singlens

Differentiable single-lens microlensing in JAX: static and annual-parallax light curves, astrometric centroid shifts, finite sources, limb darkening, and gradient-based inference.

python -m pip install singlens

Requires Python 3.12+ and JAX 0.11.2+. Install singlens[examples] for plotting and inference dependencies. Enable float64 before constructing models or JAX arrays.

Quickstart

import jax
jax.config.update("jax_enable_x64", True)
import jax.numpy as jnp
from singlens import AnnualParallax, Model, Parameters

times = jnp.linspace(7980., 8020., 501)
geometry = AnnualParallax(ra_deg=266., dec_deg=-29., t_ref=8000.)
model = Model(source="linear", parallax=geometry)
params = Parameters(t0=8000., u0=0.005, t_e=20., rho=0.01, a1=0.5,
                    pi_e_n=0.2, pi_e_e=-0.1)
A = jax.jit(model.magnification)(times, params)

Model fixes the source profile and geometry; Parameters holds fitted values. Choose source="point", "uniform", "linear", or "mu2" explicitly. Results preserve the shape of times. Parallax times are HJD − 2450000 in days. See finite-source effects for supported profiles, solver choices, and accuracy. Differentiate a parameter with JAX:

curve = jax.jit(lambda rho: model.magnification(times, params.replace(rho=rho)))
A = curve(0.01)
dA_drho = jax.jacfwd(curve)(0.01)

Use the same event parameters for the unresolved images' astrometric shift:

A, shift = jax.jit(model.joint)(times, params)
shift_mas = 1.2 * shift  # For an angular Einstein radius of 1.2 mas.

The shift is relative to the unlensed source center in the tau + 1j*beta trajectory frame, for a dark point lens and a circular source.

Documentation

• Getting started: installation, light curves, and autodiff.
• Finite-source effects: source profiles, methods, and accuracy.
• Astrometry: sky coordinates, source motion, and blended light.
• Inference examples: photometry with a GP, flux marginalization, and joint photometry/astrometry.
• API and benchmarks.

For source installation and development, see CONTRIBUTING.

Attribution

Adapted from microJAX. Table generation uses the disk-mixture method of Wang (2026); the FFT backend follows Sugiyama (2022). See attribution, NOTICE, and LICENSES for citations and the MIT, Apache-2.0, and BSD-3-Clause license notices.