An arXiv preprint reports a two-level overlapping preconditioner for time-dependent calculations in high-contrast multiscale media. In reported numerical comparisons, full NLMC required fewer iterations than the listed competing coarse spaces, while the high-permeability component delivered comparable convergence with a much smaller coarse-space dimension in one comparison.
The proposed tool is a two-level overlapping preconditioner. The time-dependent problem is discretized with implicit Euler, which treats time implicitly for stability. The analysis asks whether the resulting conditioning stays independent of coefficient contrast and mesh sizes under stated assumptions.
The paper separates the NLMC coarse space into high-permeability and low-permeability components. The part built from high-permeability features, labelled VH,1 in the paper, is proposed for suitable small time steps. For general time steps, it proposes full NLMC or an augmented space that adds a standard multiscale component.
Benchmark comparisons
The first numerical test used a time-independent permeability field on a unit square. The fine mesh was 1/200 and the coarse mesh was 1/20. The time step was 0.1 for Case 1 and 0.002 for Case 2.
In Case 1, full NLMC reached convergence in about 23 iterations, versus about 26 for GMsFEM and more than 50 for the polynomial coarse space. In the Case 2 comparison, VH,1 converged in approximately 30 iterations, comparable to GMsFEM, while using a much smaller coarse-space dimension.
The same benchmark also tested how the NLMC basis, the functions used to build the coarse space, could be constructed. An iterative construction reached almost the same convergence as direct construction within 7 iterations.
A second test used a compressible-fluid-flow model with a time-dependent permeability field on a 64 m by 64 m domain. Its fine and coarse mesh settings were 1/200 and 1/10, and the reference solution used a 201-by-201 grid. The reported contrast settings were Cr 6, 7 and 8, against a background value of 1.
Full NLMC had the lowest listed iteration counts in that comparison: 22 at Cr 6, 23 at Cr 7 and 24 at Cr 8. GMsFEM required 32, 33 and 35 iterations; MsFEM required 50, 58 and 62; the polynomial space required 84, 111 and 127.
For the smaller-time-step comparison, the high-permeability component alone, VH,1, was reported to converge independently of contrast. It required slightly more iterations than Vgms, but was more robust than Vms and standard polynomial cases.
What the analysis promises
The theoretical results focus on the condition number, a measure used to describe how difficult a linear system is for an iterative method. In the relaxed basis-construction analysis, the paper reports an order-one bound that does not depend on coefficient contrast, the fine-mesh size or the coarse-mesh size, provided the stated auxiliary-space assumption holds.
A separate result narrows the claim to a small-step regime: when the time step is proportional to the square of the coarse-mesh size, the high-permeability component can handle the global correction while the reported condition-number bound remains independent of contrast. The formal result depends on the stated coefficient and interpolation assumptions.
For every positive time step, the general theorem uses an augmented coarse space that combines the high-permeability component with a standard multiscale space. It applies under stated interpolation estimates and overlap assumptions whose constants do not depend on contrast. In the paper's strategy, this augmented option or full NLMC is the general-time-step choice.
A smaller space, within limits
The method was also tested in three dimensions on a channelized-media benchmark. Full NLMC used 32, 27, 19 and 13 iterations at Cr values 4, 6, 8 and 10, while VH,1 used 45, 41, 30 and 21. MsFEM reached 216 iterations and GAMG 121, with GAMG failing at Cr 10.
That test showed the dimension trade-off behind the component strategy. VH,1 had 130 degrees of freedom, versus 1,130 for full NLMC and 1,209 for GMsFEM, while iteration counts stayed uniformly bounded across the listed coarse-mesh resolutions. The reported scalability and iteration results remain limited to those three-dimensional configurations.
The manuscript is an arXiv version-1 preprint dated 28 Aug 2026. Its acknowledgments report support from the Tianyuan Fund for Mathematics of the National Natural Science Foundation of China and from the NSFC, through grants 12526212 and 12301559.
Paper data and sources
Original title: A Two-Level Preconditioner Based on Dominant Components for Time-Dependent Multiscale High-Contrast Problem
Authors: Yating Wang, Yibao Li, Wing Tat Leung
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text