REVIEW 4 major objections 2 minor 2 cited by
LeAP-SSN: A Semismooth Newton Method with Global Convergence Rates
T0 review · 4 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read A single parameter-free algorithm claims to guarantee global convergence on convex and nonconvex problems and then switch to superlinear local speed, all in Hilbert spaces.
desk verdict A genuinely interesting claimed unification—parameter-free Levenberg–Marquardt proximal semismooth Newton with global rates and superlinear asymptotics in Hilbert space—but the proof of how the globalization feeds the fast local phase is the thing to read before believing it. 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 central object is the LeAP-SSN iteration: a proximal semismooth Newton step regularized by an adaptive Levenberg-Marquardt parameter, combined with backtracking. The adaptive regularization keeps the Newton step well-defined and ensures a sufficient decrease without problem-specific constants, while the semismooth Newton term provides the fast local asymptotics. The Polyak-Lojasiewicz branch uses only function values and gradients, so the unknown PL constant never needs to be estimated.
What would settle it
Construct a semismooth nonconvex problem on a Hilbert space with a continuum of minimizers, choose an initial point where the backtracking line search initially stalls or takes tiny steps, and check whether the subgradient norm obeys the claimed O(1/sqrt(k)) bound and whether a superlinear phase eventually appears; a single counterexample that fails either global or local behavior would refute the central claim.
Extended reading notes
Core claim
LeAP-SSN is a proximal semismooth Newton method whose Newton steps are stabilized by an adaptive Levenberg-Marquardt regularization and a backtracking line search. The paper claims this combination is globally convergent in Hilbert spaces from arbitrary initial points, with no knowledge of constants such as the PL parameter. On convex problems, the objective gap decays as O(1/k); on nonconvex problems, the subgradient norm decays as O(1/sqrt(k)); under a Polyak-Lojasiewicz condition the convergence is linear. Under additional semismoothness and either a Dennis-Moré or partial smoothness condition at the limit point, the same algorithm achieves superlinear convergence even when the minimizer
Load-bearing premise
The superlinear local stage and even the existence of the iterates rely on semismoothness plus Dennis-Moré or partial smoothness at the limit point, and on the proximal Newton subproblems being solvable; the global rates alone do not depend on these, but the flagship speed result does.
Editorial extensions
If this is right
- Users of nonsmooth optimization get a single method that requires no tuning yet matches the worst-case guarantees of first-order methods.
- Convex and nonconvex problems in Hilbert spaces, including inverse problems and PDE-constrained settings, share a unified convergence theory.
- Under a Polyak-Lojasiewicz condition, linear convergence is achieved without knowing the PL constant.
- Superlinear convergence for non-isolated minimizers broadens the class of problems where Newton-type fast asymptotics apply.
- The combination of global rates and local superlinear speed may remove the need for separate global and local optimization phases.
Reading between the lines
- If the claims hold, similar adaptive regularized Newton schemes could be developed for stochastic or inexact settings where only noisy subgradients are available—the paper does not address noise.
- The O(1/sqrt(k)) nonconvex rate matches the classic gradient descent rate, suggesting that second-order information can be added without sacrificing the worst-case global rate.
- The non-isolated minimizer superlinear result points toward degenerate and rank-deficient problems, such as low-rank matrix recovery, though the authors do not mention this.
- A concrete test: evaluate LeAP-SSN on standard imaging or sparse logistic regression benchmarks and measure whether the superlinear phase appears near non-isolated solutions in practice.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes LeAP-SSN (Levenberg-Marquardt Adaptive Proximal Semismooth Newton method), a parameter-free optimization algorithm for Hilbert spaces. The abstract claims global nonasymptotic rates: O(1/k) for convex objectives, O(1/sqrt(k)) for nonconvex subgradient stationarity, linear convergence under a Polyak-Lojasiewicz condition, and superlinear local convergence under semismoothness and Dennis-More or partial smoothness conditions, all without knowledge of problem-specific constants. The material provided for review consists of the abstract only; no full text, theorem statements, proofs, or numerical experiments are available.
Significance. If the full manuscript substantiates the abstract, the contribution is significant: a single parameter-agnostic algorithm that simultaneously provides worst-case global rates and superlinear local asymptotics in Hilbert spaces, including for non-isolated minimizers, would be a notable advance in nonsmooth and large-scale optimization. The emphasis on avoiding problem-specific constants is attractive and, if proven, would be a practical strength. At this stage, however, these are unverified claims; the absence of precise assumptions and derivations makes the significance untestable from the submitted material.
major comments (4)
- [Abstract - global rates] The claim of convergence from 'arbitrary starting points' with O(1/sqrt(k)) subgradient rates in Hilbert spaces requires well-posedness conditions that are not stated: the objective must be bounded below, a solution or stationary point must exist, the proximal Newton subproblems must be solvable, backtracking stepsizes must be well defined, and boundedness of subgradients along iterates or bounded level sets is typically needed. Without these hypotheses, iterates may fail to exist or the rate estimate may be vacuous. The theorem statements must list these conditions explicitly.
