pith. sign in

arxiv: 2605.25306 · v3 · pith:TAETDGOWnew · submitted 2026-05-25 · 📡 eess.SY · cs.SY

Nonlinear-Gain Distributed Zeroth-Order Optimization for Networked Black-Box Control

Shengjun Zhang , Tingyi Liu , Heng Zhang , Dong Xie This is my paper

Pith reviewed 2026-06-29 21:07 UTC · model grok-4.3

classification 📡 eess.SY cs.SY
keywords distributed optimizationzeroth-order methodsnonlinear gainpowerball mapnetworked controlblack-box optimizationstochastic optimizationpeer-to-peer networks
0
0 comments X

The pith

A fractional-power powerball map applied locally to gradient estimates yields leading convergence rates for distributed zeroth-order optimization over networks.

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

The paper introduces ZOOM-PB, a coordinate-sampling distributed zeroth-order method for peer-to-peer networks where agents access only function values. It equips the method with a nonlinear feedback gain realized by a fractional-power powerball map that amplifies weak signals in flat regions and damps large stochastic estimates. Under standard smoothness, variance, and connectivity assumptions this produces the nonconvex stationarity rate O(sqrt(p/(nT))) and, under the Polyak-Lojasiewicz condition, the objective residual rate O(p/(nT)). The approach therefore keeps the known asymptotic order while altering finite-time behavior through local nonlinear control rather than added communication or refined estimators.

Core claim

ZOOM-PB preserves the known distributed zeroth-order convergence order while changing finite-time behavior through a local nonlinear control gain realized by a fractional-power powerball map on the estimated gradient; the map requires no extra transmitted states.

What carries the argument

The fractional-power powerball map acting as a nonlinear feedback gain on the estimated gradient.

If this is right

  • The method attains the leading nonconvex stationarity rate without transmitting additional states beyond the usual local updates.
  • Under the Polyak-Lojasiewicz condition the same mechanism produces the leading objective residual rate O(p/(nT)).
  • Empirical runs on black-box learning and UAV source seeking exhibit faster convergence precisely in weak-signal regimes.
  • The nonlinear gain replaces the need for primal-dual tracking or gradient-refinement steps used in prior distributed zeroth-order schemes.

Where Pith is reading between the lines

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

  • The same local nonlinear gain could be paired with other coordinate-sampling patterns to target different communication topologies.
  • Because the map is memoryless and local, it may extend directly to asynchronous or time-varying networks without redesigning the analysis.
  • Testing the powerball exponent on problems whose flatness varies with dimension would reveal how the finite-time improvement scales with p.

Load-bearing premise

The stated rates hold only when the objective is smooth, the zeroth-order oracle has bounded variance, and the communication network remains connected.

What would settle it

Measure the stationarity gap after T iterations on a smooth nonconvex test function over an n-agent connected network; the gap should scale as sqrt(p/(nT)) when p, n, and T are varied while keeping other factors fixed.

Figures

Figures reproduced from arXiv: 2605.25306 by Dong Xie, Heng Zhang, Shengjun Zhang, Tingyi Liu.

Figure 1
Figure 1. Figure 1: Training-loss comparison on the black-box binary classification [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Source-estimate trajectories under ZOOM-PB. Agents use only local [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Network-wide average concentration in the UAV source-seeking [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

This letter studies distributed stochastic optimization over a peer-to-peer network when agents can query only zeroth-order function values. We propose ZOOM-PB, a coordinate-sampling distributed zeroth-order method equipped with a fractional-power powerball map. Unlike existing distributed zeroth-order methods that mainly refine gradient estimation or introduce primal--dual tracking, the proposed mechanism acts as a nonlinear feedback gain on the estimated gradient: it amplifies weak signals in flat regions and attenuates large stochastic estimates without adding transmitted states. Under standard smoothness, oracle-variance, and network-connectivity assumptions, ZOOM-PB achieves the leading nonconvex stationarity rate $\mathcal{O}(\sqrt{p/(nT)})$, where $p$ is the decision dimension, $n$ is the number of agents, and $T$ is the iteration horizon. Under the Polyak--{\L}ojasiewicz condition, it further attains the leading objective residual rate $\mathcal{O}(p/(nT))$. Thus the method preserves the known distributed ZO order while changing the finite-time behavior through a local nonlinear control gain. Simulations on black-box learning and sensor-driven UAV source seeking show faster empirical convergence in weak-signal regimes.

