Derivative-Free Structured Updates for Muon
math.OC · cs.LG
Quick verdict
Proposes derivative-free Muon updates from structured finite differences; honest but toy-scale: V2 equals scaled coordinate FD, no theory, no code.
Summary
The paper proposes building Muon-style optimizer updates — the polar factor of a gradient-based momentum matrix — from function evaluations alone, using structured finite differences. Four variants are analyzed: entrywise coordinate recovery (mn+1 queries per iteration), random low-rank rank-one probing (r+1 queries), exhaustive basis-aligned probing (V2), and direct search over rank-one directions. The author proves V2 is exactly coordinate finite differences rescaled by h, so both give the same update in exact arithmetic. On small matrix-regression cases (8×8–16×16, single seed), exhaustive probing tracks full-gradient Muon at roughly 65× the query cost; four-probe random low-rank probing cuts evaluations 13–29× but leaves 23.7–31.3% recovery errors; direct search retains over 92% of the initial gap. Controlled noisy-gradient studies (30 realizations) locate crossovers where exact function values beat noisy gradients: entrywise DFO from σ=0.1, low-rank DFO between σ=1 and 3, and σ=0.1–0.3 for a small network. In five-seed CartPole with matched 48-episode budgets, orthogonal-probe DFO-Muon is the best zeroth-order method (1,305.6 episodes to solve) but trails REINFORCE (979.2). Key limitations: toy scale, single-seed regression tables, no convergence theory, no code release.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Muon-style updates (polar factors of a momentum matrix) can be constructed from function evaluations alone using structured rank-one finite-difference probes.
- Exhaustive basis-aligned probing (V2) is exactly coordinate finite differences scaled by h and yields the identical Muon direction because the polar factor is invariant to positive scaling.
- Random rank-one probing cuts function evaluations by 13-29x relative to exhaustive probing on matrix regression, at the cost of larger terminal recovery errors (23.7-31.3% versus 2.4-4.9% for exhaustive methods).
- When the gradient oracle is noisy, exact function values win: entrywise DFO beats noisy-gradient Muon from sigma=0.1, and low-rank DFO crosses over between sigma=1 and 3 (regression) and between 0.1 and 0.3 (network).
- On CartPole with matched 48-episode budgets, orthogonal-probe DFO-Muon is the best zeroth-order method (1,305.6 episodes to solve versus 2,707.2 for ZO-SGD) but trails REINFORCE policy gradient (979.2).
Strengths
Exact equivalence result clarifies the design space
Section 2.3 proves exhaustive V2 equals h times the coordinate finite-difference matrix (Eq. 2.4) and, via degree-zero homogeneity of the polar factor, produces identical update directions — replacing an apparent fourth method with an honest algebraic remark.
Controlled noise-crossover experiments
Sections 3.5-3.7.1 inject calibrated gradient noise (Eqs. 3.1/3.8) against exact function values over 30 realizations, and report crossovers as bounded intervals (sigma in (1,3) for regression, (0.1,0.3) for the network, (3,10) for CartPole) rather than universal thresholds.
Complete query accounting
Candor about limitations
Worked sparse example with exact query arithmetic
Weaknesses
Single-seed regression evidence
Toy scale throughout
Weak direct-search baseline and missing zeroth-order competitors
No code or data release
Low-rank surrogate is a heuristic with no probe-selection theory
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is derivative-free Muon?
It is a modification of the Muon optimizer that builds its update — the orthogonal polar factor of a momentum matrix — using only function evaluations from structured finite-difference probes, instead of gradients. This targets black-box objectives such as simulators or external solvers where derivatives are unavailable or unreliable.
Does derivative-free Muon beat gradient-based Muon?
Not when gradients are accurate and cheap: the paper shows full-gradient Muon achieves the lowest objective at zero noise. The derivative-free variants win only once the gradient oracle becomes noisy — entrywise finite differences already improve on mildly noisy gradients (sigma=0.1), and cheap low-rank probing wins at higher noise levels (crossover between sigma=1 and 3 in matrix regression).
How many function evaluations does derivative-free Muon need?
Entrywise or exhaustive basis-aligned probing costs mn+1 evaluations per iteration for an m-by-n matrix (13,001 total over 200 iterations for 8x8). Random low-rank probing uses only r+1 per iteration (e.g., 1,001 total with four probes), trading update accuracy for a 13-29x reduction in query count in the paper's experiments.
Is the V2 method different from coordinate finite differences?
Mathematically no. The paper proves exhaustive basis-aligned V2 is exactly the coordinate finite-difference matrix multiplied by the step size h, and because the polar factor is invariant to positive scaling, both methods produce the identical update direction in exact arithmetic.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures appear in the extracted text only as axis labels and legends; curve shapes, band widths, and plotted values could not be independently verified.
- Figure-dependent claims (e.g., terminal plateaus, crossover plot shapes) rest on the authors' textual descriptions rather than inspectable graphics.
- No supplementary materials or code repository are referenced, and data-generating distributions for the synthetic experiments are only partially specified in the text.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.