arxiv: 2604.26913 · v1 · submitted 2026-04-29 · 🧮 math.OC · cs.NA· math.NA· math.PR

Recognition: unknown

Generalization of Zeroth-Order Method for Quotients of Quadratic Functions

Jonas Bresch

Authors on Pith no claims yet

Pith reviewed 2026-05-07 11:19 UTC · model grok-4.3

classification 🧮 math.OC cs.NAmath.NAmath.PR

keywords zeroth-order methodgeneralized Rayleigh quotientquadratic function quotientsRiemannian optimizationspherical samplingoperator normaccelerated optimizationGram matrices

0 comments

The pith

Unconstrained sampling on the entire sphere enables closed-form zeroth-order estimates for generalized Rayleigh quotients of quadratic functions.

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

This paper establishes a generalization of zeroth-order methods to handle quotients of quadratic functions by sampling random directions freely on the sphere rather than restricting to the tangent space. The method estimates the generalized operator norm and the extremal generalized Rayleigh quotient, which in turn serve as surrogates for the Riemannian gradient and Hessian. Despite skipping the tangent space, the optimal step size remains computable in closed form. The subproblems are recast as sub-generalized Rayleigh quotient problems on specific Gram matrices. If correct, this framework supports an accelerated algorithm that achieves state-of-the-art performance and surpasses previous methods in optimization tasks involving quadratic quotients.

Core claim

The paper claims that unconstrained sampling on the entire sphere for random search directions in each iteration allows calculation of the generalized operator norm and extremal generalized Rayleigh quotient. This provides a link to zeroth-order methods for Riemannian optimization where the Riemannian gradient and Hessian are estimated by specific surrogates. The optimal step size problem is computed in closed form, and the subproblems are interpreted as sub-generalized Rayleigh quotient problems on Gram matrices, enabling construction of an accelerated algorithm with state-of-the-art behavior.

What carries the argument

Unconstrained spherical sampling of random search directions that estimates generalized operator norms and Rayleigh quotients without tangent space projections, serving as surrogates for Riemannian gradients and Hessians.

If this is right

Links zeroth-order methods directly to Riemannian first- and second-order optimization.
Enables closed-form computation of the optimal step size without using the tangent space.
Frames subproblems as sub-generalized Rayleigh quotient problems on Gram matrices.
Supports construction of an accelerated algorithm that outperforms recent works.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This approach may reduce computational cost in high-dimensional settings by avoiding tangent space calculations.
It could be extended to other optimization problems on manifolds where projections are expensive.
Experimental validation on standard benchmarks in subspace optimization would test scalability beyond the paper's claims.

Load-bearing premise

The unconstrained sampling on the entire sphere for random search directions produces reliable estimates of the generalized operator norm and extremal generalized Rayleigh quotient without requiring the tangent space.

What would settle it

A numerical experiment where the estimates from unconstrained sampling deviate significantly from those obtained with tangent-space methods or where the accelerated algorithm fails to outperform baselines on quadratic quotient problems.

Figures

Figures reproduced from arXiv: 2604.26913 by Jonas Bresch.

**Figure 3.** Figure 3: We observe that the tremendous improvement by the m-sampling approach compared to a single, uniform at random step direction on the unit sphere. Besides the observable linear convergence rate for the relative error for estimating the leading generalized eigenvalue, the minimal squared residual, as well as the estimate for the length of the true Riemannian gradient, is doing much better than a sublinear be… view at source ↗

read the original abstract

Optimization of quadratic functions and the quotient of those are relevant in subspace and iterative optimization methods. In this paper, the calculation of the generalized operator norm and extremal generalized Rayleigh quotient is considered. In contrast to recent works an unconstrained sampling approach on the entire sphere for the random search direction in each iteration is proposed. Furthermore, the link to zeroth-order methods for Riemannian first- and second-order optimization methods is provided in the sense that the Riemannian gradient and Hessian are estimated by the specific surrogates. Even though the tangent space is not used in this construction the optimal step size problem can be computed in a closed form. The subproblems of this and recent works are illuminated in the context of sub-generalized Rayleigh quotient problems on specific Gram matrices. Together the achieved theory allows to construct an accelerated algorithm which shows state-of-the-art behavior and outperforms recent works.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's core move is unconstrained sphere sampling for zeroth-order estimates on quadratic quotients, with a Gram-matrix reduction that claims to deliver closed-form steps and acceleration, but the radial bias risk is the part that needs checking.

