Pith. sign in

REVIEW 3 major objections 4 minor 13 references

The paper claims that the singularities of the Borel transform of Z(g) = ∫ e^{-L(θ)/g} dθ sit exactly at the critical values of L, turning a divergent series into a global map of the optimization landscape.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

SURGE proposes to find critical loss values from Borel singularities of a partition function and use them to scale learning rates during gradient descent.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Interesting idea, but the central theorem is wrong and the paper's own quartic example contradicts it. the 3 major comments →

arxiv 2509.01329 v1 pith:BHPITCZG submitted 2025-09-01 cs.LG math-phmath.MPmath.OC

Globally aware optimization with resurgence

Wei Bu This is my paper
classification cs.LG math-phmath.MPmath.OC
keywords resurgence theoryBorel transformdivergent asymptotic seriespartition functionnon-convex optimizationcritical pointslearning rate adaptationglobal optimization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 proposes that the divergent small-coupling expansion of the parameter-space partition function Z(g) = ∫ e^{-L(θ)/g} dθ encodes the full set of critical values of the objective function L. Its central result is a one-to-one correspondence: the singularities of the Borel transform of this expansion—obtained by dividing each coefficient by n! so that the divergent series becomes a convergent one—sit exactly at L(θ*) for every critical point θ* where ∇L(θ*) = 0. If the correspondence holds, a single analysis phase, estimating Z(g) for small g, extracting its asymptotic coefficients, and locating Borel-plane singularities, would hand a local optimizer the list of critical loss values, which the optimizer can use to decide when a lower value is nearby and scale its learning rate accordingly. The paper packages this into a wrapper algorithm (SURGE) tested on function fitting, MNIST, and a small character-level transformer, reporting consistent loss improvements over the underlying SGD, Adam, AdamW, and Muon optimizers.

Core claim

The paper claims a correspondence theorem: for Z(g) = ∫ e^{-L(θ)/g} dθ, the Borel transform of its small-g asymptotic series has singularities exactly at critical values L(θ*) where ∇L(θ*)=0. The co-area-formula argument (Appendix C) rewrites the Laplace representation as a level-set integral, B[Z](t) = ∫_{t=L(x)} dσ(x)/|∇L(x)|, so singularities occur precisely at ∇L=0. A quartic-oscillator example confirms the predicted singularities. The correspondence drives a two-phase algorithm: first estimate Z(g) at small couplings once, then run any gradient optimizer with a learning-rate scale factor that shrinks as the current loss approaches the nearest target below it.

What carries the argument

The key object is the parameter-space partition function Z(g) = ∫ e^{-L(θ)/g} dθ and its factorially divergent asymptotic series Z(g) ∼ Σ a_n g^n. The Borel transform B[Z](ζ) = Σ (a_n/n!) ζ^n converts the divergent series into a convergent one, and the paper's central identity—the co-area/level-set formula B[Z](t) = ∫_{t=L(x)} dσ(x)/|∇L(x)|—shows that the Borel transform's singularities occur exactly where |∇L(x)| vanishes. Positive-real singularities therefore give the objective values at stationary points; the algorithm reads those values as targets and uses them to scale the learning rate during gradient descent.

Load-bearing premise

The whole method rests on the assumption that a low-order polynomial fit to numerically estimated values of Z(g) at small but nonzero g recovers the true asymptotic coefficients of the divergent expansion, so the detected Borel-plane singularities are genuine rather than artifacts of the fit.

What would settle it

Compute the small-g asymptotic coefficients of Z(g) = ∫ e^{-(x²−1)²/g} dx to high order, form the Borel transform, and locate its positive-real singularities by analytic continuation (e.g., Borel–Padé). If the singularities do not approach the known critical values 0 and 1 as the number of coefficients grows, the claimed correspondence fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the correspondence is exact, the critical values of the objective function can be read off from the positive real axis of a one-dimensional Borel plane, independent of the dimension of the parameter space.
  • Any gradient-based optimizer—SGD, Adam, AdamW, Muon—can be wrapped so that its learning rate is scaled by a factor set by the gap between the current loss and the nearest target critical value.
  • If Borel analysis fails to produce targets, SURGE reduces to the underlying optimizer, so the global guidance is an additive improvement rather than a replacement of local search.
  • The correspondence reframes the factorial divergence of perturbative coefficients as the carrier of non-perturbative information, as demonstrated in the paper's trans-series treatment of the quartic oscillator.
  • Because the targets are objective-function values rather than parameter locations, the optimizer is never asked to solve the NP-hard preimage problem; it still follows local gradients, but with knowledge of whether a better critical value lies below the current loss.

