skyandtelescope.com • SEPTEMBER 2019 21Although the images of M87* are derived from only four
days of observations, EHT scientists spent years testing and
installing equipment, working in the thin air of the remote
Chilean desert, braving the cold of Antarctica. They built
computer algorithms and developed simulations of what they
might see. They installed atomic clocks so precise that they’ll
lose only 1 second in 10 million years. They did dry runs, ago-
nizing over go/no-go weather conditions at eight telescopes
at six geographic sites scattered from Hawai‘i to Spain and
Arizona to the South Pole. “In VLBI, you really only get one
shot,” says Dan Marrone (University of Arizona), who has
fl own repeatedly to the South Pole to retrofi t the telescope
there. “Everything has to be working exactly right.”
Then, in April 2017, they went for it.
As Earth turned, each telescope set its sights on M87* and
the other targets, stockpiling data. By the end of the observing
run, they’d fi lled a half ton of hard drives with 5 petabytes of
data — equivalent to 5,000 years of MP3 fi les or, Marrone quips,
“the entire selfi e collection over a lifetime for 40,000 people.”The team then fl ew these hard drives to Massachusetts
and Germany, where the stations’ observations were fed into
supercomputers and aligned by their time stamps to within
trillionths of a second. “They have to be exactly right,” says
Johnson. “If they’re even a tiny bit off, you see nothing.”
Once the researchers had calibrated their data, a subset of
them (mostly young astronomers and computer scientists just
starting their careers) split into four teams spread around the
world. “We told them, ‘Don’t talk to each other or anyone
else,’” says Marrone. “‘Choose whichever imaging algorithms
you think are best, and make images of these data.’”
“We went into a room, there were six or seven of us there,”
says Johnson, “and we actually had the fi rst picture 30 min-
utes later.”
The challenge isn’t making one image, he explains, but
understanding its subtleties. The teams had to know all the
potential images their algorithms might create and where the
codes might lead them astray. Nor did they limit their codes
to reproducing shadows: To steal a comparison from com-RegroupImage
Reconstruction
Method #1Image
Reconstruction
Method #2Image
Reconstruction
Method #3The team compares real data with
simulated shadows, blurred to the
same resolution, to deduce where the
photon ring is in the fuzzy image.Teams use different software packages
to attack data systematically, producing
three independent results.Three images are
averaged to create
a fi nal consensus
image.Baseline lengthPowerpTHE MOMENT THEY SUSPECTED Maciek Wielgus (left) runs to director Shep Doeleman to show him an initial plot of the VLBI data. The data’s
bouncing pattern indicated the array had detected ring-like structure — even before the team reconstructed an image.RE
AC
TIO
N^ P
HO
TO
S:^
HS
IN-
YU
CH
EN
,^ G
RA
PH
:^ E
HT
CO
LL
AB
OR
AT
IO
N^ /
ASTR
OPHYSICAL
JOURNAL^ LETTE
RS875 :L3 (
20
19
)