Pith. sign in

REVIEW 1 major objections 1 cited by

Distributed Riemannian Optimization in Geodesically Non-convex Environments

T0 review · 1 major / 0 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proves that the Riemannian diffusion adaptation algorithm, in which agents take stochastic gradient steps on a manifold and combine neighbors' estimates along geodesics, converges to a first-order stationary point for geodesicall

desk verdict First non-asymptotic intrinsic analysis of distributed non-convex Riemannian diffusion, but the central descent lemma currently drops the gradient-noise term. read the letter →

arxiv 2512.04915 v2 pith:655WIFPW submitted 2025-12-04 eess.SP

Xiuheng Wang , Ricardo Borsoi , Cédric Richard , Ali H. Sayed This is my paper
classification eess.SP MSC 68W1590C5690C26
keywords Riemannianoptimizationdistributeddiffusionadaptationgeodesicnon-convexityFréchetvariancePolyak-ŁojasiewiczGrassmannmanifoldstochasticgradient
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 studies a decentralized optimization problem where agents on a network minimize an average of cost functions defined over a Riemannian manifold (a curved space). It analyzes the Riemannian diffusion adaptation algorithm—each agent performs a stochastic gradient step, then combines neighbors' estimates using the manifold's geodesic geometry. The central result is that this algorithm provably reaches approximate network agreement, with the Fréchet variance of the iterates bounded by O(μ²), and converges to a first-order stationary point for general geodesically smooth but non-convex costs. Under an additional Riemannian Polyak-Łojasiewicz condition, the algorithm converges linearly up to a steady-state error. This provides the first non-asymptotic guarantees for distributed Riemannian optimization with non-convex costs, extending prior results that required geodesic convexity.

What carries the argument

The carrying objects are the Fréchet variance VF(w), a measure of how scattered the agents' iterates are around their Riemannian center of mass, and the consensus bias P(φ), the weighted sum of squared geodesic distances among the intermediate updates. Lemma 5 shows the combination step reduces Fréchet variance because the coefficient ζ₁α² − ζ₂α is negative; Lemma 6–7 and Theorem 1 turn this into a bound of order O(μ²) after O(μ⁻¹) iterations. The crucial new step is Lemma 9, a Taylor expansion of exp_{w_{k,t}}⁻¹(exp_{w_{ℓ,t}}(−μ d∇J_ℓ(w_{ℓ,t}))) that writes each neighbor's updated estimate in a common tangent space, with a remainder O(μ²). This expansion is what allows the consensus bias to

What would settle it

Run the two-step update on a unit sphere with a geodesically smooth non-convex cost, choosing the combination step α so that the geodesic average of two nearly antipodal points falls outside the ball of radius π/2. If the Fréchet variance does not decrease by the predicted factor (1−ε), then Lemma 6, and consequently Theorems 1–3, collapse.

Watch

Extended reading notes

Core claim

The paper's central claim is that, for a strongly connected network of agents minimizing an average of geodesically L-smooth functions over a compact convexity submanifold, the two-step diffusion adaptation update drives the system to approximate agreement and, with constant step sizes, to a first-order stationary point. The key technical device is a descent inequality (Lemma 8) that tracks the cost while paying a penalty proportional to the consensus bias P(φ_{t+1}). The consensus bias is then controlled via a first-order Taylor expansion of the composite exponential map (Lemma 9), which linearizes the stochastic-gradient update in the tangent space of the combination point. Combined with c

Load-bearing premise

The algorithm's iterates are assumed to remain, for every iteration, inside a compact patch of the manifold whose diameter is below the injectivity radius and the curvature bound; the proof does not show the algorithm keeps them there, but every lemma depends on that.

Editorial extensions

If this is right

  • A decentralized network can optimize a sum of geodesically non-convex costs with constant step sizes and still reach a first-order stationary point, with steady-state error O(α²)+O(α²μ²).
  • Network topology matters predictably: the agreement level improves with the spectral gap (1−λ) of the mixing matrix, so denser networks converge to consensus faster.
  • When the global cost satisfies the Riemannian PL condition, linear convergence is recovered without requiring geodesic convexity, at rate (1−μK/(4τ)).
  • The explicit appearance of curvature constants ζ₁, ζ₂, Cκ in the bounds lets users anticipate how manifold geometry slows or accelerates distributed optimization.

