Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

Horospherically Convex Optimization on Hadamard Manifolds Part I: Analysis and Algorithms

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A new kind of convexity lets curved-space optimization match Euclidean rates, using horoballs and Busemann functions in place of half-spaces and affine functions.

desk verdict A serious, mostly sound contribution to Riemannian optimization: the Euclidean-matching rates are proven for an exact-oracle model, but the unanalyzed weighted-Fréchet-mean subproblem keeps the headline claim one step short of real complexity. read the letter →

arxiv 2505.16970 v1 pith:X5DQEDGG submitted 2025-05-22 math.OC cs.NAmath.DGmath.NA

classification math.OCcs.NAmath.DGmath.NA MSC 90C4890C2552A4153C20
keywords horosphericalconvexityHadamardmanifoldsBusemannfunctionsfirst-orderoptimizationcurvature-independentratesNesterovaccelerationFrechetmeanoraclehyperbolicspace
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces a new class of functions on Hadamard manifolds, called horospherically convex (h-convex) functions, built from horoballs and Busemann functions as the curved-space analogues of half-spaces and affine functions. It proves that first-order methods for minimizing sums of h-convex functions converge at exactly the same rates as in Euclidean convex optimization: gradient descent at $L/(2N)\ d(x_0,x_*)^2$, subgradient descent at $DL/\sqrt{N+1}$, and Nesterov acceleration at $2L/N^2$, with linear rates in the strongly convex case. These rates hold without any bound on the sectional curvature, in contrast to known lower bounds for geodesically convex optimization. The paper also shows that in hyperbolic space h-convex subgradient descent can localize minimizers in $O(\log r)$ steps, which is faster than the Euclidean analogue. The main caveat, stated by the authors, is that the analysis assumes an exact weighted-Fréchet-mean oracle whose internal cost is not analyzed.

What carries the argument

The central object is the scaled Busemann function $B_{y,v}(x)$, the curved-space generalization of the affine function $\langle v,\exp_y^{-1}(x)\rangle$; its sublevel sets are horoballs, the analogues of half-spaces. h-convexity requires that at every point such a function support $f$ from below, and Theorem 1 shows that h-convex functions are exactly the pointwise suprema of scaled Busemann functions. The algorithms replace each Riemannian gradient step by minimizing a sum of quadratic models $Q^{1/s_k}_{x_k,g_i^k}$, which is exactly a weighted Fréchet mean, and the convergence proofs all rest on the 'horospherical law of cosines' (Lemma 2), which plays the role of the Euclidean law of cosines in the Lyapunov arguments.

What would settle it

For instance, take the two-term sum of Busemann functions on the hyperbolic plane used in the paper as the counterexample to additivity of h-convexity, minimize it with Algorithm (16) using an oracle that computes the exact weighted Fréchet mean to machine precision, and compare the function-value gaps against the claimed bound $L/(2N)\,d(x_0,x_*)^2$. If the gap sequence exceeds that bound by a factor that grows with the curvature or with $N$, the theorem's premise fails; if replacing the exact oracle by a cheap approximate one degrades the rate, that would show the oracle assumption is genuinely load-bearing.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that geodesic convexity was the wrong generalization of convexity for optimization complexity, and the right one is horospherical convexity. A function $f$ is h-convex if at every point $y$ there is a tangent vector $v$ with $f(x)\ge f(y)+B_{y,v}(x)$ for all $x$, where $B_{y,v}$ is a scaled Busemann function; strong h-convexity replaces $B_{y,v}$ by the squared-distance model $Q^\mu_{y,v}(x)=-\frac{1}{2\mu}\|v\|^2+\frac{\mu}{2}d(\exp_y(-v/\mu),x)^2$, which converges to $B_{y,v}$ as $\mu\to0$. Because h-convex functions are exactly suprema of scaled Busemann functions, they inherit an outer, global structure like Euclidean convexity while remaining geodesically convex. The paper proves that the natural Fréchet-mean update for a sum of h-convex functions yields the Euclidean iteration counts (Theorems 4 through 9), and that in hyperbolic space the same structure gives much faster localization of minimizers (Propositions 7 through 9).

Load-bearing premise

The convergence theorems presuppose an exact weighted-Fréchet-mean oracle, plus exact h-subgradients, at every iteration, and the cost of running this oracle is not analyzed.

Editorial extensions

If this is right

  • Tyler's M-estimator, geometric median, minimal enclosing ball, and Horn's problem become h-convex optimization problems, so the new curvature-independent rates apply to them directly.
  • Smooth strongly h-convex objectives can be minimized at a linear rate without a lower sectional-curvature bound or a feasibility ball, a guarantee that is unavailable for geodesically convex smooth functions.
  • In hyperbolic space, h-convex Lipschitz minimization needs only $O(\log r)$ subgradient steps to shrink a radius-$r$ search region to constant size, compared with $\Theta(r^2/\delta^2)$ for Euclidean convex and $\Theta(r^3/\delta^2)$ for geodesically convex problems.
  • The Moreau envelope of an h-convex function is h-convex and $(1/\lambda)$-h-smooth, so proximal regularization and smoothing work with the same constants as in Euclidean space.
  • The lower bounds that make geodesic convex optimization depend on curvature do not apply to the h-convex subclass, making h-convexity a concrete candidate escape route for scaling problems, though the paper does not resolve those applications.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The exact weighted-Fréchet-mean oracle is the true bottleneck; if that subproblem is expensive on a particular manifold, the curvature-independent iteration counts proven here may not translate into curvature-independent running times, and approximating the oracle could reintroduce geometry-dependent constants.
  • Editorial inference: The proof technique suggests a general lifting rule—replace affine supports with Busemann functions and convex combinations with Fréchet means—so other Euclidean convex algorithms whose proofs rely only on a law of cosines may admit h-convex analogues with matching rates.
  • Editorial inference: For acceleration on sums with $m>1$, the paper assumes the whole sum satisfies the descent condition rather than each summand being $L$-h-smooth; testing whether a sum of two Busemann functions satisfies this condition would delimit how far accelerated h-convex optimization extends.
  • Editorial inference: The hyperbolic-space localization result suggests that h-convexity captures objectives that 'point toward the boundary at infinity,' which may make problems like Horn's problem easier in practice than their geodesic-convexity diameter bounds suggest, but converting this into polynomial-time algorithms still requires controlling bit complexity and oracle cost.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces horospherical convexity (h-convexity) for functions on Hadamard manifolds, building on horoballs and scaled Busemann functions as analogues of half-spaces and affine functions. It develops a calculus for h-convex functions (subgradients, strong convexity, L-h-smoothness, Moreau envelopes, epigraph characterizations) and analyzes first-order methods for minimizing sums of h-convex functions. The central results are curvature-independent iteration-count bounds: gradient descent achieves L/(2N) d(x0,x*)^2 (Theorem 4), subgradient descent achieves DL/sqrt(N+1) (Theorem 6), Nesterov acceleration achieves 2L/N^2 (Theorem 8), and linear rates hold in the strongly convex case (Theorems 5, 7, 9). These guarantees are conditional on an exact weighted-Frechet-mean oracle for the update subproblems. The paper also claims faster rates in hyperbolic space using horoball localization (Propositions 7--9).