read the letter

The new piece is the switch to sampling random directions on the full sphere instead of the tangent space, plus the explicit link that turns those samples into surrogates for the Riemannian gradient and Hessian of the quotient. The paper also recasts the subproblems from recent work as generalized Rayleigh quotients on particular Gram matrices, which gives a clean way to compare them and to derive the accelerated variant. That framing is useful and seems to be the main technical contribution beyond the sampling change itself. The closed-form step-size claim follows from the same reduction, which is neat if it holds without extra assumptions. Credit for spelling out how the zeroth-order construction sits inside the Riemannian picture even when the tangent space is dropped. The soft spot is exactly the one the stress-test note raises. Directions sampled on the whole sphere are not orthogonal to the current point, so the finite-difference quotients pick up a radial component that has no meaning as a variation on the manifold. The paper asserts that the subsequent Gram-matrix step removes the bias and restores the surrogate equivalence, but without seeing the explicit bound or cancellation argument it is not obvious that the first- and second-order estimates remain accurate enough for the acceleration guarantee. If that step only works under additional restrictions that are not stated, the state-of-the-art performance claim weakens. Experiments are mentioned but not described in enough detail here to judge the practical size of any bias. This is for people who already work on subspace methods or zeroth-order Riemannian optimization and need a variant that avoids explicit tangent-space projections. A reader who cares about the generalized Rayleigh quotient setting will find the subproblem comparison worthwhile. It deserves a serious referee because the algorithmic construction is distinct and the Gram-matrix view is reproducible, even though the bias question will need a clear answer in review.

Referee Report

3 major / 1 minor

Summary. The manuscript generalizes zeroth-order methods to optimization of quotients of quadratic functions. It proposes unconstrained uniform sampling of random directions on the full sphere (without tangent-space projection) to estimate surrogates for the Riemannian gradient and Hessian, derives a closed-form optimal step size for the resulting subproblem, relates the subproblems to generalized Rayleigh quotients on Gram matrices, and constructs an accelerated algorithm asserted to achieve state-of-the-art performance that outperforms recent works.

Significance. If the central claims hold—specifically that unconstrained sphere sampling yields unbiased surrogates for the generalized operator norm and extremal Rayleigh quotient, that the closed-form step size remains valid, and that the resulting acceleration is rigorous—this approach could simplify zeroth-order Riemannian optimization for quadratic-quotient problems arising in subspace methods by removing the need for tangent-space projections. The explicit linkage to Gram-matrix subproblems and the provision of closed-form steps would be useful strengths, but the current manuscript supplies no derivations, error bounds, or experimental details to support these assertions.

major comments (3)

[Abstract] Abstract: the assertion that 'even though the tangent space is not used ... the optimal step size problem can be computed in a closed form' is load-bearing for the entire contribution. The skeptic's concern is valid on the supplied evidence: directions sampled uniformly on the sphere are not guaranteed to lie in the tangent space at the current point, so the finite-difference quotients contain an unprojected radial component. No error bound or cancellation argument via the subsequent Gram-matrix reduction is supplied, which directly undermines the claimed subproblem equivalence and the acceleration guarantee.
[Abstract] Abstract: the claim that 'together the achieved theory allows to construct an accelerated algorithm which shows state-of-the-art behavior' rests on the validity of the surrogate estimates. Without any derivation of the estimation error for the generalized operator norm or extremal generalized Rayleigh quotient under unconstrained sampling, or any comparison of the resulting convergence rate to the tangent-space-constrained baselines, the performance assertion cannot be evaluated.
[Abstract] Abstract: the subproblem equivalence to 'sub-generalized Rayleigh quotient problems on specific Gram matrices' is presented as illuminating the relation to recent works, yet no explicit equations are given showing how the unconstrained surrogate reduces to (or differs from) the tangent-space version. This equivalence is central to claiming independence from fitted quantities or self-citations and must be shown algebraically.

minor comments (1)

[Abstract] The abstract repeatedly uses the phrase 'recent works' without citation; at least the most directly comparable prior algorithms should be referenced so that the claimed outperformance can be located.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. We address each major comment below and agree that additional explicit derivations, error bounds, and algebraic details are needed to fully substantiate the claims.

read point-by-point responses

Referee: [Abstract] Abstract: the assertion that 'even though the tangent space is not used ... the optimal step size problem can be computed in a closed form' is load-bearing for the entire contribution. The skeptic's concern is valid on the supplied evidence: directions sampled uniformly on the sphere are not guaranteed to lie in the tangent space at the current point, so the finite-difference quotients contain an unprojected radial component. No error bound or cancellation argument via the subsequent Gram-matrix reduction is supplied, which directly undermines the claimed subproblem equivalence and the acceleration guarantee.

