Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Decentralized Online Riemannian Optimization for Strongly Geodesically Convex Functions

T0 review · 3 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper proves that decentralized online Riemannian optimization with strongly geodesically convex losses attains O(log T) static regret on manifolds of bounded sectional curvature, including positively curved ones, under both full-gradie

desk verdict First O(log T) decentralized Riemannian online regret result, with a load-bearing consensus-interference inequality imported from an overlapping preprint; deserves review with mandatory self-containedness fixes. read the letter →

arxiv 2607.20316 v1 pith:BV46NDWJ submitted 2026-07-22 math.OC cs.LGcs.MA

Zhanyuan Cai , Emre Sahinoglu , Shahin Shahrampour This is my paper
classification math.OCcs.LGcs.MA MSC 90C2590C4868Q3253B20
keywords decentralizedoptimizationonlinelearningRiemannianmanifoldsstronggeodesicconvexitylogarithmicregretconsensustwo-pointbanditfeedbackstatic
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

This paper asks whether the logarithmic regret improvement that strong convexity brings to Euclidean and single-agent Riemannian online optimization can survive decentralization, where agents on a curved space must align their decisions through neighbor-to-neighbor communication. The authors answer yes: for μ-strongly geodesically convex losses on manifolds with bounded sectional curvature, including positively curved spheres, a two-step protocol of projected Riemannian gradient descent followed by a curvature-aware consensus step attains O(log T) static regret, the same rate as the centralized and Euclidean minimax-optimal case up to curvature constants and the network spectral gap. The key technical move is a consensus-error analysis that tolerates the decaying step sizes needed for the logarithmic rate, instead of the fixed step sizes all prior decentralized Riemannian analyses assume. The same O(log T) rate is proven when agents see only two function evaluations per round (bandit feedback), using a smoothing radius δ = T^{-2} and a new strong-sub-g-convexity inequality. If correct, this closes the last gap between centralized and decentralized online Riemannian optimization for strongly convex objectives.

What carries the argument

The engine is the two-step update: each agent first moves along its Riemannian gradient and projects back onto the feasible set, then takes a small curvature-aware consensus step toward the tangent-space approximation of its neighbors' Fréchet mean. The analysis rests on a network-error lemma that bounds the average distance from the agents to their Fréchet mean by L times a time-weighted sum of past step sizes, ρ^{(t-k)/2} η_k, where ρ < 1 is a contraction factor depending on curvature, the consensus step size s, and the network spectral gap 1-σ_2(W). Because this convolution stays O(log T) under η_t = (μt + L/D)^{-1}, the network error does not destroy the logarithmic regret. In the bandit

What would settle it

Compute the constants C_1, C_2, C_3 on a specific positively curved manifold (for instance S^7 with the paper's radius-7π/32 ball, where D = 7π/16), plug the chosen s = C_2/(2C_1) into ρ = 1 - C_2^2(1-σ_2(W))/(2C_1(1+C_3 D^2)^2), and check whether ρ < 1 for all allowed networks; also numerically evaluate the per-round interference sum (13) on a small ring graph — if ρ ≥ 1 or the interference sum is positive, the consensus lemma and the log bound fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that, on Riemannian manifolds with bounded sectional curvature and a diameter constraint ruling out antipodal pairs, decentralized online gradient descent on strongly g-convex losses achieves static regret that grows logarithmically in the time horizon T, giving O(log T) under both full-gradient and two-point bandit feedback. Prior work on decentralized Riemannian online optimization only covered g-convex losses with fixed step sizes, yielding O(√T) regret and leaving strong convexity unexploited. The paper proves the logarithmic rate by developing a network-error analysis for time-varying (decaying) step sizes, in which the consensus error accumulates as a geome

Load-bearing premise

The logarithmic bound relies on imported geometric lemmas — the contraction factor ρ and curvature constants C_1-C_3 from a companion paper, and the projection-error bound with constant C_7 from earlier work — so if any of those carries a hidden curvature condition or wrong constant, the O(log T) guarantee collapses.

