Spectral Shrinkage of Gaussian Entropic Optimal Transport
Spectral shrinkage of Gaussian entropic OT enables direct algebraic computation, improving efficiency and understanding of infinite-dimensional degeneracy.
Key Findings
Methodology
This work employs functional calculus to analyze Gaussian entropic OT in separable Hilbert spaces, leveraging geometric notions like Proper Alignment and Schur complements. It reveals that the optimal coupling corresponds to spectral contraction of the correlation operator via a universal function fε, transforming iterative Sinkhorn schemes into direct spectral algebra. The approach allows exact cost evaluation across all ε > 0 with a single spectral decomposition, significantly enhancing multi-scale efficiency. As ε approaches zero, the analysis shows the limit converges to the most diffusive coupling—characterized as the centroid of the set of optimal Kantorovich plans—filtering out spurious correlations. Stability bounds and convergence rates are derived, with parametric rates in finite dimensions and spectral decay-dependent rates in infinite dimensions.
Key Results
- The spectral shrinkage function fε(x)=2x/（√4x²+ε²+ε） uniformly contracts eigenvalues, ensuring unique solutions. In finite dimensions, the exact entropic cost can be computed in O(d³) time once the spectral decomposition of G∗M is obtained. As ε→0, the spectral filter converges to an indicator function, leading to maximum diffusive solutions that filter out null-space correlations. Numerical experiments confirm faster convergence and stability compared to Sinkhorn, especially in high or infinite dimensions. The spectral approach simplifies computations and provides geometric insights into the limit behavior.
- Results demonstrate that in finite-dimensional settings, convergence rates are O(ε), while in infinite dimensions, rates depend on spectral decay, often slower, reflecting non-parametric behavior. The limit solutions are shown to be the maximum entropy (diffusive) couplings, providing a geometric interpretation of regularization effects. These findings have broad implications for large-scale data matching, distributional analysis, and machine learning tasks involving high-dimensional probability measures.
Significance
This research advances the theoretical understanding of Gaussian entropic OT by integrating spectral geometry and functional calculus, moving beyond matrix differential calculus. The spectral shrinkage framework simplifies computations, enabling fast, exact solutions in high-dimensional and infinite-dimensional spaces. It clarifies the geometric nature of the regularization limit, revealing that the maximum diffusive coupling emerges as the natural limit, filtering out artifacts and spurious correlations. This insight benefits applications in large-scale data analysis, generative modeling, and distribution alignment, where computational efficiency and geometric interpretability are crucial. The work also opens pathways for extending spectral methods to broader classes of distributions and non-linear settings, promising significant impact across computational optimal transport and statistical inference.
Technical Contribution
The paper introduces a novel spectral shrinkage mechanism for Gaussian EOT, grounded in functional calculus and geometric alignment. It derives a universal spectral filter fε that contracts eigenvalues of the correlation operator, enabling direct algebraic solutions via spectral decomposition. The approach transforms the problem from iterative fixed-point schemes into explicit spectral algebra, significantly reducing computational complexity. The analysis of the ε→0 limit reveals a maximum diffusive coupling characterized as the centroid of the solution set, providing a geometric interpretation of degeneracy and non-uniqueness. The derivation of stability bounds and convergence rates, especially in infinite-dimensional spaces, extends classical parametric results to non-parametric regimes, highlighting the influence of spectral decay on convergence speed.
Novelty
This work is the first to formalize spectral shrinkage as the core mechanism in Gaussian entropic OT, bridging geometric functional analysis with computational algorithms. Unlike prior methods relying on iterative schemes, it offers a closed-form spectral solution applicable in both finite and infinite dimensions. The explicit characterization of the limit as the maximum diffusive solution provides a new geometric perspective on degeneracy and non-uniqueness, which has not been addressed in previous literature. The integration of spectral analysis, proper alignment, and Fredholm determinants marks a significant conceptual advance in the field.
Limitations
- The spectral shrinkage relies on spectral decay properties; slow decay spectra may slow convergence or limit applicability in certain infinite-dimensional problems.
- In highly degenerate or non-positive definite cases, the filtering may not fully eliminate spurious correlations, requiring additional regularization strategies.
- Computational costs of spectral decomposition remain high in ultra-high-dimensional settings, necessitating further algorithmic optimization for large-scale applications.
Future Work
Future research will explore extending spectral shrinkage to non-Gaussian measures, possibly via kernel methods or non-linear spectral filters. Adaptive spectral functions could improve robustness across diverse spectral profiles. Integrating these techniques with deep learning frameworks for scalable distribution matching and generative modeling is also promising. Additionally, further theoretical work on spectral decay conditions and their impact on convergence in complex, non-linear, or non-stationary settings will deepen understanding and broaden applicability.
AI Executive Summary
Optimal transport (OT) has become a fundamental tool for measuring differences between probability distributions, with applications spanning machine learning, computer vision, and statistical inference. However, in high-dimensional and infinite-dimensional spaces, classical OT algorithms like Sinkhorn face computational challenges and slow convergence, especially as the regularization parameter ε approaches zero. This paper introduces a spectral shrinkage framework for Gaussian entropic OT, leveraging functional calculus and geometric alignment to transform the problem into a spectral contraction operation.
