🔭 astrophysics

On the Reparameterization Between Cartesian Position-Velocity Vectors and Orbital Elements in the Kepler Problem

This paper presents compact analytic expressions for the Jacobian determinants of the transformation between orbital elements and Cartesian position-velocity vectors to facilitate Bayesian orbit inference, while correcting a formal singularity in a widely used binary microlensing parameterization and demonstrating that reparameterization to Cartesian state vectors significantly improves the efficiency and robustness of astrometric orbit fitting.

Original authors: Kento Masuda, Kansuke Nunota

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kento Masuda, Kansuke Nunota

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

Imagine you are trying to describe the path of a planet orbiting a star. You have two main ways to do this, much like describing a road trip.

The Two Ways to Describe a Path

The "Orbital Elements" Way (The Recipe): This is like describing a trip by listing the specific rules of the journey: "The road is an oval shape," "It's tilted at a 30-degree angle," "The starting point is here," and "The car is moving at this speed." In astronomy, these rules are called orbital elements (like eccentricity, inclination, and period). It's a compact, elegant way to define the orbit, but it can be tricky to work with if you don't know the exact rules yet.
The "Cartesian Vector" Way (The GPS Coordinates): This is like saying, "At this exact moment, the car is at location X, Y, Z and moving with velocity Vx, Vy, Vz." This is a direct snapshot of where the object is and where it's going.

The Problem: Changing the Language

Sometimes, scientists want to switch from the "Recipe" (Orbital Elements) to the "GPS" (Position and Velocity) because it makes the math easier, especially when they have very little data and the orbit is poorly understood.

However, there's a catch. In statistics (specifically Bayesian analysis), if you change the language you use to describe your data, you have to adjust your "assumptions" (called priors). Imagine you have a bag of marbles where every color is equally likely. If you switch to describing the marbles by their weight instead of their color, you can't just assume every weight is equally likely; you have to do a specific math calculation to make sure your assumptions stay fair.

This calculation involves something called a Jacobian determinant. Think of this as a "stretching factor." When you stretch a map from one projection to another, some areas get bigger and some get smaller. The Jacobian tells you exactly how much the "space" of your assumptions stretches or shrinks during the switch.

What This Paper Did

The authors, Kento Masuda and Kansuke Nunota, did three main things:

They wrote down the exact formula for the "stretching factor."
Before this, scientists often had to use computers to guess this number or use complex software to calculate it on the fly. The authors derived a clean, simple, written-down formula (a "closed-form expression") that anyone can use instantly. It's like giving everyone a pre-calculated ruler instead of asking them to build one every time.
They fixed a mistake in a famous "recipe book."
A widely used method for studying binary stars (two stars orbiting each other) was proposed by J. Skowron and colleagues in 2011. The authors of this paper found a subtle error in how Skowron defined the "starting angle" of the orbit.
- The Analogy: Imagine trying to describe a spinning top. Skowron's method defined the "up" direction based on the top itself. If the top wobbles, your definition of "up" wobbles too, making the math break down (become "singular").
- The Fix: The authors showed that if you define "up" using a fixed point in the sky (like a distant star) instead of the spinning top itself, the math works perfectly. Interestingly, they found that Skowron's final numbers were actually correct by accident, even though the logic used to get there was flawed. They provided the correct logic to back up the correct numbers.
They proved the "GPS" method is faster and more stable.
The team ran a computer simulation to see which method was better for finding the orbit of a fake planet. They compared the "Recipe" method against the "GPS" method.
- The Result: The "GPS" method (using position and velocity) was much better. It allowed the computer to explore the possibilities 2 to 7 times faster and with fewer errors.
- Why? The "Recipe" method often creates a "bumpy" landscape where the computer gets stuck in valleys. The "GPS" method smooths out the landscape, making it much easier for the computer to find the best answer.

In Summary

This paper provides a simple, exact formula to help scientists switch between describing orbits as "rules" or as "snapshots." They corrected a long-standing confusion in how to handle the angles in binary star systems and proved that using "snapshots" (Cartesian vectors) makes the computer work much faster and more reliably when trying to figure out the orbits of planets and stars, especially when there isn't much data to go on.

Technical Summary: On the Reparameterization Between Cartesian Position–Velocity Vectors and Orbital Elements in the Kepler Problem

Problem Statement
In the Kepler problem, bound orbits are traditionally characterized by orbital elements ( $\theta$ ), which provide a compact description of size, shape, and orientation. However, observables for stellar binaries and planetary systems are often more directly related to time-dependent Cartesian position and velocity vectors, $(x, v)$ . While reparameterizing orbit inference problems from orbital elements to Cartesian state vectors is computationally advantageous—particularly when orbits are weakly constrained and the posterior geometry is non-Gaussian or highly correlated—this transformation requires rigorous handling of prior probability densities in Bayesian frameworks. Specifically, transforming priors between these spaces necessitates the determinant of the $6 \times 6$ Jacobian matrix, $J_{Kep} = \partial(x, v)/\partial\theta$ . Previous literature has often treated this transformation implicitly, used numerical approximations, or failed to account for the induced prior differences, leading to incomparable posterior distributions. Furthermore, existing formulations in specific fields, such as binary microlensing, have contained conceptual errors regarding the definition of orbital angles that render intermediate Jacobians singular.