Authors: We acknowledge that the abstract claim requires supporting analysis to be convincing. The quadratic structure of the objective ensures that radial components cancel in the surrogate construction under uniform spherical sampling, allowing the closed-form step size to hold without explicit projection. To address the concern rigorously, we will add an explicit error bound derivation and the cancellation argument (via the Gram-matrix reduction) in the revised manuscript. revision: yes
Referee: [Abstract] Abstract: the claim that 'together the achieved theory allows to construct an accelerated algorithm which shows state-of-the-art behavior' rests on the validity of the surrogate estimates. Without any derivation of the estimation error for the generalized operator norm or extremal generalized Rayleigh quotient under unconstrained sampling, or any comparison of the resulting convergence rate to the tangent-space-constrained baselines, the performance assertion cannot be evaluated.

Authors: We agree that the performance claim requires explicit supporting derivations. We will include the full derivation of the estimation error bounds for both the generalized operator norm and the extremal generalized Rayleigh quotient under unconstrained sampling. We will also add a direct comparison of the resulting convergence rates against tangent-space-constrained baselines to substantiate the state-of-the-art assertion. revision: yes
Referee: [Abstract] Abstract: the subproblem equivalence to 'sub-generalized Rayleigh quotient problems on specific Gram matrices' is presented as illuminating the relation to recent works, yet no explicit equations are given showing how the unconstrained surrogate reduces to (or differs from) the tangent-space version. This equivalence is central to claiming independence from fitted quantities or self-citations and must be shown algebraically.

Authors: We will add the missing explicit algebraic derivations in the revised manuscript. These will show step-by-step how the unconstrained surrogate reduces to (and differs from) the tangent-space version when expressed as sub-generalized Rayleigh quotient problems on the associated Gram matrices, thereby clarifying the connection to recent works. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained.

full rationale

The provided abstract and description present a proposal for unconstrained spherical sampling to estimate generalized operator norms and Rayleigh quotients, with explicit links to zeroth-order Riemannian surrogates for gradient/Hessian, closed-form step-size computation, and Gram-matrix subproblem reformulation. No equations or claims reduce a derived quantity to a fitted input by construction, no load-bearing self-citations are invoked to justify uniqueness or ansatzes, and the central acceleration claim is framed as building on (but distinct from) recent works without tautological redefinition. The sampling validity and performance assertions are independent empirical/theoretical steps that do not collapse into the paper's own inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no identifiable free parameters, axioms, or invented entities; full text would be required to audit the derivation.

pith-pipeline@v0.9.0 · 5443 in / 978 out tokens · 44954 ms · 2026-05-07T11:19:13.504129+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

51 extracted references · 42 canonical work pages

[1]

Distributed principal component analysis with limited communi- cation

F. Alimisis et al. “Distributed principal component analysis with limited communi- cation”.In:Advances in Neural Information Processing Systems (NeurIPS). Vol.34. Adv. Neural Inf. Process. Syst. Online: Curran Associates, Inc., 2021, pp. 2823– 2834.url:https://dl.acm.org/doi/10.5555/3540261.3540477

work page doi:10.5555/3540261.3540477 2021
[2]

Quasi-Newton Methods

W.-S. L. Andreas Antoniou. “Quasi-Newton Methods”. In:Practical Optimization: Algorithms and Engineering Applications. Boston, MA: Springer US, 2007, pp. 175– 202.isbn: 978-0-387-71107-2.doi:10.1007/978-0-387-71107-2_7.url:https: //doi.org/10.1007/978-0-387-71107-2%5C_7. 24

work page doi:10.1007/978-0-387-71107-2_7.url:https: 2007
[3]

Zeroth-order nonconvex stochastic opti- mization: Handling constraints, high dimensionality, and saddle points

