Brushing

Brushing is two moves welded together. Drag a rectangle across one chart and it marks a range rather than a set of objects: every record whose values fall inside it, named by the numbers on the edges instead of by identity. Then the linking half fires, and every other view of the same data repaints to show that same subset, so a band drawn on a scatter plot becomes a shorter bar in the chart beside it. Richard Becker and William Cleveland formalised the technique at Bell Labs in the 1980s, brushing scatterplot matrices so that a rectangle dragged over one panel lit the corresponding points in all the others. The full name is still brushing and linking, because neither half does anything on its own: a brush with nothing linked to it is a highlighter, and a linked view with nothing brushing it never changes.

The word to hold it against is lasso selection. A lasso selects objects on one canvas, and what it produces is a set of things you are about to act on. A brush selects a range, and what it produces is a filter every linked view answers to. That single difference explains the rest of the behaviour: a lasso’s outline dies with the gesture that drew it, while a brush stays on screen afterwards, with handles to widen it and a body to slide, because it is a standing query rather than a moment of pointing. Range select is the third neighbour and the flattest of the three, spanning a run of an ordered list with Shift and a second click, which yields a set of rows again rather than a condition on a dimension.

Most of the craft is in the edges. An empty brush should mean everything rather than nothing, so a freshly loaded view starts at its full count and the reader is filtering down rather than building up, and clearing therefore has to be its own control: a second drag would be indistinguishable from the start of a new brush. Records outside the brush are dimmed rather than deleted, because the shape the whole distribution made is the context that makes the selection mean anything, and a linked view that empties out has thrown away the comparison the reader was making. Update the linked views continuously through the drag rather than on release, since the whole appeal is watching a second chart move while your hand is still on the first one. And decide early whether a brush is one dimensional (a range along an axis, which is what most time series want) or two, because the second handle set costs more than it usually returns.

Brushing is the interaction small multiples were waiting for: once one design is repeated across a dozen panels, a brush in any panel is a question asked of all of them at once, which is exactly the coordinated-views idea that dashboards later sold under the name crossfilter. The control panel version of the same job is a faceted filter, where the ranges are picked from checkboxes and sliders instead of drawn, and it is worth shipping both: a brush is a drag over a picture, so it is unavailable to anyone not using a pointer. Publish the brushed range as two numeric fields that read and write the same state, announce the resulting count in a live region, and the chart becomes a convenience rather than the only door.