Where Pith is reading between the lines

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

  • The paper does not test whether finite-order least-squares fits to Z(g) at finite g recover the true asymptotic coefficients; a decisive experiment is to run the full pipeline on a potential with known critical values and check whether detected singularities converge to those values as the polynomial order and the number of coupling samples increase.
  • An ablation that replaces the Borel-derived targets with arbitrary decreasing values could separate the contribution of the true targets from the generic effect of periodically scaling up the learning rate; the paper reports no such ablation.
  • The correspondence mechanism is generic to Laplace-type integrals, so if it holds it should transfer beyond neural-network losses—for example to Bayesian evidence or free-energy integrals ∫ e^{-S(x)/g} dx, where the Borel singularities would locate critical values of the action S.
  • The ratio-test and thresholded-evaluation methods used for singularity detection can only see the dominant features of the fitted polynomial; monitoring how detected singularities shift with polynomial order J would provide a practical reliability check for the targets.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes SURGE, a framework that computes the statistical-mechanical partition function Z(g)=∫e^{-L(θ)/g}dθ for a neural-network objective L, extracts its small-coupling asymptotic coefficients, and identifies singularities of the Borel transform as predicted critical objective values. These targets are then used to scale the learning rate of an arbitrary base optimizer. The central assertion is a one-to-one correspondence between Borel singularities and critical values of the objective landscape (Theorem 3 / Proposition 2). The manuscript contains an analytic quartic example, a description of a numerical pipeline for extracting targets, and experiments on synthetic regression, MNIST, and a small transformer.

Significance. If the central correspondence were correct, the paper would offer a genuinely novel way to obtain global landscape information from local perturbative data, with direct algorithmic value. The author is clearly familiar with the relevant resurgence literature, and the explicit algorithms are a useful starting point. However, the main mathematical claim is not established: the proof in Appendix C is algebraically incorrect, and the paper's own analytic example in Appendix D contradicts the claimed correspondence. The numerical pipeline in Algorithm 3 is also circular with respect to the core claim because the targets are defined as singularities of the very fitted object that is meant to be validated. In its current form, the paper does not provide a sound theoretical foundation, and the experiments do not independently test the geometric prediction.

major comments (3)
  1. [Appendix C, Eqs. (49)–(51)] The proof of Theorem 3 is algebraically invalid. The co-area formula (Eq. (45)) gives ∫δ(t−S(x)) dx = ∫_{t=S(x)} dσ/|∇S|, with no factor of g. Substituting into Eq. (49) gives f(g)=∫_0^∞ dt e^{−t/g} ∫_{t=S(x)} dσ/|∇S|. Comparing with Eq. (47), B[f](t) equals the level-set integral without a prefactor g; Eq. (51) inserts a spurious g, and with that factor the Laplace inversion in Eq. (50) no longer recovers f(g). Moreover, the conclusion that ∇S=0 produces singularities of the level-set density is not generic: for a smooth S in dimension d>1, ρ(t)=∫_{t=S} dσ/|∇S| can be continuous at a critical value. Thus Theorem 3 and Proposition 2 are not established.
  2. [Appendix D, Eqs. (69)–(73)] The analytic example contradicts Proposition 2. For V(x)=x^2+x^4, the critical points are x=0 and x=±i/√2, with critical values 0 and −1/4; no real critical value is 1.0578. Yet Appendix D reports the dominant singularity as ζ1=2^{5/3}/3≈1.0578 (Eq. (73)), which is neither. The Borel coefficients b_k=(−1)^k(4k)!/(4^k(k!)^3) have |b_{k+1}/b_k|∼256k, so the Borel series has zero radius of convergence; identifying it with the entire function 0F3 (Eq. (69)) cannot be correct. Equations (70) and (72) give mutually inconsistent singularity locations. This example therefore disproves the claimed correspondence rather than supporting it.
  3. [Section 3, Algorithm 3 and Eq. (32)] The numerical pipeline is circular with respect to the central claim. The target set T is defined in Eq. (32) as singularities of the Borel transform of the fitted series; the coefficients are obtained by least-squares fitting a degree-J polynomial to estimated Z(g_s) values (Eq. (30), Algorithm 3 Steps 4–5), and Z(g_s) itself is estimated via the variational inequality in Eq. (28). No theorem or numerical evidence shows that this finite-g, finite-order fit recovers the true small-g asymptotic coefficients, nor that singularities of the fitted polynomial coincide with analytic singularities of the true Borel transform. The predicted targets are never checked against independently known critical values. The experimental gains in Section 4 are consistent with a learning-rate schedule heuristic, not with the geometric claim; the limitations acknowledged in Section 4.2 (instability, accelera
minor comments (4)
  1. [Abstract and Section 1.2] Grammatical and typographical issues: 'singularities ... corresponds' should be 'correspond'; 'Escalle' should be 'Écalle'.
  2. [Section 2.1, Eq. (11)] The denominator in the residue calculation is written as 1+ζ; it should be 1−ζ to be consistent with Eq. (7) and the preceding Borel transform.
  3. [Appendix D, Eq. (61)] The coefficient formula is internally inconsistent with the displayed values: substituting k=1 into Eq. (61) gives 6, not −3/8 as reported in Eq. (63); the Gaussian-moment derivation in Eqs. (58)–(59) also yields a different expression. Please reconcile.
  4. [Section 4] No sensitivity analysis is provided for the free parameters that determine the target set: the resurgence weight λ, polynomial order J, coupling range (gmin,gmax), singularity threshold τ, and number of samples. Figures 1–3 also lack error bars or multiple-seed information.

