This paper proposes Domain-Invariant Component Analysis (DICA), a kernel-based algorithm for domain generalization. DICA learns an invariant transformation that minimizes the dissimilarity across domains while preserving the functional relationship between input and output variables. The algorithm is motivated by a learning-theoretic analysis showing that reducing dissimilarity improves generalization ability. Experimental results on synthetic and real-world datasets demonstrate that DICA successfully learns invariant features and improves classifier performance. DICA generalizes well-known dimension reduction techniques such as kernel principal component analysis (KPCA), transfer component analysis (TCA), and covariance operator inverse regression (COIR). Theoretical analysis shows that DICA tightens generalization bounds by balancing distributional variance reduction and transformation complexity. DICA is applied to tasks such as flow cytometry data gating and Parkinson's telemonitoring, where it outperforms other methods in both supervised and unsupervised settings. The algorithm is shown to be effective in scenarios where the functional relationship between variables is stable across domains, and it can be adapted to various applications by incorporating appropriate constraints. The work highlights the importance of invariant feature learning in domain generalization and provides a theoretical foundation for its effectiveness.This paper proposes Domain-Invariant Component Analysis (DICA), a kernel-based algorithm for domain generalization. DICA learns an invariant transformation that minimizes the dissimilarity across domains while preserving the functional relationship between input and output variables. The algorithm is motivated by a learning-theoretic analysis showing that reducing dissimilarity improves generalization ability. Experimental results on synthetic and real-world datasets demonstrate that DICA successfully learns invariant features and improves classifier performance. DICA generalizes well-known dimension reduction techniques such as kernel principal component analysis (KPCA), transfer component analysis (TCA), and covariance operator inverse regression (COIR). Theoretical analysis shows that DICA tightens generalization bounds by balancing distributional variance reduction and transformation complexity. DICA is applied to tasks such as flow cytometry data gating and Parkinson's telemonitoring, where it outperforms other methods in both supervised and unsupervised settings. The algorithm is shown to be effective in scenarios where the functional relationship between variables is stable across domains, and it can be adapted to various applications by incorporating appropriate constraints. The work highlights the importance of invariant feature learning in domain generalization and provides a theoretical foundation for its effectiveness.