Linear models in astronomy and astrophysics

How it is used

Astronomy and astrophysics study objects we cannot touch: stars, planets around other stars, exploding stellar cores, bursts of radio waves from far outside our galaxy. Almost everything is learned from light or from faint ripples in space, recorded as brightness or signal strength changing over time and across wavelength. A linear model is a fitted rule that adds up measured quantities, each multiplied by a weight chosen from data, to produce an estimate or a score. A generalised version turns that sum into a probability, so the rule can answer yes-or-no questions, such as whether a dip in a star's brightness is a planet crossing in front of it.

In the work described here, linear models sit alongside more elaborate methods rather than replacing them. In one study of planet hunting in space-telescope brightness measurements, a regularised logistic regression was fitted to a handful of summary numbers describing each moment of observation, and was compared directly with trees and neural networks on the same features. Elsewhere linear fits serve as the plain reference point against which simulated signals, from collapsing stellar cores or from models of a planet's interior, are classified or matched.

Why researchers use it

Researchers reach for a linear model because it is cheap, stable and legible. The weights can be read off and inspected, so it is possible to see which measured quantity is doing the work, and whether that quantity makes physical sense or is simply an artefact of the instrument. Regularisation, which pushes weights towards zero unless the data insist otherwise, keeps the fit from chasing noise when there are many features and few confirmed examples, which is the usual situation in astronomy.

It also stands in for a baseline. Much of this work depends on simulated signals, because real detections are rare and their true answers are often unknown. A linear model shows how much of the task a simple rule already solves on such data, so any gain claimed by a larger model has something to be measured against. Where a straight line through the data is enough, it is also easier to defend to colleagues and to rerun on new observations.

In the record

Limits

A linear model assumes the answer rises or falls steadily with each measured quantity, and that the quantities act independently. Real astronomical signals rarely behave that way: a planet transit, a burst or a waveform has a shape, and shape is lost when the record is reduced to a few summary numbers. So the fit depends entirely on which numbers someone chose to compute. It also inherits the simulations it is trained on, including whatever physics they leave out, and it cannot flag the unfamiliar: anything unlike the training data is still forced into one of the available answers.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

• Linear model: the idea, across every field
• AI in astronomy and astrophysics

Every article