Hadamard Spaces: Geometry & Applications

• Hadamard spaces are complete geodesic metric spaces of nonpositive curvature that generalize Hilbert spaces and Riemannian manifolds.
• They support advanced convex analysis and optimization via unique geodesic convexity, proximal mappings, and duality frameworks.
• Applications range from geometric group theory and imaging to stochastic processes and combinatorial optimization, driving innovative research.

A Hadamard space is a complete geodesic metric space of nonpositive curvature in the sense of Alexandrov, equivalently a complete CAT(0) space. Hadamard spaces generalize Hilbert spaces and nonpositively curved Riemannian manifolds, providing a unifying nonlinear framework for convex analysis, optimization, fixed-point theory, monotone operator theory, and stochastic processes. These spaces exhibit a rich interplay between metric geometry, convexity, duality, and functional analysis, with significant implications in diverse areas such as geometric group theory, operator scaling, probabilistic geometry, combinatorial optimization, imaging, and computational phylogenetics.

1. Metric and Geometric Structure

A metric space $(X,d)$ is CAT(0) if for any triangle with vertices $x, y, z$, the squared distance between points on geodesics is bounded above by their Euclidean comparison triangle. Formally, for any $t \in [0,1]$,

$$d\bigl((1-t)x \oplus t y, z\bigr)^2 \leq (1-t) d(x,z)^2 + t d(y,z)^2 - t(1-t) d(x,y)^2.$$

A complete CAT(0) space is termed a Hadamard space. These spaces are uniquely geodesic: for any pair $x, y$, there exists a unique constant-speed geodesic joining them. Squared distance functions are strongly convex along geodesics, yielding uniqueness of metric projections on closed convex sets and firmly nonexpansive properties for projections and resolvents (Bacak, 2018).

Hadamard spaces admit a well-defined boundary at infinity $\partial X$, equipped with the cone topology and the Tits metric, supporting asymptotic geometric and functional analytic constructions (Lenze, 8 Apr 2025, Hirai, 2022).

2. Convexity, Optimization, and Duality

A function $f: X \to \mathbb{R} \cup \{+\infty\}$ is geodesically convex if $f((1-t)x \oplus t y) \leq (1-t) f(x) + t f(y)$ along all geodesics. Variational inequalities, strong convexity, and slopes can all be formulated and controlled via the CAT(0) inequality. The proximal mapping

$$\mathrm{prox}_{\lambda f}(x) = \arg\min_y \left\{ f(y) + \frac{1}{2\lambda} d(x,y)^2 \right\}$$

is single-valued and nonexpansive.

Convex analysis extends to Hadamard spaces via the asymptotic slope (“recession”) function, Busemann functions, and the Legendre–Fenchel (Busemann) conjugate: $$f^\infty(p) = \lim_{t\to\infty} \frac{f(c(t))}{t}, \qquad f^*(p) = \sup_{x} \left[ - b_p(x) - f(x) \right],$$ where $x, y, z$0 is a geodesic ray with $x, y, z$1, and $x, y, z$2 is the corresponding Busemann function. These duality notions characterize boundedness/unboundedness and enable generalized Fenchel duality and Fitzpatrick transforms for monotone operators (Hirai, 2022, Moslemipour et al., 2021).

3. Weak Convergence, Δ-Convergence, and Functional Analysis

Several weak convergence concepts are available:

• Δ-convergence: The natural “weak” convergence in CAT(0) spaces. Every bounded sequence admits a Δ-convergent subsequence; Δ-convergent sequences have unique asymptotic centers; closed convex subsets are Δ-closed; and Picard iterations of quasi-nonexpansive, Δ-demiclosed mappings Δ-converge to fixed points (Worapitpong et al., 30 Apr 2025).
• Weak topology $x, y, z$3: Characterized by convergence of metric projections onto geodesics. In a “weakly proper” Hadamard space (e.g., locally compact or Hilbert space), bounded closed convex sets are weakly compact, closed convex and weakly closed coincide, and Mazur's, Eberlein–Šmulian, and Banach–Alaoglu analogues hold (Bërdëllima, 2022).
• Comparison with Monod’s and Kakavandi's topologies: In locally compact spaces, the main weak topologies ($x, y, z$4) coincide; in general, $x, y, z$5, mirroring Hilbert space relationships.

