When a heat or chemical concentration is carried by a fast flow and diffuses only slightly (the "convection-dominated" regime), the true solution develops very thin, steep layers, and it never leaves the range set by its boundary data. This property is the maximum principle. Cheap surrogate solvers routinely break it. A reduced basis model approximates each new solution as a weighted sum of a few precomputed solutions, and that sum can overshoot near a steep front, giving negative concentrations or temperatures hotter than any heat source. This holds even when every precomputed solution was perfectly bounded. Our proposal is to build the reduced model from a discretization that provably respects the maximum principle, and then to solve the small online problem as a residual minimization subject to bound constraints. The constraints are checked only at a handful of points chosen offline. The result should be a surrogate that is as fast as a standard reduced model and never produces unphysical values.
The opportunity exists because two lines of work have not met. The first line consists of monotone schemes, discretizations whose matrix is an M-matrix. That means positive diagonal entries, non-positive off-diagonal entries, and a non-negative inverse, which together force discrete solutions to stay within bounds. These schemes now handle diffusion coefficients down to $10^{-6}$. The second line consists of stabilized reduced basis methods. These methods tame the gross instabilities of convection-dominated parametric problems, but their stabilization does not preserve bounds. The research file reports no reduced basis method that combines the two.
Core Claim
Linear combinations of bounded snapshots are not bounded. This is the same mechanism that makes moving fronts hard to compress (the slow decay of the Kolmogorov $n$-width). It can be repaired cheaply by enforcing the bound as an inequality constraint in the online solve. Because violations concentrate near layers whose locations vary smoothly with the parameter, a small offline-selected set of constraint points should suffice. If so, the online cost stays independent of the fine mesh size. A full-field a posteriori check then certifies the result, and a fallback adds points whenever that check fails.
Prior Work and Positioning
Our full-order "truth" solver is supplied by Edge-averaged virtual element methods for convection-diffusion and convection-dominated problems. That paper extends Scharfetter-Gummel edge averaging (exponential weighting of edge fluxes) to polygonal virtual element methods. It proves that its Monotone EAVE variant on Voronoi meshes with mass lumping yields an M-matrix, and it reports an $L^\infty$ error "nearly 100 times smaller than that of SUPG" at $\epsilon=10^{-6}$.
A Stable Mimetic Finite-Difference Method for Convection-Dominated Diffusion Equations gives a closely related M-matrix construction on Delaunay–Voronoi dual meshes. However, it assumes the velocity is the gradient of a potential, which the Graetz pipe flow is not. We therefore use EAVE, and we verify the M-matrix property numerically on every truth matrix rather than assuming it.
The parametric setting and the strongest existing reduced model come from the weighted reduced basis line of work: Stabilized Weighted Reduced Basis Methods for Parametrized Advection Dominated Problems with Random Inputs and its optimal-control extension Stabilized weighted reduced order methods for parametrized advection-dominated optimal control problems. These papers embed Streamline-Upwind Petrov-Galerkin (SUPG) stabilization, a residual-based artificial diffusion along streamlines, into reduced spaces, and they activate it selectively online. They motivate our proposal and also constrain it. They show that stabilization is needed online, but SUPG controls energy norms and not pointwise bounds. Overshoots therefore survive, and the file reports no overshoot measurement for these models.
The idea of enforcing physical structure inside a low-dimensional solve comes from Constraint-Satisfying Krylov Solvers for Structure-Preserving Discretizations. That work projects conservation laws onto the Krylov subspace and solves a small sequential quadratic program at the final iteration, which gives machine-precision conservation at 20–30% runtime overhead. Its constraints are equalities only. Ours are the box inequalities it explicitly does not handle, and our small subspace is a reduced basis rather than a Krylov space.
Finally, Evolve Smoothly, Fit Consistently attacks the same $n$-width barrier with a learned nonlinear decoder. It represents the main alternative philosophy, in which the basis itself is changed rather than the projection being constrained.
Research Program
Offline, we compute Monotone EAVE snapshots over the parameter domain and build a weighted greedy basis $V\in\mathbb{R}^{N\times r}$, where $N$ is the number of fine-mesh unknowns and $r$ the number of modes, following the weighted sampling of Torlo et al. Online, for a parameter $\mu$ with truth matrix $A(\mu)$ and right-hand side $f(\mu)$, we solve
$$ \min_{c\in\mathbb{R}^r}\ \|A(\mu)Vc - f(\mu)\|_{W}^2 \quad \text{subject to}\quad u_{\min}\le (Vc)_i \le u_{\max},\ i\in\mathcal{S}. $$
Here $c$ holds the reduced coefficients, $W$ is a weighting norm that we assemble affinely, and $\mathcal{S}$ is a set of $m\ll N$ constraint points. We use residual minimization rather than plain Galerkin projection because the convection operator is non-symmetric, and a minimization is what makes "satisfy the PDE as well as possible inside the bounds" well posed. The problem is a strictly convex quadratic program in $r$ variables with $2m$ constraints, which an active-set method solves in time independent of $N$.
