🌀 nonlinear sciences

Validating direct solvers for Newton's gravitational N-body problem, and the systematic comparison between IEEE floating point and Posits

This paper systematically compares IEEE 754 floating-point arithmetic and Posit implementations for solving Newton's chaotic N-body problem, finding that while half and single precisions are insufficient for individual strong encounters, current Posit software implementations either suffer from significant speed penalties or exhibit larger errors and systematic drifts compared to standard double-precision floating-point arithmetic, making them unsuitable alternatives for this specific application.

Original authors: Simon Portegies Zwart (Leiden Observatory)

Published 2026-08-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Simon Portegies Zwart (Leiden Observatory)

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Gravity is the invisible thread that stitches the universe together, pulling stars, planets, and dust into the vast, swirling structures we see in the night sky. To understand how these cosmic systems evolve, scientists build digital models that simulate the gravitational dance of countless objects. These simulations are not mere guesses; they are rigorous calculations that track the position and speed of every particle over time. However, because gravity is a chaotic force, even the tiniest error in a calculation can snowball, causing the simulated universe to drift wildly away from reality. For decades, scientists have relied on a standard system of numbers called floating-point arithmetic to perform these calculations. It is a method that has served well, but it is not perfect, and it struggles when the numbers get extremely large or extremely small. Recently, a new way of representing numbers called "Posits" has been proposed as a faster and more efficient alternative. The question facing the scientific community is whether this new tool can handle the extreme demands of simulating gravity, or if it will introduce hidden flaws that could ruin our understanding of the cosmos.

In a recent study, a researcher at Leiden Observatory set out to put this new number system to the ultimate test. The goal was to see if Posits could accurately solve the equations of motion for groups of stars and planets, a task that requires tracking the gravitational pull between every single object in a system. The researcher compared the new Posit system against the traditional floating-point numbers that have been the standard for decades, as well as against a "gold standard" of calculation that uses extreme precision to find the true answer. The simulations involved everything from a simple trio of stars locked in a stable orbit to a chaotic cluster of one hundred stars crashing and swinging past one another. The tests were designed to be unforgiving, pushing the number systems to their limits by simulating close encounters where stars pass near each other at high speeds, and by checking if the laws of physics held true even when the entire system was moving through space.

The results were clear and, for the proponents of the new system, somewhat disappointing. The study found that the new Posit system, in its current software form, is not yet ready to replace the traditional numbers for these complex gravitational problems. While the new system showed some promise in specific areas, it consistently failed to match the reliability of the standard method when dealing with the chaotic nature of gravity. In simulations involving three stars, the new system produced results that were comparable to, and in some specific metrics even slightly better than, the traditional numbers, though both drifted from the true path over time. When the researcher tested a more complex scenario with one hundred stars, the new system struggled to keep track of the stars' positions accurately, especially when they came close to one another. The errors were not just small mistakes; they were systematic, meaning the new system consistently made the same kinds of wrong turns, leading to a simulation that looked different from reality.

One of the most revealing tests involved checking a fundamental principle of physics known as Galilean invariance. This principle states that the laws of physics should look the same whether you are standing still or moving at a constant speed. In a perfect simulation, shifting the entire system of stars to a different location or giving them a uniform push should not change how they interact with each other. However, the study showed that the new number system was not immune to this issue; in fact, the traditional floating-point system suffered from the same problem. When the researcher moved the entire simulated universe, both the new system and the traditional system produced different results for how the stars orbited, whereas the "gold standard" remained consistent. This suggests that the new system introduces subtle biases depending on how the numbers are arranged, a flaw that could lead to incorrect conclusions about how real star systems behave, but it is a flaw shared by the current standard as well.