The dual space is formulated in terms of geodesic structures, generalizing the linear dual in Hilbert spaces (Bërdëllima, 2022, Moslemipour et al., 2021).

4. Algorithms and Applications in Optimization and Signal Processing

Hadamard spaces support proximal point algorithms (PPA), splitting methods, and gradient flows, generalizing classical algorithms to nonpositively curved, nonlinear settings (Bacak, 2018). Key components include:

• Proximal point iteration: Weak (Δ-) convergence to a minimizer for geodesically convex, lsc functions.
• Splitting and product resolvent methods: Allowing minimization of finite sums of convex functions or convex feasibility problems via cycling metric projections—understood using strong quasi-nonexpansivity and orbital Δ-demiclosedness (Worapitpong et al., 30 Apr 2025).
• Gradient flows: Classical evolution variational inequalities and strong continuous semigroup constructions persist, enabling applications in geometric flows such as Calabi flow in Kähler geometry.

Applications span computational phylogenetics (BHV-tree space), imaging (SPD matrix-valued variational models), and combinatorial optimization (submodular minimization via Lovász extensions in CAT(0) complexes) (Bacak, 2018).

In stochastic analysis, Hadamard spaces support well-behaved Fréchet and inductive means; strong laws of large numbers, concentration inequalities, and robustness to contamination generalize Sturm and Ziezold's laws. The geometry (e.g., “sticky” structures like open books) can yield unexpectedly high robustness properties for means, with performance modulated by intrinsic curvature (Köstenberger et al., 2023).

5. Cyclic and Alternating Projections: Regularity and Obstructions

Metric projections in Hadamard spaces are 1-Lipschitz and satisfy firm nonexpansivity via variational inequalities. For two closed convex sets, alternating projections exhibit asymptotic regularity and $x, y, z$6 decay rates, as in Hilbert spaces. However, for three or more convex sets, cyclic projections can fail to be asymptotically regular, due to the lack of global flatness and richer geometric possibilities (e.g., product-of-tripod examples), fundamentally distinguishing Hadamard spaces from Hilbert spaces (Lytchak et al., 2021).

This failure motivates ongoing inquiries into which geometric configurations or additional curvature/convexity hypotheses restore regularity, with implications for algorithmic feasibility in general CAT(0) contexts.

6. Higher Rank and Affine Rigidity

Recent work provides a metric characterization of higher-rank Hadamard spaces via non-dilational affine maps. A geodesically complete, locally compact CAT(0) space with a geometric group action admits an affine map not globally a dilation if and only if the space is a higher-rank symmetric space, a higher-rank Euclidean building, or a nontrivial metric product. Rank-one and irreducible spaces are affinely rigid: all affine maps are dilations (Lenze, 8 Apr 2025).

The presence of nontrivial closed symmetric subsets in the boundary at infinity, detected via reparametrizations along geodesic rays, links the existence of flexible affine deformations to higher-rank rigidity and classical structure theory. Explicit examples arise from symmetric spaces ($x, y, z$7), Euclidean buildings, and metric products.

7. Open Problems and Future Directions

Significant open challenges in Hadamard space theory include:

• Full development of subdifferential and duality theory beyond current Fenchel–Young analogues (Moslemipour et al., 2021, Bacak, 2018).
• Topological compactness questions for convex hulls of compact sets.
• Uniform convergence of proximal and splitting algorithms absent local compactness.
• Precise characterization of asymptotic regularity regimes for general cyclic projection schemes.
• Intrinsic geometric implications for “nonlinear” probability theory, invariant theory (operator and moment polytope problems), and Kähler geometry (Hirai, 2022, Bacak, 2018).
• Extension of regularity and convergence results to infinite dimensions and more general nonpositively curved spaces.

Hadamard spaces constitute a foundational setting that unifies geometric, analytic, and algorithmic paradigms, driving contemporary research in metric convexity, nonlinear analysis, and their broad spectrum of applications.