Reading between the lines

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

  • The most fragile part is the trajectory assumption in Assumption 2(a): the iterates are assumed to remain forever inside a compact region of bounded diameter. A testable prediction is that on a positively curved manifold such as a sphere, using a combination step α large enough to push the combined point beyond the radius π/(2√κ_max) should weaken or violate the contraction in Lemma 6.
  • The Taylor-expansion technique for the consensus bias (Lemma 9) is likely reusable for variance-reduced or accelerated variants, and for deriving tracking bounds in time-varying environments—this is my inference, not a claim in the paper.
  • Because the bounds treat α and μ independently, a natural follow-up would be to optimize a joint tuning rule that balances the O(α²) steady-state error against the O(μ⁻¹) convergence time.
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

1 major / 0 minor

Summary. The paper analyzes the Riemannian diffusion adaptation algorithm, which alternates a stochastic Riemannian gradient step with a consensus-type combination on the tangent space, for minimizing a sum of geodesically smooth, possibly geodesically non-convex costs over a network. Under assumptions on the graph, the manifold, and the stochastic gradient noise, it claims three results: approximate network agreement in Fréchet variance after O(µ^{-1}) iterations (Theorem 1), convergence of the averaged squared gradient to a first-order stationary point up to a steady-state error O(α²)+O(α²µ²) for general g-non-convex costs (Theorem 2), and linear convergence under a Riemannian PL condition up to a similar steady-state error (Theorem 3). The paper also applies the method to decentralized robust PCA on the Grassmann manifold and presents simulations on synthetic and MNIST data.

Significance. If the results are correct, this would be the first non-asymptotic analysis of an intrinsic distributed Riemannian algorithm for geodesically non-convex costs, extending the authors' earlier g-convex analysis. The paper is clearly written, the assumptions are standard for the Riemannian optimization literature, and the application to distributed robust PCA is relevant. A notable strength is that the proof strategy decouples the consensus analysis from the cost descent analysis, which is a useful structural contribution. However, the central descent lemma contains an algebraic error in Appendix E, and the stated steady-state bounds in Theorems 2 and 3 do not follow from the proof as written. The qualitative approach is promising and the error appears repairable, but the headline quantitative claims are currently unsupported.

major comments (1)
  1. [Appendix E, Eq. (55)] The equality after Eq. (55) is false. From (53), taking expectation conditional on F_t gives E J_k(ϕ_{k,t+1}) ≤ E J_k(w_{k,t}) - µ E||∇J_k(w_{k,t})||² + (Lµ²/2) E|| d∇J_k(w_{k,t})||², because E[ d∇J_k | F_t] = ∇J_k. Writing d∇J_k = ∇J_k + s_k with E[s_k|F_t]=0 and E||s_k||² ≤ σ², the correct relation is E J_k(ϕ) ≤ E J_k(w) - ϵ E||∇J_k||² + (µ-ϵ) E||s_k||², where ϵ = µ(1-Lµ/2). The proof instead replaces the right-hand side with -ϵ E|| d∇J_k||² = -ϵ E||∇J_k||² - ϵE||s_k||², dropping a nonnegative term of magnitude µE||s_k||². This error propagates through Eq. (59) and (60) into Lemma 8's descent inequality (24). Consequently, Lemma 8 lacks an additive gradient-noise term that does not vanish as α→0. When Lemma 8 is used in Theorem 2 (Eq. (28)) and Theorem 3 (Eq. (30)), the steady-state bounds gain a contribution of order O(σ²/K) in the squared-gradient bound and a corresponding noise floo

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: non-convex convergence bounds are derived, not assumed; self-citations are auxiliary.