Significance. If the results hold, the paper identifies a new structural class of non-Euclidean optimization problems whose oracle complexity matches Euclidean convex optimization, bypassing known curvature-dependent lower bounds for general geodesically convex optimization. The framework is coherent and the main convergence proofs in the appendices are self-contained; the horospherical law of cosines (Lemma 2) is a clean and useful tool. The paper is honest about its main limitation: the convergence theorems assume an exact weighted-Frechet-mean oracle, and Section 6 explicitly warns that this oracle 'may be too strong.' However, this caveat is not reflected in the abstract's 'exactly matching Euclidean convex optimization' phrasing, and the internal cost of the oracle is never analyzed. The hyperbolic-space faster-rate results are interesting but rest on a proof step currently justified by a figure rather than a derivation. Overall, the contribution is substantive and publishable in principle, but the oracle gap and the missing rigorous derivation in Proposition 8 need to be addressed.

major comments (4)
  1. [Section 4.2, Eq. (16); Section 6] The claimed 'exactly matching Euclidean convex optimization' is a statement about iteration counts conditional on an exact weighted-Frechet-mean oracle, not about total complexity. Every iteration of (16) requires minimizing a sum of squared distances to the points exp_{x_k}(-s_k g_{ik}), which is itself a geodesically convex optimization problem. For the paper's own flagship application, the Karcher mean f(x)=sum_i d(x,p_i)^2, the subproblem is another Frechet-mean problem of the same class with no closed form for m>1. The internal cost of this oracle, including its dependence on curvature and diameter, is not analyzed. Since Section 3.3 itself quotes the geodesic smoothness constant of squared distances as 2r sqrt(-K)/tanh(r sqrt(-K)), which is exactly the curvature-and-diameter dependence the outer rates are designed to eliminate, the paper should either provide an oracle-cost analysis, or explicitly and prominently reframe the contribution as iteration-complexity results for an idealized oracle model.
  2. [Remark 5] The claimed extension to approximately solving the subproblems is not quantitative: the remark never states the required additive accuracy delta as a function of N, L, D, and the target epsilon, nor does it bound the inner cost needed to reach that accuracy. As written, it is unclear whether the curvature-independent rates survive for any implementable inner solver. Please provide a concrete delta and a matching inner-complexity bound, or delete the implicit claim that the approximate version preserves the rates.
  3. [Appendix D.2, Proposition 8] The proof of Proposition 8 relies on the containment H_k ∩ H'_k ⊆ B(x_{k+1}, R_k) and the formula R_k = arccosh(e^{r_k/2}), which are asserted from 'Figure 3' and 'a simple calculation' rather than derived. Because Proposition 8 is the basis for the faster hyperbolic rates and for the ellipsoid-method complexity bound in Proposition 9, this step is load-bearing. Please replace the figure-based argument with an explicit coordinate computation (for example, in the Poincare ball model), including a proof of the claimed radius formula.
  4. [Appendix B.9] The counterexample to local-to-global h-convexity depends on the numerical verification of the connected component C (Figure 2). Since this is a theoretical counterexample, the existence of C and the claimed local h-convexity of the constructed function should be proven analytically, not justified by a plot or by referencing code.
minor comments (4)
  1. [Section 4.2, before Eq. (16)] The phrase 'If each f_i is h- 1/s -smooth' is garbled; it should read 'If each f_i is h-convex and (1/s)-h-smooth.'
  2. [Appendix D.1, Proposition 7] The proof says 'It is easy to see' for the containment B_k ∩ H_k ⊆ B(x_{k+1}, r_{k+1}); a one-sentence coordinate justification would make the argument self-contained.
  3. [Table 3] The table compares the manifold results with Euclidean guarantees but does not indicate that the manifold results are conditional on the weighted-Frechet-mean oracle; a footnote would help prevent over-interpretation.
  4. [Abstract and Section 6] The abstract's 'exactly matching those from Euclidean space' is stronger than what is proven, since Euclidean algorithms do not require a Frechet-mean oracle; the Section 6 caveat should be reflected in the abstract or in the first contribution bullet.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the convergence rates are proved self-contained from the definitions; the exact-oracle assumption is an acknowledged complexity gap, not a circular input.

full rationale

After walking the derivation chain, I find no circular step. The h-convexity and L-h-smoothness definitions are new assumptions, and Theorems 4-9 prove convergence from these definitions plus Lemma 2 (the horospherical law of cosines) via self-contained energy arguments. Theorem 1's outer characterization is essentially an immediate rewriting of Definition 1 (the 'if' direction is the definition, and the 'only if' direction uses the supporting Busemann function at each point), but this is a characterization, not a prediction or a fitted input. The self-citations [32,33,35] supply curvature lower bounds and an ellipsoid reduction used only as motivation or comparison; none of the convergence theorems depends on them. The paper explicitly flags in Section 6 that the exact weighted-Frechet-mean oracle 'may be too strong'; this is an acknowledged oracle-complexity gap (the subproblem for sums of squared distances is itself a Frechet-mean problem), not a circular reuse of the target result. Remark 5 asserts an approximate-solution extension without giving the accuracy analysis; that is an omitted proof, not circularity. Accordingly the circularity score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 1 invented entities