- [Abstract - superlinear convergence] The superlinear claim is conditional in a way that may undermine the advertised bridge from global to fast local convergence. In standard semismooth Newton analysis, the Dennis-More condition is a property of the generated step sequence, not an assumption on the problem. If the theorem merely assumes this condition on the iterates, then the algorithm is not shown to achieve superlinear convergence from arbitrary starting points. The abstract must clarify whether the condition is proved from the Levenberg-Marquardt/backtracking globalization or is an additional algorithmic assumption, and if proved, for which problem classes.
- [Abstract - Polyak-Lojasiewicz condition] Linear convergence under a Polyak-Lojasiewicz condition typically involves an unknown constant. The statement that the method 'does not require knowledge of problem-specific constants' needs clarification: the rate constant may depend on the PL constant, which is acceptable as long as the algorithm itself does not need to know it. The proof of parameter-agnostic linear convergence is delicate and is not visible in the abstract. The theorem must state the sense in which the method avoids this constant.
- [Abstract - reviewability] Only the abstract was provided for review. No definitions, assumptions, proofs, or numerical results are available, so none of the central claims can be verified. In particular, the combination of global and local rates is the paper's main selling point, but no theorem statements or derivations are included to support it. A complete manuscript is required for a substantive assessment.
minor comments (2)
- [Abstract - terminology] The acronym LeAP-SSN is defined, but the phrase 'mild semismoothness' is vague; it would help to specify whether semismoothness applies to the residual map, the objective gradient, or a generalized Jacobian selection.
- [Abstract - experiments] The claim that practical efficiency is illustrated on imaging, contact mechanics, and machine learning is not supported in the abstract; brief problem descriptions or a pointer to the experimental section would improve clarity.
Circularity Check
No circularity found: the abstract states assumptions and rates without fitting parameters or importing conclusions from self-citations.
full rationale
The paper is abstract-only, so no derivation chain or equations are available to inspect. The abstract asserts global nonasymptotic rates for convex (O(1/k)), nonconvex (O(1/sqrt(k))), and PL (linear) cases, and superlinear convergence under explicit conditions (semismoothness, Dennis-Moré, partial smoothness). These are stated as assumptions or theorem conclusions, not as fitted inputs renamed as predictions. There are no problem-specific constants that are learned from data, no self-citations that carry the argument, and no definitional equivalence between an output and an input. The Dennis-Moré condition, while potentially an algorithmic condition on iterates rather than a problem assumption, is presented as a condition under which superlinear convergence is claimed; whether the algorithm provably satisfies it from arbitrary starting points is a correctness question, not a circularity. The abstract does not claim to predict something from a model that was built using that same thing. Therefore, there is no circular step to exhibit.
Assumptions & free parameters
assumptions (4)
- domain assumption The optimization problem is posed in a Hilbert space with the stated smoothness and subdifferential structure for convex, nonconvex, and nonsmooth cases.
- domain assumption The Polyak-Lojasiewicz condition holds for the global-minimum and linear-rate claims.
- domain assumption Superlinear convergence requires semismoothness and Dennis-Moré or partial smoothness conditions at the limit point.
- standard math Standard nonsmooth analysis and semismooth Newton machinery are used without derivation.
Cite this review
Pith. "Pith review of LeAP-SSN: A Semismooth Newton Method with Global Convergence Rates." pith.science (2026). https://pith.science/paper/QGYI5LSI
@misc{pith2026250816468,
author = {Pith},
title = {Pith review of: LeAP-SSN: A Semismooth Newton Method with Global Convergence Rates},
year = {2026},
howpublished = {\url{https://pith.science/paper/QGYI5LSI}},
note = {Machine review of arXiv:2508.16468}
}
abstract
We propose LeAP-SSN (Levenberg--Marquardt Adaptive Proximal Semismooth Newton method), a semismooth Newton-type method with a simple, parameter-free globalisation strategy that guarantees convergence from arbitrary starting points in nonconvex settings to stationary points, and under a Polyak--Lojasiewicz condition, to a global minimum, in Hilbert spaces. The method employs an adaptive Levenberg--Marquardt regularisation for the Newton steps, combined with backtracking, and does not require knowledge of problem-specific constants. We establish global nonasymptotic rates: $\mathcal{O}(1/k)$ for convex problems in terms of objective values, $\mathcal{O}(1/\sqrt{k})$ under nonconvexity in terms of subgradients, and linear convergence under a Polyak--Lojasiewicz condition. The algorithm achieves superlinear convergence under mild semismoothness and Dennis--Mor\'e or partial smoothness conditions, even for non-isolated minimisers. By combining strong global guarantees with superlinear local rates in a fully parameter-agnostic framework, LeAP-SSN bridges the gap between globally convergent algorithms and the fast asymptotics of Newton's method. The practical efficiency of the method is illustrated on representative problems from imaging, contact mechanics, and machine learning.
Forward citations
Cited by 2 Pith papers
-
Mesh-dependent iteration count growth in primal-dual active set strategies
Primal-dual active set methods for obstacle problems suffer exponentially growing iteration counts under mesh refinement due to layer-by-layer active-set peeling, while Signorini-type problems grow only linearly.
-
On the Universality of Simple Trust-Region Algorithms
Classical and modified-ratio trust-region methods reach the optimal O(ε^{-1/(1+ν)}) convex and O(ε^{-(2+ν)/(1+ν)}) nonconvex complexity for any Hölder ν∈[0,1] without knowing ν.
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.