full rationale

I walked the derivation chain from Assumptions 1–4/5 through Theorems 1–3. The algorithm (2) is imported from prior work, but it is not treated as a theorem: the network-agreement, first-order-stationarity, and PL-linear-convergence conclusions are proved from the stated assumptions via Lemmas 5–11 with proofs in Appendices A–J. The cited prior results are auxiliary geometric or technical tools (Lemma 2 from [36], Lemma 3 from [38], Lemma 4 from [16], Lemma 6 'partially inspired by' [17,18], Lemma 9 analogous to Lemma 4 of [8]); none of them inject the non-convex convergence conclusion or the target bounds. No fitted constant appears in the main results: the O(µ^2), O(α^2), and O(α^2µ^2) terms follow from contraction/descent inequalities, and the simulation step sizes are direct inputs, not parameters calibrated to match a claimed prediction. The apparent Lemma 8 issue about replacing the stochastic-gradient norm with the true-gradient norm is an algebraic/correctness concern; even if an O(µσ^2) noise term were needed, that would be a repairable inequality error, not a case where a derived quantity is identical to its input by construction. Assumption 2(a) is a trajectory postulate, but postulating a favorable geometric region is an assumption, not circular reasoning. I therefore find no load-bearing circular step.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The central claim rests on five explicit modeling assumptions and several imported geometric lemmas. No new entities (particles, forces, dimensions) are introduced. The main burden is Assumption 2's unverified invariant-trajectory requirement, plus the step sizes that are hand-chosen in practice.

free parameters (2)
  • adaptation step size µ = µ=0.12/0.13 (synthetic), µ=0.006 (MNIST)
    Hand-chosen constant step sizes in experiments; all convergence rates and steady-state bounds depend on µ. Not fitted to data but chosen to make simulations work.
  • combination step size α = α=0.4 (synthetic), α=0.005/0.001 (MNIST)
    Must lie in (0, ζ2/ζ1); hand-chosen in experiments. Controls the Fréchet-variance contraction and appears in the O(α²) steady-state terms.
assumptions (7)
  • domain assumption Assumption 1: graph strongly connected, C symmetric and doubly stochastic, mixing rate λ<1
    Needed for spectral-gap arguments in Lemmas 1 and 6. Standard for diffusion adaptation.
  • domain assumption Assumption 2: iterates stay in compact convexity submanifold B with bounded curvatures, diameter B < min(inj(M), π/(2√κmax)), optimum in B
    Postulated invariant trajectory condition. If iterates leave B, Lemmas 2–4 and all main theorems fail. This is the weakest load-bearing premise.
  • domain assumption Assumption 3: each J_k is geodesically L-smooth with L-Lipschitz gradient
    Used in descent inequalities (Lemmas 8 and 11) and gradient bounds (Lemma 4).
  • domain assumption Assumption 4: stochastic gradient noise is unbiased and has bounded fourth moment σ⁴
    Used throughout stochastic bounds; implies bounded second moment.
  • domain assumption Assumption 5: global J satisfies Riemannian PL condition with constant τ
    Only used in Lemma 11 and Theorem 3; implies every stationary point is global and yields linear contraction.
  • standard math Lemmas 2–4 from prior literature: trigonometric distance bounds for curvature-bounded manifolds, Lipschitz exponential/inverse-exponential maps, bounded gradients
    Imported geometric results from [36], [38], [16]; accepted without proof and used extensively in Appendices A–G.
  • standard math Taylor's theorem, chain rule, and smoothness of exponential maps on compact Riemannian manifolds
    Used in Lemmas 9 and 10 to linearize the composite exponential map and vector transport.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Distributed Riemannian Optimization in Geodesically Non-convex Environments." pith.science (2026). https://pith.science/paper/655WIFPW

@misc{pith2026251204915,
  author       = {Pith},
  title        = {Pith review of: Distributed Riemannian Optimization in Geodesically Non-convex Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/655WIFPW}},
  note         = {Machine review of arXiv:2512.04915}
}
read the original abstract

