No. 004 · Astrophysics

The galaxy that hides by getting bigger

Take a hundred thousand stars and spread them from a three-parsec radius to a thirty-parsec radius. You have not removed a single star. You have made the object a hundred times harder to see — and the version that vanishes is the version that is fifty times heavier.

In June, NASA released a Hubble portrait of NGC 6723, a globular cluster 27,000 light-years away in Sagittarius, packed densely enough with stars to have earned the nickname the Chandelier. The coverage led with a puzzle rather than the picture: something in it does not add up. Ultraviolet observations show the cluster built its stars in two distinct bursts, the second within 634 million years of the first. A globular cluster is supposed to be a single population — one cloud, one collapse, one generation.

Three months earlier, a paper in Nature had approached the same soft spot from the opposite direction. Ethan Taylor and colleagues ran cosmological simulations resolving structures down to about three parsecs and found that globular clusters and dark-matter-rich dwarf galaxies both emerge from standard cosmology without needing separate origin stories. The paper opens by conceding the obvious:

Globular clusters are among the oldest and densest stellar systems in the Universe, yet how they form remains a mystery.

Buried in that result is a prediction. Between the two familiar categories, the simulations produce a third kind of object that nobody has ever confirmed: a globular-cluster-like dwarf. Half-light radius between 10 and 60 parsecs. Metallicity below −2.75. And a dynamical-to-stellar mass ratio — call it Υ, total mass divided by the part of it that is stars — of roughly 50, against a globular cluster’s roughly 1.

What a photograph can and cannot weigh

Sort stellar systems on a plane with size along the bottom and brightness up the side, and the three categories overlap. Globular clusters run from about 1 to 10 parsecs. The predicted new class runs 10 to 60. Dwarf galaxies start around 40 and run to a thousand. In the strip from 40 to 60 parsecs, two different kinds of object sit on top of each other, and a photograph places you somewhere on that plane and then stops. It gives you a position and no way out of the overlap.

The thing that separates them is not in the picture at all. It is how fast the stars are moving. Feed a velocity dispersion and a half-light radius into a standard estimator — the mass inside the half-light radius is four σ²R divided by G, which with parsecs and kilometres per second collapses to a tidy 930σ²R in solar masses — and you get a weight. Compare that weight with the mass the light accounts for, and the leftover is dark matter. Nothing is detected. A discrepancy is computed.

That is worth sitting with for a moment. The most consequential substance in the universe enters this analysis as a subtraction.

Ten times the radius, five magnitudes

Now the part nobody bothers to compute, because it looks too simple to be interesting.

Build two systems out of exactly the same material: a hundred thousand solar masses of old, metal-poor stars. Same stars, same colours, same total light. Give one a half-light radius of 3 parsecs and the other 30 parsecs. Ten times the radius means a hundred times the area, and the same light spread across a hundred times the area is a hundred times fainter per square arcsecond. In magnitudes, where a factor f costs 2.5 log₁₀(f), that is exactly 5.00 magnitudes. Not approximately. Exactly, because 2.5 × log₁₀(100) = 2.5 × 2.

Run it through the full surface-brightness relation and the two objects come out at 18.79 and 23.79 magnitudes per square arcsecond. Five apart, as promised.

Then add the other half. At Υ ≈ 50, the spread-out system holds 2,500,000 solar masses inside its half-light radius where the compact one holds 50,000. Forty-nine solar masses of unseen mass for every solar mass of stars.

Fifty times the mass. One hundredth of the surface brightness. The heavier object is the one you cannot see, and the reason is not exotic. It is area.

Why the class is predicted and not found

This is not a curiosity. It is a selection effect with a number attached, and the number can be worked out in about four lines.

Take a wide-field survey that detects diffuse light down to 27.5 magnitudes per square arcsecond, hunting objects of absolute magnitude −3. Solve the brightness relation backwards for the radius rather than forwards for the brightness, and the largest half-light radius it can still register is 24.3 parsecs. The predicted class occupies 10 to 60. The survey reaches 28.6% of that band and is blind to the other 71%.

To see the whole range out to 60 parsecs you would need roughly two more magnitudes of surface-brightness depth — a factor of six in sensitivity, across the entire sky. That is not a software upgrade. That is a different telescope and a different decade.

The alternative is cheaper and stranger: stop taking pictures. For the two twins above, the measured velocity dispersions come out at 4.23 and 9.47 kilometres per second, a factor of 2.24 that any modern spectrograph resolves without complaint. Spectroscopy settles in one night what imaging cannot settle at any exposure length — but spectroscopy needs a target list, and only imaging can supply one. Neither instrument is sufficient alone, which is the honest version of every observing-time proposal ever written.

Which brings the argument back to where it started. The Hubble image that opened this story is a genuinely beautiful photograph, and it is the one instrument that structurally cannot answer the question it raises.

The 634-million-year problem

NGC 6723’s two bursts are 634 million years apart. The simulations say a globular cluster should show an age spread under 10 million years. That is a factor of 63.4 — and NASA is not being loose when it calls 634 million years “relatively brief” for an object older than ten billion. It is 4.6% of the age of the universe. On a calendar where the whole history of everything is one year, it is about seventeen days.

Both statements are true. The disagreement is not about the number; it is about what the number is being compared to, which is the sort of distinction that decides arguments in science far more often than new data does.

So something has to move. Either the criterion is wrong, because an age spread measured at formation is a strange test to apply to an object that has since spent thirteen billion years accreting and merging. Or the object is misfiled, and a thing with two populations was never one globular cluster. Or the measurement is softer than it sounds — 634 million years is inferred from ultraviolet colours, not read off a clock, and “within” makes it an upper bound with an error bar nobody quoted in the headline.

Working astronomers hold all three positions. None of them is available to someone looking at the picture.

Sources

E. D. Taylor, J. I. Read, M. D. A. Orkney et al., “The emergence of globular clusters and globular-cluster-like dwarfs,” Nature 645, 327–331 (2025), doi:10.1038/s41586-025-09494-x — source of the classification criteria, the Υ values, the 3-parsec simulation resolution and the candidate list. NASA, “Hubble Spies Starry Chandelier,” science.nasa.gov, June 2026, and Live Science’s write-up of it — source of NGC 6723’s distance and the 634-million-year figure. The mass estimator is the standard Wolf et al. form, MNRAS 406, 1220 (2010). Every figure quoted here was computed and checked in code before publication.