Density Estimation for Projected Exoplanet Quantities

TL;DR

This paper introduces a novel nonparametric method to estimate the true distribution of unprojected exoplanet properties from projected measurements, improving accuracy with larger samples and aiding future survey design.

Contribution

The paper presents a new linear-equation-based density estimation technique combined with kernel smoothing, calibrated for exoplanet data, enhancing the analysis of projected exoplanet measurements.

Findings

Method accurately recovers unprojected density from projected data.

Resolution improves with larger sample sizes.

Application to real data demonstrates effectiveness.

Abstract

Exoplanet searches using radial velocity (RV) and microlensing (ML) produce samples of "projected" mass and orbital radius, respectively. We present a new method for estimating the probability density distribution (density) of the unprojected quantity from such samples. For a sample of n data values, the method involves solving n simultaneous linear equations to determine the weights of delta functions for the raw, unsmoothed density of the unprojected quantity that cause the associated cumulative distribution function (CDF) of the projected quantity to exactly reproduce the empirical CDF of the sample at the locations of the n data values. We smooth the raw density using nonparametric kernel density estimation with a normal kernel of bandwidth \sigma. We calibrate the dependence of \sigma on n by Monte Carlo experiments performed on samples drawn from a theoretical density, in which the integrated square error is minimized. We scale this calibration to the ranges of real RV samples using the Normal Reference Rule. The resolution and amplitude accuracy of the estimated density improve with n. For typical RV and ML samples, we expect the fractional noise at the PDF peak to be approximately 80 n^{-\log 2}. For illustrations, we apply the new method to 67 RV values given a similar treatment by Jorissen et al. in 2001, and to the 308 RV values listed at exoplanets.org on 20 October 2010. In addition to analyzing observational results, our methods can be used to develop measurement requirements--particularly on the minimum sample size n--for future programs, such as the microlensing survey of Earth-like exoplanets recommended by the Astro 2010 committee.