This paper studies the problem of distributed Riemannian optimization over a network of agents whose cost functions are geodesically smooth but possibly geodesically non-convex. Extending a well-known distributed optimization strategy called diffusion adaptation to Riemannian manifolds, we show that the resulting algorithm, the Riemannian diffusion adaptation, provably exhibits several desirable behaviors when minimizing a sum of geodesically smooth non-convex functions over manifolds of bounded curvature. More specifically, we establish that the algorithm can approximately achieve network agreement in the sense that Fr\'echet variance of the iterates among the agents is small. Moreover, the algorithm is guaranteed to converge to a first-order stationary point for general geodesically non-convex cost functions. When the global cost function additionally satisfies the local Riemannian Polyak-Lojasiewicz (PL) condition, we also show that it converges linearly under a constant step size up to a steady-state error. Finally, we apply this algorithm to a decentralized robust principal component analysis (PCA) problem formulated on the Grassmann manifold and low-rank matrix completion problems and illustrate its convergence and performance through numerical simulations.

Figures

Figures reproduced from arXiv: 2512.04915 by the authors.

Figure 1
Figure 1. The randomly generated graph structures used in the experiments. (a) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Illustration of MSD performance of the algorithms for distributed [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 1
Figure 1. The weights in matrix C with K = 20 agents were randomly generated by the Metropolis rule [29] and the uniform rule. For simulation on synthetic data, the MSD results are averaged over 100 independent Monte Carlo experiments. We compare our algorithm against the Riemannian non-cooperative algorithm, which independently applies R-SGD on each agent using its local data xk,t. We also provide comparisons with an extrins… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Decentralized Online Riemannian Optimization for Strongly Geodesically Convex Functions

    math.OC 2026-07 conditional novelty 6.0 of 10

    Decentralized online Riemannian gradient descent with a decaying step size achieves O(log T) static regret for strongly geodesically convex losses on manifolds with bounded sectional curvature, under full and two-poin...

Reference graph

Works this paper leans on

51 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Linear dimensionality reduction: Survey, insights, and generalizations,

    J. P. Cunningham and Z. Ghahramani, “Linear dimensionality reduction: Survey, insights, and generalizations,”The Journal of Machine Learning Research, vol. 16, no. 1, pp. 2859–2900, 2015

    2015

  2. [2]

    Matrix manifold optimization for gaussian mixtures,

    R. Hosseini and S. Sra, “Matrix manifold optimization for gaussian mixtures,” inProceedings of the 29th International Conference on Neural Information Processing Systems, vol. 1, 2015, pp. 910–918

    2015

  3. [3]

    Riemannian optimization for non-centered mixture of scaled Gaussian distributions,

    A. Collas, A. Breloy, C. Ren, G. Ginolhac, and J.-P. Ovarlez, “Riemannian optimization for non-centered mixture of scaled Gaussian distributions,” IEEE Transactions on Signal Processing, vol. 71, pp. 2475–2490, 2023

    2023

  4. [4]

    RTRMC: a riemannian trust-region method for low-rank matrix completion,

    N. Boumal and P.-A. Absil, “RTRMC: a riemannian trust-region method for low-rank matrix completion,” inAdvances in Neural Information Processing Systems, vol. 24, 2011, pp. 406–414

    2011

  5. [5]

    Low-rank matrix completion by Riemannian opti- mization,

    B. Vandereycken, “Low-rank matrix completion by Riemannian opti- mization,”SIAM Journal on Optimization, vol. 23, no. 2, pp. 1214–1236, 2013

    2013

  6. [6]

    Stochastic gradient descent on Riemannian manifolds,

    S. Bonnabel, “Stochastic gradient descent on Riemannian manifolds,” IEEE Transactions on Automatic Control, vol. 58, no. 9, pp. 2217–2229, 2013

    2013

  7. [7]

    First-order methods for geodesically convex optimization,

    H. Zhang and S. Sra, “First-order methods for geodesically convex optimization,” inConference on Learning Theory, 2016, pp. 1617–1638

    2016

  8. [8]

    Averaging stochastic gradient descent on Riemannian manifolds,

    N. Tripuraneni, N. Flammarion, F. Bach, and M. I. Jordan, “Averaging stochastic gradient descent on Riemannian manifolds,” inConference on Learning Theory. PMLR, 2018, pp. 650–687

