REVIEW 1 major objections 2 minor
An Adaptive Decentralized Quasi-Newton Method with Stepsizes Independent of the Local-Update Budget
T0 review · 1 major / 2 minor · reviewed 2026-06-27 · grok-4.3
Pith's one-line read AdaDQN decouples the allowable stepsize from the number of local updates in decentralized nonconvex optimization.
desk verdict AdaDQN claims a decoupled stepsize bound via its RIA framework and adaptive termination, but the decoupling may still depend on whether the local error analysis avoids reintroducing 1/K factors. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The Robust Inexact Algorithm (RIA) framework derived from the Majorization-Minimization lens, which analyzes the effect of inexact local updates under the chosen termination criterion and BFGS rule.
What would settle it
An explicit counterexample or numerical instance in which increasing the maximum number of local updates forces a strictly smaller stepsize to maintain convergence on a smooth nonconvex problem would falsify the decoupling claim.
Extended reading notes
Core claim
AdaDQN achieves global convergence to a first-order stationary point for smooth nonconvex decentralized problems, with the convergence stepsize bound proven to be independent of the reciprocal of the maximum number of local updates, through integration of a safeguarded termination criterion, scalable BFGS update, and event-triggered communication within the RIA framework.
Load-bearing premise
The Robust Inexact Algorithm framework must hold for the specific local-update termination criterion and BFGS update used in AdaDQN.
Editorial extensions
If this is right
- Global convergence to first-order stationary points holds without the previous stepsize penalty from extra local work.
- Communication rounds can be reduced via the event-triggered protocol while preserving the same theoretical stepsize range.
- The method applies directly to undirected networks and yields a better computation-communication tradeoff than prior decentralized schemes.
- The RIA framework supplies a template for designing other local-update methods whose stepsize bounds remain stable.
Reading between the lines
- The decoupling result may extend to directed networks or stochastic gradients if the RIA analysis can be adapted.
- In large-scale settings the independence from local-update count could allow agents to run many cheap iterations before any communication without retuning stepsizes.
- Similar majorization-minimization arguments might remove analogous bottlenecks in other inexact or asynchronous optimization algorithms.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes AdaDQN, an adaptive decentralized quasi-Newton method for smooth nonconvex optimization over undirected networks. It revisits local-update methods via a Majorization-Minimization lens to derive a Robust Inexact Algorithm (RIA) framework, then uses a safeguarded consensus-aware termination criterion, memoryless BFGS updates, and event-triggered communication. The central claim is global convergence to a first-order stationary point with a stepsize bound that is decoupled from the reciprocal of the maximum number of local updates (overcoming the standard 1/K bottleneck). Experiments show improved computation-communication tradeoffs versus prior decentralized methods.
Significance. If the decoupling result holds, the work would be significant for decentralized optimization: it removes a pessimistic theoretical restriction that has limited the practical use of multiple local updates. The RIA framework derived from the MM lens, together with the explicit construction of a termination criterion that preserves the required inexactness without reintroducing K-dependent factors, would constitute a technical contribution. The memoryless BFGS and event-triggered protocol are pragmatic additions that could translate to measurable gains in communication cost.
major comments (1)
- [§4] §4 (convergence analysis) and the RIA framework definition: the proof that the stepsize upper bound is independent of 1/K rests on the local-update error and consensus-error terms satisfying the RIA inexactness conditions without producing a multiplicative K factor in the descent inequality or effective Lipschitz constant. The manuscript must explicitly verify (via the specific termination criterion and memoryless BFGS update) that no such factor reappears; otherwise the claimed decoupling does not follow.
minor comments (2)
- [Algorithm 1] Notation for the termination threshold and the event-trigger parameter should be introduced once and used consistently; currently the same symbol appears to be overloaded in the algorithm box and the analysis.
- [§5] The experimental section would benefit from reporting the actual number of local updates per iteration (not just the maximum allowed) so readers can directly assess the computation-communication tradeoff achieved.