K. Balasubramanian and S. Ghadimi. “Zeroth-order nonconvex stochastic opti- mization: Handling constraints, high dimensionality, and saddle points”. In:Found. Comput. Math.22.1 (2022), pp. 35–76.doi:10.1007/s10208-021-09499-8

work page doi:10.1007/s10208-021-09499-8 2022
[4]

Operator inference for non-intrusive model reduction of systems with non-polynomial nonlinear terms

P. Benner et al. “Operator inference for non-intrusive model reduction of systems with non-polynomial nonlinear terms”. In:Comput. Methods Appl. Mech. Eng.372 (2020), p. 113433.doi:10.1016/j.cma.2020.113433

work page doi:10.1016/j.cma.2020.113433 2020
[5]

Cambridge University Press

N. Boumal.An Introduction to Optimization on Smooth Manifolds. 1st ed. Cam- bridge,UK:CambridgeUniversityPress,2023,p.358.doi:10.1017/9781009166164

work page doi:10.1017/9781009166164 2023
[6]

Degenerate preconditioned proximal point algorithms

K. Bredies et al. “Degenerate preconditioned proximal point algorithms”. In:SIAM J. Optim.32.3 (2022), pp. 2376–2401.doi:10.1137/21M1448112

work page doi:10.1137/21m1448112 2022
[7]

Bresch et al.Computing adjoint mismatch of linear maps

J. Bresch et al.Computing adjoint mismatch of linear maps. arXiv:2503.21361. 2025.doi:10.48550/arXiv.2503.21361

work page doi:10.48550/arxiv.2503.21361 2025
[8]

Bresch et al.Matrix-free stochastic calculation of operator norms without using adjoints

J. Bresch et al.Matrix-free stochastic calculation of operator norms without using adjoints. arXiv:2410.08297. 2024.doi:10.48550/arXiv.2410.08297

work page doi:10.48550/arxiv.2410.08297 2024
[9]

arXiv:2512.05520

J.Breschetal.Stochastic zeroth-order method for computing the generalized Rayleigh quotient. arXiv:2512.05520. 2025.doi:10.48550/arXiv.2512.05520

work page doi:10.48550/arxiv.2512.05520 2025
[10]

Two numerical methods for optimiz- ing matrix stability

J. V. Burke, A. S. Lewis, and M. L. Overton. “Two numerical methods for optimiz- ing matrix stability”. In:Linear Algebra Appl.351–352 (2002), pp. 117–145.doi: 10.1016/S0024-3795(02)00260-4

work page doi:10.1016/s0024-3795(02)00260-4 2002
[11]

Journal of Mathematical Imaging and Vision40(1), 120–145 (2010)

A. Chambolle and T. Pock. “A first-order primal-dual algorithm for convex prob- lems with applications to imaging”. In:J. Math. Imaging Vis.40 (2011), pp. 120– 145.doi:10.1007/s10851-010-0251-1

work page doi:10.1007/s10851-010-0251-1 2011
[12]

Transpose-free formulations of Lanczos-type methods for nonsymmetric linear systems

T. F. Chan, L. de Pillis, and H. van der Vorst. “Transpose-free formulations of Lanczos-type methods for nonsymmetric linear systems”. In:Numer. Algorithms 17.1-2 (1998), pp. 51–66.doi:10.1023/A:1011637511962

work page doi:10.1023/a:1011637511962 1998
[13]

One-parameter semigroups for linear evolution equa- tions

K.-J. Engel and R. Nagel. “One-parameter semigroups for linear evolution equa- tions”. In:Semigroup Forum63 (2001), pp. 278–280.doi:10.1007/s002330010042

work page doi:10.1007/s002330010042 2001
[14]

Generalized stability theory. Part II: Nonau- tonomous operators

B. F. Farrell and P. J. Ioannou. “Generalized stability theory. Part II: Nonau- tonomous operators”. In:J. Atmos. Sci.53.14 (1996), pp. 2041–2053.doi:10 . 1175/1520-0469(1996)053<2041:GSTPIN>2.0.CO;2

1996
[15]

A comprehensive study of adjoint-based optimization of non-linear systems with application to Burgers’ equation

A. Fikl et al. “A comprehensive study of adjoint-based optimization of non-linear systems with application to Burgers’ equation”. In:Internation Fluid Dynamics Conference (AIAA). Adjoints and Error Estimation. Washington, DC, USA: AIAA, 2016, p. 3805.doi:10.2514/6.2016-3805

work page doi:10.2514/6.2016-3805 2016
[16]

