Abstract Finite difference schemes for some two-point boundary value problems are analyzed. It is found that for schemes defined on nonuniform grids, the order of the local truncation error does not fully reflect the rate of convergence of the numerical approximation obtained. Numerical results are presented that indicate that this is also the case for higher-dimensional problems.
Introduction
In the numerical solution of differential equations, particularly those exhibiting rapid local variations such as boundary layers in fluid dynamics, the choice of discretization grid is critical. While uniform grids are straightforward to implement, they often require an impractical number of points to resolve sharp features. Nonuniform grids, which cluster points in regions of high activity, offer a more efficient alternative.
However, a long-standing concern in numerical analysis has been the impact of grid nonuniformity on accuracy. Standard Taylor series analysis typically suggests that moving from a uniform to a nonuniform grid reduces the local truncation error (LTE) from second-order ($O(h^2)$) to first-order ($O(h)$). This paper investigates whether this local degradation necessarily implies a loss of global accuracy.
The Local vs. Global Error Discrepancy
The central finding of this research is that the order of the LTE is a pessimistic metric for global convergence. For many common finite difference schemes, the "low-order" local errors introduced by grid irregularities possess a specific structure that allows them to cancel out when integrated across the domain.
By employing Green's functions and a technique known as summation by parts, the authors demonstrate that if the LTE follows a specific form—specifically, if the leading error term is proportional to the change in grid spacing—the global error can remain second-order. This implies that the "jaggedness" of a grid does not inherently ruin the solution's quality, potentially simplifying the design of adaptive mesh refinement strategies.
Theoretical Framework
The study formalizes this observation in Theorem 2.1, which states that for a three-point finite difference scheme, if the LTE $\tau_j$ can be expressed as: $$\tau_j = (h_j - h_{j-1})\eta(x_j) + O(h^2)$$ where $\eta$ is a smooth function, the global error remains $O(h^2)$. Remarkably, the analysis even suggests that schemes with zero-order LTE—where the local error does not vanish as the grid is refined—can still produce convergent global solutions under certain conditions.
Numerical Results
To verify these theoretical claims, the authors conducted experiments on both 1D and 2D test cases. In the 1D case, a linear second-order ODE was solved on both a uniform grid and a highly irregular nonuniform grid where spacing alternated abruptly.
As shown in Table 1, the "ratio" column tracks the reduction in error as the number of points $N$ is doubled. A ratio of 4 signifies second-order convergence ($2^2 = 4$). The data clearly shows that both the uniform and the nonuniform grids approach a ratio of 4, confirming that they share the same asymptotic rate of convergence despite the nonuniform grid's theoretically lower LTE.
Conclusion
This work demonstrates that the global rate of convergence for finite difference schemes on nonuniform grids is often higher than the local truncation error would suggest. While this study focuses on second-order ODEs and three-point schemes, the numerical evidence from 2D Poisson-like equations suggests these principles extend to higher dimensions. Future research is needed to determine if these error-cancellation effects hold for higher-order schemes and how they interact with conservation laws in computational fluid dynamics.
Summary updated