Speed was another major factor in the comparison. One of the main selling points of the new number system was that it could be calculated much faster than the traditional method. However, the study found that this speed advantage did not materialize in practice. Because the new system is not yet supported by the computer chips inside modern machines, it has to be calculated using software, which is much slower. The traditional numbers, on the other hand, are built directly into the hardware of the computer, allowing them to be processed almost instantly. The study showed that the software version of the new system was significantly slower than the traditional method, ranging from roughly 20 to 120 times slower in simpler three-star tests, and up to thousands of times slower in more complex hundred-star simulations. This massive loss of speed, combined with the lower accuracy, meant that the new system offered no practical benefit for these types of simulations.

The researcher also looked at how different levels of precision affected the results. They found that using fewer bits of information, such as sixteen or thirty-two, was simply not enough to capture the subtle details of gravitational interactions. Even with the new system, these lower precision levels led to results that were scientifically meaningless. The study confirmed that to get a reliable answer, one needs the full power of sixty-four-bit precision, and even then, the traditional method proved more robust. The new system seemed to handle numbers near the value of one very well, but as the numbers grew larger or smaller, its accuracy dropped off sharply. This is a critical weakness for gravitational simulations, where distances and masses can vary by many orders of magnitude.

Ultimately, the study concludes that while the new number system is an interesting innovation, it is not yet the ideal tool for simulating the chaotic and stiff equations that govern gravity. The traditional floating-point system, despite its age, remains the superior choice for these tasks because it offers a predictable level of accuracy and is supported by fast, dedicated computer hardware. The new system, in its current form, introduces errors that are too large and unpredictable for scientific use, and it runs far too slowly to be practical. The researcher suggests that until the new system can be implemented in hardware and its errors are better understood, scientists should continue to rely on the established methods. The quest for better tools continues, but for now, the old way of calculating the movements of the stars remains the most trustworthy path to understanding the universe.

Technical Summary: Validating Direct Solvers for Newton's Gravitational N-body Problem

Problem Statement
The paper addresses the challenge of validating direct solvers for Newton's chaotic N-body problem, specifically focusing on the systematic comparison between traditional IEEE-754 floating-point arithmetic and the newer Posit (Type III unum) number format. The gravitational N-body problem is inherently chaotic for $N > 2$ , making it highly sensitive to small numerical errors, rounding, and dynamic range limitations. The study aims to objectively evaluate whether Posits offer a viable alternative to IEEE-754 formats (fp16, bfp16, fp32, fp64, fp128) in terms of precision, speed, and reliability when integrating these stiff, chaotic ordinary differential equations.

Methodology
The author employs a rigorous validation framework comparing various numerical implementations against "ground truth" solutions generated by Brutus, an arbitrary-precision N-body code. Brutus utilizes MPFR for arbitrary-precision arithmetic combined with the Bulirsch-Stoer integrator and adaptive time-stepping to achieve convergence to a specified number of significant digits.

The experimental setup involves four distinct test cases:

Figure-8 Orbit: A planar, linearly stable 3-body problem sensitive to orbital phase and angular momentum errors.
Pythagorean Problem: A classic 3-body problem involving close encounters, sensitive to energy and angular momentum conservation.
D9 Binary System: A 3-dimensional problem involving a binary star orbiting a supermassive black hole (Sgr A*), testing secular stability, wide dynamic range, and Galilean invariance.
Plummer Sphere ( $N=100$ ): A 3-dimensional, virialized cluster of 100 particles with a power-law mass distribution, representing a computationally intensive, chaotic system.

Two specific Posit implementations are tested: CPPPosits (a specialized C++ library) and Universal (a broader ecosystem library). These are compared against hardware-accelerated IEEE-754 floats (fp16, bfp16, fp32, fp64) and software-emulated fp128. Performance is measured via wall-clock time, energy conservation, and phase-space distance from the converged solution.

