Compact finite-difference schemes on a uniform grid
A sound pulse that crosses 400 grid cells ends with an error of 61% of its height under the second-order central difference and 0.14% under the eighth-order compact scheme, on the same grid.
A simulation of the noise of an aircraft-engine fan carries sound waves across many grid cells, and a small error in each computed slope gives each wave a slightly wrong speed. Over long distances these errors add up. A workshop on benchmark problems in computational aeroacoustics, the computation of sound by computer, set common test problems that every noise-simulation code can be run on and checked against an exact answer1. Its first problem isolates this error. A smooth pulse of sound must cross 400 grid cells unchanged. The study runs it with explicit central differences, which compute the slope at a point from the values around it, and with compact schemes, which also link the slopes at neighbouring points, as presented in section 2.1 of the review by Arroyo Ramo et al.2 The order of a scheme is the power of the grid spacing at which its error shrinks; the explicit differences have orders 2, 4 and 6 and the compact schemes 4, 6 and 8.
The pulse computed with the second-order central difference
The second-order central difference estimates the slope at a point from the values at its two neighbours, and short waves, covered by few grid points, come out too slow. After 400 cells the pulse computed this way is 5.4 cells behind, its height has fallen from 0.50 to 0.27, and a train of ripples trails behind it.
Coupled derivatives
A compact scheme links the slope at a point to the slopes at its neighbours as well as to the values. Solved along a whole grid line at once, it reaches an order higher by 2 with the same neighbouring values, for 1.5 to 1.8 times the arithmetic of an explicit central difference of the same order. To compute the slope of a wave within 0.1% of its true value, the eighth-order compact scheme needs 4.5 grid points per wavelength where the second-order central difference needs 81.
The six schemes on the benchmark
After 400 cells the largest error is 61% of the pulse height with the second-order central difference, 6.2% with the sixth-order one, and 0.14% with the eighth-order compact scheme.
Cost of a 0.1% error in three dimensions
For a largest error of 0.1% of the pulse height on the benchmark, the eighth-order compact scheme can use a coarser grid. In three dimensions it needs 10 times fewer grid points than the sixth-order central difference and 5.3 times less arithmetic per derivative.
Limitations
These gains hold for smooth waves on a uniform grid. Near a wall a compact scheme needs one-sided formulas, which use values on one side only and can lower its order. Every slope on a grid line depends on the whole line, so a line split between processors needs extra communication.
References
- Tam, C. K. W. Benchmark problems and solutions. In ICASE/LaRC Workshop on Benchmark Problems in Computational Aeroacoustics (eds Hardin, J. C., Ristorcelli, J. R. & Tam, C. K. W.) 1–13 (NASA Conference Publication 3300, 1995). ntrs.nasa.gov/citations/19950023713
- Arroyo Ramo, A., Conejero, J. A., Pérez-Quiles, M. J. & Hoyas, S. Compact finite difference schemes: a review of fundamentals, applications, and practical implementation. Mathematics 14, 1958 (2026). doi:10.3390/math14111958