acceptodds
Under review as a conference paper at ICLR 2027

Parallelizable Interior-Point ADMM for Block-Structured Nonconvex Programs

Abstract

Physics is increasingly integrated into AI systems for learning, control, and predictive world modeling. This makes scalable numerical simulation an important computational primitive. Local–global optimization recasts physics simulation as a parallel nonconvex optimization problem by exploiting the splitting structure of physical models. However, existing convergence analyses of splitting algorithms do not match the empirical success of the local–global scheme. We close this gap by analyzing two variants of the ADMM algorithm, which we call Interior-Point ADMM (IP-ADMM), for block-structured composite programs with smooth nonlinear inequality and affine equality constraints. Our key idea is a one-sided finite-activation wall function that keeps the solution inside the feasible domain. As a result, the constraint blocks have closed-form update rules, each objective block requires one call to the original convex proximal oracle, and one equality-constrained quadratic solve per iteration performs the consensus update after the independent local updates. Ordinary IP-ADMM obtains a CQ-free best-iterate weak-KKT bound of , or a normalized banded Fritz–John bound of . Under Uniform Band-MFCQ, a sufficiently accurate banded Fritz–John certificate converts to KKT at a rate of . We further propose a checkpoint variant of IP-ADMM, called IP-ADMM-Ckpt, which attains a latest-checkpoint KKT bound at a rate of . Finally, we show how these upper bounds depend on the objective-block count , the constraint-block count , and the coupling norm , which reveals how the cost grows as we solve larger physical systems.

Primary area:
optimization
Keywords:
ADMM, Nonconvex constrained optimization, Physics Simulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_J7YnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IP-ADMM and a checkpoint variant for block-structured nonconvex optimization with nonlinear inequality constraints. It establishes finite-time guarantees for several stationarity certificates and evaluates the methods on small physics simulation problems.

Critical strengths

The method preserves independent local updates with closed-form constraint handling. The analysis explicitly distinguishes guarantees available without constraint qualifications from stronger KKT guarantees under Uniform Band-MFCQ, providing a useful theoretical treatment of constrained splitting methods.

Critical weaknesses

1, Insufficient relevance to ICLR. The central contribution concerns convergence analysis for general constrained optimization, making the paper appear more naturally suited to operations research or mathematical optimization. Although the introduction discusses differentiable simulation and world models, the experiments do not demonstrate an application to learning or establish a concrete consequence for machine learning. The relevance to the ICLR audience therefore remains insufficiently justified.

2, Limited connection between theory and practical performance. The main convergence experiments involve small 2D problems with only 8–24 objects or nodes. The practical parameters are not covered by the stated sufficient parameter construction, while Appendix C shows substantially worse finite-budget certificates with theorem-sufficient settings. Moreover, the KKT guarantees require an unverified Uniform Band-MFCQ condition, and neither variant reaches the KKT threshold on the larger main instances. The additional timing study shows only approximately 1.3× speedup and measures fixed sweeps rather than time to a common solution tolerance. These results leave the practical significance of the guarantees unclear.

Questions for authors

1, What concrete implication for a learning problem follows from the proposed guarantees, beyond the general motivation that simulation supports AI?

2, Can the guarantees be extended to the practical parameter settings, or can comparable practical performance be demonstrated with parameters satisfying the theoretical conditions?

Reviewer_j69hOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes IP-ADMM and a checkpoint variant for block-separable nonconvex programs. The method applies Moreau smoothing to nonsmooth objective terms, and uses a one-sided finite-activation wall for inequality blocks. The analysis gives CQ-free nearby-point weak-KKT and banded Fritz-John rates. Under Uniform Band-MFCQ, it gives KKT rates, with a faster post-checkpoint rate for IP-ADMM-Ckpt. Experiments cover capsule stacking, cloth resting, UAV trajectory planning, and a small CPU scaling study.