Simulated Author's Rebuttal
We thank the referee for the constructive feedback and for acknowledging the potential significance of the decoupling result. We address the single major comment below.
read point-by-point responses
-
Referee: [§4] §4 (convergence analysis) and the RIA framework definition: the proof that the stepsize upper bound is independent of 1/K rests on the local-update error and consensus-error terms satisfying the RIA inexactness conditions without producing a multiplicative K factor in the descent inequality or effective Lipschitz constant. The manuscript must explicitly verify (via the specific termination criterion and memoryless BFGS update) that no such factor reappears; otherwise the claimed decoupling does not follow.
Authors: We agree that explicit verification is essential for the claimed decoupling. In the proof of the main convergence result (Theorem 4.1), the safeguarded consensus-aware termination criterion is constructed so that the consensus error is bounded by a quantity independent of the maximum number of local updates K; this bound is then substituted directly into the RIA inexactness condition without introducing a multiplicative K. The memoryless BFGS update is shown to satisfy the required curvature and descent conditions while preserving the same inexactness tolerance, again without K-dependent accumulation. To make this verification more prominent, we will add a dedicated remark immediately after the proof that isolates the relevant error bounds and confirms the absence of any K factor in the descent inequality or effective Lipschitz constant. revision: partial
Circularity Check
No circularity: RIA framework and stepsize decoupling derived independently from MM lens
full rationale
The paper derives the RIA framework directly from the Majorization-Minimization lens within this work, then applies it to prove global convergence and the decoupling of the convergence stepsize bound from 1/max-local-updates for the specific termination criterion and BFGS update. No equations or claims reduce a result to a fitted parameter, self-citation, or input by construction; the central claim rests on the new framework rather than prior fitted constants or renamed patterns. The derivation is self-contained against external benchmarks.
Assumptions & free parameters
assumptions (2)
- domain assumption Objective functions are smooth
- domain assumption Communication network is undirected and connected
invented entities (1)
-
Robust Inexact Algorithm (RIA) framework
Cite this review
Pith. "Pith review of An Adaptive Decentralized Quasi-Newton Method with Stepsizes Independent of the Local-Update Budget." pith.science (2026). https://pith.science/paper/VNAQ337V
@misc{pith2026260609070,
author = {Pith},
title = {Pith review of: An Adaptive Decentralized Quasi-Newton Method with Stepsizes Independent of the Local-Update Budget},
year = {2026},
howpublished = {\url{https://pith.science/paper/VNAQ337V}},
note = {Machine review of arXiv:2606.09070}
}
abstract
This paper proposes a novel Adaptive Decentralized Quasi-Newton (AdaDQN) method for solving smooth nonconvex optimization problems over undirected networks. While standard decentralized algorithms with multiple fixed local updates typically admit a convergence stepsize inversely proportional to the number of local updates, we show that this scaling is worst-case tight for the typical unscaled fixed-local-update scheme. We revisit a class of gradient-tracking methods with this scheme from a surrogate-based perspective and establish a Robust Inexact Algorithm (RIA) framework. Inspired by this framework, AdaDQN integrates a safeguarded consensus-aware termination criterion, a standard event-triggered communication protocol, and a scalable memoryless BFGS update. We establish an $\mathcal{O}(1/T)$ best-iterate rate for first-order stationarity. For a squared stationarity tolerance $\delta$, the guaranteed gradient complexity of the fixed scheme is $\mathcal{O}(nK_g/\delta)$, whereas AdaDQN attains $\mathcal{O}(n/\delta+n\alpha\tilde\varepsilon^{-2})$, independent of the maximum local-update budget ($K_g$). Numerical experiments demonstrate that AdaDQN achieves a superior computation-communication tradeoff, outperforming state-of-the-art decentralized methods across various performance metrics.
Figures
Figures from the paper (8 more)
Reviewed June 27, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.