Annals of Eugenics , volume =

R. A. Fisher. “The Use of Multiple Measurements in Taxonomic Problems”. In: Ann. Eugen.7 (1936), pp. 179–188.doi:10.1111/j.1469-1809.1936.tb02137.x

work page doi:10.1111/j.1469-1809.1936.tb02137.x 1936
[17]

Fulton and J

W. Fulton and J. Harris.Representation Theory: A First Course. 1st ed. Graduate Texts in Mathematics. New York, NY, USA: Springer, 1991, pp. xv + 551.doi: 10.1007/978-1-4612-0979-9. 25

work page doi:10.1007/978-1-4612-0979-9 1991
[18]

Quasi-Newton Methods for Unconstrained Optimiza- tion

P. E. Gill and W. Murray. “Quasi-Newton Methods for Unconstrained Optimiza- tion”. In:IMA J. Appl. Math.9.1 (1972), pp. 91–108.doi:10.1093/imamat/9.1. 91

work page doi:10.1093/imamat/9.1 1972
[19]

M. D. Gunzburger.Perspectives in Flow Control and Optimization. 1st ed. Ad- vances in Design and Control. Berlin, Heidelberg, Germany: SIAM, 2002, pp. xiv + 258.doi:10.1137/1.9780898718720

work page doi:10.1137/1.9780898718720 2002
[20]

Halko, P

N. Halko, P. G. Martinsson, and J. A. Tropp. “Finding structure with randomness: Probabilistic algorithms for constructing approximate matrix decompositions”. In: SIAM Review53.2 (2011), pp. 217–288.doi:10.1137/090771806

work page doi:10.1137/090771806 2011
[21]

Huber.Convergence of ray-and pixel-driven discretization frameworks in the strong operator topology

R. Huber.Convergence of ray-and pixel-driven discretization frameworks in the strong operator topology. arXiv:2503.03069. 2025.doi:10 . 48550 / arXiv . 2503 . 03069

work page arXiv 2025
[22]

On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables

L. Isserlis. “On a formula for the product-moment coefficient of any order of a normal frequency distribution in any number of variables”. In:Biometrika12.1/2 (1918), pp. 134–139.doi:10.2307/2331932

work page doi:10.2307/2331932 1918
[23]

Toward the optimal preconditioned eigensolver: Locally optimal block preconditioned conjugate gradient method

A. V. Knyazev. “Toward the optimal preconditioned eigensolver: Locally optimal block preconditioned conjugate gradient method”. In:SIAM J. Sci. Comput.23.2 (2001), pp. 517–541.doi:10.1137/S1064827500366124

work page doi:10.1137/s1064827500366124 2001
[24]

Очисленномрешенииуравнения,которымвтехническихвопросах определяются частоты малых колебаний материальных систем

A.N.Krylov.“Очисленномрешенииуравнения,которымвтехническихвопросах определяются частоты малых колебаний материальных систем”. In:Izv. Akad. Nauk. SSSR Otd Mat. Estest.4 (1931), pp. 491–539.url:http://mi.mathnet. ru//eng/im5215

1931
[25]

Lanczos, An iteration method for the solution of the eigenvalue problem of linear differential and integral operators , J

C. Lanczos. “An iteration method for the solution of the eigenvalue problem of linear differential and integral operators”. In:J. Res. Natl. Bur. Stand. B45 (1950), pp. 255–282.doi:10.6028/jres.045.026

work page doi:10.6028/jres.045.026 1950
[26]

Langrenez et al.The set of Kirkwood–Dirac positive states is almost always minimal

C. Langrenez et al.The set of Kirkwood–Dirac positive states is almost always minimal. arXiv:2405.17557. 2024.doi:10.48550/arXiv.2405.17557

work page doi:10.48550/arxiv.2405.17557 2024
[27]

J. M. Lee.Introduction to Smooth Manifolds. 2nd. Vol. 218. Graduate Texts in Mathematics. New York, NY, USA: Springer, 2012.isbn: 9781489994752.doi: 10.1007/978-1-4419-9982-5

work page doi:10.1007/978-1-4419-9982-5 2012
[28]

Stochastic zeroth-order Riemannian deriva- tive estimation and optimization

J. Li, K. Balasubramanian, and S. Ma. “Stochastic zeroth-order Riemannian deriva- tive estimation and optimization”. In:Math. Oper. Res.48.2 (2023), pp. 1183–1211. doi:10.1287/moor.2022.1302

work page doi:10.1287/moor.2022.1302 2023
[29]