Critical strengths
  1. The local-global decomposition is well matched to the target problem class. Objective blocks use one proximal call, inequality blocks have closed-form updates, and the global step is a single quadratic consensus solve.
  2. The paper separates weak-KKT, banded Fritz-John, and KKT certificates. It also states explicitly where a constraint qualification is needed.
  3. The proof appendices expose dependence on the objective-block count, constraint-block count, and coupling norm. The numerical section includes exact-geometry checks and parameter sensitivity results.
Critical weaknesses
  1. The gap between theorem-sufficient and practical parameters is substantial. For cloth-8, Equation (28) gives approximately ρ0=4.7×107\rho_0 = 4.7 \times 10^7 and G=96G = 96, while the practical run uses ρ0=40\rho_0 = 40 and G=0.006G = 0.006.

  2. The baseline comparison is not fully comparable. Table 4 mixes SLSQP iteration counts with ADMM sweep counts. SLSQP has no per-iteration history, while IP-ADMM needs hundreds of thousands or millions of sweeps. BC-ADMM is unavailable for capsule stacking and is evaluated with its own fixed-margin residual.

  3. Objective values differ across methods; for capsule-16, SLSQP obtains −19.20-19.20 while IP-ADMM obtains −18.74-18.74 (Table 4). Larger instances also do not reach KKT, so the checkpoint rate is supported mainly by n=8n = 8 cases and one UAV example.

Questions for authors
  1. Can the authors give a practical parameter-selection rule that does not require the unavailable constants in Equation (28)?

  2. Can the authors estimate or verify the Uniform Band-MFCQ margins for the capsule and cloth instances?

  3. How often do surrogate-feasible solutions fail the exact segment-intersection checks? Could exact segment constraints be incorporated instead of relying on post-hoc validation?

Reviewer_FcTKOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes IP-ADMM and a checkpoint variant, IP-ADMM-Ckpt, for block-structured nonconvex composite optimization with nonlinear inequality constraints and affine coupling. Numerical evaluations cover capsule stacking, cloth resting, and a planar UAV trajectory-planning example

Critical strengths

NA

Critical weaknesses
  • The significance of the contribution for the ML/AI community is not sufficiently established. The introduction motivates the work through differentiable simulation, learned simulators, world models, and embodied AI. However, the reported applications are standalone static equilibrium and trajectory-optimization problems. Integration into learning pipelines and world models is explicitly left as future work. Consequently, the connection to AI remains primarily motivational rather than demonstrated by the proposed results. My concern is not that every optimization-theory paper must include neural-network experiments. A theoretical contribution can be valuable on its own. Here, however, the manuscript does not make a sufficiently compelling case for what important understanding or capability follows from the new guarantees. The general argument that physics simulation supports AI does not establish the significance of this particular solver or certificate hierarchy. Similarly, the existence of a previously unanalyzed combination of assumptions is not, by itself, a convincing significance argument. I would like a clearer explanation of which important limitation of existing approaches is resolved, and why resolving it matters. Table 2 compares the proposed variants with one another, but does not provide an assumption- and cost-aware comparison against prior methods that would make the magnitude of the theoretical advance easier to assess

  • The experiments do not establish a compelling practical advantage. The main convergence study contains two small, static, two-dimensional problem families with n∈{8,16,24}n\in\{8,16,24\}, using one deterministic run per family–size–method case. This is insufficient to assess robustness across initial configurations, geometries, and contact conditions. The issue is not the absence of repeated runs of an identical deterministic solver, but the narrow coverage of problem instances. The comparison with SLSQP also raises questions about practical utility. For example, on cloth-24, SLSQP terminates after 135 iterations and meets the reported KKT target, whereas IP-ADMM and IP-ADMM-Ckpt terminate after approximately 1.1 million and 2.2 million iterations, respectively, without meeting that target. These are different algorithms with different per-iteration costs, and they may reach different local solutions, so this does not establish that SLSQP is universally preferable. Nevertheless, it makes a matched wall-clock versus solution-quality comparison essential to the practical argument. Such an advantage is not demonstrated by the reported iteration counts.

  • What does the additional certification machinery enable that is consequential for the motivating applications? The manuscript establishes a formal distinction between guarantees, but does not convincingly establish the practical or scientific importance of that distinction.

Questions for authors

See above

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.