A mathematical optimization method comes with a strong promise on paper: for strongly convex problems with linear equality constraints, its continuous-time dynamics are reported to converge at a rate tied to the inverse square root of a time-scaling quantity the paper calls eta. Under a special scaling, the paper reports exponential convergence for objective error, constraint feasibility, distance from a solution and the motion of the solution trajectory. Both findings are conditional on stated assumptions and parameter choices.
The mathematical design
At its core, the method moves a candidate solution and a second set of variables associated with the linear constraints at the same time. The candidate-solution side has second-order, inertial motion, while the dual side follows a first-order rule. The Hessian is the paper's term for the curvature information that drives the damping, which the system handles implicitly.
Under its stated Lipschitz-gradient and local-integrability assumptions, the paper proves that the continuous system has a unique global strong solution from the listed initial conditions. In ordinary terms, the model has one well-defined trajectory that continues globally under those assumptions.
What the theory promises
Under its continuous-time parameter conditions, the analysis reports bounded trajectories and asymptotic orders for objective-value error, feasibility residual, distance from a solution and primal velocity. The measures cover both optimization accuracy and constraint satisfaction, along with how far the trajectory is from a solution and how quickly it is moving.
A special continuous-time scaling produces a stronger result: the paper reports exponential convergence for objective error, feasibility, trajectory and velocity. This is a special case tied to specified time-scaling and parameter assumptions, not a statement about arbitrary schedules.
The framework also reaches a nonsmooth convex case. For a proper lower-semicontinuous convex objective, the authors replace the smooth dynamics with a differential inclusion, a set-valued form of a differential equation, and extend the convergence analysis. The extension remains theoretical and conditional on its stated assumptions.
From dynamics to an algorithm
To turn the continuous design into an iterative method, the authors use implicit time discretization and derive an inertial accelerated primal-dual algorithm for the constrained problem. They state that, under mild discrete parameter conditions, the algorithm's convergence rate matches that of the continuous system. A special schedule with geometrically growing eta produces geometric convergence for objective, feasibility, solution-distance and successive-step quantities. In plain language, those quantities are predicted to shrink by a recurring proportion under that schedule.
The plots tell a narrower story
The first numerical experiment compared AL1 with ALALM-Xu and ALPDM in distributed logistic regression, with a maximum of 800 iterations. In the plotted comparisons, the authors report that AL1 had the best overall convergence. Case 2 performed best overall in practice, while Case 3 was theoretically fastest, but its rapidly growing eta schedule made subproblems difficult and degraded performance when solves were inexact.
A second experiment tested regularized least squares with a regularization parameter of 0.01. It used sigma values of 0.1 and 0.5, dimensions of 500 by 1,000 and 800 by 1,500, and maximum iteration counts of 1,000 and 1,500. The authors report that AL1 and IPAHD-SC achieved higher accuracy, faster convergence and less oscillation, while FISTA and AFBM were affected by sigma.
The results are plotted iteration-curve comparisons, not inferential statistical estimates. The analysis reports no confidence intervals or inferential tests, and no replication counts are reported for the least-squares comparisons. The evidence is therefore configuration-specific, showing how the methods behaved in the selected setups without resolving performance beyond them.
A result to read with care
The practical caveat is the gap between theorem and computation. The rate-matching result is theorem-based and assumption-dependent, while the logistic experiment reports that rapidly growing eta made subproblems difficult and that inexact solves degraded observed performance. A theoretical rate under stated conditions is therefore not the same as a guarantee for every approximate implementation.
Publication and disclosures
The manuscript is identified as arXiv:2608.25519v3, dated 29 Aug 2026, and its front matter leaves the received and accepted fields blank. It reports support from the Natural Science Foundation of Chongqing and the Team Building Project for Graduate Tutors in Chongqing. The authors say all generated or analysed data are included in the article and report no potential conflict of interest.
Paper data and sources
Original title: Inertial Primal-Dual Dynamics Methods Featuring Implicit Hessian-Driven Damping for Convex Optimization Problems in Continuous and Discrete Time
Authors: Xiangkai Sun, Zeying Gao, Liang He, Kok Lay Teo
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text