In the "friendly" geometry of Hilbert spaces, the Euclidean distance is a reliable guide. It respects the parallelogram law, allows for simple projections, and ensures that iterative algorithms behave predictably. However, many critical problems in signal processing, image reconstruction, and variational inequalities live in more complex environments: Banach spaces and multi-dimensional complex domains.
In these settings, the standard Euclidean metric becomes "blind" to the underlying geometry. To achieve strong convergence—where an algorithm actually reaches a specific target point rather than just hovering near it—researchers have developed specialized distance measures: Lyapunov functionals for Banach spaces and nonisotropic metrics for complex manifolds.
The Geometry Gap in Banach Spaces
A central challenge in functional analysis is finding fixed points for nonlinear mappings. In a Hilbert space, we can use the norm to measure progress. But as Jean-Philippe Chancelier demonstrates in Iterative schemes for computing fixed points of nonexpansive mappings in Banach spaces, robust convergence requires a framework that transcends specific geometric restrictions like strict convexity.
The problem is that in general Banach spaces, the relationship between a point and its projection onto a set is not as straightforward as a straight line. To fix this, researchers like Shih-sen Chang and Yongchun Xu replace the standard squared norm with a Lyapunov functional (often denoted as $\phi$).
As explored in Strong convergence theorems of quasi-ϕ-asymptotically nonexpansive semi-groups in Banach spaces, this functional acts as a "generalized distance." It accounts for the space's curvature and reflexivity, allowing for generalized projections. By using these projections, algorithms like the Halpern-Mann iteration or the hybrid algorithms proposed by Yongchun Xu in Modified Hybrid Algorithm for a Family of Quasi-Asymptotically Nonexpansive Mappings can force "strong convergence." This is the difference between a sequence simply getting "closer" in a vague sense and actually homing in precisely on a target point in an infinite-dimensional space.
Splitting Methods and Duality
This geometric adaptation is not just for fixed points; it extends to optimization. Wei-Bo Guan shows in The forward–backward splitting methods for variational inequalities and minimization problems in Banach spaces how the Forward-Backward Splitting (FBS) algorithm—a staple of machine learning—can be ported to Banach spaces. By leveraging duality mappings and Lyapunov distance functions, the algorithm "warps" its steps to match the space's natural geometry, ensuring stability where Euclidean methods would diverge.
Nonisotropic Metrics in Complex Analysis
A parallel problem exists in harmonic analysis on the unit ball of $n$-dimensional complex space ($\mathbb{C}^n$). Here, the geometry is nonisotropic, meaning distance does not behave the same way in all directions. If you move along the complex structure of the sphere, the "distance" should be measured differently than if you move across it.
Jaesung Lee and Kyung Soo Rim address this by utilizing a metric defined by the complex inner product: $d(z, w) = |1 - \langle z, w \rangle|^{1/2}$. In their work on the Properties of the ℳ-Harmonic Conjugate Operator, they show that this metric is essential for proving the stability of operators.
Standard Euclidean metrics fail to capture the symmetries of the complex ball. By using the nonisotropic metric, Lee and Rim demonstrate in Estimates of -Harmonic Conjugate Operator and Norm inequalities for the conjugate operator in two-weighted Lebesgue spaces that the $\mathcal{M}$-harmonic conjugate operator remains bounded. This alignment between the metric and the complex manifold's structure allows classical results from real-variable harmonic analysis (like $A_p$ weight conditions) to be successfully applied to the complex unit sphere.
Why It Works: Intuition
The shared intuition across these papers is geometric alignment.
- In Banach Spaces: The Lyapunov functional $\phi(x, y)$ replaces $\|x-y\|^2$ because it correctly interacts with the duality mapping of the space. This ensures that the "gradient" of our distance measure points in a direction that actually leads toward the target in a non-Euclidean landscape.
- In Complex Domains: The nonisotropic metric accounts for the fact that the unit sphere in $\mathbb{C}^n$ is "curved" differently depending on the complex direction. Using a Euclidean ruler on a complex sphere is like using a straight ruler to measure the surface of a ripple; it misses the essential frequency of the space.
By replacing generic distance with geometry-aware functionals, these researchers transform weak, unstable processes into strong, convergent algorithms.