The diagram that sorted the stars, by plotting two things against each other

There is no reason a scatter plot should have structure. Take two properties of a few thousand objects, plot one against the other, and the usual result is a cloud.

Plot stellar luminosity against surface temperature and the cloud does not appear. Almost every star lands on a single narrow diagonal band, with a sparse group above it to the right and another far below to the left. Nothing about the measurements forces that; it is a fact about stars, and working out what it means occupied the twentieth century.

Two axes, one expensive #

The horizontal axis is temperature, and colour supplies it from two brightness measurements. It is cheap.

The vertical axis is luminosity — total output — and getting it requires an absolute magnitude, which requires a distance. It is expensive, and in 1910 it was nearly impossible: parallaxes existed for a few dozen stars.

Hertzsprung and Russell got round it in different ways, which is why both names are on the diagram. Hertzsprung worked with clusters, where all the stars are at the same unknown distance, so apparent magnitude serves as a stand-in for absolute and the diagram’s shape appears with the vertical axis uncalibrated. Russell used the few stars with measured parallaxes and got the absolute version for a smaller sample. Both found the same structure by 1913.

The temperature axis runs backwards — hot on the left — because Russell plotted spectral type in its traditional order and the order turned out to be a temperature sequence in reverse. It has never been fixed, and every HR diagram since carries the reversal.

The vertical axis inherits an inversion of its own from the same source. Plotted as absolute magnitude it also runs backwards, since brighter means a smaller number, so a colour–magnitude diagram is upside down and back to front relative to the physical quantities it stands for. Both reversals cancel in the sense that matters — luminous stars are still at the top and hot stars still at the left — which is why nobody has ever been sufficiently annoyed to change either.

What the band is #

Ninety percent of stars sit on the main sequence, and the temptation is to read a diagonal band as a track along which stars travel. They do not.

The main sequence is a mass sequence. A star fusing hydrogen in its core sits at a position fixed almost entirely by its mass: a 0.5 solar-mass star is cool and dim, the Sun sits in the middle, a 20 solar-mass star is hot and enormously bright. Stars do not slide along the band; they sit at their mass’s position for most of their existence and then leave it, sideways. The band is narrow because the mass–luminosity relation is tight. Two stars of the same mass have nearly the same luminosity and nearly the same temperature, whatever else is true of them, so the sequence has almost no width. Its residual thickness comes from composition and from age — a star brightens slowly as it converts hydrogen to helium, so the Sun is about 30% more luminous now than when it formed.

The band is also a lifetime statement. Stars are found overwhelmingly on the main sequence because that is where they spend nearly all their time, not because most stars are somehow main-sequence stars by type. The regions off the band are sparsely populated because passing through them is quick.

The diagonals that give sizes #

Luminosity, temperature and radius are related by the Stefan–Boltzmann law:

$$L = 4\pi R^2 \sigma T^4.$$

Two of the three are the axes, so the third is fixed at every point of the diagram. Lines of constant radius run diagonally, and the dashed lines in the figures are those lines, exact.

That turns the diagram into a size chart, and the sizes are extreme. Betelgeuse is cool — about 3,500 K, so each square metre of it radiates poorly — yet it is 100,000 times more luminous than the Sun. The only way to reconcile those is an enormous surface: about 900 solar radii, which would swallow Jupiter’s orbit.

White dwarfs run the other way. Sirius B is hotter than the Sun, so it radiates fiercely per unit area, and yet it is ten thousand times fainter. It must be tiny — about the size of the Earth, containing a solar mass. That inference, made from position on the diagram alone, was the first evidence for degenerate matter, and it was so unwelcome that Eddington remarked the star was sending a message that seemed absurd.

The islands #

Two regions off the main sequence, and both are stars that have finished with core hydrogen.

Giants and supergiants, upper right: cool and very bright, so very large. When a star exhausts the hydrogen in its core, the core contracts and heats while the envelope expands and cools, and the star moves right and up. The Sun will do this in about five billion years, expanding to engulf Mercury and Venus and rendering the Earth uninhabitable well before it arrives.

White dwarfs, lower left: hot and very faint, so very small. These are the exposed cores left when a low-mass star sheds its envelope, supported not by fusion — there is none — but by electron degeneracy pressure, a quantum effect with no classical counterpart. They do not generate energy; they cool, sliding slowly down and to the right over billions of years.

Both islands are sparsely populated for the same reason, and it is a selection effect rather than a fact about how many exist. Giants are brief; white dwarfs are faint. The most common star in the galaxy is a red dwarf at the bottom right of the main sequence, and not one of them is visible to the naked eye.

Reading a cluster’s age off it #

The diagram’s most powerful use is on a star cluster, where every star formed at about the same time from the same material and sits at the same distance.