No free parameters are fitted to data: the theory is analytic and the step sizes are chosen by proof needs, not by fitting. The main assumptions are the Hadamard setting, standard Busemann-function and comparison geometry, plus exact oracles for h-subgradients and Frechet means. These oracles are the main practical weakness and are acknowledged in Section 6.

assumptions (5)
  • domain assumption M is a Hadamard manifold: complete, simply connected, non-positive sectional curvature.
    All definitions and theorems are stated for this class (Section 2.1); the results rely on Cartan-Hadamard and triangle comparison.
  • standard math Busemann functions are geodesically convex, 1-Lipschitz, C^1 with unit gradient, and the map (x, omega, y) -> B_{x,omega}(y) is continuous on M x CM(infinity) x M.
    These properties are cited from standard references [13,14,23] and used in Lemma 2, Theorem 1, and Appendix B.4.
  • standard math Triangle comparison inequality: d(x,y)^2 >= ||exp_p^{-1}(x) - exp_p^{-1}(y)||^2 (Lemma 1).
    Proved in Section 2.1 via [87, Section 6.2.2]; it underpins the horospherical law of cosines (Lemma 2) used in every convergence proof.
  • domain assumption Weighted Frechet means and the minimization of sums of Q-functions in updates (16), (20), (23) have unique solutions.
    The algorithms require exact solutions of these subproblems. Existence and uniqueness follow from strong geodesic convexity of the objective, but the computational cost of finding them is not bounded in the paper.
  • standard math Geometric containment facts in hyperbolic space: an open horoball is a union of geodesic balls (equation 31), and the intersection of two horoballs used in Proposition 8 is contained in an explicit ball.
    Equation (31) is cited from [14, p.23]; the two-horoball containment is asserted from a figure in Appendix D.2 rather than fully derived.
invented entities (1)
  • Horospherically convex function class and h-subdifferential independent evidence
    purpose: A new function class defined by Busemann-function supports, used to obtain curvature-independent convergence guarantees for optimization on Hadamard manifolds.
    The class is precisely defined in Definition 1 and can be checked for individual functions; the paper provides explicit examples (distance, squared distance, Busemann functions, Tyler M-estimator terms), and concurrent work [50] independently defined the same nonsmooth notion, providing external grounding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Horospherically Convex Optimization on Hadamard Manifolds Part I: Analysis and Algorithms." pith.science (2026). https://pith.science/paper/X5DQEDGG

@misc{pith2026250516970,
  author       = {Pith},
  title        = {Pith review of: Horospherically Convex Optimization on Hadamard Manifolds Part I: Analysis and Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X5DQEDGG}},
  note         = {Machine review of arXiv:2505.16970}
}
read the original abstract

Geodesic convexity (g-convexity) is a natural generalization of convexity to Riemannian manifolds. However, g-convexity lacks many desirable properties satisfied by Euclidean convexity. For instance, the natural notions of half-spaces and affine functions are themselves not g-convex. Moreover, recent studies have shown that the oracle complexity of geodesically convex optimization necessarily depends on the curvature of the manifold (Criscitiello and Boumal, 2022; Criscitiello and Boumal, 2023; Hamilton and Moitra, 2021), a computational bottleneck for several problems, e.g., tensor scaling. Recently, Lewis et al. (2024) addressed this challenge by proving curvature-independent convergence of subgradient descent, assuming horospherical convexity of the objective's sublevel sets. Using a similar idea, we introduce a generalization of convex functions to Hadamard manifolds, utilizing horoballs and Busemann functions as building blocks (as proxies for half-spaces and affine functions). We refer to this new notion as horospherical convexity (h-convexity). We provide algorithms for both nonsmooth and smooth h-convex optimization, which have curvature-independent guarantees exactly matching those from Euclidean space; this includes generalizations of subgradient descent and Nesterov's accelerated method. Motivated by applications, we extend these algorithms and their convergence rates to minimizing a sum of horospherically convex functions, assuming access to a weighted-Fr\'echet-mean oracle.

Figures

Figures reproduced from arXiv: 2505.16970 by the authors.

Figure 1
Figure 1. A horoball (purple) and half-space (blue) in the hyperbolic plane (Poincar´e disk model). A horoball can be defined as the limit of a sequence of geodesic balls (portrayed with dashed boundaries). In Euclidean spaces, horoballs are simply half-spaces. On general Hadamard, horoballs differ from half￾spaces. Recent progress in bridging the gap between convexity and g-convexity was made by Lewis et al. [74]. Instead of… view at source ↗
Figure 2
Figure 2. The set C is shaded gray. Q. Is it reasonable to assume that the algorithm can obtain information about f through queries to the subgradient oracle that outputs an element of ∂f(x) for given x ∈ M? One might expect a negative answer since the set ∂f(x) is not fully determined by the local behavior of f (see Appendix B.10). However, we provide a positive answer here. A. When f is h-convex, although it may not be poss… view at source ↗
Figure 3
Figure 3. Bounding the intersection of two horoballs with a ball, in the Poincar´e disk model. 59 [PITH_FULL_IMAGE:figures/full_fig_p059_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. 1-Lipschitz Neural Networks on Hadamard Manifolds

    math.NA 2026-07 conditional novelty 7.0 of 10

    Busemann-gradient-descent layers provably give 1-Lipschitz networks on Hadamard manifolds under a simple stepsize condition.

  2. Online Optimization on Hadamard Manifolds: Curvature Independent Regret Bounds on Horospherically Convex Objectives

    cs.LG 2025-09 conditional novelty 6.0 of 10

    On Hadamard manifolds, online gradient descent achieves Euclidean regret rates O(√T) and O(log T) for h-convex and strongly h-convex losses, with curvature-free constants.

