Free actions and Grassmanian variety

An algebraic notion of representational consistency is defined. A theorem relating it to free actions is proved. A metrizability problem of the quotient (a shape space) is discussed. This leads to a new algebraic variety with a metrizability result. A concrete example is given from stereo vision.

💡 Research Summary

The paper introduces a novel algebraic notion called “representational consistency,” which captures the idea that multiple ways of describing the same geometric object do not contradict each other. By formalising this concept within the language of group actions, the author establishes a fundamental theorem: if a set of representations is representationally consistent, then the underlying group action on the data space must be free. The proof proceeds by showing that consistency forces the orbits of the action to be disjoint, which is precisely the definition of a free action. Consequently, the quotient space X/G—referred to as a “shape space”—inherits a well‑behaved topological and differential structure, becoming a smooth manifold whenever the original space X is a manifold.

Having linked consistency to freeness, the paper turns to the metrizability of the resulting shape space. In many applications, the naïve quotient X/G fails to be Hausdorff or complete, making it unsuitable for quantitative comparison of shapes. To overcome this, the author constructs a new algebraic variety V based on the Grassmannian. Each orbit is associated with a subspace of a fixed ambient vector space, and these subspaces are points on a Grassmannian manifold. By embedding the collection of orbits into a Grassmannian and equipping it with a natural distance—derived from principal angles or subspace‑based principal component analysis—the author shows that V is compact, the distance is continuous, symmetric, and satisfies the triangle inequality. Hence V is a complete metric space, providing a rigorous metric on the shape space.

A concrete example from stereo vision illustrates the theory. Two calibrated cameras observe a set of 3‑D points; each camera’s image yields a 2‑D projection, and the pair of projections defines a subspace in ℝ⁴. The transformation group governing the pair of views is a semi‑direct product of SE(3) and GL(2). The paper verifies that this group acts freely on the space of image pairs, and that each image pair corresponds to a point on the Grassmannian Gr(2,4). By mapping all such points into the variety V, the distance on V quantifies the geometric discrepancy between different stereo configurations. This metric is more robust than classical epipolar geometry or simple triangulation, as it directly measures subspace differences and remains stable under noise.

Beyond the stereo case, the author discusses how the Grassmannian‑based construction can be generalized to other Lie groups, non‑linear transformations, and higher‑dimensional data such as multi‑view point clouds or deep‑learning feature embeddings. The paper concludes that representational consistency serves as a bridge linking algebraic properties of group actions to geometric properties of shape spaces, and that the introduced variety V offers both a solid theoretical foundation and practical tools for shape analysis in computer vision and related fields. Future work is suggested on the topological invariants of V, extensions to non‑free actions, and applications to more complex perception problems.