Plot a young cluster and the main sequence is complete, from the hottest stars down. Plot an old one and the top is missing: the massive stars have already left. The point where the sequence bends away — the turn-off — is at the mass whose main-sequence lifetime equals the cluster’s age.

So the turn-off is a clock. Lifetime depends steeply on mass, so measuring where the sequence ends dates the cluster, and dating globular clusters this way gives ages around 12–13 billion years. That number was for a long time in tension with the estimated age of the universe — the clusters appeared older than everything — and resolving it required both better stellar models and a revised distance scale, which is a good example of how far an error at the bottom of the ladder propagates.

What was actually plotted #

The diagram in this essay is drawn from relations. The diagram astronomers work with is drawn from measurements, and the difference between the two is worth being explicit about, because almost every feature of the real one is absent from the idealised version.

What is actually plotted is a colour–magnitude diagram: apparent or absolute magnitude in one band against the difference between two bands. Neither axis is temperature or luminosity. Both are converted, if they are converted at all, by calibrations that carry their own assumptions — and a great deal of published work simply leaves the axes as colour and magnitude and never converts, because the conversion adds error without adding information.

For most of the twentieth century the sample was small and local. The transformation came in 2018, when Gaia’s second data release published parallaxes for 1.3 billion stars, and the colour–magnitude diagram could be drawn with tens of millions of stars whose distances were individually measured rather than assumed. The result is the single most information-dense plot in astronomy, and features that had been theoretical predictions or marginal detections became obvious structures with sharp edges.

Three of them are worth naming, because none appears in the schematic. There is a second sequence parallel to the main one and about 0.75 magnitudes above it, which is the locus of unresolved binaries — two identical stars looking like one twice as bright. There is a dense knot on the giant branch, the red clump, where core-helium-burning stars of a wide range of masses all sit at nearly the same luminosity, which makes it a standard candle nobody designed. And the white dwarf sequence, previously a smudge, split into distinct branches; the split is now attributed to the release of latent heat as the carbon–oxygen interior crystallises, a phase transition inferred entirely from a bump in the density of points on a scatter plot.

The general lesson is that the diagram’s value scales with the sample in a way most measurements do not. A hundred stars give the main sequence. A hundred thousand give the giant branch and the turn-off. A hundred million give crystallisation in white dwarf interiors. Nothing about the axes changed; the structure was always there, and the only missing ingredient was enough points to see it.

What each axis costs #

The two axes of the diagram have wildly different price tags, and that asymmetry shaped how the diagram was built.

The vertical axis needs an absolute magnitude, which needs a distance. The dependency runs both ways, which is the elegant part. A cluster’s main sequence can be fitted to the calibrated one, and the vertical offset required is the distance modulus. So the diagram consumes distances and produces them: main-sequence fitting is a rung of the ladder, sitting between parallax and the Cepheids, and it exists only because the diagram has the shape it does.

The trick, stated generally #

The Hertzsprung–Russell diagram is one instance of a manoeuvre that recurs throughout astronomy, and naming the manoeuvre explains why the subject has so many diagrams of this shape.

The manoeuvre is: plot two observables against each other for a large population, and let the structure reveal a hidden parameter that neither axis measures. Here the hidden parameter is mass, which cannot be observed for a single star at all and which nonetheless organises the entire plot. The band is narrow not because the measurements are precise but because both quantities are functions of one thing.

Once stated that way the family is obvious. Leavitt’s period–luminosity relation for Cepheids is the same move — two observables, a tight relation, and a hidden parameter, which for a Cepheid is again its mass and its evolutionary state. The Tully–Fisher relation plots a galaxy’s rotation speed against its luminosity and finds a line, the hidden parameter being total mass including the part that emits nothing. The fundamental plane of elliptical galaxies is the same thing in three dimensions.

The diagnostic that distinguishes a real instance from a spurious one is whether the relation is tighter than the measurement errors would allow if the points were independent. That is the signature of a hidden parameter, and it is why the width of a sequence is often more interesting than its slope. The main sequence’s residual thickness is composition, age and rotation; the Tully–Fisher relation’s scatter is the fraction of a galaxy’s mass that is dark. In both cases the physics being sought is in the part of the plot that refuses to collapse onto a line.

The diagram’s usefulness is sharpest when the same plot is made for two populations differing in one respect, and the standard comparison is between an open cluster and a globular one.

An open cluster — the Pleiades, say — is young, a few hundred stars, and roughly solar in composition. Its main sequence is complete from the top down; its brightest members are hot blue B stars; there are no giants at all, because nothing has had time to leave.

A globular cluster is old, a few hundred thousand stars, and metal-poor by one to two orders of magnitude. Its main sequence stops abruptly at about 0.85 solar masses, and above the turn-off the diagram has a well-populated giant branch, a horizontal branch and a scattering of blue stragglers. The whole upper main sequence is missing.