    2018

Show all 51 references
  1. [9]

    No-regret online learning over Riemannian manifolds,

    X. Wang, Z. Tu, Y . Hong, Y . Wu, and G. Shi, “No-regret online learning over Riemannian manifolds,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 28 323–28 335

    2021

  2. [10]

    Riemannian stochastic optimization methods avoid strict saddle points,

    Y .-P. Hsieh, M. R. Karimi Jaghargh, A. Krause, and P. Mertikopoulos, “Riemannian stochastic optimization methods avoid strict saddle points,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 29 580–29 601

    2023

  3. [11]

    Non-parametric online change point detection on Riemannian manifolds,

    X. Wang, R. A. Borsoi, and C. Richard, “Non-parametric online change point detection on Riemannian manifolds,” inInternational Conference on Machine Learning. PMLR, 2024, pp. 50 143–50 162

    2024

  4. [12]

    Distributed optimization on Riemannian manifolds for multi-agent networks,

    S. M. Shah, “Distributed optimization on Riemannian manifolds for multi-agent networks,”arXiv:1711.11196, 2017

    2017 arXiv

  5. [13]

    Decentralized Riemannian gradient descent on the Stiefel manifold,

    S. Chen, A. Garcia, M. Hong, and S. Shahrampour, “Decentralized Riemannian gradient descent on the Stiefel manifold,” inInternational Conference on Machine Learning. PMLR, 2021, pp. 1594–1605

    2021

  6. [14]

    Incremental aggregated Riemannian gradient method for distributed PCA,

    X. Wang, Y . Jiao, H.-T. Wai, and Y . Gu, “Incremental aggregated Riemannian gradient method for distributed PCA,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2023, pp. 7492–7510

    2023

  7. [15]

    Decentralized projected Riemannian gradient method for smooth optimization on compact submanifolds embedded in the euclidean space,

    K. Deng and J. Hu, “Decentralized projected Riemannian gradient method for smooth optimization on compact submanifolds embedded in the euclidean space,”Numerische Mathematik, pp. 1–38, 2025

    2025

  8. [16]

    Riemannian diffusion adaptation for distributed optimization on manifolds,

    X. Wang, R. Borsoi, C. Richard, and A. H. Sayed, “Riemannian diffusion adaptation for distributed optimization on manifolds,” inInternational Conference on Machine Learning (ICML). PMLR, 2025

    2025

  9. [17]

    Decentralized online Riemannian optimization with dynamic environments,

    H. Chen and Q. Sun, “Decentralized online Riemannian optimization with dynamic environments,”arXiv:2410.05128, 2024

    2024 arXiv

  10. [18]

    Decentralized online Riemannian optimization beyond Hadamard manifolds,

    E. Sahinoglu and S. Shahrampour, “Decentralized online Riemannian optimization beyond Hadamard manifolds,”arXiv:2509.07779, 2025

    2025 arXiv

  11. [19]

    Consensus optimization on manifolds,

    A. Sarlette and R. Sepulchre, “Consensus optimization on manifolds,” SIAM journal on Control and Optimization, vol. 48, no. 1, pp. 56–76, 2009

    2009

  12. [20]

    Riemannian consensus for manifolds with bounded curvature,

    R. Tron, B. Afsari, and R. Vidal, “Riemannian consensus for manifolds with bounded curvature,”IEEE Transactions on Automatic Control, vol. 58, no. 4, pp. 921–934, 2012

    2012

  13. [21]

    Distributed consensus on manifolds using the Riemannian center of mass,

    S. Kraisler, S. Talebi, and M. Mesbahi, “Distributed consensus on manifolds using the Riemannian center of mass,” inIEEE Conference on Control Technology and Applications (CCTA), 2023, pp. 130–135

    2023

  14. [22]

    Diffusion adaptation strategies for distributed optimization and learning over networks,

    J. Chen and A. H. Sayed, “Diffusion adaptation strategies for distributed optimization and learning over networks,”IEEE Transactions on Signal Processing, vol. 60, no. 8, pp. 4289–4305, 2012

