Cosmic divergence, weak cosmic convergence, and fixed points at infinity

Ernest K. RyuJournal of Fixed Point Theory and Applications 2018 · DOI · arXiv

Introduction

In the fields of numerical optimization and functional analysis, fixed-point iterations of the form $x^{k+1} = T(x^k)$ are fundamental tools for finding solutions to complex problems. When the operator $T$ possesses a fixed point, the sequence typically converges toward it. However, a more complex scenario arises when $T$ is "non-expansive"—meaning it does not increase the distance between points—but lacks a fixed point entirely. In such cases, the sequence must diverge, "blowing up" toward infinity.

The "cosmic convergence conjecture" was proposed to describe the behavior of these divergent sequences. It suggested that as the sequence $x^k$ recedes to infinity, its normalized direction $x^k / \|x^k\|$ would eventually settle toward a single point on the unit sphere (the "horizon"). This paper provides a rigorous refutation of this conjecture, demonstrating that non-expansivity alone is insufficient to guarantee directional stability.

The Geometry of Divergence

To investigate the conjecture, the author focuses on operators where the "minimal displacement" is zero. If an operator has a constant "drift" (a non-zero minimal displacement vector), the sequence diverges linearly in that direction. The interesting cases occur when the operator's movement becomes infinitesimally small as it moves further out, allowing for more complex trajectories.

Constructing the Counterexample

The author disproves the conjecture by constructing a specific firmly non-expansive operator $T$ in $\mathbb{R}^2$. This is achieved using the proximal operator of a carefully designed convex function $f(x, y) = \max\{\Phi(x), \Psi(y)\}$. By engineering $\Phi$ and $\Psi$ as "staircase-like" functions with varying step sizes, the author forces the iteration sequence to follow a path that oscillates between different directions.

The result is a sequence where the normalized vectors $x^k / \|x^k\|$ do not converge to a single point. Instead, they "wander" forever, accumulating at multiple distinct directions on the unit circle—specifically $(2/\sqrt{5}, 1/\sqrt{5})$ and $(1/\sqrt{2}, 1/\sqrt{2})$. This proves that even well-behaved, non-expansive maps can produce divergent paths that "wiggle" indefinitely.

Weak Convergence in Infinite Dimensions

The paper extends this analysis to infinite-dimensional spaces, specifically $\ell^2$ (the space of square-summable sequences). Here, the author uncovers an even more counterintuitive phenomenon: weak cosmic convergence to zero.

In this setting, an operator is constructed such that the sequence $x^k$ grows in magnitude toward infinity, yet its direction $x^k / \|x^k\|$ converges weakly to the origin of the unit ball. This implies that the sequence's direction becomes so diffused across an infinite number of coordinates that it effectively "vanishes" from a weak-limit perspective. This result highlights that in infinite dimensions, the "cosmic limit" may not even reside on the boundary of the space.

Fixed Points at Infinity

Despite the failure of the general conjecture, the author provides a new geometric framework to understand these divergent points. By utilizing dual cones and separating hyperplanes, the paper establishes a positive result (Theorem 5.1).

The author shows that for any cosmic accumulation point $q$, the inequality $\langle T(x) - x, q \rangle \geq 0$ holds for all $x$. This means that while the sequence might not converge to a single direction, every accumulation point $q$ defines a half-space that the operator $T$ is constrained to move "toward."

In this sense, these directions can be viewed as "fixed points at infinity." Just as a standard fixed point $x^*$ satisfies $\|T(x) - x^*\| \leq \|x - x^*\|$, these cosmic points define a boundary that the iteration sequence cannot cross, providing a rigorous way to treat the horizon as part of the operator's domain.

Conclusion and Future Outlook

This work serves as a cautionary tale for optimization theory, proving that the intuitive assumption of directional convergence at infinity is not a universal property of non-expansive maps. While the counterexamples provided are mathematically "pathological" and may not appear in every practical application, they define the absolute limits of what can be assumed about divergent iterations.

The primary open question remains: what additional structural conditions are required to guarantee cosmic convergence? Future research may look toward the specific properties of convex functions used in proximal algorithms to find a middle ground between the general non-expansive case and the well-behaved convergence seen in practice.

Summary updated

BibTeX
@article{ryu2018cosmic,
  author = {Ernest K. Ryu},
  title = {{Cosmic divergence, weak cosmic convergence, and fixed points at infinity}},
  journal = {Journal of Fixed Point Theory and Applications},
  year = {2018},
  doi = {10.1007/s11784-018-0592-8},
}
Figure