Everything different between those two diagrams is age and composition, and each leaves a distinct signature: age moves the turn-off vertically, composition shifts the entire sequence sideways. That the two effects are separable at all is what makes cluster fitting a quantitative method rather than an impression.

The version that needs no distance #

The vertical axis is the expensive one, and there is a variant of the diagram that dispenses with it entirely.

A spectrum gives a star’s effective temperature, from the ionisation balance of its lines, and its surface gravity, from the pressure broadening of the strong ones. Neither requires a distance: both are properties of the emergent spectrum, measurable on a star whose parallax is unknown.

Plotting surface gravity against temperature therefore produces a diagram with the same structure and no distance anywhere in it. The main sequence appears as a band, the giants sit above it — at low gravity, because a giant’s surface is far from its mass — and the white dwarfs sit far below at gravities a hundred thousand times the Sun’s.

The axes are not the physical ones, and the mapping between the two versions is exact. Surface gravity is mass over radius squared and luminosity is radius squared times temperature to the fourth, so at fixed mass the two diagrams are related by a change of variable — which is why they look alike.

The version’s value is that it can be drawn for stars too distant for a parallax, which before Gaia was most of them and after Gaia is still true for anything beyond a few kiloparsecs. Large spectroscopic surveys produce it for hundreds of thousands of stars, and it is how the Galaxy’s structure is mapped in stellar populations rather than in positions.

Its weakness is that a surface gravity from a spectrum is a fitted quantity carrying a model atmosphere’s assumptions, whereas a luminosity from a parallax is nearly geometric. The two diagrams are drawn from measurements of quite different character, and where both are available they are compared — which is one of the standard checks on spectroscopic parameters.

What both versions share is that the radius diagonals are an overlay rather than a measurement, and it is worth seeing the diagram without them once.

That is the diagram’s real trick and it is easy to miss. The two axes carry the data; the diagonals carry a law, and the law is what turns a scatter plot into a physical statement. A star’s position says its luminosity and its temperature, and only the third line through that position says its radius. Every subsequent use of the diagram — the giants as swollen stars, the white dwarfs as Earth-sized, the turnoff as an age — is read through a family of curves that no telescope produced.

It is also why the diagram survived the enormous growth in the number of stars plotted on it. Adding a hundred million points changes the density of the band and changes nothing about the diagonals, so the interpretive frame is fixed while the data behind it improves by orders of magnitude. A diagram whose meaning came from its points would have had to be re-read every time the sample grew; this one only got sharper. The same is true of the axes themselves: temperature and luminosity are both derived quantities, computed from a colour and a parallax through calibrations that have been revised many times, and every revision moves the points without moving the structure. A diagram whose content is a shape rather than a set of coordinates is remarkably hard to break, which is why this one has outlasted every theory of stellar structure proposed while it was being drawn. Hertzsprung and Russell were plotting a correlation between two quantities they could not explain, a decade before anyone knew what powered a star, and the diagram they made turned out to be the natural coordinate system for everything discovered afterwards.

Where the model stops #

It is a snapshot, not a track. The single most common misreading. Stars move on this diagram, but not along the main sequence, and the band is a locus of positions rather than a path.

Distances are needed. Every vertical position depends on a distance, and errors there move stars vertically in a way that mimics real differences.

Reddening moves stars sideways. Dust makes a star look cooler and fainter, sliding it down and to the right along a direction that is uncomfortably close to the main sequence’s own slope.

Binaries. An unresolved pair looks like one star of the combined luminosity, which lifts it above the sequence by up to 0.75 magnitudes and produces the second faint band visible in good cluster diagrams — the same pairs that, resolved and timed, supply every stellar mass there is.

Not all stars. Neutron stars and black holes are nowhere on it. Brown dwarfs fall off the bottom. Objects that never fused hydrogen have no business on a diagram organised around stars that do.

The figures also share a limitation of logarithmic axes: they compress the extraordinary. The vertical axis spans ten decades — a factor of $10^{10}$ between the faintest white dwarf and the brightest supergiant — and on a log plot that looks like a comfortable page. It is not comfortable. It means one object in the picture emits ten billion times more light than another, and the diagram’s readability is bought by making that unimaginable.

The ladder from here #

Later rungs: the colour–magnitude diagram, which is what is actually plotted. Cluster diagrams and main-sequence fitting as a distance method. The turn-off as a clock. Evolutionary tracks, and why they cross the diagram rather than run along it. The instability strip and the pulsating variables in it. The Chandrasekhar limit at the bottom left. The Hayashi track for stars still contracting. The initial mass function, which says how many stars start where. And the diagram for other galaxies, where the same structure appears with different chemistry.

Hertzsprung published in a photographic journal in 1911 and was largely unread. Russell presented the same result to the Royal Astronomical Society in 1913 and it became the organising diagram of a whole science within a decade.