The black hole photo was a reconstruction problem
The Event Horizon Telescope didn't take a picture in the ordinary sense. It filled in a mostly empty measurement with priors, and was careful about which ones.
On Wednesday the Event Horizon Telescope collaboration released the first image of a black hole, the one at the center of the galaxy M87. An orange ring of light around a dark center, 55 million light years away. I stared at it for a while and then went looking for how it was made, because I knew it couldn’t have been a normal photo.
To resolve something that small and far away, you’d need a telescope the size of the Earth. So they used eight radio telescopes spread around the planet, from Hawaii to Spain to the South Pole, observing at the same time and recording data with atomic clocks for timing. The technique is very long baseline interferometry. Each pair of telescopes measures one sample of the image’s spatial frequencies, roughly one point in its Fourier transform, and the Earth’s rotation sweeps those samples around over the night.
The key fact is that this gives you a tiny, patchy fraction of the measurements you’d need to reconstruct the image directly. Most of the Fourier plane is empty. Infinitely many images are consistent with the data. To get one image, you have to add assumptions about what images are likely, which is the same thing as a prior.
What impressed me most is how careful they were about that prior. They split into four teams that worked independently for weeks, without talking to each other, using different algorithms, some traditional and some newer ones like the CHIRP method Katie Bouman worked on, which learns what patches of plausible images look like. They also tested the pipelines on synthetic data, including images that weren’t rings, to make sure the methods wouldn’t produce a ring from anything. All four teams got a ring of about the same size. That’s the evidence that the ring comes from the data and not from the assumptions.
I work on a much more ordinary version of this problem every day. Every camera image is a reconstruction. The sensor has noise and gaps, and the phone’s processing fills them in with assumptions about what images usually look like. Our face models add even stronger assumptions about what faces look like. When a model “sees” something, the question is always how much comes from the pixels and how much from what it expects.
For the black hole, the team designed the whole process around that question. They were explicit about the priors and tried to break their own result. Most machine learning systems don’t do anything like that, and I think we should borrow the habit. If a result only shows up with one set of assumptions, it’s probably the assumptions talking.