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.

Referee Report

0 major / 2 minor

Summary. The manuscript proposes ZOOM-PB, a coordinate-sampling distributed zeroth-order optimization method for peer-to-peer networks that applies a fractional-power powerball map as a local nonlinear gain to zeroth-order gradient estimates. The central claim is that, under standard smoothness, bounded oracle variance, and network connectivity assumptions, the algorithm attains the leading nonconvex stationarity rate O(sqrt(p/(nT))) and, under the Polyak-Lojasiewicz condition, the leading objective residual rate O(p/(nT)), while improving finite-time behavior in weak-signal regimes without additional transmitted states or altered assumptions. Simulations on black-box learning and sensor-driven UAV source seeking are included to demonstrate empirical gains.

Significance. If the rates are rigorously established without hidden parameter dependence in the powerball map, the work is significant for showing that a simple local nonlinear feedback can enhance practical convergence in distributed black-box settings while preserving optimal leading-order complexity from the coordinate ZO literature. The explicit preservation of known rates O(sqrt(p/(nT))) and O(p/(nT)) under standard assumptions, together with the focus on networked control applications, strengthens the contribution.

minor comments (2)
  1. The definition and properties (e.g., Lipschitz constant, range) of the fractional-power powerball map should be stated explicitly in the main text near its first introduction, rather than deferred entirely to an appendix, to allow readers to verify that it introduces no extra factors into the leading rate terms.
  2. In the simulation section, the choice of the power parameter in the powerball map and its sensitivity should be reported with at least one ablation plot or table entry, as this directly supports the claim of improved finite-time behavior.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of the manuscript, the recognition of its significance in preserving leading rates while improving finite-time behavior via local nonlinear gain, and the recommendation for minor revision. No specific major comments were raised in the report.

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The paper presents ZOOM-PB as a coordinate-sampling distributed zeroth-order method with a local fractional-power powerball map acting as nonlinear gain. Convergence rates O(sqrt(p/(nT))) (nonconvex stationarity) and O(p/(nT)) (PL residual) are stated to follow directly from standard smoothness, bounded oracle variance, and network connectivity assumptions, matching known orders for distributed coordinate ZO methods without altering the leading term. No equations or steps in the provided abstract or description reduce a claimed prediction to a fitted input by construction, invoke self-citations as load-bearing uniqueness theorems, or smuggle ansatzes via prior author work. The central claim remains independent of the paper's own fitted quantities or self-referential definitions, qualifying as a normal non-circular theoretical derivation under external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The abstract invokes standard optimization assumptions without introducing new free parameters or invented entities beyond the powerball map itself; full ledger requires the manuscript.

axioms (1)
  • domain assumption Standard smoothness, oracle-variance, and network-connectivity assumptions hold.
    Invoked to derive the stated convergence rates.
invented entities (1)
  • fractional-power powerball map no independent evidence
    purpose: Nonlinear feedback gain on estimated gradient to amplify weak signals and attenuate large estimates.
    Introduced as the core mechanism distinguishing ZOOM-PB from prior methods.

pith-pipeline@v0.9.1-grok · 5746 in / 1323 out tokens · 33424 ms · 2026-06-29T21:07:23.015247+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references

  1. [1]

    Cooperative source seeking via networked multi-vehicle systems,

    Z. Li, K. You, and S. Song, “Cooperative source seeking via networked multi-vehicle systems,”Automatica, vol. 115, p. 108853, 2020

    2020

  2. [2]

    J. C. Spall,Introduction to stochastic search and optimization: esti- mation, simulation, and control. John Wiley & Sons, 2005, vol. 65

    2005

  3. [3]

    A. R. Conn, K. Scheinberg, and L. N. Vicente,Introduction to Derivative-Free Optimization. MPS-SIAM Series on Optimization. SIAM Philadelphia, 2009

    2009

  4. [4]

    Audet and W

    C. Audet and W. Hare,Derivative-Free and Blackbox Optimization. Springer, 2017

    2017

  5. [5]

    Derivative-free optimiza- tion methods,

    J. Larson, M. Menickelly, and S. M. Wild, “Derivative-free optimiza- tion methods,”Acta Numerica, vol. 28, pp. 287–404, 2019

    2019

  6. [6]