Editorial extensions

If this is right

  • The first O(log T) static regret bound for decentralized online Riemannian optimization: strongly g-convex losses now match the centralized rate up to curvature-dependent constants and the factor (1 - σ_2(W))^{-1}.
  • The result covers positively curved manifolds, such as spheres, under the diameter condition D < (1/2) min{r_inj, π/√K_max}, extending beyond the Hadamard-manifold regime of earlier decentralized methods.
  • The O(log T) rate survives two-point bandit feedback with smoothing radius δ = T^{-2}; gradient-free agents pay only constant terms beyond the log factor, not the O(√T) rate of the g-convex baseline.
  • The network-error lemma applies to general time-varying step-size schedules, not just the specific decaying schedule used here, making it a reusable tool for other decentralized manifold optimization analyses.
  • On a sphere and on the manifold of 5×5 positive-definite matrices with real environmental data, the strongly g-convex schedule achieves lower final regret and smaller consensus error than the g-convex baseline, matching the predicted log-versus-square-root separation.

Reading between the lines

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

  • Because the step size η_t depends on knowing μ, the strong-convexity parameter, an adaptive or parameter-free variant would be a natural extension; the paper does not address what happens when μ is misspecified.
  • The non-positivity condition on the interference term (equation 13) is imported from a companion analysis and tied to the choice s = C_2/(2C_1); a self-contained proof on concrete manifolds would reveal whether there is a hidden curvature restriction on the allowed consensus step size.
  • The network-error lemma's ability to handle arbitrary decaying schedules suggests it could be combined with accelerated gossip or dynamic-regret comparators, both of which the paper lists as open directions.
  • A testable prediction is that the bandit regret curves stay separated from the O(√T) baseline even for much larger T than the experiments run, because the smoothing, shrinkage, and projection errors are engineered to be constants.
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

3 major / 5 minor

Summary. The paper studies decentralized online optimization of μ-strongly geodesically convex losses on Riemannian manifolds with bounded sectional curvature. It analyzes the same two-step gradient-plus-consensus algorithm as prior work, but with a decaying step size η_t = 1/(μt + L/D), and proves an O(log T) static regret bound in the full-information setting (Theorem 3.2). It then extends the result to two-point bandit feedback using a smoothed surrogate and a two-point gradient estimator with δ = T^{-2}, obtaining the same O(log T) rate in expectation (Theorem 4.5). The proof decomposes regret into a network-error term and an optimization-error term; the main novelty is a network-error analysis for time-varying step sizes (Lemma 3.1), in which the consensus error is bounded by a time-weighted geometric convolution of past step sizes that remains O(log T). Numerical experiments on S^7 and on the SPD manifold with FLUXNET2015 data illustrate the predicted logarithmic growth.

Significance. If the proofs are correct, the paper fills a natural gap: it provides the first decentralized online Riemannian optimization result that exploits strong geodesic convexity, matching the Euclidean and centralized Riemannian logarithmic regret rate up to curvature-dependent constants and the spectral-gap factor (1 − σ_2(W))^{-1}. The network-error analysis for decaying step sizes is a genuine technical contribution that goes beyond the fixed-step-size analyses of [18], [19], and the strong-sub-g-convexity lemma for the two-point estimator is a nontrivial addition. The paper also makes good use of comparison-based geometric inequalities and provides reproducible experiments. However, two load-bearing ingredients are imported from the overlapping-author preprint [19] without a self-contained proof: the consensus contraction factor in Lemma 3.1 and the interference-term sign condition (13). Since these are curvature-sensitive and central to the logarithmic rate, the current manuscript is not yet fully verified.

