Affine Invariant Metrics: Theory & Practice
This presentation explores affine invariant metrics—geometric and statistical measures that remain unchanged under affine transformations such as scaling, rotation, shearing, and translation. We examine their mathematical foundations on symmetric positive-definite matrices and surfaces, survey distinct metric families including classical, power, and Thompson variants, and demonstrate practical applications in image recognition, perceptual modeling, and robust statistics. The talk reveals how the principle of affine invariance enables coordinate-free comparison of complex objects across diverse domains.Script
A metric that sees through coordinate systems—that's the power of affine invariance. When you rotate, scale, or shear an object, an affine invariant metric reports the exact same distance, as if the transformation never happened.
Let's ground this idea in the geometry where it matters most.
On the manifold of symmetric positive-definite matrices, the classical affine-invariant metric is defined by a trace formula involving inverses. Geodesic distances emerge in closed form through matrix logarithms, and crucially, the entire geometry commutes with congruence transformations—any rotation or scaling of your coordinate frame leaves distances unchanged.
Affine invariance doesn't specify a single metric—it defines a family. The classical metric coexists with polar and power variants, each constructed by pulling back geometry through smooth maps. The Thompson metric, using the infinity norm, enables midpoint computation at extreme scale, while the quotient-affine metric restricts to correlation matrices by factoring out diagonal scaling.
Now let's see how these abstractions solve real problems.
On smooth surfaces, mixed volume constructions yield metrics that survive any volume-preserving affine deformation—essential for shape correspondence and symmetry detection. In computer vision, robust affine-invariant metrics handle outliers and prevent scale collapse during image matching. Yet when we measure perceptual invariance empirically, even state-of-the-art deep metrics fall short of human sensitivity to translation, rotation, and illumination changes.
Affine invariance is not a single tool but a design principle—one that lets us compare objects as they truly are, not as our coordinates happen to describe them. Visit EmergentMind.com to explore more and create your own videos.