Methodology
The authors derive compact, analytic, closed-form expressions for the Jacobian determinants of the transformation between Cartesian state vectors and orbital elements, as well as for several useful variants.

Analytic Derivation: The authors utilize the canonical nature of Delaunay variables $(l, g, h, L, G, H)$ to derive the Jacobian determinant for the full 3D Kepler problem. By relating the standard orbital elements $(a, e, i, \Omega, \omega, M)$ to Delaunay variables, they obtain a simple expression for $\det J_{Kep}$ . They also derive expressions for transformations involving alternative parameters such as orbital period $P$ or time of periastron passage $\tau$ , and for reduced Jacobians where the nodal degree of freedom is eliminated (e.g., fixing the $y$ -coordinate).
Correction of Microlensing Formulation: The paper revisits the widely used formulation for lens orbital motion in binary microlensing events by J. Skowron et al. (2011). The authors identify that Skowron et al. inadvertently defined the longitude of the ascending node ( $\Omega_{node}$ ) relative to the sky-projected binary axis at a reference epoch. Because the binary axis is determined by the orbital elements themselves, this definition makes the Jacobian formally singular (the determinant is identically zero). The authors propose a corrected derivation where the longitude of the ascending node is defined relative to an axis independent of the binary orbit (e.g., the source trajectory or celestial north).
Numerical Verification and Sampling: The analytic expressions are verified numerically using the jnkepler package and JAX automatic differentiation. To demonstrate the practical utility of the reparameterization, the authors perform a quantitative comparison of astrometric orbit fitting using a gradient-based Markov Chain Monte Carlo (MCMC) algorithm (No-U-Turn Sampler, NUTS). They simulate astrometric data for a companion (similar to HR 8799d) and compare two sampling strategies: (i) direct sampling of orbital elements with uniform priors, and (ii) sampling of Cartesian state vectors $(x, v)$ with a uniform prior, applying the analytic Jacobian determinant to correct the prior density for the orbital elements. The sampling of $(x, v)$ is performed via a specific rejection-free scheme that ensures proposals remain within the bound orbital region.

Key Contributions

Closed-Form Jacobian Expressions: The paper provides explicit, compact analytic formulas for the Jacobian determinant of the transformation between $(x, v)$ and standard orbital elements (Equation 5), as well as variants using period $P$ or periastron time $\tau$ (Equations 16–17), and reduced 5D and 2D cases (Equations 19, 23). These expressions facilitate transparent and efficient prior conversion in Bayesian inference without relying on numerical differentiation.
Correction of Microlensing Parameterization: The authors clarify a conceptual error in the Skowron et al. (2011) formulation regarding the definition of the longitude of the ascending node. They demonstrate that while the original parameterization renders the intermediate Jacobian singular, the corrected formulation (using an independent reference axis) yields a well-defined Jacobian that numerically agrees with the original result, as the determinant is independent of the longitude of the ascending node.
Quantitative Performance Analysis: Through a controlled simulation, the authors show that reparameterizing to Cartesian state vectors significantly improves the efficiency and robustness of gradient-based MCMC sampling. They demonstrate that this approach reduces the fraction of divergent transitions and increases the effective sample size (ESS) per core-minute by a factor of 2–7 compared to sampling in orbital elements, even when priors are rigorously matched.

Results

The derived analytic Jacobian determinant for the full Kepler problem is $\det J_{Kep} = \frac{1}{2} \mu^{3/2} a^{1/2} e \sin i$ .
In the microlensing context, correcting the definition of the ascending node resolves the singularity in the Jacobian calculation, confirming that the original numerical result by Skowron et al. was effectively correct despite the flawed derivation.
In astrometric fitting, reparameterization to Cartesian vectors mitigates the complex correlations present in orbital element space. This leads to a smoother posterior geometry for Hamiltonian Monte Carlo methods, resulting in faster convergence and higher sampling efficiency. The study confirms that the posterior distributions obtained via both parameterizations are identical when the Jacobian correction is applied, validating the theoretical framework.

Significance
The paper establishes that reparameterization to Cartesian state vectors is a robust and efficient strategy for orbit inference, particularly in regimes where orbits are weakly constrained. By providing analytic Jacobian expressions, the authors enable the consistent application of Bayesian priors across different parameter spaces, resolving ambiguities in previous comparative studies. The work corrects a long-standing, albeit subtle, error in microlensing literature and provides a quantitative basis for the preference of state-vector sampling in gradient-based inference algorithms. The authors note that while rejection sampling may be more efficient for low-dimensional problems with simple posteriors, the Cartesian reparameterization offers a broadly applicable alternative for higher-dimensional problems, such as those involving multiple planets or combined datasets, where the posterior geometry is complex.

More like this