Reference graph

Works this paper leans on

115 extracted references · 68 canonical work pages · cited by 2 Pith papers

  1. [1]

    Princeton University Press, 2008

    P-A Absil, Robert Mahony, and Rodolphe Sepulchre.Optimization algorithms on matrix manifolds. Princeton University Press, 2008. 19The line of research on g-convex accelerated methods is a prime example of this. See [2,4,53,61,65,75,77,78,114]. 26

  2. [2]

    From nesterov’s estimate sequence to riemannian acceleration

    Kwangjun Ahn and Suvrit Sra. From nesterov’s estimate sequence to riemannian acceleration. InConference on Learning Theory, pages 84–118. PMLR, 2020

  3. [3]

    A continuous-time perspective for modeling acceleration in riemannian optimization

    Foivos Alimisis, Antonio Orvieto, Gary B´ ecigneul, and Aurelien Lucchi. A continuous-time perspective for modeling acceleration in riemannian optimization. InInternational Conference on Artificial Intelligence and Statistics, pages 1297–1307. PMLR, 2020

  4. [4]

    Momentum improves optimization on riemannian manifolds

    Foivos Alimisis, Antonio Orvieto, Gary Becigneul, and Aurelien Lucchi. Momentum improves optimization on riemannian manifolds. InInternational conference on artificial intelligence and statistics, pages 1351–1359. PMLR, 2021

  5. [5]

    Allen-Zhu, A

    Z. Allen-Zhu, A. Garg, Y. Li, R. Oliveira, and A. Wigderson. Operator scaling via geodesically convex optimization, invariant theory and polynomial identity testing. InProceedings of the 50th Annual ACM SIGACT Symposium on the Theory of Computing (STOC 2018), 2018

  6. [6]

    Amendola, K

    C. Amendola, K. Kohn, P. Reichenbach, and A. Seigal. Invariant theory and scaling algo- rithms for maximum likelihood estimation.SIAM Journal on Applied Algebra and Geometry, 5(2):304–337, 2021

  7. [7]

    On approximating the riemannian 1-center.Computa- tional Geometry, 46(1):93–104, 2013

    Marc Arnaudon and Frank Nielsen. On approximating the riemannian 1-center.Computa- tional Geometry, 46(1):93–104, 2013

  8. [8]

    Auderset, C

    C. Auderset, C. Mazza, and E.A. Ruh. Angular Gaussian and Cauchy estimation.Journal of Multivariate Analysis, 93(1):180–197, 2005

Show all 115 references
  1. [9]

    Inf-convolution and regularization of convex functions on riemannian manifolds of nonpositive curvature.Revista Matem´ atica Complutense, 19(2):323– 345, 2006

    Daniel Azagra and Juan Ferrera. Inf-convolution and regularization of convex functions on riemannian manifolds of nonpositive curvature.Revista Matem´ atica Complutense, 19(2):323– 345, 2006

  2. [10]

    Nonsmooth analysis and hamilton– jacobi equations on riemannian manifolds.Journal of Functional Analysis, 220(2):304–361, 2005

    Daniel Azagra, Juan Ferrera, and Fernando L´ opez-Mesas. Nonsmooth analysis and hamilton– jacobi equations on riemannian manifolds.Journal of Functional Analysis, 220(2):304–361, 2005

  3. [11]

    An extension theorem for convex functions of class c1, 1 on hilbert spaces.Journal of Mathematical Analysis and Applications, 446(2):1167–1182, 2017

    Daniel Azagra and Carlos Mudarra. An extension theorem for convex functions of class c1, 1 on hilbert spaces.Journal of Mathematical Analysis and Applications, 446(2):1167–1182, 2017

  4. [12]

    Walter de Gruyter GmbH & Co KG, 2014

    Miroslav Bac´ ak.Convex analysis and optimization in Hadamard spaces, volume 22. Walter de Gruyter GmbH & Co KG, 2014

  5. [13]

    Birkh¨ auser Basel, 1995

    Werner Ballmann.Lectures on Spaces of Nonpositive Curvature. Birkh¨ auser Basel, 1995

  6. [14]

    Birkh¨ auser Boston, 1985

    Werner Ballmann, Mikhael Gromov, and Viktor” Schroeder.Manifolds of Nonpositive Cur- vature. Birkh¨ auser Boston, 1985

  7. [15]

    Potential-function proofs for gradient methods.Theory of Computing, 15(1):1–32, 2019

    Nikhil Bansal and Anupam Gupta. Potential-function proofs for gradient methods.Theory of Computing, 15(1):1–32, 2019

  8. [16]

    Fenchel conjugate via busemann function on hadamard manifolds.Applied Mathematics & Optimization, 88(3):83, 2023

    Glaydston de C Bento, Jo˜ ao Cruz Neto, and ´Italo Dowell L Melo. Fenchel conjugate via busemann function on hadamard manifolds.Applied Mathematics & Optimization, 88(3):83, 2023. 27

  9. [17]

    Horo- spherical learning with smart prototypes

    Paul Berg, Bjoern Michele, Minh-Tan Pham, Laetitia Chapel, and Nicolas Courty. Horo- spherical learning with smart prototypes. InBritish Machine Vision Conference (BMVC), 2024

  10. [18]

    Geometry of the space of phylogenetic trees.Advances in Applied Mathematics, 27(4):733–767, 2001

    Louis J Billera, Susan P Holmes, and Karen Vogtmann. Geometry of the space of phylogenetic trees.Advances in Applied Mathematics, 27(4):733–767, 2001

  11. [19]

    Hyperbolic sliced- wasserstein via geodesic and horospherical projections

    Cl´ ement Bonet, Laetitia Chapel, Lucas Drumetz, and Nicolas Courty. Hyperbolic sliced- wasserstein via geodesic and horospherical projections. InTopological, Algebraic and Geo- metric Learning Workshops 2023, pages 334–370. PMLR, 2023

  12. [20]

    Convex sets in hadamard manifolds.Differential Geometry and its Applica- tions, 17(2-3):111–121, 2002

    AA Borisenko. Convex sets in hadamard manifolds.Differential Geometry and its Applica- tions, 17(2-3):111–121, 2002

  13. [21]

    An introduction to optimization on smooth manifolds

    Nicolas Boumal. An introduction to optimization on smooth manifolds. Available online, Aug 2020

  14. [22]

    Cambridge university press, 2004

    Stephen P Boyd and Lieven Vandenberghe.Convex optimization. Cambridge university press, 2004

  15. [23]

    Springer Science & Business Media, 2013

    Martin R Bridson and Andr´ e Haefliger.Metric spaces of non-positive curvature, volume 319. Springer Science & Business Media, 2013

  16. [24]

    Convex optimization: Algorithms and complexity.Foundations and Trends®in Machine Learning, 8(3-4):231–357, 2015