major comments (3)
  1. [Appendix VII-B, Eq. (13)] This is the load-bearing step of the full-information proof. After summing (10), the authors discard the interference term ∑_i ∑_t (1/2η_t)[d^2(x_{i,t+1},x_*) − d^2(y_{i,t+1},x_*)] by asserting that, with s = C_2/(2C_1), the per-t sum over i is non-positive, 'following the argument of [19, proof of Thm. IV.2]'. No derivation is provided. This inequality is not implied by the paper's own Lemma 7.1; it concerns the nonlinear consensus step (II) on a positively curved manifold, and the Euclidean analogue is not sufficient. Because η_t = Θ(1/t), even a constant per-round positive contribution would produce an O(T^2) term and destroy the logarithmic bound. The authors must either prove (13) under Assumptions 2.1–2.3 or supply a precise, checkable statement of the invoked [19] result, including its curvature and step-size conditions.
  2. [Lemma 3.1 / Appendix VII-A] The entire network-error analysis, used in both Theorem 3.2 and Theorem 4.5, rests on the contraction inequality ∑_i d^2(x_{i,t+1}, \bar x_{t+1}) ≤ ρ ∑_i d^2(y_{i,t+1}, \bar y_{t+1}) imported from [19, Theorem III.2], with ρ = 1 − C_2^2(1−σ_2(W))/(2C_1(1+C_3D^2)^2). The constants C_1, C_2, C_3 are only defined in Appendix VII-E, and no proof or verification is given that the choice s = C_2/(2C_1) ensures ρ ∈ (0,1) for manifolds with positive sectional curvature. Since [19] is an overlapping-author preprint and the curvature sensitivity of this contraction factor is central to the claimed rate, the proof is not self-contained at a load-bearing point. Please include a self-contained proof of the contraction bound, or state the exact hypotheses and prove the specialization used here.
  3. [Appendix VII-E / VII-D] In the bandit analysis, the smoothed loss is claimed to be C_8L-Lipschitz with C_8 = cosh(√K_min), and the integral in Appendix VII-D uses cosh(√K_min δ∥u∥). When K_min < 0, which is allowed by Assumption 2.2 and is the case for the SPD/Hadamard experiment in Section V-B, √K_min is imaginary, so C_8 as defined is not a real constant; the intended comparison formula must be restated with the correct sign (e.g., involving √(−K_min) or a suitable positive-curvature expression). This makes the Lipschitz bound in the bandit proof invalid on Hadamard manifolds as written. Additionally, C_4 in Appendix VII-E depends on δ through C_6 = κ g_4(K_min, δ), while Lemma 4.3 states that C_4 depends only on K_min, K_max, and D; this dependence should be stated consistently.
minor comments (5)
  1. [Theorem 3.2] The constants R_1 and R_2 are deferred to Appendix VII-E; stating their explicit values in the theorem statement would improve reproducibility.
  2. [Appendix VII-D, Step 2 of Lemma 4.3] The Stokes identity for the two-point estimator is invoked without specifying the normalization of the uniform measure p on the ball. Please state the exact identity with the constants, or give a reference with equation numbers.
  3. [Appendix VII-E] The phrase 'C_5 find the definition in Lemma 6 of [45]' is awkward; give the definition or a precise citation with the lemma statement.
  4. [Section V-A] The bandit results are averaged over only 5 Monte-Carlo runs and no error bars are reported. Adding error bars or more runs would strengthen Figure 1.
  5. [Appendix VII-D, proof of Theorem 4.5] The display following 'By Lemma 4.3' mixes deterministic quantities with E_u-expectations without making the conditional expectation explicit. Writing all expectations on both sides would remove ambiguity.

Circularity Check

3 steps flagged · score 4.0 of 10

The O(log T) rate is not fitted or definitionally forced, but the proof's decisive network-error contraction and the sign of the interference term are imported without derivation from overlapping-author preprint [19], making the decentralized claim partly load-bearing on self-citation.

  1. self citation load bearing [Appendix VII-A, proof of Lemma 3.1 (and Section III-B before Lemma 3.1)]
    "By [19, Theorem III.2], one application of the consensus step (II) contracts the Fréchet variance of the iterates as sum_i d^2(x_{i,t+1}, bar{x}_{t+1}) ≤ ρ sum_i d^2(y_{i,t+1}, bar{y}_{t+1}), where the factor ρ ∈ (0,1)."