Concise formulas for the area and volume of a hyperspherical cap

S. Li. “Concise formulas for the area and volume of a hyperspherical cap”. In:Asian J. Math. Stat.4.1 (2011), pp. 66–70.doi:10.3923/ajms.2011.66.70

work page doi:10.3923/ajms.2011.66.70 2011
[30]

On sketching matrix norms and the top singular vector

Y. Li, H. L. Nguyen, and D. P. Woodruff. “On sketching matrix norms and the top singular vector”. In:Proceedings of the 25th annual ACM-SIAM symposium on Discrete algorithms (SODA). SIAM. Portland, OR, USA, 2014, pp. 1562–1581. doi:10.1137/1.9781611973402.114. 26

work page doi:10.1137/1.9781611973402.114 2014
[31]

The randomized Kaczmarz method with mismatched adjoint

D. A. Lorenz, S. Rose, and F. Sch¨ opfer. “The randomized Kaczmarz method with mismatched adjoint”. In:BIT Numer. Math.58.4 (2018), pp. 1079–1098.doi:10. 1007/s10543-018-0717-x

2018
[32]

Chambolle-Pock’s primal-dual method with mis- matched adjoint

D. A. Lorenz and F. Schneppe. “Chambolle-Pock’s primal-dual method with mis- matched adjoint”. In:Appl. Math. Opt.87.2 (2023), p. 22.doi:10.1007/s00245- 022-09933-5

work page doi:10.1007/s00245- 2023
[33]

Adjoint equations in stability analysis

P. Luchini and A. Bottaro. “Adjoint equations in stability analysis”. In:Annu. Rev. Fluid Mech.46 (2014), pp. 493–517.doi:10.1146/annurev-fluid-010313- 141253

work page doi:10.1146/annurev-fluid-010313- 2014
[34]

Luchini and A

P. Luchini and A. Bottaro.An introduction to adjoint problems. arXiv:2404.17304. 2024.doi:10.48550/arXiv.2404.17304

work page doi:10.48550/arxiv.2404.17304 2024
[35]

Mathematical Notes107 (2015) https://doi.org/10.1134/S0001434620030189

B. S. Mityagin. “The zero set of a real analytic function”. In:Math. Notes107.3 (2020), pp. 529–530.doi:10.1134/S0001434620030189

work page doi:10.1134/s0001434620030189 2020
[36]

B. N. Parlett.The Symmetric Eigenvalue Problem. 1st ed. Classics in Applied Mathematics. Philadelphia, PA, USA: SIAM, 1998, pp. xxiv + 391.doi:10.1137/ 1.9781611971163

1998
[37]

Sliced optimal transport on the sphere

M. Quellmalz, R. Beinert, and G. Steidl. “Sliced optimal transport on the sphere”. In:Inverse Probl.39.10 (2023), p. 105005.doi:10.1088/1361-6420/acf156

work page doi:10.1088/1361-6420/acf156 2023
[38]

Parallelly sliced optimal transport on spheres and on the rotation group

M. Quellmalz, L. Buecher, and G. Steidl. “Parallelly sliced optimal transport on spheres and on the rotation group”. In:J. Math. Imaging Vis.66.6 (2024), pp. 951– 976.doi:10.1007/s10851-024-01206-w

work page doi:10.1007/s10851-024-01206-w 2024
[39]

The Utilization of Multiple Measurements in Problems of Biological Classification

C. R. Rao. “The Utilization of Multiple Measurements in Problems of Biological Classification”. In:J. R. Stat. Soc. Ser. B (Methodol.)10.2 (1948), pp. 159–193. doi:10.1111/j.2517-6161.1948.tb00008.x

work page doi:10.1111/j.2517-6161.1948.tb00008.x 1948
[40]

R´ enyi.Wahrscheinlichkeitsrechnung

A. R´ enyi.Wahrscheinlichkeitsrechnung. 2nd. Hochschulb¨ ucher f¨ ur Mathematik. East-Berlin, GDR: VEB Deutscher Verlag der Wissenschaft, 1966.url:https : //d-nb.info/201101106

work page arXiv 1966
[41]

Howconsensus-basedoptimizationcanbeinterpretedasastochastic relaxation of gradient descent