By analyzing the correlation operators through their eigenvalues, the authors derive a universal spectral function fε that contracts eigenvalues monotonically as ε increases. This spectral filter simplifies the computation of OT costs, reducing the problem to a single spectral decomposition, which can be efficiently reused across different ε values. The approach not only accelerates multi-scale analysis but also provides deep geometric insights into the behavior of solutions in degenerate regimes.
A key contribution is the characterization of the ε→0 limit, where the spectral filter converges to an indicator function, and the optimal coupling becomes the maximum diffusive solution—interpreted as the centroid of the set of all optimal plans. This solution filters out spurious correlations, especially in the null space, offering a principled way to handle degeneracy and non-uniqueness. The authors rigorously establish stability bounds and convergence rates, showing parametric rates in finite dimensions and spectral decay-dependent rates in infinite dimensions.
Numerical experiments validate the efficiency and robustness of the spectral method, demonstrating significant improvements over traditional iterative algorithms. The framework opens new avenues for scalable, geometrically interpretable optimal transport in high-dimensional and infinite-dimensional settings, with broad implications for data science and machine learning. Future work aims to extend these spectral techniques beyond Gaussian measures and incorporate adaptive, non-linear spectral filters for even greater flexibility and applicability.
Deep Dive
Glossary
Spectral Shrinkage (谱收缩)
A process that contracts the eigenvalues of an operator via a universal function, reducing complexity and filtering noise; in this paper, it applies to correlation operators. (英文) / 一种通过普适函数收缩算子特征值的机制，用于简化和过滤相关性，本文中用于相关算子谱。 (中文)
描述高斯EOT中特征谱操作的核心机制。
Proper Alignment (正确对齐)
A geometric condition ensuring the Green’s operators satisfy G* M ⪰ 0, which guarantees the invariance of the spectrum and simplifies spectral analysis. (英文) / 保证Green算子对之间满足特定正定关系的几何条件，有助于谱分析。 (中文)
用于确保特征谱一致性和算法稳定性。
Schur Complement (Schur补)
A matrix operation that characterizes the residual information in block operators, crucial for degeneracy analysis and spectral characterization. (英文) / 描述块矩阵中残余信息的操作，用于分析退化和谱结构。 (中文)
在谱分析和极限行为研究中起关键作用。
Fredholm Determinant (弗雷德霍姆行列式)
A determinant for trace-class operators capturing spectral properties of operators, used to compute KL divergence in Gaussian couplings. (英文) / 用于描述紧算子谱性质的行列式，关键于高斯耦合的KL计算。 (中文)
在极限分析和成本计算中应用。
Maximum Diffusive Coupling (最大扩散耦合)
The limit coupling as ε→0, characterized by maximal entropy and minimal correlation, representing the most spread-out optimal plan. (英文) / 极限状态下，最大熵、最宽松的最优方案，过滤虚假相关。 (中文)
极限行为和几何筛选的核心概念。
Open Questions Unanswered questions from this research
- 1 当前谱收缩机制在非高斯分布中的推广仍未充分研究，如何在非线性和非正态场景中保持效果，是未来的重要方向。
- 2 无限维空间中谱衰减条件对收敛速度的影响尚不完全清楚，特别是在复杂的随机场或非平稳场景中。
- 3 谱方法在实际大规模数据中的实现优化仍需突破，尤其是高效的谱分解和近似算法的开发。
Applications
Immediate Applications
大规模分布匹配
在高维数据分析中，利用谱收缩快速计算分布间的距离，支持迁移学习和生成模型，提升效率和精度。
图像和信号处理
通过谱特征过滤，改善图像匹配和信号同步的鲁棒性，适用于超高维特征空间。
Long-term Vision
深度学习中的分布对齐
结合谱方法实现高效的分布迁移和生成，推动无监督学习和迁移学习的突破。
Abstract
We present a functional calculus treatment of Entropic Optimal Transport (EOT) between Gaussian measures on separable Hilbert spaces, providing a unified framework that handles infinite-dimensional degeneracy. By leveraging the notion of proper alignment and the Schur complement, we reveal that the Gaussian EOT solution operates as a precise \textit{spectral shrinkage}: the optimal coupling is uniquely determined by contracting the spectrum of the correlation operator via a universal scalar function. This geometric insight facilitates an algorithmic shift from iterative fixed-point schemes (e.g., Sinkhorn) to direct algebraic computation, enabling efficient multi-scale analysis, where a single spectral decomposition allows for the exact evaluation of entropic costs across arbitrary regularization parameters $\varepsilon > 0$ at negligible additional cost. Furthermore, we investigate the asymptotic behavior as $\varepsilon \downarrow 0$ in settings where the unregularized Optimal Transport problem admits non-unique solutions. We establish a selection principle that the regularized limit converges to the most diffusive optimal coupling --characterized as the centroid of the convex set of optimal Kantorovich plans. This demonstrates that in degenerate regimes, the entropic limit systematically rejects deterministic Monge solutions (extremal points) in favor of the optimal solution with minimal Hilbert-Schmidt correlation, effectively filtering out spurious correlations in the null space. Finally, we derive stability bounds and convergence rates, recovering established parametric rates ($\varepsilon \log(1/\varepsilon)$) in finite dimensions while identifying distinct non-parametric rates dependent on spectral decay in infinite-dimensional settings.