    This is the starting point of the network-error recursion a_{t+1} ≤ √ρ a_t + √ρ n η_t L. Without ρ ∈ (0,1), the time-weighted sum in Lemma 3.1 is not O(log T), and Theorem 3.2 has no logarithmic bound. The paper does not prove this contraction theorem; it imports it from [19], an overlapping-author preprint (Sahinoglu and Shahrampour are co-authors of the present paper). The constants C1–C3 and the formula for ρ are only named in Appendix VII-E, not derived. Thus the decentralized part of the central claim rests on a load-bearing self-citation.

  2. self citation load bearing [Appendix VII-B, proof of Theorem 3.2, 'The interference term', Eq. (13)]
    "Following the argument of [19, proof of Thm. IV.2], with s=(2C_1)^{-1}C_2 the doubly stochastic, symmetric weights make the per-t sum non-positive, i.e., sum_i 1/(2η_t)(d^2(x_{i,t+1}, x*) − d^2(y_{i,t+1}, x*)) ≤ 0. (13)"

    Inequality (13) is the only mechanism that removes the consensus-interference sum in Eq. (12). Since η_t = Θ(1/t), a constant positive contribution per round would dominate the 1/(2η_t) prefactor and destroy the logarithmic rate. The proof is not supplied here; it is deferred to the proof of Theorem IV.2 in the same overlapping-author preprint [19]. This is a curvature-sensitive sign condition on the first-order Fréchet-mean update, so importing it without derivation makes the headline regret bound contingent on an in-house unverified result rather than a self-contained argument.

1 more flagged steps
  1. self citation load bearing [Appendix VII-C, proof of Lemma 4.3, Step 4]
    "Using ∥grad f∥ ≤ L and the bound ∥Log_{x_u}(y_u) − dExp_x(u)[Log_x(y)]∥ ≤ δC_4 from [19], this is at most δC_4 L."

    This comparison bound is needed to turn the two-point estimator into the strong-sub-g-convexity inequality (6) for the smoothed surrogate. It is imported from the same overlapping-author preprint [19] without derivation. If this geometric comparison has hidden restrictions or a wrong constant, the bandit Theorem 4.5 loses the slack control needed to preserve O(log T). The import is load-bearing for the bandit result, though the strong-sub-g-convexity analysis around it is new.

full rationale

The paper's central claim is not circular in the sense of a fitted parameter being renamed a prediction or a result being true by definition. The optimization-error analysis, the shifted step size η_t = (μt + L/D)^{-1}, the telescoping argument, and the strong-sub-g-convexity framework are genuine new content, and the numerical experiments are external sanity checks. However, two decisive ingredients of the proof are imported from [19], a preprint authored by two of the present authors: (i) the consensus contraction ρ ∈ (0,1) used to derive Lemma 3.1, and (ii) the non-positivity of the interference sum, Eq. (13), without which the logarithmic rate collapses. A third import, the δC_4 distance-comparison bound in Lemma 4.3, is load-bearing for the bandit extension. These are not minor self-citations: they carry the network-error part of the main theorem. But they do not reduce the result to a by-construction equivalence, and the paper does add independent analysis around them. I therefore assign a moderate score of 4, reflecting load-bearing self-citation while acknowledging the independent content of the central derivation.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

No new physical or mathematical entities are postulated. The analysis rests on geometric assumptions (bounded curvature, diameter constraints) and imported lemmas from [19], [16], and standard Riemannian optimization literature. The paper's own contribution is the analysis under decaying step sizes and the subg-convexity lemma.