    Stochastic first-and zeroth-order methods for nonconvex stochastic programming,

    S. Ghadimi and G. Lan, “Stochastic first-and zeroth-order methods for nonconvex stochastic programming,”SIAM Journal on Optimization, vol. 23, no. 4, pp. 2341–2368, 2013

    2013

  7. [7]

    A compre- hensive linear speedup analysis for asynchronous stochastic parallel optimization from zeroth-order to first-order,

    X. Lian, H. Zhang, C.-J. Hsieh, Y . Huang, and J. Liu, “A compre- hensive linear speedup analysis for asynchronous stochastic parallel optimization from zeroth-order to first-order,” inAdvances in Neural Information Processing Systems, 2016, pp. 3054–3062

    2016

  8. [8]

    Random gradient-free minimization of convex functions,

    Y . Nesterov and V . Spokoiny, “Random gradient-free minimization of convex functions,”Foundations of Computational Mathematics, vol. 17, no. 2, pp. 527–566, 2017

    2017

  9. [9]

    Zeroth-order stochastic variance reduction for nonconvex optimiza- tion,

    S. Liu, B. Kailkhura, P.-Y . Chen, P. Ting, S. Chang, and L. Amini, “Zeroth-order stochastic variance reduction for nonconvex optimiza- tion,” inAdvances in Neural Information Processing Systems, 2018, pp. 3727–3737

    2018

  10. [10]

    Randomized gradient-free method for multia- gent optimization over time-varying networks,

    D. Yuan and D. W. Ho, “Randomized gradient-free method for multia- gent optimization over time-varying networks,”IEEE Transactions on Neural Networks and Learning Systems, vol. 26, no. 6, pp. 1342–1347, 2014

    2014

  11. [11]

    Gradient-free method for distributed multi- agent optimization via push-sum algorithms,

    D. Yuan, S. Xu, and J. Lu, “Gradient-free method for distributed multi- agent optimization via push-sum algorithms,”International Journal of Robust and Nonlinear Control, vol. 25, no. 10, pp. 1569–1580, 2015

    2015

  12. [12]

    Distributed zero-order algorithms for nonconvex multiagent optimization,

    Y . Tang, J. Zhang, and N. Li, “Distributed zero-order algorithms for nonconvex multiagent optimization,”IEEE Transactions on Control of Network Systems, vol. 8, no. 1, pp. 269–281, 2020

    2020

  13. [13]

    ZONE: Zeroth-order non- convex multiagent optimization over networks,

    D. Hajinezhad, M. Hong, and A. Garcia, “ZONE: Zeroth-order non- convex multiagent optimization over networks,”IEEE Transactions on Automatic Control, vol. 64, no. 10, pp. 3995–4010, 2019

    2019

  14. [14]

    Zeroth-order algo- rithms for stochastic distributed nonconvex optimization,

    X. Yi, S. Zhang, T. Yang, and K. H. Johansson, “Zeroth-order algo- rithms for stochastic distributed nonconvex optimization,”Automatica, vol. 142, p. 110353, 2022

    2022

  15. [15]

    Con- vergence analysis of nonconvex distributed stochastic zeroth-order coordinate method,

    S. Zhang, Y . Dong, D. Xie, L. Yao, C. P. Bailey, and S. Fu, “Con- vergence analysis of nonconvex distributed stochastic zeroth-order coordinate method,” inIEEE Conference on Decision and Control, 2021

    2021

  16. [16]

    Zeroth-order feedback optimization for cooperative multi-agent systems,

    Y . Tang, Z. Ren, and N. Li, “Zeroth-order feedback optimization for cooperative multi-agent systems,”Automatica, vol. 148, p. 110741, 2023

    2023

  17. [17]

    On the powerball method: Variants of descent methods for accelerated optimization,

    Y . Yuan, M. Li, J. Liu, and C. Tomlin, “On the powerball method: Variants of descent methods for accelerated optimization,”IEEE Control Systems Letters, vol. 3, no. 3, pp. 601–606, 2019

    2019

  18. [18]

    Accelerated zeroth-order algorithm for stochastic distributed non-convex optimization,

    S. Zhang and C. P. Bailey, “Accelerated zeroth-order algorithm for stochastic distributed non-convex optimization,” in2022 American Control Conference (ACC). IEEE, 2022, pp. 4274–4279

    2022