K.Riedletal.“Howconsensus-basedoptimizationcanbeinterpretedasastochastic relaxation of gradient descent”. In:International Conference on Machine Learning (ICML). Differentiable Almost Everything. Vienna, Austria: PMLR, 2024.url: https://openreview.net/forum?id=DkBBoKup22

2024
[42]

E. K. Ryu and W. Yin.Large-Scale Convex Optimization: Algorithms & Analy- ses via Monotone Operators. 1st ed. Cambridge, UK: Cambridge University Press, 2022.doi:10.1017/9781009160865

work page doi:10.1017/9781009160865 2022
[43]

Proximal Gradient Algorithm in the Presence of adjoint mis- match

M. Savanier et al. “Proximal Gradient Algorithm in the Presence of adjoint mis- match”. In:28th European Signal Processing Conference (EUSIPCO). Amsterdam, Netherlands: IEEE, 2021, pp. 2140–2144.doi:10 . 23919 / Eusipco47968 . 2020 . 9287430. 27

2021
[44]

Об устойчивости обратных задач

A. N. Tikhonov. “Об устойчивости обратных задач”. In:Dokl. Akad. Nauk SSSR 39 (4 1943), pp. 195–198.url:http://a-server.math.nsc.ru/IPP/BASE_WORK/ tihon_en.html

1943
[45]

A. N. Tikhonov and V. Y. Arsenin.Solutions of Ill-Posed Problems. Washington, DC, USA: V. H. Winston & Sons, 1977, pp. xiii+258.doi:10.2307/2006360

work page doi:10.2307/2006360 1977
[46]

Spectra and Pseudospectra

L. N. Trefethen. “Spectra and Pseudospectra”. In:The Graduate Student’s Guide to Numerical Analysis ’98. Springer Series in Computational Mathematics. Berlin, Heidelberg, Germany: Springer, 1999, pp. 217–250.doi:10 . 1007 / 978 - 3 - 662 - 03972-4_6

1999
[47]

Transient Effects and Nonnormal Dynamics

L. N. Trefethen and M. Embree. “Transient Effects and Nonnormal Dynamics”. In: Spectra and Pseudospectra: The Behavior of Nonnormal Matrices and Operators. Princeton, NJ, USA: Princeton University Press, 2005, pp. 133–192.doi:doi : 10.1515/9780691213101-006

work page doi:10.1515/9780691213101-006 2005
[48]

H. L. Van Trees.Optimum Array Processing: Part IV of Detection, Estimation, and Modulation Theory. New York, NY, USA: John Wiley & Sons, 2002.doi: 10.1002/0471221104

work page doi:10.1002/0471221104 2002
[49]

2018 , publisher =

R. Vershynin.High-Dimensional Probability: An Introduction with Applications in Data Science. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge, UK: Cambridge University Press, 2018.isbn: 978-1-108-41519-4.doi: 10.1017/9781108231596

work page doi:10.1017/9781108231596 2018
[50]

The evaluation of the collision matrix

G. C. Wick. “The evaluation of the collision matrix”. In:Phys. Rev.80 (2 1950), pp. 268–272.doi:10.1103/PhysRev.80.268. A. Algorithms Algorithm 4First-order Riemannian gradient ascent method GivenA∈R m×d andB∈R ℓ×d such thatrank(B) =d. Initializewithv 0 ∈S d−1. fork= 1,2,3, ...do Calculatex k =∇f(v k),selectstep sizeτ k >0,updatev k+1 =R vk(τkxk) end for ...

work page doi:10.1103/physrev.80.268 1950
[51]

Furthermore, if∥v−˜v∥< ε, it suffices for anyx∈S d−1 to takeτ= 0 such thatσ d−1(Dv) = 1. Now, since for anyy∈S d−1 andw∈R d \ {0}holds that w ∥w∥ −y ≤ w ∥w∥ −w +∥w−y∥= 1 ∥w∥ −1 ∥w∥+∥w−y∥ = 1− ∥w∥ +∥w−y∥= ∥y∥ − ∥w∥ +∥w−y∥ ≤2∥w−y∥,(30) it suffices to take ˜Dv :={x∈S d−1 | ∃τ∈R, v+τ x∈B ε/2(˜v)∪Bε/2(−˜v)} ⊂ Dv. As the latter can be trivially rewritten ˜Dv ={...