Double coordinate transformations for geoms that draw with other geoms

[has2k1]

When a geom queries the coordinate range and uses the result to create data that is then plotted by another geom, there is potentially a double transformation. i.e. coord.range() yields results in coordinate space and if followed by coord.transform(), the data may get transformed two times. Affects coords that transform the limits in coord$setup_panel_params and also do transforms in coord$transform. This includes coord_trans and coord_sf.

Affected geoms include geom_abline, geom_hline and geom_hline.

reprex

library(ggplot2)

df <- data.frame(x=1:6, y=10^(1:6))
ggplot(df, aes(x, y)) +
  geom_point() +
  geom_vline(xintercept=3, color='red') +
  coord_trans(y='log10')

The red line (at x=3) should be top to bottom.

[hadley]

I don't see any obvious way to fix this. It'll have to wait until someone is deeper into coordinate systems.

[clauswilke]

I think the correct way to fix this is to copy the line-drawing code from geom_segment():

ggplot2/R/geom-segment.r

Lines 106 to 141 in 4d2ca99

	draw_panel = function(data, panel_params, coord, arrow = NULL, arrow.fill = NULL,
	lineend = "butt", linejoin = "round", na.rm = FALSE) {
	data <- remove_missing(data, na.rm = na.rm,
	c("x", "y", "xend", "yend", "linetype", "size", "shape"),
	name = "geom_segment")
	if (empty(data)) return(zeroGrob())
	if (coord$is_linear()) {
	coord <- coord$transform(data, panel_params)
	arrow.fill <- arrow.fill %||% coord$colour
	return(segmentsGrob(coord$x, coord$y, coord$xend, coord$yend,
	default.units = "native",
	gp = gpar(
	col = alpha(coord$colour, coord$alpha),
	fill = alpha(arrow.fill, coord$alpha),
	lwd = coord$size * .pt,
	lty = coord$linetype,
	lineend = lineend,
	linejoin = linejoin
	),
	arrow = arrow
	))
	}
	data$group <- 1:nrow(data)
	starts <- subset(data, select = c(-xend, -yend))
	ends <- plyr::rename(subset(data, select = c(-x, -y)), c("xend" = "x", "yend" = "y"),
	warn_missing = FALSE)
	pieces <- rbind(starts, ends)
	pieces <- pieces[order(pieces$group),]
	GeomPath$draw_panel(pieces, panel_params, coord, arrow = arrow,
	lineend = lineend)
	},

and modify appropriately for geom_abline() etc. It's the geom's responsibility to draw correctly given the data and the coord, and if a geom can't guarantee correct drawing when using another geom's drawing code then it needs to provide its own.

[clauswilke]

I'd be happy to try to put together a PR. There's one important question, though: Is the intent that lines are straight in the data coordinate system, so they may become curved in the transformed space, or should they always be straight in the final plot?

[hadley]

I think they have to be straight in the data coordinate system, otherwise I can’t imagine how you’d interpret the parameters in (eg) polar coordinates.

[clauswilke]

Actually, I was too quick in blaming geom_abline() etc. The problem is with the coords, and specifically with the range() function. It looks like the code is inconsistent with respect to whether range() returns the range in the original data coordinates or in transformed coordinates. In several places in the code, it is assumed that range() returns original data coordinates. One is geom_abline() etc., and another location is coord_munch():

ggplot2/R/coord-munch.r

Lines 14 to 35 in 8778b48

	coord_munch <- function(coord, data, range, segment_length = 0.01) {
	if (coord$is_linear()) return(coord$transform(data, range))
	# range has theta and r values; get corresponding x and y values
	ranges <- coord$range(range)
	# Convert any infinite locations into max/min
	# Only need to work with x and y because for munching, those are the
	# only position aesthetics that are transformed
	data$x[data$x == -Inf] <- ranges$x[1]
	data$x[data$x == Inf] <- ranges$x[2]
	data$y[data$y == -Inf] <- ranges$y[1]
	data$y[data$y == Inf] <- ranges$y[2]
	# Calculate distances using coord distance metric
	dist <- coord$distance(data$x, data$y, range)
	dist[data$group[-1] != data$group[-nrow(data)]] <- NA
	# Munch and then transform result
	munched <- munch_data(data, dist, segment_length)
	coord$transform(munched, range)
	}

Thus, I can trigger a similar problem with geom_line():

library(ggplot2)
df <- data.frame(x = c(-Inf, Inf), y = c(1, 2))
p <- ggplot(df, aes(x, y)) + geom_line() + xlim(1, 2)
p

p + coord_flip()

p + coord_polar()

p + coord_trans(y = "log10")

p + coord_trans(x = "log10")
#> Warning in self$trans$x$transform(x): NaNs produced
#> Warning in trans$transform(value): NaNs produced
#> Warning: Transformation introduced infinite values in x-axis

Created on 2018-08-09 by the reprex package (v0.2.0).

Notice how coord_flip() and coord_polar() return the correct values but coord_trans() does not for x coordinates (it doesn't matter for y). One way to fix this would be to make sure that the range() function always returns the range in the original data coordinates. However, I'm not yet clear on whether there are other parts of the code that expect range() to return transformed coordinates. An alternative way to fix this would be to provide a second range function that always returns untransformed coordinates.

[lock]

This old issue has been automatically locked. If you believe you have found a related problem, please file a new issue (with reprex) and link to this issue. https://reprex.tidyverse.org/