Multilevel stochastic gradient descent (MLSGD) reached comparable accuracy 18 times faster than batched stochastic gradient descent in a reported test of optimal control for an uncertain partial differential equation, or PDE. At the same computational cost, it reduced the error by a factor of five. The reported convergence rate was about 0.5 for the multilevel method, compared with about 0.37 for batched SGD.
The comparison was made under a fixed setup
The comparison used 64 processing units. Batched SGD ran for 150 iterations with a mesh resolution of 2^-7 and a batch size of 256. The analysis also reports fewer iterations, lower bias — systematic error in an estimate — and lower variance for the multilevel method.
The result is a numerical comparison between the two methods under the stated configuration. It does not establish that MLSGD will outperform other methods on different hardware, budgets or PDE systems.
The study tests two kinds of uncertainty calculation
The study asks whether resource-aware multilevel methods can support forward uncertainty quantification and optimal control for PDEs while working within limited CPU-time and memory resources. Forward uncertainty quantification estimates uncertain model outputs when inputs are uncertain; optimal control seeks a control for a PDE under uncertainty.
For the forward problem, the budgeted multilevel Monte Carlo method is designed to compute full-field estimates of PDE systems under uncertainty. Its model combines Gaussian random-field sampling with subsurface diffusion and a hyperbolic transport PDE system.
For the control problem, the method targets a non-time-dependent elliptic diffusion equation with log-normal coefficients and homogeneous Dirichlet boundary conditions. The numerical instance used a square domain, a sinusoidal target state and a cost factor of 10^-8.
Computing limits are built into the calculation
The study treats error minimization as a knapsack problem: selecting computational work under CPU-time and memory budgets. It solves the resulting allocation problems with distributed dynamic programming, described as a parallel recursive algorithm.
The paper also states that the minimum reached by the MLSGD optimization is bounded by the imposed computational constraints. The supplied analysis does not provide complete numerical estimates for the exponents underlying that bound.
Full-field estimation stayed just inside the test budget
In a hyperbolic transport test, four nodes using 256 CPUs ran for one hour under imposed limits. Updating the full field used nearly the same memory and CPU time as scalar estimation, and total resource use remained just below the limits.
That finding applies to the reported configuration. The supplied analysis reports no confidence interval or repeated-run variability for the result.
Parallel efficiency diminished at larger scale
The main numerical experiments ran on the CPU partition of the HoreKa system with between 64 and 1,024 processing units and used about 170,000 CPU hours for development and benchmarking. The main section focused on two spatial dimensions; preliminary three-dimensional tests reached up to 16,384 processing units.
The scaling experiment compared 64 processing units on one node, 256 on four nodes and 1,024 on 16 nodes. Runs with more resources were associated with better solution quality, while the highest-resource setting also used an extra multilevel level. Parallel efficiency diminished at larger scale; the paper attributed that decline to the code’s serial fraction.
The evidence remains specific to the computer tests
The reported findings come from numerical experiments for the stated PDE models on the tested computing system and resource settings. The main results are two-dimensional, and the three-dimensional evidence is preliminary.
The supplied analysis reports no confidence intervals or repeated-run variability for the headline comparisons. Whether the reported efficiency extends to other PDE systems, computer architectures and resource regimes remains open.
The algorithms are implemented in M++. Computing time was provided by HoreKa through NHR@KIT, with support from federal and state ministries; HoreKa is partly funded by the German Research Foundation. The document is identified as arXiv:2608.25958v1 and dated 26 August 2026.
Paper data and sources
Original title: Optimized Multilevel Sampling Methods under Resource Constraints
Authors: Niklas Baumgarten
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text