free parameters (4)
  • consensus step size s = s = C_2/(2 C_1)
    Chosen from the geometric constants C1, C2 to ensure the interference term (13) is non-positive. This value is imported from [19] and is a parameter of the algorithm tuned to make the analysis go through, not fitted to data.
  • smoothing radius δ (bandit) = δ = T^{-2}
    Chosen to balance the O(δ T + τ D T) smoothing/shrinkage terms against the projection term τ D^2 Σ 1/η_t = O(δ T^2), yielding a constant. It is a design parameter of the proof, not fitted to data.
  • step-size shift c_0 (full info) = c_0 unspecified (η_t = O(1/(t+c_0)))
    The shift c_0 is introduced only to certify η_t L ≤ D on positively curved manifolds; its exact value is never specified, so the constants R_1, R_2 are not fully explicit. Not fitted to data, but a proof-construction parameter.
  • shrinkage factor τ = τ = δ θ / r
    Tied to δ and the geometric constants θ, r in Assumption 4.1; chosen so perturbed queries stay feasible. Part of the proof construction.
assumptions (6)
  • domain assumption Contraction property of consensus step (existence of ρ in (0,1) in Lemma 3.1)
    Imported from [19, Theorem III.2] without proof. The entire network-error analysis and the final regret bound rest on this contraction, which involves curvature bounds, σ_2(W), and the consensus step size s.
  • domain assumption Interference term non-positivity (inequality (13))
    Stated as 'Following the argument of [19, proof of Thm. IV.2]' with s = C_2/(2C_1). If this inequality failed, an extra O(Σ 1/η_t d^2) term would destroy the O(log T) rate. Imported without proof.
  • standard math Distance comparison lemmas (Lemma 7.1, 7.2)
    Cited from [43, Corollary 2.1], [29, Lemma 5], [16, Lemma 21], and [45]. These are standard tools in Riemannian optimization; the paper does not prove them but states the resulting constants C1-C7. Their validity requires Assumption 2.2 (bounded curvature and injectivity radius).
  • standard math Smoothed surrogate Lipschitz constant via [44, Lemma 1]
    Used in Theorem 4.5 Step 2 to bound the network term of the bandit regret. Cited from Mangoubi-Smith.
  • standard math Stokes' identity for the spherical two-point estimator (Lemma 4.3 Step 2)
    The identity relating E_u[g_δ(x)] to a δ-ball average of the gradient is cited to [37], [38]. It is standard but not proved in the paper.
  • standard math Hessian comparison for strong g-convexity of squared distance on geodesic balls
    Used in the sphere experiment to claim d^2(·, z) is μ-strongly g-convex on X; cited to Petersen. Not part of the proof of the main theorems.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Decentralized Online Riemannian Optimization for Strongly Geodesically Convex Functions." pith.science (2026). https://pith.science/paper/BV46NDWJ

@misc{pith2026260720316,
  author       = {Pith},
  title        = {Pith review of: Decentralized Online Riemannian Optimization for Strongly Geodesically Convex Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BV46NDWJ}},
  note         = {Machine review of arXiv:2607.20316}
}
abstract

We study decentralized online optimization for strongly geodesically convex (strongly g-convex) losses on Riemannian manifolds with bounded sectional curvature, including positively curved manifolds. In centralized Riemannian optimization, strong g-convexity tightens the optimal regret from $O(\sqrt{T})$ to $O(\log T)$, where $T$ is the time horizon; in the decentralized Riemannian setting, however, existing methods address only g-convex losses, leaving the strongly g-convex regime unexplored. One challenge is that the required decaying step size in the centralized regime is incompatible with existing network-error analyses, which typically assume a fixed step size. First, we provide a general network-error analysis for time-varying schedules. Next, we build on this analysis to establish the first $O(\log T)$ static regret bound for decentralized online Riemannian gradient descent, matching the minimax-optimal rate for strongly-convex Euclidean online optimization. Finally, we prove the same $O(\log T)$ regret bound for the two-point bandit feedback setting using novel strong subconvexity arguments for the smoothed versions of the loss functions.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 4 linked inside Pith

  1. [19]

    Decentralized online Rieman- nian optimization beyond Hadamard manifolds,

    E. Sahinoglu and S. Shahrampour, “Decentralized online Rieman- nian optimization beyond Hadamard manifolds,” arXiv preprint arXiv:2509.07779, 2025

    arXiv 2025

  2. [18]