Key Contributions

Systematic Benchmarking: The paper provides one of the first comprehensive benchmarks of Posits against IEEE-754 formats specifically for chaotic gravitational N-body simulations, covering bit-widths from 16 to 128.
Validation Framework: It establishes a methodology for distinguishing between "Case A" errors (acceptable deviations where the system remains within the shadowed phase space) and "Case B" errors (unacceptable deviations where the solution diverges from the true dynamical outcome).
Galilean Invariance Testing: The study introduces a specific test for Galilean invariance by offsetting initial conditions and adding systemic velocities, revealing significant failures in certain numerical representations.
Implementation Comparison: It highlights profound differences between two Posit libraries (CPPPosits and Universal), demonstrating that performance and precision are highly implementation-dependent.

Results

Low Precision (16-bit): Both standard fp16 and bfp16 (brain floating point) are insufficient for solving Newton's equations of motion. While bfp16 offers a larger dynamic range than fp16 and can complete limited orbits, neither provides scientifically meaningful results for chaotic systems.
Single Precision (32-bit): 32-bit floats and Posits (Posits<32, 2>) lead to relatively large errors in any individual strong encounter. While they could theoretically be used for statistical ensemble calculations, the paper notes that such ensemble averages are rarely validated and may be inappropriate for describing chaotic behavior; thus, they are not a reliable solution for individual trajectories.
Double Precision (64-bit):
- CPPPosits: This implementation exhibits systematic errors approximately an order of magnitude larger than fp64, with excesses exceeding two orders of magnitude in some cases. It fails to resolve close encounters reliably and exhibits significant drift in result space. Regarding speed, the 64-bit and 32-bit CPPPosits implementations were found to be 25 and 20 times slower, respectively, than hardware-accelerated fp64 calculations for the 3-body problems, while the 100-particle Plummer model showed a slowdown of approximately 120 times.
- Universal: This implementation achieves precision comparable to fp64 but suffers from a severe performance penalty, being roughly 1,800 to 2,000 times slower than fp64 due to the lack of hardware support and software implementation overhead.
Quadruple Precision (128-bit): Software-emulated fp128 successfully resolves close encounters and maintains convergence where 64-bit methods fail, though at a significant computational cost (approx. 5x slower than fp64).
Galilean Invariance: While all tested implementations show some difficulty integrating dynamical systems in moving reference frames, the results vary significantly. The paper notes that fp64 and Universal give indistinguishable results compared to the non-Galilean invariant experiment, whereas CPPPosits suffers from more pronounced errors and fails to resolve the orbit correctly in the D9 test case.
Error Distribution: In the $N=100$ Plummer model, the Universal library shows a narrower distribution of errors compared to CPPPosits but maintains a lower average number of correct decimal places than fp64. CPPPosits produces "outlier" errors that can be catastrophic for scientific interpretation.

Significance and Conclusions
The paper concludes that, in their current implementations, Posits are not the ideal alternative to fp64 for integrating chaotic or stiff ordinary differential equations such as Newton's equations of motion.

The primary limitations identified are:

Predictability: Unlike IEEE-754, which offers predictable error bounds (approx. 15 significant digits for fp64), Posits exhibit less predictable error behavior. They may perform well near unity but degrade significantly as values move away from unity, leading to occasional large rounding errors that degrade the entire integration.
Implementation Variance: The lack of a formal standard for Posits leads to significant differences in performance and accuracy between libraries (e.g., CPPPosits vs. Universal).
Performance vs. Precision Trade-off: The software implementations of Posits tested are significantly slower than hardware-accelerated fp64. The "Universal" library offers comparable precision but at a prohibitive speed cost, while "CPPPosits" offers comparable speed but inferior precision and reliability.

The author notes that while 16-bit arithmetic fails completely, bfloat16 offers improved dynamic range over standard fp16. However, for self-gravitating N-body systems, the study finds no compelling evidence to replace IEEE-754 double precision with Posits. The paper suggests that a hybrid approach (monitoring precision and correcting) might be necessary for Posits to be viable, but such an approach would likely negate any speed advantages. The study emphasizes that the reliability of chaotic system simulations depends heavily on the ability to resolve close encounters and maintain Galilean invariance, areas where current Posit implementations fall short compared to established floating-point standards.

More like this