balberg, M.

Abstract. Purpose: We test whether bilingual versus monolingual language environments in early infancy are associated with differences in intrinsic functional organization measured from resting-state fNIRS connectivity. Approach: Using the RS4 infant resting-state fNIRS cohort (HbO), we estimate shrinkage-regularized symmetric positive definite (SPD) correlation operators over fixed, non-overlapping temporal windows. Window-level SPD estimates are aggregated to a subject-level operator using a Jensen-Bregman LogDet (JBLD/Stein) barycentric mean. Each subject is represented by dominant eigenspaces interpreted as points on the Grassmann manifold and compared via canonical (principal) angles. We characterize the full principal-angle spectrum and enrich it with within-spectrum jump descriptors. In parallel, we construct a complementary learned-graph representation and analyze low-frequency eigenspaces of the symmetrically normalized Laplacian. We evaluate separability with strict leave-one-subject-out validation, building all templates and model parameters from the training fold only, and report results on a common subject set (N = 94). Results: Correlation-based Grassmannian features achieve consistent above-chance separability (ROC-AUC = 0.811). Late fusion improves performance both within modality (CORR-FUSION: BA = 0.780, F1 = 0.719, ROC-AUC = 0.838) and across modalities, where fusing correlation- and Laplacian-based subspace predictions yields the highest ROC-AUC (CROSS-FUSION: ROC-AUC = 0.871; BA = 0.767, F1 = 0.700). Conclusions: Spectral-geometric subspace markers derived from SPD connectivity capture subtle bilingualism related differences in infant resting-state fNIRS. Complementary correlation and learned-graph spectral representations provide additive information, improving robustness under strict cross-subject evaluation.