    S´ ebastien Bubeck. Convex optimization: Algorithms and complexity.Foundations and Trends®in Machine Learning, 8(3-4):231–357, 2015

  17. [25]

    A geometric alternative to nesterov’s accelerated gradient descent.arXiv preprint arXiv:1506.08187, 2015

    S´ ebastien Bubeck, Yin Tat Lee, and Mohit Singh. A geometric alternative to nesterov’s accelerated gradient descent.arXiv preprint arXiv:1506.08187, 2015

  18. [26]

    Deciding positivity of littlewood–richardson coeffi- cients.SIAM Journal on Discrete Mathematics, 27(4):1639–1681, 2013

    Peter B¨ urgisser and Christian Ikenmeyer. Deciding positivity of littlewood–richardson coeffi- cients.SIAM Journal on Discrete Mathematics, 27(4):1639–1681, 2013

  19. [27]

    B¨ urgisser, C

    P. B¨ urgisser, C. Franks, A. Garg, R. Oliveira, M. Walter, and A. Wigderson. Towards a theory of non-commutative optimization: Geodesic 1st and 2nd order methods for moment maps and polytopes. In2019 IEEE 60th Annual Symposium on Foundations of Computer Science (FOCS), pages ...

  20. [28]

    Efficient algorithms for tensor scaling, quantum marginals, and moment polytopes

    Peter B¨ urgisser, Cole Franks, Ankit Garg, Rafael Oliveira, Michael Walter, and Avi Wigder- son. Efficient algorithms for tensor scaling, quantum marginals, and moment polytopes. In 2018 IEEE 59th Annual Symposium on Foundations of Computer Science (FOCS), pages 883–897, 2018

  21. [29]

    Interior-point methods for unconstrained geometric programming and scaling problems, 2020

    Peter B¨ urgisser, Yinan Li, Harold Nieuwboer, and Michael Walter. Interior-point methods for unconstrained geometric programming and scaling problems, 2020

  22. [30]

    Horopca: Hyperbolic dimen- sionality reduction via horospherical projections

    Ines Chami, Albert Gu, Dat P Nguyen, and Christopher R´ e. Horopca: Hyperbolic dimen- sionality reduction via horospherical projections. InInternational Conference on Machine Learning, pages 1419–1429. PMLR, 2021

  23. [31]

    Geometrical and statistical properties of m-estimates of scatter on grassmann manifolds.arXiv preprint arXiv:1812.11605, 2018

    Corina Ciobotaru and Christian Mazza. Geometrical and statistical properties of m-estimates of scatter on grassmann manifolds.arXiv preprint arXiv:1812.11605, 2018. 28

  24. [32]

    Negative curvature obstructs acceleration for strongly geodesically convex optimization, even with exact first-order oracles

    Christopher Criscitiello and Nicolas Boumal. Negative curvature obstructs acceleration for strongly geodesically convex optimization, even with exact first-order oracles. InConference on Learning Theory, pages 496–542. PMLR, 2022

  25. [33]

    Curvature and complexity: Better lower bounds for geodesically convex optimization

    Christopher Criscitiello and Nicolas Boumal. Curvature and complexity: Better lower bounds for geodesically convex optimization. InThe Thirty Sixth Annual Conference on Learning Theory, pages 2969–3013. PMLR, 2023

  26. [34]

    Christopher Criscitiello, David Mart ´ ınez-Rubio, and Nicolas Boumal. Open problem: Poly- nomial linearly-convergent method for g-convex optimization? In Gergely Neu and Lorenzo Rosasco, editors,Proceedings of Thirty Sixth Conference on Learning Theory, volume 195 of Proceedi...

  27. [35]

    Open problem: Poly- nomial linearly-convergent method for geodesically convex optimization?arXiv preprint arXiv:2307.12743, 2023

    Christopher Criscitiello, David Mart ´ ınez-Rubio, and Nicolas Boumal. Open problem: Poly- nomial linearly-convergent method for geodesically convex optimization?arXiv preprint arXiv:2307.12743, 2023

  28. [36]

    Performance of first-order methods for smooth convex mini- mization: a novel approach.Mathematical Programming, 145(1):451–482, 2014

    Yoel Drori and Marc Teboulle. Performance of first-order methods for smooth convex mini- mization: a novel approach.Mathematical Programming, 145(1):451–482, 2014

  29. [37]

    An optimal first order method based on optimal quadratic averaging.SIAM Journal on Optimization, 28(1):251–271, 2018

    Dmitriy Drusvyatskiy, Maryam Fazel, and Scott Roy. An optimal first order method based on optimal quadratic averaging.SIAM Journal on Optimization, 28(1):251–271, 2018

  30. [38]

    The minimum covering sphere problem.Management science, 19(1):96–104, 1972

    D Jack Elzinga and Donald W Hearn. The minimum covering sphere problem.Management science, 19(1):96–104, 1972

  31. [39]

    Horospherical decision boundaries for large margin classification in hyperbolic space.Advances in Neural Information Processing Systems, 36, 2023

    Xiran Fan, Chun-Hao Yang, and Baba Vemuri. Horospherical decision boundaries for large margin classification in hyperbolic space.Advances in Neural Information Processing Systems, 36, 2023

  32. [40]

    Hyperbolization of cusps with convex boundary.Manuscripta mathematica, 150:475–492, 2016

    Fran¸ cois Fillastre, Ivan Izmestiev, and Giona Veronelli. Hyperbolization of cusps with convex boundary.Manuscripta mathematica, 150:475–492, 2016

  33. [41]

    Horoball hulls and extents in positive definite space

    P Thomas Fletcher, John Moeller, Jeff M Phillips, and Suresh Venkatasubramanian. Horoball hulls and extents in positive definite space. InAlgorithms and Data Structures: 12th Interna- tional Symposium, WADS 2011, New York, NY, USA, August 15-17, 2011. Proceedings 12, pages 386...