Circularity Check

2 steps flagged

The central 'critical value prediction' is the output of a fit renamed by definition: targets are defined as singularities of a Borel-transformed polynomial fitted to estimated Z(g), and then called objective values at critical points.

specific steps
  1. self definitional [Section 3.1, Definition 3 (Eq. 32) and following paragraph]
    "Definition 3 (Critical objective function Targets). The set of critical objective targets is defined as: T = {ζ ∈ S(B[Z]) : ζ ∈ R+, ζ < L0} where S(B[Z]) denotes the set of singularities of the Borel transform. ... Given a set of objective function values at critical points {ζm}M m=1,..."

    The target set is defined to be exactly the set of singularities of the Borel transform. A few sentences later the same ζm are called 'objective function values at critical points'. So the claimed output—critical objective values—is not independently computed or verified; it is, by construction, whatever singularities the Borel analysis returns. Any spurious singularity of the fitted Borel object automatically qualifies as a target, and the paper's experiments never compare returned ζm against actual critical values ∇L(θ*)=0. The prediction reduces to the definition of T.

  2. fitted input called prediction [Section 3, Algorithm 3, Steps 4-6 (Eqs. 30-31, 32); Section 3 bullet list]
    "Fit Z(g) to a power series in g: Z(g) = a0 + a1g + a2g2 + · · ·+ aJ gJ up to order J. ... Given {aj}J j=0 the coefficients of the power series (non-vanishing and factorially divergent), find Borel singularity {ζm}M m=1 in the function PJ j=0(aj/j!)ζ j on the positive real line. ... Select top M singularities: {ζm}M m=1 ⊂ S; return {ζm}M m=1 as optimization target values."

    The 'global targets' are the singularities of a finite polynomial whose coefficients are least-squares fits to estimated partition function values (which are themselves Monte Carlo/lower-bound estimates of integrals of the same objective L). No analytic continuation is performed, and no critical point or critical value is ever evaluated independently. The returned {ζm} are therefore forced by the fitted polynomial: with a different fit order, coupling range, or threshold, the targets change. Calling these fitted singularities 'critical objective function values' is a fitted input renamed as a prediction.