    2012

  15. [23]

    Diffusion strategies for adaptation and learning over networks: an examination of distributed strategies and network behavior,

    A. H. Sayed, S.-Y . Tu, J. Chen, X. Zhao, and Z. J. Towfic, “Diffusion strategies for adaptation and learning over networks: an examination of distributed strategies and network behavior,”IEEE Signal Processing Magazine, vol. 30, no. 3, pp. 155–171, 2013

    2013

  16. [24]

    Riemannian diffusion adaptation over graphs with application to online distributed PCA,

    X. Wang, R. A. Borsoi, and C. Richard, “Riemannian diffusion adaptation over graphs with application to online distributed PCA,” inIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 9736–9740

    2024

  17. [25]

    J. M. Lee,Riemannian Manifolds: An Introduction to Curvature. Springer Science & Business Media, 2006, vol. 176

    2006

  18. [26]

    M. P. Do Carmo,Differential Geometry of Curves and Surfaces. Courier Dover Publications, 2016

    2016

  19. [27]

    Absil, R

    P.-A. Absil, R. Mahony, and R. Sepulchre,Optimization Algorithms on Matrix Manifolds. Princeton University Press, 2009

    2009

  20. [28]

    Boumal,An Introduction to Optimization on Smooth Manifolds

    N. Boumal,An Introduction to Optimization on Smooth Manifolds. Cambridge University Press, 2023

    2023

  21. [29]

    Adaptation, learning, and optimization over networks,

    A. H. Sayed, “Adaptation, learning, and optimization over networks,” Foundations and Trends® in Machine Learning, vol. 7, no. 4-5, pp. 311–801, 2014

    2014

  22. [30]

    On the learning behavior of adaptive networks—Part I: Transient analysis,

    J. Chen and A. H. Sayed, “On the learning behavior of adaptive networks—Part I: Transient analysis,”IEEE Transactions on Information Theory, vol. 61, no. 6, pp. 3487–3517, 2015

    2015

  23. [31]

    Dual averaging for distributed optimization: Convergence analysis and network scaling,

    J. C. Duchi, A. Agarwal, and M. J. Wainwright, “Dual averaging for distributed optimization: Convergence analysis and network scaling,” IEEE Transactions on Automatic control, vol. 57, no. 3, pp. 592–606, 2011

    2011

  24. [32]

    On the convergence of decentralized gradient descent,

    K. Yuan, Q. Ling, and W. Yin, “On the convergence of decentralized gradient descent,”SIAM Journal on Optimization, vol. 26, no. 3, pp. 1835–1854, 2016

    2016

  25. [33]

    A unified and refined convergence analysis for non-convex decentralized learning,

    S. A. Alghunaim and K. Yuan, “A unified and refined convergence analysis for non-convex decentralized learning,”IEEE Transactions on Signal Processing, vol. 70, pp. 3264–3279, 2022

    2022

  26. [34]

    Riemannian ℓp center of mass: existence, uniqueness, and convexity,

    B. Afsari, “Riemannian ℓp center of mass: existence, uniqueness, and convexity,”Proceedings of the American Mathematical Society, vol. 139, no. 2, pp. 655–673, 2011

    2011

  27. [35]

    Riemannian SVRG: Fast stochastic op- timization on Riemannian manifolds,

    H. Zhang, S. J. Reddi, and S. Sra, “Riemannian SVRG: Fast stochastic op- timization on Riemannian manifolds,” inAdvances in Neural Information Processing Systems, 2016, pp. 4592–4600

    2016

  28. [36]

    First-order algo- rithms for min-max optimization in geodesic metric spaces,

    M. Jordan, T. Lin, and E.-V . Vlatakis-Gkaragkounis, “First-order algo- rithms for min-max optimization in geodesic metric spaces,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 6557–6574

    2022

  29. [37]

    A continuous-time perspective for modeling acceleration in Riemannian optimization,

    F. Alimisis, A. Orvieto, G. Bécigneul, and A. Lucchi, “A continuous-time perspective for modeling acceleration in Riemannian optimization,” in International Conference on Artificial Intelligence and Statistics. PMLR, 2020, pp. 1297–1307