  34. [42]

    Franks and A

    C. Franks and A. Moitra. Rigorous guarantees for tyler’s m-estimator via quantum expan- sion. In33rd Annual Conference on Learning Theory, volume 125 ofProceedings of Machine Learning Research, pages 1–32. PMLR, 2020

  35. [43]

    Franks, R

    C. Franks, R. Oliveira, A. Ramachandran, and M. Walter. Near optimal sample complexity for matrix and tensor normal models via geodesic convexity.arXiv: 2110.07583, 2021

  36. [44]

    Franks and P

    C. Franks and P. Reichenbach. Barriers for recent methods in geodesic optimization. In Proceedings of the 36th Computational Complexity Conference (CCC 2021), pages 13:1–13:54, 2021

  37. [45]

    Operator scaling with specified marginals

    Cole Franks. Operator scaling with specified marginals. InProceedings of the 50th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2018, page 190–203, New York, NY, USA, 2018. Association for Computing Machinery. 29

  38. [46]

    Linear combinations of hypersurfaces in hyperbolic space.Prepubl

    Eduardo Gallego, Gil Solanes, and Eberhard Teufel. Linear combinations of hypersurfaces in hyperbolic space.Prepubl. Univ. Auton. de Barcelona, 15:1–25, 2010

  39. [47]

    Operator scaling: Theory and applications.Foundations of Computational Mathematics, 20(2):223–290, 2020

    Ankit Garg, Leonid Gurvits, Rafael Oliveira, and Avi Wigderson. Operator scaling: Theory and applications.Foundations of Computational Mathematics, 20(2):223–290, 2020. Issue Date: April 2020

  40. [48]

    Hyperbolic busemann learning with ideal prototypes.Advances in neural information processing systems, 34:103–115, 2021

    Mina Ghadimi Atigh, Martin Keller-Ressel, and Pascal Mettes. Hyperbolic busemann learning with ideal prototypes.Advances in neural information processing systems, 34:103–115, 2021

  41. [49]

    Lewis, Genaro Lopez-Acedo, and Adriana Nicolae

    Ariel Goodwin, Adrian S. Lewis, Genaro Lopez-Acedo, and Adriana Nicolae. Recognizing weighted means in geodesic spaces, 2024

  42. [50]

    Lewis, Genaro Lopez-Acedo, and Adriana Nicolae

    Ariel Goodwin, Adrian S. Lewis, Genaro Lopez-Acedo, and Adriana Nicolae. A subgradient splitting algorithm for optimization on nonpositively curved metric spaces.arXiv preprint arXiv:2412.06730, 2024

  43. [51]

    How to conjugate c 1-close group actions.Mathema- tische Zeitschrift, 132(1):11–20, 1973

    Karsten Grove and Hermann Karcher. How to conjugate c 1-close group actions.Mathema- tische Zeitschrift, 132(1):11–20, 1973

  44. [52]

    No-go theorem for acceleration in the hyperbolic plane

    Linus Hamilton and Ankur Moitra. No-go theorem for acceleration in the hyperbolic plane. arXiv preprint arXiv:2101.05657, 2021

  45. [53]

    Riemannian accelerated gradient methods via extrapolation

    Andi Han, Bamdev Mishra, Pratik Jawanpuria, and Junbin Gao. Riemannian accelerated gradient methods via extrapolation. InInternational Conference on Artificial Intelligence and Statistics, pages 1554–1585. PMLR, 2023

  46. [54]

    American Mathematical Society, Providence, RI, 1984

    Sigurdur Helgason.Groups and Geometric Analysis: Integral Geometry, Invariant Differential Operators, and Spherical Functions, volume 83 ofMathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1984

  47. [55]

    Convex analysis on hadamard spaces and scaling problems.Foundations of Computational Mathematics, pages 1–38, 2023

    Hiroshi Hirai. Convex analysis on hadamard spaces and scaling problems.Foundations of Computational Mathematics, pages 1–38, 2023

  48. [56]

    Interior-point methods on manifolds: theory and applications

    Hiroshi Hirai, Harold Nieuwboer, and Michael Walter. Interior-point methods on manifolds: theory and applications . In2023 IEEE 64th Annual Symposium on Foundations of Computer Science (FOCS), pages 2021–2030, Los Alamitos, CA, USA, November 2023. IEEE Computer Society

  49. [57]

    Springer science & business media, 1996

    Jean-Baptiste Hiriart-Urruty and Claude Lemar´ echal.Convex analysis and minimization algorithms I: Fundamentals, volume 305. Springer science & business media, 1996

  50. [58]

    Springer Science & Business Media, 2007

    Lars H¨ ormander.Notions of convexity. Springer Science & Business Media, 2007

  51. [59]

    A. Horn. Eigenvalues of sums of hermitian matrices.Pacific Journal of Mathematics, 12:225– 241, 1962

  52. [60]

    Show that if a convexchas a supporting hyperplane at every point of its boundary, then it’s convex

    Robert Israel. Show that if a convexchas a supporting hyperplane at every point of its boundary, then it’s convex. Mathematics Stack Exchange

  53. [61]

    Understanding riemannian acceleration via a proximal extragradient framework

    Jikai Jin and Suvrit Sra. Understanding riemannian acceleration via a proximal extragradient framework. InConference on Learning Theory, pages 2924–2962. PMLR, 2022. 30

  54. [62]

    Convex functions on symmetric spaces, side lengths of polygons and the stability inequalities for weighted configurations at infinity

    Michael Kapovich, Bernhard Lee, and John Millson. Convex functions on symmetric spaces, side lengths of polygons and the stability inequalities for weighted configurations at infinity. Journal of Differential Geometry, 81(2):297–354, 2009

  55. [63]

    H. Karcher. Riemannian center of mass and mollifier smoothing.Communications on pure and applied mathematics, 30(5):509–541, 1977

  56. [64]

    Optimized first-order methods for smooth convex minimization.Mathematical programming, 159:81–107, 2016

    Donghwan Kim and Jeffrey A Fessler. Optimized first-order methods for smooth convex minimization.Mathematical programming, 159:81–107, 2016

  57. [65]

    Accelerated gradient methods for geodesically convex op- timization: Tractable algorithms and convergence analysis

    Jungbin Kim and Insoon Yang. Accelerated gradient methods for geodesically convex op- timization: Tractable algorithms and convergence analysis. InInternational Conference on Machine Learning, pages 11255–11282. PMLR, 2022

  58. [66]

    Convexity properties of the moment mapping, iii.Inventiones mathematicae, 77(3):547–552, 1984