full rationale

The main circularity is in the target-construction pipeline. Algorithm 3 and Definition 3 define the optimization targets as singularities of a Borel-transformed polynomial fitted to estimated Z(g) values, and Section 3.1 then refers to those same singularities as 'objective function values at critical points.' Since no independent check against actual critical values is made, the claimed global prediction reduces to the fitted quantity by construction. Separate mathematical errors—the invalid co-area step in Appendix C (Eq. 51 contains a spurious factor g, so the Laplace inversion in Eq. 50 does not recover f(g)) and Appendix D's dominant singularity ζ1≈1.0578, which is not a critical value of V=x²+x⁴—are correctness failures rather than circularity, but they remove the only independent support that could have made the fitted singularities meaningful. There is no load-bearing self-citation chain, so the score reflects partial, construction-level circularity rather than a fully self-referential derivation.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The method rests on free parameters chosen without specification, on assumptions that the partition function and its Borel transform behave like textbook resurgent integrals, and on the unverified identification of fitted polynomial features with true asymptotic data. No new physical entities are introduced.

free parameters (5)
  • resurgence weight lambda = not specified
    Learning-rate scaling factor in Algorithm 2, eq. 35; chosen by hand.
  • polynomial order J = not specified
    Order of power-series fit in Algorithm 3, Step 4; not stated.
  • coupling range (gmin, gmax) = selected by Algorithm 1
    Chosen among candidate ranges based on an undefined success rate and score, introducing selection bias.
  • singularity threshold tau = not specified
    Used to declare Borel singularities in Algorithm 3, Step 6.
  • series coefficients a_j = fitted to Z(g) values
    Obtained by weighted least-squares (eq. 30); the alleged asymptotic coefficients are fit, not derived from a saddle-point expansion.
axioms (4)
  • domain assumption The partition function Z(g) = ∫ e^{-L(θ)/g} dθ is finite and has an asymptotic expansion as g -> 0+.
    Used throughout Section 2.3; not proven for neural network losses on unbounded parameter space and may diverge if L is unbounded below.
  • domain assumption The Borel transform of the asymptotic series can be analytically continued with singularities only on the positive real axis.
    Invoked to define targets via eq. (32); not established for this setting.
  • standard math The coarea formula implies that singularities of the level-set integral ∫_{S=t} dσ/|∇S| occur exactly at critical values of S, and that this integral is the Borel transform of the asymptotic series.
    Used in Appendix C, eq. (51); the displayed formula carries a spurious factor g, and the conclusion conflates an exact Laplace representation with the asymptotic Borel transform.
  • ad hoc to paper Fitted polynomial coefficients a_j reproduce the factorially divergent asymptotic coefficients of Z(g).
    Algorithm 3 Steps 4-5 assumes this without justification; finite-g least-squares fits do not yield asymptotic coefficients, and the detected singularities of a polynomial are not true analytic singularities.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Globally aware optimization with resurgence." pith.science (2026). https://pith.science/paper/BHPITCZG

@misc{pith2026250901329,
  author       = {Pith},
  title        = {Pith review of: Globally aware optimization with resurgence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BHPITCZG}},
  note         = {Machine review of arXiv:2509.01329}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Modern optimization faces a fundamental challenge: local gradient-based methods provide no global information about the objective function $L$ landscape, often leading to suboptimal convergence and sensitivity to initialization. We introduce a novel optimization framework that leverages resurgence theory from complex analysis to extract global structural information from divergent asymptotic series. Our key insight is that the factorially divergent perturbative expansions of parameter space partition functions encode precise information about all critical objective function value in the landscape through their Borel transform singularities. The algorithm works by computing the statistical mechanical partition function $Z(g) = \int e^{-L(\theta)/g} d\theta$ for small coupling $g\ll 1$, extracting its asymptotic series coefficients, and identifying Borel plane singularities that correspond one-to-one with critical objective function values. These target values provide global guidance to local optimizers, enabling principled learning rate adaptation and escape from suboptimal regions. Unlike heuristic adaptive methods, targets are theoretically grounded in the geometry of the optimization landscape.

Figures

Figures reproduced from arXiv: 2509.01329 by Wei Bu.

