The figure still has the numbers
A paper went up on arXiv two days ago with a dozen earthquakes plotted in it. Small ones. Two slabs of plexiglass pressed together at an angle until the seam between them ruptures, filmed by a camera fast enough to catch the rupture crossing the field of view, which it does in about a hundred microseconds.
The results are a scatter plot. Three panels, a few hundred coloured dots, and no data file anywhere. That is normal and nobody thinks twice about it. A plot is what a paper gives you and a plot is a picture.
It is not a picture. Open the PDF, pull out the page’s content stream, and what is written there is a list of instructions for a machine that draws: move to this coordinate, curve to that one, fill. Every dot in the figure is one of those instructions, and the coordinate in it is where the dot goes, which is where the measurement was.
So the numbers never left. I read all four hundred and eighty of them back out this morning. The error in the reading is half a micron, on an axis seven centimetres long.
What you get for that is the sentences underneath the picture, weighed.
The paper’s claim is that a rupture of this kind has one degree of freedom. Whatever size it starts at, whatever stress you load the seam with, it falls onto a single curve and then creeps along it, growing or shrinking depending on where it landed. The authors write that their pulses “approximately form a monotonic line”, which is what you say when the evidence is your own eyes on your own plot.
With the coordinates in hand you can say how approximately. Every experiment sits closer to the shared curve than its own points scatter about it. In all three panels the disagreement between separate ruptures is smaller than the wobble inside any one of them, which is a much stronger result than the sentence they wrote, and it was in the figure the whole time.
Elsewhere the reading cuts the other way. Two of the panels are described as “slightly shifted upwards” from the first. They are shifted by twenty-seven per cent and fifty-six per cent. The paper puts most of the gap between those two down to the sandblasting, so a smoother seam is worth about a quarter again on the slip rate at the same rupture size. That is a big effect wearing a small word.
And then there is the part I did not go looking for.
The second and third panels each carry a washed-out grey copy of the first panel’s data, laid underneath for comparison. Those grey dots are the same measurements a second time, so they are a free check on the reading: run them through the second panel’s axes and they should land on top of the first panel’s coloured dots. Vertically they do, to the last digit the file carries. Horizontally they are all slightly too far right, and by an amount that grows with distance across the plot.
The three panels are not the same width. They carry the same axis range, so a millimetre of rupture is worth slightly fewer millimetres of paper in the second panel than in the first, and every dot that appears twice should be drawn at two different places. The grey ones are drawn at one. Somebody copied the series across instead of plotting it again.
It costs nothing. The displacement is about a millimetre on the page at its worst, one dot’s width, in a set of dots that is there so your eye can see that the new ones sit higher. No conclusion in the paper leans on it. I would never have seen it by looking, and neither would anyone else, and neither did the four people who wrote the thing.
But it is a fingerprint. It is somebody at a screen at some hour, with the first panel finished and two more to go, doing the fast thing, which is what I would have done. The figure kept that. It kept the coordinate of every measurement to a thousandth of a millimetre and it also kept the shortcut, and the shortcut is the only part of the file that is about a person.
A photograph of a plot is a picture of data. A plot in a PDF is the data, wearing a picture, and it will answer questions the paper never asked it.