    Frances Kirwan. Convexity properties of the moment mapping, iii.Inventiones mathematicae, 77(3):547–552, 1984

  59. [67]

    (MN- 31), Volume 31, volume 104

    Frances Clare Kirwan.Cohomology of Quotients in Symplectic and Algebraic Geometry. (MN- 31), Volume 31, volume 104. Princeton University Press, 1984

  60. [68]

    Klyachko

    Alexander A. Klyachko. Stable bundles, representation theory and hermitian operators.Se- lecta Mathematica, New Series, 4(3):419–445, 1998

  61. [69]

    The symplectic and algebraic geometry of horn’s problem.Linear Algebra and its Applications, 319(1):61–81, 2000

    Allen Knutson. The symplectic and algebraic geometry of horn’s problem.Linear Algebra and its Applications, 319(1):61–81, 2000

  62. [70]

    The honeycomb model of GL n(C) tensor products i: Proof of the saturation conjecture.Journal of the American Mathematical Society, 12(4):1055–1090, 1999

    Allen Knutson and Terence Tao. The honeycomb model of GL n(C) tensor products i: Proof of the saturation conjecture.Journal of the American Mathematical Society, 12(4):1055–1090, 1999

  63. [71]

    A simpler approach to obtaining an o (1/t) convergence rate for the projected stochastic subgradient method.arXiv preprint arXiv:1212.2002, 2012

    Simon Lacoste-Julien, Mark Schmidt, and Francis Bach. A simpler approach to obtaining an o (1/t) convergence rate for the projected stochastic subgradient method.arXiv preprint arXiv:1212.2002, 2012

  64. [72]

    Springer, 2018

    John M Lee.Introduction to Riemannian manifolds, volume 2. Springer, 2018

  65. [73]

    Gradient descent with a general cost.arXiv preprint arXiv:2305.04917, 2023

    Flavien L´ eger and Pierre-Cyril Aubin-Frankowski. Gradient descent with a general cost.arXiv preprint arXiv:2305.04917, 2023

  66. [74]

    Horoballs and the subgradient method.arXiv preprint arXiv:2403.15749, 2024

    Adrian S Lewis, Genaro Lopez-Acedo, and Adriana Nicolae. Horoballs and the subgradient method.arXiv preprint arXiv:2403.15749, 2024

  67. [75]

    Acceler- ated algorithms for convex and non-convex optimization on manifolds.arXiv preprint arXiv:2010.08908, 2020

    Lizhen Lin, Bayan Saparbayeva, Michael Minyi Zhang, and David B Dunson. Acceler- ated algorithms for convex and non-convex optimization on manifolds.arXiv preprint arXiv:2010.08908, 2020

  68. [76]

    Underestimate sequences via quadratic averaging.arXiv preprint arXiv:1710.03695, 2017

    Chenxin Ma, Naga Venkata C Gudapati, Majid Jahani, Rachael Tappenden, and Martin Tak´ ac. Underestimate sequences via quadratic averaging.arXiv preprint arXiv:1710.03695, 2017

  69. [77]

    Global riemannian acceleration in hyperbolic and spherical spaces

    David Mart ´ ınez-Rubio. Global riemannian acceleration in hyperbolic and spherical spaces. InInternational Conference on Algorithmic Learning Theory, pages 768–826. PMLR, 2022. 31

  70. [78]

    Accelerated riemannian optimization: Han- dling constraints with a prox to bound geometric penalties

    David Mart ´ ınez-Rubio and Sebastian Pokutta. Accelerated riemannian optimization: Han- dling constraints with a prox to bound geometric penalties. InThe Thirty Sixth Annual Conference on Learning Theory, pages 359–393. PMLR, 2023

  71. [79]

    Convergence and trade- offs in riemannian gradient descent and riemannian proximal point.arXiv preprint arXiv:2403.10429, 2024

    David Mart ´ ınez-Rubio, Christophe Roux, and Sebastian Pokutta. Convergence and trade- offs in riemannian gradient descent and riemannian proximal point.arXiv preprint arXiv:2403.10429, 2024

  72. [80]

    Horocyclically convex univalent functions.Michigan Mathe- matical Journal, 53(3):483–496, 2005

    D Mej ´ ıa and Ch Pommerenke. Horocyclically convex univalent functions.Michigan Mathe- matical Journal, 53(3):483–496, 2005

  73. [81]

    Proximit´ e et dualit´ e dans un espace hilbertien.Bulletin de la Soci´ et´ e math´ ematique de France, 93:273–299, 1965

    Jean-Jacques Moreau. Proximit´ e et dualit´ e dans un espace hilbertien.Bulletin de la Soci´ et´ e math´ ematique de France, 93:273–299, 1965

  74. [82]

    Some np-complete problems in quadratic and non- linear programming.Mathematical Programming, 39:117–129, 1987

    Katta G Murty and Santosh N Kabadi. Some np-complete problems in quadratic and non- linear programming.Mathematical Programming, 39:117–129, 1987

  75. [83]

    Nemirovski.Information-based complexity of convex programming

    A. Nemirovski.Information-based complexity of convex programming. 1994

  76. [84]

    On complexity of matrix scaling.Linear Algebra and its Applications, 302-303:435–460, 1999

    Arkadi Nemirovski and Uriel Rothblum. On complexity of matrix scaling.Linear Algebra and its Applications, 302-303:435–460, 1999

  77. [85]

    A method for solving the convex programming problem with convergence rateo(1/k 2)

    Yu E Nesterov. A method for solving the convex programming problem with convergence rateo(1/k 2). InDokl. akad. nauk Sssr, volume 269, pages 543–547, 1983

  78. [86]

    Springer, 2018

    Yurii Nesterov.Lectures on convex optimization, volume 137. Springer, 2018

  79. [87]

    Springer, 2006

    Peter Petersen.Riemannian geometry, volume 171. Springer, 2006

  80. [88]

    J. G. Ratcliffe.Foundations of Hyperbolic Manifolds. Springer International Publishing, 2019

  81. [89]

    Princeton university press, 1997

    R Tyrrell Rockafellar.Convex analysis, volume 11. Princeton university press, 1997

  82. [90]