    Decentralized online Riemannian optimization with dynamic environments,

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

    arXiv 2024

  3. [1]

    Online distributed convex optimization on dynamic networks,

    S. Hosseini, A. Chapman, and M. Mesbahi, “Online distributed convex optimization on dynamic networks,”IEEE Transactions on Automatic Control, vol. 61, no. 11, pp. 3545–3550, 2016

    2016

  4. [2]

    Distributed online optimization in dynamic environments using mirror descent,

    S. Shahrampour and A. Jadbabaie, “Distributed online optimization in dynamic environments using mirror descent,”IEEE Transactions on Automatic Control, vol. 63, no. 3, pp. 714–725, 2018

    2018

  5. [3]

    Distributed online convex optimization on time-varying directed graphs,

    M. Akbari, B. Gharesifard, and T. Linder, “Distributed online convex optimization on time-varying directed graphs,”IEEE Transactions on Control of Network Systems, vol. 4, no. 3, pp. 417–428, 2017

    2017

  6. [4]

    Distributed autonomous online learning: Regrets and intrinsic privacy-preserving properties,

    F. Yan, S. Sundaram, S. V . N. Vishwanathan, and Y . Qi, “Distributed autonomous online learning: Regrets and intrinsic privacy-preserving properties,”IEEE Transactions on Knowledge and Data Engineering, vol. 25, no. 11, pp. 2483–2493, 2013

    2013

  7. [5]

    Distributed riemannian optimization with lazy communication for collaborative geometric estimation,

    Y . Tian, A. S. Bedi, A. Koppel, M. Calvo-Fullana, D. M. Rosen, and J. P. How, “Distributed riemannian optimization with lazy communication for collaborative geometric estimation,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 4391– 4398

    2022

  8. [6]

    A linearly convergent algorithm for distributed principal component analysis,

    A. Gang and W. U. Bajwa, “A linearly convergent algorithm for distributed principal component analysis,”Signal Processing, vol. 193, p. 108408, 2022. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S016516842100445X

    2022

Show all 45 references
  1. [7]

    Riemannian optimization for active mapping with robot teams,

    A. Asgharivaskasi, F. Girke, and N. Atanasov, “Riemannian optimization for active mapping with robot teams,”IEEE Transactions on Robotics, vol. 41, pp. 1077–1097, 2025

    2025

  2. [8]

    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

  3. [9]

    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, 2013

    2013

  4. [10]

    Intrinsic decentralized stochastic rie- mannian optimization on manifolds with bounded sectional curvature,

    D. T. Nguyen and C. A. Uribe, “Intrinsic decentralized stochastic rie- mannian optimization on manifolds with bounded sectional curvature,” arXiv preprint arXiv:2603.17096, 2026

    2026

  5. [11]

    Distributed rie- mannian optimization in geodesically non-convex environments,

    X. Wang, R. Borsoi, C. Richard, and A. H. Sayed, “Distributed rie- mannian optimization in geodesically non-convex environments,”arXiv preprint arXiv:2512.04915, 2025

    2025 arXiv

  6. [12]

    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,” inProceedings of the 38th International Conference on Machine Learning (ICML), 2021, pp. 1594–1605

    2021

  7. [13]

    On the local linear rate of consensus on the Stiefel manifold,

    ——, “On the local linear rate of consensus on the Stiefel manifold,” IEEE Transactions on Automatic Control, vol. 69, no. 4, pp. 2324–2339, 2023

    2023

  8. [14]

    Decentralized projected Riemannian gradient method for smooth optimization on compact submanifolds,

    K. Deng and J. Hu, “Decentralized projected Riemannian gradient method for smooth optimization on compact submanifolds,”Numerische Mathematik, 2025

    2025

  9. [15]

    Logarithmic regret algorithms for online convex optimization,

    E. Hazan, A. Agarwal, and S. Kale, “Logarithmic regret algorithms for online convex optimization,”Machine Learning, vol. 69, no. 2–3, pp. 169–192, 2007

    2007

  10. [16]

