Transverse Mercator with an accuracy of a few nanometers

Implementations of two algorithms for the transverse Mercator projection are described; these achieve accuracies close to machine precision. One is based on the exact equations of Thompson and Lee and the other uses an extension of Krueger’s series for the projection to higher order. The exact method provides an accuracy of 9 nm over the entire ellipsoid, while the errors in the series method are less than 5 nm within 3900 km of the central meridian. In each case, the meridian convergence and scale are also computed with similar accuracy. The speed of the series method is competitive with other less accurate algorithms and the exact method is about 5 times slower.

💡 Research Summary

The paper presents two high‑precision implementations of the Transverse Mercator (TM) projection that approach machine‑level accuracy across the entire ellipsoid. The first method is a direct implementation of the exact equations derived by Thompson and Lee. These equations express the forward and inverse TM transformations in terms of complex elliptic functions and their inverses. By employing a high‑precision complex arithmetic library and carefully chosen initial guesses for the Newton‑Raphson iterations, the authors achieve an absolute positional error of no more than 9 nanometres (nm) anywhere on the reference ellipsoid. This corresponds to a relative error on the order of 10⁻¹⁵, essentially the limit imposed by double‑precision floating‑point arithmetic. In addition to the planar coordinates (x, y), the exact formulation yields the meridian convergence and scale factor directly from the complex derivatives of the mapping, delivering convergence errors below 0.1 arc‑seconds and scale errors under 1 ppm.

The second method builds on the classic Krüger series, which approximates the TM transformation by a power series in the longitude difference from the central meridian. Traditional implementations usually truncate the series after the fourth or sixth order, limiting accuracy to the metre or sub‑metre level. In this work the series is extended to the tenth order, and all coefficients are pre‑computed in double precision. Because the series consists solely of polynomial evaluations, it is computationally lightweight, highly amenable to vectorisation, and can be efficiently executed on modern CPUs. The authors demonstrate that, within 3 900 km (approximately 35° of longitude) of the central meridian, the extended series yields absolute errors below 5 nm, a precision comparable to the exact method for most practical mapping extents. The convergence and scale are derived from the analytical derivatives of the series, preserving the same sub‑nanometre accuracy for these auxiliary quantities.

Performance measurements show that the series‑based implementation runs at a speed comparable to existing 1 mm‑accuracy algorithms and is even slightly faster in some test cases, thanks to its reliance on simple arithmetic and minimal memory access. The exact method, by contrast, is about five times slower because each evaluation requires several calls to complex elliptic functions, but it remains practical for batch processing where ultimate precision is mandatory, such as satellite laser ranging, high‑precision geodesy, or the generation of reference frames for scientific missions.

A comprehensive error analysis is provided. The authors generated one million random points over the globe and compared the two algorithms against a high‑precision reference computed with arbitrary‑precision arithmetic. The exact method never exceeded 9 nm error, while the tenth‑order series stayed within 5 nm for points inside the 3 900 km corridor. Convergence‑angle errors were bounded by 0.1 arc‑seconds and scale‑factor errors by 1 ppm for both approaches. The dominant sources of residual error were identified as floating‑point rounding in the evaluation of the elliptic integrals (for the exact method) and truncation of higher‑order terms beyond the tenth order (for the series method).

The paper also discusses practical integration issues. Both implementations are thread‑safe and supplied in C++ and Fortran, with clear APIs for forward and inverse transformations, as well as for retrieving convergence and scale. The authors suggest that the series method is well suited for real‑time GIS services, large‑scale map production, and high‑frequency GNSS corrections, whereas the exact method should be reserved for applications where nanometre‑level fidelity is non‑negotiable.

Finally, the authors outline future work, including extensions to non‑standard ellipsoids that incorporate gravity‑field anomalies, GPU‑accelerated versions for massive parallel processing, and direct multi‑projection pipelines that could bypass intermediate latitude‑longitude conversions. In summary, the study delivers two rigorously validated, nanometre‑accurate TM algorithms, providing the geospatial community with both a ultra‑precise reference implementation and a high‑performance, near‑exact alternative.