    Springer Science & Business Media, 2009

    R Tyrrell Rockafellar and Roger J-B Wets.Variational analysis, volume 317. Springer Science & Business Media, 2009

  83. [91]

    Implicit riemannian opti- mism with applications to min-max problems.arXiv preprint arXiv:2501.18381, 2025

    Christophe Roux, David Mart ´ ınez-Rubio, and Sebastian Pokutta. Implicit riemannian opti- mism with applications to min-max problems.arXiv preprint arXiv:2501.18381, 2025

  84. [92]

    Springer Science & Business Media, 2013

    Alexander M Rubinov.Abstract convexity and global optimization, volume 44. Springer Science & Business Media, 2013

  85. [93]

    Rusciano

    A. Rusciano. A riemannian corollary of helly’s theorem.arXiv: 1804.10738, 2019

  86. [94]

    Computationally related problems.SIAM Journal on computing, 3(4):262–279, 1974

    Sartaj Sahni. Computationally related problems.SIAM Journal on computing, 3(4):262–279, 1974

  87. [95]

    Horocycles and convex sets in hyperbolic plane.Archiv der Mathematik, 18:529–533, 1967

    Luıs Antoni Santal´ o. Horocycles and convex sets in hyperbolic plane.Archiv der Mathematik, 18:529–533, 1967

  88. [96]

    Horospheres and convex bodies in hyperbolic space.Proceedings of the American Mathematical Society, 19(2):390–395, 1968

    Luıs Antoni Santal´ o. Horospheres and convex bodies in hyperbolic space.Proceedings of the American Mathematical Society, 19(2):390–395, 1968. 32

  89. [97]

    Singer.Abstract Convex Analysis

    I. Singer.Abstract Convex Analysis. Wiley-Interscience and Canadian Mathematics Series of Monographs and Texts. Wiley, 1997

  90. [98]

    L. T. Skovgaard. A riemannian geometry of the multivariate normal model.Scandinavian Journal of Statistics, 11(4):211–223, 1984

  91. [99]

    Fully-connected network on noncompact symmetric space and ridgelet transform based on helgason-fourier analysis

    Sho Sonoda, Isao Ishikawa, and Masahiro Ikeda. Fully-connected network on noncompact symmetric space and ridgelet transform based on helgason-fourier analysis. InInternational Conference on Machine Learning, pages 20405–20422. PMLR, 2022

  92. [100]

    Sra and R

    S. Sra and R. Hosseini. Conic geometric optimization on the manifold of positive definite matrices.SIAM Journal on Optimization, 25(1):713–739, 2015

  93. [101]

    S. Sra, N. K. Vishnoi, and O. Yildiz. On geodesically convex formulations for the brascamp- lieb constant. In Eric Blais, Klaus Jansen, Jos´ e D. P. Rolim, and David Steurer, editors,Ap- proximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, APP...

  94. [102]

    T. M. Tang and G. I. Allen. Integrated principal components analysis.arXiv: 1810.00832, 2021

  95. [103]

    Smooth strongly convex interpo- lation and exact worst-case performance of first-order methods.Mathematical Programming, 161:307–345, 2017

    Adrien B Taylor, Julien M Hendrickx, and Fran¸ cois Glineur. Smooth strongly convex interpo- lation and exact worst-case performance of first-order methods.Mathematical Programming, 161:307–345, 2017

  96. [104]

    David E. Tyler. A distribution-free m-estimator of multivariate scatter.The Annals of Statistics, 15(1):234–251, March 1987

  97. [105]

    Springer Science & Business Media, 1994

    Constantin Udriste.Convex functions and optimization methods on Riemannian manifolds, volume 297. Springer Science & Business Media, 1994

  98. [106]

    Boundary structure of convex sets in the hyperbolic space.Monatshefte f¨ ur Mathematik, 188(3):567–586, 2019

    Giona Veronelli. Boundary structure of convex sets in the hyperbolic space.Monatshefte f¨ ur Mathematik, 188(3):567–586, 2019

  99. [107]

    Springer, 2008

    C´ edric Villani et al.Optimal transport: old and new, volume 338. Springer, 2008

  100. [108]

    Entanglement poly- topes: Multiparticle entanglement from single-particle information.Science, 340(6137):1205– 1208, 2013

    Michael Walter, Brent Doran, David Gross, and Matthias Christandl. Entanglement poly- topes: Multiparticle entanglement from single-particle information.Science, 340(6137):1205– 1208, 2013

  101. [109]

    JCB Mohr (Paul Siebeck), 1922

    Alfred Weber.Ueber den standort der industrien, volume 2. JCB Mohr (Paul Siebeck), 1922

  102. [110]

    A. Wiesel. Geodesic convexity and covariance estimation.IEEE Transactions on Signal Processing, 60(12):6182–6189, 2012

  103. [111]

    Wiesel and T

    A. Wiesel and T. Zhang. Structured robust covariance estimation.Foundations and Trends in Signal Processing, 8(3):127–216, 2014

  104. [112]

    X. Yuan, W. Huang, P.-A. Absil, and K. A. Gallivan. Computing the matrix geometric mean: Riemannian versus euclidean conditioning, implementation techniques, and a rieman- nian BFGS method.Numer. Linear Algebra Appl., 27(5), 2020. 33

  105. [113]

    First-order methods for geodesically convex optimization

    Hongyi Zhang and Suvrit Sra. First-order methods for geodesically convex optimization. In Conference on learning theory, pages 1617–1638. PMLR, 2016

  106. [114]

    An estimate sequence for geodesically convex optimization

    Hongyi Zhang and Suvrit Sra. An estimate sequence for geodesically convex optimization. In Conference On Learning Theory, pages 1703–1723. PMLR, 2018

  107. [115]

    − 1 2µ ∥v∥2 + µ 2 d expy − 1 µ v , x 2# = lim s→∞ − s 2 ∥v∥2 + 1 2s d expy (−sv), x 2 (s←1/µ) = lim t→∞

    T. Zhang. Robust subspace recovery by geodesically convex optimization. 2012. A Busemann functions: Details for Section 2 A.1 Cone at infinityCM(∞)as dual space ofM The idea of consideringCM(∞) as the dual space ofMcan be found in [55]. Our approach differs from their work onl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.