    Online optimization over Riemannian manifolds,

    X. Wang, Z. Tu, Y . Hong, Y . Wu, and G. Shi, “Online optimization over Riemannian manifolds,”Journal of Machine Learning Research, vol. 24, no. 84, pp. 1–67, 2023

    2023

  11. [17]

    Optimal strate- gies and minimax lower bounds for online convex games,

    J. Abernethy, P. L. Bartlett, A. Rakhlin, and A. Tewari, “Optimal strate- gies and minimax lower bounds for online convex games,”Proceedings of the 21st Annual Conference on Learning Theory (COLT), pp. 415– 424, 2008

    2008

  12. [20]

    Finite-time analysis of stochastic nonconvex nonsmooth optimization on the Riemannian man- ifolds,

    E. Sahinoglu, Y . Sun, and S. Shahrampour, “Finite-time analysis of stochastic nonconvex nonsmooth optimization on the Riemannian man- ifolds,” inAdvances in Neural Information Processing Systems, vol. 38, 2025

    2025

  13. [21]

    Riemannian projection-free online learning,

    Z. Hu, G. Wang, and J. D. Abernethy, “Riemannian projection-free online learning,” inAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  14. [22]

    Online convex programming and generalized infinitesi- mal gradient ascent,

    M. Zinkevich, “Online convex programming and generalized infinitesi- mal gradient ascent,” inProceedings of the 20th International Confer- ence on Machine Learning (ICML), 2003, pp. 928–935

    2003

  15. [23]

    Online optimization: Competing with dynamic comparators,

    A. Jadbabaie, A. Rakhlin, S. Shahrampour, and K. Sridharan, “Online optimization: Competing with dynamic comparators,” inArtificial Intel- ligence and Statistics. PMLR, 2015, pp. 398–406

    2015

  16. [24]

    Online op- timization in dynamic environments: Improved regret rates for strongly convex problems,

    A. Mokhtari, S. Shahrampour, A. Jadbabaie, and A. Ribeiro, “Online op- timization in dynamic environments: Improved regret rates for strongly convex problems,” inProceedings of the 55th IEEE Conference on Decision and Control (CDC), 2016, pp. 7195–7201

    2016

  17. [25]

    On online optimization: Dynamic regret analysis of strongly convex and smooth problems,

    T.-J. Chang and S. Shahrampour, “On online optimization: Dynamic regret analysis of strongly convex and smooth problems,” inProceedings of the 35th AAAI Conference on Artificial Intelligence, vol. 35, no. 8, 2021, pp. 6966–6973. 12

    2021

  18. [26]

    Nearly optimal regret for decentralized online convex optimization,

    Y . Wan, T. Wei, M. Song, and L. Zhang, “Nearly optimal regret for decentralized online convex optimization,” inThe Thirty Seventh Annual Conference on Learning Theory. PMLR, 2024, pp. 4862–4888

    2024

  19. [27]

    Absil, R

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

    2009

  20. [28]

    Sato,Riemannian Optimization and Its Applications

    H. Sato,Riemannian Optimization and Its Applications. Springer, 2021, vol. 670

    2021

  21. [29]

    First-order methods for geodesically convex op- timization,

    H. Zhang and S. Sra, “First-order methods for geodesically convex op- timization,” inProceedings of the 29th Annual Conference on Learning Theory (COLT), 2016, pp. 1617–1638

    2016

  22. [30]

    Accelerated gradient methods for geodesically convex optimization: Tractable algorithms and convergence analysis,

    J. Kim and I. Yang, “Accelerated gradient methods for geodesically convex optimization: Tractable algorithms and convergence analysis,” in Proceedings of the 39th International Conference on Machine Learning (ICML), 2022

    2022

  23. [31]

    Global Riemannian acceleration in hyperbolic and spherical spaces,

    D. Mart ´ınez-Rubio, “Global Riemannian acceleration in hyperbolic and spherical spaces,” inProceedings of the 33rd International Conference on Algorithmic Learning Theory (ALT), 2022, pp. 768–826