Figure 1
Figure 1. Figure 1: 1d regression with 2 layered-MLP We evaluate the algorithm on a simple fully connected network sized (12, 10, 8) fitting some 1d function: f(x) = sin(2x) + 0.5 cos(5x) + 0.3 sin(10x) + 0.1x 2 (36) 6This is a rather crude usage of the global targets, as a first proof of concepts. We expect there to be much more profound ways of using these global targets. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MLP training on standard MNIST classification task. [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Small transformer training on standard Shakespear dataset. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

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

13 extracted references · 4 canonical work pages · 1 internal anchor

  1. [7]

    Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein

    URL https://arxiv.org/abs/1902.04811. Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks,

    Pith/arXiv arXiv 1902

  2. [9]

    doi: 10.1002/prop

    ISSN 1521-3978. doi: 10.1002/prop. 201400005. URL http://dx.doi.org/10.1002/prop.201400005. Herbert Robbins and Sutton Monro. A Stochastic Approximation Method. The Annals of Math- ematical Statistics, 22(3):400 – 407,

    doi:10.1002/prop

  3. [12]

    Jean ´Ecalle

    URL https://proceedings.neurips.cc/paper_ files/paper/2018/file/90365351ccc7437a1309dc64e4db32a3-Paper.pdf. Jean ´Ecalle. Les fonctions r ´esurgentes, Vols. I–III. Publications Math ´ematiques d’Orsay,

    arXiv 2018

  4. [1951]

    URL https: //doi.org/10.1214/aoms/1177729586

    doi: 10.1214/aoms/1177729586. URL https: //doi.org/10.1214/aoms/1177729586. Wei Shen, Ruichuan Huang, Minhui Huang, Cong Shen, and Jiawei Zhang. On the convergence analysis of muon,

    arXiv

  5. [1981]

    Foundational work introducing resurgence theory and alien calculus. 13 A I NEVITABLE DIVERGENCE OF PERTURBATIVE POWER SERIES The coefficients Zr(g) encode information about the geometry of the objective function landscape, particularly near critical points. This series is naturally asymptotic as we shall show it. According to the analysis done in Berry & ...

    work page 1991

  6. [1991]

    URL https://royalsocietypublishing.org/ doi/abs/10.1098/rspa.1991.0119

    doi: 10.1098/rspa.1991.0119. URL https://royalsocietypublishing.org/ doi/abs/10.1098/rspa.1991.0119. Arindam Bhattacharya, Jordan Cotler, Aur´elien Dersy, and Matthew D. Schwartz. Renormalons as Saddle Points. 10

    work page arXiv 1991

  7. [2014]

    Soham De, Anirbit Mukherjee, and Enayat Ullah

    URL https://arxiv.org/abs/1406.2572. Soham De, Anirbit Mukherjee, and Enayat Ullah. Convergence guarantees for rmsprop and adam in non-convex optimization and an empirical comparison to nesterov acceleration,

    Pith/arXiv arXiv

  8. [2017]

    URL https://arxiv.org/abs/1412.6980. M. Mari˜no. Lectures on non-perturbative effects in large n gauge theories, matrix models and strings. Fortschritte der Physik, 62(5–6):455–540, April

    Pith/arXiv arXiv

  9. [2018]

    Eric Delabaere and Fr´ed´eric Pham

    URL https://arxiv.org/abs/1807.06766. Eric Delabaere and Fr´ed´eric Pham. Resurgent methods in semi-classical asymptotics. Annales de l’I.H.P . Physique th´eorique, 71(1):1–94,

    Pith/arXiv arXiv

  10. [2019]

    Accessible survey for physicists covering resurgent analysis and alien calculus

    doi: 10.1016/j.aop.2019.167914. Accessible survey for physicists covering resurgent analysis and alien calculus. Greg B Fotopoulos, Paul Popovich, and Nicholas Hall Papadopoulos. Review non-convex optimiza- tion method for machine learning,

    work page arXiv 2019

  11. [2023]

    Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio

    URL https://arxiv.org/abs/2310.12317. Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, and Yoshua Bengio. Identifying and attacking the saddle point problem in high-dimensional non-convex optimization,

    Pith/arXiv arXiv

  12. [2024]

    Review Non-convex Optimization Method for Machine Learning

    URL https://arxiv.org/abs/2410.02017. Chi Jin, Praneeth Netrapalli, Rong Ge, Sham M. Kakade, and Michael I. Jordan. On nonconvex optimization for machine learning: Gradients, stochasticity, and saddle points,

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [2025]

    12 Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar

    URL https://arxiv.org/abs/2505.23737. 12 Manzil Zaheer, Sashank Reddi, Devendra Sachan, Satyen Kale, and Sanjiv Kumar. Adaptive methods for nonconvex optimization. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa- Bianchi, and R. Garnett (eds.), Advances in Neural Information Processing Systems, volume

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.