    2020

  30. [38]

    Escaping from saddle points on Riemannian manifolds,

    Y . Sun, N. Flammarion, and M. Fazel, “Escaping from saddle points on Riemannian manifolds,” inAdvances in Neural Information Processing Systems, vol. 32, 2019, pp. 7276–7286

    2019

  31. [39]

    Distributed learning in non-convex environments—Part I: Agreement at a linear rate,

    S. Vlaski and A. H. Sayed, “Distributed learning in non-convex environments—Part I: Agreement at a linear rate,”IEEE Transactions on Signal Processing, vol. 69, pp. 1242–1256, 2021

    2021

  32. [40]

    Distributed learning in non-convex environments—Part II: Polyno- mial escape from saddle-points,

    ——, “Distributed learning in non-convex environments—Part II: Polyno- mial escape from saddle-points,”IEEE Transactions on Signal Processing, vol. 69, pp. 1257–1270, 2021

    2021

  33. [41]

    Riemannian adaptive stochastic gradient algorithms on matrix manifolds,

    H. Kasai, P. Jawanpuria, and B. Mishra, “Riemannian adaptive stochastic gradient algorithms on matrix manifolds,” inInternational Conference on Machine Learning, 2019, pp. 3262–3271

    2019

  34. [42]

    On the convergence of gradient descent for finding the Riemannian center of mass,

    B. Afsari, R. Tron, and R. Vidal, “On the convergence of gradient descent for finding the Riemannian center of mass,”SIAM Journal on Control and Optimization, vol. 51, no. 3, pp. 2230–2260, 2013

    2013

  35. [43]

    An improved convergence analysis for decentralized online stochastic non-convex optimization,

    R. Xin, U. A. Khan, and S. Kar, “An improved convergence analysis for decentralized online stochastic non-convex optimization,”IEEE Transactions on Signal Processing, vol. 69, pp. 1842–1858, 2021

    2021

  36. [44]

    An overview of robust subspace recovery,

    G. Lerman and T. Maunu, “An overview of robust subspace recovery,” Proceedings of the IEEE, vol. 106, no. 8, pp. 1380–1410, 2018

    2018

  37. [45]

    Distributed robust subspace recovery,

    V . Huroyan and G. Lerman, “Distributed robust subspace recovery,”SIAM Journal on Scientific Computing, vol. 40, no. 5, pp. A3067–A3090, 2018

    2018

  38. [46]

    Trust-region methods on Riemannian manifolds,

    P.-A. Absil, C. G. Baker, and K. A. Gallivan, “Trust-region methods on Riemannian manifolds,”Foundations of Computational Mathematics, vol. 7, no. 3, pp. 303–330, 2007

    2007

  39. [47]

    Pymanopt: A python toolbox for optimization on manifolds using automatic differentiation,

    J. Townsend, N. Koep, and S. Weichwald, “Pymanopt: A python toolbox for optimization on manifolds using automatic differentiation,”Journal of Machine Learning Research, vol. 17, no. 137, pp. 1–5, 2016

    2016

  40. [48]

    The MNIST database of handwritten digits,

    Y . LeCun, “The MNIST database of handwritten digits,”http://yann. lecun. com/exdb/mnist/, 1998

    1998

  41. [49]

    Geometric wave equations,

    S. Waldmann, “Geometric wave equations,”arXiv:1208.4706, 2012

    2012 arXiv

  42. [50]

    Riemannian accelerated gradient methods via extrapolation,

    A. Han, B. Mishra, P. Jawanpuria, and J. Gao, “Riemannian accelerated gradient methods via extrapolation,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2023, pp. 1554–1585

    2023

  43. [51]

    The geometry of algorithms with orthogonality constraints,

    A. Edelman, T. A. Arias, and S. T. Smith, “The geometry of algorithms with orthogonality constraints,”SIAM journal on Matrix Analysis and Applications, vol. 20, no. 2, pp. 303–353, 1998

    1998

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.