    2022

  24. [32]

    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

  25. [33]

    Decentralized optimization on compact submanifolds by quantized riemannian gradient tracking,

    J. Chen, L. Liu, T. Zhu, Y . Liu, G. Dai, Y . Jiang, and I. W. Tsang, “Decentralized optimization on compact submanifolds by quantized riemannian gradient tracking,”IEEE Transactions on Signal Processing, 2025

    2025

  26. [34]

    Tracking and regret bounds for online zeroth-order Euclidean and Riemannian optimization,

    A. I. Maass, C. Manzie, D. Ne ˇsi´c, J. H. Manton, and I. Shames, “Tracking and regret bounds for online zeroth-order Euclidean and Riemannian optimization,”SIAM Journal on Optimization, vol. 32, no. 2, pp. 445–469, 2022

    2022

  27. [35]

    Riemannian online optimistic algorithms with dynamic regret,

    X. Wang, D. Yuan, Y . Hong, Z. Hu, L. Wang, and G. Shi, “Riemannian online optimistic algorithms with dynamic regret,”IEEE Transactions on Automatic Control, vol. 70, no. 10, pp. 6481–6496, 2025

    2025

  28. [36]

    Online optimization on Hadamard manifolds: Curvature independent regret bounds on horospherically convex objectives,

    E. Sahinoglu and S. Shahrampour, “Online optimization on Hadamard manifolds: Curvature independent regret bounds on horospherically convex objectives,” arXiv preprint arXiv:2509.11236, 2025

    2025 arXiv

  29. [37]

    Optimal algorithms for online convex optimization with multi-point bandit feedback,

    A. Agarwal, O. Dekel, and L. Xiao, “Optimal algorithms for online convex optimization with multi-point bandit feedback,” inProceedings of the 23rd Annual Conference on Learning Theory (COLT), 2010, pp. 28–40

    2010

  30. [38]

    Online convex optimization in the bandit setting: Gradient descent without a gradient,

    A. D. Flaxman, A. T. Kalai, and H. B. McMahan, “Online convex optimization in the bandit setting: Gradient descent without a gradient,” inProceedings of the 16th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), 2005, pp. 385–394

    2005

  31. [39]

    Boumal,An introduction to optimization on smooth manifolds

    N. Boumal,An introduction to optimization on smooth manifolds. Cambridge University Press, 2023

    2023

  32. [40]

    Non-existence of continuous convex functions on certain Riemannian manifolds,

    S.-T. Yau, “Non-existence of continuous convex functions on certain Riemannian manifolds,”Mathematische Annalen, vol. 207, no. 4, pp. 269–270, 1974

    1974

  33. [41]

    Petersen,Riemannian Geometry, 2nd ed., ser

    P. Petersen,Riemannian Geometry, 2nd ed., ser. Graduate Texts in Mathematics. Springer, 2006, vol. 171

    2006

  34. [42]

    Black-box reductions for decentralized online convex opti- mization in changing environments,

    Y . Wan, “Black-box reductions for decentralized online convex opti- mization in changing environments,” inProceedings of the 38th Annual Conference on Learning Theory (COLT), 2025, pp. 5605–5631

    2025

  35. [43]

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

    F. Alimisis, A. Orvieto, G. B ´ecigneul, and A. Lucchi, “A continuous- time perspective for modeling acceleration in riemannian optimization,” inProceedings of the Twenty Third International Conference on Arti- ficial Intelligence and Statistics, ser. Proceedings of Machine Le...

    2020

  36. [44]

    Rapid mixing of geodesic walks on manifolds with positive curvature,

    O. Mangoubi and A. Smith, “Rapid mixing of geodesic walks on manifolds with positive curvature,”The Annals of Applied Probability, vol. 28, no. 4, pp. 2501–2543, 2018

    2018

  37. [45]

    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 (NeurIPS), vol. 32, 2019. Zhanyuan Caiis currently a Ph.D. student in the Department of Mechanical and Industrial Engi- neering at N...

    2019

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.