The central technical piece is choosing $\mathcal{S}$. We propose a greedy procedure in the spirit of empirical interpolation. It runs the unconstrained reduced model over a training set, adds the node of worst bound violation, re-solves with constraints, and repeats until no training parameter violates the bounds anywhere on the fine mesh. We will analyze how $m$ grows with the Péclet number and with $r$. Our hypothesis is that it tracks the number of distinct layer positions, not the mesh size. Online, an $O(N)$ check of the reconstructed field certifies the bounds, and any violated node is added to $\mathcal{S}$ before a re-solve. We will measure how often this fallback fires on test parameters.
Baselines
The naive reference point is the unconstrained reduced model built on the same EAVE snapshots. This is also the ablation that isolates our key ingredient, because it shares the basis, the residual-minimization norm, and the offline cost, and differs only in the bound constraints. If it already rarely overshoots, the proposal has no reason to exist. We will report its violation rate first.
The strongest existing method is the stabilized weighted reduced basis model with selective online SUPG of Torlo et al.. We will reimplement it on the same Poiseuille-Graetz and propagating-front problems, with the same parameter distributions, the same fine mesh resolution, the same number of modes, and the same greedy tolerance. We will tune its SUPG parameter over the same grid of values that we allow for our weighting norm. Their reported figures are a truth solve of 132 seconds against 0.35 seconds for the reduced solve, an error "one order of magnitude lower" from weighting, and up to 45% cost reduction from selective stabilization. The file reports no maximum-overshoot or $L^\infty$ number for this method on Lacuna, so we will measure both ourselves. Beating it means three things together: zero bound violations on the test set, an $L^\infty$ error relative to the truth no worse than theirs, and an online time within a factor of two of theirs.
The strongest post-processing competitor is the unconstrained EAVE reduced model followed by clipping to $[u_{\min},u_{\max}]$, with an additional variant that restores the integral after clipping. This competitor is cheap and trivially bound-preserving.
As an upper reference for cost, we will also run the constrained solve with $\mathcal{S}$ equal to all nodes. This shows how much accuracy the sparse constraint set gives up, if any.
Why not just…?
Why not just clip the reduced solution afterward? This is the most serious threat, because clipping also guarantees bounds, and in $L^\infty$ it may come close to our method. The difference is that clipping ignores the PDE. It flattens overshoots without moving mass back into the layer, so it biases physically meaningful outputs such as the outlet-averaged temperature and the wall heat flux in the Graetz problem. We will compare residual norms and these output functionals between clipping and constrained minimization. If clipping matches us on outputs as well, we will report that clipping is sufficient, and the remaining contribution would be the analysis of where violations occur.
Why not just add more modes? Adding modes reduces the average error. At fronts, however, the linear-combination Gibbs effect is the $n$-width barrier itself, and the barrier is known to scale badly with the Péclet number. We will plot maximum overshoot against $r$ at several Péclet numbers to test whether violations ever vanish at practical $r$.
Why not just use a SUPG truth solver with stronger stabilization, or first-order upwinding? The EAVE comparison already reports a roughly hundredfold $L^\infty$ gap against SUPG at $\epsilon=10^{-6}$. Meanwhile, upwinding gains bounds only by smearing the layer, as the file's overview of artificial diffusion notes. We will include an upwind-truth reduced model to quantify that smearing.
Why not just use a nonlinear latent model such as the Nonlinear Fourier Ansatz decoder? Such models can compress fronts better, but nothing in Evolve Smoothly, Fit Consistently bounds their outputs. They also lose the affine offline–online structure and the residual certificate. Our constraint layer is orthogonal to that line and could later sit on top of it.
Proposed Experiments
We will use three test problems. The first is the EAVE unit-square convection-diffusion problem with $\epsilon$ parametrized over $[10^{-6},10^{-2}]$ together with the flow direction. The second is the Poiseuille-Graetz problem with Péclet number and length as random inputs. The third is the propagating front in a square, whose layer angle is stochastic and which is therefore the hardest case for a linear basis. All three use Voronoi meshes generated by Lloyd iteration, so that the truth matrices are M-matrices, and we will check this property explicitly.
The metrics are maximum overshoot and undershoot, the fraction of test parameters with any violation, the $L^\infty$ and $L^2$ errors against the EAVE truth, the error in output functionals, online wall time, the size $m$ of the constraint set, and how often the fallback fires.
Two results would settle the question. First, the constrained model must reach zero violations with $m$ growing sublinearly in $N$ as the mesh is refined at fixed Péclet number. Second, on the propagating-front problem its output-functional error must beat both clipping and SUPG reduced basis at equal $r$. If $m$ instead scales with $N$, the sparse-constraint hypothesis fails. If clipping matches us on outputs, the constraint machinery is unnecessary. We will report either outcome.
References
- Edge-Averaged Virtual Element Methods for Convection-Diffusion and Convection-Dominated Problems
- A Stable Mimetic Finite-Difference Method for Convection-Dominated Diffusion Equations
- Stabilized Weighted Reduced Basis Methods for Parametrized Advection Dominated Problems with Random Inputs
- Stabilized weighted reduced order methods for parametrized advection-dominated optimal control problems governed by partial differential equations with random inputs
- Constraint-Satisfying Krylov Solvers for Structure-Preserving Discretizations
- Evolve Smoothly, Fit Consistently: Learning Smooth Latent Dynamics For Advection-Dominated Systems